Add files using upload-large-folder tool

preprint/preprint__08d5be710fda340f8176868396f7164bafc541ee9e5230b9c4d443f3c9838375/preprint__08d5be710fda340f8176868396f7164bafc541ee9e5230b9c4d443f3c9838375.mmd

@@ -0,0 +1,613 @@
+
+# 5'-tRNAGly(GCC) halves generated by IRE1α are linked to ER stress response  
+
+Ji-Hyun Yeom Chung- Ang University  
+
+Eunkyoung Shin Chung- Ang University  
+
+Yoonjie Ha Chung- Ang University https://orcid.org/0000- 0002- 6506- 9338  
+
+Minju Joo Chung- Ang University  
+
+Hanyong Jin Yanbian University  
+
+Haifeng Liu Chung- Ang University  
+
+Daeyoung Kim Chung- Ang University  
+
+Yong- Hak Kim Daegu Catholic University School of Medicine https://orcid.org/0000- 0001- 6192- 5996  
+
+Hak Kyun Kim Chung- Ang University  
+
+Jeongkyu Kim Chung- Ang University  
+
+Hong- Man Kim NES Biotechnology  
+
+Minkyung Ryu NES Biotechnology  
+
+Keun Pil Kim Chung- Ang University  
+
+Yoonsoo Hahn Chung- Ang University https://orcid.org/0000- 0003- 4273- 9842  
+
+Jeehyeon Bae School of Pharmacy, Chung- Ang University https://orcid.org/0000- 0003- 1995- 1378  
+
+Kangseok Lee ( kangseok@cau.ac.kr) Chung- Ang University https://orcid.org/0000- 0002- 0060- 6884
+
+<--- Page Split --->
+
+## Article  
+
+Keywords: IRE1α, alternative splicing, tRNAGly(GCC), ER stress, HNRNPM, HNRNPH2, tRNA halves  
+
+Posted Date: August 3rd, 2022  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 1464849/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License
+
+<--- Page Split --->
+
+1 5'-tRNAGly(GCC) halves generated by IRE1α are linked to ER stress response
+
+<--- Page Split --->
+
+## 4 Summary  
+
+Transfer RNA (tRNA) halves (tRHs) have various biological functions. However, the biogenesis of specific \(5^{\prime}\) - tRHs under certain conditions remains unknown. Here, we report that inositol- requiring enzyme \(1\alpha\) (IRE1α) cleaves the anticodon stem- loop region of tRNA \(^{\mathrm{Gly(GCC)}}\) to produce \(5^{\prime}\) - tRHs ( \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) ) with highly selective target discrimination upon endoplasmic reticulum (ER) stress. We observed IRE1α expression- dependent \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) production in human cancer cells. Levels of \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) were positively correlated with the degree of cancer cell proliferation both in vitro and in vivo; this effect required co- expression of two nuclear ribonucleoproteins, HNRNPM and HNRNPH2, which we identified as binding proteins of \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) . In addition, \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) modulated mRNA isoform biogenesis. Furthermore, under ER stress in vivo, we observed simultaneous induction of IRE1α and \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) expression in mouse organs and a distantly related organism, Cryptococcus neoformans. Thus, collectively, our findings indicate an evolutionarily conserved function for IRE1α- generated \(5^{\prime}\) - tRH- Gly \(^{\mathrm{GCC}}\) in cellular adaptation upon ER stress.  
+
+Key words: IRE1α, alternative splicing, tRNA \(^{\mathrm{Gly(GCC)}}\) , ER stress, HNRNPM, HNRNPH2, tRNA halves
+
+<--- Page Split --->
+
+## Introduction  
+
+Transfer RNA- derived fragments (tRFs) or transfer RNA- derived small RNAs (tsRNAs) have been recognised as functional small non- coding RNAs (ncRNAs) present in most organisms<sup>1</sup>. Multiple classes of tRFs have been identified in various cell types<sup>1</sup>. In particular, 31–40 nucleotide (nt) long tRFs generated by specific cleavage in the anticodon loop of mature tRNAs are referred to as tRNA halves (tRHs). Other tRFs are 14–40 nt in length and primarily correspond to the ends of mature tRNA (5'- tRFs and 3'- CCA tRFs) or pre- tRNA (3'- U tRFs)<sup>1</sup>.  
+
+In mammalian cells, limited information exists regarding the enzymes that generate tRFs. Angiogenin (ANG), a member of the RNase A superfamily, produces tRHs under certain stress conditions<sup>2- 5</sup>. In the case of RNase Z, it cleaves pre- tRNAs and generates 3'- U tRFs containing a stretch of U residues<sup>6</sup>. Additionally, dicer induces cleavage in the D loop and T loop of tRNAs, producing 5'- tRFs and 3'- CCA tRFs, respectively<sup>7, 8</sup>. Furthermore, recent deep sequencing data suggest that dicer processes tRFs in specific tRNAs and cell types<sup>9</sup>.  
+
+Functional roles of identified tRFs in biological processes include translational regulation of gene expression<sup>10- 13</sup>, gene silencing, and regulation of ribosome synthesis<sup>6, 14</sup>. tRHs affect cell proliferation<sup>4, 14- 18</sup>, apoptosis<sup>5</sup>, and epigenetic inheritance<sup>19, 20</sup>. For instance, changes in the profiles of a subset of sperm tRFs, including 5'- tRHs of tRNA<sup>Gly(GCC) (5'- tRH- Gly<sup>GCC</sup>), were reported in mice fed a high- fat diet<sup>20</sup>. Moreover, protein restriction in mice increases 5'- tRH- Gly<sup>GCC</sup> levels<sup>19</sup>. Additionally, 5'- tRH- Gly<sup>GCC</sup>, induced by alkB homologue 3, \(\alpha\) - ketoglutarate dependent dioxygenase (ALKBH3)—a tRNA demethylase—benefits the growth and progression of cervical carcinoma<sup>16</sup>. 5'- tRH- Gly<sup>GCC</sup> levels were also upregulated in papillary thyroid carcinoma<sup>18</sup>. Although 5'- tRH- Gly<sup>GCC</sup> appears to play various roles in cellular physiology, it remains unclear which enzyme generates these tRHs.  
+
+A structural analysis of IRE1α revealed that the catalytic residues between the tRNA
+
+<--- Page Split --->
+
+endonuclease and IRE1α contain functional groups with a shared chemical nature and spatial disposition<sup>21</sup>. IRE1α—a key regulator of signalling in the unfolded protein response (UPR)—is a conserved ER- localised transmembrane protein with ribonuclease activity<sup>22</sup>. Upon ER stress, IRE1α becomes activated and cleaves specific sites in the mRNA that encodes the transcription factor X- box- binding protein 1 (XBP1)<sup>23, 24</sup>. IRE1α also participates in regulated IRE1α- dependent decay, i.e., the degradation of multiple mRNAs and miRNAs under ER stress in an XBP1- independent manner<sup>25- 27</sup>. In particular, a consensus sequence (5'- CH<sub>(U or A or C)GCM<sub>(A or C)R<sub>(G or A)- 3')</sub>) accompanied by a stem- loop structure was proposed as an IRE1α cleavage site in mRNA<sup>28</sup>.  
+
+Herein, we observed that several tRNAs bear the consensus element for IRE1α cleavage in their anticodon loop region. Considering that tRNA<sup>Gly(GCC)</sup> is one such tRNA, we hypothesised that IRE1α may participate in producing 5'- tRHs from tRNA<sup>Gly(GCC)</sup>. To test the hypothesis, we aimed to investigate the direct involvement of IRE1α in the production of 5'- tRHs from tRNA<sup>Gly(GCC)</sup>, as well as their physiological function under ER stress.  
+
+## Results  
+
+## 5'-tRH accumulation by IRE1α upregulation  
+
+To explore whether IRE1α can cleave tRNAs and produce 5'- tRHs, we compared tRF profiles in human ovarian cancer- derived KGN cells endogenously expressing IRE1α (KGN) with those in the same cells exogenously overproducing IRE1α (KGN- IRE1α<sup>oe</sup>) using small RNA- sequencing (small RNA- seq). We selected human ovarian cancer cells, as 5'- tRH- Gly<sup>GCC</sup> reportedly functions in reproductive cells<sup>16, 20</sup>. The relative abundance of 5'- tRFs from tRNA<sup>Gly(GCC)</sup> species markedly increased when IRE1α was overexpressed (Fig. 1a and Supplementary Table 1). Additionally, 5'- tRFs from tRNA<sup>Cys(GCA)</sup> appeared to accumulate,
+
+<--- Page Split --->
+
+albeit at much lower levels compared to those from tRNA \(\mathrm{Gly(GCC)}\) in KGN- IRE1 \(\alpha^{\mathrm{oe}}\) cells (Fig. 1a and Supplementary Table 1). These results support the notion that among tRNA species, only tRNA \(\mathrm{Gly(GCC)}\) and tRNA \(\mathrm{Cys(GCA)}\) bear the consensus element for IRE1 \(\alpha\) cleavage in their anticodon stem- loop region.  
+
+Among three different tRNA \(\mathrm{Gly}\) isoacceptors, containing GCC, UCC, and CCC anticodons, 5'- tRFs with their 3'- end corresponding to position 33 of tRNA \(\mathrm{Gly(GCC)}\) were most abundant and enriched in KGN- IRE1 \(\alpha^{\mathrm{oe}}\) compared to KGN cells (Fig. 1a,b and Supplementary Table 1). In addition, high levels of 5'- tRFs, with their 3'- end corresponding to positions 31 and 32 of tRNA \(\mathrm{Gly(GCC)}\) , were observed when IRE1 \(\alpha\) was overexpressed (Fig. 1b and Supplementary Table 1). These 5'- tRFs from tRNA \(\mathrm{Gly(GCC)}\) occupied approximately \(89\%\) of the total 5'- tRFs from IRE1 \(\alpha^{\mathrm{oe}}\) cells (Fig. 1b), indicating that IRE1 \(\alpha\) overexpression primarily generates 5'- tRFs from tRNA \(\mathrm{Gly(GCC)}\) . In the case of tRNA \(\mathrm{Cys(GCA)}\) , high levels of 5'- tRFs, with their 3'- end corresponding to positions 33 and 34, were observed when IRE1 \(\alpha\) was overexpressed (Fig. 1a, b and Supplementary Table 1). We also observed enrichment of 5'- tRFs with their 3'- end corresponding to position 33 of tRNA \(\mathrm{Gly(GCC)}\) in IRE1 \(\alpha\) - overexpressing cells; however, these 5'- tRFs accounted for only \(\sim 2\%\) of the total (Fig. 1a, b and Supplementary Table 1).  
+
+To validate the small RNA- seq results, tRNA fragments were analysed via northern blotting with specific probes for the 5' upstream regions of the tRNA \(\mathrm{Gly(GCC)}\) , tRNA \(\mathrm{Cys(GCA)}\) , tRNA \(\mathrm{Gly(TCC)}\) , and tRNA \(\mathrm{Lys(CTT)}\) anticodon stem- loops. An IRE1 \(\alpha\) expression- dependent increase was observed in the levels of 5'- tRFs from tRNA \(\mathrm{Gly(GCC)}\) and tRNA \(\mathrm{Cys(GCA)}\) (Fig. 1c). The relative abundances of these 5'- tRFs were \(\sim 1.2\%\) and \(\sim 0.1\%\) of full- length tRNA \(\mathrm{Gly(GCC)}\) and tRNA \(\mathrm{Cys(GCA)}\) , respectively. The length of these 5'- tRFs corresponded to 32 and 33 nt- long synthetic RNAs containing sequences of 5'- tRFs from tRNA \(\mathrm{Gly(GCC)}\) (Fig. 1c). The relative abundances of other 5'- tRFs from tRNA \(\mathrm{Gly(TCC)}\) and tRNA \(\mathrm{Lys(CTT)}\) were not significantly
+
+<--- Page Split --->
+
+changed upon IRE1α overexpression, which agreed with the small RNA- seq results. Moreover, overexpression of a catalytically inactive form of IRE1α (K599A) \(^{29}\) did not significantly impact the levels of these 5'- tRFs (Extended Data Fig. 1a), indicating IRE1α cleavage activity- dependent production of 5'- tRH- Gly \(^{GCC}\) .  
+
+To assess the size of tRFs from tRNA \(^{\mathrm{Gly(GCC)}}\) , we performed primer extension analysis on the samples used for northern blot analysis. Primer extension targeting for tRNA \(^{\mathrm{Gly(GCC)}}\) produced one distinct cDNA band in reactions prepared with RNA samples from KGN- IRE1α \(^{\mathrm{oe}}\) cells. This cDNA band was synthesized from the 3'- tRH of tRNA \(^{\mathrm{Gly(GCC)}}\) , whose 5'- end corresponded to position 34 (Fig. 1d). This 3'- tRH of tRNA \(^{\mathrm{Gly(GCC)}}\) can be generated by IRE1α cleavage of tRNA \(^{\mathrm{Gly(GCC)}}\) between positions 33 and 34 within its anticodon stem- loop region (C \(^{31}\) UG \(^{1}\) CCAC \(^{37}\) ). This cleavage can also generate a 33- nt long 5'- tRH of tRNA \(^{\mathrm{Gly(GCC)}}\) , with the same 3'- end that mapped the highest in KGN- IRE1α \(^{\mathrm{oe}}\) cells in small RNA seq analysis (Fig. 1a, b). We were not able to detect cDNA bands corresponding 31 and 32 nt- long tRFs that were indicated in small RNA- seq and northern blot analyses (Fig. 1b,c). Overexpression of angiogenin (ANG)—a ribonuclease that produces tRHs by cleaving the anticodon loop region of tRNAs \(^{4, 5}\) — resulted in high levels of 5'- tRFs from all tRNA species (tRNA \(^{\mathrm{Gly(GCC)}}\) , tRNA \(^{\mathrm{Lys(CTT)}}\) , and tRNA \(^{\mathrm{Val(TAC)}}\) ; Extended Data Fig. 1b). Collectively, these results indicate that IRE1α activity is primarily responsible for generation of 5'- tRHs from tRNA \(^{\mathrm{Gly(GCC)}}\) .  
+
+## Selective cleavage of tRNA \(^{\mathrm{Gly(GCC)}}\) by IRE1α  
+
+To assess whether IRE1α is solely responsible for the production of 5'- tRHs from tRNA \(^{\mathrm{Gly(GCC)}}\) , we isolated tRNAs from KGN cell total RNA and isolated tRNAs by size fractionation. Purified tRNAs were then incubated with human IRE1α. IRE1α was found to selectively cleave tRNA \(^{\mathrm{Gly(GCC)}}\) in vitro (Fig. 2a). Specifically, IRE1α- mediated cleavage of tRNA \(^{\mathrm{Gly(GCC)}}\)
+
+<--- Page Split --->
+
+generated one major and two minor \(5^{\prime}\) - tRNAGly(GCC) fragments (Fig. 2a). When the 3'- end of these fragments was mapped by primer extension analysis, seven distinct cDNA bands (Fig. 2b) were detected. The cDNA band was most prominent and could be synthesised from the 3'- tRH of tRNAGly(GCC) with a 5'- end corresponding to position 34 (Fig. 2b, labelled as b). This 3'- tRH could be generated by IRE1α cleavage of tRNAGly(GCC) between positions 33 and 34 within its anticodon stem- loop region (C \(^{31}\) UG \(_{1}\) CCAC \(^{37}\) ). This cleavage site corresponded to the 3'- end of the most abundant 5'- tRFs from tRNAGly(GCC) that were identified by small RNA- seq (Fig. 1b). Moreover, this cDNA band is identical to the distinct cDNA detected in the primer extension assay of tRNAGly(GCC) fragments in KGN cells following IRE1α overexpression (Fig. 1d). Another distinct band (labelled as a) also corresponded to the 3'- end of the second most abundant 5'- tRFs from tRNAGly(GCC) identified in small RNA- seq (Fig. 1b, 2b). Meanwhile, the cleavage sites deduced from other cDNA bands were not observed in the small RNA- seq analysis (Fig. 1b) or in the primer extension of in vivo generated fragments of tRNAGly(GCC) (Fig. 1d).  
+
+To further biochemically verify the ability of IRE1α to cleave tRNAGly(GCC), we conducted an in vitro IRE1α cleavage reaction using purified tRNAGly(GCC) as a substrate (Extended Data Fig. 2a–c). Two major and five minor cleavage products appeared to be dependent on IRE1α (Fig. 2c). Among them, one major cleavage product (labelled as 5) corresponded to an IRE1α cleavage product of tRNAGly(GCC) between positions 33 and 34 (Fig. 2c). This cleavage site corresponded to the 3'- end of the most abundant 5'- tRFs from tRNAGly(GCC) identified via small RNA-seq (Fig. 1b) and was the only cDNA detected in the primer extension assay of tRNAGly(GCC) in vivo generated fragments (Fig. 1d) when IRE1α was overexpressed. Five other products also corresponded to IRE1α cleavage products of tRNAGly(GCC) at sites generated by in vitro IRE1α cleavage of total tRNAs (Fig. 2b). An
+
+<--- Page Split --->
+
+additional cleavage product (labelled as 7) was also detected in small RNA- seq (Fig. 1b). However, several tRFs identified from in vitro cleavage of tRNAGly(GCC) were not detected in the small RNA- seq analyses (Fig. 1a,b) or primer extension assay of tRNAGly(GCC) fragments generated in vivo (Fig. 1d). These tRFs might have resulted from decreased IRE1α stringency in the sequence- specific cleavage of tRNAGly(GCC) in vitro, or from tRNAGly(GCC) structural alterations induced during purification or incubation. It is also possible that they arose from fragmentation of tRNAGly(GCC) cleavage products.  
+
+These clearly show the ability of IRE1α to selectively cleave tRNAGly(GCC) within the anticodon stem- loop region. Furthermore, the cleavage site (C31UG\CCAC37) deduced from both small RNA- seq (Fig. 1a,b) and primer extension analyses of in vivo generated tRNAGly(GCC) fragments (Fig. 1d) corresponded with the major in vitro IRE1α cleavage product of tRNAGly(GCC) (Fig. 2b,c). Based on these results, we designated 33- nt long 5'- tRHs generated from the cleavage of tRNAGly(GCC) at the overlapping site (C31UG\CCAC37) as 5'- tRH- GlyGCC.  
+
+## Induction of 5'- tRH-GlyGCC generation upon ER stress  
+
+Considering that IRE1α is an ER stress- activated endonuclease, we hypothesised that ER stress- induced activation of IRE1α may cause generation of 5'- tRH- GlyGCC from tRNAGly(GCC) cleavage. To test this hypothesis, we induced ER stress in KGN cells using thapsigargin (TG) or tunicamycin (TM). Western blot analysis and an XBP1 splicing assay confirmed that these agents stimulated the expression and ribonucleolytic activity of IRE1α (Fig. 3a).  
+
+Next, ER stress- induced IRE1α activation on tRNA cleavage was examined via northern blot analysis on tRNAs. In agreement with the effect of IRE1α overexpression on the generation of tRHs from tRNAGly(GCC) (Fig. 1c), 5'- tRH- GlyGCC was generated while distinct tRFs from other tRNAs were not detected following ER stress- induced IRE1α expression (Fig.
+
+<--- Page Split --->
+
+3a and Extended Data Fig. 3a). Moreover, we found that the cleavage pattern of these 5'- tRHs resembled those generated by IRE1α, which cleaves tRNAGly(GCC) between positions 33 and 34 within the anticodon stem- loop (C<sup>31</sup>UG↓CCAC<sup>37</sup>) (Fig. 3b). Furthermore, TG- or TM- induced production of 5'- tRHs was not observed in IRE1α knock out cells (IRE1α<sup>-/-</sup>; Fig. 3c and Extended Data Fig. 3b, 4). Hence, ER stress induces 5'- tRH- Gly<sup>GCC</sup> production via IRE1α- dependent tRNAGly(GCC) cleavage in KGN cells.  
+
+To investigate whether the production of 5'- tRHs from tRNAGly(GCC) is commonly coupled with ER stress in human cells, we induced ER stress in HeLa cells with TG or TM. ER stress- dependent selective generation of 5'- tRFs from tRNAGly(GCC) was also observed in HeLa cells (Extended Data Fig. 3c, d).  
+
+## Proteins interacting with 5'- tRH-Gly<sup>GCC</sup>  
+
+To investigate the functional role of 5'- tRH- Gly<sup>GCC</sup>, we characterised proteins bound to 5'- tRH- Gly<sup>GCC</sup> in KGN cells via biotinylation of the tRH 5'- and 3'- ends. Specifically, 33- nt long 5'- tRHs of tRNAGly(GCC) (5'- tRH- Gly<sup>GCC</sup> mimic) were used. To assess non- specific protein binding of biotinylated RNA with a streptavidin coated microplate, 5'- biotin- oligo A8 RNA and 3'- biotin- tRH- Gly<sup>GCC</sup> were used as controls. Two protein bands near 70 kDa and 55 kDa appeared to specifically bind to a 5'- biotin- tRH- Gly<sup>GCC</sup> in both samples of TG- treated and - untreated cells but did not bind 3'- biotin- tRH- Gly<sup>GCC</sup> or 5'- biotin- oligo A8 RNA (Extended Data Fig. 5a, right panel). The comparative tandem mass spectrometry analysis of proteins interacting with biotinylated RNA showed that the heterogeneous nuclear ribonucleoprotein M isoform b (HNRNPM) and heterogeneous nuclear ribonucleoprotein H isoform X11 (HNRNPH2) in TG- treated and - untreated samples were enriched at similar levels in the microplate containing 5'- biotin- tRH- Gly<sup>GCC</sup> (Extended Data Fig. 5b and Supplementary Table 2). These nuclear proteins
+
+<--- Page Split --->
+
+may be potential binding partners of the 5'- tRHs of tRNAGly(GCC). We were also able to detect a moderate amount of HNRNPF, while HNRNPH1 was not detected. Orthologues of HNRNP proteins have been previously identified as binding proteins of tRHGly(GCC) in mouse cells<sup>30</sup>.  
+
+We further assessed the physical interaction between 5'- tRH- Gly<sup>GCC</sup> and HNRNP proteins via electrophoretic mobility shift assay (EMSA) using purified HNRNPM and HNRNPH2 recombinant proteins and 5'- P<sup>32</sup>- labelled synthetic 5'- tRH- Gly<sup>GCC</sup> and 5'- tRH- Lys<sup>CTT</sup>. These proteins bind 5'- tRH- Gly<sup>GCC</sup> with much higher affinity than 5'- tRH- Lys<sup>CTT</sup> (Extended Data Fig. 6a), providing evidence of specific interactions between 5'- tRH- Gly<sup>GCC</sup> and HNRNP proteins.  
+
+We further tested physical interaction of 5'- tRH- Gly<sup>GCC</sup> or 3'- tRH- Gly<sup>GCC</sup> with HNRNP proteins (HNRNPM and HNRNPH2) by using the surface plasmon resonance (SPR). In SPR assay, both 5'- tRH- Gly<sup>GCC</sup> and 3'- tRH- Gly<sup>GCC</sup> showed dose- dependent binding signal to the immobilized HNRNPM and HNRNPH2 (Extended Data Fig. 6b). However, kinetic analysis indicated that 5'- tRH- Gly<sup>GCC</sup> has about \(10 \sim 35\) times higher affinities for HNRNPM and HNRNPH2 with \(K_D\) of \(86.30 \mathrm{nM}\) and \(27.07 \mathrm{nM}\) , compared to 3'- tRH- Gly<sup>GCC</sup>, which showed affinities for HNRNPM and HNRNPH2 with \(K_D\) of \(0.96 \mu \mathrm{M}\) and \(0.94 \mu \mathrm{M}\) , respectively (Extended Data Fig. 6c). These results provide clear evidence for specific and strong interaction between 5'- tRH- Gly<sup>GCC</sup> and HNRNP proteins.  
+
+## Roles of ER stress-induced 5'- tRH-Gly<sup>GCC</sup>  
+
+To investigate functional roles for 5'- tRH- Gly<sup>GCC</sup> in cancer cells that produced these 5'- tRHs upon ER stress, KGN and HeLa cells were treated with synthetic 5'- tRH- Gly<sup>GCC</sup> (5'- tRH- Gly<sup>GCC</sup> mimic) and two other control tRH mimics (5'- tRH- Lys<sup>CTT</sup> and 3'- tRH- Gly<sup>GCC</sup>). Treatment with 5'- tRH- Gly<sup>GCC</sup> mimic promoted cell survival in a manner dependent on mimic
+
+<--- Page Split --->
+
+concentrations, which increased cell survival of KGN and HeLa cells by \(34\%\) and \(25\%\) , respectively, at the highest concentration of the mimic used (Fig. 4a,b). In contrast, the two control mimics did not significantly affect cell viability with the highest concentration of \(3'\) - tRH- Gly \(^{\mathrm{GCC}}\) inducing only a moderate increase in the viability of KGN cells (Fig. 4a,b). Blocking the \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) mimic with complementary antisense DNA oligos significantly reduced the positive effects on cell viability (Extended Data Fig. 7a). The enhancement of cell viability by \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) mimic occurred due to increased proliferation of KGN cells, as no effect on apoptosis was observed (Extended Data Fig. 7b,c). In addition, tRH mimic transfection did not affect the migration capability of KGN cells (Extended Data Fig. 7d). These results suggested that \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) functions to control cancer cell proliferation.  
+
+Next, we investigated whether HNRNPM and HNRNPH2 participate in \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) - mediated cell survival. The \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) mimic- induced promotion of cell survival in KGN and HeLa cells was abolished following HNRNPM or HNRNPH2 knockdown (Fig. 4c,d). Transfection of the \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) mimics in HNRNPM- depleted KGN (Fig. 4c) or HNRNPH2- depleted HeLa (Fig. 4d) cells further reduced cell survival compared to those treated with control mimics, suggesting that HNRNPM and HNRNPH2 might have a cell line- specific role in \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) - mediated cell survival. Hence, \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) - mediated cell survival depends on its interaction with HNRNP proteins in these cancer cells. Considering that we observed analogous results following knockdown of HNRNPF/H1/H2 (Extended Data Fig. 7e,f), the \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) might interact with multiple nuclear ribonucleoprotein for cellular function.  
+
+We further examined whether IRE1α- dependent \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) generation mediates the ER stress- induced effect on cell survival. Treatment with antisense DNA oligos targeting endogenous \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) (anti- \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) ) significantly promoted TG- induced cell
+
+<--- Page Split --->
+
+death in WT KGN cells compared to those treated with 5'- tRH- Lys<sup>CTT</sup> (anti- 5'- tRH- Lys<sup>CTT</sup>) (Fig. 4e). In contrast, anti- 5'- tRH- Gly<sup>GCC</sup> did not elicit such an effect in IRE1α<sup>- /- </sup> cells (Fig. 4e). Thus, IRE1α cleavage- generated 5'- tRH- Gly<sup>GCC</sup> contributes to cellular adaptation upon ER stress.  
+
+To investigate in vivo 5'- tRH- Gly<sup>GCC</sup> function, we silenced the endogenous 5'- tRH- Gly<sup>GCC</sup> or 5'- tRH- Lys<sup>CTT</sup> by delivering antisense DNA oligos against them using a functionalized gold nanoparticle (AuNP)- based delivery system (AuNP<sup>dT</sup>) in a xenograft mouse model. As shown in Figure. 4f, tumour growth in mice treated with AuNP<sup>dT</sup> loaded with anti- 5'- tRH- Gly<sup>GCC</sup> was prominently inhibited compared with that treated with AuNP<sup>dT</sup> alone or AuNP<sup>dT</sup> loaded with anti- 5'- tRH- Lys<sup>CTT</sup> (Fig. 4f). Consistent with anti- proliferative response observed in cancer cells treated with anti- 5'- tRH- Gly<sup>GCC</sup> in vitro (Extended Data Fig. 7g), proliferating cell nuclear antigen (PCNA) expression in tumours decreased by \(\sim 44\%\) upon anti- 5'- tRH- Gly<sup>GCC</sup> treatment in xenografted tumours (Fig. 4g).  
+
+## Effect of 5'- tRH-Gly<sup>GCC</sup> in alternative splicing  
+
+To dissect the relevance of 5'- tRH- Gly<sup>GCC</sup> functioning, we performed total transcriptome analysis on KGN cells transfected with 5'- tRH- mimics. The RNA abundance of 66 genes was altered more than 1.5- fold in cells transfected with 5'- tRH- Gly<sup>GCC</sup> mimics compared to those with 5'- tRH- Lys<sup>CTT</sup> mimics (Fig. 5a and Supplementary Table 3). Functional annotation analysis further indicated that most genes were enriched in alternative splicing and phosphoproteins (Fig. 5b and Supplementary Table 4). Based on these results, and the fact that 5'- tRH- Gly<sup>GCC</sup> interacts with multiple nuclear proteins functioning in RNA splicing, we hypothesised that 5'- tRH- Gly<sup>GCC</sup> modulates alternative splicing of a target gene subset. For this
+
+<--- Page Split --->
+
+reason, we further analysed isoforms of total transcripts using nanopore sequencing and FLAIR (full- length alternative isoform analysis of RNA) modules<sup>31</sup>.  
+
+We analysed four main types of alternative splicing events (alternative 3'- and 5'- splicing, intron retention, and exon skipping events) associated with isoform formation. Compared to the control group (5'- tRH- Lys<sup>CTT</sup>), we identified 19 differential isoforms from the 17 genes in the 5'- tRH- Gly<sup>GCC</sup>- treated group, where one or more of their junctions exhibited alternative 5'/3' splice site selection or exon skipping (Supplementary Table 5). These genes had multiple alternative splicing events within their transcripts, except CFDP1 (exon skipping), PRDX4 (alternative 5'- splicing), and MAGED2 (alternative 3'- splicing) (Supplementary Table 5). Among them, the isoform usage of ELOB (Fig. 5c, upper panel) and PMSB5 (Fig. 5c, lower panel) was significantly altered between 5'- tRH- Lys<sup>CTT</sup>- and 5'- tRH- Gly<sup>GCC</sup>- treated groups. These results were confirmed by RT- qPCR using isoform specific primers (Fig. 5d). In addition, sequestering of 5'- tRH- Gly<sup>GCC</sup> by antisense DNA oligos in xenografted tumours resulted in a shift in mRNA isoform composition in an opposite direction (Fig. 5e) compared to what we observed with 5'- tRH- Gly<sup>GCC</sup> mimics treatment in cancer cells (Fig. 5d). Treatment of tumours with anti 5'- tRH- Lys<sup>CTT</sup> did not affect isoform composition of these genes (Fig. 5e). Hence, 5'- tRH- Gly<sup>GCC</sup> levels affect alternative splicing events, leading to alterations in the transcript isoform profile.  
+
+## Nucleus localisation of tRH-Gly<sup>GCC</sup>  
+
+Our results showing an interaction between tRH- Gly<sup>GCC</sup> and nuclear proteins (Extended Data Fig. 6), as well as the effect of tRH- Gly<sup>GCC</sup> mimics on transcript isoform profiles (Fig. 5), suggest that tRH- Gly<sup>GCC</sup> functions within the nucleus. Thus, to determine the subcellular distribution of ER stress- induced 5'- tRHs of tRNA<sup>Gly(GCC)</sup>, we conducted a fluorescent in situ
+
+<--- Page Split --->
+
+hybridisation assay (FISH), with a probe designed to recognise \(5'\) - tRHs of tRNAGly(GCC). This assay was performed under conditions designed to avoid the denaturation of stable mature tRNAs and hybridisation of the probe to full- length tRNAs. Fluorescent signals, obtained with the probe recognising the \(5'\) - tRHs of tRNAGly(GCC), displayed a nucleus- associated localisation pattern with higher signal intensity following TG- induced ER stress, compared to treatment with dimethylsulphoxide (DMSO; Extended Data Fig. 8a). To confirm the specificity of the hybridisation probe, we performed a series of experiments under non- denaturing or denaturing conditions using an additional control probe. This control probe (anticodon bridging probe) bridged the \(5'\) - and \(3'\) - regions spanning the nucleotides that encompass the anticodon and was designed to detect only intact full- length tRNAs with minimal complementarity for \(5'\) - tRHs of tRNAGly(GCC). The anticodon bridging probe showed a fluorescent signal under denaturing FISH conditions, while no signal was observed under non- denaturing conditions (Extended Data Fig. 8a). Hence, \(5'\) - tRHs of tRNAGly(GCC) were definitively recognised with the specific probe used under our experimental conditions. In addition, measurement of tRH- GlyGCC distribution by TaqMan assay showed that \(5'\) - tRH- GlyGCC levels were elevated by \(\sim 25\%\) in the nuclear fraction of KGN cells following TG treatment over the 6 h period (Extended Data Fig. 8b). These results indicate that \(5'\) - tRHs of tRNAGly(GCC) localise to the nucleus when cells are subjected to ER stress.  
+
+## IRE1α-dependent \(5'\) -tRHGhy(GCC) cleavage in other organisms  
+
+To investigate whether IRE1α- mediated generation of \(5'\) - tRH- GlyGCC upon ER stress occurs in other eukaryotic species, we analysed selective generation of \(5'\) - tRH- GlyGCC in an acute ER stress murine model32 and ER- stressed yeast species, Cryptococcus neoformans. IRE1
+
+<--- Page Split --->
+
+homologues of mouse and yeast show \(94.40\%\) and \(40.15\%\) sequence similarity, respectively to the protein kinase and kinase- extension nuclease domains of human IRE1α.  
+
+We observed prominent induction of IRE1α expression in the ovary, liver, epididymis, kidney, and pancreas of ER- stressed mice (Extended Data Fig. 9a). Northern blot analysis showed an increased abundance of \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) fragments in the ovary (Fig. 6a), liver (Extended Data Fig. 9b), and epididymis (Extended Data Fig. 9c) in samples taken from ER- stressed mice compared to control mice samples, while other tRNA fragments, the size of which were similar to \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) , were not detected (Extended Data Fig. 9b, d). Moreover, primer extension analysis indicates that \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) fragments in the ovary resulted from the overlapping IRE1α cleavage site identified in KGN cells, which generates a 33- nt long \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) (Fig. 6b). These results indicate that ER stress induces selective generation of \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) in mice in an IRE1α expression- dependent manner.  
+
+We also observed that high levels of \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) coincided with enhanced IRE1 expression in \(C\) . neoformans when ER stress was induced by TM treatment (Fig. 6c and Extended Data Fig. 9e). Primer extension analysis indicate that these \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) resulted from Ire1 cleavage at the site corresponding to the overlapping site identified in KGN cells and mouse ovary (Fig. 6d). A minor band, which was slightly longer than \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) , was detected in the ire1- deletion strain when treated with TM (Fig. 6c and Extended Data Fig. 9e). These results indicate the existence of an additional unknown activity for tRNA \(^{\mathrm{Gly(GCC)}}\) cleavage under ER stress in this yeast species. Taken together, IRE1α- dependent selective generation of \(5'\) - tRH- Gly \(^{\mathrm{GCC}}\) under ER stress appears to be widely conserved in eukaryotic organisms.  
+
+## Discussion
+
+<--- Page Split --->
+
+This study highlights that IRE1α- mediated selective cleavage of tRNAGly(GCC) and 5'- tRH generation upon ER stress is conserved in human, mice, and a distantly related yeast species, C. neoformans. In fact, within these organisms, IRE1α homologues selectively cleave tRNAGly(GCC) species at the same site. These results raise the question of why these organisms evolutionarily retain this biological event in response to ER stress. Perhaps, as we found in the case of human cancer cells (Fig. 4e), 5'- tRH- Gly(GCC) contributes to cellular adaptation upon ER stress.  
+
+Dicer and ANG generate different types of tRFs for multiple roles in cellular processes<sup>33</sup>. However, ANG is the only identified enzyme associated with the generation of 5'- tRHs by cleaving the anticodon stem- loop of mature tRNAs in mammalian cells<sup>2-5, 16, 34</sup>. Additional enzymes responsible for the generation of certain 5'- tRHs have not been identified under specific conditions, such as metabolic diseases, cancer, and reproductive cell maturation<sup>18-20</sup>. One such example is 5'- tRH from tRNAGly(GCC) produced in mouse sperm, which reportedly suppresses the expression of genes associated with endogenous retroelement MERVL in embryonic stem cells and embryos by regulating gene expression from specific regions of the genome<sup>19, 20</sup>. This 5'- tRH from tRNAGly(GCC) was shown to be upregulated in papillary thyroid carcinoma<sup>18</sup>. Although an increasing number of reports have revealed that tRNA- derived fragments are involved in various biological processes, its biogenesis is remains largely unknown. Here, we show that under ER stress conditions, IRE1α cleaves the anticodon stem- loop of tRNAGly(GCC) to produce 5'- tRH in human cancer cells. Moreover, generation of 5'- tRHs from tRNAGly(GCC) appears to be ER stress- specific, as it was not observed under other stress conditions tested in this study (Extended Data Fig. 10a,b).  
+
+Colicins and ANG generate tRHs by cleaving target tRNAs and the anticodon loop of most tRNAs, respectively, thereby inhibiting protein synthesis<sup>35, 36</sup>. In the case of IRE1α-
+
+<--- Page Split --->
+
+mediated generation of tRHs from tRNAGly(GCC), it is unlikely that 5'- tRH- GlyGCC affects protein synthesis efficiency, as IRE1α appears to cleave a small portion of tRNAGly(GCC) upon ER stress, and thus, does not significantly reduce the pool of mature tRNAGly(GCC) (Fig 1c and 3a). Consistent with this notion, overexpression of IRE1α did not affect expression levels of two glycine-rich proteins, which contain a high proportion of the GGC codon in their mRNA (Extended Data Fig. 10c, Supplementary Table 6).  
+
+Although the detailed modes of action for most tRFs and tRHs remain unclear, several studies indicate that they can regulate the expression and translational efficiency of endogenous target genes by interacting with binding partners, including cytochrome c, YBX1, PIWI, and the AGO family \(^{5, 10, 37 - 40}\) . In the case of 5'- tRH- GlyGCC, we found that they interact with two nuclear proteins, HNRNPM and HNRNPH2, and these interactions are required for 5'- tRH- GlyGCC to influence cancer cell survival (Fig. 4c,d). Knockdown of HNRNPF or HNRNPH1, other interacting proteins identified by Boskovic et al \(^{30}\) , showed similar effects on cancer cell survival, suggesting that 5'- tRH- GlyGCC interacts with multiple nuclear proteins to exert these effects. A recent report identified RBM17, a splicing- related RNA- binding protein, as a binding protein of 5'- tRH- GlyGCC \(^{18}\) , supporting our notion that 5'- tRH- GlyGCC functions in the nucleus.  
+
+Additionally, treatment with 5'- tRH- GlyGCC mimics altered the expression of genes primarily associated with alternative splicing, which overlaps with the function of HNRNPP proteins and RBM17 identified as 5'- tRH- GlyGCC binding proteins. Moreover, analyses of isoforms within the total transcriptome data using nanopore sequencing, together with the alternative splicing assay results, indicate that 5'- tRH- GlyGCC affects the profiles of a subset of transcript isoforms (Fig. 5). Consistent with these nuclear phenomena, we observed nuclear localisation of 5'- tRH- GlyGCC following ER stress. However, further studies are needed to elucidate the detailed mechanisms underlying the role of 5'- tRH- GlyGCC in alternative splicing
+
+<--- Page Split --->
+
+events (Fig. 6e).  
+
+Aberrant expression of tRNA fragments is reported in various human disease conditions, providing potential targets for disease detection and therapeutics. Our data showed that antisense DNA oligos- mediated \(5^{\prime}\) - tRH- Gly \(^{\text{GCC}}\) suppression inhibited tumour proliferation in a xenograft mouse model (Fig. 4f,g). Thus, we believe that understanding the regulatory role of \(5^{\prime}\) - tRH- Gly \(^{\text{GCC}}\) can be used as a novel biomarker and potential therapeutic target in cancer cells.  
+
+## References and Notes  
+
+1. Kim, H.K., Yeom, J.H. & Kay, M.A. Transfer RNA-Derived Small RNAs: Another Layer of Gene Regulation and Novel Targets for Disease Therapeutics. Mol Ther 28, 2340-2357 (2020).  
+2. Fu, H. et al. Stress induces tRNA cleavage by angiogenin in mammalian cells. FEBS letters 583, 437-442 (2009).  
+3. Yamasaki, S., Ivanov, P., Hu, G.F. & Anderson, P. Angiogenin cleaves tRNA and promotes stress-induced translational repression. The Journal of cell biology 185, 35-42 (2009).  
+4. Honda, S. et al. Sex hormone-dependent tRNA halves enhance cell proliferation in breast and prostate cancers. Proceedings of the National Academy of Sciences of the United States of America 112, E3816-3825 (2015).  
+5. Saikia, M. et al. Angiogenin-cleaved tRNA halves interact with cytochrome c, protecting cells from apoptosis during osmotic stress. Molecular and cellular biology 34, 2450-2463 (2014).  
+6. Haussecker, D. et al. Human tRNA-derived small RNAs in the global regulation of RNA silencing. RNA (New York, N.Y.) 16, 673-695 (2010).  
+7. Cole, C. et al. Filtering of deep sequencing data reveals the existence of abundant Dicer-dependent small RNAs derived from tRNAs. RNA (New York, N.Y.) 15, 2147-2160 (2009).  
+8. Maute, R.L. et al. tRNA-derived microRNA modulates proliferation and the DNA damage response and is down-regulated in B cell lymphoma. Proceedings of the National Academy of Sciences of the United States of America 110, 1404-1409 (2013).  
+9. Li, Z. et al. Extensive terminal and asymmetric processing of small RNAs from rRNAs, snoRNAs, snRNAs, and tRNAs. Nucleic acids research 40, 6787-6799 (2012).  
+10. Ivanov, P., Emara, M.M., Villen, J., Gygi, S.P. & Anderson, P. Angiogenin-induced tRNA fragments inhibit translation initiation. Molecular cell 43, 613-623 (2011).  
+11. Gebetsberger, J., Zywicki, M., Künzi, A. & Polacek, N. tRNA-derived fragments target the ribosome and function as regulatory non-coding RNA in Haloferax volcanii. Archaea (Vancouver, B.C.) 2012, 260909 (2012).
+
+<--- Page Split --->
+
+12. Luo, S. et al. Drosophila tsRNAs preferentially suppress general translation machinery via antisense pairing and participate in cellular starvation response. Nucleic acids research 46, 5250-5268 (2018).
+13. Sobala, A. & Hutvagner, G. Small RNAs derived from the 5' end of tRNA can inhibit protein translation in human cells. RNA biology 10, 553-563 (2013).
+14. Kim, H.K. et al. A transfer-RNA-derived small RNA regulates ribosome biogenesis. Nature 552, 57-62 (2017).
+15. Lee, Y.S., Shibata, Y., Malhotra, A. & Dutta, A. A novel class of small RNAs: tRNA-derived RNA fragments (tRFs). Genes & development 23, 2639-2649 (2009).
+16. Chen, Z. et al. Transfer RNA demethylase ALKBH3 promotes cancer progression via induction of tRNA-derived small RNAs. Nucleic acids research 47, 2533-2545 (2019).
+17. Shao, Y. et al. tRF-Leu-CAG promotes cell proliferation and cell cycle in non-small cell lung cancer. Chemical biology & drug design 90, 730-738 (2017).
+18. Han, L. et al. A 5'-tRNA halve, tRNA-Gly promotes cell proliferation and migration via binding to RBM17 and inducing alternative splicing in papillary thyroid cancer. J Exp Clin Cancer Res 40, 222 (2021).
+19. Sharma, U. et al. Biogenesis and function of tRNA fragments during sperm maturation and fertilization in mammals. Science (New York, N.Y.) 351, 391-396 (2016).
+20. Chen, Q. et al. Sperm tsRNAs contribute to intergenerational inheritance of an acquired metabolic disorder. Science (New York, N.Y.) 351, 397-400 (2016).
+21. Lee, K.P. et al. Structure of the dual enzyme Ire1 reveals the basis for catalysis and regulation in nonconventional RNA splicing. Cell 132, 89-100 (2008).
+22. Cox, J.S., Sham, C.E. & Walter, P. Transcriptional induction of genes encoding endoplasmic reticulum resident proteins requires a transmembrane protein kinase. Cell 73, 1197-1206 (1993).
+23. Calfon, M. et al. IRE1 couples endoplasmic reticulum load to secretory capacity by processing the XBP-1 mRNA. Nature 415, 92-96 (2002).
+24. Yoshida, H., Matsui, T., Yamamoto, A., Okada, T. & Mori, K. XBP1 mRNA is induced by ATF6 and spliced by IRE1 in response to ER stress to produce a highly active transcription factor. Cell 107, 881-891 (2001).
+25. Coelho, D.S. et al. Xbp1-independent Ire1 signaling is required for photoreceptor differentiation and rhabdomere morphogenesis in Drosophila. Cell reports 5, 791-801 (2013).
+26. Hollien, J. et al. Regulated Ire1-dependent decay of messenger RNAs in mammalian cells. The Journal of cell biology 186, 323-331 (2009).
+27. Upton, J.P. et al. IRE1α cleaves select microRNAs during ER stress to derepress translation of proapoptotic Caspase-2. Science (New York, N.Y.) 338, 818-822 (2012).
+28. Oikawa, D., Tokuda, M., Hosoda, A. & Iwawaki, T. Identification of a consensus element
+
+<--- Page Split --->
+
+451 recognized and cleaved by IRE1 alpha. Nucleic acids research 38, 6265- 6273 (2010). 452 29. Tirasophon, W., Welihinda, A.A. & Kaufman, R.J. A stress response pathway from the endoplasmic reticulum to the nucleus requires a novel bifunctional protein kinase/endoribonuclease (Ire1p) in mammalian cells. Genes & development 12, 1812- 1824 (1998). 456 30. Boskovic, A., Bing, X.Y., Kaymak, E. & Rando, O.J. Control of noncoding RNA production and histone levels by a 5' tRNA fragment. Genes & development 34, 118- 131 (2020). 457 31. Tang, A.D. et al. Full- length transcript characterization of SF3B1 mutation in chronic lymphocytic leukemia reveals downregulation of retained introns. Nature communications 11, 1438 (2020). 461 32. Abdullahi, A., Stanojcic, M., Parousis, A., Patsouris, D. & Jeschke, M.G. Modeling Acute ER Stress in Vivo and in Vitro. Shock (Augusta, Ga.) 47, 506- 513 (2017). 463 33. Kumar, P., Kuscu, C. & Dutta, A. Biogenesis and Function of Transfer RNA- Related Fragments (tRFs). Trends in biochemical sciences 41, 679- 689 (2016). 465 34. Emara, M.M. et al. Angiogenin- induced tRNA- derived stress- induced RNAs promote stress- induced stress granule assembly. The Journal of biological chemistry 285, 10959- 10968 (2010). 468 35. Ogawa, T. et al. A cytotoxic ribonuclease targeting specific transfer RNA anticodons. Science (New York, N.Y.) 283, 2097- 2100 (1999). 470 36. Tomita, K., Ogawa, T., Uozumi, T., Watanabe, K. & Masaki, H. A cytotoxic ribonuclease which specifically cleaves four isoaccepting arginine tRNAs at their anticodon loops. Proceedings of the National Academy of Sciences of the United States of America 97, 8278- 8283 (2000). 473 37. Kumar, P., Anaya, J., Mudunuri, S.B. & Dutta, A. Meta- analysis of tRNA derived RNA fragments reveals that they are evolutionarily conserved and associate with AGO proteins to recognize specific RNA targets. BMC biology 12, 78 (2014). 476 38. Goodarzi, H. et al. Endogenous tRNA- Derived Fragments Suppress Breast Cancer Progression via YBX1 Displacement. Cell 161, 790- 802 (2015). 478 39. Couvillion, M.T., Bounova, G., Purdom, E., Speed, T.P. & Collins, K. A Tetrahymena Piwi bound to mature tRNA 3' fragments activates the exonuclease Xrn2 for RNA processing in the nucleus. Molecular cell 48, 509- 520 (2012). 481 40. Martinez, G., Choudury, S.G. & Slotkin, R.K. tRNA- derived small RNAs target transposable element transcripts. Nucleic acids research 45, 5142- 5152 (2017).  
+
+## Methods  
+
+## Cell culture and reagents  
+
+Cell lines used in this study are described in Supplementary Table 7. DMSO, TG, TM, and
+
+<--- Page Split --->
+
+STF083010 were purchased from Sigma- Aldrich (St Louis, MO, USA).  
+
+## Oligonucleotides  
+
+Synthetic oligonucleotides used in this study are listed in Supplementary Table 8.  
+
+## Plasmid construction and transfection  
+
+Plasmids used in this study are listed in Supplementary Table 9. The myc- tagged IRE1α (pCMV- IRE1α) and ANG were produced by PCR amplification. The PCR products were digested with KpnI and NotI for IRE1α and EcoRI and SalI for ANG (Takara Bio, Shiga, Japan) and then ligated into pCMV- myc empty vector (Clontech, Mountain View, CA, USA). KGN cells were transfected with plasmids for IRE1α, IRE1α (K599A) and ANG using Neon transfection system (Invitrogen, Carlsbad, CA, USA) as described previously<sup>41</sup>.  
+
+## Small RNA sequencing analysis  
+
+tRNA was sequenced from two biological replicate samples. Total RNA from the KGN and KGN- IRE1α<sup>ee</sup> cells were isolated and treated with T4 polynucleotide kinase (T4 PNK; New England Biolabs) and incubated at 37 °C for 30 min. Samples were separated on a 12% polyacrylamide gel containing 8 M urea to excise the 18- 40 nt region and were visualised with SYBR Gold (Thermo Fisher Scientific). RNAs were eluted from the acrylamide bands overnight in 0.3 M NaCl and then precipitated in ethanol/glycogen. Small RNA libraries were constructed using a SMARTER<sup>®</sup> smRNA- Seq Kit for Illumina<sup>®</sup> (Takara Bio) according to the manufacturer's guidelines. Sequencing libraries were generated according to the MiSeq reagent kit v3 and single end sequencing manufacturer instructions. Small RNA- seq reads were trimmed with the cutadapt programme<sup>42</sup> with parameters recommended by the SMARTER
+
+<--- Page Split --->
+
+smRNA- Seq Kit manual. Trimmed sequences with read- lengths ranging from 15 to 42 bp were collected and mapped to the human genome and non- redundant mature tRNA sequences using the bowtie2 program<sup>43</sup> implemented in the tRAX software package (http://trna.ucsc.edu/tRAX/). Reads mapped to tRNAs were extracted and their aligned positions were obtained using the bam2bed program of the BEDOPS suite<sup>44</sup>. The final position of a read was considered a cleavage site. Number of reads ending at each position of tRNAs was calculated. When a read was mapped to multiple tRNAs, fractional counts were allocated to all mapped tRNAs. The resulting read counts were subjected to differential cleavage analysis using the DESeq2 package<sup>45</sup>. Read mapping to a single isodecoder set was assigned to individual tRF, and those mapping to multiple tRFs with identical sequences were assigned to a single tRF considering mapped read counts of their 3'- tRFs. We used a tRNA gene annotation format, such as 'W- X- Y:Z' (W: amino- acid; X: anticodon; Y: unique gene identifier; Z: cleavage site) in Gly- GCC- 1:33.  
+
+## Northern blot analysis  
+
+The procedure for northern blot analysis has been described previously<sup>46</sup>. RNA was transferred to an Immobilon Hybond- XL membrane (GE Healthcare Life Sciences, Amersham, Buckinghamshire, UK) and then hybridised with a \(^{32}\mathrm{P}\) - 5'- end- labeled probe specific for the tRNAG<sup>Gly(GCC)</sup>. The northern blot membranes were then stripped and reprobed with a radiolabelled probe specific for the tRNA<sup>Cys(GCA)</sup>, tRNAG<sup>Gly(TCC)</sup>, tRNA<sup>Lys(CTT)</sup>, tRNA<sup>Val(TAC)</sup>, or 5.8S rRNA. 5.8S rRNA was used as a loading control.  
+
+## Western blot analysis  
+
+Total proteins were extracted and analysed by western blotting as described previously<sup>47</sup>. Total
+
+<--- Page Split --->
+
+protein from yeast cells were analysed according to the method described by Bahn et al<sup>48</sup>. The antibodies used in western blot analysis are listed in Supplementary table 10.  
+
+## Primer extension analysis  
+
+Three micrograms of total RNA from KGN cells were used in primer extension reactions. The Gly- GCC- R primer was labelled at the \(5^{\prime}\) - end with \((\gamma - ^{32}\mathrm{P})\) ATP and T4 polynucleotide kinase (New England Biolabs, Ipswich, MA, USA). RNA and the labelled primers were denatured at \(70^{\circ}\mathrm{C}\) for 5 min and then annealed by cooling to \(37^{\circ}\mathrm{C}\) for 90 min. They were then extended at \(42^{\circ}\mathrm{C}\) for 1 h with 5 units (U) of avian myeloblastosis virus reverse transcriptase (AMV RTase; New England Biolabs). The products were separated on \(10\%\) polyacrylamide gel containing 8 M urea. Sequencing ladders were generated using \(5\mu \mathrm{g}\) of the PCR product amplified from the cDNA of tRNA<sup>Gly(GCC)</sup>. Images were analysed in a Bio- Rad phosphorimager using Quantity One software (Bio- Rad Laboratories).  
+
+## tRNA purification  
+
+Unfractionated tRNAs (tRNA<sup>Mix</sup>) were purified from total RNA by gel purification. In brief, total RNA from KGN cells was separated on \(10\%\) polyacrylamide gel containing 8 M urea. The tRNA fraction was eluted from the gel in RNA extraction buffer [0.5 M ammonium acetate, \(0.2\%\) sodium dodecyl sulphate, and \(0.1\mathrm{mM}\) EDTA (pH 8.0)]. The eluted tRNA<sup>Mix</sup> was purified by phenol/chloroform extraction and ethanol precipitation. For further isolation of tRNA<sup>Gly(GCC)</sup>, oligo DNA- immobilised beads were prepared according to the method described by Yokogawa et al<sup>49</sup>.
+
+<--- Page Split --->
+
+## Cleavage analysis and site mapping  
+
+Purified total tRNAMix (1 \(\mu \mathrm{g}\) ) and \(5^{\prime}\) - end \(^{32}\mathrm{P}\) - labelled tRNAGly(GCC) were incubated with 10 pmol of recombinant IRE1α (OriGene Technologies, Rockville, MD, USA) in \(20~\mu \mathrm{L}\) of cleavage buffer [0.2 M HEPES pH 7.6, 0.5 M K(OAC), \(10\mathrm{mM}\mathrm{Mg(OAC)}_2\) , \(0.5\%\) Triton X- 100, \(10\mathrm{mM}\) DTT, and \(10\mathrm{mM}\) ATP] at \(37^{\circ}\mathrm{C}\) for 30 or 120 min. The cleaved products from purified tRNAMix were recovered and used for northern blot and primer extension assays. A hydrolysis ladder was then created by incubating 2 pmol of tRNAGly(GCC) in hydrolysis buffer (50 mM \(\mathrm{NaCO_3}\) pH 9.2 and \(1\mathrm{mM}\) EDTA pH 8.0) at \(95^{\circ}\mathrm{C}\) for 10 min. RNase T1 ladder was created by incubating 2 pmol of tRNAGly(GCC) with RNase T1 (Fermentas, Waltham, MA, USA) at \(37^{\circ}\mathrm{C}\) for 2 min in reaction buffer (30 mM Tris- HCl pH 7.9, \(10\mathrm{mM}\mathrm{MgCl}_2\) , \(160\mathrm{mM}\mathrm{NaCl}\) , \(0.1\mathrm{mM}\) DTT, and \(0.1\mathrm{mM}\) EDTA pH 8.0). The cleaved products from radiolabelled tRNAGly(GCC) were separated on a \(10\%\) polyacrylamide gel containing \(8\mathrm{M}\) urea, and images were analysed in a Bio- Rad phosphorimager using the Quantity One software (Bio- Rad Laboratories).  
+
+## Semi-quantitative RT-PCR and Reverse transcription–quantitative real-time PCR (RT- qPCR)  
+
+To amplify the spliced and unspliced XBP1 mRNA, a pair of primers (XBP1 splicing- F and XBP1 splicing- R) were used to flank the splicing site and yield 473 bp and 447 bp product sizes of XBP1u and XBP1s, respectively. Products were resolved on \(2.5\%\) agarose gel. Alternative splicing was detected using isoform specific primers. To amplify the spliced and unspliced HXL1 mRNA, a pair of primers, C. deutero- F and C. deutero- R, were used as described previously \(^{55}\) , yielding PCR product sizes of 475 bp and 419 bp for HXL1u and HXL1s, respectively. These PCR products were electrophoresed on \(2.5\%\) agarose gel. ACT1
+
+<--- Page Split --->
+
+was used as a loading control. Samples for RT- qPCR were prepared and analysed as previously described<sup>47</sup>. Gene expression levels were quantified using the \(\Delta \Delta \mathrm{Ct}\) method.  
+
+## Construction of IRE1α KO cell line  
+
+IRE1α KO cells were generated as described previously<sup>53</sup>. To generate plasmids targeting IRE1α, pX458 (D10A)- IRE1α (GuideA) or pX458 (D10A)- IRE1α (GuideB) dual- guide oligonucleotide primers were cloned into the vector pX458 (D10A). Targeted IRE1α genomic DNA fragments (679 bp) were amplified using primers gIRE1α (F) and gIRE1α (R). Allelic deletion was confirmed by TOPcloner™ TA core Kit (Enzynomics, Korea) and DNA sequencing (Cosmogenetech, Korea).  
+
+## Oligonucleotide pull-down assay  
+
+KGN cells were treated for 6 h with 0.1 μM TG; untreated cells were included as controls. After washing twice with \(1 \times\) PBS, harvested cells were disrupted with ice- cold cell lysis buffer [4% CHAPS, 100 mM NaCl, 2 mM EDTA and a \(1 \times\) protease inhibition cocktail (Roche, Mannheim, Germany) in 50 mM Tris/HCl buffer, pH 7.2]. The cellular levels of IRE1α and \(\beta\) - actin were determined by western blotting. To capture proteins bound to the enriched \(5'\) - tRH<sup>Gly(GCC)</sup> from TG- treated cells, 34 nucleotide RNA baits (200 nM each), designed via biotinylation of the \(5'\) - or \(3'\) - end were utilised in a Thermo Scientific Pierce streptavidin- coated microplate containing 200 μg protein per well and an RNase inhibitor (Promega). A \(5'\) - biotin- oligo A8 RNA was included in control wells. Following incubation for 1 h at 4 °C, the contents of the protein extract were aspirated and washed with 100 mM, 300 mM, and 500 mM NaCl in 25 mM Tris/HCl (pH 7.2). The bound proteins were eluted with SDS- PAGE loading buffer
+
+<--- Page Split --->
+
+containing \(1\%\) (w/v) SDS and \(5\%\) (v/v) 2-mercaptoethanol, as well as \(10\%\) (v/v) glycerol in 25 mM Tris/HCl (pH 6.8).  
+
+## Tandem mass spectrometry analysis  
+
+Protein bands detected via SDS- PAGE, following the biotin- streptavidin method, were excised, destained, and reduced with \(50~\mathrm{mM}\) dithiothreitol at \(60^{\circ}\mathrm{C}\) for \(15\mathrm{min}\) . The reduced cysteine residues were alkylated with \(100~\mathrm{mM}\) iodoacetamide at room temperature for \(1\mathrm{h}\) in the dark. The gel pieces were then washed thrice with deionised water and dehydrated twice in acetonitrile (ACN). The dried gels were soaked in \(10~\mathrm{mM}\) ammonium bicarbonate with \(20~\mu \mathrm{g / mL}\) trypsin (Promega, Madison WI, USA) on ice. Proteins in gel were digested for \(24\mathrm{h}\) at \(37^{\circ}\mathrm{C}\) and treated again with \(20~\mu \mathrm{L}\) of trypsin solution for another \(24\mathrm{h}\) . The digested peptides were extracted from the gel pieces and analysed on an nLC Velos Pro mass instrument equipped with a PicoFrit™ column ( \(100~\mathrm{mm}\) , packed with \(5\mu \mathrm{m}\) Biobasic® C18) and an EASY- Column™ ( \(2\mathrm{cm}\) , packed with \(5\mu \mathrm{m}\) C18; Thermo Fisher Scientific). The LC conditions were as follows: \(0.3~\mu \mathrm{L / min}\) was a 45- min linear gradient from \(5\%\) to \(40\%\) ACN in a \(0.1\%\) formic acid buffer solution, followed by a \(10\mathrm{min}\) column wash with \(80\%\) ACN and \(20\mathrm{min}\) re- equilibration to the initial buffer condition. Full mass (MS1) scan was performed in the \(m / z\) 300- 2000 range in a positive ion mode. Data- dependent MS2 scans of the seven most intense ions were performed from the full scan with the options of \(1.5m / z\) isolation width, \(35\%\) normalised collision energy, and \(30~\mathrm{s}\) dynamic exclusion duration. The acquired MS2 data were primarily analysed by SEQUEST search against a human reference protein database from the National Center for Biotechnology Information (https://www.ncbi.nlm.nih.gov/genome) and common protein contaminants in the common Repository of Adventitious Proteins (https://www.thegpm.org/crap/) with the following options: maximum miscleavage of 1,
+
+<--- Page Split --->
+
+precursor mass tolerance 0.8 Da, fragment mass tolerance 1.0 Da, dynamic modification of methionine oxidation, and static modification of cysteine with iodoacetamide. The identified proteins with unique peptides are reported in Supplementary Table 2.  
+
+## Electrophoretic mobility shift assay  
+
+The synthetic 5'- mimics of tRHs were purchased from Bioneer (Daejon, Korea). The 5'- end of tRH mimics were radiolabelled using \(\gamma\) - \(^{32}\mathrm{P}\) - ATP and T4 polynucleotide kinase (Takara Bio) and purified with an Illustra MicroSpin G- 25 column (GE Healthcare Life Sciences). Prior to use, labelled and unlabelled tRH mimics were heated at \(65^{\circ}\mathrm{C}\) for 10 min and slowly cooled to room temperature. Next, \(100\mathrm{ng}\) of BSA (Takara Bio), or HNRNPM (Mybiosource, San Diego, USA) or HNRNPH2 recombinant proteins (Mybiosource), were incubated with binding buffer [10 mM Tris- HCl (pH 8.0), \(150\mathrm{mM}\) KCl, \(0.5\mathrm{mM}\) EDTA, \(0.1\%\) Triton X- 100, \(0.02\mathrm{mM}\) DTT, \(12.5\%\) glycerol] and \(3.3\mathrm{pmol}\) of cold probes for each 5'- tRH mimic for 10 min at room temperature. Samples were then incubated with \(0.033\mathrm{pmol}\) of 5'- labelled tRHs for 10 min at room temperature. Native loading dye [100 mM Tris- HCl (pH 8.0), \(8.33\%\) glycerol, \(0.002\%\) brilliant blue G] and \(8\%\) polyacrylamide gels were used to load the samples. Vacuum dried gels were exposed to an intensifying screen and images were analysed in a Bio- Rad phosphorimager using Quantity One software (Bio- Rad Laboratories).  
+
+## HNRNPM and HNRNPH2 binding constants  
+
+The binding affinities of purified HNRNPM and HNRNPH2 (Mybiosource, San Diego, USA) to synthetic 5'- tRH- Gly \(^{GCC}\) or 3'- tRH- Gly \(^{GCC}\) (Bioneer, Daejon, Korea) were measured using BIAcore T200 instrument and CM5 sensorchip (GE Healthcare Life Sciences) at \(25^{\circ}\mathrm{C}\) . Activation, immobilization, deactivation and preparation of the mock- coupled flow cell were
+
+<--- Page Split --->
+
+performed according to the manufacturer's instructions. The binding signals were generated by subtracting the signal for the mock- coupled flow cell from that for the HNRNPM- or HNRNPH2- immobilized flow cells. Calculation of equilibrium dissociation constant \((K_{D})\) from the sensorgrams were done with BIAccore T200 Evaluation software version 3.2 (GE Healthcare Life Sciences) by fitting the data to a 1:1 binding model.  
+
+## Cell viability assay  
+
+All tRH mimics were purchased from were supplied by Genolution Pharmaceuticals, Inc. (Seoul, Korea). Antisense DNA oligos were purchased from BIONICS (Seoul, Korea). Cell- viability assays were performed as previously \(^{47}\) .  
+
+## RNA interference  
+
+All siRNAs were purchased from Bioneer (Seoul, Korea). The siRNA transfection method has been described previously \(^{47}\) .  
+
+## Cell proliferation assay  
+
+KGN or HeLa cells \((1 \times 10^{4})\) were seeded in 96- well plates for \(24 \mathrm{~h}\) ; the cells were then transfected with increasing amounts of \(5'\) - or \(3'\) - tRH mimics using the lipofectamine 2000 reagent (Invitrogen). HeLa cells \((1 \times 10^{4})\) were seeded in 96- well plates for \(24 \mathrm{~h}\) ; the cells were then transfected with increasing amounts of antisense DNA oligos complementary to the \(5'\) - tRH- Gly \(^{GCC}\) or \(5'\) - tRH- Lys \(^{CTT}\) loaded onto functionalized AuNP. After a 48- h transfection, cell proliferation was measured using the Cell Proliferation ELISA, BrdU (colorimetric) kit (Sigma- Aldrich) according to the manufacturer's instructions.
+
+<--- Page Split --->
+
+## Flow cytometry analysis  
+
+To detect apoptotic cells, KGN cells \((1 \times 10^{6})\) were transfected with the indicated \(5'\) - or \(3'\) - tRHs mimic and \(48 \mathrm{~h}\) post- transfection stained with the FITC Annexin V Apoptosis Detection Kit (BD Pharmingen, San Diego, CA, USA) according to the manufacturer's instructions.  
+
+## Cell migration assay  
+
+Cell migration was assessed based on the protocol described in our previous study<sup>47</sup>. Briefly, KGN cells \((1 \times 10^{6})\) were transfected with the indicated \(5'\) - or \(3'\) - tRH mimics for \(48 \mathrm{~h}\) . Images of migrated cells were captured at \(\times 100\) magnification under a bright- field microscope (Olympus CKX41, Tokyo, Japan).  
+
+## Total transcriptome analysis  
+
+RNA was sequenced from two biological replicate samples of KGN cells transfected with the \(5'\) - tRHs mimic for \(48 \mathrm{~h}\) . In brief, total RNA samples were converted into cDNA libraries using the TruSeq Stranded mRNA Sample Prep Kit (Illumina). Starting with \(1 \mu \mathrm{g}\) of total RNA, poly- . adenylated RNA (primarily mRNA) was selected and purified using oligo- dT- conjugated magnetic beads. This mRNA was physically fragmented and converted into single- stranded cDNA using reverse transcriptase and random hexamer primers, with the addition of actinomycin D to the FSA (First Strand Synthesis Act D Mix) to suppress DNA- dependent synthesis of the second strand. Double- stranded cDNA was created by removing the RNA template and synthesising the second strand in the presence of dUTP (deoxyribouridine triphosphate) in place of dTTP (deoxythymidine triphosphate). A single A base was added to the \(3'\) end to facilitate ligation of the sequencing adapters, which contained a single T base
+
+<--- Page Split --->
+
+overhang. Adapter- ligated cDNA was amplified by PCR to increase the amount of sequence- ready library. During this amplification the polymerase stalls when it encounters a U base, rendering the second strand a poor template. Accordingly, amplified material uses the first strand as a template, thereby preserving the strand information. Final cDNA libraries were analysed for size distribution using an Agilent Bioanalyzer (DNA 1000 kit; Agilent), quantitated by qPCR (Kapa Library Quant Kit; Kapa Biosystems, Wilmington, MA), and normalised to 2 nmol/L in preparation for total transcriptome analysis.  
+
+## Alternative splicing analysis  
+
+Purified mRNA was sequenced from three biological triplicate samples of KGN cells transfected with the 5'- tRHs mimic for 48 h. Briefly, Direct RNA sequencing was performed using the Direct RNA sequencing protocol (SQK- PCS109 kit) for the MinION. All steps were followed according to the manufacturer's specification. The constructed library was loaded on a FLO- MIN106D R9.4 flow cell and sequenced on a MinION device (Oxford Nanopore Technologies). The sequencing run was terminated after 48 h. Analyses of differential isoform usage using FLAIR modules has been described previously<sup>31</sup>.  
+
+## Induction of acute ER stress in vivo  
+
+Acute ER stress was induced in vivo using a mouse model as described previously<sup>32</sup>. Briefly, immunodeficiency female or male BALB/c nu/nu mice (7- weeks- old) were purchased from Saeron Bio Inc (Uiwang, Korea) and rested for 3- 5 days. BALB/c mice were injected intraperitoneally with TG solution (1 μg/g body weight) or TM solution (0.5 μg/g body weight) as described previously. As controls, mice were injected intraperitoneally with control buffer (1× PBS containing 2% DMSO). Mice were euthanized by cervical dislocation and, major
+
+<--- Page Split --->
+
+organs were harvested at \(6\mathrm{h}\) , \(12\mathrm{h}\) , and \(24\mathrm{h}\) post- treatment, and the samples were prepared for western blot and northern blot analyses. All animal protocols were approved by the Chung- Ang University Institutional Animal Case and Use committee (IRB# CAU202000115). For \(C\) . neoformans, WT and ire1- deletion (ire1 \(\Delta\) ) strains in early log phase were treated with TM (5 \(\mu \mathrm{g / ml}\) ) and DTT (2 mM) at \(30^{\circ}\mathrm{C}\) for \(2\mathrm{h}\) .  
+
+## Fluorescence in situ hybridisation  
+
+Cells were cultured under conditions of normal growth or subjected to ER stress by treating with \(0.1\mu \mathrm{M}\) TG for \(6\mathrm{h}\) . After culture, the cells were washed thrice in PBS, fixed with \(4\%\) paraformaldehyde in PBS for \(15\mathrm{min}\) at room temperature, and washed thrice with PBS. Cells were permeabilised with \(0.2\%\) Triton X- 100 in PBS for \(15\mathrm{min}\) at room temperature and washed twice with PBS. Slides were then blocked and prehybridised for \(2\mathrm{h}\) at \(37^{\circ}\mathrm{C}\) in hybridisation buffer ( \(2\%\) bovine serum albumin, \(5\times\) Denhardt's solution, \(4\times\) SSC, and \(35\%\) deionised formamide). Hybridisation was performed overnight in a humid dark chamber at \(37^{\circ}\mathrm{C}\) in the presence of \(1\mathrm{ng / mL}\) of the indicated oligonucleotide conjugated to cyanine 3 dye (Cy3). FISH assays were also performed under denaturing conditions by heating the slides at \(75^{\circ}\mathrm{C}\) for \(5\mathrm{min}\) , immediately before the hybridisation step. After hybridisation, cells were washed once in \(2\times\) SSC containing \(50\%\) deionised formamide, once in \(2\times\) SSC, and once in \(1\times\) SSC. Cells were mounted on slides using a mounting solution containing DAPI. Fluorescence was detected with a laser scanning confocal microscope (Carl Zeiss ZEN 2011, Germany). Relative fluorescence intensities were assessed using ImageJ software (NIH, USA).  
+
+## TaqMan assay for \(5^{\prime}\) -tRH-GlyGCC  
+
+The TaqMan assay was performed as described previously \(^{47}\) . KGN cells were treated with \(0.1\)
+
+<--- Page Split --->
+
+\(\mu \mathrm{M}\) of TG for the indicated times, and fractionation of nuclear and cytosolic RNA was isolated using a Cytoplasmic and Nuclear RNA Purification Kit (Norgen Biotek, Thorold, Canada), according to the manufacturer's instructions. \(5^{\prime}\) tRH- Gly \(^{GC C}\) and U6 snRNA quantification was conducted using custom designed TaqMan microRNA assays according to manufacturer's recommended protocols (Applied Biosystems, Foster City, CA, USA).  
+
+## Mouse xenograft experiment  
+
+HeLa cells \((1 \times 10^{6})\) were subcutaneously injected into 7- week- old BALB/c nu/nu immunodeficiency mice (Saeron Bio Inc), whose weights ranged between 18 and 20 g. We randomly allocated mice to three groups. Once the HeLa cells formed tumours (tumour volume: \(\sim 0.1 \mathrm{cm}^3\) ), \(\mathrm{AuNP}^{\mathrm{dT}}\) , \(\mathrm{AuNP}^{\mathrm{dT}}\) - anti- 5'- tRH- Lys \(^{\mathrm{CTT}}\) or \(\mathrm{AuNP}^{\mathrm{dT}}\) - anti- 5'- tRH- Gly \(^{\mathrm{GC C}}\) suspended in PBS were directly injected into the tumour sites every two days. \(\mathrm{AuNP}^{\mathrm{dT}}\) - anti- 5'- tRH- Lys \(^{\mathrm{CTT}}\) and \(\mathrm{AuNP}^{\mathrm{dT}}\) - anti- 5'- tRH- Gly \(^{\mathrm{GC C}}\) were prepared by mixing \(\mathrm{AuNP}^{\mathrm{dT}}\) (NES Biotechnology, Seoul, Korea) with polyadenylated antisense DNA oligos as previously described \(^{50}\) .  
+
+Mice were weighed and sizes of the tumours were measured every other day. The volume \((\mathrm{cm}^3)\) of each tumour ((length \(\times\) width \(^2 \times \pi\) )/6) was determined over 30 days period after xenotransplantation. Tumour- bearing mice were euthanized by cervical dislocation 18 days after the first injection of the functionalized AuNP composites, and tumours were excised. The samples were prepared for RT- qPCR and western blot analyses.  
+
+## Statistical analysis  
+
+Multiple- comparison analyses of values were performed using the Student- Newman- Keuls test, and Student's \(t\) - test was used for comparisons with control samples, using SAS version 9.2 (SAS Institute, Cary, NC, USA) and SigmaPlot (Systat Software, San Jose, CA, USA). The
+
+<--- Page Split --->
+
+data are presented as mean \(\pm\) standard error of the mean (SEM); \(P < 0.05\) was considered statistically significant.  
+
+## Data availability  
+
+Small RNA- seq data was deposited in the NCBI Sequence Read Archive (SRA) under the BioProject accession number PRJNA772059. The mass spectrometry data have been deposited in the ProteomeXchange Consortium via the PRIDE \(^{51}\) partner repository with the dataset identifier PXD013798.  
+
+## References  
+
+Kim, J.H. et al. Differential apoptotic activities of wild- type FOXL2 and the adult- type granulosa cell tumor- associated mutant FOXL2 (C134W). Oncogene 30, 1653- 1663 (2011). Martin, M. Cutadapt removes adapter sequences from high- throughput sequencing reads. EMBnet/journal; Vol 17, No 1: Next Generation Sequencing Data Analysis (2011). Langmead, B. & Salzberg, S.L. Fast gapped- read alignment with Bowtie 2. Nature methods 9, 357- 359 (2012). Neph, S. et al. BEDOPS: high- performance genomic feature operations. Bioinformatics (Oxford, England) 28, 1919- 1920 (2012). Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA- seq data with DESeq2. Genome biology 15, 550 (2014). (!! INVALID CITATION !!! (Lee et al., 2002)). Shin, E. et al. An alternative miRISC targets a cancer- associated coding sequence mutation in FOXL2. The EMBO journal 39, e104719 (2020). Bahn, Y.S., Kojima, K., Cox, G.M. & Heitman, J. Specialization of the HOG pathway and its impact on differentiation and virulence of Cryptococcus neoformans. Mol Biol Cell 16, 2285- 2300 (2005). Yokogawa, T., Kitamura, Y., Nakamura, D., Ohno, S. & Nishikawa, K. Optimization of the hybridization- based method for purification of thermostable tRNAs in the presence of tetraalkylammonium salts. Nucleic acids research 38, e89 (2010). Kim, J.H. et al. A functionalized gold nanoparticles- assisted universal carrier for antisense DNA. Chem Commun (Camb) 46, 4151- 4153 (2010).
+
+<--- Page Split --->
+
+803 51. Perez- Riverol, Y. et al. The PRIDE database and related tools and resources in 2019: 804 improving support for quantification data. Nucleic acids research 47, D442- d450 (2019). 805
+
+<--- Page Split --->
+
+## Figure legends  
+
+Fig. 1. Small RNA- seq analysis of IRE1α- induced tRFs in vivo. a, Total tRF or 5'- tRF mapped read counts in KGN- WT and IRE1α- overexpressing (KGN- IRE1α<sup>oe</sup>) cells. Hatched line: reads mapped to tRNA<sup>Gly(GCC)</sup>. b, (Left) Volcano plot depicting differentially expressed 5'- tRFs in KGN- WT and KGN- IRE1α<sup>oe</sup> cells. Red dots: 5'- tRFs from tRNA<sup>Gly(GCC)</sup>; blue dots: 5'- tRFs from tRNA<sup>Cys(GCA)</sup>; black dot: 5'- tRFs from tRNA<sup>Gly(GCC)</sup> expressed at higher levels in KGN- IRE1α<sup>oe</sup> cells (red box: Log<sub>2</sub> Fold Change \(>1.5\) ; \(p < 0.001\) ). (Right) Based on small RNA- seq analysis, cleavage sites at the anticodon loop in the secondary human tRNA<sup>Gly(GCC)</sup> and tRNA<sup>Cys(GCA)</sup> structures. Red: acceptor stem at 5'- end; Purple: D loop; light green: anticodon loop; dark green: anticodon; yellow: T loop; blue: CCA tail at 3'- end. Numbering in the anticodon indicates the 3'- end positions of the tRFs; percent indicates the proportion of 5'- tRFs in total 5'- tRFs (Log<sub>2</sub> Fold Change \(>1.5\) ; \(p < 0.001\) ). c, Northern blot analysis of tRNA fragments in KGN cells following IRE1α overexpression KGN cells were transfected with plasmid encoding myc- tagged IRE1α for 24 h, total RNA was extracted for analysis of 5'- tRNA fragments by northern blotting. The expression of IRE1�� and GAPDH (loading control) was analysed by western blotting. Ribonucleolytic activity of IRE1α was confirmed XBP1 splicing assay using RT- PCR analysis of unspliced/spliced (u/s) XBP1. Red arrow: 5'- tRFs from tRNA<sup>Gly(GCC)</sup> generated by IRE1α. M: size marker. Percentage of 5'- tRF compared to full- length tRNA are shown. The data are presented as the mean \(\pm\) SEM from three independent experiments. The asterisk indicated statistically significant differences (**\(p < 0.01\) , *\(p < 0.05\) ; paired student's \(t\) - test). ns, not significant. d, (Left) Primer extension analysis of 5'- end of tRNA<sup>Gly(GCC)</sup> fragment in KGN cells. KGN cells were transfected with a plasmid encoding IRE1α or kinase defected mutant (IRE1α- K599A). (Right) Secondary structure of mature tRNA<sup>Gly(GCC)</sup> and IRE1α cleavage sites at anticodon. Numbering in the anticodon indicates the
+
+<--- Page Split --->
+
+positions of mature tRNA nucleotides. Red arrow: prominent cleaved products of the tRNAGly(GCC) generated by IRE1α.  
+
+Fig. 2. IRE1α- specific cleavage of tRNAGly(GCC) in vitro. a, Northern blotting results for tRNAGly(GCC) and tRNALys(CTT). Red arrow: prominent cleaved products of the tRNAGly(GCC) generated by IRE1α. b, Primer extension assay on tRNAGly(GCC) cleavage products in the presence of IRE1α in vitro. IRE1α cleavage sites in the tRNAGly(GCC) are denoted by different letters (a–g). c, In vitro cleavage of tRNAGly(GCC) by IRE1α. Secondary structure of mature tRNAGly(GCC) and IRE1α cleavage sites (a–g from Fig. 2b and 1–7 from Fig. 2c). Black arrows: position of the tRNAGly(GCC) cleavage site generated by IRE1α. Red arrow: major cleavage site by IRE1α.  
+
+Fig. 3. ER stress induces 5'- tRHs cleavage by tRNAGly(GCC). a, Northern blot analysis of tRNAGly(GCC)- derived fragments in KGN cells upon ER stress. KGN Cells were treated with 0.1% DMSO, TG (0.1 μM) or TM (1 μg/ml) and harvested at the indicated times. Total RNA was isolated and probed with a probe specific for the tRNAGly(GCC). Red arrow: tRHs from tRNAGly(GCC) cleaved by IRE1α. b, (Upper panel) 5'- end of tRNAGly(GCC) fragment detected in Fig. 3a (at 6 h) determined by primer extension analysis. (Lower panel) Secondary structure of mature tRNAGly(GCC) and IRE1α cleavage sites at anticodon stem loop. Red arrow: major IRE1α cleavage site. c, Northern blot analysis of tRNAGly(GCC) fragments in control KGN (WT) or IRE1α knockout-KGN cells (IRE1α-/-). Cells were treated with DMSO, TG (0.1 μM) or TM (1 μg/ml) for 6 h and harvested. Total RNA was isolated and probed with a probe specific for the tRNAGly(GCC). Relative amount of 5'- tRH-Gly(GCC) is presented in the lower panel of Fig. 3a and c, respectively. The data are presented as the mean ± SEM from three independent experiments.
+
+<--- Page Split --->
+
+The expression of IRE1α and \(\beta\) - actin (loading control) was analysed by western blotting. Ribonucleolytic activity of IRE1α was confirmed XBP1 splicing assay using RT- PCR analysis of unspliced/spliced (u/s) XBP1. Different letters denote statistically significant differences \((p\) \(< 0.0001\) ; Student- Newman- Keuls test).  
+
+Fig. 4. Functional roles of 5'- tRFs of tRNAGlu(GCC). a,b, Cell viability of KGN (a) and HeLa (b) cells following transfection with tRH mimics (5'- tRH- Lys<sup>CTT</sup>, 5'- tRH- Gly<sup>GCC</sup>, or 3'- tRH- Gly<sup>GCC</sup>). c,d, Cell viability of KGN (c) and HeLa (d) cells following transfection with small interfering RNAs (siRNAs) for HNRNPM or HNRNPH2 (200 nM) and tRH mimics (50 nM) (5'- tRH- Lys<sup>CTT</sup>, 5'- tRH- Gly<sup>GCC</sup>, or 3'- tRH- Gly<sup>GCC</sup>) (upper panel). Knockdown efficiency of HNRNPM or HNRNPH2 proteins (lower panel). e, Cell viability of WT and IRE1α<sup>-/-</sup> KGN cells following transfection with antisense DNA oligos (50 nM) targeting endogenous 5'- tRFs (anti-5'- tRH- Lys<sup>CTT</sup> or anti-5'- tRH- Gly<sup>GCC</sup>) in the absence or presence of TG (0.1 μM). (a-e) Data are presented as the mean ± SEM of three independent experiments performed in triplicate. Different letters denote statistically significant differences \((p < 0.0001\) ; Student- Newman- Keuls test). f, Volumes of tumours from mice injected with either the AuNP<sup>dT</sup> only (vehicle) as a control (n = 18), AuNP<sup>dT</sup>- anti-5'- tRH- Lys<sup>CTT</sup> (anti-5'- tRH- Lys<sup>CTT</sup>, n = 18), or AuNP<sup>dT</sup>- anti-5'- tRH- Gly<sup>GCC</sup> (anti-5'- tRH- Gly<sup>GCC</sup>, n = 18) were measured. g, Representative immunoblots and quantified data for tumours from each group are presented. (f-g) The asterisk indicated statistically significant differences (\\*p < 0.05, \\*\\*p < 0.01, \\*\\*\\*p < 0.001; paired student's t- test). ns, not significant.  
+
+Fig. 5. 5'- tRH- Gly<sup>GCC</sup> mediate alternative splicing events. a, Volcano plot of differentially expressed protein- coding genes in KGN cells transfected with 5'- tRH- Gly<sup>GCC</sup> mimic and
+
+<--- Page Split --->
+
+control KGN cells transfected with \(5^{\prime}\) - tRH- Lys<sup>CTT</sup> mimic. Blue dots: significant upregulation of target genes; red dots: significant downregulation of target genes. b, DAVID functional analysis of genes with transcript abundance altered by more than 1.5- fold. c, Differential isoform usage (left) and major isoforms (right) from the ELOB (upper panel) and PMSB5 (lower panel). Red box: alternative splicing region. d, Validation of alternative splicing events from ELOB and PMSB5 in KGN cells transfected with \(5^{\prime}\) - tRH- Lys<sup>CTT</sup> or \(5^{\prime}\) - tRH- Gly<sup>GCC</sup> mimics by RT- qPCR. e, Inhibitory effects of AuNP- conjugated antisense DNA oligos (anti- \(5^{\prime}\) - tRH- Lys<sup>CTT</sup> or anti- \(5^{\prime}\) - tRH- Gly<sup>GCC</sup>) were confirmed by validation of alternative splicing events from ELOB and PMSB5. (d- e) The asterisk indicated statistically significant differences ( \(*p < 0.05\) , \(**p < 0.01\) , \(***p < 0.001\) , \(****p < 0.0001\) ; paired student's \(t\) - test).  
+
+Fig. 6. ER stress induces generation of \(5^{\prime}\) - tRHs from tRNA<sup>Gly(GCC)</sup> in mouse and C. neoformans. a, Northern blot analysis of tRNA<sup>Gly(GCC)</sup>- derived fragments in the ovary from ER stress- induced mouse. 5.8S rRNA was used as the loading control. The expression of IRE1α and β- actin (loading control) was analysed by western blotting. b, \(5^{\prime}\) - end of tRNA<sup>Gly(GCC)</sup> fragment as determined by primer extension assay using total RNA isolated from ovaries after treatment with 0.1% DMSO or TG (left panel). Secondary structure of mouse mature tRNA<sup>Gly(GCC)</sup> and IRE1α cleavage sites at anticodon stem loop (right panel). Red arrow: TG- induced IRE1α cleavage sites c, Northern blot analysis of tRNA<sup>Gly(GCC)</sup> fragments in C. neoformans. 5.8S rRNA was used as the loading control. The expression of IRE1 and GAPDH (loading control) was analysed by western blotting. Ribonucleolytic activity of IRE1α was confirmed HXL1 splicing assay using RT- PCR analysis of unspliced/spliced (u/s) HXL1. Arrows: TM- induced IRE1 cleavage sites. d, Primer extension analysis of tRNA<sup>Gly(GCC)</sup> fragments in C. neoformans. Secondary structure of C. neoformans mature tRNA<sup>Gly(GCC)</sup> and
+
+<--- Page Split --->
+
+IRE1 cleavage sites at anticodon stem loop are illustrated (right panel). Red arrow: TM- induced IRE1 cleavage sites e, Proposed model for the IRE1α selective generation of 5'- tRH- Gly<sup>GCC</sup> that contributes to cellular adaptation upon ER stress presented in diverse eukaryotic organisms from yeast to humans.  
+
+## Extended data figure legends  
+
+Extended Data Fig. 1. Production of 5'- tRHs from tRNA<sup>Gly(GCC)</sup> in vivo via an IRE1α activity- dependent manner. a, Northern blot analysis of tRNA fragments in KGN cells. KGN cells were transfected with an empty vector (pCMV- myc) or a plasmid encoding IRE1α kinase defected mutant (IRE1α- K599A). Total RNA was isolated from the transfected cells and northern blot was performed with a \(^{32}\mathrm{P}\) - 5'- end- labeled probe specific for the tRNA<sup>Gly(GCC)</sup>. 5.8S rRNA was probed as a loading control. The expression of IRE1α and GAPDH (loading control) was analysed by western blotting. The northern blot membranes were then stripped and reprobed with a \(^{32}\mathrm{P}\) - 5'- end- labeled probe specific for the tRNA<sup>Cys(GCA)</sup>, tRNA<sup>Gly(TCC)</sup>, or tRNA<sup>Lys(CTT)</sup>. Ribonucleolytic activity of IRE1α was confirmed XBP1 splicing assay using RT- PCR analysis of unspliced/spliced (u/s) XBP1. M, size marker. Red arrow: 5'- tRFs from tRNA<sup>Gly(GCC)</sup> generated by IRE1α- K599A. b, Northern blots for tRNA<sup>Gly(GCC)</sup> fragments in KGN cells after transfection with 2.5 μg of pCMV- myc control plasmid or plasmid encoding myc- tagged ANG. 5.8S rRNA was used as the loading control. Blue arrows: prominent cleaved products of the tRNA<sup>Gly(GCC)</sup> generated by ANG.  
+
+Extended Data Fig. 2. Purification and cleavage of tRNA<sup>Gly(GCC)</sup> in vitro. a, Scheme diagram of isolation of tRNA<sup>Gly(GCC)</sup> species from purified total tRNA in vitro using the biotinylated antisense oligo DNA- conjugated streptavidin C1 beads (left panel). Red text and
+
+<--- Page Split --->
+
+lines indicate biotinylated antisense oligo DNA complementary to the secondary structure of tRNAGly(GCC) (right panel). b, Loading capacity of biotinylated antisense oligo DNA on streptavidin C1 beads in EtBr stained gel. Approximately 8 pmol of oligo DNAs were loaded onto 1 mg of streptavidin C1 beads. c, Quality of isolated tRNAGly(GCC) species determined by EtBr staining and northern blot assay.  
+
+Extended Data Fig. 3. Cleavage of tRFs induced by ER stress. a, Northern blot analysis of tRNACys(GCA), tRNAGly(TCC), or tRNALys(CTT)- derived fragments in KGN cells upon ER stress. The northern blot membranes used in Fig. 3a were stripped and re-probed with a \(^{32}\mathrm{P} - 5'\) - end- labelled probe specific for the tRNACys(GCA), tRNAGly(TCC), or tRNALys(CTT). b, Northern blot analysis of tRNA fragments in control KGN (WT) or IRE1α knockout-KGN cells (IRE1α/- ). The northern blot membranes used in Fig. 3c were stripped and re-probed with a \(^{32}\mathrm{P} - 5'\) - end- labelled probe specific for the tRNACys(GCA), tRNAGly(TCC), or tRNALys(CTT). c, Northern blot analysis of tRNAGly(GCC) fragments in HeLa cells. Red arrow: tRHs from tRNAGly(GCC) cleaved by IRE1α. Right panel: relative amount of \(5'\) - tRH-Gly(GCC). The data are presented as the mean \(\pm\) SEM from three independent experiments. \(^{**}p < 0.01\) ; paired Student's \(t\) - test. d, Northern blot membranes used in Extended data Fig. 3c were stripped and re-probed with a \(^{32}\mathrm{P} - 5'\) - end- labelled probe specific for the tRNACys(GCA), tRNAGly(TCC), or tRNALys(CTT).  
+
+Extended Data Fig. 4. Generation of IRE1α- knockout (IRE1α/- ) KGN cells using the CRISPR/Cas9 system. a, Schematic representation of the CRISPR/Cas9- nicksase strategy for IRE1α knockout in KGN cells. A pair of guide RNAs [sgIRE1α (A) and (B)] was designed by targeting the PAM sequence at the catalytic region of IRE1α. The sgRNA sequences and PAM sequences are shown in blue and red, respectively. Triangle indicates the possible cleavage sites.
+
+<--- Page Split --->
+
+b, Knockout of IRE1α validated using western blot analysis. c, T7E1 assay results to detect CRISPR/Cas9- induced modification in IRE1α<sup>-/-</sup>. Arrows indicate the positions of the expected DNA bands cleaved by T7E1. d, DNA sequences of the targeting sites in IRE1α<sup>-/-</sup> KGN cells. The regions of IRE1α sequences are shown in pink. The number of mutated nucleotides is indicated on the right. Actual chromatograms from sequencing analysis are shown.  
+
+Extended Data Fig. 5. Identification of proteins interacting with 5'- tRH- Gly<sup>GCC</sup>. a, SDS- PAGE of whole- cell lysates and 3'- /5'- biotinylated tRH- Gly<sup>(GCC)</sup>- bound proteins captured in a streptavidin microplate containing samples from KGN cells treated (+) and untreated (-) with 0.1 μM thapsigargin (TG) for 6 h. Different protein bands with molecular weights (MWs) near 70 kDa and 55 kDa in samples bound to 5'- biotin- tRH- Gly<sup>GCC</sup> are shown in an SDS- PAGE gel. b, Venn diagrams of identified proteins from the excised gel pieces at MWs near 70 kDa and 55 kDa. Normalised counts of the peptide- to- spectrum matches of identified proteins with > 2 unique peptides from tandem mass spectrometry data are compared. Tandem mass spectrometry results from each sample are shown in Supplementary Table 2.  
+
+Extended Data Fig. 6. Interaction between tRHs with HNRNP proteins. a, Electrophoretic mobility shift assay (EMSA) results for synthetic 5'- tRH- Gly<sup>GCC</sup> or 5'- tRH- Lys<sup>CTT</sup>. Synthetic 5'- tRH- Lys<sup>CTT</sup> was used as a control. b, Sensorgrams of the interaction between the immobilized HNRNP proteins (HNRNPM or HNRNPH2) and the purified 5'- tRH- Gly<sup>GCC</sup> used as analyte. Purified 5'- tRH- Gly<sup>GCC</sup> were delivered from the lowest to the highest concentration (30, 100, 300, 1000, 3000, 10000 nM). c, Summary of surface plasmon resonance (SPR) kinetic and affinity measurements using BIAcore T200. 5'- tRH- Gly<sup>GCC</sup> or 3'- tRH- Gly<sup>GCC</sup> binding to HNRNPM and HNRNPH2 measured by SPR. The equilibrium dissociation constant
+
+<--- Page Split --->
+
+\((K_{D})\) , the association constant \((k_{a})\) and the dissociation constant \((k_{d})\) are presented.  
+
+# Extended Data Fig. 7. Effects of transfecting tRHs mimics or antisense DNA oligos against  
+
+tRHs on physiology of KGN cells. a, Cell viability of KGN cells following transfection with tRHs mimics and antisense oligos to each of the tRHs mimics (mimic + antisense treated). b, Proliferation of KGN cells following transfection with tRHs mimics. Different letters denote statistically significant differences \((p < 0.0001\) ; Student- Newman- Keuls test). Data are presented as the mean \(\pm\) SEM of three independent experiments performed in triplicate. Effects of tRHs mimics on apoptosis (c) and (d) migration of KGN cells. ns, not significant. Scale bar \(= 100 \mu \mathrm{m}\) . e, f, Upper panel: cell viability of KGN (e) and HeLa (f) cells following transfection with siRNAs for HNRNPF or HNRNPH1 and tRH mimics from tRNA. Knockdown efficiency of HNRNP proteins (lower panel). Different letters denote statistically significant differences \((p < 0.0001\) ; Student- Newman- Keuls test). g, Proliferation of HeLa cells following transfection with antisense DNA oligos of tRHs. Different letters denote statistically significant differences \((p < 0.0001\) ; Student- Newman- Keuls test). Data are presented as the mean \(\pm\) SEM of three independent experiments performed in triplicate.  
+
+Extended Data Fig. 8. Subcellular localisation of 5'- tRHs from tRNAGly(GCC) during ER stress. a, Red fluorescence (Cy3): subcellular distribution of 5'- tRHs, from tRNAGly(GCC) in unstressed (right panel) and stressed cells (left panel) following TG treatment. DNA was stained with DAPI. Anticodon bridging probes were designed to anneal with the anticodon loop, and they recognised only the anticodon loop of whole tRNA to avoid significant hybridisation signals with 5'- tRHs of tRNAGly(GCC). Scale bar \(= 20 \mu \mathrm{m}\) . b, 5'- tRH- Gly(GCC) enrichment following transfection of KGN cells with 0.1 \(\mu \mathrm{M}\) of TG. The data (mean \(\pm\) SEM) are presented
+
+<--- Page Split --->
+
+as the fold enrichment calculated from three independent experiments. \(**p < 0.01\) , ns, not significant.  
+
+# Extended Data Fig. 9. Effects of ER stress-inducing agents on IRE1α expression in mouse  
+
+and C. neoformans. a, IRE1α and \(\beta\) - actin (loading control) protein expression in the organs of BALB/c mice injected with TM and TG. b, Northern blot analysis of tRNAGly(GCC)- derived fragments in the liver from ER stress- induced mice. 5.8S rRNA was used as the loading control. Red arrow: prominent cleaved products of the tRNAGly(GCC) generated by IRE1α. M, size marker from KGN cells. c, Northern blot analysis of tRNAGly(GCC)- derived fragments in the epididymis from ER stress- induced mouse. 5.8S rRNA was used as the loading control. The expression of IRE1α and \(\beta\) - actin (loading control) was analysed by western blotting. M, size marker from KGN cells. d, The northern blot membranes used in Fig. 6a were stripped and re- probed with a \(^{32}\mathrm{P} - 5'\) - end- labelled probe specific for the tRNACys(GCA), tRNAGly(TCC), or tRNALys(CTT), respectively. e, Northern blot analysis of tRNACys (GCA), tRNAGly(TCC), or tRNALys(CTT)- derived fragments in C. neoformans. The northern blot membranes used in Fig. 6c were stripped and re- probed with a \(^{32}\mathrm{P} - 5'\) - end- labelled probe specific for the tRNACys(GCA), tRNAGly(TCC), or tRNALys(CTT).  
+
+# Extended Data Fig. 10. Effects of other stresses on 5'-tRH-GlyGCC production and IRE1α overexpression on expression of GCC codon-rich genes. a, b, Northern blot analysis of KGN  
+
+cells treated with sodium arsenite (SA) (a) and Cobalt chloride (CoCl2) (b) (upper panel). IRE1α, ANG, and \(\beta\) - actin (loading control) expression (lower panel). c, Expression of FOXL2,  
+
+Loricrin, IRE1α, and \(\beta\) - actin (loading control) in KGN cells transfected with plasmid encoding  
+
+myc- tagged IRE1α and treated with STF083010 (inhibitor of the endonuclease activity of
+
+<--- Page Split --->
+
+1022 IRE1α). 
+
+1023
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+ 
+
+<center>Figure 1</center> 
+
+![](images/Figure_5.jpg)
+
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+
+<--- Page Split --->
+![](images/Figure_unknown_2.jpg)
+
+
+<--- Page Split --->
+![PLACEHOLDER_50_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_51_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_52_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_53_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_54_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_55_0]
+
+ 
+
+![PLACEHOLDER_55_1]
+
+ 
+
+![PLACEHOLDER_55_2]
+
+
+<--- Page Split --->
+![PLACEHOLDER_56_0]
+
+<center>Extended Data Figure 5 </center>  
+
+![PLACEHOLDER_56_1]
+
+
+<--- Page Split --->
+![PLACEHOLDER_57_0]
+
+ 
+
+<center>b</center> 
+
+![PLACEHOLDER_57_1]
+
+ 
+
+<center>C</center> 
+
+<table><tr><td rowspan="2"></td><td colspan="3">5'-tRH-Gly<sup>GCC</sup></td><td colspan="3">3'-tRH-Gly<sup>GCC</sup></td></tr><tr><td>\(K_D\) (nM)</td><td>\(K_a\) (M<sup>-1</sup>s<sup>-1</sup>)</td><td>\(K_d\) (s<sup>-1</sup>)</td><td>\(K_D\) (nM)</td><td>\(K_a\) (M<sup>-1</sup>s<sup>-1</sup>)</td><td>\(K_d\) (s<sup>-1</sup>)</td></tr><tr><td>HNRNPM</td><td>86.30</td><td>1.04 x 10<sup>4</sup></td><td>8.96 x 10<sup>-4</sup></td><td>959</td><td>2.74 x 10<sup>3</sup></td><td>2.63 x 10<sup>-3</sup></td></tr><tr><td>HNRNPH2</td><td>27.07</td><td>2.92 x 10<sup>4</sup></td><td>7.90 x 10<sup>-4</sup></td><td>938</td><td>4.90 x 10<sup>3</sup></td><td>4.60 x 10<sup>-3</sup></td></tr></table>
+
+<--- Page Split --->
+![PLACEHOLDER_58_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_59_0]
+
+ 
+
+<center>b</center> 
+
+![PLACEHOLDER_59_1]
+
+
+<--- Page Split --->
+![PLACEHOLDER_60_0]
+
+
+<--- Page Split --->
+![PLACEHOLDER_61_0]
+
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryTable110DB.pdf
+
+<--- Page Split --->

preprint/preprint__08ebe812ad76a570c3ee7e20f364f13b63c9040737d4f23cef0f561c4cb41a3a/images_list.json

@@ -0,0 +1,92 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1: Overview of the proposed data-driven EMS framework. (a) Comparison of three EMS paradigms: Traditional rule-based EMS relies on expert knowledge and calibration based on fixed driving cycles. Simulation-based EMS requires high-precision models and entails transitioning from simulation analysis to real-world deployment, resulting in a gap between simulated and real-world performance (sim-to-real gap). In contrast, the proposed real-world data-driven EMS learns directly from actual data. (b) China has established a three-tier EV monitoring and management system involving the state, local governments, and enterprises. The National Monitoring and Management Platform collects real-time operational data from over 20 million EVs [25]. (c) The ORL agent works with an existing EMS and continuously collects EV data to improve the EMS. (d) Overall diagram of the proposed data-driven EMS methodology.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        68,
+        880,
+        608
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2: Comparison for different datasets. (a) The simulation model for the fuel cell hybrid electric vehicle (FCEV), depicts the powertrain layout and energy flow topology. (b) The principle of EMS involves the allocation of energy flow among hybrid energy systems (fuel cell and battery) to achieve predefined objectives based on driving conditions. Here, real driving conditions are gathered and then use the simulated FCEV model to generate energy cost data for training the ORL agent. (c) Distribution of encoded actions for the four datasets, each action is normalized to the [-1,1]. D1 represents data generated by an expert policy, D2 and D3 denote suboptimal data generated by a combination of expert and random policies, D4 comprises entirely random data. (d) The states distribution of four datasets.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        72,
+        861,
+        625
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3: The ORL agent learning performance. (a) Learning curves of ORL agent for the four different datasets. (b) The comparison of absolute rewards (original rewards are negative) under three validation conditions. \"D1,\" \"D2,\" \"D3,\" and \"D4\" correspond to the best rewards achieved by ORL after learning on each respective dataset. Notably, ORL agent on various datasets closely approximate or exceed the expert policy. (c) The comparison of energy costs between the original EMS and the optimized EMS using ORL shows a notable reduction in energy costs through the data-driven learning process. (d) The action (FC power slope) distributions of optimized EMS using ORL",
+    "footnote": [],
+    "bbox": [
+      [
+        66,
+        230,
+        920,
+        747
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4: Performance analysis comparing different methods. (a) The comparison between two data-driven EMS methods. The matrix numbers represent the relative reward rates of ORL and BC compared to the expert EMS (PPO) under the same conditions, emphasizing the minimal influence of data quality on ORL's performance. (b) Comprehensive performance of different algorithms on the WTVC condition, DP representing the globally optimal EMS. (c) Comprehensive performance on the CHTC condition. (d) Comprehensive performance on the FTP condition. (e) The efficiency distribution of fuel cell power demonstrates that ORL learns a superior EMS, ensuring that fuel cell system remains within the high-efficiency range. (f) FC degradation costs under three conditions",
+    "footnote": [],
+    "bbox": [
+      [
+        68,
+        67,
+        930,
+        604
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5: ORL continuously learning and improving from data. (a) ORL for continuous learning from data in three scenarios. (b) Driving data collected from the real-world route. (c) Speed trajectories of three ZBDC conditions. (d) Comparing the total cost under ZBDC-No1, the total cost comprises hydrogen consumption, battery cost, and cell degradation cost. (e) Comparing the total cost under ZBDC-No2. (f) Comparing the total cost under ZBDC-No3. (g) Fuel cell power distribution cloud chart of Baseline EMS. (h) Fuel cell power distribution cloud chart following one data update. (i) Fuel cell power distribution cloud chart following two data updates.",
+    "footnote": [],
+    "bbox": [
+      [
+        70,
+        75,
+        940,
+        776
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6: Performance with more data. (a) Performance of PPO for 4 training data and 12 testing data. (b) The comprehensive performance comparison of different EMS methods reveals that ORL can significantly mitigate the performance degradation observed in the testing phase of PPO. (c) Performance of ORL for 12 testing data. (d) Speed distribution of four conditions. (e) The demand power of HWDC represents an extreme condition. The red dashed line indicates the maximum output power of FC system. (f) Overall performance of the 4 cases as the training data increase under test driving cycle CLTC; (g) Performance on GCDC; (h) Performance on ZNDC; (I) Performance on HWDC indicates that ORL ultimately learns a reasonable EMS in extreme driving conditions. (j) Battery SOC trajectories of different EMS under the test driving cycle HWDC.",
+    "footnote": [],
+    "bbox": [
+      [
+        68,
+        70,
+        945,
+        753
+      ]
+    ],
+    "page_idx": 11
+  }
+]

preprint/preprint__08ebe812ad76a570c3ee7e20f364f13b63c9040737d4f23cef0f561c4cb41a3a/preprint__08ebe812ad76a570c3ee7e20f364f13b63c9040737d4f23cef0f561c4cb41a3a.mmd

@@ -0,0 +1,337 @@
+
+# Learning Superior Energy Management from Electric Vehicle Data  
+
+Hongwen He hwhbit@bit.edu.cn  
+
+Beijing Institute of Technology  Yong Wang  Beijing Institute of Technology  Jingda Wu  Nanyang Technological University https://orcid.org/0000- 0002- 7336- 4492  Zhongbao Wei  Beijing Institute of Technology https://orcid.org/0000- 0003- 0051- 5648  Fengchun Sun  Beijing Institute of Technology  
+
+## Article  
+
+Keywords: Energy management, Electric vehicle data, Reinforcement learning, Fuel cell vehicles, Data- driven  
+
+Posted Date: July 3rd, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4523312/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on March 22nd, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 58192- 9.
+
+<--- Page Split --->
+
+# Learning Superior Energy Management from Electric Vehicle Data  
+
+Yong Wang \(^{a,b}\) , Jingda Wu \(^{c}\) , Hongwen He \(^{a,b}\) , Zhongbao Wei \(^{a}\) and Fengchun Sun \(^{a}\)  
+
+\(^{a}\) School of Mechanical Engineering, Beijing Institute of Technology, Beijing, 100081, China \(^{b}\) National Key Laboratory of Advanced Vehicle Integration and Control, Beijing Institute of Technology, Beijing, 100081, China \(^{c}\) The Hong Kong Polytechnic University, Hung Hom, Hong Kong  
+
+## ARTICLE INFO  
+
+## ABSTRACT  
+
+Keywords: Energy management Electric vehicle data Reinforcement learning Fuel cell vehicles Data- driven  
+
+Despite the promising potential of energy management technologies in optimizing electric vehicle (EV) performance and fostering global energy sustainability, the extensive research conducted over the past decade has yet to translate into practical applications. This discrepancy arises primarily from the reliance of existing methodologies on simulation- based development paradigms, leading to a significant disparity between simulated results and real- world efficacy. Herein, we present a pioneering real- world data- driven energy management strategies (EMS) approach that utilizes an innovative offline reinforcement learning (ORL) framework. This paradigm enables EMS to learn from diverse real- world data, obviating the need for explicit rule design or high- fidelity simulators, and allowing for seamless application of the proposed method to any existing EMS. Moreover, it continuously enhances performance even after deployment in actual energy management systems. We evaluate the proposed ORL method on fuel cell EVs, training the ORL agent to optimize energy consumption and system degradation. The EV monitoring and management platform in China provides real- world data for validating our methodology. The results demonstrate that ORL consistently learns superior EMS in various conditions. With increasing data availability, its performance improves significantly, from \(88\%\) to \(98.6\%\) relative to theoretical optimality after two data updates. After training with more than 60 million kilometers of data, the ORL agent can learn a general EMS that adapts to unseen and corner- case conditions. These results highlight the effectiveness of integrating the data- driven method with established EMS techniques to enhance performance and underline its potential to utilize large- scale data to improve vehicle energy efficiency and longevity.  
+
+## 1. Introduction  
+
+The automotive industry is undergoing a significant transformation, primarily due to the global focus on sustainability and environmental conservation. Electric vehicles (EVs) are leading this shift, playing a key role in mitigating environmental challenges and advancing sustainable transportation solutions [1, 2]. Concurrently, the emergence of hybrid energy systems (HES) within the EV powertrain represents an emerging trend, offering superior solutions over single energy systems [3]. By integrating multiple energy sources such as batteries, fuel cells, and internal combustion engines, HES improves overall efficiency, sustainability, and reliability, while also providing adaptability to a wide range of driving conditions [4]. Propelled by rapid technological advancements and supportive policies, EVs equipped with HES, including Hybrid EVs (HEV), Plug- in hybrid EVs (PHEV), and Fuel cell EVs (FCEV), are gaining traction worldwide [3]. For example, BYD, a prominent EV manufacturer, achieved sales of 3.02 million EVs in 2023, and PHEVs represent \(47.9\%\) of total sales. In the first quarter of 2024, PHEVs accounted for a notable \(51.6\%\) of total sales, indicating their growing popularity. This trend is mainly attributed to advancements in HES energy management, which enhance energy efficiency and overall performance.  
+
+The energy management strategy (EMS), responsible for allocating energy flow among HES to achieve predefined objectives, performs several vital functions essential for EVs:(1) EMS optimizes system efficiency by intelligently allocating energy flow based on driving conditions, reducing energy consumption and extending driving range [5]. (2) EMS optimizes power delivery based on driver demand, improving acceleration and responsiveness, while also ensuring smooth power delivery for improved driving experience. (3) By considering the characteristics of different power sources, EMS extends the lifespan of the HES, thus enhancing system reliability and safety [6]. Early EMS solutions utilize various rule- based approaches to achieve energy- saving objectives. Rule- based EMS involves the design of predefined rules and parameters tailored to specific driving conditions and vehicle characteristics, relying on
+
+<--- Page Split --->
+
+expert knowledge with iterative refinement based on testing feedback [7]. However, this process can be labor- intensive and time- consuming, requiring manual expertise for rule creation and extensive experimentation for parameter calibration. Moreover, its static nature and inability to adapt to dynamic driving scenarios limit its effectiveness in maximizing energy savings and overall performance.  
+
+Recently, there has been a surge in advanced algorithms for energy management of HES, including dynamic programming (DP), model predictive control (MPC), reinforcement learning (RL), and others. A common feature across these methods is the use of optimal control or machine learning (ML) algorithms to compute optimal EMS based on high- fidelity simulators. Collectively, these approaches are referred to as simulation- based EMS in our study. Although DP and MPC provide elegant solutions based on future driving data, accurately forecasting future driving conditions using historical speed and dynamic traffic information remains a significant challenge [8]. Additionally, prediction models and optimization algorithms often result in considerable computational complexity, leading to suboptimal real- time performance [3]. Consequently, applying optimal control methods in real- world vehicle settings poses considerable challenges at present. In contrast, learning- based models, exemplified by Deep RL (DRL), do not rely on knowledge of future conditions and present promising alternatives for EMS [9]. DRL involves learning optimal actions in an environment through trial and error, where the DRL agent interacts with the environment to maximize cumulative rewards over time [10]. DRL- based EMS has been a subject of active research in recent years, with recent developments in online DRL algorithms such as Deep Deterministic Policy Gradient (DDPG) [11], Soft Actor- Critic (SAC) [12], Proximal Policy Optimization (PPO) [13], and others, leading to remarkable results. However, the application of DRL- based EMS faces challenges as an online learning paradigm, particularly concerning the interaction between the DRL agent and the EV simulator, which raises safety concerns. Moreover, while existing literature assumes that simulation models accurately represent real- world conditions, the construction of high- fidelity models that encompass vehicle dynamics, powertrain, traffic scenarios, and driver behavior [14] remains a challenge [15]. This discrepancy can cause the "sim- to- real" problem, where EMS learned in simulators may not be effectively transferred to real vehicles, which leads to the complexity of EMS development [16].  
+
+In recent years, data- driven methodologies, propelled by advancements in ML techniques and the availability of large datasets, have become essential in addressing key challenges in the EV field [17]. With support from data collection platforms and open- access laboratory data, these data- driven approaches have revolutionized various aspects of battery management systems (BMS). This includes automatic discovery of complex battery aging mechanisms [18], prediction of battery safety envelopes [19], evaluation of safety conditions [20], estimation of battery state of health [21], and even enhancing battery lifetime prediction models with unlabeled data [22]. Notably, innovations in feature extraction and supervised ML techniques tailored for time- series data have greatly enhanced prediction accuracy. This success has sparked interest in exploring data- driven methods for sequential decision- making tasks, including improving energy management systems. A common approach to implementing a data- driven EMS involves using supervised learning, where the ML model is employed to capture the complex and non- linear relationships between input features and corresponding control outputs. In [23], recurrent neural networks are trained offline using a substantial amount of training data obtained from the global optimization strategy DP, yielding a sub- optimal EMS that closely approximates the DP. It is essential to recognize the differences from the prediction tasks in BMS applications, as EMS entails sequential decision- making. Although supervised learning can mimic the policy of EMS through imitation learning, its heavy reliance on expert data may result in limited generalization to new and diverse scenarios. Hence, it's crucial to explore alternative methods for learning EMS from non- expert data, which is common in automotive applications.  
+
+In this paper, we present a novel data- driven EMS paradigm, addressing key challenges discussed above by leveraging offline data generated by existing EMS strategies, without the need for explicit rule design or online interaction. Our approach integrates DRL and supervised learning techniques, departing from traditional rule- based EMS and simulation- based methods to enhance EMS capabilities, as visually depicted in Figure 1(a). Notably, our proposed method can be directly integrated with any EMS algorithm and continuously improves through the collection data. We term this approach the offline reinforcement learning (ORL) agent [24], which is trained by combining real- world data as well as carefully filtered and processed simulation data through a novel exploration scheme. To evaluate this paradigm, we focus on a fuel cell electric vehicle (FCEV) as the subject and collect data for learning power allocation strategies. The EV monitoring and management platform in China offers real- world data to validate our methodology (Figure 1(b)). Overall, the ORL agent introduced herein exhibits four key features (Figure 1(c)):
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1: Overview of the proposed data-driven EMS framework. (a) Comparison of three EMS paradigms: Traditional rule-based EMS relies on expert knowledge and calibration based on fixed driving cycles. Simulation-based EMS requires high-precision models and entails transitioning from simulation analysis to real-world deployment, resulting in a gap between simulated and real-world performance (sim-to-real gap). In contrast, the proposed real-world data-driven EMS learns directly from actual data. (b) China has established a three-tier EV monitoring and management system involving the state, local governments, and enterprises. The National Monitoring and Management Platform collects real-time operational data from over 20 million EVs [25]. (c) The ORL agent works with an existing EMS and continuously collects EV data to improve the EMS. (d) Overall diagram of the proposed data-driven EMS methodology. </center>  
+
+1) Purely data-driven EMS: By autonomously learning and optimizing from collected offline datasets, our approach facilitates the development of advanced EMS without necessitating expert knowledge or the construction of high-fidelity EV simulators. This data-driven process significantly simplifies EMS development workflows.  
+
+2) Learning from non-optimal data: Our research indicates that ORL can learn near-optimal EMS from non-optimal data and even can derive superior policies from sub-optimal. Our approach is less reliant on data quality allowing for effective learning. This practicality allows us to utilize raw data generated by actual vehicles, which is common in automotive applications.  
+
+3) Enhancement with increased data: The performance of ORL improves as more training data is utilized, demonstrating its ability to continuously adapt and enhance EMS performance. Through training across diverse
+
+<--- Page Split --->
+
+datasets, it can potentially adapt to new driving conditions and yield favorable results, even in corner- case conditions. When sufficient data are available, ORL can learn a generalized EMS.  
+
+4) Compatibility with existing EMS: Our approach seamlessly integrates with established rule-based or simulation-based EMS methods, leveraging data from onboard controllers to augment EMS performance. This ensures that baseline performance is preserved while facilitating further enhancement using ORL, making it a valuable extension to conventional EMS methodologies.  
+
+## 2. Results  
+
+### 2.1. The overview of data-driven EMS  
+
+In Figure 1(d), the framework overview of ORL for EMS is illustrated. We introduce the three phases of applying the proposed ORL algorithm to the EMS problem, which include data collection, offline learning, and evaluation. ORL is an innovative subset of RL methods rooted in data- driven approaches. Unlike most simulation- based EMS scheduling methods, the data- based learning process does not require the building of an EV simulation environment.  
+
+In the data collection phase, an available dataset \((D)\) is gathered from the existing EMS policy \((\pi_{\beta})\) . It is crucial to record various parameters such as velocity, acceleration, hydrogen consumption, electricity consumption, fuel cell degradation, battery state of charge (SOC), and others during vehicle operation. However, due to limitations in real- world data quality and privacy concerns associated with collecting data from a large number of actual vehicles, we developed an FCEV simulation model to generate laboratory datasets 2(a). This FCEV model enables the generation of extensive, high- fidelity data based on real- world driving conditions 2(b). Subsequently, we employ a tailored data processing model to encode control and state variables into standardized data formats. The data encoding process converts these parameters into a transition dataset \(D = \left(s,r,a,s^{\prime}\right)_{i}\) . Here, \(s\) , \(a\) , and \(r\) represent the state, action, and reward as described in the Methods section, respectively, while \(s^{\prime}\) denotes the subsequent state, and \(i\) indexes a transition in the dataset. These meticulously curated datasets are stored in the buffer and serve as input for the ORL agent.  
+
+In the offline learning phase, our proposed ORL agent is designed based on the Actor- Critic network and incorporates Behavior Cloning (BC) and Discriminator Blend (DB) mechanisms. In each training step, mini- batches of transitions \((s,r,a,s^{\prime})\) stored in a data buffer are sampled from the replay buffer to update the ORL networks. The Actor network, responsible for determining the action \(\mathbf{a}_{t}\sim \pi_{\beta}\left(\cdot \mid \mathbf{s}_{t}\right)\) , is updated to maximize the expected return as estimated by the Critic network. BC regularization is applied to the policy update step to encourage the policy to prioritize actions present in the dataset. Moreover, the DB component is introduced to allow actions beyond the dataset distribution similar to those included in the dataset. Through training, the agent learns to optimize its actions based on observed states and rewards. The details of the ORL agent will be elaborated upon in the Methods section in detail.  
+
+After completing the training phase, the neural network parameters representing the EMS policy of the ORL agent are saved for future use. Subsequently, the trained agent is evaluated to gauge its effectiveness and performance. Utilizing the FCEV environment established during the data collection phase, our subsequent experiments employ three standard driving cycles (WTVC, FTP, and CHTC) to evaluate energy cost. Additionally, we incorporate various real- world driving scenarios to comprehensively assess the trained EMS policy. Following the evaluation, adjustments to the agent hyperparameters or training process may be considered to improve its performance. This iterative process of training, evaluation, and refinement continues until the desired level of EMS performance is attained. Upon achieving satisfactory performance, the trained agent becomes eligible for deployment in real- world scenarios, where it can be utilized to efficiently optimize energy management systems.  
+
+### 2.2. Data for learning and analysis  
+
+We select the Proximal Policy Optimization (PPO), which demonstrates the best performance among online DRL algorithms in our EMS problem, as the expert EMS. Using PPO, we generate datasets denoted as \(D^{E}\) comprising 300K time steps. Additionally, we employ a random agent that samples actions randomly, generating datasets denoted as \(D^{R}\) , which represent poor performance. To create settings with varying levels of data quality in the suboptimal offline dataset, we combine transitions from the expert datasets \(D^{E}\) and the random datasets \(D^{R}\) in different ratios. Specifically, we consider four different dataset compositions, denoted as D1, D2, D3, and D4. These settings are defined as follows: D1 (Data- 1): Consists solely of transitions from the expert dataset \(D^{E}\) , representing the expert policy. D2 (Data- 2): Contains two- thirds of transitions from the expert dataset \(D^{E}\) and one- third of transitions from the random dataset \(D^{R}\) , representing suboptimal data. D3 (Data- 3): Comprises one- third of transitions from the expert dataset
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2: Comparison for different datasets. (a) The simulation model for the fuel cell hybrid electric vehicle (FCEV), depicts the powertrain layout and energy flow topology. (b) The principle of EMS involves the allocation of energy flow among hybrid energy systems (fuel cell and battery) to achieve predefined objectives based on driving conditions. Here, real driving conditions are gathered and then use the simulated FCEV model to generate energy cost data for training the ORL agent. (c) Distribution of encoded actions for the four datasets, each action is normalized to the [-1,1]. D1 represents data generated by an expert policy, D2 and D3 denote suboptimal data generated by a combination of expert and random policies, D4 comprises entirely random data. (d) The states distribution of four datasets. </center>  
+
+\(D^{E}\) and two- thirds of transitions from the random dataset \(D^{R}\) , representing another suboptimal data. D4 (Data- 4): Comprises solely of transitions from the random dataset \(D^{R}\) , representing the random policy.  
+
+Figure 2(a) depicts the action distributions for the four datasets. Significant differences can be observed among the four EMS policies, with the action range of D1 falling within (- 0.2, 0.5), resulting in relatively stable variations in FC power. As random policy data is introduced, the action ranges in the other datasets all fall within (- 1, 1), particularly for D4 where the action distribution is uniformly spread across (- 1, 1). This implies that this policy is noisy, denoted as a poor EMS. Figure 2(b) illustrates the state distributions for the four datasets. Note that all states have undergone post- processing and scaling to the (0, 1) interval. By comparing the box plots of the four datasets, it is evident that the SOC of D1 falls within a reasonable range (0.38- 0.7), meeting the EMS constraints regarding the battery SOC.
+
+<--- Page Split --->
+
+However, the SOC of the other datasets falls into unreasonable ranges, such as SOC in the range (0.2, 1) for D3. Additionally, with the increase in \(D^R\) data, the FC power distribution ranges of D3 and D4 become wider. Since the conditions of the four datasets are derived from fixed segments of standard driving cycles, the velocity distribution remains the same across all datasets.  
+
+Creating challenging datasets is practical because generating sub- optimal or random data is more cost- effective than collecting expert- level data from real vehicles. Therefore, an effective data- driven EMS method must be able to effectively handle and learn from these suboptimal offline datasets.  
+
+### 2.3. Learning superior EMS from non-optimal data  
+
+![](images/Figure_3.jpg)
+
+<center>Figure 3: The ORL agent learning performance. (a) Learning curves of ORL agent for the four different datasets. (b) The comparison of absolute rewards (original rewards are negative) under three validation conditions. "D1," "D2," "D3," and "D4" correspond to the best rewards achieved by ORL after learning on each respective dataset. Notably, ORL agent on various datasets closely approximate or exceed the expert policy. (c) The comparison of energy costs between the original EMS and the optimized EMS using ORL shows a notable reduction in energy costs through the data-driven learning process. (d) The action (FC power slope) distributions of optimized EMS using ORL </center>  
+
+We first study the performance of the ORL agent with different datasets. To ensure a fair comparison, the algorithm employs uniform experimental settings and network parameters across four datasets. Figure 3(a) illustrates the average reward of the training process on the four datasets. This average, computed as the mean reward over every 1000 episodes, undergoes validation across 10 iterations using three standard driving cycles: WTVC, CHTC, and FTP.
+
+<--- Page Split --->
+
+Training involves utilizing a buffer comprising 300,000 samples, with the ORL agent randomly selecting 256 data points for each training iteration, accumulating to one million training epochs. For D1, which comprises exclusively expert data, convergence is evident after approximately 210e3 episode. However, ORL exhibits a slower convergence speed during iterative learning on the other three datasets, converging at around 330e3, 600e3, and 360e3 epochs, respectively. This suggests that the data distribution influences the learning speed, but ultimately, ORL learns an effective EMS.  
+
+Figure 3(b) presents the reward performance of trained ORL across three driving cycles. It notes that the absolute reward value of ORL decreases in D1, from 323.4 to 297.3 on the CHTC, representing an improvement of \(8.8\%\) . Surprisingly, even on suboptimal datasets D2 and D3, ORL outperforms the expert strategy, with reward increases of \(1.8\%\) and \(3.4\%\) , respectively. Similarly, ORL still outperforms on the WTVC and FTP conditions, learning superior strategies from the suboptimal datasets D2 and D3. The exception is D3 on the WTVC condition, possibly due to the high- speed nature of the WTVC condition, leading to larger reward values for SOC. Despite this, the final results of the energy consumption remain within rational bounds. Particularly noteworthy is the excellent performance on the random dataset D4, where ORL closely approaches expert results across all three validation conditions, achieving rewards of 402, 325, and 395. Compared to the original average reward of 2637 for the D4 dataset, ORL has reduced the reward by \(85.8\%\) .  
+
+Figure 3(c) provides a detailed comparison of the energy costs between ORL and the original datasets. The blue dots represent the mean costs of the four original EMS datasets, accompanied by error bars indicating the maximum and minimum costs. In contrast, the red dots depict the energy costs incurred by ORL in the corresponding datasets. Despite the inclusion of random data that leads to a degradation in cost performance, ORL consistently maintains lower costs across all data sets. For instance, in the WTVC condition, the cost escalates from the initial 90 RMB in dataset D1 to 163 RMB in dataset D4. However, ORL consistently maintains costs within the range of 90- 95 RMB. In particular, the minimum cost values of the original D4 dataset exceed those achieved by ORL significantly, with ORL achieving reductions in costs that exceed \(40\%\) in all three conditions. This observation underscores ORL's ability not only to glean superior results from expert EMS but also to consistently yield excellent outcomes from progressively suboptimal datasets. Remarkably, ORL even attains expert- level EMS performance when trained solely on noisy datasets.  
+
+To elucidate the rationale behind the performance enhancements, Figure 3(d) illustrates the action distributions of optimized EMS using ORL. As different EMS policies can be reflected by the actions taken, in the context of the FCEV considered here, this pertains to the FC power slope under the same driving cycle. Comparing Figures 2(c) and 3(d), we notice significant changes in D2, D3, and D4 compared to Figure 2. In D2, D3, and D4, the action distributions closely resemble those of expert data in D1, concentrating within the range of [-0.3, 0.3], as opposed to the wider range of [-1, 1] observed in Figure 2. The change is particularly pronounced in D4, where the absence of expert data results in slight differences in the action distributions compared to D1, D2 and D3. However, all ORL policies tend to learn FC power variations with smaller ranges, ensuring smoother FC power output while satisfying power demand requirements.  
+
+In conclusion, through experimentation on three validation conditions and four datasets, our ORL agent not only learns better EMS strategies from expert strategies but also demonstrates the ability to learn near- expert strategies from entirely noisy datasets and even achieves superior results from datasets containing a mixture of expert and noisy data. This observed convergence underscores ORL's ability to enhance and optimize the original EMS through learning from data obtained from any EMS.  
+
+### 2.4. Performance by comparative evaluation  
+
+To demonstrate the superior performance of ORL, we contrast it with simulation- based and imitation learning EMS approaches. Since imitation learning and ORL are closely related, both involve learning EMS from data. We first compare the performance between ORL and BC. It's important to note that BC typically employs a supervised learning paradigm, learning from expert data, while ORL incorporates reinforcement learning with exploration mechanisms. This distinctive learning mechanism results in significant performance differences.  
+
+In Figure 4(a), we compare the testing reward in the WTVC, CHTC and FTP driving cycles, and calculate the percentage of ORL and BC costs relative to the expert EMS (PPO). In D1, both ORL and BC achieve favorable results, with ORL surpassing the original expert data by a maximum of \(9.1\%\) , while BC remains comparable to the expert. In D2, ORL maintains superiority over expert- based EMS, while BC experiences significant cost degradation (ranging from \(6\%\) to \(70\%\) ). In D3 and D4, ORL continues to outperform or closely match the expert, while BC, limited by data
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4: Performance analysis comparing different methods. (a) The comparison between two data-driven EMS methods. The matrix numbers represent the relative reward rates of ORL and BC compared to the expert EMS (PPO) under the same conditions, emphasizing the minimal influence of data quality on ORL's performance. (b) Comprehensive performance of different algorithms on the WTVC condition, DP representing the globally optimal EMS. (c) Comprehensive performance on the CHTC condition. (d) Comprehensive performance on the FTP condition. (e) The efficiency distribution of fuel cell power demonstrates that ORL learns a superior EMS, ensuring that fuel cell system remains within the high-efficiency range. (f) FC degradation costs under three conditions </center>  
+
+quality, fails to learn an optimal EMS. This underscores ORL's capability to learn superior EMS from non- expert data, while imitation learning demonstrates poorer performance and struggles to learn favorable EMS with non- expert data.  
+
+Figures 4(b- d) present detailed results of different methods under the WTVC, CHTC, and FTP conditions, with the red lines representing the percentage of cost compared to DP. It is evident that ORL learns an optimal EMS policy on the D1 dataset, achieving percentages close to \(99.9\%\) , \(99.4\%\) , and \(97.6\%\) of DP, respectively. As for PPO, a benchmark expert policy, its cost results are \(98.6\%\) , \(98.0\%\) , and \(97.6\%\) of DP, respectively. Thus, BC learns similar expert EMS in D1, but its performance significantly deteriorates on suboptimal D2 data. Another online DRL method, TD3, also demonstrates satisfactory performance, however, its overall costs are lower than those of PPO and ORL.  
+
+In Figure 4(e), the FC power distribution of the EMS learned by ORL on the D1 dataset is depicted, with the green curve representing the efficiency curve of the FC system. ORL demonstrates a superior EMS, ensuring that the
+
+<--- Page Split --->
+
+power distribution remains within the high- efficiency range, resulting in reduced hydrogen consumption. Additionally, a narrower power variation range, as shown in Figure 3(d), minimizes FC degradation costs. As illustrated in Figure 3(f), ORL incurs minimal FC degradation costs in the three conditions, with costs of 2.3, 1.4, and 1.4, respectively. Furthermore, examination of Figures 3(b- d) indicates that the battery SOC remains within a reasonable range. These findings collectively affirm that ORL has successfully learned a superior EMS.  
+
+### 2.5. Continuous learning with growing data  
+
+We have demonstrated in previous experiments that ORL can learn optimal EMS strategies from data and outperforms other methods. In this section, we further showcase ORL for continuous learning from data. We conduct experiments in three cases depicted in Figure 5(a), collecting real- vehicle data in different driving scenarios including urban road, highway, and downtown road for the three cases (Figure 5(b)).  
+
+#### 2.5.1. Case 1: Continuous learning from historical data  
+
+Take the example of driving a bus on fixed routes to illustrate Case 1. We collected real electric bus driving data in Zhengzhou, China, over three consecutive days. Figure 5(c) shows the speed trajectories of three days, labeled as ZBDC- No1, ZBDC- No2, and ZBDC- No3, respectively. Noticeably, there are variations in the speed trajectories along the same route over different days. Figures 5(d- f) show the total cost of different EMS strategies in the three scenarios, which include hydrogen consumption, battery cost, and fuel cell degradation. The baseline is the original EMS of FCEV, and using the baseline data under ZBDC- No1 driving cycle to train the ORL agent as the ORL(Z1) strategy, which is then applied to the new condition ZBDC- No2. Furthermore, we train a new ORL(Z2) EMS using data from both the baseline on ZBDC- No1 and the ORL(Z1) on ZBDC- No2, which is then validated on ZBDC- No3.  
+
+It can be observed that the baseline EMS performs poorly on the first day (ZBDC- No1), with a cost only \(88.0\%\) compared to DP. The corresponding FC power and power slope distributions are shown in Figures 5(g). On the second day as depicted in Figure 5(e) and (h), the performance of ORL(Z1) achieves a significant improvement by learning from the previous data, reaching \(96.4\%\) of the cost compared to DP on the ZBDC- No2. On the third day, continuously learning from more data, the ORL(Z2) achieves a cost of \(98.6\%\) compared to DP on the ZBDC- No3, as shown in Figure 5(f) and (i). By comparing the three power distribution, it is evident that ORL(Z1) and ORL(Z2) distribute more FC output power in the high- efficient range, resulting in lower overall energy consumption, and the smaller power slope also leads to lower system degradation costs.  
+
+In conclusion, with data updates, new data can be utilized to train ORL, leading to the evolution of batter EMS strategies. This demonstrates the ability of ORL to continuously learn and improve from historical data. Additionally, our method integrates seamlessly with established EMS methods, using real- time data from onboard controllers to increase EMS performance. This ensures that baseline performance is preserved while facilitating further enhancement using ORL, making it a valuable extension to conventional EMS methodologies.  
+
+#### 2.5.2. Case 2: Improving from simulated data  
+
+In Case 2, we address the challenge of improving EMS performance from simulated data, where historical data is absent and driving conditions are unknown. Despite the ability of simulation- based methods (such as DRL) to derive ideal EMS strategies from simulated EV models, deploying these strategies onto real vehicles often leads to performance degradation due to the stochastic and unknown nature of real- world driving conditions, a problem known as the sim- to- real gap, extensively studied in the fields of RLL.  
+
+As illustrated in Figure 6(a), during the simulation phase, the PPO algorithm is trained on the standardized driving cycle (WTVC) and three specific conditions ZBDC (Figure 5(c)) to obtain an ideal EMS, denoted as PPO (Train). Subsequently, the EMS is validated on 12 local driving conditions denoted as PPO (Test). As depicted in Figure 6(b), the cost difference between PPO (Train) and DP across the four training conditions is minimal, with an average difference of \(3.16\%\) . However, when tested on the 12 new conditions(DC- 1 to DC- 12), the average cost difference between PPO (Test) and DP rises to \(12.75\%\) . This indicates a significant performance degradation of DRL- based methods when transitioning from simulation to real- world conditions.  
+
+To mitigate the sim- to- real problem, our proposed ORL method leverages data from PPO (Test) for further learning. As shown in Figure 6(c), the ORL approach achieves a substantially lower cost across the 12 local operating conditions compared to the PPO (Train) strategy. The average cost difference between ORL and DP is merely \(1.42\%\) (Figure 6(b)). In summary, our experiments demonstrate that ORL can effectively learn from simulated data to enhance the performance of the original EMS, addressing the sim- to- real problem inherent in traditional simulation- based methods.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 5: ORL continuously learning and improving from data. (a) ORL for continuous learning from data in three scenarios. (b) Driving data collected from the real-world route. (c) Speed trajectories of three ZBDC conditions. (d) Comparing the total cost under ZBDC-No1, the total cost comprises hydrogen consumption, battery cost, and cell degradation cost. (e) Comparing the total cost under ZBDC-No2. (f) Comparing the total cost under ZBDC-No3. (g) Fuel cell power distribution cloud chart of Baseline EMS. (h) Fuel cell power distribution cloud chart following one data update. (i) Fuel cell power distribution cloud chart following two data updates. </center>
+
+<--- Page Split --->
+![](images/Figure_6.jpg)
+
+<center>Figure 6: Performance with more data. (a) Performance of PPO for 4 training data and 12 testing data. (b) The comprehensive performance comparison of different EMS methods reveals that ORL can significantly mitigate the performance degradation observed in the testing phase of PPO. (c) Performance of ORL for 12 testing data. (d) Speed distribution of four conditions. (e) The demand power of HWDC represents an extreme condition. The red dashed line indicates the maximum output power of FC system. (f) Overall performance of the 4 cases as the training data increase under test driving cycle CLTC; (g) Performance on GCDC; (h) Performance on ZNDC; (I) Performance on HWDC indicates that ORL ultimately learns a reasonable EMS in extreme driving conditions. (j) Battery SOC trajectories of different EMS under the test driving cycle HWDC. </center>
+
+<--- Page Split --->
+
+#### 2.5.3. Case 3: Learning a general EMS with large-scale data  
+
+To assess the generalization of the ORL model, especially in extreme conditions, the ORL agent trained on huge amounts of data and tested its performance on four new driving conditions. Figure 6(d) illustrates the speed distribution for these four conditions. Among these, the China Light- Duty Vehicle Test Cycle (CLTC) stands as the standard cycle, GCDC is obtained from an EV operating in a downtown road, ZBDC represents a new condition collected in Zhengzhou, and HWDC is collected from a fuel vehicle traveling on a highway. The four conditions have distinct characteristics and originate from different road and vehicle types. Particularly for HWDC, where the demand power exceeds \(250\mathrm{kW}\) . As depicted in Figure 6(e), the power demand exceeds the \(100\mathrm{kW}\) maximum output power of FC system, indicating that HWDC is an extreme condition for the FCEV studied in this work.  
+
+Firstly, we establish datasets of four different scales: 4e4 (ORL- 4), 20e4 (ORL- 20), 100e4 (ORL- 100), and 500e4 (ORL- 500) samples, excluding data from the four validation conditions obtained in previous experiments (Cases 1 and 2). The results on the four validation conditions are depicted in Figures 6(f- i). It is evident from Figure 6(f), (g), and (h) that both the reward and cost exhibit a gradual decrease with an increase in training data. With more training data, the ORL model consistently enhances its performance. Notably, the rate of performance improvement diminishes after reaching the 100e4 sample mark, with minimal disparity observed between the ORL- 100 and ORL- 500. At approximately 20e4 samples, the ORL model outperforms the DRL- TD3 algorithm (as indicated by the dashed lines in the figures). It is essential to highlight that, under the extreme HWDC condition, although ORL- 20 achieves the lowest cost (Figure 6(i)), its reward absolute value is not the lowest. This phenomenon is because the ORL- 20 prefers battery consumption during high power demands, exceeding the maximum output of the FC system, violating the SOC constraint. This phenomenon arises because ORL- 20 tends to prioritize battery consumption during periods of high power demand, the FC system fails to achieve its maximum power output. Consequently, the strategy fails to maintain the SOC within the desired range, rendering it ineffective as an EMS. Conversely, ORL- 500 demonstrates superior performance, with lower reward and cost, and SOC maintained within a reasonable range, as illustrated in Figure 6(j). Overall, after learning from 5 million data points (equivalent to over 60 million kilometers), the ORL agent can learn a general EMS adaptable to unseen and even corner- case conditions.  
+
+This result highlights two advantages of the ORL agent: First, its performance surpasses that of the original policy; second, it demonstrates that with increased data availability, learning performance improves. The ORL model can learn a general EMS from a large amount of EV data.  
+
+## 3. Methods  
+
+### 3.1. EV Environment  
+
+In this work, we evaluate EMS performance using a fuel cell hybrid electric vehicle (FCEV) within a simulation environment. Figure. 1(e) illustrates the schematic diagram of the FCEV and its components, which include a fuel cell (FC) system, a hydrogen storage tank, an electric motor (EM), and a Lithium- ion battery (LIB) pack. The FC stack serves as the primary power source to meet the energy requirements of the vehicle. The diagram also depicts the energy flow from the hydrogen storage tank to the motor. The FC system converts hydrogen energy into electricity, which then collaborates with the LIB through the high- voltage bus. This electric energy is subsequently utilized to power a single electric motor, connected to the driving wheel via a fixed- ratio final gear. According to the vehicle driving resistance equation, the driving power demand is determined by the speed and acceleration of the FCEV, and can be expressed as follows:  
+
+\[P_{d} = \frac{1}{3.6\cdot\eta_{\mathrm{me}}}\left(m g C_{f}\upsilon_{t} + m\delta \upsilon_{t}a_{t} + m g s i n(i) + \frac{C_{D}A}{21.15}\upsilon_{t}^{3}\right) \quad (1)\]  
+
+where \(\eta_{\mathrm{me}}\) is the efficiency of the vehicle drivetrain; \(m\) is the vehicle mass; \(g\) is the gravitational constant; \(C_{f}\) is the rolling resistance coefficient; \(\delta\) is the rotational mass conversion coefficient; \(C_{D}\) is the air resistance coefficient, \(A\) is the frontal area; \(\upsilon_{t}\) is the longitudinal velocity at the time step \(t\) , \(a_{t}\) is the acceleration, \(i\) is the angle of slope of the road. The power demand is provided by the FC system and the battery pack, the power balance of the FCEV can be formulated as:  
+
+\[P_{d} = \left(P_{f c}\cdot \eta_{D C / D C} + P_{b a t}\right)\cdot \eta_{D C / A C}\cdot \eta_{E M} \quad (2)\]  
+
+where \(P_{f c}\) and \(P_{b a t}\) respectively denote the output power of the fuel cell system and the LIB pack; \(\eta_{D C / D C},\eta_{D C / A C}\) and \(\eta_{\mathrm{EM}}\) represent the efficiency of the DC/DC converter, DC/AC inverter, and the electric motor, respectively. The
+
+<--- Page Split --->
+
+battery pack is modeled using an equivalent circuit model in Equation (3):  
+
+\[\left\{ \begin{array}{l l}{P_{b a t}(t) = V_{o c}(t) - R_{0}\cdot I^{2}(t)}\\ {I(t) = \frac{V_{o c}(t) - \sqrt{V_{o c}^{2}(t) - 4\cdot R_{0}\cdot P_{b a t}(t)}}{2R_{0}}}\\ {S O C(t) = \frac{Q_{0} - \int_{0}^{t}I(t)dt}{Q}} \end{array} \right. \quad (3)\]  
+
+where \(S O C\) is the battery state of charge, \(V_{o c}\) is the open- circuit voltage, \(I_{t}\) is the current at time \(t\) , \(R_{0}\) is the internal resistance, \(P_{b a t}\) is the output power in the charge- discharge cycles, \(Q_{0}\) is the initial battery capacity, \(Q\) is the nominal battery capacity.  
+
+According to the battery aging model in [12]. The degradation rate of battery operation \(\gamma_{\mathrm{bat}}\) is affected by the charge/discharge rate ( \(C_{rate}\) ). The relationship between the battery aging correction factor and \(C_{rate}\) can be fitted from experiment data:  
+
+\[\gamma_{\mathrm{bat}} = \mu_{1}\left|C_{\mathrm{rate}}\right|^{2} + \mu_{2}\left|C_{\mathrm{rate}}\right| + \mu_{3} \quad (4)\]  
+
+where \(\mu_{1}, \mu_{2}, \mu_{3}\) are the curve- fitting coefficients. LIB can operate for about 5000 full cycles in a lifetime. The battery degradation cost \(C_{bat,degr}\) can be calculated by:  
+
+\[C_{bat,degr} = \int_{0}^{t}\gamma_{bat}^{-1}P_{bat}dt\cdot PR_{bat} / (5000\cdot 3600) \quad (5)\]  
+
+where \(PR_{\mathrm{bat}}\) is the battery price per kWh that is 1500RMB/kWh.  
+
+The fuel cell system efficiency under different power conditions is obtained from experiment data. Thus, the mass flow rate of the hydrogen consumption can be calculated by:  
+
+\[\dot{m}_{H_{2}} = P_{f c s} / \left(\eta_{f c s}\cdot \mathrm{LHV}_{H_{2}}\right) \quad (6)\]  
+
+where \(\eta_{f c s}\) is the fuel cell system efficiency; \(P_{f c s}\) is the fuel cell system output power; \(\mathrm{LHV}_{H_{2}}\) is the hydrogen low calorific value. The fuel cell hydrogen cost can be calculated by:  
+
+\[C_{f c s,H_{2}} = PR_{H_{2}}\cdot \int_{0}^{t}\dot{m}_{H_{2}}dt \quad (7)\]  
+
+where \(PR_{H_{2}}\) is the hydrogen price per kilogram(60RMB/kg).  
+
+The fuel cell degrades rapidly under four typical conditions: load changing, start/stop, low power, and high power conditions. We assume that the fuel cell system continues operating until the vehicle power system is shut down, thus the start/stop condition is not considered in the EMS. The fuel cell voltage degradation rate, denoted as \(\gamma_{\mathrm{fcs}}\) , can be calculated by:  
+
+\[\gamma_{\mathrm{fcs}} = \kappa_{\mathrm{low}}\cdot T_{\mathrm{low}} + \kappa_{\mathrm{high}}\cdot T_{\mathrm{high}} + \kappa_{\mathrm{cha}}\cdot \Delta P_{\mathrm{fcs}} \quad (8)\]  
+
+where \(\kappa_{\mathrm{low}}\) is the degradation rate under low power condition; \(T_{\mathrm{low}}\) is the low power condition duration; \(\kappa_{\mathrm{high}}\) is the degradation rate under high power condition; \(T_{\mathrm{high}}\) is the high power condition duration; \(\kappa_{cha}\) is the degradation rate under load changing condition; \(\Delta P_{f c s}\) is the fuel cell power slope. Fuel cell is considered to reach the end of its life when \(10\%\) of voltage at rated power has been lost. The fuel cell operation degradation cost can be calculated by:  
+
+\[C_{f c s,degr} = k_{f c s}\cdot \gamma_{f c s}\cdot P_{f c s,\mathrm{rate}}\cdot PR_{f c s} / \left(V_{f c s,\mathrm{end}}\cdot 1000\right) \quad (9)\]  
+
+where \(k_{f c s}\) is the fuel cell life correction factor; \(V_{f c s,\mathrm{end}}\) is fuel cell voltage drop at the end- of- life; \(P_{f c}\) , rate is the rated power of the fuel cell; \(PR_{f c s}\) is the fuel cell price per kilowatt(4000RMB/kW).
+
+<--- Page Split --->
+
+### 3.2. Problem modeling  
+
+In this work, the EMS of electric vehicles is modeled as a long- term sequential decision process objective to minimize total energy cost while maintaining battery SOC within reasonable limits. The optimization objective can be formulated as:  
+
+\[J_{EMS} = \min \sum_{t = 0}^{T} cost(t) + \alpha f_{s}(SOC(t)) \quad (10)\]  
+
+where \(T\) is the total length of the driving cycle, \(cost(t)\) is the energy cost including hydrogen consumption, the cost of battery, and fuel cell degradation, \(f_{s}(SOC(t))\) is the SOC maintaining function, \(\alpha\) the tradeoff between energy cost and SOC.  
+
+To tackle the sequential decision, the energy management problem is formulated as a Markov Decision Process (MDP), which is a framework for learning the optimal EMS from interaction to minimize total energy cost. The MDP defined by a tuple \((S,A,P,R,\rho_{0},\gamma)\) , where \(S\) denotes the state space, \(A\) denotes the action space, \(P\left(s^{\prime}\mid s,a\right)\) denotes the transition distribution, \(\rho_{0}(s)\) denotes the initial state distribution, \(R(s,a)\) denotes the reward function, and \(\gamma \in (0,1)\) denotes the discount factor. The goal is to find a policy \(\pi (a\mid s)\) that maximizes the expected cumulative discounted rewards \(J(\pi) = E_{\pi ,P,\rho_{0}}\left[\sum_{t = 0}^{\infty}\gamma^{t}R\left(s_{t},a_{t}\right)\right]\) . To use this formulation for FCEV. The state space at time point \(t\) for the FCEV is defined as:  
+
+\[S = \left\{v_{t},acc_{t},SOC_{t},P_{fcs}^{t}\right\} \quad (11)\]  
+
+where \(v_{t},a_{t},P_{fcs},SOC_{t}\) are the vehicle speed, acceleration, fuel cell power, and battery SOC. The action represents the control variable, defined as allocating power to the energy sources of the vehicle. In the context of the FCEV, the action is defined as the FC power slope, denoted as \(\Delta P_{fcs}\) . The continuous action can be described as follow:  
+
+\[A = \left\{\Delta P_{fcs} = P_{fcs}^{t} - P_{fcs}^{t - 1},\Delta P_{fcs}\in [-10\mathrm{kW},10\mathrm{kW}]\right\} \quad (12)\]  
+
+The reward function \(R\) describes the reward \(R(s_{t + 1};s_{t};a_{t})\) associated with transitioning from state \(s_{t}\) to state \(s_{t} + 1\) using action \(a_{t}\) . The design of the reward function is pivotal in the learning process. For the FCEV, multiple objectives are taken into account, including hydrogen consumption, FC degradation, and battery- related costs such as electricity consumption and degradation. Additionally, it is essential to maintain the battery SOC. Therefore, the reward function is defined as the sum of energy costs while ensuring that the battery charge- sustaining constraints are maintained.  
+
+\[R = -\left\{C_{fcs,H_2} + C_{fcs,degr} + C_{bat,eh_2} + C_{bat,degr} + \alpha \left[SOC_{ref} - SOC(t)\right]^2\right\} \quad (13)\]  
+
+The battery electricity consumption \(C_{bat,eh_2}\) is calculated according to the battery charge/discharge efficiency and converted into price cost:  
+
+\[C_{bat,eh_2} = \int_0^t \left[P_{bat} / \left(\eta_{d / c}\eta_{DC / DC},\mathrm{LHV}_{H_2}\right)\right]dt\cdot PR_{H_2} \quad (14)\]  
+
+where \(\eta_{d / c}\) is the battery discharge/charge efficiency.  
+
+### 3.3. Offline RL algorithm  
+
+We employ the offline reinforcement learning (ORL) paradigm to address the MDP problem described above. The goal is to learn a policy \(\pi \sim a_{t}(\pi :S\to A)\) that maximizes the expectation of the sum of discounted rewards \(J(\pi)\) Each policy \(\pi\) has a corresponding state- action value function (also known as Q function), which denotes the expected return \(Q(s,a)\) when following the policy \(\pi\) after taking an action \(a\) in state \(s\)  
+
+\[Q(s,a) = \mathbb{E}\left[\sum_{t = t}^{\infty}\gamma^{i - t}R_{i}\mid s_{t} = s,a_{t} = a\right] \quad (15)\]  
+
+Here, we approximate the Q function \(Q(s,a)\) using deep neural networks by minimizing the squared Bellman error. The ORL algorithm utilized in our proposal is an extension of the twin delayed deep deterministic policy gradient
+
+<--- Page Split --->
+
+algorithm (TD3) [26]. TD3 is a state- of- the- art online DRL algorithm implemented under the actor- critic framework that learns a deterministic policy. For the actor part, it learns a deterministic target policy by mapping states to a specific action. The update of the Actor network of TD3 algorithm aims to maximize the estimation of the current policy by the critic network.  
+
+When offline logged datasets are available, it is reasonable to push the policy towards favoring actions contained in the dataset D. Hence, our proposed ORL algorithm augments the standard policy update step in TD3 with Behavior Cloning (BC) regularization to reinforce the policy's focus on the behaviors observed in the dataset D [27]. This regularization term encourages the policy to mimic the demonstrated behaviors more closely, leading to improved generalization and performance. Furthermore, in pursuit of improving the policy, Discriminator Blend (DB) regularization [28] is employed to enhance the flexibility of the policy constraint. This is achieved by integrating a discriminator using Generative Adversarial Networks (GANs) [29]. By incorporating a discriminator, the policy is enabled to explore actions that may not be included in the dataset \(D\) , leading to a more diverse and adaptive policy. This involves utilizing a neural network as an approximator of the policy function \(\pi\) :  
+
+\[\pi = \arg \max_{\pi}\mathbb{E}_{(s,a)\sim \mathbf{D}}\left[\lambda Q(s,\pi (s)) - (1 - \beta)(\pi (s) - a)^2 +\beta \log (D(s,\pi (s)))\right] \quad (16)\]  
+
+\(\mathbb{E}()\) is the mathematical expectation. The parameter \(\beta\) (range of 0 to 1) adjusts the balance between BC and DB constraints. The DB is trained to assess whether a given action \(a\) and state \(s\) pair belongs to the dataset \(a_{D}\) or is generated by the policy \(\pi_{\theta}\) , which acts as the generator \(G\) in GANs. This enables BD to effectively regulate the policy learning process by encouraging the policy to explore actions beyond the dataset while ensuring that these actions are plausible according to the discriminator's perception. The \(\lambda\) is a normalization term based on the average absolute value of \(Q\) to control the balance between RL and imitation, defined as:  
+
+\[\lambda = \frac{\alpha}{\frac{1}{N}\sum_{(s_i,a_i)}\left|Q(s_i,a_i)\right|} \quad (17)\]  
+
+The parameter \(\alpha\) is used to control the strength of the regularize where the larger \(\alpha\) will make the algorithm approach more RL, and \(N\) represents the number of transitions in the dataset. To normalize the characteristics of each state in the provided dataset. Let \(s_i\) be the \(i\) th feature of the state \(s\) in the dataset, let \(\mu_i,\sigma_i\) be the mean and standard deviation ( \(\eta\) is a constant value to avoid division by zero.):  
+
+\[s_i = \frac{s_i - \mu_i}{\sigma_i + \eta} \quad (18)\]  
+
+The critic part estimates the Q- value of a state- action pair. TD3 employs two critic networks to mitigate overestimation bias, with each critic having a corresponding target network. Each critic network \(Q_{1}\left(s,a\mid \theta^{Q_{1}}\right)\) and \(Q_{2}\left(s,a\mid \theta^{Q_{2}}\right)\) corresponds to a target network \(Q_{1}^{\prime}\left(s,a\mid \theta^{Q_{1}^{\prime}}\right)\) and \(Q_{2}^{\prime}\left(s,a\mid \theta^{Q_{2}^{\prime}}\right)\) respectively. The minimum Q- value among the two critics is used as the target Q- value during training. The critic network is updated by minimizing the loss function:  
+
+\[L\left(\theta^{Q_{i}}\right) = \mathbb{E}\left[\left(y_{t} - Q_{i}\left(s_{t},a_{t}\mid \theta^{Q_{i}}\right)\mid a_{t} = \mu \left(s_{t}\mid \theta^{\mu}\right)\right)^{2}\right] \quad (19)\]  
+
+where \(\theta^{Q}\) denotes the weights of the critic network. The target Q- value \(y\) is evaluated by taking the minimum of the estimates from the two Q- functions as follows:  
+
+\[y_{t} = r\left(s_{t},a_{t}\right) + \gamma \min_{i = 1,2}Q_{i}^{\prime}\left(s_{t + 1},a_{t + 1}\mid \theta^{Q_{i}^{\prime}}\right) \quad (20)\]  
+
+where \(a_{t + 1}\sim \pi_{\phi^{\prime}}\left(s_{t + 1}\right) + \epsilon ,\quad \epsilon \sim \mathrm{clip}(\mathcal{N}(0,\tilde{\sigma}), - c,c)\) is the exploration noise to smooth the value estimates and improve robustness of the learned \(Q\) functions, \(r\) is the instantaneous one- step reward, \(\gamma\) is the discounting factor.  
+
+### 3.4. Baseline Methods  
+
+We use a series of baseline EMS methods for comparatively evaluating the ORL method. The inputs and outputs of all baselines are the same as those of the proposed method.
+
+<--- Page Split --->
+
+Dynamic Programming (DP) [30]: In optimization control methods, the EMS problem is formulated as a nonlinearly constrained optimization problem, aiming to minimize the objective function presented in Equation (10). DP is an optimization control method that operates by seeking the shortest path backward in time. Its objective is to derive the minimum cost function for each grid at every stage in reverse chronological order. In our study, DP is used as the benchmark EMS policy, representing the global optimum and providing upper limits for comparison. It's important to recognize that DP requires future information as input to achieve the optimization objective.  
+
+Behavior Cloning (BC) [31]: BC, as a fundamental imitation learning approach, seeks to emulate the EMS policy by directly learning from the provided dataset, which is assumed to be generated by an expert policy or near- expert policy. It employs supervised learning techniques to train a model to map states to actions. Both BC and ORL involve learning from data for EMS applications. In this context, we establish BC as the benchmark and aim to showcase the superior performance of ORL.  
+
+Proximal Policy Optimization (PPO) [32]: PPO is a state- of- the- art online DRL algorithm, which has been extensively applied in various applications requiring sophisticated decision- making in dynamic environments. PPO offers a robust and efficient approach to training agent by leveraging on- policy learning, effective use of data through mini- batch updates, stability through policy clipping, and adaptive learning rates. Leveraging the strengths of PPO, we utilize it to generate the dataset necessary for ORL, with its policy serving as an expert (near- optimal) strategy for comparison purposes. We provide it to explore the superiority of ORL compared to the online DRL.  
+
+Twin Delayed Deep Deterministic Policy Gradient (TD3) [26]: TD3 is an advanced online DRL algorithm, stemming from the Actor- Critic framework. It has garnered significant attention due to its effectiveness in overcoming challenges associated with continuous action spaces and high- dimensional state spaces. TD3 employs twin critic networks to estimate the value of actions more accurately. By utilizing two critic networks, TD3 mitigates overestimation bias and enhances the robustness of value function estimation. We also provide it to explore the superiority of ORL compared to the online DRL.  
+
+## 4. Discussion  
+
+In conclusion, we have presented a novel data- driven EMS for hybrid energy systems in EVs. Leveraging an innovative offline reinforcement learning agent, our approach learns directly from driving data. Experimental results demonstrate that the ORL agent not only learns optimal EMS strategies from expert data but also exhibits the ability to learn superior EMS from datasets containing a mixture of expert and noisy data, and even achieves near- optimal strategies from entirely noisy datasets. Moreover, our approach demonstrates that with increased data availability, performance improves as the agent is trained with more data.  
+
+This approach offers three notable benefits. Firstly, it is sufficiently simple, as it solely relies on collected data for automatic learning by the agent, unlike the traditional EMS development process, which often requires extensive expert knowledge and repeated measurements. Furthermore, the data used in our approach are non- expert data readily available from real vehicles. Secondly, our method ensures stable performance by integrating seamlessly with existing EMS without altering the original EMS performance lower bound. Our approach continuously improves upon the baseline EMS through data- driven enhancements leveraging the strengths of both technologies. For example, to address the performance shortcomings in rule- based EMS, ORL enables incremental learning, allowing for the continual enhancement of EMS performance with historical data. Similarly, ORL addresses the sim- to- real gap problem in simulation- based methods by enhancing pretrained EMS models, thereby ensuring their effectiveness in real- world deployment scenarios. Lastly, our approach exhibits versatility, as with sufficient data, it can learn a generalized EMS applicable to various EVs and operating conditions. This aligns with the current trend of large- scale language models and similar approaches in artificial intelligence, where a single large model with large- scale data can be trained to perform well across diverse tasks and domains.  
+
+Overall, we believe that ORL could serve as a foundational framework for data- driven EMS, with potential applications extending beyond EVs to grid EMS, industrial energy management systems, and other vehicle control systems. However, a limitation of this work is that the ORL agent may require more data to further enhance its performance. Addressing this limitation could involve exploring methods to efficiently gather and utilize additional data for agent training, potentially improving its effectiveness in real- world applications.
+
+<--- Page Split --->
+
+## Acknowledgements  
+
+AcknowledgementsThis work was supported in part by the National Natural Science Foundation of China (Grant No. 52172377).  
+
+## Author Contributions  
+
+Y.W. designed the study and methodology; Y.W., J.Wu. and H.H. collected and analyzed data; Y.W. generated the figures; Y.W., J.Wu. and W.Z. wrote the manuscript; H.H., W.Z. and F.S. reviewed and edited the manuscript. All authors contributed to the paper.  
+
+## References  
+
+[1] B. Borlaug, M. Muratori, M. Gillern, D. Woody, W. Muston, T. Canada, A. Ingram, H. Gresham, C. McQueen, Heavy- duty truck electrification and the impacts of depot charging on electricity distribution systems, Nature Energy 6 (2021) 673- 682. [2] Y. Zhao, Z. Wang, Z.- J. M. Shen, F. Sun, Assessment of battery utilization and energy consumption in the large- scale development of urban electric vehicles, Proceedings of the National Academy of Sciences 118 (2021) e2017318118. [3] H. He, X. Meng, Y. Wang, A. Khajepour, X. An, R. Wang, F. Sun, Deep reinforcement learning based energy management strategies for electrified vehicles: Recent advances and perspectives, Renewable and Sustainable Energy Reviews 192 (2024) 114248. [4] Y. Li, H. He, A. Khajepour, Y. Chen, W. Huo, H. Wang, Deep reinforcement learning for intelligent energy management systems of hybrid- electric powertrains: Recent advances, open issues, and prospects, IEEE Transactions on Transportation Electrification (2024). [5] A. H. Ganesh, B. Xu, A review of reinforcement learning based energy management systems for electrified powertrains: Progress, challenge, and potential solution, Renewable and Sustainable Energy Reviews 154 (2022) 111833. [6] F. Zhang, L. Xiao, S. Coskun, H. Pang, S. Xie, K. Liu, Y. Cui, Comparative study of energy management in parallel hybrid electric vehicles considering battery ageing, Energy 264 (2023) 123219. [7] J. Wu, C. Huang, H. He, H. Huang, Confidence- aware reinforcement learning for energy management of electrified vehicles, Renewable and Sustainable Energy Reviews 191 (2024) 114154. [8] X. Tang, T. Jia, X. Hu, Y. Huang, Z. Deng, H. Pu, Naturalistic data- driven predictive energy management for plug- in hybrid electric vehicles, IEEE Transactions on Transportation Electrification 7 (2020) 497- 508. [9] Y. Wang, Y. Wu, Y. Tang, Q. Li, H. He, Cooperative energy management and eco- driving of plug- in hybrid electric vehicle via multi- agent reinforcement learning, Applied Energy 332 (2023) 120563. [10] X. Hu, T. Liu, X. Qi, M. Barth, Reinforcement learning for hybrid and plug- in hybrid electric vehicle energy management: Recent advances and prospects, IEEE Industrial Electronics Magazine 13 (2019) 16- 25. [11] Y. Wang, H. Tan, Y. Wu, J. Peng, Hybrid electric vehicle energy management with computer vision and deep reinforcement learning, IEEE Transactions on Industrial Informatics 17 (2020) 3857- 3868. [12] W. Chen, J. Peng, J. Chen, J. Zhou, Z. Wei, C. Ma, Health- considered energy management strategy for fuel cell hybrid electric vehicle based on improved soft actor critic algorithm adopted with beta policy, Energy Conversion and Management 292 (2023) 117362. [13] W. Yuankai, L. Renzong, W. Yong, L. Yi, Benchmarking deep reinforcement learning based energy management systems for hybrid electric vehicles, in: CAAI International Conference on Artificial Intelligence, Springer, 2022, pp. 613- 625. [14] Y. Wang, R. Lian, H. He, J. Betz, H. Wei, Auto- tuning dynamics parameters of intelligent electric vehicles via bayesian optimization, IEEE Transactions on Transportation Electrification (2023). [15] A. R. Mayyas, S. Kumar, P. Pisu, J. Rios, P. Jethani, Model- based design validation for advanced energy management strategies for electrified hybrid power trains using innovative vehicle hardware in the loop (vhil) approach, Applied Energy 204 (2017) 287- 302. [16] H. He, Z. Niu, Y. Wang, R. Huang, Y. Shou, Energy management optimization for connected hybrid electric vehicle using offline reinforcement learning, Journal of Energy Storage 72 (2023) 108517. [17] G. Pozzato, A. Allam, L. Pulvirenti, G. A. Negoita, W. A. Paxton, S. Onori, Analysis and key findings from real- world electric vehicle field data, Joule 7 (2023) 2035- 2053. [18] K. A. Severson, P. M. Attia, N. Jin, N. Perkins, B. Jiang, Z. Yang, M. H. Chen, M. Aykol, P. K. Herring, D. Fraggedakis, et al., Data- driven prediction of battery cycle life before capacity degradation, Nature Energy 4 (2019) 383- 391. [19] W. Li, J. Zhu, Y. Xia, M. B. Gorji, T. Wierzbicki, Data- driven safety envelope of lithium- ion batteries for electric vehicles, Joule 3 (2019) 2703- 2715. [20] J. Zhang, Y. Wang, B. Jiang, H. He, S. Huang, C. Wang, Y. Zhang, X. Han, D. Guo, G. He, et al., Realistic fault detection of li- ion battery via dynamical deep learning, Nature Communications 14 (2023) 5940. [21] D. Roman, S. Saxena, V. Robu, M. Pecht, D. Flynn, Machine learning pipeline for battery state- of- health estimation, Nature Machine Intelligence 3 (2021) 447- 456. [22] N. Guo, S. Chen, J. Tao, Y. Liu, J. Wan, X. Li, Semi- supervised learning for explainable few- shot battery lifetime prediction, Joule (2024). [23] F. Millo, L. Rolando, L. Tresca, L. Pulvirenti, Development of a neural network- based energy management system for a plug- in hybrid electric vehicle, Transportation Engineering 11 (2023) 100156. [24] R. F. Prudencio, M. R. Maximo, E. L. Colombini, A survey on offline reinforcement learning: Taxonomy, review, and open problems, IEEE Transactions on Neural Networks and Learning Systems (2023). [25] H. He, F. Sun, Z. Wang, C. Lin, C. Zhang, R. Xiong, J. Deng, X. Zhu, P. Xie, S. Zhang, et al., China's battery electric vehicles lead the world: achievements in technology system architecture and technological breakthroughs, Green Energy and Intelligent Transportation 1 (2022) 100020.
+
+<--- Page Split --->
+
+[26] S. Fujimoto, H. Hoof, D. Meger, Addressing function approximation error in actor- critic methods, in: International conference on machine learning, PMLR, 2018, pp. 1587- 1596. [27] S. Fujimoto, S. S. Gu, A minimalist approach to offline reinforcement learning, Advances in neural information processing systems 34 (2021) 20132- 20145. [28] S. Kidera, K. Shintani, T. Tsuneda, S. Yamane, Combined constraint on behavior cloning and discriminator in offline reinforcement learning, IEEE Access (2024).[29] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A. A. Bharath, Generative adversarial networks: An overview, IEEE signal processing magazine 35 (2018) 53- 65. [30] P. Saiteja, B. Ashok, Critical review on structural architecture, energy control strategies and development process towards optimal energy management in hybrid vehicles, Renewable and Sustainable Energy Reviews 157 (2022) 112038. [31] A. Block, A. Jadbabaie, D. Pfrommer, M. Simchowitz, R. Tedrake, Provable guarantees for generative behavior cloning: Bridging low- level stability and high- level behavior, Advances in Neural Information Processing Systems 36 (2024).[32] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal policy optimization algorithms, arXiv preprint arXiv:1707.06347 (2017).
+
+<--- Page Split --->

preprint/preprint__08ebe812ad76a570c3ee7e20f364f13b63c9040737d4f23cef0f561c4cb41a3a/preprint__08ebe812ad76a570c3ee7e20f364f13b63c9040737d4f23cef0f561c4cb41a3a_det.mmd

@@ -0,0 +1,463 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 955, 174]]<|/det|>
+# Learning Superior Energy Management from Electric Vehicle Data  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 257, 240]]<|/det|>
+Hongwen He hwhbit@bit.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[44, 268, 712, 476]]<|/det|>
+Beijing Institute of Technology  Yong Wang  Beijing Institute of Technology  Jingda Wu  Nanyang Technological University https://orcid.org/0000- 0002- 7336- 4492  Zhongbao Wei  Beijing Institute of Technology https://orcid.org/0000- 0003- 0051- 5648  Fengchun Sun  Beijing Institute of Technology  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 515, 103, 532]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[42, 552, 928, 595]]<|/det|>
+Keywords: Energy management, Electric vehicle data, Reinforcement learning, Fuel cell vehicles, Data- driven  
+
+<|ref|>text<|/ref|><|det|>[[44, 613, 285, 633]]<|/det|>
+Posted Date: July 3rd, 2024  
+
+<|ref|>text<|/ref|><|det|>[[42, 651, 475, 671]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4523312/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 689, 916, 732]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 750, 535, 770]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 805, 934, 848]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on March 22nd, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 58192- 9.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[68, 66, 921, 94]]<|/det|>
+# Learning Superior Energy Management from Electric Vehicle Data  
+
+<|ref|>text<|/ref|><|det|>[[70, 106, 832, 128]]<|/det|>
+Yong Wang \(^{a,b}\) , Jingda Wu \(^{c}\) , Hongwen He \(^{a,b}\) , Zhongbao Wei \(^{a}\) and Fengchun Sun \(^{a}\)  
+
+<|ref|>text<|/ref|><|det|>[[68, 140, 822, 186]]<|/det|>
+\(^{a}\) School of Mechanical Engineering, Beijing Institute of Technology, Beijing, 100081, China \(^{b}\) National Key Laboratory of Advanced Vehicle Integration and Control, Beijing Institute of Technology, Beijing, 100081, China \(^{c}\) The Hong Kong Polytechnic University, Hung Hom, Hong Kong  
+
+<|ref|>sub_title<|/ref|><|det|>[[70, 205, 228, 220]]<|/det|>
+## ARTICLE INFO  
+
+<|ref|>sub_title<|/ref|><|det|>[[370, 207, 488, 221]]<|/det|>
+## ABSTRACT  
+
+<|ref|>text<|/ref|><|det|>[[69, 232, 210, 309]]<|/det|>
+Keywords: Energy management Electric vehicle data Reinforcement learning Fuel cell vehicles Data- driven  
+
+<|ref|>text<|/ref|><|det|>[[368, 232, 928, 488]]<|/det|>
+Despite the promising potential of energy management technologies in optimizing electric vehicle (EV) performance and fostering global energy sustainability, the extensive research conducted over the past decade has yet to translate into practical applications. This discrepancy arises primarily from the reliance of existing methodologies on simulation- based development paradigms, leading to a significant disparity between simulated results and real- world efficacy. Herein, we present a pioneering real- world data- driven energy management strategies (EMS) approach that utilizes an innovative offline reinforcement learning (ORL) framework. This paradigm enables EMS to learn from diverse real- world data, obviating the need for explicit rule design or high- fidelity simulators, and allowing for seamless application of the proposed method to any existing EMS. Moreover, it continuously enhances performance even after deployment in actual energy management systems. We evaluate the proposed ORL method on fuel cell EVs, training the ORL agent to optimize energy consumption and system degradation. The EV monitoring and management platform in China provides real- world data for validating our methodology. The results demonstrate that ORL consistently learns superior EMS in various conditions. With increasing data availability, its performance improves significantly, from \(88\%\) to \(98.6\%\) relative to theoretical optimality after two data updates. After training with more than 60 million kilometers of data, the ORL agent can learn a general EMS that adapts to unseen and corner- case conditions. These results highlight the effectiveness of integrating the data- driven method with established EMS techniques to enhance performance and underline its potential to utilize large- scale data to improve vehicle energy efficiency and longevity.  
+
+<|ref|>sub_title<|/ref|><|det|>[[70, 536, 217, 555]]<|/det|>
+## 1. Introduction  
+
+<|ref|>text<|/ref|><|det|>[[68, 559, 928, 753]]<|/det|>
+The automotive industry is undergoing a significant transformation, primarily due to the global focus on sustainability and environmental conservation. Electric vehicles (EVs) are leading this shift, playing a key role in mitigating environmental challenges and advancing sustainable transportation solutions [1, 2]. Concurrently, the emergence of hybrid energy systems (HES) within the EV powertrain represents an emerging trend, offering superior solutions over single energy systems [3]. By integrating multiple energy sources such as batteries, fuel cells, and internal combustion engines, HES improves overall efficiency, sustainability, and reliability, while also providing adaptability to a wide range of driving conditions [4]. Propelled by rapid technological advancements and supportive policies, EVs equipped with HES, including Hybrid EVs (HEV), Plug- in hybrid EVs (PHEV), and Fuel cell EVs (FCEV), are gaining traction worldwide [3]. For example, BYD, a prominent EV manufacturer, achieved sales of 3.02 million EVs in 2023, and PHEVs represent \(47.9\%\) of total sales. In the first quarter of 2024, PHEVs accounted for a notable \(51.6\%\) of total sales, indicating their growing popularity. This trend is mainly attributed to advancements in HES energy management, which enhance energy efficiency and overall performance.  
+
+<|ref|>text<|/ref|><|det|>[[68, 753, 928, 883]]<|/det|>
+The energy management strategy (EMS), responsible for allocating energy flow among HES to achieve predefined objectives, performs several vital functions essential for EVs:(1) EMS optimizes system efficiency by intelligently allocating energy flow based on driving conditions, reducing energy consumption and extending driving range [5]. (2) EMS optimizes power delivery based on driver demand, improving acceleration and responsiveness, while also ensuring smooth power delivery for improved driving experience. (3) By considering the characteristics of different power sources, EMS extends the lifespan of the HES, thus enhancing system reliability and safety [6]. Early EMS solutions utilize various rule- based approaches to achieve energy- saving objectives. Rule- based EMS involves the design of predefined rules and parameters tailored to specific driving conditions and vehicle characteristics, relying on
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 71, 928, 138]]<|/det|>
+expert knowledge with iterative refinement based on testing feedback [7]. However, this process can be labor- intensive and time- consuming, requiring manual expertise for rule creation and extensive experimentation for parameter calibration. Moreover, its static nature and inability to adapt to dynamic driving scenarios limit its effectiveness in maximizing energy savings and overall performance.  
+
+<|ref|>text<|/ref|><|det|>[[67, 135, 928, 459]]<|/det|>
+Recently, there has been a surge in advanced algorithms for energy management of HES, including dynamic programming (DP), model predictive control (MPC), reinforcement learning (RL), and others. A common feature across these methods is the use of optimal control or machine learning (ML) algorithms to compute optimal EMS based on high- fidelity simulators. Collectively, these approaches are referred to as simulation- based EMS in our study. Although DP and MPC provide elegant solutions based on future driving data, accurately forecasting future driving conditions using historical speed and dynamic traffic information remains a significant challenge [8]. Additionally, prediction models and optimization algorithms often result in considerable computational complexity, leading to suboptimal real- time performance [3]. Consequently, applying optimal control methods in real- world vehicle settings poses considerable challenges at present. In contrast, learning- based models, exemplified by Deep RL (DRL), do not rely on knowledge of future conditions and present promising alternatives for EMS [9]. DRL involves learning optimal actions in an environment through trial and error, where the DRL agent interacts with the environment to maximize cumulative rewards over time [10]. DRL- based EMS has been a subject of active research in recent years, with recent developments in online DRL algorithms such as Deep Deterministic Policy Gradient (DDPG) [11], Soft Actor- Critic (SAC) [12], Proximal Policy Optimization (PPO) [13], and others, leading to remarkable results. However, the application of DRL- based EMS faces challenges as an online learning paradigm, particularly concerning the interaction between the DRL agent and the EV simulator, which raises safety concerns. Moreover, while existing literature assumes that simulation models accurately represent real- world conditions, the construction of high- fidelity models that encompass vehicle dynamics, powertrain, traffic scenarios, and driver behavior [14] remains a challenge [15]. This discrepancy can cause the "sim- to- real" problem, where EMS learned in simulators may not be effectively transferred to real vehicles, which leads to the complexity of EMS development [16].  
+
+<|ref|>text<|/ref|><|det|>[[67, 456, 928, 732]]<|/det|>
+In recent years, data- driven methodologies, propelled by advancements in ML techniques and the availability of large datasets, have become essential in addressing key challenges in the EV field [17]. With support from data collection platforms and open- access laboratory data, these data- driven approaches have revolutionized various aspects of battery management systems (BMS). This includes automatic discovery of complex battery aging mechanisms [18], prediction of battery safety envelopes [19], evaluation of safety conditions [20], estimation of battery state of health [21], and even enhancing battery lifetime prediction models with unlabeled data [22]. Notably, innovations in feature extraction and supervised ML techniques tailored for time- series data have greatly enhanced prediction accuracy. This success has sparked interest in exploring data- driven methods for sequential decision- making tasks, including improving energy management systems. A common approach to implementing a data- driven EMS involves using supervised learning, where the ML model is employed to capture the complex and non- linear relationships between input features and corresponding control outputs. In [23], recurrent neural networks are trained offline using a substantial amount of training data obtained from the global optimization strategy DP, yielding a sub- optimal EMS that closely approximates the DP. It is essential to recognize the differences from the prediction tasks in BMS applications, as EMS entails sequential decision- making. Although supervised learning can mimic the policy of EMS through imitation learning, its heavy reliance on expert data may result in limited generalization to new and diverse scenarios. Hence, it's crucial to explore alternative methods for learning EMS from non- expert data, which is common in automotive applications.  
+
+<|ref|>text<|/ref|><|det|>[[67, 731, 928, 894]]<|/det|>
+In this paper, we present a novel data- driven EMS paradigm, addressing key challenges discussed above by leveraging offline data generated by existing EMS strategies, without the need for explicit rule design or online interaction. Our approach integrates DRL and supervised learning techniques, departing from traditional rule- based EMS and simulation- based methods to enhance EMS capabilities, as visually depicted in Figure 1(a). Notably, our proposed method can be directly integrated with any EMS algorithm and continuously improves through the collection data. We term this approach the offline reinforcement learning (ORL) agent [24], which is trained by combining real- world data as well as carefully filtered and processed simulation data through a novel exploration scheme. To evaluate this paradigm, we focus on a fuel cell electric vehicle (FCEV) as the subject and collect data for learning power allocation strategies. The EV monitoring and management platform in China offers real- world data to validate our methodology (Figure 1(b)). Overall, the ORL agent introduced herein exhibits four key features (Figure 1(c)):
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[112, 68, 880, 608]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[67, 616, 930, 739]]<|/det|>
+<center>Figure 1: Overview of the proposed data-driven EMS framework. (a) Comparison of three EMS paradigms: Traditional rule-based EMS relies on expert knowledge and calibration based on fixed driving cycles. Simulation-based EMS requires high-precision models and entails transitioning from simulation analysis to real-world deployment, resulting in a gap between simulated and real-world performance (sim-to-real gap). In contrast, the proposed real-world data-driven EMS learns directly from actual data. (b) China has established a three-tier EV monitoring and management system involving the state, local governments, and enterprises. The National Monitoring and Management Platform collects real-time operational data from over 20 million EVs [25]. (c) The ORL agent works with an existing EMS and continuously collects EV data to improve the EMS. (d) Overall diagram of the proposed data-driven EMS methodology. </center>  
+
+<|ref|>text<|/ref|><|det|>[[68, 770, 928, 820]]<|/det|>
+1) Purely data-driven EMS: By autonomously learning and optimizing from collected offline datasets, our approach facilitates the development of advanced EMS without necessitating expert knowledge or the construction of high-fidelity EV simulators. This data-driven process significantly simplifies EMS development workflows.  
+
+<|ref|>text<|/ref|><|det|>[[68, 820, 928, 885]]<|/det|>
+2) Learning from non-optimal data: Our research indicates that ORL can learn near-optimal EMS from non-optimal data and even can derive superior policies from sub-optimal. Our approach is less reliant on data quality allowing for effective learning. This practicality allows us to utilize raw data generated by actual vehicles, which is common in automotive applications.  
+
+<|ref|>text<|/ref|><|det|>[[68, 884, 928, 917]]<|/det|>
+3) Enhancement with increased data: The performance of ORL improves as more training data is utilized, demonstrating its ability to continuously adapt and enhance EMS performance. Through training across diverse
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 71, 925, 105]]<|/det|>
+datasets, it can potentially adapt to new driving conditions and yield favorable results, even in corner- case conditions. When sufficient data are available, ORL can learn a generalized EMS.  
+
+<|ref|>text<|/ref|><|det|>[[68, 103, 928, 170]]<|/det|>
+4) Compatibility with existing EMS: Our approach seamlessly integrates with established rule-based or simulation-based EMS methods, leveraging data from onboard controllers to augment EMS performance. This ensures that baseline performance is preserved while facilitating further enhancement using ORL, making it a valuable extension to conventional EMS methodologies.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 193, 166, 209]]<|/det|>
+## 2. Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 215, 402, 233]]<|/det|>
+### 2.1. The overview of data-driven EMS  
+
+<|ref|>text<|/ref|><|det|>[[68, 233, 928, 299]]<|/det|>
+In Figure 1(d), the framework overview of ORL for EMS is illustrated. We introduce the three phases of applying the proposed ORL algorithm to the EMS problem, which include data collection, offline learning, and evaluation. ORL is an innovative subset of RL methods rooted in data- driven approaches. Unlike most simulation- based EMS scheduling methods, the data- based learning process does not require the building of an EV simulation environment.  
+
+<|ref|>text<|/ref|><|det|>[[68, 297, 928, 476]]<|/det|>
+In the data collection phase, an available dataset \((D)\) is gathered from the existing EMS policy \((\pi_{\beta})\) . It is crucial to record various parameters such as velocity, acceleration, hydrogen consumption, electricity consumption, fuel cell degradation, battery state of charge (SOC), and others during vehicle operation. However, due to limitations in real- world data quality and privacy concerns associated with collecting data from a large number of actual vehicles, we developed an FCEV simulation model to generate laboratory datasets 2(a). This FCEV model enables the generation of extensive, high- fidelity data based on real- world driving conditions 2(b). Subsequently, we employ a tailored data processing model to encode control and state variables into standardized data formats. The data encoding process converts these parameters into a transition dataset \(D = \left(s,r,a,s^{\prime}\right)_{i}\) . Here, \(s\) , \(a\) , and \(r\) represent the state, action, and reward as described in the Methods section, respectively, while \(s^{\prime}\) denotes the subsequent state, and \(i\) indexes a transition in the dataset. These meticulously curated datasets are stored in the buffer and serve as input for the ORL agent.  
+
+<|ref|>text<|/ref|><|det|>[[68, 473, 928, 604]]<|/det|>
+In the offline learning phase, our proposed ORL agent is designed based on the Actor- Critic network and incorporates Behavior Cloning (BC) and Discriminator Blend (DB) mechanisms. In each training step, mini- batches of transitions \((s,r,a,s^{\prime})\) stored in a data buffer are sampled from the replay buffer to update the ORL networks. The Actor network, responsible for determining the action \(\mathbf{a}_{t}\sim \pi_{\beta}\left(\cdot \mid \mathbf{s}_{t}\right)\) , is updated to maximize the expected return as estimated by the Critic network. BC regularization is applied to the policy update step to encourage the policy to prioritize actions present in the dataset. Moreover, the DB component is introduced to allow actions beyond the dataset distribution similar to those included in the dataset. Through training, the agent learns to optimize its actions based on observed states and rewards. The details of the ORL agent will be elaborated upon in the Methods section in detail.  
+
+<|ref|>text<|/ref|><|det|>[[68, 603, 928, 750]]<|/det|>
+After completing the training phase, the neural network parameters representing the EMS policy of the ORL agent are saved for future use. Subsequently, the trained agent is evaluated to gauge its effectiveness and performance. Utilizing the FCEV environment established during the data collection phase, our subsequent experiments employ three standard driving cycles (WTVC, FTP, and CHTC) to evaluate energy cost. Additionally, we incorporate various real- world driving scenarios to comprehensively assess the trained EMS policy. Following the evaluation, adjustments to the agent hyperparameters or training process may be considered to improve its performance. This iterative process of training, evaluation, and refinement continues until the desired level of EMS performance is attained. Upon achieving satisfactory performance, the trained agent becomes eligible for deployment in real- world scenarios, where it can be utilized to efficiently optimize energy management systems.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 761, 370, 779]]<|/det|>
+### 2.2. Data for learning and analysis  
+
+<|ref|>text<|/ref|><|det|>[[68, 778, 928, 925]]<|/det|>
+We select the Proximal Policy Optimization (PPO), which demonstrates the best performance among online DRL algorithms in our EMS problem, as the expert EMS. Using PPO, we generate datasets denoted as \(D^{E}\) comprising 300K time steps. Additionally, we employ a random agent that samples actions randomly, generating datasets denoted as \(D^{R}\) , which represent poor performance. To create settings with varying levels of data quality in the suboptimal offline dataset, we combine transitions from the expert datasets \(D^{E}\) and the random datasets \(D^{R}\) in different ratios. Specifically, we consider four different dataset compositions, denoted as D1, D2, D3, and D4. These settings are defined as follows: D1 (Data- 1): Consists solely of transitions from the expert dataset \(D^{E}\) , representing the expert policy. D2 (Data- 2): Contains two- thirds of transitions from the expert dataset \(D^{E}\) and one- third of transitions from the random dataset \(D^{R}\) , representing suboptimal data. D3 (Data- 3): Comprises one- third of transitions from the expert dataset
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[120, 72, 861, 625]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[68, 629, 930, 735]]<|/det|>
+<center>Figure 2: Comparison for different datasets. (a) The simulation model for the fuel cell hybrid electric vehicle (FCEV), depicts the powertrain layout and energy flow topology. (b) The principle of EMS involves the allocation of energy flow among hybrid energy systems (fuel cell and battery) to achieve predefined objectives based on driving conditions. Here, real driving conditions are gathered and then use the simulated FCEV model to generate energy cost data for training the ORL agent. (c) Distribution of encoded actions for the four datasets, each action is normalized to the [-1,1]. D1 represents data generated by an expert policy, D2 and D3 denote suboptimal data generated by a combination of expert and random policies, D4 comprises entirely random data. (d) The states distribution of four datasets. </center>  
+
+<|ref|>text<|/ref|><|det|>[[68, 767, 928, 802]]<|/det|>
+\(D^{E}\) and two- thirds of transitions from the random dataset \(D^{R}\) , representing another suboptimal data. D4 (Data- 4): Comprises solely of transitions from the random dataset \(D^{R}\) , representing the random policy.  
+
+<|ref|>text<|/ref|><|det|>[[68, 801, 928, 916]]<|/det|>
+Figure 2(a) depicts the action distributions for the four datasets. Significant differences can be observed among the four EMS policies, with the action range of D1 falling within (- 0.2, 0.5), resulting in relatively stable variations in FC power. As random policy data is introduced, the action ranges in the other datasets all fall within (- 1, 1), particularly for D4 where the action distribution is uniformly spread across (- 1, 1). This implies that this policy is noisy, denoted as a poor EMS. Figure 2(b) illustrates the state distributions for the four datasets. Note that all states have undergone post- processing and scaling to the (0, 1) interval. By comparing the box plots of the four datasets, it is evident that the SOC of D1 falls within a reasonable range (0.38- 0.7), meeting the EMS constraints regarding the battery SOC.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 71, 928, 137]]<|/det|>
+However, the SOC of the other datasets falls into unreasonable ranges, such as SOC in the range (0.2, 1) for D3. Additionally, with the increase in \(D^R\) data, the FC power distribution ranges of D3 and D4 become wider. Since the conditions of the four datasets are derived from fixed segments of standard driving cycles, the velocity distribution remains the same across all datasets.  
+
+<|ref|>text<|/ref|><|det|>[[68, 136, 928, 185]]<|/det|>
+Creating challenging datasets is practical because generating sub- optimal or random data is more cost- effective than collecting expert- level data from real vehicles. Therefore, an effective data- driven EMS method must be able to effectively handle and learn from these suboptimal offline datasets.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 198, 518, 216]]<|/det|>
+### 2.3. Learning superior EMS from non-optimal data  
+
+<|ref|>image<|/ref|><|det|>[[66, 230, 920, 747]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[67, 748, 930, 840]]<|/det|>
+<center>Figure 3: The ORL agent learning performance. (a) Learning curves of ORL agent for the four different datasets. (b) The comparison of absolute rewards (original rewards are negative) under three validation conditions. "D1," "D2," "D3," and "D4" correspond to the best rewards achieved by ORL after learning on each respective dataset. Notably, ORL agent on various datasets closely approximate or exceed the expert policy. (c) The comparison of energy costs between the original EMS and the optimized EMS using ORL shows a notable reduction in energy costs through the data-driven learning process. (d) The action (FC power slope) distributions of optimized EMS using ORL </center>  
+
+<|ref|>text<|/ref|><|det|>[[68, 862, 928, 928]]<|/det|>
+We first study the performance of the ORL agent with different datasets. To ensure a fair comparison, the algorithm employs uniform experimental settings and network parameters across four datasets. Figure 3(a) illustrates the average reward of the training process on the four datasets. This average, computed as the mean reward over every 1000 episodes, undergoes validation across 10 iterations using three standard driving cycles: WTVC, CHTC, and FTP.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 71, 928, 169]]<|/det|>
+Training involves utilizing a buffer comprising 300,000 samples, with the ORL agent randomly selecting 256 data points for each training iteration, accumulating to one million training epochs. For D1, which comprises exclusively expert data, convergence is evident after approximately 210e3 episode. However, ORL exhibits a slower convergence speed during iterative learning on the other three datasets, converging at around 330e3, 600e3, and 360e3 epochs, respectively. This suggests that the data distribution influences the learning speed, but ultimately, ORL learns an effective EMS.  
+
+<|ref|>text<|/ref|><|det|>[[68, 168, 929, 329]]<|/det|>
+Figure 3(b) presents the reward performance of trained ORL across three driving cycles. It notes that the absolute reward value of ORL decreases in D1, from 323.4 to 297.3 on the CHTC, representing an improvement of \(8.8\%\) . Surprisingly, even on suboptimal datasets D2 and D3, ORL outperforms the expert strategy, with reward increases of \(1.8\%\) and \(3.4\%\) , respectively. Similarly, ORL still outperforms on the WTVC and FTP conditions, learning superior strategies from the suboptimal datasets D2 and D3. The exception is D3 on the WTVC condition, possibly due to the high- speed nature of the WTVC condition, leading to larger reward values for SOC. Despite this, the final results of the energy consumption remain within rational bounds. Particularly noteworthy is the excellent performance on the random dataset D4, where ORL closely approaches expert results across all three validation conditions, achieving rewards of 402, 325, and 395. Compared to the original average reward of 2637 for the D4 dataset, ORL has reduced the reward by \(85.8\%\) .  
+
+<|ref|>text<|/ref|><|det|>[[68, 329, 928, 490]]<|/det|>
+Figure 3(c) provides a detailed comparison of the energy costs between ORL and the original datasets. The blue dots represent the mean costs of the four original EMS datasets, accompanied by error bars indicating the maximum and minimum costs. In contrast, the red dots depict the energy costs incurred by ORL in the corresponding datasets. Despite the inclusion of random data that leads to a degradation in cost performance, ORL consistently maintains lower costs across all data sets. For instance, in the WTVC condition, the cost escalates from the initial 90 RMB in dataset D1 to 163 RMB in dataset D4. However, ORL consistently maintains costs within the range of 90- 95 RMB. In particular, the minimum cost values of the original D4 dataset exceed those achieved by ORL significantly, with ORL achieving reductions in costs that exceed \(40\%\) in all three conditions. This observation underscores ORL's ability not only to glean superior results from expert EMS but also to consistently yield excellent outcomes from progressively suboptimal datasets. Remarkably, ORL even attains expert- level EMS performance when trained solely on noisy datasets.  
+
+<|ref|>text<|/ref|><|det|>[[68, 489, 928, 635]]<|/det|>
+To elucidate the rationale behind the performance enhancements, Figure 3(d) illustrates the action distributions of optimized EMS using ORL. As different EMS policies can be reflected by the actions taken, in the context of the FCEV considered here, this pertains to the FC power slope under the same driving cycle. Comparing Figures 2(c) and 3(d), we notice significant changes in D2, D3, and D4 compared to Figure 2. In D2, D3, and D4, the action distributions closely resemble those of expert data in D1, concentrating within the range of [-0.3, 0.3], as opposed to the wider range of [-1, 1] observed in Figure 2. The change is particularly pronounced in D4, where the absence of expert data results in slight differences in the action distributions compared to D1, D2 and D3. However, all ORL policies tend to learn FC power variations with smaller ranges, ensuring smoother FC power output while satisfying power demand requirements.  
+
+<|ref|>text<|/ref|><|det|>[[68, 634, 928, 715]]<|/det|>
+In conclusion, through experimentation on three validation conditions and four datasets, our ORL agent not only learns better EMS strategies from expert strategies but also demonstrates the ability to learn near- expert strategies from entirely noisy datasets and even achieves superior results from datasets containing a mixture of expert and noisy data. This observed convergence underscores ORL's ability to enhance and optimize the original EMS through learning from data obtained from any EMS.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 728, 454, 746]]<|/det|>
+### 2.4. Performance by comparative evaluation  
+
+<|ref|>text<|/ref|><|det|>[[68, 746, 928, 826]]<|/det|>
+To demonstrate the superior performance of ORL, we contrast it with simulation- based and imitation learning EMS approaches. Since imitation learning and ORL are closely related, both involve learning EMS from data. We first compare the performance between ORL and BC. It's important to note that BC typically employs a supervised learning paradigm, learning from expert data, while ORL incorporates reinforcement learning with exploration mechanisms. This distinctive learning mechanism results in significant performance differences.  
+
+<|ref|>text<|/ref|><|det|>[[68, 826, 928, 907]]<|/det|>
+In Figure 4(a), we compare the testing reward in the WTVC, CHTC and FTP driving cycles, and calculate the percentage of ORL and BC costs relative to the expert EMS (PPO). In D1, both ORL and BC achieve favorable results, with ORL surpassing the original expert data by a maximum of \(9.1\%\) , while BC remains comparable to the expert. In D2, ORL maintains superiority over expert- based EMS, while BC experiences significant cost degradation (ranging from \(6\%\) to \(70\%\) ). In D3 and D4, ORL continues to outperform or closely match the expert, while BC, limited by data
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[68, 67, 930, 604]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[67, 620, 930, 727]]<|/det|>
+<center>Figure 4: Performance analysis comparing different methods. (a) The comparison between two data-driven EMS methods. The matrix numbers represent the relative reward rates of ORL and BC compared to the expert EMS (PPO) under the same conditions, emphasizing the minimal influence of data quality on ORL's performance. (b) Comprehensive performance of different algorithms on the WTVC condition, DP representing the globally optimal EMS. (c) Comprehensive performance on the CHTC condition. (d) Comprehensive performance on the FTP condition. (e) The efficiency distribution of fuel cell power demonstrates that ORL learns a superior EMS, ensuring that fuel cell system remains within the high-efficiency range. (f) FC degradation costs under three conditions </center>  
+
+<|ref|>text<|/ref|><|det|>[[68, 760, 928, 794]]<|/det|>
+quality, fails to learn an optimal EMS. This underscores ORL's capability to learn superior EMS from non- expert data, while imitation learning demonstrates poorer performance and struggles to learn favorable EMS with non- expert data.  
+
+<|ref|>text<|/ref|><|det|>[[68, 793, 928, 890]]<|/det|>
+Figures 4(b- d) present detailed results of different methods under the WTVC, CHTC, and FTP conditions, with the red lines representing the percentage of cost compared to DP. It is evident that ORL learns an optimal EMS policy on the D1 dataset, achieving percentages close to \(99.9\%\) , \(99.4\%\) , and \(97.6\%\) of DP, respectively. As for PPO, a benchmark expert policy, its cost results are \(98.6\%\) , \(98.0\%\) , and \(97.6\%\) of DP, respectively. Thus, BC learns similar expert EMS in D1, but its performance significantly deteriorates on suboptimal D2 data. Another online DRL method, TD3, also demonstrates satisfactory performance, however, its overall costs are lower than those of PPO and ORL.  
+
+<|ref|>text<|/ref|><|det|>[[68, 889, 928, 923]]<|/det|>
+In Figure 4(e), the FC power distribution of the EMS learned by ORL on the D1 dataset is depicted, with the green curve representing the efficiency curve of the FC system. ORL demonstrates a superior EMS, ensuring that the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 72, 928, 153]]<|/det|>
+power distribution remains within the high- efficiency range, resulting in reduced hydrogen consumption. Additionally, a narrower power variation range, as shown in Figure 3(d), minimizes FC degradation costs. As illustrated in Figure 3(f), ORL incurs minimal FC degradation costs in the three conditions, with costs of 2.3, 1.4, and 1.4, respectively. Furthermore, examination of Figures 3(b- d) indicates that the battery SOC remains within a reasonable range. These findings collectively affirm that ORL has successfully learned a superior EMS.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 166, 448, 184]]<|/det|>
+### 2.5. Continuous learning with growing data  
+
+<|ref|>text<|/ref|><|det|>[[68, 184, 928, 249]]<|/det|>
+We have demonstrated in previous experiments that ORL can learn optimal EMS strategies from data and outperforms other methods. In this section, we further showcase ORL for continuous learning from data. We conduct experiments in three cases depicted in Figure 5(a), collecting real- vehicle data in different driving scenarios including urban road, highway, and downtown road for the three cases (Figure 5(b)).  
+
+<|ref|>title<|/ref|><|det|>[[68, 261, 516, 278]]<|/det|>
+#### 2.5.1. Case 1: Continuous learning from historical data  
+
+<|ref|>text<|/ref|><|det|>[[68, 278, 928, 406]]<|/det|>
+Take the example of driving a bus on fixed routes to illustrate Case 1. We collected real electric bus driving data in Zhengzhou, China, over three consecutive days. Figure 5(c) shows the speed trajectories of three days, labeled as ZBDC- No1, ZBDC- No2, and ZBDC- No3, respectively. Noticeably, there are variations in the speed trajectories along the same route over different days. Figures 5(d- f) show the total cost of different EMS strategies in the three scenarios, which include hydrogen consumption, battery cost, and fuel cell degradation. The baseline is the original EMS of FCEV, and using the baseline data under ZBDC- No1 driving cycle to train the ORL agent as the ORL(Z1) strategy, which is then applied to the new condition ZBDC- No2. Furthermore, we train a new ORL(Z2) EMS using data from both the baseline on ZBDC- No1 and the ORL(Z1) on ZBDC- No2, which is then validated on ZBDC- No3.  
+
+<|ref|>text<|/ref|><|det|>[[68, 405, 928, 535]]<|/det|>
+It can be observed that the baseline EMS performs poorly on the first day (ZBDC- No1), with a cost only \(88.0\%\) compared to DP. The corresponding FC power and power slope distributions are shown in Figures 5(g). On the second day as depicted in Figure 5(e) and (h), the performance of ORL(Z1) achieves a significant improvement by learning from the previous data, reaching \(96.4\%\) of the cost compared to DP on the ZBDC- No2. On the third day, continuously learning from more data, the ORL(Z2) achieves a cost of \(98.6\%\) compared to DP on the ZBDC- No3, as shown in Figure 5(f) and (i). By comparing the three power distribution, it is evident that ORL(Z1) and ORL(Z2) distribute more FC output power in the high- efficient range, resulting in lower overall energy consumption, and the smaller power slope also leads to lower system degradation costs.  
+
+<|ref|>text<|/ref|><|det|>[[68, 534, 928, 615]]<|/det|>
+In conclusion, with data updates, new data can be utilized to train ORL, leading to the evolution of batter EMS strategies. This demonstrates the ability of ORL to continuously learn and improve from historical data. Additionally, our method integrates seamlessly with established EMS methods, using real- time data from onboard controllers to increase EMS performance. This ensures that baseline performance is preserved while facilitating further enhancement using ORL, making it a valuable extension to conventional EMS methodologies.  
+
+<|ref|>title<|/ref|><|det|>[[68, 628, 434, 645]]<|/det|>
+#### 2.5.2. Case 2: Improving from simulated data  
+
+<|ref|>text<|/ref|><|det|>[[68, 645, 928, 726]]<|/det|>
+In Case 2, we address the challenge of improving EMS performance from simulated data, where historical data is absent and driving conditions are unknown. Despite the ability of simulation- based methods (such as DRL) to derive ideal EMS strategies from simulated EV models, deploying these strategies onto real vehicles often leads to performance degradation due to the stochastic and unknown nature of real- world driving conditions, a problem known as the sim- to- real gap, extensively studied in the fields of RLL.  
+
+<|ref|>text<|/ref|><|det|>[[68, 725, 928, 838]]<|/det|>
+As illustrated in Figure 6(a), during the simulation phase, the PPO algorithm is trained on the standardized driving cycle (WTVC) and three specific conditions ZBDC (Figure 5(c)) to obtain an ideal EMS, denoted as PPO (Train). Subsequently, the EMS is validated on 12 local driving conditions denoted as PPO (Test). As depicted in Figure 6(b), the cost difference between PPO (Train) and DP across the four training conditions is minimal, with an average difference of \(3.16\%\) . However, when tested on the 12 new conditions(DC- 1 to DC- 12), the average cost difference between PPO (Test) and DP rises to \(12.75\%\) . This indicates a significant performance degradation of DRL- based methods when transitioning from simulation to real- world conditions.  
+
+<|ref|>text<|/ref|><|det|>[[68, 838, 928, 919]]<|/det|>
+To mitigate the sim- to- real problem, our proposed ORL method leverages data from PPO (Test) for further learning. As shown in Figure 6(c), the ORL approach achieves a substantially lower cost across the 12 local operating conditions compared to the PPO (Train) strategy. The average cost difference between ORL and DP is merely \(1.42\%\) (Figure 6(b)). In summary, our experiments demonstrate that ORL can effectively learn from simulated data to enhance the performance of the original EMS, addressing the sim- to- real problem inherent in traditional simulation- based methods.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[70, 75, 940, 776]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[67, 788, 930, 880]]<|/det|>
+<center>Figure 5: ORL continuously learning and improving from data. (a) ORL for continuous learning from data in three scenarios. (b) Driving data collected from the real-world route. (c) Speed trajectories of three ZBDC conditions. (d) Comparing the total cost under ZBDC-No1, the total cost comprises hydrogen consumption, battery cost, and cell degradation cost. (e) Comparing the total cost under ZBDC-No2. (f) Comparing the total cost under ZBDC-No3. (g) Fuel cell power distribution cloud chart of Baseline EMS. (h) Fuel cell power distribution cloud chart following one data update. (i) Fuel cell power distribution cloud chart following two data updates. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[68, 70, 945, 753]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[67, 763, 930, 884]]<|/det|>
+<center>Figure 6: Performance with more data. (a) Performance of PPO for 4 training data and 12 testing data. (b) The comprehensive performance comparison of different EMS methods reveals that ORL can significantly mitigate the performance degradation observed in the testing phase of PPO. (c) Performance of ORL for 12 testing data. (d) Speed distribution of four conditions. (e) The demand power of HWDC represents an extreme condition. The red dashed line indicates the maximum output power of FC system. (f) Overall performance of the 4 cases as the training data increase under test driving cycle CLTC; (g) Performance on GCDC; (h) Performance on ZNDC; (I) Performance on HWDC indicates that ORL ultimately learns a reasonable EMS in extreme driving conditions. (j) Battery SOC trajectories of different EMS under the test driving cycle HWDC. </center>
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[68, 72, 555, 90]]<|/det|>
+#### 2.5.3. Case 3: Learning a general EMS with large-scale data  
+
+<|ref|>text<|/ref|><|det|>[[68, 88, 929, 217]]<|/det|>
+To assess the generalization of the ORL model, especially in extreme conditions, the ORL agent trained on huge amounts of data and tested its performance on four new driving conditions. Figure 6(d) illustrates the speed distribution for these four conditions. Among these, the China Light- Duty Vehicle Test Cycle (CLTC) stands as the standard cycle, GCDC is obtained from an EV operating in a downtown road, ZBDC represents a new condition collected in Zhengzhou, and HWDC is collected from a fuel vehicle traveling on a highway. The four conditions have distinct characteristics and originate from different road and vehicle types. Particularly for HWDC, where the demand power exceeds \(250\mathrm{kW}\) . As depicted in Figure 6(e), the power demand exceeds the \(100\mathrm{kW}\) maximum output power of FC system, indicating that HWDC is an extreme condition for the FCEV studied in this work.  
+
+<|ref|>text<|/ref|><|det|>[[68, 216, 929, 475]]<|/det|>
+Firstly, we establish datasets of four different scales: 4e4 (ORL- 4), 20e4 (ORL- 20), 100e4 (ORL- 100), and 500e4 (ORL- 500) samples, excluding data from the four validation conditions obtained in previous experiments (Cases 1 and 2). The results on the four validation conditions are depicted in Figures 6(f- i). It is evident from Figure 6(f), (g), and (h) that both the reward and cost exhibit a gradual decrease with an increase in training data. With more training data, the ORL model consistently enhances its performance. Notably, the rate of performance improvement diminishes after reaching the 100e4 sample mark, with minimal disparity observed between the ORL- 100 and ORL- 500. At approximately 20e4 samples, the ORL model outperforms the DRL- TD3 algorithm (as indicated by the dashed lines in the figures). It is essential to highlight that, under the extreme HWDC condition, although ORL- 20 achieves the lowest cost (Figure 6(i)), its reward absolute value is not the lowest. This phenomenon is because the ORL- 20 prefers battery consumption during high power demands, exceeding the maximum output of the FC system, violating the SOC constraint. This phenomenon arises because ORL- 20 tends to prioritize battery consumption during periods of high power demand, the FC system fails to achieve its maximum power output. Consequently, the strategy fails to maintain the SOC within the desired range, rendering it ineffective as an EMS. Conversely, ORL- 500 demonstrates superior performance, with lower reward and cost, and SOC maintained within a reasonable range, as illustrated in Figure 6(j). Overall, after learning from 5 million data points (equivalent to over 60 million kilometers), the ORL agent can learn a general EMS adaptable to unseen and even corner- case conditions.  
+
+<|ref|>text<|/ref|><|det|>[[68, 473, 928, 523]]<|/det|>
+This result highlights two advantages of the ORL agent: First, its performance surpasses that of the original policy; second, it demonstrates that with increased data availability, learning performance improves. The ORL model can learn a general EMS from a large amount of EV data.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 544, 180, 561]]<|/det|>
+## 3. Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 568, 258, 585]]<|/det|>
+### 3.1. EV Environment  
+
+<|ref|>text<|/ref|><|det|>[[68, 585, 929, 730]]<|/det|>
+In this work, we evaluate EMS performance using a fuel cell hybrid electric vehicle (FCEV) within a simulation environment. Figure. 1(e) illustrates the schematic diagram of the FCEV and its components, which include a fuel cell (FC) system, a hydrogen storage tank, an electric motor (EM), and a Lithium- ion battery (LIB) pack. The FC stack serves as the primary power source to meet the energy requirements of the vehicle. The diagram also depicts the energy flow from the hydrogen storage tank to the motor. The FC system converts hydrogen energy into electricity, which then collaborates with the LIB through the high- voltage bus. This electric energy is subsequently utilized to power a single electric motor, connected to the driving wheel via a fixed- ratio final gear. According to the vehicle driving resistance equation, the driving power demand is determined by the speed and acceleration of the FCEV, and can be expressed as follows:  
+
+<|ref|>equation<|/ref|><|det|>[[113, 732, 925, 775]]<|/det|>
+\[P_{d} = \frac{1}{3.6\cdot\eta_{\mathrm{me}}}\left(m g C_{f}\upsilon_{t} + m\delta \upsilon_{t}a_{t} + m g s i n(i) + \frac{C_{D}A}{21.15}\upsilon_{t}^{3}\right) \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 779, 928, 860]]<|/det|>
+where \(\eta_{\mathrm{me}}\) is the efficiency of the vehicle drivetrain; \(m\) is the vehicle mass; \(g\) is the gravitational constant; \(C_{f}\) is the rolling resistance coefficient; \(\delta\) is the rotational mass conversion coefficient; \(C_{D}\) is the air resistance coefficient, \(A\) is the frontal area; \(\upsilon_{t}\) is the longitudinal velocity at the time step \(t\) , \(a_{t}\) is the acceleration, \(i\) is the angle of slope of the road. The power demand is provided by the FC system and the battery pack, the power balance of the FCEV can be formulated as:  
+
+<|ref|>equation<|/ref|><|det|>[[113, 868, 925, 890]]<|/det|>
+\[P_{d} = \left(P_{f c}\cdot \eta_{D C / D C} + P_{b a t}\right)\cdot \eta_{D C / A C}\cdot \eta_{E M} \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 896, 928, 931]]<|/det|>
+where \(P_{f c}\) and \(P_{b a t}\) respectively denote the output power of the fuel cell system and the LIB pack; \(\eta_{D C / D C},\eta_{D C / A C}\) and \(\eta_{\mathrm{EM}}\) represent the efficiency of the DC/DC converter, DC/AC inverter, and the electric motor, respectively. The
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[67, 71, 614, 90]]<|/det|>
+battery pack is modeled using an equivalent circuit model in Equation (3):  
+
+<|ref|>equation<|/ref|><|det|>[[111, 98, 926, 188]]<|/det|>
+\[\left\{ \begin{array}{l l}{P_{b a t}(t) = V_{o c}(t) - R_{0}\cdot I^{2}(t)}\\ {I(t) = \frac{V_{o c}(t) - \sqrt{V_{o c}^{2}(t) - 4\cdot R_{0}\cdot P_{b a t}(t)}}{2R_{0}}}\\ {S O C(t) = \frac{Q_{0} - \int_{0}^{t}I(t)dt}{Q}} \end{array} \right. \quad (3)\]  
+
+<|ref|>text<|/ref|><|det|>[[67, 199, 928, 250]]<|/det|>
+where \(S O C\) is the battery state of charge, \(V_{o c}\) is the open- circuit voltage, \(I_{t}\) is the current at time \(t\) , \(R_{0}\) is the internal resistance, \(P_{b a t}\) is the output power in the charge- discharge cycles, \(Q_{0}\) is the initial battery capacity, \(Q\) is the nominal battery capacity.  
+
+<|ref|>text<|/ref|><|det|>[[68, 247, 928, 298]]<|/det|>
+According to the battery aging model in [12]. The degradation rate of battery operation \(\gamma_{\mathrm{bat}}\) is affected by the charge/discharge rate ( \(C_{rate}\) ). The relationship between the battery aging correction factor and \(C_{rate}\) can be fitted from experiment data:  
+
+<|ref|>equation<|/ref|><|det|>[[111, 306, 926, 333]]<|/det|>
+\[\gamma_{\mathrm{bat}} = \mu_{1}\left|C_{\mathrm{rate}}\right|^{2} + \mu_{2}\left|C_{\mathrm{rate}}\right| + \mu_{3} \quad (4)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 340, 928, 376]]<|/det|>
+where \(\mu_{1}, \mu_{2}, \mu_{3}\) are the curve- fitting coefficients. LIB can operate for about 5000 full cycles in a lifetime. The battery degradation cost \(C_{bat,degr}\) can be calculated by:  
+
+<|ref|>equation<|/ref|><|det|>[[111, 385, 926, 425]]<|/det|>
+\[C_{bat,degr} = \int_{0}^{t}\gamma_{bat}^{-1}P_{bat}dt\cdot PR_{bat} / (5000\cdot 3600) \quad (5)\]  
+
+<|ref|>text<|/ref|><|det|>[[93, 433, 586, 451]]<|/det|>
+where \(PR_{\mathrm{bat}}\) is the battery price per kWh that is 1500RMB/kWh.  
+
+<|ref|>text<|/ref|><|det|>[[68, 450, 928, 484]]<|/det|>
+The fuel cell system efficiency under different power conditions is obtained from experiment data. Thus, the mass flow rate of the hydrogen consumption can be calculated by:  
+
+<|ref|>equation<|/ref|><|det|>[[111, 492, 926, 525]]<|/det|>
+\[\dot{m}_{H_{2}} = P_{f c s} / \left(\eta_{f c s}\cdot \mathrm{LHV}_{H_{2}}\right) \quad (6)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 532, 928, 567]]<|/det|>
+where \(\eta_{f c s}\) is the fuel cell system efficiency; \(P_{f c s}\) is the fuel cell system output power; \(\mathrm{LHV}_{H_{2}}\) is the hydrogen low calorific value. The fuel cell hydrogen cost can be calculated by:  
+
+<|ref|>equation<|/ref|><|det|>[[111, 575, 926, 614]]<|/det|>
+\[C_{f c s,H_{2}} = PR_{H_{2}}\cdot \int_{0}^{t}\dot{m}_{H_{2}}dt \quad (7)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 622, 528, 640]]<|/det|>
+where \(PR_{H_{2}}\) is the hydrogen price per kilogram(60RMB/kg).  
+
+<|ref|>text<|/ref|><|det|>[[68, 639, 928, 705]]<|/det|>
+The fuel cell degrades rapidly under four typical conditions: load changing, start/stop, low power, and high power conditions. We assume that the fuel cell system continues operating until the vehicle power system is shut down, thus the start/stop condition is not considered in the EMS. The fuel cell voltage degradation rate, denoted as \(\gamma_{\mathrm{fcs}}\) , can be calculated by:  
+
+<|ref|>equation<|/ref|><|det|>[[111, 715, 926, 736]]<|/det|>
+\[\gamma_{\mathrm{fcs}} = \kappa_{\mathrm{low}}\cdot T_{\mathrm{low}} + \kappa_{\mathrm{high}}\cdot T_{\mathrm{high}} + \kappa_{\mathrm{cha}}\cdot \Delta P_{\mathrm{fcs}} \quad (8)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 745, 928, 812]]<|/det|>
+where \(\kappa_{\mathrm{low}}\) is the degradation rate under low power condition; \(T_{\mathrm{low}}\) is the low power condition duration; \(\kappa_{\mathrm{high}}\) is the degradation rate under high power condition; \(T_{\mathrm{high}}\) is the high power condition duration; \(\kappa_{cha}\) is the degradation rate under load changing condition; \(\Delta P_{f c s}\) is the fuel cell power slope. Fuel cell is considered to reach the end of its life when \(10\%\) of voltage at rated power has been lost. The fuel cell operation degradation cost can be calculated by:  
+
+<|ref|>equation<|/ref|><|det|>[[111, 821, 926, 846]]<|/det|>
+\[C_{f c s,degr} = k_{f c s}\cdot \gamma_{f c s}\cdot P_{f c s,\mathrm{rate}}\cdot PR_{f c s} / \left(V_{f c s,\mathrm{end}}\cdot 1000\right) \quad (9)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 853, 928, 889]]<|/det|>
+where \(k_{f c s}\) is the fuel cell life correction factor; \(V_{f c s,\mathrm{end}}\) is fuel cell voltage drop at the end- of- life; \(P_{f c}\) , rate is the rated power of the fuel cell; \(PR_{f c s}\) is the fuel cell price per kilowatt(4000RMB/kW).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[68, 70, 270, 88]]<|/det|>
+### 3.2. Problem modeling  
+
+<|ref|>text<|/ref|><|det|>[[68, 87, 928, 137]]<|/det|>
+In this work, the EMS of electric vehicles is modeled as a long- term sequential decision process objective to minimize total energy cost while maintaining battery SOC within reasonable limits. The optimization objective can be formulated as:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 142, 926, 191]]<|/det|>
+\[J_{EMS} = \min \sum_{t = 0}^{T} cost(t) + \alpha f_{s}(SOC(t)) \quad (10)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 198, 928, 248]]<|/det|>
+where \(T\) is the total length of the driving cycle, \(cost(t)\) is the energy cost including hydrogen consumption, the cost of battery, and fuel cell degradation, \(f_{s}(SOC(t))\) is the SOC maintaining function, \(\alpha\) the tradeoff between energy cost and SOC.  
+
+<|ref|>text<|/ref|><|det|>[[68, 246, 928, 362]]<|/det|>
+To tackle the sequential decision, the energy management problem is formulated as a Markov Decision Process (MDP), which is a framework for learning the optimal EMS from interaction to minimize total energy cost. The MDP defined by a tuple \((S,A,P,R,\rho_{0},\gamma)\) , where \(S\) denotes the state space, \(A\) denotes the action space, \(P\left(s^{\prime}\mid s,a\right)\) denotes the transition distribution, \(\rho_{0}(s)\) denotes the initial state distribution, \(R(s,a)\) denotes the reward function, and \(\gamma \in (0,1)\) denotes the discount factor. The goal is to find a policy \(\pi (a\mid s)\) that maximizes the expected cumulative discounted rewards \(J(\pi) = E_{\pi ,P,\rho_{0}}\left[\sum_{t = 0}^{\infty}\gamma^{t}R\left(s_{t},a_{t}\right)\right]\) . To use this formulation for FCEV. The state space at time point \(t\) for the FCEV is defined as:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 365, 926, 397]]<|/det|>
+\[S = \left\{v_{t},acc_{t},SOC_{t},P_{fcs}^{t}\right\} \quad (11)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 403, 928, 455]]<|/det|>
+where \(v_{t},a_{t},P_{fcs},SOC_{t}\) are the vehicle speed, acceleration, fuel cell power, and battery SOC. The action represents the control variable, defined as allocating power to the energy sources of the vehicle. In the context of the FCEV, the action is defined as the FC power slope, denoted as \(\Delta P_{fcs}\) . The continuous action can be described as follow:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 460, 926, 496]]<|/det|>
+\[A = \left\{\Delta P_{fcs} = P_{fcs}^{t} - P_{fcs}^{t - 1},\Delta P_{fcs}\in [-10\mathrm{kW},10\mathrm{kW}]\right\} \quad (12)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 501, 928, 585]]<|/det|>
+The reward function \(R\) describes the reward \(R(s_{t + 1};s_{t};a_{t})\) associated with transitioning from state \(s_{t}\) to state \(s_{t} + 1\) using action \(a_{t}\) . The design of the reward function is pivotal in the learning process. For the FCEV, multiple objectives are taken into account, including hydrogen consumption, FC degradation, and battery- related costs such as electricity consumption and degradation. Additionally, it is essential to maintain the battery SOC. Therefore, the reward function is defined as the sum of energy costs while ensuring that the battery charge- sustaining constraints are maintained.  
+
+<|ref|>equation<|/ref|><|det|>[[112, 590, 926, 623]]<|/det|>
+\[R = -\left\{C_{fcs,H_2} + C_{fcs,degr} + C_{bat,eh_2} + C_{bat,degr} + \alpha \left[SOC_{ref} - SOC(t)\right]^2\right\} \quad (13)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 630, 928, 666]]<|/det|>
+The battery electricity consumption \(C_{bat,eh_2}\) is calculated according to the battery charge/discharge efficiency and converted into price cost:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 671, 926, 714]]<|/det|>
+\[C_{bat,eh_2} = \int_0^t \left[P_{bat} / \left(\eta_{d / c}\eta_{DC / DC},\mathrm{LHV}_{H_2}\right)\right]dt\cdot PR_{H_2} \quad (14)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 720, 457, 739]]<|/det|>
+where \(\eta_{d / c}\) is the battery discharge/charge efficiency.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 750, 293, 767]]<|/det|>
+### 3.3. Offline RL algorithm  
+
+<|ref|>text<|/ref|><|det|>[[68, 767, 928, 833]]<|/det|>
+We employ the offline reinforcement learning (ORL) paradigm to address the MDP problem described above. The goal is to learn a policy \(\pi \sim a_{t}(\pi :S\to A)\) that maximizes the expectation of the sum of discounted rewards \(J(\pi)\) Each policy \(\pi\) has a corresponding state- action value function (also known as Q function), which denotes the expected return \(Q(s,a)\) when following the policy \(\pi\) after taking an action \(a\) in state \(s\)  
+
+<|ref|>equation<|/ref|><|det|>[[112, 838, 926, 888]]<|/det|>
+\[Q(s,a) = \mathbb{E}\left[\sum_{t = t}^{\infty}\gamma^{i - t}R_{i}\mid s_{t} = s,a_{t} = a\right] \quad (15)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 895, 928, 930]]<|/det|>
+Here, we approximate the Q function \(Q(s,a)\) using deep neural networks by minimizing the squared Bellman error. The ORL algorithm utilized in our proposal is an extension of the twin delayed deep deterministic policy gradient
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 70, 928, 137]]<|/det|>
+algorithm (TD3) [26]. TD3 is a state- of- the- art online DRL algorithm implemented under the actor- critic framework that learns a deterministic policy. For the actor part, it learns a deterministic target policy by mapping states to a specific action. The update of the Actor network of TD3 algorithm aims to maximize the estimation of the current policy by the critic network.  
+
+<|ref|>text<|/ref|><|det|>[[68, 135, 929, 283]]<|/det|>
+When offline logged datasets are available, it is reasonable to push the policy towards favoring actions contained in the dataset D. Hence, our proposed ORL algorithm augments the standard policy update step in TD3 with Behavior Cloning (BC) regularization to reinforce the policy's focus on the behaviors observed in the dataset D [27]. This regularization term encourages the policy to mimic the demonstrated behaviors more closely, leading to improved generalization and performance. Furthermore, in pursuit of improving the policy, Discriminator Blend (DB) regularization [28] is employed to enhance the flexibility of the policy constraint. This is achieved by integrating a discriminator using Generative Adversarial Networks (GANs) [29]. By incorporating a discriminator, the policy is enabled to explore actions that may not be included in the dataset \(D\) , leading to a more diverse and adaptive policy. This involves utilizing a neural network as an approximator of the policy function \(\pi\) :  
+
+<|ref|>equation<|/ref|><|det|>[[112, 287, 926, 320]]<|/det|>
+\[\pi = \arg \max_{\pi}\mathbb{E}_{(s,a)\sim \mathbf{D}}\left[\lambda Q(s,\pi (s)) - (1 - \beta)(\pi (s) - a)^2 +\beta \log (D(s,\pi (s)))\right] \quad (16)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 326, 928, 425]]<|/det|>
+\(\mathbb{E}()\) is the mathematical expectation. The parameter \(\beta\) (range of 0 to 1) adjusts the balance between BC and DB constraints. The DB is trained to assess whether a given action \(a\) and state \(s\) pair belongs to the dataset \(a_{D}\) or is generated by the policy \(\pi_{\theta}\) , which acts as the generator \(G\) in GANs. This enables BD to effectively regulate the policy learning process by encouraging the policy to explore actions beyond the dataset while ensuring that these actions are plausible according to the discriminator's perception. The \(\lambda\) is a normalization term based on the average absolute value of \(Q\) to control the balance between RL and imitation, defined as:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 429, 926, 475]]<|/det|>
+\[\lambda = \frac{\alpha}{\frac{1}{N}\sum_{(s_i,a_i)}\left|Q(s_i,a_i)\right|} \quad (17)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 481, 928, 549]]<|/det|>
+The parameter \(\alpha\) is used to control the strength of the regularize where the larger \(\alpha\) will make the algorithm approach more RL, and \(N\) represents the number of transitions in the dataset. To normalize the characteristics of each state in the provided dataset. Let \(s_i\) be the \(i\) th feature of the state \(s\) in the dataset, let \(\mu_i,\sigma_i\) be the mean and standard deviation ( \(\eta\) is a constant value to avoid division by zero.):  
+
+<|ref|>equation<|/ref|><|det|>[[112, 553, 926, 589]]<|/det|>
+\[s_i = \frac{s_i - \mu_i}{\sigma_i + \eta} \quad (18)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 595, 928, 689]]<|/det|>
+The critic part estimates the Q- value of a state- action pair. TD3 employs two critic networks to mitigate overestimation bias, with each critic having a corresponding target network. Each critic network \(Q_{1}\left(s,a\mid \theta^{Q_{1}}\right)\) and \(Q_{2}\left(s,a\mid \theta^{Q_{2}}\right)\) corresponds to a target network \(Q_{1}^{\prime}\left(s,a\mid \theta^{Q_{1}^{\prime}}\right)\) and \(Q_{2}^{\prime}\left(s,a\mid \theta^{Q_{2}^{\prime}}\right)\) respectively. The minimum Q- value among the two critics is used as the target Q- value during training. The critic network is updated by minimizing the loss function:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 707, 926, 741]]<|/det|>
+\[L\left(\theta^{Q_{i}}\right) = \mathbb{E}\left[\left(y_{t} - Q_{i}\left(s_{t},a_{t}\mid \theta^{Q_{i}}\right)\mid a_{t} = \mu \left(s_{t}\mid \theta^{\mu}\right)\right)^{2}\right] \quad (19)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 747, 928, 782]]<|/det|>
+where \(\theta^{Q}\) denotes the weights of the critic network. The target Q- value \(y\) is evaluated by taking the minimum of the estimates from the two Q- functions as follows:  
+
+<|ref|>equation<|/ref|><|det|>[[112, 788, 926, 822]]<|/det|>
+\[y_{t} = r\left(s_{t},a_{t}\right) + \gamma \min_{i = 1,2}Q_{i}^{\prime}\left(s_{t + 1},a_{t + 1}\mid \theta^{Q_{i}^{\prime}}\right) \quad (20)\]  
+
+<|ref|>text<|/ref|><|det|>[[68, 832, 928, 867]]<|/det|>
+where \(a_{t + 1}\sim \pi_{\phi^{\prime}}\left(s_{t + 1}\right) + \epsilon ,\quad \epsilon \sim \mathrm{clip}(\mathcal{N}(0,\tilde{\sigma}), - c,c)\) is the exploration noise to smooth the value estimates and improve robustness of the learned \(Q\) functions, \(r\) is the instantaneous one- step reward, \(\gamma\) is the discounting factor.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 880, 263, 897]]<|/det|>
+### 3.4. Baseline Methods  
+
+<|ref|>text<|/ref|><|det|>[[68, 897, 928, 930]]<|/det|>
+We use a series of baseline EMS methods for comparatively evaluating the ORL method. The inputs and outputs of all baselines are the same as those of the proposed method.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[68, 71, 928, 170]]<|/det|>
+Dynamic Programming (DP) [30]: In optimization control methods, the EMS problem is formulated as a nonlinearly constrained optimization problem, aiming to minimize the objective function presented in Equation (10). DP is an optimization control method that operates by seeking the shortest path backward in time. Its objective is to derive the minimum cost function for each grid at every stage in reverse chronological order. In our study, DP is used as the benchmark EMS policy, representing the global optimum and providing upper limits for comparison. It's important to recognize that DP requires future information as input to achieve the optimization objective.  
+
+<|ref|>text<|/ref|><|det|>[[68, 168, 928, 250]]<|/det|>
+Behavior Cloning (BC) [31]: BC, as a fundamental imitation learning approach, seeks to emulate the EMS policy by directly learning from the provided dataset, which is assumed to be generated by an expert policy or near- expert policy. It employs supervised learning techniques to train a model to map states to actions. Both BC and ORL involve learning from data for EMS applications. In this context, we establish BC as the benchmark and aim to showcase the superior performance of ORL.  
+
+<|ref|>text<|/ref|><|det|>[[68, 248, 928, 346]]<|/det|>
+Proximal Policy Optimization (PPO) [32]: PPO is a state- of- the- art online DRL algorithm, which has been extensively applied in various applications requiring sophisticated decision- making in dynamic environments. PPO offers a robust and efficient approach to training agent by leveraging on- policy learning, effective use of data through mini- batch updates, stability through policy clipping, and adaptive learning rates. Leveraging the strengths of PPO, we utilize it to generate the dataset necessary for ORL, with its policy serving as an expert (near- optimal) strategy for comparison purposes. We provide it to explore the superiority of ORL compared to the online DRL.  
+
+<|ref|>text<|/ref|><|det|>[[68, 344, 928, 443]]<|/det|>
+Twin Delayed Deep Deterministic Policy Gradient (TD3) [26]: TD3 is an advanced online DRL algorithm, stemming from the Actor- Critic framework. It has garnered significant attention due to its effectiveness in overcoming challenges associated with continuous action spaces and high- dimensional state spaces. TD3 employs twin critic networks to estimate the value of actions more accurately. By utilizing two critic networks, TD3 mitigates overestimation bias and enhances the robustness of value function estimation. We also provide it to explore the superiority of ORL compared to the online DRL.  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 466, 197, 484]]<|/det|>
+## 4. Discussion  
+
+<|ref|>text<|/ref|><|det|>[[68, 488, 928, 586]]<|/det|>
+In conclusion, we have presented a novel data- driven EMS for hybrid energy systems in EVs. Leveraging an innovative offline reinforcement learning agent, our approach learns directly from driving data. Experimental results demonstrate that the ORL agent not only learns optimal EMS strategies from expert data but also exhibits the ability to learn superior EMS from datasets containing a mixture of expert and noisy data, and even achieves near- optimal strategies from entirely noisy datasets. Moreover, our approach demonstrates that with increased data availability, performance improves as the agent is trained with more data.  
+
+<|ref|>text<|/ref|><|det|>[[68, 584, 928, 794]]<|/det|>
+This approach offers three notable benefits. Firstly, it is sufficiently simple, as it solely relies on collected data for automatic learning by the agent, unlike the traditional EMS development process, which often requires extensive expert knowledge and repeated measurements. Furthermore, the data used in our approach are non- expert data readily available from real vehicles. Secondly, our method ensures stable performance by integrating seamlessly with existing EMS without altering the original EMS performance lower bound. Our approach continuously improves upon the baseline EMS through data- driven enhancements leveraging the strengths of both technologies. For example, to address the performance shortcomings in rule- based EMS, ORL enables incremental learning, allowing for the continual enhancement of EMS performance with historical data. Similarly, ORL addresses the sim- to- real gap problem in simulation- based methods by enhancing pretrained EMS models, thereby ensuring their effectiveness in real- world deployment scenarios. Lastly, our approach exhibits versatility, as with sufficient data, it can learn a generalized EMS applicable to various EVs and operating conditions. This aligns with the current trend of large- scale language models and similar approaches in artificial intelligence, where a single large model with large- scale data can be trained to perform well across diverse tasks and domains.  
+
+<|ref|>text<|/ref|><|det|>[[68, 793, 928, 875]]<|/det|>
+Overall, we believe that ORL could serve as a foundational framework for data- driven EMS, with potential applications extending beyond EVs to grid EMS, industrial energy management systems, and other vehicle control systems. However, a limitation of this work is that the ORL agent may require more data to further enhance its performance. Addressing this limitation could involve exploring methods to efficiently gather and utilize additional data for agent training, potentially improving its effectiveness in real- world applications.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[70, 75, 250, 92]]<|/det|>
+## Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[92, 97, 900, 115]]<|/det|>
+AcknowledgementsThis work was supported in part by the National Natural Science Foundation of China (Grant No. 52172377).  
+
+<|ref|>sub_title<|/ref|><|det|>[[70, 138, 276, 155]]<|/det|>
+## Author Contributions  
+
+<|ref|>text<|/ref|><|det|>[[70, 161, 928, 210]]<|/det|>
+Y.W. designed the study and methodology; Y.W., J.Wu. and H.H. collected and analyzed data; Y.W. generated the figures; Y.W., J.Wu. and W.Z. wrote the manuscript; H.H., W.Z. and F.S. reviewed and edited the manuscript. All authors contributed to the paper.  
+
+<|ref|>sub_title<|/ref|><|det|>[[70, 234, 172, 250]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[68, 250, 930, 930]]<|/det|>
+[1] B. Borlaug, M. Muratori, M. Gillern, D. Woody, W. Muston, T. Canada, A. Ingram, H. Gresham, C. McQueen, Heavy- duty truck electrification and the impacts of depot charging on electricity distribution systems, Nature Energy 6 (2021) 673- 682. [2] Y. Zhao, Z. Wang, Z.- J. M. Shen, F. Sun, Assessment of battery utilization and energy consumption in the large- scale development of urban electric vehicles, Proceedings of the National Academy of Sciences 118 (2021) e2017318118. [3] H. He, X. Meng, Y. Wang, A. Khajepour, X. An, R. Wang, F. Sun, Deep reinforcement learning based energy management strategies for electrified vehicles: Recent advances and perspectives, Renewable and Sustainable Energy Reviews 192 (2024) 114248. [4] Y. Li, H. He, A. Khajepour, Y. Chen, W. Huo, H. Wang, Deep reinforcement learning for intelligent energy management systems of hybrid- electric powertrains: Recent advances, open issues, and prospects, IEEE Transactions on Transportation Electrification (2024). [5] A. H. Ganesh, B. Xu, A review of reinforcement learning based energy management systems for electrified powertrains: Progress, challenge, and potential solution, Renewable and Sustainable Energy Reviews 154 (2022) 111833. [6] F. Zhang, L. Xiao, S. Coskun, H. Pang, S. Xie, K. Liu, Y. Cui, Comparative study of energy management in parallel hybrid electric vehicles considering battery ageing, Energy 264 (2023) 123219. [7] J. Wu, C. Huang, H. He, H. Huang, Confidence- aware reinforcement learning for energy management of electrified vehicles, Renewable and Sustainable Energy Reviews 191 (2024) 114154. [8] X. Tang, T. Jia, X. Hu, Y. Huang, Z. Deng, H. Pu, Naturalistic data- driven predictive energy management for plug- in hybrid electric vehicles, IEEE Transactions on Transportation Electrification 7 (2020) 497- 508. [9] Y. Wang, Y. Wu, Y. Tang, Q. Li, H. He, Cooperative energy management and eco- driving of plug- in hybrid electric vehicle via multi- agent reinforcement learning, Applied Energy 332 (2023) 120563. [10] X. Hu, T. Liu, X. Qi, M. Barth, Reinforcement learning for hybrid and plug- in hybrid electric vehicle energy management: Recent advances and prospects, IEEE Industrial Electronics Magazine 13 (2019) 16- 25. [11] Y. Wang, H. Tan, Y. Wu, J. Peng, Hybrid electric vehicle energy management with computer vision and deep reinforcement learning, IEEE Transactions on Industrial Informatics 17 (2020) 3857- 3868. [12] W. Chen, J. Peng, J. Chen, J. Zhou, Z. Wei, C. Ma, Health- considered energy management strategy for fuel cell hybrid electric vehicle based on improved soft actor critic algorithm adopted with beta policy, Energy Conversion and Management 292 (2023) 117362. [13] W. Yuankai, L. Renzong, W. Yong, L. Yi, Benchmarking deep reinforcement learning based energy management systems for hybrid electric vehicles, in: CAAI International Conference on Artificial Intelligence, Springer, 2022, pp. 613- 625. [14] Y. Wang, R. Lian, H. He, J. Betz, H. Wei, Auto- tuning dynamics parameters of intelligent electric vehicles via bayesian optimization, IEEE Transactions on Transportation Electrification (2023). [15] A. R. Mayyas, S. Kumar, P. Pisu, J. Rios, P. Jethani, Model- based design validation for advanced energy management strategies for electrified hybrid power trains using innovative vehicle hardware in the loop (vhil) approach, Applied Energy 204 (2017) 287- 302. [16] H. He, Z. Niu, Y. Wang, R. Huang, Y. Shou, Energy management optimization for connected hybrid electric vehicle using offline reinforcement learning, Journal of Energy Storage 72 (2023) 108517. [17] G. Pozzato, A. Allam, L. Pulvirenti, G. A. Negoita, W. A. Paxton, S. Onori, Analysis and key findings from real- world electric vehicle field data, Joule 7 (2023) 2035- 2053. [18] K. A. Severson, P. M. Attia, N. Jin, N. Perkins, B. Jiang, Z. Yang, M. H. Chen, M. Aykol, P. K. Herring, D. Fraggedakis, et al., Data- driven prediction of battery cycle life before capacity degradation, Nature Energy 4 (2019) 383- 391. [19] W. Li, J. Zhu, Y. Xia, M. B. Gorji, T. Wierzbicki, Data- driven safety envelope of lithium- ion batteries for electric vehicles, Joule 3 (2019) 2703- 2715. [20] J. Zhang, Y. Wang, B. Jiang, H. He, S. Huang, C. Wang, Y. Zhang, X. Han, D. Guo, G. He, et al., Realistic fault detection of li- ion battery via dynamical deep learning, Nature Communications 14 (2023) 5940. [21] D. Roman, S. Saxena, V. Robu, M. Pecht, D. Flynn, Machine learning pipeline for battery state- of- health estimation, Nature Machine Intelligence 3 (2021) 447- 456. [22] N. Guo, S. Chen, J. Tao, Y. Liu, J. Wan, X. Li, Semi- supervised learning for explainable few- shot battery lifetime prediction, Joule (2024). [23] F. Millo, L. Rolando, L. Tresca, L. Pulvirenti, Development of a neural network- based energy management system for a plug- in hybrid electric vehicle, Transportation Engineering 11 (2023) 100156. [24] R. F. Prudencio, M. R. Maximo, E. L. Colombini, A survey on offline reinforcement learning: Taxonomy, review, and open problems, IEEE Transactions on Neural Networks and Learning Systems (2023). [25] H. He, F. Sun, Z. Wang, C. Lin, C. Zhang, R. Xiong, J. Deng, X. Zhu, P. Xie, S. Zhang, et al., China's battery electric vehicles lead the world: achievements in technology system architecture and technological breakthroughs, Green Energy and Intelligent Transportation 1 (2022) 100020.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[66, 72, 930, 262]]<|/det|>
+[26] S. Fujimoto, H. Hoof, D. Meger, Addressing function approximation error in actor- critic methods, in: International conference on machine learning, PMLR, 2018, pp. 1587- 1596. [27] S. Fujimoto, S. S. Gu, A minimalist approach to offline reinforcement learning, Advances in neural information processing systems 34 (2021) 20132- 20145. [28] S. Kidera, K. Shintani, T. Tsuneda, S. Yamane, Combined constraint on behavior cloning and discriminator in offline reinforcement learning, IEEE Access (2024).[29] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A. A. Bharath, Generative adversarial networks: An overview, IEEE signal processing magazine 35 (2018) 53- 65. [30] P. Saiteja, B. Ashok, Critical review on structural architecture, energy control strategies and development process towards optimal energy management in hybrid vehicles, Renewable and Sustainable Energy Reviews 157 (2022) 112038. [31] A. Block, A. Jadbabaie, D. Pfrommer, M. Simchowitz, R. Tedrake, Provable guarantees for generative behavior cloning: Bridging low- level stability and high- level behavior, Advances in Neural Information Processing Systems 36 (2024).[32] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal policy optimization algorithms, arXiv preprint arXiv:1707.06347 (2017).
+
+<--- Page Split --->

preprint/preprint__0913196261d535330330d1e998de6b5f2eff345031b264d7786eff4a9e3ac6a7/preprint__0913196261d535330330d1e998de6b5f2eff345031b264d7786eff4a9e3ac6a7.mmd

@@ -0,0 +1,478 @@
+
+# Mapping Decidualization Resistance in Former Severe Preeclampsia Patients at Multi-Omic Levels  
+
+Tamara Garrido- Gómez tgarrideo@fundacioncarlossimon.com  
+
+Carlos Simon Foundation - INCLIVA Health Research Institute Irene Muñoz- Blat Carlos Simon Foundation - INCLIVA Health Research Institute Raul Pérez- Moraga Carlos Simon Foundation - INCLIVA Health Research Institute https://orcid.org/0000- 0002- 9611- 9123  
+
+Nerea Castillo Marco Carlos Simon Foundation - INCLIVA Health Research Institute  
+
+Nerea Castillo Marco Carlos Simon Foundation - INCLIVA Health Research Institute  
+
+Ana Ochando Carlos Simon Foundation - INCLIVA Health Research Institute  
+
+Sheila Ortega Carlos Simon Foundation - INCLIVA Health Research Institute  
+
+Marcos Parras Carlos Simon Foundation - INCLIVA Health Research Institute  
+
+Rogelio Monfort Hospital Universitario y Politecnico La Fe  
+
+Elena Satorres- Perez Hospital Universitario y Politecnico La Fe  
+
+Blanca Novillo Hospital Universitario y Politecnico La Fe  
+
+Alfredo Perales Hospital Universitario La Fe, Valencia Spain.  
+
+Matthew Gormley University California San Francisco  
+
+Beatriz Roson Carlos Simon Foundation- INCLIVA Health Research Institute, 46012 Valencia, Spain https://orcid.org/0000- 0002- 9851- 2025  
+
+Susan Fisher University California San Francisco
+
+<--- Page Split --->
+
+# Carlos Simon  
+
+Carlos SimonCarlos Simon Foundation- INCLIVA Health Research Institute https://orcid.org/0000- 0003- 0902- 9531  
+
+## Article  
+
+# Keywords:  
+
+Posted Date: May 6th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4331532/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+## Additional Declarations:  
+
+There is NO Competing Interest.  
+
+Supplementary Figure 1 is not available with this version.  
+
+Version of Record: A version of this preprint was published at Nature Medicine on January 7th, 2025. See the published version at https://doi.org/10.1038/s41591- 024- 03407- 7.
+
+<--- Page Split --->
+
+# Mapping Decidualization Resistance in Former Severe Preeclampsia Patients at Multi-Omic Levels  
+
+1 Mapping Decidualization Resistance in Former Severe Preeclampsia 2 Patients at Multi-Omic Levels 3 4 Irene Muñoz- Blat<sup>1,2,8</sup>, Raul Perez- Moraga<sup>1,3,8</sup>, Nerea Castillo- Marco<sup>1,2,8</sup>, Teresa Cordero<sup>1,2</sup>, Ana 5 Ochando<sup>1</sup>, Sheila Ortega<sup>1</sup>, Marcos Parras<sup>1,2</sup>, Rogelio Monfort<sup>4</sup>, Elena Satorres- Perez<sup>4</sup>, Blanca Novillo<sup>4</sup>, 6 Alfredo Perales<sup>4</sup>, Matthew Gormley<sup>5</sup>, Beatriz Roson<sup>1,2</sup>, Susan Fisher<sup>5</sup>, Carlos Simón<sup>1,6,7</sup>, Tamara 7 Garrido- Gómez<sup>1,2</sup> 8 9 <sup>1</sup>Carlos Simon Foundation, Valencia, Spain 10 <sup>2</sup> INCLIVA Health Research Institute, Valencia, Spain 11 <sup>3</sup> R&D Department, Igenomix, Valencia, Spain 12 <sup>4</sup> Hospital Universitario y Politecnico La Fe, Valencia, Spain 13 <sup>5</sup> University California San Francisco, San Francisco, CA, USA 14 <sup>6</sup>Department of Pediatrics, Obstetrics and Gynecology, University of Valencia, Valencia, Spain 15 <sup>7</sup>Department of Obstetrics and Gynecology, Beth Israel Deaconess Medical Center, Harvard Medical 16 School, Boston, MA, USA 17 <sup>8</sup>These authors contributed equally 18 <sup>#</sup> Correspondence; carlos.simon@uv.es; tgarrido@fundacioncarlossimon.com
+
+<--- Page Split --->
+
+Endometrial decidualization resistance (DR) is implicated in various gynaecological and obstetric conditions. Employing a multi- omic strategy, we unraveled the cellular and molecular characteristics of DR in patients that have suffered severe preeclampsia (sPE). Morphological analysis unveiled significant glandular anatomical abnormalities, confirmed histologically. Single- cell RNA sequencing (scRNA- seq) of endometrial samples from sPE cases (n=11) and controls (n=12) revealed sPE- associated shifts in cell composition, manifesting as a stromal mosaic state characterized by proliferative stromal cells (MMP11, SFRP1+) alongside IGFBP1+ decidualized cells, with concurrent epithelial mosaicism and a dearth of epithelial- stromal transition associated with decidualization. Cell- cell communication network mapping underscored aberrant crosstalk among specific cell types, implicating crucial pathways such as endoglin and WNT. Spatial transcriptomics in a replication cohort validated DR- associated features. Laser capture microdissection/mass spectrometry in a second replication cohort corroborated several scRNA- seq findings, notably the absence of stromal to epithelial transition at a pathway level, indicating disrupted response to steroid hormones, particularly estrogens. These insights shed light on potential molecular mechanisms underpinning DR pathogenesis in the context of sPE.
+
+<--- Page Split --->
+
+## Introduction  
+
+Pregnancy health is shaped during the periconceptional period due to the interplay between the implanting embryo and the maternal endometrium (1). Decidualization entails the functional and morphological changes that occur within the endometrium transforming the maternal uterine lining to accommodate the invasive trophoblast (2- 4). The endometrial stromal cells (ESCs) transformation is hormonally regulated, driven by increasing progesterone levels and local cAMP production (5, 6), which stimulate the synthesis of a complex network of intracellular and secreted proteins (7). It starts in the early secretory phase of the menstrual cycle, independent of the presence of a conceptus, in areas adjacent to the uterine spiral arteries thereafter spreading throughout the endometrium (8).  
+
+Decidualization resistance (DR) refers to the inability of the endometrium to undergo these specific changes and has been documented in reproductive disorders including endometriosis (9- 11), miscarriage (12), recurrent pregnancy loss (13, 14) and the great obstetrical syndromes. The latter include preeclampsia (PE) (15- 17), intrauterine growth restriction (IUGR) (18), and/or placenta accrete spectrum disorder (19).  
+
+Preeclampsia is a major- obstetric complication affecting \(8\%\) of first- time pregnancies. It is characterized by the new onset of hypertension, proteinuria, and other signs such as vascular endothelial damage (20). The life- threatening condition known as severe PE (sPE) is diagnosed based on higher blood pressure criteria (systolic \(\geq 160 \mathrm{mm} \mathrm{Hg}\) or diastolic of \(\leq 100 \mathrm{mm} \mathrm{Hg}\) ), symptoms of central nervous system dysfunction, hepatic abnormalities, thrombocytopenia, renal abnormalities, and/or pulmonary edema (21). sPE is a placental insufficiency syndrome mediated by early shallow cytotrophoblast (CTB) invasion of uterine decidua and spiral arterioles, leading to incomplete endovascular invasion and altered uteroplacental perfusion (20, 22, 23). We provided evidence of a decidualization defect in women with sPE, detected at the time of delivery and persisting years after the affected pregnancy (15). Endometrial bulk RNA- seq results from affected women post- sPE revealed altered ovarian hormone receptor signaling pathways (16).  
+
+Here, we initially observed gross and microscopic endometrial glandular defects in late secretory endometrium from former sPE patients. Then, we assemble a spatially resolved single- cell multionic characterization of the DR using sPE as a clinical model. Our objective is to gain insights into this pathological condition, combining different omics techniques including single- cell RNA sequencing (scRNA seq), spatial transcriptomics and laser capture microdissection coupled mass spectrometry [LCM- MS]) with different samples sets. Thus, we provide the first detailed description of the whole status of cells, cell communications, perturbations in specific cell communication pathways at scRNA- seq, spatial and proteomic resolution.
+
+<--- Page Split --->
+
+## Results  
+
+Decidualization resistance (DR) in former sPE patients: associated changes at gross, microscopic and single cell levels.  
+
+Endometrial tissue was collected during the late secretory phase of the cycle from women with a previous sPE pregnancy (Fig. 1a- d) and control individuals who had normal obstetric outcomes (Fig. 1e- h). Initial examination of the samples (n=8/group) (Supplementary Table 1a) consistently showed distinct morphological features of the endometrium from cases that were not observed in control specimens. Specifically, the openings of the glands were dilated in the cases (Fig. 1a,b) vs. the control group (Fig. 1e,f). Histological analysis of H&E stained, longitudinal and sagittal tissue sections confirmed this finding and showed that the dilation involved the entire gland (compare Fig. 1c,d vs. Fig. 1g,h). We took these findings as preliminary evidence of DR endometrial defects, during the late secretory phase of the cycle, among the former sPE patients.  
+
+Thus, we used a scRNA- seq approach to uncover the molecular correlates of the observed morphological differences. We profiled biopsies of late secretory phase endometrium from former sPE patients (n=11) as compared to equivalent samples from control individuals who had normal obstetric outcomes (n=12) (Supplementary Table 1b). The scRNA- seq workflow is outlined in Fig. 1i. Altogether we profiled 65,381 high quality cells (see Methods) and integrated the transcriptomes in a uniform manifold approximation and projection (UMAP). Of these 28,154 were from the sPE group (Fig. 1j) and 37,227 were from the control group (Fig. 1k).  
+
+The cells were clustered using previously published scRNA- seq datasets and markers (2, 3, 24). Nine major cell types were identified: epithelium, stroma, cycling stroma, perivascular, endothelium, NK cells, T cells, macrophages, and B cells. The expression of canonical markers of each cell type is shown as a dotplot in Extended Data Fig. 1a. Global integration of the merged cases and control datasets enabled an initial comparison of the cellular composition of the endometrial biopsies from the two groups (Fig. 1l). Differences in the clustering of stromal and epithelial cells was immediately apparent, the samples from the former sPE patients lacked subpopulations that were more prominent in the controls, and vice versa. This finding prompted us to perform a detailed comparison of these cell types from the two sources.  
+
+## Endometrial cell type-specific differentiation defects in former sPE patients: Stroma.  
+
+Integration of the stromal and perivascular cell transcriptomes identified several subpopulations, including three types of decidual cells (Fig. 2a). Decidualized stroma 1 cells expressed genes related to ribosome activity (RPS29, RPL37A) and progesterone- associated endometrial protein (PAEP).
+
+<--- Page Split --->
+
+Decidualized stroma 2 cells co- expressed markers of the pre- decidual and decidual states, \(EGR1\) and \(CXCL2\) (25, 26), respectively. Decidualized stroma 3 cells were characterized by well- known decidualization markers such as \(IGFBP1\) and \(TIMP3\) (27), \(ATF3\) (28), and transcription factors that are associated with late decidualization, including \(CSRNP1\) (3).  
+
+We also identified a stroma EMT subpopulation (stromal transition), this cluster expressed \(PDGFRA\) that is involved in the transition that occurs during endometrial regeneration after menstruation and decidualization at the maternal- fetal interface (27). Another cluster is the proliferative stroma expressing markers of the proliferative phase of the menstrual cycle: \(SFRP4\) (29), \(MMP11\) (30, 31), \(DIO2\) (32), \(SFRP1\) (33) and \(PGRMC1\) (34). We also captured two distinct perivascular cell subpopulations that maintain a close relationship with stromal cells during decidualization. The perivascular 1 (STEAP4+) and perivascular 2 (MYH11+) subpopulations clustered near cells that regulate stromal angiogenesis, that expressed \(CCBE1\) , a regulator of \(VEGFC\) (35, 36). The complete set of stromal cell subtypes and their differential gene expression is shown in Extended data Fig. 1b.  
+
+Mapping the identity of the cells enabled identification of subpopulations that were differentially abundant in sPE or control samples (Fig. 2b). Specifically, decidualized stroma 1 and 2 as well as the stromal transition subpopulation were enriched in control samples, whereas decidualized stroma 3 and the proliferative stroma were increased in endometrium from sPE samples. This mosaic state—proliferative stromal cells (MMP11, \(SFRP1+\) ) coexisting with \(IGFBP1+\) decidualized cells—was the hallmark of DR in sPE. We graphed the data as neighborhoods in which dots represent neighborhood sizes and the weight of the connecting lines depicts the number of cells shared among neighborhoods (Fig. 2c). The results confirmed a statistically significant increase in the number of cells that defined this mosaic state as well as the decrease of decidualized stromal 1, 2 and stromal transitioning cells in sPE samples. The neighborhood data were also visualized as a Beeswarm plot (Fig. 2d).  
+
+Differential gene expression analysis revealed a significant number of dysregulated genes (DEGs) associated with DR in all the affected stromal subpopulations (Log2FC>0.5, and p. adjusted<0.05) (Extended Data Fig. 2a). In cases, decidualized stroma 1 and 2 cells upregulated \(TIMP3\) , and decidual stroma 3 cells upregulated \(IGFBP1\) and \(IGFBP6\) . Also in sPE cases, the proliferative stroma subpopulation was characterized by upregulation of \(SFRP4\) present in the proliferative phase (37) and \(FOS\) that is present in the endometrium of endometriosis patients (38). Additionally, in sPE cases endothelial cells showed an upregulation of \(C2CD4B\) (39), a marker of acute inflammatory response (Extended Data Fig. 2b). Consistently, immune cells from sPE, including macrophages, natural killer and B cells showed an aberrant transcriptomic profile indicative of inflammatory dysregulation, highlighted by higher expression of \(IL1B\) , \(CCL5\) (40), \(SCGB1D2\) (41), respectively (Extended Data Fig. 2c). These results point to an imbalance in stromal populations in sPE with a predominant aberrant stage of proliferating stromal cells coexisting with decidualised cells in a proinflammatory niche.
+
+<--- Page Split --->
+
+## Endometrial cell type-specific differentiation defects in former sPE patients: Epithelia  
+
+We integrated the epithelial cell transcriptomes, which identified the following epithelial cell subtypes (Fig. 2e): ciliated epithelium (which expressed TPP3, RSPH1, and C1orf194), additionally, these cells expressed FOXJ1 and PIFO as previously shown (3). Preciliated epithelial cells expressed CDC20B and CCNO. We also identified specific PDGFRA+ ciliated epithelial cells which were only present in former sPE patients (sPE ciliated epithelium). The population transitioning between epithelium and stroma (epithelial transition) expressed genes identified in the stromal cells, including TIMP3 and DCN. Luminal epithelial cells were identified by published markers: TGS1 and MSLN (3). Glandular secretory cells—the most abundant epithelial cell type—expressed PAEP, CXCL14, and SPP1. As with the stroma, we identified a subpopulation of epithelial cells expressing markers of the proliferative phase, such as IHH and EMID1. The complete set of epithelial cell subtypes from cases and controls and their differentially expressed genes is shown in Extended data Fig. 1c.  
+
+Mapping the identity of the cells enabled identification of subpopulations that were differentially abundant in sPE or control samples (Fig. 2f). Specifically, the proliferative epithelium, ciliated epithelium and sPE ciliated epithelium subpopulations were more numerous in sPE. It has been reported that the numbers of ciliated epithelial cells increase during the proliferative phase (42), supporting the concept of abnormal epithelial differentiation in DR. Cell types that were enriched in the control group (glandular epithelial cells and epithelial transition) comprised a small fraction of the sPE samples. Next, we constructed neighbourhood graphs of the data (Fig. 2g). The results confirmed the statistically significant, differential abundances of the afore mentioned cell types. The neighbourhood data were also visualized as a Beeswarm plot (Fig. 2h).  
+
+Analysis of gene expression revealed significant dysregulation associated with DR in the differentially abundant epithelial subpopulations (Log2FC \(>0.5\) and Spatial FDR \(< 0.1\) ) (Extended data Fig. 2d). VIM, PGR, and MMP7 were upregulated in the proliferative epithelium of samples from former sPE patients. The sPE epithelium subpopulation was characterized by upregulation of SRFP4 and IGF1, and downregulation of CXCL14, PAEP and TIMP3; the ciliated epithelial cells downregulated MT1G, CXCL14 and GPX3. As to the subpopulations that were less abundant in former sPE patients, the glandular epithelium cells from these individuals had dysregulated expression of several secretoglobins and reduced mRNA levels of the transcriptional regulator, MECOM. Also epithelium transition had higher expression of SERF4, FOS, JUN, and lower expression of CXCL14 and TIMP3, among other.  
+
+Altogether, we confirm that there is a notable imbalance in epithelial populations in former sPE patients with a predominant aberrant stage of proliferative epithelium and ciliated epithelium coexisting with glandular secretory and luminal epithelium together with various aberrant epithelial cell types that might contribute to the epithelial phenotype observed macroscopically and microscopically (Fig 1a- h).
+
+<--- Page Split --->
+
+## Endometrial cell type-specific differentiation defects in former sPE patients: Epithelia to Stroma Transition.  
+
+We integrated the transcriptomes of the subpopulations that were either the precursors or the products of the epithelia to stroma transition zone (Fig. 3a; Extended data Fig. 3). This analysis included: stroma, epithelium, epithelium- stromal transition, ciliated epithelium and cycling cells. The complete set of epithelial cell subtypes from both sample groups and their differentially expressed genes is shown in Extended data Fig. 1d. Comparing both conditions (Fig. 3b) there was a notable absence of all the subtypes in sPE samples, including those in the transition zone. This conclusion was substantiated by neighbourhood analysis (Fig. 3c) and the data were visualized as a beeswarm plot (Fig. 3d). Thus, cells involved in the epithelia- to- mesenchymal transition (EMT), which were abundant in the control samples, were greatly reduced in the endometrium from former sPE patients.  
+
+To delve deeper into understanding of the epithelial- to- stromal transition, we performed RNA velocity analysis followed by cellDancer correction of the combined datasets were impacted in late secretory phase endometrial biopsies from former sPE patients (Extended Data Fig. 3a). The results revealed a temporal dynamic transcriptional process originating from epithelial cells and extending to stromal cells. We depicted this dynamic using cellCondiments obtaining two lineages (Extended Data Fig. 3b). Linage 1, representing the most interesting progression alongside the epithelium, epithelium- stromal transition and stroma cell populations, was projected in both sPE and control samples (Fig. 3e). The differentiation vector map for lineage 1 shows a clear transition from secretory glandular epithelium to stroma describing an EMT associated with late secretory decidual state in controls. In contrast, we found significant density disturbances that are associated with DR in formerly sPE patients (Extended Data Fig. 3c and d). We identify genes whose expression patterns change along the pseudotime (Fig. 3f; Extended Data Fig. 4). Epithelial cells reported with the expression of genes (DUSP2, IL17C, and FTH1), transition zone expressed genes associated with the EMT (SNAI2 and NOTCH3) and finally stromal cells are identified by the expression of collagens (COL6A1, COL6A2), GRIA3 and B3GALT5. Finally, differential expression between conditions highlighted the substantial contribution of controls to the transition (e.g., IGFBP2 and ACTA2) (Fig. 3g). These results suggested not only a decrease but also a dysregulation of fundamental aspects of stromal cell transition to an epithelial phenotype in DR endometrial samples.  
+
+## Cell-cell communications are dysregulated among endometrial cells in sPE  
+
+A myriad of cell- cell communication (CCC) networks is essential for properly responding to changing hormone levels during decidualization. We used CellChat to infer the ligand- receptor pairs that were involved in endometrial CCC networks in the control vs. the sPE datasets with statistical significance.
+
+<--- Page Split --->
+
+The relative and absolute flow of information, calculated using the total interaction probability among all cell subpopulations in both conditions, is depicted in Extended Data Fig. 5a,b. Signals that were stronger in samples from sPE samples included components of the NRG, BMP, CX3C, and EGF pathways. Signals that were stronger in control samples included components of the IL1 and MADCAM pathways. In general, there was a notable increase in signals flowing among cells obtained from former sPE patients. Next, we mapped the perturbations in communication networks to specific cell types (Extended Data Fig. 6). The affected cells included glandular secretory epithelium, stroma, stroma early secretory, and stromal transition. Consistent with the overall pattern of information flow (Extended Data Fig. 6a), many of the outgoing (Extended Data Fig. 6b) and incoming signals (Extended Data Fig. 6c) in cells that comprised the sPE samples were largely absent in the comparable cell subpopulations from the control samples. Overall, mapping of the affected networks to particular cell types reaffirmed the phase shift of specific stromal cell subpopulations in the endometrium from former sPE patients.  
+
+Next, we investigated the ligand- receptor pairs contributing to the sPE- associated CCC network perturbations and the cell types that were involved. The most statistically significant pathways and those that play particularly important roles in decidualization are displayed as chord plots (Fig. 4a,c,e,g,i,j) with the relative contribution of specific ligand receptor pairs shown as bar graphs (Fig. 4b,d,f,h).  
+
+In control samples, endothelin- mediated signals are directed between the endothelium and glandular secretory epithelium to the various types of stromal cell subtypes (Fig. 4a). In DR cases, however, the signals are mainly redistributed between epithelial early proliferative and different stromal cell subpopulations. This dispersion involves weakening of EDN1- EDNRA signals and the new appearance of signaling mediated by EDN3- EDNRB (Fig. 4b) in sPE. Canonical and non- canonical WNT (ncWNT) signals are key regulators of decidualization. As to the latter, in control samples, the stromal transition subpopulation is strongly autoregulating and communicating with endothelial cells and the perivascular subpopulations (Fig. 4c). In samples from sPE cases, DR signals emerging from the stromal transition subpopulation with endothelium and the stroma are diminished and autocrine stroma early secretory regulation emerged (Fig. 4c), also involving new ligand- receptor pairs such as WNT5A- FZD10 (Fig. 4d). In control samples, canonical WNT signalling primarily involves autocrine regulation of the stromal transition cell type and communication between this subpopulation and stromal decidual cells (Fig. 4e). In cases, stromal transition cells and early secretory stromal cells signal to undifferentiated stromal cells (Fig. 4e), involving at least 4 canonical WNT pathways that were not active in control samples (Fig. 4f).  
+
+There were also obvious alterations in signaling via SEMA3A, which is highly expressed in the proliferative phase of the cycle (24). In control samples, the stromal transition subpopulation communicates with endothelial cells and stroma (Fig. 4g). In DR samples, signals had a very different pattern, primarily distributed between proliferative epithelial cells and numerous other types (e.g.,
+
+<--- Page Split --->
+
+decidual cells, proliferative stromal cells, and endothelial cells; Fig. 4g). The latter signals included many pathways that were not observed in control samples (Fig. 4h and Extended data Fig. 7 and 8). We also investigated communication pathways involving SPP1, which is highly expressed during the mid- late secretory phase by epithelial cells (43). In control samples, SPP1 signals were weakly distributed among numerous cell types (Fig. 4i). In samples from the DR cases, strong autocrine (glandular epithelium) and paracrine signals (between glandular epithelium and numerous stromal a decidual cell types; Fig. 4i) emerged. Finally, dramatic changes in signaling via tenascin, an extracellular matrix protein, were also evident (Fig. 4j). This molecule is highly express in endometriosis (44) and in endometrium during the proliferative phase (45). In controls, communication was primary between stromal decidual cell type 3 and the epithelial subtypes. In DR cases, new circuitry arose involving glandular secretory epithelium and stromal cells as well as the various subtypes of decidual cells. In conclusion, the data suggest that in the endometrium of the cases with DR there is a very significant rewiring of signaling pathways in terms of molecules and cell types. Altogether the proliferative phenotype of cells from the cases leads to aberrant decidualization.  
+
+## Spatial Resolution of DR in sPE  
+
+We spatially resolved the transcriptomic changes discovered by scRNA- seq by spatial transcriptomics on formalin- fixed paraffin- embedded endometrial tissue biopsied during the late secretory phase, analysing 95 regions of interest (ROIs) in sPE (n=8) and control (n=8) patients (Fig. 5). Those regions fell into three categories based on the endometrial tissue architecture: (i) enriched in stromal (VIM+) and endothelial cells (CD31+) (Fig. 5a), (ii) enriched in glandular epithelium (PanCK+) (Fig. 5c) and (iii) luminal epithelial regions enriched in luminal epithelium and stromal cells (PanCK+, VIM+) (Fig. 5e). Differential expression analysis (cases vs. controls) revealed: (i) 430 DEGs in stroma, (ii) 575 DEGs in glandular epithelium, and (iii) 456 DEGs in luminal ROIs. Heatmaps of the DEGs for each compartment were constructed by unsupervised hierarchical clustering based on Pearson distances of the normalized data z- scores. With a few exceptions, sPE and control samples clustered separately.  
+
+We coalesced the DEGs from the single cell and spatial transcriptomics analyses (Extended data Fig. 9). In general, the DEGs had the same pattern of up- or downregulation in both analyses, which cross- validated the findings. In the enriched stromal ROIs (Fig. 5b and Extended data Fig. 9a), DEGs identified by both technologies included those involved in cytoskeletal organization. TPM1, an actin binding protein involved in contractility and cytoskeleton dynamics (46), was downregulated as were MYL9, which increases decidual contractility, and ANXA2 previously associated with DR (15). SEMA3, identified in our analysis of cell- cell communications, was upregulated as was KDM6b, which is involved in regulating decidual DNA methylation and modulating target gene expression (47).
+
+<--- Page Split --->
+
+The glandular epithelium ROI genes (Fig. 5d and Extended data Fig. 9b) were upregulated in DR cases and shared with the single cell analysis of epithelium including TNFRSF21 and the TNF receptor. They are also upregulated in women with recurrent pregnancy loss (48).  
+
+The luminal epithelium ROI genes (Fig. 5f and Extended data Fig. 9c) that were upregulated in DR cases and shared with the single cell analysis of stroma included LIMA, LIM domain and Actin Binding 1, a cytoskeleton- associated protein that inhibits actin filament depolymerization, is essential for proper mitochondrial function and a key effector mediating pluripotency (49). SCRAB1, a scavenger receptor involved in the removal of apoptotic cells in degenerated decidua tissue, was downregulated. The luminal epithelium ROI genes (Fig. 5f) that were upregulated in cases and shared with the single cell analysis of the epithelium included MMP7. In controls, its expression is associated with the proliferative phase of the menstrual cycle and downregulated during the late secretory phase (30). DUSP2 and ACADSB also shared this pattern. DUSP2 interacts with IL- 6 and its overexpression activates pathways that regulate inflammation and proliferation (50). ACADSB participates in cell migration, invasion and proliferation and cancer cells (51).  
+
+Altogether, coalescing the single cell and spatial transcriptomics data, pointed to an imbalance of cells in the epithelial and stromal compartments of samples from the DR cases that was attributable to late secretory endometrium retaining characteristics of the proliferative phase of the cycle. More specifically, our results suggested dysregulation of fundamental aspects of epithelial cell transition to a stromal phenotype during decidualization.  
+
+## Spatial proteomic mapping of DR in sPE by laser capture microdissection-mass spectrometry  
+
+We used LCM- MS to analyze the glandular epithelium, luminal epithelium and stroma of endometrial biopsies obtained during the late secretory phase of the cycle from an independent cohort of donors (sPE, n=7; controls, n=10) (Fig. 6a). To decipher the underlying functions of the identified proteins, overrepresentation analysis was performed, and protein- protein interaction networks were built to identify perturbations in endometria from the cases. In addition, estrogen receptor 1 (ESR1) and progesterone receptor (PGR) were included in the network to assess their associations with the DR phenotype.  
+
+The stromal compartment had the highest number of DE proteins. In this regard, 439 (43.9%) were specific to sPE samples and 53 (5.8%) were unique to control samples (Fig. 6b). Among proteins specific to sPE samples, was STAT3—which participate in decidualization downstream of PGR and is aberrantly increased in endometriosis (52, 53)—. Response to steroid hormones, particularly estrogens, is enriched in the sPE group, as is aging and cell growth (p.adjusted<0.05; Fig. 6c), consistent with increased proliferation and supported by the differential expression of markers of cell proliferation in
+
+<--- Page Split --->
+
+tumorigenesis such as ARG1 and B2M. Controls group showed an enrichment for extracellular matrix organization and EMT. The analysis of the protein- protein interaction network identified ESR1 and STAT3 as hub proteins, supporting the central role of hormonal signaling in the sPE affected pathways (Fig. 6d).  
+
+The glandular epithelium of control samples had 401 (27%) unique proteins while 103 (6.9%) were only present in sPE samples and 982 were common to both groups (Fig. 6b). The proteome from controls was enriched for hormone secretion (p.adjusted \(< 0.05\) ), whereas the sPE proteome was enriched for cell survival in response to ROS and epithelial cell differentiation (p.adjusted \(< 0.05\) ; Fig. 6c). The protein- protein interaction network revealed that these pathways are significantly interconnected with ESR1 and PGR (Fig. 6e). Thus, disrupted hormonal signaling could drive an imbalance in proliferation/differentiation, affecting the secretory function of endometrial glands by mechanisms that include the generation of ROS and defective cytoskeletal organization. Notable proteins related to this phenotype were specific to controls such as EGFR and LRP1.  
+
+The majority of luminal epithelium proteins 562 (63.3%) were shared between cases and controls; 67 (7.6%) were unique to controls and 254 (28.8%) were specific to cases (Fig. 6b). This proteome was enriched for response to steroids, specifically estrogens (p.adjusted \(< 0.05\) ; Fig. 6c), the major hormone during the proliferative phase. Accordingly, proteins specific to sPE were involved in cell proliferation (e.g., NONO, TRIM25), increased metabolism (e.g., PSMD2, PSMD7) and, consequently, inhibition of apoptosis signaling (e.g. SOD2, WFS1). In contrast, proteins unique to controls were enriched for extracellular matrix organization (e.g. CAV1, LAMB2). Then, we built a network with proteins involved in representative features of these pathways that were significantly disturbed in sPE (Fig. 6f). The results supported strong interconnections among these pathways and the key role of ESR1 and PGR.  
+
+Finally, we identified intersections between the biological processes identified by the scRNA- seq and LCM- MS analyses (Fig. 6g). There was substantial concordance between the two datasets in terms of those that were specific to sPE samples or controls. In many cases, there was also an overlap in the related biological processes that mapped to two or more compartments. The sPE group was characterized by processes/molecules involved in epidermal regulation, responses to oxidative stress/ROS and aging. The control group was characterized by processes such as protein secretion, cell growth, extracellular matrix organization, leukocyte adhesion and IL2 production. The sPE group was also notable for the absence of processes that were significantly represented in the control group such as gland development, response to LIF and EMT.  
+
+Thus, by overlapping the proteomic- based pathways associated with sPE and control samples with the DEGs from the single- cell dataset, we confirmed the key molecular features that shape DR at a multi- omic level. Overall, the main features are an imbalance in proliferation/differentiation in various regions
+
+<--- Page Split --->
+
+of the endometrium, associated with a disturbed response to steroid hormones, particularly estrogens, which could ultimately affect epithelial- stromal crosstalk and the secretory function of glands. Altogether, these results provide the deepest characterization of DR to date with possible clinical implications not only in the understanding of the pathogenesis of sPE but also in endometriosis and other pathological conditions related with DR.  
+
+## Discussion  
+
+Most pregnancy disorders originate from alterations in the periconeptional period, at the time where trophoectoderm invades the decidualized endometrium (1). Defective trophoblast invasion is the primary underlying cause of the great obstetrical syndromes such as preeclampsia, IUGR or preterm labor (54). Trying to answer why is placentation abnormal in these relevant pathologies our group and others have identified a decidualization defect in the endometrium (the soil) of these patients as a contributing factor in these obstetrical conditions (4), including sPE (15, 16, 55), IUGR or placenta.  
+
+Our present study offers a comprehensive exploration of the DR condition in former sPE patients, shedding light on the associated changes at gross, microscopic, scRNA seq, spatial transcriptomics and proteomics by LCM- MS. Our findings underscore the intricate molecular landscape underlying DR, implicating not only alterations in stroma and epithelium but also in the epithelial- to- stromal transition.  
+
+At the morphological level, distinct features including glandular dilation, hinted at underlying DR. Deeper investigation using scRNA- seq revealed significant alterations in cellular composition and differentiation patterns. At the stromal level, our analysis delineated specific subpopulations associated with different decidualization stages, revealing a shift in the balance leading to a mosaic of proliferative and decidualized stromal cells in former sPE patients, which causes cellular communications to be affected. Dysregulated gene expression signatures further corroborated the aberrant cellular states, highlighting the upregulation of markers indicative of proliferative activity and downregulation of those associated with decidualization. Additionally, endothelial and immune cells showed a disturbed transcriptome consistent with inflammatory dysregulation.  
+
+Similarly, an imbalance was evident in the epithelial cell subpopulations, with increased abundance of proliferative and ciliated epithelial cells in DR samples from former sPE patients. Dysregulated gene expression patterns mirrored this imbalance, with upregulation of genes linked to proliferation and downregulation of those associated with secretory functions indicating a potential dysregulation in the normal differentiation and maturation processes of uterine epithelial cells. This feature is relevant because glandular epithelia is the source of the uterine fluid or "uterine milk" that will provide nutrients and antimicrobial protection for the implanting conceptus.
+
+<--- Page Split --->
+
+Our analysis also uncovered perturbations in the epithelial- to- stromal transition, emphasizing a significant reduction in cells involved in this transition in former sPE patients. Furthermore, dysregulated cell- cell communication networks were evident, with notable alterations in signaling pathways critical for decidualization as WNT or SPP1, angiogenesis, and hormonal regulation.  
+
+Spatial transcriptomics and proteomic analyses provided additional layers of insight, highlighting spatially specific dysregulation in both epithelial and stromal compartments, mainly related to proliferation/differentiation imbalance. Key pathways implicated in DR, including hormonal signaling —particularly response to estrogens—, oxidative stress response, and cytoskeletal organization, were identified, further corroborating findings from single- cell analyses.  
+
+The comprehensive characterization of DR presented in this study not only advances our understanding of the pathogenesis of sPE but also has broader implications for other great obstetrical syndromes prevalent gynecological diseases such as endometriosis (9, 10), and recurrent miscarriage (13). The identified molecular signatures may serve as potential biomarkers for preconceptional detection of DR and intervention strategies before pregnancy is established. Preconceptional care is emerging as a key component of reproductive care not only to reduce perinatal morbidity and mortality, but to optimize health for mothers and children (56).  
+
+In summary, our multi- omic approach offers a nuanced depiction of DR, revealing an imbalance in proliferation/differentiation within the endometrium, disrupted response to steroid hormones—especially estrogens—and potential impacts on epithelial- stromal communication and glandular secretory function. These findings provide valuable insights into molecular underpinnings of DR and highlighting potential preconceptional therapeutic targets for mitigating the occurrence of the great obstetric syndromes.  
+
+## Material and Methods  
+
+## Study design  
+
+A total of 23 non- pregnant women who had a previous pregnancy were enrolled in this study for single- cell RNA- Sequencing analysis. From those, 12 were women with a healthy previous pregnancy as control cases and 11 were women diagnosed with Severe Preeclampsia in their last pregnancy clinically classified based the ACOG guidelines: high blood pressure (systolic \(>160\) or diastolic \(>100 \mathrm{mmHg}\) ) or thrombocytopenia, impaired liver function, progressive renal insufficiency, pulmonary edema, or the inset cerebral or visual disturbances. Endometrial biopsies were collected during the late secretory phase of the menstrual cycle (1 to 3 days before menstruation) and tissue were segmented in different portions, one of the were embedded in paraffin to obtain histological sections for the spatial transcriptomics approach, other section were included in OCT for the laser capture and mass spectrometry approach
+
+<--- Page Split --->
+
+and another portion was processed to obtain the cell isolation that will be used to the single- cell RNA sequencing using 10X genomics technologies. The maternal parameters were recruited in the Supplementary table 1. A two tailed Student's T- test was applied between sPE and Control variables based on the normal distribution of data.  
+
+## Sample collection  
+
+All human endometrial samples of this study were collected in University and Polytechnic La Fe Hospital, Valencia, Spain and pass their Clinical Research Ethics Committee. Samples were collected from women aged 18- 40 without any medical condition who had been pregnant 1- 8 years earlier. All donors had regular menstrual cycles (25- 31 days) without underlying pathological condition and had no received hormonal therapy in the 3 months preceding sample collection. Endometrial biopsies were obtained by pipeline catheter (Genetics Hamont- Achel, Belgium) under sterile conditions and were maintained in a preservation solution HypoThermosol® FRS (Stemcell Technologies) at \(4^{\circ}\mathrm{C}\) until their processing.  
+
+## Sample Processing  
+
+Endometrial biopsies were washed with PBS and mince. Afterwards, samples were enzymatically digested with a solution containing \(1\mathrm{mg / mL}\) Collagenase V (C9263, Sigma- Aldrich), \(100\mu \mathrm{g / mL}\) DNase Type I (03724751103, Roche) and \(10\%\) inactivated FBS in RPMI media for \(45\mathrm{min}\) at \(37^{\circ}\mathrm{C}\) and \(175\mathrm{rpm}\) . The enzymatic reaction was inactivated by adding 1 volume of \(10\%\) inactivated FBS in RPMI media (Complete medium), the solution filtered through a \(100 - \mu \mathrm{m}\) cell strainer, and the strainer washed with \(5\mathrm{mL}\) of complete medium. The flow- through solution contained the stromal fraction, whereas the undigested tissue retained in the cell strainer we the epithelial fraction of the endometrium.  
+
+The stromal fraction was centrifuged \(5\mathrm{min}\) at \(2000\mathrm{rpm}\) , and the pellet was washed with PBS. 1x RBC Lysis Buffer (eBioscience, ThermoFisher Scientific) \(5\mathrm{min}\) at RT was applied to eliminate blood from the sample following manufacturer recommendation. The cell fraction was washed with PBS centrifuged \(5\mathrm{min}\) at \(2000\mathrm{rpm}\) and the pellet resuspended with \(200\mu \mathrm{L}\) of \(0.04\%\) BSA in PBS (w/v). Finally, sample was filtered using a \(40 - \mu \mathrm{m}\) flowmi filter (BAH136800040- 50EA, Merk) to obtain a single cell suspension of the stromal fraction.  
+
+Regarding the epithelial fraction, the \(100 - \mu \mathrm{m}\) cell strainer mentioned above contained the epithelial cells. They were recovered by flushing \(15\mathrm{mL}\) PBS to the inverted cell strainer into a \(50\mathrm{- mL}\) falcon tube and centrifuged \(5\mathrm{min}\) at \(2000\mathrm{rpm}\) . For digestion of the epithelial tissue, the pellet was incubated with \(5\mathrm{mL}\) of \(100\mu \mathrm{g / mL}\) DNase Type I (03724751103, Roche) in \(0.25\%\) Trypsin- EDTA solution (25200072, Life Technologies) \(10\mathrm{min}\) at \(37^{\circ}\mathrm{C}\) and \(175\mathrm{rpm}\) shacking. After inactivation with complete media, the suspension was filtered through a \(100 - \mu \mathrm{m}\) cell strainer, the filter was washed with \(5\mathrm{mL}\) PBS and the solution centrifuged \(5\mathrm{min}\) at \(2000\mathrm{rpm}\) . If needed, the pellet was treated with RBC lysis solution as
+
+<--- Page Split --->
+
+mentioned before, washed again with PBS, resuspended with \(200~\mu \mathrm{L}\) of \(0.04\%\) BSA in PBS (w/v) and filtered using \(40 - \mu \mathrm{m}\) flowmi filter.  
+
+An aliquot of stromal and epithelial cell suspensions was stained with Tripan Blue dye and counted for alive cell concentration in an automatic cell counter (EVETM, NanoEnTek). Same number of both stromal and epithelial cells will be mixed and a total of 17,000 cells were loaded a 10X Chromium as explained in the following section.  
+
+## Single cell processing  
+
+scRNA- seq analysis of the endometrial processed samples were performed using the 10X Chromium technology (10X Genomics, Pleasanton, CA, USA). As mentioned, 17,000 cells were loaded onto a 10X G Chip to obtain Gel Bead- in- emulsions (GEMs) containing an individual cell. GEMs were used to generate barcoded cDNA libraries following the manufacturer's instructions (Single Cell 3' Reagent Kit v3.1, 10X Genomics) and quantified using the TapeStation High Sensitivity D5000 kit (Agilent, Germany). Following, cDNA (1- 100 ng) was obtained to construct gene expression libraries that were quantified using the TapeStation High Sensitivity D1000 kit (Agilent, Germany) determining the average fragment size and library concentration. Libraries were normalized, diluted, and sequenced on the Illumina NovaSeq 6000 system (Illumina, USA) according to the manufacturer's instructions.  
+
+## scRNA-seq data processing and filtering  
+
+Raw sequences were demultiplexed, aligned, and counted using the CellRanger software suite (v 6.0.2) for whole cell gene expression calculations, which takes advantage of intronic reads to improve sensitivity and sequencing depth (human reference genome GRCh38- 2020- A). Low- quality droplets and barcodes were filtered out in four quality control- based consecutive steps throughout the analysis: (i) low UMI- count barcode removal using an EmptyDrops- based method; (ii) cells marked as doublets by DoubletFinder (2.0.3) and scds (1.6.0) tools – the hybrid approach from the scds R package was used to avoid removing false- positive doublets; (iii) cells with median absolute deviation (MAD) \(>3\) in two of three basic quality control metrics: number of detected features, number of counts and mitochondrial ratio. These cell- to- count matrices were integrated and corrected using Seurat and scVI functions, as described below. A final filtering step, (iv), was applied alongside different rounds of clustering, where the obtained clustered cells with less than 750 features/cell, more than \(25\%\) mitochondrial ratio, and/or showing a pattern of high doublet- scoring plus no gene marker associated expression (during manual cell type annotations), were also removed (Supplementary figure 1). A total of 65,381 high quality cells were integrated within the uniform manifold approximation and projection (UMAP). From those 28,154 cells formed the sPE group and 37,227 cells the control group.  
+
+## Integration of single cells across conditions and clustering
+
+<--- Page Split --->
+
+As a first clustering analysis approach, read count matrices per sample were merged and processed following Seurat's default pipeline (package version 4.1.3). After normalization, the first thirty principal components on the 4,000 highly variable genes were used for dimensional reduction; cells were clustered and projected onto the UMAP. FindNeighbors and FindClusters functions were then applied for graph- based clustering by constructing a KNN graph using Euclidean distance in the principal component analysis space, which was then defined into clusters using the Louvain algorithm to optimize the standard modularity function. cluster (R package v0.4.4) was applied to select the most stable clustering resolution. The first output of sample distribution in clusters and cluster marker genes was then explored to evaluate biases from our data batches. Next, the scVI python package (v0.19.0) was used to remove patient origin inter- individual biases. The top thirty scVI components were used to embed and plot cells in the new reduced dimension space. These matrices were then input into Seurat's clustering and differential expression protocol. The clustering of different primary cell types for fine- grain cell type annotations was equally computed, following all the above- described steps.  
+
+## Tissue cell composition analysis and annotation of scRNA-seq datasets  
+
+The identification and labeling of major cell types based on differentially expressed genes per cluster compared to the remaining clusters using the Wilcoxon Rank Sum test and adjusted p- values for multiple comparisons with the false discovery rate (FDR) method (57). The primary cell populations were labeled by revising the expression of reported canonical markers from each cluster. Cell clustering was conducted under single- cell published datasets(2, 3, 24), manual curation of canonical markers published in human protein atlas (HPA)(58). The main cell types were then subset separately: epithelium, stromal fibroblast, perivascular, cycling stroma, endothelium, macrophages, Becells, Tecells, Nk cells. A new clustering was performed on each to create the different zoom- ins that describe the contained sub- populations. The clustered zoom- ins were manually annotated by an extensive review of differentially expressed genes in the Human Protein Atlas database, single- cell atlases, and scientific literature(51). Genes with strong cluster- specificity, as determined by a p- value below 0.01, and the highest rank fold change and percentage of expressing cells were considered.  
+
+## Cell trajectories on the transcriptional pseudotimes  
+
+The inference of cell trajectories was based on the predicted velocity of mRNA synthesis in each cell (RNA velocity). The method for calculating RNA velocity begins with obtaining the spliced/unspliced matrix using the Velocytool (59). RNA velocities were separately obtained for each zoom- in clustering. As already stated by(60), calculations using Velocytool are sensitive to genes that undergo sudden large changes in their expression patterns, described as Transcription Bursts. To avoid this issue, velocities computed by scVelo (v0.2.5) were corrected using CellDancer (v1.17). In scVelo, we filtered the genes with fewer than 20 counts shared by cells in the filter genes function and used only the top
+
+<--- Page Split --->
+
+2000 most variable genes in the filter_gene_s_dispersion. We used the dynamic mode to compute cellular and velocity dynamics. In CellDancer (61), we computed the velocities and projected the vector field predictions computed by the model onto the previously computed UMAP embedding space. After identifying the start and ends of the trajectories, we used Slingshot (62) to compute the genes related to changes across pseudotime in the inferred trajectories.  
+
+## Analysis of differential cell abundances and differential gene expression  
+
+To identify differential abundance in cell populations between sPE and control endometria we used MiloR (version 1.6.0) tool described by Dann et al.(63). This approach supports the grouping of cells on a k- nearest neighbor graph and evaluates the change in cell abundance between conditions, p values were adjusted using spatial FDR of \(< 0.1\) . Further differential gene expression analysis was performed using the Model- based Analysis of Single Cell Transcriptomics (MAST), where a contrast test was established for each cell type (64).  
+
+## Analysis of cell-to-cell communication networks  
+
+In this study, we employed the CellChat R package (v1.1.3) to discern potential cell interactions between distinct cell populations within both control and preeclampsia samples. This tool employs a curated database to infer the overall interaction probability and communication information flows based on the expression of specific ligand- receptor pairs. To elucidate, the total interaction probability measures the likelihood of communication between two cell types, where one serves as the sender and the other as the receiver. This estimation is based on the number of interactor molecules expressed (i.e., ligand- receptor pairs) and the strength of this interaction (expression level).  
+
+Subsequently, the cumulative communication probability of all pairs in a pathway network is employed to compute the communication information flow. To ensure robust results, pathways with fewer than ten cells were excluded from the analysis ( \(10 <\) cells per subpopulation). Additionally, we accounted for the impact of varying cell population sizes by enabling the 'population.size' argument in the 'computeCommunProb' function, which properly scales the communication weight of each subpopulation (set to TRUE).  
+
+Finally, we conducted a differential CCC analysis between control and preeclampsia samples using the 'ranknet' function, applying a significance threshold of 0.05.  
+
+## Spatial Transcriptomics sample processing  
+
+To achieve this objective, Nanostring® technology was utilized, specifically employing the GeoMx® Human Whole Transcriptome Atlas panel. This panel comprehensively covers protein- coding
+
+<--- Page Split --->
+
+genes, allowing the spatial analysis of any tissue within a chosen biological area. In this study, specific biological regions were selected to capture the higher cellular diversity commonly found in the late secretory phase of the endometrium, encompassing various cell types such as stromal, endothelial, luminal epithelial, and glandular epithelial cells, among others.  
+
+We followed established experimental procedures of Nanostring with some modifications, as detailed below. In summary, we prepared \(5\mu \mathrm{m}\) thick FFPE sections from 16 specimens. These consecutive sections were then subjected to H&E and WTA (Whole Transcriptome Analysis).  
+
+The H&E slides was used as guide to select the ROIs. A total of 95 ROI, distributed equally in 16 patients were selected covering the main cell types of the endometrium.  
+
+For the WTA, the slides were first incubated at \(60^{\circ}\mathrm{C}\) for an hour, then deparaffinized using limonene (Merck Chemicals And Life Science, S.A.U.), and subsequently rehydrated. Antigen retrieval was performed by subjecting the slides to \(1\mathrm{x}\) Tris- EDTA/pH 9 in a steamer for \(20\mathrm{min}\) at \(100^{\circ}\mathrm{C}\) . Next, proteinase K (Thermo Fisher Scientific, AM2548) was used to digest the samples at a concentration of \(1\mathrm{Xg / ml}\) for \(15\mathrm{min}\) at \(37^{\circ}\mathrm{C}\) . Following that, the samples were postfixed in neutral- buffered formalin for \(5\mathrm{min}\) .  
+
+Then samples underwent hybridization with a UV- photocleavable barcode- conjugated RNA in situ hybridization probe set, targeting 18,269 genes WTA. This hybridization process was conducted overnight at \(37^{\circ}\mathrm{C}\) . Afterward, the samples were thoroughly washed to eliminate any non- specific probes, and morphology markers were then applied to facilitate counterstaining. This counterstaining step was carried out for \(1\mathrm{h}\) at room temperature.  
+
+Then, morphology markers utilized were a combination of fluorescent agents: 1:10 SYTO13 (Nanostring), 1:50 anti- panCK- Alexa Fluor 532 (Nanostring), 1:500 anti- vimentin Alexa Fluor 594 (Santa Cruz Biotechnology; cat.no: sc- 373717 AF594), 1:50 anti- CD31 Alexa Fluor 647 (Abcam; cat.no: ab215912) in blocking buffer W (NanoString). These markers were applied in blocking buffer W (NanoString). The immunofluorescence images, region of interest (ROI) selection, marker- specific area of interest (AOI) segmentation, and spatially indexed barcode cleavage and collection were performed using a GeoMx DSP instrument (NanoString). The typical exposure times were \(50\mathrm{ms}\) for SYTO13, \(200\mathrm{ms}\) for anti- panCK AF532, \(200\mathrm{ms}\) for anti- vimentin AF 594, and \(200\mathrm{ms}\) for anti- CD31 AF647. Approximately 96 ROIs were collected per specimen.  
+
+For library preparation, the manufacturer's instructions were followed, involving PCR amplification to add Illumina adapter sequences and unique dual sample indexes. Each single molecule in the sequencing library corresponded to one cluster on the flow cell. The paired- end read strategy recommended by NanoString resulted in two paired- end reads, forward and reverse, per cluster, which can be described as one read pair. To achieve a minimum sequencing depth of 150- 200 reads per square
+
+<--- Page Split --->
+
+micron of illumination area, all WTA AOIs were sequenced on a NextSeq 2000 (Read type: Paired- end, read length: read 1: 27bp, read 2: 27bp, index 1 (i7): 8bp, index 2 (i7): 8bp).  
+
+## QC Of Spatial transcriptomics  
+
+Inn every ROI, in segments with \(< 1000\) raw reads are removed, as well as segments with less than \(80\%\) of alignment, trimmed or stitched were removed. Regarding the \(\%\) of sequencing saturation defined as ([1- deduplicated reads/aligned reads] \(\%\) ) ROIs with less that \(50\%\) of this parameter are also removed. The geometric mean of several unique negative probes included in the GeoMx panel is calculated and it stablished the background count level per segment,  
+
+The No Template Control (NTC) count: values \(>1000\) could indicate contamination for the segments associated with this NTC; however, in cases where the NTC count is between 1000- 10000, the segments may be used if the NTC data is uniformly low (e.g. 0- 2 counts for all probes). Quality control for nuclei: \(>100\) nuclei per segment is generally recommended;. All segments of the study pass those qc filters.  
+
+We also remove outlier negative control probes from the data to refine our estimation of background and downstream gene detection. We remove outlier probes either entirely from the study (global) or from specific segments (local). There are 18807 probes that passed QC. 1 Global outlier and 7 Local outliers were detected  
+
+Regarding the limit of quantification (LOQ) per ROI/AOI segment based on the negative control probes to guide the filtering of segments and genes with low signal relative to background. The formula for calculating the LOQ in the \(\mathrm{i}^{\mathrm{th}}\) segment at n standard deviations ( \(\mathrm{n} = 2\) for this study) is:  
+
+\[LOQ_{i} = g e o m e a n(N e g P r o b e_{i})*g e o S D(N e g P r o b e_{i})^{n}\]  
+
+In this dataset, we choose to remove segments with fewer than \(10\%\) of the genes detected. As a result, 0 segments were flagged and removed. In total, 0 segments were removed from the study from either segment QC or filtering.  
+
+10,239 targets were detected above LOQ in \(10\%\) or more of the segments. Including the genes of interest list, 10,489 targets in total were analyzed further. We filtered down to this number of targets  
+
+## Normalization of spatial transcriptomics  
+
+Two common methods for normalization of GeoMx® WTA/CTA data are i) Quartile 3 (Q3) or ii) background. Both methods estimate a normalization factor per ROI/AOI segment to bring the segment data distributions together. Q3 is typically the preferred approach. High correlation between the geometric mean of the negative control probe counts and 75th quantile (Q3) of expression. Q3 normalization was used downstream for analysis.
+
+<--- Page Split --->
+
+## Contrasting groups with Linear Mixed Models in Spatial transcriptomics  
+
+Ia given contrast, we are interested in comparing the magnitude of differences between groups and quantifying the significance. In many GeoMx experiments, multiple areas of illumination (AOIs) or regions of interest (ROIs) are sampled in a given slide. To account for the nested, non- independent sampling, we often use a linear mixed effect model (LMM) with slide as a random effect. The LMM accounts for the subsampling per tissue, allowing us to adjust for the fact that the multiple regions of interest placed per tissue section are not independent observations, as is the assumption with other traditional statistical tests.  
+
+Overall, there are two main types of the LMM models when used with GeoMx data: A) with random slope and B) without random slope.  
+
+When comparing features that co- exist in a given tissue section, a random slope is included in the LMM model. When comparing features that are mutually exclusive in a given tissue section the LMM model does not require a random slope.  
+
+## Spatial proteomics: laser capture microdissection; protein isolation/digestion and mass spectrometry.  
+
+Three regions of each endometrial biopsy were isolated by laser capture microdissection (LMD): glandular epithelium, luminal epithelium and stroma. They were separately processed and analyzed using methods we previously published (65). Briefly, Frozen blocks of the tissue were sectioned \((- 20^{\circ}\mathrm{C})\) using a Leica CM3050 cryostat. Sections \((20\mu \mathrm{m})\) were mounted on Director slides (Expression Pathology; Hembrough et al., 2012). Slides with sections were kept under dry ice until LMD the same day. Sections on slides were manually defrosted in room air \((30\mathrm{~s})\) , immersed in PBS until the OCT was completely removed \((\sim 2\mathrm{~min})\) , dipped in \(0.1\%\) Toluidine Blue for \(30\mathrm{~s}\) , washed in ice- cold PBS, dehydrated (30 s/treatment) in a graded ethanol series \((75\% , 95\% , 100\%)\) , then rapidly dried with compressed nitrogen.  
+
+Following capture, the samples were incubated in an alkaline surfactant- containing solution at \(60^{\circ}\mathrm{C}\) with rigorous vortexing every 15 min for 1 h. Iodoacetamide (Thermo Fisher Scientific) was added to \(15\mathrm{mM}\) and incubated at room temperature in the dark for \(30\mathrm{min}\) . Proteins were digested overnight at \(37^{\circ}\mathrm{C}\) with trypsin \((20\mathrm{ng / \mu l}\) , Pierce), then centrifuged at \(16,000\mathrm{g}\) (Eppendorf) for \(10\mathrm{min}\) . Trifluoroacetic acid (Pierce) was added to the supernatant such that the final concentration was \(0.5\%\) . Duplicate technical replicates were performed along with a control that consisted of a blank DIRECTOR slide that was carried through the entire sample preparation protocol.  
+
+Samples were analyzed by reverse- phase HPLC- ESI- MS/MS using an Eksigent Ultra Plus nano- LC 2D HPLC system directly connected to an orthogonal quadrupole time- of- flight SCIEX TripleTOF 6600 mass spectrometer (SCIEX). Peptide and protein identifications were determined using the Paragon
+
+<--- Page Split --->
+
+algorithm within the ProteinPilot search engine (v.5.0.2, SCIEX) against the corresponding proteome FASTA files obtained from UniProt.  
+
+## Enrichment analysis  
+
+Gene Ontology (GO) analyses were conducted to obtain biological processes using enrichGO function from clusterProfiler R package (v. 4.2.2). The input proteins were those described for both sPE and control groups per endometrial region (glandular epithelium, luminal epithelium, and stroma). Specific pathways of a group refer to enriched pathways that are present in the controls or in the sPE, but are not present in the other group. The input genes were the differential expressed genes grouped by endometrial region. The p- value adjustment method was FDR with a cutoff of 0.05.  
+
+## Protein-protein interaction network  
+
+The protein- protein interaction networks were created using the functional analysis suit String and visualized using Cytoscape software. Hub genes were extracted using the maximal clique centrality (MCC) and maximum neighborhood component (MNC) of the cytoHubba plugin.  
+
+Data availability  
+
+The single- cell RNA- sequencing data generated for this manuscript has been uploaded to GEO under accession number GSE265862. The uploaded data includes i) H5ad files containing the aggregated count matrices and metadata of each cell studied in the major cell populations and subpopulations; ii) Raw count matrices processed by Cell Ranger. The raw sequences are not publicly available due to privacy concerns. However, they are available from the corresponding authors (C.S, carlos.simon@uv.es; TG, tgarrido@fundacioncarlosimon.com) upon reasonable request and with permission of the Institutional Review Board of the Spanish hospitals involved.  
+
+# Figures
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+
+<--- Page Split --->
+
+Figure 1. Morphological features in decidualization resistance reflected in single cell atlas in sPE and control condition. (a) Representative endometrial tissue collected during late secretory phase from women with a previous sPE \((n = 7)\) . (b) Zoom - in of the macroscopical glands of the endometrial tissue of figure 1 a. (c) Representative H&E slides staining of cross-section endometrial tissue of an sPE sample. (d) Representative H&E slides staining of longitudinal-section endometrial tissue of an sPE sample. (e) Representative endometrial tissue collected during late secretory phase from control women \((n = 10)\) . (f) Zoom - in of the macroscopical glands of the endometrial tissue of fig. 1c. (g) Representative H&E slides staining of cross-section endometrial tissue of a control sample. (h) Representative H&E slides staining of longitudinal-section endometrial tissue of a control sample. (i) Single-cell sequencing workflow. (j) Uniform manifold approximation and projection (UMAP) of single-cell integration of high-quality cells sPE cells (28,154) of the major cell types of the endometrium during late secretory phase. (k) UMAP of single-cell integration of high-quality cells control cells (37,227) of the major cell types of the endometrium during late secretory phase. (l) UMAP of the 65,381 high quality cells of types of merged cells of both sPE (blue) and control (red) samples.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Figure 2. Altered stromal and epithelial cell differentiation states of DR in sPE condition. (a) UMAP of cell subpopulation identification of stromal and perivascular fraction of endometrium in late secretory phase. (b) UMAP of stromal merged cells of both sPE (blue) and control (red) samples. (c) Neighbourhood graph represents the differential abundance of stromal cell in late secretory endometrium. Dot size represents neighbourhoods, while edges thickness (weight) depicts the </center>
+
+<--- Page Split --->
+
+number of cells shared between neighbourhoods. Neighbourhoods coloured in blue represent those with a significant decrease in cell abundance in sPE and red highlight cells enriched in sPE. (d) Beeswarm plot of differential cell abundance by stromal cell subtypes. X-axis represents the log2-fold change in abundance of sPE. Each dot represents a neighbourhood; neighbourhoods coloured in blue represent those with a significant decrease in cell abundance in sPE condition while red dots are enriched in sPE samples. (e) Cell subpopulation identification of epithelial fraction of endometrium in late secretory phase. (f) UMAP of epithelial merged cells of both sPE (blue) and control (red) samples. (g) Neighbourhood graph represents the differential abundance of epithelial cell in late secretory endometrium. Dot size represents neighbourhoods, while edges depict the number of cells shared between neighbourhoods. Neighbourhoods coloured in blue represent those with a significant decrease in cell abundance in sPE and red highlight cells enriched in sPE. (h) Beeswarm plot of differential cell abundance by epithelial cell subtypes. X-axis represents the log2-fold change in abundance of sPE. Each dot represents a neighbourhood; neighbourhoods coloured in blue represent those with a significant decrease in cell abundance in sPE condition while red dots are enriched in sPE samples.
+
+<--- Page Split --->
+![PLACEHOLDER_27_0]
+
+
+<--- Page Split --->
+
+Figure 3. Absence of EMT in endometria with DR from sPE condition. (a) Zoom in of EMT and cell subpopulation identification of endometrium in late secretory phase. Circle highlights the epithelial- to- stroma transition subpopulation. (b) UMAP of EMT merged cells of both sPE (blue) and control (red) samples. (c) Neighborhood graph represents the differential cell abundance of MET cells in late secretory endometrium. Dot size represents neighborhoods, while edges depict the number of cells shared between neighborhoods. Neighborhoods colored in blue represent those with a significant decrease in cell abundance in sPE and red highlight cells enriched in sPE. (d) Beeswarm plot of differential cell abundance by MET cell subtypes. X- axis represents the log2- fold change in abundance of sPE. Each dot represents a neighborhood; neighborhoods colored in blue represent those with a significant decrease in cell abundance in sPE condition while red dots are enriched in sPE samples. (e) RNA velocity generated with scVelo of sPE and control samples of epithelial transition, epithelium- to- stromal transition and stromal transition subpopulations. Ciliated cell subtype was removed from trajectory inferences downstream analysis. Arrows represents the cell trajectories across clusters inferring differentiation cell trajectories. (f) Pattern of gene expression along the pseudotime from epithelium to stroma in the inferred trajectory. (g) Dotplot of differentially expressed genes in sPE vs controls identified across pseudotime associated to stroma, epithelial- to- stroma transition, and epithelium (color represents the average expression and dot size refers to the percentage of cells of each cluster expressing each marker).
+
+<--- Page Split --->
+![PLACEHOLDER_29_0]
+
+
+<--- Page Split --->
+
+Figure 4. Dysfunctional cell- to- cell communication networks associated with altered cell composition in DR. (a, c, e, g, i, j) Chords plots displaying the CCC network of Endoglin (EDN), noncanonical WNT (ncWNT), canonical WNT, Semaphorin (SEMA3), SPP1 and TENASCIN in sPE and control. Each coloured dot represents a cell subtype. Colour arrow represents de incoming signalling and the thickness of the lines refers to the strength of the signal between cell subtypes. (b, d, f, h) Barplot of each ligand- receptor pair contributing to the CCC of Endoglin, ncWNT, WNT and SEMA3. Red colour represents sPE expression pair and blue the Control.
+
+<--- Page Split --->
+![PLACEHOLDER_31_0]
+
+<center>Figure 5. Decidualization resistance in sPE confirmed with spatial transcriptomics. (a) Immunofluorescence of enriched stromal ROIs selected of one representative sPE sample (PanCK in green, Vimentin in yellow, CD31 in red and nucli in blue) and unsupervised hierarchical clustering based on Pearson distances of the normalized data z-scores of the top genes of enriched stromal ROIs. (b) Volcano plots depicting DEGs between sPE and controls within stromal ROIs. (c) Immunofluorescence of enriched glandular epithelial ROIs selected of one representative sPE sample and unsupervised hierarchical clustering based on Pearson distances of the normalized data z-scores of the top genes of enriched glandular epithelium ROIs. (d) Volcano plots depicting DEGs between sPE and controls within Glandular epithelial ROIs. (e) Immunofluorescence of enriched luminal epithelial </center>
+
+<--- Page Split --->
+
+761 ROI selected of one representative sPE sample and unsupervised hierarchical clustering based on 762 Pearson distances of the normalized data z-scores of the top genes of enriched luminal epithelium ROIs. 763 (f) Volcano plots depicting DEGs between sPE and controls within luminar epithelial ROIs. Heatmap 764 legend reflects info of group (sPE in red and control in blue), and colours represented the ROI of each 765 participant. Volcano plot legend represent significant genes ( \(\mathrm{p}< 0.05\) ) and non-significant genes (NS < 766 0).
+
+<--- Page Split --->
+![PLACEHOLDER_33_0]
+  
+
+![PLACEHOLDER_33_1]
+  
+
+![PLACEHOLDER_33_2]
+  
+
+![PLACEHOLDER_33_3]
+  
+
+![PLACEHOLDER_33_4]
+
+
+<--- Page Split --->
+
+Figure 6. Differential endometrial proteome associated with decidualization resistance in sPE compared to controls. (a) Endometrial section before laser capture microdissection, and after isolating the regions of interest (glandular epithelium, luminal epithelium, and stromal compartment). (b) Venn diagrams showing the total number of proteins identified between controls and sPE in the stromal compartment, glandular epithelium, and luminal epithelium, respectively. (c) Differential expressed pathways specific to controls and sPE per region analysed. Color gradient shows the protein ratio (%), which refers to what proportion of all the proteins detected in the region are proteins involved in the pathway. All pathways included are significantly enriched (adjusted. \(\mathrm{P< 0.05}\) ). (d) Protein-protein interaction network including those proteins involved in highlighted pathways in the stromal compartment. Colour shows the specificity of proteins (red, proteins unique to sPE; blue, proteins unique to controls; purple, proteins shared by the two groups; green, ESR1 and PGR). Shape shows the pathway (circle, response to steroids; hexagon, aging; rhombus, cell growth; rectangle, extracellular matrix organization). (e) Protein-protein interaction network including those proteins involved in highlighted pathways in the glandular epithelium. Colour shows the specificity of proteins (red, proteins unique to sPE; blue, proteins unique to controls; purple, proteins shared by the two groups; green, ESR1 and PGR). Shape shows the pathway (circle, hormone secretion; hexagon, response to reactive oxygen species (ROS); square, cell survival in response to ROS; rhombus, epidermal cell differentiation; rectangle, regulation of actin cytoskeleton organization. (f) Protein-protein interaction network including those proteins involved in highlighted pathways in the luminal epithelium. Colour shows the specificity of proteins (red, proteins unique to sPE; blue, proteins unique to controls; purple, proteins shared by the two groups; green, ESR1 and PGR). Shape shows the pathway (circle, response to steroids; hexagon, negative regulation of apoptotic signaling pathway; rhombus, proteasomal protein catabolic process; rectangle, extracellular matrix organization) (PPI enrichment refers to protein-protein interaction enrichment). (g) Dot plot showing the functional enrichment coincides between the spatial proteome and the differentially expressed genes at single-cell resolution. Left panel, sPE-specific pathways. Right panel, controls-specific pathways. Pink, scRNA-seq data; purple, LCM-MS.  
+
+## References  
+
+1. Moreno I, Capalbo A, Mas A, Garrido-Gomez T, Roson B, Poli M, et al. The human periconceptional maternal-embryonic space in health and disease. Physiol Rev. 2023;103(3):1965-2038. Epub 20230216. doi: 10.1152/physrev.00050.2021. PubMed PMID: 36796099.  
+2. Wang W, Vilella F, Alama P, Moreno I, Mignardi M, Isakova A, et al. Single-cell transcriptomic atlas of the human endometrium during the menstrual cycle. Nature Medicine. 2020;26(10):1644-53. doi: 10.1038/s41591-020-1040-z.  
+3. Garcia-Alonso L, Handfield LF, Roberts K, Nikolakopoulou K, Fernando RC, Gardner L, et al. Mapping the temporal and spatial dynamics of the human endometrium in vivo and in vitro. Nat Genet. 2021;53(12):1698-711. Epub 20211202. doi: 10.1038/s41588-021-00972-2. PubMed PMID: 34857954; PubMed Central PMCID: PMC8648563.
+
+<--- Page Split --->
+
+4. Ng SW, Norwitz GA, Pavlicev M, Tilburgs T, Simón C, Norwitz ER. Endometrial Decidualization: The Primary Driver of Pregnancy Health. Int J Mol Sci. 2020;21(11). Epub 2020/06/08. doi: 10.3390/ijms21114092. PubMed PMID: 32521725; PubMed Central PMCID: PMC7312091.  
+5. Tabanelli S, Tang B, Gurpide E. In vitro decidualization of human endometrial stromal cells. J Steroid Biochem Mol Biol. 1992;42(3-4):337-44. doi: 10.1016/0960-0760(92)90137-8. PubMed PMID: 1534990.  
+6. Irwin JC, Utian WH, Eckert RL. Sex steroids and growth factors differentially regulate the growth and differentiation of cultured human endometrial stromal cells. Endocrinology. 1991;129(5):2385-92. doi: 10.1210/endo-129-5-2385. PubMed PMID: 1935772.  
+7. Garrido-Gomez T, Dominguez F, Lopez JA, Camafeita E, Quiñonero A, Martinez-Conejero JA, et al. Modeling human endometrial decidualization from the interaction between proteome and secretome. J Clin Endocrinol Metab. 2011;96(3):706-16. Epub 2010/12/29. doi: 10.1210/jc.2010-1825. PubMed PMID: 21190976.  
+8. Ramathal C, Bagchi I, Taylor R, Bagchi M. Endometrial Decidualization: Of Mice and Men. Seminars in Reproductive Medicine. 2010;28(01):017-26. doi: 10.1055/s-0029-1242989.  
+9. Su R-W, Strug MR, Joshi NR, Jeong J-W, Miele L, Lessey BA, et al. Decreased Notch Pathway Signaling in the Endometrium of Women With Endometriosis Impairs Decidualization. The Journal of Clinical Endocrinology & Metabolism. 2015;100(3):E433-E42. doi: 10.1210/jc.2014-3720.  
+10. Leone Roberti Maggiore U, Ferrero S, Mangili G, Bergamini A, Inversetti A, Giorgione V, et al. A systematic review on endometriosis during pregnancy: diagnosis, misdiagnosis, complications and outcomes. Human Reproduction Update. 2016;22(1):70-103. doi: 10.1093/humupd/dmv045.  
+11. Klemmt PAB, Carver JG, Kennedy SH, Koninckx PR, Mardon HJ. Stromal cells from endometriotic lesions and endometrium from women with endometriosis have reduced decidualization capacity. Fertility and Sterility. 2006;85(3):564-72. doi: 10.1016/j.fertnstert.2005.08.046.  
+12. Gellersen B, Brosens JJ. Cyclic decidualization of the human endometrium in reproductive health and failure. Endocr Rev. 2014;35(6):851-905. Epub 2014/08/20. doi: 10.1210/er.2014-1045. PubMed PMID: 25141152.  
+13. Lucas ES, Vrljicak P, Muter J, Diniz-da-Costa MM, Brighton PJ, Kong CS, et al. Recurrent pregnancy loss is associated with a pro-senescent decidual response during the peri-implantation window. Commun Biol. 2020;3(1):37. Epub 20200121. doi: 10.1038/s42003-020-0763-1. PubMed PMID: 31965050; PubMed Central PMCID: PMC6972755.  
+14. Salker MS, Christian M, Steel JH, Nautiyal J, Lavery S, Trew G, et al. Deregulation of the serum-and glucocorticoid-inducible kinase SGK1 in the endometrium causes reproductive failure. Nat Med. 2011;17(11):1509-13. Epub 2011/10/16. doi: 10.1038/nm.2498. PubMed PMID: 22001908.  
+15. Garrido-Gomez T, Dominguez F, Quiñonero A, Diaz-Gimeno P, Kapidzic M, Gormley M, et al. Defective decidualization during and after severe preeclampsia reveals a possible maternal contribution to the etiology. Proc Natl Acad Sci U S A. 2017;114(40):E8468-E77. Epub 2017/09/18. doi: 10.1073/pnas.1706546114. PubMed PMID: 28923940; PubMed Central PMCID: PMC5635883.  
+16. Garrido-Gomez T, Castillo-Marco N, Clemente-Ciscar M, Cordero T, Muñoz-Blat I, Amadoz A, et al. Disrupted PGR-B and ESR1 signaling underlies defective decidualization linked to severe preeclampsia. Elife. 2021;10. Epub 20211028. doi: 10.7554/elife.70753. PubMed PMID: 34709177; PubMed Central PMCID: PMC8553341.  
+17. Garrido-Gomez T, Quiñonero A, Dominguez F, Rubert L, Perales A, Hajjar KA, et al. Preeclampsia: a defect in decidualization is associated with deficiency of Annexin A2. Am J Obstet Gynecol. 2020a;222(4):376.e1-.e17. Epub 2019/11/15. doi: 10.1016/j.ajog.2019.11.1250. PubMed PMID: 31738896.  
+18. Gupta MB, Abu Shehab M, Nygard K, Biggar K, Singal SS, Santoro N, et al. IUGR Is Associated With Marked Hyperphosphorylation of Decidual and Maternal Plasma IGFBP-1. The Journal of Clinical Endocrinology & Metabolism. 2019;104(2):408-22. doi: 10.1210/jc.2018-00820.
+
+<--- Page Split --->
+
+19. Jauniaux E, Jurkovic D, Hussein AM, Burton GJ. New insights into the etiopathology of placenta accreta spectrum. Am J Obstet Gynecol. 2022;227(3):384-91. Epub 20220303. doi: 10.1016/j.ajog.2022.02.038. PubMed PMID: 35248577.  
+20. Roberts JM, Cooper DW. Pathogenesis and genetics of pre-eclampsia. Lancet. 2001;357(9249):53-6. PubMed PMID: 11197372.  
+21. Gestational Hypertension and Preeclampsia: ACOG Practice Bulletin, Number 222. Obstet Gynecol. 2020;135(6):e237-e60. doi: 10.1097/aog.0000000000003891. PubMed PMID: 32443079.  
+22. Staff AC, Fjeldstad HE, Fosheim IK, Moe K, Turowski G, Johnsen GM, et al. Failure of physiological transformation and spiral artery atherosis: their roles in preeclampsia. Am J Obstet Gynecol. 2020. Epub 2020/09/21. doi: 10.1016/j.ajog.2020.09.026. PubMed PMID: 32971013.  
+23. Fisher SJ. Why is placentation abnormal in preeclampsia? American Journal of Obstetrics and Gynecology. 2015;213(4):S115-S22. doi: 10.1016/j.ajog.2015.08.042.  
+24. Lv H, Zhao G, Jiang P, Wang H, Wang Z, Yao S, et al. Deciphering the endometrial niche of human thin endometrium at single-cell resolution. Proceedings of the National Academy of Sciences. 2022;119(8). doi: 10.1073/pnas.2115912119.  
+25. Liu H, Wang LL, Xu QH, Wang J, Zhang YJ, Luo J, et al. UHRF1 shapes both the trophoblast invasion and decidual macrophage differentiation in early pregnancy. Faseb j. 2022;36(4):e22247. doi: 10.1096/fj.202101647RR. PubMed PMID: 35262949.  
+26. Hess AP, Hamilton AE, Talbi S, Dosiou C, Nyegaard M, Nayak N, et al. Decidual Stromal Cell Response to Paracrine Signals from the Trophoblast: Amplification of Immune and Angiogenic Modulators1. Biology of Reproduction. 2007;76(1):102-17. doi: 10.1095/biolreprod.106.054791.  
+27. Vento-Tormo R, Efremova M, Botting RA, Turco MY, Vento-Tormo M, Meyer KB, et al. Single-cell reconstruction of the early maternal-fetal interface in humans. Nature. 2018;563(7731):347-53. Epub 2018/11/14. doi: 10.1038/s41586-018-0698-6. PubMed PMID: 30429548.  
+28. Wang Z, Liu Y, Liu J, Kong N, Jiang Y, Jiang R, et al. ATF3 deficiency impairs the proliferative–secretory phase transition and decidualization in RIF patients. Cell Death & Disease. 2021;12(4). doi: 10.1038/s41419-021-03679-8.  
+29. Wu B, Li Y, Nie N, Shen X, Jiang W, Liu Y, et al. SFRP4+ stromal cell subpopulation with IGF1 signaling in human endometrial regeneration. Cell Discovery. 2022;8(1). doi: 10.1038/s41421-022-00438-7.  
+30. Hisamatsu Y, Murata H, Tsubokura H, Hashimoto Y, Kitada M, Tanaka S, et al. Matrix Metalloproteinases in Human Decidualized Endometrial Stromal Cells. Current Issues in Molecular Biology. 2021;43(3):2111-23. doi: 10.3390/cimb43030146.  
+31. Osteen KG, Rodgers WH, Gaire M, Hargrove JT, Gorstein F, Matrisian LM. Stromal-epithelial interaction mediates steroidal regulation of metalloproteinase expression in human endometrium. Proc Natl Acad Sci U S A. 1994;91(21):10129-33. doi: 10.1073/pnas.91.21.10129. PubMed PMID: 7937850; PubMed Central PMCID: PMC44971.  
+32. Kirkwood PM, Gibson DA, Shaw I, Dobie R, Kelepouri O, Henderson NC, et al. Single-cell RNA sequencing and lineage tracing confirm mesenchymal to epithelial transformation (MET) contributes to repair of the endometrium at menstruation. Elife. 2022;11. Epub 20221216. doi: 10.7554/eLife.77663. PubMed PMID: 36524724; PubMed Central PMCID: PMC9873258.  
+33. Cheng C-W, Smith SK, Charnock-Jones DS. Transcript profile and localization of Wnt signaling–related molecules in human endometrium. Fertility and Sterility. 2008;90(1):201-4. doi: 10.1016/j.fertnstert.2007.05.077.  
+34. Salsano S, Quiñonero A, Pérez S, Garrido Gómez T, Simón C, Dominguez F. Dynamic expression of PGRMC1 and SERBP1 in human endometrium: an implication in the human decidualization process. Fertility and Sterility. 2017;108(5):832-42. e1. doi: 10.1016/j.fertnstert.2017.07.1163.  
+35. Roukens MG, Peterson-Maduro J, Padberg Y, Jeltsch M, Leppänen V-M, Bos FL, et al. Functional Dissection of the CCBE1 Protein. Circulation Research. 2015;116(10):1660-9. doi: 10.1161/circresaha.116.304949.
+
+<--- Page Split --->
+
+36. Bos FL, Caunt M, Peterson-Maduro J, Planas-Paz L, Kowalski J, Karpanen T, et al. CCBE1 is essential for mammalian lymphatic vascular development and enhances the lymphangiogenic effect of vascular endothelial growth factor-C in vivo. Circ Res. 2011;109(5):486-91. Epub 20110721. doi: 10.1161/CIRCRESAHA.111.250738. PubMed PMID: 21778431.
+37. Poli-Neto OB, Meola J, Rosa ESJC, Tiezzi D. Transcriptome meta-analysis reveals differences of immune profile between eutopic endometrium from stage I-II and III-IV endometriosis independently of hormonal milieu. Sci Rep. 2020;10(1):313. Epub 20200115. doi: 10.1038/s41598-019-57207-y. PubMed PMID: 31941945; PubMed Central PMCID: PMC6962450.
+38. Pan H, Sheng J-Z, Tang L, Zhu R, Zhou T-H, Huang H-F. Increased expression of c-fos protein associated with increased matrix metalloproteinase-9 protein expression in the endometrium of endometriotic patients. Fertility and Sterility. 2008;90(4):1000-7. doi: 10.1016/j.fertnstert.2007.07.1386.
+39. Bult CJ, Sternberg PW. The alliance of genome resources: transforming comparative genomics. Mamm Genome. 2023;34(4):531-44. Epub 20230904. doi: 10.1007/s00335-023-10015-2. PubMed PMID: 37666946; PubMed Central PMCID: PMC10628019.
+40. Zeng Z, Lan T, Wei Y, Wei X. CCL5/CCR5 axis in human diseases and related treatments. Genes Dis. 2022;9(1):12-27. Epub 20210826. doi: 10.1016/j.gendis.2021.08.004. PubMed PMID: 34514075; PubMed Central PMCID: PMC8423937.
+41. Mootz M, Jakwerth CA, Schmidt-Weber CB, Zissler UM. Secretoglobins in the big picture of immunoregulation in airway diseases. Allergy. 2022;77(3):767-77. doi: 10.1111/all.15033.
+42. Masterton R, Armstrong EM, More IA. The cyclical variation in the percentage of ciliated cells in the normal human endometrium. J Reprod Fertil. 1975;42(3):537-40. doi: 10.1530/jrf.0.0420537. PubMed PMID: 47390.
+43. Von Wolff M, Strowitzki T, Becker V, Zepf C, Tabibzadeh S, Thaler CJ. Endometrial osteopontin, a ligand of β3-integrin, is maximally expressed around the time of the “implantation window”. Fertility and Sterility. 2001;76(4):775-81. doi: 10.1016/s0015-0282(01)02015-5.
+44. Tan O, Ornek T, Seval Y, Sati L, Arici A. Tenascin is highly expressed in endometriosis and its expression is upregulated by estrogen. Fertility and Sterility. 2008;89(5):1082-9. doi: 10.1016/j.fertnstert.2007.05.028.
+45. Harrington DJ, Lessey BA, Rai V, Bergqvist A, Kennedy S, Manek S, et al. Tenascin is differentially expressed in endometrium and endometriosis. J Pathol. 1999;187(2):242-8. doi: 10.1002/(sici)1096-9896(199901)187:2<242::aid-path221>3.0. co;2-t. PubMed PMID: 10365101.
+46. Gunning PW, Hardeman EC. Tropomyosins. Current Biology. 2017;27(1):R8-R13. doi: 10.1016/j.cub.2016.11.033.
+47. Nancy P, Siewiera J, Rizzuto G, Tagliani E, Osokine I, Manandhar P, et al. H3K27me3 dynamics dictate evolving uterine states in pregnancy and parturition. Journal of Clinical Investigation. 2017;128(1):233-47. doi: 10.1172/jci95937.
+48. Li Y, Wang R, Wang M, Huang W, Liu C, Fang Z, et al. RNA Sequencing of Decidua Reveals Differentially Expressed Genes in Recurrent Pregnancy Loss. Reproductive Sciences. 2021;28(8):2261-9. doi: 10.1007/s43032-021-00482-w.
+49. Duethorn B, Groll F, Rieger B, Drexler HCA, Brinkmann H, Kremer L, et al. Lima1 mediates the pluripotency control of membrane dynamics and cellular metabolism. Nature Communications. 2022;13(1). doi: 10.1038/s41467-022-28139-5.
+50. Xu Y, Wu D, Hui B, Shu L, Tang X, Wang C, et al. A novel regulatory mechanism network mediated by lncRNA TUG1 that induces the impairment of spiral artery remodeling in preeclampsia. Molecular Therapy. 2022;30(4):1692-705. doi: 10.1016/j.ymthe.2022.01.043.
+51. Uhlen M, Fagerberg L, Hallström BM, Lindskog C, Oksvold P, Mardinoglu A, et al. Proteomics. Tissue-based map of the human proteome. Science. 2015;347(6220):1260419. doi: 10.1126/science.1260419. PubMed PMID: 25613900.
+52. Lee JH, Kim TH, Oh SJ, Yoo JY, Akira S, Ku BJ, et al. Signal transducer and activator of transcription-3 (Stat3) plays a critical role in implantation via progesterone receptor in uterus. Faseb
+
+<--- Page Split --->
+
+957 j. 2013;27(7):2553- 63. Epub 20130326. doi: 10.1096/fj.12- 225664. PubMed PMID: 23531596; 958 PubMed Central PMCID: PMC3688751. 959 53. Kim BG, Yoo JY, Kim TH, Shin JH, Langenheim JF, Ferguson SD, et al. Aberrant activation of 960 signal transducer and activator of transcription- 3 (STAT3) signaling in endometriosis. Hum Reprod. 961 2015;30(5):1069- 78. Epub 20150306. doi: 10.1093/humrep/dev050. PubMed PMID: 25750101; 962 PubMed Central PMCID: PMC4400199. 963 54. Brosens I, Pijnenborg R, Vercruysse L, Romero R. The "Great Obstetrical Syndromes" are 964 associated with disorders of deep placentation. American Journal of Obstetrics and Gynecology. 965 2011;204(3):193- 201. doi: 10.1016/j.ajog.2010.08.009. 966 55. Garrido- Gómez T, Dominguez F, Quiñonero A, Estella C, Vilella F, Pellicer A, et al. Annexin A2 967 is critical for embryo adhesiveness to the human endometrium by RhoA activation through F- actin 968 regulation. FASEB J. 2012;26(9):3715- 27. Epub 2012/05/29. doi: 10.1096/fj.12- 204008. PubMed 969 PMID: 22645245. 970 56. Fleming TP, Watkins AJ, Velazquez MA, Mathers JC, Prentice AM, Stephenson J, et al. Origins 971 of lifetime health around the time of conception: causes and consequences. The Lancet. 972 2018;391(10132):1842- 52. doi: 10.1016/s0140- 6736(18)30312- x. 973 57. Benjamini Y, Hochberg Y. Controlling the False Discovery Rate: A Practical and Powerful 974 Approach to Multiple Testing. Journal of the Royal Statistical Society Series B (Methodological). 975 1995;57(1):289- 300. 976 58. Karlsson M, Zhang C, Mear L, Zhong W, Digre A, Katona B, et al. A single- cell type 977 transcriptomics map of human tissues. Science Advances. 2021;7(31):eabh2169. doi: 978 10.1126/sciadv.abh2169. 979 59. La Manno G, Soldatov R, Zeisel A, Braun E, Hochgerner H, Petukhov V, et al. RNA velocity of 980 single cells. Nature. 2018;560(7719):494- 8. Epub 20180808. doi: 10.1038/s41586- 018- 0414- 6. 981 PubMed PMID: 30089906; PubMed Central PMCID: PMC6130801. 982 60. Bergen V, Lange M, Peidli S, Wolf FA, Theis FJ. Generalizing RNA velocity to transient cell states 983 through dynamical modeling. Nat Biotechnol. 2020;38(12):1408- 14. Epub 20200803. doi: 984 10.1038/s41587- 020- 0591- 3. PubMed PMID: 32747759. 985 61. Li S, Zhang P, Chen W, Ye L, Brannan KW, Le N- T, et al. A relay velocity model infers cell- 986 dependent RNA velocity. Nature Biotechnology. 2024;42(1):99- 108. doi: 10.1038/s41587- 023- 01728- 987 5. 988 62. Street K, Risso D, Fletcher RB, Das D, Ngai J, Yosef N, et al. Slingshot: cell lineage and 989 pseudotime inference for single- cell transcriptomics. BMC Genomics. 2018;19(1). doi: 990 10.1186/s12864- 018- 4772- 0. 991 63. Dann E, Henderson NC, Teichmann SA, Morgan MD, Marioni JC. Differential abundance testing 992 on single- cell data using k- nearest neighbor graphs. Nature Biotechnology. 2022;40(2):245- 53. doi: 993 10.1038/s41587- 021- 01033- z. 994 64. Finak G, Mcdavid A, Yajima M, Deng J, Gersuk V, Shalek AK, et al. MAST: a flexible statistical 995 framework for assessing transcriptional changes and characterizing heterogeneity in single- cell RNA 996 sequencing data. Genome Biology. 2015;16(1). doi: 10.1186/s13059- 015- 0844- 5. 997 65. Gormley M, Oliverio O, Kapidzic M, Ona K, Hall S, Fisher SJ. RNA profiling of laser 998 microdissected human trophoblast subtypes at mid- gestation reveals a role for cannabinoid signaling 999 in invasion. Development. 2021;148(20). doi: 10.1242/dev.199626.
+
+<--- Page Split --->
+
+# 1. 1. 1. 1. 1. 1. 1. 2. 2. 2. 2. 2. 2. 2.
+
+<--- Page Split --->

preprint/preprint__0917f47f7945ea8fabfadaf8a77e22df7840ae52c0729198350fb1b6dd32dbf5/images_list.json

@@ -0,0 +1,92 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1: Landscape of somatic alterations in acral and sun-exposed melanomas. a-b, Genomic and clinical characterization of acral and sun-exposed melanoma samples sequenced in this work. a, The number of nonsynonymous SNVs and indels per melanoma exome (columns), cohort, age, frequently mutated genes (at least \\(5\\%\\) recurrence frequency in either melanoma subtype), nonsynonymous base substitution frequencies, and dominant COSMIC mutational signatures \\(^{92,93}\\) . Sig., signature; 5mC, 5-methylcytosine. b, The number of significant focal amplifications and deletions (GISTIC \\(Q < 0.05\\) ) per melanoma exome (columns), ordered identically to panel a. Cytobands with focal amplifications or deletions with at least \\(10\\%\\) recurrence frequency in either melanoma subtype are shown (GISTIC \\(Q < 10^{-5}\\) ), ordered by the relative difference in recurrence frequency in acral versus sun-exposed melanoma. c, Genes are plotted according to the fraction of acral (y-axis) or sun-exposed (x-axis) tumors where they are present with \\(4+\\) copies. Considering the genome-wide distribution of differences in recurrence frequencies between melanoma subtypes, genes are identified as significantly recurrent in acral or sun-exposed melanomas if their \\(|z\\text{-score}|\\geq 3\\) (dashed lines). Significantly recurrent genes are colored according to their cytoband location (inset). For clarity, a small amount of jitter was added to distinguish overlapping genes.",
+    "footnote": [],
+    "bbox": [
+      [
+        113,
+        90,
+        884,
+        536
+      ]
+    ],
+    "page_idx": 15
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2: Focal amplifications in 22q11.21 are linked to shorter survival time, regional metastasis, and depletion of immunomodulatory programs in acral melanoma. a, Association between recurrent somatic alterations and overall survival (OS) in acral melanoma. Z-scores with positive and negative values indicated shorter and longer survival time, respectively \\((|Z| > 1.96\\) is significant at \\(P< 0.05\\) ; Methods). OS was calculated from the date of diagnosis (x-axis) and the date of tumor resection (y-axis). Shown are genes with a nonsynonymous mutation frequency of at least \\(5\\%\\) in either melanoma subtype and focal copy number events with at least \\(10\\%\\) recurrence frequency in each acral cohort (Supplementary Table 6). Focal amplifications in",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        90,
+        884,
+        734
+      ]
+    ],
+    "page_idx": 16
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3: Cell proliferation in response to suppression of LZTR1. a, Impact of LZTR1 knockdown on cell proliferation in nine primary melanoma cell lines and two normal human melanocyte lines (NBMEL). Key mutations are indicated (Supplementary Table 9). Bar plots depict fold change between the \\(3^{\\text{rd}}\\) and \\(6^{\\text{th}}\\) day after infection with LZTR1 shRNA (numbered), as compared to control (scrambled) shRNA ('C'). All shRNAs significantly reduced proliferation relative to control ( \\(P < 0.05\\) ; two-sided \\(t\\) test with unequal variance). b, Western blot showing the efficiency of LZTR1 knockdown in primary acral and sun-exposed melanoma cell lines, and in normal melanocytes, related to panel a. c, Cell proliferation (left) and LZTR1 expression (right) of a sun-exposed melanoma cell line that lost one LZTR1 allele in response to genomic modification by different CRISPR-Cas9 sgRNAs targeting",
+    "footnote": [],
+    "bbox": [
+      [
+        190,
+        95,
+        812,
+        707
+      ]
+    ],
+    "page_idx": 18
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4: Changes in MAPK signaling in response to LZTR1 knockdown. a, b, Impact of LZTR1 loss on (a) RAS-GTP activity as measured by RAS-GTPase Activation ELISA assay, and (b) RAS levels, five days after shLZTR1 infection. c, Effect of LZTR1 loss on MAPK activity. Key somatic events are indicated below. WT, wildtype. d, e Increased levels of BRAF activity in response to shLZTR1 is associated with increase pERK. In panel e, cells were incubated with RAF the kinase inhibitors PLX4032 (500 nM) or LY3009120 (100 nM) for four hours at the end of treatment with shRNA. f, RAS translocation to the cytoplasm in response to shLZTR1; a-d, RAS is visualized by staining with magenta; Green (Cy2) indicates GM130 (a, b) and calnexin (c, d). Scale",
+    "footnote": [],
+    "bbox": [
+      [
+        155,
+        108,
+        860,
+        737
+      ]
+    ],
+    "page_idx": 20
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5: Impact of LZTR1 knockdown on apoptosis, Rb signaling, and pigmentation. a, Gene Set Enrichment Analysis (GSEA)94 showing concordance in hallmark pathways among bulk acral melanoma tumors, acral melanoma single-cell",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        90,
+        835,
+        840
+      ]
+    ],
+    "page_idx": 22
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6: LZTR1 and CRKL confer properties consistent with malignant cell transformation and metastasis initiation. a, Morphological changes and spheroid formations in early passage normal human melanocytes (NBMEL C1220) overexpressing LZTR1 and/or CRKL. Top: Phase-contrast images of parental and infected cells in 2D culture. LZTR1 images were taken after two days induction with doxycycline (200 ng/ml), CRKL after three days of infection with PLX304-CRKL, and LZTR1+CRKL after six days infection of LZTR1 melanocytes with PLX304-CRKL and three days stimulation with doxycycline. Insert in LZTR1 shows that colonies were",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        90,
+        881,
+        750
+      ]
+    ],
+    "page_idx": 24
+  }
+]

preprint/preprint__09855f388022d98f6d42aad5171260d00fc3df0376deedaa6931278a022350c0/images_list.json

@@ -0,0 +1,115 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "FIG. 1: Patterns of transcriptional activity. A. Schematic of the model. Twenty TFs (pink) that switch between on/off states at rate \\(\\alpha = 10^{-5} \\tau_{\\mathrm{B}}^{-1}\\) (with \\(\\tau_{\\mathrm{B}}\\) the Brownian time, see SI, Supplementary Note 1) or \\(0.001 \\mathrm{~s}^{-1}\\) bind specifically to 39 TUs (red beads) randomly positioned along the chain, and non-specifically to other beads (blue). A TU is considered transcriptionally active if associated with a TF. B. Example conformation (TFs not shown). Some beads cluster and form loops; one TU not in a cluster (and not transcribed) is green, and another that is in a cluster (and transcribed) is yellow. Inset: zoom of boxed region. C. Transcriptional activity for each TU bead averaged over 1000 simulations (each lasting \\(10^{5} \\tau_{\\mathrm{B}}\\) ). TUs are grouped according to activity, with red, green and blue bars showing high \\((>70\\%)\\) , medium \\((20 - 70\\%)\\) and low \\((< 20\\%)\\) activity, respectively. This gives a population-level measure of activity. D. Variation of activity across simulations (reflecting cell-to-cell variation) for 3 representative TUs with high (red), medium (green), or low (blue) average activity (defined as in C).",
+    "footnote": [],
+    "bbox": [
+      [
+        92,
+        63,
+        525,
+        415
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "FIG. 2: Transcriptional bursting. A. Snapshots showing a 100-bead section of the simulated chain taken at different times. Initially, none of the 5 TUs (red) are in clusters and are inactive; later, 4 TUs join a cluster and are close to TFs – and so are transcribed. B. Kymograph where each row shows the changing transcription state of one TU during a simulation; pixels are colored red if the bead is associated with a TF and so transcribed, or black otherwise. White rectangle: example burst.",
+    "footnote": [],
+    "bbox": [
+      [
+        90,
+        61,
+        530,
+        312
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "FIG. 3: Regulatory networks formed by TU beads are percolating at low TF concentrations. Simulations (as Fig. 1, with \\(\\geq 800\\) simulations/condition) with different average numbers of active TFs \\((n)\\) and switching rate \\((\\alpha)\\) . Networks were constructed by calculating the Pearson correlation between the transcription time-series for all pairs of TUs; nodes represent each of the 39 TUs and edges are placed between nodes where there is a significant correlation \\((>0.15\\) in absolute value, corresponding to \\(p < 10^{-6}\\) ). A. Effect of TF concentration and switching. 39 nodes are shown around the perimeter, and thick black and grey lines denote positive and negative correlations between transcriptional activities of bead pairs. B. Effect of \\(n\\) on the fraction of nodes in the largest connected component.",
+    "footnote": [],
+    "bbox": [
+      [
+        545,
+        61,
+        969,
+        310
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "FIG. 4: Modelling SNPs and eQTL action. Sets of simulations ( \\(\\geq 800\\) simulations/condition) where each of the 39 TU beads is made non-binding in turn (to represent 39 different SNPs in regulatory elements) are compared with those with the \"wild-type\" chain (as Fig. 1). A. Chain with mutant (nonbinding) TU bead 930. (i) Snapshot. TFs not shown (inset: same structure without blue beads). (ii) Transcriptional rates of the 17 TUs with significantly different values in mutant first compared with the wild type one ( \\(p \\simeq 0.046\\) ; Students t-test). (iii) Regulatory network inferred from the matrix of Pearson correlations between activities of TUs. (iv) Change in Pearson correlation between TUs. B. Results from simulations where each TU bead is mutated in turn, and the \"transcriptional difference\" from the wild type (see text) determined. (i) Transcriptional difference versus position along the chain. (ii) Positive correlation of transcriptional difference with TU activity in wild-type. The plot shows that if we mutate a TU with high transcriptional activity, this leads to a larger difference.",
+    "footnote": [],
+    "bbox": [
+      [
+        93,
+        60,
+        520,
+        392
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "FIG. 5: Looping subtly affects transcriptional activity. Results of two sets of simulations are compared; one set as Fig. 1, in the other the chain contains 8 permanent loops (to represent convergent loops stabilized by cohesin/CTCF). A. Snapshot (beads within loops are magenta; TFs not shown; inset – same structure with only TUs shown). B. Average transcriptional activity for each TU in the looped chain (magenta bars – values for TUs in loops; magenta arcs – loop positions). C. Comparison between activity in wild-type and looped configuration for the 25 TUs with significantly different values in the two sets ( \\(p \\simeq 0.003\\) ; Students t-test). D. Regulatory network inferred from the matrix of Pearson correlations between expression of TUs (as Fig. 3A). E. Change in Pearson correlation between TUs due to loops.",
+    "footnote": [],
+    "bbox": [
+      [
+        88,
+        202,
+        520,
+        501
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "FIG. 6: Neighboring heterochromatin affects transcriptional activity. Results from two sets of simulations (at least 800 runs for each condition) are compared; one set as Fig. 1, in the other beads around TU beads 905, 907, 930 and 931 (from bead 901 to 940) are non-binding (to represent embedding the TU beads in heterochromatin). A. Snapshot with heterochromatic beads shown in gray (TFs not shown; inset – the same structure with only TUs). B. Average transcriptional activity for each TU. C. Comparison of average transcriptional activity with respect to wild-type for the 22 TUs with significantly different values in the two sets ( \\(p \\simeq 0.003\\) ; Students t-test). D. Regulatory network inferred from the matrix of Pearson correlations between activities of TUs (as Fig. 3A). E. Change in Pearson correlation between TUs due to heterochromatin.",
+    "footnote": [],
+    "bbox": [
+      [
+        540,
+        58,
+        976,
+        404
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_7.jpg",
+    "caption": "FIG. 7: Comparison of transcriptional activities of TUs on HSA14 in HUVECs determined using simulations and GRO-seq. A. Workflow (DHS model). Simulations (244 runs) involve a chain (35784 beads) representing HSA14, and 1700 switchable TFs confined in an ellipsoidal territory. Beads are classified as TUs (red, strong-binding), euchromatic (blue, weak-binding) or heterochromatic (grey, non-binding). Transcriptional activities from simulations are compared with those of GRO-seq data, by measuring the Spearman rank correlation. B. (i) Snapshot (TFs not shown). (ii, iii) TU beads and TFs in this configuration. C. Comparison of transcriptional activities of TUs from simulations and GRO-seq (ranked from \\(0 - 100\\%\\) , then binned in quintiles and showed as a heat map). A scatter plot of unbinned ranks of beads corresponding to SEs are superimposed (white circles). D. Comparison of transcriptional activities from simulations (for both DHS and HMM models) and GRO-seq for all 3 kb regions/beads, only TUs, and only connected patches of binding beads (see text). All correlations are significant \\((p < 10^{-6}\\) , indicated by grey lines). E (i-ii) Capture-HiC-like contact maps obtained from simulations and experiments [42] showing logarithm of number of contacts between 30 kbp bins which contain TUs.",
+    "footnote": [],
+    "bbox": [
+      [
+        585,
+        64,
+        927,
+        562
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_8.jpg",
+    "caption": "FIG. 8: Modelling effects of the DiGeorge deletion in HSA22. A. Workflow. Simulations (800 simulations/condition) for wild type (17102 beads) and deletion (16250 beads, where wild-type beads 6305 - 7156 are cut, corresponding to a deletion of chr22:18912231 - 21465672 in hg19). [Agreement between predicted transcriptional activity and GRO-seq in HSA22 is similar to that found for HSA1 (here, Spearman correlation is \\(r \\sim 0.29\\) , \\(p < 10^{-6}\\) ]. B. (i) Manhattan plot showing \\(-\\log_{10}\\) (p-value) as a function of genomic position along HSA22 (position given in Mbp), for changes in TU transcriptional activities between wild-type and deletion. (ii) Quantile-quantile plot showing expected versus observed values for \\(-\\log_{10}\\) (p-value) for the same data in (i). Expected values are computed from the normal distribution (these correspond to the null hypothesis according to which the change in transcriptional activities in the deletion is purely due to random variation). C. Regulatory networks of two 3 Mbp segments in chromosome 22 inferred from the Pearson correlation matrix. Edges show positive correlations \\(> 0.12\\) ( \\(p = 0.0007\\) ). Segments chosen have roughly the same number of nodes in 3 Mbp as the short fragment (Fig. 3Aii).",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 9
+  }
+]

preprint/preprint__09855f388022d98f6d42aad5171260d00fc3df0376deedaa6931278a022350c0/preprint__09855f388022d98f6d42aad5171260d00fc3df0376deedaa6931278a022350c0.mmd

@@ -0,0 +1,269 @@
+
+# Complex small-world regulatory networks emerge from the 3D organisation of the human genome  
+
+Chris Brackley  University of Edinburgh  
+
+Nick Gilbert  University of Edinburgh  https://orcid.org/0000- 0003- 0505- 6081  
+
+Davide Michieletto  University of Edinburgh  
+
+Argyris Papantonis  University of Göttingen  
+
+Maria Pereira  University of Edinburgh  
+
+Peter Cook  Oxford University  https://orcid.org/0000- 0002- 6639- 188X  Davide Marenduzzo (dmarendu@ph.ed.ac.uk)  University of Edinburgh  https://orcid.org/0000- 0003- 3974- 4915  
+
+## Article  
+
+Keywords: regulatory networks, transcription factors, human genome  
+
+Posted Date: June 12th, 2021  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 566854/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License.  Read Full License  
+
+Version of Record: A version of this preprint was published at Nature Communications on October 1st, 2021. See the published version at https://doi.org/10.1038/s41467- 021- 25875-y.
+
+<--- Page Split --->
+
+# Complex small-world regulatory networks emerge from the 3D organisation of the human genome  
+
+C. 
+A. Brackley<sup>1</sup>, 
+N. Gilbert<sup>2</sup>, 
+D. Michieletto<sup>1,2</sup>, 
+A. Papantonis<sup>3</sup>, 
+M. 
+C. 
+F. Pereira<sup>1</sup>, 
+P. 
+R. Cook<sup>4</sup>, 
+D. Marenduzzo<sup>1</sup>  <sup>1</sup>SUPA, School of Physics and Astronomy, University of Edinburgh, Peter Guthrie Tait Road, Edinburgh EH9 3FD, UK  <sup>2</sup>MRC Human Genetics Unit, MRC Institute of Genetics & Molecular Medicine, University of Edinburgh, Western General Hospital, Edinburgh EH4 2XU, UK  <sup>3</sup>Institute of Pathology, University Medical Center, Georg-August University of Gottingen, 37075 Gottingen, Germany  <sup>4</sup>Sir William Dunn School of Pathology, University of Oxford, South Parks Road, Oxford OX1 3RE, UK  
+
+The discovery that overexpressing one or a few critical transcription factors can switch cell state suggests that gene regulatory networks are relatively simple. In contrast, genome- wide association studies (GWAS) point to complex phenotypes being determined by hundreds of loci that rarely encode transcription factors and which individually have small effects. Here, we use computer simulations and a simple fitting- free polymer model of chromosomes to show that spatial correlations arising from 3D genome organisation naturally lead to stochastic and bursty transcription as well as complex small- world regulatory networks (where the transcriptional activity of each genomic region subtly affects almost all others). These effects require factors to be present at sub- saturating levels; increasing levels dramatically simplifies networks as more transcription units are pressed into use. Consequently, results from GWAS can be reconciled with those involving overexpression. We apply this pan- genomic model to predict patterns of transcriptional activity in whole human chromosomes, and, as an example, the effects of the deletion causing the diGeorge syndrome.  
+
+## INTRODUCTION  
+
+Transcription - the copying of DNA into RNA - is tightly regulated. Early insights into regulatory mechanisms came from work on binary on/off genetic switches controlled by one (or just a few) transcription factors such as the lambda and lac repressor in Escherichia coli [1]. Similar regulatory mechanisms are present in eukaryotes, albeit with additional complexity. For instance, a fibroblast cell can be reprogrammed into a muscle cell by a single master regulator (MYOD) [2, 3], or into pluripotent stem cells by four Yamanaka factors (Oct4, Sox2, c- Myc, Klf4) [4].  
+
+Genome- wide association studies (GWAS) lead to quite a different view: gene regulation is widely distributed and involves interactions between hundreds (perhaps thousands) of loci scattered around the genome [5, 6]. GWAS allow quantitative trait loci (QTLs) affecting any measurable genetic trait to be ranked in an unbiased way. With complex traits like human height, and diseases such as schizophrenia and type II diabetes, the top ten QTLs in the rank order combine to yield only modest effects, while the top one- hundred still account for less than half of the total genetic effect. Hundred more QTLs are expected to be identified as sample sizes and data resolution improve [5- 7]. Expression QTLs (eQTLs) are QTLs affecting transcription of other DNA regions. Perhaps surprisingly, these are rarely found in genes encoding transcription factors or other proteins; instead, they usually involve single- nucleotide changes in non- coding elements that bind transcription factors such as active enhancers and promot  
+
+ers [8- 10].  
+
+Results from GWAS lead to the view that most generegulatory networks are incredibly complex, with the activity of a given gene being affected by a panoply of eQTLs, each having a tiny effect. This is captured by the "omnic gene" model, which is based on a set of gene- interaction equations [5, 6] such that the activity of almost any gene affects that of almost every other one. This model provides a useful and appealing framework to view GWAS results. However, it is difficult to compare its outputs with experimental data because it contains many parameters that are currently unknown and require fitting to training datasets.  
+
+In general, existing models for gene regulation traditionally assume post- transcriptional and biochemically mediated interactions between different genes [11, 12], and disregard the role of 3D chromatin structure. Here we propose an alternative but complementary framework that links transcriptional regulation directly to 3D genome structure, deliberately neglecting downstream biochemical regulation to enable unambiguous interpretation of our results. This framework is motivated by experiments showing that chromatin folding can lead to contacts between enhancers and promoters affecting transcription, and that 3D structure changes in disease [13, 14]. Additionally, because our modelling is essentially fitting- free, its output can be directly compared to experiments. When the agreement is good, our model is validated; when poor, it points to some missing ingredient (such as biochemical feedback) that could be included in future models.  
+
+Here, we use stochastic computer simulations of a polymer model for chromosome organization, in which a chain
+
+<--- Page Split --->
+
+of beads represents a chromatin fibre, and a set of spheres complexes of transcription factors and RNA polymerases - which we will call "TFs" for short. Some chromatin beads are identified as transcription units (TUs), and we call them TU beads. They contain binding sites for TFs, and can be sites of transcriptional initiation (we do not discriminate between genic and non- genic promoters). As a simple starting point we only consider one type of TF that binds specifically and multivalently to TU beads, and non- specifically (i.e., with weak affinity) to every other bead. We perform 3D Brownian dynamics simulations that evolve the diffusive dynamics of the chain and associated factors. We previously showed that similar polymer models yield structures resembling those seen using chromosome- conformation- capture (3C) [15- 19] and microscopy [20]. Here, we link 3D structure to expression and transcriptional dynamics by measuring how often a TU bead is transcribed - which we do by computing the fraction of time it binds a TF. To establish the methodology, we model a 3 Mbp chromatin fragment, before going on to simulate whole human chromosomes.  
+
+Our simulations capture many features of eukaryotic regulation. For example, transcription is stochastic and bursty (in agreement with single- cell transcriptomics data), and the predicted pattern of transcriptional activity in human chromosomes correlates significantly with that observed experimentally. We also find that small- world (percolating) networks that encapsulate much of the rich complexity observed in GWAS emerge through spatial effects alone. In other words, the activity of most (probably all) TUs in our model is affected by the activity of most (probably all) other segments in the genome. We find such pan- genomic regulation critically requires non- saturating concentrations of TFs - as normally found in vivo - and that increasing concentrations dramatically simplifies the networks. This enables us to reconcile the GWAS- based view that regulatory networks are complicated with the observation that overexpressing one or a few TFs can decisively alter cell state.  
+
+## RESULTS  
+
+We first consider a simple system where a 3 Mbp chromatin fragment is represented by a chain of 1000 beads (each \(30\mathrm{nm}\) in diameter, and corresponding to \(3\mathrm{kbp}\) ). We select at random \(N = 39\) beads and identify them as TUs (Fig. 1A; full details in SI, Supplementary Notes 1 and 2). The linear density of TU in the fragment is similar to that in human chromosome 22. Additionally, \(n\) spheres (also 30 nm in diameter) represent TFs (recall these are complexes of transcription factors and RNA polymerase II). TFs bind reversibly to TUs via a strong attractive interaction, and to all other beads weakly and non- specifically. An important feature is that TFs switch between active (binding) and inactive (non- binding) state at rate \(\alpha\) . Many factors switch like this in vivo (e.g., due to phosphorylation and de- phosphorylation), and switching is required to account for the rapid exchange of factors and polymerases between bound and free states seen in live- cell photobleaching experiments [21]. As \(\sim 7\) out of 8 polymerases attempting to initiate at promoters dissociate with a half- life of \(\sim 2.4\) s [22], our complexes generally behave like those in vivo.   
+
+While our results refer to a single patterning of TUs along the fibre, they are representative of any arbitrary random positioning of TUs: in other words the qualitative trends we present below are robust and do not depend on the particular choice of the 1D pattern of TU along the fibre in any way.  
+
+We say a TU bead is transcribed whenever a TF lies close to it (SI, Supplementary Note 1), and the transcriptional activity of a TU is then the fraction of time it is transcribed during a simulation. To reflect the situation in mammalian cells (Supplementary Note 4 and [23]), we typically assume there are fewer TFs than TU beads (i.e., \(n = 10\) TFs in the active binding state at any time, compared to 39 TUs).  
+
+By interrogating TF- chromatin interactions at regular time intervals over hundreds of simulations we build up a population picture of transcription. A typical configuration of the 3 Mbp fragment is shown in Figure 1B. Strikingly, bound TFs spontaneously cluster, despite there being no attractive interactions between TUs or between TFs. Such clustering is driven by the "bridging- induced attraction" [16, 24, 25] that arises due to a positive feedback: when a TF forms a molecular bridge between two chromatin regions and forms a loop, the local chromatin concentration increases, making further TF binding more likely. Clusters then grow until limited by entropic costs of crowding (Fig. S1A). Most of the non- trivial phenomena described below result from such clustering. Clustering requires TF multivalency, as monovalent factors do not cluster [24]. However, the assumption of multivalency, which is common in the polymer physics literature [15], is well- founded. Several TFs are known to be bivalent or multivalent [26], and, more importantly, our spheres represent complexes of TFs and polymerases, so they will behave as multivalent binders even when the individual TFs in the complex are monovalent. Although clustering does not require any interactions between TFs, adding a weak attraction between them, as might arise for instance due to macromolecular crowding or electrostatic interactions between intrinsically disordered regions, should not qualitatively change any of the results discussed here (at least as long as TFs still microphase separate into clusters rather than undergoing macroscopic phase separation).  
+
+The clusters we observe, and which emerge through the bridging- induced attraction, are qualitatively similar to those seen in vivo, which are variously described as transcriptional compartments, hubs, super- enhancer (SE) clusters, phase- separated droplets/condensates, and factories [7, 10, 27- 29]. They are also similar to the contact domains seen in microC [30], which are formed by acces
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>FIG. 1: Patterns of transcriptional activity. A. Schematic of the model. Twenty TFs (pink) that switch between on/off states at rate \(\alpha = 10^{-5} \tau_{\mathrm{B}}^{-1}\) (with \(\tau_{\mathrm{B}}\) the Brownian time, see SI, Supplementary Note 1) or \(0.001 \mathrm{~s}^{-1}\) bind specifically to 39 TUs (red beads) randomly positioned along the chain, and non-specifically to other beads (blue). A TU is considered transcriptionally active if associated with a TF. B. Example conformation (TFs not shown). Some beads cluster and form loops; one TU not in a cluster (and not transcribed) is green, and another that is in a cluster (and transcribed) is yellow. Inset: zoom of boxed region. C. Transcriptional activity for each TU bead averaged over 1000 simulations (each lasting \(10^{5} \tau_{\mathrm{B}}\) ). TUs are grouped according to activity, with red, green and blue bars showing high \((>70\%)\) , medium \((20 - 70\%)\) and low \((< 20\%)\) activity, respectively. This gives a population-level measure of activity. D. Variation of activity across simulations (reflecting cell-to-cell variation) for 3 representative TUs with high (red), medium (green), or low (blue) average activity (defined as in C). </center>  
+
+sible DNA sites clustering together in 3D space. Clustering arising through the bridging- induced attraction has recently been found in vitro for systems of DNA and cohesin (which binds multivalently to DNA) [31].  
+
+## Transcriptional activity varies along the chromatin fibre and is highly stochastic  
+
+As TFs have the same affinity for all TUs, one might expect each TU to be bound with equal likelihood; however,  
+
+transcriptional activity (the fraction of time a TU is transcribed) varies from \(\sim 10 - 90\%\) (Fig. 1C). What causes this variation? As TF copy number is limiting, and as bound TFs cluster, most transcription occurs in clusters – as is the case in vivo [7, 32–34]. Since TUs are positioned irregularly along the fragment, some have closer neighbours in 1D sequence space than others, and these are inevitably the ones most likely to cluster and be transcribed. Instead, those far from their neighbours are less likely to cluster and are less active. Accordingly, the transcriptional activity of a TU anticorrelates with distance to the nearest TU along the fibre (Fig. S1B; the Spearman correlation is \(r \simeq - 0.94\) , p- value \(p < 10^{- 12}\) ).  
+
+Whilst Figure 1C pertains to population averages of 1000 simulations, it is informative to consider each simulation independently (as in single- cell transcriptomics). Such analysis shows that transcriptional activity is stochastic, varying substantially from simulation to simulation: a TU active in some simulations may be silent in others (Fig. 1D).  
+
+## Transcriptional bursting  
+
+During a simulation, chromatin conformation can change dramatically (Fig. 2A). Such changes often yield transcriptional "bursts" – periods of continued activity followed by silent periods (Fig. 2B) – as TUs with intermediate levels of activity repeatedly join a cluster to give a burst and then dissociate. Notably, TUs lying close to each other in sequence space often start and stop bursts coordinately due to the intrinsic positive feedback in the system (Fig. S1A).  
+
+These results are consistent with experimental observations: single cell Hi- C [35] and transcriptomics [36] show that the structure and function of each individual cell is unique, and bursting is well documented [37–40] with nearby promoters often firing together [38].  
+
+## Local chromatin architecture creates small-world percolating transcription networks  
+
+To investigate correlations between transcriptional activities of different TUs, we compute the Pearson correlation matrix between the activity of all possible TU pairs, and identify an emergent regulatory network in which TUs form nodes (Fig. 3A and Fig. S2). Specifically, we draw an edge between two TUs whenever there is a statistically significant positive or negative correlation between their transcriptional dynamics (Fig. 3A). This network arises only due to spatial interactions, as we assume no underlying biochemical regulation.  
+
+The network shows a striking property. With \(n = 10\) active TFs, most nodes are connected (Fig. 3Aii), and the fraction of TUs participating in the largest connected com
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>FIG. 2: Transcriptional bursting. A. Snapshots showing a 100-bead section of the simulated chain taken at different times. Initially, none of the 5 TUs (red) are in clusters and are inactive; later, 4 TUs join a cluster and are close to TFs – and so are transcribed. B. Kymograph where each row shows the changing transcription state of one TU during a simulation; pixels are colored red if the bead is associated with a TF and so transcribed, or black otherwise. White rectangle: example burst. </center>  
+
+ponent is close to 1 (Fig. 3B). Such a network is said to be "percolating", which means that any two nodes are connected by a path along edges. Our percolating networks are also "small- world", which means that most nodes can be reached from every other node by a small number of steps [41] – we provide quantitative measurements of the small world- ness of our networks in the SI (Supplementary Note 5). The small- world phenomenology is consistent with the multitude of small- effect eQTLs detected by GWAS [5, 6]. Notably, the regulation we observe act at the transcriptional level, and not post- transcriptionally as envisaged by the omnigenic model [5, 6].  
+
+How might our simple model give rise to complex regulatory networks? By analysing simulation trajectories, we noted that TUs lying near each other in 1D sequence space often joined the same cluster in 3D. As a result, the activity of these clustered beads is highly positively correlated. At the same time, cluster formation sequesters TFs and so reduces the likelihood that another cluster forms elsewhere. As a result, most long- range correlations are negative (Fig. 3A).  
+
+Crucially, these network properties depend on there being a low TF copy- number (as in vivo [23]) so TU beads do not become saturated. We therefore reasoned that increasing copy number should suppress correlations as more rarely- transcribed TUs are pressed into use. Indeed, increasing \(n\) reduces long- range negative correlations  
+
+![](images/Figure_3.jpg)
+
+<center>FIG. 3: Regulatory networks formed by TU beads are percolating at low TF concentrations. Simulations (as Fig. 1, with \(\geq 800\) simulations/condition) with different average numbers of active TFs \((n)\) and switching rate \((\alpha)\) . Networks were constructed by calculating the Pearson correlation between the transcription time-series for all pairs of TUs; nodes represent each of the 39 TUs and edges are placed between nodes where there is a significant correlation \((>0.15\) in absolute value, corresponding to \(p < 10^{-6}\) ). A. Effect of TF concentration and switching. 39 nodes are shown around the perimeter, and thick black and grey lines denote positive and negative correlations between transcriptional activities of bead pairs. B. Effect of \(n\) on the fraction of nodes in the largest connected component. </center>  
+
+(Fig. 3Aiii,iv), and the fraction of nodes in the largest- connected component falls (Fig. 3B). Another way to think about this result is: if resources are plentiful, there is no need for sharing or competition, and all TUs can bind a TF independently of each other. If TFs do not switch and are permanently in the binding state (and \(n = 10\) ), the network becomes even more highly connected (Fig. 3Ai).  
+
+## Modelling effect of mutations and SNPs in regulatory elements  
+
+GWAS reveals that single- nucleotide polymorphisms (SNPs) in regulatory elements and TUs can lead to many small changes in transcriptional activity across the genome. To model this, we abrogate TF binding to one TU in the chain. Bead 930 is chosen first because it is usually highly active (Fig. 1C). This single "knock- out" affects in a statistically significant way the activity of almost half of the other TUs, both near and far away in sequence space (Fig. 4Aii). The immediately adjacent TU (i.e., bead 931) is down- regulated the most, while more distant ones are
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>FIG. 4: Modelling SNPs and eQTL action. Sets of simulations ( \(\geq 800\) simulations/condition) where each of the 39 TU beads is made non-binding in turn (to represent 39 different SNPs in regulatory elements) are compared with those with the "wild-type" chain (as Fig. 1). A. Chain with mutant (nonbinding) TU bead 930. (i) Snapshot. TFs not shown (inset: same structure without blue beads). (ii) Transcriptional rates of the 17 TUs with significantly different values in mutant first compared with the wild type one ( \(p \simeq 0.046\) ; Students t-test). (iii) Regulatory network inferred from the matrix of Pearson correlations between activities of TUs. (iv) Change in Pearson correlation between TUs. B. Results from simulations where each TU bead is mutated in turn, and the "transcriptional difference" from the wild type (see text) determined. (i) Transcriptional difference versus position along the chain. (ii) Positive correlation of transcriptional difference with TU activity in wild-type. The plot shows that if we mutate a TU with high transcriptional activity, this leads to a larger difference. </center>  
+
+up- regulated (due to loss of a strong competitor). This knock- out also rewires the whole network, even though it still retains its small- world character (Fig. 4Aiii). Both positive and negative interactions are affected along the whole chain, as shown by a heat map of the change in Pearson correlation between TU transcriptional activities (Fig. 4Aiv).  
+
+We next systematically knock out each TU in turn. To quantify global effects, we define a "transcriptional difference" between the wild- type and each knock- out based on a standard Euclidian- distance metric (SI, Supplementary Note 3); the larger this quantity, the more different the two states are. This difference varies \(> 10\) - fold between different mutations (Fig. 4Bi).  
+
+Together, these observations are reminiscent of the behaviour of SNPs and eQTLs. Thus, each TU mutant can be seen as a SNP underlying an eQTL; then, those with low and high transcriptional differences (Fig. 4Bi,ii) are low- and high- effect eQTLs (low- effect mutants are often isolated in sequence space), and those with wide effects (e.g., bead 930 in Fig. 4A) may be viewed as omnigenic.  
+
+## Modelling loops, heterochromatin and euchromatin  
+
+In mammalian genomes, promoter- enhancer pairs are often contained in loops stabilized by cohesin and the CCCTC- binding factor (CTCF) [42- 44]. To investigate how such loops might affect transcription, we incorporated eight permanent and non- overlapping loops at different positions in the chain (Fig. 5A, loops \(a - h\) ). In reality, such loops may arise from extrusion by cohesin halted at convergent CTCF loops [42]. Our assumption of stable, permanent loops is quantitatively accurate in the limit in which the interaction between cohesin and CTCF is strong and long- lived. However, we expect the trends to be qualitatively similar for more transient loops consistent with the loop extrusion model as in [19, 43].  
+
+The inclusion of stable loops has subtle effects. For example, loop \(h\) encompasses three TUs (beads 905, 907, 930), and expression of one is slightly boosted compared to the unlooped case (Figs. 5B,C). This is consistent with the idea that looping switches on some genes during development [45], and increases enhancer- promoter interactions [46]. However, up- regulation requires appropriate positioning of a TU within the loop. For instance, loop \(d\) encompasses two TUs (beads 396, 404), and has no effect on their activity. Broadly speaking, looping up- regulates activity, but not invariably so, and - perhaps surprisingly - two of the three most up- regulated TUs (beads 33, 886) are not contained in loops (Fig. 5C). Looping also extensively rewires the regulatory network (Fig. 5D,E). Globally, the increase in activity is modest, as incorporating all beads into closely- packed loops only increases total activity by \(\sim 10\%\) , with - once again - some TUs being down- as well as up- regulated (Fig. S3). This is consistent with experiments showing that the interplay between looping and expression is complex [47] but slight (e.g., knocking down human cohesin leaves expression of \(87\%\) genes unaffected, with global levels changing \(< 30\%\) [48]).  
+
+In simulations thus far, TFs bind strongly to TU beads, and weakly to all others to model binding to open euchromatin [19, 49]. To investigate the effects of heterochromatin - which binds few TFs, carries few histone marks [50], and is gene poor and traditionally viewed as transcriptionally inert - we perform simulations where four of the most- active TUs (905, 907, 930 and 931) are embedded in a non- binding segment (running from bead 901 - 940). This has a dramatic effect (Figs. 6A- C): the activity of the TU beads now embedded in the non- binding
+
+<--- Page Split --->
+
+island are at least halved, some nearby neighbors are downregulated, and more distant ones up- regulated (again due to a reduction in competition; Figs. 6B,C). The regulatory network is also rewired (Figs. 5D,E).  
+
+Just as embedment in a non- binding segment downregulates a TU bead, embedment in a weak- binding (euchromatic) one up- regulates it (Fig. S4). This shows our model effectively captures position effects where the local chromatin context strongly influences activity [51].  
+
+![](images/Figure_5.jpg)
+
+<center>FIG. 5: Looping subtly affects transcriptional activity. Results of two sets of simulations are compared; one set as Fig. 1, in the other the chain contains 8 permanent loops (to represent convergent loops stabilized by cohesin/CTCF). A. Snapshot (beads within loops are magenta; TFs not shown; inset – same structure with only TUs shown). B. Average transcriptional activity for each TU in the looped chain (magenta bars – values for TUs in loops; magenta arcs – loop positions). C. Comparison between activity in wild-type and looped configuration for the 25 TUs with significantly different values in the two sets ( \(p \simeq 0.003\) ; Students t-test). D. Regulatory network inferred from the matrix of Pearson correlations between expression of TUs (as Fig. 3A). E. Change in Pearson correlation between TUs due to loops. </center>  
+
+## Modelling a whole human chromosome  
+
+We next model a whole mid- sized human chromosome (HSA 14, length 107 Mbp; Fig. 7A), in a well- characterized and differentiated diploid cell (HUVEC, human umbilical vein endothelial cell). Now, multivalent and switchable TFs (20% active at any moment) at a non- saturating concentration bind to a string with 35784 beads. As chromosome territories are often ellipsoidal, simulations are performed in an ellipsoid of appropriate size[7, 52]; consequently, chromatin density is now higher than in simulations detailed above, with volume fractions comparable to those in vivo ( \(\sim 14\%\) ).  
+
+![](images/Figure_6.jpg)
+
+<center>FIG. 6: Neighboring heterochromatin affects transcriptional activity. Results from two sets of simulations (at least 800 runs for each condition) are compared; one set as Fig. 1, in the other beads around TU beads 905, 907, 930 and 931 (from bead 901 to 940) are non-binding (to represent embedding the TU beads in heterochromatin). A. Snapshot with heterochromatic beads shown in gray (TFs not shown; inset – the same structure with only TUs). B. Average transcriptional activity for each TU. C. Comparison of average transcriptional activity with respect to wild-type for the 22 TUs with significantly different values in the two sets ( \(p \simeq 0.003\) ; Students t-test). D. Regulatory network inferred from the matrix of Pearson correlations between activities of TUs (as Fig. 3A). E. Change in Pearson correlation between TUs due to heterochromatin. </center>   
+
+Chromatin beads are classified using DNase- hypersensitivity data and ChIP- seq data for H3K27ac. DNase- hypersensitive sites (DHS) are excellent markers to locate promoters and enhancers (and so TF binding sites [19, 53]), whereas H3K27ac modifications strongly correlate with open chromatin [19]. Therefore, if the 3 kbp region corresponding to a chromatin bead has a DHS, then that bead is a TU; if it has H3K27ac, it is a euchromatin bead, and all other beads are non- binding (heterochromatic). We call this the "DHS" model. As properties of different chromatin segments have been catalogued using "hidden- Markov models" (HMMs) applied to many data sets [50], we alternatively classify beads
+
+<--- Page Split --->
+
+according to HMM state; we call this the "HMM model" (Fig. S5). For more details, see Supplementary Note 4.  
+
+Simulations using the DHS model again yield clusters enriched in TUs and TFs (Fig. 7B). As before, aggregating data from many simulations allows determination of transcriptional activities of every bead, which we compare with those of corresponding regions determined experimentally [54] by GRO- seq (global run- on sequencing [55]); activities of all 3 kbp regions are ranked from high to low, binned into quintiles, and compared. In Figure 7C, squares near the diagonal from bottom- left to top- right have high ranks (shown as red and yellow) compared to those off- diagonal (blue and purple) indicating good concordance between simulations and data. A specific sub- set of beads corresponding to super- enhancers (SEs) – which are highly active in vivo [56] – are also highly active in simulations (shown as white dots concentrated at top right). Plots showing the rank of transcriptional activities in simulations and experiments in selected genomic regions are shown in Fig. S6. Simulations yield patterns qualitatively closer to those obtained with GRO- seq than those given by poly(A) \(^+\) RNA- seq, as the latter only include gene transcription.  
+
+Concordance between results from simulations and GRO- seq is confirmed by the Spearman rank correlation \((\sim 0.38\) for all beads; \(p< 10^{- 12}\) ; this measure is used because it is less sensitive to outliers; Fig. 7D). Restricting analysis just to TUs provides a more stringent comparison (as all TUs bind TFs with equal affinity); it still yields a significant correlation ( \(r\simeq 0.32\) , \(p< 10^{- 12}\) ; Fig. 7D). As neighbouring high- affinity regions tend to have roughly similar transcriptional rates in both simulations and data, we also average rates found in active "patches" (contiguous sets of beads which are either all TUs or all labelled as euchromatin), but found this has no significant effect (Fig. 7D). Concordance was confirmed using our HMM model (Fig. 7D, right, and Fig. S5). Adding cohesin- mediated looping to simulations involving the DHS model did not significantly change agreement with experimental data (e.g., for TUs only, \(r\simeq 0.33\) , \(p< 10^{- 12}\) ). Similar agreement with GRO- seq data was obtained from simulations applied to the H1 human embryonic stem- cell line (for TUs using the DHS model, \(r\simeq 0.29\) , \(p< 10^{- 12}\) ), and to the GM12878 cell line (DHS model, \(r\simeq 0.33\) , \(p< 10^{- 12}\) ).  
+
+As in the chromosome fragment simulations (Fig. S1B), the transcriptional activity of a TU in our model anticorrelates with the distance to the nearest TU. In our HSA14 simulations, the presence of heterochromatin slightly reduces the absolute value of the correlation, which however remains highly significant (Spearman correlation \(r\simeq - 0.83\) , \(p< 10^{- 12}\) ). Interestingly, the experimental GRO- seq signal of a DHS also anticorrelates with the distance to the nearest DHS in a significant way, although more weakly than in simulations (Fig. S7; over the whole genome the Spearman correlation is \(r\sim - 0.23\) , \(p< 10^{- 12}\) ).  
+
+## Networks inferred from simulations are qualitatively similar to experimental ones  
+
+Regulatory networks emerging from our whole chromosome simulations are again small- world and highly connected (Fig. S8, and Supplementary Note 5). To facilitate comparison with previous results, we select four segments of HSA14 that have the same length as the one considered in Fig. 4 (i.e., 3 Mbp), and roughly the same density of TUs; all four segments again have highly- connected components (compare Figs. S8 and Fig. 4). However, patterns in real chromosomes and artificial fragments are quite different. In HSA14 networks, there are more positive interactions between sets of adjacent TUs and other sets that are \(>10\) beads distant in sequence space (black lines across the middle of circles in Fig. S8).  
+
+Whole- chromosome networks also have the following statistical properties. First, their node- degree distribution decays exponentially (Fig. S9A) – as found in gene networks [57] but not in transcription factor interaction networks, which are often scale- free [58]. Second, they are modular (as clusters arising due to the bridging- induced attraction are the basic co- regulated building blocks) – again as found in gene [57] and eQTL [59] networks. [Modularity is apparent from the blocks visible in the correlation matrices, such as in Fig. S2.] Third, node degree broadly correlates with transcriptional activity (Spearman correlation 0.59, p- value \(< 10^{- 12}\) ) – as in gene coregulation networks [57].  
+
+## Contact maps found by simulations are qualitatively similar to Hi-C  
+
+We previously showed [16] that simulations involving two different TFs (binding to active and inactive regions, respectively) yield contact maps much like those found with Hi- C [42]. Therefore, we expected the present simulations to reflect Hi- C data poorly as they involve only one TF binding to the minor (i.e., active) fraction of the genome, so contacts made by this structured minority would be obscured by those due to the unstructured majority. Even so, simulations yield contact maps broadly similar to those obtained by Hi- C (Fig. 7E). To measure the agreement, we use a comparison based on contact maps restricted to TUs as anchors – which may be considered as equivalent to interactions obtained by promoter- capture HiC [60]. These yield good concordance (Fig. 7E; Pearson coefficient \(r = 0.82\) ; \(r = 0.47\) when monitoring only long- range contacts between TUs at least 300 kbp away, \(p< 10^{- 6}\) in both cases). The exponent with which contact probability decays with 1D distance is \(\sim - 1.1\) in experiments, and \(\sim - 0.8\) in simulations (fitted for 1D distances between \(\sim 30\) kbp and 1.5 Mbp), both broadly consistent with the \(- 1\) value expected for a fractal globule [61]. The small discrepancy may point to our simulations slightly overestimating the weight of long- range contacts, perhaps
+
+<--- Page Split --->
+
+because we do not include loop extrusion.  
+
+Overall the results obtained in our HSA14 simulations show that a simple model based on 3D chromatin organisation captures much of the complexity in 3D structure and transcription of a whole human chromosome.  
+
+## Modelling chromosome 22 carrying the diGeorge deletion  
+
+Our approach can, in principle, be applied to study any chromosome providing appropriate genomic data are available (e.g., on DNase hypersensitivity and histone acetylation). As a proof of principle, we studied the effect of deleting \(\sim 2.55\) Mbp from HSA22 - an alteration which is associated with the diGeorge syndrome (Fig. 8A) [65]. This syndrome affects \(\sim 1\) in 4000 people, and the variable symptoms include congenital heart problems, frequent infections, developmental delays, and learning problems.  
+
+We predict a multitude of small effects in TU activity, both near and far away from the deletion (see the Manhattan plot in Fig. 8Bi). In particular, most TUs are slightly up- regulated, as fewer TUs compete for the same number of factors, and the TUs which change the most have intermediate transcriptional activities in the wildtype (Fig. S10). The p- values associated with the change in transcriptional activities vary widely, and comparison of the observed distribution with the null hypothesis (indicating that changes in measured transcription are due to random variation) shows the observed is highly enriched in small p- values (Fig. 8Bii), as is generally the case with results from GWAS [5, 6]. The regulatory network is also re- wired (Fig. 8C). Results are consistent with measurements of differential gene expressions in patients, which showed both a large number of up- regulated and downregulated genes [62]. A more quantitative comparison between experiments and simulations would benefit from having GRO- seq data that include non- genic transcription.  
+
+Clearly, this approach opens up a rich new field of study. For instance, while there may be processes which occur in vivo which are not represented in our model, it could still give an indication of the genes most likely to be affected by any chromosome rearrangement.  
+
+## DISCUSSION AND CONCLUSIONS  
+
+We have described a parsimonious 3D stochastic model for transcriptional dynamics based on multivalent binding of factors and polymerases (TFs) to genic and non- genic transcriptional units (TUs) in a chain representing a chromatin fibre. A distinctive feature of our framework is that it is fitting- free, which means the model is truly predictive and can provide a mechanistic understanding of the phenomena we observe. On the other hand, the absence of fitting renders it challenging to obtain a fully quantitative agreement between modelling and experiment.  
+
+![](images/Figure_7.jpg)
+
+<center>FIG. 7: Comparison of transcriptional activities of TUs on HSA14 in HUVECs determined using simulations and GRO-seq. A. Workflow (DHS model). Simulations (244 runs) involve a chain (35784 beads) representing HSA14, and 1700 switchable TFs confined in an ellipsoidal territory. Beads are classified as TUs (red, strong-binding), euchromatic (blue, weak-binding) or heterochromatic (grey, non-binding). Transcriptional activities from simulations are compared with those of GRO-seq data, by measuring the Spearman rank correlation. B. (i) Snapshot (TFs not shown). (ii, iii) TU beads and TFs in this configuration. C. Comparison of transcriptional activities of TUs from simulations and GRO-seq (ranked from \(0 - 100\%\) , then binned in quintiles and showed as a heat map). A scatter plot of unbinned ranks of beads corresponding to SEs are superimposed (white circles). D. Comparison of transcriptional activities from simulations (for both DHS and HMM models) and GRO-seq for all 3 kb regions/beads, only TUs, and only connected patches of binding beads (see text). All correlations are significant \((p < 10^{-6}\) , indicated by grey lines). E (i-ii) Capture-HiC-like contact maps obtained from simulations and experiments [42] showing logarithm of number of contacts between 30 kbp bins which contain TUs. </center>
+
+<--- Page Split --->
+
+## A workflow  
+
+HSA22 wild- type (16,250 beads)  
+
+![](images/Figure_8.jpg)
+  
+
+B effects on transcription at other sites are:  
+
+i ... widely scattered ii ... small  
+
+![PLACEHOLDER_9_1]
+  
+
+![PLACEHOLDER_9_2]
+
+<center>FIG. 8: Modelling effects of the DiGeorge deletion in HSA22. A. Workflow. Simulations (800 simulations/condition) for wild type (17102 beads) and deletion (16250 beads, where wild-type beads 6305 - 7156 are cut, corresponding to a deletion of chr22:18912231 - 21465672 in hg19). [Agreement between predicted transcriptional activity and GRO-seq in HSA22 is similar to that found for HSA1 (here, Spearman correlation is \(r \sim 0.29\) , \(p < 10^{-6}\) ]. B. (i) Manhattan plot showing \(-\log_{10}\) (p-value) as a function of genomic position along HSA22 (position given in Mbp), for changes in TU transcriptional activities between wild-type and deletion. (ii) Quantile-quantile plot showing expected versus observed values for \(-\log_{10}\) (p-value) for the same data in (i). Expected values are computed from the normal distribution (these correspond to the null hypothesis according to which the change in transcriptional activities in the deletion is purely due to random variation). C. Regulatory networks of two 3 Mbp segments in chromosome 22 inferred from the Pearson correlation matrix. Edges show positive correlations \(> 0.12\) ( \(p = 0.0007\) ). Segments chosen have roughly the same number of nodes in 3 Mbp as the short fragment (Fig. 3Aii). </center>  
+
+In our simulations two types of fibres were considered: a 3 Mbp fragment with randomly- positioned TUs, which is useful to exemplify emerging trends, and human chromosomes 14 and 22 where TUs were appropriately positioned according to bioinformatic data. Despite deliberately excluding any explicit underlying network of biochemical regulation, our model nevertheless yields some notable results. These depend on having a low TF copy- number - a feature compatible with observations in vivo [23]. First, since TFs bind with the same affinity to all TUs, one might expect the latter to all be transcribed similarly, but they are not (Fig. 1). This is largely due to inter- TU spacing; TUs lying close together in 1D sequence space tend to be the most active (Fig. 1C) with positively- correlated dynamics reminiscent of transcriptional bursting (Fig. 2B). This is because they often cluster into structures which are analogous to the phase- separated transcription hubs/factories seen experimentally [7, 10], or to contact domains formed by accessible DNA sites found by high- resolution mapping of chromatin interactions by microC [30]. Second, switching off binding at any TU significantly affects the activity of many others, both near and far away in sequence space (Fig. 4). Third, introducing stable loops has subtle effects (Fig. 5), consistent with the modest changes in expression seen experimentally in cohesin knock- outs and degrons [48]. Fourth, transcriptional activity of a TU is strongly affected by the local environment in ways that are reminiscent of the silencing of a gene by incorporation into heterochromatin [51] (Fig. 6), or activation by embedment in euchromatin (Fig. S4). Fifth, the stochasticity seen in individual simulations reflects that detected by single- cell transcriptomics and single- cell Hi- C. Nevertheless, this variability does not prevent emergence of robust phenotypes in a cell population. Sixth, our simple fitting- free model predicts patterns of transcriptional activity in human chromosomes that promisingly and significantly correlate with experimental GRO- seq data (Fig. 7). This suggests that chromatin structure significantly constrains transcriptional activity. We hypothesise that additional downstream biochemical regulation, not included in our model, may provide a tool to adjust this underlying "structural" pattern of activity in a way which may be required for appropriate biological function.   
+
+Finally, our results enable us to reconcile two conflicting sets of data, namely that regulatory networks are both complex (as GWAS shows that thousands of loci around the genome control complex phenotypes [5, 6]) and simple (as over- expressing just four Yamanaka factors switches cell fate [4]). Thus, our simulations reveal complex small- world networks of mutual up- and down- regulation (Figs. 3 and S8), consistent with GWAS results. However, increasing TF copy- number dramatically simplifies network structure (Fig. 3). We suggest such a simplification occurs when a fibroblast is reprogrammed into a pluripotent stem cell by over- expressing the Yamanaka factors; the high factor concentration simplifies the network so that the factors can combine to switch the phenotype (Fig. S11).  
+
+Taken together, these results suggest the activity - or inactivity - of every genomic region affects that of every other region to some extent. We describe our framework as "pan- genomic" (Fig. S8). This is reminiscent of the omni- . genic model [5, 6] in the sense that many loci are involved, all having small effects. However, it differs as it provides an underlying mechanism for pangenomic effects, by posit
+
+<--- Page Split --->
+
+ing a direct and immediate effect of structure on regulation at the transcriptional level, which contrasts with the nontrivial post- transcriptional pathways envisioned by the on- nigenic model. Additionally, our pangenomic model yields a natural framework to qualitatively understand mutually exclusive gene expression, when switching on one gene in a family turns off all others (as in developing olfactory neurons [63]). The current model to explain this phenomenon postulates a coupling between cis- acting up- regulation and trans- acting down- regulation. The pangenomic networks we find provide exactly this type of regulatory interactions (Fig. 3). On the other hand, it is challenging within our current model to account for local negative feedback mechanisms leading to noise reduction or oscillations [11], as these are more likely to arise biochemically (an example is the p53- Mdm2 system which achieves stabilisation of the cellular concentration of p53 via a negative feedback loop [64]).  
+
+In conclusion, we have developed a framework that can be applied to predict the transcriptional activity of any genomic fragment in health or disease (Figs. 7, 8) providing appropriate experimental data is available. Predictive power can be enhanced by incorporating additional TFs, and more suitable datasets of histone marks. Other features which can improve correlations between experiments and simulations are a more accurate modelling of cohesin loop formation by loop extrusion, and of the heteromorphic nature of chromatin [19]. We hope to report on work incorporating the latter two features in the future.  
+
+We thank the European Research Council (ERC CoG 648050 THREEDCELLPHYSICS) for support.  
+
+## Code availability  
+
+The code used for the simulation is LAMMPS, which is publicly available at https://lammps.sandia.gov/. Custom codes written to analyse data are available from the corresponding author upon request.  
+
+## Data availability  
+
+The datasets generated during and/or analysed during the current study are available from the corresponding author upon request.  
+
+## Author contributions  
+
+C. A. B., N. G., D. Mi., A. P., P. R. C., M. C. F. P., D. Ma. designed research; C. A. B., M. C. F. P., D. Ma. performed research; C. A. B., N. G., D. Mi., A. P., M. C. F. P., P. R. C., D. Ma. analysed the data and wrote the manuscript.  
+
+[3] A. Dall'Agnese, L. Caputo, C. Nicoletti, J. di Iulio, A. Schmitt, S. Gatto, Y. Diao, Z. Ye, M. Forcato, R. Perera, et al., Mol. Cell 76, 453 (2019). [4] K. Takahashi, K. Tanabe, M. Ohnuki, M. Narita, T. Ichisaka, K. Tomoda, and S. Yamanaka, Cell 131, 861 (2007). [5] E. A. Boyle, Y. I. Li, and J. K. Pritchard, Cell 169, 1177 (2017). [6] X. Liu, Y. I. Li, and J. K. Pritchard, Cell 177, 1022 (2019). [7] P. R. Cook and D. Marenduzzo, Nucleic Acids Res. 46, 9895 (2018). [8] R. Andersson, C. Gebhard, I. Miguel-Escalada, I. Hoof, J. Bornholdt, M. Boyd, Y. Chen, X. Zhao, C. Schmidl, T. Suzuki, et al., Nature 507, 455 (2014). [9] B. M. Javierre, O. S. Burren, S. P. Wilder, R. Kreuzhuber, S. M. Hill, S. Sewitz, J. Cairns, S. W. Wingett, C. Varnai, M. J. Thiecke, et al., Cell 167, 1369 (2016). [10] P. Cramer, Nature 573, 45 (2019). [11] K. Sneppen, S. Krishna, and S. Semsey, Annu. Rev. Biophys. 39, 43 (2010). [12] P. Smolen, D. A. Baxter, and J. H. Byrne, Bull. Math. Biol. 62, 247 (2000). [13] A. Pombo and N. Dillon, Nat. Rev. Mol. Cell Biol. 16, 245 (2015). [14] M. Spielmann, D. G. Lupiáñez, and S. Mundlos, Nat. Rev. Genet. 19, 453 (2018). [15] M. Barbieri, M. Chotalia, J. Fraser, L.- M. Lavitas, J. Dostie, A. Pombo, and M. Nicodemi, Proc. Natl. Acad. Sci. USA 109, 16173 (2012). [16] C. A. Brackley, J. Johnson, S. Kelly, P. R. Cook, and D. Marenduzzo, Nucleic Acids Res. 44, 3503 (2016). [17] N. Gilbert and D. Marenduzzo, Chromosome Res. 25, 1 (2017). [18] M. C. F. Pereira, C. A. Brackley, D. Michieletto, C. Annunziabella, S. Bianco, A. M. Chiariello, M. Nicodemi, and D. Marenduzzo, bioRxiv, 305359 (2018). [19] A. Buckle, C. A. Brackley, S. Boyle, D. Marenduzzo, and N. Gilbert, Mol. Cell 72, 786 (2018). [20] E. H. Finn, G. Pegoraro, H. B. Brandao, A.- L. Valton, M. E. Oomen, J. Dekker, L. Mirny, and T. Misteli, Cell 176, 1502 (2019). [21] C. A. Brackley, B. Liebchen, D. Michieletto, F. L. Mouvet, P. R. Cook, and D. Marenduzzo, Biophys. J. 28, 1085 (2017). [22] B. Steurer, R. C. Janssens, B. Geverts, M. E. Geijer, F. Wienholz, A. F. Theil, J. Chang, S. Dealy, J. Pothof, W. A. van Cappellen, et al., Proc. Natl. Acad. Sci. USA 115, E4368 (2018). [23] R. C. Brewster, F. M. Weinert, H. G. Garcia, D. Song, M. Rydenfelt, and R. Phillips, Cell 156, 1312 (2014). [24] C. A. Brackley, S. Taylor, A. Papantonis, P. R. Cook, and D. Marenduzzo, Proc. Natl. Acad. Sci. USA 110, E3605 (2013). [25] C. Brackley, J. Phys.: Condens. Matter 32, 314002 (2020). [26] S. Kilic, A. L. Bachmann, L. C. Bryan, and B. Fierz, Nat. Comm. 6, 7313 (2015). [27] P. R. Cook, Science 284, 1790 (1999). [28] A. Papantonis, T. Kohro, S. Baboo, J. D. Larkin, B. Deng, P. Short, S. Tsutsumi, S. Taylor, Y. Kanki, M. Kobayashi, et al., EMBO J. 31, 4404 (2012). [29] K. Shrinivas, B. R. Sabari, E. L. Coffey, I. A. Klein, A. Boija, A. V. Zamudio, J. Schuijers, N. M. Hannett, P. A. Sharp, R. A. Young, et al., Mol. Cell 75, 549 (2019). [30] T.- H. S. Hsieh, C. Cattoglio, E. Slobodyanuk, A. S.
+
+<--- Page Split --->
+
+Hansen, O. J. Rando, R. Tjian, and X. Darzacq, Mol. Cell 78, 539 (2020).[31] J.- K. Ryu, C. Bouchoux, H. W. Liu, E. Kim, M. Minamino, R. de Groot, A. J. Katan, A. Bonato, D. Marenduzzo, D. Michieletto, F. Uhlmann, and C. Dekker, Sci. Adv. 7, eabe5905 (2021).[32] A. Pombo, D. A. Jackson, M. Hollinshead, Z. Wang, R. G. Roeder, and P. R. Cook, EMBO J. 18, 2241 (1999).[33] I. Faro-Trindade and P. R. Cook, Mol. Biol. Cell 17, 2910 (2006).[34] R. A. Beagrie, A. Scialdone, M. Schueler, D. C. Kraemer, M. Chotalia, S. Q. Xie, M. Barbieri, I. de Santiago, L.- M. Lavitas, M. R. Branco, et al., Nature 543, 519 (2017).[35] T. Nagano, Y. Lublin, T. J. Stevens, S. Schoenfelder, E. Yaffe, W. Dean, E. D. Laue, A. Tanay, and P. Fraser, Nature 502, 59 (2013).[36] I. C. Macaulay and T. Voet, PLoS Genet. 10, e1004126 (2014).[37] F. Muerdter and A. Stark, Curr. Biol. 26, R895 (2016).[38] T. Fukaya, B. Lim, and M. Levine, Cell 166, 358 (2016).[39] C. R. Bartman, S. C. Hsu, C. C.- S. Hsiung, A. Raj, and G. A. Blobel, Mol. Cell 62, 237 (2016).[40] D. M. Suter, N. Molina, D. Gattield, K. Schneider, U. Schibler, and F. Naef, Science 332, 472 (2011).[41] M. D. Humphries and K. Gurney, PLoS ONE 3, e0002051 (2008).[42] S. S. Rao, M. H. Huntley, N. C. Durand, E. K. Stamenova, I. D. Bochkov, J. T. Robinson, A. L. Sanborn, I. Machol, A. D. Omer, E. S. Lander, and E. LiebermanAiden, Cell 159, 1665 (2014).[43] G. Fudenberg, M. Imakaev, C. Lu, A. Goloborodko, N. Abdennur, and L. A. Mirny, Cell Rep. 15, 2038 (2016).[44] C. A. Brackley, J. Johnson, D. Michieletto, A. N. Morozov, M. Nicodemi, P. R. Cook, and D. Marenduzzo, Phys. Rev. Lett. 119, 138101 (2017).[45] M. Oti, J. Falck, M. A. Huynen, and H. Zhou, BMC Genomics 17, 252 (2016).[46] D. Sasca, H. Yun, G. Giotopoulos, J. Szybinski, T. Evan, N. K. Wilson, M. Gerstung, P. Gallipoli, A. R. Green, R. Hills, et al., Blood 134, 2195 (2019).[47] M. I. Robson, A. R. Ringel, and S. Mundlos, Mol. Cell 74, 1110 (2019).[48] S. S. Rao, S.- C. Huang, B. G. St Hilaire, J. M. Engreitz,  
+
+E. M. Perez, K.- R. Kieffer-Kwon, A. L. Sanborn, S. E. Johnstone, G. D. Bascom, I. D. Bochkov, et al., Cell 171, 305 (2017).[49] N. Gilbert, S. Boyle, H. Fiegler, K. Woodfine, N. P. Carter, and W. A. Bickmore, Cell 118, 555 (2004).[50] J. Ernst, P. Kheradpour, T. S. Mikkelsen, N. Shoresh, L. D. Ward, C. B. Epstein, X. Zhang, L. Wang, R. Issner, M. Coyne, et al., Nature 473, 43 (2011).[51] R. T. Timms, I. A. Tchasovnikarova, and P. J. Lehner, BioEssays 38, 333 (2016).[52] Y. Wang, M. Nagarajan, C. Uhler, and G. Shivashankar, Mol. Biol. Cell 28, 1997 (2017).[53] E. P. Consortium, Nature 489, 57 (2012).[54] H. Niskanen, I. Tuszynskax, R. Zaborowski, M. Heinaniemi, S. Yla-Herttuala, B. Wilczynski, and M. U. Kaikkonen, Nucleic Acids Res. 46, 1724 (2017).[55] A. Jordán-Pla, M. E. Pérez-Martínez, and J. E. Pérez-Ortín, Methods 159, 177 (2019).[56] A. Khan and X. Zhang, Nucleic Acids Res. 44, D164 (2015).[57] V. Belcastro, V. Siciliano, F. Gregoretti, P. Mithbaokar, G. Dharmalingam, S. Berlingieri, F. Iorio, G. Oliva, R. Polishchuk, N. Brunetti-Pierri, et al., Nucleic Acids Res. 39, 8677 (2011).[58] W. Z. Ouma, K. Pogacar, and E. Grotewold, PLoS Comput. Biol. 14, e1006098 (2018).[59] M. Fagny, J. N. Paulson, M. L. Kuijjer, A. R. Sonawane, C.- Y. Chen, C. M. Lopes-Ramos, K. Glass, J. Quackenbush, and J. Platig, Proc. Natl. Acad. Sci. USA 114, E7841 (2017).[60] B. Mifsud, F. Tavares-Cadete, A. N. Young, R. Sugar, S. Schoenfelder, L. Ferreira, S. W. Wingett, S. Andrews, W. Grey, P. A. Ewels, et al., Nat. Genet. 47, 598 (2015).[61] L. A. Mirny, Chromosome Res. 19, 37 (2011).[62] M. Jalbrzikowski, M. T. Lazaro, F. Gao, A. Huang, C. Chow, D. H. Geschwind, G. Coppola, and C. E. Bearden, PLoS ONE 10, e0132542 (2015).[63] A. K. Alsing and K. Sneppen, Nucleic Acids Res. 41, 4755 (2013).[64] S. L. Harris and A. J. Levine, Oncogene 24, 2899 (2005).[65] https://dosage.clinicalgenome.org/clingen_region.cgi?id=ISCA-37446
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- Sl.pangenomic.revised.pdf
+
+<--- Page Split --->

preprint/preprint__09855f388022d98f6d42aad5171260d00fc3df0376deedaa6931278a022350c0/preprint__09855f388022d98f6d42aad5171260d00fc3df0376deedaa6931278a022350c0_det.mmd

@@ -0,0 +1,356 @@
+<|ref|>title<|/ref|><|det|>[[44, 107, 912, 177]]<|/det|>
+# Complex small-world regulatory networks emerge from the 3D organisation of the human genome  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 260, 238]]<|/det|>
+Chris Brackley  University of Edinburgh  
+
+<|ref|>text<|/ref|><|det|>[[44, 243, 615, 285]]<|/det|>
+Nick Gilbert  University of Edinburgh  https://orcid.org/0000- 0003- 0505- 6081  
+
+<|ref|>text<|/ref|><|det|>[[44, 290, 260, 332]]<|/det|>
+Davide Michieletto  University of Edinburgh  
+
+<|ref|>text<|/ref|><|det|>[[44, 337, 259, 378]]<|/det|>
+Argyris Papantonis  University of Göttingen  
+
+<|ref|>text<|/ref|><|det|>[[44, 383, 260, 425]]<|/det|>
+Maria Pereira  University of Edinburgh  
+
+<|ref|>text<|/ref|><|det|>[[44, 430, 616, 517]]<|/det|>
+Peter Cook  Oxford University  https://orcid.org/0000- 0002- 6639- 188X  Davide Marenduzzo (dmarendu@ph.ed.ac.uk)  University of Edinburgh  https://orcid.org/0000- 0003- 3974- 4915  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 558, 102, 575]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 594, 640, 614]]<|/det|>
+Keywords: regulatory networks, transcription factors, human genome  
+
+<|ref|>text<|/ref|><|det|>[[44, 633, 300, 652]]<|/det|>
+Posted Date: June 12th, 2021  
+
+<|ref|>text<|/ref|><|det|>[[44, 670, 463, 690]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 566854/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 707, 910, 750]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License.  Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 787, 925, 830]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on October 1st, 2021. See the published version at https://doi.org/10.1038/s41467- 021- 25875-y.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[120, 59, 936, 93]]<|/det|>
+# Complex small-world regulatory networks emerge from the 3D organisation of the human genome  
+
+<|ref|>text<|/ref|><|det|>[[110, 105, 941, 210]]<|/det|>
+C. 
+A. Brackley<sup>1</sup>, 
+N. Gilbert<sup>2</sup>, 
+D. Michieletto<sup>1,2</sup>, 
+A. Papantonis<sup>3</sup>, 
+M. 
+C. 
+F. Pereira<sup>1</sup>, 
+P. 
+R. Cook<sup>4</sup>, 
+D. Marenduzzo<sup>1</sup>  <sup>1</sup>SUPA, School of Physics and Astronomy, University of Edinburgh, Peter Guthrie Tait Road, Edinburgh EH9 3FD, UK  <sup>2</sup>MRC Human Genetics Unit, MRC Institute of Genetics & Molecular Medicine, University of Edinburgh, Western General Hospital, Edinburgh EH4 2XU, UK  <sup>3</sup>Institute of Pathology, University Medical Center, Georg-August University of Gottingen, 37075 Gottingen, Germany  <sup>4</sup>Sir William Dunn School of Pathology, University of Oxford, South Parks Road, Oxford OX1 3RE, UK  
+
+<|ref|>text<|/ref|><|det|>[[191, 219, 864, 394]]<|/det|>
+The discovery that overexpressing one or a few critical transcription factors can switch cell state suggests that gene regulatory networks are relatively simple. In contrast, genome- wide association studies (GWAS) point to complex phenotypes being determined by hundreds of loci that rarely encode transcription factors and which individually have small effects. Here, we use computer simulations and a simple fitting- free polymer model of chromosomes to show that spatial correlations arising from 3D genome organisation naturally lead to stochastic and bursty transcription as well as complex small- world regulatory networks (where the transcriptional activity of each genomic region subtly affects almost all others). These effects require factors to be present at sub- saturating levels; increasing levels dramatically simplifies networks as more transcription units are pressed into use. Consequently, results from GWAS can be reconciled with those involving overexpression. We apply this pan- genomic model to predict patterns of transcriptional activity in whole human chromosomes, and, as an example, the effects of the deletion causing the diGeorge syndrome.  
+
+<|ref|>sub_title<|/ref|><|det|>[[227, 416, 373, 430]]<|/det|>
+## INTRODUCTION  
+
+<|ref|>text<|/ref|><|det|>[[88, 448, 513, 604]]<|/det|>
+Transcription - the copying of DNA into RNA - is tightly regulated. Early insights into regulatory mechanisms came from work on binary on/off genetic switches controlled by one (or just a few) transcription factors such as the lambda and lac repressor in Escherichia coli [1]. Similar regulatory mechanisms are present in eukaryotes, albeit with additional complexity. For instance, a fibroblast cell can be reprogrammed into a muscle cell by a single master regulator (MYOD) [2, 3], or into pluripotent stem cells by four Yamanaka factors (Oct4, Sox2, c- Myc, Klf4) [4].  
+
+<|ref|>text<|/ref|><|det|>[[88, 606, 513, 859]]<|/det|>
+Genome- wide association studies (GWAS) lead to quite a different view: gene regulation is widely distributed and involves interactions between hundreds (perhaps thousands) of loci scattered around the genome [5, 6]. GWAS allow quantitative trait loci (QTLs) affecting any measurable genetic trait to be ranked in an unbiased way. With complex traits like human height, and diseases such as schizophrenia and type II diabetes, the top ten QTLs in the rank order combine to yield only modest effects, while the top one- hundred still account for less than half of the total genetic effect. Hundred more QTLs are expected to be identified as sample sizes and data resolution improve [5- 7]. Expression QTLs (eQTLs) are QTLs affecting transcription of other DNA regions. Perhaps surprisingly, these are rarely found in genes encoding transcription factors or other proteins; instead, they usually involve single- nucleotide changes in non- coding elements that bind transcription factors such as active enhancers and promot  
+
+<|ref|>text<|/ref|><|det|>[[543, 415, 616, 430]]<|/det|>
+ers [8- 10].  
+
+<|ref|>text<|/ref|><|det|>[[543, 431, 966, 590]]<|/det|>
+Results from GWAS lead to the view that most generegulatory networks are incredibly complex, with the activity of a given gene being affected by a panoply of eQTLs, each having a tiny effect. This is captured by the "omnic gene" model, which is based on a set of gene- interaction equations [5, 6] such that the activity of almost any gene affects that of almost every other one. This model provides a useful and appealing framework to view GWAS results. However, it is difficult to compare its outputs with experimental data because it contains many parameters that are currently unknown and require fitting to training datasets.  
+
+<|ref|>text<|/ref|><|det|>[[543, 591, 966, 830]]<|/det|>
+In general, existing models for gene regulation traditionally assume post- transcriptional and biochemically mediated interactions between different genes [11, 12], and disregard the role of 3D chromatin structure. Here we propose an alternative but complementary framework that links transcriptional regulation directly to 3D genome structure, deliberately neglecting downstream biochemical regulation to enable unambiguous interpretation of our results. This framework is motivated by experiments showing that chromatin folding can lead to contacts between enhancers and promoters affecting transcription, and that 3D structure changes in disease [13, 14]. Additionally, because our modelling is essentially fitting- free, its output can be directly compared to experiments. When the agreement is good, our model is validated; when poor, it points to some missing ingredient (such as biochemical feedback) that could be included in future models.  
+
+<|ref|>text<|/ref|><|det|>[[543, 832, 965, 859]]<|/det|>
+Here, we use stochastic computer simulations of a polymer model for chromosome organization, in which a chain
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[88, 60, 513, 358]]<|/det|>
+of beads represents a chromatin fibre, and a set of spheres complexes of transcription factors and RNA polymerases - which we will call "TFs" for short. Some chromatin beads are identified as transcription units (TUs), and we call them TU beads. They contain binding sites for TFs, and can be sites of transcriptional initiation (we do not discriminate between genic and non- genic promoters). As a simple starting point we only consider one type of TF that binds specifically and multivalently to TU beads, and non- specifically (i.e., with weak affinity) to every other bead. We perform 3D Brownian dynamics simulations that evolve the diffusive dynamics of the chain and associated factors. We previously showed that similar polymer models yield structures resembling those seen using chromosome- conformation- capture (3C) [15- 19] and microscopy [20]. Here, we link 3D structure to expression and transcriptional dynamics by measuring how often a TU bead is transcribed - which we do by computing the fraction of time it binds a TF. To establish the methodology, we model a 3 Mbp chromatin fragment, before going on to simulate whole human chromosomes.  
+
+<|ref|>text<|/ref|><|det|>[[88, 360, 513, 614]]<|/det|>
+Our simulations capture many features of eukaryotic regulation. For example, transcription is stochastic and bursty (in agreement with single- cell transcriptomics data), and the predicted pattern of transcriptional activity in human chromosomes correlates significantly with that observed experimentally. We also find that small- world (percolating) networks that encapsulate much of the rich complexity observed in GWAS emerge through spatial effects alone. In other words, the activity of most (probably all) TUs in our model is affected by the activity of most (probably all) other segments in the genome. We find such pan- genomic regulation critically requires non- saturating concentrations of TFs - as normally found in vivo - and that increasing concentrations dramatically simplifies the networks. This enables us to reconcile the GWAS- based view that regulatory networks are complicated with the observation that overexpressing one or a few TFs can decisively alter cell state.  
+
+<|ref|>sub_title<|/ref|><|det|>[[260, 643, 340, 656]]<|/det|>
+## RESULTS  
+
+<|ref|>text<|/ref|><|det|>[[88, 674, 513, 858], [542, 61, 966, 162]]<|/det|>
+We first consider a simple system where a 3 Mbp chromatin fragment is represented by a chain of 1000 beads (each \(30\mathrm{nm}\) in diameter, and corresponding to \(3\mathrm{kbp}\) ). We select at random \(N = 39\) beads and identify them as TUs (Fig. 1A; full details in SI, Supplementary Notes 1 and 2). The linear density of TU in the fragment is similar to that in human chromosome 22. Additionally, \(n\) spheres (also 30 nm in diameter) represent TFs (recall these are complexes of transcription factors and RNA polymerase II). TFs bind reversibly to TUs via a strong attractive interaction, and to all other beads weakly and non- specifically. An important feature is that TFs switch between active (binding) and inactive (non- binding) state at rate \(\alpha\) . Many factors switch like this in vivo (e.g., due to phosphorylation and de- phosphorylation), and switching is required to account for the rapid exchange of factors and polymerases between bound and free states seen in live- cell photobleaching experiments [21]. As \(\sim 7\) out of 8 polymerases attempting to initiate at promoters dissociate with a half- life of \(\sim 2.4\) s [22], our complexes generally behave like those in vivo.   
+
+<|ref|>text<|/ref|><|det|>[[542, 163, 966, 246]]<|/det|>
+While our results refer to a single patterning of TUs along the fibre, they are representative of any arbitrary random positioning of TUs: in other words the qualitative trends we present below are robust and do not depend on the particular choice of the 1D pattern of TU along the fibre in any way.  
+
+<|ref|>text<|/ref|><|det|>[[542, 247, 966, 346]]<|/det|>
+We say a TU bead is transcribed whenever a TF lies close to it (SI, Supplementary Note 1), and the transcriptional activity of a TU is then the fraction of time it is transcribed during a simulation. To reflect the situation in mammalian cells (Supplementary Note 4 and [23]), we typically assume there are fewer TFs than TU beads (i.e., \(n = 10\) TFs in the active binding state at any time, compared to 39 TUs).  
+
+<|ref|>text<|/ref|><|det|>[[542, 348, 966, 757]]<|/det|>
+By interrogating TF- chromatin interactions at regular time intervals over hundreds of simulations we build up a population picture of transcription. A typical configuration of the 3 Mbp fragment is shown in Figure 1B. Strikingly, bound TFs spontaneously cluster, despite there being no attractive interactions between TUs or between TFs. Such clustering is driven by the "bridging- induced attraction" [16, 24, 25] that arises due to a positive feedback: when a TF forms a molecular bridge between two chromatin regions and forms a loop, the local chromatin concentration increases, making further TF binding more likely. Clusters then grow until limited by entropic costs of crowding (Fig. S1A). Most of the non- trivial phenomena described below result from such clustering. Clustering requires TF multivalency, as monovalent factors do not cluster [24]. However, the assumption of multivalency, which is common in the polymer physics literature [15], is well- founded. Several TFs are known to be bivalent or multivalent [26], and, more importantly, our spheres represent complexes of TFs and polymerases, so they will behave as multivalent binders even when the individual TFs in the complex are monovalent. Although clustering does not require any interactions between TFs, adding a weak attraction between them, as might arise for instance due to macromolecular crowding or electrostatic interactions between intrinsically disordered regions, should not qualitatively change any of the results discussed here (at least as long as TFs still microphase separate into clusters rather than undergoing macroscopic phase separation).  
+
+<|ref|>text<|/ref|><|det|>[[542, 760, 966, 858]]<|/det|>
+The clusters we observe, and which emerge through the bridging- induced attraction, are qualitatively similar to those seen in vivo, which are variously described as transcriptional compartments, hubs, super- enhancer (SE) clusters, phase- separated droplets/condensates, and factories [7, 10, 27- 29]. They are also similar to the contact domains seen in microC [30], which are formed by acces
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[92, 63, 525, 415]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[88, 429, 515, 666]]<|/det|>
+<center>FIG. 1: Patterns of transcriptional activity. A. Schematic of the model. Twenty TFs (pink) that switch between on/off states at rate \(\alpha = 10^{-5} \tau_{\mathrm{B}}^{-1}\) (with \(\tau_{\mathrm{B}}\) the Brownian time, see SI, Supplementary Note 1) or \(0.001 \mathrm{~s}^{-1}\) bind specifically to 39 TUs (red beads) randomly positioned along the chain, and non-specifically to other beads (blue). A TU is considered transcriptionally active if associated with a TF. B. Example conformation (TFs not shown). Some beads cluster and form loops; one TU not in a cluster (and not transcribed) is green, and another that is in a cluster (and transcribed) is yellow. Inset: zoom of boxed region. C. Transcriptional activity for each TU bead averaged over 1000 simulations (each lasting \(10^{5} \tau_{\mathrm{B}}\) ). TUs are grouped according to activity, with red, green and blue bars showing high \((>70\%)\) , medium \((20 - 70\%)\) and low \((< 20\%)\) activity, respectively. This gives a population-level measure of activity. D. Variation of activity across simulations (reflecting cell-to-cell variation) for 3 representative TUs with high (red), medium (green), or low (blue) average activity (defined as in C). </center>  
+
+<|ref|>text<|/ref|><|det|>[[88, 697, 513, 754]]<|/det|>
+sible DNA sites clustering together in 3D space. Clustering arising through the bridging- induced attraction has recently been found in vitro for systems of DNA and cohesin (which binds multivalently to DNA) [31].  
+
+<|ref|>sub_title<|/ref|><|det|>[[102, 787, 499, 813]]<|/det|>
+## Transcriptional activity varies along the chromatin fibre and is highly stochastic  
+
+<|ref|>text<|/ref|><|det|>[[88, 830, 513, 858]]<|/det|>
+As TFs have the same affinity for all TUs, one might expect each TU to be bound with equal likelihood; however,  
+
+<|ref|>text<|/ref|><|det|>[[541, 60, 966, 247]]<|/det|>
+transcriptional activity (the fraction of time a TU is transcribed) varies from \(\sim 10 - 90\%\) (Fig. 1C). What causes this variation? As TF copy number is limiting, and as bound TFs cluster, most transcription occurs in clusters – as is the case in vivo [7, 32–34]. Since TUs are positioned irregularly along the fragment, some have closer neighbours in 1D sequence space than others, and these are inevitably the ones most likely to cluster and be transcribed. Instead, those far from their neighbours are less likely to cluster and are less active. Accordingly, the transcriptional activity of a TU anticorrelates with distance to the nearest TU along the fibre (Fig. S1B; the Spearman correlation is \(r \simeq - 0.94\) , p- value \(p < 10^{- 12}\) ).  
+
+<|ref|>text<|/ref|><|det|>[[541, 247, 966, 346]]<|/det|>
+Whilst Figure 1C pertains to population averages of 1000 simulations, it is informative to consider each simulation independently (as in single- cell transcriptomics). Such analysis shows that transcriptional activity is stochastic, varying substantially from simulation to simulation: a TU active in some simulations may be silent in others (Fig. 1D).  
+
+<|ref|>sub_title<|/ref|><|det|>[[660, 374, 850, 388]]<|/det|>
+## Transcriptional bursting  
+
+<|ref|>text<|/ref|><|det|>[[541, 402, 966, 532]]<|/det|>
+During a simulation, chromatin conformation can change dramatically (Fig. 2A). Such changes often yield transcriptional "bursts" – periods of continued activity followed by silent periods (Fig. 2B) – as TUs with intermediate levels of activity repeatedly join a cluster to give a burst and then dissociate. Notably, TUs lying close to each other in sequence space often start and stop bursts coordinately due to the intrinsic positive feedback in the system (Fig. S1A).  
+
+<|ref|>text<|/ref|><|det|>[[541, 533, 966, 604]]<|/det|>
+These results are consistent with experimental observations: single cell Hi- C [35] and transcriptomics [36] show that the structure and function of each individual cell is unique, and bursting is well documented [37–40] with nearby promoters often firing together [38].  
+
+<|ref|>sub_title<|/ref|><|det|>[[561, 630, 945, 656]]<|/det|>
+## Local chromatin architecture creates small-world percolating transcription networks  
+
+<|ref|>text<|/ref|><|det|>[[541, 675, 966, 816]]<|/det|>
+To investigate correlations between transcriptional activities of different TUs, we compute the Pearson correlation matrix between the activity of all possible TU pairs, and identify an emergent regulatory network in which TUs form nodes (Fig. 3A and Fig. S2). Specifically, we draw an edge between two TUs whenever there is a statistically significant positive or negative correlation between their transcriptional dynamics (Fig. 3A). This network arises only due to spatial interactions, as we assume no underlying biochemical regulation.  
+
+<|ref|>text<|/ref|><|det|>[[541, 817, 966, 858]]<|/det|>
+The network shows a striking property. With \(n = 10\) active TFs, most nodes are connected (Fig. 3Aii), and the fraction of TUs participating in the largest connected com
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[90, 61, 530, 312]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[88, 325, 515, 440]]<|/det|>
+<center>FIG. 2: Transcriptional bursting. A. Snapshots showing a 100-bead section of the simulated chain taken at different times. Initially, none of the 5 TUs (red) are in clusters and are inactive; later, 4 TUs join a cluster and are close to TFs – and so are transcribed. B. Kymograph where each row shows the changing transcription state of one TU during a simulation; pixels are colored red if the bead is associated with a TF and so transcribed, or black otherwise. White rectangle: example burst. </center>  
+
+<|ref|>text<|/ref|><|det|>[[88, 468, 515, 640]]<|/det|>
+ponent is close to 1 (Fig. 3B). Such a network is said to be "percolating", which means that any two nodes are connected by a path along edges. Our percolating networks are also "small- world", which means that most nodes can be reached from every other node by a small number of steps [41] – we provide quantitative measurements of the small world- ness of our networks in the SI (Supplementary Note 5). The small- world phenomenology is consistent with the multitude of small- effect eQTLs detected by GWAS [5, 6]. Notably, the regulation we observe act at the transcriptional level, and not post- transcriptionally as envisaged by the omnigenic model [5, 6].  
+
+<|ref|>text<|/ref|><|det|>[[88, 643, 515, 772]]<|/det|>
+How might our simple model give rise to complex regulatory networks? By analysing simulation trajectories, we noted that TUs lying near each other in 1D sequence space often joined the same cluster in 3D. As a result, the activity of these clustered beads is highly positively correlated. At the same time, cluster formation sequesters TFs and so reduces the likelihood that another cluster forms elsewhere. As a result, most long- range correlations are negative (Fig. 3A).  
+
+<|ref|>text<|/ref|><|det|>[[88, 775, 515, 858]]<|/det|>
+Crucially, these network properties depend on there being a low TF copy- number (as in vivo [23]) so TU beads do not become saturated. We therefore reasoned that increasing copy number should suppress correlations as more rarely- transcribed TUs are pressed into use. Indeed, increasing \(n\) reduces long- range negative correlations  
+
+<|ref|>image<|/ref|><|det|>[[545, 61, 969, 310]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[540, 335, 967, 510]]<|/det|>
+<center>FIG. 3: Regulatory networks formed by TU beads are percolating at low TF concentrations. Simulations (as Fig. 1, with \(\geq 800\) simulations/condition) with different average numbers of active TFs \((n)\) and switching rate \((\alpha)\) . Networks were constructed by calculating the Pearson correlation between the transcription time-series for all pairs of TUs; nodes represent each of the 39 TUs and edges are placed between nodes where there is a significant correlation \((>0.15\) in absolute value, corresponding to \(p < 10^{-6}\) ). A. Effect of TF concentration and switching. 39 nodes are shown around the perimeter, and thick black and grey lines denote positive and negative correlations between transcriptional activities of bead pairs. B. Effect of \(n\) on the fraction of nodes in the largest connected component. </center>  
+
+<|ref|>text<|/ref|><|det|>[[541, 538, 966, 639]]<|/det|>
+(Fig. 3Aiii,iv), and the fraction of nodes in the largest- connected component falls (Fig. 3B). Another way to think about this result is: if resources are plentiful, there is no need for sharing or competition, and all TUs can bind a TF independently of each other. If TFs do not switch and are permanently in the binding state (and \(n = 10\) ), the network becomes even more highly connected (Fig. 3Ai).  
+
+<|ref|>sub_title<|/ref|><|det|>[[541, 672, 964, 699]]<|/det|>
+## Modelling effect of mutations and SNPs in regulatory elements  
+
+<|ref|>text<|/ref|><|det|>[[541, 717, 966, 858]]<|/det|>
+GWAS reveals that single- nucleotide polymorphisms (SNPs) in regulatory elements and TUs can lead to many small changes in transcriptional activity across the genome. To model this, we abrogate TF binding to one TU in the chain. Bead 930 is chosen first because it is usually highly active (Fig. 1C). This single "knock- out" affects in a statistically significant way the activity of almost half of the other TUs, both near and far away in sequence space (Fig. 4Aii). The immediately adjacent TU (i.e., bead 931) is down- regulated the most, while more distant ones are
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[93, 60, 520, 392]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[88, 403, 515, 628]]<|/det|>
+<center>FIG. 4: Modelling SNPs and eQTL action. Sets of simulations ( \(\geq 800\) simulations/condition) where each of the 39 TU beads is made non-binding in turn (to represent 39 different SNPs in regulatory elements) are compared with those with the "wild-type" chain (as Fig. 1). A. Chain with mutant (nonbinding) TU bead 930. (i) Snapshot. TFs not shown (inset: same structure without blue beads). (ii) Transcriptional rates of the 17 TUs with significantly different values in mutant first compared with the wild type one ( \(p \simeq 0.046\) ; Students t-test). (iii) Regulatory network inferred from the matrix of Pearson correlations between activities of TUs. (iv) Change in Pearson correlation between TUs. B. Results from simulations where each TU bead is mutated in turn, and the "transcriptional difference" from the wild type (see text) determined. (i) Transcriptional difference versus position along the chain. (ii) Positive correlation of transcriptional difference with TU activity in wild-type. The plot shows that if we mutate a TU with high transcriptional activity, this leads to a larger difference. </center>  
+
+<|ref|>text<|/ref|><|det|>[[88, 658, 513, 757]]<|/det|>
+up- regulated (due to loss of a strong competitor). This knock- out also rewires the whole network, even though it still retains its small- world character (Fig. 4Aiii). Both positive and negative interactions are affected along the whole chain, as shown by a heat map of the change in Pearson correlation between TU transcriptional activities (Fig. 4Aiv).  
+
+<|ref|>text<|/ref|><|det|>[[88, 760, 513, 859]]<|/det|>
+We next systematically knock out each TU in turn. To quantify global effects, we define a "transcriptional difference" between the wild- type and each knock- out based on a standard Euclidian- distance metric (SI, Supplementary Note 3); the larger this quantity, the more different the two states are. This difference varies \(> 10\) - fold between different mutations (Fig. 4Bi).  
+
+<|ref|>text<|/ref|><|det|>[[541, 61, 966, 162]]<|/det|>
+Together, these observations are reminiscent of the behaviour of SNPs and eQTLs. Thus, each TU mutant can be seen as a SNP underlying an eQTL; then, those with low and high transcriptional differences (Fig. 4Bi,ii) are low- and high- effect eQTLs (low- effect mutants are often isolated in sequence space), and those with wide effects (e.g., bead 930 in Fig. 4A) may be viewed as omnigenic.  
+
+<|ref|>sub_title<|/ref|><|det|>[[550, 189, 956, 202]]<|/det|>
+## Modelling loops, heterochromatin and euchromatin  
+
+<|ref|>text<|/ref|><|det|>[[541, 220, 966, 404]]<|/det|>
+In mammalian genomes, promoter- enhancer pairs are often contained in loops stabilized by cohesin and the CCCTC- binding factor (CTCF) [42- 44]. To investigate how such loops might affect transcription, we incorporated eight permanent and non- overlapping loops at different positions in the chain (Fig. 5A, loops \(a - h\) ). In reality, such loops may arise from extrusion by cohesin halted at convergent CTCF loops [42]. Our assumption of stable, permanent loops is quantitatively accurate in the limit in which the interaction between cohesin and CTCF is strong and long- lived. However, we expect the trends to be qualitatively similar for more transient loops consistent with the loop extrusion model as in [19, 43].  
+
+<|ref|>text<|/ref|><|det|>[[541, 405, 966, 717]]<|/det|>
+The inclusion of stable loops has subtle effects. For example, loop \(h\) encompasses three TUs (beads 905, 907, 930), and expression of one is slightly boosted compared to the unlooped case (Figs. 5B,C). This is consistent with the idea that looping switches on some genes during development [45], and increases enhancer- promoter interactions [46]. However, up- regulation requires appropriate positioning of a TU within the loop. For instance, loop \(d\) encompasses two TUs (beads 396, 404), and has no effect on their activity. Broadly speaking, looping up- regulates activity, but not invariably so, and - perhaps surprisingly - two of the three most up- regulated TUs (beads 33, 886) are not contained in loops (Fig. 5C). Looping also extensively rewires the regulatory network (Fig. 5D,E). Globally, the increase in activity is modest, as incorporating all beads into closely- packed loops only increases total activity by \(\sim 10\%\) , with - once again - some TUs being down- as well as up- regulated (Fig. S3). This is consistent with experiments showing that the interplay between looping and expression is complex [47] but slight (e.g., knocking down human cohesin leaves expression of \(87\%\) genes unaffected, with global levels changing \(< 30\%\) [48]).  
+
+<|ref|>text<|/ref|><|det|>[[541, 718, 966, 859]]<|/det|>
+In simulations thus far, TFs bind strongly to TU beads, and weakly to all others to model binding to open euchromatin [19, 49]. To investigate the effects of heterochromatin - which binds few TFs, carries few histone marks [50], and is gene poor and traditionally viewed as transcriptionally inert - we perform simulations where four of the most- active TUs (905, 907, 930 and 931) are embedded in a non- binding segment (running from bead 901 - 940). This has a dramatic effect (Figs. 6A- C): the activity of the TU beads now embedded in the non- binding
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[88, 60, 513, 120]]<|/det|>
+island are at least halved, some nearby neighbors are downregulated, and more distant ones up- regulated (again due to a reduction in competition; Figs. 6B,C). The regulatory network is also rewired (Figs. 5D,E).  
+
+<|ref|>text<|/ref|><|det|>[[88, 120, 513, 190]]<|/det|>
+Just as embedment in a non- binding segment downregulates a TU bead, embedment in a weak- binding (euchromatic) one up- regulates it (Fig. S4). This shows our model effectively captures position effects where the local chromatin context strongly influences activity [51].  
+
+<|ref|>image<|/ref|><|det|>[[88, 202, 520, 501]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[88, 512, 515, 688]]<|/det|>
+<center>FIG. 5: Looping subtly affects transcriptional activity. Results of two sets of simulations are compared; one set as Fig. 1, in the other the chain contains 8 permanent loops (to represent convergent loops stabilized by cohesin/CTCF). A. Snapshot (beads within loops are magenta; TFs not shown; inset – same structure with only TUs shown). B. Average transcriptional activity for each TU in the looped chain (magenta bars – values for TUs in loops; magenta arcs – loop positions). C. Comparison between activity in wild-type and looped configuration for the 25 TUs with significantly different values in the two sets ( \(p \simeq 0.003\) ; Students t-test). D. Regulatory network inferred from the matrix of Pearson correlations between expression of TUs (as Fig. 3A). E. Change in Pearson correlation between TUs due to loops. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 729, 453, 742]]<|/det|>
+## Modelling a whole human chromosome  
+
+<|ref|>text<|/ref|><|det|>[[88, 759, 513, 858], [541, 614, 966, 671]]<|/det|>
+We next model a whole mid- sized human chromosome (HSA 14, length 107 Mbp; Fig. 7A), in a well- characterized and differentiated diploid cell (HUVEC, human umbilical vein endothelial cell). Now, multivalent and switchable TFs (20% active at any moment) at a non- saturating concentration bind to a string with 35784 beads. As chromosome territories are often ellipsoidal, simulations are performed in an ellipsoid of appropriate size[7, 52]; consequently, chromatin density is now higher than in simulations detailed above, with volume fractions comparable to those in vivo ( \(\sim 14\%\) ).  
+
+<|ref|>image<|/ref|><|det|>[[540, 58, 976, 404]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[540, 415, 966, 590]]<|/det|>
+<center>FIG. 6: Neighboring heterochromatin affects transcriptional activity. Results from two sets of simulations (at least 800 runs for each condition) are compared; one set as Fig. 1, in the other beads around TU beads 905, 907, 930 and 931 (from bead 901 to 940) are non-binding (to represent embedding the TU beads in heterochromatin). A. Snapshot with heterochromatic beads shown in gray (TFs not shown; inset – the same structure with only TUs). B. Average transcriptional activity for each TU. C. Comparison of average transcriptional activity with respect to wild-type for the 22 TUs with significantly different values in the two sets ( \(p \simeq 0.003\) ; Students t-test). D. Regulatory network inferred from the matrix of Pearson correlations between activities of TUs (as Fig. 3A). E. Change in Pearson correlation between TUs due to heterochromatin. </center>   
+
+<|ref|>text<|/ref|><|det|>[[541, 672, 966, 859]]<|/det|>
+Chromatin beads are classified using DNase- hypersensitivity data and ChIP- seq data for H3K27ac. DNase- hypersensitive sites (DHS) are excellent markers to locate promoters and enhancers (and so TF binding sites [19, 53]), whereas H3K27ac modifications strongly correlate with open chromatin [19]. Therefore, if the 3 kbp region corresponding to a chromatin bead has a DHS, then that bead is a TU; if it has H3K27ac, it is a euchromatin bead, and all other beads are non- binding (heterochromatic). We call this the "DHS" model. As properties of different chromatin segments have been catalogued using "hidden- Markov models" (HMMs) applied to many data sets [50], we alternatively classify beads
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[88, 60, 513, 90]]<|/det|>
+according to HMM state; we call this the "HMM model" (Fig. S5). For more details, see Supplementary Note 4.  
+
+<|ref|>text<|/ref|><|det|>[[88, 91, 514, 387]]<|/det|>
+Simulations using the DHS model again yield clusters enriched in TUs and TFs (Fig. 7B). As before, aggregating data from many simulations allows determination of transcriptional activities of every bead, which we compare with those of corresponding regions determined experimentally [54] by GRO- seq (global run- on sequencing [55]); activities of all 3 kbp regions are ranked from high to low, binned into quintiles, and compared. In Figure 7C, squares near the diagonal from bottom- left to top- right have high ranks (shown as red and yellow) compared to those off- diagonal (blue and purple) indicating good concordance between simulations and data. A specific sub- set of beads corresponding to super- enhancers (SEs) – which are highly active in vivo [56] – are also highly active in simulations (shown as white dots concentrated at top right). Plots showing the rank of transcriptional activities in simulations and experiments in selected genomic regions are shown in Fig. S6. Simulations yield patterns qualitatively closer to those obtained with GRO- seq than those given by poly(A) \(^+\) RNA- seq, as the latter only include gene transcription.  
+
+<|ref|>text<|/ref|><|det|>[[88, 388, 514, 686]]<|/det|>
+Concordance between results from simulations and GRO- seq is confirmed by the Spearman rank correlation \((\sim 0.38\) for all beads; \(p< 10^{- 12}\) ; this measure is used because it is less sensitive to outliers; Fig. 7D). Restricting analysis just to TUs provides a more stringent comparison (as all TUs bind TFs with equal affinity); it still yields a significant correlation ( \(r\simeq 0.32\) , \(p< 10^{- 12}\) ; Fig. 7D). As neighbouring high- affinity regions tend to have roughly similar transcriptional rates in both simulations and data, we also average rates found in active "patches" (contiguous sets of beads which are either all TUs or all labelled as euchromatin), but found this has no significant effect (Fig. 7D). Concordance was confirmed using our HMM model (Fig. 7D, right, and Fig. S5). Adding cohesin- mediated looping to simulations involving the DHS model did not significantly change agreement with experimental data (e.g., for TUs only, \(r\simeq 0.33\) , \(p< 10^{- 12}\) ). Similar agreement with GRO- seq data was obtained from simulations applied to the H1 human embryonic stem- cell line (for TUs using the DHS model, \(r\simeq 0.29\) , \(p< 10^{- 12}\) ), and to the GM12878 cell line (DHS model, \(r\simeq 0.33\) , \(p< 10^{- 12}\) ).  
+
+<|ref|>text<|/ref|><|det|>[[88, 687, 514, 842]]<|/det|>
+As in the chromosome fragment simulations (Fig. S1B), the transcriptional activity of a TU in our model anticorrelates with the distance to the nearest TU. In our HSA14 simulations, the presence of heterochromatin slightly reduces the absolute value of the correlation, which however remains highly significant (Spearman correlation \(r\simeq - 0.83\) , \(p< 10^{- 12}\) ). Interestingly, the experimental GRO- seq signal of a DHS also anticorrelates with the distance to the nearest DHS in a significant way, although more weakly than in simulations (Fig. S7; over the whole genome the Spearman correlation is \(r\sim - 0.23\) , \(p< 10^{- 12}\) ).  
+
+<|ref|>sub_title<|/ref|><|det|>[[548, 62, 958, 88]]<|/det|>
+## Networks inferred from simulations are qualitatively similar to experimental ones  
+
+<|ref|>text<|/ref|><|det|>[[541, 105, 965, 290]]<|/det|>
+Regulatory networks emerging from our whole chromosome simulations are again small- world and highly connected (Fig. S8, and Supplementary Note 5). To facilitate comparison with previous results, we select four segments of HSA14 that have the same length as the one considered in Fig. 4 (i.e., 3 Mbp), and roughly the same density of TUs; all four segments again have highly- connected components (compare Figs. S8 and Fig. 4). However, patterns in real chromosomes and artificial fragments are quite different. In HSA14 networks, there are more positive interactions between sets of adjacent TUs and other sets that are \(>10\) beads distant in sequence space (black lines across the middle of circles in Fig. S8).  
+
+<|ref|>text<|/ref|><|det|>[[541, 291, 965, 476]]<|/det|>
+Whole- chromosome networks also have the following statistical properties. First, their node- degree distribution decays exponentially (Fig. S9A) – as found in gene networks [57] but not in transcription factor interaction networks, which are often scale- free [58]. Second, they are modular (as clusters arising due to the bridging- induced attraction are the basic co- regulated building blocks) – again as found in gene [57] and eQTL [59] networks. [Modularity is apparent from the blocks visible in the correlation matrices, such as in Fig. S2.] Third, node degree broadly correlates with transcriptional activity (Spearman correlation 0.59, p- value \(< 10^{- 12}\) ) – as in gene coregulation networks [57].  
+
+<|ref|>sub_title<|/ref|><|det|>[[545, 490, 958, 515]]<|/det|>
+## Contact maps found by simulations are qualitatively similar to Hi-C  
+
+<|ref|>text<|/ref|><|det|>[[541, 532, 965, 858]]<|/det|>
+We previously showed [16] that simulations involving two different TFs (binding to active and inactive regions, respectively) yield contact maps much like those found with Hi- C [42]. Therefore, we expected the present simulations to reflect Hi- C data poorly as they involve only one TF binding to the minor (i.e., active) fraction of the genome, so contacts made by this structured minority would be obscured by those due to the unstructured majority. Even so, simulations yield contact maps broadly similar to those obtained by Hi- C (Fig. 7E). To measure the agreement, we use a comparison based on contact maps restricted to TUs as anchors – which may be considered as equivalent to interactions obtained by promoter- capture HiC [60]. These yield good concordance (Fig. 7E; Pearson coefficient \(r = 0.82\) ; \(r = 0.47\) when monitoring only long- range contacts between TUs at least 300 kbp away, \(p< 10^{- 6}\) in both cases). The exponent with which contact probability decays with 1D distance is \(\sim - 1.1\) in experiments, and \(\sim - 0.8\) in simulations (fitted for 1D distances between \(\sim 30\) kbp and 1.5 Mbp), both broadly consistent with the \(- 1\) value expected for a fractal globule [61]. The small discrepancy may point to our simulations slightly overestimating the weight of long- range contacts, perhaps
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[88, 62, 395, 75]]<|/det|>
+because we do not include loop extrusion.  
+
+<|ref|>text<|/ref|><|det|>[[87, 76, 515, 133]]<|/det|>
+Overall the results obtained in our HSA14 simulations show that a simple model based on 3D chromatin organisation captures much of the complexity in 3D structure and transcription of a whole human chromosome.  
+
+<|ref|>sub_title<|/ref|><|det|>[[108, 147, 494, 172]]<|/det|>
+## Modelling chromosome 22 carrying the diGeorge deletion  
+
+<|ref|>text<|/ref|><|det|>[[88, 189, 515, 312]]<|/det|>
+Our approach can, in principle, be applied to study any chromosome providing appropriate genomic data are available (e.g., on DNase hypersensitivity and histone acetylation). As a proof of principle, we studied the effect of deleting \(\sim 2.55\) Mbp from HSA22 - an alteration which is associated with the diGeorge syndrome (Fig. 8A) [65]. This syndrome affects \(\sim 1\) in 4000 people, and the variable symptoms include congenital heart problems, frequent infections, developmental delays, and learning problems.  
+
+<|ref|>text<|/ref|><|det|>[[88, 312, 515, 585]]<|/det|>
+We predict a multitude of small effects in TU activity, both near and far away from the deletion (see the Manhattan plot in Fig. 8Bi). In particular, most TUs are slightly up- regulated, as fewer TUs compete for the same number of factors, and the TUs which change the most have intermediate transcriptional activities in the wildtype (Fig. S10). The p- values associated with the change in transcriptional activities vary widely, and comparison of the observed distribution with the null hypothesis (indicating that changes in measured transcription are due to random variation) shows the observed is highly enriched in small p- values (Fig. 8Bii), as is generally the case with results from GWAS [5, 6]. The regulatory network is also re- wired (Fig. 8C). Results are consistent with measurements of differential gene expressions in patients, which showed both a large number of up- regulated and downregulated genes [62]. A more quantitative comparison between experiments and simulations would benefit from having GRO- seq data that include non- genic transcription.  
+
+<|ref|>text<|/ref|><|det|>[[88, 586, 515, 656]]<|/det|>
+Clearly, this approach opens up a rich new field of study. For instance, while there may be processes which occur in vivo which are not represented in our model, it could still give an indication of the genes most likely to be affected by any chromosome rearrangement.  
+
+<|ref|>sub_title<|/ref|><|det|>[[153, 683, 448, 696]]<|/det|>
+## DISCUSSION AND CONCLUSIONS  
+
+<|ref|>text<|/ref|><|det|>[[88, 714, 515, 854]]<|/det|>
+We have described a parsimonious 3D stochastic model for transcriptional dynamics based on multivalent binding of factors and polymerases (TFs) to genic and non- genic transcriptional units (TUs) in a chain representing a chromatin fibre. A distinctive feature of our framework is that it is fitting- free, which means the model is truly predictive and can provide a mechanistic understanding of the phenomena we observe. On the other hand, the absence of fitting renders it challenging to obtain a fully quantitative agreement between modelling and experiment.  
+
+<|ref|>image<|/ref|><|det|>[[585, 64, 927, 562]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[540, 577, 968, 851]]<|/det|>
+<center>FIG. 7: Comparison of transcriptional activities of TUs on HSA14 in HUVECs determined using simulations and GRO-seq. A. Workflow (DHS model). Simulations (244 runs) involve a chain (35784 beads) representing HSA14, and 1700 switchable TFs confined in an ellipsoidal territory. Beads are classified as TUs (red, strong-binding), euchromatic (blue, weak-binding) or heterochromatic (grey, non-binding). Transcriptional activities from simulations are compared with those of GRO-seq data, by measuring the Spearman rank correlation. B. (i) Snapshot (TFs not shown). (ii, iii) TU beads and TFs in this configuration. C. Comparison of transcriptional activities of TUs from simulations and GRO-seq (ranked from \(0 - 100\%\) , then binned in quintiles and showed as a heat map). A scatter plot of unbinned ranks of beads corresponding to SEs are superimposed (white circles). D. Comparison of transcriptional activities from simulations (for both DHS and HMM models) and GRO-seq for all 3 kb regions/beads, only TUs, and only connected patches of binding beads (see text). All correlations are significant \((p < 10^{-6}\) , indicated by grey lines). E (i-ii) Capture-HiC-like contact maps obtained from simulations and experiments [42] showing logarithm of number of contacts between 30 kbp bins which contain TUs. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[103, 60, 218, 77]]<|/det|>
+## A workflow  
+
+<|ref|>text<|/ref|><|det|>[[105, 81, 319, 94]]<|/det|>
+HSA22 wild- type (16,250 beads)  
+
+<|ref|>image<|/ref|><|det|>[[130, 97, 515, 150]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[103, 164, 500, 179]]<|/det|>
+B effects on transcription at other sites are:  
+
+<|ref|>text<|/ref|><|det|>[[103, 180, 450, 195]]<|/det|>
+i ... widely scattered ii ... small  
+
+<|ref|>image<|/ref|><|det|>[[103, 198, 520, 360]]<|/det|>  
+
+<|ref|>image<|/ref|><|det|>[[106, 363, 520, 450]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[87, 462, 513, 725]]<|/det|>
+<center>FIG. 8: Modelling effects of the DiGeorge deletion in HSA22. A. Workflow. Simulations (800 simulations/condition) for wild type (17102 beads) and deletion (16250 beads, where wild-type beads 6305 - 7156 are cut, corresponding to a deletion of chr22:18912231 - 21465672 in hg19). [Agreement between predicted transcriptional activity and GRO-seq in HSA22 is similar to that found for HSA1 (here, Spearman correlation is \(r \sim 0.29\) , \(p < 10^{-6}\) ]. B. (i) Manhattan plot showing \(-\log_{10}\) (p-value) as a function of genomic position along HSA22 (position given in Mbp), for changes in TU transcriptional activities between wild-type and deletion. (ii) Quantile-quantile plot showing expected versus observed values for \(-\log_{10}\) (p-value) for the same data in (i). Expected values are computed from the normal distribution (these correspond to the null hypothesis according to which the change in transcriptional activities in the deletion is purely due to random variation). C. Regulatory networks of two 3 Mbp segments in chromosome 22 inferred from the Pearson correlation matrix. Edges show positive correlations \(> 0.12\) ( \(p = 0.0007\) ). Segments chosen have roughly the same number of nodes in 3 Mbp as the short fragment (Fig. 3Aii). </center>  
+
+<|ref|>text<|/ref|><|det|>[[88, 744, 513, 858], [541, 60, 966, 558]]<|/det|>
+In our simulations two types of fibres were considered: a 3 Mbp fragment with randomly- positioned TUs, which is useful to exemplify emerging trends, and human chromosomes 14 and 22 where TUs were appropriately positioned according to bioinformatic data. Despite deliberately excluding any explicit underlying network of biochemical regulation, our model nevertheless yields some notable results. These depend on having a low TF copy- number - a feature compatible with observations in vivo [23]. First, since TFs bind with the same affinity to all TUs, one might expect the latter to all be transcribed similarly, but they are not (Fig. 1). This is largely due to inter- TU spacing; TUs lying close together in 1D sequence space tend to be the most active (Fig. 1C) with positively- correlated dynamics reminiscent of transcriptional bursting (Fig. 2B). This is because they often cluster into structures which are analogous to the phase- separated transcription hubs/factories seen experimentally [7, 10], or to contact domains formed by accessible DNA sites found by high- resolution mapping of chromatin interactions by microC [30]. Second, switching off binding at any TU significantly affects the activity of many others, both near and far away in sequence space (Fig. 4). Third, introducing stable loops has subtle effects (Fig. 5), consistent with the modest changes in expression seen experimentally in cohesin knock- outs and degrons [48]. Fourth, transcriptional activity of a TU is strongly affected by the local environment in ways that are reminiscent of the silencing of a gene by incorporation into heterochromatin [51] (Fig. 6), or activation by embedment in euchromatin (Fig. S4). Fifth, the stochasticity seen in individual simulations reflects that detected by single- cell transcriptomics and single- cell Hi- C. Nevertheless, this variability does not prevent emergence of robust phenotypes in a cell population. Sixth, our simple fitting- free model predicts patterns of transcriptional activity in human chromosomes that promisingly and significantly correlate with experimental GRO- seq data (Fig. 7). This suggests that chromatin structure significantly constrains transcriptional activity. We hypothesise that additional downstream biochemical regulation, not included in our model, may provide a tool to adjust this underlying "structural" pattern of activity in a way which may be required for appropriate biological function.   
+
+<|ref|>text<|/ref|><|det|>[[542, 560, 966, 757]]<|/det|>
+Finally, our results enable us to reconcile two conflicting sets of data, namely that regulatory networks are both complex (as GWAS shows that thousands of loci around the genome control complex phenotypes [5, 6]) and simple (as over- expressing just four Yamanaka factors switches cell fate [4]). Thus, our simulations reveal complex small- world networks of mutual up- and down- regulation (Figs. 3 and S8), consistent with GWAS results. However, increasing TF copy- number dramatically simplifies network structure (Fig. 3). We suggest such a simplification occurs when a fibroblast is reprogrammed into a pluripotent stem cell by over- expressing the Yamanaka factors; the high factor concentration simplifies the network so that the factors can combine to switch the phenotype (Fig. S11).  
+
+<|ref|>text<|/ref|><|det|>[[542, 759, 966, 858]]<|/det|>
+Taken together, these results suggest the activity - or inactivity - of every genomic region affects that of every other region to some extent. We describe our framework as "pan- genomic" (Fig. S8). This is reminiscent of the omni- . genic model [5, 6] in the sense that many loci are involved, all having small effects. However, it differs as it provides an underlying mechanism for pangenomic effects, by posit
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[88, 60, 515, 319]]<|/det|>
+ing a direct and immediate effect of structure on regulation at the transcriptional level, which contrasts with the nontrivial post- transcriptional pathways envisioned by the on- nigenic model. Additionally, our pangenomic model yields a natural framework to qualitatively understand mutually exclusive gene expression, when switching on one gene in a family turns off all others (as in developing olfactory neurons [63]). The current model to explain this phenomenon postulates a coupling between cis- acting up- regulation and trans- acting down- regulation. The pangenomic networks we find provide exactly this type of regulatory interactions (Fig. 3). On the other hand, it is challenging within our current model to account for local negative feedback mechanisms leading to noise reduction or oscillations [11], as these are more likely to arise biochemically (an example is the p53- Mdm2 system which achieves stabilisation of the cellular concentration of p53 via a negative feedback loop [64]).  
+
+<|ref|>text<|/ref|><|det|>[[88, 321, 515, 476]]<|/det|>
+In conclusion, we have developed a framework that can be applied to predict the transcriptional activity of any genomic fragment in health or disease (Figs. 7, 8) providing appropriate experimental data is available. Predictive power can be enhanced by incorporating additional TFs, and more suitable datasets of histone marks. Other features which can improve correlations between experiments and simulations are a more accurate modelling of cohesin loop formation by loop extrusion, and of the heteromorphic nature of chromatin [19]. We hope to report on work incorporating the latter two features in the future.  
+
+<|ref|>text<|/ref|><|det|>[[88, 479, 515, 508]]<|/det|>
+We thank the European Research Council (ERC CoG 648050 THREEDCELLPHYSICS) for support.  
+
+<|ref|>sub_title<|/ref|><|det|>[[106, 512, 230, 526]]<|/det|>
+## Code availability  
+
+<|ref|>text<|/ref|><|det|>[[88, 529, 515, 585]]<|/det|>
+The code used for the simulation is LAMMPS, which is publicly available at https://lammps.sandia.gov/. Custom codes written to analyse data are available from the corresponding author upon request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[106, 589, 228, 602]]<|/det|>
+## Data availability  
+
+<|ref|>text<|/ref|><|det|>[[88, 606, 515, 650]]<|/det|>
+The datasets generated during and/or analysed during the current study are available from the corresponding author upon request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[106, 655, 260, 668]]<|/det|>
+## Author contributions  
+
+<|ref|>text<|/ref|><|det|>[[88, 671, 515, 744]]<|/det|>
+C. A. B., N. G., D. Mi., A. P., P. R. C., M. C. F. P., D. Ma. designed research; C. A. B., M. C. F. P., D. Ma. performed research; C. A. B., N. G., D. Mi., A. P., M. C. F. P., P. R. C., D. Ma. analysed the data and wrote the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[544, 60, 969, 861]]<|/det|>
+[3] A. Dall'Agnese, L. Caputo, C. Nicoletti, J. di Iulio, A. Schmitt, S. Gatto, Y. Diao, Z. Ye, M. Forcato, R. Perera, et al., Mol. Cell 76, 453 (2019). [4] K. Takahashi, K. Tanabe, M. Ohnuki, M. Narita, T. Ichisaka, K. Tomoda, and S. Yamanaka, Cell 131, 861 (2007). [5] E. A. Boyle, Y. I. Li, and J. K. Pritchard, Cell 169, 1177 (2017). [6] X. Liu, Y. I. Li, and J. K. Pritchard, Cell 177, 1022 (2019). [7] P. R. Cook and D. Marenduzzo, Nucleic Acids Res. 46, 9895 (2018). [8] R. Andersson, C. Gebhard, I. Miguel-Escalada, I. Hoof, J. Bornholdt, M. Boyd, Y. Chen, X. Zhao, C. Schmidl, T. Suzuki, et al., Nature 507, 455 (2014). [9] B. M. Javierre, O. S. Burren, S. P. Wilder, R. Kreuzhuber, S. M. Hill, S. Sewitz, J. Cairns, S. W. Wingett, C. Varnai, M. J. Thiecke, et al., Cell 167, 1369 (2016). [10] P. Cramer, Nature 573, 45 (2019). [11] K. Sneppen, S. Krishna, and S. Semsey, Annu. Rev. Biophys. 39, 43 (2010). [12] P. Smolen, D. A. Baxter, and J. H. Byrne, Bull. Math. Biol. 62, 247 (2000). [13] A. Pombo and N. Dillon, Nat. Rev. Mol. Cell Biol. 16, 245 (2015). [14] M. Spielmann, D. G. Lupiáñez, and S. Mundlos, Nat. Rev. Genet. 19, 453 (2018). [15] M. Barbieri, M. Chotalia, J. Fraser, L.- M. Lavitas, J. Dostie, A. Pombo, and M. Nicodemi, Proc. Natl. Acad. Sci. USA 109, 16173 (2012). [16] C. A. Brackley, J. Johnson, S. Kelly, P. R. Cook, and D. Marenduzzo, Nucleic Acids Res. 44, 3503 (2016). [17] N. Gilbert and D. Marenduzzo, Chromosome Res. 25, 1 (2017). [18] M. C. F. Pereira, C. A. Brackley, D. Michieletto, C. Annunziabella, S. Bianco, A. M. Chiariello, M. Nicodemi, and D. Marenduzzo, bioRxiv, 305359 (2018). [19] A. Buckle, C. A. Brackley, S. Boyle, D. Marenduzzo, and N. Gilbert, Mol. Cell 72, 786 (2018). [20] E. H. Finn, G. Pegoraro, H. B. Brandao, A.- L. Valton, M. E. Oomen, J. Dekker, L. Mirny, and T. Misteli, Cell 176, 1502 (2019). [21] C. A. Brackley, B. Liebchen, D. Michieletto, F. L. Mouvet, P. R. Cook, and D. Marenduzzo, Biophys. J. 28, 1085 (2017). [22] B. Steurer, R. C. Janssens, B. Geverts, M. E. Geijer, F. Wienholz, A. F. Theil, J. Chang, S. Dealy, J. Pothof, W. A. van Cappellen, et al., Proc. Natl. Acad. Sci. USA 115, E4368 (2018). [23] R. C. Brewster, F. M. Weinert, H. G. Garcia, D. Song, M. Rydenfelt, and R. Phillips, Cell 156, 1312 (2014). [24] C. A. Brackley, S. Taylor, A. Papantonis, P. R. Cook, and D. Marenduzzo, Proc. Natl. Acad. Sci. USA 110, E3605 (2013). [25] C. Brackley, J. Phys.: Condens. Matter 32, 314002 (2020). [26] S. Kilic, A. L. Bachmann, L. C. Bryan, and B. Fierz, Nat. Comm. 6, 7313 (2015). [27] P. R. Cook, Science 284, 1790 (1999). [28] A. Papantonis, T. Kohro, S. Baboo, J. D. Larkin, B. Deng, P. Short, S. Tsutsumi, S. Taylor, Y. Kanki, M. Kobayashi, et al., EMBO J. 31, 4404 (2012). [29] K. Shrinivas, B. R. Sabari, E. L. Coffey, I. A. Klein, A. Boija, A. V. Zamudio, J. Schuijers, N. M. Hannett, P. A. Sharp, R. A. Young, et al., Mol. Cell 75, 549 (2019). [30] T.- H. S. Hsieh, C. Cattoglio, E. Slobodyanuk, A. S.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 60, 515, 600]]<|/det|>
+Hansen, O. J. Rando, R. Tjian, and X. Darzacq, Mol. Cell 78, 539 (2020).[31] J.- K. Ryu, C. Bouchoux, H. W. Liu, E. Kim, M. Minamino, R. de Groot, A. J. Katan, A. Bonato, D. Marenduzzo, D. Michieletto, F. Uhlmann, and C. Dekker, Sci. Adv. 7, eabe5905 (2021).[32] A. Pombo, D. A. Jackson, M. Hollinshead, Z. Wang, R. G. Roeder, and P. R. Cook, EMBO J. 18, 2241 (1999).[33] I. Faro-Trindade and P. R. Cook, Mol. Biol. Cell 17, 2910 (2006).[34] R. A. Beagrie, A. Scialdone, M. Schueler, D. C. Kraemer, M. Chotalia, S. Q. Xie, M. Barbieri, I. de Santiago, L.- M. Lavitas, M. R. Branco, et al., Nature 543, 519 (2017).[35] T. Nagano, Y. Lublin, T. J. Stevens, S. Schoenfelder, E. Yaffe, W. Dean, E. D. Laue, A. Tanay, and P. Fraser, Nature 502, 59 (2013).[36] I. C. Macaulay and T. Voet, PLoS Genet. 10, e1004126 (2014).[37] F. Muerdter and A. Stark, Curr. Biol. 26, R895 (2016).[38] T. Fukaya, B. Lim, and M. Levine, Cell 166, 358 (2016).[39] C. R. Bartman, S. C. Hsu, C. C.- S. Hsiung, A. Raj, and G. A. Blobel, Mol. Cell 62, 237 (2016).[40] D. M. Suter, N. Molina, D. Gattield, K. Schneider, U. Schibler, and F. Naef, Science 332, 472 (2011).[41] M. D. Humphries and K. Gurney, PLoS ONE 3, e0002051 (2008).[42] S. S. Rao, M. H. Huntley, N. C. Durand, E. K. Stamenova, I. D. Bochkov, J. T. Robinson, A. L. Sanborn, I. Machol, A. D. Omer, E. S. Lander, and E. LiebermanAiden, Cell 159, 1665 (2014).[43] G. Fudenberg, M. Imakaev, C. Lu, A. Goloborodko, N. Abdennur, and L. A. Mirny, Cell Rep. 15, 2038 (2016).[44] C. A. Brackley, J. Johnson, D. Michieletto, A. N. Morozov, M. Nicodemi, P. R. Cook, and D. Marenduzzo, Phys. Rev. Lett. 119, 138101 (2017).[45] M. Oti, J. Falck, M. A. Huynen, and H. Zhou, BMC Genomics 17, 252 (2016).[46] D. Sasca, H. Yun, G. Giotopoulos, J. Szybinski, T. Evan, N. K. Wilson, M. Gerstung, P. Gallipoli, A. R. Green, R. Hills, et al., Blood 134, 2195 (2019).[47] M. I. Robson, A. R. Ringel, and S. Mundlos, Mol. Cell 74, 1110 (2019).[48] S. S. Rao, S.- C. Huang, B. G. St Hilaire, J. M. Engreitz,  
+
+<|ref|>text<|/ref|><|det|>[[540, 60, 968, 590]]<|/det|>
+E. M. Perez, K.- R. Kieffer-Kwon, A. L. Sanborn, S. E. Johnstone, G. D. Bascom, I. D. Bochkov, et al., Cell 171, 305 (2017).[49] N. Gilbert, S. Boyle, H. Fiegler, K. Woodfine, N. P. Carter, and W. A. Bickmore, Cell 118, 555 (2004).[50] J. Ernst, P. Kheradpour, T. S. Mikkelsen, N. Shoresh, L. D. Ward, C. B. Epstein, X. Zhang, L. Wang, R. Issner, M. Coyne, et al., Nature 473, 43 (2011).[51] R. T. Timms, I. A. Tchasovnikarova, and P. J. Lehner, BioEssays 38, 333 (2016).[52] Y. Wang, M. Nagarajan, C. Uhler, and G. Shivashankar, Mol. Biol. Cell 28, 1997 (2017).[53] E. P. Consortium, Nature 489, 57 (2012).[54] H. Niskanen, I. Tuszynskax, R. Zaborowski, M. Heinaniemi, S. Yla-Herttuala, B. Wilczynski, and M. U. Kaikkonen, Nucleic Acids Res. 46, 1724 (2017).[55] A. Jordán-Pla, M. E. Pérez-Martínez, and J. E. Pérez-Ortín, Methods 159, 177 (2019).[56] A. Khan and X. Zhang, Nucleic Acids Res. 44, D164 (2015).[57] V. Belcastro, V. Siciliano, F. Gregoretti, P. Mithbaokar, G. Dharmalingam, S. Berlingieri, F. Iorio, G. Oliva, R. Polishchuk, N. Brunetti-Pierri, et al., Nucleic Acids Res. 39, 8677 (2011).[58] W. Z. Ouma, K. Pogacar, and E. Grotewold, PLoS Comput. Biol. 14, e1006098 (2018).[59] M. Fagny, J. N. Paulson, M. L. Kuijjer, A. R. Sonawane, C.- Y. Chen, C. M. Lopes-Ramos, K. Glass, J. Quackenbush, and J. Platig, Proc. Natl. Acad. Sci. USA 114, E7841 (2017).[60] B. Mifsud, F. Tavares-Cadete, A. N. Young, R. Sugar, S. Schoenfelder, L. Ferreira, S. W. Wingett, S. Andrews, W. Grey, P. A. Ewels, et al., Nat. Genet. 47, 598 (2015).[61] L. A. Mirny, Chromosome Res. 19, 37 (2011).[62] M. Jalbrzikowski, M. T. Lazaro, F. Gao, A. Huang, C. Chow, D. H. Geschwind, G. Coppola, and C. E. Bearden, PLoS ONE 10, e0132542 (2015).[63] A. K. Alsing and K. Sneppen, Nucleic Acids Res. 41, 4755 (2013).[64] S. L. Harris and A. J. Levine, Oncogene 24, 2899 (2005).[65] https://dosage.clinicalgenome.org/clingen_region.cgi?id=ISCA-37446
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[61, 130, 317, 150]]<|/det|>
+- Sl.pangenomic.revised.pdf
+
+<--- Page Split --->

preprint/preprint__0996185957de22ec96c13ddd4eeca9080a098f778fc62996a2333ee389185663/images_list.json

@@ -0,0 +1,77 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 | The schematical diagram of the presented BSTCM. The BSTCM system integrates STC metasurface into SSVEP-based BCI systems. The STC metasurface can correctly support the visual stimuli for the SSVEP-based BCI and facilitate the information interaction with the external environment. Based on the BSTCM system, high-security encrypted wireless communication systems are realized for the first time by combining with the variant HSKs and VSS method. In this way, smart devices can be controlled by the human mind with high security.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        85,
+        880,
+        394
+      ]
+    ],
+    "page_idx": 20
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2 | The diagram of interaction and data processing procedures of the SSVEP-based BCI system. a. The interaction procedures for the SSVEP-based BCI involve the user wearing an EEG cap and focusing their attention on the intended target LED stimuli. Once the user's selected frequency is recognized, the interaction commands are transmitted to the STC metasurface. b. The feature extraction process includes the transformation of the input BCI signals into the frequency domain. The resulting Signal FFT is then subjected to outer product with pre-defined Template FFTs, yielding eight feature graphs. c. The lightweight classification model. Four convolutional and two fully connected layers were employed. d. An example of frequency responses of SSVEP signals for the four target frequencies, showing distinctive amplitude peaks corresponding to different frequencies.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        84,
+        868,
+        620
+      ]
+    ],
+    "page_idx": 21
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3 | The details of the STC metasurface. a. The geometrical structure of the STC metasurface element. b, c. The reflective amplitude and phase of the element, respectively. d-f. The optimized STC matrices for different scattering angles at different harmonic frequencies. g-i. The far-field results corresponding to the optimized STC matrices. j-m. The encoding schemes for the transmitting symbol “0” or “1” for two users: the symbols “0” and “1” for Users 1 (j and k); the symbols “0” and “1” for Users 2 (l and m).",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        96,
+        880,
+        555
+      ]
+    ],
+    "page_idx": 22
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4 | The encrypted wireless communication system based on the BSTCM platform. a. The encrypted encoding scheme combining the harmonic-secret keys and the VSS method. b. The experimental scenarios of the encrypted wireless communication system, in which the operator equipped with EEG cap transmits the encrypted information to two users via the BSTCM platform, respectively. c. The receiving signals corresponding to the encoding scheme in Fig 3(j-m). d, e. The decoded information of VSK1 and VSK2 from the receiving signals based on the encrypted coding scheme. f. The correct transmitting information is decoded by the extracted VSK1 and VSK2.",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        92,
+        884,
+        540
+      ]
+    ],
+    "page_idx": 23
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5 | Experiment on the wireless remote mind control of smart devices based on the BSTCM platform. a. The system architecture of wireless remote control. b. The experimental scenarios of the wireless remote-control system by the human brain. c, d. The theoretical and experimental results for far-filed patterns at \\(\\pm 2\\mathrm{th}\\) harmonic frequencies. f. The temporal waveform of the output voltages from four detectors when the operator lights up four devices in sequence. e. The variations between the input power of the detectors and output voltage with respect to the different distances between the receiving antenna and the metasurface center.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        88,
+        880,
+        428
+      ]
+    ],
+    "page_idx": 24
+  }
+]

preprint/preprint__0996185957de22ec96c13ddd4eeca9080a098f778fc62996a2333ee389185663/preprint__0996185957de22ec96c13ddd4eeca9080a098f778fc62996a2333ee389185663.mmd

@@ -0,0 +1,239 @@
+
+# Secure wireless communication of brain-computer interface and mind control of smart devices enabled by space-time-coding metasurface  
+
+Tie Jun Cui  t.jcui@seu.edu.cn  
+
+Southeast University https://orcid.org/0000- 0002- 5862- 1497  
+
+Qiang Xiao  Southeast University  
+
+Lin Han Fan  Southeast University  
+
+Qian Ma  Southeast University https://orcid.org/0000- 0002- 4662- 8667  
+
+Yu Ming Ning  Southeast University  
+
+Ze Gu  Southeast University  
+
+Long Chen  Southeast University https://orcid.org/0009- 0007- 1533- 0319  
+
+Lianlin Li  peking university https://orcid.org/0000- 0001- 9394- 3638  
+
+Jian Wei You  Southeast University https://orcid.org/0000- 0001- 5761- 9507  
+
+Ya Feng Niu  Southeast University  
+
+## Article  
+
+Keywords: Space- time- coding metasurface, brain- computer interface, secure wireless communication, human- machine interactions  
+
+Posted Date: August 23rd, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4860006/v1
+
+<--- Page Split --->
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on August 25th, 2025. See the published version at https://doi.org/10.1038/s41467-025-63326-0.
+
+<--- Page Split --->
+
+# Secure wireless communication of brain-computer interface and mind control of smart devices enabled by space-time-coding metasurface  
+
+Qiang Xiao \(^{1,\dagger}\) , Lin Han Fan \(^{2,\dagger}\) , Qian Ma \(^{1,\dagger}\) , \(*\) , Yu Ming Ning \(^{1}\) , Ze Gu \(^{1}\) , Long Chen \(^{1}\) , Lianlin Li \(^{3}\) , Jian Wei You \(^{1}\) , Ya Feng Niu \(^{2,\ast}\) and Tie Jun Cui \(^{1,\ast}\)  
+
+\(^{1}\) State Key Laboratory of Millimeter Waves and Institute of Electromagnetic Space, Southeast University, Nanjing 210096, China  
+
+\(^{2}\) School of Mechanical Engineering, Southeast University, Nanjing 210096, China  
+
+\(^{3}\) State Key Laboratory of Advanced Optical Communication Systems and Networks, Department of  
+
+Electronics, Peking University, 100871 Beijing, China  
+
+\*E- mail: maqian@seu.edu.cn, nyf@seu.edu.cn, tjcui@seu.edu.cn  
+
+\(^{\dagger}\) These authors contributed equally to this work  
+
+## Abstract  
+
+Brain- computer interface (BCI) provides an interconnected pathway between human brain and external devices and paves a potential route for mind manipulations. However, most of existing BCI technologies are based on simple signal transmission and independent of other interface devices, owing to the considerations of reliability and safety of human brain's information interaction in the complicated wireless environment. To address the formidable limitation, we present a brain space- time- coding metasurface (BSTCM) system to deeply fuse the visual stimulation and electromagnetic manipulation for reliable and secure information transfer between human brain and external devices. Here, we innovatively integrate the BCI flashing frame and electromagnetic encoding sequence in the BSTCM system, and the STC metasurface ensures the secure wireless communications by using the harmonic- encrypted beams. We design and fabricate a proof- of- principle demonstration system and experimentally show that the proposed wireless BCI scheme could establish a remote but safeguarded paradigm for human- machine interactions, intelligent metasurfaces, and potential applications in metaverse, as a prominently scrutinized domain in the future 6G wireless communications.  
+
+KEYWORDS: Space- time- coding metasurface, brain- computer interface, secure wireless communication, human- machine interactions
+
+<--- Page Split --->
+
+## Introduction  
+
+Brain- computer interface (BCI) has emerged as a cutting- edge technology in human- machine interaction, and demonstrates promising applications such as metaverse interaction \(^1\) and smart homes \(^2\) . Electroencephalography (EEG) signals remain the predominant input signal modality in BCI systems, with notable implementations including motor imagery \(^3\) , P300 \(^4\) , and steady- state visually evoked potential (SSVEP) \(^5\) . The SSVEP- based BCI systems utilize the brain's SSVEP response to the fixed- frequency visual stimuli for mind recognition and interaction \(^6\) , offering a significant advantage of high information transfer rate (ITR). The BCIs have been widely used in various fields such as augmented reality (AR) \(^7,8\) , spelling input \(^9,10\) , medical rehabilitation and equipment control \(^11,12\) . Recently, some high- performance BCI systems have been continuously proposed \(^13,18\) . The advent and progression of 6G wireless communication technology substantially broaden the application prospects of BCIs, particularly in Internet of Things (IoT), virtual reality (VR) devices, and beyond \(^19,21\) . Hence, ensuring high security and privacy preservation \(^22\) becomes imperative when facilitating intelligent interactions within the constructed communication environment.  
+
+However, most existing BCI systems lack in- depth research in terms of security. During the wireless transmissions of brain signals in a BCI system, there exists a vulnerability to theft and attack, potentially leading to the generation of inaccurate device control commands and unauthorized disclosure of personal privacy. Though some methodologies have been proposed to enhance the security and privacy of data in the BCI systems \(^{23 - 25}\) , there is a conspicuous absence to investigate the encryption mechanisms specifically tailored for the BCI systems. More importantly, the implementation of visual stimulation remains to be isolated from the back- end information processing systems, lacking deep information fusion and interaction. With the increasing demand for high security in the BCI systems, it is essential to develop intelligent interactions in the secure and reliable communication environment. The frequency- dependent SSVEP response and programmable harmonic characteristics of space- time- coding (STC) metasurfaces display notable similarities. Therefore, the STC metasurfaces can be used as a promising method that not only provides the visual stimulation but also ensures security of the BCI systems at the physical layer owing to their powerful ability to flexibly manipulate
+
+<--- Page Split --->
+
+electromagnetic (EM) waves in both time and space domains<sup>26</sup>.  
+
+The metasurfaces are composed of specific unit structures arranged in periodic or quasi- periodic arrays, which can flexibly control the EM waves at the subwavelength scale and yield a large number of unusual physical phenomena and novel devices<sup>27- 30</sup>. The proposal of digital coding and programmable metasurfaces has established a profound connection between the EM fields and digital information under the control of a high- speed field programmable gate array (FPGA)<sup>31</sup>. Recently, the exploration of STC metasurfaces has sparked a surge of research interest due to the excellent ability to manipulate the EM waves and process digital information in both temporal and spatial dimensions<sup>32- 37</sup>, resulting in many novel physical phenomena that cannot be realized by the traditional spatially modulated metasurfaces. More importantly, the utilization of STC metasurfaces holds the capability to precisely control the amplitudes, phases, and polarizations at different harmonic frequencies independently by specially designing the STC matrices<sup>38</sup>, which opens up avenues to develop advanced communication schemes with enhanced efficiency and reliability<sup>39- 45</sup>. Hence, the STC metasurface is a potential candidate for deep information modulation and interaction in the SSVEP- based BCI system owing to its capability to process information and interact with the frequency- dependent visual stimulation.  
+
+In this article, we report a brain space- time- coding metasurface (BSTCM) system, which combines the human brain intelligence with the flexible capability of EM manipulations by the STC metasurface. To the best of our knowledge, this is the first instance to integrate the visual stimulation and EM encoding regulation deeply in a BCI system through metasurface. Here, the STC metasurface is used to effectively support visual stimulation for SSVEP- based BCIs while enabling the information interaction with the external environment. A lightweight deep- learning classification model is implemented to maximize the recognition performance. To enhance the BCI security, we propose a high- security encrypted wireless communication system based on the diverse harmonic regulation characteristics of BSTCM, which effectively combines the variant harmonic- based secret keys (HSKs) and visual secret sharing (VSS) method<sup>46- 48</sup>. In this way, decoding by eavesdroppers is only possible through the simultaneous interception of confidential information across all harmonic frequencies in the communication process and the acquisition of complex secret keys, which proves the system's exceptional
+
+<--- Page Split --->
+
+levels of security. Finally, smart device control is presented by the BSTCM system, realizing intelligent human- machine interactions. The proposed BSTCM can establish a new paradigm of human machine interactions, metasurface- based wireless communications, and potential applications in metaverse.  
+
+## System configuration  
+
+The proposed BSTCM is schematically demonstrated in Fig. 1, which consists of an STC metasurface, a brain- computer interface based on SSVEP, and an FPGA control module. The metasurface element contains a meta- structure to regulate the EM wave, and a light- emitting diode (LED) stimulator for the SSVEP paradigm. The STC metasurface is divided into four LED flickering partitions, operating at different frequencies: \(8.5\mathrm{Hz}\) , \(10\mathrm{Hz}\) , \(11.5\mathrm{Hz}\) , and \(7\mathrm{Hz}\) . The corresponding SSVEP- based EEG signals can be accurately extracted by a head- worn EEG cap when the operator is exposed to the LED flickering stimulus at distinct frequencies with the assistance of FPGA. To enhance the efficiency and accuracy of the EEG signal identification, we propose a deep learning model to recognize the extracted EEG signals, as shown in Fig. 2c. Upon completion of the identification process, the recognized signals are transmitted to FPGA, enabling to control the STC metasurface to update the corresponding STC matrices in the four LED flickering frequencies without disrupting the EEG stimulation. The BSTCM system empowers individuals to customize and modulate the EM waves by the human mind.  
+
+Leveraging the abundant harmonic modulation capabilities and the advantages of human brain intelligence, the BSTCM system can respectively implement a high- security encrypted wireless communication system and a smart device controlled by human mind. In the high- security encrypted wireless communications, the harmonic frequencies are adopted as the secret keys and the VSS method is employed to encrypt the transmitted information. Firstly, the information is encrypted into two visual secret ciphertexts that depend on the corresponding HSKs. As one of the interface devices in the metaverse, BCI plays a crucial role in enabling individuals to interact in the virtual world, and it requires high- security performance. The presented encrypted wireless communication is envisioned to fulfill the requirement for secure communication in the metaverse. For instance, in a metaverse scenario shown in Fig. 1, a legitimate transmitter (Alice) equipped with an EEG cap intentionally transfers the secret information to two legitimate receivers (Bob and Carol) at two harmonic frequencies based on the encrypted communication. The eavesdropper (Eve) cannot decrypt the transmitted
+
+<--- Page Split --->
+
+information unless she simultaneously obtains the secret keys, two ciphertexts, and encryption mechanisms, which is nearly impossible. Hence, the proposed system ensures high security and concealment by generating a large number of HSKs by modulating the STC metasurface. Ingeniously camouflaged as an LED stimulator, the STC metasurface effectively obstructs the eavesdroppers from detecting the information interaction. On the other hand, the presented system also enables wireless control of smart devices in real environments, allowing for manipulating the devices directly based on the user's brain intention without the requirement of physical actions. Hence the BSTCM can serve as a stimulator for SSVEP, and effectively manipulate the EM waves for secure BCI wireless communications.  
+
+## SSVEP signal recognition  
+
+The EEG signals can be accurately recognized in a BCI system through the detection of SSVEP components. The target flickering stimuli with four frequencies are supported with the STC metasurface integrated with LEDs, and the commands can be output by simply fixating on the corresponding visual stimuli. The STC metasurface is composed of \(32 \times 32\) elements, and segmented into four partitions, with each region consisting of \(16 \times 16\) elements. In these partitions, LEDs operate at four distinct frequencies (8.5Hz, 10Hz, 11.5Hz, and 7Hz) and are implemented to evoke four distinct SSVEP signals. As depicted in Fig. 2a, the SSVEP- BCI is segmented into three distinct stages: preparation stage, signal acquisition stage, and signal recognition stage. During the preparation stage, the participant wears an EEG cap equipped with electrodes positioned over the O1 and O2 regions and focuses the attention on one of the regions on the STC metasurface. In the signal acquisition stage, the participant maintains their gaze while an EEG amplifier collects the EEG data (for details, see Supplementary Note 1). Fig. 2d illustrates the frequency distribution of the EEG signals for the four target frequencies after the FFT analysis (prefiltered using a Butterworth bandpass filter with a frequency range of 5- 40Hz). As illustrated in Fig. 2b, the signal undergoes an initial pre- processing procedure during the signal recognition stage, involving signal decomposition, spectrum calculation and weighted summation, output signal amplitude spectrum (SigSpec). Subsequently, an outer product operation is executed with the pre- calculated reference spectra (RefSpec), yielding eight feature graphs (comprising of twochannel and four target frequencies), each sized by \(160 \times 160\) pixels. A discernible square grid pattern emerges in the feature graph corresponding to the frequency of the anticipated classification result (for additional information, refer to Supplemental Information Notes 2- 5). The conventional SSVEP classification and recognition algorithms include canonical correlation analysis (CCA) \(^{49}\) , filter bank canonical correlation
+
+<--- Page Split --->
+
+analysis (FBCCA) \(^{50}\) , task- related component analysis (TRCA) \(^{9}\) , support vector machine (SVM) \(^{51}\) , and convolutional neural network (CNN) \(^{52}\) . However, some limitations still exist in the above- mentioned algorithms. Here, we propose a deep learning classification algorithm that embeds the SSVEP components in the signal recognition stage (refer to Supplemental Information Note 6). The above- mentioned eight feature graphs are input into a lightweight CNN classification model containing four convolutional and two liner layers, as shown in Fig. 2c. The feature graphs are initially processed through a convolutional input layer. Subsequently, a MaxPooling operation with a \(2 \times 2\) kernel is applied following the activation of the data via the ReLU function. To facilitate the propagation of original data, a residual connection is strategically integrated between the second and third convolutional layers. Each convolutional layer is followed by the same \(2 \times 2\) MaxPooling operation and ReLU activation. Upon the extraction of features from the eight feature graphs, a series of two successive linear layers are deployed to categorize the features into four distinct target frequencies. Finally, an output layer with sigmoid activation is employed to reshape the output of the model into 4 confidence scores. Ultimately, the model achieved a classification accuracy of \(96.67\%\) on the validation set, with precision and recall rates surpassing \(90\%\) . Such results underscore the efficacy of the SSVEP component classification algorithm proposed in this work.  
+
+## Design of STC metasurface  
+
+Fig. 3a showcases the detailed configuration of the 1- bit STC metasurface element integrated with an LED, which consists of three metallic layers and multiple substrate layers. The top metallic layer has two slotted rectangular patches with a width of \(w = 7.3 \mathrm{mm}\) and a length of \(l = 11 \mathrm{mm}\) , in which the middle slot has a width of \(w_{1} = 1.5 \mathrm{mm}\) and a length of \(l_{1} = 7 \mathrm{mm}\) . The outer metal frame with a width of \(0.15 \mathrm{mm}\) is used to weaken the coupling effect between elements. The top F4B dielectric substrate has a relative permittivity of 2.65 and a loss tangent of 0.003 with the thickness \(h = 3 \mathrm{mm}\) and period \(p = 19 \mathrm{mm}\) . Two PIN diodes are serially loaded on the middle of two rectangular patches, and the two rectangular patches connected with metallic bars serve as positive and negative poles, in which two inductors are used as radio- frequency (RF) chokes to effectively isolate the RF signal from direct current (DC). An LED serving as the flash stimulation is placed outside the metallic patches to avoid affecting the scattering performance of the meta- atom and shares the same voltage with PIN diodes. To evaluate the performance of the 1- bit meta- atom, full- wave simulations are conducted using the commercial software CST Microwave Studio. As shown in Figs. 3b and 3c, the scattering characteristics of the meta- atom can be altered by switching the states of two PIN diodes
+
+<--- Page Split --->
+
+corresponding to distinct RLC equivalent circuit models, hence realizing the 1- bit phase regulation. The diodes with OFF state are assigned as coding “0”, and ON state as coding “1”. When the meta-atom is subjected to normal illumination by a \(y\) - polarized plane wave, a phase difference of \(180^{\circ}\) between “0” and “1” states is observed at approximately \(6.9\mathrm{GHz}\) , and both states exhibit reflective amplitudes exceeding - 1 dB, indicating excellent performance of the 1- bit programmable meta-atom. Hence the designed meta-atom can be used to build the 1- bit STC metasurface. The detailed prototype design, modeling and characterization are given in Methods and Supplementary Note 9.  
+
+To establish a dual- user encrypted wireless communication system, we implement an amplitude- shift keying modulation scheme using the STC metasurface, enabling simultaneous encrypted transmissions via space and frequency multiplexing. The control voltages, originating from FPGA, are applied to the PIN diodes, enabling periodic switching to the reflection phases of the 1- bit STC element with a time period of \(T_0 = 1 / f_0\) . Consequently, each STC element possesses a periodic time- coding sequence and constitutes part of an STC matrix, making it possible to flexibly control the EM waves in both space and time domains. The flexibility of STC matrices can generate a series of new harmonic spectra at the frequency interval of \(f_0\) , which are further optimized with specific requirements (see Supplementary Notes 7 and 8). The space- time modulation enables abundant harmonic characteristics to construct harmonic- based encrypted wireless communications. Using the 2D STC matrices, we can modulate the EM waves to steer to the desired directions for target users. The STC matrices encompass 16 rows of elements, each possessing periodic time- coding sequences with 11 intervals. As illustrative examples, three radiation patterns corresponding to three STC matrices (M0, M1, and M2) are shown in Fig. 3. The radiation pattern of the optimized STC matrix M1 in Fig. 3e is strategically manipulated to yield a main beam with a deflected angle of \(\theta = - 15^{\circ}\) at the \(+1\) st harmonic frequency, as showcased in Fig. 3h. Similarly, the radiation pattern of the STC matrix M2 (Fig. 3f) is optimized to direct the beam towards the angle \(\theta = +30^{\circ}\) , in correspondence with the - 1st harmonic frequency, as demonstrated in Fig. 3i. Furthermore, the time- invariant STC matrix M0 (Fig. 3d) facilitates the positioning of main beam towards a direction of \(\theta = 0^{\circ}\) , associated with the fundamental frequency, as illustrated in Fig. 3g. As an example of amplitude- shift keying modulation scheme, the left panels ( \(f_1 = 8.5\mathrm{Hz}\) and \(f_2 = 10\mathrm{Hz}\) ) of the STC metasurface are utilized to send digital symbols “0” and “1” to User1 at the position of \(\theta = - 15^{\circ}\) when the transmitting information is encoded as binary data streams. The right panels ( \(f_3 = 11.5\mathrm{Hz}\) and \(f_4 = 7\mathrm{Hz}\) ) of the STC metasurface are utilized to send digital
+
+<--- Page Split --->
+
+symbols “0” and “1” to User2 at the position of \(\theta = +30^{\circ}\) . Experiments are carried out in a microwave anechoic chamber to measure the far- field radiation patterns of the STC metasurface, as illustrated in Supplementary Notes 10 and 11.  
+
+Constrained by the requirement of the SSVEP- based BCI to gather data and detect EEG signals at a single frequency, a viable encoding strategy is proposed to enhance the symbol error rate (SER), as depicted in Fig. 3j- m. The transmission of the digital symbols “0/1” involves to encode a series of repeated data frames with a period of 4 frames as synchronous frames. In the encoding scheme, the first bit “1” serves as the start bit, signaling the initiation of the transmission process. The second bit signifies the harmonic frequency (+1st or - 1st) used for transmission, and the third bit corresponds to the transmission of the data itself. Finally, the fourth bit functions as the stop bit, indicating the end to transmit one data frame. Based on the encoding process, the digital symbol “0” or “1” for User1 is encoded as the repeated data frames “10001000” or “10101010...”, respectively. Similarly, the digital symbol “0/1” to User2 is encoded as repeated data frames “11001100...” or “11101110...”, respectively. Hence, the STC matrix M1 operating with a switching frequency \(f_0\) , is precisely injected into the region of high amplitude associated with the flicker frequencies \(f_1\) and \(f_2\) , as illustrated in Fig. 3j and 3k. The STC matrices are updated when the information is sent to User1 via the BSTCM system. Similarly, the STC matrix M2 with a period switch frequency \(f_0\) , undergoes injection into the high amplitude attributed to the flicker frequencies \(f_3\) and \(f_4\) , as displayed in Fig. 3j and 3k, representing the transmission of information to User2. Remarkably, the update of STC matrices does not affect the flickering frequency of LEDs used for SSVEP stimuli, ensuring that BCI signals can be synchronously detected and recognized.  
+
+## Secure encrypted wireless communication system  
+
+The conventional VSS method is employed to encrypt the individual pixels of confidential information, where each pixel is divided into two sub- pixels. It is worth noting that each share derived from this process remains devoid of any discernible information pertaining to the original image content. As a result, a secret message possesses two visual shared keys (VSKs), which effectively guarantee that the disclosure of any single VSK does not compromise the confidentiality of the original secret information. Here, we present a novel encryption strategy by combining the harmonic characteristics of STC metasurface with the VSS method, which is designed to realize high- security encrypted wireless communication directly controlled by the
+
+<--- Page Split --->
+
+human mind. In the presented encryption strategy, two VSKs are not simply stacked with each other to recover the original information, but rely on HSKs. In the described encryption process, the confidential information to be transmitted is firstly encrypted to two random VSKs based on the presented encrypted encoding scheme. Then, two generated VSKs are further encoded into two random HSKs for the \(\pm 1\) harmonic frequencies, respectively. During decryption, two VSKs attached to each secret can be reconstructed with the pre- customized variant HSK sequences. The details of encryption and decryption methods can be found in Supplementary Note 12. The target image in question can be decrypted through the incoherent superposition of two extracted VSKs, aided by the presence of HSKs. By imbuing addressable shared information bits with the harmonic dynamics, a robust level of information security is achieved. As an illustrative example in Fig. 4a, the image “H”, which consists of \(5 \times 5\) grid of black or white pixels, is encrypted into two random VSK1 and VSK2 based on the presented encryption method. Two bitstream sequences of VSKs can be successively transmitted to two designated Users (Bob and Carol) by the BCI operator (Alice) using the presented BSTCM system.  
+
+To validate the encrypted communication scheme, we experimentally set up an encrypted wireless communication system using the BSTCM, as demonstrated in Fig. 4b. The prototype of 1- bit STC metasurface is used to construct the encrypted communication system, as shown in Supplementary Figure 15. A commercial BCI device (actiCHamp of Brain Product GmbH company) composed of an EEG cap and amplifier is used to acquire the EEG signals when the operator stares at different partitions on the STC metasurface. The corresponding EEG signals can be accurately recognized by the proposed deep learning algorithm. The detailed description of the algorithm is provided in Supplementary Note 6. A linearly polarized horn antenna, connected to a signal generator (Keysight E8267D), is placed at a vertical distance of 4 m away from the metasurface center to excite a monochromatic plane wave at the frequency of 6.9 GHz. The receiving terminals are composed of two horn antennas (severed as two users) and a spectrum analyzer (Keysight N9040D), which is used to demodulate the received signal. Two receiving horn antennas are located at the angles of \(\theta = - 15^{\circ}\) and \(\theta = +30^{\circ}\) with relative to the metasurface normal, respectively. In the transmission process, the information undergoes encryption, resulting in the generation of two harmonic- based cipher texts according to the
+
+<--- Page Split --->
+
+encryption scheme. The transmission process is achieved by ensuring that the BCI operator maintains direct visual focus on the STC metasurface in alignment with the binary cipher texts. In addition, the system enables prompt acquisition, recognition, and translation of EEG signals into appropriate control signals of FPGA, which in turn drive the STC metasurface. As a result, the accurate transmission of binary cipher texts is accomplished, facilitating reliable and efficient communications between the BCI operator and the intended recipients.  
+
+In the receiving process, the radiated signals propagating through free space are received individually by the two horn antennas, and these signals are subsequently demodulated using the spectrum analyzer. As shown in Fig. 4c, the demodulated signals represent four encoding schemes that correspond to different digital symbols ("0" or "1") for the two users and then are processed to the decoded results shown in Fig. 4d by the FFT method. The decoded results clearly exhibit four repeated data frames, providing strong evidence to support the capability of the proposed system to recover the transmitting information successfully. Hence two binary cipher texts are transmitted to two specific users by the BCI operator based on the proposed wireless communication system. The transmitted information is received and the measured results are comprehensively showcased in Supplementary Note 13 and Supplementary Video 1. By employing the encoding method described earlier, the data streams corresponding to the specific users are extracted from the received signals, as depicted in Figs. 4e- f. Subsequently, these data streams are decrypted accurately by the method outlined in Supplementary Note 12. The deciphered information is displayed in Fig. 4g. The successful decryption and retrieval of the transmitted information affirm the robustness and efficacy of the proposed system in maintaining high levels of security and confidentiality. The system is not easily deciphered or cracked, proving its ability to protect sensitive information during transmissions.  
+
+## Mind control to smart devices  
+
+To experimentally verify the performance of the proposed BSTCM system, we conducted experiments to show the remote control capabilities of smart devices through human intention, which holds significant potential to enhance the quality of life for individuals with disabilities.
+
+<--- Page Split --->
+
+The system flowchart depicted in Fig. 5a outlines the key steps involved in the experimental setup. Initially, the EEG signals are acquired and subsequently processed by the BSTCM system for command recognition. The recognized commands are then translated to update the STC matrices of metasurface by an FPGA. The four partitions of the STC metasurface are controlled by FPGA to generate different beams at four deflected angles corresponding to the four harmonic frequencies, enabling the remote control of smart devices. Notably, the main beam at the - 2nd harmonic frequency exhibits a pronounced peak in the direction of \(\theta = +10^{\circ}\) , corresponding to the device assigned to User 3. Conversely, the main beam at the +2nd harmonic frequency demonstrates a strong peak at \(\theta = - 45^{\circ}\) , corresponding to the device assigned to User 1. As displayed in Figs. 5c and d, the measured radiation patterns exhibit good consistency with theoretical predictions associated to the corresponding STC matrices. Additionally, the main beams were deflected at the other two angles (- 15° and 30°) using the +1st and - 1st harmonic frequencies (Users 2 and 4), as shown in Figs. 3h and i. The excellent harmonic beam manipulation of the STC metasurface lays a solid theory foundation for smart device control. The receiving terminals are composed of four horn antennas, four RF energy detectors (LMH2110), four Microcontroller Unit (MCU) modules (Arduino), and four LED modules (serve as smart devices), as depicted in Fig. 5b. The RF energy from harmonic beams captured by the detectors can be converted into DC voltages, which are subsequently fed into the MCU modules. Once the corresponding DC voltage is detected, the MCU modules send the control commands to activate the corresponding LED module, thereby implementing the remote control of the smart devices. The experimental setup is shown in Fig. 5b, where the STC metasurface is illuminated by a monochromatic frequency signal of 6.9GHz.  
+
+To verify the robustness of the brain- control smart devices system, the position of the receiving antenna is sequentially placed at different distances from the metasurface center while maintaining specific harmonic direction angles with respect to the metasurface normal. The detection of RF energy from the emitted harmonic beams is carried out across varying distances and the corresponding DC voltages can be accurately converted by the RF detectors, as illustrated in Fig. 5e (see Supplementary Note 14 for details). We observe that the RF energies of four users relatively decrease and the converted DC voltages also decrease accordingly as
+
+<--- Page Split --->
+
+the distance increases. To guarantee the attainment of the far- field region and maintain system stability, the four receiving antennas are positioned at a distance of approximately \(300\mathrm{cm}\) from the metasurface. Consequently, the detection voltage threshold is set as \(0.15\mathrm{V}\) , \(0.4\mathrm{V}\) , \(0.4\mathrm{V}\) , and \(0.2\mathrm{V}\) in MCU modules, respectively. We recorded a video of the brain- controlled multiple smart devices, as presented in Supplementary Movie 2. The sequence diagrams shown in Fig. 5f provide a clear visual representation of the sequential control process, where the smart devices (LED modules) are successively controlled through human intention using the BSTCM system. The experiment proves the feasibility of brain- controlled smart devices. By harnessing the power of BCI technology, individuals can exert control over smart devices without needing physical interaction. The BSTCM system enables seamless and intuitive manipulations of various smart devices, facilitating greater independence and convenience for people in daily life. The experimental results serve as a compelling validation supporting the significant benefits that the BSTCM can offer in improving the quality of life for individuals facing physical challenges.  
+
+## Conclusions  
+
+We presented a BSTCM system, which combines the human brain intelligence and flexible EM control capabilities of the STC metasurface. A series of visual stimuli encoded by specific coding schemes was employed on the metasurface to realize deep information processing and interaction in the EM field for reliable and stable communication environments. In addition, a lightweight deep- learning algorithm was proposed to accurately recognize the brain signals extracted by the BCI device. We further demonstrated a high- security encrypted wireless communication system based on diverse harmonic frequency characteristics, which combines the variant HSKs and VSS methods. The proposed system can enable the system to have high security. Finally, remote control of smart devices was demonstrated by the BSTCM system, realizing intelligent human- machine interactions. The proposed BSTCM, integrating the EM manipulation capability with the intelligence of the human brain, provides a new paradigm of interactions among human machine interface in metaverse, and future 6G wireless communication applications.
+
+<--- Page Split --->
+
+## Methods and materials  
+
+Details on STC metasurface. A metasurface prototype was fabricated by the standard printed circuit board (PCB) process. The entire STC metasurface is composed of \(32 \times 32\) digital elements and an effective size of \(608 \times 608 \mathrm{~mm}^2\) . For the convenience of processing and welding, the entire metasurface was disassembled into 8 sub- metasurfaces with \(8 \times 16\) digital elements and an effective size of \(152 \times 304 \mathrm{~mm}^2\) . A total number of \(2 \times 32 \times 32\) PIN diodes (SMP1320- 079LF from SKYWORKS) and RF inductors of \(10 \mathrm{nH}\) were elaborately welded on the metasurfaces with the surface mount technology, which is a mature engineering method to weld small components on PCB using the batch solder- reflow processes in a dedicated machine. Meanwhile, an LED is embedded into each metasurface element.  
+
+The brain signal acquisition. As depicted in Fig. 2a, the SSVEP- BCI is delineated into three discrete distinct stages: the preparation stage, the signal acquisition stage, and the signal recognition stage. The EEG data in this study were acquired using the actiCHamp EEG amplifier manufactured by Brain Products GmbH. This amplifier offers a maximum capacity of 64 channels, 24- bit precision, and a sampling rate of \(100 \mathrm{kHz}\) . It exhibits a high common mode rejection ratio (CMRR) of 100, making it suitable for capturing and detecting SSVEP signals effectively. This study uses a standard 10- 20 EEG cap to obtain the signal. Since SSVEP is a local potential that predominantly originates from the occipital lobe \(^5\) , the O1 and O2 electrodes located on the occipital region were selected for EEG data acquisition, with a reference electrode placed at the vertex (Cz) region. The sample time in this study was set to four seconds. With a sampling rate of \(512 \mathrm{~Hz}\) and excluding the reference electrode data, each sample yielded 4096 sampling points (4 seconds \(\times 512 \mathrm{~Hz} \times 2\) channels). Square wave signals were utilized to present visual stimuli on the LEDs of the STC metasurface units, which emitted the red light, thereby ensuring adequate visual contrast. To minimize the harmonic interference, four target visual stimulus frequencies were: top- left at \(8.5 \mathrm{~Hz}\) , bottom- left at \(10 \mathrm{~Hz}\) , top- right at \(11.5 \mathrm{~Hz}\) , and bottom- right at \(7 \mathrm{~Hz}\) . To mitigate mutual interference between two closely spaced flickering stimuli pairs, the frequencies \(7 \mathrm{~Hz}\) and \(8.5 \mathrm{~Hz}\) , as well as \(10 \mathrm{~Hz}\) and \(11.5 \mathrm{~Hz}\) ,
+
+<--- Page Split --->
+
+were placed diagonally opposite to one another. The example of the collected brain wave signal data for four target frequencies is shown in Supplementary Figure 1. Data processing and model training were conducted in Python and PyTorch environments. Signal processing primarily involved utilizing scipy package for filtering and FFT calculations. Given that neither the absolute value of the EEG signal nor its temporal characteristics were pertinent to this study, all sampled signals were linearly transformed to have a mean of zero and a standard deviation of one.  
+
+Classification model training. Initially, a dataset comprising 240 instances of raw SSVEP signals was collected from two participants. Subsequently, a subset consisting of 80 instances of randomly selected raw signals was employed to create RefSpec. The remaining 160 instances of raw signals were randomly selected to serve as training and validation datasets for classification model training. The two subsets have an equally divided number of signals across four stimuli frequencies. As for the training and validation dataset, a subset of 60 instances was extracted from the initial subset of 160 instances to serve as the validation set. The remaining 100 instances underwent a data augmentation process, which resulted in an expanded dataset of 1700 instances. These augmented instances constituted the training set, which was then utilized to train the model for 300 epochs. The validation operation was executed every two training epochs.  
+
+## Acknowledgements  
+
+The work is supported by the National Natural Science Foundation of China (62288101, 92167202, 72171044), the Major Project of Natural Science Foundation of Jiangsu Province (BK20212002), the State Key Laboratory of Millimeter Waves, Southeast University, China (K201924), the Fundamental Research Funds for the Central Universities (2242018R30001), the 111 Project (111- 2- 05), the China Postdoctoral Science Foundation (2021M700761), and ZhiShan Scholar Program of Southeast University (2242022R40004).  
+
+## Author contributions
+
+<--- Page Split --->
+
+T.J.C. suggested the designs and planned and supervised the work, in consultation with Q.M., Y.F.N., Q.X., and J.W.Y. conceived the idea and carried out the theoretical analysis and numerical simulations. Q.X., L.H.F., Y.M.N., and Z.G. built the system and performed the experimental measurements. Q.X., L.H.F. and L.C. performed the data analysis. Q.X. and L.H.F. wrote the manuscript. Q.M., L.L., J.W. Y, and T.J.C. reviewed the manuscript. All authors discussed the theoretical aspects and numerical simulations, interpreted the results and reviewed the manuscript.  
+
+## Competing interests  
+
+The authors declare no competing financial interest.  
+
+## Data availability  
+
+The data that support the findings of this study are available from the corresponding author upon request.  
+
+## Code availability  
+
+The code that supports the findings of this study are available from the corresponding author upon reasonable request.  
+
+## References  
+
+Wu, D. et al. Virtual-Reality Interpromotion Technology for Metaverse: A Survey. IEEE Internet of Things Journal 10, 15788- 15809, doi:10.1109/jiot.2023.3265848 (2023). Park, S., Ha, J., Park, J., Lee, K. & Im, C.- H. Brain- Controlled, AR- Based Home Automation System Using SSVEP- Based Brain- Computer Interface and EOG- Based Eye Tracker: A Feasibility Study for the Elderly End User. IEEE Transactions on Neural Systems and Rehabilitation Engineering 31, 544- 553, doi:10.1109/tnsre.2022.3228124 (2023). Al- Saegh, A., Dawwd, S. A. & Abdul- Jabbar, J. M. Deep learning for motor imagery EEG- based classification: A review. Biomedical Signal Processing and Control 63, doi:10.1016/j.bspc.2020.102172 (2021). Lenhardt, A., Kaper, M. & Ritter, H. J. An Adaptive P300- Based Online Brain- Computer Interface. IEEE Transactions on Neural Systems and Rehabilitation Engineering 16, 121- 130, doi:10.1109/tnsre.2007.912816 (2008). Vialatte, F.- B., Maurice, M., Dauwels, J. & Cichocki, A. Steady- state visually evoked potentials: Focus on essential paradigms and future perspectives. Progress in Neurobiology 90, 418- 438, doi:10.1016/j.pneurobio.2009.11.005 (2010). Niu, Y. et al. Improving SSVEP- BCI System Interaction Efficiency: Design Recommendations for Shape of Visual Stimuli and Number of Auxiliary Stimuli. International Journal of Human- Computer Interaction, 1- 22, doi:10.1080/10447318.2023.2188540 (2023).
+
+<--- Page Split --->
+
+464 7 Zhao, X., Du, Y. & Zhang, R. A CNN- based multi- target fast classification method for AR- SSVEP. 465 Computers in Biology and Medicine 141, doi:10.1016/j.compbiomed.2021.105042 (2022). 466 Ravi, A., Lu, J., Pearce, S. & Jiang, N. Enhanced System Robustness of Asynchronous BCI in 467 Augmented Reality Using Steady- State Motion Visual Evoked Potential. IEEE Transactions on 468 Neural Systems and Rehabilitation Engineering 30, 85- 95, doi:10.1109/tnsre.2022.3140772 (2022). 469 Nakanishi, M. et al. Enhancing Detection of SSVEPs for a High- Speed Brain Speller Using Task- 470 Related Component Analysis. IEEE Transactions on Biomedical Engineering 65, 104- 112, 471 doi:10.1109/tbme.2017.2694818 (2018). 472 Bin, G., Gao, X., Wang, Y., Hong, B. & Gao, S. VEP- based brain- computer interfaces: time, 473 frequency, and code modulations [Research Frontier. IEEE Computational Intelligence Magazine 4, 474 22- 26, doi:10.1109/mci.2009.934562 (2009). 475 Li, Y. Q., Pan, J. H., Wang, F. & Yu, Z. L. A Hybrid BCI System Combining P300 and SSVEP and 476 Its Application to Wheelchair Control. Ieee Transactions on Biomedical Engineering 60, 3156- 3166, 477 doi:10.1109/tbme.2013.2270283 (2013). 478 Muller- Putz, G. R. & Pfurtscheller, G. Control of an Electrical Prosthesis With an SSVEP- Based 479 BCI. IEEE Transactions on Biomedical Engineering 55, 361- 364, doi:10.1109/tbme.2007.897815 480 (2008). 481 Willett, F. R., Avansino, D. T., Hochberg, L. R., Henderson, J. M. & Shenoy, K. V. High- performance 482 brain- to- text communication via handwriting. Nature 593, 249- 254, doi:10.1038/ s41586- 021- 483 03506- 2 (2021). 484 Roy, A. M. Adaptive transfer learning- based multiscale feature fused deep convolutional neural 485 network for EEG MI multiclassification in brain- computer interface. Engineering Applications of 486 Artificial Intelligence 116, doi:10.1016/j.engappai.2022.105347 (2022). 487 Ma, Q. et al. Directly wireless communication of human minds via non- invasive brain- computer- 488 metasurface platform. eLight 2, doi:10.1186/s43593- 022- 00019- x (2022). 489 Wandelt, S. K. et al. Representation of internal speech by single neurons in human supramarginal 490 gyrus. Nature Human Behaviour, doi:10.1038/s41562- 024- 01867- y (2024). 491 Xiao, Q. et al. Electromagnetic Brain- Computer- Metasurface Holography. ACS Photonics, 492 doi:10.1021/acsphotonics.2c01349 (2023). 493 Zhu, R. et al. Remotely Mind- Controlled Metasurface via Brainwaves. eLight 2, 10 (2022). 494 Jiang, W., Han, B., Habibi, M. A. & Schotten, H. D. The Road Towards 6G: A Comprehensive Survey. 495 IEEE Open Journal of the Communications Society 2, 334- 366, doi:10.1109/ojcoms.2021.3057679 496 (2021). 497 Saad, W., Bennis, M. & Chen, M. A Vision of 6G Wireless Systems: Applications, Trends, 498 Technologies, and Open Research Problems. IEEE Network 34, 134- 142, doi:10.1109/mnet.001. 499 1900287 (2020). 500 Zhang, Z. et al. 6G Wireless Networks: Vision, Requirements, Architecture, and Key Technologies. 501 IEEE Vehicular Technology Magazine 14, 28- 41, doi:10.1109/mvt.2019.2921208 (2019). 502 Dang, S., Amin, O., Shihada, B. & Alouini, M.- S. What should 6G be? Nature Electronics 3, 20- 29, 503 doi:10.1038/s41928- 019- 0355- 6 (2020). 504 Brocal, F. Brain- computer interfaces in safety and security fields: Risks and applications. Safety 505 Science 160, doi:10.1016/j.ssci.2022.106051 (2023). 506 Bhalerao, S., Ansari, I. A. & Kumar, A. Protection of BCI system via reversible watermarking of 507 EEG signal. Electronics Letters 56, 1389- 1392, doi:10.1049/el.2020.2532 (2020).
+
+<--- Page Split --->
+
+509 25 Bernal, S. L., Celdrán, A. H., Pérez, G. M., Barros, M. T. & Balasubramaniam, S. Security in Brain- 509 Computer Interfaces. ACM Computing Surveys 54, 1- 35, doi:10.1145/3427376 (2021). 510 26 Wei, M., Zhao, H., Galdi, V., Li, L. & Cui, T. J. Metasurface- enabled smart wireless attacks at the 511 physical layer. Nature Electronics 6, 610- 618, doi:10.1038/s41928- 023- 01011- 0 (2023). 512 27 Pendry, J. B. Negative refraction makes a perfect lens. Phys Rev Lett 85, 3966- 3969, 513 doi:10.1103/PhysRevLett.85.3966 (2000). 514 28 Pendry, J. B., Schurig, D. & Smith, D. R. Controlling Electromagnetic Fields. Science 312, 1780- 515 1782, doi:10.1126/science.1125907 (2006). 516 29 Liu, R. et al. Broadband ground- plane cloak. Science 323, 366- 369 (2009). 517 30 Smith, D. R. & Kroll, N. Negative refractive index in left- handed materials. Phys Rev Lett 85, 2933- 518 2936, doi:10.1103/PhysRevLett.85.2933 (2000). 519 31 Cui, T. J., Qi, M. Q., Wan, X., Zhao, J. & Cheng, Q. Coding metamaterials, digital metamaterials 520 and programmable metamaterials. Light: Science & Applications 3, e218- e218, doi:10.1038/lsa. 521 2014.99 (2014). 522 32 Zhang, L. et al. Space- Time- Coding Digital Metasurfaces. Nat. Commun. 9, 4334 (2018). 523 33 Wu, H. et al. Harmonic information transitions of spatiotemporal metasurfaces. Light Sci Appl 9, 524 198, doi:10.1038/s41377- 020- 00441- 1 (2020). 525 34 Wu, G.- B., Dai, J. Y., Cheng, Q., Cui, T. J. & Chan, C. H. Sideband- free space- time- coding 526 metasurface antennas. Nature Electronics 5, 808- 819, doi:10.1038/s41928- 022- 00857- 0 (2022). 527 35 Zhang, L. et al. Co- Prime Modulation for Space- Time- Coding Digital Metasurfaces with Ultralow- 528 Scattering Characteristics. Advanced Functional Materials, doi:10.1002/adfm.202314110 (2024). 529 36 Zhang, L. et al. A wireless communication scheme based on space- and frequency- division 530 multiplexing using digital metasurfaces. Nature Electronics 4, 218- 227, doi:10.1038/s41928- 021- 531 00554- 4 (2021). 532 37 Shaltout, A. M., Shalaev, V. M. & Brongersma, M. L. Spatiotemporal light control with active 533 metasurfaces. Science 364, doi:10.1126/science.aat3100 (2019). 534 38 Wu, G.- B. et al. A universal metasurface antenna to manipulate all fundamental characteristics of 535 electromagnetic waves. Nature Communications 14, doi:10.1038/s41467- 023- 40717- 9 (2023). 536 39 Zhao, J. et al. Programmable time- domain digital- coding metasurface for non- linear harmonic 537 manipulation and new wireless communication systems. National Science Review 6, 231- 238, 538 doi:10.1093/nsr/nwy135 (2019). 539 40 Dai, J. Y. et al. Wireless Communication Based on Information Metasurfaces. IEEE Transactions on 540 Microwave Theory and Techniques 69, 1493- 1510, doi:10.1109/tmtt.2021.3054662 (2021). 541 41 Dai, J. Y. et al. Realization of Multi- Modulation Schemes for Wireless Communication by Time- 542 Domain Digital Coding Metasurface. IEEE Transactions on Antennas and Propagation 68, 1618- 543 1627, doi:10.1109/tap.2019.2952460 (2020). 544 42 Dai, J. Y. et al. Wireless Communications through a Simplified Architecture Based on Time- Domain 545 Digital Coding Metasurface. Advanced Materials Technologies 4, doi:10.1002/admt. 201900044 546 (2019). 547 43 Zhang, L. et al. A Wireless Communication Scheme Based on Space- and Frequency- Division 548 Multiplexing Using Digital Metasurfaces. Nat. Electron. 4, 218 (2021). 549 44 Tang, W. et al. Wireless Communications with Programmable Metasurface: New Paradigms, 550 Opportunities, and Challenges on Transceiver Design. IEEE Wireless Communications 27, 180- 187, 551 doi:10.1109/mwc.001.1900308 (2020).
+
+<--- Page Split --->
+
+552 45 Tang, W. et al. Programmable metasurface- based RF chain- free 8PSK wireless transmitter. Electronics Letters 55, 417- 420, doi:10.1049/el.2019.0400 (2019). 553 46 Chen, T.- H. & Tsao, K.- H. Visual secret sharing by random grids revisited. Pattern Recognition 42, 2203- 2217, doi:10.1016/j.patcog.2008.11.015 (2009). 555 47 Yang, C.- N. New visual secret sharing schemes using probabilistic method. Pattern Recognition Letters 25, 481- 494, doi:10.1016/j.patrec.2003.12.011 (2004). 556 48 Wang, D., Yi, F. & Li, X. Probabilistic visual secret sharing schemes for grey- scale images and color images. Information Sciences 181, 2189- 2208, doi:10.1016/j.ins.2011.01.019 (2011). 557 49 Bin, G., Gao, X., Yan, Z., Hong, B. & Gao, S. An online multi- channel SSVEP- based brain- computer interface using a canonical correlation analysis method. Journal of Neural Engineering 6, doi:10.1088/1741- 2560/6/4/046002 (2009). 558 50 Chen, X., Wang, Y., Gao, S., Jung, T.- P. & Gao, X. Filter bank canonical correlation analysis for implementing a high- speed SSVEP- based brain- computer interface. Journal of Neural Engineering 12, doi:10.1088/1741- 2560/12/4/046008 (2015). 559 51 Singla, R., Khosla, A. & Jha, R. Influence of stimuli colour in SSVEP- based BCI wheelchair control using support vector machines. Journal of Medical Engineering & Technology 38, 125- 134, doi:10.3109/03091902.2014.884179 (2014). 560 52 Ding, W. et al. Filter Bank Convolutional Neural Network for Short Time- Window Steady- State Visual Evoked Potential Classification. IEEE Transactions on Neural Systems and Rehabilitation Engineering 29, 2615- 2624, doi:10.1109/tnsre.2021.3132162 (2021).
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Fig. 1 | The schematical diagram of the presented BSTCM. The BSTCM system integrates STC metasurface into SSVEP-based BCI systems. The STC metasurface can correctly support the visual stimuli for the SSVEP-based BCI and facilitate the information interaction with the external environment. Based on the BSTCM system, high-security encrypted wireless communication systems are realized for the first time by combining with the variant HSKs and VSS method. In this way, smart devices can be controlled by the human mind with high security. </center>
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2 | The diagram of interaction and data processing procedures of the SSVEP-based BCI system. a. The interaction procedures for the SSVEP-based BCI involve the user wearing an EEG cap and focusing their attention on the intended target LED stimuli. Once the user's selected frequency is recognized, the interaction commands are transmitted to the STC metasurface. b. The feature extraction process includes the transformation of the input BCI signals into the frequency domain. The resulting Signal FFT is then subjected to outer product with pre-defined Template FFTs, yielding eight feature graphs. c. The lightweight classification model. Four convolutional and two fully connected layers were employed. d. An example of frequency responses of SSVEP signals for the four target frequencies, showing distinctive amplitude peaks corresponding to different frequencies. </center>
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Fig. 3 | The details of the STC metasurface. a. The geometrical structure of the STC metasurface element. b, c. The reflective amplitude and phase of the element, respectively. d-f. The optimized STC matrices for different scattering angles at different harmonic frequencies. g-i. The far-field results corresponding to the optimized STC matrices. j-m. The encoding schemes for the transmitting symbol “0” or “1” for two users: the symbols “0” and “1” for Users 1 (j and k); the symbols “0” and “1” for Users 2 (l and m). </center>
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Fig. 4 | The encrypted wireless communication system based on the BSTCM platform. a. The encrypted encoding scheme combining the harmonic-secret keys and the VSS method. b. The experimental scenarios of the encrypted wireless communication system, in which the operator equipped with EEG cap transmits the encrypted information to two users via the BSTCM platform, respectively. c. The receiving signals corresponding to the encoding scheme in Fig 3(j-m). d, e. The decoded information of VSK1 and VSK2 from the receiving signals based on the encrypted coding scheme. f. The correct transmitting information is decoded by the extracted VSK1 and VSK2. </center>
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Fig. 5 | Experiment on the wireless remote mind control of smart devices based on the BSTCM platform. a. The system architecture of wireless remote control. b. The experimental scenarios of the wireless remote-control system by the human brain. c, d. The theoretical and experimental results for far-filed patterns at \(\pm 2\mathrm{th}\) harmonic frequencies. f. The temporal waveform of the output voltages from four detectors when the operator lights up four devices in sequence. e. The variations between the input power of the detectors and output voltage with respect to the different distances between the receiving antenna and the metasurface center. </center>
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryMaterials.pdf SupplementaryVideos.zip
+
+<--- Page Split --->

preprint/preprint__0996185957de22ec96c13ddd4eeca9080a098f778fc62996a2333ee389185663/preprint__0996185957de22ec96c13ddd4eeca9080a098f778fc62996a2333ee389185663_det.mmd

@@ -0,0 +1,304 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 952, 208]]<|/det|>
+# Secure wireless communication of brain-computer interface and mind control of smart devices enabled by space-time-coding metasurface  
+
+<|ref|>text<|/ref|><|det|>[[44, 230, 235, 275]]<|/det|>
+Tie Jun Cui  t.jcui@seu.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[44, 303, 598, 323]]<|/det|>
+Southeast University https://orcid.org/0000- 0002- 5862- 1497  
+
+<|ref|>text<|/ref|><|det|>[[44, 328, 238, 368]]<|/det|>
+Qiang Xiao  Southeast University  
+
+<|ref|>text<|/ref|><|det|>[[44, 374, 238, 414]]<|/det|>
+Lin Han Fan  Southeast University  
+
+<|ref|>text<|/ref|><|det|>[[44, 420, 598, 460]]<|/det|>
+Qian Ma  Southeast University https://orcid.org/0000- 0002- 4662- 8667  
+
+<|ref|>text<|/ref|><|det|>[[44, 466, 238, 506]]<|/det|>
+Yu Ming Ning  Southeast University  
+
+<|ref|>text<|/ref|><|det|>[[44, 512, 238, 551]]<|/det|>
+Ze Gu  Southeast University  
+
+<|ref|>text<|/ref|><|det|>[[44, 558, 598, 598]]<|/det|>
+Long Chen  Southeast University https://orcid.org/0009- 0007- 1533- 0319  
+
+<|ref|>text<|/ref|><|det|>[[44, 604, 565, 644]]<|/det|>
+Lianlin Li  peking university https://orcid.org/0000- 0001- 9394- 3638  
+
+<|ref|>text<|/ref|><|det|>[[44, 650, 598, 690]]<|/det|>
+Jian Wei You  Southeast University https://orcid.org/0000- 0001- 5761- 9507  
+
+<|ref|>text<|/ref|><|det|>[[44, 696, 238, 736]]<|/det|>
+Ya Feng Niu  Southeast University  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 780, 104, 798]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 818, 933, 860]]<|/det|>
+Keywords: Space- time- coding metasurface, brain- computer interface, secure wireless communication, human- machine interactions  
+
+<|ref|>text<|/ref|><|det|>[[44, 879, 322, 898]]<|/det|>
+Posted Date: August 23rd, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 916, 475, 935]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4860006/v1
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 44, 916, 87]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 105, 535, 125]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 160, 936, 204]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on August 25th, 2025. See the published version at https://doi.org/10.1038/s41467-025-63326-0.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[67, 88, 872, 140]]<|/det|>
+# Secure wireless communication of brain-computer interface and mind control of smart devices enabled by space-time-coding metasurface  
+
+<|ref|>text<|/ref|><|det|>[[113, 155, 860, 195]]<|/det|>
+Qiang Xiao \(^{1,\dagger}\) , Lin Han Fan \(^{2,\dagger}\) , Qian Ma \(^{1,\dagger}\) , \(*\) , Yu Ming Ning \(^{1}\) , Ze Gu \(^{1}\) , Long Chen \(^{1}\) , Lianlin Li \(^{3}\) , Jian Wei You \(^{1}\) , Ya Feng Niu \(^{2,\ast}\) and Tie Jun Cui \(^{1,\ast}\)  
+
+<|ref|>text<|/ref|><|det|>[[115, 215, 872, 255]]<|/det|>
+\(^{1}\) State Key Laboratory of Millimeter Waves and Institute of Electromagnetic Space, Southeast University, Nanjing 210096, China  
+
+<|ref|>text<|/ref|><|det|>[[115, 261, 707, 279]]<|/det|>
+\(^{2}\) School of Mechanical Engineering, Southeast University, Nanjing 210096, China  
+
+<|ref|>text<|/ref|><|det|>[[115, 284, 832, 302]]<|/det|>
+\(^{3}\) State Key Laboratory of Advanced Optical Communication Systems and Networks, Department of  
+
+<|ref|>text<|/ref|><|det|>[[115, 307, 505, 323]]<|/det|>
+Electronics, Peking University, 100871 Beijing, China  
+
+<|ref|>text<|/ref|><|det|>[[115, 329, 582, 345]]<|/det|>
+\*E- mail: maqian@seu.edu.cn, nyf@seu.edu.cn, tjcui@seu.edu.cn  
+
+<|ref|>text<|/ref|><|det|>[[115, 352, 456, 368]]<|/det|>
+\(^{\dagger}\) These authors contributed equally to this work  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 400, 195, 415]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[113, 421, 884, 802]]<|/det|>
+Brain- computer interface (BCI) provides an interconnected pathway between human brain and external devices and paves a potential route for mind manipulations. However, most of existing BCI technologies are based on simple signal transmission and independent of other interface devices, owing to the considerations of reliability and safety of human brain's information interaction in the complicated wireless environment. To address the formidable limitation, we present a brain space- time- coding metasurface (BSTCM) system to deeply fuse the visual stimulation and electromagnetic manipulation for reliable and secure information transfer between human brain and external devices. Here, we innovatively integrate the BCI flashing frame and electromagnetic encoding sequence in the BSTCM system, and the STC metasurface ensures the secure wireless communications by using the harmonic- encrypted beams. We design and fabricate a proof- of- principle demonstration system and experimentally show that the proposed wireless BCI scheme could establish a remote but safeguarded paradigm for human- machine interactions, intelligent metasurfaces, and potential applications in metaverse, as a prominently scrutinized domain in the future 6G wireless communications.  
+
+<|ref|>text<|/ref|><|det|>[[115, 842, 881, 880]]<|/det|>
+KEYWORDS: Space- time- coding metasurface, brain- computer interface, secure wireless communication, human- machine interactions
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 93, 247, 112]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[113, 123, 885, 505]]<|/det|>
+Brain- computer interface (BCI) has emerged as a cutting- edge technology in human- machine interaction, and demonstrates promising applications such as metaverse interaction \(^1\) and smart homes \(^2\) . Electroencephalography (EEG) signals remain the predominant input signal modality in BCI systems, with notable implementations including motor imagery \(^3\) , P300 \(^4\) , and steady- state visually evoked potential (SSVEP) \(^5\) . The SSVEP- based BCI systems utilize the brain's SSVEP response to the fixed- frequency visual stimuli for mind recognition and interaction \(^6\) , offering a significant advantage of high information transfer rate (ITR). The BCIs have been widely used in various fields such as augmented reality (AR) \(^7,8\) , spelling input \(^9,10\) , medical rehabilitation and equipment control \(^11,12\) . Recently, some high- performance BCI systems have been continuously proposed \(^13,18\) . The advent and progression of 6G wireless communication technology substantially broaden the application prospects of BCIs, particularly in Internet of Things (IoT), virtual reality (VR) devices, and beyond \(^19,21\) . Hence, ensuring high security and privacy preservation \(^22\) becomes imperative when facilitating intelligent interactions within the constructed communication environment.  
+
+<|ref|>text<|/ref|><|det|>[[113, 514, 885, 895]]<|/det|>
+However, most existing BCI systems lack in- depth research in terms of security. During the wireless transmissions of brain signals in a BCI system, there exists a vulnerability to theft and attack, potentially leading to the generation of inaccurate device control commands and unauthorized disclosure of personal privacy. Though some methodologies have been proposed to enhance the security and privacy of data in the BCI systems \(^{23 - 25}\) , there is a conspicuous absence to investigate the encryption mechanisms specifically tailored for the BCI systems. More importantly, the implementation of visual stimulation remains to be isolated from the back- end information processing systems, lacking deep information fusion and interaction. With the increasing demand for high security in the BCI systems, it is essential to develop intelligent interactions in the secure and reliable communication environment. The frequency- dependent SSVEP response and programmable harmonic characteristics of space- time- coding (STC) metasurfaces display notable similarities. Therefore, the STC metasurfaces can be used as a promising method that not only provides the visual stimulation but also ensures security of the BCI systems at the physical layer owing to their powerful ability to flexibly manipulate
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 629, 107]]<|/det|>
+electromagnetic (EM) waves in both time and space domains<sup>26</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[113, 115, 884, 528]]<|/det|>
+The metasurfaces are composed of specific unit structures arranged in periodic or quasi- periodic arrays, which can flexibly control the EM waves at the subwavelength scale and yield a large number of unusual physical phenomena and novel devices<sup>27- 30</sup>. The proposal of digital coding and programmable metasurfaces has established a profound connection between the EM fields and digital information under the control of a high- speed field programmable gate array (FPGA)<sup>31</sup>. Recently, the exploration of STC metasurfaces has sparked a surge of research interest due to the excellent ability to manipulate the EM waves and process digital information in both temporal and spatial dimensions<sup>32- 37</sup>, resulting in many novel physical phenomena that cannot be realized by the traditional spatially modulated metasurfaces. More importantly, the utilization of STC metasurfaces holds the capability to precisely control the amplitudes, phases, and polarizations at different harmonic frequencies independently by specially designing the STC matrices<sup>38</sup>, which opens up avenues to develop advanced communication schemes with enhanced efficiency and reliability<sup>39- 45</sup>. Hence, the STC metasurface is a potential candidate for deep information modulation and interaction in the SSVEP- based BCI system owing to its capability to process information and interact with the frequency- dependent visual stimulation.  
+
+<|ref|>text<|/ref|><|det|>[[112, 533, 884, 887]]<|/det|>
+In this article, we report a brain space- time- coding metasurface (BSTCM) system, which combines the human brain intelligence with the flexible capability of EM manipulations by the STC metasurface. To the best of our knowledge, this is the first instance to integrate the visual stimulation and EM encoding regulation deeply in a BCI system through metasurface. Here, the STC metasurface is used to effectively support visual stimulation for SSVEP- based BCIs while enabling the information interaction with the external environment. A lightweight deep- learning classification model is implemented to maximize the recognition performance. To enhance the BCI security, we propose a high- security encrypted wireless communication system based on the diverse harmonic regulation characteristics of BSTCM, which effectively combines the variant harmonic- based secret keys (HSKs) and visual secret sharing (VSS) method<sup>46- 48</sup>. In this way, decoding by eavesdroppers is only possible through the simultaneous interception of confidential information across all harmonic frequencies in the communication process and the acquisition of complex secret keys, which proves the system's exceptional
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 883, 190]]<|/det|>
+levels of security. Finally, smart device control is presented by the BSTCM system, realizing intelligent human- machine interactions. The proposed BSTCM can establish a new paradigm of human machine interactions, metasurface- based wireless communications, and potential applications in metaverse.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 221, 302, 238]]<|/det|>
+## System configuration  
+
+<|ref|>text<|/ref|><|det|>[[113, 251, 884, 567]]<|/det|>
+The proposed BSTCM is schematically demonstrated in Fig. 1, which consists of an STC metasurface, a brain- computer interface based on SSVEP, and an FPGA control module. The metasurface element contains a meta- structure to regulate the EM wave, and a light- emitting diode (LED) stimulator for the SSVEP paradigm. The STC metasurface is divided into four LED flickering partitions, operating at different frequencies: \(8.5\mathrm{Hz}\) , \(10\mathrm{Hz}\) , \(11.5\mathrm{Hz}\) , and \(7\mathrm{Hz}\) . The corresponding SSVEP- based EEG signals can be accurately extracted by a head- worn EEG cap when the operator is exposed to the LED flickering stimulus at distinct frequencies with the assistance of FPGA. To enhance the efficiency and accuracy of the EEG signal identification, we propose a deep learning model to recognize the extracted EEG signals, as shown in Fig. 2c. Upon completion of the identification process, the recognized signals are transmitted to FPGA, enabling to control the STC metasurface to update the corresponding STC matrices in the four LED flickering frequencies without disrupting the EEG stimulation. The BSTCM system empowers individuals to customize and modulate the EM waves by the human mind.  
+
+<|ref|>text<|/ref|><|det|>[[113, 581, 884, 895]]<|/det|>
+Leveraging the abundant harmonic modulation capabilities and the advantages of human brain intelligence, the BSTCM system can respectively implement a high- security encrypted wireless communication system and a smart device controlled by human mind. In the high- security encrypted wireless communications, the harmonic frequencies are adopted as the secret keys and the VSS method is employed to encrypt the transmitted information. Firstly, the information is encrypted into two visual secret ciphertexts that depend on the corresponding HSKs. As one of the interface devices in the metaverse, BCI plays a crucial role in enabling individuals to interact in the virtual world, and it requires high- security performance. The presented encrypted wireless communication is envisioned to fulfill the requirement for secure communication in the metaverse. For instance, in a metaverse scenario shown in Fig. 1, a legitimate transmitter (Alice) equipped with an EEG cap intentionally transfers the secret information to two legitimate receivers (Bob and Carol) at two harmonic frequencies based on the encrypted communication. The eavesdropper (Eve) cannot decrypt the transmitted
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 82, 884, 299]]<|/det|>
+information unless she simultaneously obtains the secret keys, two ciphertexts, and encryption mechanisms, which is nearly impossible. Hence, the proposed system ensures high security and concealment by generating a large number of HSKs by modulating the STC metasurface. Ingeniously camouflaged as an LED stimulator, the STC metasurface effectively obstructs the eavesdroppers from detecting the information interaction. On the other hand, the presented system also enables wireless control of smart devices in real environments, allowing for manipulating the devices directly based on the user's brain intention without the requirement of physical actions. Hence the BSTCM can serve as a stimulator for SSVEP, and effectively manipulate the EM waves for secure BCI wireless communications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 319, 339, 337]]<|/det|>
+## SSVEP signal recognition  
+
+<|ref|>text<|/ref|><|det|>[[111, 345, 884, 914]]<|/det|>
+The EEG signals can be accurately recognized in a BCI system through the detection of SSVEP components. The target flickering stimuli with four frequencies are supported with the STC metasurface integrated with LEDs, and the commands can be output by simply fixating on the corresponding visual stimuli. The STC metasurface is composed of \(32 \times 32\) elements, and segmented into four partitions, with each region consisting of \(16 \times 16\) elements. In these partitions, LEDs operate at four distinct frequencies (8.5Hz, 10Hz, 11.5Hz, and 7Hz) and are implemented to evoke four distinct SSVEP signals. As depicted in Fig. 2a, the SSVEP- BCI is segmented into three distinct stages: preparation stage, signal acquisition stage, and signal recognition stage. During the preparation stage, the participant wears an EEG cap equipped with electrodes positioned over the O1 and O2 regions and focuses the attention on one of the regions on the STC metasurface. In the signal acquisition stage, the participant maintains their gaze while an EEG amplifier collects the EEG data (for details, see Supplementary Note 1). Fig. 2d illustrates the frequency distribution of the EEG signals for the four target frequencies after the FFT analysis (prefiltered using a Butterworth bandpass filter with a frequency range of 5- 40Hz). As illustrated in Fig. 2b, the signal undergoes an initial pre- processing procedure during the signal recognition stage, involving signal decomposition, spectrum calculation and weighted summation, output signal amplitude spectrum (SigSpec). Subsequently, an outer product operation is executed with the pre- calculated reference spectra (RefSpec), yielding eight feature graphs (comprising of twochannel and four target frequencies), each sized by \(160 \times 160\) pixels. A discernible square grid pattern emerges in the feature graph corresponding to the frequency of the anticipated classification result (for additional information, refer to Supplemental Information Notes 2- 5). The conventional SSVEP classification and recognition algorithms include canonical correlation analysis (CCA) \(^{49}\) , filter bank canonical correlation
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 81, 884, 496]]<|/det|>
+analysis (FBCCA) \(^{50}\) , task- related component analysis (TRCA) \(^{9}\) , support vector machine (SVM) \(^{51}\) , and convolutional neural network (CNN) \(^{52}\) . However, some limitations still exist in the above- mentioned algorithms. Here, we propose a deep learning classification algorithm that embeds the SSVEP components in the signal recognition stage (refer to Supplemental Information Note 6). The above- mentioned eight feature graphs are input into a lightweight CNN classification model containing four convolutional and two liner layers, as shown in Fig. 2c. The feature graphs are initially processed through a convolutional input layer. Subsequently, a MaxPooling operation with a \(2 \times 2\) kernel is applied following the activation of the data via the ReLU function. To facilitate the propagation of original data, a residual connection is strategically integrated between the second and third convolutional layers. Each convolutional layer is followed by the same \(2 \times 2\) MaxPooling operation and ReLU activation. Upon the extraction of features from the eight feature graphs, a series of two successive linear layers are deployed to categorize the features into four distinct target frequencies. Finally, an output layer with sigmoid activation is employed to reshape the output of the model into 4 confidence scores. Ultimately, the model achieved a classification accuracy of \(96.67\%\) on the validation set, with precision and recall rates surpassing \(90\%\) . Such results underscore the efficacy of the SSVEP component classification algorithm proposed in this work.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 517, 355, 534]]<|/det|>
+## Design of STC metasurface  
+
+<|ref|>text<|/ref|><|det|>[[112, 547, 884, 912]]<|/det|>
+Fig. 3a showcases the detailed configuration of the 1- bit STC metasurface element integrated with an LED, which consists of three metallic layers and multiple substrate layers. The top metallic layer has two slotted rectangular patches with a width of \(w = 7.3 \mathrm{mm}\) and a length of \(l = 11 \mathrm{mm}\) , in which the middle slot has a width of \(w_{1} = 1.5 \mathrm{mm}\) and a length of \(l_{1} = 7 \mathrm{mm}\) . The outer metal frame with a width of \(0.15 \mathrm{mm}\) is used to weaken the coupling effect between elements. The top F4B dielectric substrate has a relative permittivity of 2.65 and a loss tangent of 0.003 with the thickness \(h = 3 \mathrm{mm}\) and period \(p = 19 \mathrm{mm}\) . Two PIN diodes are serially loaded on the middle of two rectangular patches, and the two rectangular patches connected with metallic bars serve as positive and negative poles, in which two inductors are used as radio- frequency (RF) chokes to effectively isolate the RF signal from direct current (DC). An LED serving as the flash stimulation is placed outside the metallic patches to avoid affecting the scattering performance of the meta- atom and shares the same voltage with PIN diodes. To evaluate the performance of the 1- bit meta- atom, full- wave simulations are conducted using the commercial software CST Microwave Studio. As shown in Figs. 3b and 3c, the scattering characteristics of the meta- atom can be altered by switching the states of two PIN diodes
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 81, 883, 273]]<|/det|>
+corresponding to distinct RLC equivalent circuit models, hence realizing the 1- bit phase regulation. The diodes with OFF state are assigned as coding “0”, and ON state as coding “1”. When the meta-atom is subjected to normal illumination by a \(y\) - polarized plane wave, a phase difference of \(180^{\circ}\) between “0” and “1” states is observed at approximately \(6.9\mathrm{GHz}\) , and both states exhibit reflective amplitudes exceeding - 1 dB, indicating excellent performance of the 1- bit programmable meta-atom. Hence the designed meta-atom can be used to build the 1- bit STC metasurface. The detailed prototype design, modeling and characterization are given in Methods and Supplementary Note 9.  
+
+<|ref|>text<|/ref|><|det|>[[112, 285, 884, 904]]<|/det|>
+To establish a dual- user encrypted wireless communication system, we implement an amplitude- shift keying modulation scheme using the STC metasurface, enabling simultaneous encrypted transmissions via space and frequency multiplexing. The control voltages, originating from FPGA, are applied to the PIN diodes, enabling periodic switching to the reflection phases of the 1- bit STC element with a time period of \(T_0 = 1 / f_0\) . Consequently, each STC element possesses a periodic time- coding sequence and constitutes part of an STC matrix, making it possible to flexibly control the EM waves in both space and time domains. The flexibility of STC matrices can generate a series of new harmonic spectra at the frequency interval of \(f_0\) , which are further optimized with specific requirements (see Supplementary Notes 7 and 8). The space- time modulation enables abundant harmonic characteristics to construct harmonic- based encrypted wireless communications. Using the 2D STC matrices, we can modulate the EM waves to steer to the desired directions for target users. The STC matrices encompass 16 rows of elements, each possessing periodic time- coding sequences with 11 intervals. As illustrative examples, three radiation patterns corresponding to three STC matrices (M0, M1, and M2) are shown in Fig. 3. The radiation pattern of the optimized STC matrix M1 in Fig. 3e is strategically manipulated to yield a main beam with a deflected angle of \(\theta = - 15^{\circ}\) at the \(+1\) st harmonic frequency, as showcased in Fig. 3h. Similarly, the radiation pattern of the STC matrix M2 (Fig. 3f) is optimized to direct the beam towards the angle \(\theta = +30^{\circ}\) , in correspondence with the - 1st harmonic frequency, as demonstrated in Fig. 3i. Furthermore, the time- invariant STC matrix M0 (Fig. 3d) facilitates the positioning of main beam towards a direction of \(\theta = 0^{\circ}\) , associated with the fundamental frequency, as illustrated in Fig. 3g. As an example of amplitude- shift keying modulation scheme, the left panels ( \(f_1 = 8.5\mathrm{Hz}\) and \(f_2 = 10\mathrm{Hz}\) ) of the STC metasurface are utilized to send digital symbols “0” and “1” to User1 at the position of \(\theta = - 15^{\circ}\) when the transmitting information is encoded as binary data streams. The right panels ( \(f_3 = 11.5\mathrm{Hz}\) and \(f_4 = 7\mathrm{Hz}\) ) of the STC metasurface are utilized to send digital
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 82, 885, 151]]<|/det|>
+symbols “0” and “1” to User2 at the position of \(\theta = +30^{\circ}\) . Experiments are carried out in a microwave anechoic chamber to measure the far- field radiation patterns of the STC metasurface, as illustrated in Supplementary Notes 10 and 11.  
+
+<|ref|>text<|/ref|><|det|>[[113, 165, 884, 629]]<|/det|>
+Constrained by the requirement of the SSVEP- based BCI to gather data and detect EEG signals at a single frequency, a viable encoding strategy is proposed to enhance the symbol error rate (SER), as depicted in Fig. 3j- m. The transmission of the digital symbols “0/1” involves to encode a series of repeated data frames with a period of 4 frames as synchronous frames. In the encoding scheme, the first bit “1” serves as the start bit, signaling the initiation of the transmission process. The second bit signifies the harmonic frequency (+1st or - 1st) used for transmission, and the third bit corresponds to the transmission of the data itself. Finally, the fourth bit functions as the stop bit, indicating the end to transmit one data frame. Based on the encoding process, the digital symbol “0” or “1” for User1 is encoded as the repeated data frames “10001000” or “10101010...”, respectively. Similarly, the digital symbol “0/1” to User2 is encoded as repeated data frames “11001100...” or “11101110...”, respectively. Hence, the STC matrix M1 operating with a switching frequency \(f_0\) , is precisely injected into the region of high amplitude associated with the flicker frequencies \(f_1\) and \(f_2\) , as illustrated in Fig. 3j and 3k. The STC matrices are updated when the information is sent to User1 via the BSTCM system. Similarly, the STC matrix M2 with a period switch frequency \(f_0\) , undergoes injection into the high amplitude attributed to the flicker frequencies \(f_3\) and \(f_4\) , as displayed in Fig. 3j and 3k, representing the transmission of information to User2. Remarkably, the update of STC matrices does not affect the flickering frequency of LEDs used for SSVEP stimuli, ensuring that BCI signals can be synchronously detected and recognized.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 650, 543, 668]]<|/det|>
+## Secure encrypted wireless communication system  
+
+<|ref|>text<|/ref|><|det|>[[115, 686, 884, 899]]<|/det|>
+The conventional VSS method is employed to encrypt the individual pixels of confidential information, where each pixel is divided into two sub- pixels. It is worth noting that each share derived from this process remains devoid of any discernible information pertaining to the original image content. As a result, a secret message possesses two visual shared keys (VSKs), which effectively guarantee that the disclosure of any single VSK does not compromise the confidentiality of the original secret information. Here, we present a novel encryption strategy by combining the harmonic characteristics of STC metasurface with the VSS method, which is designed to realize high- security encrypted wireless communication directly controlled by the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 885, 469]]<|/det|>
+human mind. In the presented encryption strategy, two VSKs are not simply stacked with each other to recover the original information, but rely on HSKs. In the described encryption process, the confidential information to be transmitted is firstly encrypted to two random VSKs based on the presented encrypted encoding scheme. Then, two generated VSKs are further encoded into two random HSKs for the \(\pm 1\) harmonic frequencies, respectively. During decryption, two VSKs attached to each secret can be reconstructed with the pre- customized variant HSK sequences. The details of encryption and decryption methods can be found in Supplementary Note 12. The target image in question can be decrypted through the incoherent superposition of two extracted VSKs, aided by the presence of HSKs. By imbuing addressable shared information bits with the harmonic dynamics, a robust level of information security is achieved. As an illustrative example in Fig. 4a, the image “H”, which consists of \(5 \times 5\) grid of black or white pixels, is encrypted into two random VSK1 and VSK2 based on the presented encryption method. Two bitstream sequences of VSKs can be successively transmitted to two designated Users (Bob and Carol) by the BCI operator (Alice) using the presented BSTCM system.  
+
+<|ref|>text<|/ref|><|det|>[[113, 486, 885, 895]]<|/det|>
+To validate the encrypted communication scheme, we experimentally set up an encrypted wireless communication system using the BSTCM, as demonstrated in Fig. 4b. The prototype of 1- bit STC metasurface is used to construct the encrypted communication system, as shown in Supplementary Figure 15. A commercial BCI device (actiCHamp of Brain Product GmbH company) composed of an EEG cap and amplifier is used to acquire the EEG signals when the operator stares at different partitions on the STC metasurface. The corresponding EEG signals can be accurately recognized by the proposed deep learning algorithm. The detailed description of the algorithm is provided in Supplementary Note 6. A linearly polarized horn antenna, connected to a signal generator (Keysight E8267D), is placed at a vertical distance of 4 m away from the metasurface center to excite a monochromatic plane wave at the frequency of 6.9 GHz. The receiving terminals are composed of two horn antennas (severed as two users) and a spectrum analyzer (Keysight N9040D), which is used to demodulate the received signal. Two receiving horn antennas are located at the angles of \(\theta = - 15^{\circ}\) and \(\theta = +30^{\circ}\) with relative to the metasurface normal, respectively. In the transmission process, the information undergoes encryption, resulting in the generation of two harmonic- based cipher texts according to the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 884, 247]]<|/det|>
+encryption scheme. The transmission process is achieved by ensuring that the BCI operator maintains direct visual focus on the STC metasurface in alignment with the binary cipher texts. In addition, the system enables prompt acquisition, recognition, and translation of EEG signals into appropriate control signals of FPGA, which in turn drive the STC metasurface. As a result, the accurate transmission of binary cipher texts is accomplished, facilitating reliable and efficient communications between the BCI operator and the intended recipients.  
+
+<|ref|>text<|/ref|><|det|>[[113, 263, 885, 730]]<|/det|>
+In the receiving process, the radiated signals propagating through free space are received individually by the two horn antennas, and these signals are subsequently demodulated using the spectrum analyzer. As shown in Fig. 4c, the demodulated signals represent four encoding schemes that correspond to different digital symbols ("0" or "1") for the two users and then are processed to the decoded results shown in Fig. 4d by the FFT method. The decoded results clearly exhibit four repeated data frames, providing strong evidence to support the capability of the proposed system to recover the transmitting information successfully. Hence two binary cipher texts are transmitted to two specific users by the BCI operator based on the proposed wireless communication system. The transmitted information is received and the measured results are comprehensively showcased in Supplementary Note 13 and Supplementary Video 1. By employing the encoding method described earlier, the data streams corresponding to the specific users are extracted from the received signals, as depicted in Figs. 4e- f. Subsequently, these data streams are decrypted accurately by the method outlined in Supplementary Note 12. The deciphered information is displayed in Fig. 4g. The successful decryption and retrieval of the transmitted information affirm the robustness and efficacy of the proposed system in maintaining high levels of security and confidentiality. The system is not easily deciphered or cracked, proving its ability to protect sensitive information during transmissions.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 784, 377, 801]]<|/det|>
+## Mind control to smart devices  
+
+<|ref|>text<|/ref|><|det|>[[115, 820, 883, 893]]<|/det|>
+To experimentally verify the performance of the proposed BSTCM system, we conducted experiments to show the remote control capabilities of smart devices through human intention, which holds significant potential to enhance the quality of life for individuals with disabilities.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 85, 884, 696]]<|/det|>
+The system flowchart depicted in Fig. 5a outlines the key steps involved in the experimental setup. Initially, the EEG signals are acquired and subsequently processed by the BSTCM system for command recognition. The recognized commands are then translated to update the STC matrices of metasurface by an FPGA. The four partitions of the STC metasurface are controlled by FPGA to generate different beams at four deflected angles corresponding to the four harmonic frequencies, enabling the remote control of smart devices. Notably, the main beam at the - 2nd harmonic frequency exhibits a pronounced peak in the direction of \(\theta = +10^{\circ}\) , corresponding to the device assigned to User 3. Conversely, the main beam at the +2nd harmonic frequency demonstrates a strong peak at \(\theta = - 45^{\circ}\) , corresponding to the device assigned to User 1. As displayed in Figs. 5c and d, the measured radiation patterns exhibit good consistency with theoretical predictions associated to the corresponding STC matrices. Additionally, the main beams were deflected at the other two angles (- 15° and 30°) using the +1st and - 1st harmonic frequencies (Users 2 and 4), as shown in Figs. 3h and i. The excellent harmonic beam manipulation of the STC metasurface lays a solid theory foundation for smart device control. The receiving terminals are composed of four horn antennas, four RF energy detectors (LMH2110), four Microcontroller Unit (MCU) modules (Arduino), and four LED modules (serve as smart devices), as depicted in Fig. 5b. The RF energy from harmonic beams captured by the detectors can be converted into DC voltages, which are subsequently fed into the MCU modules. Once the corresponding DC voltage is detected, the MCU modules send the control commands to activate the corresponding LED module, thereby implementing the remote control of the smart devices. The experimental setup is shown in Fig. 5b, where the STC metasurface is illuminated by a monochromatic frequency signal of 6.9GHz.  
+
+<|ref|>text<|/ref|><|det|>[[115, 708, 884, 894]]<|/det|>
+To verify the robustness of the brain- control smart devices system, the position of the receiving antenna is sequentially placed at different distances from the metasurface center while maintaining specific harmonic direction angles with respect to the metasurface normal. The detection of RF energy from the emitted harmonic beams is carried out across varying distances and the corresponding DC voltages can be accurately converted by the RF detectors, as illustrated in Fig. 5e (see Supplementary Note 14 for details). We observe that the RF energies of four users relatively decrease and the converted DC voltages also decrease accordingly as
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 87, 884, 469]]<|/det|>
+the distance increases. To guarantee the attainment of the far- field region and maintain system stability, the four receiving antennas are positioned at a distance of approximately \(300\mathrm{cm}\) from the metasurface. Consequently, the detection voltage threshold is set as \(0.15\mathrm{V}\) , \(0.4\mathrm{V}\) , \(0.4\mathrm{V}\) , and \(0.2\mathrm{V}\) in MCU modules, respectively. We recorded a video of the brain- controlled multiple smart devices, as presented in Supplementary Movie 2. The sequence diagrams shown in Fig. 5f provide a clear visual representation of the sequential control process, where the smart devices (LED modules) are successively controlled through human intention using the BSTCM system. The experiment proves the feasibility of brain- controlled smart devices. By harnessing the power of BCI technology, individuals can exert control over smart devices without needing physical interaction. The BSTCM system enables seamless and intuitive manipulations of various smart devices, facilitating greater independence and convenience for people in daily life. The experimental results serve as a compelling validation supporting the significant benefits that the BSTCM can offer in improving the quality of life for individuals facing physical challenges.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 518, 223, 535]]<|/det|>
+## Conclusions  
+
+<|ref|>text<|/ref|><|det|>[[113, 557, 884, 910]]<|/det|>
+We presented a BSTCM system, which combines the human brain intelligence and flexible EM control capabilities of the STC metasurface. A series of visual stimuli encoded by specific coding schemes was employed on the metasurface to realize deep information processing and interaction in the EM field for reliable and stable communication environments. In addition, a lightweight deep- learning algorithm was proposed to accurately recognize the brain signals extracted by the BCI device. We further demonstrated a high- security encrypted wireless communication system based on diverse harmonic frequency characteristics, which combines the variant HSKs and VSS methods. The proposed system can enable the system to have high security. Finally, remote control of smart devices was demonstrated by the BSTCM system, realizing intelligent human- machine interactions. The proposed BSTCM, integrating the EM manipulation capability with the intelligence of the human brain, provides a new paradigm of interactions among human machine interface in metaverse, and future 6G wireless communication applications.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 117, 317, 134]]<|/det|>
+## Methods and materials  
+
+<|ref|>text<|/ref|><|det|>[[115, 152, 885, 395]]<|/det|>
+Details on STC metasurface. A metasurface prototype was fabricated by the standard printed circuit board (PCB) process. The entire STC metasurface is composed of \(32 \times 32\) digital elements and an effective size of \(608 \times 608 \mathrm{~mm}^2\) . For the convenience of processing and welding, the entire metasurface was disassembled into 8 sub- metasurfaces with \(8 \times 16\) digital elements and an effective size of \(152 \times 304 \mathrm{~mm}^2\) . A total number of \(2 \times 32 \times 32\) PIN diodes (SMP1320- 079LF from SKYWORKS) and RF inductors of \(10 \mathrm{nH}\) were elaborately welded on the metasurfaces with the surface mount technology, which is a mature engineering method to weld small components on PCB using the batch solder- reflow processes in a dedicated machine. Meanwhile, an LED is embedded into each metasurface element.  
+
+<|ref|>text<|/ref|><|det|>[[114, 429, 884, 895]]<|/det|>
+The brain signal acquisition. As depicted in Fig. 2a, the SSVEP- BCI is delineated into three discrete distinct stages: the preparation stage, the signal acquisition stage, and the signal recognition stage. The EEG data in this study were acquired using the actiCHamp EEG amplifier manufactured by Brain Products GmbH. This amplifier offers a maximum capacity of 64 channels, 24- bit precision, and a sampling rate of \(100 \mathrm{kHz}\) . It exhibits a high common mode rejection ratio (CMRR) of 100, making it suitable for capturing and detecting SSVEP signals effectively. This study uses a standard 10- 20 EEG cap to obtain the signal. Since SSVEP is a local potential that predominantly originates from the occipital lobe \(^5\) , the O1 and O2 electrodes located on the occipital region were selected for EEG data acquisition, with a reference electrode placed at the vertex (Cz) region. The sample time in this study was set to four seconds. With a sampling rate of \(512 \mathrm{~Hz}\) and excluding the reference electrode data, each sample yielded 4096 sampling points (4 seconds \(\times 512 \mathrm{~Hz} \times 2\) channels). Square wave signals were utilized to present visual stimuli on the LEDs of the STC metasurface units, which emitted the red light, thereby ensuring adequate visual contrast. To minimize the harmonic interference, four target visual stimulus frequencies were: top- left at \(8.5 \mathrm{~Hz}\) , bottom- left at \(10 \mathrm{~Hz}\) , top- right at \(11.5 \mathrm{~Hz}\) , and bottom- right at \(7 \mathrm{~Hz}\) . To mitigate mutual interference between two closely spaced flickering stimuli pairs, the frequencies \(7 \mathrm{~Hz}\) and \(8.5 \mathrm{~Hz}\) , as well as \(10 \mathrm{~Hz}\) and \(11.5 \mathrm{~Hz}\) ,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 884, 275]]<|/det|>
+were placed diagonally opposite to one another. The example of the collected brain wave signal data for four target frequencies is shown in Supplementary Figure 1. Data processing and model training were conducted in Python and PyTorch environments. Signal processing primarily involved utilizing scipy package for filtering and FFT calculations. Given that neither the absolute value of the EEG signal nor its temporal characteristics were pertinent to this study, all sampled signals were linearly transformed to have a mean of zero and a standard deviation of one.  
+
+<|ref|>text<|/ref|><|det|>[[113, 310, 885, 607]]<|/det|>
+Classification model training. Initially, a dataset comprising 240 instances of raw SSVEP signals was collected from two participants. Subsequently, a subset consisting of 80 instances of randomly selected raw signals was employed to create RefSpec. The remaining 160 instances of raw signals were randomly selected to serve as training and validation datasets for classification model training. The two subsets have an equally divided number of signals across four stimuli frequencies. As for the training and validation dataset, a subset of 60 instances was extracted from the initial subset of 160 instances to serve as the validation set. The remaining 100 instances underwent a data augmentation process, which resulted in an expanded dataset of 1700 instances. These augmented instances constituted the training set, which was then utilized to train the model for 300 epochs. The validation operation was executed every two training epochs.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 645, 286, 662]]<|/det|>
+## Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[114, 678, 884, 820]]<|/det|>
+The work is supported by the National Natural Science Foundation of China (62288101, 92167202, 72171044), the Major Project of Natural Science Foundation of Jiangsu Province (BK20212002), the State Key Laboratory of Millimeter Waves, Southeast University, China (K201924), the Fundamental Research Funds for the Central Universities (2242018R30001), the 111 Project (111- 2- 05), the China Postdoctoral Science Foundation (2021M700761), and ZhiShan Scholar Program of Southeast University (2242022R40004).  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 860, 301, 876]]<|/det|>
+## Author contributions
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 82, 885, 250]]<|/det|>
+T.J.C. suggested the designs and planned and supervised the work, in consultation with Q.M., Y.F.N., Q.X., and J.W.Y. conceived the idea and carried out the theoretical analysis and numerical simulations. Q.X., L.H.F., Y.M.N., and Z.G. built the system and performed the experimental measurements. Q.X., L.H.F. and L.C. performed the data analysis. Q.X. and L.H.F. wrote the manuscript. Q.M., L.L., J.W. Y, and T.J.C. reviewed the manuscript. All authors discussed the theoretical aspects and numerical simulations, interpreted the results and reviewed the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 265, 320, 285]]<|/det|>
+## Competing interests  
+
+<|ref|>text<|/ref|><|det|>[[117, 302, 538, 320]]<|/det|>
+The authors declare no competing financial interest.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 352, 285, 371]]<|/det|>
+## Data availability  
+
+<|ref|>text<|/ref|><|det|>[[117, 380, 881, 422]]<|/det|>
+The data that support the findings of this study are available from the corresponding author upon request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 451, 290, 470]]<|/det|>
+## Code availability  
+
+<|ref|>text<|/ref|><|det|>[[117, 479, 881, 521]]<|/det|>
+The code that supports the findings of this study are available from the corresponding author upon reasonable request.  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 551, 228, 569]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[115, 579, 885, 911]]<|/det|>
+Wu, D. et al. Virtual-Reality Interpromotion Technology for Metaverse: A Survey. IEEE Internet of Things Journal 10, 15788- 15809, doi:10.1109/jiot.2023.3265848 (2023). Park, S., Ha, J., Park, J., Lee, K. & Im, C.- H. Brain- Controlled, AR- Based Home Automation System Using SSVEP- Based Brain- Computer Interface and EOG- Based Eye Tracker: A Feasibility Study for the Elderly End User. IEEE Transactions on Neural Systems and Rehabilitation Engineering 31, 544- 553, doi:10.1109/tnsre.2022.3228124 (2023). Al- Saegh, A., Dawwd, S. A. & Abdul- Jabbar, J. M. Deep learning for motor imagery EEG- based classification: A review. Biomedical Signal Processing and Control 63, doi:10.1016/j.bspc.2020.102172 (2021). Lenhardt, A., Kaper, M. & Ritter, H. J. An Adaptive P300- Based Online Brain- Computer Interface. IEEE Transactions on Neural Systems and Rehabilitation Engineering 16, 121- 130, doi:10.1109/tnsre.2007.912816 (2008). Vialatte, F.- B., Maurice, M., Dauwels, J. & Cichocki, A. Steady- state visually evoked potentials: Focus on essential paradigms and future perspectives. Progress in Neurobiology 90, 418- 438, doi:10.1016/j.pneurobio.2009.11.005 (2010). Niu, Y. et al. Improving SSVEP- BCI System Interaction Efficiency: Design Recommendations for Shape of Visual Stimuli and Number of Auxiliary Stimuli. International Journal of Human- Computer Interaction, 1- 22, doi:10.1080/10447318.2023.2188540 (2023).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[57, 75, 888, 900]]<|/det|>
+464 7 Zhao, X., Du, Y. & Zhang, R. A CNN- based multi- target fast classification method for AR- SSVEP. 465 Computers in Biology and Medicine 141, doi:10.1016/j.compbiomed.2021.105042 (2022). 466 Ravi, A., Lu, J., Pearce, S. & Jiang, N. Enhanced System Robustness of Asynchronous BCI in 467 Augmented Reality Using Steady- State Motion Visual Evoked Potential. IEEE Transactions on 468 Neural Systems and Rehabilitation Engineering 30, 85- 95, doi:10.1109/tnsre.2022.3140772 (2022). 469 Nakanishi, M. et al. Enhancing Detection of SSVEPs for a High- Speed Brain Speller Using Task- 470 Related Component Analysis. IEEE Transactions on Biomedical Engineering 65, 104- 112, 471 doi:10.1109/tbme.2017.2694818 (2018). 472 Bin, G., Gao, X., Wang, Y., Hong, B. & Gao, S. VEP- based brain- computer interfaces: time, 473 frequency, and code modulations [Research Frontier. IEEE Computational Intelligence Magazine 4, 474 22- 26, doi:10.1109/mci.2009.934562 (2009). 475 Li, Y. Q., Pan, J. H., Wang, F. & Yu, Z. L. A Hybrid BCI System Combining P300 and SSVEP and 476 Its Application to Wheelchair Control. Ieee Transactions on Biomedical Engineering 60, 3156- 3166, 477 doi:10.1109/tbme.2013.2270283 (2013). 478 Muller- Putz, G. R. & Pfurtscheller, G. Control of an Electrical Prosthesis With an SSVEP- Based 479 BCI. IEEE Transactions on Biomedical Engineering 55, 361- 364, doi:10.1109/tbme.2007.897815 480 (2008). 481 Willett, F. R., Avansino, D. T., Hochberg, L. R., Henderson, J. M. & Shenoy, K. V. High- performance 482 brain- to- text communication via handwriting. Nature 593, 249- 254, doi:10.1038/ s41586- 021- 483 03506- 2 (2021). 484 Roy, A. M. Adaptive transfer learning- based multiscale feature fused deep convolutional neural 485 network for EEG MI multiclassification in brain- computer interface. Engineering Applications of 486 Artificial Intelligence 116, doi:10.1016/j.engappai.2022.105347 (2022). 487 Ma, Q. et al. Directly wireless communication of human minds via non- invasive brain- computer- 488 metasurface platform. eLight 2, doi:10.1186/s43593- 022- 00019- x (2022). 489 Wandelt, S. K. et al. Representation of internal speech by single neurons in human supramarginal 490 gyrus. Nature Human Behaviour, doi:10.1038/s41562- 024- 01867- y (2024). 491 Xiao, Q. et al. Electromagnetic Brain- Computer- Metasurface Holography. ACS Photonics, 492 doi:10.1021/acsphotonics.2c01349 (2023). 493 Zhu, R. et al. Remotely Mind- Controlled Metasurface via Brainwaves. eLight 2, 10 (2022). 494 Jiang, W., Han, B., Habibi, M. A. & Schotten, H. D. The Road Towards 6G: A Comprehensive Survey. 495 IEEE Open Journal of the Communications Society 2, 334- 366, doi:10.1109/ojcoms.2021.3057679 496 (2021). 497 Saad, W., Bennis, M. & Chen, M. A Vision of 6G Wireless Systems: Applications, Trends, 498 Technologies, and Open Research Problems. IEEE Network 34, 134- 142, doi:10.1109/mnet.001. 499 1900287 (2020). 500 Zhang, Z. et al. 6G Wireless Networks: Vision, Requirements, Architecture, and Key Technologies. 501 IEEE Vehicular Technology Magazine 14, 28- 41, doi:10.1109/mvt.2019.2921208 (2019). 502 Dang, S., Amin, O., Shihada, B. & Alouini, M.- S. What should 6G be? Nature Electronics 3, 20- 29, 503 doi:10.1038/s41928- 019- 0355- 6 (2020). 504 Brocal, F. Brain- computer interfaces in safety and security fields: Risks and applications. Safety 505 Science 160, doi:10.1016/j.ssci.2022.106051 (2023). 506 Bhalerao, S., Ansari, I. A. & Kumar, A. Protection of BCI system via reversible watermarking of 507 EEG signal. Electronics Letters 56, 1389- 1392, doi:10.1049/el.2020.2532 (2020).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[58, 87, 886, 900]]<|/det|>
+509 25 Bernal, S. L., Celdrán, A. H., Pérez, G. M., Barros, M. T. & Balasubramaniam, S. Security in Brain- 509 Computer Interfaces. ACM Computing Surveys 54, 1- 35, doi:10.1145/3427376 (2021). 510 26 Wei, M., Zhao, H., Galdi, V., Li, L. & Cui, T. J. Metasurface- enabled smart wireless attacks at the 511 physical layer. Nature Electronics 6, 610- 618, doi:10.1038/s41928- 023- 01011- 0 (2023). 512 27 Pendry, J. B. Negative refraction makes a perfect lens. Phys Rev Lett 85, 3966- 3969, 513 doi:10.1103/PhysRevLett.85.3966 (2000). 514 28 Pendry, J. B., Schurig, D. & Smith, D. R. Controlling Electromagnetic Fields. Science 312, 1780- 515 1782, doi:10.1126/science.1125907 (2006). 516 29 Liu, R. et al. Broadband ground- plane cloak. Science 323, 366- 369 (2009). 517 30 Smith, D. R. & Kroll, N. Negative refractive index in left- handed materials. Phys Rev Lett 85, 2933- 518 2936, doi:10.1103/PhysRevLett.85.2933 (2000). 519 31 Cui, T. J., Qi, M. Q., Wan, X., Zhao, J. & Cheng, Q. Coding metamaterials, digital metamaterials 520 and programmable metamaterials. Light: Science & Applications 3, e218- e218, doi:10.1038/lsa. 521 2014.99 (2014). 522 32 Zhang, L. et al. Space- Time- Coding Digital Metasurfaces. Nat. Commun. 9, 4334 (2018). 523 33 Wu, H. et al. Harmonic information transitions of spatiotemporal metasurfaces. Light Sci Appl 9, 524 198, doi:10.1038/s41377- 020- 00441- 1 (2020). 525 34 Wu, G.- B., Dai, J. Y., Cheng, Q., Cui, T. J. & Chan, C. H. Sideband- free space- time- coding 526 metasurface antennas. Nature Electronics 5, 808- 819, doi:10.1038/s41928- 022- 00857- 0 (2022). 527 35 Zhang, L. et al. Co- Prime Modulation for Space- Time- Coding Digital Metasurfaces with Ultralow- 528 Scattering Characteristics. Advanced Functional Materials, doi:10.1002/adfm.202314110 (2024). 529 36 Zhang, L. et al. A wireless communication scheme based on space- and frequency- division 530 multiplexing using digital metasurfaces. Nature Electronics 4, 218- 227, doi:10.1038/s41928- 021- 531 00554- 4 (2021). 532 37 Shaltout, A. M., Shalaev, V. M. & Brongersma, M. L. Spatiotemporal light control with active 533 metasurfaces. Science 364, doi:10.1126/science.aat3100 (2019). 534 38 Wu, G.- B. et al. A universal metasurface antenna to manipulate all fundamental characteristics of 535 electromagnetic waves. Nature Communications 14, doi:10.1038/s41467- 023- 40717- 9 (2023). 536 39 Zhao, J. et al. Programmable time- domain digital- coding metasurface for non- linear harmonic 537 manipulation and new wireless communication systems. National Science Review 6, 231- 238, 538 doi:10.1093/nsr/nwy135 (2019). 539 40 Dai, J. Y. et al. Wireless Communication Based on Information Metasurfaces. IEEE Transactions on 540 Microwave Theory and Techniques 69, 1493- 1510, doi:10.1109/tmtt.2021.3054662 (2021). 541 41 Dai, J. Y. et al. Realization of Multi- Modulation Schemes for Wireless Communication by Time- 542 Domain Digital Coding Metasurface. IEEE Transactions on Antennas and Propagation 68, 1618- 543 1627, doi:10.1109/tap.2019.2952460 (2020). 544 42 Dai, J. Y. et al. Wireless Communications through a Simplified Architecture Based on Time- Domain 545 Digital Coding Metasurface. Advanced Materials Technologies 4, doi:10.1002/admt. 201900044 546 (2019). 547 43 Zhang, L. et al. A Wireless Communication Scheme Based on Space- and Frequency- Division 548 Multiplexing Using Digital Metasurfaces. Nat. Electron. 4, 218 (2021). 549 44 Tang, W. et al. Wireless Communications with Programmable Metasurface: New Paradigms, 550 Opportunities, and Challenges on Transceiver Design. IEEE Wireless Communications 27, 180- 187, 551 doi:10.1109/mwc.001.1900308 (2020).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[58, 84, 884, 700]]<|/det|>
+552 45 Tang, W. et al. Programmable metasurface- based RF chain- free 8PSK wireless transmitter. Electronics Letters 55, 417- 420, doi:10.1049/el.2019.0400 (2019). 553 46 Chen, T.- H. & Tsao, K.- H. Visual secret sharing by random grids revisited. Pattern Recognition 42, 2203- 2217, doi:10.1016/j.patcog.2008.11.015 (2009). 555 47 Yang, C.- N. New visual secret sharing schemes using probabilistic method. Pattern Recognition Letters 25, 481- 494, doi:10.1016/j.patrec.2003.12.011 (2004). 556 48 Wang, D., Yi, F. & Li, X. Probabilistic visual secret sharing schemes for grey- scale images and color images. Information Sciences 181, 2189- 2208, doi:10.1016/j.ins.2011.01.019 (2011). 557 49 Bin, G., Gao, X., Yan, Z., Hong, B. & Gao, S. An online multi- channel SSVEP- based brain- computer interface using a canonical correlation analysis method. Journal of Neural Engineering 6, doi:10.1088/1741- 2560/6/4/046002 (2009). 558 50 Chen, X., Wang, Y., Gao, S., Jung, T.- P. & Gao, X. Filter bank canonical correlation analysis for implementing a high- speed SSVEP- based brain- computer interface. Journal of Neural Engineering 12, doi:10.1088/1741- 2560/12/4/046008 (2015). 559 51 Singla, R., Khosla, A. & Jha, R. Influence of stimuli colour in SSVEP- based BCI wheelchair control using support vector machines. Journal of Medical Engineering & Technology 38, 125- 134, doi:10.3109/03091902.2014.884179 (2014). 560 52 Ding, W. et al. Filter Bank Convolutional Neural Network for Short Time- Window Steady- State Visual Evoked Potential Classification. IEEE Transactions on Neural Systems and Rehabilitation Engineering 29, 2615- 2624, doi:10.1109/tnsre.2021.3132162 (2021).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 85, 880, 394]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 400, 883, 511]]<|/det|>
+<center>Fig. 1 | The schematical diagram of the presented BSTCM. The BSTCM system integrates STC metasurface into SSVEP-based BCI systems. The STC metasurface can correctly support the visual stimuli for the SSVEP-based BCI and facilitate the information interaction with the external environment. Based on the BSTCM system, high-security encrypted wireless communication systems are realized for the first time by combining with the variant HSKs and VSS method. In this way, smart devices can be controlled by the human mind with high security. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 84, 868, 620]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 622, 881, 806]]<|/det|>
+<center>Fig. 2 | The diagram of interaction and data processing procedures of the SSVEP-based BCI system. a. The interaction procedures for the SSVEP-based BCI involve the user wearing an EEG cap and focusing their attention on the intended target LED stimuli. Once the user's selected frequency is recognized, the interaction commands are transmitted to the STC metasurface. b. The feature extraction process includes the transformation of the input BCI signals into the frequency domain. The resulting Signal FFT is then subjected to outer product with pre-defined Template FFTs, yielding eight feature graphs. c. The lightweight classification model. Four convolutional and two fully connected layers were employed. d. An example of frequency responses of SSVEP signals for the four target frequencies, showing distinctive amplitude peaks corresponding to different frequencies. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[123, 96, 880, 555]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 565, 881, 675]]<|/det|>
+<center>Fig. 3 | The details of the STC metasurface. a. The geometrical structure of the STC metasurface element. b, c. The reflective amplitude and phase of the element, respectively. d-f. The optimized STC matrices for different scattering angles at different harmonic frequencies. g-i. The far-field results corresponding to the optimized STC matrices. j-m. The encoding schemes for the transmitting symbol “0” or “1” for two users: the symbols “0” and “1” for Users 1 (j and k); the symbols “0” and “1” for Users 2 (l and m). </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[123, 92, 884, 540]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 548, 881, 693]]<|/det|>
+<center>Fig. 4 | The encrypted wireless communication system based on the BSTCM platform. a. The encrypted encoding scheme combining the harmonic-secret keys and the VSS method. b. The experimental scenarios of the encrypted wireless communication system, in which the operator equipped with EEG cap transmits the encrypted information to two users via the BSTCM platform, respectively. c. The receiving signals corresponding to the encoding scheme in Fig 3(j-m). d, e. The decoded information of VSK1 and VSK2 from the receiving signals based on the encrypted coding scheme. f. The correct transmitting information is decoded by the extracted VSK1 and VSK2. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[128, 88, 880, 428]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 436, 883, 565]]<|/det|>
+<center>Fig. 5 | Experiment on the wireless remote mind control of smart devices based on the BSTCM platform. a. The system architecture of wireless remote control. b. The experimental scenarios of the wireless remote-control system by the human brain. c, d. The theoretical and experimental results for far-filed patterns at \(\pm 2\mathrm{th}\) harmonic frequencies. f. The temporal waveform of the output voltages from four detectors when the operator lights up four devices in sequence. e. The variations between the input power of the detectors and output voltage with respect to the different distances between the receiving antenna and the metasurface center. </center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 43, 312, 71]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[42, 93, 768, 114]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 308, 177]]<|/det|>
+SupplementaryMaterials.pdf SupplementaryVideos.zip
+
+<--- Page Split --->

preprint/preprint__09af231471af1cc00448f4d0e910132c71d4ce10f09b6fc16f3a37da4472e364/images_list.json

@@ -0,0 +1,152 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 (a) U.S. automotive sector employment consists of assembly and parts manufacturing jobs, with assembly comprising \\(33\\%\\) of total automotive jobs. During periods of economic downturn, the percentage of jobs lost in the automotive sector exceeded that of the overall U.S. manufacturing sector, as seen during the recession of 2008 and the pandemic of 2019. (b) The transition from ICEV to BEV production will create shifts in the types and quantities of jobs in both assembly and parts manufacturing. BEVs will require workers to manufacture battery cells and battery packs instead of engines. Jobs categories are classified based on the North American Industrial Classification System (NAICS) codes. Assembly jobs: NAICS 3361. Engine parts jobs: NAICS 33631. Powertrain parts jobs: NAICS 33635. Other parts: NAICS codes 3363(x) according to the parenthesized values in (b).",
+    "footnote": [],
+    "bbox": [
+      [
+        207,
+        204,
+        828,
+        400
+      ]
+    ],
+    "page_idx": 2
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2 Transition plants: vehicle assembly plants that have fully transitioned from assembling ICEVs to BEVs. (a) Map of U.S. automotive vehicle assembly activity as of 2022. Colors show the number of workers classified under \"motor vehicle manufacturing\" (NAICS code 3361) for each state. Markers highlight major manufacturing plants in each state. Three \"transition plants\" have been identified as examples of complete ICEV to BEV production transformations: (1) Alameda County, CA, representing the transition of the former New United Motors Manufacturing, Inc. (NUMMI) vehicle assembly plant, which assembled ICEV passenger vehicles, to the Tesla plant which assembles BEVs; (2) Oakland County, Michigan, representing the transition from the production of General Motors ICEVs to the Chevy Bolt BEV over six years; and (3) McLean County, Illinois, representing the transition of the former Mitsubishi vehicle assembly plant to the Rivian plant which assembles electric pick-up trucks. (b) Possible trajectories of vehicle assembly workforce size that this work seeks to clarify.",
+    "footnote": [],
+    "bbox": [
+      [
+        210,
+        88,
+        824,
+        260
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3 (a) Vehicle production history in Alameda County, California. Before 2010, the New United Motors Manufacturing Incorporated (NUMMI) factory produced ICEVs including the Corolla and the Tacoma. The factory has since been taken over by Tesla which has now been producing BEVs for over a decade. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via government NAICS 3361 data from the U.S. Quarterly Workforce Indicators (QWI), as well as via news reports. (d) Labor intensity in Alameda compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum.",
+    "footnote": [],
+    "bbox": [
+      [
+        208,
+        87,
+        832,
+        450
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4 (a) Vehicle production history in Oakland County, Michigan, home to the Orion Township assembly plant owned by General Motors. Before the 2008 recession, plants in the cities of Pontiac and Wixom were also actively producing vehicles, but both plants shut down in 2010, leaving the Orion plant as the sole operating plant in the county. In 2016, Orion began producing the Chevy Bolt BEV alongside GM ICEVs, before exclusively producing the Bolt as of 2021. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via averaging two government NAICS 3361 datasets: the U.S. Quarterly Workforce Indicators (QWI) and the U.S. Quarterly Census of Employment and Wages (QCEW). (d) Labor intensity in Oakland compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum.",
+    "footnote": [],
+    "bbox": [
+      [
+        205,
+        85,
+        832,
+        450
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5 (a) Vehicle production history in McLean County, Illinois. Before 2016, Mitsubishi owned an ICEV production plant in the town of Normal. The factory has since been purchased by Rivian, which began producing electric SUVs and pick-up trucks in 2021. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via news reports because government NAICS 3361 data was suppressed for McLean County. (d) Labor intensity in McLean compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum.",
+    "footnote": [],
+    "bbox": [
+      [
+        203,
+        85,
+        832,
+        450
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Fig. A1 U.S.-level vehicle production (a), assembly workers (b), and labor intensity (c). Vehicle production data was obtained from the Automotive News Research & Data Center, while employment data was obtained from QCEW.",
+    "footnote": [],
+    "bbox": [
+      [
+        283,
+        128,
+        666,
+        545
+      ]
+    ],
+    "page_idx": 23
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "Fig. A2 Battery cell and pack manufacturing increases the labor intensity of making BEVs. (a) Annual vehicle production volume in Alameda, CA. (b) Employment in Alameda, CA and Sparks, NV, based on news reports (see Table A2). Employment in Sparks, NV, reflects additional workers for battery cell and pack manufacturing at Tesla Gigafactory 1. Data is shown only when Alameda and Sparks data are available for the same production year. (c) Comparison of labor intensity (WPV) with and without including workers from Sparks, NV.",
+    "footnote": [],
+    "bbox": [
+      [
+        234,
+        102,
+        792,
+        416
+      ]
+    ],
+    "page_idx": 24
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "Fig. A3 Comparison of vehicle sales prices (MSRP) in Alameda (a), Oakland (b), and McLean (c). Dollar values are inflation-adjusted to 2023 dollars based on the Consumer Price Index (CPI) for U.S. new vehicles [61]. MSRP for all available vehicle trims are shown.",
+    "footnote": [],
+    "bbox": [
+      [
+        205,
+        520,
+        833,
+        680
+      ]
+    ],
+    "page_idx": 24
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "Fig. A4 Concept illustration: vertical integration creates more workforce co-location at the site of vehicle assembly.",
+    "footnote": [],
+    "bbox": [
+      [
+        171,
+        140,
+        788,
+        321
+      ]
+    ],
+    "page_idx": 25
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "Fig. B5 Comparison of engine manufacturing parts jobs in 2022 against projected battery cell manufacturing jobs assuming 100% BEV uptake and under various assumptions of jobs per GWh.",
+    "footnote": [],
+    "bbox": [
+      [
+        268,
+        263,
+        763,
+        533
+      ]
+    ],
+    "page_idx": 28
+  }
+]

preprint/preprint__09af231471af1cc00448f4d0e910132c71d4ce10f09b6fc16f3a37da4472e364/preprint__09af231471af1cc00448f4d0e910132c71d4ce10f09b6fc16f3a37da4472e364.mmd

@@ -0,0 +1,435 @@
+
+# "30% fewer workers for electric vehicle assembly": harbinger or myth?  
+
+Andrew Weng  
+
+asweng@umich.edu  
+
+University of Michigan  Omar Ahmed  University of Michigan  Gabriel Ehrlich  University of Michigan  Anna Stefanopoulou  University of Michigan  
+
+## Article  
+
+Keywords: electric vehicle, manufacturing jobs, labor intensity  
+
+Posted Date: April 12th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4237003/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on September 16th, 2024. See the published version at https://doi.org/10.1038/s41467- 024- 52435- x.
+
+<--- Page Split --->
+
+# "30% fewer workers for electric vehicle assembly": harbinger or myth?  
+
+Andrew Weng \(^{1\dagger}\) , Omar Y. Ahmed \(^{1\dagger}\) , Gabriel Ehrlich \(^{2}\) , Anna Stefanopoulou \(^{1*}\)  
+
+\(^{1}\) Department of Mechanical Engineering, University of Michigan - Ann Arbor, 1231 Beal Ave, Ann Arbor, 48109, MI, U.S. \(^{2}\) Department of Economics, University of Michigan - Ann Arbor, 611 Tappan Ave, Ann Arbor, 48109, MI, U.S.  
+
+\*Corresponding author(s). E- mail(s): annastef@umich.edu; Contributing authors: asweng@umich.edu; oyahmed@umich.edu; gehrlich@umich.edu; \(^{\dagger}\) These authors contributed equally to this work.  
+
+## Abstract  
+
+It has been widely hypothesized that the transition to battery electric vehicles will require \(30\%\) fewer assembly workers compared to internal combustion engine vehicles. This work uses publicly available datasets on vehicle production and employment to show that vehicle assembly plants in the U.S. that have previously assembled internal combustion vehicles but have since fully transitioned to assembling battery electric vehicles have required more, not fewer, workers to assemble the same number of vehicles. Our study suggests that widespread loss of employment at electric vehicle assembly sites is a smaller risk than many fear. Moreover, our study serves as a call for more regionally- focused analyses of the transition's effects on labor using data- driven and macro- level surveying approaches.  
+
+Keywords: electric vehicle, manufacturing jobs, labor intensity
+
+<--- Page Split --->
+
+## 1 Introduction  
+
+The automotive industry employs 13 million workers in the U.S., including nearly 1 million in the manufacturing sector [1- 3] (Figure 1). Most of these workers are engaged in the production of internal combustion engine vehicles (ICEVs) today, but a rapid shift towards battery electric vehicles (BEVs) is underway as major automakers set ambitious targets to phase out ICEV production within the next two decades [4, 5].  
+
+![](images/Figure_1.jpg)
+
+<center>Fig. 1 (a) U.S. automotive sector employment consists of assembly and parts manufacturing jobs, with assembly comprising \(33\%\) of total automotive jobs. During periods of economic downturn, the percentage of jobs lost in the automotive sector exceeded that of the overall U.S. manufacturing sector, as seen during the recession of 2008 and the pandemic of 2019. (b) The transition from ICEV to BEV production will create shifts in the types and quantities of jobs in both assembly and parts manufacturing. BEVs will require workers to manufacture battery cells and battery packs instead of engines. Jobs categories are classified based on the North American Industrial Classification System (NAICS) codes. Assembly jobs: NAICS 3361. Engine parts jobs: NAICS 33631. Powertrain parts jobs: NAICS 33635. Other parts: NAICS codes 3363(x) according to the parenthesized values in (b). </center>  
+
+How will the transition to BEV production affect the overall number of jobs in the automotive sector? The answer to this question is at the core of a "Just Transition" which secures the future and livelihoods of workers and their communities in the transition to a low- carbon economy [6- 10]. For many U.S. auto workers, the possibility of job loss is not theoretical but experienced. During the 2008 global recession, automotive manufacturing employment declined by \(23\%\) within a year (Figure 1a). The overall U.S. manufacturing sector lost \(12\%\) of employment over the same period, suggesting that the automotive sector is particularly vulnerable to job losses during periods of economic downturn. Although some industry analysts note the potential for the transition to BEVs to create new U.S. jobs [11], the potential for reduced labor demand in the transition to BEV production has raised concerns among automotive labor groups that the transition may be disruptive for U.S. workers [6].  
+
+Despite the importance of understanding the BEV transition's effect on the number of automotive jobs, existing reports have been scarce and contradictory. A common
+
+<--- Page Split --->
+
+narrative is that BEV powertrains have fewer parts compared to ICEV powertrains and thus take fewer workers to assemble, so the BEV transition will result in a net loss in automotive jobs. The claim of “30% fewer workers for EV assembly” entered public discourse as early as 2017 [12] and remains a central claim as part of ongoing debates on the effect of the BEV transition on jobs [6, 13–16]. The Economic Policy Institute found that, without policies promoting local production of electric vehicle powertrain components, 75,000 jobs could be lost in the U.S. by 2030. However, the same report also predicted that employment could rise by 150,000 jobs given local production [17]. A study by the Boston Consulting Group found that BEV labor requirements are about 1% less than those for ICEVs after accounting for all production process differences [18]. A study by Cotterman et al. (2022) found that the labor intensity required for BEV powertrain manufacturing can be more than twice as high as that for ICEVs if battery cell manufacturing is included and when industry shop floor data is used instead of academic models [19, 20]. The lack of consensus from these existing reports underscores the difficulty of estimating the future trajectory of BEV jobs based solely on technical assumptions [21] and expert judgment [22].  
+
+This work studies data from existing vehicle assembly plants based in the U.S. that have already fully transitioned from ICEV production to BEV production. By studying data from existing assembly plants, we show the effect of the ongoing BEV transition on automotive assembly jobs in the U.S. without making a priori assumptions of labor intensity, battery manufacturing location, and reliance on expert judgment. In all three assembly plants studied, we found that BEVs require more workers per vehicle produced than ICEVs.  
+
+## 2 Identifying BEV transition plants  
+
+For this work, we first identified U.S. counties in which there existed a single historic ICEV assembly plant that has since been converted to produce BEVs (Figure 2a). These sites, termed “transition plants,” provide the most direct comparison of labor intensity differences before, during, and after the BEV transition. Two attributes define transition plants. First, the plant must have fully transitioned from the assembly of ICEVs to BEVs. A full transition ensures that the latest labor intensity figures correspond to BEV production without considering the effect of simultaneous production of BEVs and ICEVs. Second, the plant must have been producing BEVs for at least two years and at a volume of more than 10,000 vehicles per year. This helps to exclude BEV assembly plants that are in the very early stages of production ramp- up, where production volumes are far from equilibrium conditions and with scarce data.  
+
+To understand the effect of the BEV transition on the workforce size at each transition plant, we frame the transition as occurring in several stages, each of which bears consequences for the workforce size at the plant (Figure 2b). In the first stage, ICEV lines are retired, resulting in lower vehicle production volumes and a reduced workforce. In the second stage, assembly lines are re- tooled for BEV production, renewing the demand for workers. In the third and final stage, the BEV assembly plant approaches a steady- state in operational capabilities and production volumes. At this stage, the workforce size is expected to stabilize.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2 Transition plants: vehicle assembly plants that have fully transitioned from assembling ICEVs to BEVs. (a) Map of U.S. automotive vehicle assembly activity as of 2022. Colors show the number of workers classified under "motor vehicle manufacturing" (NAICS code 3361) for each state. Markers highlight major manufacturing plants in each state. Three "transition plants" have been identified as examples of complete ICEV to BEV production transformations: (1) Alameda County, CA, representing the transition of the former New United Motors Manufacturing, Inc. (NUMMI) vehicle assembly plant, which assembled ICEV passenger vehicles, to the Tesla plant which assembles BEVs; (2) Oakland County, Michigan, representing the transition from the production of General Motors ICEVs to the Chevy Bolt BEV over six years; and (3) McLean County, Illinois, representing the transition of the former Mitsubishi vehicle assembly plant to the Rivian plant which assembles electric pick-up trucks. (b) Possible trajectories of vehicle assembly workforce size that this work seeks to clarify. </center>  
+
+We identified three transition plants for this study: Alameda County in California (Figure 3), Oakland County in Michigan (Figure 4), and McLean County in Illinois (Figure 5). Alameda was chosen as the site of the historic New United Motor Manufacturing Incorporated (NUMMI) plant, a joint venture between General Motors and Toyota, which closed in 2010 and has subsequently been owned and operated by Tesla to produce BEVs. Alameda represents a 'near- steady- state' BEV assembly case: Tesla has now been producing BEVs from this site for over a decade, and its annual production volume of BEVs now exceeds that of the NUMMI plant at its peak. Oakland was next identified as home to the General Motors (GM) Orion assembly plant, which began producing the Chevy Bolt BEV in 2016 concurrently with ICEVs [23]. As of 2021 the plant was exclusively making the Bolt, before ending its production in December 2023 [24]. Oakland thus provides a case study in which the same workforce transitioned from making ICEVs to making BEVs over a period of five years. Finally, McLean was identified as the home to a former ICEV plant owned by Mitsubishi, which has since been taken over by Rivian to produce mass- market electric light- duty trucks. McLean represents the case of a burgeoning BEV manufacturer at the early stages of vehicle production ramp- up.
+
+<--- Page Split --->
+
+## 3 Understanding labor intensity through workers per vehicle  
+
+At each transition plant, we study the labor intensity of vehicle assembly before, during, and after the transition to BEVs. Labor intensity is defined by the number of assembly workers needed to produce 1,000 vehicles (WPV) according to:  
+
+\[\mathrm{WPV}(k) = \frac{W(k)}{V(k)}\times 1000, \quad (1)\]  
+
+where \(W(k)\) is the number of auto assembly workers employed at the site averaged over four quarters of year \(k\) and \(V(k)\) is the total number of light- duty vehicles produced at the site during year \(k\) . WPV normalizes the workforce size by the production volume and is thus a measure of labor intensity. WPV can be converted to units of hours worked per vehicle by assuming a total annual hours worked per worker (see Section 8.3). However, since the hours worked are not publicly known at each transition plant, we chose to report labor intensity as WPV for this work.  
+
+Vehicle production data is obtained from Automotive News Research & Data Center [25], which details vehicle production volumes per make, model, and assembly location (see Section 8.1). Automotive assembly worker data reflects employment under the North American Industrial Classification System (NAICS) code 3361, Motor Vehicle Manufacturing. Worker data is corroborated by combining data from from two publicly available government databases - Quarterly Census of Employment and Wages (QCEW) and Quarterly Workforce Indicators (QWI) - as well as from local news reports (see Section 8.2). National vehicle production and employment data (Figure A1) provides a reference estimate of average ICEV labor intensity, since BEV production in the U.S. had not surpassed 7% as of 2022. In the past two decades, national labor intensity has ranged between approximately 17 and 28 WPV. Labor intensity reached its lowest point of 17 WPV in 2015 before steadily climbing toward a high of 28 WPV in 2021.  
+
+## 4 Alameda: high labor intensity despite a decade of BEV production  
+
+Alameda County, California, is home to the vehicle assembly plant historically owned by North America Motor Manufacturing Incorporated (NUMMI) [26]. The plant was in operation from 1984 to 2010 as a joint venture between General Motors and Toyota, producing ca. 429,000 vehicles per year at its peak. The plant produced midsize economy passenger vehicles (Toyota Corolla and Pontiac Vibe) and light- duty trucks (Toyota Tacoma) (Figure 3a). The plant closed in 2010 after General Motors pulled out of the partnership in the aftermath of the global recession [27, 28]. The plant closure resulted in the direct loss of 4,700 manufacturing jobs [27] as well as the closure of 34 businesses in Alameda that supplied parts to the factory [29].  
+
+Following the NUMMI plant closure, Tesla purchased the factory and began to retool the lines, first to produce the Model S sedan and Model X SUV starting in 2012
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Fig. 3 (a) Vehicle production history in Alameda County, California. Before 2010, the New United Motors Manufacturing Incorporated (NUMMI) factory produced ICEVs including the Corolla and the Tacoma. The factory has since been taken over by Tesla which has now been producing BEVs for over a decade. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via government NAICS 3361 data from the U.S. Quarterly Workforce Indicators (QWI), as well as via news reports. (d) Labor intensity in Alameda compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum. </center>  
+
+and 2015, respectively, followed by the Model 3 and Model Y mass- market vehicles starting in 2017 and 2020, respectively. By 2019, the annual vehicle production volume surpassed that of the NUMMI plant at its peak, and by 2022, the plant was producing more than 450,000 units per year (Figure 3b). In the years between 2012 and 2022, Tesla has maintained a net hiring rate of ca. 2,500 manufacturing workers per year, totaling over 25,000 workers as of 2022 (Figure 3c). These employment numbers are corroborated by both government data and local news reports (see Methods section).  
+
+The NUMMI factory reached peak labor efficiency, i.e. its lowest labor intensity, in 2006 at 15 WPV (Figure 3d). Since then, labor intensity has risen year- on- year as vehicle production volume declined. After Tesla acquired the plant in 2012, labor intensity stayed above 50 WPV over the next five years, coinciding with the production of Model S and Model X. Labor intensity then began to decrease starting in 2017,
+
+<--- Page Split --->
+
+when Tesla began production of the Model 3, its first mass- market BEV. During the period between 2019 and 2022, labor intensity averaged at 51 WPV. Overall, during the decade since Tesla began to build BEVs, the labor intensity stayed above 45 WPV.  
+
+To the best of our knowledge, the BEV labor force in Alameda includes labor for battery pack assembly for the Model S/X [30] but excludes battery pack assembly for the Model 3/Y which are reportedly manufactured off- site in Sparks, Nevada [31- 35]. This BEV labor force further excludes battery cell manufacturing since battery cells for the Model S and X are sourced from Panasonic in Japan [36], and cells for the Tesla Model 3 and Y are reportedly made in Sparks [31- 34]. Starting in 2017, Tesla also began to make Model 3 electric motors [35, 37] and battery packs in Sparks. The data after 2017 thus primarily reflects the labor intensity for assembling Model 3 and Y vehicles, excluding electric motor manufacturing, battery cell manufacturing, and battery pack manufacturing, with the exception of Model S/X battery packs which presumably continued to be made on- site. When battery manufacturing workers from Sparks, Nevada are included, labor intensity further increases to 67 WPV (Figure A2) in 2022, reflecting an additional 50% increase in labor intensity.  
+
+Overall, Alameda highlights one example of an ICEV to BEV transition in which each BEV took more than twice as many workers to assemble, even before considering the additional labor needed to manufacture battery cells and electric motors. Tesla, now with more than a decade of BEV production, has reached annual production volumes exceeding its former ICEV counterparts. Yet, labor intensity between 2019 and 2022 (51 WPV) remained more than double that of the NUMMI plant during its peak productivity year in 2006 (16 WPV).  
+
+## 5 Oakland: same plant owner, similar labor intensity  
+
+Oakland County, Michigan, is the home to the Orion Assembly plant owned by General Motors (GM). Before the 2008 recession, the plant produced the Chevy Malibu and Pontiac G6 passenger sedans with a production rate peaking at ca. 456,000 per year in 2004 (Figure 4a). Production numbers declined in the proceeding years, eventually reaching zero in 2008 when the plant was idled as GM declared bankruptcy [38] (Figure 4b). As the economy recovered from the recession, the Orion plant re- opened to produce the Buick Verano and Chevy Sonic passenger sedans [39, 40]. In 2016, GM began to convert its assembly lines to produce the Chevy Bolt BEV [23]. By 2021, the plant had transitioned to exclusively assembling the Bolt, with the production rate peaking at ca. 42,000 units per year. GM ended production of the Bolt in December 2023 with plans to renovate the Orion plant to make electric trucks starting in 2025 [24]. Since 2016, vehicle assembly employment in Oakland County mirrored the vehicle production rate: as vehicle production declined, so did employment (Figure 4c).  
+
+Data before 2016 reflects the labor intensity of ICEV assembly, which varied widely (Figure 4d). Before the 2008 recession, labor intensity varied between 24 WPV and 35 WPV. Following the factory shutdown in 2010, labor intensity decreased to 17 WPV. As the plant began to make the Chevy Bolt BEV in 2016, labor intensity began to rise. However, over the same period, the baseline national labor intensity (i.e. baseline
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Fig. 4 (a) Vehicle production history in Oakland County, Michigan, home to the Orion Township assembly plant owned by General Motors. Before the 2008 recession, plants in the cities of Pontiac and Wixom were also actively producing vehicles, but both plants shut down in 2010, leaving the Orion plant as the sole operating plant in the county. In 2016, Orion began producing the Chevy Bolt BEV alongside GM ICEVs, before exclusively producing the Bolt as of 2021. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via averaging two government NAICS 3361 datasets: the U.S. Quarterly Workforce Indicators (QWI) and the U.S. Quarterly Census of Employment and Wages (QCEW). (d) Labor intensity in Oakland compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum. </center>  
+
+ICEV labor intensity) rose by a similar amount, which we attribute to a general market shift towards larger vehicle types [41].  
+
+Oakland thus represents a case in which the transition to BEV assembly did not appreciably change the labor intensity trajectory compared to the rest of the U.S., which continued to assemble mostly ICEVs during the same period. However, while the labor intensity remained similar, the total number of jobs declined due to the reduced vehicle production volumes (Figure 4b,c). Finally, we note that the labor intensity corresponding to Chevy Bolt BEV assembly excludes battery pack assembly labor to
+
+<--- Page Split --->
+
+the best of our knowledge<sup>1</sup>. Battery cell manufacturing labor is also excluded since the battery cells are manufactured off- site<sup>2</sup>. The labor intensity for the Bolt is expected to increase if either battery pack assembly or cell manufacturing activity is included.  
+
+## 6 McLean: ten-fold increase in labor intensity during BEV factory production ramp  
+
+In this final case study, we turn to McLean County, Illinois, home to the former Mitsubishi vehicle assembly plant [44] (Figure 5). During Mitsubishi's ownership, vehicles produced included the Eclipse sedan and Outlander sport utility vehicle, with vehicle production plateauing ca. 69,000 vehicles in 2014. In the same year, the plant employed 1,250 workers. In 2015, Mitsubishi shut down operations due to global competitive pressures [45]. The plant was subsequently purchased by Rivian to be re- tooled for assembling electric pickup trucks [46]. In 2022, Rivian produced ca. 18,000 electric vehicles while employing 6,000 workers.  
+
+We thus calculate labor intensity in McLean to be 18 WPV during ICEV production in 2014, compared with 316 WPV during BEV production in 2022. The high labor intensity seen in 2022 mirrors the similarly high levels seen during Alameda's first five years of BEV production (Figure 3d). Both cases represent periods during which fledgling BEV makers undergo rapid production ramp- up and in which production levels have not yet reached a steady state. Rivian also reportedly manufactures battery packs on- site [47], contributing to an additional use of labor which is included in the labor intensity calculation. Note, however, that battery cell manufacturing is not factored into the labor intensity calculation to the best of our knowledge since Rivian is reportedly using cells supplied by Samsung that are not manufactured on- site [47].  
+
+## 7 Discussion  
+
+### 7.1 More workers for BEV assembly, not fewer  
+
+The three case studies in Alameda, Oakland, and McLean counties collectively suggest that each BEV requires just as many, if not more, workers to assemble than each ICEV. We summarize this finding in Table 1, which compares the labor intensity before and after the BEV transition at each transition plant. For each comparison, we report the numbers corresponding to the year with the highest peak labor productivity (the inverse of labor intensity). In Alameda, labor intensity rose three- fold from 16 WPV to 46 WPV after the BEV transition. In Oakland, labor intensity rose two- fold from 17 WPV to 31 WPV. Finally, in McLean, labor intensity rose from 18 WPV to 316 WPV. In all three cases, the labor intensity increased following the BEV transition.
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Fig. 5 (a) Vehicle production history in McLean County, Illinois. Before 2016, Mitsubishi owned an ICEV production plant in the town of Normal. The factory has since been purchased by Rivian, which began producing electric SUVs and pick-up trucks in 2021. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via news reports because government NAICS 3361 data was suppressed for McLean County. (d) Labor intensity in McLean compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum. </center>  
+
+Alameda highlighted a labor scenario for a maturing BEV assembly process. After ten years of BEV production, labor intensity in Alameda remained more than twice as high compared to its ICEV counterparts. McLean showed how labor intensity may increase ten- fold during the early years of vehicle production ramp towards volume production, especially for a new automaker. Oakland highlighted a case in which the labor intensity for making BEVs was no different from the baseline labor intensity for making ICEVs around the U.S. during the same time period.  
+
+### 7.2 Explaining higher labor intensity in BEV assembly  
+
+We postulate several factors that may influence the labor intensity of BEV assembly: the degree of investment in manufacturing technology development, the higher complexity of premium- class vehicles, and the extent of vertical integration. These factors
+
+<--- Page Split --->
+
+
+Table 1 More workers per vehicle needed for BEV assembly for all three counties studied. Production volume, employment numbers, and labor intensity correspond to the year of minimum labor intensity. The lists of vehicle models highlight certain high-volume production models and are not exhaustive. (\*) Battery pack assembly for Model S and X reportedly took place in Alameda [30]; Battery pack assembly for Model 3 and Y reportedly took place offsite in Gigafactory 1 located in Sparks, Nevada [31-35].   
+
+<table><tr><td>ICEV Assembly</td><td>BEV Assembly</td></tr><tr><td>Alameda, CA</td><td></td></tr><tr><td>Owner</td><td>NUMMI</td></tr><tr><td>Vehicle models</td><td>Tacoma, Corolla, Vibe</td></tr><tr><td>Peak productivity year</td><td>2006</td></tr><tr><td>Production volume</td><td>429,000</td></tr><tr><td>Employment</td><td>6,700</td></tr><tr><td>Labor intensity</td><td>16 WPV</td></tr><tr><td>Includes pack assembly?</td><td>—</td></tr><tr><td>Includes cell manuf.?</td><td>—</td></tr><tr><td>Oakland, MI</td><td></td></tr><tr><td>Owner</td><td>General Motors</td></tr><tr><td>Vehicle models</td><td>Sonic, Verano, Malibu</td></tr><tr><td>Peak productivity year</td><td>2013</td></tr><tr><td>Production volume</td><td>159,000</td></tr><tr><td>Employment</td><td>2,600</td></tr><tr><td>Labor intensity</td><td>17 WPV</td></tr><tr><td>Includes pack assembly?</td><td>—</td></tr><tr><td>Includes cell manuf.?</td><td>—</td></tr><tr><td>McLean, IL</td><td></td></tr><tr><td>Owner</td><td>Mitsubishi</td></tr><tr><td>Vehicle models</td><td>Outlander, Galant, Eclipse</td></tr><tr><td>Peak productivity year</td><td>2014</td></tr><tr><td>Production volume</td><td>69,000</td></tr><tr><td>Employment</td><td>1,300</td></tr><tr><td>Labor intensity</td><td>18 WPV</td></tr><tr><td>Includes pack assembly?</td><td>—</td></tr><tr><td>Includes cell manuf.?</td><td>—</td></tr></table>  
+
+may explain why the labor intensity for BEV assembly today exceeds the labor intensity of past ICEV assembly, and, furthermore, why labor intensity outcomes are not uniform but vary by region. We discuss each of these factors in turn.  
+
+Investment in manufacturing technology development. Increased levels of investment in technology development have the counter- intuitive effect of suppressing near- term labor productivity as companies hire a larger workforce to improve the technology. Such may be the case in Alameda, where Tesla's investment in BEV manufacturing technology [48] has resulted in a workforce of salaried engineers who are co- located within the assembly plant [49]. This claim is corroborated by government occupation data, which shows that engineering occupations accounted for \(7\%\) of total assembly employment in California in 2021 compared to \(5\%\) at the national level and \(3\%\) in Michigan (Table A1). Wages data tells a similar story: the average income of assembly workers in Alameda, CA, more than doubled from 2013 to 2021, indicating a growing presence of high- paid engineering occupations. Comparatively, in Oakland, MI, the average income increased by only \(20\%\) over the same period. We note that
+
+<--- Page Split --->
+
+the presence of the engineering and other non- production workforce does not appear to be the major driver of the increase in labor intensity of BEV assembly compared to ICEV assembly in Alameda. Even under the extreme assumption that all workers in Alameda prior to the BEV transition were strictly in production roles, Alameda's BEV labor intensity in 2021 would remain over 30 WPV, twice as high as during peak ICEV production in 2006<sup>3</sup>.  
+
+Vehicle complexity. First- time BEV makers tend to produce and sell premium- class vehicles before they produce and sell economy- class vehicles. For example, the first mass- produced BEVs from Tesla (Model S) and Rivian (R1T) are both premium- class vehicles that sell for more than \$80,000 USD<sup>4</sup>. By comparison, before the BEV transition, the median sales price of ICEVs produced at the same plants was only \$28,000 USD, since production was dominated by mass- market vehicles such as the Toyota Tacoma and the Mitsubishi Outlander (Figure A3). The present- day bias towards more premium- class BEVs may thus partly explain the higher labor intensity observed for BEV assembly currently. As BEV manufacturers move towards offering more economy- class BEVs, labor intensity may yet decrease. This trend may already be seen in Oakland, where the Chevy Bolt EV's median sales price nearly matched that of its ICEV predecessors, as did labor intensity.  
+
+Vertical integration. The rise in BEV labor productivity also parallels a growing industry trend towards consolidating workers who have historically worked off- site (i.e. at the site of parts suppliers) but who now work within the assembly plant (Figure A4). This consolidation is most clearly evident in Alameda, where Tesla has chosen to design and manufacture many vehicle components in- house, including electric motors [50], semiconductors [30], and seats [51]. In contrast, legacy automakers, over a period between the late 1990s and early 2000s, have opted for a strategy of outsourcing in which many vehicle components, such as the engine and transmission, are designed and manufactured off- site by parts suppliers [48, 52]. Consequently, parts manufacturing workers (e.g. engine and transmission manufacturing workers) that supported vehicle assembly tended not to be co- located at the site of vehicle assembly [53]. A transition to BEVs does not necessarily guarantee that an automaker will move towards greater vertical integration. For example, in the Oakland assembly plant, GM reportedly assembled battery packs off- site, so the workers for battery pack assembly are not included from our measure of labor intensity for the Chevy Bolt.  
+
+### 7.3 Parts manufacturing  
+
+While this study focused on the trajectory of automotive assembly jobs, the fate of automotive parts manufacturing jobs warrants further study. Parts manufacturing jobs comprised 66% of all auto manufacturing jobs in the U.S. in 2022 (Figure 1). For some parts manufacturing activities such as electrical, steering, suspension, brakes, seats, and interior trim, worker demand will persist within the context of BEV production. However, disruption in transmission and engine- related parts manufacturing jobs is expected since these components are simplified or absent in BEV powertrains
+
+<--- Page Split --->
+
+[54]. Engine manufacturing jobs will especially be impacted, considering the lack of combustion engines in BEVs.  
+
+[54]. Engine manufacturing jobs will especially be impacted, considering the lack of combustion engines in BEVs.In the U.S., engine manufacturing jobs accounted for \(7\%\) of all U.S. auto manufacturing jobs in 2022, or 56,486 workers. With the BEV transition, these workers face job losses but also the opportunity for re- employment in other parts of the BEV value chain. The most immediate source of worker re- employment is in battery cell manufacturing which accounts for up to \(75\%\) of the total labor intensity for producing a BEV powertrain [19]. Employment data on existing BEV OEMs also suggests that battery cell manufacturing including pack assembly can increase labor needed per BEV by over \(50\%\) (Figure A2 and Table A2) \(^{5}\) . Ultimately, whether new jobs in battery cell manufacturing will replace lost jobs in automotive parts manufacturing depends on the labor intensity needed to make battery cells, the geographical co- location of jobs, and skills [55- 58], which should be further investigated (Figure B5).  
+
+### 7.4 Outlook  
+
+Our study challenges the narrative that BEVs require \(30\%\) less labor to assemble than ICEVs. The observed data in three transition plants indicates that BEVs are more labor intensive to assemble than ICEVs, rather than less. These results suggest that the path towards greater BEV manufacturing will require a workforce size at the assembly plant that matches or exceeds the size of the ICEV workforce. The demand for workers at BEV assembly sites is spurred by a continued need to innovate and improve upon existing BEV manufacturing technology, a drive towards greater vertical integration, and the present- day tendency towards the production of higher- cost BEVs. Our analysis suggests that rapid, widespread job displacement during the BEV transition is a smaller risk than many fear.  
+
+## 8 Methods  
+
+### 8.1 Vehicle Production Data  
+
+Vehicle production data \(V\) was obtained from the Automotive News Research & Data Center [25], which collates North American light- duty (i.e., passenger vehicles and pick- up trucks) vehicle sales, production, and inventory data on a monthly basis and organizes the data by automaker, vehicle make and model, and manufacturing plant location. Only vehicle production in the U.S. was considered for this work. Electric vehicle models were manually identified by cross- referencing publicly available lists of BEV makes and models. The counties in which manufacturing plants reside were manually identified using the counties' Federal Information Processing System (FIPS) codes to enable linking vehicle production data with county- level employment data (see Section 8.2).
+
+<--- Page Split --->
+
+### 8.2 Employment Data  
+
+County- level automotive manufacturing employment data \(W\) was sought from two government data sources, the Quarterly Census of Employment and Wages (QCEW) and the Quarterly Workforce Indicators (QWI), as well as from local news reports.  
+
+The QCEW dataset, administered by the U.S. Bureau of Labor Statistics (BLS), comprises a quarterly count of employment and wages for workers covered by unemployment insurance programs, which totals more than \(95\%\) of U.S. workers [59]. The data is aggregated and classified by industry according to the North American Industry Classification System (NAICS), and provided for county, metropolitan statistical area (MSA), state, and national levels. This dataset provides employment data at the level of an "establishment", defined as a single physical worksite engaged predominantly in one type of economic activity, e.g., making automotive vehicles.  
+
+The QWI dataset is administered by the U.S. Census Bureau and consists of administrative data from the Longitudinal Employer- Household Dynamics (LEHD) program, including Unemployment Insurance Wage Records, data from the Census Bureau, and data from the Office of Personnel Management (OPM). This allows the QWI dataset to provide both firm- level and worker- level data. The QWI data covers more than just those eligible for unemployment insurance benefits and includes all employers and their employees for which the administrative records are available. It provides detailed breakdowns by industry (using NAICS) and worker demographics (gender, age, education, race, and ethnicity), as well as earnings and various measures of job and worker flows. Unlike QCEW, which focuses on the establishment level, QWI produces insights into both the employer side (job creation, destruction, etc.) and employee side (turnover rates, accessions, separations) of labor dynamics.  
+
+This work sought national, state- level, and county- level employment data from both QCEW and QWI under NAICS code 3361, which corresponds to Motor Vehicle Manufacturing and encompasses assembly workers. Parts manufacturing labor is excluded from these counts because it has its own distinct NAICS code. Occupation data from the Occupational Employment and Wage Statistics (OEWS) program of BLS was used to confirm that approximately \(60 - 80\%\) of NAICS 3361 workers are in production occupations (Standard Occupational Classification code 51- 0000, which includes assemblers and fabricators), depending on the region of the U.S. For reference, Table A1 summarizes occupation- industry data for NAICS 3361 in Michigan, California, and the U.S.  
+
+As described in Section 8.3, NAICS 3361 data was sometimes suppressed at the county level from QCEW, QWI, or both databases. For this reason, local news reports, found via internet searches, provided another independent estimate of employment data for specific automotive factory sites. This data was used to substitute employment data if QCEW and QWI data were both suppressed, or to corroborate available QCEW and QWI data. Table 2 summarizes which data sources were used for each of the three counties.
+
+<--- Page Split --->
+
+
+<table><tr><td>County</td><td>QCEW (Gov)</td><td>QWI (Gov)</td><td>News</td><td>Notes</td></tr><tr><td>Alameda</td><td></td><td>✓</td><td>✓</td><td>QCEW data was suppressed</td></tr><tr><td>Oakland</td><td>✓</td><td>✓</td><td></td><td>Average of QCEW and QWI data was used</td></tr><tr><td>McLean</td><td></td><td></td><td>✓</td><td>QCEW and QWI data were both suppressed</td></tr></table>
+
+Table 2 Summary of data sources used to study employment in the three transition counties.  
+
+## 8.3 Limitations  
+
+Regional (state- level and county- level) employment data for a particular industry is often suppressed by QCEW and QWI databases to maintain employer anonymity. This most often occurs when a single large employer comprises a majority of the data for a particular industry in a particular region. See Table 2 for a description of the suppression instances relevant to this work. Using two government databases as well as local news reports ensured a sufficient level of redundancy to circumvent these suppression instances. In most cases, the employment numbers from these data sources agree with each other, improving confidence in the reported numbers.  
+
+Another limitation is that employment data corresponding to NAICS code 3361 covers both light- and heavy- duty vehicle manufacturing employment. The vehicle production data from Automotive News only covers light- duty vehicles, so combining the two datasets requires assuming a negligible volume of heavy- duty vehicle manufacturing presence within the selected counties. NAICS sub- code 33611 covers employment specifically for light- duty manufacturing, but that data is suppressed for the states and counties examined for this work.  
+
+Additionally, there are discontinuities in the WPV calculations shown in Fig. 3- 5. These discontinuities correspond to periods of zero and extremely low vehicle production. For Alameda and McLean, periods of zero vehicle production occurred during ownership transition from their respective ICE- making to BEV- making companies. For Oakland, this period occurred during the Great Recession. We also observed that in each county, such a small number of vehicles were produced in the first year resuming production that the WPV metric disproportionately inflated for that year. Thus, for each county, we discard labor intensity calculation for the time period in which zero vehicles were produced, plus the first subsequent year that production resumed.  
+
+We also acknowledge that another common metric for labor intensity is hours worked per vehicle. While the data on hours worked is available at the national level through surveys such as Current Employment Statistics (CES), county- level data is not made public by government agencies. For this reason, we chose to report labor intensity in units of workers per vehicle. From workers per vehicle, hours worked per vehicle can be approximated using the formula:  
+
+\[\mathrm{HPV} = \frac{W\times t}{V} \quad (2)\]  
+
+where HPV is the hours worked per vehicle, \(W\) is the number of workers, \(V\) is the total vehicles produced, and \(t\) is the annualized hours worked per worker. The time input \(t\) can be estimated to be 2,236 hours, assuming an average of 43 hours worked per week for vehicle manufacturing workers according to BLS [60].
+
+<--- Page Split --->
+
+## Resource Availability  
+
+Resource AvailabilityFurther information and requests should be directed to and will be fulfilled by Anna Stefanopoulou (annastef@umich.edu).  
+
+## Acknowledgements  
+
+AcknowledgementsThe authors thank Kristin Dziczek for her insights and guidance on the auto manufacturing landscape in Michigan. The authors also thank Katelyn Freese and Katerina Freudenberg for their assistance with processing vehicle production data. We also appreciate help from Rebecca Gao for editing the manuscript.  
+
+## Author Contributions  
+
+Author ContributionsA.W.: conceptualization; methodology; writing - original draft; writing - review and editing; visualization. O.Y.A: methodology; software; investigation; data curation; visualization; writing - original draft; writing - review and editing. G.E.: methodology; writing - review and editing. A.S: conceptualization; writing - review and editing; funding acquisition; project administration.  
+
+## Glossary of Terms  
+
+Glossary of TermsBEV Battery Electric VehicleGM General MotorsICEV Internal Combustion Engine VehicleOEWS Occupational Employment and Wage StatisticsNAICS North American Industry Classification SystemNUMMI New United Motors Manufacturing Inc.QCEW Quarterly Census of Employment and WagesQWI Quarterly Workforce Indicators
+
+<--- Page Split --->
+
+## References  
+
+[1] International Labor Organization: COVID- 19 and the automotive industry. Technical report, International Labour Organization, Sectoral Policies Department (April 2020)  
+
+[2] U.S. Bureau of Labor Statistics: Automotive Industry: Employment, Earnings, and Hours. https://www.bls.gov/iag/tgs/iagauto.htm. Accessed: 2023- 10- 16 (2023)  
+
+[3] Klier, T., Rubenstein, J.: Charging Ahead—The Electrification of the Auto Industry. https://www.chicagofed.org/publications/blogs/chicago- fed- insights/2021/charging- ahead- electrification- auto- industry. Accessed: 2024- 3- 13 (2021)  
+
+[4] International Energy Agency: Corporate strategy – Global EV Outlook 2023 – Analysis. https://www.iea.org/reports/global- ev- outlook- 2023/corporate- strategy. Accessed: 2023- 10- 24 (2023)  
+
+[5] Committee on Assessment of Technologies for Improving Fuel Economy of Light- Duty Vehicles- Phase: Assessment of technologies for improving Light- Duty vehicle fuel economy—2025- 2035. Technical report, National Academies of Sciences (2021)  
+
+[6] UAW Research Department: Taking the high road: Strategies for a fair EV future. Technical report, UAW (January 2020)  
+
+[7] Emden, J., Murphy, L.: COP26: A just transition? – workshop summary. Technical report, Institute for Public Policy Research (January 2022)  
+
+[8] Romero- Lankao, P., Rosner, N., Brandtner, C., Rea, C., Mejia- Montero, A., Pilo, F., Dokshin, F., Castan- Broto, V., Burch, S., Schnur, S.: A framework to centre justice in energy transition innovations. Nature Energy, 1- 7 (2023)  
+
+[9] Just Transition Initiative: A framework for just transitions. Technical report, Center for Strategic & International Studies, Climate Investment Funds (January 2021)  
+
+[10] Lim, J., Aklin, M., Frank, M.R.: Location is a major barrier for transferring US fossil fuel employment to green jobs. Nat. Commun. 14(1), 5711 (2023)  
+
+[11] Laska, A., Hughes- Cromwick, E.: Electric vehicles: Policies to help america lead. Technical report, Third Way (November 2022)  
+
+[12] Ford Motor Company: Ford motor company - CEO strategic update. Technical report, Ford Motor Company (October 2017)  
+
+[13] Vellequette, L.P.: VW accelerates electric push with more models, more production. Technical report, Automotive News (March 2019)
+
+<--- Page Split --->
+
+[14] Levin, T.: Ford is slashing thousands of jobs as it goes electric. experts say a tidal wave of layoffs will rock the industry as it undergoes a seismic shift. Business Insider (2022)  
+
+[15] Charette, R.N.: How EVs are reshaping labor markets. Technical report, IEEE Spectrum (January 2023)  
+
+[16] Fichera, A.: Trump autoworkers speech fact check: What of electric vehicles? The New York Times (2023)  
+
+[17] Barrett, J., Bivens, J.: The stakes for workers in how policymakers manage the coming shift to all- electric vehicles. Technical report, Economic Policy Institute (September 2021)  
+
+[18] Küpper, D., Kuhlmann, K., Tominaga, K., Arora, A., Schlageter, J.: Shifting gears in auto manufacturing. Technical report, Boston Consulting Group (September 2020)  
+
+[19] Cotterman, T., Fuchs, E.R.H., Whitefoot, K.: The transition to electrified vehicles: Evaluating the labor demand of manufacturing conventional versus battery electric vehicle powertrains (2022)  
+
+[20] Cotterman, T.L.: Technology transitions in the electricity and automotive sectors: Embracing political, social, and economic constraints. PhD thesis, Carnegie Mellon University, Ann Arbor, United States (2022)  
+
+[21] Sakti, A., Azevedo, I.M.L., Fuchs, E.R.H., Michalek, J.J., Gallagher, K.G., Whitacre, J.F.: Consistency and robustness of forecasting for emerging technologies: The case of li- ion batteries for electric vehicles. Energy Policy 106, 415- 426 (2017)  
+
+[22] Funk, P., Davis, A., Vaishnav, P., Dewitt, B., Fuchs, E.: Individual inconsistency and aggregate rationality: Overcoming inconsistencies in expert judgment at the technical frontier. Technol. Forecast. Soc. Change 155, 119984 (2020)  
+
+[23] Burden, M.: GM Orion readies for Chevy Bolt EV production. https://www.detroitnews.com/story/business/autos/general-motors/2016/09/13/gm- orion- readies- chevy- bolt- ev- production/90294594/. Accessed: 2024- 3- 4 (2016)  
+
+[24] Noble, B., Hall, K.: GM laying off hundreds of workers as Chevrolet Camaro, Bolt production end. https://www.detroitnews.com/story/business/autos/chrysler/2023/12/14/gm- layoffs- autoworkers- uaw- chevrolet- camaro- bolt- production- end/71923598007/. Accessed: 2024- 3- 4 (2023)  
+
+[25] Automotive News: Automotive News Research & Data Center. Title of the
+
+<--- Page Split --->
+
+publication associated with this dataset: Automotive News (2023)  
+
+[26] Austenfeld, J.R.B.: NUMMI - the great experiment. Technical report (2007)  
+
+[27] Shaiken, H.: Commitment is a Two- Way street: Toyota, california and NUMMI. Technical Report 202- 10, Institute for Research on Labor and Employment, UC Berkeley (2010)  
+
+[28] Troncoso, J.N.: End of the line: Reassembling the legacy of NUMMI, the american middle class in the era of globalization and recession. PhD thesis, University of California, Berkeley (2012)  
+
+[29] Johnston, A.: NUMMI, five years later: Picking up the pieces. https://www.kalw.org/show/crosscurrents/2015- 06- 01/ nummi- five- years- later- picking- up- the- pieces. Accessed: 2023- 10- 26 (2015)  
+
+[30] Cohen, J.: An Automotive Insider's Tour Of The Tesla Fremont Factory. https://cleantechnica.com/2014/06/25/automotive- insiders- tour- tesla- fremont- factory/. Accessed: 2023- 12- 5 (2014)  
+
+[31] Field, K.: Tesla Model 3 Battery Pack Cell Teardown Highlights Performance Improvements. https://cleantechnica.com/2019/01/28/ tesla- model- 3- battery- pack- cell- teardown- highlights- performance- improvements/. Accessed: 2023- 12- 4 (2019)  
+
+[32] Hawley, G.: Understanding Tesla's lithium ion batteries. https://evannex.com/blogs/news/understanding- teslas- lithium- ion- batteries. Accessed: 2023- 12- 4 (2023)  
+
+[33] Kane, M.: What Batteries Are Tesla Using In Its Electric Cars? https://insideevs.com/news/587455/batteries- tesla- using- electric- cars/. Accessed: 2023- 12- 4 (2022)  
+
+[34] The Tesla Team: Battery Cell Production Begins at the Gigafactory. https://www.tesla.com/blog/battery- cell- production- begins- gigafactory. Accessed: 2023- 12- 4 (2017)  
+
+[35] Korosec, K.: Tesla Will Make Electric Motors for the Model 3 at Its Massive Gigafactory. https://fortune.com/2017/01/17/ tesla- model- 3- motor- gigafactory/. Accessed: 2023- 12- 4 (2017)  
+
+[36] The Tesla Team: Panasonic Enters into Supply Agreement with Tesla Motors to Supply Automotive- Grade Battery Cells. https://www.tesla.com/blog/panasonic- enters- supply- agreement- tesla- motors- supply- automotivegrade- battery- c. Accessed: 2023- 12- 4 (2011)
+
+<--- Page Split --->
+
+[37] Lambert, F.: Tesla is building new 'drive unit production lines' at the Gigafactory, will not only manufacture battery packs. https://electrek.co/2016/10/15/tesla- drive- unit- production- lines- gigafactory- model- 3/. Accessed: 2023- 12- 5 (2016)  
+
+[38] Goldstein, A.: Janesville: An American Story. Simon & Schuster (2017)  
+
+[39] Vlasic, B.: With sonic, G.M. stands atomaking on its head. The New York Times (2011)  
+
+[40] Nadrowski, K.: Orion Assembly Site. https://web.archive.org/web/20110523235653/http://media.gm.com/content/media/us/en/news/news_detail.detail.brand.GM.html/content/Pages/news/Plant_Facts/Assembly/Orion. Accessed: 2023- 12- 5 (2011)  
+
+[41] Hula, A., Maguire, A., Bunker, A., Rojcek, T., Harrison, S.: The 2023 EPA automotive trends report. Technical Report EPA- 420- R- 23- 033, United States Environmental Protection Agency (December 2023)  
+
+[42] Motors, G.: Chevrolet Bolt EV Battery Production Resumes. https://news.gm.com/newsroom.detail.html/Pages/news/us/en/2021/sep/0920- bolt.html. Accessed: 2023- 12- 5 (2021)  
+
+[43] Motors, G.: Chevrolet Bolt EV Battery Production Resumes. https://media.chevrolet.com/media/us/en/chevrolet/home.detail.html/content/Pages/news/us/en/2021/sep/0920- bolt.html. Accessed: 2023- 12- 6 (2021)  
+
+[44] Motors, M.: Mitsubishi Motors North America, Inc. - Manufacturing Division. https://web.archive.org/web/20060506083640/http://media.mitsubishicars.com/detail?mid=MIT2004101847405&mime=ASC. Accessed: 2023- 12- 5 (2004)  
+
+[45] Yerak, B., Cancino, A.: Mitsubishi closing normal plant in illinois, ending U.S. production. Chicago Tribune (2015)  
+
+[46] The Detroit News: Rivian builds electric pickup truck and SUV. The Detroit News (2022)  
+
+[47] Weintraub, S.: Rivian R1T first drive: Easily the best pickup I've ever driven, both off- road and on. https://electrek.co/2021/09/28/rivian- r1t- first- drive- easily- the- best- pickup- ive- ever- driven- both- off- road- and- on/. Accessed: 2023- 12- 6 (2021)  
+
+[48] Cutcher- Gershenfeld, J., Brooks, D., Mulloy, M.: The decline and resurgence of the U.S. auto industry. Technical Report 399, Economic Policy Institute (May 2015)  
+
+[49] Furr, N., Dyer, J.: Lessons from tesla's approach to innovation. Harvard Business
+
+<--- Page Split --->
+
+Review (2020)  
+
+[50] Bellon, T., White, J.: Focus: Build or buy? automakers chasing tesla rethink dependence on suppliers. Reuters (2022)  
+
+[51] Field, K.: Behind The Scenes At Tesla's Seat Factory #CleanTechnica Field Trip. https://cleantechnica.com/2019/04/27/behind-the-scenes-at-teslas-seat- factory/. Accessed: 2023- 12- 4 (2019)  
+
+[52] Chen, Y., Chowdhury, S.D., Donada, C.: Mirroring hypothesis and integrality: Evidence from tesla motors. J. Eng. Tech. Manage. 54, 41- 55 (2019)  
+
+[53] Herrigel, G., Wittke, V.: Varieties of vertical disintegration: The global trend toward heterogeneous supply relations and the reproduction of difference in US and german manufacturing. Industry Studies Association 2004- 15 (2004)  
+
+[54] Ruyter, A., Weller, S., Henry, I., Raimnie, A., Bentley, G., Nielsen, B.: Enabling a just transition in automotive: Evidence from the west midlands and south australia. Technical report, The British Academy (June 2022)  
+
+[55] Krusemark, L., Ganguly, S., Harp, T., Kulicki, A., Smith, C., Prasad, K.V.: Examining workforce needs for north america: Battery industry education and training needs assessment (BIETNA). Technical report, Center for Automotive Research (2024)  
+
+[56] Combemale, C., Whitefoot, K.S., Ales, L., Fuchs, E.R.H.: Not all technological change is equal: how the separability of tasks mediates the effect of technology change on skill demand. Ind Corp Change 30(6), 1361- 1387 (2022)  
+
+[57] Weaver, A., Osterman, P.: Skill demands and mismatch in U.S. manufacturing. ILR Review 70(2), 275- 307 (2017)  
+
+[58] Cotterman, T., Fuchs, E.R.H., Small, M.J., Whitefoot, K.: The Transition to Electrified Vehicles: Implications for the Future of Automotive Manufacturing and Worker Skills and Occupations (2022)  
+
+[59] Sadeghi, A.: The births and deaths of business establishments in the united states. Mon. Labor Rev. December 2008(1), 1- 18 (2008)  
+
+[60] U.S. Bureau of Labor Statistics: Automotive Industry: Employment, Earnings, and Hours. https://www.bls.gov/iag/tgs/iagauto.htm. Accessed: 2023- 10- 26 (2022)  
+
+[61] Federal Reserve Bank of St. Louis: Consumer Price Index for All Urban Consumers: New Vehicles in U.S. City Average (2024)  
+
+[62] Campagnol, N., Pfeiffer, A., Tryggestad, C.: Capturing the battery value- chain opportunity. Technical Report 1, McKinsey & Company (January 2022)
+
+<--- Page Split --->
+
+[63] Lambert, F.: Tesla Gigafactory 1 now employs over 3,000 workers as it becomes biggest battery factory in the world. https://electrek.co/2018/08/21/tesla- gigafactory- 1- 3000- workers/. Accessed: 2023- 12- 11 (2018)  
+
+[64] Knehr, K.W., Kubal, J.J., Nelson, P.A., Ahmed, S.: Battery performance and cost modeling for electric vehicles - a manual for BatPaC v5.0. Technical Report ANL/CSE- 22/1, Argonne National Laboratory (July 2022)
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+
+<center>Fig. A1 U.S.-level vehicle production (a), assembly workers (b), and labor intensity (c). Vehicle production data was obtained from the Automotive News Research & Data Center, while employment data was obtained from QCEW. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>Fig. A2 Battery cell and pack manufacturing increases the labor intensity of making BEVs. (a) Annual vehicle production volume in Alameda, CA. (b) Employment in Alameda, CA and Sparks, NV, based on news reports (see Table A2). Employment in Sparks, NV, reflects additional workers for battery cell and pack manufacturing at Tesla Gigafactory 1. Data is shown only when Alameda and Sparks data are available for the same production year. (c) Comparison of labor intensity (WPV) with and without including workers from Sparks, NV. </center>  
+
+![](images/Figure_unknown_2.jpg)
+
+<center>Fig. A3 Comparison of vehicle sales prices (MSRP) in Alameda (a), Oakland (b), and McLean (c). Dollar values are inflation-adjusted to 2023 dollars based on the Consumer Price Index (CPI) for U.S. new vehicles [61]. MSRP for all available vehicle trims are shown. </center>
+
+<--- Page Split --->
+![](images/Figure_unknown_3.jpg)
+
+<center>Fig. A4 Concept illustration: vertical integration creates more workforce co-location at the site of vehicle assembly. </center>  
+
+<table><tr><td></td><td>Location</td><td>2013</td><td>2021</td></tr><tr><td rowspan="2">% of NAICS 3361 workers in production</td><td>California (State)</td><td>66%</td><td>62%</td></tr><tr><td>Michigan (State)</td><td>74%</td><td>81%</td></tr><tr><td></td><td>U.S.</td><td>74%</td><td>76%</td></tr><tr><td rowspan="2">% of NAICS 3361 workers in engineering</td><td>California (State)</td><td>4%</td><td>7%</td></tr><tr><td>Michigan (State)</td><td>5%</td><td>3%</td></tr><tr><td></td><td>U.S.</td><td>5%</td><td>5%</td></tr><tr><td rowspan="2">NAICS 3361 average monthly pay</td><td>Alameda, CA</td><td>$6,243</td><td>$16,462</td></tr><tr><td>Oakland, MI</td><td>$7,557</td><td>$8,907</td></tr><tr><td></td><td>U.S.</td><td>$6,660</td><td>$6,864</td></tr></table>  
+
+Table A1 Proportion of NAICS 3361 workers in production (SOC code 51- 0000) and architecture/engineering occupations (SOC code 17- 0000), and average monthly pay of NAICS 3361 workers for California, Michigan, and the U.S. Occupation data was obtained from the Occupational Employment and Wage Statistics (OEWS). Income data for Alameda, CA was obtained from QWI. Income data for Oakland, MI was obtained from averaging QWI and QCEW data. Income data for the U.S. was obtained from QCEW.
+
+<--- Page Split --->
+
+
+<table><tr><td>Location</td><td>Date</td><td>News Source</td><td>Reported Employment</td></tr><tr><td rowspan="6">Tesla (Alameda)</td><td>Jun 2012</td><td>SFGATE</td><td>1,000</td></tr><tr><td>Jul 2013</td><td>Wired</td><td>3,000</td></tr><tr><td>Jun 2016</td><td>TheCountryCaller</td><td>6,000</td></tr><tr><td>Oct 2017</td><td>The Mercury News</td><td>10,000</td></tr><tr><td>Mar 2019</td><td>Forbes</td><td>15,000</td></tr><tr><td>Jun 2022</td><td>Tesla</td><td>22,000</td></tr><tr><td rowspan="3">Tesla/PENA (Sparks)</td><td>2016</td><td>Electrek</td><td>850</td></tr><tr><td>2017</td><td>Electrek</td><td>3,249</td></tr><tr><td>2018</td><td>The Associated Press</td><td>7,059</td></tr><tr><td></td><td>2022</td><td>Tesla</td><td>12,000</td></tr><tr><td rowspan="3">NUMMI (Alameda)</td><td>Jan 2002</td><td>SFGATE</td><td>5,739</td></tr><tr><td>Mar 2006</td><td>East Bay Times</td><td>5,500</td></tr><tr><td>Apr 2010</td><td>Recordnet.com</td><td>4,700</td></tr><tr><td rowspan="4">Rivian (Normal)</td><td>Oct 2021</td><td>WGLT</td><td>3,000</td></tr><tr><td>Apr 2022</td><td>CIPROUD</td><td>5,000</td></tr><tr><td>Jun 2022</td><td>Energy News Network</td><td>5,600</td></tr><tr><td>Jul 2022</td><td>CIPROUD</td><td>6,000</td></tr><tr><td></td><td>Mar 2023</td><td>WGLT</td><td>7,400</td></tr><tr><td rowspan="4">Mitsubishi (Normal)</td><td>2004</td><td>Chicago Tribune</td><td>3,150</td></tr><tr><td>2014</td><td>Local Wiki</td><td>1,250</td></tr><tr><td>2015</td><td>Chicago Tribune</td><td>1,280</td></tr><tr><td>2016</td><td>WQAD8</td><td>1,200</td></tr><tr><td rowspan="3">GM (Orion)</td><td>2013</td><td>CarGroup.org</td><td>2,561</td></tr><tr><td>2022</td><td>GM</td><td>1,238</td></tr><tr><td>2023</td><td>Wards Auto</td><td>1,270</td></tr></table>
+
+Table A2 List of news reports used to corroborate factory employment numbers. PENA: Panasonic Energy of North America.
+
+<--- Page Split --->
+
+## Appendix B Workers per GWh  
+
+McKinsey reported that, on average, new battery factories add approximately 80 jobs for every GWh of capacity, i.e. 80 workers per GWh [62]. This number carries some uncertainty since differences in value- chain coverage, e.g. battery- cell production only versus local module and pack production or co- location of R&D facilities, are unclear.  
+
+Tesla's Gigafactory 1 reportedly employed 3,249 people when the factory was producing 20 GWh of annual output [63]. Among these workers, 1,201 were employed by Panasonic, the main battery cell manufacturer, 93 are employed by Heitkamp & Thumann Group (H&T), a battery cell can supplier, and 1,955 were employed by Tesla. Assuming those employed by Panasonic and H&T are responsible for battery cell manufacturing, we infer that 1,294 workers are involved with producing 20 GWh of annual output, or 65 workers per GWh. If the employees from Tesla are included, then the calculation yields 162 workers per GWh.  
+
+The BatPaC v5.0 baseline factory model reported an annual labor of 3,876,000 hours per year to produce 50 GWh of output [64]. Assuming each worker works 2,236 hours per year and (equivalent to a 43- hour work- week, the U.S. average for automotive manufacturing [60]), this amounts to 35 workers per GWh.  
+
+Cotterman et al. [19] reported labor intensity per BEV powertrain assuming a 60kWh battery pack which varied depending on the data source and whether the labor was broken down between cell and pack/module assembly. For data sources where this breakdown was available, labor intensity ranged between 11 to 16 hours per 60kWh for industry data sources and 6 to 15 hours per 60kWh for public data sources. Assuming again 2,236 hours per year worked per worker, the range of labor demand is equivalent to 44 to 119 workers per GWh.
+
+<--- Page Split --->
+![](images/Figure_unknown_4.jpg)
+
+<center>Fig. B5 Comparison of engine manufacturing parts jobs in 2022 against projected battery cell manufacturing jobs assuming 100% BEV uptake and under various assumptions of jobs per GWh. </center>
+
+<--- Page Split --->

preprint/preprint__09af231471af1cc00448f4d0e910132c71d4ce10f09b6fc16f3a37da4472e364/preprint__09af231471af1cc00448f4d0e910132c71d4ce10f09b6fc16f3a37da4472e364_det.mmd

@@ -0,0 +1,585 @@
+<|ref|>title<|/ref|><|det|>[[43, 107, 912, 177]]<|/det|>
+# "30% fewer workers for electric vehicle assembly": harbinger or myth?  
+
+<|ref|>text<|/ref|><|det|>[[43, 196, 175, 214]]<|/det|>
+Andrew Weng  
+
+<|ref|>text<|/ref|><|det|>[[53, 223, 234, 240]]<|/det|>
+asweng@umich.edu  
+
+<|ref|>text<|/ref|><|det|>[[42, 269, 252, 426]]<|/det|>
+University of Michigan  Omar Ahmed  University of Michigan  Gabriel Ehrlich  University of Michigan  Anna Stefanopoulou  University of Michigan  
+
+<|ref|>sub_title<|/ref|><|det|>[[43, 469, 103, 486]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[42, 506, 580, 526]]<|/det|>
+Keywords: electric vehicle, manufacturing jobs, labor intensity  
+
+<|ref|>text<|/ref|><|det|>[[44, 545, 300, 563]]<|/det|>
+Posted Date: April 12th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[42, 583, 475, 601]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4237003/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 619, 914, 662]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 680, 535, 700]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 736, 920, 779]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on September 16th, 2024. See the published version at https://doi.org/10.1038/s41467- 024- 52435- x.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[191, 157, 767, 208]]<|/det|>
+# "30% fewer workers for electric vehicle assembly": harbinger or myth?  
+
+<|ref|>text<|/ref|><|det|>[[234, 229, 718, 264]]<|/det|>
+Andrew Weng \(^{1\dagger}\) , Omar Y. Ahmed \(^{1\dagger}\) , Gabriel Ehrlich \(^{2}\) , Anna Stefanopoulou \(^{1*}\)  
+
+<|ref|>text<|/ref|><|det|>[[193, 273, 763, 339]]<|/det|>
+\(^{1}\) Department of Mechanical Engineering, University of Michigan - Ann Arbor, 1231 Beal Ave, Ann Arbor, 48109, MI, U.S. \(^{2}\) Department of Economics, University of Michigan - Ann Arbor, 611 Tappan Ave, Ann Arbor, 48109, MI, U.S.  
+
+<|ref|>text<|/ref|><|det|>[[216, 365, 737, 430]]<|/det|>
+\*Corresponding author(s). E- mail(s): annastef@umich.edu; Contributing authors: asweng@umich.edu; oyahmed@umich.edu; gehrlich@umich.edu; \(^{\dagger}\) These authors contributed equally to this work.  
+
+<|ref|>sub_title<|/ref|><|det|>[[443, 456, 512, 469]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[206, 473, 750, 604]]<|/det|>
+It has been widely hypothesized that the transition to battery electric vehicles will require \(30\%\) fewer assembly workers compared to internal combustion engine vehicles. This work uses publicly available datasets on vehicle production and employment to show that vehicle assembly plants in the U.S. that have previously assembled internal combustion vehicles but have since fully transitioned to assembling battery electric vehicles have required more, not fewer, workers to assemble the same number of vehicles. Our study suggests that widespread loss of employment at electric vehicle assembly sites is a smaller risk than many fear. Moreover, our study serves as a call for more regionally- focused analyses of the transition's effects on labor using data- driven and macro- level surveying approaches.  
+
+<|ref|>text<|/ref|><|det|>[[206, 615, 605, 628]]<|/det|>
+Keywords: electric vehicle, manufacturing jobs, labor intensity
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[208, 83, 384, 101]]<|/det|>
+## 1 Introduction  
+
+<|ref|>text<|/ref|><|det|>[[207, 111, 831, 184]]<|/det|>
+The automotive industry employs 13 million workers in the U.S., including nearly 1 million in the manufacturing sector [1- 3] (Figure 1). Most of these workers are engaged in the production of internal combustion engine vehicles (ICEVs) today, but a rapid shift towards battery electric vehicles (BEVs) is underway as major automakers set ambitious targets to phase out ICEV production within the next two decades [4, 5].  
+
+<|ref|>image<|/ref|><|det|>[[207, 204, 828, 400]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[206, 416, 832, 520]]<|/det|>
+<center>Fig. 1 (a) U.S. automotive sector employment consists of assembly and parts manufacturing jobs, with assembly comprising \(33\%\) of total automotive jobs. During periods of economic downturn, the percentage of jobs lost in the automotive sector exceeded that of the overall U.S. manufacturing sector, as seen during the recession of 2008 and the pandemic of 2019. (b) The transition from ICEV to BEV production will create shifts in the types and quantities of jobs in both assembly and parts manufacturing. BEVs will require workers to manufacture battery cells and battery packs instead of engines. Jobs categories are classified based on the North American Industrial Classification System (NAICS) codes. Assembly jobs: NAICS 3361. Engine parts jobs: NAICS 33631. Powertrain parts jobs: NAICS 33635. Other parts: NAICS codes 3363(x) according to the parenthesized values in (b). </center>  
+
+<|ref|>text<|/ref|><|det|>[[207, 540, 831, 712]]<|/det|>
+How will the transition to BEV production affect the overall number of jobs in the automotive sector? The answer to this question is at the core of a "Just Transition" which secures the future and livelihoods of workers and their communities in the transition to a low- carbon economy [6- 10]. For many U.S. auto workers, the possibility of job loss is not theoretical but experienced. During the 2008 global recession, automotive manufacturing employment declined by \(23\%\) within a year (Figure 1a). The overall U.S. manufacturing sector lost \(12\%\) of employment over the same period, suggesting that the automotive sector is particularly vulnerable to job losses during periods of economic downturn. Although some industry analysts note the potential for the transition to BEVs to create new U.S. jobs [11], the potential for reduced labor demand in the transition to BEV production has raised concerns among automotive labor groups that the transition may be disruptive for U.S. workers [6].  
+
+<|ref|>text<|/ref|><|det|>[[206, 712, 830, 740]]<|/det|>
+Despite the importance of understanding the BEV transition's effect on the number of automotive jobs, existing reports have been scarce and contradictory. A common
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[165, 87, 790, 316]]<|/det|>
+narrative is that BEV powertrains have fewer parts compared to ICEV powertrains and thus take fewer workers to assemble, so the BEV transition will result in a net loss in automotive jobs. The claim of “30% fewer workers for EV assembly” entered public discourse as early as 2017 [12] and remains a central claim as part of ongoing debates on the effect of the BEV transition on jobs [6, 13–16]. The Economic Policy Institute found that, without policies promoting local production of electric vehicle powertrain components, 75,000 jobs could be lost in the U.S. by 2030. However, the same report also predicted that employment could rise by 150,000 jobs given local production [17]. A study by the Boston Consulting Group found that BEV labor requirements are about 1% less than those for ICEVs after accounting for all production process differences [18]. A study by Cotterman et al. (2022) found that the labor intensity required for BEV powertrain manufacturing can be more than twice as high as that for ICEVs if battery cell manufacturing is included and when industry shop floor data is used instead of academic models [19, 20]. The lack of consensus from these existing reports underscores the difficulty of estimating the future trajectory of BEV jobs based solely on technical assumptions [21] and expert judgment [22].  
+
+<|ref|>text<|/ref|><|det|>[[165, 316, 790, 414]]<|/det|>
+This work studies data from existing vehicle assembly plants based in the U.S. that have already fully transitioned from ICEV production to BEV production. By studying data from existing assembly plants, we show the effect of the ongoing BEV transition on automotive assembly jobs in the U.S. without making a priori assumptions of labor intensity, battery manufacturing location, and reliance on expert judgment. In all three assembly plants studied, we found that BEVs require more workers per vehicle produced than ICEVs.  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 429, 593, 450]]<|/det|>
+## 2 Identifying BEV transition plants  
+
+<|ref|>text<|/ref|><|det|>[[165, 458, 790, 614]]<|/det|>
+For this work, we first identified U.S. counties in which there existed a single historic ICEV assembly plant that has since been converted to produce BEVs (Figure 2a). These sites, termed “transition plants,” provide the most direct comparison of labor intensity differences before, during, and after the BEV transition. Two attributes define transition plants. First, the plant must have fully transitioned from the assembly of ICEVs to BEVs. A full transition ensures that the latest labor intensity figures correspond to BEV production without considering the effect of simultaneous production of BEVs and ICEVs. Second, the plant must have been producing BEVs for at least two years and at a volume of more than 10,000 vehicles per year. This helps to exclude BEV assembly plants that are in the very early stages of production ramp- up, where production volumes are far from equilibrium conditions and with scarce data.  
+
+<|ref|>text<|/ref|><|det|>[[165, 615, 790, 730]]<|/det|>
+To understand the effect of the BEV transition on the workforce size at each transition plant, we frame the transition as occurring in several stages, each of which bears consequences for the workforce size at the plant (Figure 2b). In the first stage, ICEV lines are retired, resulting in lower vehicle production volumes and a reduced workforce. In the second stage, assembly lines are re- tooled for BEV production, renewing the demand for workers. In the third and final stage, the BEV assembly plant approaches a steady- state in operational capabilities and production volumes. At this stage, the workforce size is expected to stabilize.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[210, 88, 824, 260]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[206, 280, 833, 405]]<|/det|>
+<center>Fig. 2 Transition plants: vehicle assembly plants that have fully transitioned from assembling ICEVs to BEVs. (a) Map of U.S. automotive vehicle assembly activity as of 2022. Colors show the number of workers classified under "motor vehicle manufacturing" (NAICS code 3361) for each state. Markers highlight major manufacturing plants in each state. Three "transition plants" have been identified as examples of complete ICEV to BEV production transformations: (1) Alameda County, CA, representing the transition of the former New United Motors Manufacturing, Inc. (NUMMI) vehicle assembly plant, which assembled ICEV passenger vehicles, to the Tesla plant which assembles BEVs; (2) Oakland County, Michigan, representing the transition from the production of General Motors ICEVs to the Chevy Bolt BEV over six years; and (3) McLean County, Illinois, representing the transition of the former Mitsubishi vehicle assembly plant to the Rivian plant which assembles electric pick-up trucks. (b) Possible trajectories of vehicle assembly workforce size that this work seeks to clarify. </center>  
+
+<|ref|>text<|/ref|><|det|>[[207, 423, 832, 666]]<|/det|>
+We identified three transition plants for this study: Alameda County in California (Figure 3), Oakland County in Michigan (Figure 4), and McLean County in Illinois (Figure 5). Alameda was chosen as the site of the historic New United Motor Manufacturing Incorporated (NUMMI) plant, a joint venture between General Motors and Toyota, which closed in 2010 and has subsequently been owned and operated by Tesla to produce BEVs. Alameda represents a 'near- steady- state' BEV assembly case: Tesla has now been producing BEVs from this site for over a decade, and its annual production volume of BEVs now exceeds that of the NUMMI plant at its peak. Oakland was next identified as home to the General Motors (GM) Orion assembly plant, which began producing the Chevy Bolt BEV in 2016 concurrently with ICEVs [23]. As of 2021 the plant was exclusively making the Bolt, before ending its production in December 2023 [24]. Oakland thus provides a case study in which the same workforce transitioned from making ICEVs to making BEVs over a period of five years. Finally, McLean was identified as the home to a former ICEV plant owned by Mitsubishi, which has since been taken over by Rivian to produce mass- market electric light- duty trucks. McLean represents the case of a burgeoning BEV manufacturer at the early stages of vehicle production ramp- up.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[165, 81, 754, 121]]<|/det|>
+## 3 Understanding labor intensity through workers per vehicle  
+
+<|ref|>text<|/ref|><|det|>[[165, 130, 790, 174]]<|/det|>
+At each transition plant, we study the labor intensity of vehicle assembly before, during, and after the transition to BEVs. Labor intensity is defined by the number of assembly workers needed to produce 1,000 vehicles (WPV) according to:  
+
+<|ref|>equation<|/ref|><|det|>[[380, 185, 788, 218]]<|/det|>
+\[\mathrm{WPV}(k) = \frac{W(k)}{V(k)}\times 1000, \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[165, 230, 790, 330]]<|/det|>
+where \(W(k)\) is the number of auto assembly workers employed at the site averaged over four quarters of year \(k\) and \(V(k)\) is the total number of light- duty vehicles produced at the site during year \(k\) . WPV normalizes the workforce size by the production volume and is thus a measure of labor intensity. WPV can be converted to units of hours worked per vehicle by assuming a total annual hours worked per worker (see Section 8.3). However, since the hours worked are not publicly known at each transition plant, we chose to report labor intensity as WPV for this work.  
+
+<|ref|>text<|/ref|><|det|>[[165, 330, 790, 515]]<|/det|>
+Vehicle production data is obtained from Automotive News Research & Data Center [25], which details vehicle production volumes per make, model, and assembly location (see Section 8.1). Automotive assembly worker data reflects employment under the North American Industrial Classification System (NAICS) code 3361, Motor Vehicle Manufacturing. Worker data is corroborated by combining data from from two publicly available government databases - Quarterly Census of Employment and Wages (QCEW) and Quarterly Workforce Indicators (QWI) - as well as from local news reports (see Section 8.2). National vehicle production and employment data (Figure A1) provides a reference estimate of average ICEV labor intensity, since BEV production in the U.S. had not surpassed 7% as of 2022. In the past two decades, national labor intensity has ranged between approximately 17 and 28 WPV. Labor intensity reached its lowest point of 17 WPV in 2015 before steadily climbing toward a high of 28 WPV in 2021.  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 530, 785, 568]]<|/det|>
+## 4 Alameda: high labor intensity despite a decade of BEV production  
+
+<|ref|>text<|/ref|><|det|>[[165, 576, 790, 707]]<|/det|>
+Alameda County, California, is home to the vehicle assembly plant historically owned by North America Motor Manufacturing Incorporated (NUMMI) [26]. The plant was in operation from 1984 to 2010 as a joint venture between General Motors and Toyota, producing ca. 429,000 vehicles per year at its peak. The plant produced midsize economy passenger vehicles (Toyota Corolla and Pontiac Vibe) and light- duty trucks (Toyota Tacoma) (Figure 3a). The plant closed in 2010 after General Motors pulled out of the partnership in the aftermath of the global recession [27, 28]. The plant closure resulted in the direct loss of 4,700 manufacturing jobs [27] as well as the closure of 34 businesses in Alameda that supplied parts to the factory [29].  
+
+<|ref|>text<|/ref|><|det|>[[165, 707, 790, 736]]<|/det|>
+Following the NUMMI plant closure, Tesla purchased the factory and began to retool the lines, first to produce the Model S sedan and Model X SUV starting in 2012
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[208, 87, 832, 450]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[206, 465, 832, 548]]<|/det|>
+<center>Fig. 3 (a) Vehicle production history in Alameda County, California. Before 2010, the New United Motors Manufacturing Incorporated (NUMMI) factory produced ICEVs including the Corolla and the Tacoma. The factory has since been taken over by Tesla which has now been producing BEVs for over a decade. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via government NAICS 3361 data from the U.S. Quarterly Workforce Indicators (QWI), as well as via news reports. (d) Labor intensity in Alameda compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum. </center>  
+
+<|ref|>text<|/ref|><|det|>[[207, 563, 832, 666]]<|/det|>
+and 2015, respectively, followed by the Model 3 and Model Y mass- market vehicles starting in 2017 and 2020, respectively. By 2019, the annual vehicle production volume surpassed that of the NUMMI plant at its peak, and by 2022, the plant was producing more than 450,000 units per year (Figure 3b). In the years between 2012 and 2022, Tesla has maintained a net hiring rate of ca. 2,500 manufacturing workers per year, totaling over 25,000 workers as of 2022 (Figure 3c). These employment numbers are corroborated by both government data and local news reports (see Methods section).  
+
+<|ref|>text<|/ref|><|det|>[[207, 665, 832, 737]]<|/det|>
+The NUMMI factory reached peak labor efficiency, i.e. its lowest labor intensity, in 2006 at 15 WPV (Figure 3d). Since then, labor intensity has risen year- on- year as vehicle production volume declined. After Tesla acquired the plant in 2012, labor intensity stayed above 50 WPV over the next five years, coinciding with the production of Model S and Model X. Labor intensity then began to decrease starting in 2017,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[165, 86, 790, 129]]<|/det|>
+when Tesla began production of the Model 3, its first mass- market BEV. During the period between 2019 and 2022, labor intensity averaged at 51 WPV. Overall, during the decade since Tesla began to build BEVs, the labor intensity stayed above 45 WPV.  
+
+<|ref|>text<|/ref|><|det|>[[165, 130, 790, 315]]<|/det|>
+To the best of our knowledge, the BEV labor force in Alameda includes labor for battery pack assembly for the Model S/X [30] but excludes battery pack assembly for the Model 3/Y which are reportedly manufactured off- site in Sparks, Nevada [31- 35]. This BEV labor force further excludes battery cell manufacturing since battery cells for the Model S and X are sourced from Panasonic in Japan [36], and cells for the Tesla Model 3 and Y are reportedly made in Sparks [31- 34]. Starting in 2017, Tesla also began to make Model 3 electric motors [35, 37] and battery packs in Sparks. The data after 2017 thus primarily reflects the labor intensity for assembling Model 3 and Y vehicles, excluding electric motor manufacturing, battery cell manufacturing, and battery pack manufacturing, with the exception of Model S/X battery packs which presumably continued to be made on- site. When battery manufacturing workers from Sparks, Nevada are included, labor intensity further increases to 67 WPV (Figure A2) in 2022, reflecting an additional 50% increase in labor intensity.  
+
+<|ref|>text<|/ref|><|det|>[[165, 315, 790, 415]]<|/det|>
+Overall, Alameda highlights one example of an ICEV to BEV transition in which each BEV took more than twice as many workers to assemble, even before considering the additional labor needed to manufacture battery cells and electric motors. Tesla, now with more than a decade of BEV production, has reached annual production volumes exceeding its former ICEV counterparts. Yet, labor intensity between 2019 and 2022 (51 WPV) remained more than double that of the NUMMI plant during its peak productivity year in 2006 (16 WPV).  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 429, 789, 450]]<|/det|>
+## 5 Oakland: same plant owner, similar labor intensity  
+
+<|ref|>text<|/ref|><|det|>[[165, 458, 790, 644]]<|/det|>
+Oakland County, Michigan, is the home to the Orion Assembly plant owned by General Motors (GM). Before the 2008 recession, the plant produced the Chevy Malibu and Pontiac G6 passenger sedans with a production rate peaking at ca. 456,000 per year in 2004 (Figure 4a). Production numbers declined in the proceeding years, eventually reaching zero in 2008 when the plant was idled as GM declared bankruptcy [38] (Figure 4b). As the economy recovered from the recession, the Orion plant re- opened to produce the Buick Verano and Chevy Sonic passenger sedans [39, 40]. In 2016, GM began to convert its assembly lines to produce the Chevy Bolt BEV [23]. By 2021, the plant had transitioned to exclusively assembling the Bolt, with the production rate peaking at ca. 42,000 units per year. GM ended production of the Bolt in December 2023 with plans to renovate the Orion plant to make electric trucks starting in 2025 [24]. Since 2016, vehicle assembly employment in Oakland County mirrored the vehicle production rate: as vehicle production declined, so did employment (Figure 4c).  
+
+<|ref|>text<|/ref|><|det|>[[165, 644, 790, 715]]<|/det|>
+Data before 2016 reflects the labor intensity of ICEV assembly, which varied widely (Figure 4d). Before the 2008 recession, labor intensity varied between 24 WPV and 35 WPV. Following the factory shutdown in 2010, labor intensity decreased to 17 WPV. As the plant began to make the Chevy Bolt BEV in 2016, labor intensity began to rise. However, over the same period, the baseline national labor intensity (i.e. baseline
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[205, 85, 832, 450]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[205, 462, 832, 568]]<|/det|>
+<center>Fig. 4 (a) Vehicle production history in Oakland County, Michigan, home to the Orion Township assembly plant owned by General Motors. Before the 2008 recession, plants in the cities of Pontiac and Wixom were also actively producing vehicles, but both plants shut down in 2010, leaving the Orion plant as the sole operating plant in the county. In 2016, Orion began producing the Chevy Bolt BEV alongside GM ICEVs, before exclusively producing the Bolt as of 2021. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via averaging two government NAICS 3361 datasets: the U.S. Quarterly Workforce Indicators (QWI) and the U.S. Quarterly Census of Employment and Wages (QCEW). (d) Labor intensity in Oakland compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum. </center>  
+
+<|ref|>text<|/ref|><|det|>[[206, 584, 830, 613]]<|/det|>
+ICEV labor intensity) rose by a similar amount, which we attribute to a general market shift towards larger vehicle types [41].  
+
+<|ref|>text<|/ref|><|det|>[[206, 613, 832, 699]]<|/det|>
+Oakland thus represents a case in which the transition to BEV assembly did not appreciably change the labor intensity trajectory compared to the rest of the U.S., which continued to assemble mostly ICEVs during the same period. However, while the labor intensity remained similar, the total number of jobs declined due to the reduced vehicle production volumes (Figure 4b,c). Finally, we note that the labor intensity corresponding to Chevy Bolt BEV assembly excludes battery pack assembly labor to
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[165, 87, 790, 130]]<|/det|>
+the best of our knowledge<sup>1</sup>. Battery cell manufacturing labor is also excluded since the battery cells are manufactured off- site<sup>2</sup>. The labor intensity for the Bolt is expected to increase if either battery pack assembly or cell manufacturing activity is included.  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 144, 712, 184]]<|/det|>
+## 6 McLean: ten-fold increase in labor intensity during BEV factory production ramp  
+
+<|ref|>text<|/ref|><|det|>[[165, 192, 790, 306]]<|/det|>
+In this final case study, we turn to McLean County, Illinois, home to the former Mitsubishi vehicle assembly plant [44] (Figure 5). During Mitsubishi's ownership, vehicles produced included the Eclipse sedan and Outlander sport utility vehicle, with vehicle production plateauing ca. 69,000 vehicles in 2014. In the same year, the plant employed 1,250 workers. In 2015, Mitsubishi shut down operations due to global competitive pressures [45]. The plant was subsequently purchased by Rivian to be re- tooled for assembling electric pickup trucks [46]. In 2022, Rivian produced ca. 18,000 electric vehicles while employing 6,000 workers.  
+
+<|ref|>text<|/ref|><|det|>[[165, 307, 790, 450]]<|/det|>
+We thus calculate labor intensity in McLean to be 18 WPV during ICEV production in 2014, compared with 316 WPV during BEV production in 2022. The high labor intensity seen in 2022 mirrors the similarly high levels seen during Alameda's first five years of BEV production (Figure 3d). Both cases represent periods during which fledgling BEV makers undergo rapid production ramp- up and in which production levels have not yet reached a steady state. Rivian also reportedly manufactures battery packs on- site [47], contributing to an additional use of labor which is included in the labor intensity calculation. Note, however, that battery cell manufacturing is not factored into the labor intensity calculation to the best of our knowledge since Rivian is reportedly using cells supplied by Samsung that are not manufactured on- site [47].  
+
+<|ref|>sub_title<|/ref|><|det|>[[166, 464, 317, 483]]<|/det|>
+## 7 Discussion  
+
+<|ref|>sub_title<|/ref|><|det|>[[166, 493, 633, 510]]<|/det|>
+### 7.1 More workers for BEV assembly, not fewer  
+
+<|ref|>text<|/ref|><|det|>[[165, 516, 790, 644]]<|/det|>
+The three case studies in Alameda, Oakland, and McLean counties collectively suggest that each BEV requires just as many, if not more, workers to assemble than each ICEV. We summarize this finding in Table 1, which compares the labor intensity before and after the BEV transition at each transition plant. For each comparison, we report the numbers corresponding to the year with the highest peak labor productivity (the inverse of labor intensity). In Alameda, labor intensity rose three- fold from 16 WPV to 46 WPV after the BEV transition. In Oakland, labor intensity rose two- fold from 17 WPV to 31 WPV. Finally, in McLean, labor intensity rose from 18 WPV to 316 WPV. In all three cases, the labor intensity increased following the BEV transition.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[203, 85, 832, 450]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[206, 464, 832, 536]]<|/det|>
+<center>Fig. 5 (a) Vehicle production history in McLean County, Illinois. Before 2016, Mitsubishi owned an ICEV production plant in the town of Normal. The factory has since been purchased by Rivian, which began producing electric SUVs and pick-up trucks in 2021. (b) Annual vehicle production volumes. (c) Employment numbers, sourced via news reports because government NAICS 3361 data was suppressed for McLean County. (d) Labor intensity in McLean compared to the U.S. average, measured in workers needed to produce 1,000 vehicles per annum. </center>  
+
+<|ref|>text<|/ref|><|det|>[[207, 552, 832, 651]]<|/det|>
+Alameda highlighted a labor scenario for a maturing BEV assembly process. After ten years of BEV production, labor intensity in Alameda remained more than twice as high compared to its ICEV counterparts. McLean showed how labor intensity may increase ten- fold during the early years of vehicle production ramp towards volume production, especially for a new automaker. Oakland highlighted a case in which the labor intensity for making BEVs was no different from the baseline labor intensity for making ICEVs around the U.S. during the same time period.  
+
+<|ref|>sub_title<|/ref|><|det|>[[207, 666, 752, 683]]<|/det|>
+### 7.2 Explaining higher labor intensity in BEV assembly  
+
+<|ref|>text<|/ref|><|det|>[[207, 689, 832, 732]]<|/det|>
+We postulate several factors that may influence the labor intensity of BEV assembly: the degree of investment in manufacturing technology development, the higher complexity of premium- class vehicles, and the extent of vertical integration. These factors
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[237, 90, 718, 420]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[237, 421, 714, 502]]<|/det|>
+Table 1 More workers per vehicle needed for BEV assembly for all three counties studied. Production volume, employment numbers, and labor intensity correspond to the year of minimum labor intensity. The lists of vehicle models highlight certain high-volume production models and are not exhaustive. (\*) Battery pack assembly for Model S and X reportedly took place in Alameda [30]; Battery pack assembly for Model 3 and Y reportedly took place offsite in Gigafactory 1 located in Sparks, Nevada [31-35].   
+
+<table><tr><td>ICEV Assembly</td><td>BEV Assembly</td></tr><tr><td>Alameda, CA</td><td></td></tr><tr><td>Owner</td><td>NUMMI</td></tr><tr><td>Vehicle models</td><td>Tacoma, Corolla, Vibe</td></tr><tr><td>Peak productivity year</td><td>2006</td></tr><tr><td>Production volume</td><td>429,000</td></tr><tr><td>Employment</td><td>6,700</td></tr><tr><td>Labor intensity</td><td>16 WPV</td></tr><tr><td>Includes pack assembly?</td><td>—</td></tr><tr><td>Includes cell manuf.?</td><td>—</td></tr><tr><td>Oakland, MI</td><td></td></tr><tr><td>Owner</td><td>General Motors</td></tr><tr><td>Vehicle models</td><td>Sonic, Verano, Malibu</td></tr><tr><td>Peak productivity year</td><td>2013</td></tr><tr><td>Production volume</td><td>159,000</td></tr><tr><td>Employment</td><td>2,600</td></tr><tr><td>Labor intensity</td><td>17 WPV</td></tr><tr><td>Includes pack assembly?</td><td>—</td></tr><tr><td>Includes cell manuf.?</td><td>—</td></tr><tr><td>McLean, IL</td><td></td></tr><tr><td>Owner</td><td>Mitsubishi</td></tr><tr><td>Vehicle models</td><td>Outlander, Galant, Eclipse</td></tr><tr><td>Peak productivity year</td><td>2014</td></tr><tr><td>Production volume</td><td>69,000</td></tr><tr><td>Employment</td><td>1,300</td></tr><tr><td>Labor intensity</td><td>18 WPV</td></tr><tr><td>Includes pack assembly?</td><td>—</td></tr><tr><td>Includes cell manuf.?</td><td>—</td></tr></table>  
+
+<|ref|>text<|/ref|><|det|>[[165, 525, 790, 568]]<|/det|>
+may explain why the labor intensity for BEV assembly today exceeds the labor intensity of past ICEV assembly, and, furthermore, why labor intensity outcomes are not uniform but vary by region. We discuss each of these factors in turn.  
+
+<|ref|>text<|/ref|><|det|>[[165, 568, 790, 740]]<|/det|>
+Investment in manufacturing technology development. Increased levels of investment in technology development have the counter- intuitive effect of suppressing near- term labor productivity as companies hire a larger workforce to improve the technology. Such may be the case in Alameda, where Tesla's investment in BEV manufacturing technology [48] has resulted in a workforce of salaried engineers who are co- located within the assembly plant [49]. This claim is corroborated by government occupation data, which shows that engineering occupations accounted for \(7\%\) of total assembly employment in California in 2021 compared to \(5\%\) at the national level and \(3\%\) in Michigan (Table A1). Wages data tells a similar story: the average income of assembly workers in Alameda, CA, more than doubled from 2013 to 2021, indicating a growing presence of high- paid engineering occupations. Comparatively, in Oakland, MI, the average income increased by only \(20\%\) over the same period. We note that
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[207, 87, 831, 172]]<|/det|>
+the presence of the engineering and other non- production workforce does not appear to be the major driver of the increase in labor intensity of BEV assembly compared to ICEV assembly in Alameda. Even under the extreme assumption that all workers in Alameda prior to the BEV transition were strictly in production roles, Alameda's BEV labor intensity in 2021 would remain over 30 WPV, twice as high as during peak ICEV production in 2006<sup>3</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[207, 173, 831, 343]]<|/det|>
+Vehicle complexity. First- time BEV makers tend to produce and sell premium- class vehicles before they produce and sell economy- class vehicles. For example, the first mass- produced BEVs from Tesla (Model S) and Rivian (R1T) are both premium- class vehicles that sell for more than \$80,000 USD<sup>4</sup>. By comparison, before the BEV transition, the median sales price of ICEVs produced at the same plants was only \$28,000 USD, since production was dominated by mass- market vehicles such as the Toyota Tacoma and the Mitsubishi Outlander (Figure A3). The present- day bias towards more premium- class BEVs may thus partly explain the higher labor intensity observed for BEV assembly currently. As BEV manufacturers move towards offering more economy- class BEVs, labor intensity may yet decrease. This trend may already be seen in Oakland, where the Chevy Bolt EV's median sales price nearly matched that of its ICEV predecessors, as did labor intensity.  
+
+<|ref|>text<|/ref|><|det|>[[207, 344, 831, 556]]<|/det|>
+Vertical integration. The rise in BEV labor productivity also parallels a growing industry trend towards consolidating workers who have historically worked off- site (i.e. at the site of parts suppliers) but who now work within the assembly plant (Figure A4). This consolidation is most clearly evident in Alameda, where Tesla has chosen to design and manufacture many vehicle components in- house, including electric motors [50], semiconductors [30], and seats [51]. In contrast, legacy automakers, over a period between the late 1990s and early 2000s, have opted for a strategy of outsourcing in which many vehicle components, such as the engine and transmission, are designed and manufactured off- site by parts suppliers [48, 52]. Consequently, parts manufacturing workers (e.g. engine and transmission manufacturing workers) that supported vehicle assembly tended not to be co- located at the site of vehicle assembly [53]. A transition to BEVs does not necessarily guarantee that an automaker will move towards greater vertical integration. For example, in the Oakland assembly plant, GM reportedly assembled battery packs off- site, so the workers for battery pack assembly are not included from our measure of labor intensity for the Chevy Bolt.  
+
+<|ref|>sub_title<|/ref|><|det|>[[208, 569, 451, 585]]<|/det|>
+### 7.3 Parts manufacturing  
+
+<|ref|>text<|/ref|><|det|>[[207, 592, 831, 693]]<|/det|>
+While this study focused on the trajectory of automotive assembly jobs, the fate of automotive parts manufacturing jobs warrants further study. Parts manufacturing jobs comprised 66% of all auto manufacturing jobs in the U.S. in 2022 (Figure 1). For some parts manufacturing activities such as electrical, steering, suspension, brakes, seats, and interior trim, worker demand will persist within the context of BEV production. However, disruption in transmission and engine- related parts manufacturing jobs is expected since these components are simplified or absent in BEV powertrains
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[165, 86, 790, 115]]<|/det|>
+[54]. Engine manufacturing jobs will especially be impacted, considering the lack of combustion engines in BEVs.  
+
+<|ref|>text<|/ref|><|det|>[[165, 116, 790, 272]]<|/det|>
+[54]. Engine manufacturing jobs will especially be impacted, considering the lack of combustion engines in BEVs.In the U.S., engine manufacturing jobs accounted for \(7\%\) of all U.S. auto manufacturing jobs in 2022, or 56,486 workers. With the BEV transition, these workers face job losses but also the opportunity for re- employment in other parts of the BEV value chain. The most immediate source of worker re- employment is in battery cell manufacturing which accounts for up to \(75\%\) of the total labor intensity for producing a BEV powertrain [19]. Employment data on existing BEV OEMs also suggests that battery cell manufacturing including pack assembly can increase labor needed per BEV by over \(50\%\) (Figure A2 and Table A2) \(^{5}\) . Ultimately, whether new jobs in battery cell manufacturing will replace lost jobs in automotive parts manufacturing depends on the labor intensity needed to make battery cells, the geographical co- location of jobs, and skills [55- 58], which should be further investigated (Figure B5).  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 286, 286, 302]]<|/det|>
+### 7.4 Outlook  
+
+<|ref|>text<|/ref|><|det|>[[165, 309, 790, 452]]<|/det|>
+Our study challenges the narrative that BEVs require \(30\%\) less labor to assemble than ICEVs. The observed data in three transition plants indicates that BEVs are more labor intensive to assemble than ICEVs, rather than less. These results suggest that the path towards greater BEV manufacturing will require a workforce size at the assembly plant that matches or exceeds the size of the ICEV workforce. The demand for workers at BEV assembly sites is spurred by a continued need to innovate and improve upon existing BEV manufacturing technology, a drive towards greater vertical integration, and the present- day tendency towards the production of higher- cost BEVs. Our analysis suggests that rapid, widespread job displacement during the BEV transition is a smaller risk than many fear.  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 467, 297, 486]]<|/det|>
+## 8 Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 497, 452, 514]]<|/det|>
+### 8.1 Vehicle Production Data  
+
+<|ref|>text<|/ref|><|det|>[[165, 520, 790, 662]]<|/det|>
+Vehicle production data \(V\) was obtained from the Automotive News Research & Data Center [25], which collates North American light- duty (i.e., passenger vehicles and pick- up trucks) vehicle sales, production, and inventory data on a monthly basis and organizes the data by automaker, vehicle make and model, and manufacturing plant location. Only vehicle production in the U.S. was considered for this work. Electric vehicle models were manually identified by cross- referencing publicly available lists of BEV makes and models. The counties in which manufacturing plants reside were manually identified using the counties' Federal Information Processing System (FIPS) codes to enable linking vehicle production data with county- level employment data (see Section 8.2).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[207, 85, 429, 101]]<|/det|>
+### 8.2 Employment Data  
+
+<|ref|>text<|/ref|><|det|>[[207, 107, 831, 150]]<|/det|>
+County- level automotive manufacturing employment data \(W\) was sought from two government data sources, the Quarterly Census of Employment and Wages (QCEW) and the Quarterly Workforce Indicators (QWI), as well as from local news reports.  
+
+<|ref|>text<|/ref|><|det|>[[207, 151, 831, 265]]<|/det|>
+The QCEW dataset, administered by the U.S. Bureau of Labor Statistics (BLS), comprises a quarterly count of employment and wages for workers covered by unemployment insurance programs, which totals more than \(95\%\) of U.S. workers [59]. The data is aggregated and classified by industry according to the North American Industry Classification System (NAICS), and provided for county, metropolitan statistical area (MSA), state, and national levels. This dataset provides employment data at the level of an "establishment", defined as a single physical worksite engaged predominantly in one type of economic activity, e.g., making automotive vehicles.  
+
+<|ref|>text<|/ref|><|det|>[[207, 265, 832, 435]]<|/det|>
+The QWI dataset is administered by the U.S. Census Bureau and consists of administrative data from the Longitudinal Employer- Household Dynamics (LEHD) program, including Unemployment Insurance Wage Records, data from the Census Bureau, and data from the Office of Personnel Management (OPM). This allows the QWI dataset to provide both firm- level and worker- level data. The QWI data covers more than just those eligible for unemployment insurance benefits and includes all employers and their employees for which the administrative records are available. It provides detailed breakdowns by industry (using NAICS) and worker demographics (gender, age, education, race, and ethnicity), as well as earnings and various measures of job and worker flows. Unlike QCEW, which focuses on the establishment level, QWI produces insights into both the employer side (job creation, destruction, etc.) and employee side (turnover rates, accessions, separations) of labor dynamics.  
+
+<|ref|>text<|/ref|><|det|>[[207, 435, 832, 577]]<|/det|>
+This work sought national, state- level, and county- level employment data from both QCEW and QWI under NAICS code 3361, which corresponds to Motor Vehicle Manufacturing and encompasses assembly workers. Parts manufacturing labor is excluded from these counts because it has its own distinct NAICS code. Occupation data from the Occupational Employment and Wage Statistics (OEWS) program of BLS was used to confirm that approximately \(60 - 80\%\) of NAICS 3361 workers are in production occupations (Standard Occupational Classification code 51- 0000, which includes assemblers and fabricators), depending on the region of the U.S. For reference, Table A1 summarizes occupation- industry data for NAICS 3361 in Michigan, California, and the U.S.  
+
+<|ref|>text<|/ref|><|det|>[[207, 578, 832, 677]]<|/det|>
+As described in Section 8.3, NAICS 3361 data was sometimes suppressed at the county level from QCEW, QWI, or both databases. For this reason, local news reports, found via internet searches, provided another independent estimate of employment data for specific automotive factory sites. This data was used to substitute employment data if QCEW and QWI data were both suppressed, or to corroborate available QCEW and QWI data. Table 2 summarizes which data sources were used for each of the three counties.
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[165, 84, 792, 141]]<|/det|>
+
+<table><tr><td>County</td><td>QCEW (Gov)</td><td>QWI (Gov)</td><td>News</td><td>Notes</td></tr><tr><td>Alameda</td><td></td><td>✓</td><td>✓</td><td>QCEW data was suppressed</td></tr><tr><td>Oakland</td><td>✓</td><td>✓</td><td></td><td>Average of QCEW and QWI data was used</td></tr><tr><td>McLean</td><td></td><td></td><td>✓</td><td>QCEW and QWI data were both suppressed</td></tr></table>
+
+<|ref|>table_footnote<|/ref|><|det|>[[165, 145, 757, 157]]<|/det|>
+Table 2 Summary of data sources used to study employment in the three transition counties.  
+
+<|ref|>sub_title<|/ref|><|det|>[[165, 182, 320, 197]]<|/det|>
+## 8.3 Limitations  
+
+<|ref|>text<|/ref|><|det|>[[165, 204, 790, 318]]<|/det|>
+Regional (state- level and county- level) employment data for a particular industry is often suppressed by QCEW and QWI databases to maintain employer anonymity. This most often occurs when a single large employer comprises a majority of the data for a particular industry in a particular region. See Table 2 for a description of the suppression instances relevant to this work. Using two government databases as well as local news reports ensured a sufficient level of redundancy to circumvent these suppression instances. In most cases, the employment numbers from these data sources agree with each other, improving confidence in the reported numbers.  
+
+<|ref|>text<|/ref|><|det|>[[165, 318, 790, 418]]<|/det|>
+Another limitation is that employment data corresponding to NAICS code 3361 covers both light- and heavy- duty vehicle manufacturing employment. The vehicle production data from Automotive News only covers light- duty vehicles, so combining the two datasets requires assuming a negligible volume of heavy- duty vehicle manufacturing presence within the selected counties. NAICS sub- code 33611 covers employment specifically for light- duty manufacturing, but that data is suppressed for the states and counties examined for this work.  
+
+<|ref|>text<|/ref|><|det|>[[165, 418, 790, 547]]<|/det|>
+Additionally, there are discontinuities in the WPV calculations shown in Fig. 3- 5. These discontinuities correspond to periods of zero and extremely low vehicle production. For Alameda and McLean, periods of zero vehicle production occurred during ownership transition from their respective ICE- making to BEV- making companies. For Oakland, this period occurred during the Great Recession. We also observed that in each county, such a small number of vehicles were produced in the first year resuming production that the WPV metric disproportionately inflated for that year. Thus, for each county, we discard labor intensity calculation for the time period in which zero vehicles were produced, plus the first subsequent year that production resumed.  
+
+<|ref|>text<|/ref|><|det|>[[165, 547, 790, 632]]<|/det|>
+We also acknowledge that another common metric for labor intensity is hours worked per vehicle. While the data on hours worked is available at the national level through surveys such as Current Employment Statistics (CES), county- level data is not made public by government agencies. For this reason, we chose to report labor intensity in units of workers per vehicle. From workers per vehicle, hours worked per vehicle can be approximated using the formula:  
+
+<|ref|>equation<|/ref|><|det|>[[421, 641, 787, 672]]<|/det|>
+\[\mathrm{HPV} = \frac{W\times t}{V} \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[165, 682, 790, 740]]<|/det|>
+where HPV is the hours worked per vehicle, \(W\) is the number of workers, \(V\) is the total vehicles produced, and \(t\) is the annualized hours worked per worker. The time input \(t\) can be estimated to be 2,236 hours, assuming an average of 43 hours worked per week for vehicle manufacturing workers according to BLS [60].
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[207, 81, 460, 101]]<|/det|>
+## Resource Availability  
+
+<|ref|>text<|/ref|><|det|>[[207, 111, 831, 141]]<|/det|>
+Resource AvailabilityFurther information and requests should be directed to and will be fulfilled by Anna Stefanopoulou (annastef@umich.edu).  
+
+<|ref|>sub_title<|/ref|><|det|>[[208, 155, 433, 175]]<|/det|>
+## Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[207, 184, 831, 242]]<|/det|>
+AcknowledgementsThe authors thank Kristin Dziczek for her insights and guidance on the auto manufacturing landscape in Michigan. The authors also thank Katelyn Freese and Katerina Freudenberg for their assistance with processing vehicle production data. We also appreciate help from Rebecca Gao for editing the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[208, 256, 466, 275]]<|/det|>
+## Author Contributions  
+
+<|ref|>text<|/ref|><|det|>[[207, 284, 832, 356]]<|/det|>
+Author ContributionsA.W.: conceptualization; methodology; writing - original draft; writing - review and editing; visualization. O.Y.A: methodology; software; investigation; data curation; visualization; writing - original draft; writing - review and editing. G.E.: methodology; writing - review and editing. A.S: conceptualization; writing - review and editing; funding acquisition; project administration.  
+
+<|ref|>sub_title<|/ref|><|det|>[[207, 370, 424, 390]]<|/det|>
+## Glossary of Terms  
+
+<|ref|>text<|/ref|><|det|>[[216, 399, 650, 515]]<|/det|>
+Glossary of TermsBEV Battery Electric VehicleGM General MotorsICEV Internal Combustion Engine VehicleOEWS Occupational Employment and Wage StatisticsNAICS North American Industry Classification SystemNUMMI New United Motors Manufacturing Inc.QCEW Quarterly Census of Employment and WagesQWI Quarterly Workforce Indicators
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[165, 81, 295, 101]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[170, 110, 792, 152]]<|/det|>
+[1] International Labor Organization: COVID- 19 and the automotive industry. Technical report, International Labour Organization, Sectoral Policies Department (April 2020)  
+
+<|ref|>text<|/ref|><|det|>[[172, 164, 791, 209]]<|/det|>
+[2] U.S. Bureau of Labor Statistics: Automotive Industry: Employment, Earnings, and Hours. https://www.bls.gov/iag/tgs/iagauto.htm. Accessed: 2023- 10- 16 (2023)  
+
+<|ref|>text<|/ref|><|det|>[[172, 220, 791, 264]]<|/det|>
+[3] Klier, T., Rubenstein, J.: Charging Ahead—The Electrification of the Auto Industry. https://www.chicagofed.org/publications/blogs/chicago- fed- insights/2021/charging- ahead- electrification- auto- industry. Accessed: 2024- 3- 13 (2021)  
+
+<|ref|>text<|/ref|><|det|>[[172, 275, 791, 319]]<|/det|>
+[4] International Energy Agency: Corporate strategy – Global EV Outlook 2023 – Analysis. https://www.iea.org/reports/global- ev- outlook- 2023/corporate- strategy. Accessed: 2023- 10- 24 (2023)  
+
+<|ref|>text<|/ref|><|det|>[[172, 329, 791, 387]]<|/det|>
+[5] Committee on Assessment of Technologies for Improving Fuel Economy of Light- Duty Vehicles- Phase: Assessment of technologies for improving Light- Duty vehicle fuel economy—2025- 2035. Technical report, National Academies of Sciences (2021)  
+
+<|ref|>text<|/ref|><|det|>[[172, 398, 789, 428]]<|/det|>
+[6] UAW Research Department: Taking the high road: Strategies for a fair EV future. Technical report, UAW (January 2020)  
+
+<|ref|>text<|/ref|><|det|>[[172, 438, 789, 468]]<|/det|>
+[7] Emden, J., Murphy, L.: COP26: A just transition? – workshop summary. Technical report, Institute for Public Policy Research (January 2022)  
+
+<|ref|>text<|/ref|><|det|>[[172, 479, 791, 523]]<|/det|>
+[8] Romero- Lankao, P., Rosner, N., Brandtner, C., Rea, C., Mejia- Montero, A., Pilo, F., Dokshin, F., Castan- Broto, V., Burch, S., Schnur, S.: A framework to centre justice in energy transition innovations. Nature Energy, 1- 7 (2023)  
+
+<|ref|>text<|/ref|><|det|>[[172, 534, 790, 577]]<|/det|>
+[9] Just Transition Initiative: A framework for just transitions. Technical report, Center for Strategic & International Studies, Climate Investment Funds (January 2021)  
+
+<|ref|>text<|/ref|><|det|>[[165, 588, 790, 618]]<|/det|>
+[10] Lim, J., Aklin, M., Frank, M.R.: Location is a major barrier for transferring US fossil fuel employment to green jobs. Nat. Commun. 14(1), 5711 (2023)  
+
+<|ref|>text<|/ref|><|det|>[[165, 629, 789, 659]]<|/det|>
+[11] Laska, A., Hughes- Cromwick, E.: Electric vehicles: Policies to help america lead. Technical report, Third Way (November 2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 669, 790, 699]]<|/det|>
+[12] Ford Motor Company: Ford motor company - CEO strategic update. Technical report, Ford Motor Company (October 2017)  
+
+<|ref|>text<|/ref|><|det|>[[165, 710, 789, 739]]<|/det|>
+[13] Vellequette, L.P.: VW accelerates electric push with more models, more production. Technical report, Automotive News (March 2019)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[205, 85, 833, 129]]<|/det|>
+[14] Levin, T.: Ford is slashing thousands of jobs as it goes electric. experts say a tidal wave of layoffs will rock the industry as it undergoes a seismic shift. Business Insider (2022)  
+
+<|ref|>text<|/ref|><|det|>[[206, 141, 832, 170]]<|/det|>
+[15] Charette, R.N.: How EVs are reshaping labor markets. Technical report, IEEE Spectrum (January 2023)  
+
+<|ref|>text<|/ref|><|det|>[[206, 181, 832, 210]]<|/det|>
+[16] Fichera, A.: Trump autoworkers speech fact check: What of electric vehicles? The New York Times (2023)  
+
+<|ref|>text<|/ref|><|det|>[[206, 222, 832, 265]]<|/det|>
+[17] Barrett, J., Bivens, J.: The stakes for workers in how policymakers manage the coming shift to all- electric vehicles. Technical report, Economic Policy Institute (September 2021)  
+
+<|ref|>text<|/ref|><|det|>[[206, 277, 832, 320]]<|/det|>
+[18] Küpper, D., Kuhlmann, K., Tominaga, K., Arora, A., Schlageter, J.: Shifting gears in auto manufacturing. Technical report, Boston Consulting Group (September 2020)  
+
+<|ref|>text<|/ref|><|det|>[[206, 331, 832, 374]]<|/det|>
+[19] Cotterman, T., Fuchs, E.R.H., Whitefoot, K.: The transition to electrified vehicles: Evaluating the labor demand of manufacturing conventional versus battery electric vehicle powertrains (2022)  
+
+<|ref|>text<|/ref|><|det|>[[206, 386, 832, 429]]<|/det|>
+[20] Cotterman, T.L.: Technology transitions in the electricity and automotive sectors: Embracing political, social, and economic constraints. PhD thesis, Carnegie Mellon University, Ann Arbor, United States (2022)  
+
+<|ref|>text<|/ref|><|det|>[[206, 440, 832, 498]]<|/det|>
+[21] Sakti, A., Azevedo, I.M.L., Fuchs, E.R.H., Michalek, J.J., Gallagher, K.G., Whitacre, J.F.: Consistency and robustness of forecasting for emerging technologies: The case of li- ion batteries for electric vehicles. Energy Policy 106, 415- 426 (2017)  
+
+<|ref|>text<|/ref|><|det|>[[206, 509, 832, 552]]<|/det|>
+[22] Funk, P., Davis, A., Vaishnav, P., Dewitt, B., Fuchs, E.: Individual inconsistency and aggregate rationality: Overcoming inconsistencies in expert judgment at the technical frontier. Technol. Forecast. Soc. Change 155, 119984 (2020)  
+
+<|ref|>text<|/ref|><|det|>[[206, 563, 832, 620]]<|/det|>
+[23] Burden, M.: GM Orion readies for Chevy Bolt EV production. https://www.detroitnews.com/story/business/autos/general-motors/2016/09/13/gm- orion- readies- chevy- bolt- ev- production/90294594/. Accessed: 2024- 3- 4 (2016)  
+
+<|ref|>text<|/ref|><|det|>[[206, 632, 832, 704]]<|/det|>
+[24] Noble, B., Hall, K.: GM laying off hundreds of workers as Chevrolet Camaro, Bolt production end. https://www.detroitnews.com/story/business/autos/chrysler/2023/12/14/gm- layoffs- autoworkers- uaw- chevrolet- camaro- bolt- production- end/71923598007/. Accessed: 2024- 3- 4 (2023)  
+
+<|ref|>text<|/ref|><|det|>[[206, 716, 832, 730]]<|/det|>
+[25] Automotive News: Automotive News Research & Data Center. Title of the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[198, 87, 686, 101]]<|/det|>
+publication associated with this dataset: Automotive News (2023)  
+
+<|ref|>text<|/ref|><|det|>[[165, 111, 761, 128]]<|/det|>
+[26] Austenfeld, J.R.B.: NUMMI - the great experiment. Technical report (2007)  
+
+<|ref|>text<|/ref|><|det|>[[165, 138, 790, 181]]<|/det|>
+[27] Shaiken, H.: Commitment is a Two- Way street: Toyota, california and NUMMI. Technical Report 202- 10, Institute for Research on Labor and Employment, UC Berkeley (2010)  
+
+<|ref|>text<|/ref|><|det|>[[165, 193, 790, 236]]<|/det|>
+[28] Troncoso, J.N.: End of the line: Reassembling the legacy of NUMMI, the american middle class in the era of globalization and recession. PhD thesis, University of California, Berkeley (2012)  
+
+<|ref|>text<|/ref|><|det|>[[165, 247, 790, 290]]<|/det|>
+[29] Johnston, A.: NUMMI, five years later: Picking up the pieces. https://www.kalw.org/show/crosscurrents/2015- 06- 01/ nummi- five- years- later- picking- up- the- pieces. Accessed: 2023- 10- 26 (2015)  
+
+<|ref|>text<|/ref|><|det|>[[165, 301, 790, 344]]<|/det|>
+[30] Cohen, J.: An Automotive Insider's Tour Of The Tesla Fremont Factory. https://cleantechnica.com/2014/06/25/automotive- insiders- tour- tesla- fremont- factory/. Accessed: 2023- 12- 5 (2014)  
+
+<|ref|>text<|/ref|><|det|>[[165, 356, 800, 413]]<|/det|>
+[31] Field, K.: Tesla Model 3 Battery Pack Cell Teardown Highlights Performance Improvements. https://cleantechnica.com/2019/01/28/ tesla- model- 3- battery- pack- cell- teardown- highlights- performance- improvements/. Accessed: 2023- 12- 4 (2019)  
+
+<|ref|>text<|/ref|><|det|>[[165, 424, 790, 467]]<|/det|>
+[32] Hawley, G.: Understanding Tesla's lithium ion batteries. https://evannex.com/blogs/news/understanding- teslas- lithium- ion- batteries. Accessed: 2023- 12- 4 (2023)  
+
+<|ref|>text<|/ref|><|det|>[[165, 479, 790, 522]]<|/det|>
+[33] Kane, M.: What Batteries Are Tesla Using In Its Electric Cars? https://insideevs.com/news/587455/batteries- tesla- using- electric- cars/. Accessed: 2023- 12- 4 (2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 534, 790, 577]]<|/det|>
+[34] The Tesla Team: Battery Cell Production Begins at the Gigafactory. https://www.tesla.com/blog/battery- cell- production- begins- gigafactory. Accessed: 2023- 12- 4 (2017)  
+
+<|ref|>text<|/ref|><|det|>[[165, 589, 790, 632]]<|/det|>
+[35] Korosec, K.: Tesla Will Make Electric Motors for the Model 3 at Its Massive Gigafactory. https://fortune.com/2017/01/17/ tesla- model- 3- motor- gigafactory/. Accessed: 2023- 12- 4 (2017)  
+
+<|ref|>text<|/ref|><|det|>[[165, 644, 816, 700]]<|/det|>
+[36] The Tesla Team: Panasonic Enters into Supply Agreement with Tesla Motors to Supply Automotive- Grade Battery Cells. https://www.tesla.com/blog/panasonic- enters- supply- agreement- tesla- motors- supply- automotivegrade- battery- c. Accessed: 2023- 12- 4 (2011)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[204, 85, 833, 145]]<|/det|>
+[37] Lambert, F.: Tesla is building new 'drive unit production lines' at the Gigafactory, will not only manufacture battery packs. https://electrek.co/2016/10/15/tesla- drive- unit- production- lines- gigafactory- model- 3/. Accessed: 2023- 12- 5 (2016)  
+
+<|ref|>text<|/ref|><|det|>[[204, 154, 767, 170]]<|/det|>
+[38] Goldstein, A.: Janesville: An American Story. Simon & Schuster (2017)  
+
+<|ref|>text<|/ref|><|det|>[[204, 181, 832, 211]]<|/det|>
+[39] Vlasic, B.: With sonic, G.M. stands atomaking on its head. The New York Times (2011)  
+
+<|ref|>text<|/ref|><|det|>[[205, 221, 832, 280]]<|/det|>
+[40] Nadrowski, K.: Orion Assembly Site. https://web.archive.org/web/20110523235653/http://media.gm.com/content/media/us/en/news/news_detail.detail.brand.GM.html/content/Pages/news/Plant_Facts/Assembly/Orion. Accessed: 2023- 12- 5 (2011)  
+
+<|ref|>text<|/ref|><|det|>[[205, 290, 832, 335]]<|/det|>
+[41] Hula, A., Maguire, A., Bunker, A., Rojcek, T., Harrison, S.: The 2023 EPA automotive trends report. Technical Report EPA- 420- R- 23- 033, United States Environmental Protection Agency (December 2023)  
+
+<|ref|>text<|/ref|><|det|>[[205, 344, 831, 389]]<|/det|>
+[42] Motors, G.: Chevrolet Bolt EV Battery Production Resumes. https://news.gm.com/newsroom.detail.html/Pages/news/us/en/2021/sep/0920- bolt.html. Accessed: 2023- 12- 5 (2021)  
+
+<|ref|>text<|/ref|><|det|>[[205, 399, 831, 444]]<|/det|>
+[43] Motors, G.: Chevrolet Bolt EV Battery Production Resumes. https://media.chevrolet.com/media/us/en/chevrolet/home.detail.html/content/Pages/news/us/en/2021/sep/0920- bolt.html. Accessed: 2023- 12- 6 (2021)  
+
+<|ref|>text<|/ref|><|det|>[[205, 453, 831, 498]]<|/det|>
+[44] Motors, M.: Mitsubishi Motors North America, Inc. - Manufacturing Division. https://web.archive.org/web/20060506083640/http://media.mitsubishicars.com/detail?mid=MIT2004101847405&mime=ASC. Accessed: 2023- 12- 5 (2004)  
+
+<|ref|>text<|/ref|><|det|>[[205, 508, 831, 539]]<|/det|>
+[45] Yerak, B., Cancino, A.: Mitsubishi closing normal plant in illinois, ending U.S. production. Chicago Tribune (2015)  
+
+<|ref|>text<|/ref|><|det|>[[205, 549, 831, 579]]<|/det|>
+[46] The Detroit News: Rivian builds electric pickup truck and SUV. The Detroit News (2022)  
+
+<|ref|>text<|/ref|><|det|>[[205, 589, 848, 648]]<|/det|>
+[47] Weintraub, S.: Rivian R1T first drive: Easily the best pickup I've ever driven, both off- road and on. https://electrek.co/2021/09/28/rivian- r1t- first- drive- easily- the- best- pickup- ive- ever- driven- both- off- road- and- on/. Accessed: 2023- 12- 6 (2021)  
+
+<|ref|>text<|/ref|><|det|>[[205, 658, 832, 702]]<|/det|>
+[48] Cutcher- Gershenfeld, J., Brooks, D., Mulloy, M.: The decline and resurgence of the U.S. auto industry. Technical Report 399, Economic Policy Institute (May 2015)  
+
+<|ref|>text<|/ref|><|det|>[[205, 712, 831, 728]]<|/det|>
+[49] Furr, N., Dyer, J.: Lessons from tesla's approach to innovation. Harvard Business
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[198, 87, 306, 100]]<|/det|>
+Review (2020)  
+
+<|ref|>text<|/ref|><|det|>[[164, 111, 790, 142]]<|/det|>
+[50] Bellon, T., White, J.: Focus: Build or buy? automakers chasing tesla rethink dependence on suppliers. Reuters (2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 152, 790, 197]]<|/det|>
+[51] Field, K.: Behind The Scenes At Tesla's Seat Factory #CleanTechnica Field Trip. https://cleantechnica.com/2019/04/27/behind-the-scenes-at-teslas-seat- factory/. Accessed: 2023- 12- 4 (2019)  
+
+<|ref|>text<|/ref|><|det|>[[165, 207, 790, 237]]<|/det|>
+[52] Chen, Y., Chowdhury, S.D., Donada, C.: Mirroring hypothesis and integrality: Evidence from tesla motors. J. Eng. Tech. Manage. 54, 41- 55 (2019)  
+
+<|ref|>text<|/ref|><|det|>[[165, 247, 790, 291]]<|/det|>
+[53] Herrigel, G., Wittke, V.: Varieties of vertical disintegration: The global trend toward heterogeneous supply relations and the reproduction of difference in US and german manufacturing. Industry Studies Association 2004- 15 (2004)  
+
+<|ref|>text<|/ref|><|det|>[[165, 301, 790, 346]]<|/det|>
+[54] Ruyter, A., Weller, S., Henry, I., Raimnie, A., Bentley, G., Nielsen, B.: Enabling a just transition in automotive: Evidence from the west midlands and south australia. Technical report, The British Academy (June 2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 356, 790, 415]]<|/det|>
+[55] Krusemark, L., Ganguly, S., Harp, T., Kulicki, A., Smith, C., Prasad, K.V.: Examining workforce needs for north america: Battery industry education and training needs assessment (BIETNA). Technical report, Center for Automotive Research (2024)  
+
+<|ref|>text<|/ref|><|det|>[[165, 425, 790, 470]]<|/det|>
+[56] Combemale, C., Whitefoot, K.S., Ales, L., Fuchs, E.R.H.: Not all technological change is equal: how the separability of tasks mediates the effect of technology change on skill demand. Ind Corp Change 30(6), 1361- 1387 (2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 480, 789, 510]]<|/det|>
+[57] Weaver, A., Osterman, P.: Skill demands and mismatch in U.S. manufacturing. ILR Review 70(2), 275- 307 (2017)  
+
+<|ref|>text<|/ref|><|det|>[[165, 520, 790, 565]]<|/det|>
+[58] Cotterman, T., Fuchs, E.R.H., Small, M.J., Whitefoot, K.: The Transition to Electrified Vehicles: Implications for the Future of Automotive Manufacturing and Worker Skills and Occupations (2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 575, 789, 605]]<|/det|>
+[59] Sadeghi, A.: The births and deaths of business establishments in the united states. Mon. Labor Rev. December 2008(1), 1- 18 (2008)  
+
+<|ref|>text<|/ref|><|det|>[[165, 615, 790, 660]]<|/det|>
+[60] U.S. Bureau of Labor Statistics: Automotive Industry: Employment, Earnings, and Hours. https://www.bls.gov/iag/tgs/iagauto.htm. Accessed: 2023- 10- 26 (2022)  
+
+<|ref|>text<|/ref|><|det|>[[165, 670, 790, 700]]<|/det|>
+[61] Federal Reserve Bank of St. Louis: Consumer Price Index for All Urban Consumers: New Vehicles in U.S. City Average (2024)  
+
+<|ref|>text<|/ref|><|det|>[[165, 710, 790, 741]]<|/det|>
+[62] Campagnol, N., Pfeiffer, A., Tryggestad, C.: Capturing the battery value- chain opportunity. Technical Report 1, McKinsey & Company (January 2022)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[206, 86, 832, 130]]<|/det|>
+[63] Lambert, F.: Tesla Gigafactory 1 now employs over 3,000 workers as it becomes biggest battery factory in the world. https://electrek.co/2018/08/21/tesla- gigafactory- 1- 3000- workers/. Accessed: 2023- 12- 11 (2018)  
+
+<|ref|>text<|/ref|><|det|>[[206, 141, 832, 185]]<|/det|>
+[64] Knehr, K.W., Kubal, J.J., Nelson, P.A., Ahmed, S.: Battery performance and cost modeling for electric vehicles - a manual for BatPaC v5.0. Technical Report ANL/CSE- 22/1, Argonne National Laboratory (July 2022)
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[283, 128, 666, 545]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[163, 549, 790, 585]]<|/det|>
+<center>Fig. A1 U.S.-level vehicle production (a), assembly workers (b), and labor intensity (c). Vehicle production data was obtained from the Automotive News Research & Data Center, while employment data was obtained from QCEW. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[234, 102, 792, 416]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[205, 421, 833, 491]]<|/det|>
+<center>Fig. A2 Battery cell and pack manufacturing increases the labor intensity of making BEVs. (a) Annual vehicle production volume in Alameda, CA. (b) Employment in Alameda, CA and Sparks, NV, based on news reports (see Table A2). Employment in Sparks, NV, reflects additional workers for battery cell and pack manufacturing at Tesla Gigafactory 1. Data is shown only when Alameda and Sparks data are available for the same production year. (c) Comparison of labor intensity (WPV) with and without including workers from Sparks, NV. </center>  
+
+<|ref|>image<|/ref|><|det|>[[205, 520, 833, 680]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[205, 695, 833, 731]]<|/det|>
+<center>Fig. A3 Comparison of vehicle sales prices (MSRP) in Alameda (a), Oakland (b), and McLean (c). Dollar values are inflation-adjusted to 2023 dollars based on the Consumer Price Index (CPI) for U.S. new vehicles [61]. MSRP for all available vehicle trims are shown. </center>
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[171, 140, 788, 321]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[163, 340, 792, 364]]<|/det|>
+<center>Fig. A4 Concept illustration: vertical integration creates more workforce co-location at the site of vehicle assembly. </center>  
+
+<|ref|>table<|/ref|><|det|>[[208, 475, 745, 595]]<|/det|>
+
+<table><tr><td></td><td>Location</td><td>2013</td><td>2021</td></tr><tr><td rowspan="2">% of NAICS 3361 workers in production</td><td>California (State)</td><td>66%</td><td>62%</td></tr><tr><td>Michigan (State)</td><td>74%</td><td>81%</td></tr><tr><td></td><td>U.S.</td><td>74%</td><td>76%</td></tr><tr><td rowspan="2">% of NAICS 3361 workers in engineering</td><td>California (State)</td><td>4%</td><td>7%</td></tr><tr><td>Michigan (State)</td><td>5%</td><td>3%</td></tr><tr><td></td><td>U.S.</td><td>5%</td><td>5%</td></tr><tr><td rowspan="2">NAICS 3361 average monthly pay</td><td>Alameda, CA</td><td>$6,243</td><td>$16,462</td></tr><tr><td>Oakland, MI</td><td>$7,557</td><td>$8,907</td></tr><tr><td></td><td>U.S.</td><td>$6,660</td><td>$6,864</td></tr></table>  
+
+<|ref|>text<|/ref|><|det|>[[210, 597, 745, 677]]<|/det|>
+Table A1 Proportion of NAICS 3361 workers in production (SOC code 51- 0000) and architecture/engineering occupations (SOC code 17- 0000), and average monthly pay of NAICS 3361 workers for California, Michigan, and the U.S. Occupation data was obtained from the Occupational Employment and Wage Statistics (OEWS). Income data for Alameda, CA was obtained from QWI. Income data for Oakland, MI was obtained from averaging QWI and QCEW data. Income data for the U.S. was obtained from QCEW.
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[238, 228, 797, 560]]<|/det|>
+
+<table><tr><td>Location</td><td>Date</td><td>News Source</td><td>Reported Employment</td></tr><tr><td rowspan="6">Tesla (Alameda)</td><td>Jun 2012</td><td>SFGATE</td><td>1,000</td></tr><tr><td>Jul 2013</td><td>Wired</td><td>3,000</td></tr><tr><td>Jun 2016</td><td>TheCountryCaller</td><td>6,000</td></tr><tr><td>Oct 2017</td><td>The Mercury News</td><td>10,000</td></tr><tr><td>Mar 2019</td><td>Forbes</td><td>15,000</td></tr><tr><td>Jun 2022</td><td>Tesla</td><td>22,000</td></tr><tr><td rowspan="3">Tesla/PENA (Sparks)</td><td>2016</td><td>Electrek</td><td>850</td></tr><tr><td>2017</td><td>Electrek</td><td>3,249</td></tr><tr><td>2018</td><td>The Associated Press</td><td>7,059</td></tr><tr><td></td><td>2022</td><td>Tesla</td><td>12,000</td></tr><tr><td rowspan="3">NUMMI (Alameda)</td><td>Jan 2002</td><td>SFGATE</td><td>5,739</td></tr><tr><td>Mar 2006</td><td>East Bay Times</td><td>5,500</td></tr><tr><td>Apr 2010</td><td>Recordnet.com</td><td>4,700</td></tr><tr><td rowspan="4">Rivian (Normal)</td><td>Oct 2021</td><td>WGLT</td><td>3,000</td></tr><tr><td>Apr 2022</td><td>CIPROUD</td><td>5,000</td></tr><tr><td>Jun 2022</td><td>Energy News Network</td><td>5,600</td></tr><tr><td>Jul 2022</td><td>CIPROUD</td><td>6,000</td></tr><tr><td></td><td>Mar 2023</td><td>WGLT</td><td>7,400</td></tr><tr><td rowspan="4">Mitsubishi (Normal)</td><td>2004</td><td>Chicago Tribune</td><td>3,150</td></tr><tr><td>2014</td><td>Local Wiki</td><td>1,250</td></tr><tr><td>2015</td><td>Chicago Tribune</td><td>1,280</td></tr><tr><td>2016</td><td>WQAD8</td><td>1,200</td></tr><tr><td rowspan="3">GM (Orion)</td><td>2013</td><td>CarGroup.org</td><td>2,561</td></tr><tr><td>2022</td><td>GM</td><td>1,238</td></tr><tr><td>2023</td><td>Wards Auto</td><td>1,270</td></tr></table>
+
+<|ref|>table_footnote<|/ref|><|det|>[[238, 562, 756, 585]]<|/det|>
+Table A2 List of news reports used to corroborate factory employment numbers. PENA: Panasonic Energy of North America.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[166, 81, 558, 102]]<|/det|>
+## Appendix B Workers per GWh  
+
+<|ref|>text<|/ref|><|det|>[[166, 111, 790, 168]]<|/det|>
+McKinsey reported that, on average, new battery factories add approximately 80 jobs for every GWh of capacity, i.e. 80 workers per GWh [62]. This number carries some uncertainty since differences in value- chain coverage, e.g. battery- cell production only versus local module and pack production or co- location of R&D facilities, are unclear.  
+
+<|ref|>text<|/ref|><|det|>[[166, 168, 790, 280]]<|/det|>
+Tesla's Gigafactory 1 reportedly employed 3,249 people when the factory was producing 20 GWh of annual output [63]. Among these workers, 1,201 were employed by Panasonic, the main battery cell manufacturer, 93 are employed by Heitkamp & Thumann Group (H&T), a battery cell can supplier, and 1,955 were employed by Tesla. Assuming those employed by Panasonic and H&T are responsible for battery cell manufacturing, we infer that 1,294 workers are involved with producing 20 GWh of annual output, or 65 workers per GWh. If the employees from Tesla are included, then the calculation yields 162 workers per GWh.  
+
+<|ref|>text<|/ref|><|det|>[[166, 281, 790, 338]]<|/det|>
+The BatPaC v5.0 baseline factory model reported an annual labor of 3,876,000 hours per year to produce 50 GWh of output [64]. Assuming each worker works 2,236 hours per year and (equivalent to a 43- hour work- week, the U.S. average for automotive manufacturing [60]), this amounts to 35 workers per GWh.  
+
+<|ref|>text<|/ref|><|det|>[[166, 339, 790, 438]]<|/det|>
+Cotterman et al. [19] reported labor intensity per BEV powertrain assuming a 60kWh battery pack which varied depending on the data source and whether the labor was broken down between cell and pack/module assembly. For data sources where this breakdown was available, labor intensity ranged between 11 to 16 hours per 60kWh for industry data sources and 6 to 15 hours per 60kWh for public data sources. Assuming again 2,236 hours per year worked per worker, the range of labor demand is equivalent to 44 to 119 workers per GWh.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[268, 263, 763, 533]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[204, 540, 832, 565]]<|/det|>
+<center>Fig. B5 Comparison of engine manufacturing parts jobs in 2022 against projected battery cell manufacturing jobs assuming 100% BEV uptake and under various assumptions of jobs per GWh. </center>
+
+<--- Page Split --->

preprint/preprint__09b079d80b3efdeef69e111bf52a9dde06da16423667f03702d57d11bbbeec9b/images_list.json

@@ -0,0 +1,77 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1. Conceptual outline and proposed concept for the meta borylation of phenol. A. Meta functionalization of phenol, a, electrophilic approach. b, template approach. c, transient directing group approach. B. Conceptual Background. a, common structural skeleton for O-Si interaction. b-e, known literature reports of O-Si interactions. C. Hypothesis, D. Reaction design. a, ligand design by structural modification, b, proposed approach for meta borylation.",
+    "footnote": [],
+    "bbox": [
+      [
+        105,
+        85,
+        884,
+        636
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2. Ligand design and optimization of reaction conditions. A. Optimization of steering group, B. Deleterious results with other ligands. C. Origin of meta selectivity, D. Proof of concept. Reactions are on 0.2 mmol scales. \\(^{a}\\) Conversion was reported. In parenthesis, isolated yields are reported. See SI for details.",
+    "footnote": [],
+    "bbox": [
+      [
+        104,
+        115,
+        888,
+        570
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3. Substrates scope for substituted arenes. Reactions are in 0.5 mmol scale. <sup>a</sup>Conversion was reported. <sup>b</sup>1.5 equiv. B<sub>2pin</sub> was used. <sup>c</sup>2.0 equiv. B<sub>2pin</sub> was used. See SI for details.",
+    "footnote": [],
+    "bbox": [
+      [
+        102,
+        207,
+        895,
+        802
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4. Substrates scope for the 4-substituted arenes and C6 borylation of indoles. Reactions are in 0.5 mmol scale. aConversions were reported. b2.0 equiv. B2pin2. See SI for details.",
+    "footnote": [],
+    "bbox": [
+      [
+        105,
+        194,
+        884,
+        775
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5. Applications, Catalyst Preparation and Testing. A. Late-stage meta-C-H borylation, B. Removal of silane group, C. Catalyst synthesis, D. Test of reactivity of catalyst 10. Reactions are in 0.5 mmol scale. aConversions were reported. See SI for details.",
+    "footnote": [],
+    "bbox": [
+      [
+        103,
+        84,
+        888,
+        448
+      ]
+    ],
+    "page_idx": 8
+  }
+]

preprint/preprint__09b079d80b3efdeef69e111bf52a9dde06da16423667f03702d57d11bbbeec9b/preprint__09b079d80b3efdeef69e111bf52a9dde06da16423667f03702d57d11bbbeec9b.mmd

@@ -0,0 +1,164 @@
+
+# Meta Selective C-H Borylation Directed by Secondary Silicon Oxygen Interaction  
+
+Buddhadeb Chattopadhyay ( \(\boxed{ \begin{array}{r l} \end{array} }\) buddhadeb.c@cbmr.res.in) Center of Biomedical Research (CBMR) https://orcid.org/0000- 0001- 8473- 2695  
+
+Saikat Guria Center of Biomedical Research (CBMR)  
+
+Mirja Md Hassan Centre of Biomedical Research (CBMR)  
+
+Sayan Dey Center of Biomedical Research (CBMR)  
+
+Article  
+
+Keywords:  
+
+Posted Date: August 17th, 2022  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 1837437/v1  
+
+License: \(\circledcirc\) This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest. We declare that the authors have no competing interests except that We have filled an Indian Patent (Patent Application No: 202211036590) based on this work (including the ligand and catalyst).
+
+<--- Page Split --->
+
+# Meta Selective C-H Borylation Directed by Secondary Silicon Oxygen Interaction  
+
+Saikat Guria†, Mirja Md Mahamudul Hassan†, Sayan Dey†, Buddhadeb Chattopadhyay†\*  
+
+†Department of Biological & Synthetic Chemistry, Centre of Biomedical Research, SGPGIMS Campus, Raebareli Road, Lucknow 226014, Uttar Pradesh, India \*Correspondence to: buddhadeb.c@cbmr.res.in  
+
+Summary Paragraph: Remote meta selective C- H functionalization<sup>1,2,3</sup> of aromatic compounds remains a challenging problem in chemical synthesis. Here, we report an iridium catalyst bearing a bidentate pyridine- pyridone (PY- PYRI) ligand framework that efficiently catalyzes this meta selective borylation reaction. We demonstrate that the developed concept can be employed to introduce a boron functionality at the remote meta position of phenols, phenol containing bioactive and drug molecules, which was an extraordinary challenge. Moreover, we have demonstrated that the method can also be applied for the remote C6 borylation of indole derivatives including tryptophan that was the key synthetic precursor for the total synthesis of Verruculogen and Fumitremorgin A alkaloids. The origin of the remote meta selectivity was described as a secondary silicon oxygen interaction<sup>4</sup> that was never used in C- H functionalization chemistry.
+
+<--- Page Split --->
+
+Transition metal- catalyzed C- H bond activation and functionalization \(^{5,6,7,8,9,10,11}\) of aromatic compounds has been branded as one of the most significant chemical transformations. This has a profound impact in modern synthetic organic chemistry, ranging from laboratory methods to industrial deployment. \(^{12,13}\) However, the key underlying principles for the success of the metal catalysis lies on the two important factors, such as: (i) design and synthesis of new generation ligand framework that can produce highly reactive catalyst system \(^{14,15}\) and (ii) substrates' structure modifications \(^{16}\) by which site selectivity could be controlled by the steric crowding \(^{17,18,19,20,21}\) or various weak interactions \(^{22,23,24,25}\) of the aromatic compounds among several similar type of C- H bonds via the ligand- substrate pre- organization \(^{26,27}\) . In recent times, many elegant approaches \(^{28}\) have been developed for the functionalization of proximal \(^{15,29,30,31}\) and remote C- H bonds \(^{1,3,32,33,34,35,36,37,38}\) of arenes by the design of either new ligand frameworks with an extended architectures featuring a weak coordinating functional groups \(^{39}\) or templates \(^{40}\) as well as transient mediators \(^{41}\) or transient directing groups \(^{42}\) attached with the substrates. While ligand having an extended architecture or template approaches are extremely important to functionalize the remotely located C- H bonds of arenes, but requirement of multi- step preparation of the linkers of the ligands and templates of the aromatic substrates significantly limit the wide application of the methods. \(^{43}\)  
+
+Among numerous aromatic substrates, phenols are the most widespread aromatic compounds that acquired household products including several bioactive to important drug molecules. \(^{44}\) Moreover, it is well- documented that \(10\%\) of the top 200 selling pharmaceuticals contain a phenol and several others employ phenols as synthetic intermediates. \(^{45}\) Furthermore, phenols are also key components of the biopolymers melanin, lignin, resins and polyphenylene oxides. \(^{44,45,46}\) In industry, phenol is routinely used as a raw material to make numerous important components by means of its diversification via the synthetic manipulation. \(^{44,45}\) Thus, direct functionalization of phenols would be a significant development for the rapid access of numerous important products. \(^{46}\) In this context, traditional electrophilic substitution is an alternative methods that affords variously substituted phenols (Fig. 1A, a). \(^{47}\) Employing this method, one can easily access ortho and para substituted phenol derivatives, although often remain a chance to have mixture of isomers. However, functionalization of the remote meta C- H bonds of phenols is extremely difficult because of the extreme inertness of the meta C- H bonds. Several pioneering approaches have been developed by Yu and others either using template method \(^{48,49,50}\) or transient directing group by Larrosa \(^{2}\) (Fig 1A, b, c). But, achieving the meta functionalized products using these methods, it is essential to have specialized substrates that limits the application of the methods.  
+
+Having tremendous importance of catalytic C- H borylation \(^{51,52,53,54,55,56}\) in organic synthesis, we report here a concept for the meta selective C- H borylation of phenols via an unprecedented Si- O interaction that has never been utilized in C- H functionalization chemistry. Literature reports revealed \(^{4}\) that the most common structural motif for this O- Si interaction can be found in the amide skeleton, where a filled p- orbital of oxygen atom can interact with the vacant d- orbital of the tetracoordinated silicon atom consisting of at least one electronegative atom. The role of this electronegative atom is to make silicon atom more electropositive by developing a partial positive charge on the silicon atom (Fig. 1B, a). Notably, while various reports of O- Si weak interactions have been shown in a number of intermediates (Fig 1B, b- e), \(^{4,57,58}\) there was no report of utilization of these interactions in the catalysis research. Inspired from this background reports, \(^{4,57,58}\) we have proposed a hypothesis where phenol is protected with an easily removable electropositive silane group and silane protected phenol meet all the necessary criteria (having electropositive silane with attached electronegative oxygen atom) for the O- Si weak interaction with 2- pyridine moiety having amide skeleton (4). (Fig. 1C). The reaction design (ligand and catalyst design) is shown in Fig. 1D, a, b. The designed ligand (PY- PYRI) consists of two parts, one part is the simple pyridine unit (PY) and another part is a 2- pyridine (PYRI) unit \(^{59}\) which was redesigned by the skeletal modification of bipyridine core structure. The origin of the remote meta selectivity is presented in Fig. 1D, b. We hypothesized that the designed ligand (PY- PYRI) would control the remote meta selectivity owing to the following two considerations. Firstly, in presence of [Ir(cod)(OMe)] \(_{2}\) , the ligand (PY- PYRI) will generate a complex (Int- A) without tautomerization of the 2- pyridine unit. Secondly, the p- orbital of the oxygen atom of the 2- pyridine unit will interact with the vacant d- orbital of the tetracoordinated electropositive silicon atom of the substrate (Fig. 1D, b).
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Fig. 1. Conceptual outline and proposed concept for the meta borylation of phenol. A. Meta functionalization of phenol, a, electrophilic approach. b, template approach. c, transient directing group approach. B. Conceptual Background. a, common structural skeleton for O-Si interaction. b-e, known literature reports of O-Si interactions. C. Hypothesis, D. Reaction design. a, ligand design by structural modification, b, proposed approach for meta borylation. </center>  
+
+The designed ligand (PY- PYRI) was prepared by the known synthetic methods (SI, for details), which was employed for the reaction optimization of steering silane group attached with phenol (Fig. 2A). We started our initial studies with the substrate (1a- I) featuring SiMe3 as the steering group under iridium- catalyzed conditions using the designed ligand (L1: PY- PYRI) at \(40^{\circ}C\) temperature, which afforded good meta/para selectivity ( \(\mathrm{m} / \mathrm{p} = 87 / 13\) ) with \(68\%\) conversion. Evaluating other silane based steering groups under the same reaction conditions, we observed that \(\mathrm{Si}(\mathrm{Pr})_3\) produced best meta selectivity (entry 2a: \(\mathrm{m} / \mathrm{p} = 94 / 6\) ) with excellent conversion ( \(95\%\) ). With this optimized \(\mathrm{Si}(\mathrm{Pr})_3\) as steering group, different other ligands have also been tested to observe the effect on the selectivity (Fig. 2B). It was found that replacing the tert- butyl group with methyl group, ligand (L2) gave \(91\%\) meta selectivity with less conversions ( \(68\%\) ). Similar selectivity was obtained when the reaction was performed with the ligand (L3) without any substituents on the pyridine unit of the ligand. Notably, the meta selectivity and conversion was found to be less using the bipyridine ligand (L4) compared to the ligands (L1- L3), which indicated the important role of the
+
+<--- Page Split --->
+
+2- pyridine unit of the designed (PY- PYRI) ligand. Employment of other ligands (L5- L8) resulted in no reaction except the ligands (L9 & L10), which gave moderate meta selectivity.  
+
+![](images/Figure_2.jpg)
+
+<center>Fig. 2. Ligand design and optimization of reaction conditions. A. Optimization of steering group, B. Deleterious results with other ligands. C. Origin of meta selectivity, D. Proof of concept. Reactions are on 0.2 mmol scales. \(^{a}\) Conversion was reported. In parenthesis, isolated yields are reported. See SI for details. </center>  
+
+At the outset, we proposed the tentative hypothesis that an O- Si secondary interaction between the oxygen atom of the 2- pyridine unit of the ligand and the silicon atom of the substrate's steering group would interact each other via the filled p- orbital and empty d- orbital (Fig. 2C). \(^{4}\) Moreover, due to the high electronegativity difference between oxygen and silicon, the O- Si bond will be highly polarized, thereby may interact via a weak O- Si interaction. To prove this hypothesis for the origin of the meta selectivity, we performed a reaction with substrates featuring C- Si bond (- CH2SiPr3: 1a- VII and - CH2SiMe3: 1a- VIII, Fig. 2D), in which lacking of electronegative atom attached with silicon atom causes non- polarised bond, resulted in very low meta selective borylation due to loss of interaction between 2- pyridine of catalyst and substrates silicon atom. This experiment indicated an "O- Si" secondary interaction between the ligand and substrate that guides the selectivity of the borylation.  
+
+Using L1 (PY- PYRI) as ligand and Si(Pr)3 as steering group, we next performed the iridium- catalyzed meta borylation of a variety of phenols that afforded excellent meta selectivity and yields of the isolated borylated products (Fig. 3). For example, we first tested 2- chlorophenol for the borylation reaction, while our designed ligand (L1) gave high meta selectivity (m/p = 92/8), traditional dtbpy ligand provided poor meta selectivity (m/p = 63/37), which clearly demonstrated the utility of the designed (L1: PY- PYRI) ligand. Other 2- substituted phenols, such as 2- bromo (1c) and 2- iodo (1d) afforded high meta selectivity that have great synthetic values owing to the two different types of handles on the phenols. Likewise, phenols bearing various alkyl chain ranging from methyl to pentyl (1e- 1g) at the
+
+<--- Page Split --->
+
+2- positions along with trifluoromethyl (1h), isopropyl (1i) and trifluoromethoxy (1l) smoothly underwent meta borylation irrespective of the nature of the substituent. Amino phenol (1k), substrate of momentous importance for the chemical and pharmaceutical industries<sup>60</sup> is borylated with high meta selectivity \((m / p = 97 / 3)\) without borylation next to the amino group, which is known to give ortho borylation under iridium- catalyzed borylation conditions via in situ generation of NHBpin group.<sup>61</sup> Thioether (1l) that usually directs borylation at the ortho position<sup>62</sup> also underwent borylation with good meta selectivity. We observed that phenols containing functional groups such as cyano (1m), Bpin (1n), cyclic amine (1o), cyclohexyl (1p), ketomethyl (1q) and homologous ester (1r) afforded high level of meta selectivity and tolerated well under the employed reaction conditions.  
+
+![](images/Figure_3.jpg)
+
+<center>Fig. 3. Substrates scope for substituted arenes. Reactions are in 0.5 mmol scale. <sup>a</sup>Conversion was reported. <sup>b</sup>1.5 equiv. B<sub>2pin</sub> was used. <sup>c</sup>2.0 equiv. B<sub>2pin</sub> was used. See SI for details. </center>  
+
+Amide functionalities (1s & 1t) that are known to undergo numerous synthetic transformations<sup>63</sup> exhibited excellent meta selective borylation. Borylation of phenols having \(\mathrm{CF}_3\) (1h) and CN (1m) substituents at the ortho position afforded exclusively meta borylation, the same substituents at the meta position of phenols (1u & 1v) also gave meta selective borylation, which indicated the generality of the developed method. Moreover, fluoro- substituted
+
+<--- Page Split --->
+
+arene, which typically gives borylation next to the fluorine atom under standard iridium- catalyzed conditions, in this case, 3- fluorophenol (1w) gave meta borylation as the major product. Several disubstituted phenols (1x- 1ag) were also examined under the developed conditions that reacted smoothly to afford variously substituted meta borylated products in high yields. 2,2'- Biphenol, compound of paramount importance in medicinal chemistry as well as in chemical industry64 can selectively be mono- and diborylation (2ah & 2ai) by tuning the amount of boron reagent. A bulky substituent at the ortho positions (1aj) did not hamper the reaction that gave 96% meta borylated product with 90% isolated yield.  
+
+![](images/Figure_4.jpg)
+
+<center>Fig. 4. Substrates scope for the 4-substituted arenes and C6 borylation of indoles. Reactions are in 0.5 mmol scale. aConversions were reported. b2.0 equiv. B2pin2. See SI for details. </center>  
+
+Next, we focused on the meta borylation of those phenols bearing a substituent at the para position (Fig. 4). Because, borylation at the remote meta position in presence of a para substituent remains an extraordinary challenge due to the steric reason. Moreover, we selected those substituents at the para position that already provided exclusive meta borylation of phenols when they were located at either ortho or meta positions. The reason for this selection is mainly to observe the overall effects of the borylation by the same substituents. For the
+
+<--- Page Split --->
+
+testification, we begun with the 4- methyl phenol (3a) that afforded \(91\%\) meta selective borylation. Increasing the chain length from small methyl group to the relatively bulkier alkyl groups such as, ethyl (3b), pentyl (3c), hexyl (3d) and isopropyl (3e), the borylation underwent smoothly with further enhancement of the meta selectivity from \(91\%\) to \(100\%\) . Para- substituted ethers and thioethers bearing electronically different substituents (3f- 3j) reacted with \(100\%\) meta selectivity, which revealed that the scope of the meta borylation is very general regardless of the nature of the substituents. While 2- CN, 3- CN as well as 2- CF3 and 3- CF3 bearing phenols resulted in excellent meta borylation, the same substituents at the para position reacted to yield \(100\%\) meta borylation. Likewise, we also observed that chloro (3m) and bromo (3n) containing phenols reacted to give the meta borylation products solely irrespective of their position in the phenol. Moreover, it has been found that the phenols featuring bulky substituents at the para position (3o- 3r) also gave exclusively meta borylation, although conversion was moderate in case of the cyclohexyl group.  
+
+In 2015, Baran et al. reported \(^{65}\) the first total synthesis of Verruculogen and Fumitremorgin A enabled by ligand- controlled C- H borylation as the key step of TIPS protected tryptophan. We were curious if our designed ligand system could provide the remote C6 borylation of TIPS protected indoles and TIPS protected tryptophan. For that reason we performed borylation of TIPS- protected tryptophan (3s) (synthetic key precursor of bioactive alkaloids Verruculogen and Fumitremorgin A) which provided C6 borylation with \(91\%\) selectivity with excellent conversions. We also found that TIPS- protected other indole derivatives (3t & 3u) and TIPS- protected carbazole (3v) smoothly underwent remote borylation affording excellent selectivity and conversion. This developed method provided a simple way to borylate the 3- substituted indoles derivatives that might be beneficial for the total synthesis or the late- stage functionalization of several bioactive molecules.  
+
+Late- stage functionalization \(^{66}\) of complex bioactive and medicinally important molecules by the site selective C- H activation is a powerful method for the development of new drug candidates. \(^{67}\) In this context, introducing a boron functionality into the bioactive and medicinally important molecules would further enhance the identification of new lead molecules not only for the enormous importance of the boron- bearing small molecules \(^{68}\) but also for the uniqueness of the boron group towards the diverse derivatization towards numerous other functional groups. Thus, we tested our developed method for several commercially available bioactive and drug molecules (Fig. 5A). For example, cannabinoid core (5a: used as a psychoactive drug), methyl salicylate derivatives (5b: an anti- inflammatory and analgesic agent), tyrosol derivatives (5c: an antioxidant), eugenol derivatives (5d: a flavouring agent), sesamol derivatives (5e: an antioxidant), naproxen derivatives (5f: a nonsteroidal anti- inflammatory drug, NSAID), deoxyarbutin derivatives (5g: used for treatment of hyperpigmentation disorders) and homosalate (5h: used as a sunscreen) were meta borylated with high yield and selectivity. The steering silane group from the borylated phenols has been removed under a very mild reaction conditions at room temperature (ethylene glycol, KF, 1h) that afforded the meta borylated phenols in high yields (Fig. 5B). Notably, the meta borylated phenols can further be transformed to a number of substituted phenols/resorcinols that are difficult to prepare by otherwise.  
+
+Next, we aimed to prepare the active catalyst (10) that was proposed to form in situ between the reaction of the designed ligand (L1: PY- PYRI) and [Ir(cod)(OME)]2 during the meta selective borylation conditions. Accordingly, we performed the reaction and isolated the catalyst [10: Ir(cod)(PY- PYRI)] in \(90\%\) yield (Fig. 5C). The catalyst structure was confirmed by X- ray crystallography and other spectroscopic data. The catalytic efficiency of this catalyst [10: Ir(cod)(PY- PYRI)] was further tested in the meta borylation reactions, which exhibited same level of meta selectivity with better product conversion (compared to the in situ generation) (Fig. 5D). Moreover, we checked the stability of the catalyst [10: Ir(cod)(PY- PYRI)] and found highly stable that can even be stored in open air. Furthermore, to verify the broad utility of this air stable catalyst [10: Ir(cod)(PY- PYRI)], we performed several test experiments using this catalyst [10: Ir(cod)(PY- PYRI)] that was stored in open air and found no loss of catalytic activity even after 30 days (Fig. 5D, SI for details).
+
+<--- Page Split --->
+![](images/Figure_5.jpg)
+
+<center>Fig. 5. Applications, Catalyst Preparation and Testing. A. Late-stage meta-C-H borylation, B. Removal of silane group, C. Catalyst synthesis, D. Test of reactivity of catalyst 10. Reactions are in 0.5 mmol scale. aConversions were reported. See SI for details. </center>  
+
+In conclusion, we report a new class of ligand and catalyst that has demonstrated remarkable efficiency for the remote meta selective borylation of phenols featuring all types of substitutions at the arene ring. In addition, we have seen that our developed ligand system is beneficial for the remote C6- borylation of indoles derivatives including tryptophan which is a synthetic precursor of bioactive alkaloids (Verruculogen and Fumitremorgin A). Several late- stage meta borylations have been showcased with bioactive and drug molecules that might be useful for repurposing medicines and identification of new lead drug candidates. For the first time, an "O- Si" secondary interaction has been employed to tune the remote selectivity. We anticipate that the designed ligand and catalyst will also find wide application in the context of other C- H functionalization reactions.  
+
+## References and notes  
+
+1. Zhang, Z., Tanaka, K. & Yu, J-Q. Remote site-selective C-H activation directed by a catalytic bifunctional template. Nature 543, 538-542 (2017).  
+2. Luo, J., Preciado, S. & Larrosa, I. Overriding Ortho-Para Selectivity via a Traceless Directing Group Relay Strategy: The Meta-Selective Arylation of Phenols, J. Am. Chem. Soc. 136, 4109-4112 (2014).  
+3. Sinha, S. K., Guin, S., Maiti, S., Biswas, J. P., Porey, S. & Maiti, D. Toolbox for Distal C-H Bond Functionalizations in Organic Molecules. Chem. Rev. 122, 5682-5841 (2022).  
+4. Lazareva, N. F., Sterkhova, I. V. & Vashchenko, A. V. N-[difluoro(methyl)silyl]carboxamides: Synthesis, structural features and theoretical estimating of Si<O- dative bond energy. Journal of Molecular Structure 1225, 129130 (2021).  
+5. Lyons, T. W. & Sanford, M. S. Palladium-Catalyzed Ligand-Directed C-H Functionalization Reactions. Chem. Rev. 110, 1147-1169 (2010).  
+6. Davies, H. M. L., Bois, J. D. & Yu, J-Q. C-H Functionalization in organic synthesis. Chem. Soc. Rev. 40, 1855-1856 (2011).
+
+<--- Page Split --->
+
+7. Arockiam, P. B., Bruneau, C. & Dixneuf, P. H. Ruthenium (II)-Catalyzed C-H Bond Activation and Functionalization. Chem. Rev. 112, 5879-5918 (2012).8. Sambiagio, C. et al. A comprehensive overview of directing groups applied in metal-catalysed C-H functionalisation chemistry. Chem. Soc. Rev. 47, 6603-6743 (2018).9. Crabtree, R. H. & Lei, A. Introduction: CH Activation. Chem. Rev. 117, 8481-8482 (2017).10. Shilov, A. E. & Shul'pin, G. B. Activation of C-H Bonds by Metal Complexes. Chem. Rev. 97, 2879-2932 (1997).11. Rogge, T. et al. C-H activation. Nat Rev Methods Primers 1, (2021). doi.org/10.1038/s43586-021-00041-2.12. Yamaguchi, J., Yamaguchi, A. D. & K. Itami, C-H Bond Functionalization: Emerging Synthetic Tools for Natural Products and Pharmaceuticals. Angew. Chem. Int. Ed. 51, 8960-9009 (2012).13. Dalton, T., Faber, T. & Glorius, F. C-H Activation: Toward Sustainability and Applications. ACS Cent. Sci. 7, 245-261 (2021).14. Liao, K. et al. Design of catalysts for site-selective and enantioselective functionalization of non-activated primary C-H bonds. Nature Chem 10, 1048-1055 (2018).15. Hoque, M. E., Hassan, M. M. M. & Chattopadhyay, B. Remarkably Efficient Iridium Catalysts for Directed C(sp2)-H and C(sp3)-H Borylation of Diverse Classes of Substrates. J. Am. Chem. Soc. 143, 5022-5037 (2021).16. Hoveyda, A. H., Evans, D. A. & Fu, G. C. Substrate-directable chemical reactions. Chem. Rev. 93, 1307-1370 (1993).17. Cheng, C. & Hartwig, J. F. Rhodium-Catalyzed Intermolecular C-H Silylation of Arenes with High Steric Regiocontrol. Science, 343, 853-857 (2014).18. Ramadoss, B., Jin, Y., Asako, S. & Ilies, L. Remote steric control for undirected meta-selective C-H activation of arenes, Science, 375, 658-663 (2022).19. Cho, J. Y., Tse, M. K., Holmes, D., Maleczka, R. E. & Smith, M. R. Remarkably selective iridium catalysts for the elaboration of aromatic C-H bonds. Science 295, 305-308 (2002).20. Saito, Y., Segawa Y. & Itami, K. para-C-H Borylation of Benzene Derivatives by a Bulky Iridium Catalyst. J. Am. Chem. Soc., 137, 5193-5198 (2015).21. Mondal, A., Chen, H., Flamig, L., Wedi, P. & Gemmeren, M. V. Sterically Controlled Late-Stage C-H Alkynylation of Arenes. J. Am. Chem. Soc. 141, 18662-18667 (2019).22. Kuninobu, Y., Ida, H., Nishi, M. & Kanai, M. A meta-selective C-H borylation directed by a secondary interaction between ligand and substrate. Nat. Chem. 7, 712-717 (2015).23. Fanourakis, A., Docherty, P. J., Chuentragool, P. & Phipps, R. J. Recent Developments in Enantioselective Transition Metal Catalysis Featuring Attractive Noncovalent Interactions between Ligand and Substrate, ACS Catalysis, 10, 10672-10714 (2020).24. Zhang, T., Luan, Y., Lam, N. Y. S., Li, J., Li, Y., Ye, M. Yu, & J-Q. A directive Ni catalyst overrides conventional site selectivity in pyridine C-H alkenylation. Nat. Chem. 13, 1207-1213 (2021).25. Hoque, M. E., Bisht, R., Haldar, C. & Chattopadhyay, B. Noncovalent Interactions in Ir-Catalyzed C-H Activation: L-Shaped Ligand for Para-Selective Borylation of Aromatic Esters. J. Am. Chem. Soc. 139, 7745-7748 (2017).26. Dydio, P. & Reek, J. N. H. Supramolecular control of selectivity in transitionmetal catalysis through substrate preorganization. Chem. Sci., 5, 2135-2145 (2014).27. Lou, Y., Wei, J., Li, M. & Zhu, Y. Distal Ionic Substrate-Catalyst Interactions Enable Long-Range Stereocontrol: Access to Remote Quaternary Stereocenters through a Desymmetrizing Suzuki-Miyaura Reaction. J. Am. Chem. Soc. 144, 123-129 (2022).28. Bisht, R., Haldar, C., Hassan, M. M. M., Hoque, M. E., Chaturvedi, J. & Chattopadhyay, B. Metal-catalysed C-H bond activation and borylation. Chem. Soc. Rev., 51, 5042-5100 (2022).29. Ros, A., Fernandez, R. & Lassaletta, J. M. Functional group directed C-H borylation. Chem. Soc. Rev. 43, 3229-3243 (2014).30. Kawamorita, S., Ohmiya, H., Hara, K., Fukuoka, A. & Sawamura, M. Directed Ortho Borylation of Functionalized Arenes Catalyzed by a Silica-Supported Compact Phosphine-Iridium System. J. Am. Chem. Soc., 131, 5058-5059 (2019).31. Boebel, T. A. & Hartwig, J. F. Silly-Directed, Iridium-Catalyzed ortho-Borylation of Arenes. A One-Pot ortho-Borylation of Phenols, Arylamines, and Alkylarenes. J. Am. Chem. Soc., 130, 7534-7535 (2008).32. Bisht R. & Chattopadhyay, B. Formal Ir-Catalyzed Ligand-Enabled Ortho and Meta Borylation of Aromatic Aldehydes via in Situ-Generated Imines. J. Am. Chem. Soc., 138, 84-87 (2016).33. Yang, L., Uemura, N. & Nakao, Y. meta-Selective C-H Borylation of Benzamides and Pyridines by an Iridium-Lewis Acid Bifunctional Catalyst. J. Am. Chem. Soc. 141, 7972-7979 (2019).
+
+<--- Page Split --->
+
+34. Davis, H. J., Madalina, M. T. & Phipps, R. J. Ion Pair-Directed Regiocontrol in Transition-Metal Catalysis: A Meta-Selective C-H Borylation of Aromatic Quaternary Ammonium Salts. J. Am. Chem. Soc. 138, 12759-12762 (2016).  
+35. Chaturvedi, J., Haldar, C., Bisht, R., Pandey, G. & Chattopadhyay, B. Meta Selective C-H Borylation of Sterically Biased and Unbiased Substrates Directed by Electrostatic Interaction. J. Am. Chem. Soc. 143, 7604-7611 (2021).  
+36. Mihai, M. T., Williams, B. D. & Phipps, R. J. Para-Selective C-H Borylation of Common Arene Building Blocks Enabled by Ion-Pairing with a Bulky Counteraction. J. Am. Chem. Soc. 141, 15477-15482 (2019).  
+37. Bastidas, J. R. M., Oleskey, T. J., Miller, S. L., Smith, M. R. & Maleczka, R. E. Para-Selective, Iridium-Catalyzed C-H Borylations of Sulfated Phenols, Benzyl Alcohols, and Anilines Directed by Ion-Pair Electrostatic Interactions. J. Am. Chem. Soc. 141, 15483-15487 (2019).  
+38. Chang, W. et. al. Computationally designed ligands enable tunable borylation of remote C-H bonds in arenes. Chem, 8, 1775-1788 (2022).  
+39. Engle, K. M., Mei, T., Wasa, M. & Yu, J-Q. Weak Coordination as a Powerful Means for Developing Broadly Useful C-H Functionalization Reactions. Acc. Chem. Res. 45, 788-802 (2012).  
+40. Leow, D., Li, G., Mei, T.-S. & Yu, J-Q. Activation of remote meta-C-H bonds assisted by an end-on template. Nature 486, 518-522 (2012).  
+41. Shi, H., Herron, A. N., Shao, Y., Shao, Q. & Yu, J-Q. Enantioselective remote meta-C-H arylation and alkylation via a chiral transient mediator, Nature 558, 581-585 (2018).  
+42. Gandeepan, P. & Ackermann, L. Transient Directing Groups for Transformative C-H Activation by Synergistic Metal Catalysis. Chem 4, 199-222 (2018).  
+43. Meng, G. et al. Achieving Site-Selectivity for C-H Activation Processes Based on Distance and Geometry: A Carpenter's Approach. J. Am. Chem. Soc. 142, 10571-10591 (2020).  
+44. Scott, K. A., Cox, P. B. & Njardarson, J. T. Phenols in Pharmaceuticals: Analysis of a Recurring Motif. J. Med. Chem. 65, 7044-7072 (2022).  
+45. Bartolomei, B., Gentile, G., Rosso, C., Filippini, G. & Prato, M. Turning the Light on Phenols: New Opportunities in Organic Synthesis. Chem. Eur. J. 27, 16062-16070 (2021).  
+46. Quideau, S., Deffieux, D., Douat-Casassus, C. & Pouységou, L. Angew. Chem. Int. Ed. 50, 586-621 (2011).  
+47. Huang, Z. & Lumb, J-P. Phenol-Directed C-H Functionalization. ACS Catal. 9, 521-555 (2019).  
+48. Dai, H-X., Li, G., Zhang, X-G., Stepan, A. F. & Yu, J-Q. Pd(II)-Catalyzed ortho- or meta-C-H Olefination of Phenol Derivatives. J. Am. Chem. Soc. 135, 7567-7571 (2013).  
+49. Wan, L., Dastbaravardeh, N., Li, G. & Yu, J-Q. Cross-Coupling of Remote meta-C-H Bonds Directed by a U-Shaped Template. J. Am. Chem. Soc. 135, 18056-18059 (2013).  
+50. Xu, J. et al. Sequential Functionalization of meta-C-H and ipso-C-O Bonds of Phenols, J. Am. Chem. Soc. 141, 1903-1907 (2019).  
+51. Iverson, C. N. & Smith, M. R. III., Stoichiometric and Catalytic B-C Bond Formation from Unactivated Hydrocarbons and Boranes. J. Am. Chem. Soc. 121, 7696-7697 (1999).  
+52. Ishiyama, T. et al. Mild iridium-catalyzed borylation of arenes. High turnover numbers, room temperature reactions, and isolation of a potential intermediate. J. Am. Chem. Soc. 124, 390-391 (2002).  
+53. Mkhalid, I. A. I., Barnard, J. H., Marder, T. B., Murphy, J. M. & Hartwig, J. F. C-H Activation for the Construction of C-B Bonds. Chem. Rev. 110, 890-931 (2010).  
+54. Hartwig, J. F. Regioselectivity of the borylation of alkanes and arenes. Chem. Soc. Rev. 40, 1992-2002 (2011).  
+55. Boller, T. M., Murphy, J. M., Hapke, M., Ishiyama, T., Miyaura, N. & Hartwig, J. F. Mechanism of the Mild Functionalization of Arenes by Diboron Reagents Catalyzed by Iridium Complexes. Intermediacy and Chemistry of Bipyridine-Ligated Iridium Trisboryl Complexes. J. Am. Chem. Soc. 127, 14263-14278 (2005).  
+56. Haldar, C., Hoque, M. E., Chaturvedi, J., Hassan, M. M. M. & Chattopadhyay, B. Ir-catalyzed proximal and distal C-H borylation of arenes. Chem. Commun., 57, 13059-13074 (2021).  
+57. Muhammad, S., Bassindale, A. R., Taylor, P. G., Male, L., Coles, S. J. & Hursthouse, M. B. Study of Binuclear Silicon Complexes of Diketopiperazine at SN2 Reaction Profile. Organometallics 30, 564-571 (2011).  
+58. Sohail, M. et al. Synthesis and Hydrolysis-Condensation Study of Water-Soluble Self-Assembled Pentacoordinate Polysilylamides. Organometallics 32, 1721-1731 (2013).  
+59. Li, Z. et. al. A tautomeric ligand enables directed C-H hydroxylation with molecular oxygen. Science 372, 1452-1457 (2021).  
+60. Lajiness, J. P. et. al. Design, Synthesis, and Evaluation of Duocarmycin O-Amino Phenol Prodrugs Subject to Tunable Reductive Activation. J. Med. Chem. 53, 7731-7738 (2010)
+
+<--- Page Split --->
+
+61. Preshlock S. M. et al. A Traceless Directing Group for C-H Borylation. Angew. Chem., Int. Ed. 52, 12915-12919 (2013).  
+62. Li, H. L., Kuninobu, Y. & Kanai, M. Lewis Acid-Base Interaction-Controlled ortho-Selective C-H Borylation of Aryl Sulfides. Angew. Chem., Int. Ed. 56, 1495-1499 (2017).  
+63. Sun, W. et al. Chemodivergent transformations of amides using gem-diborylalkanes as pro-nucleophiles. Nat Commun 11, 3113 (2020).  
+64. Hua, Z., Vassar, V. C., Choi, H. & Ojima, I. New biphenol-based, fine-tunable monodentate phosphoramidite ligands for catalytic asymmetric transformations. Proc Natl Acad Sci, 101, 5411-5416 (2004).  
+65. Feng, Y., Holte, D., Zoller, J., Umemiya, S., Simke, L. R., Baran, P. S. Total Synthesis of Verruculogen and Fumitremorgin A Enabled by Ligand-Controlled C-H Borylation. J. Am. Chem. Soc. 137, 10160-10163 (2015).  
+66. Zhang, L. & Ritter, T. A Perspective on Late-Stage Aromatic C-H Bond Functionalization. J. Am. Chem. Soc. 144, 2399-2414 (2022).  
+67. Guillemard, L., Kaplaneris, N., Ackermann, L. & Johansson, M. J. Late-stage C-H functionalization offers new opportunities in drug discovery. Nat. Rev. Chem. 5, 522-545 (2021).  
+68. Thareja, S., Zhu, M., Ji, X., Wang, B. Boron-based small molecules in disease detection and treatment (2013-2016). Heterocycl. Commun. 23, 137-153 (2017).  
+
+## Acknowledgements  
+
+We thank Centre of Biomedical Research (CBMR) for providing research facility. We also thank IIT Kanpur for the X- ray crystallography data collection. Funding: This work was supported by SERB- SUPRA grant (SPR/2019/000158). SD and SG thank CSIR for their JRF, MMMH thanks UGC for an SRF. Author contributions: BC conceived the concept. SG developed the ligand. SG, MMMH and SD performed the experiments. BC supervised the project. All authors contributed to writing and proofreading of manuscript and SI. Competing interests: We have filled an Indian Patent (Patent Application No: 202211036590) based on this work (including the ligand and catalyst). Data and materials availability: X- ray dataset for catalyst 10 is freely available at the Cambridge Crystallographic Data Centre under deposition number 2180880.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupportingInformation.pdf
+
+<--- Page Split --->

preprint/preprint__09b079d80b3efdeef69e111bf52a9dde06da16423667f03702d57d11bbbeec9b/preprint__09b079d80b3efdeef69e111bf52a9dde06da16423667f03702d57d11bbbeec9b_det.mmd

@@ -0,0 +1,194 @@
+<|ref|>title<|/ref|><|det|>[[44, 107, 785, 177]]<|/det|>
+# Meta Selective C-H Borylation Directed by Secondary Silicon Oxygen Interaction  
+
+<|ref|>text<|/ref|><|det|>[[44, 195, 755, 238]]<|/det|>
+Buddhadeb Chattopadhyay ( \(\boxed{ \begin{array}{r l} \end{array} }\) buddhadeb.c@cbmr.res.in) Center of Biomedical Research (CBMR) https://orcid.org/0000- 0001- 8473- 2695  
+
+<|ref|>text<|/ref|><|det|>[[44, 243, 395, 285]]<|/det|>
+Saikat Guria Center of Biomedical Research (CBMR)  
+
+<|ref|>text<|/ref|><|det|>[[44, 290, 395, 333]]<|/det|>
+Mirja Md Hassan Centre of Biomedical Research (CBMR)  
+
+<|ref|>text<|/ref|><|det|>[[44, 338, 395, 378]]<|/det|>
+Sayan Dey Center of Biomedical Research (CBMR)  
+
+<|ref|>text<|/ref|><|det|>[[44, 417, 102, 435]]<|/det|>
+Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 455, 137, 474]]<|/det|>
+Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 493, 320, 512]]<|/det|>
+Posted Date: August 17th, 2022  
+
+<|ref|>text<|/ref|><|det|>[[44, 531, 474, 551]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 1837437/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 568, 910, 611]]<|/det|>
+License: \(\circledcirc\) This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 629, 944, 695]]<|/det|>
+Additional Declarations: There is NO Competing Interest. We declare that the authors have no competing interests except that We have filled an Indian Patent (Patent Application No: 202211036590) based on this work (including the ligand and catalyst).
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[152, 80, 842, 100]]<|/det|>
+# Meta Selective C-H Borylation Directed by Secondary Silicon Oxygen Interaction  
+
+<|ref|>text<|/ref|><|det|>[[181, 115, 812, 134]]<|/det|>
+Saikat Guria†, Mirja Md Mahamudul Hassan†, Sayan Dey†, Buddhadeb Chattopadhyay†\*  
+
+<|ref|>text<|/ref|><|det|>[[140, 149, 857, 200]]<|/det|>
+†Department of Biological & Synthetic Chemistry, Centre of Biomedical Research, SGPGIMS Campus, Raebareli Road, Lucknow 226014, Uttar Pradesh, India \*Correspondence to: buddhadeb.c@cbmr.res.in  
+
+<|ref|>text<|/ref|><|det|>[[102, 215, 896, 359]]<|/det|>
+Summary Paragraph: Remote meta selective C- H functionalization<sup>1,2,3</sup> of aromatic compounds remains a challenging problem in chemical synthesis. Here, we report an iridium catalyst bearing a bidentate pyridine- pyridone (PY- PYRI) ligand framework that efficiently catalyzes this meta selective borylation reaction. We demonstrate that the developed concept can be employed to introduce a boron functionality at the remote meta position of phenols, phenol containing bioactive and drug molecules, which was an extraordinary challenge. Moreover, we have demonstrated that the method can also be applied for the remote C6 borylation of indole derivatives including tryptophan that was the key synthetic precursor for the total synthesis of Verruculogen and Fumitremorgin A alkaloids. The origin of the remote meta selectivity was described as a secondary silicon oxygen interaction<sup>4</sup> that was never used in C- H functionalization chemistry.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[102, 78, 895, 304]]<|/det|>
+Transition metal- catalyzed C- H bond activation and functionalization \(^{5,6,7,8,9,10,11}\) of aromatic compounds has been branded as one of the most significant chemical transformations. This has a profound impact in modern synthetic organic chemistry, ranging from laboratory methods to industrial deployment. \(^{12,13}\) However, the key underlying principles for the success of the metal catalysis lies on the two important factors, such as: (i) design and synthesis of new generation ligand framework that can produce highly reactive catalyst system \(^{14,15}\) and (ii) substrates' structure modifications \(^{16}\) by which site selectivity could be controlled by the steric crowding \(^{17,18,19,20,21}\) or various weak interactions \(^{22,23,24,25}\) of the aromatic compounds among several similar type of C- H bonds via the ligand- substrate pre- organization \(^{26,27}\) . In recent times, many elegant approaches \(^{28}\) have been developed for the functionalization of proximal \(^{15,29,30,31}\) and remote C- H bonds \(^{1,3,32,33,34,35,36,37,38}\) of arenes by the design of either new ligand frameworks with an extended architectures featuring a weak coordinating functional groups \(^{39}\) or templates \(^{40}\) as well as transient mediators \(^{41}\) or transient directing groups \(^{42}\) attached with the substrates. While ligand having an extended architecture or template approaches are extremely important to functionalize the remotely located C- H bonds of arenes, but requirement of multi- step preparation of the linkers of the ligands and templates of the aromatic substrates significantly limit the wide application of the methods. \(^{43}\)  
+
+<|ref|>text<|/ref|><|det|>[[102, 320, 896, 544]]<|/det|>
+Among numerous aromatic substrates, phenols are the most widespread aromatic compounds that acquired household products including several bioactive to important drug molecules. \(^{44}\) Moreover, it is well- documented that \(10\%\) of the top 200 selling pharmaceuticals contain a phenol and several others employ phenols as synthetic intermediates. \(^{45}\) Furthermore, phenols are also key components of the biopolymers melanin, lignin, resins and polyphenylene oxides. \(^{44,45,46}\) In industry, phenol is routinely used as a raw material to make numerous important components by means of its diversification via the synthetic manipulation. \(^{44,45}\) Thus, direct functionalization of phenols would be a significant development for the rapid access of numerous important products. \(^{46}\) In this context, traditional electrophilic substitution is an alternative methods that affords variously substituted phenols (Fig. 1A, a). \(^{47}\) Employing this method, one can easily access ortho and para substituted phenol derivatives, although often remain a chance to have mixture of isomers. However, functionalization of the remote meta C- H bonds of phenols is extremely difficult because of the extreme inertness of the meta C- H bonds. Several pioneering approaches have been developed by Yu and others either using template method \(^{48,49,50}\) or transient directing group by Larrosa \(^{2}\) (Fig 1A, b, c). But, achieving the meta functionalized products using these methods, it is essential to have specialized substrates that limits the application of the methods.  
+
+<|ref|>text<|/ref|><|det|>[[102, 558, 896, 862]]<|/det|>
+Having tremendous importance of catalytic C- H borylation \(^{51,52,53,54,55,56}\) in organic synthesis, we report here a concept for the meta selective C- H borylation of phenols via an unprecedented Si- O interaction that has never been utilized in C- H functionalization chemistry. Literature reports revealed \(^{4}\) that the most common structural motif for this O- Si interaction can be found in the amide skeleton, where a filled p- orbital of oxygen atom can interact with the vacant d- orbital of the tetracoordinated silicon atom consisting of at least one electronegative atom. The role of this electronegative atom is to make silicon atom more electropositive by developing a partial positive charge on the silicon atom (Fig. 1B, a). Notably, while various reports of O- Si weak interactions have been shown in a number of intermediates (Fig 1B, b- e), \(^{4,57,58}\) there was no report of utilization of these interactions in the catalysis research. Inspired from this background reports, \(^{4,57,58}\) we have proposed a hypothesis where phenol is protected with an easily removable electropositive silane group and silane protected phenol meet all the necessary criteria (having electropositive silane with attached electronegative oxygen atom) for the O- Si weak interaction with 2- pyridine moiety having amide skeleton (4). (Fig. 1C). The reaction design (ligand and catalyst design) is shown in Fig. 1D, a, b. The designed ligand (PY- PYRI) consists of two parts, one part is the simple pyridine unit (PY) and another part is a 2- pyridine (PYRI) unit \(^{59}\) which was redesigned by the skeletal modification of bipyridine core structure. The origin of the remote meta selectivity is presented in Fig. 1D, b. We hypothesized that the designed ligand (PY- PYRI) would control the remote meta selectivity owing to the following two considerations. Firstly, in presence of [Ir(cod)(OMe)] \(_{2}\) , the ligand (PY- PYRI) will generate a complex (Int- A) without tautomerization of the 2- pyridine unit. Secondly, the p- orbital of the oxygen atom of the 2- pyridine unit will interact with the vacant d- orbital of the tetracoordinated electropositive silicon atom of the substrate (Fig. 1D, b).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[105, 85, 884, 636]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[102, 639, 896, 698]]<|/det|>
+<center>Fig. 1. Conceptual outline and proposed concept for the meta borylation of phenol. A. Meta functionalization of phenol, a, electrophilic approach. b, template approach. c, transient directing group approach. B. Conceptual Background. a, common structural skeleton for O-Si interaction. b-e, known literature reports of O-Si interactions. C. Hypothesis, D. Reaction design. a, ligand design by structural modification, b, proposed approach for meta borylation. </center>  
+
+<|ref|>text<|/ref|><|det|>[[102, 713, 896, 888]]<|/det|>
+The designed ligand (PY- PYRI) was prepared by the known synthetic methods (SI, for details), which was employed for the reaction optimization of steering silane group attached with phenol (Fig. 2A). We started our initial studies with the substrate (1a- I) featuring SiMe3 as the steering group under iridium- catalyzed conditions using the designed ligand (L1: PY- PYRI) at \(40^{\circ}C\) temperature, which afforded good meta/para selectivity ( \(\mathrm{m} / \mathrm{p} = 87 / 13\) ) with \(68\%\) conversion. Evaluating other silane based steering groups under the same reaction conditions, we observed that \(\mathrm{Si}(\mathrm{Pr})_3\) produced best meta selectivity (entry 2a: \(\mathrm{m} / \mathrm{p} = 94 / 6\) ) with excellent conversion ( \(95\%\) ). With this optimized \(\mathrm{Si}(\mathrm{Pr})_3\) as steering group, different other ligands have also been tested to observe the effect on the selectivity (Fig. 2B). It was found that replacing the tert- butyl group with methyl group, ligand (L2) gave \(91\%\) meta selectivity with less conversions ( \(68\%\) ). Similar selectivity was obtained when the reaction was performed with the ligand (L3) without any substituents on the pyridine unit of the ligand. Notably, the meta selectivity and conversion was found to be less using the bipyridine ligand (L4) compared to the ligands (L1- L3), which indicated the important role of the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[102, 81, 895, 113]]<|/det|>
+2- pyridine unit of the designed (PY- PYRI) ligand. Employment of other ligands (L5- L8) resulted in no reaction except the ligands (L9 & L10), which gave moderate meta selectivity.  
+
+<|ref|>image<|/ref|><|det|>[[104, 115, 888, 570]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[102, 574, 895, 618]]<|/det|>
+<center>Fig. 2. Ligand design and optimization of reaction conditions. A. Optimization of steering group, B. Deleterious results with other ligands. C. Origin of meta selectivity, D. Proof of concept. Reactions are on 0.2 mmol scales. \(^{a}\) Conversion was reported. In parenthesis, isolated yields are reported. See SI for details. </center>  
+
+<|ref|>text<|/ref|><|det|>[[102, 634, 895, 778]]<|/det|>
+At the outset, we proposed the tentative hypothesis that an O- Si secondary interaction between the oxygen atom of the 2- pyridine unit of the ligand and the silicon atom of the substrate's steering group would interact each other via the filled p- orbital and empty d- orbital (Fig. 2C). \(^{4}\) Moreover, due to the high electronegativity difference between oxygen and silicon, the O- Si bond will be highly polarized, thereby may interact via a weak O- Si interaction. To prove this hypothesis for the origin of the meta selectivity, we performed a reaction with substrates featuring C- Si bond (- CH2SiPr3: 1a- VII and - CH2SiMe3: 1a- VIII, Fig. 2D), in which lacking of electronegative atom attached with silicon atom causes non- polarised bond, resulted in very low meta selective borylation due to loss of interaction between 2- pyridine of catalyst and substrates silicon atom. This experiment indicated an "O- Si" secondary interaction between the ligand and substrate that guides the selectivity of the borylation.  
+
+<|ref|>text<|/ref|><|det|>[[102, 793, 895, 905]]<|/det|>
+Using L1 (PY- PYRI) as ligand and Si(Pr)3 as steering group, we next performed the iridium- catalyzed meta borylation of a variety of phenols that afforded excellent meta selectivity and yields of the isolated borylated products (Fig. 3). For example, we first tested 2- chlorophenol for the borylation reaction, while our designed ligand (L1) gave high meta selectivity (m/p = 92/8), traditional dtbpy ligand provided poor meta selectivity (m/p = 63/37), which clearly demonstrated the utility of the designed (L1: PY- PYRI) ligand. Other 2- substituted phenols, such as 2- bromo (1c) and 2- iodo (1d) afforded high meta selectivity that have great synthetic values owing to the two different types of handles on the phenols. Likewise, phenols bearing various alkyl chain ranging from methyl to pentyl (1e- 1g) at the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[102, 80, 896, 207]]<|/det|>
+2- positions along with trifluoromethyl (1h), isopropyl (1i) and trifluoromethoxy (1l) smoothly underwent meta borylation irrespective of the nature of the substituent. Amino phenol (1k), substrate of momentous importance for the chemical and pharmaceutical industries<sup>60</sup> is borylated with high meta selectivity \((m / p = 97 / 3)\) without borylation next to the amino group, which is known to give ortho borylation under iridium- catalyzed borylation conditions via in situ generation of NHBpin group.<sup>61</sup> Thioether (1l) that usually directs borylation at the ortho position<sup>62</sup> also underwent borylation with good meta selectivity. We observed that phenols containing functional groups such as cyano (1m), Bpin (1n), cyclic amine (1o), cyclohexyl (1p), ketomethyl (1q) and homologous ester (1r) afforded high level of meta selectivity and tolerated well under the employed reaction conditions.  
+
+<|ref|>image<|/ref|><|det|>[[102, 207, 895, 802]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[101, 803, 896, 833]]<|/det|>
+<center>Fig. 3. Substrates scope for substituted arenes. Reactions are in 0.5 mmol scale. <sup>a</sup>Conversion was reported. <sup>b</sup>1.5 equiv. B<sub>2pin</sub> was used. <sup>c</sup>2.0 equiv. B<sub>2pin</sub> was used. See SI for details. </center>  
+
+<|ref|>text<|/ref|><|det|>[[102, 850, 895, 914]]<|/det|>
+Amide functionalities (1s & 1t) that are known to undergo numerous synthetic transformations<sup>63</sup> exhibited excellent meta selective borylation. Borylation of phenols having \(\mathrm{CF}_3\) (1h) and CN (1m) substituents at the ortho position afforded exclusively meta borylation, the same substituents at the meta position of phenols (1u & 1v) also gave meta selective borylation, which indicated the generality of the developed method. Moreover, fluoro- substituted
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[102, 80, 896, 192]]<|/det|>
+arene, which typically gives borylation next to the fluorine atom under standard iridium- catalyzed conditions, in this case, 3- fluorophenol (1w) gave meta borylation as the major product. Several disubstituted phenols (1x- 1ag) were also examined under the developed conditions that reacted smoothly to afford variously substituted meta borylated products in high yields. 2,2'- Biphenol, compound of paramount importance in medicinal chemistry as well as in chemical industry64 can selectively be mono- and diborylation (2ah & 2ai) by tuning the amount of boron reagent. A bulky substituent at the ortho positions (1aj) did not hamper the reaction that gave 96% meta borylated product with 90% isolated yield.  
+
+<|ref|>image<|/ref|><|det|>[[105, 194, 884, 775]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[102, 777, 894, 808]]<|/det|>
+<center>Fig. 4. Substrates scope for the 4-substituted arenes and C6 borylation of indoles. Reactions are in 0.5 mmol scale. aConversions were reported. b2.0 equiv. B2pin2. See SI for details. </center>  
+
+<|ref|>text<|/ref|><|det|>[[103, 823, 895, 902]]<|/det|>
+Next, we focused on the meta borylation of those phenols bearing a substituent at the para position (Fig. 4). Because, borylation at the remote meta position in presence of a para substituent remains an extraordinary challenge due to the steric reason. Moreover, we selected those substituents at the para position that already provided exclusive meta borylation of phenols when they were located at either ortho or meta positions. The reason for this selection is mainly to observe the overall effects of the borylation by the same substituents. For the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[102, 80, 895, 256]]<|/det|>
+testification, we begun with the 4- methyl phenol (3a) that afforded \(91\%\) meta selective borylation. Increasing the chain length from small methyl group to the relatively bulkier alkyl groups such as, ethyl (3b), pentyl (3c), hexyl (3d) and isopropyl (3e), the borylation underwent smoothly with further enhancement of the meta selectivity from \(91\%\) to \(100\%\) . Para- substituted ethers and thioethers bearing electronically different substituents (3f- 3j) reacted with \(100\%\) meta selectivity, which revealed that the scope of the meta borylation is very general regardless of the nature of the substituents. While 2- CN, 3- CN as well as 2- CF3 and 3- CF3 bearing phenols resulted in excellent meta borylation, the same substituents at the para position reacted to yield \(100\%\) meta borylation. Likewise, we also observed that chloro (3m) and bromo (3n) containing phenols reacted to give the meta borylation products solely irrespective of their position in the phenol. Moreover, it has been found that the phenols featuring bulky substituents at the para position (3o- 3r) also gave exclusively meta borylation, although conversion was moderate in case of the cyclohexyl group.  
+
+<|ref|>text<|/ref|><|det|>[[102, 271, 895, 415]]<|/det|>
+In 2015, Baran et al. reported \(^{65}\) the first total synthesis of Verruculogen and Fumitremorgin A enabled by ligand- controlled C- H borylation as the key step of TIPS protected tryptophan. We were curious if our designed ligand system could provide the remote C6 borylation of TIPS protected indoles and TIPS protected tryptophan. For that reason we performed borylation of TIPS- protected tryptophan (3s) (synthetic key precursor of bioactive alkaloids Verruculogen and Fumitremorgin A) which provided C6 borylation with \(91\%\) selectivity with excellent conversions. We also found that TIPS- protected other indole derivatives (3t & 3u) and TIPS- protected carbazole (3v) smoothly underwent remote borylation affording excellent selectivity and conversion. This developed method provided a simple way to borylate the 3- substituted indoles derivatives that might be beneficial for the total synthesis or the late- stage functionalization of several bioactive molecules.  
+
+<|ref|>text<|/ref|><|det|>[[102, 430, 895, 654]]<|/det|>
+Late- stage functionalization \(^{66}\) of complex bioactive and medicinally important molecules by the site selective C- H activation is a powerful method for the development of new drug candidates. \(^{67}\) In this context, introducing a boron functionality into the bioactive and medicinally important molecules would further enhance the identification of new lead molecules not only for the enormous importance of the boron- bearing small molecules \(^{68}\) but also for the uniqueness of the boron group towards the diverse derivatization towards numerous other functional groups. Thus, we tested our developed method for several commercially available bioactive and drug molecules (Fig. 5A). For example, cannabinoid core (5a: used as a psychoactive drug), methyl salicylate derivatives (5b: an anti- inflammatory and analgesic agent), tyrosol derivatives (5c: an antioxidant), eugenol derivatives (5d: a flavouring agent), sesamol derivatives (5e: an antioxidant), naproxen derivatives (5f: a nonsteroidal anti- inflammatory drug, NSAID), deoxyarbutin derivatives (5g: used for treatment of hyperpigmentation disorders) and homosalate (5h: used as a sunscreen) were meta borylated with high yield and selectivity. The steering silane group from the borylated phenols has been removed under a very mild reaction conditions at room temperature (ethylene glycol, KF, 1h) that afforded the meta borylated phenols in high yields (Fig. 5B). Notably, the meta borylated phenols can further be transformed to a number of substituted phenols/resorcinols that are difficult to prepare by otherwise.  
+
+<|ref|>text<|/ref|><|det|>[[102, 669, 895, 829]]<|/det|>
+Next, we aimed to prepare the active catalyst (10) that was proposed to form in situ between the reaction of the designed ligand (L1: PY- PYRI) and [Ir(cod)(OME)]2 during the meta selective borylation conditions. Accordingly, we performed the reaction and isolated the catalyst [10: Ir(cod)(PY- PYRI)] in \(90\%\) yield (Fig. 5C). The catalyst structure was confirmed by X- ray crystallography and other spectroscopic data. The catalytic efficiency of this catalyst [10: Ir(cod)(PY- PYRI)] was further tested in the meta borylation reactions, which exhibited same level of meta selectivity with better product conversion (compared to the in situ generation) (Fig. 5D). Moreover, we checked the stability of the catalyst [10: Ir(cod)(PY- PYRI)] and found highly stable that can even be stored in open air. Furthermore, to verify the broad utility of this air stable catalyst [10: Ir(cod)(PY- PYRI)], we performed several test experiments using this catalyst [10: Ir(cod)(PY- PYRI)] that was stored in open air and found no loss of catalytic activity even after 30 days (Fig. 5D, SI for details).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[103, 84, 888, 448]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[102, 450, 896, 494]]<|/det|>
+<center>Fig. 5. Applications, Catalyst Preparation and Testing. A. Late-stage meta-C-H borylation, B. Removal of silane group, C. Catalyst synthesis, D. Test of reactivity of catalyst 10. Reactions are in 0.5 mmol scale. aConversions were reported. See SI for details. </center>  
+
+<|ref|>text<|/ref|><|det|>[[103, 510, 896, 639]]<|/det|>
+In conclusion, we report a new class of ligand and catalyst that has demonstrated remarkable efficiency for the remote meta selective borylation of phenols featuring all types of substitutions at the arene ring. In addition, we have seen that our developed ligand system is beneficial for the remote C6- borylation of indoles derivatives including tryptophan which is a synthetic precursor of bioactive alkaloids (Verruculogen and Fumitremorgin A). Several late- stage meta borylations have been showcased with bioactive and drug molecules that might be useful for repurposing medicines and identification of new lead drug candidates. For the first time, an "O- Si" secondary interaction has been employed to tune the remote selectivity. We anticipate that the designed ligand and catalyst will also find wide application in the context of other C- H functionalization reactions.  
+
+<|ref|>sub_title<|/ref|><|det|>[[104, 656, 264, 671]]<|/det|>
+## References and notes  
+
+<|ref|>text<|/ref|><|det|>[[101, 687, 897, 890]]<|/det|>
+1. Zhang, Z., Tanaka, K. & Yu, J-Q. Remote site-selective C-H activation directed by a catalytic bifunctional template. Nature 543, 538-542 (2017).  
+2. Luo, J., Preciado, S. & Larrosa, I. Overriding Ortho-Para Selectivity via a Traceless Directing Group Relay Strategy: The Meta-Selective Arylation of Phenols, J. Am. Chem. Soc. 136, 4109-4112 (2014).  
+3. Sinha, S. K., Guin, S., Maiti, S., Biswas, J. P., Porey, S. & Maiti, D. Toolbox for Distal C-H Bond Functionalizations in Organic Molecules. Chem. Rev. 122, 5682-5841 (2022).  
+4. Lazareva, N. F., Sterkhova, I. V. & Vashchenko, A. V. N-[difluoro(methyl)silyl]carboxamides: Synthesis, structural features and theoretical estimating of Si<O- dative bond energy. Journal of Molecular Structure 1225, 129130 (2021).  
+5. Lyons, T. W. & Sanford, M. S. Palladium-Catalyzed Ligand-Directed C-H Functionalization Reactions. Chem. Rev. 110, 1147-1169 (2010).  
+6. Davies, H. M. L., Bois, J. D. & Yu, J-Q. C-H Functionalization in organic synthesis. Chem. Soc. Rev. 40, 1855-1856 (2011).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[100, 92, 897, 905]]<|/det|>
+7. Arockiam, P. B., Bruneau, C. & Dixneuf, P. H. Ruthenium (II)-Catalyzed C-H Bond Activation and Functionalization. Chem. Rev. 112, 5879-5918 (2012).8. Sambiagio, C. et al. A comprehensive overview of directing groups applied in metal-catalysed C-H functionalisation chemistry. Chem. Soc. Rev. 47, 6603-6743 (2018).9. Crabtree, R. H. & Lei, A. Introduction: CH Activation. Chem. Rev. 117, 8481-8482 (2017).10. Shilov, A. E. & Shul'pin, G. B. Activation of C-H Bonds by Metal Complexes. Chem. Rev. 97, 2879-2932 (1997).11. Rogge, T. et al. C-H activation. Nat Rev Methods Primers 1, (2021). doi.org/10.1038/s43586-021-00041-2.12. Yamaguchi, J., Yamaguchi, A. D. & K. Itami, C-H Bond Functionalization: Emerging Synthetic Tools for Natural Products and Pharmaceuticals. Angew. Chem. Int. Ed. 51, 8960-9009 (2012).13. Dalton, T., Faber, T. & Glorius, F. C-H Activation: Toward Sustainability and Applications. ACS Cent. Sci. 7, 245-261 (2021).14. Liao, K. et al. Design of catalysts for site-selective and enantioselective functionalization of non-activated primary C-H bonds. Nature Chem 10, 1048-1055 (2018).15. Hoque, M. E., Hassan, M. M. M. & Chattopadhyay, B. Remarkably Efficient Iridium Catalysts for Directed C(sp2)-H and C(sp3)-H Borylation of Diverse Classes of Substrates. J. Am. Chem. Soc. 143, 5022-5037 (2021).16. Hoveyda, A. H., Evans, D. A. & Fu, G. C. Substrate-directable chemical reactions. Chem. Rev. 93, 1307-1370 (1993).17. Cheng, C. & Hartwig, J. F. Rhodium-Catalyzed Intermolecular C-H Silylation of Arenes with High Steric Regiocontrol. Science, 343, 853-857 (2014).18. Ramadoss, B., Jin, Y., Asako, S. & Ilies, L. Remote steric control for undirected meta-selective C-H activation of arenes, Science, 375, 658-663 (2022).19. Cho, J. Y., Tse, M. K., Holmes, D., Maleczka, R. E. & Smith, M. R. Remarkably selective iridium catalysts for the elaboration of aromatic C-H bonds. Science 295, 305-308 (2002).20. Saito, Y., Segawa Y. & Itami, K. para-C-H Borylation of Benzene Derivatives by a Bulky Iridium Catalyst. J. Am. Chem. Soc., 137, 5193-5198 (2015).21. Mondal, A., Chen, H., Flamig, L., Wedi, P. & Gemmeren, M. V. Sterically Controlled Late-Stage C-H Alkynylation of Arenes. J. Am. Chem. Soc. 141, 18662-18667 (2019).22. Kuninobu, Y., Ida, H., Nishi, M. & Kanai, M. A meta-selective C-H borylation directed by a secondary interaction between ligand and substrate. Nat. Chem. 7, 712-717 (2015).23. Fanourakis, A., Docherty, P. J., Chuentragool, P. & Phipps, R. J. Recent Developments in Enantioselective Transition Metal Catalysis Featuring Attractive Noncovalent Interactions between Ligand and Substrate, ACS Catalysis, 10, 10672-10714 (2020).24. Zhang, T., Luan, Y., Lam, N. Y. S., Li, J., Li, Y., Ye, M. Yu, & J-Q. A directive Ni catalyst overrides conventional site selectivity in pyridine C-H alkenylation. Nat. Chem. 13, 1207-1213 (2021).25. Hoque, M. E., Bisht, R., Haldar, C. & Chattopadhyay, B. Noncovalent Interactions in Ir-Catalyzed C-H Activation: L-Shaped Ligand for Para-Selective Borylation of Aromatic Esters. J. Am. Chem. Soc. 139, 7745-7748 (2017).26. Dydio, P. & Reek, J. N. H. Supramolecular control of selectivity in transitionmetal catalysis through substrate preorganization. Chem. Sci., 5, 2135-2145 (2014).27. Lou, Y., Wei, J., Li, M. & Zhu, Y. Distal Ionic Substrate-Catalyst Interactions Enable Long-Range Stereocontrol: Access to Remote Quaternary Stereocenters through a Desymmetrizing Suzuki-Miyaura Reaction. J. Am. Chem. Soc. 144, 123-129 (2022).28. Bisht, R., Haldar, C., Hassan, M. M. M., Hoque, M. E., Chaturvedi, J. & Chattopadhyay, B. Metal-catalysed C-H bond activation and borylation. Chem. Soc. Rev., 51, 5042-5100 (2022).29. Ros, A., Fernandez, R. & Lassaletta, J. M. Functional group directed C-H borylation. Chem. Soc. Rev. 43, 3229-3243 (2014).30. Kawamorita, S., Ohmiya, H., Hara, K., Fukuoka, A. & Sawamura, M. Directed Ortho Borylation of Functionalized Arenes Catalyzed by a Silica-Supported Compact Phosphine-Iridium System. J. Am. Chem. Soc., 131, 5058-5059 (2019).31. Boebel, T. A. & Hartwig, J. F. Silly-Directed, Iridium-Catalyzed ortho-Borylation of Arenes. A One-Pot ortho-Borylation of Phenols, Arylamines, and Alkylarenes. J. Am. Chem. Soc., 130, 7534-7535 (2008).32. Bisht R. & Chattopadhyay, B. Formal Ir-Catalyzed Ligand-Enabled Ortho and Meta Borylation of Aromatic Aldehydes via in Situ-Generated Imines. J. Am. Chem. Soc., 138, 84-87 (2016).33. Yang, L., Uemura, N. & Nakao, Y. meta-Selective C-H Borylation of Benzamides and Pyridines by an Iridium-Lewis Acid Bifunctional Catalyst. J. Am. Chem. Soc. 141, 7972-7979 (2019).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[100, 92, 897, 920]]<|/det|>
+34. Davis, H. J., Madalina, M. T. & Phipps, R. J. Ion Pair-Directed Regiocontrol in Transition-Metal Catalysis: A Meta-Selective C-H Borylation of Aromatic Quaternary Ammonium Salts. J. Am. Chem. Soc. 138, 12759-12762 (2016).  
+35. Chaturvedi, J., Haldar, C., Bisht, R., Pandey, G. & Chattopadhyay, B. Meta Selective C-H Borylation of Sterically Biased and Unbiased Substrates Directed by Electrostatic Interaction. J. Am. Chem. Soc. 143, 7604-7611 (2021).  
+36. Mihai, M. T., Williams, B. D. & Phipps, R. J. Para-Selective C-H Borylation of Common Arene Building Blocks Enabled by Ion-Pairing with a Bulky Counteraction. J. Am. Chem. Soc. 141, 15477-15482 (2019).  
+37. Bastidas, J. R. M., Oleskey, T. J., Miller, S. L., Smith, M. R. & Maleczka, R. E. Para-Selective, Iridium-Catalyzed C-H Borylations of Sulfated Phenols, Benzyl Alcohols, and Anilines Directed by Ion-Pair Electrostatic Interactions. J. Am. Chem. Soc. 141, 15483-15487 (2019).  
+38. Chang, W. et. al. Computationally designed ligands enable tunable borylation of remote C-H bonds in arenes. Chem, 8, 1775-1788 (2022).  
+39. Engle, K. M., Mei, T., Wasa, M. & Yu, J-Q. Weak Coordination as a Powerful Means for Developing Broadly Useful C-H Functionalization Reactions. Acc. Chem. Res. 45, 788-802 (2012).  
+40. Leow, D., Li, G., Mei, T.-S. & Yu, J-Q. Activation of remote meta-C-H bonds assisted by an end-on template. Nature 486, 518-522 (2012).  
+41. Shi, H., Herron, A. N., Shao, Y., Shao, Q. & Yu, J-Q. Enantioselective remote meta-C-H arylation and alkylation via a chiral transient mediator, Nature 558, 581-585 (2018).  
+42. Gandeepan, P. & Ackermann, L. Transient Directing Groups for Transformative C-H Activation by Synergistic Metal Catalysis. Chem 4, 199-222 (2018).  
+43. Meng, G. et al. Achieving Site-Selectivity for C-H Activation Processes Based on Distance and Geometry: A Carpenter's Approach. J. Am. Chem. Soc. 142, 10571-10591 (2020).  
+44. Scott, K. A., Cox, P. B. & Njardarson, J. T. Phenols in Pharmaceuticals: Analysis of a Recurring Motif. J. Med. Chem. 65, 7044-7072 (2022).  
+45. Bartolomei, B., Gentile, G., Rosso, C., Filippini, G. & Prato, M. Turning the Light on Phenols: New Opportunities in Organic Synthesis. Chem. Eur. J. 27, 16062-16070 (2021).  
+46. Quideau, S., Deffieux, D., Douat-Casassus, C. & Pouységou, L. Angew. Chem. Int. Ed. 50, 586-621 (2011).  
+47. Huang, Z. & Lumb, J-P. Phenol-Directed C-H Functionalization. ACS Catal. 9, 521-555 (2019).  
+48. Dai, H-X., Li, G., Zhang, X-G., Stepan, A. F. & Yu, J-Q. Pd(II)-Catalyzed ortho- or meta-C-H Olefination of Phenol Derivatives. J. Am. Chem. Soc. 135, 7567-7571 (2013).  
+49. Wan, L., Dastbaravardeh, N., Li, G. & Yu, J-Q. Cross-Coupling of Remote meta-C-H Bonds Directed by a U-Shaped Template. J. Am. Chem. Soc. 135, 18056-18059 (2013).  
+50. Xu, J. et al. Sequential Functionalization of meta-C-H and ipso-C-O Bonds of Phenols, J. Am. Chem. Soc. 141, 1903-1907 (2019).  
+51. Iverson, C. N. & Smith, M. R. III., Stoichiometric and Catalytic B-C Bond Formation from Unactivated Hydrocarbons and Boranes. J. Am. Chem. Soc. 121, 7696-7697 (1999).  
+52. Ishiyama, T. et al. Mild iridium-catalyzed borylation of arenes. High turnover numbers, room temperature reactions, and isolation of a potential intermediate. J. Am. Chem. Soc. 124, 390-391 (2002).  
+53. Mkhalid, I. A. I., Barnard, J. H., Marder, T. B., Murphy, J. M. & Hartwig, J. F. C-H Activation for the Construction of C-B Bonds. Chem. Rev. 110, 890-931 (2010).  
+54. Hartwig, J. F. Regioselectivity of the borylation of alkanes and arenes. Chem. Soc. Rev. 40, 1992-2002 (2011).  
+55. Boller, T. M., Murphy, J. M., Hapke, M., Ishiyama, T., Miyaura, N. & Hartwig, J. F. Mechanism of the Mild Functionalization of Arenes by Diboron Reagents Catalyzed by Iridium Complexes. Intermediacy and Chemistry of Bipyridine-Ligated Iridium Trisboryl Complexes. J. Am. Chem. Soc. 127, 14263-14278 (2005).  
+56. Haldar, C., Hoque, M. E., Chaturvedi, J., Hassan, M. M. M. & Chattopadhyay, B. Ir-catalyzed proximal and distal C-H borylation of arenes. Chem. Commun., 57, 13059-13074 (2021).  
+57. Muhammad, S., Bassindale, A. R., Taylor, P. G., Male, L., Coles, S. J. & Hursthouse, M. B. Study of Binuclear Silicon Complexes of Diketopiperazine at SN2 Reaction Profile. Organometallics 30, 564-571 (2011).  
+58. Sohail, M. et al. Synthesis and Hydrolysis-Condensation Study of Water-Soluble Self-Assembled Pentacoordinate Polysilylamides. Organometallics 32, 1721-1731 (2013).  
+59. Li, Z. et. al. A tautomeric ligand enables directed C-H hydroxylation with molecular oxygen. Science 372, 1452-1457 (2021).  
+60. Lajiness, J. P. et. al. Design, Synthesis, and Evaluation of Duocarmycin O-Amino Phenol Prodrugs Subject to Tunable Reductive Activation. J. Med. Chem. 53, 7731-7738 (2010)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[100, 95, 896, 328]]<|/det|>
+61. Preshlock S. M. et al. A Traceless Directing Group for C-H Borylation. Angew. Chem., Int. Ed. 52, 12915-12919 (2013).  
+62. Li, H. L., Kuninobu, Y. & Kanai, M. Lewis Acid-Base Interaction-Controlled ortho-Selective C-H Borylation of Aryl Sulfides. Angew. Chem., Int. Ed. 56, 1495-1499 (2017).  
+63. Sun, W. et al. Chemodivergent transformations of amides using gem-diborylalkanes as pro-nucleophiles. Nat Commun 11, 3113 (2020).  
+64. Hua, Z., Vassar, V. C., Choi, H. & Ojima, I. New biphenol-based, fine-tunable monodentate phosphoramidite ligands for catalytic asymmetric transformations. Proc Natl Acad Sci, 101, 5411-5416 (2004).  
+65. Feng, Y., Holte, D., Zoller, J., Umemiya, S., Simke, L. R., Baran, P. S. Total Synthesis of Verruculogen and Fumitremorgin A Enabled by Ligand-Controlled C-H Borylation. J. Am. Chem. Soc. 137, 10160-10163 (2015).  
+66. Zhang, L. & Ritter, T. A Perspective on Late-Stage Aromatic C-H Bond Functionalization. J. Am. Chem. Soc. 144, 2399-2414 (2022).  
+67. Guillemard, L., Kaplaneris, N., Ackermann, L. & Johansson, M. J. Late-stage C-H functionalization offers new opportunities in drug discovery. Nat. Rev. Chem. 5, 522-545 (2021).  
+68. Thareja, S., Zhu, M., Ji, X., Wang, B. Boron-based small molecules in disease detection and treatment (2013-2016). Heterocycl. Commun. 23, 137-153 (2017).  
+
+<|ref|>sub_title<|/ref|><|det|>[[104, 343, 238, 357]]<|/det|>
+## Acknowledgements  
+
+<|ref|>text<|/ref|><|det|>[[100, 360, 896, 530]]<|/det|>
+We thank Centre of Biomedical Research (CBMR) for providing research facility. We also thank IIT Kanpur for the X- ray crystallography data collection. Funding: This work was supported by SERB- SUPRA grant (SPR/2019/000158). SD and SG thank CSIR for their JRF, MMMH thanks UGC for an SRF. Author contributions: BC conceived the concept. SG developed the ligand. SG, MMMH and SD performed the experiments. BC supervised the project. All authors contributed to writing and proofreading of manuscript and SI. Competing interests: We have filled an Indian Patent (Patent Application No: 202211036590) based on this work (including the ligand and catalyst). Data and materials availability: X- ray dataset for catalyst 10 is freely available at the Cambridge Crystallographic Data Centre under deposition number 2180880.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[61, 130, 318, 150]]<|/det|>
+SupportingInformation.pdf
+
+<--- Page Split --->

preprint/preprint__09b108cd9ad0034506ae629f2fb23d3bf7f4cd8c58a23feee09b7b89729c5397/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1: Thermo-hydrodynamic manipulation of Au NPs in NaCl solution. a, The sample consists of two glass slides that confine a \\(3\\mu \\mathrm{m}\\) thin liquid film of gold nanoparticles (AuNPs) dispersed in aqueous NaCl solution. The lower glass slide carries a \\(50\\mathrm{nm}\\) Au film that is locally heated by optical absorption of a focused laser of \\(\\lambda = 532\\mathrm{nm}\\) wavelength. b, The experimental setup comprises an inverted optical microscope equipped with an acusto-optical deflector controlled, steerable focused laser with a wavelength of \\(\\lambda = 532\\mathrm{nm}\\) . The AuNPs are observed using darkfield illumination with an oil-immersion dark-field condenser (NA 1.2) and a \\(100\\times\\) oil-immersion objective set to NA 0.6. Images are recorded with an EMCCD camera.",
+    "footnote": [],
+    "bbox": [
+      [
+        80,
+        81,
+        900,
+        333
+      ]
+    ],
+    "page_idx": 2
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2: DLVO potential, lateral diffusion analysis, temperature distribution and thermo-osmotic flow field. a, Plot of the DLVO potential, equation (1), between a \\(250~\\mathrm{nm}\\) Au NP and a \\(50~\\mathrm{nm}\\) Au film on a glass surface as function of surface-surface distance \\(d\\) for different NaCl concentrations \\(c_0\\) . The shaded curves display the calculated probability density for finding the particle at this distance at the different salt concentrations (see Supplementary Information for details). The vertical dashed lines correspond to the mean distance of the particle as calculated from the probability density for a \\(3\\mu \\mathrm{m}\\) liquid film height. b, The measured diffusion coefficient \\(D_1 / D_0\\) parallel to the Au film with respect to the bulk diffusion coefficient \\(D_0\\) as function of the NaCl concentration \\(c_0\\) . Symbols correspond to the experimental values. The lines reflect the theoretical prediction including a distance dependent diffusion coefficient for three different Hamaker constants (dotted: \\(A_H = 4\\cdot 10^{-20}\\mathrm{J}\\) , dash-dotted: \\(5\\cdot 10^{-20}\\mathrm{J}\\) and dashed: \\(6\\cdot 10^{-20}\\mathrm{J}\\) ) of gold according to a Boltzmann weighting (see text). c, Relation between the mean distance \\(\\langle d\\rangle\\) and the NaCl concentration \\(c_0\\) . The symbols and the horizontal lines denote the calculated distances for measured concentrations. d, Simulation of the relative temperature increment in the \\(xz\\) -plane of the sample. e, Experimentally obtained temperature increment \\(\\Delta T_{\\mathrm{max}}\\) as a function of the incident laser power \\(P_0\\) (green data points) compared to the simulated values (green curve). f, Measured thermo-osmotic flow field in the \\(xy\\) -plane in close proximity to the gold film ( \\(z< 500\\mathrm{nm}\\) ). g, Measured thermo-osmotic flow field in the \\(xz\\) -plane. h, Illustration of the measured flow field planes in f and g. i, j, The \\(x\\) - and \\(z\\) -component of the measured flow velocities compared to the simulation results in k and l.",
+    "footnote": [],
+    "bbox": [
+      [
+        78,
+        81,
+        920,
+        576
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3: Forces on trapped NPs in \\(10\\mathrm{mM}\\) NaCl. a, The lateral trap stiffness obtained from the experimental position histograms (blue data points) as function the laser power \\(P_0\\) compared to the simulation result (blue solid line). b, The \\(x\\) -component of the thermo-osmotic drag force \\(F_{\\mathrm{T}}^{\\mathrm{TF}}\\) (blue line), the optical force \\(F_{\\mathrm{Z}}^{\\mathrm{OF}}\\) (green line) and the total force \\(F_{\\mathrm{Z}}^{\\mathrm{OF}} + F_{\\mathrm{Z}}^{\\mathrm{TF}}\\) (black dashed line) as function the incident laser power \\(P_0\\) for a NP located at \\(x = 0\\) , \\(d = 30\\) ( \\(z = d + R\\) ). The attractive DLVO force \\(F_{\\mathrm{DLVO}}^{\\mathrm{DLVO}}\\) is independent of the incident laser power and depicted as horizontal, red line. c, Trajectory of a AuNP for a heating laser power of \\(1.25\\mathrm{mW}\\) (see Supplementary Video 2 for details). The inset shows the corresponding lateral distribution histogram. d, Time traces of the \\(z\\) -position for three different laser powers. e, Trajectory of a AuNP for a heating laser power of \\(2.5\\mathrm{mW}\\) , which is above the threshold power of \\(2.25\\mathrm{mW}\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        80,
+        81,
+        916,
+        253
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4: Manipulation of NPs over a Au film in NaCl and SDS solution. a, A AuNP with \\(50 \\mathrm{nm}\\) radius trapped at a NaCl of \\(30 \\mathrm{mM}\\) (Supplementary Video 5). b, Manipulation of two AuNPs by a multiplexed laser beam (Supplementary Video 6). c, Control of three AuNPs (Supplementary Video 7). d, Actuation of a single AuNP on a circular trajectory by a steerable laser beam (Supplementary Video 8). The green and white, dashed arrows denote the moving direction of the laser focus and particle, respectively. e, Generation of thermo-viscous flows by rotating the laser focus on a circle with a rotation frequency of \\(f = 500 \\mathrm{Hz}\\) at high laser powers (Supplementary Video 9). Note, that laser movement (green arrow) and the thermo-viscous flow (white, dashed arrow) and have opposite directions. f, Attraction of a AuNP and PS NPs in \\(5 \\mathrm{mM}\\) SDS due to depletion (Supplementary Video 10). g, An AuNP (125 nm radius) trapped in an ensemble of polystyrene (PS) NPs of the same size at \\(10 \\mathrm{mM}\\) NaCl (Supplementary Video 11), where the PS-particles are repelled due to thermophoresis. h, Attraction of PS ellipsoids (2.39 \\(\\mu \\mathrm{m}\\) major-axis length, \\(0.34 \\mu \\mathrm{m}\\) minor-axis length) in \\(5 \\mathrm{mM}\\) SDS (Supplementary Video 12).",
+    "footnote": [],
+    "bbox": [
+      [
+        80,
+        460,
+        916,
+        784
+      ]
+    ],
+    "page_idx": 6
+  }
+]

preprint/preprint__09b108cd9ad0034506ae629f2fb23d3bf7f4cd8c58a23feee09b7b89729c5397/preprint__09b108cd9ad0034506ae629f2fb23d3bf7f4cd8c58a23feee09b7b89729c5397.mmd

@@ -0,0 +1,185 @@
+
+# Hydrodynamic manipulation of nano-objects by optically induced thermo-osmotic flows  
+
+Martin Fränzl  Leipzig University https://orcid.org/0000- 0001- 6754- 8554  Frank Cichos (cichos@physik.uni- leipzig.de)  Leipzig University https://orcid.org/0000- 0002- 9803- 4975  
+
+## Article  
+
+Keywords: nano- objects, microfluidics, hydrodynamics, thermo- osti  
+
+Posted Date: September 24th, 2021  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 879955/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License.  Read Full License  
+
+Version of Record: A version of this preprint was published at Nature Communications on February 3rd, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 28212- z.
+
+<--- Page Split --->
+
+# Hydrodynamic manipulation of nano-objects by optically induced thermo-osmotic flows  
+
+Martin Fränzl \(^{1}\) and Frank Cichos \(^{1,*}\)  
+
+\(^{1}\) Peter Debye Institute for Soft Matter Physics, Molecular Nanophotonics Group, Universität Leipzig, Linnéstr. 5, 04103 Leipzig, Germany.  
+
+\* cichos@physik.uni- leipzig.de  
+
+The manipulation of nano- objects at the microscale is of great technological significance to construct new functional materials, to manipulate tiny amounts of liquids, to reconfigure sensorial systems or to detect minute concentrations of analytes in medical screening. It is commonly approached by the generation of potential energy landscapes, for example, with optical fields or by using pressure driven microfluidics. Here we show that strong hydrodynamic boundary flows enable the trapping and manipulation of nano- objects near surfaces. These thermo- osmotic flows are induced by modulating the van der Waals interaction at a solid- liquid interface with optically generated temperature fields. We use a thin gold film on a glass substrate to provide localized but reconfigurable point- like optical heating. Convergent boundary flows with velocities of tens of micrometres per second are observed and substantiated by a quantitative physical model. The hydrodynamic forces acting on suspended nanoparticles and attractive van der Waals or depletion induced forces enable precise positioning and guiding of the nanoparticles. Fast multiplexing of flow fields further provides the means for parallel manipulation of many nano- objects and the generation of complex flow fields. Our findings have direct consequences for the field of plasmonic nano- tweezers as well as other thermo- plasmonic trapping schemes and pave the way for a general scheme of nanoscopic manipulation with boundary flows.  
+
+The control and manipulation of nano- objects is a key element for future nanophotonics \(^{1 - 5}\) , material science \(^{4,6,7}\) , biotechnology \(^{2,8,9}\) or even quantum sensing \(^{10}\) . Analytes dissolved in liquids, for example, need to be delivered, concentrated, separated or locally confined for further studies to become eventually processed and removed. Photonic elements including plasmonic nano- structures require precise positioning or controlled rearrangements to serve as adaptive functional structures. Key elements of the control at the micro- and nanoscale are often either pressure driven fluidics transporting liquid volume and solutes or the generation of potential energy landscapes or force fields. The latter is achieved with optical \(^{3}\) and plasmonic tweezers \(^{11,12}\) , magnetic fields \(^{13}\) , or using electrokinetic \(^{14}\) or opto- electronic \(^{15}\) effects. Especially in the field of plasmonic tweezers and nanoantennas where light is used to excite collective electron motion in noble- metals, the Joule losses lead to the unavoidable generation of heat at boundaries as an unwanted side effect \(^{16,17}\) . Yet, such optically generated temperature fields seem also suitable for the manipulation of nano- objects in liquids, for example, for the trapping of nanoparticles \(^{18}\) and single molecules \(^{19}\) or protein aggregates \(^{20}\) as well as for manufacturing active particles \(^{21 - 24}\) . Those techniques rely on a drift of molecules and particles in optically generated temperature gradients termed thermophoresis or suggest thermoelectric effects \(^{25}\) relying on a thermally induced charge separation. In addition, thermo- electrohydrodynamic effects using time- varying electric fields have been proposed for rapid particle transport \(^{26,27}\) and convective effects that arise from temperature- induced density changes in the large liquid cells have been reported \(^{28 - 31}\) .  
+
+Here we report on a fundamental physical process that  
+
+is able to provide a versatile trapping and manipulation of nano- objects and fluids near surfaces in the simplest geometries. Contrary to most other techniques, our scheme is based on hydrodynamic flows generated by optically induced thermo- osmosis. Thermo- osmosis relies on a perturbation of the interfacial interactions at a solid- liquid boundary and is present in all experiments involving temperature gradients in plasmonic structures including plasmonic tweezers. We show that local temperature gradients on a thin gold film induce strong interfacial flows of several 10 to \(100 \mu \mathrm{m} \mathrm{s}^{- 1}\) in its direct vicinity (10 nm) that results in a flow pattern reminiscent of convection. Based on a fully quantitative analysis of our experimental results we reveal that these thermo- osmotic flows on gold- water interfaces are induced by a temperature- induced perturbation of the van der Waals (vdW) interactions. Nano- objects suspended in the liquid are therefore dragged by the hydrodynamic forces originating from these flows. Utilizing attractive vdW interactions of the nano- object with the gold surface or temperature- induced depletion, we trap and manipulate different types of nano- objects near the surface. The fast heating at small scales allows us to multiplex flow fields and to manipulate multiple objects with great precision. Our detailed analysis of the flow fields, the localization accuracy of nano- objects, and a comparison with numerical and theoretical predictions provide a quantitative understanding of these effects and paves the way for controlling boundary layer dynamics to manipulate objects at the smallest length scales in solutions.  
+
+Experimental configuration and working principle Our experiments rely on a simple sample geometry with a gold film (50 nm) that is deposited on a microscopy glass cov
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Fig. 1: Thermo-hydrodynamic manipulation of Au NPs in NaCl solution. a, The sample consists of two glass slides that confine a \(3\mu \mathrm{m}\) thin liquid film of gold nanoparticles (AuNPs) dispersed in aqueous NaCl solution. The lower glass slide carries a \(50\mathrm{nm}\) Au film that is locally heated by optical absorption of a focused laser of \(\lambda = 532\mathrm{nm}\) wavelength. b, The experimental setup comprises an inverted optical microscope equipped with an acusto-optical deflector controlled, steerable focused laser with a wavelength of \(\lambda = 532\mathrm{nm}\) . The AuNPs are observed using darkfield illumination with an oil-immersion dark-field condenser (NA 1.2) and a \(100\times\) oil-immersion objective set to NA 0.6. Images are recorded with an EMCCD camera. </center>  
+
+erslip (Fig. 1a). The sample chamber contains a suspension of gold nanoparticles (AuNPs) or other nano- objects (polystyrene NPs and ellipsoids) with a controlled amount of salt (NaCl), surfactants (SDS, ...) or polymers (PEG). The gold film is heated locally in an inverted microscopy setup by a highly focused laser (532 nm) using beam steering optics (Acusto- Optic- Deflector, AOD). The nano- objects are observed using darkfield illumination with an oil- immersion darkfield condenser (NA 1.2) and a \(100\times\) oil- immersion objective set to NA 0.6 (Fig. 1b). Additional details of the experimental setup and sample preparation are provided in the Methods section.  
+
+The trapping of nano- objects as detailed in the following is comprising two effects. i) The vertical confinement of the suspended objects as achieved by an attractive interaction of the suspended nano- objects with the gold surface, which is found to be the vdW interaction for gold nanoparticles and can be replaced by depletion forces for other materials. ii) The generation of thermo- osmotic boundary flows that are induced by the local heating and the corresponding perturbation of the liquid- solid interactions. This boundary flow is directed radially inwards to the heated spot and provides a confining hydrodynamic force on suspended objects at the heating spot.  
+
+Dynamics of AuNPs close to a Au film Consider a single AuNP with a radius of \(R = 125\mathrm{nm}\) that is suspended in an aqueous solution of NaCl at a \(10\mathrm{mM}\) concentration and diffusing in a thin liquid film of about \(3\mu \mathrm{m}\) thickness over a \(50\mathrm{nm}\) Au film (Fig. 1a). Exploring the diffusion of the particle we observe a restriction of the \(z\) - positions to a thin layer close to the gold film. The gold particle never defocuses under these conditions while it does in deionized (DI) water (Supplementary Video 1). This restricted out- of- plane motion is the result of interactions comprising an attractive vdW contribution, a repulsion of the electrostatic double layers of the particle and surface32 as described by the DLVO theory and the gravitational potential (see Supplementary Information for details).   
+
+\[V(d,c_{0}) = V_{\mathrm{E}}(d,c_{0}) + V_{\mathrm{vdW}}(d) + V_{\mathrm{G}}(d). \quad (1)\]  
+
+The total potential for the experimental situation is depicted in Fig. 2a for different salt concentrations (see SI for parameters) as a function of the surface- to- surface distance \(d = z - R\) . The stronger screening of the surface charges at the gold film and the AuNP at higher salt concentration increase the importance of attractive vdW interactions to create this secondary minimum in the DLVO part of the potential. This potential influences the observed dynamics as well as the particle couples with its hydrodynamic flow field to the solid boundary33. The in- plane \(D_{\parallel}\) (equation 2) and out- of- plane \(D_{\perp}\) diffusion coefficient (see Supplementary Information for details) are modulated with the distance \(z\) of the particle from the wall.  
+
+\[\frac{D_{\parallel}(z)}{D_0}\approx 1 - \frac{9}{16}\frac{R}{z} +\frac{1}{8}\left(\frac{R}{z}\right)^3\pm \dots := \gamma_{\parallel}^{-1}(z) \quad (2)\]  
+
+Over the course of a diffusion trajectory, the particle samples different regions with different diffusion coefficients according to its probability density \(p(d) \propto \exp (- V(d, c_0) / (k_B T))\) to be at a distance \(d\) from the surface (filled regions in Fig. 2a). The observed in- plane diffusion coefficient is thus a weighted average of the diffusion coefficient over the different vertical positions \(d\) . Using \(p(d)\) we can calculate the corresponding salt concentration dependence of the in- plane diffusion coefficient and compare that to the experimental results. Fig. 2b shows that the experimentally observed \(D_{\parallel}\) is decreasing with increasing salt concentration due to the hydrodynamic coupling in fair
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2: DLVO potential, lateral diffusion analysis, temperature distribution and thermo-osmotic flow field. a, Plot of the DLVO potential, equation (1), between a \(250~\mathrm{nm}\) Au NP and a \(50~\mathrm{nm}\) Au film on a glass surface as function of surface-surface distance \(d\) for different NaCl concentrations \(c_0\) . The shaded curves display the calculated probability density for finding the particle at this distance at the different salt concentrations (see Supplementary Information for details). The vertical dashed lines correspond to the mean distance of the particle as calculated from the probability density for a \(3\mu \mathrm{m}\) liquid film height. b, The measured diffusion coefficient \(D_1 / D_0\) parallel to the Au film with respect to the bulk diffusion coefficient \(D_0\) as function of the NaCl concentration \(c_0\) . Symbols correspond to the experimental values. The lines reflect the theoretical prediction including a distance dependent diffusion coefficient for three different Hamaker constants (dotted: \(A_H = 4\cdot 10^{-20}\mathrm{J}\) , dash-dotted: \(5\cdot 10^{-20}\mathrm{J}\) and dashed: \(6\cdot 10^{-20}\mathrm{J}\) ) of gold according to a Boltzmann weighting (see text). c, Relation between the mean distance \(\langle d\rangle\) and the NaCl concentration \(c_0\) . The symbols and the horizontal lines denote the calculated distances for measured concentrations. d, Simulation of the relative temperature increment in the \(xz\) -plane of the sample. e, Experimentally obtained temperature increment \(\Delta T_{\mathrm{max}}\) as a function of the incident laser power \(P_0\) (green data points) compared to the simulated values (green curve). f, Measured thermo-osmotic flow field in the \(xy\) -plane in close proximity to the gold film ( \(z< 500\mathrm{nm}\) ). g, Measured thermo-osmotic flow field in the \(xz\) -plane. h, Illustration of the measured flow field planes in f and g. i, j, The \(x\) - and \(z\) -component of the measured flow velocities compared to the simulation results in k and l. </center>  
+
+agreement with the theoretical predictions (for three different Hamaker constants for the AuNP gold surface interaction). Calculating in addition the mean distance \(\langle d\rangle\) of the particle from the surface reveals that only in the case of \(c_0 = 10\mathrm{mM}\) the particle is confined in the DLVO potential well, while at lower salt concentrations an enhanced probability (Fig. 2a) of finding the particle near the surface is the cause of the observed diffusion coefficient. These calculations help us to estimate the mean distance \(\langle d\rangle\) of the particle from the surface, which is about \(1.5\mu \mathrm{m}\) and \(0.9\mu \mathrm{m}\) for the lowest NaCl concentrations (Fig. 2c). At a concentration of \(c_0 = 10\mathrm{mM}\) the particle is hovering at a distance  
+
+of \(\langle d\rangle = 20\mathrm{nm}\) surface. Note that this corresponds to values of \(z / (2R)\approx 0.58\) , which is far below the commonly explored region of the hydrodynamic coupling of colloids to walls<sup>33</sup> allowing to experimentally explore new terrains also in the field of hydrodynamic wall coupling of colloids.  
+
+Hydrodynamic Particle Confinement When tightly focusing the light of \(532\mathrm{nm}\) wavelength to the gold film, a part of the incident energy (about \(30\%\) ) is absorbed and converted into heat that perturbs the liquid- solid interactions. The temperature rise at the gold surface can be determined using a thin nematic liquid crystal (5CB) film and substan
+
+<--- Page Split --->
+
+tiated by finite element simulations with the complete threedimensional temperature profile in the solution (see Fig. 2d, e and Supplementary Information for details).  
+
+These local temperature perturbations of the solidliquid interactions at the interface induce a thermo- osmotic flow34,35. Taking a liquid volume element close to the solid from the cold side and exchanging that with one at the hot side would not only transport heat since the liquid volumes have different temperatures, but also additional free energy as the liquid has a different interaction with the solid in these regions. The flow is induced in an ultrathin boundary layer corresponding in thickness to the length scale of liquid- solid interactions. Since the characteristic interaction length of liquid- solid interactions is only a few nanometers, the boundary flow on the substrate can be collapsed into a quasi- slip hydrodynamic boundary condition:  
+
+\[v_{\parallel} = -\frac{1}{\eta}\int_{0}^{\infty}z h(z)\mathrm{d}z\frac{\nabla_{\parallel}T}{T} = \chi \frac{\nabla_{\parallel}T}{T}, \quad (3)\]  
+
+where \(h(z)\) is the excess enthalpy, \(T\) the temperature and \(\nabla_{\parallel}T\) is the temperature gradient parallel to the surface. The integral can be summarized to a thermo- osmotic coefficient \(\chi\) . The thermo- osmotic coefficient \(\chi\) , therefore, contains all information about the interfacial interaction between the liquid and the solid. If \(\chi < 0\) the liquid is driven to the cold, whereas for \(\chi > 0\) , the liquid is driven to the hot. These boundary flows are present at all liquid- solid interfaces with tangential temperature gradients, though, they are commonly overlooked. They become particularly important for plasmonic and thermo- plasmonic trapping16, as those techniques rely on the dynamics of molecules and particles in the direct vicinity of plasmonic nanostructures. The boundary flow drives the flow field inside the fluid film. The resulting volumetric flow field can be tracked experimentally by single AuNPs in DI water, where the particles are not confined to a surface layer as reported above. We analyze the in- plane \((xy)\) position of the particle and its \(z\) - position, where the latter is estimated from the radius \(r_0\) of the defocussed particle images (see Supplementary Video 2 and Supplementary Information for details). The measured velocity distributions in the \(xy\) - plane near the gold layer and in the \(xz\) - plane are shown in Fig. 2f and g, respectively. The \(x\) - and \(z\) - component of the measured flow velocities are depicted in Fig. 2i, j and compare well to simulation results in Fig. 2k, l. From these measurements, we extract a thermo- osmotic coefficient on the order of \(\chi \sim 10\cdot 10^{- 10}\mathrm{m}^2\mathrm{s}^{- 1}\) (see Supplementary Information for details). We can break down the contributions to this value with equations (4) and (5) to estimate the double layer and vdW contributions using the experimental parameters. Note that AuNP do not show thermophoresis due to their high thermal conductivity and thus isothermal surface.  
+
+\[\chi_{\mathrm{E}} = \frac{\epsilon\zeta^2}{8\eta}\approx 0.8\cdot 10^{-10}\mathrm{m}^2 /\mathrm{s}^{-1}. \quad (4)\]  
+
+For the electrostatic contribution we used \(\zeta = - 30\mathrm{mV}^{36}\) and \(\epsilon = 80\epsilon_0\) (see Methods section for details). An estimate of the vdW contribution can be given by  
+
+\[\chi_{\mathrm{vdW}} = \frac{A_{\mathrm{H}}\beta T}{3\pi\eta d_0}\approx 9.3\cdot 10^{-10}\mathrm{m}^2\mathrm{s}^{-1}, \quad (5)\]  
+
+with \(\beta = 0.2\cdot 10^{- 3}\mathrm{K}^{- 1}\) being the thermal expansion coefficient of water and \(d_0 = 0.2\mathrm{nm}\) for the cut- off parameter34 (see Supplementary Information for details). The sum of both contributions \(\chi = \chi_{\mathrm{E}} + \chi_{\mathrm{vdW}} = 10.1\cdot 10^{- 10}\mathrm{m}^2\mathrm{s}^{- 1}\) matches well the experimental result and suggests that thermo- osmosis at gold- water interfaces is governed by vdW interactions. The obtained quasi slip velocities are ranging up to \(80\mu \mathrm{m / s}\) and provide, due to their omnipresence, a unique tool for nanofluidics. These thermo- osmotic flows are induced without any external pressure difference. They can be controlled by the light intensity heating laser and are quickly switched due to the extremely fast heat conduction at these length scales. Moreover the finding of the vdW dominated thermo- osmotic flows suggest that such contributions must be present in any plasmonic trapping experiment with extended gold structures12,16,17,27.  
+
+Using \(F_T^{\mathrm{eff}} = 6\pi \eta R\gamma_{\parallel}v_x\) and \(F_T^{\mathrm{eff}} = 6\pi \eta R\gamma_{\perp}v_z\) , where \(\gamma_{\parallel}\) and \(\gamma_{\perp}\) are the correction factors for the friction coefficient of a sphere close to a surface we are able to extract the hydrodynamic forces that are exerted on the AuNP tracers (see equation (2) and Supplementary Information for details). The lateral forces allow to confine objects at the heating spot, yet the hydrodynamic force normal to the surface ( \(z\) - direction) is repulsive without any additional interaction. Finally, such boundary flows with substantial vertical velocity gradients also exhibit a vorticity (see Supplementary Information for details) that generates a torque on suspended objects causing them to rotate37.  
+
+Single particle trapping and flow field multiplexing The repulsive normal component caused by the hydrodynamic drag is now superimposed with the attractive force due to the DLVO potential when increasing the NaCl concentration. The surface- to- surface distance between AuNP and gold film and the depth of the appearing secondary DLVO potential minimum can be controlled by the NaCl concentration. At a NaCl concentration of about \(c_0 = 10\mathrm{mM}\) , the attractive potential has a depth of about \(10k_{\mathrm{B}}T\) (see Fig. 2a) and is strong enough to compete with the vertical drag force and additional optical forces on the AuNP to trap the particle above the heating spot.  
+
+Supplementary Video 3 demonstrates this trapping of an AuNP above the hot spot on the Au surface. This is purely the result of the hydrodynamic drag forces generated by the thermo- osmotic flow and the attractive vdW interaction between the AuNP and the Au film. This observation is substantiated by a quantitative evaluation of the lateral trap stiffness and vertical forces, as depicted in Fig. 3a, b. The fluctuations of the particle in the hydrodynamic flow arise from a balance of the restoring hydrodynamic currents and the diffusive currents. Analysis of the lateral position histograms (inset in Fig. 3c for \(1.25\mathrm{mW}\) ) yields an effective stiffness of the trap (Fig. 3a) that well matches the predictions based on the thermo- osmotic flow (Fig. 2i, j). The hydrodynamic trapping stiffness increases linearly up to a heating power of about \(1.8\mathrm{mW}\) . At this power, the vertical forces become strong enough to let the particle escape the secondary DLVO minimum, which is visible from the \(z\) - position time traces displayed in Fig. 3d. The AuNPs are then observed to move vertically out of the DLVO potential to follow the flow inside the sample and to eventually return
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Fig. 3: Forces on trapped NPs in \(10\mathrm{mM}\) NaCl. a, The lateral trap stiffness obtained from the experimental position histograms (blue data points) as function the laser power \(P_0\) compared to the simulation result (blue solid line). b, The \(x\) -component of the thermo-osmotic drag force \(F_{\mathrm{T}}^{\mathrm{TF}}\) (blue line), the optical force \(F_{\mathrm{Z}}^{\mathrm{OF}}\) (green line) and the total force \(F_{\mathrm{Z}}^{\mathrm{OF}} + F_{\mathrm{Z}}^{\mathrm{TF}}\) (black dashed line) as function the incident laser power \(P_0\) for a NP located at \(x = 0\) , \(d = 30\) ( \(z = d + R\) ). The attractive DLVO force \(F_{\mathrm{DLVO}}^{\mathrm{DLVO}}\) is independent of the incident laser power and depicted as horizontal, red line. c, Trajectory of a AuNP for a heating laser power of \(1.25\mathrm{mW}\) (see Supplementary Video 2 for details). The inset shows the corresponding lateral distribution histogram. d, Time traces of the \(z\) -position for three different laser powers. e, Trajectory of a AuNP for a heating laser power of \(2.5\mathrm{mW}\) , which is above the threshold power of \(2.25\mathrm{mW}\) . </center>  
+
+to the boundary flow via sedimentation (Fig. 3e, Supplementary Video 4). The forces which eject the particle from the potential comprise the hydrodynamic and the optical forces due to the radiation pressure from the heating laser leaked through the film. We have evaluated the individual contributions in simulations. They are shown together with the hydrodynamic force and the total vertical force as compared to the attractive force of the DLVO potential (Fig. 3b) and provide quantitative agreement (threshold heating power of \(2.25\mathrm{mW}\) ) with our experimental results. Note that while the stationary distribution of particles in the vertical direction is not influenced by the diffusive dynamics, the escape rate from the potential well is heavily altered by the fact that the vertical diffusion coefficient \(D_{\perp}\) of the particle is decreasing to zero when approaching the gold film. This is enhancing the trapping times considerably (see Supplementary Information for details) but also increases the time required for the particle to enter the DLVO minimum by diffusion.  
+
+The observed trapping is, hence, a vdW assisted thermodynamic process. Vertical confinement is achieved by vdW attraction and double layer repulsion, while lateral confinement is the result of thermo- osmotic flows induced in an ultra- thin sheet of liquid at the interface. No additional contributions, for example, due to convective flows with similar flow patterns (see Supplementary Information for details) or thermo- electric effects are required for a quantitative description \(^{25,31,38,39}\) . Precise tuning of the DLVO potential enables the trapping of even smaller Au NPs (Fig. 4a, Supplementary Video 5).  
+
+The speed of heat diffusion, which is about 4 orders of magnitude faster than the particle diffusion \(^{40}\) allows us to introduce a flow field multiplexing. We switch the heating location between different positions inducing thermo- osmotic flow fields for time periods of about \(100\mu \mathrm{s}\) . With the help of this multiplexing, we are able to hold multiple \(R = 125\mathrm{nm}\) AuNPs (Fig. 4b, c) at distances of less than \(1\mu \mathrm{m}\) , which would not be possible with continuous heating of close- by locations (Supplementary Videos 6 and 7). A trapped AuNP can also be guided along the predefined path over the Au film as fast as \(10\mu \mathrm{m}\mathrm{s}^{- 1}\) (Fig. 4d,  
+
+Supplementary Video 8). At larger manipulation speeds ( \(f > 100\mathrm{Hz}\) ) and higher heating power ( \(P_0 > 10\mathrm{mW}\) ) the thermo- osmotic attraction to the heating spot is combined with thermo- viscous flows \(^{41,42}\) . These flows originate from the temperature dependent viscosity \(\eta (T)\) of the liquid and are directed opposite to the scanning direction of the laser \(^{42}\) . The result of this combination of thermo- osmosis and thermo- viscous flows is a rotating ring- like particle structure (Fig. 4e and Supplementary Video 9).  
+
+These different effects that can be exploited in a simple planar geometry give rise to numerous applications including for example a freely configurable nanoparticle on mirror geometry for plasmonic sensing \(^{5,43}\) . The multiplexing of local flow field may be be helpful to construct more complex effective flow fields for an efficient transport of analytes without external pressure.  
+
+Beyond thermo- osmotic van der Waals trapping So far, the presented manipulation is based on thermo- osmotic flows that drive the lateral motion of suspended colloids and a vertical confinement due to the secondary minimum of the DLVO potential between AuNP and Au film. While the thermo- osmotic flows are characteristic for all systems containing a heated gold/water interface including all previous studies on thermo- plasmonic trapping, the DLVO potential minimum is much weaker for other materials like polymer colloids or macromolecules due to their smaller vdW attraction. Often, those system even show a repulsion from the heat source due to thermophoresis, which is not present for AuNP. A more generalized strategy therefore needs additional attractive contributions, which confine suspended colloids or molecules to regions close to the gold surface to take advantage of the thermo- osmotic flow.  
+
+Such attractive contributions can arise from depletion interactions \(^{21,44}\) . Thereby a temperature gradient repels dissolved molecules from the heated regions generating a concentration gradient that drives suspended nano- objects to the heating spot. To demonstrate this effect we use the surfactant sodium dodecyl sulfate (SDS) at a concentration of \(5\mathrm{mM}\) well below the critical micelle concentration
+
+<--- Page Split --->
+
+(8.2 mM) to avoid complications of micelle formation. We suspend additional polystyrene particles(PS) and AuNPs of the same size ( \(R = 125 \mathrm{nm}\) ) in the solution and compare their dynamics to a solution with AuNPs and PS particles without SDS but \(10 \mathrm{mM} \mathrm{NaCl}\) . Remarkably, the heated spot is attractive for both AuNPs and for PS NPs (Fig. 4f, Supplementary Video 10) in the SDS solution showing even PS colloidal crystal growth, while only the AuNP is trapped in the NaCl solution and the PS particles are repelled by thermophoresis (Fig. 4g, Supplementary Video 11).  
+
+The observations in NaCl are readily explained by the fact that the AuNP is confined in the DLVO minimum as demonstrated above but the PS particle is not due to a 10 times lower Hamaker constant (see Supplementary Information for details). The PS particle samples the whole liquid film thickness equally and not preferentially the region close to the Au film and experiences an additional thermophoretic drift velocity given by  
+
+\[\pmb {u} = -\frac{2}{3}\chi \frac{\nabla T}{T} = -D_{\mathrm{T}}\nabla T, \quad (6)\]  
+
+where \(D_{\mathrm{T}}\) is the thermophoretic mobility \(^{34}\) and \(\nabla T\) the temperature gradient (Fig. 4g, Supplementary Video 11). For \(\chi >0\) the particle is driven to the cold. From equation (4) we find \(\chi \approx \chi_{\mathrm{E}} = 1.28\cdot 10^{- 10} \mathrm{m}^{2} \mathrm{s}^{- 1}\) and \(D_{\mathrm{T}} \approx 0.3 \mu \mathrm{m}^{2} \mathrm{K}^{- 1} \mathrm{s}^{- 1}\) , where we have used a measured zeta potential of \(\zeta \approx - 38 \mathrm{mV}\) . The vdW contribution, \(\chi_{\mathrm{vdW}}\) to  
+
+either the thermophoretic drift or the attraction to the gold surface can be neglected due to the smaller Hamaker constant of PS. From the stationary probability distribution of the PS NP we find a Soret coefficient of \(S_{\mathrm{T}} \approx 0.24 \mathrm{K}^{- 1}\) (see Supplementary Information for details) in agreement with our theoretical prediction \(S_{\mathrm{T}} = D_{\mathrm{T}} / D_{\parallel} \approx 0.21 \mathrm{K}^{- 1}\) .  
+
+In the SDS solution, the additional surfactant molecules now undergo thermophoresis to yield a concentration gradient in which suspended colloidal particles drift. The lower concentration in the heated regions promotes an effective attractive interaction of suspended colloids with the gold surface due to depletion forces. The drift velocity is described by an additional term to the thermodiffusion coefficient \(D_{\mathrm{T}}\) , that is, the second term in brackets in equation (7) \(^{21,34,44}\) .  
+
+\[\pmb {u} = -\left(D_{\mathrm{T}} - \frac{k_{\mathrm{B}}}{3\eta} R^{2}c_{0}N_{\mathrm{A}}\left(T S_{\mathrm{DS}}^{\mathrm{SDS}} - 1\right)\right)\nabla T \quad (7)\]  
+
+Here \(R\) is the size of the SDS molecule, \(c\) the concentration in units of mol/l and \(S_{\mathrm{DS}}^{\mathrm{SDS}}\) the Soret coefficient of SDS. For \(R = 2 \mathrm{nm}^{45}\) , \(c_{0} = 5 \mathrm{mM}\) and \(S_{\mathrm{DS}}^{\mathrm{SDS}} = 0.03 \mathrm{K}^{- 146}\) we find \(- 0.43 \mu \mathrm{m}^{2} \mathrm{K}^{- 1} \mathrm{s}^{- 1}\) for the additional depletion contribution, which exceeds the thermophoretic mobility, \(D_{\mathrm{T}} \approx 0.3 \mu \mathrm{m}^{2} \mathrm{K}^{- 1} \mathrm{s}^{- 1}\) , rendering the overall mobility negative. The PS NPs and the AuNPs are thus driven to the the heated Au film surface (Fig. 4f, Supplementary Video 10) which allows for further transport in the thermo- osmotic  
+
+![](images/Figure_4.jpg)
+
+<center>Fig. 4: Manipulation of NPs over a Au film in NaCl and SDS solution. a, A AuNP with \(50 \mathrm{nm}\) radius trapped at a NaCl of \(30 \mathrm{mM}\) (Supplementary Video 5). b, Manipulation of two AuNPs by a multiplexed laser beam (Supplementary Video 6). c, Control of three AuNPs (Supplementary Video 7). d, Actuation of a single AuNP on a circular trajectory by a steerable laser beam (Supplementary Video 8). The green and white, dashed arrows denote the moving direction of the laser focus and particle, respectively. e, Generation of thermo-viscous flows by rotating the laser focus on a circle with a rotation frequency of \(f = 500 \mathrm{Hz}\) at high laser powers (Supplementary Video 9). Note, that laser movement (green arrow) and the thermo-viscous flow (white, dashed arrow) and have opposite directions. f, Attraction of a AuNP and PS NPs in \(5 \mathrm{mM}\) SDS due to depletion (Supplementary Video 10). g, An AuNP (125 nm radius) trapped in an ensemble of polystyrene (PS) NPs of the same size at \(10 \mathrm{mM}\) NaCl (Supplementary Video 11), where the PS-particles are repelled due to thermophoresis. h, Attraction of PS ellipsoids (2.39 \(\mu \mathrm{m}\) major-axis length, \(0.34 \mu \mathrm{m}\) minor-axis length) in \(5 \mathrm{mM}\) SDS (Supplementary Video 12). </center>
+
+<--- Page Split --->
+
+boundary flow. Additional contributions, as for example thermo- electric fields may even enhance the attractive components. Overall, this concept is readily transferred to other objects as shown in Figure 4h and Supplementary Video 12, where we have trapped ellipsoidal PS particles in a \(5\mathrm{mM}\) solution of SDS. Note that as compared to other schemes, our approach always includes thermo- osmotic boundary flows.  
+
+Conclusion In conclusion, we have demonstrated that thermo- hydrodynamic boundary flows can manipulate nano- objects with unprecedented flexibility in a very simple sample geometry. These flows are the key for future thermo- optofluidic implementations with an extensive range of applications in the fields of i) nanoparticle sorting and separation; ii) assembly of nanophotonic circuits and plasmonic quantum sensors; iii) biotechnology on- chip laboratories and iv) manufacturing of nanomaterials and functional nanosurfaces. We have substantiated our experimental findings of thermo- osmotic flow assisted trapping with a quantitative theoretical description. A flow field multiplexing scheme has been further developed to allow for the simultaneous manipulation of many individual nano- objects and the generation of complex effective flow patterns. Our concept can be combined with other thermally induced effects such as thermophoresis, depletion forces and thermoviscous flows to form a fully- featured nanofluidic system- on- a- chip. Besides direct consequences for the field of plasmonic nano- tweezers and other thermoplasmonic trapping schemes, the use of thermo- hydrodynamic flows as a tool for nanofluidic applications will extend the limits at the forefront of nanotechnology and help to develop AI and feedback controlled schemes for the material synthesis.  
+
+## References  
+
+[1] O. M. Maragò, P. H. Jones, P. G. Gucciardi, G. Volpe, and A. C. Ferrari, Optical Trapping and Manipulation of Nanostructures, Nat. Nanotechnol. 8, 807- 819 (2013). [2] D. Gao, W. Ding, M. Nieto- Vesperinas, X. Ding, M. Rahman, T. Zhang, C. Lim, and C.- W. Qiu, Optical Manipulation from the Microscale to the Nanoscale: Fundamentals, Advances and Prospects, Light Sci. Appl. 6, 17039 (2017). [3] C. Bradac, Nanoscale Optical Trapping: A Review, Adv. Opt. Mat. 6, 1800005 (2018). [4] J. Xavier, S. Vincent, F. Meder, and F. Vollmer, Advances in Optoplasmonic Sensors - Combining Optical Nano/Microcavities and Photonic Crystals with Plasmonic Nanostructures and Nanoparticles, Nanophotonics 7, 1- 38 (2018). [5] G.- C. Li, Q. Zhang, S. A. Maier, and D. Lei, Plasmonic Particle- on- Film Nanocavities: A Versatile Platform for Plasmon- Enhanced Spectroscopy and Photochemistry, Nanophotonics 7, 1865- 1889 (2018). [6] L. Lin, X. Peng, and Y. Zheng, Reconfigurable Opto- Thermoelectric Printing of Colloidal Particles, Chem. Commun. 53, 7357- 7360 (2017). [7] Y. Xie, J. Rufo, R. Zhong, J. Rich, P. Li, K. W. Leong, and T. J. Huang, Microfluidic Isolation and Enrichment of Nanoparticles, ACS Nano 14, 16220- 16240 (2020). [8] I. A. Favre- Bulle, A. B. Stilgoe, E. K. Scott, and H. Rubinsztein- Dunlop, Optical Trapping in Vivo: Theory, Practice, and Applications, Nanophotonics 8, 1023- 1040 (2019).  
+
+[9] D. Choudhary, A. Mossa, M. Jadhav, and C. Cecconi, Bio- Molecular Applications of Recent Developments in Optical Tweezers, Biomolecules 9, 23 (2019). [10] M. Kayci, H.- C. Chang, and A. Radenovic, Electron Spin Resonance of Nitrogen- Vacancy Defects Embedded in Single Nanodiamonds in an ABEL Trap, Nano Lett. 14, 5335- 5341 (2014). [11] M. L. Juan, M. Righini, and R. Quidant, Plasmon Nano- Optical Tweezers, Nat. Photonics 5, 349- 356 (2011). [12] Y. Zhang, C. Min, X. Dou, X. Wang, H. P. Urbach, M. G. Somekh, and X. Yuan, Plasmonic Tweezers: For Nanoscale Optical Trapping and Beyond, Light Sci. Appl. 10, 59 (2021). [13] I. De Vlaminck and C. Dekker, Recent Advances in Magnetic Tweezers, Annu. Rev. Biophys. 41, 453- 472 (2012). [14] Q. Wang and W. E. Moerner, An Adaptive Anti- Brownian Electrokinetic Trap with Real- Time Information on Single- Molecule Diffusivity and Mobility, ACS Nano 5, 5792- 5799 (2011). [15] M. C. Wu, Optoelectronic Tweezers, Nat. Photonics 5, 322- 324 (2011). [16] A. Kotnala and R. Gordon, Quantification of High- Efficiency Trapping of Nanoparticles in a Double Nanohole Optical Tweezers, Nano Lett. 14, 853- 856 (2014). [17] Q. Jiang, B. Rogez, J.- B. Claude, G. Baffou, and J. Wenger, Quantifying the Role of the Surfactant and the Thermophoretic Force in Plasmonic Nano- Optical Trapping, Nano Lett. 20, 8811- 8817 (2020). [18] M. Braun and F. Cichos, Optically Controlled Thermophoretic Trapping of Single Nano- Objects, ACS Nano 7, 11200- 11208 (2013). [19] M. Braun, A. P. Bregulla, K. Günther, M. Mertig, and F. Cichos, Single Molecules Trapped by Dynamic Inhomogeneous Temperature Fields, Nano Lett. 15, 5499- 5505 (2015). [20] M. Fränzl, T. Thalheim, J. Adler, D. Huster, J. Possekardt, M. Mertig, and F. Cichos, Thermophoretic Trap for Single Amyloid Fibril and Protein Aggregation Studies, Nat. Methods 16, 611- 614 (2019). [21] H.- R. Jiang, H. Wada, N. Yoshinaga, and M. Sano, Manipulation of Colloids by a Nonequilibrium Depletion Force in a Temperature Gradient, Phys. Rev. Lett. 102, 208301 (2009). [22] A. P. Bregulla, H. Yang, and F. Cichos, Stochastic Localization of Microswimmers by Photon Nudging, ACS Nano 8, 6542- 6550 (2014). [23] U. Khadka, V. Holubec, H. Yang, and F. Cichos, Active Particles Bound by Information Flows, Nat. Commun. 9, 3864 (2018). [24] M. Fränzl, S. Muños- Landin, V. Holubec, and F. Cichos, Fully Steerable Symmetric Thermoplasmonic Microswimmers, ACS Nano 15, 3434- 3440 (2021). [25] L. Lin, M. Wang, X. Peng, E. N. Lissek, Z. Mao, L. Scarabelli, E. Adkins, S. Coskun, H. E. Unalan, B. A. Korgel et al., Opto- Thermoelectric Nanotweezers, Nat. Photonics 12, 195- 201 (2018). [26] J. C. Ndukaife, A. V. Kildishev, A. G. A. Nnanna, V. M. Shalaev, S. T. Wereley, and A. Boltasseva, Long- Range and Rapid Transport of Individual Nano- Objects by a Hybrid Electrorheromoplasmonic Nanotweezer, Nat. Nanotechnol. 11, 53- 59 (2016). [27] C. Hong, S. Yang, and J. C. Ndukaife, Stand- off Trapping and Manipulation of Sub- 10 Nm Objects and Biomolecules Using Opto- Thermo- Electrohydrodynamic Tweezers, Nat. Nanotechnol. 15, 908- 913 (2020).
+
+<--- Page Split --->
+
+[28] J. S. Donner, G. Baffou, D. McCloskey, and R. Quidant, Plasmon- Assisted Optofluidics, ACS Nano 5, 5457- 5462 (2011).[29] B. J. Roxworthy, A. M. Bhuiya, S. P. Vanka, and K. C. Toussaint, Understanding and Controlling Plasmon- Induced Convection, Nat. Commun. 5, 3173 (2014).[30] J. Chen, J. F.- C. Loo, D. Wang, Y. Zhang, S.- K. Kong, and H.- P. Ho, Thermal Optofluidics: Principles and Applications, Adv. Opt. Mater. 8, 1900829 (2020).[31] B. Ciraulo, J. Garcia- Guirado, I. de Miguel, J. Ortega Arroyo, and R. Quidant, Long- Range Optofluidic Control with Plasmon Heating, Nat. Commun. 12, 2001 (2021).[32] J. N. Israelachvili, Intermolecular and Surface Forces, 3rd Edition (Elsevier, Academic Press, Amsterdam, 2011).[33] B. Lin, J. Yu, and S. A. Rice, Direct Measurements of Constrained Brownian Motion of an Isolated Sphere between Two Walls, Phys. Rev. E 62, 3909- 3919 (2000).[34] A. Wurger, Thermal Non- Equilibrium Transport in Colloids, Rep. Prog. Phys. 73, 126601 (2010).[35] A. P. Bregulla, A. Wurger, K. Gunther, M. Mertig, and F. Cichos, Thermo- Osmotic Flow in Thin Films, Phys. Rev. Lett. 116, 188303 (2016).[36] M. Giesbers, J. Kleijn, and M. A. Cohen Stuart, The Electrical Double Layer on Gold Probed by Electrokinetic and Surface Force Measurements, J. Colloid Interface Sci. 248, 88- 95 (2002).[37] J. J. Bluemink, D. Lohse, A. Prosperetti, and L. Van Wijngaarden, A Sphere in a Uniformly Rotating or Shearing Flow, J. Fluid Mech. 600, 201- 233 (2008).[38] L. Lin, X. Peng, M. Wang, L. Scarabelli, Z. Mao, L. M. Liz- Marzán, M. F. Becker, and Y. Zheng, Light- Directed Reversible Assembly of Plasmonic Nanoparticles Using Plasmon- Enhanced Thermophoresis, ACS Nano 10, 9659- 9668 (2016).[39] L. Lin, J. Zhang, X. Peng, Z. Wu, A. C. H. Coughlan, Z. Mao, M. A. Bevan, and Y. Zheng, Opto- Thermophoretic Assembly of Colloidal Matter, Sci. Adv. 3, 1700458 (2017).[40] G. Baffou, F. Cichos, and R. Quidant, Applications and Challenges of Thermoplasmonics, Nat. Mater. 19, 946- 958 (2020).[41] F. M. Weinert, J. A. Kraus, T. Franosch, and D. Braun, Microscale Fluid Flow Induced by Thermoviscous Expansion Along a Traveling Wave, Phys. Rev. Lett. 100, 164501 (2008).[42] F. M. Weinert, C. B. Mast, and D. Braun, Optical Fluid and Biomolecule Transport with Thermal Fields, Phys. Chem. Chem. Phys. 13, 9918 (2011).[43] R. Chikkaraddy, B. de Nijs, F. Benz, S. J. Barrow, O. A. Scherman, E. Rosta, A. Demetriadou, P. Fox, O. Hess, and J. J. Baumberg, Single- Molecule Strong Coupling at Room Temperature in Plasmonic Nanocavities, Nature 535, 127- 130 (2016).[44] Y. T. Maeda, A. Buguin, and A. Libchaber, Thermal Separation: Interplay between the Soret Effect and Entropic Force Gradient, Phys. Rev. Lett. 107, 038301 (2011).[45] O. Syshchyk, D. Afanasenkau, Z. Wang, H. Kriegs, J. Buitenhuis, and S. Wiegand, Influence of Temperature and Charge Effects on Thermophoresis of Polystyrene Beads, Eur. Phys. J. E 39, 129 (2016).[46] D. Vigolo, S. Buzzaccaro, and R. Piazza, Thermophoresis and Thermoelectricity in Surfactant Solutions, Langmuir 26, 7792- 7801 (2010).[47] Q. Zhang, H. Yu, M. Barbiero, B. Wang, and M. Gu, Artificial Neural Networks Enabled by Nanophotonics, Light Sci Appl 8, 42 (2019).  
+
+[48] J. Xavier, D. Yu, C. Jones, E. Zossimova, and F. Vollmer, Quantum Nanophotonic and Nanoplasmonic Sensing: Towards Quantum Optical Bioscience Laboratories on Chip, Nanophotonics 10, 1387- 1435 (2021).[49] J. Peng, H.- H. Jeong, Q. Lin, S. Cormier, H.- L. Liang, M. F. L. De Volder, S. Vignolini, and J. J. Baumberg, Scalable Electrochromic Nanopixels Using Plasmonics, Sci. Adv. 5, 10.1126/sciadv.aaw2205 (2019).[50] E. Mohammadi, A. Tittl, K. L. Tsakmakidis, T. V. Raziman, and A. G. Curto, Dual Nanoresonators for Ultrasonic Optical Detection, ACS Photonics 8, 1754- 1762 (2021).  
+
+## Methods  
+
+Experimental setup The experimental setup (Figure S19) consists of an inverted microscope (Olympus, IX71) with a mounted piezo translation stage (Physik Instrumente, P- 733.3). The microparticles are heated by a focused, continuous- wave laser at a wavelength of \(532 \mathrm{nm}\) (CNI, MGL- III- 532). The beam diameter is increased by a beam expander and sent to an acousto- optic deflector (AA Opto- Electronic, DTSXY- 400- 532) and a lens system to steer the laser focus in the sample plane. The deflected beam is focused by an oil- immersion objective (Olympus, UPlanApo \(\times 100 / 1.35\) , Oil, Iris, NA 0.5 - 1.35) to the sample plane \((w_0 \approx 0.8 \mu \mathrm{m}\) beam waist in the sample plane). The sample is illuminated with an oil- immersion darkfield condenser (Olympus, U- DCW, NA 1.2 - 1.4) and a white- light LED (Thorlabs, SOLIS- 3C). The scattered light is imaged by the objective and a tube lens ( \(250 \mathrm{mm}\) ) to an EMCCD (electron- multiplying charge- coupled device) camera (Andor, iXon DV885LC). The variable numerical aperture of the objective was set to a value below the minimum aperture of the darkfield condenser. The dichroic beam splitter (D) was selected to reflect the laser wavelength (Omega Optical, 560DRLP) and a notch filter (F) is used to block any remaining back reflections from the laser (Thorlabs, NF533- 17). The acousto- optic deflector (AOD), as well as the piezo stage, are driven by an AD/DA (analogue- digital/digital- analogue) converter (Jäger Messtechnik, ADwin- Gold II). A LabVIEW program running on a desktop PC (Intel Core i7 2600 4 \(\times 3.40 \mathrm{GHz}\) CPU) is used to record and process the images as well as to control the AOD feedback via the AD/DA converter.  
+
+Sample preparation The sample consists of two glass coverslips \((22 \mathrm{mm} \times 22 \mathrm{mm})\) confining a thin liquid film. First, the coverslips were thoroughly cleaned by rinsing successively with acetone, isopropyl and Milli- Q water and dried with a nitrogen gun. Subsequently, the edges of one coverslip were covered with a thin layer of PDMS (polydimethylsiloxane) for sealing. The particle solution used for the experiments was prepared by dispersing \(0.250 \mu \mathrm{m}\) diameter gold particles (Cytodiagnostics) in ... solution. Finally, \(0.5 \mu \mathrm{l}\) of the mixed particle suspension is pipetted in the middle of one of the coverslips and the other is placed on top. Depending on the area covered by the liquid, typically about \((10 \mathrm{mm} \times 10 \mathrm{mm})\) , the resulting liquid film height is about \(5 \mu \mathrm{m}\) .  
+
+## Acknowledgement  
+
+The authors acknowledge financial support by ...  
+
+## Author contributions  
+
+M.F. and F.C. designed the experiments. M.F. performed the experiments. M.F. and F.C. analyzed the experimental data. M.F. and F.C. implemented and evaluated the numerical calculations. M.F. and F.C. wrote the manuscript. All authors discussed the results and commented on the manuscript.
+
+<--- Page Split --->
+
+## Competing interests  
+
+The authors declare no competing interest.  
+
+## Additional information  
+
+Supplementary information is available for this paper at https://doi.org ...  
+
+Correspondence and requests for materials should be addressed to F.C.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- Video6.mp4- Sl.pdf- Video3.mp4- Video2.mp4- Video8.mp4- Video7.mp4- Video1.mp4- Video5.mp4- Video4.mp4
+
+<--- Page Split --->

preprint/preprint__09b108cd9ad0034506ae629f2fb23d3bf7f4cd8c58a23feee09b7b89729c5397/preprint__09b108cd9ad0034506ae629f2fb23d3bf7f4cd8c58a23feee09b7b89729c5397_det.mmd

@@ -0,0 +1,252 @@
+<|ref|>title<|/ref|><|det|>[[44, 106, 872, 175]]<|/det|>
+# Hydrodynamic manipulation of nano-objects by optically induced thermo-osmotic flows  
+
+<|ref|>text<|/ref|><|det|>[[44, 195, 570, 283]]<|/det|>
+Martin Fränzl  Leipzig University https://orcid.org/0000- 0001- 6754- 8554  Frank Cichos (cichos@physik.uni- leipzig.de)  Leipzig University https://orcid.org/0000- 0002- 9803- 4975  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 325, 102, 342]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 362, 673, 382]]<|/det|>
+Keywords: nano- objects, microfluidics, hydrodynamics, thermo- osti  
+
+<|ref|>text<|/ref|><|det|>[[44, 400, 350, 419]]<|/det|>
+Posted Date: September 24th, 2021  
+
+<|ref|>text<|/ref|><|det|>[[44, 438, 463, 458]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 879955/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 476, 910, 519]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License.  Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[44, 554, 935, 597]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on February 3rd, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 28212- z.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[81, 95, 828, 136]]<|/det|>
+# Hydrodynamic manipulation of nano-objects by optically induced thermo-osmotic flows  
+
+<|ref|>text<|/ref|><|det|>[[82, 154, 348, 169]]<|/det|>
+Martin Fränzl \(^{1}\) and Frank Cichos \(^{1,*}\)  
+
+<|ref|>text<|/ref|><|det|>[[81, 187, 770, 214]]<|/det|>
+\(^{1}\) Peter Debye Institute for Soft Matter Physics, Molecular Nanophotonics Group, Universität Leipzig, Linnéstr. 5, 04103 Leipzig, Germany.  
+
+<|ref|>text<|/ref|><|det|>[[82, 223, 280, 235]]<|/det|>
+\* cichos@physik.uni- leipzig.de  
+
+<|ref|>text<|/ref|><|det|>[[80, 244, 916, 413]]<|/det|>
+The manipulation of nano- objects at the microscale is of great technological significance to construct new functional materials, to manipulate tiny amounts of liquids, to reconfigure sensorial systems or to detect minute concentrations of analytes in medical screening. It is commonly approached by the generation of potential energy landscapes, for example, with optical fields or by using pressure driven microfluidics. Here we show that strong hydrodynamic boundary flows enable the trapping and manipulation of nano- objects near surfaces. These thermo- osmotic flows are induced by modulating the van der Waals interaction at a solid- liquid interface with optically generated temperature fields. We use a thin gold film on a glass substrate to provide localized but reconfigurable point- like optical heating. Convergent boundary flows with velocities of tens of micrometres per second are observed and substantiated by a quantitative physical model. The hydrodynamic forces acting on suspended nanoparticles and attractive van der Waals or depletion induced forces enable precise positioning and guiding of the nanoparticles. Fast multiplexing of flow fields further provides the means for parallel manipulation of many nano- objects and the generation of complex flow fields. Our findings have direct consequences for the field of plasmonic nano- tweezers as well as other thermo- plasmonic trapping schemes and pave the way for a general scheme of nanoscopic manipulation with boundary flows.  
+
+<|ref|>text<|/ref|><|det|>[[80, 446, 481, 860]]<|/det|>
+The control and manipulation of nano- objects is a key element for future nanophotonics \(^{1 - 5}\) , material science \(^{4,6,7}\) , biotechnology \(^{2,8,9}\) or even quantum sensing \(^{10}\) . Analytes dissolved in liquids, for example, need to be delivered, concentrated, separated or locally confined for further studies to become eventually processed and removed. Photonic elements including plasmonic nano- structures require precise positioning or controlled rearrangements to serve as adaptive functional structures. Key elements of the control at the micro- and nanoscale are often either pressure driven fluidics transporting liquid volume and solutes or the generation of potential energy landscapes or force fields. The latter is achieved with optical \(^{3}\) and plasmonic tweezers \(^{11,12}\) , magnetic fields \(^{13}\) , or using electrokinetic \(^{14}\) or opto- electronic \(^{15}\) effects. Especially in the field of plasmonic tweezers and nanoantennas where light is used to excite collective electron motion in noble- metals, the Joule losses lead to the unavoidable generation of heat at boundaries as an unwanted side effect \(^{16,17}\) . Yet, such optically generated temperature fields seem also suitable for the manipulation of nano- objects in liquids, for example, for the trapping of nanoparticles \(^{18}\) and single molecules \(^{19}\) or protein aggregates \(^{20}\) as well as for manufacturing active particles \(^{21 - 24}\) . Those techniques rely on a drift of molecules and particles in optically generated temperature gradients termed thermophoresis or suggest thermoelectric effects \(^{25}\) relying on a thermally induced charge separation. In addition, thermo- electrohydrodynamic effects using time- varying electric fields have been proposed for rapid particle transport \(^{26,27}\) and convective effects that arise from temperature- induced density changes in the large liquid cells have been reported \(^{28 - 31}\) .  
+
+<|ref|>text<|/ref|><|det|>[[97, 869, 480, 881]]<|/det|>
+Here we report on a fundamental physical process that  
+
+<|ref|>text<|/ref|><|det|>[[515, 446, 916, 821]]<|/det|>
+is able to provide a versatile trapping and manipulation of nano- objects and fluids near surfaces in the simplest geometries. Contrary to most other techniques, our scheme is based on hydrodynamic flows generated by optically induced thermo- osmosis. Thermo- osmosis relies on a perturbation of the interfacial interactions at a solid- liquid boundary and is present in all experiments involving temperature gradients in plasmonic structures including plasmonic tweezers. We show that local temperature gradients on a thin gold film induce strong interfacial flows of several 10 to \(100 \mu \mathrm{m} \mathrm{s}^{- 1}\) in its direct vicinity (10 nm) that results in a flow pattern reminiscent of convection. Based on a fully quantitative analysis of our experimental results we reveal that these thermo- osmotic flows on gold- water interfaces are induced by a temperature- induced perturbation of the van der Waals (vdW) interactions. Nano- objects suspended in the liquid are therefore dragged by the hydrodynamic forces originating from these flows. Utilizing attractive vdW interactions of the nano- object with the gold surface or temperature- induced depletion, we trap and manipulate different types of nano- objects near the surface. The fast heating at small scales allows us to multiplex flow fields and to manipulate multiple objects with great precision. Our detailed analysis of the flow fields, the localization accuracy of nano- objects, and a comparison with numerical and theoretical predictions provide a quantitative understanding of these effects and paves the way for controlling boundary layer dynamics to manipulate objects at the smallest length scales in solutions.  
+
+<|ref|>text<|/ref|><|det|>[[515, 842, 915, 881]]<|/det|>
+Experimental configuration and working principle Our experiments rely on a simple sample geometry with a gold film (50 nm) that is deposited on a microscopy glass cov
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[80, 81, 900, 333]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[80, 345, 917, 395]]<|/det|>
+<center>Fig. 1: Thermo-hydrodynamic manipulation of Au NPs in NaCl solution. a, The sample consists of two glass slides that confine a \(3\mu \mathrm{m}\) thin liquid film of gold nanoparticles (AuNPs) dispersed in aqueous NaCl solution. The lower glass slide carries a \(50\mathrm{nm}\) Au film that is locally heated by optical absorption of a focused laser of \(\lambda = 532\mathrm{nm}\) wavelength. b, The experimental setup comprises an inverted optical microscope equipped with an acusto-optical deflector controlled, steerable focused laser with a wavelength of \(\lambda = 532\mathrm{nm}\) . The AuNPs are observed using darkfield illumination with an oil-immersion dark-field condenser (NA 1.2) and a \(100\times\) oil-immersion objective set to NA 0.6. Images are recorded with an EMCCD camera. </center>  
+
+<|ref|>text<|/ref|><|det|>[[80, 415, 481, 570]]<|/det|>
+erslip (Fig. 1a). The sample chamber contains a suspension of gold nanoparticles (AuNPs) or other nano- objects (polystyrene NPs and ellipsoids) with a controlled amount of salt (NaCl), surfactants (SDS, ...) or polymers (PEG). The gold film is heated locally in an inverted microscopy setup by a highly focused laser (532 nm) using beam steering optics (Acusto- Optic- Deflector, AOD). The nano- objects are observed using darkfield illumination with an oil- immersion darkfield condenser (NA 1.2) and a \(100\times\) oil- immersion objective set to NA 0.6 (Fig. 1b). Additional details of the experimental setup and sample preparation are provided in the Methods section.  
+
+<|ref|>text<|/ref|><|det|>[[80, 572, 481, 728]]<|/det|>
+The trapping of nano- objects as detailed in the following is comprising two effects. i) The vertical confinement of the suspended objects as achieved by an attractive interaction of the suspended nano- objects with the gold surface, which is found to be the vdW interaction for gold nanoparticles and can be replaced by depletion forces for other materials. ii) The generation of thermo- osmotic boundary flows that are induced by the local heating and the corresponding perturbation of the liquid- solid interactions. This boundary flow is directed radially inwards to the heated spot and provides a confining hydrodynamic force on suspended objects at the heating spot.  
+
+<|ref|>text<|/ref|><|det|>[[80, 751, 481, 880], [516, 415, 916, 468]]<|/det|>
+Dynamics of AuNPs close to a Au film Consider a single AuNP with a radius of \(R = 125\mathrm{nm}\) that is suspended in an aqueous solution of NaCl at a \(10\mathrm{mM}\) concentration and diffusing in a thin liquid film of about \(3\mu \mathrm{m}\) thickness over a \(50\mathrm{nm}\) Au film (Fig. 1a). Exploring the diffusion of the particle we observe a restriction of the \(z\) - positions to a thin layer close to the gold film. The gold particle never defocuses under these conditions while it does in deionized (DI) water (Supplementary Video 1). This restricted out- of- plane motion is the result of interactions comprising an attractive vdW contribution, a repulsion of the electrostatic double layers of the particle and surface32 as described by the DLVO theory and the gravitational potential (see Supplementary Information for details).   
+
+<|ref|>equation<|/ref|><|det|>[[576, 480, 915, 496]]<|/det|>
+\[V(d,c_{0}) = V_{\mathrm{E}}(d,c_{0}) + V_{\mathrm{vdW}}(d) + V_{\mathrm{G}}(d). \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[516, 508, 916, 675]]<|/det|>
+The total potential for the experimental situation is depicted in Fig. 2a for different salt concentrations (see SI for parameters) as a function of the surface- to- surface distance \(d = z - R\) . The stronger screening of the surface charges at the gold film and the AuNP at higher salt concentration increase the importance of attractive vdW interactions to create this secondary minimum in the DLVO part of the potential. This potential influences the observed dynamics as well as the particle couples with its hydrodynamic flow field to the solid boundary33. The in- plane \(D_{\parallel}\) (equation 2) and out- of- plane \(D_{\perp}\) diffusion coefficient (see Supplementary Information for details) are modulated with the distance \(z\) of the particle from the wall.  
+
+<|ref|>equation<|/ref|><|det|>[[555, 686, 915, 716]]<|/det|>
+\[\frac{D_{\parallel}(z)}{D_0}\approx 1 - \frac{9}{16}\frac{R}{z} +\frac{1}{8}\left(\frac{R}{z}\right)^3\pm \dots \coloneqq \gamma_{\parallel}^{-1}(z) \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[516, 726, 916, 880]]<|/det|>
+Over the course of a diffusion trajectory, the particle samples different regions with different diffusion coefficients according to its probability density \(p(d) \propto \exp (- V(d, c_0) / (k_B T))\) to be at a distance \(d\) from the surface (filled regions in Fig. 2a). The observed in- plane diffusion coefficient is thus a weighted average of the diffusion coefficient over the different vertical positions \(d\) . Using \(p(d)\) we can calculate the corresponding salt concentration dependence of the in- plane diffusion coefficient and compare that to the experimental results. Fig. 2b shows that the experimentally observed \(D_{\parallel}\) is decreasing with increasing salt concentration due to the hydrodynamic coupling in fair
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[78, 81, 920, 576]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[80, 584, 917, 700]]<|/det|>
+<center>Fig. 2: DLVO potential, lateral diffusion analysis, temperature distribution and thermo-osmotic flow field. a, Plot of the DLVO potential, equation (1), between a \(250~\mathrm{nm}\) Au NP and a \(50~\mathrm{nm}\) Au film on a glass surface as function of surface-surface distance \(d\) for different NaCl concentrations \(c_0\) . The shaded curves display the calculated probability density for finding the particle at this distance at the different salt concentrations (see Supplementary Information for details). The vertical dashed lines correspond to the mean distance of the particle as calculated from the probability density for a \(3\mu \mathrm{m}\) liquid film height. b, The measured diffusion coefficient \(D_1 / D_0\) parallel to the Au film with respect to the bulk diffusion coefficient \(D_0\) as function of the NaCl concentration \(c_0\) . Symbols correspond to the experimental values. The lines reflect the theoretical prediction including a distance dependent diffusion coefficient for three different Hamaker constants (dotted: \(A_H = 4\cdot 10^{-20}\mathrm{J}\) , dash-dotted: \(5\cdot 10^{-20}\mathrm{J}\) and dashed: \(6\cdot 10^{-20}\mathrm{J}\) ) of gold according to a Boltzmann weighting (see text). c, Relation between the mean distance \(\langle d\rangle\) and the NaCl concentration \(c_0\) . The symbols and the horizontal lines denote the calculated distances for measured concentrations. d, Simulation of the relative temperature increment in the \(xz\) -plane of the sample. e, Experimentally obtained temperature increment \(\Delta T_{\mathrm{max}}\) as a function of the incident laser power \(P_0\) (green data points) compared to the simulated values (green curve). f, Measured thermo-osmotic flow field in the \(xy\) -plane in close proximity to the gold film ( \(z< 500\mathrm{nm}\) ). g, Measured thermo-osmotic flow field in the \(xz\) -plane. h, Illustration of the measured flow field planes in f and g. i, j, The \(x\) - and \(z\) -component of the measured flow velocities compared to the simulation results in k and l. </center>  
+
+<|ref|>text<|/ref|><|det|>[[80, 721, 481, 877]]<|/det|>
+agreement with the theoretical predictions (for three different Hamaker constants for the AuNP gold surface interaction). Calculating in addition the mean distance \(\langle d\rangle\) of the particle from the surface reveals that only in the case of \(c_0 = 10\mathrm{mM}\) the particle is confined in the DLVO potential well, while at lower salt concentrations an enhanced probability (Fig. 2a) of finding the particle near the surface is the cause of the observed diffusion coefficient. These calculations help us to estimate the mean distance \(\langle d\rangle\) of the particle from the surface, which is about \(1.5\mu \mathrm{m}\) and \(0.9\mu \mathrm{m}\) for the lowest NaCl concentrations (Fig. 2c). At a concentration of \(c_0 = 10\mathrm{mM}\) the particle is hovering at a distance  
+
+<|ref|>text<|/ref|><|det|>[[515, 721, 916, 787]]<|/det|>
+of \(\langle d\rangle = 20\mathrm{nm}\) surface. Note that this corresponds to values of \(z / (2R)\approx 0.58\) , which is far below the commonly explored region of the hydrodynamic coupling of colloids to walls<sup>33</sup> allowing to experimentally explore new terrains also in the field of hydrodynamic wall coupling of colloids.  
+
+<|ref|>text<|/ref|><|det|>[[515, 804, 915, 881]]<|/det|>
+Hydrodynamic Particle Confinement When tightly focusing the light of \(532\mathrm{nm}\) wavelength to the gold film, a part of the incident energy (about \(30\%\) ) is absorbed and converted into heat that perturbs the liquid- solid interactions. The temperature rise at the gold surface can be determined using a thin nematic liquid crystal (5CB) film and substan
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 83, 480, 122]]<|/det|>
+tiated by finite element simulations with the complete threedimensional temperature profile in the solution (see Fig. 2d, e and Supplementary Information for details).  
+
+<|ref|>text<|/ref|><|det|>[[80, 123, 481, 292]]<|/det|>
+These local temperature perturbations of the solidliquid interactions at the interface induce a thermo- osmotic flow34,35. Taking a liquid volume element close to the solid from the cold side and exchanging that with one at the hot side would not only transport heat since the liquid volumes have different temperatures, but also additional free energy as the liquid has a different interaction with the solid in these regions. The flow is induced in an ultrathin boundary layer corresponding in thickness to the length scale of liquid- solid interactions. Since the characteristic interaction length of liquid- solid interactions is only a few nanometers, the boundary flow on the substrate can be collapsed into a quasi- slip hydrodynamic boundary condition:  
+
+<|ref|>equation<|/ref|><|det|>[[150, 298, 480, 338]]<|/det|>
+\[v_{\parallel} = -\frac{1}{\eta}\int_{0}^{\infty}z h(z)\mathrm{d}z\frac{\nabla_{\parallel}T}{T} = \chi \frac{\nabla_{\parallel}T}{T}, \quad (3)\]  
+
+<|ref|>text<|/ref|><|det|>[[80, 345, 481, 775]]<|/det|>
+where \(h(z)\) is the excess enthalpy, \(T\) the temperature and \(\nabla_{\parallel}T\) is the temperature gradient parallel to the surface. The integral can be summarized to a thermo- osmotic coefficient \(\chi\) . The thermo- osmotic coefficient \(\chi\) , therefore, contains all information about the interfacial interaction between the liquid and the solid. If \(\chi < 0\) the liquid is driven to the cold, whereas for \(\chi > 0\) , the liquid is driven to the hot. These boundary flows are present at all liquid- solid interfaces with tangential temperature gradients, though, they are commonly overlooked. They become particularly important for plasmonic and thermo- plasmonic trapping16, as those techniques rely on the dynamics of molecules and particles in the direct vicinity of plasmonic nanostructures. The boundary flow drives the flow field inside the fluid film. The resulting volumetric flow field can be tracked experimentally by single AuNPs in DI water, where the particles are not confined to a surface layer as reported above. We analyze the in- plane \((xy)\) position of the particle and its \(z\) - position, where the latter is estimated from the radius \(r_0\) of the defocussed particle images (see Supplementary Video 2 and Supplementary Information for details). The measured velocity distributions in the \(xy\) - plane near the gold layer and in the \(xz\) - plane are shown in Fig. 2f and g, respectively. The \(x\) - and \(z\) - component of the measured flow velocities are depicted in Fig. 2i, j and compare well to simulation results in Fig. 2k, l. From these measurements, we extract a thermo- osmotic coefficient on the order of \(\chi \sim 10\cdot 10^{- 10}\mathrm{m}^2\mathrm{s}^{- 1}\) (see Supplementary Information for details). We can break down the contributions to this value with equations (4) and (5) to estimate the double layer and vdW contributions using the experimental parameters. Note that AuNP do not show thermophoresis due to their high thermal conductivity and thus isothermal surface.  
+
+<|ref|>equation<|/ref|><|det|>[[163, 778, 480, 806]]<|/det|>
+\[\chi_{\mathrm{E}} = \frac{\epsilon\zeta^2}{8\eta}\approx 0.8\cdot 10^{-10}\mathrm{m}^2 /\mathrm{s}^{-1}. \quad (4)\]  
+
+<|ref|>text<|/ref|><|det|>[[80, 812, 481, 852]]<|/det|>
+For the electrostatic contribution we used \(\zeta = - 30\mathrm{mV}^{36}\) and \(\epsilon = 80\epsilon_0\) (see Methods section for details). An estimate of the vdW contribution can be given by  
+
+<|ref|>equation<|/ref|><|det|>[[152, 856, 480, 884]]<|/det|>
+\[\chi_{\mathrm{vdW}} = \frac{A_{\mathrm{H}}\beta T}{3\pi\eta d_0}\approx 9.3\cdot 10^{-10}\mathrm{m}^2\mathrm{s}^{-1}, \quad (5)\]  
+
+<|ref|>text<|/ref|><|det|>[[515, 82, 916, 288]]<|/det|>
+with \(\beta = 0.2\cdot 10^{- 3}\mathrm{K}^{- 1}\) being the thermal expansion coefficient of water and \(d_0 = 0.2\mathrm{nm}\) for the cut- off parameter34 (see Supplementary Information for details). The sum of both contributions \(\chi = \chi_{\mathrm{E}} + \chi_{\mathrm{vdW}} = 10.1\cdot 10^{- 10}\mathrm{m}^2\mathrm{s}^{- 1}\) matches well the experimental result and suggests that thermo- osmosis at gold- water interfaces is governed by vdW interactions. The obtained quasi slip velocities are ranging up to \(80\mu \mathrm{m / s}\) and provide, due to their omnipresence, a unique tool for nanofluidics. These thermo- osmotic flows are induced without any external pressure difference. They can be controlled by the light intensity heating laser and are quickly switched due to the extremely fast heat conduction at these length scales. Moreover the finding of the vdW dominated thermo- osmotic flows suggest that such contributions must be present in any plasmonic trapping experiment with extended gold structures12,16,17,27.  
+
+<|ref|>text<|/ref|><|det|>[[515, 289, 916, 445]]<|/det|>
+Using \(F_T^{\mathrm{eff}} = 6\pi \eta R\gamma_{\parallel}v_x\) and \(F_T^{\mathrm{eff}} = 6\pi \eta R\gamma_{\perp}v_z\) , where \(\gamma_{\parallel}\) and \(\gamma_{\perp}\) are the correction factors for the friction coefficient of a sphere close to a surface we are able to extract the hydrodynamic forces that are exerted on the AuNP tracers (see equation (2) and Supplementary Information for details). The lateral forces allow to confine objects at the heating spot, yet the hydrodynamic force normal to the surface ( \(z\) - direction) is repulsive without any additional interaction. Finally, such boundary flows with substantial vertical velocity gradients also exhibit a vorticity (see Supplementary Information for details) that generates a torque on suspended objects causing them to rotate37.  
+
+<|ref|>text<|/ref|><|det|>[[515, 465, 916, 620]]<|/det|>
+Single particle trapping and flow field multiplexing The repulsive normal component caused by the hydrodynamic drag is now superimposed with the attractive force due to the DLVO potential when increasing the NaCl concentration. The surface- to- surface distance between AuNP and gold film and the depth of the appearing secondary DLVO potential minimum can be controlled by the NaCl concentration. At a NaCl concentration of about \(c_0 = 10\mathrm{mM}\) , the attractive potential has a depth of about \(10k_{\mathrm{B}}T\) (see Fig. 2a) and is strong enough to compete with the vertical drag force and additional optical forces on the AuNP to trap the particle above the heating spot.  
+
+<|ref|>text<|/ref|><|det|>[[515, 621, 916, 881]]<|/det|>
+Supplementary Video 3 demonstrates this trapping of an AuNP above the hot spot on the Au surface. This is purely the result of the hydrodynamic drag forces generated by the thermo- osmotic flow and the attractive vdW interaction between the AuNP and the Au film. This observation is substantiated by a quantitative evaluation of the lateral trap stiffness and vertical forces, as depicted in Fig. 3a, b. The fluctuations of the particle in the hydrodynamic flow arise from a balance of the restoring hydrodynamic currents and the diffusive currents. Analysis of the lateral position histograms (inset in Fig. 3c for \(1.25\mathrm{mW}\) ) yields an effective stiffness of the trap (Fig. 3a) that well matches the predictions based on the thermo- osmotic flow (Fig. 2i, j). The hydrodynamic trapping stiffness increases linearly up to a heating power of about \(1.8\mathrm{mW}\) . At this power, the vertical forces become strong enough to let the particle escape the secondary DLVO minimum, which is visible from the \(z\) - position time traces displayed in Fig. 3d. The AuNPs are then observed to move vertically out of the DLVO potential to follow the flow inside the sample and to eventually return
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[80, 81, 916, 253]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[80, 264, 916, 328]]<|/det|>
+<center>Fig. 3: Forces on trapped NPs in \(10\mathrm{mM}\) NaCl. a, The lateral trap stiffness obtained from the experimental position histograms (blue data points) as function the laser power \(P_0\) compared to the simulation result (blue solid line). b, The \(x\) -component of the thermo-osmotic drag force \(F_{\mathrm{T}}^{\mathrm{TF}}\) (blue line), the optical force \(F_{\mathrm{Z}}^{\mathrm{OF}}\) (green line) and the total force \(F_{\mathrm{Z}}^{\mathrm{OF}} + F_{\mathrm{Z}}^{\mathrm{TF}}\) (black dashed line) as function the incident laser power \(P_0\) for a NP located at \(x = 0\) , \(d = 30\) ( \(z = d + R\) ). The attractive DLVO force \(F_{\mathrm{DLVO}}^{\mathrm{DLVO}}\) is independent of the incident laser power and depicted as horizontal, red line. c, Trajectory of a AuNP for a heating laser power of \(1.25\mathrm{mW}\) (see Supplementary Video 2 for details). The inset shows the corresponding lateral distribution histogram. d, Time traces of the \(z\) -position for three different laser powers. e, Trajectory of a AuNP for a heating laser power of \(2.5\mathrm{mW}\) , which is above the threshold power of \(2.25\mathrm{mW}\) . </center>  
+
+<|ref|>text<|/ref|><|det|>[[81, 349, 481, 595]]<|/det|>
+to the boundary flow via sedimentation (Fig. 3e, Supplementary Video 4). The forces which eject the particle from the potential comprise the hydrodynamic and the optical forces due to the radiation pressure from the heating laser leaked through the film. We have evaluated the individual contributions in simulations. They are shown together with the hydrodynamic force and the total vertical force as compared to the attractive force of the DLVO potential (Fig. 3b) and provide quantitative agreement (threshold heating power of \(2.25\mathrm{mW}\) ) with our experimental results. Note that while the stationary distribution of particles in the vertical direction is not influenced by the diffusive dynamics, the escape rate from the potential well is heavily altered by the fact that the vertical diffusion coefficient \(D_{\perp}\) of the particle is decreasing to zero when approaching the gold film. This is enhancing the trapping times considerably (see Supplementary Information for details) but also increases the time required for the particle to enter the DLVO minimum by diffusion.  
+
+<|ref|>text<|/ref|><|det|>[[81, 596, 481, 739]]<|/det|>
+The observed trapping is, hence, a vdW assisted thermodynamic process. Vertical confinement is achieved by vdW attraction and double layer repulsion, while lateral confinement is the result of thermo- osmotic flows induced in an ultra- thin sheet of liquid at the interface. No additional contributions, for example, due to convective flows with similar flow patterns (see Supplementary Information for details) or thermo- electric effects are required for a quantitative description \(^{25,31,38,39}\) . Precise tuning of the DLVO potential enables the trapping of even smaller Au NPs (Fig. 4a, Supplementary Video 5).  
+
+<|ref|>text<|/ref|><|det|>[[81, 740, 481, 880]]<|/det|>
+The speed of heat diffusion, which is about 4 orders of magnitude faster than the particle diffusion \(^{40}\) allows us to introduce a flow field multiplexing. We switch the heating location between different positions inducing thermo- osmotic flow fields for time periods of about \(100\mu \mathrm{s}\) . With the help of this multiplexing, we are able to hold multiple \(R = 125\mathrm{nm}\) AuNPs (Fig. 4b, c) at distances of less than \(1\mu \mathrm{m}\) , which would not be possible with continuous heating of close- by locations (Supplementary Videos 6 and 7). A trapped AuNP can also be guided along the predefined path over the Au film as fast as \(10\mu \mathrm{m}\mathrm{s}^{- 1}\) (Fig. 4d,  
+
+<|ref|>text<|/ref|><|det|>[[516, 349, 916, 467]]<|/det|>
+Supplementary Video 8). At larger manipulation speeds ( \(f > 100\mathrm{Hz}\) ) and higher heating power ( \(P_0 > 10\mathrm{mW}\) ) the thermo- osmotic attraction to the heating spot is combined with thermo- viscous flows \(^{41,42}\) . These flows originate from the temperature dependent viscosity \(\eta (T)\) of the liquid and are directed opposite to the scanning direction of the laser \(^{42}\) . The result of this combination of thermo- osmosis and thermo- viscous flows is a rotating ring- like particle structure (Fig. 4e and Supplementary Video 9).  
+
+<|ref|>text<|/ref|><|det|>[[516, 468, 916, 556]]<|/det|>
+These different effects that can be exploited in a simple planar geometry give rise to numerous applications including for example a freely configurable nanoparticle on mirror geometry for plasmonic sensing \(^{5,43}\) . The multiplexing of local flow field may be be helpful to construct more complex effective flow fields for an efficient transport of analytes without external pressure.  
+
+<|ref|>text<|/ref|><|det|>[[516, 581, 916, 788]]<|/det|>
+Beyond thermo- osmotic van der Waals trapping So far, the presented manipulation is based on thermo- osmotic flows that drive the lateral motion of suspended colloids and a vertical confinement due to the secondary minimum of the DLVO potential between AuNP and Au film. While the thermo- osmotic flows are characteristic for all systems containing a heated gold/water interface including all previous studies on thermo- plasmonic trapping, the DLVO potential minimum is much weaker for other materials like polymer colloids or macromolecules due to their smaller vdW attraction. Often, those system even show a repulsion from the heat source due to thermophoresis, which is not present for AuNP. A more generalized strategy therefore needs additional attractive contributions, which confine suspended colloids or molecules to regions close to the gold surface to take advantage of the thermo- osmotic flow.  
+
+<|ref|>text<|/ref|><|det|>[[516, 791, 916, 880]]<|/det|>
+Such attractive contributions can arise from depletion interactions \(^{21,44}\) . Thereby a temperature gradient repels dissolved molecules from the heated regions generating a concentration gradient that drives suspended nano- objects to the heating spot. To demonstrate this effect we use the surfactant sodium dodecyl sulfate (SDS) at a concentration of \(5\mathrm{mM}\) well below the critical micelle concentration
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 83, 481, 213]]<|/det|>
+(8.2 mM) to avoid complications of micelle formation. We suspend additional polystyrene particles(PS) and AuNPs of the same size ( \(R = 125 \mathrm{nm}\) ) in the solution and compare their dynamics to a solution with AuNPs and PS particles without SDS but \(10 \mathrm{mM} \mathrm{NaCl}\) . Remarkably, the heated spot is attractive for both AuNPs and for PS NPs (Fig. 4f, Supplementary Video 10) in the SDS solution showing even PS colloidal crystal growth, while only the AuNP is trapped in the NaCl solution and the PS particles are repelled by thermophoresis (Fig. 4g, Supplementary Video 11).  
+
+<|ref|>text<|/ref|><|det|>[[80, 214, 481, 318]]<|/det|>
+The observations in NaCl are readily explained by the fact that the AuNP is confined in the DLVO minimum as demonstrated above but the PS particle is not due to a 10 times lower Hamaker constant (see Supplementary Information for details). The PS particle samples the whole liquid film thickness equally and not preferentially the region close to the Au film and experiences an additional thermophoretic drift velocity given by  
+
+<|ref|>equation<|/ref|><|det|>[[187, 328, 480, 355]]<|/det|>
+\[\pmb {u} = -\frac{2}{3}\chi \frac{\nabla T}{T} = -D_{\mathrm{T}}\nabla T, \quad (6)\]  
+
+<|ref|>text<|/ref|><|det|>[[80, 362, 481, 442]]<|/det|>
+where \(D_{\mathrm{T}}\) is the thermophoretic mobility \(^{34}\) and \(\nabla T\) the temperature gradient (Fig. 4g, Supplementary Video 11). For \(\chi >0\) the particle is driven to the cold. From equation (4) we find \(\chi \approx \chi_{\mathrm{E}} = 1.28\cdot 10^{- 10} \mathrm{m}^{2} \mathrm{s}^{- 1}\) and \(D_{\mathrm{T}} \approx 0.3 \mu \mathrm{m}^{2} \mathrm{K}^{- 1} \mathrm{s}^{- 1}\) , where we have used a measured zeta potential of \(\zeta \approx - 38 \mathrm{mV}\) . The vdW contribution, \(\chi_{\mathrm{vdW}}\) to  
+
+<|ref|>text<|/ref|><|det|>[[515, 83, 916, 162]]<|/det|>
+either the thermophoretic drift or the attraction to the gold surface can be neglected due to the smaller Hamaker constant of PS. From the stationary probability distribution of the PS NP we find a Soret coefficient of \(S_{\mathrm{T}} \approx 0.24 \mathrm{K}^{- 1}\) (see Supplementary Information for details) in agreement with our theoretical prediction \(S_{\mathrm{T}} = D_{\mathrm{T}} / D_{\parallel} \approx 0.21 \mathrm{K}^{- 1}\) .  
+
+<|ref|>text<|/ref|><|det|>[[515, 163, 916, 277]]<|/det|>
+In the SDS solution, the additional surfactant molecules now undergo thermophoresis to yield a concentration gradient in which suspended colloidal particles drift. The lower concentration in the heated regions promotes an effective attractive interaction of suspended colloids with the gold surface due to depletion forces. The drift velocity is described by an additional term to the thermodiffusion coefficient \(D_{\mathrm{T}}\) , that is, the second term in brackets in equation (7) \(^{21,34,44}\) .  
+
+<|ref|>equation<|/ref|><|det|>[[560, 285, 915, 316]]<|/det|>
+\[\pmb {u} = -\left(D_{\mathrm{T}} - \frac{k_{\mathrm{B}}}{3\eta} R^{2}c_{0}N_{\mathrm{A}}\left(T S_{\mathrm{DS}}^{\mathrm{SDS}} - 1\right)\right)\nabla T \quad (7)\]  
+
+<|ref|>text<|/ref|><|det|>[[515, 324, 916, 441]]<|/det|>
+Here \(R\) is the size of the SDS molecule, \(c\) the concentration in units of mol/l and \(S_{\mathrm{DS}}^{\mathrm{SDS}}\) the Soret coefficient of SDS. For \(R = 2 \mathrm{nm}^{45}\) , \(c_{0} = 5 \mathrm{mM}\) and \(S_{\mathrm{DS}}^{\mathrm{SDS}} = 0.03 \mathrm{K}^{- 146}\) we find \(- 0.43 \mu \mathrm{m}^{2} \mathrm{K}^{- 1} \mathrm{s}^{- 1}\) for the additional depletion contribution, which exceeds the thermophoretic mobility, \(D_{\mathrm{T}} \approx 0.3 \mu \mathrm{m}^{2} \mathrm{K}^{- 1} \mathrm{s}^{- 1}\) , rendering the overall mobility negative. The PS NPs and the AuNPs are thus driven to the the heated Au film surface (Fig. 4f, Supplementary Video 10) which allows for further transport in the thermo- osmotic  
+
+<|ref|>image<|/ref|><|det|>[[80, 460, 916, 784]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[80, 800, 916, 883]]<|/det|>
+<center>Fig. 4: Manipulation of NPs over a Au film in NaCl and SDS solution. a, A AuNP with \(50 \mathrm{nm}\) radius trapped at a NaCl of \(30 \mathrm{mM}\) (Supplementary Video 5). b, Manipulation of two AuNPs by a multiplexed laser beam (Supplementary Video 6). c, Control of three AuNPs (Supplementary Video 7). d, Actuation of a single AuNP on a circular trajectory by a steerable laser beam (Supplementary Video 8). The green and white, dashed arrows denote the moving direction of the laser focus and particle, respectively. e, Generation of thermo-viscous flows by rotating the laser focus on a circle with a rotation frequency of \(f = 500 \mathrm{Hz}\) at high laser powers (Supplementary Video 9). Note, that laser movement (green arrow) and the thermo-viscous flow (white, dashed arrow) and have opposite directions. f, Attraction of a AuNP and PS NPs in \(5 \mathrm{mM}\) SDS due to depletion (Supplementary Video 10). g, An AuNP (125 nm radius) trapped in an ensemble of polystyrene (PS) NPs of the same size at \(10 \mathrm{mM}\) NaCl (Supplementary Video 11), where the PS-particles are repelled due to thermophoresis. h, Attraction of PS ellipsoids (2.39 \(\mu \mathrm{m}\) major-axis length, \(0.34 \mu \mathrm{m}\) minor-axis length) in \(5 \mathrm{mM}\) SDS (Supplementary Video 12). </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 84, 481, 175]]<|/det|>
+boundary flow. Additional contributions, as for example thermo- electric fields may even enhance the attractive components. Overall, this concept is readily transferred to other objects as shown in Figure 4h and Supplementary Video 12, where we have trapped ellipsoidal PS particles in a \(5\mathrm{mM}\) solution of SDS. Note that as compared to other schemes, our approach always includes thermo- osmotic boundary flows.  
+
+<|ref|>text<|/ref|><|det|>[[80, 191, 481, 512]]<|/det|>
+Conclusion In conclusion, we have demonstrated that thermo- hydrodynamic boundary flows can manipulate nano- objects with unprecedented flexibility in a very simple sample geometry. These flows are the key for future thermo- optofluidic implementations with an extensive range of applications in the fields of i) nanoparticle sorting and separation; ii) assembly of nanophotonic circuits and plasmonic quantum sensors; iii) biotechnology on- chip laboratories and iv) manufacturing of nanomaterials and functional nanosurfaces. We have substantiated our experimental findings of thermo- osmotic flow assisted trapping with a quantitative theoretical description. A flow field multiplexing scheme has been further developed to allow for the simultaneous manipulation of many individual nano- objects and the generation of complex effective flow patterns. Our concept can be combined with other thermally induced effects such as thermophoresis, depletion forces and thermoviscous flows to form a fully- featured nanofluidic system- on- a- chip. Besides direct consequences for the field of plasmonic nano- tweezers and other thermoplasmonic trapping schemes, the use of thermo- hydrodynamic flows as a tool for nanofluidic applications will extend the limits at the forefront of nanotechnology and help to develop AI and feedback controlled schemes for the material synthesis.  
+
+<|ref|>sub_title<|/ref|><|det|>[[80, 529, 172, 543]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[90, 552, 483, 880]]<|/det|>
+[1] O. M. Maragò, P. H. Jones, P. G. Gucciardi, G. Volpe, and A. C. Ferrari, Optical Trapping and Manipulation of Nanostructures, Nat. Nanotechnol. 8, 807- 819 (2013). [2] D. Gao, W. Ding, M. Nieto- Vesperinas, X. Ding, M. Rahman, T. Zhang, C. Lim, and C.- W. Qiu, Optical Manipulation from the Microscale to the Nanoscale: Fundamentals, Advances and Prospects, Light Sci. Appl. 6, 17039 (2017). [3] C. Bradac, Nanoscale Optical Trapping: A Review, Adv. Opt. Mat. 6, 1800005 (2018). [4] J. Xavier, S. Vincent, F. Meder, and F. Vollmer, Advances in Optoplasmonic Sensors - Combining Optical Nano/Microcavities and Photonic Crystals with Plasmonic Nanostructures and Nanoparticles, Nanophotonics 7, 1- 38 (2018). [5] G.- C. Li, Q. Zhang, S. A. Maier, and D. Lei, Plasmonic Particle- on- Film Nanocavities: A Versatile Platform for Plasmon- Enhanced Spectroscopy and Photochemistry, Nanophotonics 7, 1865- 1889 (2018). [6] L. Lin, X. Peng, and Y. Zheng, Reconfigurable Opto- Thermoelectric Printing of Colloidal Particles, Chem. Commun. 53, 7357- 7360 (2017). [7] Y. Xie, J. Rufo, R. Zhong, J. Rich, P. Li, K. W. Leong, and T. J. Huang, Microfluidic Isolation and Enrichment of Nanoparticles, ACS Nano 14, 16220- 16240 (2020). [8] I. A. Favre- Bulle, A. B. Stilgoe, E. K. Scott, and H. Rubinsztein- Dunlop, Optical Trapping in Vivo: Theory, Practice, and Applications, Nanophotonics 8, 1023- 1040 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[520, 85, 916, 865]]<|/det|>
+[9] D. Choudhary, A. Mossa, M. Jadhav, and C. Cecconi, Bio- Molecular Applications of Recent Developments in Optical Tweezers, Biomolecules 9, 23 (2019). [10] M. Kayci, H.- C. Chang, and A. Radenovic, Electron Spin Resonance of Nitrogen- Vacancy Defects Embedded in Single Nanodiamonds in an ABEL Trap, Nano Lett. 14, 5335- 5341 (2014). [11] M. L. Juan, M. Righini, and R. Quidant, Plasmon Nano- Optical Tweezers, Nat. Photonics 5, 349- 356 (2011). [12] Y. Zhang, C. Min, X. Dou, X. Wang, H. P. Urbach, M. G. Somekh, and X. Yuan, Plasmonic Tweezers: For Nanoscale Optical Trapping and Beyond, Light Sci. Appl. 10, 59 (2021). [13] I. De Vlaminck and C. Dekker, Recent Advances in Magnetic Tweezers, Annu. Rev. Biophys. 41, 453- 472 (2012). [14] Q. Wang and W. E. Moerner, An Adaptive Anti- Brownian Electrokinetic Trap with Real- Time Information on Single- Molecule Diffusivity and Mobility, ACS Nano 5, 5792- 5799 (2011). [15] M. C. Wu, Optoelectronic Tweezers, Nat. Photonics 5, 322- 324 (2011). [16] A. Kotnala and R. Gordon, Quantification of High- Efficiency Trapping of Nanoparticles in a Double Nanohole Optical Tweezers, Nano Lett. 14, 853- 856 (2014). [17] Q. Jiang, B. Rogez, J.- B. Claude, G. Baffou, and J. Wenger, Quantifying the Role of the Surfactant and the Thermophoretic Force in Plasmonic Nano- Optical Trapping, Nano Lett. 20, 8811- 8817 (2020). [18] M. Braun and F. Cichos, Optically Controlled Thermophoretic Trapping of Single Nano- Objects, ACS Nano 7, 11200- 11208 (2013). [19] M. Braun, A. P. Bregulla, K. Günther, M. Mertig, and F. Cichos, Single Molecules Trapped by Dynamic Inhomogeneous Temperature Fields, Nano Lett. 15, 5499- 5505 (2015). [20] M. Fränzl, T. Thalheim, J. Adler, D. Huster, J. Possekardt, M. Mertig, and F. Cichos, Thermophoretic Trap for Single Amyloid Fibril and Protein Aggregation Studies, Nat. Methods 16, 611- 614 (2019). [21] H.- R. Jiang, H. Wada, N. Yoshinaga, and M. Sano, Manipulation of Colloids by a Nonequilibrium Depletion Force in a Temperature Gradient, Phys. Rev. Lett. 102, 208301 (2009). [22] A. P. Bregulla, H. Yang, and F. Cichos, Stochastic Localization of Microswimmers by Photon Nudging, ACS Nano 8, 6542- 6550 (2014). [23] U. Khadka, V. Holubec, H. Yang, and F. Cichos, Active Particles Bound by Information Flows, Nat. Commun. 9, 3864 (2018). [24] M. Fränzl, S. Muños- Landin, V. Holubec, and F. Cichos, Fully Steerable Symmetric Thermoplasmonic Microswimmers, ACS Nano 15, 3434- 3440 (2021). [25] L. Lin, M. Wang, X. Peng, E. N. Lissek, Z. Mao, L. Scarabelli, E. Adkins, S. Coskun, H. E. Unalan, B. A. Korgel et al., Opto- Thermoelectric Nanotweezers, Nat. Photonics 12, 195- 201 (2018). [26] J. C. Ndukaife, A. V. Kildishev, A. G. A. Nnanna, V. M. Shalaev, S. T. Wereley, and A. Boltasseva, Long- Range and Rapid Transport of Individual Nano- Objects by a Hybrid Electrorheromoplasmonic Nanotweezer, Nat. Nanotechnol. 11, 53- 59 (2016). [27] C. Hong, S. Yang, and J. C. Ndukaife, Stand- off Trapping and Manipulation of Sub- 10 Nm Objects and Biomolecules Using Opto- Thermo- Electrohydrodynamic Tweezers, Nat. Nanotechnol. 15, 908- 913 (2020).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[80, 85, 483, 870]]<|/det|>
+[28] J. S. Donner, G. Baffou, D. McCloskey, and R. Quidant, Plasmon- Assisted Optofluidics, ACS Nano 5, 5457- 5462 (2011).[29] B. J. Roxworthy, A. M. Bhuiya, S. P. Vanka, and K. C. Toussaint, Understanding and Controlling Plasmon- Induced Convection, Nat. Commun. 5, 3173 (2014).[30] J. Chen, J. F.- C. Loo, D. Wang, Y. Zhang, S.- K. Kong, and H.- P. Ho, Thermal Optofluidics: Principles and Applications, Adv. Opt. Mater. 8, 1900829 (2020).[31] B. Ciraulo, J. Garcia- Guirado, I. de Miguel, J. Ortega Arroyo, and R. Quidant, Long- Range Optofluidic Control with Plasmon Heating, Nat. Commun. 12, 2001 (2021).[32] J. N. Israelachvili, Intermolecular and Surface Forces, 3rd Edition (Elsevier, Academic Press, Amsterdam, 2011).[33] B. Lin, J. Yu, and S. A. Rice, Direct Measurements of Constrained Brownian Motion of an Isolated Sphere between Two Walls, Phys. Rev. E 62, 3909- 3919 (2000).[34] A. Wurger, Thermal Non- Equilibrium Transport in Colloids, Rep. Prog. Phys. 73, 126601 (2010).[35] A. P. Bregulla, A. Wurger, K. Gunther, M. Mertig, and F. Cichos, Thermo- Osmotic Flow in Thin Films, Phys. Rev. Lett. 116, 188303 (2016).[36] M. Giesbers, J. Kleijn, and M. A. Cohen Stuart, The Electrical Double Layer on Gold Probed by Electrokinetic and Surface Force Measurements, J. Colloid Interface Sci. 248, 88- 95 (2002).[37] J. J. Bluemink, D. Lohse, A. Prosperetti, and L. Van Wijngaarden, A Sphere in a Uniformly Rotating or Shearing Flow, J. Fluid Mech. 600, 201- 233 (2008).[38] L. Lin, X. Peng, M. Wang, L. Scarabelli, Z. Mao, L. M. Liz- Marzán, M. F. Becker, and Y. Zheng, Light- Directed Reversible Assembly of Plasmonic Nanoparticles Using Plasmon- Enhanced Thermophoresis, ACS Nano 10, 9659- 9668 (2016).[39] L. Lin, J. Zhang, X. Peng, Z. Wu, A. C. H. Coughlan, Z. Mao, M. A. Bevan, and Y. Zheng, Opto- Thermophoretic Assembly of Colloidal Matter, Sci. Adv. 3, 1700458 (2017).[40] G. Baffou, F. Cichos, and R. Quidant, Applications and Challenges of Thermoplasmonics, Nat. Mater. 19, 946- 958 (2020).[41] F. M. Weinert, J. A. Kraus, T. Franosch, and D. Braun, Microscale Fluid Flow Induced by Thermoviscous Expansion Along a Traveling Wave, Phys. Rev. Lett. 100, 164501 (2008).[42] F. M. Weinert, C. B. Mast, and D. Braun, Optical Fluid and Biomolecule Transport with Thermal Fields, Phys. Chem. Chem. Phys. 13, 9918 (2011).[43] R. Chikkaraddy, B. de Nijs, F. Benz, S. J. Barrow, O. A. Scherman, E. Rosta, A. Demetriadou, P. Fox, O. Hess, and J. J. Baumberg, Single- Molecule Strong Coupling at Room Temperature in Plasmonic Nanocavities, Nature 535, 127- 130 (2016).[44] Y. T. Maeda, A. Buguin, and A. Libchaber, Thermal Separation: Interplay between the Soret Effect and Entropic Force Gradient, Phys. Rev. Lett. 107, 038301 (2011).[45] O. Syshchyk, D. Afanasenkau, Z. Wang, H. Kriegs, J. Buitenhuis, and S. Wiegand, Influence of Temperature and Charge Effects on Thermophoresis of Polystyrene Beads, Eur. Phys. J. E 39, 129 (2016).[46] D. Vigolo, S. Buzzaccaro, and R. Piazza, Thermophoresis and Thermoelectricity in Surfactant Solutions, Langmuir 26, 7792- 7801 (2010).[47] Q. Zhang, H. Yu, M. Barbiero, B. Wang, and M. Gu, Artificial Neural Networks Enabled by Nanophotonics, Light Sci Appl 8, 42 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[515, 85, 916, 201]]<|/det|>
+[48] J. Xavier, D. Yu, C. Jones, E. Zossimova, and F. Vollmer, Quantum Nanophotonic and Nanoplasmonic Sensing: Towards Quantum Optical Bioscience Laboratories on Chip, Nanophotonics 10, 1387- 1435 (2021).[49] J. Peng, H.- H. Jeong, Q. Lin, S. Cormier, H.- L. Liang, M. F. L. De Volder, S. Vignolini, and J. J. Baumberg, Scalable Electrochromic Nanopixels Using Plasmonics, Sci. Adv. 5, 10.1126/sciadv.aaw2205 (2019).[50] E. Mohammadi, A. Tittl, K. L. Tsakmakidis, T. V. Raziman, and A. G. Curto, Dual Nanoresonators for Ultrasonic Optical Detection, ACS Photonics 8, 1754- 1762 (2021).  
+
+<|ref|>sub_title<|/ref|><|det|>[[515, 250, 590, 263]]<|/det|>
+## Methods  
+
+<|ref|>text<|/ref|><|det|>[[515, 277, 916, 565]]<|/det|>
+Experimental setup The experimental setup (Figure S19) consists of an inverted microscope (Olympus, IX71) with a mounted piezo translation stage (Physik Instrumente, P- 733.3). The microparticles are heated by a focused, continuous- wave laser at a wavelength of \(532 \mathrm{nm}\) (CNI, MGL- III- 532). The beam diameter is increased by a beam expander and sent to an acousto- optic deflector (AA Opto- Electronic, DTSXY- 400- 532) and a lens system to steer the laser focus in the sample plane. The deflected beam is focused by an oil- immersion objective (Olympus, UPlanApo \(\times 100 / 1.35\) , Oil, Iris, NA 0.5 - 1.35) to the sample plane \((w_0 \approx 0.8 \mu \mathrm{m}\) beam waist in the sample plane). The sample is illuminated with an oil- immersion darkfield condenser (Olympus, U- DCW, NA 1.2 - 1.4) and a white- light LED (Thorlabs, SOLIS- 3C). The scattered light is imaged by the objective and a tube lens ( \(250 \mathrm{mm}\) ) to an EMCCD (electron- multiplying charge- coupled device) camera (Andor, iXon DV885LC). The variable numerical aperture of the objective was set to a value below the minimum aperture of the darkfield condenser. The dichroic beam splitter (D) was selected to reflect the laser wavelength (Omega Optical, 560DRLP) and a notch filter (F) is used to block any remaining back reflections from the laser (Thorlabs, NF533- 17). The acousto- optic deflector (AOD), as well as the piezo stage, are driven by an AD/DA (analogue- digital/digital- analogue) converter (Jäger Messtechnik, ADwin- Gold II). A LabVIEW program running on a desktop PC (Intel Core i7 2600 4 \(\times 3.40 \mathrm{GHz}\) CPU) is used to record and process the images as well as to control the AOD feedback via the AD/DA converter.  
+
+<|ref|>text<|/ref|><|det|>[[515, 583, 916, 725]]<|/det|>
+Sample preparation The sample consists of two glass coverslips \((22 \mathrm{mm} \times 22 \mathrm{mm})\) confining a thin liquid film. First, the coverslips were thoroughly cleaned by rinsing successively with acetone, isopropyl and Milli- Q water and dried with a nitrogen gun. Subsequently, the edges of one coverslip were covered with a thin layer of PDMS (polydimethylsiloxane) for sealing. The particle solution used for the experiments was prepared by dispersing \(0.250 \mu \mathrm{m}\) diameter gold particles (Cytodiagnostics) in ... solution. Finally, \(0.5 \mu \mathrm{l}\) of the mixed particle suspension is pipetted in the middle of one of the coverslips and the other is placed on top. Depending on the area covered by the liquid, typically about \((10 \mathrm{mm} \times 10 \mathrm{mm})\) , the resulting liquid film height is about \(5 \mu \mathrm{m}\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[515, 747, 666, 761]]<|/det|>
+## Acknowledgement  
+
+<|ref|>text<|/ref|><|det|>[[515, 770, 818, 781]]<|/det|>
+The authors acknowledge financial support by ...  
+
+<|ref|>sub_title<|/ref|><|det|>[[515, 803, 689, 817]]<|/det|>
+## Author contributions  
+
+<|ref|>text<|/ref|><|det|>[[515, 828, 916, 880]]<|/det|>
+M.F. and F.C. designed the experiments. M.F. performed the experiments. M.F. and F.C. analyzed the experimental data. M.F. and F.C. implemented and evaluated the numerical calculations. M.F. and F.C. wrote the manuscript. All authors discussed the results and commented on the manuscript.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[82, 83, 252, 97]]<|/det|>
+## Competing interests  
+
+<|ref|>text<|/ref|><|det|>[[82, 104, 350, 116]]<|/det|>
+The authors declare no competing interest.  
+
+<|ref|>sub_title<|/ref|><|det|>[[82, 132, 270, 146]]<|/det|>
+## Additional information  
+
+<|ref|>text<|/ref|><|det|>[[80, 155, 480, 178]]<|/det|>
+Supplementary information is available for this paper at https://doi.org ...  
+
+<|ref|>text<|/ref|><|det|>[[80, 183, 480, 205]]<|/det|>
+Correspondence and requests for materials should be addressed to F.C.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 93, 765, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[59, 130, 192, 363]]<|/det|>
+- Video6.mp4- Sl.pdf- Video3.mp4- Video2.mp4- Video8.mp4- Video7.mp4- Video1.mp4- Video5.mp4- Video4.mp4
+
+<--- Page Split --->

preprint/preprint__09b35197b5ad4345c614483e5c88cb8e2e52505854eacc518df23d1042b94777/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1",
+    "footnote": [],
+    "bbox": [
+      [
+        68,
+        55,
+        910,
+        650
+      ]
+    ],
+    "page_idx": 16
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2",
+    "footnote": [],
+    "bbox": [
+      [
+        40,
+        40,
+        825,
+        777
+      ]
+    ],
+    "page_idx": 17
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3",
+    "footnote": [],
+    "bbox": [
+      [
+        45,
+        135,
+        945,
+        308
+      ]
+    ],
+    "page_idx": 18
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4",
+    "footnote": [],
+    "bbox": [
+      [
+        55,
+        50,
+        936,
+        565
+      ]
+    ],
+    "page_idx": 19
+  }
+]

preprint/preprint__09b35197b5ad4345c614483e5c88cb8e2e52505854eacc518df23d1042b94777/preprint__09b35197b5ad4345c614483e5c88cb8e2e52505854eacc518df23d1042b94777.mmd

@@ -0,0 +1,340 @@
+
+# Deep Learning Approach for Evaluating Lumbar Intervertebral Disc Degeneration: Achieving High Accurate Segmentation for Quantitative Analysis on MRI  
+
+Hua- dong Zheng Shanghai University Yue- li Sun ( \(\boxed{\pi}\) yueli_sun@foxmail.com)  
+
+Longhua Hospital, Shanghai University of Traditional Chinese Medicine De- wei Kong  
+
+Depart ment of Radiology \(\boxed{\pi}\) Longhua Hospital of Shanghai University of TCM  
+
+Meng- chen Yin Longhua Hospital, Shanghai University of TCM  
+
+Jiang Chen Dongzhimen Hospital of BeijingUniversity of Chinese Medicine  
+
+Yong- peng Lin Dongzhimen Hospital, Beijing University of Chinese Medicine  
+
+Xue- feng Ma Shenzhen Pingle Orthopedics Hospital (Shenzhen Pingshan District Hospital of TCM)  
+
+Hong- shen Wang Guangdong Provincial Hospital of Chinese Medicine  
+
+Guangjie Yuan Shanghai University  
+
+Min Yao Longhua Hospital, Shanghai University of TCM  
+
+Xuejun Cui Longhua Hospital, Shanghai University of TCM  
+
+Yingzhong Tian Shanghai University  
+
+Yongjun Wang Shanghai University of Traditional Medicine https://orcid.org/0000- 0001- 9333- 2423  
+
+# Article
+
+<--- Page Split --->
+
+Keywords: lumbar disc degeneration, intervertebral disc degeneration, MRI, deep learning and image processing technology  
+
+Posted Date: September 2nd, 2021  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 864336/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Version of Record: A version of this preprint was published at Nature Communications on February 11th, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 28387- 5.
+
+<--- Page Split --->
+
+## Abstract  
+
+Purpose: Using deep learning and image processing technology, a standardized automatic segmentation and quantitation network of lumbar disc degeneration based on T2MRI was proposed to help residents accurately evaluate the intervertebral disc (IVD) degeneration.  
+
+Materials and Methods: A semantic segmentation network (BianqueNet) consist of self- attention mechanism skip connection module and deep feature extraction module was proposed to achieve highprecision segmentation of IVD related areas. A quantitative method was used to calculate the signal intensity difference (△SI) in IVD, average disc height (DH), disc height index (DHI), and disc height- to- diameter ratio (DHR). Quantitative ranges for these IVD parameters in a larger population was established among the 1051 MRI images collected from four hospitals around China.  
+
+Results: The average dice coefficients of BianqueNet for vertebral bodies and intervertebral discs segmentation are \(97.04\%\) and \(94.76\%\) , respectively. This procedure was suitable for different MRI centers and different resolution of lumbar spine T2MRI (ICC \(= .874 \sim .958\) ). These geographic parameters of IVD degeneration have a significant negative correlation with the modified Pfirrmann Grade, while signal intensity in IVD degeneration had excellent reliability according to the modified Pfirrmann Grade (macroF1 \(= 90.63\% \sim 92.02\%\) ).  
+
+Conclusion: we developed a fully automated deep learning- based lumbar spine segmentation network, which demonstrated strong versatility and high reliability to assist residents on IVD degeneration evaluating by means of IVD degeneration quantitation.  
+
+Implication for Patient Care: Deep learning- based approaches have the potential to maximize diagnostic performance for detecting disc degeneration and assessing risk of disc herniation while reducing subjectivity, variability, and errors due to distraction and fatigue associated with human interpretation.  
+
+## Introduction  
+
+The intervertebral disc (IVD) plays an important role in distributing loads and absorbing shock in the spine, which is comprised of a gel- like nucleus pulposus (NP), collagenous annulus fibrosis (AF) layers, and ring- like cartilaginous endplates (EP). Identifying IVD structural changes, including IVD deformation, NP dehydration and EP ossification, due to chronic degeneration or acute injury, in patients undergoing MRI of the lumbar spine has many important clinical implications \(^{1}\) . It has been determined that IVD degeneration is a consequence of aging. Accumulated compressive overload usually lead to functional fatigue fractures in endplates and subsequently IVD herniation \(^{2 - 4}\) , which may lead to increased inflammation \(^{5}\) , nerve compression \(^{6}\) and release of pain factors \(^{7}\) . Lifestyle modifications and surgical interventions are likely to be most effective for treating IVD degeneration or herniation, but it is more important to initiate screening and prevention during the earliest stages of the disease process.
+
+<--- Page Split --->
+
+MRI with morphologic cartilage imaging sequences has been shown to have high specificity but only moderate sensitivity for detecting dehydration and deformation within the IVD degeneration \(^{2,4}\) . Diagnostic performance is highly dependent on the level of reader expertise, and only moderate interobserver agreement between readers has been reported in most studies \(^{2}\) . Quantitative analysis is efficient and comprehensive in evaluating IVD degeneration by measuring the signal intensity and geometric information. Early research on quantitative measurement of intervertebral discs used general image processing programs for manual measurement \(^{8 - 10}\) . However, there is still no universal automatic IVD degeneration analysis tool in this field. The lack of a universal and widely accepted standard definition of IVD degeneration is one of the main reasons.  
+
+There has been much recent interest in using deep learning methods in medical imaging \(^{11}\) . With the wide- spread application of convolutional neural network classifiers in medical images, many studies use the rectangular box surrounding the lumbar IVD as input, and the corresponding degeneration level as the label to train the classifier for learning degenerative features by neural network. However, the input rectangular bounding box of the intervertebral disc needs to be segmented artificially or detected automatically using complex algorithms \(^{1,12 - 17}\) . There are also some studies on the quantitative measurement of intervertebral discs based on deep learning, which did not use quantitative data to evaluate intervertebral disc degeneration \(^{18,19}\) .  
+
+In this study, a fully automated deep learning- based lumbar spine segmentation network (LSSN) has been developed at our institution by using a deep convolutional neural network (CNN) with the self- attention skip connection, deep feature extraction module and the corresponding loss function. According to IVD degeneration features (water content loss and height decrease) \(^{20}\) , signal intensity difference and geometric parameters of IVD are calculated and validated with the modified Pfirrmann grading system. Finally, baseline ranges of lumbar IVD parameters among different gender and age and lumbar level was established based on a large population around China for quantitative and structured report. The diagram of this study is illustrated in Fig. 1.  
+
+## Materials And Methods  
+
+## MRI Data Sets  
+
+This study was approved by Institutional Review Board (IRB) in all the participating sites. All retrospective subject data were obtained with a waiver of consent under IRB approval. The data were anonymized before being shared.  
+
+## Data sets for segmentation training (Data set A & B)  
+
+Training and validation of the proposed lumbar spine semantic segmentation method was carried out by performing an institutional review board- approved retrospective analysis of lumbar spine images from
+
+<--- Page Split --->
+
+286 subjects who underwent MR imaging in the Longhua Hospital, Shanghai University of TCM between January 1, 2019, to December 31, 2020. Among these, there're 223 subjects using a 1.5- T MRI unit (MAGNETOM Aera XJ, SIEMENS) and 63 subjects using another 1.5- T MRI unit (MAGNETOM Avanto, SIEMENS), which were trained two separate segmentation networks for different resolution of 512\*512 (Data set A) and 320\*320 (Data set B). Mid- sagittal T2 images of different resolution were exported from Data set A and Data set B respectively, being randomly allocated into each training set or test set (Fig.1). All images in the segmentation data set were labeled by LabelMe (version 3.3.6, CSAIL, Massachusetts Institute of Technology) 21. Based on the structural features mentioned in the modified Pfirrmann grading system, the segmentation area of 14 parts, included 5 vertebral bodies (L1- L5), 5 lumbar IVDs (L1/L2- L5/S1), sacrum (S1), pre- iliac fat area, cerebrospinal fluid area in the spinal canal, and background as Fig. 2a.  
+
+## Data set for quantitative analysis (Data set C)  
+
+The proposed LSSN was used to extracted 1051 lumbar spine images as Data set C in four hospitals around China, including Longhua Hospital, Shanghai University of TCM, Guangdong Provincial Hospital of Chinese Medicine, Shenzhen Pingle Orthopedics Hospital, and Dongzhimen Hospital, Beijing University of Chinese Medicine between January 1, 2019, and March 30, 2021. The imaging parameters of all sites are summarized in Table 1.  
+
+Table 1 Imaging Parameters for the MRI Sequences in the 4 Sites
+
+<--- Page Split --->
+
+
+<table><tr><td>Site</td><td>City</td><td>Strength of the Magnet</td><td>Company</td><td>Model</td><td>Coil</td></tr><tr><td>Longhua Hospital, Shanghai University of TCM</td><td>Shanghai</td><td>1.5-Tesla</td><td>SIEMENS</td><td>MAGNETOM Aera XJ</td><td>18-channel Spine Tim 4G coil</td></tr><tr><td>Guangdong Provincial Hospital of Chinese Medicine</td><td>Guangzhou</td><td>3-Tesla</td><td>SIEMENS</td><td>TIM Systems</td><td>32-channel Spine Tim coil</td></tr><tr><td>Shenzhen Pingle Orthopedics Hospital</td><td>Shenzhen</td><td>1.5-Tesla</td><td>SIEMENS</td><td>MAGNETOM Essenza</td><td>8-channel quadrature body coil</td></tr><tr><td>Dongzhimen Hospital, Beijing University of Chinese Medicine</td><td>Beijing</td><td>1.5-Tesla</td><td>SIEMENS</td><td>MAGNETOM Amira</td><td>24-channel quadrature body coil</td></tr></table>
+
+A research team, composed of a 4-year radiology resident (DW Kong), two 8-year orthopedic resident (J Chen, XF Ma), two 4-year orthopedic resident (YL Sun, YP Lin) and a 2-year orthopedic resident (MC Yin),discussed together for the final segment and Pfirrmann grade for each MR image.
+
+# Lumbar Spine Segmentation from MR Images
+
+# Convolutional Neural Network (CNN) Training
+
+The critical component of LSSN is an improved deeplabv3+ segmentation network22 with backbone ResNet-10123, called BianqueNet. The BianqueNet was built on the basis of deep feature extraction to extract richer semantic information and denser features. An illustration of this semantic segmentation network is shown in Fig. 2. The entire network consists of a swin transform skip connection (ST-SC)module and a deep feature extraction (DFE) module. Swin Transform is a hierarchical transform calculated by shifting the window, which has the advantages of high efficiency and low complexity 24.The skip connections structure designed in this study uses two successive Swin-Transformer blocks, with 1*1 convolutional layers in parallel at the same time, and finally the two output features are spliced.Through the pyramid pooling module, feature information of different depths through pooling operations
+
+<--- Page Split --->
+
+of different scales can be obtained. By repeating check with feature map of 4096 channels multi- scale information \(^{25}\) , the network can achieve efficient features extraction with a dense semantic feature map of 256 channels.  
+
+32- depth BMP images were exported from raw MRI to train the LSSN as input. In the upsampling phase, a modified upsampling operation with a deconvolution decoder was used to recover more detailed features of the segmentation target. In the feature extraction phase, the feature maps of different resolutions were obtained by down- sampling and output to the ST- SC module, which splices images and extracts features from different resolutions. According to feature pyramid \(^{26}\) , feature maps with low- resolution and high- resolution were integrated to extract more semantic and spatial information, in which a \(3 \times 3\) double convolutional layer was used for the fused feature map to improve the feature. Finally, a double upsampling operation was performed to obtain a dense prediction image.  
+
+## Weighted Dice Loss Function  
+
+A weighted dice loss function as below was proposed to enhance segmentation performance by estimating difficulties in difference images with typical or atypical structure, which ensured consistent in segmentation:  
+
+\[L w d i c e = \frac{1}{c}\sum_{j = 1}^{C}\xi_{j}\left(1 - \frac{2\sum_{i = 1}^{N}p_{1i}g_{1i}}{\sum_{i = 1}^{N}p_{1i}g_{1i} + \sum_{i = 1}^{N}p_{0i}g_{1i} + \sum_{i = 1}^{N}p_{1i}g_{0i}}\right) \quad (1)\]  
+
+This formula was used in the output of the softmax layer, where the is the probability of voxel (target) and is the probability of voxel (non- target). So was for and . represents different segmentation areas, represents the total number of channels, which is taken as 14. represent the weight of different segmentation channels. According to the experimental analysis results, channels weight was set to 0.9, 0.8 and 1 for vertebral body, IVD and the other respectively, which may achieve the best segmentation performance.  
+
+For avoiding that the subsequent feature extraction operations are affected, corrosion and expansion operations were used to remove the burrs (Fig. 2b).  
+
+## Lumbar IVD Quantitative Analysis  
+
+## Parameters Calculation based on Pfirrmann Grading System  
+
+Based on previous studies \(^{18,25 - 27}\) , some extraction and calculation methods were modified with histogram features of IVD. signal intensity difference (ΔSI) was obtained to quantify the blurring degree of boundary between NP and AF, which indicating water content in IVD. Average disc height (DH), disc
+
+<--- Page Split --->
+
+height index (DHI), and disc height- to- diameter ratio (DHR) were obtained to quantify structural collapse in IVD degeneration. Specific calculation methods for each parameter are described in the Supplement File 1.  
+
+## Versatility Test for Images with Different Origins  
+
+IVD parameters extracted by LSSN in mid- sagittal lumbar MR images with different resolutions were compared with each other. In the data set B, 46 images with resolution of \(320*320\) were randomly selected to be segmented and quantified by model B. Meanwhile, these images were adjusted to \(512*512\) for segmentation and quantitation by model A. IVD parameters extracted from these two models were used for versatility test. If IVD parameters from LSSN shows good consistency under different origins of imaging, LSSN will be used into a larger population (Data set C) with different machines and models to establish degenerative ranges of IVD parameters in different people.  
+
+## Quantitation for IVD Degeneration  
+
+Relationship between IVD parameters (including \(\Delta \mathrm{SI}\) , DH, DHI, and HDR) and demographic information (including gender, age, and segment) and correlation between IVD parameters and IVD degeneration (based on the modified Pfirrmann grading system) were analyzed respectively in Data set C. Based on these results, a baseline of IVD degeneration in larger population was established, which may indicate a qualitative IVD degeneration in different population with accordance to the modified Pfirrmann grading system. Details in quantitative protocols were shown in the Supplement File 2.  
+
+## Statistical analysis  
+
+The intraclass correlation coefficient (ICC) was used to analyze the consistency between the IVD parameters calculated using the original resolution ( \(320*320\) ) from the data set B and the adjusted resolution ( \(512*512\) ). The macroF1- score and the Kendall correlation coefficient were used to test sensitivity and specificity in IVD degeneration grading performance among deep learning methods and 3 residents in the two data sets with different resolution according to the modified Pfirrmann grading system. An absolute value of r of 0- 0.4 was considered as weak correlation, 0.4- 0.6 as moderate correlation, and greater than 0.6 as strong correlation.  
+
+Spearman rank correlation coefficient between IVD signal intensity and grading score has been calculated via SPSS (version 26, IBM, USA). Multiple regression analysis was performed on IVD quantitative parameters (DH, DHI, HDR) and baseline information (including gender, age, segment) to describe some characters in larger population via Stata (version 15.1, USA).  
+
+## Results
+
+<--- Page Split --->
+
+## Segmentation Performance  
+
+The BianqueNet provided good segmentation performance of IVD- related areas. The mean Dice coefficients (mDice) and mean intersection over Union (mIoU) were \(94.45\%\) and \(89.88\%\) for whole lumbar spine, \(96.71\%\) and \(93.66\%\) for vertebral body, \(94.38\%\) and \(89.43\%\) for IVD. Based on deeplabv3+, Adding the three modules of DFE, ST- SC and FPN improved segmentation significantly as shown in Table 2. The average training time for BianqueNet was 10 hours in each data set. Segmentation of vertebral bodies and IVDs on the mid- sagittal MRI image for a patient took approximately 1 seconds with the trained network.  
+
+Table 2 BianqueNet shows superior segmentation effectiveness demonstrated by the pixel-level Dice and IoU coefficient   
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">DFE</td><td colspan="2">Module</td><td colspan="2">Vertebral body</td><td colspan="2">IVD</td><td colspan="2">Lumbar spine</td></tr><tr><td>ST- SC</td><td>FPN</td><td>mDice</td><td>mIoU</td><td>mDice</td><td>mIoU</td><td>mDice</td><td>mIoV</td></tr><tr><td>DeepLabv3+</td><td></td><td></td><td></td><td>0.9671</td><td>0.9366</td><td>0.9438</td><td>0.8943</td><td>0.9445</td><td>0.8988</td></tr><tr><td>DeepLabv3++DFE</td><td>√</td><td></td><td></td><td>0.9681</td><td>0.9384</td><td>0.9444</td><td>0.8960</td><td>0.9455</td><td>0.9006</td></tr><tr><td>DeepLabv3++DFE+ST-SC</td><td>√</td><td>√</td><td></td><td>0.9692</td><td>0.9405</td><td>0.9458</td><td>0.8982</td><td>0.9468</td><td>0.9028</td></tr><tr><td>DeepLabv3++DFE+ST-SC+FPN (BianqueNet)</td><td>√</td><td>√</td><td>√</td><td>0.9703</td><td>0.9425</td><td>0.9480</td><td>0.9019</td><td>0.9470</td><td>0.9035</td></tr></table>  
+
+## Versatility test for different resolution  
+
+A total of 230 IVDS and 276 vertebral bodies of 46 subjects were segmented after resolution of MRI- exported images had been adjusted from \(320*320\) to \(512*512\) . The results showed a good consistency in using different parameter calculation algorithms for different resolution of MRI- exported images. Among them, the measurement of intervertebral disc geometric parameters DHI and DWR have extremely high ICC values, which are 0.958 (p=0.000) and 0.956 (p=0.000), respectively, and the ICC value of the \(\Delta SI\) is 0.874 (p=0.000), as shown in Table 3.  
+
+Table 3 Consistency analysis of intervertebral disc parameters calculated by MRI of different sizes   
+
+<table><tr><td>Measure</td><td>Intraclass Correlationb</td></tr><tr><td></td><td>ICCa 95%CI</td></tr><tr><td>ΔSI</td><td>.874*** 0.8400.9020</td></tr><tr><td>DHI</td><td>.958*** 0.9430.9680</td></tr><tr><td>HDR</td><td>.956*** 0.8860.9780</td></tr></table>
+
+<--- Page Split --->
+
+Two-way mixed effects model where people effects are random and measures effects are fixed. ICC, intraclass correlation coefficient; 95% CI, 95% confidence interval;  
+
+a. The estimator is the same, whether the interaction effect is present or not.  
+
+b. Type A intraclass correlation coefficients using an absolute agreement definition.  
+
+## Characteristics of IVD Parameters in a Larger Population  
+
+After screening 1508 MRI images in 4 sites around China, a total of 1051 individuals were collected, in which there're 144 excluded for imaging quality and 313 excluded for irregular structures (especially in vertebral bodies). The demographic information (including age and gender) distributed evenly as shown in Table 4, which were integrated to conduct correlation analysis with IVD parameters.  
+
+Table 4 Included Patient Demographic Information from the Four Sites around China   
+
+<table><tr><td rowspan="2">Site</td><td rowspan="2">Number</td><td colspan="6">Age(F/M)</td></tr><tr><td>20-29</td><td>30-39</td><td>40-49</td><td>50-59</td><td>60-69</td><td>70-89</td></tr><tr><td>Longhua Hospital, Shanghai University of TCM</td><td>433</td><td>32/21</td><td>52/51</td><td>49/45</td><td>34/35</td><td>53/39</td><td>12/10</td></tr><tr><td>Shenzhen Pingle Orthopedics Hospital</td><td>222</td><td>16/18</td><td>20/20</td><td>19/20</td><td>18/21</td><td>13/23</td><td>9/25</td></tr><tr><td>Guangdong Provincial Hospital of Chinese Medicine</td><td>246</td><td>19/24</td><td>20/15</td><td>23/17</td><td>22/17</td><td>18/15</td><td>22/34</td></tr><tr><td>Dongzhimen Hospital, Beijing University of Chinese Medicine</td><td>150</td><td>7/8</td><td>13/18</td><td>21/17</td><td>13/8</td><td>12/11</td><td>8/14</td></tr><tr><td>Total</td><td>1051</td><td>74/71</td><td>105/104</td><td>112/99</td><td>87/81</td><td>96/88</td><td>51/83</td></tr></table>  
+
+Supplement Figure (1- 4) and Table 5 shows comprehensive distribution of IVD parameters in a larger population and multiple regression analysis result among IVD parameters and each demographic information respectively. \(\Delta S / \mathrm{in}\) IVDs decreased with age, while DH, DHI and DHR of IVDs increased with age, reaching peak at the age of 50- 60 ( \(P< 0.01\) ). There're no significant different between male and
+
+<--- Page Split --->
+
+female in in IVDs, while DH, DHI and DHR of IVDs were significantly higher in males than those in females \((P< 0.01)\) . In additions, DH, DHI and DHR were significantly higher in lower segmental IVDs (L3-L4, L4-L5 and L5-S1) than upper ones (L1-L2 and L2-L3), and disc height of L4-L5 IVDs was highest \((P< 0.01)\) . In the further analysis of all the IVD height parameters, the influence of segments on the parameters is greater than those of age. For the IVD height, the influence of gender is greater than age. For DHI and HDR, gender and age have similar effects.  
+
+Table 5 The results of multiple regression analysis of signal intensity peak difference, DH, DHI, HDR and gender, different ages, and different disc positions   
+
+<table><tr><td>N16151</td><td>ΔS/</td><td>DH</td><td>DHI</td><td>HDR</td></tr><tr><td>female</td><td>-.0279</td><td>-.2541***</td><td>-.1121***</td><td>.1115***</td></tr><tr><td>male</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td></tr><tr><td>20-30</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.00</td></tr><tr><td>30-40</td><td>-.1669***</td><td>.0796***</td><td>.0557*</td><td>.1100***</td></tr><tr><td>40-50</td><td>-.3802***</td><td>.1110***</td><td>.0927***</td><td>.0980***</td></tr><tr><td>50-60</td><td>-.4826***</td><td>.1612***</td><td>.1577***</td><td>.0440</td></tr><tr><td>60-70</td><td>-.6002***</td><td>.1427***</td><td>.1687***</td><td>.0099</td></tr><tr><td>70-90</td><td>-.5137***</td><td>.0328</td><td>.0806***</td><td>-.0674***</td></tr><tr><td>L1-L2</td><td>.2800***</td><td>-.7181***</td><td>-.6708***</td><td>-.4932***</td></tr><tr><td>L2-L3</td><td>.1719***</td><td>-.3832***</td><td>-.4155***</td><td>-.2912***</td></tr><tr><td>L3-L4</td><td>.0907***</td><td>-.1593***</td><td>-.1942***</td><td>-.1122***</td></tr><tr><td>L4-L5</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td></tr><tr><td>L5-S1</td><td>.1526***</td><td>-.0520**</td><td>-.0312</td><td>.1105***</td></tr></table>
+
+\\*\\*\\* \(\mathsf{p}< 0.01\) \\*\\* \(\mathsf{p}< 0.05\) \\* \(\mathsf{p}< 0.1\)  
+
+## Correlation with IVD Degeneration Grading  
+
+Considering structural collapse with IVD degeneration according to the modified Pfirrmann grading system, a regression analysis was conducted to investigate correlation between its certain grading (For analyzing the correlation between degeneration segments and \(\Delta SI,\) the corresponding grading were 1, 2, 3, 4, and (5- 8). For analyzing the correlation between the degeneration segments and geometric parameters, the corresponding grading are (1- 5), 6, 7, 8)) and IVD parameters in different age, gender and segments. As shown in Table 6, IVD parameters showed a good accordance to the modified Pfirrmann grade.
+
+<--- Page Split --->
+
+Table 6 Correlations between IVD Parameters and Modified Pfirrmann Grading
+
+<table><tr><td rowspan="2">lumbar level</td><td rowspan="2">\(\Delta S/\)</td><td colspan="2">DH</td><td colspan="2">DHI</td><td colspan="2">HDR</td></tr><tr><td>-f</td><td>female</td><td>male</td><td>female</td><td>male</td><td>female</td></tr><tr><td>L1/L2</td><td></td><td>-.421***</td><td>-.296***</td><td>-.304***</td><td>-.235***</td><td>-.473***</td><td>-.397***</td></tr><tr><td>L2/L3</td><td></td><td>-.481***</td><td>-.417***</td><td>-.354***</td><td>-.398***</td><td>-.575***</td><td>-.455***</td></tr><tr><td>L3/L4</td><td></td><td>-.639***</td><td>-.470***</td><td>-.530***</td><td>-.443***</td><td>-.626***</td><td>-.539***</td></tr><tr><td>L4/L5</td><td></td><td>-.656***</td><td>-.696***</td><td>-.560***</td><td>-.665***</td><td>-.709***</td><td>-.758***</td></tr><tr><td>L5/S1</td><td></td><td>-.701***</td><td>-.687***</td><td>-.641***</td><td>-.664***</td><td>-.744***</td><td>-.778***</td></tr></table>
+
+*** p&lt;0.01 ** p&lt;0.05 * p&lt;0.1
+
+r, Spearman rank correlation coefficients
+
+Regarding water content loss with IVD degeneration, result from a further regression analysis showed a stronger correlation between the modified Pfirrmann grade (1,2,3,4, and (5-8)) and \(\Delta S/\) (R=-0.966, \(P=0.000\) ). Specific ranges of according to the modified Pfirrmann grade (1,2,3,4, and (5-8)) were calculated and listed in **Table 7.**
+
+Table 7 Quantitative ranges of $\Delta S/$ according to the modified Pfirrmann Grade (1-8)
+
+<table><tr><td>modified Pfirrmann Grade</td><td>1</td><td>2</td><td>3</td><td>4</td><td>5-8</td></tr><tr><td>Number</td><td>154</td><td>1130</td><td>1622</td><td>1315</td><td>1034</td></tr><tr><td>(mean±SD)</td><td></td><td>121.97±9.96</td><td>95.34±7.20</td><td>72.34±7.81</td><td>44.63±8.49 20.60±9.28</td></tr></table>
+
+According to the results of multiple regression analysis, gender and segements have significant correlations with \(\Delta S/\), while age, gender and segments have significant correlations with geometric parameters. **Fig.3** and **Supplement Table (1-4)** showed comprehensive distribution of IVD parameters in a larger population.
+
+# Discussion
+
+<--- Page Split --->
+
+Our study described a fully automated deep learning–based lumbar IVD quantitative system utilizing a CNN with the self-attention skip connection, deep feature extraction module and the corresponding loss function for segmenting IVD-related areas to extract geometric and signal parameters. The proposed deep learning approach achieved high accuracy for segmentation and measurement. More specifically, our method showed high consistency with the modified Pfirrmann grading system. 
+
+Compared with previously reported conventional image processing methods for lumbar spine MRI, our method is focus on quantitative measurement other than degeneration grade classification. Standard and accurate ranges of in IVD was established to quantify IVD degeneration, which has strong applicability and accuracy for grading IVD degeneration (macroF1: 92.02% and 90.63% in two data sets) as shown in Table 8. 
+
+Table 8 Accuracy of IVD degeneration grading with ΔS/in IVD 
+
+<table><tr><td>Modified Pfirrmann Grade</td><td>1</td><td>2</td><td>3</td><td>4</td><td>5-8</td><td>macro-average (%)</td><td>macroF1(%)</td></tr><tr><td rowspan="2">Data set A</td><td>Precision (%)</td><td>60.76</td><td>97.28</td><td>99.40</td><td>97.89</td><td>89.08</td><td>88.89</td><td>92.02</td></tr><tr><td>Recall (%)</td><td>100</td><td>90.96</td><td>97.84</td><td>90.05</td><td>98.15</td><td>95.40</td><td></td></tr><tr><td rowspan="2">Data set B</td><td>Precision (%)</td><td>/</td><td>81.82</td><td>93.55</td><td>100</td><td>85.71</td><td>90.27</td><td>90.63</td></tr><tr><td>Recall (%)</td><td>/</td><td>90.00</td><td>90.63</td><td>83.33</td><td>100</td><td>90.99</td><td></td></tr></table>
+
+By means of LSSN, all the IVD parameters will be extracted and quantified from MR images in about half of second, which may describe water content loss and structural collapse in IVD, indicating degeneration process. Fig.4 shows a potential application for structural MRI report output from the IVD degeneration quantitative analysis. 
+
+Pfirrmann grading system, as the most used IVD imaging grading method, was designed based on symptomatic patients with an average age of about 40 years old 15. Therefore, its reliability for early IVD degeneration or IVD degeneration in the elderly people may be unsatisfied. This study proposed an automatic quantitative method for IVD degeneration assessment in asymptomatic patients of different ages. In addition, multiple quantitative sequences in imaging are generally used together to accurately evaluate IVD degeneration or lesions, which are too time-consuming to popularize MR imaging quantitative analysis in IVD. Our LSSN may meet both patients' affordability and clinical diagnosis needs. 
+
+Our study has limitations. First, we only included MR images with relatively regular outline in IVD-related areas, most of which were accurately segmented by LSSN. Second, accuracy of LSSN is dependent on the depth of BMP exported from MR imaging system. Third, the subjects included in this study did not take symptoms (such as low back pain) into account, lacking clinical validation on IVD degeneration.
+
+<--- Page Split --->
+
+Finally, as a retrospective study, IVD parameters extracted from Data set C did not exactly represent the real- world setting. Further research is needed to determine the applicability of this LSSN in a prospective multi- institutional study in patients with low back pain.  
+
+In conclusion, we developed a fully automated deep learning- based lumbar spine segmentation network, which demonstrated strong versatility and high reliability to assist residents on IVD degeneration grading by means of IVD degeneration quantitation.  
+
+## Declarations  
+
+## Acknowledged  
+
+This study was supported by the National Natural Science Foundation of China (81930116, 81804115, 81873317, and 81704096).  
+
+## Author contributions  
+
+Guarantor of integrity of entire study, YL Sun, YJ Wang; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, HD Zheng, MC Yin, M Yao, XJ Cui; clinical studies, YL Sun, DW Kong, MC Yin, J Chen, YP Lin, XF Ma; experimental studies, YZ Tian, HS Wang, GJ Yuan; statistical analysis, HD Zheng, M Yao, XJ Cui; and manuscript editing, HD Zheng, YL Sun, YJ Wang.  
+
+## Disclosure of Conflict of Interest  
+
+All author disclosed no relevant relationships.  
+
+## References  
+
+1. Schnake, K. J., Putzier, M., Haas, N. P. & Kandziora, F. Mechanical concepts for disc regeneration. Eur. Spine J. 15, 354-360 (2006).  
+2. Myers, E. R. & Wilson, S. E. Biomechanics of osteoporosis and vertebral fracture. Spine vol. 22, 25S-31S (1997).  
+3. Chu, J. Y., Skrzypiec, D., Pollintine, P. & Adams, M. A. Can compressive stress be measured experimentally within the annulus fibrous of degenerated intervertebral discs? Proc. Inst. Mech. Eng. Part H J. Eng. Med. 222, 161-170 (2008).
+
+<--- Page Split --->
+
+4. Zhao, F. D., Pollintine, P., Hole, B. D., Adams, M. A. & Dolan, P. Vertebral fractures usually affect the cranial endplate because it is thinner and supported by less-dense trabecular bone. Bone 44, 372–379 (2009).  
+
+5. Richardson, S. M. et al. Degenerate Human Nucleus Pulposus Cells Promote Neurite Outgrowth in Neural Cells. PLoS One 7, (2012).  
+
+6. Stefanakis, M. et al. Annulus fissures are mechanically and chemically conducive to the ingrowth of nerves and blood vessels. Spine (Phila. Pa. 1976). 37, 1883–1891 (2012).  
+
+7. Khan, A. N. et al. Inflammatory biomarkers of low back pain and disc degeneration: a review. Ann. N. Y. Acad. Sci. 1410, 68–84 (2017).  
+
+8. Christian W.A. Pfirrmann, Alexander Metzdorf, Achim Elfering, Juerg Hodler, N. B. Effect of Aging and Degeneration on Disc Volume and Shape: A Quantitative Study in Asymptomatic Volunteers. J. Orthop. Res. Sept. 25, 1121–1127 (2007).  
+
+9. Ma, J. et al. Is fractal dimension a reliable imaging biomarker for the quantitative classification of an intervertebral disk? Eur. Spine J. 29, 1175–1180 (2020).  
+
+10. Jarman, J. P. et al. Intervertebral disc height loss demonstrates the threshold of major pathological changes during degeneration. Eur. Spine J. 24, 1944–1950 (2015).  
+
+11. Lecun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).  
+
+12. Jamaludin, A., Kadir, T. & Zisserman, A. SpineNet: Automated classification and evidence visualization in spinal MRIs. Med. Image Anal. 41, 63–73 (2017).  
+
+13. Luoma, K. et al. Low back pain in relation to lumbar disc degeneration. Spine (Phila. Pa. 1976). 25, 487–492 (2000).  
+
+14. Lootus, M., Kadir, T. & Zisserman, A. Automated radiological grading of spinal MRI. Lect. Notes Comput. Vis. Biomech. 20, 119–130 (2015).  
+
+15. Pfirrmann, C. W. A., Metzdorf, A., Zanetti, M., Hodler, J. & Boos, N. Magnetic resonance classification of lumbar intervertebral disc degeneration. Spine (Phila. Pa. 1976). 26, 1873–1878 (2001).  
+
+16. Castro-Mateos, I., Hua, R., Pozo, J. M., Lazary, A. & Frangi, A. F. Intervertebral disc classification by its degree of degeneration from T2-weighted magnetic resonance images. Eur. Spine J. 25, 2721–2727 (2016).  
+
+17. Shinde, J. V., Joshi, Y. V. & Manthalkar, R. R. Multidomain Feature Level Fusion for Classification of Lumbar Intervertebral Disc Using Spine MR Images. IETE J. Res. 0, 1–14 (2020).  
+
+18. Huang, J. et al. Spine Explorer: a deep learning based fully automated program for efficient and reliable quantifications of the vertebrae and discs on sagittal lumbar spine MR images. Spine J. 20, 590–599 (2020).  
+
+19. Pang, S. et al. Direct automated quantitative measurement of spine by cascade amplifier regression network with manifold regularization. Med. Image Anal. 55, 103–115 (2019).  
+
+20. Griffith, J. F. et al. Modified Pfirrmann grading system for lumbar intervertebral disc degeneration. Spine (Phila. Pa. 1976). 32, 708–712 (2007).
+
+<--- Page Split --->
+
+21. Marois, B. & Syssau, P. Pratiques des banques françaises en termes d'analyse du risque-pays. Rev. française Gest. 32, 77-94 (2006).22. Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. Proceedings of the European conference on computer vision (ECCV). 801-818 (2018).23. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. 770-778 (2016).24. Liu, Z. et al. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. arXiv preprint arXiv:2103.14030 (2021).25. Zhao, H., Shi, J., Qi, X., Wang, X. & Jia, J. Pyramid scene parsing network. Proc. - 30th IEEE Conf. Comput. Vis. Pattern Recognition, CVPR 2017. 2881-2890 (2017).26. Li, X. et al. Weighted feature pyramid networks for object detection. Proc. - 2019 IEEE Int'l Conf Parallel Distrib. Process. with Appl. Big Data Cloud Comput. Sustain. Comput. Commun. Soc. Comput. Networking, ISPA/BDCloud/SustainCom/SocialCom. 1500-1504 (2019)27. Waldenberg, C., Hebella, H., Brisby, H. & Lagerstrand, K. M. MRI histogram analysis enables objective and continuous classification of intervertebral disc degeneration. Eur. Spine J. 27, 1042-1048 (2018).28. Abdollah V, Parent EC, Battie MC. Reliability and validity of lumbar disc height quantification methods using magnetic resonance images. Biomed Tech (Berl). 64, 111-117 (2018).29. Dabbs VM, Dabbs LG. Correlation between disc height narrowing and low-back pain. Spine (Phila Pa 1976). 15, 1366-9 (1990).  
+
+## Figures
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1 </center>  
+
+The flowchart of the study process
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2 </center>  
+
+The fully automatic IVD quantitative analysis system based on semantic segmentation network. The proposed method consisted of segmentation CNNs with SW- SC module and DFE module, histogram- based signal intensity quantization, and area- based fully automated geometric measurement. (a) Segmentation label of lumbar spine related area, (b) Each image channel output by the model corresponds to a segmentation area, (c) The outline of the segmented area is displayed on the original
+
+<--- Page Split --->
+
+image. Signal intensity histogram calculation (d) Cerebrospinal fluid area, (e) Presacral fat area, (f) L3L4 intervertebral disc area. (g) Intervertebral disc parameter calculation, (h) Vertebral body corner detection result (red point) and feature point calculation result (green point), (i) 80% area extraction result of the intervertebral disc center.  
+
+![](images/Figure_3.jpg)
+
+<center>Figure 3 </center>  
+
+Characteristics of IVD Parameters ((a) The mean and standard deviation (σ) of the \(\Delta \mathrm{SI}\) of each the Modified Pfirrmann Grading System (level 1, 2, 3, 4, 5), \(\Delta \mathrm{SI}\) (b), DH (c), DHI (d) and HDR (e)) in different age, gender, and segments
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4 </center>  
+
+Quantitative analysis results of typical cases. (a) 23- year- old male(Longhua Hospital); (b) 49- year- old female(Dongzhimen Hospital, Beijing University of Chinese Medicine); (c) 63- year- old male(Guangdong Provincial Hospital of Chinese Medicine); (d) 81- year- old male(Shenzhen Pingle Orthopedics Hospital).  
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- SupplementFigures.pdf- SupplementFile1signalintensityandgeographicmeasurementIVD.docx- SupplementFile2IVDquantitativeanalysiswithPfirmannGrading.docx- SupplementTable1.docx- SupplementTable2.docx
+
+<--- Page Split --->
+
+- SupplementTable3.docx- SupplementTable4.docx
+
+<--- Page Split --->

preprint/preprint__09b35197b5ad4345c614483e5c88cb8e2e52505854eacc518df23d1042b94777/preprint__09b35197b5ad4345c614483e5c88cb8e2e52505854eacc518df23d1042b94777_det.mmd

@@ -0,0 +1,465 @@
+<|ref|>title<|/ref|><|det|>[[44, 106, 945, 241]]<|/det|>
+# Deep Learning Approach for Evaluating Lumbar Intervertebral Disc Degeneration: Achieving High Accurate Segmentation for Quantitative Analysis on MRI  
+
+<|ref|>text<|/ref|><|det|>[[44, 263, 715, 330]]<|/det|>
+Hua- dong Zheng Shanghai University Yue- li Sun ( \(\boxed{\pi}\) yueli_sun@foxmail.com)  
+
+<|ref|>text<|/ref|><|det|>[[44, 332, 715, 375]]<|/det|>
+Longhua Hospital, Shanghai University of Traditional Chinese Medicine De- wei Kong  
+
+<|ref|>text<|/ref|><|det|>[[50, 378, 714, 398]]<|/det|>
+Depart ment of Radiology \(\boxed{\pi}\) Longhua Hospital of Shanghai University of TCM  
+
+<|ref|>text<|/ref|><|det|>[[44, 403, 465, 444]]<|/det|>
+Meng- chen Yin Longhua Hospital, Shanghai University of TCM  
+
+<|ref|>text<|/ref|><|det|>[[44, 449, 600, 490]]<|/det|>
+Jiang Chen Dongzhimen Hospital of BeijingUniversity of Chinese Medicine  
+
+<|ref|>text<|/ref|><|det|>[[44, 495, 586, 537]]<|/det|>
+Yong- peng Lin Dongzhimen Hospital, Beijing University of Chinese Medicine  
+
+<|ref|>text<|/ref|><|det|>[[44, 542, 794, 584]]<|/det|>
+Xue- feng Ma Shenzhen Pingle Orthopedics Hospital (Shenzhen Pingshan District Hospital of TCM)  
+
+<|ref|>text<|/ref|><|det|>[[44, 589, 510, 630]]<|/det|>
+Hong- shen Wang Guangdong Provincial Hospital of Chinese Medicine  
+
+<|ref|>text<|/ref|><|det|>[[44, 635, 231, 676]]<|/det|>
+Guangjie Yuan Shanghai University  
+
+<|ref|>text<|/ref|><|det|>[[44, 681, 465, 723]]<|/det|>
+Min Yao Longhua Hospital, Shanghai University of TCM  
+
+<|ref|>text<|/ref|><|det|>[[44, 728, 465, 769]]<|/det|>
+Xuejun Cui Longhua Hospital, Shanghai University of TCM  
+
+<|ref|>text<|/ref|><|det|>[[44, 774, 231, 815]]<|/det|>
+Yingzhong Tian Shanghai University  
+
+<|ref|>text<|/ref|><|det|>[[44, 820, 794, 862]]<|/det|>
+Yongjun Wang Shanghai University of Traditional Medicine https://orcid.org/0000- 0001- 9333- 2423  
+
+<|ref|>title<|/ref|><|det|>[[44, 902, 101, 920]]<|/det|>
+# Article
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 44, 905, 88]]<|/det|>
+Keywords: lumbar disc degeneration, intervertebral disc degeneration, MRI, deep learning and image processing technology  
+
+<|ref|>text<|/ref|><|det|>[[44, 105, 344, 125]]<|/det|>
+Posted Date: September 2nd, 2021  
+
+<|ref|>text<|/ref|><|det|>[[44, 144, 463, 164]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 864336/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 181, 910, 225]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 259, 945, 303]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on February 11th, 2022. See the published version at https://doi.org/10.1038/s41467- 022- 28387- 5.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 159, 68]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[42, 84, 949, 150]]<|/det|>
+Purpose: Using deep learning and image processing technology, a standardized automatic segmentation and quantitation network of lumbar disc degeneration based on T2MRI was proposed to help residents accurately evaluate the intervertebral disc (IVD) degeneration.  
+
+<|ref|>text<|/ref|><|det|>[[42, 166, 955, 302]]<|/det|>
+Materials and Methods: A semantic segmentation network (BianqueNet) consist of self- attention mechanism skip connection module and deep feature extraction module was proposed to achieve highprecision segmentation of IVD related areas. A quantitative method was used to calculate the signal intensity difference (△SI) in IVD, average disc height (DH), disc height index (DHI), and disc height- to- diameter ratio (DHR). Quantitative ranges for these IVD parameters in a larger population was established among the 1051 MRI images collected from four hospitals around China.  
+
+<|ref|>text<|/ref|><|det|>[[42, 318, 951, 454]]<|/det|>
+Results: The average dice coefficients of BianqueNet for vertebral bodies and intervertebral discs segmentation are \(97.04\%\) and \(94.76\%\) , respectively. This procedure was suitable for different MRI centers and different resolution of lumbar spine T2MRI (ICC \(= .874 \sim .958\) ). These geographic parameters of IVD degeneration have a significant negative correlation with the modified Pfirrmann Grade, while signal intensity in IVD degeneration had excellent reliability according to the modified Pfirrmann Grade (macroF1 \(= 90.63\% \sim 92.02\%\) ).  
+
+<|ref|>text<|/ref|><|det|>[[42, 470, 930, 538]]<|/det|>
+Conclusion: we developed a fully automated deep learning- based lumbar spine segmentation network, which demonstrated strong versatility and high reliability to assist residents on IVD degeneration evaluating by means of IVD degeneration quantitation.  
+
+<|ref|>text<|/ref|><|det|>[[42, 553, 951, 620]]<|/det|>
+Implication for Patient Care: Deep learning- based approaches have the potential to maximize diagnostic performance for detecting disc degeneration and assessing risk of disc herniation while reducing subjectivity, variability, and errors due to distraction and fatigue associated with human interpretation.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 642, 208, 669]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[41, 681, 940, 914]]<|/det|>
+The intervertebral disc (IVD) plays an important role in distributing loads and absorbing shock in the spine, which is comprised of a gel- like nucleus pulposus (NP), collagenous annulus fibrosis (AF) layers, and ring- like cartilaginous endplates (EP). Identifying IVD structural changes, including IVD deformation, NP dehydration and EP ossification, due to chronic degeneration or acute injury, in patients undergoing MRI of the lumbar spine has many important clinical implications \(^{1}\) . It has been determined that IVD degeneration is a consequence of aging. Accumulated compressive overload usually lead to functional fatigue fractures in endplates and subsequently IVD herniation \(^{2 - 4}\) , which may lead to increased inflammation \(^{5}\) , nerve compression \(^{6}\) and release of pain factors \(^{7}\) . Lifestyle modifications and surgical interventions are likely to be most effective for treating IVD degeneration or herniation, but it is more important to initiate screening and prevention during the earliest stages of the disease process.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[40, 44, 944, 253]]<|/det|>
+MRI with morphologic cartilage imaging sequences has been shown to have high specificity but only moderate sensitivity for detecting dehydration and deformation within the IVD degeneration \(^{2,4}\) . Diagnostic performance is highly dependent on the level of reader expertise, and only moderate interobserver agreement between readers has been reported in most studies \(^{2}\) . Quantitative analysis is efficient and comprehensive in evaluating IVD degeneration by measuring the signal intensity and geometric information. Early research on quantitative measurement of intervertebral discs used general image processing programs for manual measurement \(^{8 - 10}\) . However, there is still no universal automatic IVD degeneration analysis tool in this field. The lack of a universal and widely accepted standard definition of IVD degeneration is one of the main reasons.  
+
+<|ref|>text<|/ref|><|det|>[[41, 272, 952, 455]]<|/det|>
+There has been much recent interest in using deep learning methods in medical imaging \(^{11}\) . With the wide- spread application of convolutional neural network classifiers in medical images, many studies use the rectangular box surrounding the lumbar IVD as input, and the corresponding degeneration level as the label to train the classifier for learning degenerative features by neural network. However, the input rectangular bounding box of the intervertebral disc needs to be segmented artificially or detected automatically using complex algorithms \(^{1,12 - 17}\) . There are also some studies on the quantitative measurement of intervertebral discs based on deep learning, which did not use quantitative data to evaluate intervertebral disc degeneration \(^{18,19}\) .  
+
+<|ref|>text<|/ref|><|det|>[[41, 472, 958, 655]]<|/det|>
+In this study, a fully automated deep learning- based lumbar spine segmentation network (LSSN) has been developed at our institution by using a deep convolutional neural network (CNN) with the self- attention skip connection, deep feature extraction module and the corresponding loss function. According to IVD degeneration features (water content loss and height decrease) \(^{20}\) , signal intensity difference and geometric parameters of IVD are calculated and validated with the modified Pfirrmann grading system. Finally, baseline ranges of lumbar IVD parameters among different gender and age and lumbar level was established based on a large population around China for quantitative and structured report. The diagram of this study is illustrated in Fig. 1.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 676, 354, 704]]<|/det|>
+## Materials And Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 718, 262, 747]]<|/det|>
+## MRI Data Sets  
+
+<|ref|>text<|/ref|><|det|>[[43, 763, 952, 829]]<|/det|>
+This study was approved by Institutional Review Board (IRB) in all the participating sites. All retrospective subject data were obtained with a waiver of consent under IRB approval. The data were anonymized before being shared.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 858, 722, 887]]<|/det|>
+## Data sets for segmentation training (Data set A & B)  
+
+<|ref|>text<|/ref|><|det|>[[43, 902, 949, 945]]<|/det|>
+Training and validation of the proposed lumbar spine semantic segmentation method was carried out by performing an institutional review board- approved retrospective analysis of lumbar spine images from
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[40, 45, 951, 295]]<|/det|>
+286 subjects who underwent MR imaging in the Longhua Hospital, Shanghai University of TCM between January 1, 2019, to December 31, 2020. Among these, there're 223 subjects using a 1.5- T MRI unit (MAGNETOM Aera XJ, SIEMENS) and 63 subjects using another 1.5- T MRI unit (MAGNETOM Avanto, SIEMENS), which were trained two separate segmentation networks for different resolution of 512\*512 (Data set A) and 320\*320 (Data set B). Mid- sagittal T2 images of different resolution were exported from Data set A and Data set B respectively, being randomly allocated into each training set or test set (Fig.1). All images in the segmentation data set were labeled by LabelMe (version 3.3.6, CSAIL, Massachusetts Institute of Technology) 21. Based on the structural features mentioned in the modified Pfirrmann grading system, the segmentation area of 14 parts, included 5 vertebral bodies (L1- L5), 5 lumbar IVDs (L1/L2- L5/S1), sacrum (S1), pre- iliac fat area, cerebrospinal fluid area in the spinal canal, and background as Fig. 2a.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 323, 640, 352]]<|/det|>
+## Data set for quantitative analysis (Data set C)  
+
+<|ref|>text<|/ref|><|det|>[[42, 366, 951, 480]]<|/det|>
+The proposed LSSN was used to extracted 1051 lumbar spine images as Data set C in four hospitals around China, including Longhua Hospital, Shanghai University of TCM, Guangdong Provincial Hospital of Chinese Medicine, Shenzhen Pingle Orthopedics Hospital, and Dongzhimen Hospital, Beijing University of Chinese Medicine between January 1, 2019, and March 30, 2021. The imaging parameters of all sites are summarized in Table 1.  
+
+<|ref|>text<|/ref|><|det|>[[44, 496, 608, 518]]<|/det|>
+Table 1 Imaging Parameters for the MRI Sequences in the 4 Sites
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[42, 42, 955, 490]]<|/det|>
+
+<table><tr><td>Site</td><td>City</td><td>Strength of the Magnet</td><td>Company</td><td>Model</td><td>Coil</td></tr><tr><td>Longhua Hospital, Shanghai University of TCM</td><td>Shanghai</td><td>1.5-Tesla</td><td>SIEMENS</td><td>MAGNETOM Aera XJ</td><td>18-channel Spine Tim 4G coil</td></tr><tr><td>Guangdong Provincial Hospital of Chinese Medicine</td><td>Guangzhou</td><td>3-Tesla</td><td>SIEMENS</td><td>TIM Systems</td><td>32-channel Spine Tim coil</td></tr><tr><td>Shenzhen Pingle Orthopedics Hospital</td><td>Shenzhen</td><td>1.5-Tesla</td><td>SIEMENS</td><td>MAGNETOM Essenza</td><td>8-channel quadrature body coil</td></tr><tr><td>Dongzhimen Hospital, Beijing University of Chinese Medicine</td><td>Beijing</td><td>1.5-Tesla</td><td>SIEMENS</td><td>MAGNETOM Amira</td><td>24-channel quadrature body coil</td></tr></table>
+
+<|ref|>text<|/ref|><|det|>[[42, 545, 945, 607]]<|/det|>
+A research team, composed of a 4-year radiology resident (DW Kong), two 8-year orthopedic resident (J Chen, XF Ma), two 4-year orthopedic resident (YL Sun, YP Lin) and a 2-year orthopedic resident (MC Yin),discussed together for the final segment and Pfirrmann grade for each MR image.
+
+<|ref|>title<|/ref|><|det|>[[44, 642, 743, 668]]<|/det|>
+# Lumbar Spine Segmentation from MR Images
+
+<|ref|>title<|/ref|><|det|>[[42, 700, 641, 725]]<|/det|>
+# Convolutional Neural Network (CNN) Training
+
+<|ref|>text<|/ref|><|det|>[[42, 746, 953, 950]]<|/det|>
+The critical component of LSSN is an improved deeplabv3+ segmentation network22 with backbone ResNet-10123, called BianqueNet. The BianqueNet was built on the basis of deep feature extraction to extract richer semantic information and denser features. An illustration of this semantic segmentation network is shown in Fig. 2. The entire network consists of a swin transform skip connection (ST-SC)module and a deep feature extraction (DFE) module. Swin Transform is a hierarchical transform calculated by shifting the window, which has the advantages of high efficiency and low complexity 24.The skip connections structure designed in this study uses two successive Swin-Transformer blocks, with 1*1 convolutional layers in parallel at the same time, and finally the two output features are spliced.Through the pyramid pooling module, feature information of different depths through pooling operations
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 44, 940, 112]]<|/det|>
+of different scales can be obtained. By repeating check with feature map of 4096 channels multi- scale information \(^{25}\) , the network can achieve efficient features extraction with a dense semantic feature map of 256 channels.  
+
+<|ref|>text<|/ref|><|det|>[[41, 130, 955, 312]]<|/det|>
+32- depth BMP images were exported from raw MRI to train the LSSN as input. In the upsampling phase, a modified upsampling operation with a deconvolution decoder was used to recover more detailed features of the segmentation target. In the feature extraction phase, the feature maps of different resolutions were obtained by down- sampling and output to the ST- SC module, which splices images and extracts features from different resolutions. According to feature pyramid \(^{26}\) , feature maps with low- resolution and high- resolution were integrated to extract more semantic and spatial information, in which a \(3 \times 3\) double convolutional layer was used for the fused feature map to improve the feature. Finally, a double upsampling operation was performed to obtain a dense prediction image.  
+
+<|ref|>sub_title<|/ref|><|det|>[[42, 340, 429, 370]]<|/det|>
+## Weighted Dice Loss Function  
+
+<|ref|>text<|/ref|><|det|>[[42, 384, 936, 451]]<|/det|>
+A weighted dice loss function as below was proposed to enhance segmentation performance by estimating difficulties in difference images with typical or atypical structure, which ensured consistent in segmentation:  
+
+<|ref|>equation<|/ref|><|det|>[[60, 476, 840, 530]]<|/det|>
+\[L w d i c e = \frac{1}{c}\sum_{j = 1}^{C}\xi_{j}\left(1 - \frac{2\sum_{i = 1}^{N}p_{1i}g_{1i}}{\sum_{i = 1}^{N}p_{1i}g_{1i} + \sum_{i = 1}^{N}p_{0i}g_{1i} + \sum_{i = 1}^{N}p_{1i}g_{0i}}\right) \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[41, 553, 930, 690]]<|/det|>
+This formula was used in the output of the softmax layer, where the is the probability of voxel (target) and is the probability of voxel (non- target). So was for and . represents different segmentation areas, represents the total number of channels, which is taken as 14. represent the weight of different segmentation channels. According to the experimental analysis results, channels weight was set to 0.9, 0.8 and 1 for vertebral body, IVD and the other respectively, which may achieve the best segmentation performance.  
+
+<|ref|>text<|/ref|><|det|>[[42, 707, 919, 752]]<|/det|>
+For avoiding that the subsequent feature extraction operations are affected, corrosion and expansion operations were used to remove the burrs (Fig. 2b).  
+
+<|ref|>sub_title<|/ref|><|det|>[[42, 780, 551, 813]]<|/det|>
+## Lumbar IVD Quantitative Analysis  
+
+<|ref|>sub_title<|/ref|><|det|>[[45, 839, 830, 869]]<|/det|>
+## Parameters Calculation based on Pfirrmann Grading System  
+
+<|ref|>text<|/ref|><|det|>[[42, 883, 940, 952]]<|/det|>
+Based on previous studies \(^{18,25 - 27}\) , some extraction and calculation methods were modified with histogram features of IVD. signal intensity difference (ΔSI) was obtained to quantify the blurring degree of boundary between NP and AF, which indicating water content in IVD. Average disc height (DH), disc
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 45, 951, 110]]<|/det|>
+height index (DHI), and disc height- to- diameter ratio (DHR) were obtained to quantify structural collapse in IVD degeneration. Specific calculation methods for each parameter are described in the Supplement File 1.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 140, 671, 169]]<|/det|>
+## Versatility Test for Images with Different Origins  
+
+<|ref|>text<|/ref|><|det|>[[42, 183, 953, 342]]<|/det|>
+IVD parameters extracted by LSSN in mid- sagittal lumbar MR images with different resolutions were compared with each other. In the data set B, 46 images with resolution of \(320*320\) were randomly selected to be segmented and quantified by model B. Meanwhile, these images were adjusted to \(512*512\) for segmentation and quantitation by model A. IVD parameters extracted from these two models were used for versatility test. If IVD parameters from LSSN shows good consistency under different origins of imaging, LSSN will be used into a larger population (Data set C) with different machines and models to establish degenerative ranges of IVD parameters in different people.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 370, 490, 399]]<|/det|>
+## Quantitation for IVD Degeneration  
+
+<|ref|>text<|/ref|><|det|>[[42, 413, 940, 549]]<|/det|>
+Relationship between IVD parameters (including \(\Delta \mathrm{SI}\) , DH, DHI, and HDR) and demographic information (including gender, age, and segment) and correlation between IVD parameters and IVD degeneration (based on the modified Pfirrmann grading system) were analyzed respectively in Data set C. Based on these results, a baseline of IVD degeneration in larger population was established, which may indicate a qualitative IVD degeneration in different population with accordance to the modified Pfirrmann grading system. Details in quantitative protocols were shown in the Supplement File 2.  
+
+<|ref|>sub_title<|/ref|><|det|>[[45, 579, 330, 610]]<|/det|>
+## Statistical analysis  
+
+<|ref|>text<|/ref|><|det|>[[42, 624, 953, 783]]<|/det|>
+The intraclass correlation coefficient (ICC) was used to analyze the consistency between the IVD parameters calculated using the original resolution ( \(320*320\) ) from the data set B and the adjusted resolution ( \(512*512\) ). The macroF1- score and the Kendall correlation coefficient were used to test sensitivity and specificity in IVD degeneration grading performance among deep learning methods and 3 residents in the two data sets with different resolution according to the modified Pfirrmann grading system. An absolute value of r of 0- 0.4 was considered as weak correlation, 0.4- 0.6 as moderate correlation, and greater than 0.6 as strong correlation.  
+
+<|ref|>text<|/ref|><|det|>[[44, 799, 925, 888]]<|/det|>
+Spearman rank correlation coefficient between IVD signal intensity and grading score has been calculated via SPSS (version 26, IBM, USA). Multiple regression analysis was performed on IVD quantitative parameters (DH, DHI, HDR) and baseline information (including gender, age, segment) to describe some characters in larger population via Stata (version 15.1, USA).  
+
+<|ref|>sub_title<|/ref|><|det|>[[45, 911, 144, 936]]<|/det|>
+## Results
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 44, 468, 75]]<|/det|>
+## Segmentation Performance  
+
+<|ref|>text<|/ref|><|det|>[[41, 89, 950, 245]]<|/det|>
+The BianqueNet provided good segmentation performance of IVD- related areas. The mean Dice coefficients (mDice) and mean intersection over Union (mIoU) were \(94.45\%\) and \(89.88\%\) for whole lumbar spine, \(96.71\%\) and \(93.66\%\) for vertebral body, \(94.38\%\) and \(89.43\%\) for IVD. Based on deeplabv3+, Adding the three modules of DFE, ST- SC and FPN improved segmentation significantly as shown in Table 2. The average training time for BianqueNet was 10 hours in each data set. Segmentation of vertebral bodies and IVDs on the mid- sagittal MRI image for a patient took approximately 1 seconds with the trained network.  
+
+<|ref|>table<|/ref|><|det|>[[44, 325, 886, 481]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[41, 263, 950, 305]]<|/det|>
+Table 2 BianqueNet shows superior segmentation effectiveness demonstrated by the pixel-level Dice and IoU coefficient   
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">DFE</td><td colspan="2">Module</td><td colspan="2">Vertebral body</td><td colspan="2">IVD</td><td colspan="2">Lumbar spine</td></tr><tr><td>ST- SC</td><td>FPN</td><td>mDice</td><td>mIoU</td><td>mDice</td><td>mIoU</td><td>mDice</td><td>mIoV</td></tr><tr><td>DeepLabv3+</td><td></td><td></td><td></td><td>0.9671</td><td>0.9366</td><td>0.9438</td><td>0.8943</td><td>0.9445</td><td>0.8988</td></tr><tr><td>DeepLabv3++DFE</td><td>√</td><td></td><td></td><td>0.9681</td><td>0.9384</td><td>0.9444</td><td>0.8960</td><td>0.9455</td><td>0.9006</td></tr><tr><td>DeepLabv3++DFE+ST-SC</td><td>√</td><td>√</td><td></td><td>0.9692</td><td>0.9405</td><td>0.9458</td><td>0.8982</td><td>0.9468</td><td>0.9028</td></tr><tr><td>DeepLabv3++DFE+ST-SC+FPN (BianqueNet)</td><td>√</td><td>√</td><td>√</td><td>0.9703</td><td>0.9425</td><td>0.9480</td><td>0.9019</td><td>0.9470</td><td>0.9035</td></tr></table>  
+
+<|ref|>sub_title<|/ref|><|det|>[[43, 518, 613, 550]]<|/det|>
+## Versatility test for different resolution  
+
+<|ref|>text<|/ref|><|det|>[[41, 563, 953, 699]]<|/det|>
+A total of 230 IVDS and 276 vertebral bodies of 46 subjects were segmented after resolution of MRI- exported images had been adjusted from \(320*320\) to \(512*512\) . The results showed a good consistency in using different parameter calculation algorithms for different resolution of MRI- exported images. Among them, the measurement of intervertebral disc geometric parameters DHI and DWR have extremely high ICC values, which are 0.958 (p=0.000) and 0.956 (p=0.000), respectively, and the ICC value of the \(\Delta SI\) is 0.874 (p=0.000), as shown in Table 3.  
+
+<|ref|>table<|/ref|><|det|>[[331, 751, 666, 937]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[43, 715, 894, 737]]<|/det|>
+Table 3 Consistency analysis of intervertebral disc parameters calculated by MRI of different sizes   
+
+<table><tr><td>Measure</td><td>Intraclass Correlationb</td></tr><tr><td></td><td>ICCa 95%CI</td></tr><tr><td>ΔSI</td><td>.874*** 0.8400.9020</td></tr><tr><td>DHI</td><td>.958*** 0.9430.9680</td></tr><tr><td>HDR</td><td>.956*** 0.8860.9780</td></tr></table>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 46, 910, 88]]<|/det|>
+Two-way mixed effects model where people effects are random and measures effects are fixed. ICC, intraclass correlation coefficient; 95% CI, 95% confidence interval;  
+
+<|ref|>text<|/ref|><|det|>[[42, 106, 697, 127]]<|/det|>
+a. The estimator is the same, whether the interaction effect is present or not.  
+
+<|ref|>text<|/ref|><|det|>[[42, 144, 758, 165]]<|/det|>
+b. Type A intraclass correlation coefficients using an absolute agreement definition.  
+
+<|ref|>sub_title<|/ref|><|det|>[[42, 194, 904, 226]]<|/det|>
+## Characteristics of IVD Parameters in a Larger Population  
+
+<|ref|>text<|/ref|><|det|>[[42, 240, 941, 330]]<|/det|>
+After screening 1508 MRI images in 4 sites around China, a total of 1051 individuals were collected, in which there're 144 excluded for imaging quality and 313 excluded for irregular structures (especially in vertebral bodies). The demographic information (including age and gender) distributed evenly as shown in Table 4, which were integrated to conduct correlation analysis with IVD parameters.  
+
+<|ref|>table<|/ref|><|det|>[[40, 380, 956, 805]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[42, 347, 775, 368]]<|/det|>
+Table 4 Included Patient Demographic Information from the Four Sites around China   
+
+<table><tr><td rowspan="2">Site</td><td rowspan="2">Number</td><td colspan="6">Age(F/M)</td></tr><tr><td>20-29</td><td>30-39</td><td>40-49</td><td>50-59</td><td>60-69</td><td>70-89</td></tr><tr><td>Longhua Hospital, Shanghai University of TCM</td><td>433</td><td>32/21</td><td>52/51</td><td>49/45</td><td>34/35</td><td>53/39</td><td>12/10</td></tr><tr><td>Shenzhen Pingle Orthopedics Hospital</td><td>222</td><td>16/18</td><td>20/20</td><td>19/20</td><td>18/21</td><td>13/23</td><td>9/25</td></tr><tr><td>Guangdong Provincial Hospital of Chinese Medicine</td><td>246</td><td>19/24</td><td>20/15</td><td>23/17</td><td>22/17</td><td>18/15</td><td>22/34</td></tr><tr><td>Dongzhimen Hospital, Beijing University of Chinese Medicine</td><td>150</td><td>7/8</td><td>13/18</td><td>21/17</td><td>13/8</td><td>12/11</td><td>8/14</td></tr><tr><td>Total</td><td>1051</td><td>74/71</td><td>105/104</td><td>112/99</td><td>87/81</td><td>96/88</td><td>51/83</td></tr></table>  
+
+<|ref|>text<|/ref|><|det|>[[42, 854, 940, 944]]<|/det|>
+Supplement Figure (1- 4) and Table 5 shows comprehensive distribution of IVD parameters in a larger population and multiple regression analysis result among IVD parameters and each demographic information respectively. \(\Delta S / \mathrm{in}\) IVDs decreased with age, while DH, DHI and DHR of IVDs increased with age, reaching peak at the age of 50- 60 ( \(P< 0.01\) ). There're no significant different between male and
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[41, 45, 955, 179]]<|/det|>
+female in in IVDs, while DH, DHI and DHR of IVDs were significantly higher in males than those in females \((P< 0.01)\) . In additions, DH, DHI and DHR were significantly higher in lower segmental IVDs (L3-L4, L4-L5 and L5-S1) than upper ones (L1-L2 and L2-L3), and disc height of L4-L5 IVDs was highest \((P< 0.01)\) . In the further analysis of all the IVD height parameters, the influence of segments on the parameters is greater than those of age. For the IVD height, the influence of gender is greater than age. For DHI and HDR, gender and age have similar effects.  
+
+<|ref|>table<|/ref|><|det|>[[225, 253, 774, 696]]<|/det|>
+<|ref|>table_caption<|/ref|><|det|>[[41, 196, 941, 240]]<|/det|>
+Table 5 The results of multiple regression analysis of signal intensity peak difference, DH, DHI, HDR and gender, different ages, and different disc positions   
+
+<table><tr><td>N16151</td><td>ΔS/</td><td>DH</td><td>DHI</td><td>HDR</td></tr><tr><td>female</td><td>-.0279</td><td>-.2541***</td><td>-.1121***</td><td>.1115***</td></tr><tr><td>male</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td></tr><tr><td>20-30</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.00</td></tr><tr><td>30-40</td><td>-.1669***</td><td>.0796***</td><td>.0557*</td><td>.1100***</td></tr><tr><td>40-50</td><td>-.3802***</td><td>.1110***</td><td>.0927***</td><td>.0980***</td></tr><tr><td>50-60</td><td>-.4826***</td><td>.1612***</td><td>.1577***</td><td>.0440</td></tr><tr><td>60-70</td><td>-.6002***</td><td>.1427***</td><td>.1687***</td><td>.0099</td></tr><tr><td>70-90</td><td>-.5137***</td><td>.0328</td><td>.0806***</td><td>-.0674***</td></tr><tr><td>L1-L2</td><td>.2800***</td><td>-.7181***</td><td>-.6708***</td><td>-.4932***</td></tr><tr><td>L2-L3</td><td>.1719***</td><td>-.3832***</td><td>-.4155***</td><td>-.2912***</td></tr><tr><td>L3-L4</td><td>.0907***</td><td>-.1593***</td><td>-.1942***</td><td>-.1122***</td></tr><tr><td>L4-L5</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td></tr><tr><td>L5-S1</td><td>.1526***</td><td>-.0520**</td><td>-.0312</td><td>.1105***</td></tr></table>
+
+<|ref|>table_footnote<|/ref|><|det|>[[44, 696, 287, 714]]<|/det|>
+\\*\\*\\* \(\mathsf{p}< 0.01\) \\*\\* \(\mathsf{p}< 0.05\) \\* \(\mathsf{p}< 0.1\)  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 745, 684, 777]]<|/det|>
+## Correlation with IVD Degeneration Grading  
+
+<|ref|>text<|/ref|><|det|>[[42, 792, 957, 927]]<|/det|>
+Considering structural collapse with IVD degeneration according to the modified Pfirrmann grading system, a regression analysis was conducted to investigate correlation between its certain grading (For analyzing the correlation between degeneration segments and \(\Delta SI,\) the corresponding grading were 1, 2, 3, 4, and (5- 8). For analyzing the correlation between the degeneration segments and geometric parameters, the corresponding grading are (1- 5), 6, 7, 8)) and IVD parameters in different age, gender and segments. As shown in Table 6, IVD parameters showed a good accordance to the modified Pfirrmann grade.
+
+<--- Page Split --->
+<|ref|>table_caption<|/ref|><|det|>[[42, 45, 720, 63]]<|/det|>
+Table 6 Correlations between IVD Parameters and Modified Pfirrmann Grading
+
+<|ref|>table<|/ref|><|det|>[[111, 80, 886, 327]]<|/det|>
+
+<table><tr><td rowspan="2">lumbar level</td><td rowspan="2">\(\Delta S/\)</td><td colspan="2">DH</td><td colspan="2">DHI</td><td colspan="2">HDR</td></tr><tr><td>-f</td><td>female</td><td>male</td><td>female</td><td>male</td><td>female</td></tr><tr><td>L1/L2</td><td></td><td>-.421***</td><td>-.296***</td><td>-.304***</td><td>-.235***</td><td>-.473***</td><td>-.397***</td></tr><tr><td>L2/L3</td><td></td><td>-.481***</td><td>-.417***</td><td>-.354***</td><td>-.398***</td><td>-.575***</td><td>-.455***</td></tr><tr><td>L3/L4</td><td></td><td>-.639***</td><td>-.470***</td><td>-.530***</td><td>-.443***</td><td>-.626***</td><td>-.539***</td></tr><tr><td>L4/L5</td><td></td><td>-.656***</td><td>-.696***</td><td>-.560***</td><td>-.665***</td><td>-.709***</td><td>-.758***</td></tr><tr><td>L5/S1</td><td></td><td>-.701***</td><td>-.687***</td><td>-.641***</td><td>-.664***</td><td>-.744***</td><td>-.778***</td></tr></table>
+
+<|ref|>text<|/ref|><|det|>[[42, 343, 287, 359]]<|/det|>
+*** p&lt;0.01 ** p&lt;0.05 * p&lt;0.1
+
+<|ref|>text<|/ref|><|det|>[[42, 382, 395, 397]]<|/det|>
+r, Spearman rank correlation coefficients
+
+<|ref|>text<|/ref|><|det|>[[42, 458, 936, 543]]<|/det|>
+Regarding water content loss with IVD degeneration, result from a further regression analysis showed a stronger correlation between the modified Pfirrmann grade (1,2,3,4, and (5-8)) and \(\Delta S/\) (R=-0.966, \(P=0.000\) ). Specific ranges of according to the modified Pfirrmann grade (1,2,3,4, and (5-8)) were calculated and listed in **Table 7.**
+
+<|ref|>table_caption<|/ref|><|det|>[[42, 563, 760, 580]]<|/det|>
+Table 7 Quantitative ranges of $\Delta S/$ according to the modified Pfirrmann Grade (1-8)
+
+<|ref|>table<|/ref|><|det|>[[47, 597, 952, 754]]<|/det|>
+
+<table><tr><td>modified Pfirrmann Grade</td><td>1</td><td>2</td><td>3</td><td>4</td><td>5-8</td></tr><tr><td>Number</td><td>154</td><td>1130</td><td>1622</td><td>1315</td><td>1034</td></tr><tr><td>(mean±SD)</td><td></td><td>121.97±9.96</td><td>95.34±7.20</td><td>72.34±7.81</td><td>44.63±8.49 20.60±9.28</td></tr></table>
+
+<|ref|>text<|/ref|><|det|>[[42, 795, 953, 878]]<|/det|>
+According to the results of multiple regression analysis, gender and segements have significant correlations with \(\Delta S/\), while age, gender and segments have significant correlations with geometric parameters. **Fig.3** and **Supplement Table (1-4)** showed comprehensive distribution of IVD parameters in a larger population.
+
+<|ref|>title<|/ref|><|det|>[[42, 905, 192, 926]]<|/det|>
+# Discussion
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 45, 936, 156]]<|/det|>
+Our study described a fully automated deep learning–based lumbar IVD quantitative system utilizing a CNN with the self-attention skip connection, deep feature extraction module and the corresponding loss function for segmenting IVD-related areas to extract geometric and signal parameters. The proposed deep learning approach achieved high accuracy for segmentation and measurement. More specifically, our method showed high consistency with the modified Pfirrmann grading system. 
+
+<|ref|>text<|/ref|><|det|>[[42, 177, 944, 284]]<|/det|>
+Compared with previously reported conventional image processing methods for lumbar spine MRI, our method is focus on quantitative measurement other than degeneration grade classification. Standard and accurate ranges of in IVD was established to quantify IVD degeneration, which has strong applicability and accuracy for grading IVD degeneration (macroF1: 92.02% and 90.63% in two data sets) as shown in Table 8. 
+
+<|ref|>table_caption<|/ref|><|det|>[[42, 304, 576, 323]]<|/det|>
+Table 8 Accuracy of IVD degeneration grading with ΔS/in IVD 
+
+<|ref|>table<|/ref|><|det|>[[42, 339, 952, 544]]<|/det|>
+<table><tr><td>Modified Pfirrmann Grade</td><td>1</td><td>2</td><td>3</td><td>4</td><td>5-8</td><td>macro-average (%)</td><td>macroF1(%)</td></tr><tr><td rowspan="2">Data set A</td><td>Precision (%)</td><td>60.76</td><td>97.28</td><td>99.40</td><td>97.89</td><td>89.08</td><td>88.89</td><td>92.02</td></tr><tr><td>Recall (%)</td><td>100</td><td>90.96</td><td>97.84</td><td>90.05</td><td>98.15</td><td>95.40</td><td></td></tr><tr><td rowspan="2">Data set B</td><td>Precision (%)</td><td>/</td><td>81.82</td><td>93.55</td><td>100</td><td>85.71</td><td>90.27</td><td>90.63</td></tr><tr><td>Recall (%)</td><td>/</td><td>90.00</td><td>90.63</td><td>83.33</td><td>100</td><td>90.99</td><td></td></tr></table>
+
+<|ref|>text<|/ref|><|det|>[[42, 583, 950, 669]]<|/det|>
+By means of LSSN, all the IVD parameters will be extracted and quantified from MR images in about half of second, which may describe water content loss and structural collapse in IVD, indicating degeneration process. Fig.4 shows a potential application for structural MRI report output from the IVD degeneration quantitative analysis. 
+
+<|ref|>text<|/ref|><|det|>[[42, 689, 950, 847]]<|/det|>
+Pfirrmann grading system, as the most used IVD imaging grading method, was designed based on symptomatic patients with an average age of about 40 years old 15. Therefore, its reliability for early IVD degeneration or IVD degeneration in the elderly people may be unsatisfied. This study proposed an automatic quantitative method for IVD degeneration assessment in asymptomatic patients of different ages. In addition, multiple quantitative sequences in imaging are generally used together to accurately evaluate IVD degeneration or lesions, which are too time-consuming to popularize MR imaging quantitative analysis in IVD. Our LSSN may meet both patients' affordability and clinical diagnosis needs. 
+
+<|ref|>text<|/ref|><|det|>[[42, 867, 940, 954]]<|/det|>
+Our study has limitations. First, we only included MR images with relatively regular outline in IVD-related areas, most of which were accurately segmented by LSSN. Second, accuracy of LSSN is dependent on the depth of BMP exported from MR imaging system. Third, the subjects included in this study did not take symptoms (such as low back pain) into account, lacking clinical validation on IVD degeneration.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 45, 944, 111]]<|/det|>
+Finally, as a retrospective study, IVD parameters extracted from Data set C did not exactly represent the real- world setting. Further research is needed to determine the applicability of this LSSN in a prospective multi- institutional study in patients with low back pain.  
+
+<|ref|>text<|/ref|><|det|>[[42, 128, 949, 194]]<|/det|>
+In conclusion, we developed a fully automated deep learning- based lumbar spine segmentation network, which demonstrated strong versatility and high reliability to assist residents on IVD degeneration grading by means of IVD degeneration quantitation.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 216, 213, 242]]<|/det|>
+## Declarations  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 257, 283, 291]]<|/det|>
+## Acknowledged  
+
+<|ref|>text<|/ref|><|det|>[[42, 307, 932, 350]]<|/det|>
+This study was supported by the National Natural Science Foundation of China (81930116, 81804115, 81873317, and 81704096).  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 382, 406, 415]]<|/det|>
+## Author contributions  
+
+<|ref|>text<|/ref|><|det|>[[41, 431, 940, 590]]<|/det|>
+Guarantor of integrity of entire study, YL Sun, YJ Wang; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, HD Zheng, MC Yin, M Yao, XJ Cui; clinical studies, YL Sun, DW Kong, MC Yin, J Chen, YP Lin, XF Ma; experimental studies, YZ Tian, HS Wang, GJ Yuan; statistical analysis, HD Zheng, M Yao, XJ Cui; and manuscript editing, HD Zheng, YL Sun, YJ Wang.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 619, 613, 654]]<|/det|>
+## Disclosure of Conflict of Interest  
+
+<|ref|>text<|/ref|><|det|>[[44, 671, 438, 691]]<|/det|>
+All author disclosed no relevant relationships.  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 715, 195, 740]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[57, 757, 951, 922]]<|/det|>
+1. Schnake, K. J., Putzier, M., Haas, N. P. & Kandziora, F. Mechanical concepts for disc regeneration. Eur. Spine J. 15, 354-360 (2006).  
+2. Myers, E. R. & Wilson, S. E. Biomechanics of osteoporosis and vertebral fracture. Spine vol. 22, 25S-31S (1997).  
+3. Chu, J. Y., Skrzypiec, D., Pollintine, P. & Adams, M. A. Can compressive stress be measured experimentally within the annulus fibrous of degenerated intervertebral discs? Proc. Inst. Mech. Eng. Part H J. Eng. Med. 222, 161-170 (2008).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[57, 44, 930, 110]]<|/det|>
+4. Zhao, F. D., Pollintine, P., Hole, B. D., Adams, M. A. & Dolan, P. Vertebral fractures usually affect the cranial endplate because it is thinner and supported by less-dense trabecular bone. Bone 44, 372–379 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[57, 116, 930, 160]]<|/det|>
+5. Richardson, S. M. et al. Degenerate Human Nucleus Pulposus Cells Promote Neurite Outgrowth in Neural Cells. PLoS One 7, (2012).  
+
+<|ref|>text<|/ref|><|det|>[[57, 166, 944, 211]]<|/det|>
+6. Stefanakis, M. et al. Annulus fissures are mechanically and chemically conducive to the ingrowth of nerves and blood vessels. Spine (Phila. Pa. 1976). 37, 1883–1891 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[57, 216, 945, 260]]<|/det|>
+7. Khan, A. N. et al. Inflammatory biomarkers of low back pain and disc degeneration: a review. Ann. N. Y. Acad. Sci. 1410, 68–84 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[57, 265, 945, 333]]<|/det|>
+8. Christian W.A. Pfirrmann, Alexander Metzdorf, Achim Elfering, Juerg Hodler, N. B. Effect of Aging and Degeneration on Disc Volume and Shape: A Quantitative Study in Asymptomatic Volunteers. J. Orthop. Res. Sept. 25, 1121–1127 (2007).  
+
+<|ref|>text<|/ref|><|det|>[[57, 337, 951, 382]]<|/det|>
+9. Ma, J. et al. Is fractal dimension a reliable imaging biomarker for the quantitative classification of an intervertebral disk? Eur. Spine J. 29, 1175–1180 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 386, 936, 431]]<|/det|>
+10. Jarman, J. P. et al. Intervertebral disc height loss demonstrates the threshold of major pathological changes during degeneration. Eur. Spine J. 24, 1944–1950 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[55, 435, 745, 457]]<|/det|>
+11. Lecun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[55, 462, 855, 506]]<|/det|>
+12. Jamaludin, A., Kadir, T. & Zisserman, A. SpineNet: Automated classification and evidence visualization in spinal MRIs. Med. Image Anal. 41, 63–73 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[55, 511, 933, 555]]<|/det|>
+13. Luoma, K. et al. Low back pain in relation to lumbar disc degeneration. Spine (Phila. Pa. 1976). 25, 487–492 (2000).  
+
+<|ref|>text<|/ref|><|det|>[[55, 560, 904, 604]]<|/det|>
+14. Lootus, M., Kadir, T. & Zisserman, A. Automated radiological grading of spinal MRI. Lect. Notes Comput. Vis. Biomech. 20, 119–130 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[55, 609, 940, 654]]<|/det|>
+15. Pfirrmann, C. W. A., Metzdorf, A., Zanetti, M., Hodler, J. & Boos, N. Magnetic resonance classification of lumbar intervertebral disc degeneration. Spine (Phila. Pa. 1976). 26, 1873–1878 (2001).  
+
+<|ref|>text<|/ref|><|det|>[[55, 658, 945, 725]]<|/det|>
+16. Castro-Mateos, I., Hua, R., Pozo, J. M., Lazary, A. & Frangi, A. F. Intervertebral disc classification by its degree of degeneration from T2-weighted magnetic resonance images. Eur. Spine J. 25, 2721–2727 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[55, 729, 935, 775]]<|/det|>
+17. Shinde, J. V., Joshi, Y. V. & Manthalkar, R. R. Multidomain Feature Level Fusion for Classification of Lumbar Intervertebral Disc Using Spine MR Images. IETE J. Res. 0, 1–14 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 780, 940, 848]]<|/det|>
+18. Huang, J. et al. Spine Explorer: a deep learning based fully automated program for efficient and reliable quantifications of the vertebrae and discs on sagittal lumbar spine MR images. Spine J. 20, 590–599 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[55, 852, 940, 897]]<|/det|>
+19. Pang, S. et al. Direct automated quantitative measurement of spine by cascade amplifier regression network with manifold regularization. Med. Image Anal. 55, 103–115 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[55, 901, 920, 946]]<|/det|>
+20. Griffith, J. F. et al. Modified Pfirrmann grading system for lumbar intervertebral disc degeneration. Spine (Phila. Pa. 1976). 32, 708–712 (2007).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[45, 44, 936, 560]]<|/det|>
+21. Marois, B. & Syssau, P. Pratiques des banques françaises en termes d'analyse du risque-pays. Rev. française Gest. 32, 77-94 (2006).22. Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. Proceedings of the European conference on computer vision (ECCV). 801-818 (2018).23. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. 770-778 (2016).24. Liu, Z. et al. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. arXiv preprint arXiv:2103.14030 (2021).25. Zhao, H., Shi, J., Qi, X., Wang, X. & Jia, J. Pyramid scene parsing network. Proc. - 30th IEEE Conf. Comput. Vis. Pattern Recognition, CVPR 2017. 2881-2890 (2017).26. Li, X. et al. Weighted feature pyramid networks for object detection. Proc. - 2019 IEEE Int'l Conf Parallel Distrib. Process. with Appl. Big Data Cloud Comput. Sustain. Comput. Commun. Soc. Comput. Networking, ISPA/BDCloud/SustainCom/SocialCom. 1500-1504 (2019)27. Waldenberg, C., Hebella, H., Brisby, H. & Lagerstrand, K. M. MRI histogram analysis enables objective and continuous classification of intervertebral disc degeneration. Eur. Spine J. 27, 1042-1048 (2018).28. Abdollah V, Parent EC, Battie MC. Reliability and validity of lumbar disc height quantification methods using magnetic resonance images. Biomed Tech (Berl). 64, 111-117 (2018).29. Dabbs VM, Dabbs LG. Correlation between disc height narrowing and low-back pain. Spine (Phila Pa 1976). 15, 1366-9 (1990).  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 575, 143, 600]]<|/det|>
+## Figures
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[68, 55, 910, 650]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[44, 694, 115, 712]]<|/det|>
+<center>Figure 1 </center>  
+
+<|ref|>text<|/ref|><|det|>[[44, 737, 346, 756]]<|/det|>
+The flowchart of the study process
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[40, 40, 825, 777]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[44, 800, 118, 820]]<|/det|>
+<center>Figure 2 </center>  
+
+<|ref|>text<|/ref|><|det|>[[42, 841, 925, 955]]<|/det|>
+The fully automatic IVD quantitative analysis system based on semantic segmentation network. The proposed method consisted of segmentation CNNs with SW- SC module and DFE module, histogram- based signal intensity quantization, and area- based fully automated geometric measurement. (a) Segmentation label of lumbar spine related area, (b) Each image channel output by the model corresponds to a segmentation area, (c) The outline of the segmented area is displayed on the original
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[41, 45, 944, 133]]<|/det|>
+image. Signal intensity histogram calculation (d) Cerebrospinal fluid area, (e) Presacral fat area, (f) L3L4 intervertebral disc area. (g) Intervertebral disc parameter calculation, (h) Vertebral body corner detection result (red point) and feature point calculation result (green point), (i) 80% area extraction result of the intervertebral disc center.  
+
+<|ref|>image<|/ref|><|det|>[[45, 135, 945, 308]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[43, 346, 117, 365]]<|/det|>
+<center>Figure 3 </center>  
+
+<|ref|>text<|/ref|><|det|>[[42, 387, 933, 455]]<|/det|>
+Characteristics of IVD Parameters ((a) The mean and standard deviation (σ) of the \(\Delta \mathrm{SI}\) of each the Modified Pfirrmann Grading System (level 1, 2, 3, 4, 5), \(\Delta \mathrm{SI}\) (b), DH (c), DHI (d) and HDR (e)) in different age, gender, and segments
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[55, 50, 936, 565]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[44, 593, 118, 613]]<|/det|>
+<center>Figure 4 </center>  
+
+<|ref|>text<|/ref|><|det|>[[42, 635, 933, 701]]<|/det|>
+Quantitative analysis results of typical cases. (a) 23- year- old male(Longhua Hospital); (b) 49- year- old female(Dongzhimen Hospital, Beijing University of Chinese Medicine); (c) 63- year- old male(Guangdong Provincial Hospital of Chinese Medicine); (d) 81- year- old male(Shenzhen Pingle Orthopedics Hospital).  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 723, 311, 751]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 774, 765, 795]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 812, 700, 939]]<|/det|>
+- SupplementFigures.pdf- SupplementFile1signalintensityandgeographicmeasurementIVD.docx- SupplementFile2IVDquantitativeanalysiswithPfirmannGrading.docx- SupplementTable1.docx- SupplementTable2.docx
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 46, 297, 92]]<|/det|>
+- SupplementTable3.docx- SupplementTable4.docx
+
+<--- Page Split --->

preprint/preprint__09bb925dea8007aaf7d14d8f79096de27fcb7fe802393ba1d044ff47a75f57c2/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1. Synthesis ZIF films from ultradilute solutions. a, Schematic of the ZIF film synthesis approach. b, Composition diagram comparing the precursor solution composition used in this study with those reported in the literature (Supplementary Table 1). AFM (c) and the corresponding height profile (d) of a monolayer ZIF film on HOPG. e, Monolayer and multilayer ZIF films on HOPG with discrete thicknesses as a function of synthesis time. Error",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        80,
+        875,
+        770
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2. Structure determination of 2DZIF films. a, Bright-field TEM image of the 2DZIF film supported on suspended graphene, and (b) its corresponding SAED pattern. The pattern from graphene is identified with green circles and those from 2DZIF with white circles. c, In-plane GIXRD data from an aZIF film (top) and a 2DZIF film (middle) prepared on \\(\\mathrm{Si / SiO_2}\\) and graphene/Si/SiO2, respectively, along with a radially integrated trace (bottom) of the SAED pattern shown in (B). d, N1s XPS spectra from ZIF-8, ZIF-L, aZIF and 2DZIF films. The N-Zn and N-H coordination environments are shown on the right. e, DFT-relaxed structure of the 2DZIF and a visualization of the 6-MR (right). f, HRTEM image of the 2DZIF film lying flat on the \\(hk0\\) plane, resting on suspended graphene, and (g) corresponding Fourier transform compared with the simulated diffraction pattern from the proposed structure oriented along the \\(c\\) -out-of-plane direction. h, Left: CTF-corrected image of the highlighted area in (f) based on a defocus value of -130 nm analyzed from the Thon rings in the Fourier transform pattern. Right: simulated projected potential map along the [001] direction of 2DZIF.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        80,
+        875,
+        586
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_12.jpg",
+    "caption": "Fig. 3. 2DZIF structure and its relationship with ZIF-L. a, Schematic contrasting the arrangement of layers within a ZIF-L crystal with that of a monolayer 2DZIF. b, Registry between 2DZIF and supercell of graphene based on SAED data in Supplementary Fig. 12. c, Structures of a ZIF-L layer (left) and 2DZIF (right) viewed along the [001], [100], and [010] directions. d, Schematic illustrating the etching of 2DZIF in water. SEM (e) and AFM (f) images of the triangular grains of 2DZIF obtained by a short etching in water. g, AFM height profile corresponding to the line in (f).",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        81,
+        884,
+        450
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4. Applications of 2DZIF and aZIF. a, Schematic of a 2DZIF film supported on nanoporous graphene reinforced with PTMSP. b, \\(\\mathrm{H}_2\\) , \\(\\mathrm{CO}_2\\) , \\(\\mathrm{N}_2\\) and \\(\\mathrm{CH}_4\\) permeances of the support film (PTMSP-reinforced NG) and the supported 2DZIF film. c, 2DZIF membrane separation performance for an equimolar \\(\\mathrm{H}_2 / \\mathrm{N}_2\\) mixed feed. d, Comparison of the \\(\\mathrm{H}_2 / \\mathrm{N}_2\\) separation performance of 2DZIF membranes with the state-of-the-art (Supplementary Table",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        75,
+        881,
+        760
+      ]
+    ],
+    "page_idx": 13
+  }
+]

preprint/preprint__09bb925dea8007aaf7d14d8f79096de27fcb7fe802393ba1d044ff47a75f57c2/preprint__09bb925dea8007aaf7d14d8f79096de27fcb7fe802393ba1d044ff47a75f57c2.mmd

@@ -0,0 +1,383 @@
+
+# Nanometer-thick crystalline and amorphous zeolitic imidazolate framework films for membrane and patterning applications  
+
+Qi Liu École Polytechnique Fédérale de Lausanne Yurun Miao  
+
+Johns Hopkins University https://orcid.org/0000- 0001- 6429- 8297  
+
+Luis Francisco Villalobos Yale University https://orcid.org/0000- 0002- 0745- 4246  
+
+Shaoxian Li École Polytechnique Fédérale de Lausanne  
+
+Deepu J. Babu École Polytechnique Fédérale de Lausanne  
+
+Cailing Chen King Abdullah University of Science and Technology https://orcid.org/0000- 0003- 2598- 1354  
+
+Heng- Yu Chi École Polytechnique Fédérale de Lausanne  
+
+Mohammad Tohidi Vahdat École Polytechnique Fédérale de Lausanne  
+
+Jian Hao École Polytechnique Fédérale de Lausanne  
+
+Shuqing Song École Polytechnique Fédérale de Lausanne  
+
+Yu Han King Abdullah University of Science and Technology https://orcid.org/0000- 0003- 1462- 1118  
+
+Michael Tsapatsis Johns Hopkins University https://orcid.org/0000- 0001- 5610- 3525  
+
+Kumar Varoon Agrawal ( kumar.agrawal@epfl.ch) École Polytechnique Fédérale de Lausanne https://orcid.org/0000- 0002- 5170- 6412  
+
+Article  
+
+Keywords:
+
+<--- Page Split --->
+
+**Posted Date:** March 14th, 2023  
+
+**DOI:** https://doi.org/10.21203/rs.3.rs-2666142/v1  
+
+**License:** © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+**Additional Declarations:** Yes there is potential Competing Interest. A patent application based on the finding reported in the manuscript is filed.  
+
+**Version of Record:** A version of this preprint was published at Nature Materials on September 21st, 2023. See the published version at https://doi.org/10.1038/s41563-023-01669-z.
+
+<--- Page Split --->
+
+# Nanometer-thick crystalline and amorphous zeolitic imidazolate framework films for membrane and patterning applications  
+
+Qi Liu<sup>1†</sup>, Yurun Miao<sup>2</sup>, Luis Francisco Villalobos<sup>1</sup>, Shaoxian Li<sup>1</sup>, Deepu J. Babu<sup>1†</sup>, Cailing Chen<sup>4</sup>, Heng- Yu Chi<sup>1</sup>, Mohammad Tohidi Vahdat<sup>1,5</sup>, Jian Hao<sup>1</sup>, Shuqing Song<sup>1</sup>, Yu Han<sup>4</sup>, Michael Tsapatsis<sup>2,3</sup>, Kumar Varoon Agrawal<sup>1\*</sup>  
+
+1. Laboratory of Advanced Separations, École Polytechnique Fédérale de Lausanne (EPFL), 1950 Sion, Switzerland.  
+
+2. Department of Chemical and Biomolecular, Engineering & Institute for NanoBioTechnology, Johns Hopkins University, Baltimore, MD, USA  
+
+3. Applied Physics Laboratory, Johns Hopkins University, Laurel, MD, USA  
+
+4. Advanced Membranes and Porous Materials Center, Physical Sciences and Engineering Division, King Abdullah University of Science and Technology, Thuwal 23955-6900, Saudi Arabia.  
+
+5. Theory and Simulation of Materials (THEOS) and National Centre for Computational Design and Discovery of Novel Materials (MARVEL), EPFL, Lausanne, Switzerland.  
+
+<sup>†</sup>Present address: College of Chemistry, Chemical Engineering and Materials Science, Soochow University, Suzhou, 215123, China.  
+
+<sup>†</sup>Present address: Materials Science and Metallurgical Engineering, Indian Institute of Technology, Hyderabad, Telangana 502 284, India.  
+
+\*Corresponding author: kumar.agrawal@epfl.ch
+
+<--- Page Split --->
+
+## Abstract  
+
+Zeolitic imidazolate frameworks (ZIFs) are a subset of metal- organic frameworks (MOFs) with more than 200 characterized crystalline and amorphous networks made of divalent transition metal centers (e.g., \(\mathrm{Zn^{2 + }}\) and \(\mathrm{Co^{2 + }}\) ) linked by imidazolate linkers. ZIF thin films have been pursued intensively motivated by the desire to prepare membranes for selective gas and liquid separations. To achieve membranes with high throughput, as in A- scale biological channels with nanometer- scale pathlengths, ZIF films with the minimum possible thickness, down to just one unit cell, are highly desired. Control of ZIF film thickness at the 10- nm- scale may also enable emerging, MOF- inspired, applications including patterned crystalline MOF films, and amorphous organic- inorganic resists for high- resolution electron- beam (e- beam) and extreme UV (EUV) lithography. However, the state- of- the- art methods yield ZIF films with thicknesses exceeding 40 nanometers. Here, we report a deposition method from ultra- dilute precursor mixtures that within minutes yields uniform ZIF deposits with nm- scale thickness control. On crystalline substrate such as graphene, two- dimensional crystalline ZIF (2DZIF) film with thickness of a unit- cell could be achieved, which composed of a six- membered zinc- imidazolate coordination ring enabling record- high \(\mathrm{H}_2\) permselective separation performance. Deposition under identical conditions on amorphous substrates yields macroscopically smooth amorphous ZIF (aZIF) films, which can be used as negative- and positive- tone resists yielding pattern features down to \(20 \mathrm{nm}\) . The method reported here will likely accelerate the development of 2D crystalline and ultrathin amorphous MOF films for applications ranging from separation membranes to sensors and patterning for microelectronic applications.  
+
+## Main Text:  
+
+ZIFs \(^{1,2}\) are a class of MOFs that hold promise for applications in molecular separations \(^{3 - 10}\) , patterning \(^{11 - 15}\) and sensing \(^{16}\) . Their chemical and physical properties have been widely explored as a function of framework flexibility \(^{17 - 20}\) as well as structural defects \(^{21,22}\) . The realization of two- dimensional (2D) ZIF films with thickness down to that afforded by a single structural building unit is highly desired to make ZIF analogues to graphene and related 2D materials with an added advantage; the intrinsic nanoporosity of ZIF can be used to separate molecules while maximizing the permselective flux \(^{23}\) . Another highly desirable feature is a nanometer- scale control over the film thickness, which can allow one to fabricate nanoscale patterns, which are desirable for incorporation in microelectronic devices \(^{24}\) , and for the development of
+
+<--- Page Split --->
+
+next generation organic- inorganic high- resolution photo- and e- beam- resists<sup>17</sup>. However, the realization of 2D crystalline and ultrathin amorphous ZIF films has remained elusive. While layered ZIFs such as ZIF- L<sup>25</sup>, \(\mathrm{Zn_2(bim)_4^{26}}\) , and analogs<sup>27</sup> have been reported, the individual ZIF layers in these materials have a small aspect ratio which prevents the realization of continuous 2D ZIF films with structural uniformity over a macroscopic (e.g., wafer) length scale. The state- of- the- art of ZIF deposition methods yields polycrystalline films with thickness larger than \(40 \mathrm{nm}\) nanometer<sup>28- 31</sup>. This is mainly due to difficulty in achieving in- plane film growth without film thickening.  
+
+Considerable knowledge exists on ZIF/MOF crystal nucleation and growth in solution<sup>32- 37</sup>. Based upon data from synchrotron X- ray scattering, density- functional theory (DFT) and molecular dynamics simulations, and other techniques, it is generally accepted that ZIF formation involves a sequence of events starting from the formation of small ( \(\sim 1 \mathrm{nm}\) ) metastable prenucleation clusters, which evolve through aggregation followed by intra- aggregate ZIF nucleation and growth. Recent studies on surface- directed MOF growth<sup>38- 46</sup> indicate that the diffusion of MOF precursors in the vicinity of the 2D material, and the MOF- 2D material interactions, are key to regulate the crystallinity of the MOF film and the ability to maintain in- plane/horizontal growth (desired for ultrathin films) versus out- of- plane/vertical (undesired) growth.  
+
+Herein, we report macroscopically uniform 2D ZIF films with exquisite nanometer- scale control over the film thickness by suppressing the out- of- plane growth by using an ultradilute growth solution. The ultralow precursor concentration restricts homogeneous nucleation in the solution and facilitates the growth of nanometer- thick films over an immersed substrate with deposition timescales of a few minutes. The film crystallinity is determined by the interaction of molecular precursors with the substrate ranging from substrate- registry- determined ordered film to amorphous films in the absence of any crystallographic registry. The film thickness could be controlled with a resolution of a single layer by controlling the deposition time and number of coatings.  
+
+The ZIF films were synthesized by immersing a substrate in an ultradilute precursor solution ( \(\leq 2 \mathrm{mM} \mathrm{Zn}^{2 + }\) and \(\leq 16 \mathrm{mM} 2\) - methylimidazole (2- mIm), respectively) for a few minutes (Fig. 1a). The use of such ultradilute solutions for the growth of ZIF films has not been reported before (Fig. 1b, Supplementary Table 1). They were used here in an effort to suppress the homogeneous nucleation in the bulk solution. With a diminished nuclei population in the bulk
+
+<--- Page Split --->
+
+solution, the attachment of preformed nuclei to the substrate can be reduced or eliminated. This is expected to promote film growth by the assembly of molecular precursors on the substrate. Since this thin film growth mode is anticipated to be sensitive on the type of the substrate, we carried out synthesis using distinct substrates: (i) graphitic substrates with atomically smooth terrace such as highly oriented pyrolytic graphite (HOPG) or graphene, (ii) Si/SiO₂ wafer with a 300- nm- thick oxide layer, and (iii) single crystal sapphire (Al₂O₃).  
+
+ZIF films prepared on HOPG using a growth solution of 1 mM \(\mathrm{Zn^{2 + }}\) and 8 mM 2- mM and reaction time of 5 min were examined by optical and scanning electron microscopy (SEM) (Supplementary Fig. 1). A sharp change in contrast was observed at the air/precursor- solution interface beyond which the film had a uniform contrast indicating that the film was smooth, continuous, and macroscopically uniform. Atomic force microscopy (AFM) imaging near the interface confirmed that the ZIF film is indeed continuous and has a thickness of approximately 2 nm (Fig. 1c and d). When the synthesis time was reduced to 2 min, we observed a submonolayer film with micrometer- sized domains (Supplementary Fig. 2). The domains were faceted and had thickness of 2 nm, consistent with the thickness of the continuous film indicating that the film is crystalline consisting of micron- sized grains. We could obtain 4 and 6 nm thick films by increasing the growth time from 5 min to 10 and 15 min, respectively (Fig. 1e and Supplementary Fig. 3a- f). Increasing growth time to 20 min further did not lead to thicker film, indicating precursor depletion (Supplementary Fig. 3g- i). Thicker (8 nm) films could be obtained by doubling the precursor concentration (Supplementary Fig. 3j- l). A discrete, 2 nm, increase in film thickness further suggests a crystalline order. A fitting of film thickness with the number of probable layers yielded a monolayer thickness of 2 nm (Fig. 1e). Macroscopically large ZIF films spanning several centimeters in width could be obtained on chemical vapor deposition (CVD) derived graphene film resting on a Cu foil (Fig. 1f and 1g).  
+
+We also obtained macroscopically smooth, continuous, and uniform ZIF films on Si/SiO₂ wafers (Fig. 1h). AFM of one of these films, prepared using 2 mM \(\mathrm{Zn^{2 + }}\) and 16 mM 2- mM and deposition time of 10 s, indicated that the film is smooth with thickness near 8 nm (Fig. 1i and 1j). Ellipsometry of several ZIF films on Si/SiO₂ wafers, prepared by varying the synthesis time, indicated that the film thickness could be tuned in the range of 8- 18 nm consistent with the corresponding AFM data (Supplementary Fig. 4).
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Fig. 1. Synthesis ZIF films from ultradilute solutions. a, Schematic of the ZIF film synthesis approach. b, Composition diagram comparing the precursor solution composition used in this study with those reported in the literature (Supplementary Table 1). AFM (c) and the corresponding height profile (d) of a monolayer ZIF film on HOPG. e, Monolayer and multilayer ZIF films on HOPG with discrete thicknesses as a function of synthesis time. Error </center>
+
+<--- Page Split --->
+
+bars in this figure represent the standard deviation of measurement. f, Optical image of a 2DZIF film on CVD graphene resting on a Cu foil. g, Scanning electron microscopy (SEM) image of a 2DZIF film on CVD graphene. The image was compiled by combining \(1155 (35 \times 33)\) images by scanning the whole surface of the large sample. h, SEM and optical images of a ZIF film on \(\mathrm{Si / SiO_2}\) . AFM (i) and the corresponding height profile (j) of the ZIF film on \(\mathrm{Si / SiO_2}\) .  
+
+Graphene supported ZIF film could be suspended on a holey transmission electron microscopy (TEM) grid (Fig. 2a). The film was devoid of large crystals and appeared uniform. Selected area electron diffraction (SAED) from a micrometer- sized area yielded three sets of diffraction patterns (Fig. 2b). The first two sets ((01), highlighted with green circles) had six- fold symmetry originating from two slightly misoriented (by \(3.0^{\circ}\) ) grains of graphene, while the last set had two- fold symmetry and belonged to a single grain of ZIF (highlighted with white circles), confirming that ZIF prepared on graphene was crystalline with grains at least a micron in size consistent with the AFM- based imaging of grains in the submonolayer film (Supplementary Fig. 2). Hereafter, the ZIF films on graphitic substrates are referred to as 2DZIF. The fact that a single 2DZIF grain could grow over two slightly misoriented graphene grains indicates that the growth could accommodate a small mismatch in its registry with the substrate. Diffraction pattern from 2DZIF, typically representing a single grain, was observed from every single spot over a large area. Based on the diffraction pattern, \(a\) and \(b\) lattice parameters of 2.4 and \(2.0 \mathrm{nm}\) , respectively, could be fitted.  
+
+We carried out synchrotron grazing incidence X- ray diffraction (GIXRD) of a 10- nm- thick ZIF film on graphene resting on a \(\mathrm{Si / SiO_2}\) wafer (Supplementary Fig. 5- 7). The in- plane GIXRD pattern revealed sharp diffraction peaks, consistent with the peak positions obtained by the radial integration of the SAED pattern (Fig. 2c and Supplementary Table 2), confirming that the film formed on the graphitic substrate exhibits crystalline order. In contrast, we did not observe diffraction from the ZIF films prepared directly on \(\mathrm{Si / SiO_2}\) wafer without the graphene layer indicating that these films were amorphous (hereafter referred to as aZIF films, Fig. 2c). When a single crystal sapphire ( \(\mathrm{Al_2O_3}\) ) was used as the substrate, highly crystalline 2DZIF film was formed (Supplementary Fig. 8 and 9). The presence of the order in the ZIF film when prepared over a crystalline substrate and the lack of an order when prepared over the amorphous substrate indicates a strong role of substrate registry in the formation of the ordered 2DZIF film.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Fig. 2. Structure determination of 2DZIF films. a, Bright-field TEM image of the 2DZIF film supported on suspended graphene, and (b) its corresponding SAED pattern. The pattern from graphene is identified with green circles and those from 2DZIF with white circles. c, In-plane GIXRD data from an aZIF film (top) and a 2DZIF film (middle) prepared on \(\mathrm{Si / SiO_2}\) and graphene/Si/SiO2, respectively, along with a radially integrated trace (bottom) of the SAED pattern shown in (B). d, N1s XPS spectra from ZIF-8, ZIF-L, aZIF and 2DZIF films. The N-Zn and N-H coordination environments are shown on the right. e, DFT-relaxed structure of the 2DZIF and a visualization of the 6-MR (right). f, HRTEM image of the 2DZIF film lying flat on the \(hk0\) plane, resting on suspended graphene, and (g) corresponding Fourier transform compared with the simulated diffraction pattern from the proposed structure oriented along the \(c\) -out-of-plane direction. h, Left: CTF-corrected image of the highlighted area in (f) based on a defocus value of -130 nm analyzed from the Thon rings in the Fourier transform pattern. Right: simulated projected potential map along the [001] direction of 2DZIF. </center>
+
+<--- Page Split --->
+
+X- ray photoelectron spectroscopy (XPS) of the 2DZIF and aZIF films was carried out to gain further insights into their coordination environments (Fig. 2d). The N1s XPS data of 2DZIF when compared to that of ZIF- L layers and a prototypical non- layered ZIF (ZIF- 8) revealed that both 2DZIF and ZIF- L yield two peaks (399.0 and \(400.2\mathrm{eV}\) corresponding to N- Zn and N- H bonds, respectively) in contrast to a single peak (399.0 eV) from the ZIF- 8 crystals. This is consistent with the presence of abundant surface terminations (N- H) in the 2DZIF layers. In comparison, the population of N- H species was significantly diminished for aZIF indicating a nonlayered amorphous structure. The Zn2p XPS spectra was similar for all samples (Supplementary Fig. 10) indicating a similar coordination environment for Zn.  
+
+To gain insight into the structure of 2DZIF, structural relaxation based on DFT was carried out starting with the \(a\) and \(b\) lattice parameters obtained by SAED. The \(c\) - axis parameter was estimated by the AFM measurements (2 nm), and was subsequently relaxed by DFT calculations. The relaxed structure has an orthorhombic space group Cmce with the following structural parameters; \(a = 24.196\mathrm{\AA}\) , \(b = 19.719\mathrm{\AA}\) , \(c = 20.908\mathrm{\AA}\) , \(\alpha = 90^{\circ}\) , \(\beta = 90^{\circ}\) , and \(\gamma = 90^{\circ}\) (Supplementary Table 2). The layer in 2DZIF is composed of alternating 4- member ring (MR) and 6- MR chains while terminal 2- mIm linkers are present on both sides of the layer (Fig. 2e and 3c). The pore aperture of 2DZIF is constituted by the 6- MR (Fig. 2e) which is attractive for gas separation.  
+
+Aberration- corrected high- resolution TEM (AC- HRTEM) imaging of the 2DZIF film suspended on a TEM grid was carried out along the [001] crystallographic direction (Fig. 2f). The imaging was carried out using a low- dose beam condition<sup>47</sup> to minimize damages to the beam sensitive 2DZIF lattice. Indeed, the obtained HRTEM image revealed the high crystallinity of the 2DZIF film. The corresponding Fourier transform validated the \(c\) - out- of- plane orientation of the film and was consistent with the simulated electron diffraction pattern from a film lying flat along the same orientation (Fig. 2g). Projection along the \(c\) - out- of- plane axis from the contrast transfer function (CTF) corrected image revealed alternating chains of 4- MR and 6- MR (Fig. 2h, left), consistent with the simulated [001]- projected electrostatic potential map of 2DZIF structure obtained by DFT structural relaxation (Fig. 2h, right).  
+
+The registry or the lack of registry of 2DZIF with the underlying substrate plays an important role in determining its structure and morphology especially when contrasted against the closely related material, ZIF- L. Fig. 3a highlights the morphological differences in ZIF- L and 2DZIF. While the layers in ZIF- L and 2DZIF are stacked along the \(c\) - axis, the former grows as a leaf
+
+<--- Page Split --->
+
+shaped layered crystals (Supplementary Fig. 11) whereas the latter can form macroscopically uniform monolayer films. The unique leaf shape is formed because ZIF- L layers stack with first progressively increasing and then progressively decreasing lateral size along the \(b\) - axis. An analysis of orientation of 2DZIF grain over graphene by SAED showed that the 2DZIF films crystallized maintaining a fixed set of orientation with graphene (Supplementary Fig. 12), indicating that substrate registry indeed plays a role in promoting in- plane growth of 2DZIF (Fig. 3b and Supplementary Note 1). While both ZIF- L and 2DZIF have orthorhombic lattices, the unit- cell parameters of 2DZIF grown on graphene, obtained by DFT relaxation, are distinct from those of ZIF- L where the latter has a significantly shorter parameter along the \(b\) (17.060 Å) axis compared to the former (19.719 Å, Fig. 3c and Supplementary Table 2).  
+
+The grains of 2DZIF could be visualized by partial etching of 2DZIF films based on the well documented dissolution of ZIFs in water (Fig. 3d)48. After partial dissolution, the grain shape was triangular with a lateral size of 1- 2 \(\mu \mathrm{m}\) (Fig. 3e) consistent with earlier observations of domains in the sub- monolayer film. The three sides of the triangular grains could be assigned to be (110), (110) and (100) lattice planes, respectively, reported to be the minimum surface energy planes for ZIF layers (Supplementary Fig. 13)49. AFM images (Fig. 3f) confirmed that the grains have uniform thickness of \(\sim \mathrm{ca. 2 nm}\) consistent with the structure of 2DZIF. SAED of this sample showed the same pattern with 2DZIF confirming that the triangular domains were indeed 2DZIF (Supplementary Fig. 14).
+
+<--- Page Split --->
+![](images/Supplementary_Figure_12.jpg)
+
+<center>Fig. 3. 2DZIF structure and its relationship with ZIF-L. a, Schematic contrasting the arrangement of layers within a ZIF-L crystal with that of a monolayer 2DZIF. b, Registry between 2DZIF and supercell of graphene based on SAED data in Supplementary Fig. 12. c, Structures of a ZIF-L layer (left) and 2DZIF (right) viewed along the [001], [100], and [010] directions. d, Schematic illustrating the etching of 2DZIF in water. SEM (e) and AFM (f) images of the triangular grains of 2DZIF obtained by a short etching in water. g, AFM height profile corresponding to the line in (f). </center>  
+
+The \(3.2\mathrm{\AA}\) gap in the 6- MR of 2DZIF is attractive for sieving \(\mathrm{H}_2\) (kinetic diameter of \(2.89\mathrm{\AA}\) from larger gas molecules such as \(\mathrm{CO_2}\) ( \(3.30\mathrm{\AA}\) ), \(\mathrm{N}_2\) ( \(3.64\mathrm{\AA}\) ), and \(\mathrm{CH_4}\) ( \(3.80\mathrm{\AA}\) ) \(^{50,51}\) . The 2DZIF film is mechanically robust with Young's modulus of \(8.1\pm 2.1\) GPa (Supplementary Fig. 15), comparable to that of the three- dimensional analogs \(^{52}\) . Therefore, we probed the \(\mathrm{H}_2\) - sieving performance on the 2DZIF film. For this, 2DZIF films were grown on nanoporous graphene (NG, Supplementary Fig. 16) mechanically reinforced with a dense 250- nm- thick poly[1- (trimethylsilyl)propyne] (PTMSP) film (Supplementary Fig. 17 and 18) where the NG/PTMSP film acts as a support film (Fig. 4a). The pores in NG were intentionally designed to be large \((1.8\pm 1.2\mathrm{nm})^{53}\) to rule out any molecular sieving from NG and to allow the determination of \(\mathrm{H}_2\) - sieving from the 2DZIF film. The 2DZIF films, resting on a macroporous metal foil support (pore size of \(5\mu \mathrm{m}\) , area of \(1\mathrm{mm}^2\) ), exhibited a molecular cut- off for molecules larger than \(\mathrm{H}_2\) ,
+
+<--- Page Split --->
+
+indicating that gas transport was controlled by the 6- MR of 2DZIF (Fig. 4b and Supplementary Fig. 19 and 20). The \(\mathrm{H}_2\) permeance was large ( \(>15000\) gas permeation units or GPU; 1 GPU = \(3.35 \times 10^{- 10} \mathrm{~mol} \mathrm{~m}^{- 2} \mathrm{~s}^{- 1} \mathrm{~Pa}^{- 1}\) ) similar to that from the support film (Supplementary Table 3 and Supplementary Fig. 19), indicating a negligible transport resistance from the 2DZIF layer.  
+
+When an equimolar \(\mathrm{H}_2:\mathrm{N}_2\) mixture was probed with feed pressure of 2 bar, a \(\mathrm{H}_2\) permeance of 17300 GPU with a \(\mathrm{H}_2 / \mathrm{N}_2\) separation factor of 115 could be obtained (Fig. 4c). Another membrane when tested under a high- pressure feed (8 bar), exhibited a high \(\mathrm{H}_2\) flux of 2.8 mol \(\mathrm{m}^{- 2} \mathrm{~s}^{- 1}\) and \(\mathrm{H}_2 / \mathrm{N}_2\) separation factor of 52 (Supplementary Fig. 21). This performance constitutes one of the best combinations of \(\mathrm{H}_2\) flux and \(\mathrm{H}_2 / \mathrm{N}_2\) separation factor (Fig. 4d and Supplementary Fig. 22, 23; Supplementary Table 5). Larger area (centimeter- scale) 2DZIF membrane could be also prepared (Supplementary Fig. 24), thanks to the highly uniform deposition of 2DZIF films on graphene (Fig. 1f and g). They also show attractive \(\mathrm{H}_2\) permselective separation performance (Supplementary Table 4), in agreement with the smaller- area membranes (Supplementary Note 2).
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Fig. 4. Applications of 2DZIF and aZIF. a, Schematic of a 2DZIF film supported on nanoporous graphene reinforced with PTMSP. b, \(\mathrm{H}_2\) , \(\mathrm{CO}_2\) , \(\mathrm{N}_2\) and \(\mathrm{CH}_4\) permeances of the support film (PTMSP-reinforced NG) and the supported 2DZIF film. c, 2DZIF membrane separation performance for an equimolar \(\mathrm{H}_2 / \mathrm{N}_2\) mixed feed. d, Comparison of the \(\mathrm{H}_2 / \mathrm{N}_2\) separation performance of 2DZIF membranes with the state-of-the-art (Supplementary Table </center>
+
+<--- Page Split --->
+
+5). GO, CMP, and HOF refer to graphene oxide, conjugated microporous polymers and hydrogen-bonded organic frameworks, respectively. The dash line indicates Knudsen selectivity for the \(\mathrm{H}_2 / \mathrm{N}_2\) pair. e, Schematic of the patterning process. TEM (f) and AFM (g) images of nanoscale patterns made on an aZIF film. h, AFM height profile corresponding to the line in (g).  
+
+Amorphous MOFs exhibit unique physical and chemical properties due to the absence of anisotropy and crystalline grains<sup>1,54</sup>. On one hand, they may not have the well- defined pore structures of crystalline MOFs required for certain molecular sieving applications, but at the same time, they do not exhibit grain boundaries and structural anisotropies of crystalline MOFs, which can create film non- uniformities. A potential use of organic- inorganic films is in next generation resists for photolithography<sup>55</sup> in place of currently used polymeric resists, and, for this application, MOF- inspired metal- organic clusters have been proposed for high resolution patterning<sup>13</sup>. As a demonstration of the potential of our deposition method in this emerging application, an aZIF film was deposited on a silicon nitride support and subsequently exposed to a direct- write electron beam using 1:1 line- and space- patterns ranging from 10 to 40 nm in line width (or half pitch) (Fig. 4e). The aZIF films behave similarly to ZIF- L crystals, for which e- beam treatment can induce contrast in water dissolution behavior based on framework densification and disintegration of the ligand molecular structure<sup>56- 58</sup>. After development in water, the irradiated area was preserved while the non- irradiated area was dissolved (Fig. 4f), confirming aZIF as a negative- tone resist. The thickness of the remaining aZIF structure was determined to be \(\sim 25 \mathrm{nm}\) by AFM (Fig. 4g and h). The resolution of the resulting pattern, as exemplified by the well- resolved lines at 20 nm half pitch, is comparable to the state- of- the- art metal- containing resists<sup>13,59</sup>, which are an emerging class of material that hold promise in extreme ultraviolet lithography and electron beam lithography<sup>57- 61</sup>. aZIF can also be patterned in positive- tone mode by a vapor phase ligand pretreatment. The as- deposited aZIF is exposed to the sublimated vapor of 4,5- dichloroimidazole (dcIm) at \(75^{\circ} \mathrm{C}\) for 1.5 h, during which the aZIF matrix is partially exchanged or infiltrated with dcIm ligand. The dcIm- treated film is then exposed to a direct- write electron beam. After development in organic solvents, the irradiated area is removed while non- irradiated area is preserved, which showed similar sensitivity compared to the reported data (Supplementary Fig. 25 and Supplementary Table S6). Furthermore, to improve compatibility with microfabrication processes, aZIF films are spin- coated on silicon wafers, and their thicknesses can be controlled by spin speed (Supplementary Fig. 26- 28). The simple fabrication for ultrathin ZIF films reported in this
+
+<--- Page Split --->
+
+study could accelerate the development of new ZIF- based resist materials for lithographic applications<sup>62- 64</sup>.  
+
+The method reported here can be extended to other promising MOF structures. 2D film of UiO- 66- NH<sub>2</sub> can also be deposited by on graphene (Supplementary Fig. 29). It makes this approach reported here broad and interesting to develop a number of 2D MOF films in the future. In conclusion, we report the synthesis of ZIFs as macroscopically uniform amorphous and crystalline 2D films from an ultradilute solution. The 2DZIF film yields exceptional H<sub>2</sub>- sieving performance, thanks to the ordered 2D structure with a high density of 6- MR hosting H<sub>2</sub>- selective gap, making such a film the ultimate selective layer for membrane application. In the absence of substrate registry, ultrathin amorphous films are demonstrated, which are promising for advancing the limit of nanoscale patterning.  
+
+## Online content  
+
+Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available online.  
+
+## References  
+
+1. Bennett, T. D. & Cheetham, A. K. Amorphous Metal–Organic Frameworks, Acc. Chem. Res. 47, 1555-1562 (2014).  
+2. Banerjee, R. et al. High-Throughput Synthesis of Zeolitic Imidazolate Frameworks and Application to CO<sub>2</sub> Capture, Science 319, 939-943 (2008).  
+3. Li, Y-S. et al. Molecular Sieve Membrane: Supported Metal–Organic Framework with High Hydrogen Selectivity, Angew. Chem. 122, 558-561 (2010).  
+4. Brown, A. J. et al. Interfacial microfluidic processing of metal-organic framework hollow fiber membranes, Science 345, 72-75 (2014).  
+5. Ma, X. et al. Zeolitic imidazolate framework membranes made by ligand-induced permselectivation, Science 361, 1008-1011 (2018).  
+6. Gúcúyener, C., van den Bergh, J., Gascon, J. & Kapteijn, F. Ethane/Ethene Separation Turned on Its Head: Selective Ethane Adsorption on the Metal–Organic Framework ZIF-7 through a Gate-Opening Mechanism, J. Am. Chem. Soc. 132, 17704-17706 (2010).
+
+<--- Page Split --->
+
+7. Kwon, H. T. & Jeong, H-K. In Situ Synthesis of Thin Zeolitic–Imidazolate Framework ZIF-8 Membranes Exhibiting Exceptionally High Propylene/Propane Separation, J. Am. Chem. Soc. 125, 10763-10768 (2013).  
+
+8. Zhou, S. et al. Paralyzed Membrane: Current-Driven Synthesis of a Metal-Organic Framework with Sharpened Propene/Propane Separation. Sci. Adv. 4:eaau1393 (2018).  
+
+9. Huang, K. et al. A ZIF-71 Hollow Fiber Membrane Fabricated by Contra-Diffusion. ACS Appl. Mater. Interfaces 7, 16157-16160 (2015).  
+
+10. van den Bergh, J. et al. Understanding the Anomalous Alkane Selectivity of ZIF-7 in the Separation of Light Alkane/Alkene Mixtures. Chem. Eur. J. 17, 8832-8840 (2011).  
+
+11. Tu, M. et al. Direct X-ray and electron-beam lithography of halogenated zeolitic imidazolate frameworks, Nat. Mater. 20, 93-99 (2021).  
+
+12. Miao, Y. et al. Solvent-free bottom-up patterning of zeolitic imidazolate frameworks, Nat. Commun. 13, 420 (2022).  
+
+13. Dalstein, O. et al. Nanoimprinted, Submicrometric, MOF-Based 2D Photonic Structures: Toward Easy Selective Vapors Sensing by a Smartphone Camera, Adv. Funct. Mater. 26, 81-90 (2016).  
+
+14. Xu, H. et al. Metal-Organic Framework-Inspired Metal-Containing Clusters for High-Resolution Patterning, Chem. Mater. 30, 4124-4133 (2018).  
+
+15. Stassen, I. et al. Solvent-free synthesis of supported ZIF-8 films and patterns through transformation of deposited zinc oxide precursors, CrystEngComm 15, 9308-9311 (2013).  
+
+16. Lu, G. & Hupp, J. T. Metal-Organic Frameworks as Sensors: A ZIF-8 Based Fabry-Pérot Device as a Selective Sensor for Chemical Vapors and Gases, J. Am. Chem. Soc. 132, 7832-7833 (2010).  
+
+17. Moggach, S. A., Bennett, T. D. & Cheetham, A. K. The Effect of Pressure on ZIF-8: Increasing Pore Size with Pressure and the Formation of a High-Pressure Phase at 1.47 GPa, Angew. Chem. Int. Ed. 48, 7087-7089 (2009).  
+
+18. Zhang, C. et al. Unexpected Molecular Sieving Properties of Zeolitic Imidazolate Framework-8, J. Phys. Chem. Lett. 3, 2130-2134 (2012).  
+
+19. Knebel, A. et al. Defibrillation of soft porous metal-organic frameworks with electric fields, Science 358, 347-351 (2017).  
+
+20. Zhang, K. et al. Exploring the Framework Hydrophobicity and Flexibility of ZIF-8: From Biofuel Recovery to Hydrocarbon Separations. J. Phys. Chem. Lett. 4, 3618-3622 (2013).  
+
+21. Bennett, T. D. Cheetham, A. K., Fuchs, A. H. & Coudert, F-X. Interplay between defects, disorder and flexibility in metal-organic frameworks, Nat. Chem. 9, 11-16 (2017).  
+
+22. Babu, D. J. et al. Restricting Lattice Flexibility in Polycrystalline Metal-Organic Framework Membranes for Carbon Capture, Adv. Mater. 31, 1900855 (2019).  
+
+23. Gascon, J. & Kapteijn, F. Metal-Organic Framework Membranes—High Potential, Bright Future? Angew. Chem. Int. Ed. 49, 1530-1532 (2010).
+
+<--- Page Split --->
+
+24. Stassen, I. et al. An updated roadmap for the integration of metal–organic frameworks with electronic devices and chemical sensors. Chem. Soc. Rev. 46, 3185-3241 (2017).25. Chen, R. et al. A two-dimensional zeolitic imidazolate framework with a cushion-shaped cavity for \(\mathrm{CO_2}\) adsorption, Chem. Commun. 49, 9500-9502 (2013).26. Peng, Y. et al. Metal-organic framework nanosheets as building blocks for molecular sieving membranes, Science 346, 1356-1359 (2014).27. Peng, Y., Li, Y., Ban, Y. & Yang, W. Two-Dimensional Metal–Organic Framework Nanosheets for Membrane-Based Gas Separation. Angew. Chem. 129, 9889-9893 (2017).28. He, G., Dakhchoune, M., Zhao, J., Huang, S. & Agrawal, K. V. Electrophoretic Nuclei Assembly for Crystallization of High-Performance Membranes on Unmodified Supports, Adv. Funct. Mater. 28, 1707427 (2018).29. Wei, R. et al. Aqueously Cathodic Deposition of ZIF-8 Membranes for Superior Propylene/Propane Separation, Adv. Funct. Mater. 30, 1907089 (2020).30. Eum, K. et al. ZIF-8 Membrane Separation Performance Tuning by Vapor Phase Ligand Treatment, Angew. Chem. 131, 16542-16546 (2019).31. Chen, Z. et al. Large-Area Crystalline Zeolitic Imidazolate Framework Thin Films. Angew. Chem. Int. Ed. 61, 14124-14130 (2021).32. Van Vleet, M. J., Weng, T., Li, X. & Schmidt, J. R. In Situ, Time-Resolved, and Mechanistic Studies of Metal-Organic Framework Nucleation and Growth. Chem. Rev. 118, 3681-3721 (2018).33. Cravillon, J. et al. Rapid Room-Temperature Synthesis and Characterization of Nanocrystals of a Prototypical Zeolitic Imidazolate Framework. Chem. Mater. 21, 1410-1412 (2009).34. Cravillon, J. et al. Fast Nucleation and Growth of ZIF-8 Nanocrystals Monitored by Time-Resolved In Situ Small-Angle and Wide-Angle X-Ray Scattering. Angew. Chem. Int. Ed. 50, 8067-8071 (2011).35. Jian, M. et al. Water-based synthesis of zeolitic imidazolate framework-8 with high morphology level at room temperature. RSC Adv. 5, 48433-48441 (2015).36. Terban, M. W. et al. Early stage structural development of prototypical zeolitic imidazolate framework (ZIF) in solution. Nanoscale 10, 4291-4300 (2018).37. Filez, M. et al. Elucidation of the pre-nucleation phase directing metal-organic framework formation. Cell Rep. Phys. Sci. 2, 100680 (2021).38. Doustkhah, E., Hassandoost, R., Khataee, A., Luque, R. & Assadi, M. H. N. Hard-templated metal–organic frameworks for advanced applications. Chem. Soc. Rev. 50, 2927-2953 (2021).39. Wang, K., Hui, K. N., San Hui, K., Peng, S. & Xu, Y. Recent progress in metal–organic framework/graphene-derived materials for energy storage and conversion: design, preparation, and application. Chem. Sci. 12, 5737-5766 (2021).
+
+<--- Page Split --->
+
+40. Wang, J. et al. Zeolitic Imidazolate Framework/Graphene Oxide Hybrid Nanosheets Functionalized Thin Film Nanocomposite Membrane for Enhanced Antimicrobial Performance. ACS Appl. Mater. Interfaces 8, 25508-25519 (2016).  
+
+41. Pokhrel, J. et al. \(\mathrm{CO_2}\) adsorption behavior of amine-functionalized ZIF-8, graphene oxide, and ZIF-8/graphene oxide composites under dry and wet conditions. Microporous Mesoporous Mater. 267, 53-67 (2018).  
+
+42. Li, S. et al. Unconventional Nucleation and Oriented Growth of ZIF-8 Crystals on Non-Polar Surface. Adv. Mater. 24, 5954-5958 (2012).  
+
+43. Yang, H. et al. Vacuum-assisted assembly of ZIF-8@GO composite membranes on ceramic tube with enhanced organic solvent nanofiltration performance. J. Membr. Sci. 545, 158-166 (2018).  
+
+44. Choi, E. et al. Pore Tuning of Metal-Organic Framework Membrane Anchored on Graphene-Oxide Nanoribbon. Adv. Funct. Mater. 31, 2011146 (2021).  
+
+45. Sikdar, A. et al. Diffusion driven nanostructuring of metal-organic frameworks (MOFs) for graphene hydrogel based tunable heterostructures: highly active electrocatalysts for efficient water oxidation. J. Mater. Chem. A 9, 7640-7649 (2021).  
+
+46. Tao, J. et al. Controlling Metal-Organic Framework/ZnO Heterostructure Kinetics through Selective Ligand Binding to ZnO Surface Steps. Chem. Mater. 32, 6666-6675, (2020).  
+
+47. Zhu, Y. et al. Unravelling surface and interfacial structures of a metal-organic framework by transmission electron microscopy, Nat. Mater. 16, 532-536 (2017).  
+
+48. Zhang, H., Zhao, M. & Lin, Y. S. Stability of ZIF-8 in water under ambient conditions, Microporous Mesoporous Mater. 279, 201-210 (2019).  
+
+49. Motevalli, B., Taherifar, N., Wang, H. & Liu, J. Z. Ab Initio Simulations to Understand the Leaf-Shape Crystal Morphology of ZIF-L with Two-Dimensional Layered Network, J. Phys. Chem. C 121, 2221-2227 (2017).  
+
+50. Zhu, W., Li, X., Sun, Y., Guo, R., & Ding, S. Introducing hydrophilic ultra-thin ZIF-L into mixed matrix membranes for \(\mathrm{CO_2 / CH_4}\) separation, RSC Adv. 9, 2339-23399 (2019).  
+
+51. Yang, K. et al. ZIF-L membrane with a membrane-interlocked-support composite architecture for \(\mathrm{H_2 / CO_2}\) separation, Sci. Bull. 66, 1869-1876 (2021).  
+
+52. Tan, J. C., Bennett, T. D., Cheetham, A. K. Chemical structure, network topology, and porosity effects on the mechanical properties of Zeolitic Imidazolate Frameworks, Proc. Natl. Acad. Sci. 107, 9938-9943 (2010).  
+
+53. He, G. et al. High-permeance polymer-functionalized single-layer graphene membranes that surpass the postcombustion carbon capture target, Energy Environ. Sci. 12, 3305-3312 (2019).  
+
+54. Bennett, T. D. & Horike, S. Liquid, glass and amorphous solid states of coordination polymers and metal–organic frameworks, Nat. Rev. Mater. 3, 431-440 (2018).
+
+<--- Page Split --->
+
+55. Kosma, V., De Simone, D., Vandenberghe, G. Metal Based Materials for EUV Lithography, J. Photopolymer. Sci. Technol. 32, 179-183 (2019).  
+
+56. Conrad, S. et al. Controlling Dissolution and Transformation of Zeolitic Imidazolate Frameworks by using Electron-Beam-Induced Amorphization, Angew. Chem. Int. Ed. 57, 13592-13597 (2018).  
+
+57. Stowers, J. & Keszler, D. A. High resolution, high sensitivity inorganic resists, Microelectron. Eng. 86, 730-733 (2009).  
+
+58. Miao, Y., Tsapatsis, M. Electron Beam Patterning of Metal-Organic Frameworks. Chem. Mater. 33, 754-760 (2021).  
+
+59. Oleksak, R. P. et al. Chemical and Structural Investigation of High-Resolution Patterning with HafSOx, ACS Appl. Mater. Interfaces 6, 2917-2921 (2014).  
+
+60. Ghosh, S. et al. Chem. Mater. Two distinct stages of structural modification of ZIF-L MOF under electron-beam irradiation. 33, 5681-5689 (2021).  
+
+61. Luo, C. et al. Review of recent advances in inorganic photoresists, RSC Adv. 10, 8385-8395 (2020).  
+
+62. Manouras, T. & Argitis, P. High Sensitivity Resists for EUV Lithography: A Review of Material Design Strategies and Performance Results, Nanomaterials 10, 1593 (2020).  
+
+63. Gangaik, A. S., Georgiev, Y. M. & Holmes, J. D. New Generation Electron Beam Resists: A Review, Chem. Mater. 29, 1898-1917 (2017).  
+
+64. Ghash, S. et al. Two Distinct Stages of Structural Modification of ZIF-L MOF under Electron-Beam Irradiation, Chem. Mater. 33, 5681-5689 (2021).  
+
+## Methods  
+
+## Chemicals  
+
+Zinc nitrate hexahydrate \(\mathrm{(Zn(NO_3)_2\cdot 6H_2O)}\) was purchased from Sigma-Aldrich. 2- methylimidazole (2-mIm) was from Chemie Brunschwig AG. HCl (32 wt%) was purchased from Reactolab S.A.. poly[1-(trimethylsilyl)-1-propylene] (PTMSP) was from ABCR. \(\mathrm{FeCl_3}\) (97%) and \(\mathrm{Na_2S_2O_8}\) was bought from Sigma-Aldrich. Cu foil (50 mm, 99.9%) were purchased from STREM. Toluene (AR) and methanol (AR) were from Fischer. All chemicals were used without further purifications. \(\mathrm{Si / SiO_2}\) wafers were purchased from University Wafer Inc. \(\mathrm{Si / SiO_2}\) wafer with single layer graphene was bought from Ted Pella. Highly oriented pyrolytic graphite (HOPG) (ZYA quality, GRAS/1.0x7x7) was purchased from ScanSens. Silicon nitride TEM supports (50 nm silicon nitride film on a \(200\mu \mathrm{m}\) silicon frame with nine viewing
+
+<--- Page Split --->
+
+windows, each \(0.1 \times 0.1 \mathrm{mm}\) were purchased from Ted Pella. Sapphire (Al₂O₃) wafer (Z- cut, \(10 \mathrm{cm} \times 0.5 \mathrm{mm}\) ) was purchased from MTI.  
+
+## Characterizations and Measurements  
+
+SEM measurements were performed on a Teneo Scanning Electron Microscope operating at 1 kV. PXRD data were collected at a Bruker D8 Discover diffractometer with a Lynxeye XE detector, operated at \(40 \mathrm{kV}\) , \(400 \mathrm{mA}\) for Cu Kα \((\lambda = 1.5406 \mathrm{\AA})\) at ambient temperature and pressure. Bright- field TEM images and selected- area electron diffraction (SAED) images were obtained with a Talos F200X microscope operated at \(200 \mathrm{kV}\) . TEM images for patterns on silicon nitride were obtained on a ThermoFisher TF30 TEM operating at \(300 \mathrm{kV}\) .  
+
+Low- dose AC- HRTEM was performed on a Cs- corrected FEI cubed G2 Titan 60- 300 electron microscope at \(300 \mathrm{kV}\) , using a Gatan K2 direct- detection camera in electron counting mode. The AC- HRTEM images were acquired with the dose fractionation function, and each image stack is composed of 120 frames with \(0.05 \mathrm{s}\) exposure for each frame, with a total electron dose of \(\sim 60 \mathrm{e}^{- }\mathrm{\AA}^{- 2}\) . The raw image was denoised by using an average background subtraction filter (ABSF). The CTF correction was performed based on the defocus value determined from the amorphous thon rings in the Fourier transform, and the projected electrostatic potential was simulated by the QSTEM software (QSTEM V2, 31). Simulated ED pattern was carried out by using Singlecrystal module of CrystalMaker software.  
+
+AFM images and modulus measurement were recorded on a Bruker MultiMode 8 AFM. For modulus measurement, a Bruker Tap525A rectangular probe was used and calibrated with standard sample sapphire, polystyrene and HOPG. XPS were carried out on an Axis Supra (Kratos Analytical) using the monochromated K x- ray line of an aluminum anode. Synchrotron GIXRD was carried out at beamline BM01, Swiss- Norwegian beamline (SNBL) at the European Synchrotron Radiation Facility (ESRF) with wavelength of \(0.683 \mathrm{\AA}\) and \(0.960 \mathrm{\AA}\) .  
+
+The gas separation performance of all membranes was recorded on a homemade permeation set- up. The pressure on the feed side was maintained at 2- 8 bar and on the permeate side at 1 bar during the measurements. All measurements were done after reaching the steady state with argon as the sweep gas. The membranes were sealed with stainless- steel gasket. The composition of permeate was analysed using an online Hiden Analytical HPR- 20 mass spectrometer.  
+
+The permeances, \(J_{i}\) , of gas \(i\) was calculated by Eq. S1
+
+<--- Page Split --->
+
+\[J_{i} = X_{i} / (A\cdot \Delta P_{i}) \quad (S1)\]  
+
+where \(X_{i}\) is the molar flow rate of component \(i\) across the membrane area, \(A\) and \(\Delta P_{i}\) is the transmembrane pressure difference for the component \(i\) . The selectivity \(\alpha_{ij}\) of two gases ( \(i\) and \(j\) , where \(i\) is the faster permeating gas) was calculated by Eq. S2  
+
+\[\alpha_{ij} = J_{i} / J_{j} \quad (S2)\]  
+
+## Synthesis of 2DZIF/aZIF film  
+
+In a petri dish containing \(29~\mathrm{ml}\) of \(\mathrm{Zn(NO_3)_2}\) aqueous solution at room temperature, the substrate (HOPG, graphene/Si/SiO \(_2\) for 2DZIF and Si/SiO \(_2\) for aZIF) was partially immersed. Then \(1\mathrm{ml}\) of \(2\mathrm{- mlm}\) aqueous solution was added. After a certain time, the substrate was removed to stop the reaction.  
+
+## Synthesis of 2DZIF membrane for gas separation  
+
+First, single- layer graphene was synthesized by using low- pressure CVD of methane on copper foil following the literature. Before the synthesis, the copper foil was annealed at \(1077^{\circ}\mathrm{C}\) in a \(\mathrm{H}_2 / \mathrm{Ar}\) atmosphere for \(60~\mathrm{min}\) . Then, \(\mathrm{CO_2}\) ( \(100~\mathrm{mL / min}\) ) and \(\mathrm{H}_2\) ( \(8~\mathrm{mL / min}\) ) flow was introduced successively, each for \(30~\mathrm{min}\) , to remove the contaminations. At last, \(\mathrm{CH_4}\) ( \(24~\mathrm{mL / min}\) ) and \(\mathrm{H}_2\) ( \(8~\mathrm{mL / min}\) ) flow was used to grow single- layer graphene on copper film for \(30~\mathrm{min}\) at pressure of \(460~\mathrm{mTorr}\) .  
+
+After the synthesis of single- layer graphene, an \(\mathrm{O_2}\) plasma (MTI plasma cleaner, EQ- PCE- 3, \(13.56\mathrm{MHz}\) , \(17\mathrm{W}\) ) was carried out to introduce nanopores for application of 2DZIF film as selective layer in the membrane. Briefly, the atmosphere in the plasma chamber was exchanged by \(\mathrm{O_2}\) flow to pressure around \(50\mathrm{mTorr}\) . Then a plasma was generated for \(4\mathrm{s}\) to etch SLG to obtain NG. After the plasma, a solution of PTMSP in toluene ( \(1.25\mathrm{wt}\%\) ) was spin- coated on NG at \(1000\mathrm{rpm}\) for \(30\mathrm{s}\) and \(2000\mathrm{rpm}\) for \(30\mathrm{s}\) , respectively. The sample was dried in ambient air at room temperature overnight. Then, copper foil was etched by a combination of \(\mathrm{FeCl_3}\) ( \(0.5\mathrm{M}\) in water), HCl ( \(0.1\mathrm{M}\) in water) and water. The floating graphene/PTMSP film was transferred on the surface of a \(29\mathrm{ml}\) of aqueous solution of \(\mathrm{Zn(NO_3)_2}\) ( \(2\mathrm{mmol / L}\) ) in a petri dish, and an aqueous solution of \(2\mathrm{- mlm}\) ( \(1\mathrm{mL}\) , \(0.5\mathrm{mol / L}\) ) was added in. After \(2\mathrm{min}\) of reaction the resulted 2DZIF/graphene/PTMSP film was transferred to substrate (e.g., macroporous W support hosting \(5\mu \mathrm{m}\) holes over a \(1\mathrm{mm}^2\) area for membranes) for further characterizations or applications. For the synthesis of centimeter- scale 2DZIF membrane, \(1\mathrm{wt}\%\) of Teflon AF in GOLDEN perfluorinated fluid was spin- coated on NG at \(300\mathrm{rpm}\) for \(60\mathrm{s}\) , and heated at \(60^{\circ}\mathrm{C}\)
+
+<--- Page Split --->
+
+for \(3\mathrm{h}\) . After that, the sample was put into a homemade membrane module (Supplementary Fig. 19) while Cu foil facing up. This allowed etching of Cu foil in the membrane module by \(10\mathrm{wt}\% \mathrm{Na}_2\mathrm{S}_2\mathrm{O}_8\) aqueous solution exposing NG surface for 2DZIF synthesis. Finally, 2DZIF film was synthesized by exposing NG to growth solution for \(10\mathrm{min}\) at room temperature (2 mM \(\mathrm{Zn(NO_3)_2}\) and \(16\mathrm{mM}2\mathrm{- }\mathrm{mM}\) ).  
+
+## Sample preparation of 2DZIF on graphene for AFM  
+
+For the AFM sample preparation of 2DZIF film on graphene/PTMSP substrate, the 2DZIF/graphene/PTMSP film was transferred on \(\mathrm{Si / SiO_2}\) wafer with 2DZIF layer facing the wafer. Then the sample was annealed at \(70^{\circ}\mathrm{C}\) for \(4\mathrm{h}\) , to increase the adhesion between film and \(\mathrm{Si / SiO_2}\) wafer. After that, the sample was immersed in toluene for \(12\mathrm{h}\) , to remove PTMSP layer.  
+
+The AFM image of 2DZIF film with triangular morphology after \(5\mathrm{min}\) of water etching was collected directly on the film with 2DZIF layer facing up. Specifically, the 2DZIF/graphene/PTMSP film with triangular morphology was first scooped by glass slide with 2DZIF layer facing the slide, and a \(\mathrm{Si / SiO_2}\) wafer attached with double- sided carbon tape was pressed onto PTMSP layer. As a result, the 2DZIF/graphene/PTMSP film was transferred onto \(\mathrm{Si / SiO_2}\) wafer, resulting in 2DZIF layer facing up. AFM measurement of 2DZIF film on HOPG, \(\mathrm{Si / SiO_2}\) and graphene/ \(\mathrm{Si / SiO_2}\) was carried out directly on the sample without any treatment.  
+
+## Sample preparation for TEM  
+
+Similar to sample preparation for AFM, 2DZIF/graphene/PTMSP film was transferred on TEM grid with 2DZIF layer facing the TEM grid. Then the sample was annealed at \(70^{\circ}\mathrm{C}\) for \(4\mathrm{h}\) , to increase the adhesion between film and TEM grid. After that, the sample was immersed in toluene for \(12\mathrm{h}\) , to remove the PTMSP layer.  
+
+For the electron patterning sample, aZIF layer was grown on a silicon nitride supports. Prior to the ZIF deposition, the silicon nitride supports were treated with oxygen plasma for \(10\mathrm{min}\) (29.6 W, \(400\mathrm{mTorr}\) oxygen pressure) in a plasma cleaner (Harrick Plasma) to improve the surface reactivity.  
+
+## Electron beam patterning of aZIF films  
+
+For negative tone patterning, aZIF films were exposed to electron beam using a Thermo Fisher Helios G4 UC Dual Beam microscope operating at \(20\mathrm{kV}\) accelerating voltage and \(400\mathrm{pA}\)
+
+<--- Page Split --->
+
+beam current. The areal doses were \(80\mathrm{mC / cm^2}\) for all patterns. After exposure, the aZIF films were developed in deionized water for \(24\mathrm{h}\) and blow dried in a stream of nitrogen gas.  
+
+For positive tone patterning, the aZIF films were first placed in a \(60\mathrm{mL}\) PFA vessel with a bed of \(0.2\mathrm{g}4,5\) - dichloroimidazole (dcIm) and heated at \(75^{\circ}\mathrm{C}\) for \(1.5\mathrm{h}\) . The dcIm- treated aZIF was then exposed to electron beam using a Thermo Fisher Helios G4 UC Dual Beam microscope operating at \(20\mathrm{kV}\) accelerating voltage and \(100\mathrm{pA}\) beam current. The areal doses were \(2\mathrm{mC / cm^2}\) for all patterns. After exposure, the aZIF films were immersed in NMP and acetone each for \(10\mathrm{s}\) and blow dried in a stream of nitrogen gas.  
+
+## Spin-coating of aZIF films on silicon wafers  
+
+Aqueous solutions of \(\mathrm{Zn(NO_3)_2}\) (4 mmol/L) and 2- Im (32 mmol/L) were mixed in a T connector (0.5 mm I.D.) at \(1\mathrm{mL / min}\) injection rate, respectively, using a two- channel syringe pump. The mixed solution was spin- coated on silicon wafers while spinning at speeds between 500 and 2000 rpm for \(1\mathrm{min}\) , followed by \(10\mathrm{s}\) spinning at 2000 rpm to completely dry the film. The spin- coated films were exposed to electron beam using line patterns of \(100\mathrm{nm}\) width and \(400\mathrm{nm}\) spacing and developed in deionized water for \(24\mathrm{h}\) . The height of the line patterns were measured to determine the thickness of spin- coated films.  
+
+## Structural simulation  
+
+The simulation of 2DZIF structure was carried out by the DFT calculations. At first, the reported ZIF- L structure was imported into Forcite module in Material Studio software (Accelrys, San Diego, CA) to calculate the initial structural model, and unit- cell was set to be orthorhombic and \(a = 24.0\mathrm{\AA}\) , \(b = 20.0\mathrm{\AA}\) , \(c = 20.0\mathrm{\AA}\) , \(\alpha = 90^{\circ}\) , \(\beta = 90^{\circ}\) , \(\gamma = 90^{\circ}\) , respectively, while the connectivity was kept. The calculation task was geometry optimization. The quality was set to be fine, and 'smart' algorithm was selected. After that, van der Waals DFT calculations were performed using the Quantum ESPRESSO package<sup>66,67</sup>. The Brillouin zone was sampled at the gamma point. An energy cutoff of 60 Ry was used for the plane wave expansion of the wavefunctions. A kinetic energy cutoff of 480 Ry on the charge was used together with ultra- soft pseudopotentials<sup>68,69</sup>. The relaxation was performed with the Perdew- Burke- Ernzerhof (PBE) functional<sup>70</sup>. The system was relaxed to the lowest energy configuration of atoms. The surface geometry had been optimized with the convergence thresholds of \(1 \times 10^{-4}\mathrm{Ry}\) and \(3.28 \times 10^{-3}\mathrm{Ry / Bohr}\) for the total energy and forces, respectively.  
+
+## Data availability
+
+<--- Page Split --->
+
+Data supporting the findings in the present work are available in the manuscript or Supplementary Information. Additional data are available from the corresponding author upon reasonable request.  
+
+65. Huang, S. et al. Millisecond lattice gasification for high-density \(\mathrm{CO_2}\) - and \(\mathrm{O_2}\) -sieving nanopores in single-layer graphene. Sci. Adv. 7: eabf0116 (2021).  
+
+66. Giannozzi, P. et al. QUANTUM ESPRESSO: a modular and open-source software project for quantum simulations of materials. J. Phys. Condens. Matter 21, 390052 (2009).  
+
+67. Giannozzi, P. et al. Advanced capabilities for materials modelling with Quantum ESPRESSO. J. Phys. Condens. Matter 29, 465901 (2017).  
+
+68. Lejaeghere, K. et al. Reproducibility in density functional theory calculations of solids. Science 351, 6280 (2016).  
+
+69. Prandini, G., Marrazzo, A., Castelli, I. E., Mounet, N. & Marzari, N. Precision and efficiency in solid-state pseudopotential calculations. npj Comput. Mater. 4, 72 (2018).  
+
+70. Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865-3868 (1996).  
+
+Acknowledgments The authors acknowledge Dr. Dmitry Chernyshov and Dr. Diadkin Vadim at beamline BM01, the-Swiss-Norwegian Beamlines (SNBL), European Synchrotron Radiation Facility (ESRF) for assistance with synchrotron GIXRD experiments (doi:10.15151/ESRF-ES-670011338. ), Dr. Pascal Alexander Schouwink from EPFL for the help of XRD data. Dr. Mounir Mensi and Mojtaba Rezaei from EPFL for the help of XPS and AFM. This project is primarily supported by European Research Council Starting Grant (805437-UltimateMembranes). Parts of the work was supported by Swiss National Science Foundation (SNSF) Assistant Professor Energy Grant (PYAPP2_173645) and SNSF project (514601); Y. M. and M. T. acknowledge funding by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Division of Chemical Sciences, Geosciences and Biosciences under Award DE-SC0021212.  
+
+Author contributions K. A. and Q. L. conceived the project and wrote the manuscript. Q. L., Y. M., D. B., and J. H. prepared the samples. Q. L., Y. M. and S. L. performed AFM
+
+<--- Page Split --->
+
+measurements. Q. L. and Y. M. collected SEM data. Q. L. performed XRD measurement. S. L. collected XPS data. L. V. and H. C. collected TEM images and ED data. Q. L. and M. V. carried out structural simulations. C. C. and Y. H. helped with AC- HRTEM images. Q. L. and S. S. carried out gas permeance experiments. M. T. conceived and Y. M. performed the aZIF film deposition and e- beam patterning experiments. All authors discussed the results and commented on the manuscript.  
+
+Competing interests: A patent application based on the findings reported here has been filed.  
+
+## Additional information  
+
+Supplementary information The online version contains supplementary material available at Correspondence and requests for materials should be addressed to Kumar Varoon Agrawal. Peer review information Nature thanks anonymous reviewer(s) for their contribution to the peer review of this work.  
+
+Reprints and permissions information is available at http://www.nature.com/reprints.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+- Sl.pdf
+
+<--- Page Split --->

preprint/preprint__09bb925dea8007aaf7d14d8f79096de27fcb7fe802393ba1d044ff47a75f57c2/preprint__09bb925dea8007aaf7d14d8f79096de27fcb7fe802393ba1d044ff47a75f57c2_det.mmd

@@ -0,0 +1,516 @@
+<|ref|>title<|/ref|><|det|>[[44, 106, 945, 210]]<|/det|>
+# Nanometer-thick crystalline and amorphous zeolitic imidazolate framework films for membrane and patterning applications  
+
+<|ref|>text<|/ref|><|det|>[[44, 229, 636, 293]]<|/det|>
+Qi Liu École Polytechnique Fédérale de Lausanne Yurun Miao  
+
+<|ref|>text<|/ref|><|det|>[[55, 297, 635, 316]]<|/det|>
+Johns Hopkins University https://orcid.org/0000- 0001- 6429- 8297  
+
+<|ref|>text<|/ref|><|det|>[[44, 322, 544, 363]]<|/det|>
+Luis Francisco Villalobos Yale University https://orcid.org/0000- 0002- 0745- 4246  
+
+<|ref|>text<|/ref|><|det|>[[44, 369, 430, 410]]<|/det|>
+Shaoxian Li École Polytechnique Fédérale de Lausanne  
+
+<|ref|>text<|/ref|><|det|>[[44, 416, 430, 457]]<|/det|>
+Deepu J. Babu École Polytechnique Fédérale de Lausanne  
+
+<|ref|>text<|/ref|><|det|>[[44, 463, 867, 504]]<|/det|>
+Cailing Chen King Abdullah University of Science and Technology https://orcid.org/0000- 0003- 2598- 1354  
+
+<|ref|>text<|/ref|><|det|>[[44, 509, 430, 550]]<|/det|>
+Heng- Yu Chi École Polytechnique Fédérale de Lausanne  
+
+<|ref|>text<|/ref|><|det|>[[44, 555, 430, 596]]<|/det|>
+Mohammad Tohidi Vahdat École Polytechnique Fédérale de Lausanne  
+
+<|ref|>text<|/ref|><|det|>[[44, 602, 430, 643]]<|/det|>
+Jian Hao École Polytechnique Fédérale de Lausanne  
+
+<|ref|>text<|/ref|><|det|>[[44, 649, 430, 689]]<|/det|>
+Shuqing Song École Polytechnique Fédérale de Lausanne  
+
+<|ref|>text<|/ref|><|det|>[[44, 695, 867, 736]]<|/det|>
+Yu Han King Abdullah University of Science and Technology https://orcid.org/0000- 0003- 1462- 1118  
+
+<|ref|>text<|/ref|><|det|>[[44, 741, 636, 782]]<|/det|>
+Michael Tsapatsis Johns Hopkins University https://orcid.org/0000- 0001- 5610- 3525  
+
+<|ref|>text<|/ref|><|det|>[[44, 787, 787, 829]]<|/det|>
+Kumar Varoon Agrawal ( kumar.agrawal@epfl.ch) École Polytechnique Fédérale de Lausanne https://orcid.org/0000- 0002- 5170- 6412  
+
+<|ref|>text<|/ref|><|det|>[[44, 869, 101, 886]]<|/det|>
+Article  
+
+<|ref|>text<|/ref|><|det|>[[44, 906, 135, 924]]<|/det|>
+Keywords:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[44, 46, 314, 64]]<|/det|>
+**Posted Date:** March 14th, 2023  
+
+<|ref|>text<|/ref|><|det|>[[44, 84, 474, 103]]<|/det|>
+**DOI:** https://doi.org/10.21203/rs.3.rs-2666142/v1  
+
+<|ref|>text<|/ref|><|det|>[[44, 120, 910, 164]]<|/det|>
+**License:** © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[44, 182, 910, 225]]<|/det|>
+**Additional Declarations:** Yes there is potential Competing Interest. A patent application based on the finding reported in the manuscript is filed.  
+
+<|ref|>text<|/ref|><|det|>[[44, 259, 945, 303]]<|/det|>
+**Version of Record:** A version of this preprint was published at Nature Materials on September 21st, 2023. See the published version at https://doi.org/10.1038/s41563-023-01669-z.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[176, 90, 823, 140]]<|/det|>
+# Nanometer-thick crystalline and amorphous zeolitic imidazolate framework films for membrane and patterning applications  
+
+<|ref|>text<|/ref|><|det|>[[130, 167, 870, 238]]<|/det|>
+Qi Liu<sup>1†</sup>, Yurun Miao<sup>2</sup>, Luis Francisco Villalobos<sup>1</sup>, Shaoxian Li<sup>1</sup>, Deepu J. Babu<sup>1†</sup>, Cailing Chen<sup>4</sup>, Heng- Yu Chi<sup>1</sup>, Mohammad Tohidi Vahdat<sup>1,5</sup>, Jian Hao<sup>1</sup>, Shuqing Song<sup>1</sup>, Yu Han<sup>4</sup>, Michael Tsapatsis<sup>2,3</sup>, Kumar Varoon Agrawal<sup>1\*</sup>  
+
+<|ref|>text<|/ref|><|det|>[[118, 293, 880, 336]]<|/det|>
+1. Laboratory of Advanced Separations, École Polytechnique Fédérale de Lausanne (EPFL), 1950 Sion, Switzerland.  
+
+<|ref|>text<|/ref|><|det|>[[118, 351, 880, 395]]<|/det|>
+2. Department of Chemical and Biomolecular, Engineering & Institute for NanoBioTechnology, Johns Hopkins University, Baltimore, MD, USA  
+
+<|ref|>text<|/ref|><|det|>[[118, 407, 744, 427]]<|/det|>
+3. Applied Physics Laboratory, Johns Hopkins University, Laurel, MD, USA  
+
+<|ref|>text<|/ref|><|det|>[[118, 439, 881, 508]]<|/det|>
+4. Advanced Membranes and Porous Materials Center, Physical Sciences and Engineering Division, King Abdullah University of Science and Technology, Thuwal 23955-6900, Saudi Arabia.  
+
+<|ref|>text<|/ref|><|det|>[[118, 520, 880, 565]]<|/det|>
+5. Theory and Simulation of Materials (THEOS) and National Centre for Computational Design and Discovery of Novel Materials (MARVEL), EPFL, Lausanne, Switzerland.  
+
+<|ref|>text<|/ref|><|det|>[[118, 577, 880, 620]]<|/det|>
+<sup>†</sup>Present address: College of Chemistry, Chemical Engineering and Materials Science, Soochow University, Suzhou, 215123, China.  
+
+<|ref|>text<|/ref|><|det|>[[118, 633, 881, 677]]<|/det|>
+<sup>†</sup>Present address: Materials Science and Metallurgical Engineering, Indian Institute of Technology, Hyderabad, Telangana 502 284, India.  
+
+<|ref|>text<|/ref|><|det|>[[118, 690, 515, 708]]<|/det|>
+\*Corresponding author: kumar.agrawal@epfl.ch
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 86, 198, 101]]<|/det|>
+## Abstract  
+
+<|ref|>text<|/ref|><|det|>[[115, 105, 883, 600]]<|/det|>
+Zeolitic imidazolate frameworks (ZIFs) are a subset of metal- organic frameworks (MOFs) with more than 200 characterized crystalline and amorphous networks made of divalent transition metal centers (e.g., \(\mathrm{Zn^{2 + }}\) and \(\mathrm{Co^{2 + }}\) ) linked by imidazolate linkers. ZIF thin films have been pursued intensively motivated by the desire to prepare membranes for selective gas and liquid separations. To achieve membranes with high throughput, as in A- scale biological channels with nanometer- scale pathlengths, ZIF films with the minimum possible thickness, down to just one unit cell, are highly desired. Control of ZIF film thickness at the 10- nm- scale may also enable emerging, MOF- inspired, applications including patterned crystalline MOF films, and amorphous organic- inorganic resists for high- resolution electron- beam (e- beam) and extreme UV (EUV) lithography. However, the state- of- the- art methods yield ZIF films with thicknesses exceeding 40 nanometers. Here, we report a deposition method from ultra- dilute precursor mixtures that within minutes yields uniform ZIF deposits with nm- scale thickness control. On crystalline substrate such as graphene, two- dimensional crystalline ZIF (2DZIF) film with thickness of a unit- cell could be achieved, which composed of a six- membered zinc- imidazolate coordination ring enabling record- high \(\mathrm{H}_2\) permselective separation performance. Deposition under identical conditions on amorphous substrates yields macroscopically smooth amorphous ZIF (aZIF) films, which can be used as negative- and positive- tone resists yielding pattern features down to \(20 \mathrm{nm}\) . The method reported here will likely accelerate the development of 2D crystalline and ultrathin amorphous MOF films for applications ranging from separation membranes to sensors and patterning for microelectronic applications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 644, 216, 660]]<|/det|>
+## Main Text:  
+
+<|ref|>text<|/ref|><|det|>[[115, 682, 884, 899]]<|/det|>
+ZIFs \(^{1,2}\) are a class of MOFs that hold promise for applications in molecular separations \(^{3 - 10}\) , patterning \(^{11 - 15}\) and sensing \(^{16}\) . Their chemical and physical properties have been widely explored as a function of framework flexibility \(^{17 - 20}\) as well as structural defects \(^{21,22}\) . The realization of two- dimensional (2D) ZIF films with thickness down to that afforded by a single structural building unit is highly desired to make ZIF analogues to graphene and related 2D materials with an added advantage; the intrinsic nanoporosity of ZIF can be used to separate molecules while maximizing the permselective flux \(^{23}\) . Another highly desirable feature is a nanometer- scale control over the film thickness, which can allow one to fabricate nanoscale patterns, which are desirable for incorporation in microelectronic devices \(^{24}\) , and for the development of
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 883, 274]]<|/det|>
+next generation organic- inorganic high- resolution photo- and e- beam- resists<sup>17</sup>. However, the realization of 2D crystalline and ultrathin amorphous ZIF films has remained elusive. While layered ZIFs such as ZIF- L<sup>25</sup>, \(\mathrm{Zn_2(bim)_4^{26}}\) , and analogs<sup>27</sup> have been reported, the individual ZIF layers in these materials have a small aspect ratio which prevents the realization of continuous 2D ZIF films with structural uniformity over a macroscopic (e.g., wafer) length scale. The state- of- the- art of ZIF deposition methods yields polycrystalline films with thickness larger than \(40 \mathrm{nm}\) nanometer<sup>28- 31</sup>. This is mainly due to difficulty in achieving in- plane film growth without film thickening.  
+
+<|ref|>text<|/ref|><|det|>[[117, 295, 883, 536]]<|/det|>
+Considerable knowledge exists on ZIF/MOF crystal nucleation and growth in solution<sup>32- 37</sup>. Based upon data from synchrotron X- ray scattering, density- functional theory (DFT) and molecular dynamics simulations, and other techniques, it is generally accepted that ZIF formation involves a sequence of events starting from the formation of small ( \(\sim 1 \mathrm{nm}\) ) metastable prenucleation clusters, which evolve through aggregation followed by intra- aggregate ZIF nucleation and growth. Recent studies on surface- directed MOF growth<sup>38- 46</sup> indicate that the diffusion of MOF precursors in the vicinity of the 2D material, and the MOF- 2D material interactions, are key to regulate the crystallinity of the MOF film and the ability to maintain in- plane/horizontal growth (desired for ultrathin films) versus out- of- plane/vertical (undesired) growth.  
+
+<|ref|>text<|/ref|><|det|>[[117, 558, 883, 775]]<|/det|>
+Herein, we report macroscopically uniform 2D ZIF films with exquisite nanometer- scale control over the film thickness by suppressing the out- of- plane growth by using an ultradilute growth solution. The ultralow precursor concentration restricts homogeneous nucleation in the solution and facilitates the growth of nanometer- thick films over an immersed substrate with deposition timescales of a few minutes. The film crystallinity is determined by the interaction of molecular precursors with the substrate ranging from substrate- registry- determined ordered film to amorphous films in the absence of any crystallographic registry. The film thickness could be controlled with a resolution of a single layer by controlling the deposition time and number of coatings.  
+
+<|ref|>text<|/ref|><|det|>[[117, 796, 883, 915]]<|/det|>
+The ZIF films were synthesized by immersing a substrate in an ultradilute precursor solution ( \(\leq 2 \mathrm{mM} \mathrm{Zn}^{2 + }\) and \(\leq 16 \mathrm{mM} 2\) - methylimidazole (2- mIm), respectively) for a few minutes (Fig. 1a). The use of such ultradilute solutions for the growth of ZIF films has not been reported before (Fig. 1b, Supplementary Table 1). They were used here in an effort to suppress the homogeneous nucleation in the bulk solution. With a diminished nuclei population in the bulk
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 883, 225]]<|/det|>
+solution, the attachment of preformed nuclei to the substrate can be reduced or eliminated. This is expected to promote film growth by the assembly of molecular precursors on the substrate. Since this thin film growth mode is anticipated to be sensitive on the type of the substrate, we carried out synthesis using distinct substrates: (i) graphitic substrates with atomically smooth terrace such as highly oriented pyrolytic graphite (HOPG) or graphene, (ii) Si/SiO₂ wafer with a 300- nm- thick oxide layer, and (iii) single crystal sapphire (Al₂O₃).  
+
+<|ref|>text<|/ref|><|det|>[[115, 247, 884, 686]]<|/det|>
+ZIF films prepared on HOPG using a growth solution of 1 mM \(\mathrm{Zn^{2 + }}\) and 8 mM 2- mM and reaction time of 5 min were examined by optical and scanning electron microscopy (SEM) (Supplementary Fig. 1). A sharp change in contrast was observed at the air/precursor- solution interface beyond which the film had a uniform contrast indicating that the film was smooth, continuous, and macroscopically uniform. Atomic force microscopy (AFM) imaging near the interface confirmed that the ZIF film is indeed continuous and has a thickness of approximately 2 nm (Fig. 1c and d). When the synthesis time was reduced to 2 min, we observed a submonolayer film with micrometer- sized domains (Supplementary Fig. 2). The domains were faceted and had thickness of 2 nm, consistent with the thickness of the continuous film indicating that the film is crystalline consisting of micron- sized grains. We could obtain 4 and 6 nm thick films by increasing the growth time from 5 min to 10 and 15 min, respectively (Fig. 1e and Supplementary Fig. 3a- f). Increasing growth time to 20 min further did not lead to thicker film, indicating precursor depletion (Supplementary Fig. 3g- i). Thicker (8 nm) films could be obtained by doubling the precursor concentration (Supplementary Fig. 3j- l). A discrete, 2 nm, increase in film thickness further suggests a crystalline order. A fitting of film thickness with the number of probable layers yielded a monolayer thickness of 2 nm (Fig. 1e). Macroscopically large ZIF films spanning several centimeters in width could be obtained on chemical vapor deposition (CVD) derived graphene film resting on a Cu foil (Fig. 1f and 1g).  
+
+<|ref|>text<|/ref|><|det|>[[117, 707, 884, 849]]<|/det|>
+We also obtained macroscopically smooth, continuous, and uniform ZIF films on Si/SiO₂ wafers (Fig. 1h). AFM of one of these films, prepared using 2 mM \(\mathrm{Zn^{2 + }}\) and 16 mM 2- mM and deposition time of 10 s, indicated that the film is smooth with thickness near 8 nm (Fig. 1i and 1j). Ellipsometry of several ZIF films on Si/SiO₂ wafers, prepared by varying the synthesis time, indicated that the film thickness could be tuned in the range of 8- 18 nm consistent with the corresponding AFM data (Supplementary Fig. 4).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 80, 875, 770]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 777, 883, 893]]<|/det|>
+<center>Fig. 1. Synthesis ZIF films from ultradilute solutions. a, Schematic of the ZIF film synthesis approach. b, Composition diagram comparing the precursor solution composition used in this study with those reported in the literature (Supplementary Table 1). AFM (c) and the corresponding height profile (d) of a monolayer ZIF film on HOPG. e, Monolayer and multilayer ZIF films on HOPG with discrete thicknesses as a function of synthesis time. Error </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 883, 201]]<|/det|>
+bars in this figure represent the standard deviation of measurement. f, Optical image of a 2DZIF film on CVD graphene resting on a Cu foil. g, Scanning electron microscopy (SEM) image of a 2DZIF film on CVD graphene. The image was compiled by combining \(1155 (35 \times 33)\) images by scanning the whole surface of the large sample. h, SEM and optical images of a ZIF film on \(\mathrm{Si / SiO_2}\) . AFM (i) and the corresponding height profile (j) of the ZIF film on \(\mathrm{Si / SiO_2}\) .  
+
+<|ref|>text<|/ref|><|det|>[[116, 222, 883, 563]]<|/det|>
+Graphene supported ZIF film could be suspended on a holey transmission electron microscopy (TEM) grid (Fig. 2a). The film was devoid of large crystals and appeared uniform. Selected area electron diffraction (SAED) from a micrometer- sized area yielded three sets of diffraction patterns (Fig. 2b). The first two sets ((01), highlighted with green circles) had six- fold symmetry originating from two slightly misoriented (by \(3.0^{\circ}\) ) grains of graphene, while the last set had two- fold symmetry and belonged to a single grain of ZIF (highlighted with white circles), confirming that ZIF prepared on graphene was crystalline with grains at least a micron in size consistent with the AFM- based imaging of grains in the submonolayer film (Supplementary Fig. 2). Hereafter, the ZIF films on graphitic substrates are referred to as 2DZIF. The fact that a single 2DZIF grain could grow over two slightly misoriented graphene grains indicates that the growth could accommodate a small mismatch in its registry with the substrate. Diffraction pattern from 2DZIF, typically representing a single grain, was observed from every single spot over a large area. Based on the diffraction pattern, \(a\) and \(b\) lattice parameters of 2.4 and \(2.0 \mathrm{nm}\) , respectively, could be fitted.  
+
+<|ref|>text<|/ref|><|det|>[[116, 584, 883, 872]]<|/det|>
+We carried out synchrotron grazing incidence X- ray diffraction (GIXRD) of a 10- nm- thick ZIF film on graphene resting on a \(\mathrm{Si / SiO_2}\) wafer (Supplementary Fig. 5- 7). The in- plane GIXRD pattern revealed sharp diffraction peaks, consistent with the peak positions obtained by the radial integration of the SAED pattern (Fig. 2c and Supplementary Table 2), confirming that the film formed on the graphitic substrate exhibits crystalline order. In contrast, we did not observe diffraction from the ZIF films prepared directly on \(\mathrm{Si / SiO_2}\) wafer without the graphene layer indicating that these films were amorphous (hereafter referred to as aZIF films, Fig. 2c). When a single crystal sapphire ( \(\mathrm{Al_2O_3}\) ) was used as the substrate, highly crystalline 2DZIF film was formed (Supplementary Fig. 8 and 9). The presence of the order in the ZIF film when prepared over a crystalline substrate and the lack of an order when prepared over the amorphous substrate indicates a strong role of substrate registry in the formation of the ordered 2DZIF film.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 80, 875, 586]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 586, 881, 900]]<|/det|>
+<center>Fig. 2. Structure determination of 2DZIF films. a, Bright-field TEM image of the 2DZIF film supported on suspended graphene, and (b) its corresponding SAED pattern. The pattern from graphene is identified with green circles and those from 2DZIF with white circles. c, In-plane GIXRD data from an aZIF film (top) and a 2DZIF film (middle) prepared on \(\mathrm{Si / SiO_2}\) and graphene/Si/SiO2, respectively, along with a radially integrated trace (bottom) of the SAED pattern shown in (B). d, N1s XPS spectra from ZIF-8, ZIF-L, aZIF and 2DZIF films. The N-Zn and N-H coordination environments are shown on the right. e, DFT-relaxed structure of the 2DZIF and a visualization of the 6-MR (right). f, HRTEM image of the 2DZIF film lying flat on the \(hk0\) plane, resting on suspended graphene, and (g) corresponding Fourier transform compared with the simulated diffraction pattern from the proposed structure oriented along the \(c\) -out-of-plane direction. h, Left: CTF-corrected image of the highlighted area in (f) based on a defocus value of -130 nm analyzed from the Thon rings in the Fourier transform pattern. Right: simulated projected potential map along the [001] direction of 2DZIF. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[116, 82, 883, 300]]<|/det|>
+X- ray photoelectron spectroscopy (XPS) of the 2DZIF and aZIF films was carried out to gain further insights into their coordination environments (Fig. 2d). The N1s XPS data of 2DZIF when compared to that of ZIF- L layers and a prototypical non- layered ZIF (ZIF- 8) revealed that both 2DZIF and ZIF- L yield two peaks (399.0 and \(400.2\mathrm{eV}\) corresponding to N- Zn and N- H bonds, respectively) in contrast to a single peak (399.0 eV) from the ZIF- 8 crystals. This is consistent with the presence of abundant surface terminations (N- H) in the 2DZIF layers. In comparison, the population of N- H species was significantly diminished for aZIF indicating a nonlayered amorphous structure. The Zn2p XPS spectra was similar for all samples (Supplementary Fig. 10) indicating a similar coordination environment for Zn.  
+
+<|ref|>text<|/ref|><|det|>[[116, 320, 883, 538]]<|/det|>
+To gain insight into the structure of 2DZIF, structural relaxation based on DFT was carried out starting with the \(a\) and \(b\) lattice parameters obtained by SAED. The \(c\) - axis parameter was estimated by the AFM measurements (2 nm), and was subsequently relaxed by DFT calculations. The relaxed structure has an orthorhombic space group Cmce with the following structural parameters; \(a = 24.196\mathrm{\AA}\) , \(b = 19.719\mathrm{\AA}\) , \(c = 20.908\mathrm{\AA}\) , \(\alpha = 90^{\circ}\) , \(\beta = 90^{\circ}\) , and \(\gamma = 90^{\circ}\) (Supplementary Table 2). The layer in 2DZIF is composed of alternating 4- member ring (MR) and 6- MR chains while terminal 2- mIm linkers are present on both sides of the layer (Fig. 2e and 3c). The pore aperture of 2DZIF is constituted by the 6- MR (Fig. 2e) which is attractive for gas separation.  
+
+<|ref|>text<|/ref|><|det|>[[116, 558, 883, 800]]<|/det|>
+Aberration- corrected high- resolution TEM (AC- HRTEM) imaging of the 2DZIF film suspended on a TEM grid was carried out along the [001] crystallographic direction (Fig. 2f). The imaging was carried out using a low- dose beam condition<sup>47</sup> to minimize damages to the beam sensitive 2DZIF lattice. Indeed, the obtained HRTEM image revealed the high crystallinity of the 2DZIF film. The corresponding Fourier transform validated the \(c\) - out- of- plane orientation of the film and was consistent with the simulated electron diffraction pattern from a film lying flat along the same orientation (Fig. 2g). Projection along the \(c\) - out- of- plane axis from the contrast transfer function (CTF) corrected image revealed alternating chains of 4- MR and 6- MR (Fig. 2h, left), consistent with the simulated [001]- projected electrostatic potential map of 2DZIF structure obtained by DFT structural relaxation (Fig. 2h, right).  
+
+<|ref|>text<|/ref|><|det|>[[117, 820, 881, 915]]<|/det|>
+The registry or the lack of registry of 2DZIF with the underlying substrate plays an important role in determining its structure and morphology especially when contrasted against the closely related material, ZIF- L. Fig. 3a highlights the morphological differences in ZIF- L and 2DZIF. While the layers in ZIF- L and 2DZIF are stacked along the \(c\) - axis, the former grows as a leaf
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[116, 82, 884, 325]]<|/det|>
+shaped layered crystals (Supplementary Fig. 11) whereas the latter can form macroscopically uniform monolayer films. The unique leaf shape is formed because ZIF- L layers stack with first progressively increasing and then progressively decreasing lateral size along the \(b\) - axis. An analysis of orientation of 2DZIF grain over graphene by SAED showed that the 2DZIF films crystallized maintaining a fixed set of orientation with graphene (Supplementary Fig. 12), indicating that substrate registry indeed plays a role in promoting in- plane growth of 2DZIF (Fig. 3b and Supplementary Note 1). While both ZIF- L and 2DZIF have orthorhombic lattices, the unit- cell parameters of 2DZIF grown on graphene, obtained by DFT relaxation, are distinct from those of ZIF- L where the latter has a significantly shorter parameter along the \(b\) (17.060 Å) axis compared to the former (19.719 Å, Fig. 3c and Supplementary Table 2).  
+
+<|ref|>text<|/ref|><|det|>[[116, 345, 884, 564]]<|/det|>
+The grains of 2DZIF could be visualized by partial etching of 2DZIF films based on the well documented dissolution of ZIFs in water (Fig. 3d)48. After partial dissolution, the grain shape was triangular with a lateral size of 1- 2 \(\mu \mathrm{m}\) (Fig. 3e) consistent with earlier observations of domains in the sub- monolayer film. The three sides of the triangular grains could be assigned to be (110), (110) and (100) lattice planes, respectively, reported to be the minimum surface energy planes for ZIF layers (Supplementary Fig. 13)49. AFM images (Fig. 3f) confirmed that the grains have uniform thickness of \(\sim \mathrm{ca. 2 nm}\) consistent with the structure of 2DZIF. SAED of this sample showed the same pattern with 2DZIF confirming that the triangular domains were indeed 2DZIF (Supplementary Fig. 14).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 81, 884, 450]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 460, 883, 628]]<|/det|>
+<center>Fig. 3. 2DZIF structure and its relationship with ZIF-L. a, Schematic contrasting the arrangement of layers within a ZIF-L crystal with that of a monolayer 2DZIF. b, Registry between 2DZIF and supercell of graphene based on SAED data in Supplementary Fig. 12. c, Structures of a ZIF-L layer (left) and 2DZIF (right) viewed along the [001], [100], and [010] directions. d, Schematic illustrating the etching of 2DZIF in water. SEM (e) and AFM (f) images of the triangular grains of 2DZIF obtained by a short etching in water. g, AFM height profile corresponding to the line in (f). </center>  
+
+<|ref|>text<|/ref|><|det|>[[117, 640, 884, 909]]<|/det|>
+The \(3.2\mathrm{\AA}\) gap in the 6- MR of 2DZIF is attractive for sieving \(\mathrm{H}_2\) (kinetic diameter of \(2.89\mathrm{\AA}\) from larger gas molecules such as \(\mathrm{CO_2}\) ( \(3.30\mathrm{\AA}\) ), \(\mathrm{N}_2\) ( \(3.64\mathrm{\AA}\) ), and \(\mathrm{CH_4}\) ( \(3.80\mathrm{\AA}\) ) \(^{50,51}\) . The 2DZIF film is mechanically robust with Young's modulus of \(8.1\pm 2.1\) GPa (Supplementary Fig. 15), comparable to that of the three- dimensional analogs \(^{52}\) . Therefore, we probed the \(\mathrm{H}_2\) - sieving performance on the 2DZIF film. For this, 2DZIF films were grown on nanoporous graphene (NG, Supplementary Fig. 16) mechanically reinforced with a dense 250- nm- thick poly[1- (trimethylsilyl)propyne] (PTMSP) film (Supplementary Fig. 17 and 18) where the NG/PTMSP film acts as a support film (Fig. 4a). The pores in NG were intentionally designed to be large \((1.8\pm 1.2\mathrm{nm})^{53}\) to rule out any molecular sieving from NG and to allow the determination of \(\mathrm{H}_2\) - sieving from the 2DZIF film. The 2DZIF films, resting on a macroporous metal foil support (pore size of \(5\mu \mathrm{m}\) , area of \(1\mathrm{mm}^2\) ), exhibited a molecular cut- off for molecules larger than \(\mathrm{H}_2\) ,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 82, 883, 177]]<|/det|>
+indicating that gas transport was controlled by the 6- MR of 2DZIF (Fig. 4b and Supplementary Fig. 19 and 20). The \(\mathrm{H}_2\) permeance was large ( \(>15000\) gas permeation units or GPU; 1 GPU = \(3.35 \times 10^{- 10} \mathrm{~mol} \mathrm{~m}^{- 2} \mathrm{~s}^{- 1} \mathrm{~Pa}^{- 1}\) ) similar to that from the support film (Supplementary Table 3 and Supplementary Fig. 19), indicating a negligible transport resistance from the 2DZIF layer.  
+
+<|ref|>text<|/ref|><|det|>[[116, 190, 884, 432]]<|/det|>
+When an equimolar \(\mathrm{H}_2:\mathrm{N}_2\) mixture was probed with feed pressure of 2 bar, a \(\mathrm{H}_2\) permeance of 17300 GPU with a \(\mathrm{H}_2 / \mathrm{N}_2\) separation factor of 115 could be obtained (Fig. 4c). Another membrane when tested under a high- pressure feed (8 bar), exhibited a high \(\mathrm{H}_2\) flux of 2.8 mol \(\mathrm{m}^{- 2} \mathrm{~s}^{- 1}\) and \(\mathrm{H}_2 / \mathrm{N}_2\) separation factor of 52 (Supplementary Fig. 21). This performance constitutes one of the best combinations of \(\mathrm{H}_2\) flux and \(\mathrm{H}_2 / \mathrm{N}_2\) separation factor (Fig. 4d and Supplementary Fig. 22, 23; Supplementary Table 5). Larger area (centimeter- scale) 2DZIF membrane could be also prepared (Supplementary Fig. 24), thanks to the highly uniform deposition of 2DZIF films on graphene (Fig. 1f and g). They also show attractive \(\mathrm{H}_2\) permselective separation performance (Supplementary Table 4), in agreement with the smaller- area membranes (Supplementary Note 2).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 75, 881, 760]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 777, 883, 896]]<|/det|>
+<center>Fig. 4. Applications of 2DZIF and aZIF. a, Schematic of a 2DZIF film supported on nanoporous graphene reinforced with PTMSP. b, \(\mathrm{H}_2\) , \(\mathrm{CO}_2\) , \(\mathrm{N}_2\) and \(\mathrm{CH}_4\) permeances of the support film (PTMSP-reinforced NG) and the supported 2DZIF film. c, 2DZIF membrane separation performance for an equimolar \(\mathrm{H}_2 / \mathrm{N}_2\) mixed feed. d, Comparison of the \(\mathrm{H}_2 / \mathrm{N}_2\) separation performance of 2DZIF membranes with the state-of-the-art (Supplementary Table </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 883, 201]]<|/det|>
+5). GO, CMP, and HOF refer to graphene oxide, conjugated microporous polymers and hydrogen-bonded organic frameworks, respectively. The dash line indicates Knudsen selectivity for the \(\mathrm{H}_2 / \mathrm{N}_2\) pair. e, Schematic of the patterning process. TEM (f) and AFM (g) images of nanoscale patterns made on an aZIF film. h, AFM height profile corresponding to the line in (g).  
+
+<|ref|>text<|/ref|><|det|>[[115, 210, 883, 902]]<|/det|>
+Amorphous MOFs exhibit unique physical and chemical properties due to the absence of anisotropy and crystalline grains<sup>1,54</sup>. On one hand, they may not have the well- defined pore structures of crystalline MOFs required for certain molecular sieving applications, but at the same time, they do not exhibit grain boundaries and structural anisotropies of crystalline MOFs, which can create film non- uniformities. A potential use of organic- inorganic films is in next generation resists for photolithography<sup>55</sup> in place of currently used polymeric resists, and, for this application, MOF- inspired metal- organic clusters have been proposed for high resolution patterning<sup>13</sup>. As a demonstration of the potential of our deposition method in this emerging application, an aZIF film was deposited on a silicon nitride support and subsequently exposed to a direct- write electron beam using 1:1 line- and space- patterns ranging from 10 to 40 nm in line width (or half pitch) (Fig. 4e). The aZIF films behave similarly to ZIF- L crystals, for which e- beam treatment can induce contrast in water dissolution behavior based on framework densification and disintegration of the ligand molecular structure<sup>56- 58</sup>. After development in water, the irradiated area was preserved while the non- irradiated area was dissolved (Fig. 4f), confirming aZIF as a negative- tone resist. The thickness of the remaining aZIF structure was determined to be \(\sim 25 \mathrm{nm}\) by AFM (Fig. 4g and h). The resolution of the resulting pattern, as exemplified by the well- resolved lines at 20 nm half pitch, is comparable to the state- of- the- art metal- containing resists<sup>13,59</sup>, which are an emerging class of material that hold promise in extreme ultraviolet lithography and electron beam lithography<sup>57- 61</sup>. aZIF can also be patterned in positive- tone mode by a vapor phase ligand pretreatment. The as- deposited aZIF is exposed to the sublimated vapor of 4,5- dichloroimidazole (dcIm) at \(75^{\circ} \mathrm{C}\) for 1.5 h, during which the aZIF matrix is partially exchanged or infiltrated with dcIm ligand. The dcIm- treated film is then exposed to a direct- write electron beam. After development in organic solvents, the irradiated area is removed while non- irradiated area is preserved, which showed similar sensitivity compared to the reported data (Supplementary Fig. 25 and Supplementary Table S6). Furthermore, to improve compatibility with microfabrication processes, aZIF films are spin- coated on silicon wafers, and their thicknesses can be controlled by spin speed (Supplementary Fig. 26- 28). The simple fabrication for ultrathin ZIF films reported in this
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 128]]<|/det|>
+study could accelerate the development of new ZIF- based resist materials for lithographic applications<sup>62- 64</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[117, 142, 883, 358]]<|/det|>
+The method reported here can be extended to other promising MOF structures. 2D film of UiO- 66- NH<sub>2</sub> can also be deposited by on graphene (Supplementary Fig. 29). It makes this approach reported here broad and interesting to develop a number of 2D MOF films in the future. In conclusion, we report the synthesis of ZIFs as macroscopically uniform amorphous and crystalline 2D films from an ultradilute solution. The 2DZIF film yields exceptional H<sub>2</sub>- sieving performance, thanks to the ordered 2D structure with a high density of 6- MR hosting H<sub>2</sub>- selective gap, making such a film the ultimate selective layer for membrane application. In the absence of substrate registry, ultrathin amorphous films are demonstrated, which are promising for advancing the limit of nanoscale patterning.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 380, 248, 397]]<|/det|>
+## Online content  
+
+<|ref|>text<|/ref|><|det|>[[118, 421, 881, 514]]<|/det|>
+Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available online.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 537, 214, 554]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[112, 576, 884, 856]]<|/det|>
+1. Bennett, T. D. & Cheetham, A. K. Amorphous Metal–Organic Frameworks, Acc. Chem. Res. 47, 1555-1562 (2014).  
+2. Banerjee, R. et al. High-Throughput Synthesis of Zeolitic Imidazolate Frameworks and Application to CO<sub>2</sub> Capture, Science 319, 939-943 (2008).  
+3. Li, Y-S. et al. Molecular Sieve Membrane: Supported Metal–Organic Framework with High Hydrogen Selectivity, Angew. Chem. 122, 558-561 (2010).  
+4. Brown, A. J. et al. Interfacial microfluidic processing of metal-organic framework hollow fiber membranes, Science 345, 72-75 (2014).  
+5. Ma, X. et al. Zeolitic imidazolate framework membranes made by ligand-induced permselectivation, Science 361, 1008-1011 (2018).  
+6. Gúcúyener, C., van den Bergh, J., Gascon, J. & Kapteijn, F. Ethane/Ethene Separation Turned on Its Head: Selective Ethane Adsorption on the Metal–Organic Framework ZIF-7 through a Gate-Opening Mechanism, J. Am. Chem. Soc. 132, 17704-17706 (2010).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 85, 881, 137]]<|/det|>
+7. Kwon, H. T. & Jeong, H-K. In Situ Synthesis of Thin Zeolitic–Imidazolate Framework ZIF-8 Membranes Exhibiting Exceptionally High Propylene/Propane Separation, J. Am. Chem. Soc. 125, 10763-10768 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[115, 147, 880, 183]]<|/det|>
+8. Zhou, S. et al. Paralyzed Membrane: Current-Driven Synthesis of a Metal-Organic Framework with Sharpened Propene/Propane Separation. Sci. Adv. 4:eaau1393 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 192, 880, 227]]<|/det|>
+9. Huang, K. et al. A ZIF-71 Hollow Fiber Membrane Fabricated by Contra-Diffusion. ACS Appl. Mater. Interfaces 7, 16157-16160 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[115, 236, 880, 272]]<|/det|>
+10. van den Bergh, J. et al. Understanding the Anomalous Alkane Selectivity of ZIF-7 in the Separation of Light Alkane/Alkene Mixtures. Chem. Eur. J. 17, 8832-8840 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[115, 281, 880, 316]]<|/det|>
+11. Tu, M. et al. Direct X-ray and electron-beam lithography of halogenated zeolitic imidazolate frameworks, Nat. Mater. 20, 93-99 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 325, 880, 361]]<|/det|>
+12. Miao, Y. et al. Solvent-free bottom-up patterning of zeolitic imidazolate frameworks, Nat. Commun. 13, 420 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[117, 370, 880, 423]]<|/det|>
+13. Dalstein, O. et al. Nanoimprinted, Submicrometric, MOF-Based 2D Photonic Structures: Toward Easy Selective Vapors Sensing by a Smartphone Camera, Adv. Funct. Mater. 26, 81-90 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[117, 433, 880, 468]]<|/det|>
+14. Xu, H. et al. Metal-Organic Framework-Inspired Metal-Containing Clusters for High-Resolution Patterning, Chem. Mater. 30, 4124-4133 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[117, 478, 880, 514]]<|/det|>
+15. Stassen, I. et al. Solvent-free synthesis of supported ZIF-8 films and patterns through transformation of deposited zinc oxide precursors, CrystEngComm 15, 9308-9311 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[117, 523, 880, 576]]<|/det|>
+16. Lu, G. & Hupp, J. T. Metal-Organic Frameworks as Sensors: A ZIF-8 Based Fabry-Pérot Device as a Selective Sensor for Chemical Vapors and Gases, J. Am. Chem. Soc. 132, 7832-7833 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[117, 585, 880, 638]]<|/det|>
+17. Moggach, S. A., Bennett, T. D. & Cheetham, A. K. The Effect of Pressure on ZIF-8: Increasing Pore Size with Pressure and the Formation of a High-Pressure Phase at 1.47 GPa, Angew. Chem. Int. Ed. 48, 7087-7089 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[117, 648, 880, 683]]<|/det|>
+18. Zhang, C. et al. Unexpected Molecular Sieving Properties of Zeolitic Imidazolate Framework-8, J. Phys. Chem. Lett. 3, 2130-2134 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[117, 693, 880, 728]]<|/det|>
+19. Knebel, A. et al. Defibrillation of soft porous metal-organic frameworks with electric fields, Science 358, 347-351 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[117, 738, 880, 773]]<|/det|>
+20. Zhang, K. et al. Exploring the Framework Hydrophobicity and Flexibility of ZIF-8: From Biofuel Recovery to Hydrocarbon Separations. J. Phys. Chem. Lett. 4, 3618-3622 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[117, 783, 880, 818]]<|/det|>
+21. Bennett, T. D. Cheetham, A. K., Fuchs, A. H. & Coudert, F-X. Interplay between defects, disorder and flexibility in metal-organic frameworks, Nat. Chem. 9, 11-16 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[117, 828, 880, 863]]<|/det|>
+22. Babu, D. J. et al. Restricting Lattice Flexibility in Polycrystalline Metal-Organic Framework Membranes for Carbon Capture, Adv. Mater. 31, 1900855 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 873, 880, 908]]<|/det|>
+23. Gascon, J. & Kapteijn, F. Metal-Organic Framework Membranes—High Potential, Bright Future? Angew. Chem. Int. Ed. 49, 1530-1532 (2010).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 83, 885, 901]]<|/det|>
+24. Stassen, I. et al. An updated roadmap for the integration of metal–organic frameworks with electronic devices and chemical sensors. Chem. Soc. Rev. 46, 3185-3241 (2017).25. Chen, R. et al. A two-dimensional zeolitic imidazolate framework with a cushion-shaped cavity for \(\mathrm{CO_2}\) adsorption, Chem. Commun. 49, 9500-9502 (2013).26. Peng, Y. et al. Metal-organic framework nanosheets as building blocks for molecular sieving membranes, Science 346, 1356-1359 (2014).27. Peng, Y., Li, Y., Ban, Y. & Yang, W. Two-Dimensional Metal–Organic Framework Nanosheets for Membrane-Based Gas Separation. Angew. Chem. 129, 9889-9893 (2017).28. He, G., Dakhchoune, M., Zhao, J., Huang, S. & Agrawal, K. V. Electrophoretic Nuclei Assembly for Crystallization of High-Performance Membranes on Unmodified Supports, Adv. Funct. Mater. 28, 1707427 (2018).29. Wei, R. et al. Aqueously Cathodic Deposition of ZIF-8 Membranes for Superior Propylene/Propane Separation, Adv. Funct. Mater. 30, 1907089 (2020).30. Eum, K. et al. ZIF-8 Membrane Separation Performance Tuning by Vapor Phase Ligand Treatment, Angew. Chem. 131, 16542-16546 (2019).31. Chen, Z. et al. Large-Area Crystalline Zeolitic Imidazolate Framework Thin Films. Angew. Chem. Int. Ed. 61, 14124-14130 (2021).32. Van Vleet, M. J., Weng, T., Li, X. & Schmidt, J. R. In Situ, Time-Resolved, and Mechanistic Studies of Metal-Organic Framework Nucleation and Growth. Chem. Rev. 118, 3681-3721 (2018).33. Cravillon, J. et al. Rapid Room-Temperature Synthesis and Characterization of Nanocrystals of a Prototypical Zeolitic Imidazolate Framework. Chem. Mater. 21, 1410-1412 (2009).34. Cravillon, J. et al. Fast Nucleation and Growth of ZIF-8 Nanocrystals Monitored by Time-Resolved In Situ Small-Angle and Wide-Angle X-Ray Scattering. Angew. Chem. Int. Ed. 50, 8067-8071 (2011).35. Jian, M. et al. Water-based synthesis of zeolitic imidazolate framework-8 with high morphology level at room temperature. RSC Adv. 5, 48433-48441 (2015).36. Terban, M. W. et al. Early stage structural development of prototypical zeolitic imidazolate framework (ZIF) in solution. Nanoscale 10, 4291-4300 (2018).37. Filez, M. et al. Elucidation of the pre-nucleation phase directing metal-organic framework formation. Cell Rep. Phys. Sci. 2, 100680 (2021).38. Doustkhah, E., Hassandoost, R., Khataee, A., Luque, R. & Assadi, M. H. N. Hard-templated metal–organic frameworks for advanced applications. Chem. Soc. Rev. 50, 2927-2953 (2021).39. Wang, K., Hui, K. N., San Hui, K., Peng, S. & Xu, Y. Recent progress in metal–organic framework/graphene-derived materials for energy storage and conversion: design, preparation, and application. Chem. Sci. 12, 5737-5766 (2021).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 84, 881, 137]]<|/det|>
+40. Wang, J. et al. Zeolitic Imidazolate Framework/Graphene Oxide Hybrid Nanosheets Functionalized Thin Film Nanocomposite Membrane for Enhanced Antimicrobial Performance. ACS Appl. Mater. Interfaces 8, 25508-25519 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[115, 146, 881, 199]]<|/det|>
+41. Pokhrel, J. et al. \(\mathrm{CO_2}\) adsorption behavior of amine-functionalized ZIF-8, graphene oxide, and ZIF-8/graphene oxide composites under dry and wet conditions. Microporous Mesoporous Mater. 267, 53-67 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 209, 880, 243]]<|/det|>
+42. Li, S. et al. Unconventional Nucleation and Oriented Growth of ZIF-8 Crystals on Non-Polar Surface. Adv. Mater. 24, 5954-5958 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[115, 254, 881, 305]]<|/det|>
+43. Yang, H. et al. Vacuum-assisted assembly of ZIF-8@GO composite membranes on ceramic tube with enhanced organic solvent nanofiltration performance. J. Membr. Sci. 545, 158-166 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[115, 316, 880, 351]]<|/det|>
+44. Choi, E. et al. Pore Tuning of Metal-Organic Framework Membrane Anchored on Graphene-Oxide Nanoribbon. Adv. Funct. Mater. 31, 2011146 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 361, 880, 413]]<|/det|>
+45. Sikdar, A. et al. Diffusion driven nanostructuring of metal-organic frameworks (MOFs) for graphene hydrogel based tunable heterostructures: highly active electrocatalysts for efficient water oxidation. J. Mater. Chem. A 9, 7640-7649 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 423, 880, 458]]<|/det|>
+46. Tao, J. et al. Controlling Metal-Organic Framework/ZnO Heterostructure Kinetics through Selective Ligand Binding to ZnO Surface Steps. Chem. Mater. 32, 6666-6675, (2020).  
+
+<|ref|>text<|/ref|><|det|>[[115, 468, 880, 503]]<|/det|>
+47. Zhu, Y. et al. Unravelling surface and interfacial structures of a metal-organic framework by transmission electron microscopy, Nat. Mater. 16, 532-536 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[115, 513, 880, 548]]<|/det|>
+48. Zhang, H., Zhao, M. & Lin, Y. S. Stability of ZIF-8 in water under ambient conditions, Microporous Mesoporous Mater. 279, 201-210 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 558, 880, 610]]<|/det|>
+49. Motevalli, B., Taherifar, N., Wang, H. & Liu, J. Z. Ab Initio Simulations to Understand the Leaf-Shape Crystal Morphology of ZIF-L with Two-Dimensional Layered Network, J. Phys. Chem. C 121, 2221-2227 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[115, 620, 880, 655]]<|/det|>
+50. Zhu, W., Li, X., Sun, Y., Guo, R., & Ding, S. Introducing hydrophilic ultra-thin ZIF-L into mixed matrix membranes for \(\mathrm{CO_2 / CH_4}\) separation, RSC Adv. 9, 2339-23399 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 666, 880, 700]]<|/det|>
+51. Yang, K. et al. ZIF-L membrane with a membrane-interlocked-support composite architecture for \(\mathrm{H_2 / CO_2}\) separation, Sci. Bull. 66, 1869-1876 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 711, 880, 762]]<|/det|>
+52. Tan, J. C., Bennett, T. D., Cheetham, A. K. Chemical structure, network topology, and porosity effects on the mechanical properties of Zeolitic Imidazolate Frameworks, Proc. Natl. Acad. Sci. 107, 9938-9943 (2010).  
+
+<|ref|>text<|/ref|><|det|>[[115, 774, 880, 825]]<|/det|>
+53. He, G. et al. High-permeance polymer-functionalized single-layer graphene membranes that surpass the postcombustion carbon capture target, Energy Environ. Sci. 12, 3305-3312 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[115, 836, 880, 871]]<|/det|>
+54. Bennett, T. D. & Horike, S. Liquid, glass and amorphous solid states of coordination polymers and metal–organic frameworks, Nat. Rev. Mater. 3, 431-440 (2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 84, 885, 120]]<|/det|>
+55. Kosma, V., De Simone, D., Vandenberghe, G. Metal Based Materials for EUV Lithography, J. Photopolymer. Sci. Technol. 32, 179-183 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[117, 128, 881, 183]]<|/det|>
+56. Conrad, S. et al. Controlling Dissolution and Transformation of Zeolitic Imidazolate Frameworks by using Electron-Beam-Induced Amorphization, Angew. Chem. Int. Ed. 57, 13592-13597 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[117, 192, 880, 227]]<|/det|>
+57. Stowers, J. & Keszler, D. A. High resolution, high sensitivity inorganic resists, Microelectron. Eng. 86, 730-733 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[117, 236, 880, 272]]<|/det|>
+58. Miao, Y., Tsapatsis, M. Electron Beam Patterning of Metal-Organic Frameworks. Chem. Mater. 33, 754-760 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[117, 281, 880, 317]]<|/det|>
+59. Oleksak, R. P. et al. Chemical and Structural Investigation of High-Resolution Patterning with HafSOx, ACS Appl. Mater. Interfaces 6, 2917-2921 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[117, 326, 880, 362]]<|/det|>
+60. Ghosh, S. et al. Chem. Mater. Two distinct stages of structural modification of ZIF-L MOF under electron-beam irradiation. 33, 5681-5689 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[117, 371, 880, 407]]<|/det|>
+61. Luo, C. et al. Review of recent advances in inorganic photoresists, RSC Adv. 10, 8385-8395 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[117, 415, 881, 452]]<|/det|>
+62. Manouras, T. & Argitis, P. High Sensitivity Resists for EUV Lithography: A Review of Material Design Strategies and Performance Results, Nanomaterials 10, 1593 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[117, 460, 880, 496]]<|/det|>
+63. Gangaik, A. S., Georgiev, Y. M. & Holmes, J. D. New Generation Electron Beam Resists: A Review, Chem. Mater. 29, 1898-1917 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[117, 505, 880, 541]]<|/det|>
+64. Ghash, S. et al. Two Distinct Stages of Structural Modification of ZIF-L MOF under Electron-Beam Irradiation, Chem. Mater. 33, 5681-5689 (2021).  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 590, 196, 606]]<|/det|>
+## Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 628, 211, 644]]<|/det|>
+## Chemicals  
+
+<|ref|>text<|/ref|><|det|>[[117, 655, 883, 870]]<|/det|>
+Zinc nitrate hexahydrate \(\mathrm{(Zn(NO_3)_2\cdot 6H_2O)}\) was purchased from Sigma-Aldrich. 2- methylimidazole (2-mIm) was from Chemie Brunschwig AG. HCl (32 wt%) was purchased from Reactolab S.A.. poly[1-(trimethylsilyl)-1-propylene] (PTMSP) was from ABCR. \(\mathrm{FeCl_3}\) (97%) and \(\mathrm{Na_2S_2O_8}\) was bought from Sigma-Aldrich. Cu foil (50 mm, 99.9%) were purchased from STREM. Toluene (AR) and methanol (AR) were from Fischer. All chemicals were used without further purifications. \(\mathrm{Si / SiO_2}\) wafers were purchased from University Wafer Inc. \(\mathrm{Si / SiO_2}\) wafer with single layer graphene was bought from Ted Pella. Highly oriented pyrolytic graphite (HOPG) (ZYA quality, GRAS/1.0x7x7) was purchased from ScanSens. Silicon nitride TEM supports (50 nm silicon nitride film on a \(200\mu \mathrm{m}\) silicon frame with nine viewing
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 880, 127]]<|/det|>
+windows, each \(0.1 \times 0.1 \mathrm{mm}\) were purchased from Ted Pella. Sapphire (Al₂O₃) wafer (Z- cut, \(10 \mathrm{cm} \times 0.5 \mathrm{mm}\) ) was purchased from MTI.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 143, 446, 161]]<|/det|>
+## Characterizations and Measurements  
+
+<|ref|>text<|/ref|><|det|>[[117, 177, 883, 319]]<|/det|>
+SEM measurements were performed on a Teneo Scanning Electron Microscope operating at 1 kV. PXRD data were collected at a Bruker D8 Discover diffractometer with a Lynxeye XE detector, operated at \(40 \mathrm{kV}\) , \(400 \mathrm{mA}\) for Cu Kα \((\lambda = 1.5406 \mathrm{\AA})\) at ambient temperature and pressure. Bright- field TEM images and selected- area electron diffraction (SAED) images were obtained with a Talos F200X microscope operated at \(200 \mathrm{kV}\) . TEM images for patterns on silicon nitride were obtained on a ThermoFisher TF30 TEM operating at \(300 \mathrm{kV}\) .  
+
+<|ref|>text<|/ref|><|det|>[[116, 333, 883, 550]]<|/det|>
+Low- dose AC- HRTEM was performed on a Cs- corrected FEI cubed G2 Titan 60- 300 electron microscope at \(300 \mathrm{kV}\) , using a Gatan K2 direct- detection camera in electron counting mode. The AC- HRTEM images were acquired with the dose fractionation function, and each image stack is composed of 120 frames with \(0.05 \mathrm{s}\) exposure for each frame, with a total electron dose of \(\sim 60 \mathrm{e}^{- }\mathrm{\AA}^{- 2}\) . The raw image was denoised by using an average background subtraction filter (ABSF). The CTF correction was performed based on the defocus value determined from the amorphous thon rings in the Fourier transform, and the projected electrostatic potential was simulated by the QSTEM software (QSTEM V2, 31). Simulated ED pattern was carried out by using Singlecrystal module of CrystalMaker software.  
+
+<|ref|>text<|/ref|><|det|>[[117, 564, 883, 707]]<|/det|>
+AFM images and modulus measurement were recorded on a Bruker MultiMode 8 AFM. For modulus measurement, a Bruker Tap525A rectangular probe was used and calibrated with standard sample sapphire, polystyrene and HOPG. XPS were carried out on an Axis Supra (Kratos Analytical) using the monochromated K x- ray line of an aluminum anode. Synchrotron GIXRD was carried out at beamline BM01, Swiss- Norwegian beamline (SNBL) at the European Synchrotron Radiation Facility (ESRF) with wavelength of \(0.683 \mathrm{\AA}\) and \(0.960 \mathrm{\AA}\) .  
+
+<|ref|>text<|/ref|><|det|>[[117, 721, 883, 863]]<|/det|>
+The gas separation performance of all membranes was recorded on a homemade permeation set- up. The pressure on the feed side was maintained at 2- 8 bar and on the permeate side at 1 bar during the measurements. All measurements were done after reaching the steady state with argon as the sweep gas. The membranes were sealed with stainless- steel gasket. The composition of permeate was analysed using an online Hiden Analytical HPR- 20 mass spectrometer.  
+
+<|ref|>text<|/ref|><|det|>[[118, 879, 551, 897]]<|/det|>
+The permeances, \(J_{i}\) , of gas \(i\) was calculated by Eq. S1
+
+<--- Page Split --->
+<|ref|>equation<|/ref|><|det|>[[367, 83, 692, 102]]<|/det|>
+\[J_{i} = X_{i} / (A\cdot \Delta P_{i}) \quad (S1)\]  
+
+<|ref|>text<|/ref|><|det|>[[115, 117, 881, 186]]<|/det|>
+where \(X_{i}\) is the molar flow rate of component \(i\) across the membrane area, \(A\) and \(\Delta P_{i}\) is the transmembrane pressure difference for the component \(i\) . The selectivity \(\alpha_{ij}\) of two gases ( \(i\) and \(j\) , where \(i\) is the faster permeating gas) was calculated by Eq. S2  
+
+<|ref|>equation<|/ref|><|det|>[[392, 200, 696, 220]]<|/det|>
+\[\alpha_{ij} = J_{i} / J_{j} \quad (S2)\]  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 235, 376, 253]]<|/det|>
+## Synthesis of 2DZIF/aZIF film  
+
+<|ref|>text<|/ref|><|det|>[[117, 268, 882, 362]]<|/det|>
+In a petri dish containing \(29~\mathrm{ml}\) of \(\mathrm{Zn(NO_3)_2}\) aqueous solution at room temperature, the substrate (HOPG, graphene/Si/SiO \(_2\) for 2DZIF and Si/SiO \(_2\) for aZIF) was partially immersed. Then \(1\mathrm{ml}\) of \(2\mathrm{- mlm}\) aqueous solution was added. After a certain time, the substrate was removed to stop the reaction.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 377, 545, 396]]<|/det|>
+## Synthesis of 2DZIF membrane for gas separation  
+
+<|ref|>text<|/ref|><|det|>[[117, 410, 883, 553]]<|/det|>
+First, single- layer graphene was synthesized by using low- pressure CVD of methane on copper foil following the literature. Before the synthesis, the copper foil was annealed at \(1077^{\circ}\mathrm{C}\) in a \(\mathrm{H}_2 / \mathrm{Ar}\) atmosphere for \(60~\mathrm{min}\) . Then, \(\mathrm{CO_2}\) ( \(100~\mathrm{mL / min}\) ) and \(\mathrm{H}_2\) ( \(8~\mathrm{mL / min}\) ) flow was introduced successively, each for \(30~\mathrm{min}\) , to remove the contaminations. At last, \(\mathrm{CH_4}\) ( \(24~\mathrm{mL / min}\) ) and \(\mathrm{H}_2\) ( \(8~\mathrm{mL / min}\) ) flow was used to grow single- layer graphene on copper film for \(30~\mathrm{min}\) at pressure of \(460~\mathrm{mTorr}\) .  
+
+<|ref|>text<|/ref|><|det|>[[116, 567, 883, 907]]<|/det|>
+After the synthesis of single- layer graphene, an \(\mathrm{O_2}\) plasma (MTI plasma cleaner, EQ- PCE- 3, \(13.56\mathrm{MHz}\) , \(17\mathrm{W}\) ) was carried out to introduce nanopores for application of 2DZIF film as selective layer in the membrane. Briefly, the atmosphere in the plasma chamber was exchanged by \(\mathrm{O_2}\) flow to pressure around \(50\mathrm{mTorr}\) . Then a plasma was generated for \(4\mathrm{s}\) to etch SLG to obtain NG. After the plasma, a solution of PTMSP in toluene ( \(1.25\mathrm{wt}\%\) ) was spin- coated on NG at \(1000\mathrm{rpm}\) for \(30\mathrm{s}\) and \(2000\mathrm{rpm}\) for \(30\mathrm{s}\) , respectively. The sample was dried in ambient air at room temperature overnight. Then, copper foil was etched by a combination of \(\mathrm{FeCl_3}\) ( \(0.5\mathrm{M}\) in water), HCl ( \(0.1\mathrm{M}\) in water) and water. The floating graphene/PTMSP film was transferred on the surface of a \(29\mathrm{ml}\) of aqueous solution of \(\mathrm{Zn(NO_3)_2}\) ( \(2\mathrm{mmol / L}\) ) in a petri dish, and an aqueous solution of \(2\mathrm{- mlm}\) ( \(1\mathrm{mL}\) , \(0.5\mathrm{mol / L}\) ) was added in. After \(2\mathrm{min}\) of reaction the resulted 2DZIF/graphene/PTMSP film was transferred to substrate (e.g., macroporous W support hosting \(5\mu \mathrm{m}\) holes over a \(1\mathrm{mm}^2\) area for membranes) for further characterizations or applications. For the synthesis of centimeter- scale 2DZIF membrane, \(1\mathrm{wt}\%\) of Teflon AF in GOLDEN perfluorinated fluid was spin- coated on NG at \(300\mathrm{rpm}\) for \(60\mathrm{s}\) , and heated at \(60^{\circ}\mathrm{C}\)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 882, 200]]<|/det|>
+for \(3\mathrm{h}\) . After that, the sample was put into a homemade membrane module (Supplementary Fig. 19) while Cu foil facing up. This allowed etching of Cu foil in the membrane module by \(10\mathrm{wt}\% \mathrm{Na}_2\mathrm{S}_2\mathrm{O}_8\) aqueous solution exposing NG surface for 2DZIF synthesis. Finally, 2DZIF film was synthesized by exposing NG to growth solution for \(10\mathrm{min}\) at room temperature (2 mM \(\mathrm{Zn(NO_3)_2}\) and \(16\mathrm{mM}2\mathrm{- }\mathrm{mM}\) ).  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 216, 572, 235]]<|/det|>
+## Sample preparation of 2DZIF on graphene for AFM  
+
+<|ref|>text<|/ref|><|det|>[[117, 250, 882, 367]]<|/det|>
+For the AFM sample preparation of 2DZIF film on graphene/PTMSP substrate, the 2DZIF/graphene/PTMSP film was transferred on \(\mathrm{Si / SiO_2}\) wafer with 2DZIF layer facing the wafer. Then the sample was annealed at \(70^{\circ}\mathrm{C}\) for \(4\mathrm{h}\) , to increase the adhesion between film and \(\mathrm{Si / SiO_2}\) wafer. After that, the sample was immersed in toluene for \(12\mathrm{h}\) , to remove PTMSP layer.  
+
+<|ref|>text<|/ref|><|det|>[[117, 382, 882, 549]]<|/det|>
+The AFM image of 2DZIF film with triangular morphology after \(5\mathrm{min}\) of water etching was collected directly on the film with 2DZIF layer facing up. Specifically, the 2DZIF/graphene/PTMSP film with triangular morphology was first scooped by glass slide with 2DZIF layer facing the slide, and a \(\mathrm{Si / SiO_2}\) wafer attached with double- sided carbon tape was pressed onto PTMSP layer. As a result, the 2DZIF/graphene/PTMSP film was transferred onto \(\mathrm{Si / SiO_2}\) wafer, resulting in 2DZIF layer facing up. AFM measurement of 2DZIF film on HOPG, \(\mathrm{Si / SiO_2}\) and graphene/ \(\mathrm{Si / SiO_2}\) was carried out directly on the sample without any treatment.  
+
+<|ref|>sub_title<|/ref|><|det|>[[119, 565, 375, 583]]<|/det|>
+## Sample preparation for TEM  
+
+<|ref|>text<|/ref|><|det|>[[118, 597, 882, 690]]<|/det|>
+Similar to sample preparation for AFM, 2DZIF/graphene/PTMSP film was transferred on TEM grid with 2DZIF layer facing the TEM grid. Then the sample was annealed at \(70^{\circ}\mathrm{C}\) for \(4\mathrm{h}\) , to increase the adhesion between film and TEM grid. After that, the sample was immersed in toluene for \(12\mathrm{h}\) , to remove the PTMSP layer.  
+
+<|ref|>text<|/ref|><|det|>[[118, 706, 882, 798]]<|/det|>
+For the electron patterning sample, aZIF layer was grown on a silicon nitride supports. Prior to the ZIF deposition, the silicon nitride supports were treated with oxygen plasma for \(10\mathrm{min}\) (29.6 W, \(400\mathrm{mTorr}\) oxygen pressure) in a plasma cleaner (Harrick Plasma) to improve the surface reactivity.  
+
+<|ref|>sub_title<|/ref|><|det|>[[119, 814, 461, 832]]<|/det|>
+## Electron beam patterning of aZIF films  
+
+<|ref|>text<|/ref|><|det|>[[118, 847, 881, 891]]<|/det|>
+For negative tone patterning, aZIF films were exposed to electron beam using a Thermo Fisher Helios G4 UC Dual Beam microscope operating at \(20\mathrm{kV}\) accelerating voltage and \(400\mathrm{pA}\)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 881, 127]]<|/det|>
+beam current. The areal doses were \(80\mathrm{mC / cm^2}\) for all patterns. After exposure, the aZIF films were developed in deionized water for \(24\mathrm{h}\) and blow dried in a stream of nitrogen gas.  
+
+<|ref|>text<|/ref|><|det|>[[117, 141, 883, 284]]<|/det|>
+For positive tone patterning, the aZIF films were first placed in a \(60\mathrm{mL}\) PFA vessel with a bed of \(0.2\mathrm{g}4,5\) - dichloroimidazole (dcIm) and heated at \(75^{\circ}\mathrm{C}\) for \(1.5\mathrm{h}\) . The dcIm- treated aZIF was then exposed to electron beam using a Thermo Fisher Helios G4 UC Dual Beam microscope operating at \(20\mathrm{kV}\) accelerating voltage and \(100\mathrm{pA}\) beam current. The areal doses were \(2\mathrm{mC / cm^2}\) for all patterns. After exposure, the aZIF films were immersed in NMP and acetone each for \(10\mathrm{s}\) and blow dried in a stream of nitrogen gas.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 299, 496, 318]]<|/det|>
+## Spin-coating of aZIF films on silicon wafers  
+
+<|ref|>text<|/ref|><|det|>[[117, 332, 883, 500]]<|/det|>
+Aqueous solutions of \(\mathrm{Zn(NO_3)_2}\) (4 mmol/L) and 2- Im (32 mmol/L) were mixed in a T connector (0.5 mm I.D.) at \(1\mathrm{mL / min}\) injection rate, respectively, using a two- channel syringe pump. The mixed solution was spin- coated on silicon wafers while spinning at speeds between 500 and 2000 rpm for \(1\mathrm{min}\) , followed by \(10\mathrm{s}\) spinning at 2000 rpm to completely dry the film. The spin- coated films were exposed to electron beam using line patterns of \(100\mathrm{nm}\) width and \(400\mathrm{nm}\) spacing and developed in deionized water for \(24\mathrm{h}\) . The height of the line patterns were measured to determine the thickness of spin- coated films.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 516, 306, 533]]<|/det|>
+## Structural simulation  
+
+<|ref|>text<|/ref|><|det|>[[117, 548, 883, 862]]<|/det|>
+The simulation of 2DZIF structure was carried out by the DFT calculations. At first, the reported ZIF- L structure was imported into Forcite module in Material Studio software (Accelrys, San Diego, CA) to calculate the initial structural model, and unit- cell was set to be orthorhombic and \(a = 24.0\mathrm{\AA}\) , \(b = 20.0\mathrm{\AA}\) , \(c = 20.0\mathrm{\AA}\) , \(\alpha = 90^{\circ}\) , \(\beta = 90^{\circ}\) , \(\gamma = 90^{\circ}\) , respectively, while the connectivity was kept. The calculation task was geometry optimization. The quality was set to be fine, and 'smart' algorithm was selected. After that, van der Waals DFT calculations were performed using the Quantum ESPRESSO package<sup>66,67</sup>. The Brillouin zone was sampled at the gamma point. An energy cutoff of 60 Ry was used for the plane wave expansion of the wavefunctions. A kinetic energy cutoff of 480 Ry on the charge was used together with ultra- soft pseudopotentials<sup>68,69</sup>. The relaxation was performed with the Perdew- Burke- Ernzerhof (PBE) functional<sup>70</sup>. The system was relaxed to the lowest energy configuration of atoms. The surface geometry had been optimized with the convergence thresholds of \(1 \times 10^{-4}\mathrm{Ry}\) and \(3.28 \times 10^{-3}\mathrm{Ry / Bohr}\) for the total energy and forces, respectively.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 878, 262, 895]]<|/det|>
+## Data availability
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 881, 151]]<|/det|>
+Data supporting the findings in the present work are available in the manuscript or Supplementary Information. Additional data are available from the corresponding author upon reasonable request.  
+
+<|ref|>text<|/ref|><|det|>[[117, 171, 881, 216]]<|/det|>
+65. Huang, S. et al. Millisecond lattice gasification for high-density \(\mathrm{CO_2}\) - and \(\mathrm{O_2}\) -sieving nanopores in single-layer graphene. Sci. Adv. 7: eabf0116 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[117, 234, 881, 279]]<|/det|>
+66. Giannozzi, P. et al. QUANTUM ESPRESSO: a modular and open-source software project for quantum simulations of materials. J. Phys. Condens. Matter 21, 390052 (2009).  
+
+<|ref|>text<|/ref|><|det|>[[117, 297, 881, 342]]<|/det|>
+67. Giannozzi, P. et al. Advanced capabilities for materials modelling with Quantum ESPRESSO. J. Phys. Condens. Matter 29, 465901 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[117, 361, 880, 405]]<|/det|>
+68. Lejaeghere, K. et al. Reproducibility in density functional theory calculations of solids. Science 351, 6280 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[117, 424, 881, 469]]<|/det|>
+69. Prandini, G., Marrazzo, A., Castelli, I. E., Mounet, N. & Marzari, N. Precision and efficiency in solid-state pseudopotential calculations. npj Comput. Mater. 4, 72 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[117, 488, 880, 532]]<|/det|>
+70. Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865-3868 (1996).  
+
+<|ref|>text<|/ref|><|det|>[[115, 551, 883, 816]]<|/det|>
+Acknowledgments The authors acknowledge Dr. Dmitry Chernyshov and Dr. Diadkin Vadim at beamline BM01, the-Swiss-Norwegian Beamlines (SNBL), European Synchrotron Radiation Facility (ESRF) for assistance with synchrotron GIXRD experiments (doi:10.15151/ESRF-ES-670011338. ), Dr. Pascal Alexander Schouwink from EPFL for the help of XRD data. Dr. Mounir Mensi and Mojtaba Rezaei from EPFL for the help of XPS and AFM. This project is primarily supported by European Research Council Starting Grant (805437-UltimateMembranes). Parts of the work was supported by Swiss National Science Foundation (SNSF) Assistant Professor Energy Grant (PYAPP2_173645) and SNSF project (514601); Y. M. and M. T. acknowledge funding by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Division of Chemical Sciences, Geosciences and Biosciences under Award DE-SC0021212.  
+
+<|ref|>text<|/ref|><|det|>[[117, 860, 881, 905]]<|/det|>
+Author contributions K. A. and Q. L. conceived the project and wrote the manuscript. Q. L., Y. M., D. B., and J. H. prepared the samples. Q. L., Y. M. and S. L. performed AFM
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 882, 225]]<|/det|>
+measurements. Q. L. and Y. M. collected SEM data. Q. L. performed XRD measurement. S. L. collected XPS data. L. V. and H. C. collected TEM images and ED data. Q. L. and M. V. carried out structural simulations. C. C. and Y. H. helped with AC- HRTEM images. Q. L. and S. S. carried out gas permeance experiments. M. T. conceived and Y. M. performed the aZIF film deposition and e- beam patterning experiments. All authors discussed the results and commented on the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[118, 238, 880, 258]]<|/det|>
+Competing interests: A patent application based on the findings reported here has been filed.  
+
+<|ref|>sub_title<|/ref|><|det|>[[119, 302, 320, 320]]<|/det|>
+## Additional information  
+
+<|ref|>text<|/ref|><|det|>[[117, 333, 882, 444]]<|/det|>
+Supplementary information The online version contains supplementary material available at Correspondence and requests for materials should be addressed to Kumar Varoon Agrawal. Peer review information Nature thanks anonymous reviewer(s) for their contribution to the peer review of this work.  
+
+<|ref|>text<|/ref|><|det|>[[117, 454, 830, 473]]<|/det|>
+Reprints and permissions information is available at http://www.nature.com/reprints.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[44, 92, 765, 112]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[61, 130, 137, 149]]<|/det|>
+- Sl.pdf
+
+<--- Page Split --->

preprint/preprint__09caa31641cd2fa62399cf94c4caec3caa098ba1b127a492b84d0583324ff5f2/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. Overview of the scaffold hopping approach. SAR and crystal structures of selected benzyl and aniline analogs. A) Crystal structure of compound 127 (yellow sticks) with 14-3-3o (grey surface) and phospho-ERα peptide (orange sticks). Interacting aminoacids are shown as sticks and water molecules as red spheres (PDB 8ALW). B) Ligand overlay of compound 127 and docking pose of the new MCR scaffold. C) General MCR scaffold and main points of variation. D) Overview of general synthetic routes. Detailed experimental conditions are described in the SI. E) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. F) Crystal structure overlay for compounds 1 (cyan sticks) and 127 (yellow sticks) bound to 14-3-3o (grey surface) /ERα (orange sticks). G) Crystal structure of compound 1 (cyan sticks) with 14-3-3o/ERα. Interacting aminoacid residues are shown as sticks and interacting water molecules as red spheres. H-I) Structural overlays of compounds 2 (brown sticks), and 10 (dark red sticks) with compound 1 (cyan sticks).",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        145,
+        890,
+        797
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2. SAR and crystal structures of selected double-ortho substituted analogs. A) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-E) Crystal structures of 14-3-3o/ERα with compounds 17 (dark pink sticks), 19 (teal sticks) and overlays of crystal structures for compounds 17 (dark pink sticks), 20 (pale green sticks), 21 (pale purple sticks) and 25 (bright yellow sticks).",
+    "footnote": [],
+    "bbox": [
+      [
+        116,
+        135,
+        886,
+        441
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3. SAR and crystal structures of analogs substituted in positions X and Y. A) MS bar graphs at 1 \\(\\mu \\mathrm{M}\\) . For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-E) Crystal structure of 14-3-3o/ERα with compound 28 (dark yellow sticks), and overlays of crystal structures for compounds 28 (dark yellow sticks), 32 (pink sticks) and 33 (emerald green sticks) with 21 (pale purple sticks).",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        320,
+        884,
+        636
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4. SAR and crystal structures of 2,6-di-Me analogs substituted in positions X and W. Biophysical data (MS, TR-FRET, SPR) and cell data (NanoBRET). A) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-C) Overlays of crystal structures of 14-3-3α/ERα ternary complexes with compounds 40 (orange sticks), 41 (red sticks), and 42 (blue sticks). D) TR-FRET schematic and protein titration data for representative compounds at 100 μM compound or DMSO. E) SPR data for the binary 14-3-3α/ERα interaction and ternary interactions with 181, 17 and 41 (mean +/- SD, n=2). F-G) 14-3-3α-HaloTag/Nuc-ERα NanoBRET assay in HEK293T cells with compound titrations (1:2 dilution, starting at 40 μM). Data points excluded where compound dosage was toxic to the cells. MCR compounds compared to the previously described stabilizer 181 and 85, an inactive compound as the negative control. Bar graphs quantifying pEC50 values.",
+    "footnote": [],
+    "bbox": [
+      [
+        95,
+        90,
+        925,
+        740
+      ]
+    ],
+    "page_idx": 9
+  }
+]

preprint/preprint__09caa31641cd2fa62399cf94c4caec3caa098ba1b127a492b84d0583324ff5f2/preprint__09caa31641cd2fa62399cf94c4caec3caa098ba1b127a492b84d0583324ff5f2.mmd

@@ -0,0 +1,292 @@
+
+# Rapid scaffold-hopping for molecular glues: from fragments to cell-active probes targeting the 14-3-3/ERα complex  
+
+Michelle Arkin michelle.arkin@ucsf.edu  
+
+University of California at San Francisco https://orcid.org/0000- 0002- 9366- 6770  
+
+Markella Konstantinidou University of California San Francisco https://orcid.org/0000- 0001- 5972- 4140  
+
+Marios Zingiridis University of Crete https://orcid.org/0009- 0008- 1150- 2926  
+
+Marloes Pennings Eindhoven University of Technology https://orcid.org/0000- 0002- 3366- 0238  
+
+Michael Fragkiadakis University of Crete  
+
+Johanna Virta University of California San Francisco  
+
+Jezrael Revalde University of California, San Francisco  
+
+Emira Visser TU Eindhoven  
+
+Christian Ottmann Eindhoven University of Technology https://orcid.org/0000- 0001- 7315- 0315  
+
+Luc Brunsveld TU Eindhoven https://orcid.org/0000- 0001- 5675- 511X  
+
+Constantinos Neochoritis University of Crete https://orcid.org/0000- 0001- 5098- 5504  
+
+## Article  
+
+Keywords: covalent, estrogen receptor, MCR, molecular glue, 14- 3- 3  
+
+Posted Date: February 28th, 2025  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 6051794/v1
+
+<--- Page Split --->
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: Yes there is potential Competing Interest. M.R.A, C.O, and L.B. are co- founders of Ambagon Therapeutics. C.O. is an employee of Ambagon Therapeutics.  
+
+Version of Record: A version of this preprint was published at Nature Communications on July 14th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 61176- 4.
+
+<--- Page Split --->
+
+# Rapid scaffold-hopping for molecular glues: from fragments to cell-active probes targeting the 14-3-3/ERα complex  
+
+Markella Konstantinidou\*[a], Marios Zingiridis[b], Marloes A.M. Pennings[c], Michael Fragkiadakis[b], Johanna M. Virta[a], Jezrael L. Revalde[a], Emira J. Visser[c], Christian Ottmann[c], Luc Brunsveld[c], Constantinos G. Neochoritis\*[b], Michelle R. Arkin\*[a]  
+
+[a] M. Konstantinidou, J.M. Virta, J.L. Revalde, M.R. Arkin  Department of Pharmaceutical Chemistry and Small Molecule Discovery Centre (SMDC)  University of California San Francisco (UCSF)  CA 94143 (USA)  E- mail: markella.constantinidou@ucsf.edu, michelle.arkin@ucsf.edu  [b] M. Zingiridis, M. Fragkiadakis, C.G. Neochoritis  Department of Chemistry  University of Crete  Voutes, Heraklion, 70013, Greece  E- mail: kneochor@uoc.gr  [c] M.A.M. Pennings, E.J. Visser, C. Ottmann, L. Brunsveld  Laboratory of Chemical Biology, Department of Biomedical Engineering and Institute for Complex Molecular Systems (ICMS)  Eindhoven University of Technology  5600 MB Eindhoven, The Netherlands  
+
+Abstract: Molecular glues, small molecules that bind cooperatively at a protein- protein interface, have emerged as powerful modalities for the modulation of protein- protein interactions (PPIs) and "undruggable" targets. The systematic identification of new chemical matter with a molecular glue mechanism of action remains a significant challenge in drug discovery. Here, we present a scaffold hopping approach, using as a starting point our previously developed molecular glues for the native 14- 3- 3/estrogen receptor alpha (ERα) complex. The novel, computationally designed scaffold was based on the Groebke- Blackburn- Bienaymé multi- component reaction (MCR), leading to drug- like analogs with multiple points of variation, thus enabling the rapid derivatization and optimization of the scaffold. Structure- activity relationships (SAR) were developed using intact mass spectrometry and TR- FRET. Rational structure- guided optimization was facilitated by crystal structures of ternary complexes with the glues, 14- 3- 3 and phospho- peptides mimicking the highly disordered C- terminus of ERα. We measured the kinetics of 14- 3- 3/ERα peptide binding by SPR, using a format in which a 14- 3- 3/molecular glue complex was immobilized on the SPR chip. The most potent compounds stabilized the complex by 100- fold and increased the residence time by 14- fold. Cellular stabilization of 14- 3- 3/ERα for the most potent analogs was confirmed using a NanoBRET assay with full- length proteins in live cells (EC50 = 2.7 – 5 μM). Our approach highlights the potential of MCR chemistry, combined with scaffold hopping, to drive the development and optimization of unprecedented molecular glue scaffolds.  
+
+## Introduction  
+
+The stabilization of native protein- protein interactions (PPIs) with small molecules offers an attractive strategy for the activation or inhibition of signaling pathways in a therapeutic context. \(^{1,2}\) PPIs were traditionally considered difficult targets due to the lack of well- defined pockets and the presence of large, hydrophobic surfaces. \(^{3 - 5}\) The fundamental understanding of the mechanism of action of molecular glues – small molecules that bind cooperatively at PPI interfaces and strengthen weak, pre- existing interactions – has enabled the stabilization of PPIs by taking into account the elements of cooperativity, molecular recognition and shape complementarity. \(^{6,7}\)  
+
+A particularly challenging class of PPIs includes proteins that are intrinsically disordered and only become partially structured when bound to a protein partner, such as a chaperone binding to a client protein. \(^{8,9}\) 14- 3- 3 is an abundant scaffolding protein that recognizes specific phospho- serine or phospho- threonine motifs on disordered domains of the client and upon binding creates a structured binding interface. \(^{10}\) Molecular glues targeting 14- 3- 3/client complexes bind
+
+<--- Page Split --->
+
+to the composite surface formed at this interface; the inherent cooperativity of this approach yields molecular glues with selectivity and potency. Of note, 14- 3- 3 proteins lack function – the function is instead encoded on the client protein and in particular on the phospho- site that is being recognized, leading either to activation or inhibition of signaling pathways. \(^{11}\)  
+
+Among the extensive interactome of 14- 3- 3, here we focus on its native interaction with the hormone regulated transcription factor estrogen receptor \(\alpha\) (ERα). 14- 3- 3 recognizes the protein sequence surrounding phospho- T594 on the disordered C- terminus on the F- domain of ERα and acts as a negative regulator by inhibiting ERα binding to chromatin and blocking ERα- mediated transcription. \(^{12,13}\) To date, ERα small molecule drugs, acting either as inhibitors or degraders, target the adjacent ligand binding domain (LBD). \(^{14 - 17}\) However, mutations in the LBD are often associated with acquired endocrine resistance. \(^{18}\) Thus, stabilization of the native 14- 3- 3/ERα PPI could be useful as an alternative strategy to block ERα transcriptional activity in ERα positive breast cancer, especially in cases of acquired endocrine resistance. The feasibility of this approach, targeting the F- domain, is corroborated by studies using the natural product fusicoccin- A (FC- A) and its semi- synthetic analogs that stabilize the interaction between 14- 3- 3 and the C- terminus of ERα. \(^{19}\) We now require drug- like chemical probes to define the biological impact of targeting the F- domain to inhibit ERα in hormone- positive breast cancer.  
+
+We have applied different strategies for the identification of chemical matter to stabilize the 14- 3- 3/ERα complex. We used a site- directed fragment- based technology, termed "disulfide tethering" with intact mass spectrometry as the readout to identify reversible fragments bound at the native cysteine (C38) of 14- 3- 3α in the presence of a phosphorylated peptide that represented the disordered C- terminus of ERα. \(^{20}\) Rational, structure- guided optimization of the reversible disulfide fragments led to irreversibly covalent, selective molecular glues that bound at the composite surface of 14- 3- 3α/ERα. \(^{21}\) For the development of non- covalent molecular glues, we used a fragment- linking approach, derived from the crystal structures of two diverse fragments that were identified in crystallographic and disulfide tethering screens. \(^{22}\) Thus, the 14- 3- 3/ERα PPI has served as a valuable system to test diverse molecular- glue discovery strategies.  
+
+Here, we present a scaffold- hopping approach based on multi- component reaction chemistry (MCR). Multi- component reactions are defined as synthetic approaches where at least three starting materials react in a single step to form a complex scaffold, where most of the atoms contribute to the newly formed product. This broad definition covers reactions with various synthetic mechanisms. \(^{23,24}\) MCR chemistry, due to its highly divergent character, is an enticing strategy for developing new scaffolds and rapid structure- activity relationships (SAR), as it allows the combination of short synthetic routes with high diversity and complexity. MCR has emerged as an attractive alternative to multistep linear convergent synthetic approaches and has been successfully applied to the synthesis of active pharmaceutical ingredients (API). \(^{25 - 30}\) Here, we describe our strategy for the development of a drug- like MCR scaffold stabilizing the 14- 3- 3α/ERα complex. The most potent analogs of the series showed efficacy in orthogonal biophysical assays and cell- based PPI stabilization in the low micromolar range.  
+
+## Results  
+
+## Structure activity relationships (SAR)  
+
+Structurally, 14- 3- 3 binds ERα by recognizing phospho- T594, the penultimate residue on the C- terminus of ERα. This creates a large, open, solvent- exposed pocket that can accommodate a small molecule. Although steric factors are not an issue for targeting the composite surface of the 14- 3- 3/ERα complex, we found it was important to rigidify an initially flexible scaffold to maximize the stabilization effect. \(^{21}\) Our aim in this work was to design a scaffold that would be more rigid from the beginning, locking in a favorable three- dimensional shape complementary to the large pocket.  
+
+To this end, we used the freely accessible software AnchorQuery™, which performs pharmacophore- based screening of approximately 31 million compounds that are readily synthesizable through one- step multi- component reaction (MCR) chemistry. \(^{31,32}\) Although AnchorQuery™ was originally developed for PPI inhibitors \(^{33}\) , in this case it was successful in proposing MCR scaffolds for PPI stabilizers. AnchorQuery™ requires a ligand- bound crystal structure or docked binding pose as a starting point. We used the crystal structure of the previously disclosed compound 127 (PDB 8ALW) that was bound at the composite surface of the 14- 3- 3α/ERα complex, with a favorable ligand conformation, based on our biophysical data. \(^{21}\) The compound formed multiple favorable interactions both with 14- 3- 3α and the phospho- ERα peptide (Fig 1A). In the co- crystal structure, the irreversible chloroacetamide warhead of compound 127
+
+<--- Page Split --->
+
+formed a covalent bond with C38 of 14- 3- 3o. The \(p\) - chloro-phenyl ring occupied a small hydrophobic pocket that formed a halogen bond with K122 of 14- 3- 3. The tetrahydropyrane ring adopted a favorable conformation that allowed the formation of hydrophobic interactions with 14- 3- 3 residues (L218, I219), the terminal Val595 of ERα, and a water-  
+
+![](images/Figure_1.jpg)
+
+<center>Figure 1. Overview of the scaffold hopping approach. SAR and crystal structures of selected benzyl and aniline analogs. A) Crystal structure of compound 127 (yellow sticks) with 14-3-3o (grey surface) and phospho-ERα peptide (orange sticks). Interacting aminoacids are shown as sticks and water molecules as red spheres (PDB 8ALW). B) Ligand overlay of compound 127 and docking pose of the new MCR scaffold. C) General MCR scaffold and main points of variation. D) Overview of general synthetic routes. Detailed experimental conditions are described in the SI. E) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. F) Crystal structure overlay for compounds 1 (cyan sticks) and 127 (yellow sticks) bound to 14-3-3o (grey surface) /ERα (orange sticks). G) Crystal structure of compound 1 (cyan sticks) with 14-3-3o/ERα. Interacting aminoacid residues are shown as sticks and interacting water molecules as red spheres. H-I) Structural overlays of compounds 2 (brown sticks), and 10 (dark red sticks) with compound 1 (cyan sticks). </center>
+
+<--- Page Split --->
+
+mediated hydrogen bond from the oxygen atom in the ring, which was part of a large water network. The overall ligand conformation also led to a key water- mediated hydrogen bond between the aniline nitrogen and the terminal carboxylic acid of Val595 of ERα, significantly contributing to molecular recognition. In this compound series, the introduction of large non- aromatic rings, such as the tetrahydropyrane, combined with aniline rings were necessary to limit the multiple ligand conformations and improved the affinity to the complex.  
+
+Understanding the binding mode of 127, we were able to use AnchorQueryTM for a scaffold- hopping approach. The software required an anchor motif on the ligand that was bioisosteric to an amino acid residue and was kept constant for all pharmacophore- based searches of the database. A suitable anchor in our case was the \(p\) - chloro- phenyl ring that was deeply buried at the PPI interface in the small pocket surrounding K122 of 14- 3- 3 (Fig S1). This motif served as the "phenylalanine anchor". Selection of three additional pharmacophore points on different parts of the ligand were necessary for running queries and could include all possible ligand- protein interactions, such as hydrogen bond donors or acceptors, aromatic rings, hydrophobic residues or charged ions. We applied an additional filter, limiting the molecular weight of the hits to 400 Da. We screened all 27 MCR reactions, but interestingly all our top hits were based on the Groebke- Blackburn- Bienaymé (GBB) three- component reaction (GBB- 3CR).34- 36 The GBB- 3CR utilized aldehydes, 2- aminopyridines and isocyanides, leading to imidazo[1,2- a]pyridines.37,38 This privileged scaffold has been present in several clinical candidates and marketed drugs, such as zolpidem, mipropfen, minodronic acid, and olprinone.39  
+
+Comparison of the docking pose of the proposed GBB compounds with the co- crystal structure of compound 127 revealed a considerable 3D shape complementarity of the two scaffolds (Fig 1B). Additionally, the GBB scaffold was drug- like and more rigid compared to our original ligand, potentially restricting the possible ligand conformations. Further docking and selection of suitable substituents to achieve favorable ligand – protein interactions were performed with the docking software SeeSAR [version 14.0.0; BioSolveIT GmbH, Sankt Augustin, Germany, 2024, www.biosolveit.de/SeeSAR].  
+
+Once compounds were selected for synthesis, synthetic routes were established using commercially available aldehydes and 2- aminopyridines, whereas isocyanides were synthesized from primary aromatic amines (Fig 1C- 1D). Two general, three- step synthetic routes were developed based on the type of aldehyde building blocks used. In the first synthetic route (route A), Boc- protected aldehydes reacted with 2- aminopyridines and isocyanides in methanol, using scandium triflate as a Lewis acid catalyst to form the GBB intermediate. The Boc- protecting group was removed under acidic conditions and the chloroacetamide warhead was introduced with an amidation reaction, either with chloroacetyl chloride or an amide coupling. In the second synthetic route (route B), to reduce the cost of certain Boc- protected aldehydes, nitro- substituted aromatic aldehydes were used instead. The GBB- 3CR was performed under the same experimental conditions, followed by a nitro- reduction. The nitro- group was reduced either using iron trichloride and zinc or using ammoniotrihydroborate and gold catalysts with Au/TiO2.40 The two reduction methods led to comparable, almost quantitative yields. The last step, as previously, was the introduction of the electrophilic warhead. Thus, the GBB- 3CR allowed multiple combinations of the three main building blocks, facilitating the synthesis of analogs. Synthetic details are provided in the Supplementary Information.  
+
+To test whether this scaffold was suitable for the development of 14- 3- 3α/ERα stabilizers, we synthesized a small set of derivatives varying only the isocyanide position, which according to our design was expected to interact with K122 of 14- 3- 3. We included benzyl and phenyl isocyanides with a diverse substitution pattern on the aryl ring. We maintained the covalent chloroacetamide warhead, based on our extensive investigation of electrophiles in our previous work.21  
+
+The compounds were tested in an intact mass spectrometry (MS) assay, which monitored the formation of the covalent bond with C38 of 14- 3- 3α. Binding measurements were made in the presence or absence of the phospho- ERα peptide and as a function of time to distinguish between cooperative stabilizers and neutral binders and to select compounds with fast binding kinetics. The phospho- ERα peptide was used at a concentration 2- fold above the dissociation constant (Kd). The assay was performed as a compound titration (Fig S2- S3). We additionally quantified compound binding at 1 μM (10:1 compound:protein) over several time points (table S2). Throughout this work, we reference the time- course data as bar graphs, using grey bars to represent binding to 14- 3- 3α alone ("apo") and colored bars to represent binding to 14- 3- 3 α/ERα complex.
+
+<--- Page Split --->
+
+Testing the first analogs in the MS assay revealed striking differences among benzyl and phenyl analogs (Fig 1E, Fig S4). The benzyl analog 1, which lacked aryl ring substitutions, was a neutral binder. Introduction of electron withdrawing groups, such as halogens or nitro- groups in the o-, m- or p- position (compounds 2- 9) reduced binding in the presence of ERα, even in the case of a small F- substitution. Interestingly, the non- substituted phenyl analog 10 showed increased binding to 14- 3- 3o, but still lacked cooperativity, since it showed increased binding in the presence or absence of ERα. Introduction of halogens in the p- position significantly reduced apo binding, improving compound cooperativity (compounds 11 and 12). However, in contrast to the SAR observed in our previous series<sup>21</sup>, less binding was observed for the p- Cl analog (11), compared to the non- substituted analog (10). Remarkably, a Me- group in the o- position (13) led to the first molecular glue of the series; compound 13 showed binding in the presence of ERα and very low apo binding. Bigger substitutions in the o- position significantly decreased the observed binding, indicating unfavorable steric effects (14- 16) (Fig S4).  
+
+We solved co- crystal structures of the 14- 3- 3o/ERα complex with compounds 1, 2 and 10 to elucidate their binding modes (Fig 1F- I). The crystal structure of compound 1 revealed a comparable binding mode to the original ligand 127, supporting the design and docking hypothesis (Fig 1F- G). The chloroacetamide moiety of 1 bound covalently to C38 of 14- 3- 3o, and two water- mediated hydrogen bonds were formed between the carbonyl group and R41 and the backbone of E115, and one direct hydrogen bond with N42 of 14- 3- 3 at 3A. The imidazopyridine ring was positioned towards the hydrophobic residues (L218, I219, G171 and L222) at the "roof" of the 14- 3- 3 pocket. In addition, the nitrogen of the five- membered ring formed a hydrogen bond with D215 of 14- 3- 3, which adopted two conformations upon binding of compound 1. Importantly, a large water network was formed between the benzylamine of 1 to N42, S45 of 14- 3- 3, which reached the terminal carboxylic acid of V595 and the pT594 of ERα, and K122 of 14- 3- 3. The benzyl ring, adjacent to the electrophilic warhead was positioned between the hydrophobic 14- 3- 3 residues F119 and I168, (Fig 1G). The structural overlay of analogs 1 and 2 showed an identical binding mode with the additional o- F substituent forming a halogen bond with I219 of 14- 3- 3 (Fig 1H). Replacing the benzyl ring (1) with a phenyl ring (10) resulted in a slight turn of the compound that placed the aryl rings of both analogs in the same position in the pocket surrounding K122. The water network that connected the amine of compound 1 to residues of 14- 3- 3 and ERα was conserved upon modification of the benzyl ring (1) to a phenyl ring (10). The removal of the methylene group positioned the aniline nitrogen of (10) closer the terminal carboxylic acid of ERα, in the position previously occupied by the methylene group of (1). Additionally, it led to increased axial rotation of the bond between the nitrogen and the aryl ring, which might have resulted in the increased apo binding in the MS assay, since this axial rotation was not possible for the benzyl analogs (Fig 1I).  
+
+Since the o- Me substitution correlated with improved binding in the presence of ERα in the MS assay and taking into account the rotational nature of the aniline - phenyl ring bond, we designed and synthesized a series of analogs with double o- substitutions (Fig 2A, Fig S5). The symmetrical 2,6- di- Me analog 17 showed significantly faster binding in the presence of ERα and almost no apo binding, indicating molecular- glue- like binding. Analog 18 with an o- Et group in addition to o- Me, was weaker, indicating unfavorable steric effects (Fig S5). Analogs 19, 20, and 21 with additional o- OMe, o- F and o- Cl groups respectively, showed similar, rapid binding to the symmetric analog 17 and low apo binding. In agreement with our previous observation, the introduction of additional substitutions in the p- position (- Cl, - Me, - OH) significantly reduced binding (compounds 22- 25), especially for the triple substituted analogs (double o- and p- positions).  
+
+Co- crystal structures of the 14- 3- 3o/ERα complex with compounds 17, 19, 20, 21, and 25 revealed differences in the positions of the double- ortho substituents (Fig 2B- E). In the case of the symmetric analog 17 one o- Me group was oriented in the back of the pocket, which was primarily hydrophobic, forming hydrophobic interactions with I219 (3.2A) and facing Val595 of ERα. The second o- Me group was oriented in the front of the pocket and formed hydrophobic interactions with F119 (3.8 A). The formation of these interactions seemed to favorably restrict the axial rotation of the aniline- phenyl bond, locking its position in a highly complementary shape with the composite 14- 3- 3/ERα surface. In analog 19 the larger o- OMe group was positioned in the front of the pocket and formed, together with the aniline nitrogen, a water- mediated hydrogen bond with the terminal carboxylic acid of Val595 of ERα. The orientation of the warhead amide differed in two analogs, but in both cases a direct hydrogen bond with N42 of 14- 3- 3 was formed (3.0 A). For the halogen- containing analogs o- F (20) and o- Cl (21) the halogens interacted with I219 and the common o- Me group with F119, whereas the aniline nitrogen interacted with the terminal carboxylic acid of Val595 of ERα via a water mediated- hydrogen bond. The binding mode of the triple substituted analog 25, although overall comparable with the double substituted analog 17, showed an additional hydrogen bond between the p- OH group and the backbone
+
+<--- Page Split --->
+
+carbonyl group of I168 (3.2 Å). The presence of this direct interaction seemed to negatively affect the axial rotation of the aniline-phenyl bond, resulting in significantly reduced binding in the MS assay and loss of cooperativity.  
+
+![](images/Figure_2.jpg)
+
+<center>Figure 2. SAR and crystal structures of selected double-ortho substituted analogs. A) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-E) Crystal structures of 14-3-3o/ERα with compounds 17 (dark pink sticks), 19 (teal sticks) and overlays of crystal structures for compounds 17 (dark pink sticks), 20 (pale green sticks), 21 (pale purple sticks) and 25 (bright yellow sticks). </center>  
+
+We synthesized analogs of compound 21, maintaining the 2- Cl,6- Me substitutions and varied positions X and Y on the scaffold (Fig 3A, Fig S6). Position X referred to substitutions on the imidazopyridine ring and was expected to contribute to additional hydrophobic interactions with Val595 of ERα. Position Y included modifications that were expected to improve interactions with the rim of the 14- 3- 3 pocket, primarily with variations of the aldehyde building blocks and to a smaller extent with different electrophilic warheads. In both cases, even the introduction of small groups led to surprising effects in the binding modes, as revealed by crystal structures.  
+
+For position X, aiming for hydrophobic interactions with Val595 of ERα, we introduced groups with a range of radii (- F (26), - Cl (27), - Me (28), - i- Pr (29), - CF3 (30)). Steric effects seemed to play a bigger role than electronic effects, since the best tolerated substitutions were - Cl (27) and - Me (28). For position Y, derived from the aldehyde building blocks, the o- F analog (31) was tolerated, but was weaker compared to the unsubstituted (21); a plausible explanation being the rotational nature of the aryl ring- imidazopyridine ring bond, which lost axial symmetry with the presence of the o- F group. Introduction of a methylene linker (32) or replacement of the aryl ring with a cyclohexyl ring (33) that had a similar size made the compounds almost unable to bind in the MS assay. Modifications in the position of the electrophilic warhead were also not tolerated. The less- reactive ester analog (34) was inactive, whereas halogenated chloroacetamides (35- 37) did not lead to the expected mass adducts, showing instability in the MS and unclear chemical reactivity with 14- 3- 3. A plausible explanation was their increased reactivity, which correlated with reduced stability. The sulfamate warhead (38) was tolerated in the MS assay, but was significantly weaker than the chloroacetamide analog, in addition to being less atom efficient.  
+
+Crystal structures were solved for 28, 32, and 33 and were compared as overlays with analog 21 (Fig 3B- E). The - Me group on the imidazopyridine ring of 28 was expected to contribute to hydrophobic interactions with Val595 of ERα. However, it disrupted the previously observed binding mode and instead moved the imidazopyridine ring upwards, in the hydrophobic pocket of 14- 3- 3 formed by L218, I219 and L222. This upward movement affected the conformation of L218, which moved upwards but maintained contact with 28. The movement in the pocket also altered the interactions of the double- ortho- substituted aryl ring; while the o- Cl still interacted with I219 on the roof of the binding site, the o- Me
+
+<--- Page Split --->
+
+group moved further away from F119 in the bottom of the pocket, and the interaction was lost. The direct hydrogen bond between the amide and N42 and the water- mediated bond between the aniline nitrogen and the carboxylic acid of Val595 were maintained. Overall, the changes in binding mode were associated with slower binding, but a similar apparent Kd in the MS assay (Fig. S3). Analog 32, with an additional methylene group on the electrophilic warhead linker showed a slightly disrupted binding mode. While the biggest difference was the position of the aryl ring next to the longer linker, the imidazopyridine ring also moved further away from ERα; the latter correlated negatively with the binding observed in the MS assay. In contrast to the previous analogs, the o- Cl group on the aryl ring of 32 pointed out of the 14- 3- 3 pocket and due to increased distance, was unable to interact with F119. Thus, the presence of an additional methylene group on the linker was sufficient to make the compound unable to stabilize the complex. Analog 33 bearing a cyclohexyl ring instead of an aromatic ring maintained the interaction between the o- Cl group of the aryl ring and Ile219, but not the interaction of the o- Me group with F119. The orientation of the amide bond next to the warhead also differed; however, the hydrogen bond with N42 was still formed. The presence of the cyclohexyl ring, which had the possibility of adopting more conformations compared to the more rigid aryl ring, correlated with reduced binding to the 14- 3- 3/ERα complex in the MS assay.  
+
+![](images/Figure_3.jpg)
+
+<center>Figure 3. SAR and crystal structures of analogs substituted in positions X and Y. A) MS bar graphs at 1 \(\mu \mathrm{M}\) . For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-E) Crystal structure of 14-3-3o/ERα with compound 28 (dark yellow sticks), and overlays of crystal structures for compounds 28 (dark yellow sticks), 32 (pink sticks) and 33 (emerald green sticks) with 21 (pale purple sticks). </center>  
+
+Investigation of modifications in positions X and Y provided valuable input in the groups that were tolerated; however, these single- point substitutions did not significantly improve the stabilization effect in the MS assay. Based on the crystallographic input, we synthesized four analogs combining the favorable substitutions from positions X and Y with the symmetric double o- Me analog 17, hypothesizing that a symmetric rotatable bond would correlate with improved potency (Fig 4A). This hypothesis was readily confirmed by evaluating the symmetric analogs 39- 42 in the MS assay. Analogs 39 and 40 included the - Me group on the imidazopyridine ring (X position) and the - F group on aryl ring (referred to as W position), respectively. Analogs 41 and 42 both had the F group in W position and a - Me or - Cl group in X position, respectively. All analogs showed comparable, fast, cooperative binding to 14- 3- 3/ERα in the MS assay, indicating that the symmetry of the 2,6- di- Me- aniline ring was more favorable compared to the asymmetric 2- Cl,6- Me. All four analogs showed low apo binding with analog 41 showing the lowest.
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Figure 4. SAR and crystal structures of 2,6-di-Me analogs substituted in positions X and W. Biophysical data (MS, TR-FRET, SPR) and cell data (NanoBRET). A) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-C) Overlays of crystal structures of 14-3-3α/ERα ternary complexes with compounds 40 (orange sticks), 41 (red sticks), and 42 (blue sticks). D) TR-FRET schematic and protein titration data for representative compounds at 100 μM compound or DMSO. E) SPR data for the binary 14-3-3α/ERα interaction and ternary interactions with 181, 17 and 41 (mean +/- SD, n=2). F-G) 14-3-3α-HaloTag/Nuc-ERα NanoBRET assay in HEK293T cells with compound titrations (1:2 dilution, starting at 40 μM). Data points excluded where compound dosage was toxic to the cells. MCR compounds compared to the previously described stabilizer 181 and 85, an inactive compound as the negative control. Bar graphs quantifying pEC50 values. </center>
+
+<--- Page Split --->
+
+Crystal structures were solved for the symmetric analogs 40, 41 and 42, highlighting the significance of two rotational bonds in the scaffold: the aniline- aryl ring bond, as previously discussed and the aryl ring- imidazopyridine ring bond (Fig 4B- C). The o- F substituent in W position, adjacent to the aryl ring- imidazopyridine rotational bond, adopted two different orientations; an inward orientation for analog 40 and an outward orientation for 41 and 42, which each had an additional substituent on the imidazopyridine ring. With the inward conformation of analog 40, the o- F substituent interacted with P167 of 14- 3- 3, whereas for 41 and 42 the o- F group was oriented to made multiple stabilizing interactions, being approximately 3 Å from N42 of 14- 3- 3, and the o- Me substituent and aniline nitrogen of the compound itself, thus contributing both to ligand- protein interactions and intramolecular interactions, restricting the axial bond rotation. In agreement with previous observations, the presence of a substituent in the X position (41 and 42) led to an upward movement of the imidazopyridine ring, orienting the substituent toward the hydrophobic residues I219 and L222. Overall, the presence of the additional substituent on the imidazopyridine ring in the last two analogs, even though in a distant position compared to the o- F- substituted ring, correlated favorably with the restriction of the rotational bond on the F- aryl ring. The different substituents of the imidazopyridine ring, Me (41) or Cl (42), did not affect the position of the stabilizer in the crystal structures; however, in the MS assay the Me group showed lower apo binding to 14- 3- 3, resulting in higher cooperativity of the ternary complex.  
+
+## TR-FRET  
+
+Up to this point, the SAR was developed using the intact MS assay and supported by crystallography. In our previous work \(^{20 - 22}\) , we relied on a fluorescence anisotropy assay (FA) to confirm cooperativity. The FA assay typically used (5- carboxylfluorescein) FAM- labeled phospho- peptides to quantify peptide binding to the compound/14- 3- 3 complex. For the MCR scaffold, however, this assay proved to be unsuitable, since the extensively conjugated ring system was intrinsically fluorescent in the same wavelengths as the FAM- labeled peptide (480 nm and 520 nm). To circumvent this issue, we turned our attention to far- red fluorescent dyes. The cy5- labeled ERα- peptide with excitation wavelength of 651 nm and emission wavelength of 670 nm, significantly affected the Kd of the 14- 3- 3/ERα complex. The reported Kd using the acetylated ERα peptide in an isothermal calorimetry (ITC) experiment or the FAM- labeled- ERα in an FA assay is in the range of 1- 2 μM. \(^{21}\) The cy5- ERα- peptide, however, resulted in a Kd of 6 nM, indicating significant dye binding (Fig S7).  
+
+As a suitable alternative to the FA assay, we developed a TR- FRET assay \(^{41}\) , using HIS- tagged- full- length 14- 3- 3σ, biotin- labeled- ERα peptide, an anti- HIS- tag monoclonal antibody conjugated with a Tb(III) criptate as the donor and streptavidin conjugated with the D2 dye as the acceptor. The observed Kd for this system was 30 nM. To ensure that the observed difference in Kd from ITC or the FAM- labeled- ERα in an FA assay was not related to the biotin- tag on the peptide, we performed a competition assay with the FAM- labeled peptide. The Kd of the biotin- peptide was 2.5 μM, in good agreement with the FAM- peptide's Kd (Fig S7). This result suggested that the lower Kd in the TR- FRET assay was due to avidity \(^{42}\) ; since 14- 3- 3 and the antibody were both dimers and streptavidin was tetravalent, different multivalent complexes could form, resulting in differences in the observed binding affinity between the labeled and the unlabeled proteins.  
+
+We performed assay optimization with 2D- titrations of 14- 3- 3, biotin- ERα and donor/acceptor ratios. We then tested the synthesized compounds using the optimized experimental conditions: 200 nM 14- 3- 3 (top concentration, 2- fold dilution), 50 nM biotin- ERα, 0.166 nM MAb anti- 6HIS Tb and 6.25 nM SA- D2. The compounds were tested at 100 μM and DMSO was used a negative control. 14- 3- 3 was titrated using an Echo acoustic dispenser, followed by the addition of the compounds and the peptide; after 1h incubation at room temperature, the donor and acceptor were added. In contrast to the MS assay, maximum signal was obtained after 2h rather than 16h - a plausible explanation again being avidity. The tight Kd in the TR- FRET assay resulted in a small assay window and a hook effect was observed at higher protein concentrations (ca. \(50 - 100\) nM 14- 3- 3).  
+
+Nevertheless, the TR- FRET assay was sensitive to the addition of molecular glues. In addition to quantifying fold stabilization for the 14- 3- 3/ERα complex \((AppKd_{(compound)} / AppKd_{(DMSO)})\) , we also quantified the fold increase in the TR- FRET signal (the ratio of observed Emax and Emin). To validate the TR- FRET assay, we included two positive controls: the natural product FC- A and our previously described stabilizer compound \(181^{21}\) . The two compounds gave comparable results. FC- A had an \(AppKd_{(compound)}\) of 11 nM, fold- stabilization of 3.09, and fold- increase of 5.52. Compound 181 had an \(AppKd_{(compound)}\) of 8.6 nM, fold- stabilization of 3.95 and fold- increase of 5.52 (Fig S8, table S3).
+
+<--- Page Split --->
+
+In good agreement with the MS data, neutral binders 1 and 10 showed only a small signal shift compared to the DMSO control (Fig 4D, Fig S8). The 2- Me analog 13 showed 4.19- fold- stabilization and 2.36- fold increase, whereas 2,6- di- Me analog 17 showed 4.78- fold- stabilization and 2.58- fold increase, the same rank order as the MS assay. The more sterically hindered analog 18 (o- Me, o- Et) showed weaker stabilization (3.06- fold- stabilization and 2.14- fold increase). Analogs 19, 20, and 21, all bearing the o- Me group and additional o- OMe, o- F and o- Cl groups respectively, showed comparable stabilization (4.41 – 5.23- fold stabilization and 2.1 – 2.48- fold increase). The highest stabilization effect was observed for the final symmetric analogs 39 – 42. Analog 39 with a - Me group in X position had an AppKd(compound) of 7.8 nM, fold- stabilization of 4.35 and fold- increase of 2.61, whereas 40 with the - F group in W position had an AppKd(compound) of 8.4 nM, fold- stabilization of 4.04 and fold- increase of 3.12. Analog 41, which had both the - Me group in X position and the - F group in the W position had the lowest AppKd(compound) (5.2 nM) and the highest fold- stabilization (6.53) and fold- increase (3.71) of the series. Analog 42, which differed only in the X position (- Cl instead of - Me group) appeared weaker (AppKd(compound) 8.8 nM, fold- stabilization of 3.87 and fold- increase of 2.97). In summary, while the fold- changes were dampened by avidity, 14- 3- 3/ERα molecular glues showed the same rank- order in the TR- FRET assay as in the mass spectrometry assay used for initial SAR.  
+
+## SPR  
+
+Surface Plasmon Resonance (SPR) was then used to analyze the kinetic parameters of the ERα peptide binding to 14- 3- 3α in the presence of compounds 181, 17 and 41, and to compare the AppKd(compound) and kinetics to the binary ERα/14- 3- 3α interaction (Fig 4E, Fig S9). Here, 14- 3- 3α tagged with a Twinstrept- tag was captured on a SPR chip coated with Strep- Tactin XT, after which a 2- fold dilution series of acetylated ERα phospho- peptide was injected. For the binary interaction, the fast dissociation rate (koff) reached the limit of detection of the SPR instrument, due to a relative weak interaction with a Kd of 1.1 μM, which was in line with the ITC and FA experiments. The covalent bond between the chloroacetamide warhead of the compounds and C38 of 14- 3- 3α was formed after overnight incubation in the presence of ERα. Immobilization of this complex on the chip, followed by extensive washing to remove the bound ERα peptide, allowed us to determine the kinetics of ERα binding to the 14- 3- 3α/stabilizer complex. The previously described stabilizer 181 decreased the off- rate to 0.016 s- 1 and simultaneously increased the association rate (kOn) by a factor of 16, resulting in a low nanomolar affinity constant for the ERαpeptide/14- 3- 3 complex (3.9 nM; stabilization = 282- fold). The analogs of the newly designed scaffold, 17 and 41, both led to an 8- fold increase in association rate compared to the binary interaction. The binding of compound 41 induced a stronger decrease in dissociation rate of ERα compared to 17, which resulted in Kd values of 10.1 nM and 15.1 nM for 41 and 17, respectively. The higher stabilizing potency of 41 compared to 17 (stabilization = 110- fold for 41 and 71- fold for 17) were in rank- order agreement with the TR- FRET data. The decreased dissociation rates in the presence of the stabilizers, especially for 41 and 181, increased the residence time of ERα binding to 14- 3- 3 from approximately 3.4 s to 47.6 s and 62.5s, respectively.  
+
+## NanoBRET  
+
+To test the effects of the compounds on the full- length PPI, we used a NanoBRET assay we developed previously43. Compounds were tested using a C- terminal fusion14- 3- 3α- HaloTag and full length, N- terminal fusion NanoLuc- ERα (Fig 4F- G; Table S4). Briefly, NanoLuc- ERα and 14- 3- 3α- HaloTag plasmids were transfected in 1:10 ratio in hormone deprived HEK293T cells. After 48 hours post transfection, cells were seeded in assay plates with the experimental wells treated with 100 nM HaloTag NanoBRET 618 ligand (Promega) and no- ligand control wells treated with DMSO (v/v). Cells were treated for 24 hours with compounds in 1:2 dilution series starting at 40 μM. The BRET signal was read and normalized against DMSO treated samples. All active compounds resulted in an increase in BRET signal compared to the negative control, 85. The previously described compound 18121 stabilized the 14- 3- 3α/ERα complex in cells with an EC50 value of 5.2 μM and a 1.7- fold increase in the BRET signal. Of the compounds tested, the neutral binder 10 was less effective than 181 (EC50 value > 100 μM, 1.6- fold increase). Compound 13 showed an improved EC50 and fold- increase compared to 10 (12 μM and 1.8- fold). Compound 17 had the lowest EC50 value of 2.7 μM and showed 1.8- fold increase in BRET signal, a slight improvement compared to 181. Compounds 40 and 42 had the same 2- fold- increases in BRET signal, though exhibited slightly different EC50 values (7.2 and 4.6 μM, respectively). Compound 41 resulted in the highest increase in BRET signal, a 2.6- fold increase; however, it had a similar EC50 value to 181 of 5 μM. In the NanoBRET assay, compounds 17 and 41 performed most effectively, based on the EC50 value (17) and fold- increase in BRET signal (41) of the panel of compounds tested. The majority of MCR stabilizers showed similar EC50 values to the previous scaffold represented by 181, but consistently showed improved fold- increase in BRET signal.
+
+<--- Page Split --->
+
+Compounds 181, 17, and 41, along with the negative control, 85, were tested in a NanoBRET assay where the cysteine of interest in 14- 3- 3α C38, was mutated to an asparagine, 14- 3- 3σ<sup>C38N</sup>- HaloTag (Fig S10). The BRET signal did not increase for 85, 181, or 17 with increasing compound concentration. There was a minimal increase in BRET signal for compound 41, from 2.5 to 3.5 mBU at 20 μM 41. In comparison to the assays done with 14- 3- 3σ<sup>WT</sup>- HaloTag, the non- normalized BRET signal increased from 6.7 to 16.1 mBU at the same concentration of 41 (data shown as normalized). In summary, when C38 was not present in 14- 3- 3σ, the compounds were unable to bind and stabilize the full- length 14- 3- 3σ/ERα complex in cells (Fig S10).  
+
+## Discussion  
+
+Using MCR chemistry, we demonstrated the potential of a scaffold- hopping approach for molecular glues stabilizing a native 14- 3- 3/cient PPI. Our approach combined computational de novo design with multi- component reaction chemistry, which included short, efficient synthetic routes with multiple points of variation, accelerating the synthesis of analogs. Thus two distinct optimization approaches led to two highly diverse chemical series, converging on similar efficacies as 14- 3- 3σ/ERα molecular glues.  
+
+In the MCR approach, structure- activity relationships were established with an intact mass spectrometry assay, which allowed the distinction between neutral binders and stabilizers - cooperative molecular glues. Overall, the SAR showed that the introduction of even small substituents in the case of molecular glues could profoundly affect their potency and cooperativity. The best compounds, eg, 41, was already \(50\%\) bound at \(1\mu M\) after \(1\mathrm{hr}\) of incubation. Crystal structures of ternary complexes were crucial in elucidating small changes in the binding modes of otherwise highly similar analogs. Starting from the weak neutral binder 1 and by removing one methylene group, we significantly increased binding to the complex for compound 10. Introduction of substituents in the o- position was sufficient to reduce apo binding and turn the compounds from neutral binders to cooperative molecular glues. Although the number of rotational bonds is generally kept to a minimum, in this case we took advantage of two rotational bonds in the scaffold and with appropriate substituents achieved favorable ligand conformations resulting in increased potency, as shown for analogs 17 and 41.  
+
+Several biophysical assays provided complementary strategies for evaluating these novel molecular glues. A TR- FRET assay circumvented the issue of scaffold fluorescence, which could have been a limiting factor in establishing SAR and allowed the rank- ordering of analogs. A new SPR assay, in which the covalent ligand was pre- associated with 14- 3- 3σ before immobilization, provided an insightful analysis of binding/unbinding kinetics and started to hint at differences between the two series of molecular glues. Saturating concentrations of compound 41, for instance, stabilized the ERα/14- 3- 3 complex by 110- fold and slowing the koff by 14- fold. Taken together, biophysical assays allowed the quantification of cooperativity (MS, TR- FRET) and kinetics (SPR); importantly, the observed SAR was consistent among these three assays. The NanoBRET assay further showed good correlation with the biophysical protein/phosphopeptide assays while demonstrating stabilization of the full- length proteins with an EC50 value of \(2.7 - 5\mu M\) in intact cells. Overall, the optimized, cell- active MCR scaffold will facilitate chemical biology approaches to study the 14- 3- 3/ERα interaction, which has so far been unexploited for drug discovery.  
+
+## Methods  
+
+## PROTEIN EXPRESSION AND PURIFICATION  
+
+The 14- 3- 3 α isoform (full- length for mass spectrometry and TR- FRET assays, \(\Delta C\) for crystallography) with an N- terminal His6 tag was expressed in Rosetta™ 2(DE3)PLysS competent E. coli (Novagen) from a pPROEX HTb expression vector. After transformation following manufacturer's instructions, single colonies were picked to inoculate \(30~\mathrm{mL}\) precultures (LB), which were added to \(1.5\mathrm{L}\) terrific broth (TB) medium after overnight growth at \(37^{\circ}C\) , 250 rpm. Expression was induced upon reaching \(\mathrm{OD}_{600}\) \(1.9 - 2.1\) by adding \(400~\mu \mathrm{M}\) IPTG. After overnight expression at \(30^{\circ}C\) , \(150~\mathrm{rpm}\) , cells were harvested by centrifugation at \(6,500~\mathrm{rpm}\) , resuspended in lysis buffer ( \(50~\mathrm{mM}\) HEPES pH 7.5, 500 mM NaCl, \(20~\mathrm{mM}\) imidazole, \(10\%\) glycerol, \(1\mathrm{mM}\) TCEP), and lysed by sonication. The His6- tagged protein was purified by Ni- affinity chromatography (Ni- NTA Agarose, Invitrogen) (Wash buffer \(50~\mathrm{mM}\) HEPES pH 7.5, \(500~\mathrm{mM}\) NaCl, \(20~\mathrm{mM}\) imidazole, \(1\mathrm{mM}\) TCEP; Elution buffer \(50~\mathrm{mM}\) HEPES pH 7.5, \(500~\mathrm{mM}\) NaCl, \(500~\mathrm{mM}\) imidazole, \(1\mathrm{mM}\) TCEP) and analyzed for purity by SDS- PAGE and Q- Tof LC/MS. The protein was buffer exchanged (Storage buffer \(25~\mathrm{mM}\) HEPES pH 7.5, \(150~\mathrm{mM}\) NaCl, \(1~\mathrm{mM}\) TCEP) and concentrated to \(\sim 16~\mathrm{mg / mL}\) and aliquots flash- frozen for storage at \(- 80^{\circ}C\) . The \(\Delta C\) variant was truncated at the C- terminus after T231 to enhance crystallization and after the first Ni- affinity chromatography column, the construct was treated with TEV protease to cleave off the His6 tag during dialysis ( \(25~\mathrm{mM}\)
+
+<--- Page Split --->
+
+HEPES, pH 7.5, 200 mM NaCl, 5% glycerol, 10 mM MgCl2, 250 μM TCEP) overnight at \(4^{\circ}C\) . The flow- through of a second Ni- affinity column was subjected to a final purification step by size exclusion chromatography (Superdex 75 pg 16/60 size exclusion column (GE Life Science) (SEC buffer: 25 mM HEPES pH 7.5, 100 mM NaCl, 10 mM MgCl2, 250 μM TCEP). The protein was concentrated to \(\sim 60\) mg/mL, analyzed for purity by SDS- PAGE and Q- Tof LC/MS and aliquots flash- frozen for storage at \(- 80^{\circ}C\) .  
+
+## PEPTIDES  
+
+Peptides for mass spectrometry, fluorescence anisotropy and TR- FRET assays were purchased from Elim Biopharmaceuticals, Inc. (Hayward, CA). Peptides for SPR and X- ray crystallography was purchased from GenScript Biotech Corp. The following peptides were used:  
+
+Ac- KYYITGEAEGFPA(pT)V- COOH (MS assay, 15- mer), 5- FAM- AEGFPA(pT)V- COOH (FA assay, 8mer ERα- pp), cy5- KYYITGEAEGFPA(pT)V- COOH (FA assay, 15- mer), biotin- KYYITGEAEGFPA(pT)V- COOH (TR- FRET assay, 15- mer), Ac- EGFPA(pT)V- COOH (crystallography and SPR, 7- mer)  
+
+## INTACT MASS SPECTROMETRY ASSAY  
+
+Mass spectrometry dose response assays were performed on a Waters Acquity UPLC/ Xevo G2- XS Q- Tof mass spectrometer. A Waters UPLC Protein BEH- C4 Column (300 A, \(1.7 \mu \mathrm{m}\) , \(2.1 \mathrm{mm} \times 50 \mathrm{mm}\) ) was used to desalt the samples prior to application on the mass spectrometer. For 19- point MS dose responses, \(50 \mathrm{mM}\) compound stocks in DMSO were serially diluted in 3- fold increment in a master plate, then \(1000 \mathrm{nl}\) of the compounds were transferred in the assay plates. Master mixes containing \(100 \mathrm{nM}\) full- length wild type 14- 3- 3α in the absence or presence of \(2 \mu \mathrm{M}\) ERα were then dispensed into 384 well plates (Greiner Bio- One, catalog number 784201). Assay buffer was TRIS (10 mM, pH 8.0) and final volume per well was \(50 \mu \mathrm{l}\) , with final top concentration of compounds dose response series at 1 mM. The reaction mixtures were incubated for 1h at rt before subjected to MS. Four measurements (1h, 8h, 16h, 24h) were performed for time- course experiments. The injection volume for each sample was \(6 \mu \mathrm{l}\) . \(24 \mu \mathrm{l}\) of sample were needed for the time- course experiments, so the total volume in the assay plate was adjusted to \(50 \mu \mathrm{l}\) , to account for the dead volume in the injections. Data collection and automated processing followed a custom workflow, as previously described. \(^{44}\) z Plots were created using GraphPad Prism with the log(agonist) vs. response (variable slope, four parameters) fitting model.  
+
+## \(\mathbf{K}_{\mathrm{D}}\) DETERMINATION FOR FAM-, cy5- AND BIOTIN-LABELED ERα PEPTIDES  
+
+For \(\mathrm{K}_{\mathrm{D}}\) determination, N- terminal fluorescein- labeled ERα peptide (5- FAM) or cy5- labeled ERα peptide and HIS- tag FL 14- 3- 3α were diluted in buffer (10mM HEPES pH 7.5, 150mM NaCl, \(0.05\%\) tween 20, \(0.05\%\) BGG (bovine gamma globulin)). Two- fold dilution series of 14- 3- 3 were made in black, round- bottom 384- microwell plates (Greiner Bio- one 784900) in a final sample volume of \(10 \mu \mathrm{L}\) in triplicates. FAM- or cy5- labeled ERα peptides (final assay concentration 10nM) were dissolved in assay buffer and mixed with the protein dilution series on the plates. Fluorescence anisotropy measurements were performed after 1h incubation at room temperature on an Envision HTS Dual Detector 2105 plate reader (for FAM- labeled ERα peptide: filter set lex: 480, lem: 535, and D505fp/D538 advanced dual mirror). For cy5- labeled ERα peptide filter set lex: 620, lem: 688 nm, and D658fp/D688 advanced dual mirror). Data were reported at endpoint. Prism 10 (GraphPad) was used to generate plots using the [agonist] vs. response (variable slope, four parameters) fitting model to determine \(\mathrm{K}_{\mathrm{D}}\) values.  
+
+For \(\mathrm{K}_{\mathrm{D}}\) determination of the biotin- labeled ERα peptide a competition assay was performed. 5- FAM- and biotin- labeled ERα peptide were diluted in buffer (10mM HEPES pH 7.5, 150mM NaCl, \(0.05\%\) tween 20, \(0.05\%\) BGG (bovine gamma globulin)). Two- fold dilution series of biotin- labeled ERα peptide were made in black, round- bottom 384- microwell plates (Greiner Bio- one 784900) in a final sample volume of \(10 \mu \mathrm{L}\) in triplicates. A mastermix of 14- 3- 3α and 5- FAM- ERα was dispensed on the assay plate (final assay concentrations: \(6 \mu \mathrm{M}\) 14- 3- 3α (IC80) and \(10 \mathrm{nM}\) 5- FAM- ERα). Fluorescence anisotropy measurements were performed after 1h incubation at room temperature using a Molecular Devices ID5 plate reader (filter set lex: \(485 \pm 20 \mathrm{nm}\) , lem: \(535 \pm 25 \mathrm{nm}\) ; integration time: \(50 \mathrm{ms}\) ; settle time: \(0 \mathrm{ms}\) ; shake 5 sec, medium, read height \(3.00 \mathrm{mm}\) , G- factor \(= 1\) ). Data were reported at endpoint. Prism 10 (GraphPad) was used to generate plots using the [agonist] vs. response (variable slope, four parameters) fitting model to determine \(\mathrm{K}_{\mathrm{D}}\) values. The obtained \(\mathrm{K}_{\mathrm{D}}\) value was corrected using the Cheng- Prusoff equation.  
+
+\(\mathrm{K}_{\mathrm{d}} = \mathrm{IC}_{50} / (1 + [\mathrm{S}] / \mathrm{Km})\) \(\mathrm{K}_{\mathrm{d}} = 6.9 / (1 + 50 \mathrm{nM} / 30 \mathrm{nM}) = 2.5 \mu \mathrm{M}\)
+
+<--- Page Split --->
+
+## TR-FRET PROTEIN TITRATIONS  
+
+TR- FRET PROTEIN TITRATIONSFor assay optimization, 2D titrations of biotin- labeled ERα peptide, HIS- tag FL 14- 3- 3α and streptavidin- D2 were performed in assay buffer (10mM HEPES pH 7.5, 150mM NaCl, 0.05% tween 20, 0.05% BGG (bovine gamma globulin)). The donor (MAb anti- 6HIS Tb cryoplate gold) concentration was kept constant at 0.166 nM. For TR- FRET protein titrations, biotin- labeled ERα peptide (50 nM), the compounds or DMSO (100 μM), MAb anti- 6HIS Tb cryoplate gold (0.166 nM) and streptavidin- D2 (6.25 nM) were mixed in assay buffer (10mM HEPES pH 7.5, 150mM NaCl, 0.05% tween 20, 0.05% BGG (bovine gamma globulin)). 2- fold serial dilutions of HIS- tag FL 14- 3- 3α were performed (200 nM top assay concentration, 12- point dilution series). The assay was performed in 384- well microplates (Corning 4513, low volume white) at a volume of 10 μl per well. The following procedure was used: The compounds (50 mM stocks in DMSO) were transferred in echo LDV masterplates. 20 nL were transferred from the masterplate to the assay plate to achieve 100 μM compound concentration in the assay using Echo acoustic dispensing. The biotin- labeled ERα peptide was dissolved in assay buffer (10mM HEPES pH 7.5, 150mM NaCl, 0.05% tween 20, 0.05% BGG (bovine gamma globulin)) and dispensed in the assay plate using Dragonfly. 14- 3- 3 dilution series were prepared using Echo acoustic dispensing. Assay plates were incubated for 1hr at room temperature before the addition of a mastermix containing the donor (MAb anti- 6HIS Tb cryoplate gold) and acceptor (streptavidin- D2) in assay buffer. The mastermix was dispensed with Dragonfly. Assay plates were incubated for 2hr at room temperature prior to TR- FRET measurements using the Envision HTS Dual Detector 2105 plate reader equipped with the TR- FRET filter set (320/615/665 nm) and a D407/D630 advanced dual mirror. A 50 μs delay was employed to reduce background fluorescence. The TR- FRET signal was obtained through calculating the ratio of 665 nm to 615 nm fluorescence (x 1000), and Prism 10 (GraphPad) was used to generate plots using the [agonist] vs. response (variable slope, four parameters) fitting model. At least two independent experiments were performed.  
+
+## SPR  
+
+SPRThe SPR experiments were performed at \(25^{\circ}C\) using a Biacore X100 and a 200 nm Strep- Tactic XT derivatized linear polycarboxylate hydrogel chip, medium charge density (XanTec Bioanalytics). All proteins and peptides were dissolved in fresh running buffer prepared with ultrapure water and filtered through a \(0.2 \mu m\) filter (10 mM HEPES pH 7.4, 200 mM NaCl, 50 μM EDTA, 0.005% P20). First the surface was conditioned with a 1 min injection of 3 M Guanidine HCl. Then, the recombinant 14- 3- 3α- Twinstrept protein (250 nM) was captured on flow cell 2 of the sensor chip at a flow rate of \(10 \mu L / min\) for 2 minutes, which resulted in a capture level of 1000 RU. For the ternary interaction, 14- 3- 3α- Twinstrept protein (250 nM) was first incubated overnight with \(1 \mu M\) Ac- ERα peptide and \(20 \mu M\) compound prior to immobilization to the chip. The bound ERα peptide was washed away using running buffer flowed over the chip for 15 min. Flow cell 1 was left blank as a reference surface. After immobilization of the protein, the Biacore X100 was primed with running buffer. Multi- cycle kinetic measurements were conducted at a flow rate of \(30 \mu L / min\) . A 2- fold dilution series of analyte (Ac- ERα peptide) in running buffer were injected over the sensor chip for 2 min, followed by dissociation of 3 min (binary interaction), 7 or 13 minutes (ternary interaction). For the binary interaction, the highest concentration of ERα was 50 μM, and for the ternary interactions this was 250 nM. Between cycles of one multi- cycle measurement, no regeneration step was performed due to complete dissociation of the analyte. After a measurement, the chip was regenerated by 2 times 30 sec injections of 3 M Guanidine HCl. The data was corrected by double subtracting to the reference surface (flow cell 1) and buffer injection and analyzed using 1:1 interaction fitting model with the BIA evaluation software (2020).  
+
+## X-RAY CRYSTALLOGRAPHY DATA COLLECTION AND REFINEMENT  
+
+X- RAY CRYSTALLOGRAPHY DATA COLLECTION AND REFINEMENTThe 14- 3- 3αΔC protein, acetylated ERα and compounds (50 mM stock in DMSO) were dissolved in complexation buffer (25 mM HEPES pH=7.5, 2 mM MgCl₂ and 100 μM TCEP) and mixed in a 1:2:3 or 1:2:5 molecular stoichiometry (protein : peptide : compound) with a final protein concentration of 12 mg/mL. The complex was set- up for sitting- drop crystallization after overnight incubation at 4 °C, in a custom crystallization liquor (0.05 M HEPES (pH 7.1, 7.3, 7.5, 7.7), 0.19 M CaCl₂, 24- 29% PEG400, and 5% (v/v) glycerol). Crystals grew within 10- 14 days at 4 °C. Crystals were fished and flash- cooled in liquid nitrogen. X- ray diffraction (XRD) data were collected at the European Synchrotron Radiation Facility (ESRF Grenoble, France, beamline ID23- 1, ID30A- 3/MASSIF- 3, or ID23- 2) or at the Deutsches Elektronen- Synchrotron (DESY Hamburg, Germany, beamline PETRA III). Data was processed using CCP4i2 suite (version 8.0.019). After indexing and integrating the data, scaling was done using AIMLESS. The data was phased with MolRep, using PDB 4JC3 as template. The presence of co- crystallized ligands was verified by visual inspection of the Fo- Fc and 2Fo- Fc electron density maps in COOT (version 0.9.8.93). If electron density corresponding to the co- crystallized ligand was present, its structure, restraints, and covalent bond were generated using AceDRG. After
+
+<--- Page Split --->
+
+building in the ligand, model rebuilding and refinement was performed using REFMAC5. The PDB REDO server (pdb- redo.edu) was used to complete the model building and refinement. The images were created using the PyMOL Molecular Graphics System (Schrödinger LLC, version 4.6.0). See SI table S10 for data collection and refinement statistics.  
+
+The structures were deposited in the protein data bank (PDB) with IDs: 916S (28), 916T (32), 916U (33), 916V (40), 916W (41), 916X (42), 916Y (1), 916Z (2), 9170 (17), 9171 (19), 9172 (10), 9173 (20), 9174 (21), and 9175 (25).  
+
+## NanoBRET  
+
+NanoBRET assays were performed as previously described.43 HEK293T cells were cultured DMEM, high glucose (Gibco) supplemented with \(10\%\) charcoal stripped Fetal Bovine Serum (FBS; Gibco) and \(1\%\) penicillin/streptomycin. Cells were transfected with a 1:10 ratio of Nanoluc- ERa:14- 3- 3o- HaloTag plasmid for 48 hours using jetOPTIMUS transfection reagent (Polyplus). Cells were then seeded at 8,000 cells per well in a 384- well plate (Corning #3570) in FluoroBrite DMEM (phenol red- free; Gibco) with \(4\%\) charcoal stripped FBS and treated with \(100~\mathrm{mM}\) HaloTag NanoBRET 618 Ligand (Promega) or equivalent volume of DMSO as a no ligand negative control. Following plating, cells were treated for 24 hours with compound in 1:2 dilution series starting at \(40~\mu \mathrm{M}\) (0.35% DMSO final concentration). After 24 hours, the BRET signal was read using an EnVision XCite 2105 plate reader at 618 nm (HaloTag) and 460 nm (NIuc). The final corrected NanoBRET ratio was calculated using the following equation:  
+
+\[CorrectedBRETratio = \left(\frac{618nm}{460nm}\right)_{HaloTagLigand} - \left(\frac{618nm}{460nm}\right)_{NoLigandcontrol}\]  
+
+The BRET ratios were normalized to samples treated with DMSO.  
+
+## DOCKING  
+
+Computational design for SAR optimization and docking was performed with SeeSAR version 14.0.0; BioSolveIT GmbH, Sankt Augustin, Germany, 2022, www.biosolveit.de/SeeSAR  
+
+## SOFTWARE VERSIONS  
+
+Prism (10.2.1), Illustrator (22.1 (64- bit)), Biorender (64- bit), Pymol (4.6.0), CCP4i2 (8.0.003), COOT (0.9.8.1), Phenix (1.19.2- 4158)  
+
+Supporting Information. Supplementary figures and tables, synthetic procedures, compound characterization, NMR spectra, crystallography data (PDF).  
+
+## References  
+
+1. Andrei, S. A. et al. Stabilization of protein-protein interactions in drug discovery. Expert Opin Drug Discov 12, 925-940 (2017).  
+2. Bier, D., Thiel, P., Briels, J. & Ottmann, C. Stabilization of Protein-Protein Interactions in chemical biology and drug discovery. Progress in Biophysics and Molecular Biology 119, 10-19 (2015).  
+3. Arkin, M. R. & Wells, J. A. Small-molecule inhibitors of protein-protein interactions: progressing towards the dream. Nat Rev Drug Discov 3, 301-317 (2004).  
+4. Arkin, M. R., Tang, Y. & Wells, J. A. Small-Molecule Inhibitors of Protein-Protein Interactions: Progressing toward the Reality. Chemistry & Biology 21, 1102-1114 (2014).  
+5. Smith, M. C. & Gestwicki, J. E. Features of protein-protein interactions that translate into potent inhibitors: topology, surface area and affinity. Expert Rev. Mol. Med. 14, e16 (2012).  
+6. Schreiber, S. L. Molecular glues and bifunctional compounds: Therapeutic modalities based on induced proximity. Cell Chemical Biology 31, 1050-1063 (2024).  
+7. Konstantinidou, M. & Arkin, M. R. Molecular glues for protein-protein interactions: Progressing toward a new dream. Cell Chem Biol 31, 1064-1088 (2024).  
+8. Ruan, H., Sun, Q., Zhang, W., Liu, Y. & Lai, L. Targeting intrinsically disordered proteins at the edge of chaos. Drug Discovery Today 24, 217-227 (2019).  
+9. Santofimia-Castaño, P. et al. Targeting intrinsically disordered proteins involved in cancer. Cell. Mol. Life Sci. 77, 1695-1707 (2020).  
+10. Aitken, A. 14-3-3 proteins: a historic overview. Semin Cancer Biol 16, 162-172 (2006).  
+11. Pennington, K. L., Chan, T. Y., Torres, M. P. & Andersen, J. L. The dynamic and stress-adaptive signaling hub of 14-3-3: emerging mechanisms of regulation and context-dependent protein-protein interactions. Oncogene 37, 5587-5604 (2018).
+
+<--- Page Split --->
+
+12. De Vries-van Leeuwen, I. J. et al. Interaction of 14-3-3 proteins with the estrogen receptor alpha F domain provides a drug target interface. Proc Natl Acad Sci U S A 110, 8894-8899 (2013).13. Somsen, B. A. et al. Molecular basis and dual ligand regulation of tetrameric estrogen receptor \(\alpha /14-3-3\zeta\) protein complex. Journal of Biological Chemistry 299, 104855 (2023).14. Liu, Y., Ma, H. & Yao, J. ERα, A Key Target for Cancer Therapy: A Review. OTT Volume 13, 2183-2191 (2020).15. Paterni, I., Bertini, S., Granchi, C., Macchia, M. & Minutolo, F. Estrogen receptor ligands: a patent review update. Expert Opinion on Therapeutic Patents 23, 1247-1271 (2013).16. Scott, J. S. & Barlaam, B. Selective estrogen receptor degraders (SERDs) and covalent antagonists (SERCAs): a patent review (2015-present). Expert Opinion on Therapeutic Patents 32, 131-151 (2022).17. Scott, J. S. & Klinowska, T. C. M. Selective estrogen receptor degraders (SERDs) and covalent antagonists (SERCAs): a patent review (July 2021-December 2023). Expert Opinion on Therapeutic Patents 34, 333-350 (2024).18. Jeselsohn, R., Buchwalter, G., De Angelis, C., Brown, M. & Schiff, R. ESR1 mutations—a mechanism for acquired endocrine resistance in breast cancer. Nat Rev Clin Oncol 12, 573-583 (2015).19. Visser, E. J. et al. Estrogen Receptor \(\alpha /14-3-3\) molecular glues as alternative treatment strategy for endocrine resistant breast cancer. Preprint at https://doi.org/10.1101/2024.04.25.591105 (2024).20. Kenanova, D. N. et al. A Systematic Approach to the Discovery of Protein-Protein Interaction Stabilizers. ACS Cent. Sci. 9, 937-946 (2023).21. Konstantinidou, M. et al. Structure-Based Optimization of Covalent, Small-Molecule Stabilizers of the 14-3-3/ERα Protein-Protein Interaction from Nonselective Fragments. J. Am. Chem. Soc. 145, 20328-20343 (2023).22. Visser, E. J. et al. From Tethered to Freestanding Stabilizers of 14-3-3 Protein-Protein Interactions through Fragment Linking. Angew Chem Int Ed 62, e202308004 (2023).23. Dömling, A., Wang, W. & Wang, K. Chemistry and Biology Of Multicomponent Reactions. Chem. Rev. 112, 3083-3135 (2012).24. Zarganes-Tzitzikas, T., Chandgude, A. L. & Dömling, A. Multicomponent Reactions, Union of MCRs and Beyond. The Chemical Record 15, 981-996 (2015).25. Fotopoulou, E., Anastasiou, P. K., Tomza, C. & Neochoritis, C. G. The Ugi reaction as the green alternative towards active pharmaceutical ingredients. Tetrahedron Green Chem 3, 100044 (2024).26. Li, X., Zarganes-Tzitzikas, T., Kurpiewska, K. & Dömling, A. Amenamevir by Ugi-4CR. Green Chem. 25, 1322-1325 (2023).27. Zarganes-Tzitzikas, T., Neochoritis, C. G. & Dömling, A. Atorvastatin (Lipitor) by MCR. ACS Med. Chem. Lett. 10, 389-392 (2019).28. Znabet, A. et al. A highly efficient synthesis of telaprevir by strategic use of biocatalysis and multicomponent reactions. Chem. Commun. 46, 7918 (2010).29. Wehlan, H., Oehme, J., Schäfer, A. & Rossen, K. Development of Scalable Conditions for the Ugi Reaction—Application to the Synthesis of (R)-Lacosamide. Org. Process Res. Dev. 19, 1980-1986 (2015).30. Váradi, A. et al. Synthesis of Carfentanil Amide Opioids Using the Ugi Multicomponent Reaction. ACS Chem. Neurosci. 6, 1570-1577 (2015).31. Koes, D. et al. Enabling Large-Scale Design, Synthesis and Validation of Small Molecule Protein-Protein Antagonists. PLoS ONE 7, e32839 (2012).32. Koes, D. R., Dömling, A. & Camacho, C. J. AnchorQuery: Rapid online virtual screening for small-molecule protein-protein interaction inhibitors. Protein Sci 27, 229-232 (2018).33. Neochoritis, C. G. et al. Hitting on the move: Targeting intrinsically disordered protein states of the MDM2-p53 interaction. European Journal of Medicinal Chemistry 182, 111588 (2019).34. Groebke, K., Weber, L. & Mehlin, F. Synthesis of Imidazo[1,2-a] annulated Pyridines, Pyrazines and Pyrimidines by a Novel Three-Component Condensation. Synlett 1998, 661-663 (1998).35. Blackburn, C., Guan, B., Fleming, P., Shiosaki, K. & Tsai, S. Parallel synthesis of 3-aminoimidazo[1,2-a]pyridines and pyrazines by a new three-component condensation. Tetrahedron Letters 39, 3635-3638 (1998).36. Bienaymé, H. & Bouzid, K. A New Heterocyclic Multicomponent Reaction For the Combinatorial Synthesis of Fused 3-Aminoimidazoles. Angewandte Chemie International Edition 37, 2234-2237 (1998).37. Boltjes, A. & Dömling, A. The Groebke-Blackburn-Bienaymé Reaction. Eur J Org Chem 2019, 7007-7049 (2019).38. Shukla, P., Azad, C. S., Deswal, D. & Narula, A. K. Revisiting the GBB reaction and redefining its relevance in medicinal chemistry: A review. Drug Discovery Today 29, 104237 (2024).39. Devi, N., Singh, D., K. Rawal, R., Barwal, J. & Singh, V. Medicinal Attributes of Imidazo[1,2-a]pyridine Derivatives: An Update. CTMC 16, 2963-2994 (2016).40. Vasilikogiannaki, E., Gryparis, C., Kotzabasaki, V., Lykakis, I. N. & Stratakis, M. Facile Reduction of Nitroarenes into Anilines and Nitroalkanes into Hydroxylamines via the Rapid Activation of Ammonia- Borane Complex by Supported Gold Nanoparticles. Adv Synth Catal 355, 907-911 (2013).41. Degorce, F. HTRF: A Technology Tailored for Drug Discovery - A Review of Theoretical Aspects and Recent Applications. TOCHGENJ 3, 22-32 (2009).
+
+<--- Page Split --->
+
+42. Vauquelin, G. & Charlton, S. J. Exploring avidity: understanding the potential gains in functional affinity and target residence time of bivalent and heterobivalent ligands. British J Pharmacology 168, 1771-1785 (2013). 
+43. Vickery, H. R., Virta, J. M., Konstantinidou, M. & Arkin, M. R. Development of a NanoBRET assay for evaluation of 14-3-3α molecular glues. SLAS Discov 29, 100165 (2024). 
+44. Hallenbeck, K. K. et al. A Liquid Chromatography/Mass Spectrometry Method for Screening Disulfide Tethering Fragments. SLAS Discov 2472555217732072 (2017) doi:10.1177/2472555217732072.  
+
+Acknowledgements. This research was funded by the Ono Pharma Foundation Breakthrough Science Initiative Award, NIH/NIGMS GM147696 and the Netherlands Organization for Scientific Research (NWO) through Gravity program 024.001.035 and ENW M-grant OCENW.M20.200. We acknowledge Foundation for Research and Innovation (H.F.R.I.) under the "2nd Call for H.F.R.I. Research Projects to support Post-Doctoral Researchers" (Project Number: 0911) and Emperikion Idryma (to C.G.N). We thank Amanda Paulson for the automated mass spec data processing infrastructure in the SMDC. We acknowledge the European Synchrotron Radiation Facility (ESRF) for provision of synchrotron radiation facilities, and we would like to thank David Flot and Max Nanao for assistance and support in using beamlines ID23-1, ID23-2, ID30A-3 (mx2407 and mx2526). We thank DESY (Hamburg, Germany), a member of the Helmholtz Association HGF, for the provision of experimental facilities. Parts of this research were carried out at PETRA III.  
+
+## Author contributions.  
+
+M.K. conceived the work, designed the compounds, performed the MS and TR-FRET experiments, and analyzed the data with contributions from C.O, L.C., C.G.N. and M.R.A. M.Z and M.F synthesized and characterized compounds. M.A.M.P solved most of the crystal structures and performed the SPR experiments. J.M.V. performed the NanoBRET assay. J.L.R. was involved in the development and optimization of the TR-FRET assay. E.M.J. solved the initial crystal structures. M.K., C.G.N., M.R.A, C.O. and L.B. supervised the project. M.K. wrote the manuscript with contributions from all authors.  
+
+Conflict of interest. Michelle R. Arkin, Christian Ottmann and Luc Brunsveld are co- founders of Ambagon Therapeutics.  
+
+Keywords: covalent • estrogen receptor • MCR • molecular glue • 14- 3- 3  
+
+TOC  
+
+![PLACEHOLDER_17_0]
+  
+
+We describe a scaffold- hopping approach suitable for the identification of molecular glues stabilizing the 14- 3- 3α/ERα complex. The multi- component reaction- based scaffold was rapidly optimized, and validated in biophysical assays, while several co- crystal structures elucidated the binding modes. The best compounds were tested in a cellular NanoBRET assay, showing low micromolar potency.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+20250210MCRSupplement.pdf
+
+<--- Page Split --->

preprint/preprint__09caa31641cd2fa62399cf94c4caec3caa098ba1b127a492b84d0583324ff5f2/preprint__09caa31641cd2fa62399cf94c4caec3caa098ba1b127a492b84d0583324ff5f2_det.mmd

@@ -0,0 +1,385 @@
+<|ref|>title<|/ref|><|det|>[[42, 106, 912, 210]]<|/det|>
+# Rapid scaffold-hopping for molecular glues: from fragments to cell-active probes targeting the 14-3-3/ERα complex  
+
+<|ref|>text<|/ref|><|det|>[[42, 230, 315, 277]]<|/det|>
+Michelle Arkin michelle.arkin@ucsf.edu  
+
+<|ref|>text<|/ref|><|det|>[[42, 301, 760, 323]]<|/det|>
+University of California at San Francisco https://orcid.org/0000- 0002- 9366- 6770  
+
+<|ref|>text<|/ref|><|det|>[[42, 327, 744, 370]]<|/det|>
+Markella Konstantinidou University of California San Francisco https://orcid.org/0000- 0001- 5972- 4140  
+
+<|ref|>text<|/ref|><|det|>[[42, 374, 578, 417]]<|/det|>
+Marios Zingiridis University of Crete https://orcid.org/0009- 0008- 1150- 2926  
+
+<|ref|>text<|/ref|><|det|>[[42, 421, 728, 464]]<|/det|>
+Marloes Pennings Eindhoven University of Technology https://orcid.org/0000- 0002- 3366- 0238  
+
+<|ref|>text<|/ref|><|det|>[[42, 468, 228, 508]]<|/det|>
+Michael Fragkiadakis University of Crete  
+
+<|ref|>text<|/ref|><|det|>[[42, 514, 384, 556]]<|/det|>
+Johanna Virta University of California San Francisco  
+
+<|ref|>text<|/ref|><|det|>[[42, 561, 386, 602]]<|/det|>
+Jezrael Revalde University of California, San Francisco  
+
+<|ref|>text<|/ref|><|det|>[[42, 608, 179, 647]]<|/det|>
+Emira Visser TU Eindhoven  
+
+<|ref|>text<|/ref|><|det|>[[42, 653, 728, 696]]<|/det|>
+Christian Ottmann Eindhoven University of Technology https://orcid.org/0000- 0001- 7315- 0315  
+
+<|ref|>text<|/ref|><|det|>[[42, 700, 536, 742]]<|/det|>
+Luc Brunsveld TU Eindhoven https://orcid.org/0000- 0001- 5675- 511X  
+
+<|ref|>text<|/ref|><|det|>[[42, 746, 578, 788]]<|/det|>
+Constantinos Neochoritis University of Crete https://orcid.org/0000- 0001- 5098- 5504  
+
+<|ref|>sub_title<|/ref|><|det|>[[42, 827, 103, 845]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[42, 865, 628, 886]]<|/det|>
+Keywords: covalent, estrogen receptor, MCR, molecular glue, 14- 3- 3  
+
+<|ref|>text<|/ref|><|det|>[[42, 904, 336, 923]]<|/det|>
+Posted Date: February 28th, 2025  
+
+<|ref|>text<|/ref|><|det|>[[42, 941, 475, 961]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 6051794/v1
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[42, 44, 915, 87]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 105, 936, 149]]<|/det|>
+Additional Declarations: Yes there is potential Competing Interest. M.R.A, C.O, and L.B. are co- founders of Ambagon Therapeutics. C.O. is an employee of Ambagon Therapeutics.  
+
+<|ref|>text<|/ref|><|det|>[[42, 184, 912, 227]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on July 14th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 61176- 4.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[115, 88, 881, 125]]<|/det|>
+# Rapid scaffold-hopping for molecular glues: from fragments to cell-active probes targeting the 14-3-3/ERα complex  
+
+<|ref|>text<|/ref|><|det|>[[115, 138, 881, 191]]<|/det|>
+Markella Konstantinidou\*[a], Marios Zingiridis[b], Marloes A.M. Pennings[c], Michael Fragkiadakis[b], Johanna M. Virta[a], Jezrael L. Revalde[a], Emira J. Visser[c], Christian Ottmann[c], Luc Brunsveld[c], Constantinos G. Neochoritis\*[b], Michelle R. Arkin\*[a]  
+
+<|ref|>text<|/ref|><|det|>[[115, 210, 784, 420]]<|/det|>
+[a] M. Konstantinidou, J.M. Virta, J.L. Revalde, M.R. Arkin  Department of Pharmaceutical Chemistry and Small Molecule Discovery Centre (SMDC)  University of California San Francisco (UCSF)  CA 94143 (USA)  E- mail: markella.constantinidou@ucsf.edu, michelle.arkin@ucsf.edu  [b] M. Zingiridis, M. Fragkiadakis, C.G. Neochoritis  Department of Chemistry  University of Crete  Voutes, Heraklion, 70013, Greece  E- mail: kneochor@uoc.gr  [c] M.A.M. Pennings, E.J. Visser, C. Ottmann, L. Brunsveld  Laboratory of Chemical Biology, Department of Biomedical Engineering and Institute for Complex Molecular Systems (ICMS)  Eindhoven University of Technology  5600 MB Eindhoven, The Netherlands  
+
+<|ref|>text<|/ref|><|det|>[[114, 465, 882, 695]]<|/det|>
+Abstract: Molecular glues, small molecules that bind cooperatively at a protein- protein interface, have emerged as powerful modalities for the modulation of protein- protein interactions (PPIs) and "undruggable" targets. The systematic identification of new chemical matter with a molecular glue mechanism of action remains a significant challenge in drug discovery. Here, we present a scaffold hopping approach, using as a starting point our previously developed molecular glues for the native 14- 3- 3/estrogen receptor alpha (ERα) complex. The novel, computationally designed scaffold was based on the Groebke- Blackburn- Bienaymé multi- component reaction (MCR), leading to drug- like analogs with multiple points of variation, thus enabling the rapid derivatization and optimization of the scaffold. Structure- activity relationships (SAR) were developed using intact mass spectrometry and TR- FRET. Rational structure- guided optimization was facilitated by crystal structures of ternary complexes with the glues, 14- 3- 3 and phospho- peptides mimicking the highly disordered C- terminus of ERα. We measured the kinetics of 14- 3- 3/ERα peptide binding by SPR, using a format in which a 14- 3- 3/molecular glue complex was immobilized on the SPR chip. The most potent compounds stabilized the complex by 100- fold and increased the residence time by 14- fold. Cellular stabilization of 14- 3- 3/ERα for the most potent analogs was confirmed using a NanoBRET assay with full- length proteins in live cells (EC50 = 2.7 – 5 μM). Our approach highlights the potential of MCR chemistry, combined with scaffold hopping, to drive the development and optimization of unprecedented molecular glue scaffolds.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 717, 212, 731]]<|/det|>
+## Introduction  
+
+<|ref|>text<|/ref|><|det|>[[115, 741, 882, 831]]<|/det|>
+The stabilization of native protein- protein interactions (PPIs) with small molecules offers an attractive strategy for the activation or inhibition of signaling pathways in a therapeutic context. \(^{1,2}\) PPIs were traditionally considered difficult targets due to the lack of well- defined pockets and the presence of large, hydrophobic surfaces. \(^{3 - 5}\) The fundamental understanding of the mechanism of action of molecular glues – small molecules that bind cooperatively at PPI interfaces and strengthen weak, pre- existing interactions – has enabled the stabilization of PPIs by taking into account the elements of cooperativity, molecular recognition and shape complementarity. \(^{6,7}\)  
+
+<|ref|>text<|/ref|><|det|>[[115, 844, 882, 904]]<|/det|>
+A particularly challenging class of PPIs includes proteins that are intrinsically disordered and only become partially structured when bound to a protein partner, such as a chaperone binding to a client protein. \(^{8,9}\) 14- 3- 3 is an abundant scaffolding protein that recognizes specific phospho- serine or phospho- threonine motifs on disordered domains of the client and upon binding creates a structured binding interface. \(^{10}\) Molecular glues targeting 14- 3- 3/client complexes bind
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 882, 149]]<|/det|>
+to the composite surface formed at this interface; the inherent cooperativity of this approach yields molecular glues with selectivity and potency. Of note, 14- 3- 3 proteins lack function – the function is instead encoded on the client protein and in particular on the phospho- site that is being recognized, leading either to activation or inhibition of signaling pathways. \(^{11}\)  
+
+<|ref|>text<|/ref|><|det|>[[115, 161, 882, 326]]<|/det|>
+Among the extensive interactome of 14- 3- 3, here we focus on its native interaction with the hormone regulated transcription factor estrogen receptor \(\alpha\) (ERα). 14- 3- 3 recognizes the protein sequence surrounding phospho- T594 on the disordered C- terminus on the F- domain of ERα and acts as a negative regulator by inhibiting ERα binding to chromatin and blocking ERα- mediated transcription. \(^{12,13}\) To date, ERα small molecule drugs, acting either as inhibitors or degraders, target the adjacent ligand binding domain (LBD). \(^{14 - 17}\) However, mutations in the LBD are often associated with acquired endocrine resistance. \(^{18}\) Thus, stabilization of the native 14- 3- 3/ERα PPI could be useful as an alternative strategy to block ERα transcriptional activity in ERα positive breast cancer, especially in cases of acquired endocrine resistance. The feasibility of this approach, targeting the F- domain, is corroborated by studies using the natural product fusicoccin- A (FC- A) and its semi- synthetic analogs that stabilize the interaction between 14- 3- 3 and the C- terminus of ERα. \(^{19}\) We now require drug- like chemical probes to define the biological impact of targeting the F- domain to inhibit ERα in hormone- positive breast cancer.  
+
+<|ref|>text<|/ref|><|det|>[[115, 339, 882, 474]]<|/det|>
+We have applied different strategies for the identification of chemical matter to stabilize the 14- 3- 3/ERα complex. We used a site- directed fragment- based technology, termed "disulfide tethering" with intact mass spectrometry as the readout to identify reversible fragments bound at the native cysteine (C38) of 14- 3- 3α in the presence of a phosphorylated peptide that represented the disordered C- terminus of ERα. \(^{20}\) Rational, structure- guided optimization of the reversible disulfide fragments led to irreversibly covalent, selective molecular glues that bound at the composite surface of 14- 3- 3α/ERα. \(^{21}\) For the development of non- covalent molecular glues, we used a fragment- linking approach, derived from the crystal structures of two diverse fragments that were identified in crystallographic and disulfide tethering screens. \(^{22}\) Thus, the 14- 3- 3/ERα PPI has served as a valuable system to test diverse molecular- glue discovery strategies.  
+
+<|ref|>text<|/ref|><|det|>[[115, 486, 882, 636]]<|/det|>
+Here, we present a scaffold- hopping approach based on multi- component reaction chemistry (MCR). Multi- component reactions are defined as synthetic approaches where at least three starting materials react in a single step to form a complex scaffold, where most of the atoms contribute to the newly formed product. This broad definition covers reactions with various synthetic mechanisms. \(^{23,24}\) MCR chemistry, due to its highly divergent character, is an enticing strategy for developing new scaffolds and rapid structure- activity relationships (SAR), as it allows the combination of short synthetic routes with high diversity and complexity. MCR has emerged as an attractive alternative to multistep linear convergent synthetic approaches and has been successfully applied to the synthesis of active pharmaceutical ingredients (API). \(^{25 - 30}\) Here, we describe our strategy for the development of a drug- like MCR scaffold stabilizing the 14- 3- 3α/ERα complex. The most potent analogs of the series showed efficacy in orthogonal biophysical assays and cell- based PPI stabilization in the low micromolar range.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 651, 175, 665]]<|/det|>
+## Results  
+
+<|ref|>sub_title<|/ref|><|det|>[[117, 680, 378, 694]]<|/det|>
+## Structure activity relationships (SAR)  
+
+<|ref|>text<|/ref|><|det|>[[115, 695, 882, 770]]<|/det|>
+Structurally, 14- 3- 3 binds ERα by recognizing phospho- T594, the penultimate residue on the C- terminus of ERα. This creates a large, open, solvent- exposed pocket that can accommodate a small molecule. Although steric factors are not an issue for targeting the composite surface of the 14- 3- 3/ERα complex, we found it was important to rigidify an initially flexible scaffold to maximize the stabilization effect. \(^{21}\) Our aim in this work was to design a scaffold that would be more rigid from the beginning, locking in a favorable three- dimensional shape complementary to the large pocket.  
+
+<|ref|>text<|/ref|><|det|>[[115, 784, 882, 904]]<|/det|>
+To this end, we used the freely accessible software AnchorQuery™, which performs pharmacophore- based screening of approximately 31 million compounds that are readily synthesizable through one- step multi- component reaction (MCR) chemistry. \(^{31,32}\) Although AnchorQuery™ was originally developed for PPI inhibitors \(^{33}\) , in this case it was successful in proposing MCR scaffolds for PPI stabilizers. AnchorQuery™ requires a ligand- bound crystal structure or docked binding pose as a starting point. We used the crystal structure of the previously disclosed compound 127 (PDB 8ALW) that was bound at the composite surface of the 14- 3- 3α/ERα complex, with a favorable ligand conformation, based on our biophysical data. \(^{21}\) The compound formed multiple favorable interactions both with 14- 3- 3α and the phospho- ERα peptide (Fig 1A). In the co- crystal structure, the irreversible chloroacetamide warhead of compound 127
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 884, 136]]<|/det|>
+formed a covalent bond with C38 of 14- 3- 3o. The \(p\) - chloro-phenyl ring occupied a small hydrophobic pocket that formed a halogen bond with K122 of 14- 3- 3. The tetrahydropyrane ring adopted a favorable conformation that allowed the formation of hydrophobic interactions with 14- 3- 3 residues (L218, I219), the terminal Val595 of ERα, and a water-  
+
+<|ref|>image<|/ref|><|det|>[[123, 145, 890, 797]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[140, 796, 894, 901]]<|/det|>
+<center>Figure 1. Overview of the scaffold hopping approach. SAR and crystal structures of selected benzyl and aniline analogs. A) Crystal structure of compound 127 (yellow sticks) with 14-3-3o (grey surface) and phospho-ERα peptide (orange sticks). Interacting aminoacids are shown as sticks and water molecules as red spheres (PDB 8ALW). B) Ligand overlay of compound 127 and docking pose of the new MCR scaffold. C) General MCR scaffold and main points of variation. D) Overview of general synthetic routes. Detailed experimental conditions are described in the SI. E) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. F) Crystal structure overlay for compounds 1 (cyan sticks) and 127 (yellow sticks) bound to 14-3-3o (grey surface) /ERα (orange sticks). G) Crystal structure of compound 1 (cyan sticks) with 14-3-3o/ERα. Interacting aminoacid residues are shown as sticks and interacting water molecules as red spheres. H-I) Structural overlays of compounds 2 (brown sticks), and 10 (dark red sticks) with compound 1 (cyan sticks). </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 883, 164]]<|/det|>
+mediated hydrogen bond from the oxygen atom in the ring, which was part of a large water network. The overall ligand conformation also led to a key water- mediated hydrogen bond between the aniline nitrogen and the terminal carboxylic acid of Val595 of ERα, significantly contributing to molecular recognition. In this compound series, the introduction of large non- aromatic rings, such as the tetrahydropyrane, combined with aniline rings were necessary to limit the multiple ligand conformations and improved the affinity to the complex.  
+
+<|ref|>text<|/ref|><|det|>[[114, 175, 883, 358]]<|/det|>
+Understanding the binding mode of 127, we were able to use AnchorQueryTM for a scaffold- hopping approach. The software required an anchor motif on the ligand that was bioisosteric to an amino acid residue and was kept constant for all pharmacophore- based searches of the database. A suitable anchor in our case was the \(p\) - chloro- phenyl ring that was deeply buried at the PPI interface in the small pocket surrounding K122 of 14- 3- 3 (Fig S1). This motif served as the "phenylalanine anchor". Selection of three additional pharmacophore points on different parts of the ligand were necessary for running queries and could include all possible ligand- protein interactions, such as hydrogen bond donors or acceptors, aromatic rings, hydrophobic residues or charged ions. We applied an additional filter, limiting the molecular weight of the hits to 400 Da. We screened all 27 MCR reactions, but interestingly all our top hits were based on the Groebke- Blackburn- Bienaymé (GBB) three- component reaction (GBB- 3CR).34- 36 The GBB- 3CR utilized aldehydes, 2- aminopyridines and isocyanides, leading to imidazo[1,2- a]pyridines.37,38 This privileged scaffold has been present in several clinical candidates and marketed drugs, such as zolpidem, mipropfen, minodronic acid, and olprinone.39  
+
+<|ref|>text<|/ref|><|det|>[[115, 369, 883, 460]]<|/det|>
+Comparison of the docking pose of the proposed GBB compounds with the co- crystal structure of compound 127 revealed a considerable 3D shape complementarity of the two scaffolds (Fig 1B). Additionally, the GBB scaffold was drug- like and more rigid compared to our original ligand, potentially restricting the possible ligand conformations. Further docking and selection of suitable substituents to achieve favorable ligand – protein interactions were performed with the docking software SeeSAR [version 14.0.0; BioSolveIT GmbH, Sankt Augustin, Germany, 2024, www.biosolveit.de/SeeSAR].  
+
+<|ref|>text<|/ref|><|det|>[[114, 471, 883, 668]]<|/det|>
+Once compounds were selected for synthesis, synthetic routes were established using commercially available aldehydes and 2- aminopyridines, whereas isocyanides were synthesized from primary aromatic amines (Fig 1C- 1D). Two general, three- step synthetic routes were developed based on the type of aldehyde building blocks used. In the first synthetic route (route A), Boc- protected aldehydes reacted with 2- aminopyridines and isocyanides in methanol, using scandium triflate as a Lewis acid catalyst to form the GBB intermediate. The Boc- protecting group was removed under acidic conditions and the chloroacetamide warhead was introduced with an amidation reaction, either with chloroacetyl chloride or an amide coupling. In the second synthetic route (route B), to reduce the cost of certain Boc- protected aldehydes, nitro- substituted aromatic aldehydes were used instead. The GBB- 3CR was performed under the same experimental conditions, followed by a nitro- reduction. The nitro- group was reduced either using iron trichloride and zinc or using ammoniotrihydroborate and gold catalysts with Au/TiO2.40 The two reduction methods led to comparable, almost quantitative yields. The last step, as previously, was the introduction of the electrophilic warhead. Thus, the GBB- 3CR allowed multiple combinations of the three main building blocks, facilitating the synthesis of analogs. Synthetic details are provided in the Supplementary Information.  
+
+<|ref|>text<|/ref|><|det|>[[115, 679, 882, 739]]<|/det|>
+To test whether this scaffold was suitable for the development of 14- 3- 3α/ERα stabilizers, we synthesized a small set of derivatives varying only the isocyanide position, which according to our design was expected to interact with K122 of 14- 3- 3. We included benzyl and phenyl isocyanides with a diverse substitution pattern on the aryl ring. We maintained the covalent chloroacetamide warhead, based on our extensive investigation of electrophiles in our previous work.21  
+
+<|ref|>text<|/ref|><|det|>[[114, 751, 883, 873]]<|/det|>
+The compounds were tested in an intact mass spectrometry (MS) assay, which monitored the formation of the covalent bond with C38 of 14- 3- 3α. Binding measurements were made in the presence or absence of the phospho- ERα peptide and as a function of time to distinguish between cooperative stabilizers and neutral binders and to select compounds with fast binding kinetics. The phospho- ERα peptide was used at a concentration 2- fold above the dissociation constant (Kd). The assay was performed as a compound titration (Fig S2- S3). We additionally quantified compound binding at 1 μM (10:1 compound:protein) over several time points (table S2). Throughout this work, we reference the time- course data as bar graphs, using grey bars to represent binding to 14- 3- 3α alone ("apo") and colored bars to represent binding to 14- 3- 3 α/ERα complex.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[114, 88, 883, 255]]<|/det|>
+Testing the first analogs in the MS assay revealed striking differences among benzyl and phenyl analogs (Fig 1E, Fig S4). The benzyl analog 1, which lacked aryl ring substitutions, was a neutral binder. Introduction of electron withdrawing groups, such as halogens or nitro- groups in the o-, m- or p- position (compounds 2- 9) reduced binding in the presence of ERα, even in the case of a small F- substitution. Interestingly, the non- substituted phenyl analog 10 showed increased binding to 14- 3- 3o, but still lacked cooperativity, since it showed increased binding in the presence or absence of ERα. Introduction of halogens in the p- position significantly reduced apo binding, improving compound cooperativity (compounds 11 and 12). However, in contrast to the SAR observed in our previous series<sup>21</sup>, less binding was observed for the p- Cl analog (11), compared to the non- substituted analog (10). Remarkably, a Me- group in the o- position (13) led to the first molecular glue of the series; compound 13 showed binding in the presence of ERα and very low apo binding. Bigger substitutions in the o- position significantly decreased the observed binding, indicating unfavorable steric effects (14- 16) (Fig S4).  
+
+<|ref|>text<|/ref|><|det|>[[114, 266, 883, 555]]<|/det|>
+We solved co- crystal structures of the 14- 3- 3o/ERα complex with compounds 1, 2 and 10 to elucidate their binding modes (Fig 1F- I). The crystal structure of compound 1 revealed a comparable binding mode to the original ligand 127, supporting the design and docking hypothesis (Fig 1F- G). The chloroacetamide moiety of 1 bound covalently to C38 of 14- 3- 3o, and two water- mediated hydrogen bonds were formed between the carbonyl group and R41 and the backbone of E115, and one direct hydrogen bond with N42 of 14- 3- 3 at 3A. The imidazopyridine ring was positioned towards the hydrophobic residues (L218, I219, G171 and L222) at the "roof" of the 14- 3- 3 pocket. In addition, the nitrogen of the five- membered ring formed a hydrogen bond with D215 of 14- 3- 3, which adopted two conformations upon binding of compound 1. Importantly, a large water network was formed between the benzylamine of 1 to N42, S45 of 14- 3- 3, which reached the terminal carboxylic acid of V595 and the pT594 of ERα, and K122 of 14- 3- 3. The benzyl ring, adjacent to the electrophilic warhead was positioned between the hydrophobic 14- 3- 3 residues F119 and I168, (Fig 1G). The structural overlay of analogs 1 and 2 showed an identical binding mode with the additional o- F substituent forming a halogen bond with I219 of 14- 3- 3 (Fig 1H). Replacing the benzyl ring (1) with a phenyl ring (10) resulted in a slight turn of the compound that placed the aryl rings of both analogs in the same position in the pocket surrounding K122. The water network that connected the amine of compound 1 to residues of 14- 3- 3 and ERα was conserved upon modification of the benzyl ring (1) to a phenyl ring (10). The removal of the methylene group positioned the aniline nitrogen of (10) closer the terminal carboxylic acid of ERα, in the position previously occupied by the methylene group of (1). Additionally, it led to increased axial rotation of the bond between the nitrogen and the aryl ring, which might have resulted in the increased apo binding in the MS assay, since this axial rotation was not possible for the benzyl analogs (Fig 1I).  
+
+<|ref|>text<|/ref|><|det|>[[115, 564, 882, 702]]<|/det|>
+Since the o- Me substitution correlated with improved binding in the presence of ERα in the MS assay and taking into account the rotational nature of the aniline - phenyl ring bond, we designed and synthesized a series of analogs with double o- substitutions (Fig 2A, Fig S5). The symmetrical 2,6- di- Me analog 17 showed significantly faster binding in the presence of ERα and almost no apo binding, indicating molecular- glue- like binding. Analog 18 with an o- Et group in addition to o- Me, was weaker, indicating unfavorable steric effects (Fig S5). Analogs 19, 20, and 21 with additional o- OMe, o- F and o- Cl groups respectively, showed similar, rapid binding to the symmetric analog 17 and low apo binding. In agreement with our previous observation, the introduction of additional substitutions in the p- position (- Cl, - Me, - OH) significantly reduced binding (compounds 22- 25), especially for the triple substituted analogs (double o- and p- positions).  
+
+<|ref|>text<|/ref|><|det|>[[115, 713, 883, 910]]<|/det|>
+Co- crystal structures of the 14- 3- 3o/ERα complex with compounds 17, 19, 20, 21, and 25 revealed differences in the positions of the double- ortho substituents (Fig 2B- E). In the case of the symmetric analog 17 one o- Me group was oriented in the back of the pocket, which was primarily hydrophobic, forming hydrophobic interactions with I219 (3.2A) and facing Val595 of ERα. The second o- Me group was oriented in the front of the pocket and formed hydrophobic interactions with F119 (3.8 A). The formation of these interactions seemed to favorably restrict the axial rotation of the aniline- phenyl bond, locking its position in a highly complementary shape with the composite 14- 3- 3/ERα surface. In analog 19 the larger o- OMe group was positioned in the front of the pocket and formed, together with the aniline nitrogen, a water- mediated hydrogen bond with the terminal carboxylic acid of Val595 of ERα. The orientation of the warhead amide differed in two analogs, but in both cases a direct hydrogen bond with N42 of 14- 3- 3 was formed (3.0 A). For the halogen- containing analogs o- F (20) and o- Cl (21) the halogens interacted with I219 and the common o- Me group with F119, whereas the aniline nitrogen interacted with the terminal carboxylic acid of Val595 of ERα via a water mediated- hydrogen bond. The binding mode of the triple substituted analog 25, although overall comparable with the double substituted analog 17, showed an additional hydrogen bond between the p- OH group and the backbone
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 883, 120]]<|/det|>
+carbonyl group of I168 (3.2 Å). The presence of this direct interaction seemed to negatively affect the axial rotation of the aniline-phenyl bond, resulting in significantly reduced binding in the MS assay and loss of cooperativity.  
+
+<|ref|>image<|/ref|><|det|>[[116, 135, 886, 441]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[124, 444, 881, 495]]<|/det|>
+<center>Figure 2. SAR and crystal structures of selected double-ortho substituted analogs. A) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-E) Crystal structures of 14-3-3o/ERα with compounds 17 (dark pink sticks), 19 (teal sticks) and overlays of crystal structures for compounds 17 (dark pink sticks), 20 (pale green sticks), 21 (pale purple sticks) and 25 (bright yellow sticks). </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 516, 882, 608]]<|/det|>
+We synthesized analogs of compound 21, maintaining the 2- Cl,6- Me substitutions and varied positions X and Y on the scaffold (Fig 3A, Fig S6). Position X referred to substitutions on the imidazopyridine ring and was expected to contribute to additional hydrophobic interactions with Val595 of ERα. Position Y included modifications that were expected to improve interactions with the rim of the 14- 3- 3 pocket, primarily with variations of the aldehyde building blocks and to a smaller extent with different electrophilic warheads. In both cases, even the introduction of small groups led to surprising effects in the binding modes, as revealed by crystal structures.  
+
+<|ref|>text<|/ref|><|det|>[[115, 620, 882, 802]]<|/det|>
+For position X, aiming for hydrophobic interactions with Val595 of ERα, we introduced groups with a range of radii (- F (26), - Cl (27), - Me (28), - i- Pr (29), - CF3 (30)). Steric effects seemed to play a bigger role than electronic effects, since the best tolerated substitutions were - Cl (27) and - Me (28). For position Y, derived from the aldehyde building blocks, the o- F analog (31) was tolerated, but was weaker compared to the unsubstituted (21); a plausible explanation being the rotational nature of the aryl ring- imidazopyridine ring bond, which lost axial symmetry with the presence of the o- F group. Introduction of a methylene linker (32) or replacement of the aryl ring with a cyclohexyl ring (33) that had a similar size made the compounds almost unable to bind in the MS assay. Modifications in the position of the electrophilic warhead were also not tolerated. The less- reactive ester analog (34) was inactive, whereas halogenated chloroacetamides (35- 37) did not lead to the expected mass adducts, showing instability in the MS and unclear chemical reactivity with 14- 3- 3. A plausible explanation was their increased reactivity, which correlated with reduced stability. The sulfamate warhead (38) was tolerated in the MS assay, but was significantly weaker than the chloroacetamide analog, in addition to being less atom efficient.  
+
+<|ref|>text<|/ref|><|det|>[[115, 814, 882, 905]]<|/det|>
+Crystal structures were solved for 28, 32, and 33 and were compared as overlays with analog 21 (Fig 3B- E). The - Me group on the imidazopyridine ring of 28 was expected to contribute to hydrophobic interactions with Val595 of ERα. However, it disrupted the previously observed binding mode and instead moved the imidazopyridine ring upwards, in the hydrophobic pocket of 14- 3- 3 formed by L218, I219 and L222. This upward movement affected the conformation of L218, which moved upwards but maintained contact with 28. The movement in the pocket also altered the interactions of the double- ortho- substituted aryl ring; while the o- Cl still interacted with I219 on the roof of the binding site, the o- Me
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 883, 300]]<|/det|>
+group moved further away from F119 in the bottom of the pocket, and the interaction was lost. The direct hydrogen bond between the amide and N42 and the water- mediated bond between the aniline nitrogen and the carboxylic acid of Val595 were maintained. Overall, the changes in binding mode were associated with slower binding, but a similar apparent Kd in the MS assay (Fig. S3). Analog 32, with an additional methylene group on the electrophilic warhead linker showed a slightly disrupted binding mode. While the biggest difference was the position of the aryl ring next to the longer linker, the imidazopyridine ring also moved further away from ERα; the latter correlated negatively with the binding observed in the MS assay. In contrast to the previous analogs, the o- Cl group on the aryl ring of 32 pointed out of the 14- 3- 3 pocket and due to increased distance, was unable to interact with F119. Thus, the presence of an additional methylene group on the linker was sufficient to make the compound unable to stabilize the complex. Analog 33 bearing a cyclohexyl ring instead of an aromatic ring maintained the interaction between the o- Cl group of the aryl ring and Ile219, but not the interaction of the o- Me group with F119. The orientation of the amide bond next to the warhead also differed; however, the hydrogen bond with N42 was still formed. The presence of the cyclohexyl ring, which had the possibility of adopting more conformations compared to the more rigid aryl ring, correlated with reduced binding to the 14- 3- 3/ERα complex in the MS assay.  
+
+<|ref|>image<|/ref|><|det|>[[120, 320, 884, 636]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[120, 639, 884, 688]]<|/det|>
+<center>Figure 3. SAR and crystal structures of analogs substituted in positions X and Y. A) MS bar graphs at 1 \(\mu \mathrm{M}\) . For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-E) Crystal structure of 14-3-3o/ERα with compound 28 (dark yellow sticks), and overlays of crystal structures for compounds 28 (dark yellow sticks), 32 (pink sticks) and 33 (emerald green sticks) with 21 (pale purple sticks). </center>  
+
+<|ref|>text<|/ref|><|det|>[[114, 725, 883, 877]]<|/det|>
+Investigation of modifications in positions X and Y provided valuable input in the groups that were tolerated; however, these single- point substitutions did not significantly improve the stabilization effect in the MS assay. Based on the crystallographic input, we synthesized four analogs combining the favorable substitutions from positions X and Y with the symmetric double o- Me analog 17, hypothesizing that a symmetric rotatable bond would correlate with improved potency (Fig 4A). This hypothesis was readily confirmed by evaluating the symmetric analogs 39- 42 in the MS assay. Analogs 39 and 40 included the - Me group on the imidazopyridine ring (X position) and the - F group on aryl ring (referred to as W position), respectively. Analogs 41 and 42 both had the F group in W position and a - Me or - Cl group in X position, respectively. All analogs showed comparable, fast, cooperative binding to 14- 3- 3/ERα in the MS assay, indicating that the symmetry of the 2,6- di- Me- aniline ring was more favorable compared to the asymmetric 2- Cl,6- Me. All four analogs showed low apo binding with analog 41 showing the lowest.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[95, 90, 925, 740]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[97, 740, 912, 825]]<|/det|>
+<center>Figure 4. SAR and crystal structures of 2,6-di-Me analogs substituted in positions X and W. Biophysical data (MS, TR-FRET, SPR) and cell data (NanoBRET). A) MS bar graphs at 1 μM. For each compound, time course experiments were performed with measurements at 1h, 8h, 16h and 24h. ERα data are shown with different colors for each compound, and apo data in grey. B-C) Overlays of crystal structures of 14-3-3α/ERα ternary complexes with compounds 40 (orange sticks), 41 (red sticks), and 42 (blue sticks). D) TR-FRET schematic and protein titration data for representative compounds at 100 μM compound or DMSO. E) SPR data for the binary 14-3-3α/ERα interaction and ternary interactions with 181, 17 and 41 (mean +/- SD, n=2). F-G) 14-3-3α-HaloTag/Nuc-ERα NanoBRET assay in HEK293T cells with compound titrations (1:2 dilution, starting at 40 μM). Data points excluded where compound dosage was toxic to the cells. MCR compounds compared to the previously described stabilizer 181 and 85, an inactive compound as the negative control. Bar graphs quantifying pEC50 values. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[114, 114, 883, 343]]<|/det|>
+Crystal structures were solved for the symmetric analogs 40, 41 and 42, highlighting the significance of two rotational bonds in the scaffold: the aniline- aryl ring bond, as previously discussed and the aryl ring- imidazopyridine ring bond (Fig 4B- C). The o- F substituent in W position, adjacent to the aryl ring- imidazopyridine rotational bond, adopted two different orientations; an inward orientation for analog 40 and an outward orientation for 41 and 42, which each had an additional substituent on the imidazopyridine ring. With the inward conformation of analog 40, the o- F substituent interacted with P167 of 14- 3- 3, whereas for 41 and 42 the o- F group was oriented to made multiple stabilizing interactions, being approximately 3 Å from N42 of 14- 3- 3, and the o- Me substituent and aniline nitrogen of the compound itself, thus contributing both to ligand- protein interactions and intramolecular interactions, restricting the axial bond rotation. In agreement with previous observations, the presence of a substituent in the X position (41 and 42) led to an upward movement of the imidazopyridine ring, orienting the substituent toward the hydrophobic residues I219 and L222. Overall, the presence of the additional substituent on the imidazopyridine ring in the last two analogs, even though in a distant position compared to the o- F- substituted ring, correlated favorably with the restriction of the rotational bond on the F- aryl ring. The different substituents of the imidazopyridine ring, Me (41) or Cl (42), did not affect the position of the stabilizer in the crystal structures; however, in the MS assay the Me group showed lower apo binding to 14- 3- 3, resulting in higher cooperativity of the ternary complex.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 355, 182, 368]]<|/det|>
+## TR-FRET  
+
+<|ref|>text<|/ref|><|det|>[[114, 369, 883, 519]]<|/det|>
+Up to this point, the SAR was developed using the intact MS assay and supported by crystallography. In our previous work \(^{20 - 22}\) , we relied on a fluorescence anisotropy assay (FA) to confirm cooperativity. The FA assay typically used (5- carboxylfluorescein) FAM- labeled phospho- peptides to quantify peptide binding to the compound/14- 3- 3 complex. For the MCR scaffold, however, this assay proved to be unsuitable, since the extensively conjugated ring system was intrinsically fluorescent in the same wavelengths as the FAM- labeled peptide (480 nm and 520 nm). To circumvent this issue, we turned our attention to far- red fluorescent dyes. The cy5- labeled ERα- peptide with excitation wavelength of 651 nm and emission wavelength of 670 nm, significantly affected the Kd of the 14- 3- 3/ERα complex. The reported Kd using the acetylated ERα peptide in an isothermal calorimetry (ITC) experiment or the FAM- labeled- ERα in an FA assay is in the range of 1- 2 μM. \(^{21}\) The cy5- ERα- peptide, however, resulted in a Kd of 6 nM, indicating significant dye binding (Fig S7).  
+
+<|ref|>text<|/ref|><|det|>[[114, 532, 883, 669]]<|/det|>
+As a suitable alternative to the FA assay, we developed a TR- FRET assay \(^{41}\) , using HIS- tagged- full- length 14- 3- 3σ, biotin- labeled- ERα peptide, an anti- HIS- tag monoclonal antibody conjugated with a Tb(III) criptate as the donor and streptavidin conjugated with the D2 dye as the acceptor. The observed Kd for this system was 30 nM. To ensure that the observed difference in Kd from ITC or the FAM- labeled- ERα in an FA assay was not related to the biotin- tag on the peptide, we performed a competition assay with the FAM- labeled peptide. The Kd of the biotin- peptide was 2.5 μM, in good agreement with the FAM- peptide's Kd (Fig S7). This result suggested that the lower Kd in the TR- FRET assay was due to avidity \(^{42}\) ; since 14- 3- 3 and the antibody were both dimers and streptavidin was tetravalent, different multivalent complexes could form, resulting in differences in the observed binding affinity between the labeled and the unlabeled proteins.  
+
+<|ref|>text<|/ref|><|det|>[[114, 683, 883, 805]]<|/det|>
+We performed assay optimization with 2D- titrations of 14- 3- 3, biotin- ERα and donor/acceptor ratios. We then tested the synthesized compounds using the optimized experimental conditions: 200 nM 14- 3- 3 (top concentration, 2- fold dilution), 50 nM biotin- ERα, 0.166 nM MAb anti- 6HIS Tb and 6.25 nM SA- D2. The compounds were tested at 100 μM and DMSO was used a negative control. 14- 3- 3 was titrated using an Echo acoustic dispenser, followed by the addition of the compounds and the peptide; after 1h incubation at room temperature, the donor and acceptor were added. In contrast to the MS assay, maximum signal was obtained after 2h rather than 16h - a plausible explanation again being avidity. The tight Kd in the TR- FRET assay resulted in a small assay window and a hook effect was observed at higher protein concentrations (ca. \(50 - 100\) nM 14- 3- 3).  
+
+<|ref|>text<|/ref|><|det|>[[114, 817, 883, 907]]<|/det|>
+Nevertheless, the TR- FRET assay was sensitive to the addition of molecular glues. In addition to quantifying fold stabilization for the 14- 3- 3/ERα complex \((AppKd_{(compound)} / AppKd_{(DMSO)})\) , we also quantified the fold increase in the TR- FRET signal (the ratio of observed Emax and Emin). To validate the TR- FRET assay, we included two positive controls: the natural product FC- A and our previously described stabilizer compound \(181^{21}\) . The two compounds gave comparable results. FC- A had an \(AppKd_{(compound)}\) of 11 nM, fold- stabilization of 3.09, and fold- increase of 5.52. Compound 181 had an \(AppKd_{(compound)}\) of 8.6 nM, fold- stabilization of 3.95 and fold- increase of 5.52 (Fig S8, table S3).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[114, 88, 883, 300]]<|/det|>
+In good agreement with the MS data, neutral binders 1 and 10 showed only a small signal shift compared to the DMSO control (Fig 4D, Fig S8). The 2- Me analog 13 showed 4.19- fold- stabilization and 2.36- fold increase, whereas 2,6- di- Me analog 17 showed 4.78- fold- stabilization and 2.58- fold increase, the same rank order as the MS assay. The more sterically hindered analog 18 (o- Me, o- Et) showed weaker stabilization (3.06- fold- stabilization and 2.14- fold increase). Analogs 19, 20, and 21, all bearing the o- Me group and additional o- OMe, o- F and o- Cl groups respectively, showed comparable stabilization (4.41 – 5.23- fold stabilization and 2.1 – 2.48- fold increase). The highest stabilization effect was observed for the final symmetric analogs 39 – 42. Analog 39 with a - Me group in X position had an AppKd(compound) of 7.8 nM, fold- stabilization of 4.35 and fold- increase of 2.61, whereas 40 with the - F group in W position had an AppKd(compound) of 8.4 nM, fold- stabilization of 4.04 and fold- increase of 3.12. Analog 41, which had both the - Me group in X position and the - F group in the W position had the lowest AppKd(compound) (5.2 nM) and the highest fold- stabilization (6.53) and fold- increase (3.71) of the series. Analog 42, which differed only in the X position (- Cl instead of - Me group) appeared weaker (AppKd(compound) 8.8 nM, fold- stabilization of 3.87 and fold- increase of 2.97). In summary, while the fold- changes were dampened by avidity, 14- 3- 3/ERα molecular glues showed the same rank- order in the TR- FRET assay as in the mass spectrometry assay used for initial SAR.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 313, 150, 326]]<|/det|>
+## SPR  
+
+<|ref|>text<|/ref|><|det|>[[114, 328, 883, 585]]<|/det|>
+Surface Plasmon Resonance (SPR) was then used to analyze the kinetic parameters of the ERα peptide binding to 14- 3- 3α in the presence of compounds 181, 17 and 41, and to compare the AppKd(compound) and kinetics to the binary ERα/14- 3- 3α interaction (Fig 4E, Fig S9). Here, 14- 3- 3α tagged with a Twinstrept- tag was captured on a SPR chip coated with Strep- Tactin XT, after which a 2- fold dilution series of acetylated ERα phospho- peptide was injected. For the binary interaction, the fast dissociation rate (koff) reached the limit of detection of the SPR instrument, due to a relative weak interaction with a Kd of 1.1 μM, which was in line with the ITC and FA experiments. The covalent bond between the chloroacetamide warhead of the compounds and C38 of 14- 3- 3α was formed after overnight incubation in the presence of ERα. Immobilization of this complex on the chip, followed by extensive washing to remove the bound ERα peptide, allowed us to determine the kinetics of ERα binding to the 14- 3- 3α/stabilizer complex. The previously described stabilizer 181 decreased the off- rate to 0.016 s- 1 and simultaneously increased the association rate (kOn) by a factor of 16, resulting in a low nanomolar affinity constant for the ERαpeptide/14- 3- 3 complex (3.9 nM; stabilization = 282- fold). The analogs of the newly designed scaffold, 17 and 41, both led to an 8- fold increase in association rate compared to the binary interaction. The binding of compound 41 induced a stronger decrease in dissociation rate of ERα compared to 17, which resulted in Kd values of 10.1 nM and 15.1 nM for 41 and 17, respectively. The higher stabilizing potency of 41 compared to 17 (stabilization = 110- fold for 41 and 71- fold for 17) were in rank- order agreement with the TR- FRET data. The decreased dissociation rates in the presence of the stabilizers, especially for 41 and 181, increased the residence time of ERα binding to 14- 3- 3 from approximately 3.4 s to 47.6 s and 62.5s, respectively.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 599, 196, 612]]<|/det|>
+## NanoBRET  
+
+<|ref|>text<|/ref|><|det|>[[114, 613, 883, 867]]<|/det|>
+To test the effects of the compounds on the full- length PPI, we used a NanoBRET assay we developed previously43. Compounds were tested using a C- terminal fusion14- 3- 3α- HaloTag and full length, N- terminal fusion NanoLuc- ERα (Fig 4F- G; Table S4). Briefly, NanoLuc- ERα and 14- 3- 3α- HaloTag plasmids were transfected in 1:10 ratio in hormone deprived HEK293T cells. After 48 hours post transfection, cells were seeded in assay plates with the experimental wells treated with 100 nM HaloTag NanoBRET 618 ligand (Promega) and no- ligand control wells treated with DMSO (v/v). Cells were treated for 24 hours with compounds in 1:2 dilution series starting at 40 μM. The BRET signal was read and normalized against DMSO treated samples. All active compounds resulted in an increase in BRET signal compared to the negative control, 85. The previously described compound 18121 stabilized the 14- 3- 3α/ERα complex in cells with an EC50 value of 5.2 μM and a 1.7- fold increase in the BRET signal. Of the compounds tested, the neutral binder 10 was less effective than 181 (EC50 value > 100 μM, 1.6- fold increase). Compound 13 showed an improved EC50 and fold- increase compared to 10 (12 μM and 1.8- fold). Compound 17 had the lowest EC50 value of 2.7 μM and showed 1.8- fold increase in BRET signal, a slight improvement compared to 181. Compounds 40 and 42 had the same 2- fold- increases in BRET signal, though exhibited slightly different EC50 values (7.2 and 4.6 μM, respectively). Compound 41 resulted in the highest increase in BRET signal, a 2.6- fold increase; however, it had a similar EC50 value to 181 of 5 μM. In the NanoBRET assay, compounds 17 and 41 performed most effectively, based on the EC50 value (17) and fold- increase in BRET signal (41) of the panel of compounds tested. The majority of MCR stabilizers showed similar EC50 values to the previous scaffold represented by 181, but consistently showed improved fold- increase in BRET signal.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 883, 195]]<|/det|>
+Compounds 181, 17, and 41, along with the negative control, 85, were tested in a NanoBRET assay where the cysteine of interest in 14- 3- 3α C38, was mutated to an asparagine, 14- 3- 3σ<sup>C38N</sup>- HaloTag (Fig S10). The BRET signal did not increase for 85, 181, or 17 with increasing compound concentration. There was a minimal increase in BRET signal for compound 41, from 2.5 to 3.5 mBU at 20 μM 41. In comparison to the assays done with 14- 3- 3σ<sup>WT</sup>- HaloTag, the non- normalized BRET signal increased from 6.7 to 16.1 mBU at the same concentration of 41 (data shown as normalized). In summary, when C38 was not present in 14- 3- 3σ, the compounds were unable to bind and stabilize the full- length 14- 3- 3σ/ERα complex in cells (Fig S10).  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 208, 204, 223]]<|/det|>
+## Discussion  
+
+<|ref|>text<|/ref|><|det|>[[115, 232, 883, 308]]<|/det|>
+Using MCR chemistry, we demonstrated the potential of a scaffold- hopping approach for molecular glues stabilizing a native 14- 3- 3/cient PPI. Our approach combined computational de novo design with multi- component reaction chemistry, which included short, efficient synthetic routes with multiple points of variation, accelerating the synthesis of analogs. Thus two distinct optimization approaches led to two highly diverse chemical series, converging on similar efficacies as 14- 3- 3σ/ERα molecular glues.  
+
+<|ref|>text<|/ref|><|det|>[[115, 319, 883, 471]]<|/det|>
+In the MCR approach, structure- activity relationships were established with an intact mass spectrometry assay, which allowed the distinction between neutral binders and stabilizers - cooperative molecular glues. Overall, the SAR showed that the introduction of even small substituents in the case of molecular glues could profoundly affect their potency and cooperativity. The best compounds, eg, 41, was already \(50\%\) bound at \(1\mu M\) after \(1\mathrm{hr}\) of incubation. Crystal structures of ternary complexes were crucial in elucidating small changes in the binding modes of otherwise highly similar analogs. Starting from the weak neutral binder 1 and by removing one methylene group, we significantly increased binding to the complex for compound 10. Introduction of substituents in the o- position was sufficient to reduce apo binding and turn the compounds from neutral binders to cooperative molecular glues. Although the number of rotational bonds is generally kept to a minimum, in this case we took advantage of two rotational bonds in the scaffold and with appropriate substituents achieved favorable ligand conformations resulting in increased potency, as shown for analogs 17 and 41.  
+
+<|ref|>text<|/ref|><|det|>[[115, 483, 883, 652]]<|/det|>
+Several biophysical assays provided complementary strategies for evaluating these novel molecular glues. A TR- FRET assay circumvented the issue of scaffold fluorescence, which could have been a limiting factor in establishing SAR and allowed the rank- ordering of analogs. A new SPR assay, in which the covalent ligand was pre- associated with 14- 3- 3σ before immobilization, provided an insightful analysis of binding/unbinding kinetics and started to hint at differences between the two series of molecular glues. Saturating concentrations of compound 41, for instance, stabilized the ERα/14- 3- 3 complex by 110- fold and slowing the koff by 14- fold. Taken together, biophysical assays allowed the quantification of cooperativity (MS, TR- FRET) and kinetics (SPR); importantly, the observed SAR was consistent among these three assays. The NanoBRET assay further showed good correlation with the biophysical protein/phosphopeptide assays while demonstrating stabilization of the full- length proteins with an EC50 value of \(2.7 - 5\mu M\) in intact cells. Overall, the optimized, cell- active MCR scaffold will facilitate chemical biology approaches to study the 14- 3- 3/ERα interaction, which has so far been unexploited for drug discovery.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 661, 184, 676]]<|/det|>
+## Methods  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 686, 437, 700]]<|/det|>
+## PROTEIN EXPRESSION AND PURIFICATION  
+
+<|ref|>text<|/ref|><|det|>[[115, 701, 883, 896]]<|/det|>
+The 14- 3- 3 α isoform (full- length for mass spectrometry and TR- FRET assays, \(\Delta C\) for crystallography) with an N- terminal His6 tag was expressed in Rosetta™ 2(DE3)PLysS competent E. coli (Novagen) from a pPROEX HTb expression vector. After transformation following manufacturer's instructions, single colonies were picked to inoculate \(30~\mathrm{mL}\) precultures (LB), which were added to \(1.5\mathrm{L}\) terrific broth (TB) medium after overnight growth at \(37^{\circ}C\) , 250 rpm. Expression was induced upon reaching \(\mathrm{OD}_{600}\) \(1.9 - 2.1\) by adding \(400~\mu \mathrm{M}\) IPTG. After overnight expression at \(30^{\circ}C\) , \(150~\mathrm{rpm}\) , cells were harvested by centrifugation at \(6,500~\mathrm{rpm}\) , resuspended in lysis buffer ( \(50~\mathrm{mM}\) HEPES pH 7.5, 500 mM NaCl, \(20~\mathrm{mM}\) imidazole, \(10\%\) glycerol, \(1\mathrm{mM}\) TCEP), and lysed by sonication. The His6- tagged protein was purified by Ni- affinity chromatography (Ni- NTA Agarose, Invitrogen) (Wash buffer \(50~\mathrm{mM}\) HEPES pH 7.5, \(500~\mathrm{mM}\) NaCl, \(20~\mathrm{mM}\) imidazole, \(1\mathrm{mM}\) TCEP; Elution buffer \(50~\mathrm{mM}\) HEPES pH 7.5, \(500~\mathrm{mM}\) NaCl, \(500~\mathrm{mM}\) imidazole, \(1\mathrm{mM}\) TCEP) and analyzed for purity by SDS- PAGE and Q- Tof LC/MS. The protein was buffer exchanged (Storage buffer \(25~\mathrm{mM}\) HEPES pH 7.5, \(150~\mathrm{mM}\) NaCl, \(1~\mathrm{mM}\) TCEP) and concentrated to \(\sim 16~\mathrm{mg / mL}\) and aliquots flash- frozen for storage at \(- 80^{\circ}C\) . The \(\Delta C\) variant was truncated at the C- terminus after T231 to enhance crystallization and after the first Ni- affinity chromatography column, the construct was treated with TEV protease to cleave off the His6 tag during dialysis ( \(25~\mathrm{mM}\)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 883, 164]]<|/det|>
+HEPES, pH 7.5, 200 mM NaCl, 5% glycerol, 10 mM MgCl2, 250 μM TCEP) overnight at \(4^{\circ}C\) . The flow- through of a second Ni- affinity column was subjected to a final purification step by size exclusion chromatography (Superdex 75 pg 16/60 size exclusion column (GE Life Science) (SEC buffer: 25 mM HEPES pH 7.5, 100 mM NaCl, 10 mM MgCl2, 250 μM TCEP). The protein was concentrated to \(\sim 60\) mg/mL, analyzed for purity by SDS- PAGE and Q- Tof LC/MS and aliquots flash- frozen for storage at \(- 80^{\circ}C\) .  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 180, 191, 193]]<|/det|>
+## PEPTIDES  
+
+<|ref|>text<|/ref|><|det|>[[115, 195, 883, 241]]<|/det|>
+Peptides for mass spectrometry, fluorescence anisotropy and TR- FRET assays were purchased from Elim Biopharmaceuticals, Inc. (Hayward, CA). Peptides for SPR and X- ray crystallography was purchased from GenScript Biotech Corp. The following peptides were used:  
+
+<|ref|>text<|/ref|><|det|>[[115, 241, 540, 315]]<|/det|>
+Ac- KYYITGEAEGFPA(pT)V- COOH (MS assay, 15- mer), 5- FAM- AEGFPA(pT)V- COOH (FA assay, 8mer ERα- pp), cy5- KYYITGEAEGFPA(pT)V- COOH (FA assay, 15- mer), biotin- KYYITGEAEGFPA(pT)V- COOH (TR- FRET assay, 15- mer), Ac- EGFPA(pT)V- COOH (crystallography and SPR, 7- mer)  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 330, 395, 344]]<|/det|>
+## INTACT MASS SPECTROMETRY ASSAY  
+
+<|ref|>text<|/ref|><|det|>[[115, 345, 883, 540]]<|/det|>
+Mass spectrometry dose response assays were performed on a Waters Acquity UPLC/ Xevo G2- XS Q- Tof mass spectrometer. A Waters UPLC Protein BEH- C4 Column (300 A, \(1.7 \mu \mathrm{m}\) , \(2.1 \mathrm{mm} \times 50 \mathrm{mm}\) ) was used to desalt the samples prior to application on the mass spectrometer. For 19- point MS dose responses, \(50 \mathrm{mM}\) compound stocks in DMSO were serially diluted in 3- fold increment in a master plate, then \(1000 \mathrm{nl}\) of the compounds were transferred in the assay plates. Master mixes containing \(100 \mathrm{nM}\) full- length wild type 14- 3- 3α in the absence or presence of \(2 \mu \mathrm{M}\) ERα were then dispensed into 384 well plates (Greiner Bio- One, catalog number 784201). Assay buffer was TRIS (10 mM, pH 8.0) and final volume per well was \(50 \mu \mathrm{l}\) , with final top concentration of compounds dose response series at 1 mM. The reaction mixtures were incubated for 1h at rt before subjected to MS. Four measurements (1h, 8h, 16h, 24h) were performed for time- course experiments. The injection volume for each sample was \(6 \mu \mathrm{l}\) . \(24 \mu \mathrm{l}\) of sample were needed for the time- course experiments, so the total volume in the assay plate was adjusted to \(50 \mu \mathrm{l}\) , to account for the dead volume in the injections. Data collection and automated processing followed a custom workflow, as previously described. \(^{44}\) z Plots were created using GraphPad Prism with the log(agonist) vs. response (variable slope, four parameters) fitting model.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 555, 648, 569]]<|/det|>
+## \(\mathbf{K}_{\mathrm{D}}\) DETERMINATION FOR FAM-, cy5- AND BIOTIN-LABELED ERα PEPTIDES  
+
+<|ref|>text<|/ref|><|det|>[[115, 570, 883, 720]]<|/det|>
+For \(\mathrm{K}_{\mathrm{D}}\) determination, N- terminal fluorescein- labeled ERα peptide (5- FAM) or cy5- labeled ERα peptide and HIS- tag FL 14- 3- 3α were diluted in buffer (10mM HEPES pH 7.5, 150mM NaCl, \(0.05\%\) tween 20, \(0.05\%\) BGG (bovine gamma globulin)). Two- fold dilution series of 14- 3- 3 were made in black, round- bottom 384- microwell plates (Greiner Bio- one 784900) in a final sample volume of \(10 \mu \mathrm{L}\) in triplicates. FAM- or cy5- labeled ERα peptides (final assay concentration 10nM) were dissolved in assay buffer and mixed with the protein dilution series on the plates. Fluorescence anisotropy measurements were performed after 1h incubation at room temperature on an Envision HTS Dual Detector 2105 plate reader (for FAM- labeled ERα peptide: filter set lex: 480, lem: 535, and D505fp/D538 advanced dual mirror). For cy5- labeled ERα peptide filter set lex: 620, lem: 688 nm, and D658fp/D688 advanced dual mirror). Data were reported at endpoint. Prism 10 (GraphPad) was used to generate plots using the [agonist] vs. response (variable slope, four parameters) fitting model to determine \(\mathrm{K}_{\mathrm{D}}\) values.  
+
+<|ref|>text<|/ref|><|det|>[[115, 720, 883, 870]]<|/det|>
+For \(\mathrm{K}_{\mathrm{D}}\) determination of the biotin- labeled ERα peptide a competition assay was performed. 5- FAM- and biotin- labeled ERα peptide were diluted in buffer (10mM HEPES pH 7.5, 150mM NaCl, \(0.05\%\) tween 20, \(0.05\%\) BGG (bovine gamma globulin)). Two- fold dilution series of biotin- labeled ERα peptide were made in black, round- bottom 384- microwell plates (Greiner Bio- one 784900) in a final sample volume of \(10 \mu \mathrm{L}\) in triplicates. A mastermix of 14- 3- 3α and 5- FAM- ERα was dispensed on the assay plate (final assay concentrations: \(6 \mu \mathrm{M}\) 14- 3- 3α (IC80) and \(10 \mathrm{nM}\) 5- FAM- ERα). Fluorescence anisotropy measurements were performed after 1h incubation at room temperature using a Molecular Devices ID5 plate reader (filter set lex: \(485 \pm 20 \mathrm{nm}\) , lem: \(535 \pm 25 \mathrm{nm}\) ; integration time: \(50 \mathrm{ms}\) ; settle time: \(0 \mathrm{ms}\) ; shake 5 sec, medium, read height \(3.00 \mathrm{mm}\) , G- factor \(= 1\) ). Data were reported at endpoint. Prism 10 (GraphPad) was used to generate plots using the [agonist] vs. response (variable slope, four parameters) fitting model to determine \(\mathrm{K}_{\mathrm{D}}\) values. The obtained \(\mathrm{K}_{\mathrm{D}}\) value was corrected using the Cheng- Prusoff equation.  
+
+<|ref|>text<|/ref|><|det|>[[115, 870, 375, 900]]<|/det|>
+\(\mathrm{K}_{\mathrm{d}} = \mathrm{IC}_{50} / (1 + [\mathrm{S}] / \mathrm{Km})\) \(\mathrm{K}_{\mathrm{d}} = 6.9 / (1 + 50 \mathrm{nM} / 30 \mathrm{nM}) = 2.5 \mu \mathrm{M}\)
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[116, 103, 340, 118]]<|/det|>
+## TR-FRET PROTEIN TITRATIONS  
+
+<|ref|>text<|/ref|><|det|>[[114, 118, 883, 418]]<|/det|>
+TR- FRET PROTEIN TITRATIONSFor assay optimization, 2D titrations of biotin- labeled ERα peptide, HIS- tag FL 14- 3- 3α and streptavidin- D2 were performed in assay buffer (10mM HEPES pH 7.5, 150mM NaCl, 0.05% tween 20, 0.05% BGG (bovine gamma globulin)). The donor (MAb anti- 6HIS Tb cryoplate gold) concentration was kept constant at 0.166 nM. For TR- FRET protein titrations, biotin- labeled ERα peptide (50 nM), the compounds or DMSO (100 μM), MAb anti- 6HIS Tb cryoplate gold (0.166 nM) and streptavidin- D2 (6.25 nM) were mixed in assay buffer (10mM HEPES pH 7.5, 150mM NaCl, 0.05% tween 20, 0.05% BGG (bovine gamma globulin)). 2- fold serial dilutions of HIS- tag FL 14- 3- 3α were performed (200 nM top assay concentration, 12- point dilution series). The assay was performed in 384- well microplates (Corning 4513, low volume white) at a volume of 10 μl per well. The following procedure was used: The compounds (50 mM stocks in DMSO) were transferred in echo LDV masterplates. 20 nL were transferred from the masterplate to the assay plate to achieve 100 μM compound concentration in the assay using Echo acoustic dispensing. The biotin- labeled ERα peptide was dissolved in assay buffer (10mM HEPES pH 7.5, 150mM NaCl, 0.05% tween 20, 0.05% BGG (bovine gamma globulin)) and dispensed in the assay plate using Dragonfly. 14- 3- 3 dilution series were prepared using Echo acoustic dispensing. Assay plates were incubated for 1hr at room temperature before the addition of a mastermix containing the donor (MAb anti- 6HIS Tb cryoplate gold) and acceptor (streptavidin- D2) in assay buffer. The mastermix was dispensed with Dragonfly. Assay plates were incubated for 2hr at room temperature prior to TR- FRET measurements using the Envision HTS Dual Detector 2105 plate reader equipped with the TR- FRET filter set (320/615/665 nm) and a D407/D630 advanced dual mirror. A 50 μs delay was employed to reduce background fluorescence. The TR- FRET signal was obtained through calculating the ratio of 665 nm to 615 nm fluorescence (x 1000), and Prism 10 (GraphPad) was used to generate plots using the [agonist] vs. response (variable slope, four parameters) fitting model. At least two independent experiments were performed.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 432, 148, 445]]<|/det|>
+## SPR  
+
+<|ref|>text<|/ref|><|det|>[[115, 446, 883, 688]]<|/det|>
+SPRThe SPR experiments were performed at \(25^{\circ}C\) using a Biacore X100 and a 200 nm Strep- Tactic XT derivatized linear polycarboxylate hydrogel chip, medium charge density (XanTec Bioanalytics). All proteins and peptides were dissolved in fresh running buffer prepared with ultrapure water and filtered through a \(0.2 \mu m\) filter (10 mM HEPES pH 7.4, 200 mM NaCl, 50 μM EDTA, 0.005% P20). First the surface was conditioned with a 1 min injection of 3 M Guanidine HCl. Then, the recombinant 14- 3- 3α- Twinstrept protein (250 nM) was captured on flow cell 2 of the sensor chip at a flow rate of \(10 \mu L / min\) for 2 minutes, which resulted in a capture level of 1000 RU. For the ternary interaction, 14- 3- 3α- Twinstrept protein (250 nM) was first incubated overnight with \(1 \mu M\) Ac- ERα peptide and \(20 \mu M\) compound prior to immobilization to the chip. The bound ERα peptide was washed away using running buffer flowed over the chip for 15 min. Flow cell 1 was left blank as a reference surface. After immobilization of the protein, the Biacore X100 was primed with running buffer. Multi- cycle kinetic measurements were conducted at a flow rate of \(30 \mu L / min\) . A 2- fold dilution series of analyte (Ac- ERα peptide) in running buffer were injected over the sensor chip for 2 min, followed by dissociation of 3 min (binary interaction), 7 or 13 minutes (ternary interaction). For the binary interaction, the highest concentration of ERα was 50 μM, and for the ternary interactions this was 250 nM. Between cycles of one multi- cycle measurement, no regeneration step was performed due to complete dissociation of the analyte. After a measurement, the chip was regenerated by 2 times 30 sec injections of 3 M Guanidine HCl. The data was corrected by double subtracting to the reference surface (flow cell 1) and buffer injection and analyzed using 1:1 interaction fitting model with the BIA evaluation software (2020).  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 701, 596, 716]]<|/det|>
+## X-RAY CRYSTALLOGRAPHY DATA COLLECTION AND REFINEMENT  
+
+<|ref|>text<|/ref|><|det|>[[115, 717, 883, 896]]<|/det|>
+X- RAY CRYSTALLOGRAPHY DATA COLLECTION AND REFINEMENTThe 14- 3- 3αΔC protein, acetylated ERα and compounds (50 mM stock in DMSO) were dissolved in complexation buffer (25 mM HEPES pH=7.5, 2 mM MgCl₂ and 100 μM TCEP) and mixed in a 1:2:3 or 1:2:5 molecular stoichiometry (protein : peptide : compound) with a final protein concentration of 12 mg/mL. The complex was set- up for sitting- drop crystallization after overnight incubation at 4 °C, in a custom crystallization liquor (0.05 M HEPES (pH 7.1, 7.3, 7.5, 7.7), 0.19 M CaCl₂, 24- 29% PEG400, and 5% (v/v) glycerol). Crystals grew within 10- 14 days at 4 °C. Crystals were fished and flash- cooled in liquid nitrogen. X- ray diffraction (XRD) data were collected at the European Synchrotron Radiation Facility (ESRF Grenoble, France, beamline ID23- 1, ID30A- 3/MASSIF- 3, or ID23- 2) or at the Deutsches Elektronen- Synchrotron (DESY Hamburg, Germany, beamline PETRA III). Data was processed using CCP4i2 suite (version 8.0.019). After indexing and integrating the data, scaling was done using AIMLESS. The data was phased with MolRep, using PDB 4JC3 as template. The presence of co- crystallized ligands was verified by visual inspection of the Fo- Fc and 2Fo- Fc electron density maps in COOT (version 0.9.8.93). If electron density corresponding to the co- crystallized ligand was present, its structure, restraints, and covalent bond were generated using AceDRG. After
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 883, 149]]<|/det|>
+building in the ligand, model rebuilding and refinement was performed using REFMAC5. The PDB REDO server (pdb- redo.edu) was used to complete the model building and refinement. The images were created using the PyMOL Molecular Graphics System (Schrödinger LLC, version 4.6.0). See SI table S10 for data collection and refinement statistics.  
+
+<|ref|>text<|/ref|><|det|>[[115, 163, 881, 194]]<|/det|>
+The structures were deposited in the protein data bank (PDB) with IDs: 916S (28), 916T (32), 916U (33), 916V (40), 916W (41), 916X (42), 916Y (1), 916Z (2), 9170 (17), 9171 (19), 9172 (10), 9173 (20), 9174 (21), and 9175 (25).  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 210, 195, 222]]<|/det|>
+## NanoBRET  
+
+<|ref|>text<|/ref|><|det|>[[115, 223, 883, 359]]<|/det|>
+NanoBRET assays were performed as previously described.43 HEK293T cells were cultured DMEM, high glucose (Gibco) supplemented with \(10\%\) charcoal stripped Fetal Bovine Serum (FBS; Gibco) and \(1\%\) penicillin/streptomycin. Cells were transfected with a 1:10 ratio of Nanoluc- ERa:14- 3- 3o- HaloTag plasmid for 48 hours using jetOPTIMUS transfection reagent (Polyplus). Cells were then seeded at 8,000 cells per well in a 384- well plate (Corning #3570) in FluoroBrite DMEM (phenol red- free; Gibco) with \(4\%\) charcoal stripped FBS and treated with \(100~\mathrm{mM}\) HaloTag NanoBRET 618 Ligand (Promega) or equivalent volume of DMSO as a no ligand negative control. Following plating, cells were treated for 24 hours with compound in 1:2 dilution series starting at \(40~\mu \mathrm{M}\) (0.35% DMSO final concentration). After 24 hours, the BRET signal was read using an EnVision XCite 2105 plate reader at 618 nm (HaloTag) and 460 nm (NIuc). The final corrected NanoBRET ratio was calculated using the following equation:  
+
+<|ref|>equation<|/ref|><|det|>[[252, 371, 744, 404]]<|/det|>
+\[CorrectedBRETratio = \left(\frac{618nm}{460nm}\right)_{HaloTagLigand} - \left(\frac{618nm}{460nm}\right)_{NoLigandcontrol}\]  
+
+<|ref|>text<|/ref|><|det|>[[115, 420, 545, 435]]<|/det|>
+The BRET ratios were normalized to samples treated with DMSO.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 451, 187, 464]]<|/det|>
+## DOCKING  
+
+<|ref|>text<|/ref|><|det|>[[115, 464, 881, 493]]<|/det|>
+Computational design for SAR optimization and docking was performed with SeeSAR version 14.0.0; BioSolveIT GmbH, Sankt Augustin, Germany, 2022, www.biosolveit.de/SeeSAR  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 504, 280, 517]]<|/det|>
+## SOFTWARE VERSIONS  
+
+<|ref|>text<|/ref|><|det|>[[115, 517, 881, 545]]<|/det|>
+Prism (10.2.1), Illustrator (22.1 (64- bit)), Biorender (64- bit), Pymol (4.6.0), CCP4i2 (8.0.003), COOT (0.9.8.1), Phenix (1.19.2- 4158)  
+
+<|ref|>text<|/ref|><|det|>[[115, 558, 881, 588]]<|/det|>
+Supporting Information. Supplementary figures and tables, synthetic procedures, compound characterization, NMR spectra, crystallography data (PDF).  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 603, 197, 616]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[112, 616, 880, 905]]<|/det|>
+1. Andrei, S. A. et al. Stabilization of protein-protein interactions in drug discovery. Expert Opin Drug Discov 12, 925-940 (2017).  
+2. Bier, D., Thiel, P., Briels, J. & Ottmann, C. Stabilization of Protein-Protein Interactions in chemical biology and drug discovery. Progress in Biophysics and Molecular Biology 119, 10-19 (2015).  
+3. Arkin, M. R. & Wells, J. A. Small-molecule inhibitors of protein-protein interactions: progressing towards the dream. Nat Rev Drug Discov 3, 301-317 (2004).  
+4. Arkin, M. R., Tang, Y. & Wells, J. A. Small-Molecule Inhibitors of Protein-Protein Interactions: Progressing toward the Reality. Chemistry & Biology 21, 1102-1114 (2014).  
+5. Smith, M. C. & Gestwicki, J. E. Features of protein-protein interactions that translate into potent inhibitors: topology, surface area and affinity. Expert Rev. Mol. Med. 14, e16 (2012).  
+6. Schreiber, S. L. Molecular glues and bifunctional compounds: Therapeutic modalities based on induced proximity. Cell Chemical Biology 31, 1050-1063 (2024).  
+7. Konstantinidou, M. & Arkin, M. R. Molecular glues for protein-protein interactions: Progressing toward a new dream. Cell Chem Biol 31, 1064-1088 (2024).  
+8. Ruan, H., Sun, Q., Zhang, W., Liu, Y. & Lai, L. Targeting intrinsically disordered proteins at the edge of chaos. Drug Discovery Today 24, 217-227 (2019).  
+9. Santofimia-Castaño, P. et al. Targeting intrinsically disordered proteins involved in cancer. Cell. Mol. Life Sci. 77, 1695-1707 (2020).  
+10. Aitken, A. 14-3-3 proteins: a historic overview. Semin Cancer Biol 16, 162-172 (2006).  
+11. Pennington, K. L., Chan, T. Y., Torres, M. P. & Andersen, J. L. The dynamic and stress-adaptive signaling hub of 14-3-3: emerging mechanisms of regulation and context-dependent protein-protein interactions. Oncogene 37, 5587-5604 (2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[110, 90, 884, 888]]<|/det|>
+12. De Vries-van Leeuwen, I. J. et al. Interaction of 14-3-3 proteins with the estrogen receptor alpha F domain provides a drug target interface. Proc Natl Acad Sci U S A 110, 8894-8899 (2013).13. Somsen, B. A. et al. Molecular basis and dual ligand regulation of tetrameric estrogen receptor \(\alpha /14-3-3\zeta\) protein complex. Journal of Biological Chemistry 299, 104855 (2023).14. Liu, Y., Ma, H. & Yao, J. ERα, A Key Target for Cancer Therapy: A Review. OTT Volume 13, 2183-2191 (2020).15. Paterni, I., Bertini, S., Granchi, C., Macchia, M. & Minutolo, F. Estrogen receptor ligands: a patent review update. Expert Opinion on Therapeutic Patents 23, 1247-1271 (2013).16. Scott, J. S. & Barlaam, B. Selective estrogen receptor degraders (SERDs) and covalent antagonists (SERCAs): a patent review (2015-present). Expert Opinion on Therapeutic Patents 32, 131-151 (2022).17. Scott, J. S. & Klinowska, T. C. M. Selective estrogen receptor degraders (SERDs) and covalent antagonists (SERCAs): a patent review (July 2021-December 2023). Expert Opinion on Therapeutic Patents 34, 333-350 (2024).18. Jeselsohn, R., Buchwalter, G., De Angelis, C., Brown, M. & Schiff, R. ESR1 mutations—a mechanism for acquired endocrine resistance in breast cancer. Nat Rev Clin Oncol 12, 573-583 (2015).19. Visser, E. J. et al. Estrogen Receptor \(\alpha /14-3-3\) molecular glues as alternative treatment strategy for endocrine resistant breast cancer. Preprint at https://doi.org/10.1101/2024.04.25.591105 (2024).20. Kenanova, D. N. et al. A Systematic Approach to the Discovery of Protein-Protein Interaction Stabilizers. ACS Cent. Sci. 9, 937-946 (2023).21. Konstantinidou, M. et al. Structure-Based Optimization of Covalent, Small-Molecule Stabilizers of the 14-3-3/ERα Protein-Protein Interaction from Nonselective Fragments. J. Am. Chem. Soc. 145, 20328-20343 (2023).22. Visser, E. J. et al. From Tethered to Freestanding Stabilizers of 14-3-3 Protein-Protein Interactions through Fragment Linking. Angew Chem Int Ed 62, e202308004 (2023).23. Dömling, A., Wang, W. & Wang, K. Chemistry and Biology Of Multicomponent Reactions. Chem. Rev. 112, 3083-3135 (2012).24. Zarganes-Tzitzikas, T., Chandgude, A. L. & Dömling, A. Multicomponent Reactions, Union of MCRs and Beyond. The Chemical Record 15, 981-996 (2015).25. Fotopoulou, E., Anastasiou, P. K., Tomza, C. & Neochoritis, C. G. The Ugi reaction as the green alternative towards active pharmaceutical ingredients. Tetrahedron Green Chem 3, 100044 (2024).26. Li, X., Zarganes-Tzitzikas, T., Kurpiewska, K. & Dömling, A. Amenamevir by Ugi-4CR. Green Chem. 25, 1322-1325 (2023).27. Zarganes-Tzitzikas, T., Neochoritis, C. G. & Dömling, A. Atorvastatin (Lipitor) by MCR. ACS Med. Chem. Lett. 10, 389-392 (2019).28. Znabet, A. et al. A highly efficient synthesis of telaprevir by strategic use of biocatalysis and multicomponent reactions. Chem. Commun. 46, 7918 (2010).29. Wehlan, H., Oehme, J., Schäfer, A. & Rossen, K. Development of Scalable Conditions for the Ugi Reaction—Application to the Synthesis of (R)-Lacosamide. Org. Process Res. Dev. 19, 1980-1986 (2015).30. Váradi, A. et al. Synthesis of Carfentanil Amide Opioids Using the Ugi Multicomponent Reaction. ACS Chem. Neurosci. 6, 1570-1577 (2015).31. Koes, D. et al. Enabling Large-Scale Design, Synthesis and Validation of Small Molecule Protein-Protein Antagonists. PLoS ONE 7, e32839 (2012).32. Koes, D. R., Dömling, A. & Camacho, C. J. AnchorQuery: Rapid online virtual screening for small-molecule protein-protein interaction inhibitors. Protein Sci 27, 229-232 (2018).33. Neochoritis, C. G. et al. Hitting on the move: Targeting intrinsically disordered protein states of the MDM2-p53 interaction. European Journal of Medicinal Chemistry 182, 111588 (2019).34. Groebke, K., Weber, L. & Mehlin, F. Synthesis of Imidazo[1,2-a] annulated Pyridines, Pyrazines and Pyrimidines by a Novel Three-Component Condensation. Synlett 1998, 661-663 (1998).35. Blackburn, C., Guan, B., Fleming, P., Shiosaki, K. & Tsai, S. Parallel synthesis of 3-aminoimidazo[1,2-a]pyridines and pyrazines by a new three-component condensation. Tetrahedron Letters 39, 3635-3638 (1998).36. Bienaymé, H. & Bouzid, K. A New Heterocyclic Multicomponent Reaction For the Combinatorial Synthesis of Fused 3-Aminoimidazoles. Angewandte Chemie International Edition 37, 2234-2237 (1998).37. Boltjes, A. & Dömling, A. The Groebke-Blackburn-Bienaymé Reaction. Eur J Org Chem 2019, 7007-7049 (2019).38. Shukla, P., Azad, C. S., Deswal, D. & Narula, A. K. Revisiting the GBB reaction and redefining its relevance in medicinal chemistry: A review. Drug Discovery Today 29, 104237 (2024).39. Devi, N., Singh, D., K. Rawal, R., Barwal, J. & Singh, V. Medicinal Attributes of Imidazo[1,2-a]pyridine Derivatives: An Update. CTMC 16, 2963-2994 (2016).40. Vasilikogiannaki, E., Gryparis, C., Kotzabasaki, V., Lykakis, I. N. & Stratakis, M. Facile Reduction of Nitroarenes into Anilines and Nitroalkanes into Hydroxylamines via the Rapid Activation of Ammonia- Borane Complex by Supported Gold Nanoparticles. Adv Synth Catal 355, 907-911 (2013).41. Degorce, F. HTRF: A Technology Tailored for Drug Discovery - A Review of Theoretical Aspects and Recent Applications. TOCHGENJ 3, 22-32 (2009).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 872, 170]]<|/det|>
+42. Vauquelin, G. & Charlton, S. J. Exploring avidity: understanding the potential gains in functional affinity and target residence time of bivalent and heterobivalent ligands. British J Pharmacology 168, 1771-1785 (2013). 
+43. Vickery, H. R., Virta, J. M., Konstantinidou, M. & Arkin, M. R. Development of a NanoBRET assay for evaluation of 14-3-3α molecular glues. SLAS Discov 29, 100165 (2024). 
+44. Hallenbeck, K. K. et al. A Liquid Chromatography/Mass Spectrometry Method for Screening Disulfide Tethering Fragments. SLAS Discov 2472555217732072 (2017) doi:10.1177/2472555217732072.  
+
+<|ref|>text<|/ref|><|det|>[[115, 194, 882, 325]]<|/det|>
+Acknowledgements. This research was funded by the Ono Pharma Foundation Breakthrough Science Initiative Award, NIH/NIGMS GM147696 and the Netherlands Organization for Scientific Research (NWO) through Gravity program 024.001.035 and ENW M-grant OCENW.M20.200. We acknowledge Foundation for Research and Innovation (H.F.R.I.) under the "2nd Call for H.F.R.I. Research Projects to support Post-Doctoral Researchers" (Project Number: 0911) and Emperikion Idryma (to C.G.N). We thank Amanda Paulson for the automated mass spec data processing infrastructure in the SMDC. We acknowledge the European Synchrotron Radiation Facility (ESRF) for provision of synchrotron radiation facilities, and we would like to thank David Flot and Max Nanao for assistance and support in using beamlines ID23-1, ID23-2, ID30A-3 (mx2407 and mx2526). We thank DESY (Hamburg, Germany), a member of the Helmholtz Association HGF, for the provision of experimental facilities. Parts of this research were carried out at PETRA III.  
+
+<|ref|>sub_title<|/ref|><|det|>[[116, 336, 266, 349]]<|/det|>
+## Author contributions.  
+
+<|ref|>text<|/ref|><|det|>[[115, 349, 882, 427]]<|/det|>
+M.K. conceived the work, designed the compounds, performed the MS and TR-FRET experiments, and analyzed the data with contributions from C.O, L.C., C.G.N. and M.R.A. M.Z and M.F synthesized and characterized compounds. M.A.M.P solved most of the crystal structures and performed the SPR experiments. J.M.V. performed the NanoBRET assay. J.L.R. was involved in the development and optimization of the TR-FRET assay. E.M.J. solved the initial crystal structures. M.K., C.G.N., M.R.A, C.O. and L.B. supervised the project. M.K. wrote the manuscript with contributions from all authors.  
+
+<|ref|>text<|/ref|><|det|>[[115, 440, 881, 470]]<|/det|>
+Conflict of interest. Michelle R. Arkin, Christian Ottmann and Luc Brunsveld are co- founders of Ambagon Therapeutics.  
+
+<|ref|>text<|/ref|><|det|>[[116, 483, 592, 498]]<|/det|>
+Keywords: covalent • estrogen receptor • MCR • molecular glue • 14- 3- 3  
+
+<|ref|>text<|/ref|><|det|>[[116, 525, 142, 537]]<|/det|>
+TOC  
+
+<|ref|>image<|/ref|><|det|>[[118, 556, 737, 766]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[128, 774, 728, 844]]<|/det|>
+We describe a scaffold- hopping approach suitable for the identification of molecular glues stabilizing the 14- 3- 3α/ERα complex. The multi- component reaction- based scaffold was rapidly optimized, and validated in biophysical assays, while several co- crystal structures elucidated the binding modes. The best compounds were tested in a cellular NanoBRET assay, showing low micromolar potency.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 93, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 355, 150]]<|/det|>
+20250210MCRSupplement.pdf
+
+<--- Page Split --->

preprint/preprint__09d33aed00231d67d90a8ec99efc3d40d229c84279128ce1b4364a1aa9264ebf/images_list.json

@@ -0,0 +1,152 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_1.jpg",
+    "caption": "Figure 1. Embryonic stem cells and cancer cells share compositionally similar epipheroperomes. a Schematic illustrating the biochemical and functional distinctions between epipheroperomes, defined as long-lasting heterooligomeric assemblies composed of tightly associated chaperones and co-chaperones, and traditional chaperones. Unlike chaperones, which assist in protein folding or assembly, epipheroperomes sequester proteins, reshaping protein-protein interactions, and consequently altering cellular phenotypes. The schematic also outlines key principles for the use of PU-probes in epipheroperome analysis. b Detection of epipheroperome components (chaperones and co-chaperones) through SDS-PAGE (bottom, total protein levels) and native-PAGE (top), followed by immunoblotting. See also Supplementary Fig. 1. c Visualization of HSP90 in epipheroperomes using the PU-TCO click probe. See also Supplementary Fig. 2. Gel images are representative of three independent experiments. d Epipheroperome constituent chaperones and co-chaperones identified through mass spectrometry analyses of PU-beads cargo. Representative data of two independent experiments. See Supplementary Fig. 3 for the GA-cargo. e Illustration of an isobaric, discriminant peptide pair from ESC lysate samples and HSP90 captured by PU- and GA-beads. Representative data of two independent experiments. f Schematic summary. Both cancer cells and pluripotent stem cells harbor epipheroperomes. These epipheroperomes undergo disassembly during differentiation processes. Source data are provided in Supplementary Data 1 and in Source data file.",
+    "footnote": [],
+    "bbox": [
+      [
+        75,
+        72,
+        936,
+        666
+      ]
+    ],
+    "page_idx": 34
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2",
+    "footnote": [],
+    "bbox": [
+      [
+        95,
+        35,
+        880,
+        768
+      ]
+    ],
+    "page_idx": 35
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3. Phosphorylation of key residues located in the charged linker supports HSP90 incorporation into epicarpemores. a Experiment outline and expected outcomes. b Tandem MS spectra of HSP90 Ser226 (bottom) and Ser255 (top) phosphorylated peptides are presented, supporting the sequence and phosphorylation site identification. c Comparison of the extracted ion chromatogram of HSP90 Ser255 phosphopeptide in the PU-bead cargo (red trace, left panel) and ESC lysate (black trace, left panel) with a representative unmodified tryptic peptide in the PU-bead cargo (blue trace, right panel) and ESC lysate (black trace, right panel). d Ion intensity values of all phosphopeptides and the ratio of mean peptide intensity for each phosphosite in the samples described in panel a (n = 4 Ca and n = 2, NT). e Ratio of individual peptide intensity for each phosphosite in the samples described in the schematic (S255 n = 5; S226 n = 4; S263 n = 8; S231 n = 5). Source data are provided as Source Data file and as Supplementary Data 3,6.",
+    "footnote": [],
+    "bbox": [
+      [
+        70,
+        40,
+        888,
+        817
+      ]
+    ],
+    "page_idx": 36
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Figure 4. Phosphorylation of key residues located in the charged linker of HSP90 leads to a conformational shift in the linker, exposing the middle domain of the protein. A Model of the HSP90-HSP90-HSP70-HSP70-HOP assembly used for the molecular dynamics simulations. A and B, protomers A and B, respectively. b Protein secondary structure elements (SSE) like alpha-helices and beta-strands of the charged linker of protomer A of ATP-bound HSP90 monitored throughout the MD simulation. WT (HSP90 S226/S255), phosphomimetic (HSP90 S226E/S255E) and non-phosphorylatable (HSP90 S226A/S255A) mutants were analyzed. The plot on the left reports SSE distribution by residue index throughout the charged linker and the plot on the right monitors each residue and its SSE assignment over time. Schematic illustrating the primary structure of the full-length HSP90 with color-coded domains is also shown: NTD, N-terminal domain; MD, middle domain and CTD, C-terminal domain. The charged linker (CL) and the location of the two key serine residues are also shown (top inset). The gray bar indicates the CL segment encompassing residues 218 to 232. c Cartoon representation of ATP-bound HSP90 protomer A in assemblies containing the phosphomimetic (HSP90 S226E/S255E) or the non-phosphorylatable (HSP90 S226A/S255A) mutants is shown. Green, reference trajectory; gray, representative trajectories of \\(n = 1,000\\) . The inset illustrates the surfaces available for the interaction between HSP90 A and HSP70 A when the CL is in the 'up' conformation. A blue arrow indicates the location of the key beta-strand in the charged linker. See also Supplementary Figs. 5 and 6.",
+    "footnote": [],
+    "bbox": [
+      [
+        50,
+        60,
+        911,
+        666
+      ]
+    ],
+    "page_idx": 37
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5. Phosphorylation of key residues located in the charged linker of HSP90 facilitates assembly motions conducive to epichepaterome core formation. A Calculated dynamic cross-correlation matrix of Ca atoms around their mean positions for 100 ns molecular dynamics simulations. ATP-bound WT (HSP90 S226/ S255), phosphoimimetic (HSP90 S226E/S255E) and non-phosphorylatable (HSP90 S226A/S255A) mutantcontaining HSP90-HSP90-HSP70-HSP70-HOP assemblies were analyzed. The cartoon below captures the key motions among the different domains of the individual assembly components. Extents of correlated motions and anti-correlated motions are color-coded from blue to red, which represent positive and negative correlations, respectively. The assembly contains two full-length HSP90beta proteins (protomer A and protomer B). The two HSP70 proteins (HSP70 A and HSP70 B) and the HOP protein are of sizes reported, and as per the constructs used in 7KW7. b Cartoon showing assemblies that are preferentially formed when the HSP90 charged linker is either phosphorylated (as in the EE mutant) or not phosphorylated (as in the WT protein).",
+    "footnote": [],
+    "bbox": [
+      [
+        52,
+        54,
+        911,
+        737
+      ]
+    ],
+    "page_idx": 38
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6",
+    "footnote": [],
+    "bbox": [
+      [
+        110,
+        264,
+        911,
+        560
+      ]
+    ],
+    "page_idx": 39
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_7.jpg",
+    "caption": "Figure 7. Phosphorylation of key residues located in the charged linker supports HSP90 incorporation into epicheporemes. a Overview of the experimental design and expected outcomes. b Analysis of transfection efficacy in cells transfected with HSP90β mutants, as indicated in panel a. c Detection of epicheporeme components (chaperones and co-chaperones) through SDS-PAGE (bottom, total protein levels) and native-PAGE (top), followed by immunoblotting. Blue brackets indicate the approximate position of epicheporeme-incorporated chaperones. Data are presented as mean ± s.e.m., \\(n = 3\\) , one-way ANOVA with Sidak's post-hoc, EE vs AA. d Visualization of HSP90 in epicheporemes using the PU-TCO probe clicked to Cy5 (left) and the mCherry tag (middle). Right, merged images. MWM, molecular weight marker. e Detection and quantification of epicheporeme components through PU-beads capture as indicated in panel a. Protein amount loaded for 'Input' represents 2% of the protein amount incubated with the beads. Data are presented as mean ± s.e.m., \\(n = 3\\) , unpaired two-tailed t-test. Gel images are representative of three independent experiments. Source data are provided as Source data file.",
+    "footnote": [],
+    "bbox": [
+      [
+        52,
+        75,
+        888,
+        666
+      ]
+    ],
+    "page_idx": 40
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_8.jpg",
+    "caption": "Figure 8. Phosphorylation of key residues located in the HSP90 charged linker favors ESC proliferation and self-renewal potential. a ESC proliferation at 60 h post-transfection in E14 cells transfected with either the phosphomimetic HSP90βS226E,S255E (EE) or the nonphosphorylatable HSP90S226A,S255A (AA) mutant. Medium (1x) or high (2x) plasmid concentrations were employed. Data are presented as mean ± s.e.m., \\(n = 6\\) , one-way ANOVA with Sidak's post-hoc, EE vs AA. b Representative spectra ( \\(n = 3\\) independent experiments) of phosphopeptides, S255P (left) and S226P (right), and a representative unmodified tryptic peptide (middle) in mCherry-tagged WT HSP90β affinity-purified from ESC (red) or differentiated trophoblast (black) cells. c Representative spectra ( \\(n = 3\\) independent experiments) of a tryptic peptide from Oct4 protein co-purified from ESCs labeled with heavy or light isotope lysine and arginine expressing either the phosphomimetic (EE) or the nonphosphorylatable (AA) HSP90 mutant. Quantitative analysis via mass spectrometry (MS) to determine protein abundance is shown. d Overview of the experimental design and expected outcomes (panels e,f). e,f Detection and quantification of Oct4 protein expressed in cells transfected with the indicated HSP90 mutants or vector control (panel e) and sequestered into the epicatherome platforms (identified through PU-beads capture, panel f). (e) Data are presented as mean ± s.e.m., \\(n = 5\\) AA, \\(n = 5\\) EE, \\(n = 3\\) WT, \\(n = 3\\) empty vector, one-way ANOVA with Dunnett's post-hoc, EE vs AA, WT vs AA, empty vector vs AA. (f) Data are presented as mean ± s.e.m., \\(n = 3\\) , unpaired two-tailed t-test. Source data are provided as Source Data file and Supplementary Data 5.",
+    "footnote": [],
+    "bbox": [
+      [
+        91,
+        102,
+        875,
+        666
+      ]
+    ],
+    "page_idx": 41
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_9.jpg",
+    "caption": "Figure 9",
+    "footnote": [],
+    "bbox": [
+      [
+        90,
+        100,
+        800,
+        549
+      ]
+    ],
+    "page_idx": 42
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_10.jpg",
+    "caption": "Figure 10. Human tissues positive for epicarpemores exhibit p-Ser226 HSP90β positivity, and conversely, those negative for epicarpemores show no or negligible p-Ser226 signal within HSP90's charged linker. a Cartoon illustrating the processing of human tissue for biochemical analyses. Both tumor (T) and tumor adjacent (TA) tissues, determined by gross pathological evaluation to be potentially non-cancerous, were harvested and analyzed. b MDA-MB-468 breast cancer cells (epicarpemore-high) and ASPC1 pancreatic cancer cells (epicarpemore-low) served as controls for assessing p-Ser226 HSP90 levels. c The graph presents the relationship between epicarpemore positivity and HSP90 Ser226 phosphorylation for tissues described in panel a. Data represent mean ± s.e.m., with \\(n = 9\\) tumor (T) and \\(n = 9\\) paired tumor-adjacent (TA) tissues classified based on epicarpemore positivity or negativity, as determined by Native PAGE (see panel d); unpaired two-tailed t-test. d Detection of epicarpemores through native-PAGE (top), and of p-Ser226 HSP90 (middle) and total HSP90 (bottom) by SDS-PAGE, followed by immunoblotting, in tissues from the indicated patient specimens, as in panel a. Blue brackets indicate the approximate position of epicarpemore-incorporated HSP90. Note: Obtaining genuinely \"normal\" tissue adjacent to tumors presents challenges, especially in the case of pancreatic tissue. The relatively small size of the organ and the nature of surgical procedures for pancreatic cancer often lead to the collection of normal samples in close proximity to the tumor. It's crucial to acknowledge that, due to these challenges, we designate potentially normal tissue as tumor-adjacent tissue, recognizing that it may not entirely reflect a truly normal tissue state. PDAC, Pancreatic Ductal Adenocarcinoma; IDC, Invasive Ductal Carcinoma; ILC, Invasive Lobular Carcinoma; ER, Estrogen Receptor; PR, Progesterone Receptor. Source data are provided as Source data file.",
+    "footnote": [],
+    "bbox": [
+      [
+        137,
+        55,
+        840,
+        700
+      ]
+    ],
+    "page_idx": 43
+  }
+]

preprint/preprint__09e9f783051f85aba41aac3aa9df36479c9f31b14fd396fee288daa064471a28/images_list.json

@@ -0,0 +1,55 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "a",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 34
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "b",
+    "footnote": [],
+    "bbox": [
+      [
+        66,
+        60,
+        999,
+        245
+      ]
+    ],
+    "page_idx": 36
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "d",
+    "footnote": [],
+    "bbox": [
+      [
+        68,
+        315,
+        433,
+        472
+      ]
+    ],
+    "page_idx": 38
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "c",
+    "footnote": [],
+    "bbox": [
+      [
+        540,
+        293,
+        790,
+        455
+      ]
+    ],
+    "page_idx": 40
+  }
+]

preprint/preprint__09f1a55d00104ac396ababe3986f197810df2a7652eb1fe992b357090dd1b42c/images_list.json

@@ -0,0 +1,115 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1: Overview of the Proposed TrafficSafe Framework. (a) The U.S. faces one of the highest crash risks among developed countries, with a rising trend. However, analyzing and addressing this issue is challenging due to the heterogeneous factors involved in crash events, including traffic conditions, human behavior, environmental impacts, and driver characteristics. To tackle this, we propose TrafficSafe, a framework designed for two key tasks: 1) Predicting crash outcomes and 2) Attributing crash factors with conditional risk analysis. By addressing questions such as why crashes occur and how to mitigate crash risks, TrafficSafe seeks to deliver optimal policy for safety improvement, aligning with the Vision Zero goal. (b) The TrafficSafe workflow incorporates multi-modal data, including driver behavior, vehicle details, infrastructure, and environmental conditions, represented through textual reports, satellite imagery, and other formats. Leveraging an AI-expert cooperative method, the crash data is transformed into textual prompts, resulting in the TrafficSafe Event dataset comprising 58,903 prompts. TrafficSafe LLM is created with accurate and trustworthy forecasting abilities for further analysis. Building on this pipeline, TrafficSafe Attribution operates across three dimensions: 1) Event-level risk analysis to identify feature contributions, 2) Conditional risk analysis to assess state-level risks under varying conditions, and 3) Data collection guidance to optimize the data acquisition process. The results of TrafficSafe Attribution provide actionable insights to enhance data analysis and collection, fostering a more comprehensive understanding of crash data and events.",
+    "footnote": [],
+    "bbox": [
+      [
+        140,
+        87,
+        848,
+        690
+      ]
+    ],
+    "page_idx": 2
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2: TrafficSafe Crash Outcomes Prediction Pipeline. Multi-modal crash data is collected and organized into textual prompts through an AI-expert cooperative process. The HSIS crash data, satellite images, and infrastructure data are used to extract general and infrastructure information, including the crash time, location, the road level, and so on. The vehicle data, and person data are converted into the event information and the unit information, including vehicle movements, driver characteristics (e.g., age, gender, alcohol use), vehicle attributes (e.g., manufacture year), and so on. TrafficSafe Event dataset is created with three prediction targets: Injury, Severity, and Type. The Injury task predicts the number of people injured in the crash event, the Severity task estimates the severity level of the crash, such as no apparent injury or fatal, and the Type task classifies type of crash, such as single vehicle with object or angle impacts right (The crash event consequences classification are provided in Supplementary Table 2 and Supplementary Table 3. The TrafficSafe LLM is fine-tuned using the TrafficSafe Event dataset. To reframe the crash outcomes prediction from a classification task to a language inference task, TrafficSafe LLM is fine-tuned by adding prediction targets as special tokens in its vocabulary and adjusting parameters using Low-Rank Adaptations (LoRA) \\(^{29}\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        140,
+        100,
+        880,
+        433
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure 3: TrafficSafe LLM Provides Accurate and Trustworthy Predictions. TrafficSafe LLM produces robust confusion matrices for both the (a) Washington and (b) Illinois datasets (we select the best results for each task by F1-score). In contrast, baseline models tend to predict the most frequent category across both the (c) Washington and (d) Illinois datasets (we show baseline models with the best F1-score. The performances for other baseline models can be found in Extended Data Figure 3). Meanwhile, TrafficSafe LLM produces trustworthy predictions for both the (e) Washington and (f) Illinois datasets. Higher confidence levels in the model's predictions correspond to an increased likelihood of accuracy. Furthermore, (g) The TrafficSafe LLM achieves higher precision for fatal crash predictions. (h) Fatal crash predictions also exhibit higher confidence compared to average predictions in Illinois dataset. The Washington dataset is not shown due to limited fatal cases. (i) For fatal crashes, the TrafficSafe LLM achieves near-perfect precision (97.61%) when the confidence score exceeds 0.6, indicating that the TrafficSafe LLM can deliver highly accurate and trustworthy predictions for fatal crashes.",
+    "footnote": [],
+    "bbox": [
+      [
+        105,
+        92,
+        899,
+        720
+      ]
+    ],
+    "page_idx": 9
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "(a) Sentence-based Feature Attribution Results for a Crash Resulting in Serious Injuries in Washington Dataset.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5: Conditional Risk Analysis for the Serious Injury and Fatal Crashes. Higher confidence scores in TrafficSafe LLM's predictions correspond to greater accuracy, allowing the confidence score (calculated as the sum of feature contributions for all data components) to serve as an indicator of risk level for serious and fatal crashes. (a) The estimated risk levels for various feature combinations are presented, with each level corresponding to the average confidence score of TrafficSafe LLM's predictions under the same conditions. Each column represents a specific combination of conditions (marked by dark dots) alongside the corresponding feature contribution for selected factors. (b) Feature contributions of five key factors and their proportions relative to all factors are visualized. The inner circle represents the average feature contribution of each factor across different values, while the outer circle shows the percentage share of each factor in the total average feature contribution. The unit for BAC (Blood Alcohol Content) is \"mg/L\", which is omitted in this figure. (c) Average feature contribution for each factor under specific values. Bars are marked in pink if the value exceeds the corresponding factor's average shown in (a) and in blue if it does not. (d) The strong correlation between the number of risk factors and the risk level of crashes. The high-risk factors are defined as driving after drinking (both BAC <= 80 mg/L and BAC > 80 mg/L), driving in work zones, driving on freeways, pedestrian-involved crashes, and high-risk driver behaviors (aggressive or impairment-related). For each case, we tallied the number of these risk factors and calculated the average risk level for all cases sharing the same count. (e) Feature contributions of different data components during the training stage for Washington and Illinois datasets.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        100,
+        866,
+        586
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Extended Data Figure 3: The Confusion Matrix for TrafficSafe LLM and the Traditional Methods in (a) Washington Dataset and (b) Illinois Dataset.",
+    "footnote": [],
+    "bbox": [
+      [
+        100,
+        83,
+        890,
+        680
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "TrafficSafe Attribution - #EC36495",
+    "footnote": [],
+    "bbox": [
+      [
+        485,
+        120,
+        888,
+        420
+      ]
+    ],
+    "page_idx": 31
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Extended Data Figure 5: One Example of Sentence-based Feature Attribution Results for A Crash Resulting in No Apparent Injury in Illinois Dataset.",
+    "footnote": [],
+    "bbox": [
+      [
+        102,
+        270,
+        896,
+        660
+      ]
+    ],
+    "page_idx": 32
+  }
+]

preprint/preprint__09f1a55d00104ac396ababe3986f197810df2a7652eb1fe992b357090dd1b42c/preprint__09f1a55d00104ac396ababe3986f197810df2a7652eb1fe992b357090dd1b42c.mmd

@@ -0,0 +1,628 @@
+
+# Customizing Large Language Models for Reliable and Interpretable Traffic Crash Prediction and Safety Interventions  
+
+Hao Frank Yang haofrankyang@jhu.edu  
+
+Johns Hopkins University https://orcid.org/0000- 0001- 6431- 8956  
+
+Yang Zhao Johns Hopkins University Pu Wang Johns Hopkins University Yibo Zhao Johns Hopkins University Hongru Du Johns Hopkins university  
+
+Article  
+
+Keywords:  
+
+Posted Date: April 29th, 2025  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 5947574/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on October 7th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 64574- w.
+
+<--- Page Split --->
+
+# Customizing Large Language Models for Reliable and Interpretable Traffic Crash Prediction and Safety Interventions  
+
+Yang Zhao \(^{1, 2 + }\) , Pu Wang \(^{1, 2 + }\) , Yibo Zhao \(^{1, 2}\) , Hongru Du \(^{1, 2}\) , and Hao (Frank) Yang \(^{1, 2*}\)  
+
+\(^{1}\) Center for Systems Science and Engineering, Johns Hopkins University, Baltimore, MD, USA. \(^{2}\) Department of Civil and Systems Engineering, Johns Hopkins University, Baltimore, MD, USA. \(^{+}\) The authors contributed equally. \(^{*}\) The corresponding authors information: haofrankyang@jhu.edu  
+
+## ABSTRACT  
+
+Predicting crash events is crucial for understanding crash distributions and their contributing factors, thereby enabling the design of proactive traffic safety policy interventions. However, existing methods struggle to interpret the complex interplay among various sources of traffic crash data, including numeric characteristics, textual reports, crash imagery, environmental conditions, and driver behavior records. As a result, they often fail to capture the rich semantic information and intricate interrelationships embedded in these diverse data sources, limiting their ability to identify critical crash risk factors. In this research, we propose TrafficSafe, a framework that adapts Large Language Models (LLMs) to reframe crash prediction and feature attribution as text- based reasoning. A multi- modal crash dataset including 58,903 real- world reports together with belonged infrastructure, environmental, driver, and vehicle information is collected and textualized into TrafficSafe Event dataset (totaling 12.74 million words). By customizing and fine- tuning state- of- the- art LLMs on this dataset, the proposed TrafficSafe LLM achieves a \(42\%\) average improvement in F1- score over baselines across multiple crash prediction tasks, particularly for severe crashes. To interpret these predictions and uncover contributing factors, we introduce TrafficSafe Attribution, a sentence- level feature attribution framework enabling conditional risk analysis. Findings show that alcohol- impaired driving is the leading factor in severe crashes, with aggressive and impairment- related behaviors having nearly twice the contribution for severe crashes compared to other driver behaviors. In addition, the co- occurrence of crash- contributing factors, such as alcohol- impaired driving, work zones, improper driving behaviors and other factors can significantly elevate risk levels. Furthermore, TrafficSafe Attribution highlights pivotal features during model training, guiding strategic crash data collection for iterative performance improvements. The proposed TrafficSafe offers a transformative leap in traffic safety research based on foundation models, providing a blueprint for translating advanced artificial intelligence technologies into responsible, actionable, and life- saving outcomes. It is now reshaping how traffic researchers and policymakers approach the road safety.  
+
+## 1 Introduction  
+
+Predicting traffic crash outcomes at the event level can greatly improve our understanding of crashes contributing factors and support the safety policy interventions. Currently, the United States has one of the highest traffic crash risks among developed countries (see Figure 1a), with 42,795 fatalities reported in \(2022^{2}\) . The number of fatalities still shows a persistent upward trend over recent decades, highlighting the urgent need for innovative
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>Figure 1: Overview of the Proposed TrafficSafe Framework. (a) The U.S. faces one of the highest crash risks among developed countries, with a rising trend. However, analyzing and addressing this issue is challenging due to the heterogeneous factors involved in crash events, including traffic conditions, human behavior, environmental impacts, and driver characteristics. To tackle this, we propose TrafficSafe, a framework designed for two key tasks: 1) Predicting crash outcomes and 2) Attributing crash factors with conditional risk analysis. By addressing questions such as why crashes occur and how to mitigate crash risks, TrafficSafe seeks to deliver optimal policy for safety improvement, aligning with the Vision Zero goal. (b) The TrafficSafe workflow incorporates multi-modal data, including driver behavior, vehicle details, infrastructure, and environmental conditions, represented through textual reports, satellite imagery, and other formats. Leveraging an AI-expert cooperative method, the crash data is transformed into textual prompts, resulting in the TrafficSafe Event dataset comprising 58,903 prompts. TrafficSafe LLM is created with accurate and trustworthy forecasting abilities for further analysis. Building on this pipeline, TrafficSafe Attribution operates across three dimensions: 1) Event-level risk analysis to identify feature contributions, 2) Conditional risk analysis to assess state-level risks under varying conditions, and 3) Data collection guidance to optimize the data acquisition process. The results of TrafficSafe Attribution provide actionable insights to enhance data analysis and collection, fostering a more comprehensive understanding of crash data and events. </center>
+
+<--- Page Split --->
+
+approaches to uncover the major causes of crashes and provide actionable insights for policy interventions. An effective data- driven crash prediction model can learn from historical crash- related data, and offer potential guidance for reducing crash risk by identifying the leading factors of crashes<sup>3</sup>. Current research for crash prediction can be grouped into two groups: 1) macroscopic (statistic- level) prediction<sup>4–8</sup> and 2) microscopic (event- level) prediction<sup>9–13</sup>. Macroscopic prediction typically relies on statistical methods to gain a general understanding of safety levels, compare the safety performance of different areas and time frames, identify high- risk zones, and track safety trends<sup>4,5</sup>. While these methods can partially predict when and where crashes are more likely to occur, they fail to forecast who is involved, what types will likely to be, why crashes happen, and how to mitigate risks at event granularity<sup>6,14</sup>. To address this limitation, microscopic crash prediction, which focuses on specific traffic conditions and circumstances, has been developed to predict the crashes consequences using machine learning (ML) approaches<sup>11,13</sup>. Despite their potential in answering who and what, these models face limitations in crash prediction precision and generalization<sup>5,6</sup>. Moreover, integrating multi- modal traffic crash data and interpreting model’s outputs (together with contributing factors) remain challenging. Consequently, existing crash prediction models struggle to accurately forecast crash outcomes and effectively incorporate their insights into the design of policy intervention.  
+
+Crash data are inherently heterogeneous, making accurate prediction a significant challenge. After the crash happened, first responders compile textual and numerical on- site details, often supplemented by images, driver behavior data, and licensing records. Although these diverse sources hold immense potential for crash prediction and feature attribution, three key obstacles must be addressed: 1) Data Integration. Existing approaches often reduce multi- modal data to one- hot embeddings for classification tasks<sup>15–17</sup>. However, these approaches often neglect the valuable information contained within textual and behavioral data, potentially limiting the accuracy and reliability of crash prediction models<sup>11,13</sup>. 2) Method Generalization. Scaling crash- event prediction models to new data remains a complex endeavor due to the large variety of features, the complexity of representation extraction and encoding, and the diverse formats in which crash data appear<sup>18,19</sup>. Current machine learning solutions are often tailored to specific data types, limiting their adaptability when new cases or additional data modalities arise<sup>6,20–22</sup>. 3) Feature Learning and Attribution. Multi- modal crash data regularly include partially overlapping information, such as road attributes recorded in both on- site images and textual crash reports, complicating the accurate assessment of each feature’s unique contribution.  
+
+Recent advancements in Large Language Models (LLMs), such as GPT- 4<sup>23</sup> and LLaMA 3<sup>24</sup>, have demonstrated their potential for deriving complex crash patterns from multi- modal data<sup>25</sup> and addressing persistent challenges. However, fully adapting LLMs to predict crash outcomes and inform effective safety interventions requires overcoming three primary technical hurdles in data, model, and interpretability. From a data perspective, diverse crash records, including images, textual notes, and driver behavior logs, must be reformatted into textual inputs suitable for LLM processing. In terms of modeling, the generative nature of LLMs, which have extensive output vocabularies (e.g., LLaMA 3’s 128,256 tokens), poses challenges for discriminative learning tasks and raises concerns about trustworthiness, particularly when crash outcomes (e.g., crash type or severity) are well- defined by public agencies into finite categories. Furthermore, interpreting LLM’s outputs for crash prediction becomes difficult, as it remains unclear how much we can trust the forecasting results and how individual factors contribute to crash outcomes. This lack of interpretability and robustness analysis hinders the development of data- driven,
+
+<--- Page Split --->
+
+actionable plans for mitigating the crash risks.  
+
+This study advances traditional traffic safety analysis by shifting from aggregate- level considerations to event- level crash prediction. We propose TrafficSafe (see Figure 1b), a novel LLM- driven framework designed for addressing these challenges to provide a comprehensive understanding of crash events. TrafficSafe comprises three main components: TrafficSafe Event dataset for multi- modal crash data integration, TrafficSafe LLM for crash outcomes prediction, and TrafficSafe Attribution for conditional risk analysis. Together, these components enable accurate and trustworthy crash consequences prediction and risk attribution, answering the when, where, who, what, why, and how to support targeted traffic safety interventions. By reframing crash outcome prediction as a text- based reasoning task, TrafficSafe exploits the inherent language reasoning capabilities of LLMs to offer actionable insights for crash prevention, ultimately paving the way for data- driven safety solutions.  
+
+## 2 Novelties and Contributions  
+
+This study advances traditional traffic safety analysis by shifting from aggregate- level considerations to event- level crash prediction. In particular, we customize LLMs to forecast expected crash consequences and attribute relevant features with enhanced accuracy and interpretability. Our proposed framework, TrafficSafe, supports reliable and accountable learning from multi- modal crash data, facilitating a deeper understanding of crash events. Key contributions and findings include:  
+
+Unlocking multi- modal data integration and text reasoning for crash consequence prediction. We introduce the TrafficSafe framework to extend the LLMs for crash outcomes prediction. Rather than treating crash features as isolated numerical inputs, TrafficSafe integrates them into the broader semantic context of traffic data. To effectively utilize and integrate the multi- modal crash data, the TrafficSafe Event dataset is constructed with 58,903 textual prompts totaling over 12 million words. The TrafficSafe LLM is then fine- tuned by framing crash outcome prediction as a task specific token generation task. This approach yields a \(41.7\%\) increase in average F1- score across multiple crash consequence prediction tasks.  
+
+Integrating traffic safety priors in LLMs for trustworthy crash predictions. Compared with existing LLMs, we incorporate crash- domain knowledge and priors into the model's vocabulary as special tokens. This addition allows us to tailor the output to specific crash categories, including crash type, severity, and number of injuries, thereby providing a direct way to measure the model's trustworthiness and link to targeted interventions. Experimental results of TrafficSafe LLM show a strong correlation between increasing confidence in the model's output and higher prediction accuracy, achieving over \(70\%\) accuracy when the confidence score exceeds \(60\%\) . Notably, the model reaches more than \(95\%\) precision for fatal crash predictions when the confidence score surpasses \(60\%\) . This feature offers quantitative evidence to support safety- oriented decision- making and helps close the gap regarding how to trust the model's predictive results.  
+
+Advancing feature interpretation for conditional risk analysis and policy intervention, even in unseen scenarios. The TrafficSafe Attribution framework is proposed for conditional risk analysis, which is supported by a novel sentence- level feature contributions calculation method, enabling event- level feature attribution for textual inputs of TrafficSafe LLM. Then, the "what- if" conditional analysis can further identify and analyze
+
+<--- Page Split --->
+
+the most critical risk conditions and their combinations. For instance, alcohol- impaired driving consistently emerges as a leading contributor to serious and fatal crashes. While driving in a work zone under sober conditions poses minimal risk, combining these conditions with alcohol consumption drastically increases danger, making it one of the most hazardous scenarios for severe crashes. Furthermore, aggressive and impairment- related behaviors demonstrate nearly double the impact on severe crashes compared to other driver behaviors. These insights lay the groundwork for implementing targeted traffic safety policies and interventions<sup>26</sup>.  
+
+- Guiding optimal data collection for efficient model evolution and lifelong learning. A longstanding challenge in crash modeling is determining how to select valuable data from heterogeneous sources and prioritize which information first responders should capture during incident documentation. The proposed TrafficSafe Attribution addresses this by estimating the contributions of multi-modal data during training, then quantifying which data types have the greatest impact on model performance. Such insights guide more effective traffic safety data collection, improving crash prediction accuracy while also supporting efficient, continuous model evolution through a targeted, data-driven strategy.  
+
+## 3 Results  
+
+### 3.1 Multi-modal Crash Data  
+
+Our cleaned dataset comprises crash data from Washington State in 2022, totaling 16,188 records, and from Illinois in 2022, totaling 42,715 records, after excluding cases with missing key attributes related to vehicle or crash object status. Primary sources include the Highway Safety Information System (HSIS) crash data<sup>27</sup> and satellite images<sup>28</sup>. The HSIS crash data contains four major components: crash data, infrastructure data, vehicle data, and the person data. Crash data provides detailed descriptions of crashes, such as location, time, and injury severity. Infrastructure data includes information about road layouts and traffic characteristics, such as road level and speed limits. Vehicle data contains details such as manufacturing year and reported defects of the involved vehicles, while person data captures demographic and other relevant details about drivers and passengers, such as age and gender. Satellite images complement the HSIS data by providing additional visual context, including information about lanes, intersections, and other roadway attributes. Further information on raw data formats and types is available in Section 5.1.1.  
+
+### 3.2 TrafficSafe Crash Outcomes Prediction Pipeline  
+
+To leverage the multi- modal crash data described in Section 3.1 for crash prediction, we developed the TrafficSafe crash outcomes prediction pipeline, which transforms crash outcomes prediction into a text- based reasoning task. To achieve this, the raw crash data is organized into the textual TrafficSafe Event dataset, which is then used to fine- tune the TrafficSafe LLM. Figure 2 presents an overview of the TrafficSafe crash outcomes prediction pipeline, with subsequent sections detailing each stage of the pipeline.  
+
+### 3.2.1 Constructing Prompts and Prediction Targets  
+
+The TrafficSafe Event dataset is created through an AI- expert cooperative textualization process, organizing multimodal raw data for effective crash prediction. The detailed information about the raw data feature engineering and
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Figure 2: TrafficSafe Crash Outcomes Prediction Pipeline. Multi-modal crash data is collected and organized into textual prompts through an AI-expert cooperative process. The HSIS crash data, satellite images, and infrastructure data are used to extract general and infrastructure information, including the crash time, location, the road level, and so on. The vehicle data, and person data are converted into the event information and the unit information, including vehicle movements, driver characteristics (e.g., age, gender, alcohol use), vehicle attributes (e.g., manufacture year), and so on. TrafficSafe Event dataset is created with three prediction targets: Injury, Severity, and Type. The Injury task predicts the number of people injured in the crash event, the Severity task estimates the severity level of the crash, such as no apparent injury or fatal, and the Type task classifies type of crash, such as single vehicle with object or angle impacts right (The crash event consequences classification are provided in Supplementary Table 2 and Supplementary Table 3. The TrafficSafe LLM is fine-tuned using the TrafficSafe Event dataset. To reframe the crash outcomes prediction from a classification task to a language inference task, TrafficSafe LLM is fine-tuned by adding prediction targets as special tokens in its vocabulary and adjusting parameters using Low-Rank Adaptations (LoRA) \(^{29}\) . </center>  
+
+117 the textualization process are available in Section 5.1.2. As shown in Figure 2, the constructed prompts are divided into five parts: one system prompt and four content parts, with each content part containing approximately 100 words. These parts include:  
+
+- System Prompt: Provides an introduction and task-specific instructions.  
+
+- General Information: Includes general information about the time and location of the prediction region and the roadway category.  
+
+- Infrastructure Information: Describes road infrastructure, encompassing static features like the number of lanes and speed limits, as well as dynamic elements such as work zones, lighting, and road surface
+
+<--- Page Split --->
+
+conditions.  
+
+- Event Information: Contains detailed descriptions of crash events, such as the number of vehicles involved and their directions of movement.  
+
+- Unit Information: Provides vehicle and individual details relevant for crash prediction, such as airbag status and the driver's age.  
+
+The prediction targets consist of three variables: Injury, Severity, and Type (see Figure 2) \(^{30 - 32}\) . Specifically, Injury task predicts the number of people injured in the given crash event. Injury task is treated as a classification task with four categories: zero, one, two, and three or more than three, where crashes involving more than two injured people are grouped into a single category due to the limited number of such cases. The Severity task assesses the level of injury severity in a crash, classified into five levels from no apparent injury to fatal. Type task predicts the type of crash, such as the rear-end collision or collision with object, with 14 crash type categories in the Washington dataset and 16 in the Illinois dataset. Detailed information on the defined targets is available in Section 5.1.3.  
+
+For each crash event, we perform the feature engineering and textualization process, organize the textualized data as input, and process labels corresponding to three tasks. The complete prompt examples are presented in Extended Data Figure 1 and Extended Data Figure 2. Ultimately, after filtering out data items with missing information, the TrafficSafe Event dataset merges the complementary information from multi- modal data sources and contains 58,903 crash records with approximately 12.74 million words. These records are split into training, validation, and test sets in a 7:1.5:1.5 ratio.  
+
+#### 3.2.2 Adapting LLM for Crash Prediction  
+
+Although vanilla LLMs like Llama 3 possess broad general knowledge and strong text reasoning capabilities, they demonstrate limited effectiveness on crash prediction tasks without the fine- tuning process (see Supplementary Section 1). To address this, we developed TrafficSafe LLM, a specialized model fine- tuned on the processed TrafficSafe Event dataset. This fine- tuning process enhances the LLM's comprehension of crash events and enables accurate outcome prediction. Specifically, special tokens are introduced into the LLM vocabulary as prediction targets (Number of Injury, Severity, and Crash Type), fine- tuning the model to generate these tokens during prediction. The details of the fine- tuning are provided in Section 5.2.  
+
+### 3.3 Performance Evaluation of TrafficSafe LLM  
+
+In this section, we evaluate the performance of TrafficSafe LLM and compare its performance with other baselines (see Section 5.2.6). The fine- tuning process is based on two vanilla LLMs with different sizes: Llama 3.1 8B and Llama 3.1 70B. Accuracy, precision, and F1- score are used as the evaluation metrics, the detail information is available in Section 5.2.5.  
+
+TrafficSafe LLM provides accurate crash predictions, even in zero- shot scenarios. Table 1 compares the performances of TrafficSafe LLM and adopted baselines. The results show that the TrafficSafe LLM outperforms all the baselines in each task setting with an average F1- score improvement of 41.7% across multiple tasks. Specifically, in the crash Type prediction task in Washington dataset, the TrafficSafe LLM achieves F1- score of
+
+<--- Page Split --->
+
+
+Table 1: Performance Comparison of the three Crash Prediction Tasks on Washington Dataset and Illinois Dataset. We present quality metrics along with model rankings by averaging the column-wise rank. The zero-shot results for North Carolina and Maine were derived using the model trained on the Illinois dataset. In supervised finetuning experiments on the Washington and Illinois datasets, TrafficSafe LLM outperforms all other methods, with TrafficSafe 70B achieving the best performance. Additionally, TrafficSafe LLM demonstrates strong generalization capabilities in zero-shot experiments on the North Carolina and Maine datasets.   
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Model</td><td colspan="3">Injury</td><td colspan="3">Severity</td><td colspan="3">Type</td><td rowspan="2">Rank</td></tr><tr><td>Accuracy</td><td>Precision</td><td>F1-score</td><td>Accuracy</td><td>Precision</td><td>F1-score</td><td>Accuracy</td><td>Precision</td><td></td></tr><tr><td rowspan="7">Washington</td><td>RandomForest33</td><td>0.522</td><td>0.649</td><td>0.545</td><td>0.628</td><td>0.546</td><td>0.549</td><td>0.740</td><td>0.398</td><td>0.274</td><td>4 (4.11)</td></tr><tr><td>AdaBoost34</td><td>0.495</td><td>0.245</td><td>0.328</td><td>0.492</td><td>0.245</td><td>0.328</td><td>0.563</td><td>0.249</td><td>0.302</td><td>8 (6.00)</td></tr><tr><td>CatBoost35</td><td>0.495</td><td>0.245</td><td>0.328</td><td>0.492</td><td>0.245</td><td>0.328</td><td>0.715</td><td>0.400</td><td>0.329</td><td>6 (5.22)</td></tr><tr><td>DecisionTree36</td><td>0.495</td><td>0.245</td><td>0.328</td><td>0.528</td><td>0.428</td><td>0.372</td><td>0.628</td><td>0.406</td><td>0.323</td><td>5 (4.67)</td></tr><tr><td>LogisticRegression37</td><td>0.495</td><td>0.245</td><td>0.328</td><td>0.492</td><td>0.245</td><td>0.328</td><td>0.547</td><td>0.401</td><td>0.309</td><td>7 (5.67)</td></tr><tr><td>XGBoost38</td><td>0.566</td><td>0.665</td><td>0.469</td><td>0.534</td><td>0.428</td><td>0.367</td><td>0.739</td><td>0.413</td><td>0.298</td><td>3 (4.00)</td></tr><tr><td>National Baseline39</td><td>0.343</td><td>0.555</td><td>0.424</td><td>0.353</td><td>0.547</td><td>0.429</td><td>/</td><td>/</td><td>/</td><td>/</td></tr><tr><td rowspan="3"></td><td>TrafficSafe 8B</td><td>0.622</td><td>0.630</td><td>0.618</td><td>0.640</td><td>0.636</td><td>0.634</td><td>0.756</td><td>0.763</td><td>0.755</td><td>2 (2.22)</td></tr><tr><td>TrafficSafe 70B</td><td>0.630</td><td>0.682</td><td>0.649</td><td>0.648</td><td>0.644</td><td>0.644</td><td>0.760</td><td>0.775</td><td>0.759</td><td>1 (1.00)</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td rowspan="7">Illinois</td><td>RandomForest33</td><td>0.462</td><td>0.554</td><td>0.383</td><td>0.430</td><td>0.452</td><td>0.338</td><td>0.610</td><td>0.670</td><td>0.632</td><td>3 (4.11)</td></tr><tr><td>AdaBoost34</td><td>0.403</td><td>0.183</td><td>0.251</td><td>0.318</td><td>0.147</td><td>0.200</td><td>0.109</td><td>0.083</td><td>0.083</td><td>8 (8.00)</td></tr><tr><td>CatBoost35</td><td>0.457</td><td>0.543</td><td>0.388</td><td>0.454</td><td>0.446</td><td>0.404</td><td>0.535</td><td>0.656</td><td>0.579</td><td>4 (4.22)</td></tr><tr><td>DecisionTree36</td><td>0.426</td><td>0.514</td><td>0.410</td><td>0.417</td><td>0.398</td><td>0.361</td><td>0.504</td><td>0.624</td><td>0.548</td><td>6 (5.33)</td></tr><tr><td>LogisticRegression37</td><td>0.413</td><td>0.439</td><td>0.410</td><td>0.360</td><td>0.385</td><td>0.355</td><td>0.379</td><td>0.477</td><td>0.400</td><td>7 (6.33)</td></tr><tr><td>XGBoost38</td><td>0.442</td><td>0.575</td><td>0.340</td><td>0.405</td><td>0.419</td><td>0.278</td><td>0.678</td><td>0.694</td><td>0.683</td><td>5 (4.56)</td></tr><tr><td>National Baseline39</td><td>0.369</td><td>0.136</td><td>0.199</td><td>0.442</td><td>0.195</td><td>0.271</td><td>/</td><td>/</td><td>/</td><td>/</td></tr><tr><td rowspan="2"></td><td>TrafficSafe 8B</td><td>0.529</td><td>0.529</td><td>0.533</td><td>0.578</td><td>0.584</td><td>0.571</td><td>0.701</td><td>0.768</td><td>0.721</td><td>2 (1.89)</td></tr><tr><td>TrafficSafe 70B</td><td>0.534</td><td>0.587</td><td>0.543</td><td>0.554</td><td>0.561</td><td>0.548</td><td>0.727</td><td>0.767</td><td>0.737</td><td>1 (1.44)</td></tr><tr><td>North Carolina</td><td>TrafficSafe 8B (zero-shot)</td><td>0.511</td><td>0.776</td><td>0.468</td><td>0.549</td><td>0.638</td><td>0.487</td><td>0.691</td><td>0.775</td><td>0.672</td><td>/</td></tr><tr><td>Maine</td><td>TrafficSafe 8B (zero-shot)</td><td>0.521</td><td>0.573</td><td>0.457</td><td>0.542</td><td>0.582</td><td>0.493</td><td>0.701</td><td>0.622</td><td>0.613</td><td>/</td></tr></table>  
+
+0.759, which is more than \(130\%\) higher than all other comparative methods. TrafficSafe LLM performs well on both the Washington and Illinois datasets, demonstrating its stability across diverse geographical regions. Moreover, as shown in the confusion matrix in Figure 3a and Figure 3b, beyond improved metrics, TrafficSafe LLM demonstrates a more balanced prediction distribution. In contrast, as shown in Figure 3c and Figure 3d, traditional machine learning models tend to predict the dominant categories (e.g., zero under Injury prediction task, no apparent injury under Severity prediction task). The complete confusion matrix is shown in Extended Data Figure 3. Moreover, the ability of a model to generalize across unseen scenarios is vital to ensuring its robustness and applicability in real- world contexts. We tested TrafficSafe LLM generalization ability by using the TrafficSafe LLM trained on Illinois dataset and evaluating its performance on the unseen Maine and North Carolina datasets. Notably, without additional fine- tuning, TrafficSafe LLM achieved F1- scores averaging 0.542 in North Carolina and 0.521 in Maine, closely matching its performance in Illinois (see Table 1). This underscores TrafficSafe LLM's ability to generalize well to previously unseen datasets, further validating its potential for real- world applications.  
+
+TrafficSafe LLM provides trustworthy crash predictions, where a higher confidence score links to higher accuracy. TrafficSafe LLM tailors LLMs for discriminative crash outcomes prediction tasks, generating predictions accompanied by confidence scores that represent the probabilities associated with specific special tokens. Figure 3e and Figure 3f illustrate the trend of accuracy in relation to the confidence scores of TrafficSafe LLM's predictions for the Washington and Illinois datasets. The results indicate that our model achieves greater accuracy at higher confidence levels. For instance, for the Injury prediction task in the Washington dataset, when
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 3: TrafficSafe LLM Provides Accurate and Trustworthy Predictions. TrafficSafe LLM produces robust confusion matrices for both the (a) Washington and (b) Illinois datasets (we select the best results for each task by F1-score). In contrast, baseline models tend to predict the most frequent category across both the (c) Washington and (d) Illinois datasets (we show baseline models with the best F1-score. The performances for other baseline models can be found in Extended Data Figure 3). Meanwhile, TrafficSafe LLM produces trustworthy predictions for both the (e) Washington and (f) Illinois datasets. Higher confidence levels in the model's predictions correspond to an increased likelihood of accuracy. Furthermore, (g) The TrafficSafe LLM achieves higher precision for fatal crash predictions. (h) Fatal crash predictions also exhibit higher confidence compared to average predictions in Illinois dataset. The Washington dataset is not shown due to limited fatal cases. (i) For fatal crashes, the TrafficSafe LLM achieves near-perfect precision (97.61%) when the confidence score exceeds 0.6, indicating that the TrafficSafe LLM can deliver highly accurate and trustworthy predictions for fatal crashes. </center>
+
+<--- Page Split --->
+
+the model's confidence score exceeds 0.40, the accuracy rises above 0.65, and with confidence scores over 0.60, the accuracy surpasses 0.80. The strong positive correlation between confidence scores and accuracy showcases the trustworthiness of the TrafficSafe framework. By providing reliable confidence scores alongside predictions, the framework empowers informed decision- making in real- world applications.  
+
+### 3.4 TrafficSafe Attribution and Result Interpretation  
+
+Understanding how TrafficSafe LLM generates accurate predictions and how various components of the input prompt influence the outcomes is fundamental to enabling evidence- based decision- making. As shown in figure 3e and 3f, the TrafficSafe LLM's confidence score strongly correlates with its predictive accuracy for fatal and serious injury crashes, therefore, we can use the confidence score to represent a case's real- world risk level. Notably, the TrafficSafe LLM's confidence scores tend to be lower than their corresponding precision values (see figure 3e and 3f, indicating that the confidence score is a conservative estimate of risk).  
+
+Within the TrafficSafe Attribution framework, a sentence- based feature contributions calculation method was proposed to identify how each sentence contributes to the LLM's outputs based on Shapley theory which is recognized as a systematic and equitable method for attributing the contribution of each feature to a model's output40,41, thereby revealing crash- related factors at the event level (see Section 5.3 for details). In essence, each feature's contribution represents its share of responsibility for the model's confidence in a particular prediction. The sum of all feature contributions equals the confidence score itself. Figure 4 illustrates sentence- level feature contributions for the severity of individual crash events, using one crash from Washington and one from Illinois as examples. In the Washington crash example (Figure 4a), Driver Behavior (e.g., reckless driving or speeding) is the primary factor contributing to serious injury crashes with the feature contribution of 0.258. Person Info (e.g., no seatbelt use) also shows a substantial impact with the feature contribution of 0.149. By contrast, Dynamic Info (daylight and dry roads) lowers the probability of crash with serious injuries with a negative feature contribution of - 0.009. While, in the Illinois example (Figure 4b), an elevated BAC (Blood Alcohol Content, with feature contribution of 0.284) and the presence of a Work Zone (feature contribution of 0.462) notably increase the likelihood of fatal crash outcomes. Beyond the above examples, more additional sentence- level feature attribution analysis can be found in Extended Data Figure 4, 5, Supplementary Section 5 and 6.  
+
+The following sections utilize TrafficSafe Attribution framework to examine feature importance from two perspectives: 1) at the inference stage, to identify key factors influencing crash predictions under various conditions and high- risk scenarios, and 2) at the training stage, to understand which data components are most important for model learning.  
+
+### 3.4.1 Factor Attribution at Inference Stage for Conditional Risk Analysis  
+
+Conditional analysis evaluates crash outcomes across various scenarios, such as driving with or without alcohol consumption, to quantify the risk factors associated with each scenario. Severe crashes (serious injuries and fatal crashes) were prioritized in the conditional analysis due to their critical importance for traffic safety. These crashes, particularly fatal ones, were predicted accurately and reliably by TrafficSafe LLM (see Figures 3g, 3h, and 3i). Five key contributing factors were identified for this conditional analysis: Driver BAC (BAC = 0 or not offered / BAC < 80 / BAC >= 80), Roadway Type (Highway / not highway), Work Zone (Work zone / not work zone), User Type (Pedalcyclist or pedestrian / not pedalcyclist or pedestrian), and Driver Behavior (Aggressive driving
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+  
+
+![](images/Figure_5.jpg)
+
+<center>(a) Sentence-based Feature Attribution Results for a Crash Resulting in Serious Injuries in Washington Dataset. </center>  
+
+(b) Sentence-based Feature Attribution Results for a Crash Resulting in Fatalities in Illinois Dataset.  
+
+Figure 4: Single Case Feature Attribution Results for Severity Task. The left part displays the full prompt from (a) Washington and (b) Illinois, with different colors representing various semantic text sequences. The right part illustrates the feature contribution assigned to each text sequence. Positive contributions signify a supportive role in the model's prediction, whereas negative contributions indicate a detracting influence. The absolute value of these contributions represents the importance of each sequence to the model's output.
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>Figure 5: Conditional Risk Analysis for the Serious Injury and Fatal Crashes. Higher confidence scores in TrafficSafe LLM's predictions correspond to greater accuracy, allowing the confidence score (calculated as the sum of feature contributions for all data components) to serve as an indicator of risk level for serious and fatal crashes. (a) The estimated risk levels for various feature combinations are presented, with each level corresponding to the average confidence score of TrafficSafe LLM's predictions under the same conditions. Each column represents a specific combination of conditions (marked by dark dots) alongside the corresponding feature contribution for selected factors. (b) Feature contributions of five key factors and their proportions relative to all factors are visualized. The inner circle represents the average feature contribution of each factor across different values, while the outer circle shows the percentage share of each factor in the total average feature contribution. The unit for BAC (Blood Alcohol Content) is "mg/L", which is omitted in this figure. (c) Average feature contribution for each factor under specific values. Bars are marked in pink if the value exceeds the corresponding factor's average shown in (a) and in blue if it does not. (d) The strong correlation between the number of risk factors and the risk level of crashes. The high-risk factors are defined as driving after drinking (both BAC <= 80 mg/L and BAC > 80 mg/L), driving in work zones, driving on freeways, pedestrian-involved crashes, and high-risk driver behaviors (aggressive or impairment-related). For each case, we tallied the number of these risk factors and calculated the average risk level for all cases sharing the same count. (e) Feature contributions of different data components during the training stage for Washington and Illinois datasets. </center>
+
+<--- Page Split --->
+
+/ impairment- related behavior / traffic rules violations / improper driving / others). Collectively, these factors accounted for an average of \(79.33\%\) of the model's overall attribution in predicting serious and fatal crashes (see Figure 5b). A summary of key findings is provided:  
+
+- The BAC record emerges as a critical determinant in predicting serious and fatal crashes. Among all contributing factors, BAC accounts for \(25.26\%\) of the total contribution to serious and fatal crash prediction (see Figure 5b). Notably, its contribution substantially increases when a driver consumes alcohol, irrespective of the amount. When drivers are under the influence of alcohol even if their BAC does not exceed the legal intoxication limit of \(80 \mathrm{mg / L}^{42,43}\) , this factor's feature contribution still reaches approximately 0.45, surpassing that of most other factors in many cases (see Figure 5a). Conversely, when a driver's BAC is recorded as "zero or not offered," its contribution approaches zero, indicating minimal impact on the model's predictions.  
+
+- Driving in a work zone is already risky under sober conditions, but alcohol consumption significantly increases the danger, making it one of the most hazardous scenarios for severe-injury crashes. As shown in Figure 5a, driving in a work zone while sober ("Work Zone-Yes" and "BAC = 0 or not offered") contributes little to severe crash outcomes, with an average feature contribution of 0.03. However, after consuming alcohol (whether "BAC >= 80" or "BAC < 80"), the work zone feature contribution rises more than seven time to an average of 0.22. Furthermore, the overall crash risk increases substantially when driving in a work zone after drinking, as indicated by an average risk level of 0.78, compared to 0.44 under sober conditions. These findings indicate that work zones become especially hazardous when alcohol consumption is involved, creating one of the highest-risk scenarios for severe crash outcomes. Potential drunk driving warnings and risk mitigation strategies shall be closely linked with work-zone areas.  
+
+- Aggressive driving and impairment-related behavior exhibit the highest contributions among driver behaviors. Furthermore, combined with other conditions, aggressive and impairment-related behaviors pose nearly twice the risk for severe crash outcomes compared to other driver behaviors. As illustrated in Figure 5c, aggressive driving emerges as the most significant contributor between driver behaviors, with feature contribution of 0.14. Impairment-related behavior, including driving under the influence of alcohol or drugs, also has a substantial influence, with average feature contribution of 0.11. In comparison, other improper driver behaviors, such as traffic rule violations (feature contribution of 0.07) and distractions like mobile phone use (categorized under improper driving, with feature contribution of 0.03), show below-average contributions to serious and fatal crashes. The "other" category, which includes normal driving and unknown behaviors, has the smallest impact, with feature contribution of 0.03.  
+
+- The co-occurrence of risk factors significantly increases the expected crash risk level. As illustrated in Figure 5d, our analysis reveals a strong correlation between the number of risk factors present in a crash and the expected risk level for severe crash outcomes. When only one risk factor is involved, the risk level for severe crash outcomes is estimated at 0.59. This value increases to 0.62 with two risk factors, surges to 0.78 with three, and escalates to 0.94 when four risk factors co-occur. Notably, scenarios with three or more risk factors are markedly more dangerous than those with one or two. For example, while a combination of a BAC
+
+<--- Page Split --->
+
+exceeding \(80\mathrm{mg / L}\) and driving in a non- freeway work zone yields an average risk level of 0.51, substituting the non- freeway work zone with a freeway work zone increases the average risk level dramatically to 0.97. Such elevated risks indicate that the synergy among multiple factors is far from merely additive; instead, they appear to compound one another, amplifying the potential for severe outcomes. These findings show that transportation agencies need to prioritize multi- faceted interventions tailored specifically to scenarios with overlapping high- risk conditions.  
+
+### 3.4.2 Factor Attribution at Training Stage for Effective Data Collection and Model Development  
+
+Event information and unit information are the most important components for the model training. While feature contributions at the inference stage reveal which features drive critical crash outcomes, understanding feature contributions during training provides deeper insights into which data components most effectively enhance model accuracy. As shown in Figure 5e, the feature contributions of each component in the Washington and Illinois datasets are shown, demonstrating their impact on the model's performance during training (see Supplementary Table 7 for detailed results and Section 5.3 for calculation details). The results indicate that in both the Washington and Illinois datasets, for the Severity task, the unit information describing attributes of the primary entities involved in the crash has the highest contribution to the model's performance (0.314 in Washington, 0.234 in Illinois). Event information, which provides information on the vehicle's movement prior to the crash, is followed by unit information and has the second highest contribution (0.110 in Washington, 0.158 in Illinois). For the Crash Type Prediction task, the event information has the highest contribution (0.388 in Washington, 0.283 in Illinois), followed by the unit information (0.257 in Washington, 0.279 in Illinois) and other components. These results can provide preliminary guidance on prioritizing the information collection for crash events, thereby improving crash prediction and feature attributions for better safety decision support.  
+
+## 4 Discussion  
+
+Deciphering traffic crash modeling as a linguistic learning task is a promising way for future safety research. Most of the existing ML models for crash prediction typically treat various factors as independent numerical input variables<sup>11,44</sup>. However, this approach fails to capture information richness from the textual crash reports, such as detailed descriptions of behaviors, vehicle movements prior to the crash, and the traffic conditions. To address these issues, we employ an AI- expert cooperative prompt design approach to process diverse data types, including crash reports (textual), satellite and crash images (visual), and infrastructure characteristics (categorical), into a textual TrafficSafe Event dataset and use LLM for prediction. This transformation reframes the task of crash prediction into a text reasoning problem, enabling the use of LLMs to analyze and predict outcomes while preserving the rich, detailed textual information in crash reports, rather than reducing it to simplistic numerical representations. As demonstrated by the results in Section 3.3, with our customization process, the TrafficSafe LLM outperforms all the baseline models, highlighting the advantages of reasoning through textual representations. Building on this, TrafficSafe Attribution extends the framework by enabling conditional analysis of textual prompts, quantifying the contribution of specific factors to crash outcomes under various scenarios. As shown in Section 3.4, this approach effectively identifies key contributors to crashes and high- risk scenarios, and offers data collection guidance for the iterative improvements in the future.
+
+<--- Page Split --->
+
+Providing reliable and interpretable predictions with quantifiable trustworthiness. As illustrated in Figures 3e and 3f, TrafficSafe LLM demonstrates a deep understanding of input- output correlations, yielding predictions whose accuracy increases alongside higher confidence scores. In all tasks across Washington and Illinois dataset, when the TrafficSafe LLM's confidence score exceeds \(60\%\) , which captures over \(70\%\) of the crash events under consideration. Furthermore, the confidence scores for fatal crash predictions are notably higher than those of other crash categories ( \(60\%\) of model confident score leading to over \(95\%\) of the real- world occurrence risk, see Figures 3g, 3h, and 3i). This trackable confidence- precision correlation can provide decision- makers with a robust tool for forecasting crashes under quantifiable uncertainty. Beyond predictive trustworthiness, the TrafficSafe Attribution framework provides interpretable feature attribution by quantifying each feature's contribution to the confidence score (see Section 5.3). A higher feature contribution translates into a higher confidence score, which in turn yields greater prediction accuracy for the severe crash outcomes. Thus the factor with higher feature contribution value has higher impact to the model's prediction. For instance, alcohol- impaired driving (BAC \(>0\) ) increases the confidence score for severe crash predictions by more than 0.47, serving as a critical indicator for the likelihood of these severe crash outcomes.  
+
+Identifying high- risk traffic crashes through conditional factors attribution even in unseen scenarios. The TrafficSafe framework enables a detailed, sentence- level analysis of crash factors through conditional attribution, yielding critical insights into high- risk scenarios. In data- rich situations where sufficient data is available for each condition, TrafficSafe can rank the risk levels associated with various conditions, offering a prioritized list of scenarios that pose the highest danger. This capability supports targeted policy interventions by identifying specific conditions that substantially increase crash risks. For instance, as shown in Figure 5a, driving in a work zone under sober conditions poses low level of risk; however, alcohol consumption in the same setting dramatically amplifies this risk, creating one of the most hazardous scenarios for severe crashes. This insight suggests potential policy interventions, such as mandatory BAC testing in work zones, to mitigate these risks. Likewise, Figure 5c highlights that aggressive driving and impairment- related behaviors markedly increase the likelihood of serious or fatal outcomes, emphasizing the importance of driver education to discourage aggressive behavior and driving under the influence. Moreover, TrafficSafe can be generalized to predict and understand data- sparse scenarios through "what- if" analysis, allowing hypothetical changes to specific conditions to be tested and their potential impact evaluated. For example, while this study lacked sufficient data to comprehensively analyze the effects of user type (e.g., pedestrians or not pedestrians) or roadway type (e.g., freeway or not freeway), TrafficSafe provides a reliable mechanism to simulate and analyze such conditions.  
+
+Assessing the impact of data utility for improved future data collection and life- long learning. Traffic crash data are inherently complex and multi- modal, making it crucial to identify which components are most informative and how critical they are for future traffic safety data collection. During the training stage, data attribution analysis revealed that unit information (e.g., driver behavior and vehicle details) and event information (e.g., vehicle movement and environmental conditions) exert the greatest influence on crash prediction performance. Specifically, for the Severity task in the Illinois dataset, these features contributed 0.173 and 0.287, respectively, to the model's prediction confidence (see Figure 5e). These findings underscore the importance of prioritizing the collection of detailed, high- quality movement and behavior data in crash events, such as precise records of alcohol use, vehicle defects, vulnerable users' status, and road conditions. In contrast, although general information
+
+<--- Page Split --->
+
+(feature contribution of 0.038) and infrastructure information (feature contribution of 0.019) remain valuable, their impact on some tasks is comparatively smaller. By directing data collection efforts toward gathering richer, more consistent information format in these critical safety domains, the accuracy of TrafficSafe can be further improved. Limitations of the proposed TrafficSafe. A primary limitation relates to the handling of multi- modal data. In the TrafficSafe framework, satellite images were processed into textual descriptions and incorporated into prompts. While this approach offers flexibility, advancements in multi- modal foundation models and increasing research on integrating multi- modal data with LLMs present promising alternatives. Leveraging specialized image encoders or utilizing multi- modal foundation models for processing image data are compelling directions. Another potential limitation lies in the efficiency of model training and attribution. Fine- tuning LLMs and computing feature contributions have always required substantial resources and time. Although we employed LoRA fine- tuning and a stratified sampling technique to enhance efficiency, implementing the complete framework still demands significant resources. This poses certain limitations when resources are scarce or in situations demanding rapid model deployment.  
+
+## Reference  
+
+1. Federal Highway Administration (FHWA). Vision Zero Action Plans Accessed: 2025-01-15. 2025.  
+2. International Transport Forum (ITF). Road Safety Annual Report 2023 (OECD Publishing, Paris, 2023).  
+3. Islam, M. R., Wang, D. & Abdel-Aty, M. Calibrated confidence learning for large-scale real-time crash and severity prediction. npj Sustainable Mobility and Transport 1, 1 (2024).  
+4. Bougna, T., Hundal, G. & Taniform, P. Quantitative analysis of the social costs of road traffic crashes literature. Accident Analysis & Prevention 165, 106282 (2022).  
+5. Wen, X., Xie, Y., Jiang, L., Pu, Z. & Ge, T. Applications of machine learning methods in traffic crash severity modelling: current status and future directions. Transport reviews 41, 855-879 (2021).  
+6. Mannering, F., Bhat, C. R., Shankar, V. & Abdel-Aty, M. Big data, traditional data and the tradeoffs between prediction and causality in highway-safety analysis. Analytic methods in accident research 25, 100113 (2020).  
+7. Yan, X., Zou, Z., Feng, S., Zhu, H., Sun, H. & Liu, H. X. Learning naturalistic driving environment with statistical realism. Nature communications 14, 2037 (2023).  
+8. Tazul Islam, M., Thue, L. & Grekul, J. Understanding traffic safety culture: implications for increasing traffic safety. Transportation Research Record 2635, 79-89 (2017).  
+9. Mannering, F. L. & Bhat, C. R. Analytic methods in accident research: Methodological frontier and future directions. Analytic methods in accident research 1, 1-22 (2014).  
+10. Dong, C., Shao, C., Li, J. & Xiong, Z. An improved deep learning model for traffic crash prediction. Journal of Advanced Transportation 2018, 3869106 (2018).  
+11. Theofilatos, A., Chen, C. & Antoniou, C. Comparing machine learning and deep learning methods for real-time crash prediction. Transportation research record 2673, 169-178 (2019).
+
+<--- Page Split --->
+
+12. Rahim, M. A. & Hassan, H. M. A deep learning based traffic crash severity prediction framework. \*Accident Analysis & Prevention\* \*\*154\*\*, 106090 (2021).  
+
+13. Sattar, K., Chikh Oughali, F., Assi, K., Ratrout, N., Jamal, A. & Masirur Rahman, S. Transparent deep machine learning framework for predicting traffic crash severity. \*Neural Computing and Applications\* \*\*35\*\*, 1535–1547 (2023).  
+
+14. Carrodano, C. Data-driven risk analysis of nonlinear factor interactions in road safety using Bayesian networks. \*Scientific Reports\* \*\*14\*\*, 18948 (2024).  
+
+15. Sharma, A., Zheng, Z., Kim, J., Bhaskar, A. & Haque, M. M. Is an informed driver a better decision maker? A grouped random parameters with heterogeneity-in-means approach to investigate the impact of the connected environment on driving behaviour in safety-critical situations. \*Analytic Methods in Accident Research\* \*\*27\*\*, 100127. ISSN: 2213-6657 (2020).  
+
+16. Xu, C., Ding, Z., Wang, C. & Li, Z. Statistical analysis of the patterns and characteristics of connected and autonomous vehicle involved crashes. \*Journal of safety research\* \*\*71\*\*, 41–47 (2019).  
+
+17. Abdel-Aty, M. & Ding, S. A matched case-control analysis of autonomous vs human-driven vehicle accidents. \*Nature Communications\* \*\*15\*\*, 4931 (2024).  
+
+18. Boggs, A. M., Wali, B. & Khattak, A. J. Exploratory analysis of automated vehicle crashes in California: A text analytics & hierarchical Bayesian heterogeneity-based approach. \*Accident Analysis & Prevention\* \*\*135\*\*, 105354 (2020).  
+
+19. Wali, B., Khattak, A. J. & Karnowski, T. The relationship between driving volatility in time to collision and crash-injury severity in a naturalistic driving environment. \*Analytic methods in accident research\* \*\*28\*\*, 100136 (2020).  
+
+20. Ahmed, S. S., Cohen, J. & Anastasopoulos, P. C. A correlated random parameters with heterogeneity in means approach of deer-vehicle collisions and resulting injury-severities. \*Analytic methods in accident research\* \*\*30\*\*, 100160 (2021).  
+
+21. Liu, Z., Chen, Y., Xia, F., Bian, J., Zhu, B., Shen, G. & Kong, X. Tap: Traffic accident profiling via multi-task spatio-temporal graph representation learning. \*ACM Transactions on Knowledge Discovery from Data\* \*\*17\*\*, 1–25 (2023).  
+
+22. Lu, J., Grembek, O. & Hansen, M. Learning the representation of surrogate safety measures to identify traffic conflict. \*Accident Analysis & Prevention\* \*\*174\*\*, 106755 (2022).  
+
+23. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. \*arXiv preprint arXiv:2303.08774\* (2023).  
+
+24. Meta, A., Jauhi, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. \*arXiv preprint arXiv:2407.21783\* \*\*2\*\* (2024).  
+
+25. Gao, C., Lan, X., Lu, Z., Mao, J., Piao, J., Wang, H., Jin, D. & Li, Y. S3: Social-network Simulation System with Large Language Model-Empowered Agents. \*arXiv preprint arXiv:2307.14984\* (2023).
+
+<--- Page Split --->
+
+26. Rezapour, M. & Ksabati, K. Identification of factors associated with various types of impaired driving. Humanities and social sciences communications 9, 1–11 (2022).  
+
+27. U.S. Department of Transportation, Federal Highway Administration. Highway Safety Information System (HSIS) https://highways.dot.gov. Accessed: January 13, 2025. 2025.  
+
+28. Developers, G. Google Maps Static API Documentation https://developers.google.com/maps/documentation/maps-static. Accessed: January 13, 2025. 2025.  
+
+29. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L. & Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021).  
+
+30. Abdel-Aty, M., Keller, J. & Brady, P. A. Analysis of types of crashes at signalized intersections by using complete crash data and tree-based regression. Transportation Research Record 1908, 37–45 (2005).  
+
+31. Iranitalab, A. & Khattak, A. Comparison of four statistical and machine learning methods for crash severity prediction. Accident Analysis & Prevention 108, 27–36. ISSN: 0001-4575 (2017).  
+
+32. Savolainen, P. T., Mannering, F. L., Lord, D. & Quddus, M. A. The statistical analysis of highway crash-injury severities: A review and assessment of methodological alternatives. Accident Analysis & Prevention 43, 1666–1676. ISSN: 0001-4575 (2011).  
+
+33. Breiman, L. Random Forests. Machine Learning 45, 5–32 (Oct. 2001).  
+
+34. Freund, Y. & Schapire, R. E. A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting. Journal of Computer and System Sciences 55, 119–139. ISSN: 0022-0000 (1997).  
+
+35. Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V. & Gulin, A. CatBoost: unbiased boosting with categorical features 2019. arXiv: 1706.09516 [cs.LG].  
+
+36. Quinlan, J. R. Induction of decision trees. Machine learning 1, 81–106 (1986).  
+
+37. Cox, D. R. The regression analysis of binary sequences. Journal of the Royal Statistical Society Series B: Statistical Methodology 20, 215–232 (1958).  
+
+38. Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Association for Computing Machinery, San Francisco, California, USA, 2016), 785–794. ISBN: 9781450342322.  
+
+39. Lord, D., Geedipally, S., Pratt, M., Park, E. S., Khazraee, S. & Fitzpatrick, K. Safety Prediction Models for Six-Lane and One-Way Urban and Suburban Arterials ISBN: 978-0-309-29560-4 (Mar. 2022).  
+
+40. Bordt, S. & von Luxburg, U. From shapley values to generalized additive models and back in International Conference on Artificial Intelligence and Statistics (2023), 709–745.  
+
+41. Shapley, L. S. in Contributions to the Theory of Games II (eds Kuhn, H. W. & Tucker, A. W.) 307–317 (Princeton University Press, Princeton, 1953).  
+
+42. Washington State Legislature. Revised Code of Washington: Driving under the influence https://app.leg.wa.gov/rcw/default.aspx?cite=46.61.502. Accessed: 2025-01-19. 2025.
+
+<--- Page Split --->
+
+43. Illinois Secretary of State. Driving Under the Influence (DUI) https://www.ilsos.gov/departments/drivers/traffic_safety/DUI/home.html. Accessed: 2025-01-19. 2025. 
+44. Bhuiyan, H., Ara, J., Hasib, K. M., Sourav, M. I. H., Karim, F. B., Sik-Lanyi, C., Governatori, G., Rakotoni-rainy, A. & Yasmin, S. Crash severity analysis and risk factors identification based on an alternate data source: a case study of developing country. Scientific reports 12, 21243 (2022). 
+45. Zhang, J., Huang, J., Jin, S. & Lu, S. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024). 
+46. Zhang, J., Sun, Q., Liu, J., Xiong, L., Pei, J. & Ren, K. Efficient Sampling Approaches to Shapley Value Approximation. Proc. ACM Manag. Data 1 (May 2023).
+
+<--- Page Split --->
+
+## 5 Methods  
+
+### 5.1 TrafficSafe Event Dataset Construction  
+
+As introduced in Section 3.1, the raw crash data is multi- modal, and integrated from various sources. To adapt the raw data for LLMs' fine- tuning process, we employ the feature engineering and textualization process to generate textual inputs. In this section, we will discuss the formats of raw data and the textualization process (see Extended Data Figure 6).  
+
+### 5.1.1 Raw Data  
+
+The raw crash data used in this study was obtained from the HSIS \(^{27}\) and Google Maps \(^{28}\) . Data from the HSIS, sourced from multiple systems, encompasses a variety of formats, including categorical, numerical, and textual. In total, four main datasets were used:  
+
+- Crash data. This dataset captures the essential spatio-temporal and contextual attributes of each crash. It includes crash date, time, day of the week, and month, along with location details such as route number, milepost, and the surrounding area's classification (e.g., rural or urban). Higher-level planning attributes (e.g., roadway and functional classifications, intersection-related indicators) are also recorded. In addition, it documents the dynamic circumstances leading up to the event, including the number of vehicles and pedestrians involved, vehicle travel directions (increasing or decreasing milepost), and any maneuvers performed (e.g., lane changes, straight-line movement).  
+
+- Infrastructure data. This dataset details the physical and infrastructural features of the crash site. Key elements include the type of road surface (e.g., asphalt or concrete), average annual daily traffic (AADT), posted speed limits, and access control mechanisms. It also encompasses dimensions such as total road width, right and left shoulder widths, and median width (including median barriers if present), as well as road surface conditions (e.g., dry or wet) and ambient lighting at the time of the crash (e.g., daylight or dusk).  
+
+- Vehicle data. This dataset consolidates information on the vehicles involved in each crash, including vehicle type (e.g., passenger car or truck), intended use (e.g., commercial or private), mechanical condition (e.g., defects), and relevant driver actions (e.g., lane changes or stopping). Additional information on airbag deployment and occupant ejection status provides further granularity.  
+
+- Person data. This dataset compiles information about individuals involved in the crash, detailing demographic characteristics such as age, gender, and seating position. It also includes the use of safety equipment (e.g., seat belts or helmets) and any contributing factors, such as driver distraction or impairment.  
+
+The satellite images obtained from Google Maps serve as a supplementary data source to complement the HSIS dataset. Details of the integration process are provided in Section 5.1.2. Overall, we collect 16,188 crash events data from Washington State and 42,715 events from Illinois State for further analysis.  
+
+### 5.1.2 Feature Engineering and Textualization of Crash Data  
+
+To adapt the multi- modal data to the input of LLMs, we followed the following process to generate textual prompt from raw data entry:
+
+<--- Page Split --->
+
+- Data mapping and organization. For each crash, we associated the crash report with the involved vehicles and individuals using the crash ID, thus obtaining descriptions of the crash and the persons involved. The route ID and milepost were used to identify the specific road segment where the crash occurred, allowing us to gather related road and environment information from infrastructure data. The integrated data was then systematically organized into four categories: general information, infrastructure information, event information, and unit information, aligning with the components outlined in Section 3.2.1.  
+
+- Satellite images textualization. The HSIS datasets provide GPS coordinates for crash locations in Washington and Illinois. To address missing information such as the number of road lanes, high-resolution satellite images (512 × 512 pixels at a zoom level of 19) were retrieved using these GPS coordinates via the Google Maps API. These images supplement the crash dataset with crucial infrastructure and environmental context. Descriptive textual annotations were generated from the satellite images using GPT-4, filling key gaps in the original dataset. These annotations include information such as the number of lanes at the crash site, whether the crash occurred at an intersection, and whether the surrounding area is residential.  
+
+- Dimensionality reduction. Raw data include abundant attributes with rich and varied descriptions. However, some features suffer from insufficient distinction between attribute values due to the original classification's complexity. To address this, we performed dimensionality reduction on these attributes by combining domain experts' insights with GPT-4o clustering results. For example, similar classifications like "pedalyclist struck by vehicle" and "pedalyclist strikes vehicle" were clustered under a broader category such as "pedalyclist collisions". This process generalized the data and reduced redundancy.  
+
+- Prompt generation using AI-expert textualization method. To generate logically coherent and continuous textual data suitable for LLM training, we transformed each category of data into text format using GPT-4. All data are organized as key-value pairs and we get four parts of the key-value pairs for each event case. Then GPT-4o is used to generate the text prompt for each section of the key-value pairs individually. For each part, we apply straightforward prompt to GPT-4o, such as "Please translate a python dictionary to paragraph, act as a crash data interpreter". The text content is extracted from GPT-4o's response for each part consisting of approximately 100 words. By linking four parts of text, we obtain a comprehensive textual description for each crash event case. The detailed process is shown in Extended Data Figure 7.  
+
+#### 5.1.3 Define Prediction Targets  
+
+We select three variables as the prediction targets: Injury, Severity, and crash Type. The three targets are defined as:  
+
+- The Injury \(n_{l}^{\mathcal{P}}\in \{f(l)|l = 0,1,2,\dots \}\) , where \(i\) denotes the \(i\) -th data in the dataset, \(\mathcal{D}\in \{\mathcal{W},\mathcal{I}\}\) denotes the Washington dataset \(\mathcal{W}\) or the Illinois dataset \(\mathcal{I}\) , \(l\) represents the number of people injured, and \(f(l)\) denotes the label when the injured people is \(l\) .  
+
+- The Severity \(s_{l}^{\mathcal{P}}\in \{S_{k}|k = 1,2,\dots \}\) (define on the KABCO scale \(^1\) ), where \(S_{k}\) is the \(k\) -th level of crash severity.
+
+<--- Page Split --->
+
+- The Type \(t_{i}^{\mathcal{O}} \in \{T_{k}^{\mathcal{O}} | k = 1, 2, \dots \}\) , where \(T_{k}^{\mathcal{O}}\) is the \(k\) -th label of crash type in dataset \(\mathcal{D}\) .  
+
+We utilize these three variables to describe the crash result \(\mathrm{CR}_{i}^{\mathcal{O}}\) . The crash outcome can be presented in the following format: \(\mathrm{CR}_{i}^{\mathcal{O}} = (n_{i}^{\mathcal{O}}, s_{i}^{\mathcal{O}}, t_{i}^{\mathcal{O}})\) . For numerical variables, the function \(f(l)\) describes the number of people injured in crash as follows: "zero" if \(l = 0\) , "one" if \(l = 1\) , "two" if \(l = 2\) , and "three and more than three" if \(l \geq 3\) , the values for \(S_{k}\) and \(T_{k}^{\mathcal{O}}\) are provided in the Supplementary Table 2 and Supplementary Table 3.  
+
+### 5.2 TrafficSafe LLM  
+
+We fine- tune TrafficSafe LLM by adapting LLaMa 3.1<sup>24</sup> to crash prediction tasks to enhance the LLMs' capabilities in interpreting crash data, identifying critical factors, and conducting feature attribution analysis to offer insights for crash prevention. In this section, we will introduce detailed information of the fine- tuning process.  
+
+#### 5.2.1 Construct Training Data for LLMs  
+
+In the training of LLMs, a single input consists of three components: the system prompt, the user prompt, and the target prompt. The system prompt introduces the task, for example: "You are a helpful assistant designed to predict the severity of a traffic crash ...". The user prompt comprises the four content parts detailed in Section 5.1.2 for each case. The target prompt represents the expected output. Examples of these prompts are shown in Extended Data Figure 1, Extended Data Figure 2, and Supplementary Section 3. We tokenize the text inputs using LLaMA 3.1's tokenizer.  
+
+#### 5.2.2 Additional Special Tokens for Classification  
+
+To adapt the LLM as a crash classifier, additional tokens have been incorporated into the tokenizer's vocabulary, and the detailed crash attributes categories are listed in Supplementary Table 2 and Supplementary Table 3. Specifically, for predicting the number of people Injuries of Washington dataset and Illinois dataset, we have introduced four special tokens: <ZERO>, <ONE>, <TWO>, and <THREE AND MORE THAN THREE>. Similarly, for predicting the Crash Severity of Washington dataset and Illinois dataset, we use five additional tokens: \(S_{k}\) , where \(1 \leq k \leq 5\) , corresponding to different levels of severity. The Type task differs slightly between the Washington and Illinois datasets. For Washington datasets, we utilize 14 special tokens: \(T_{k}^{\mathcal{W}}\) , where \(1 \leq k \leq 14\) , each representing a specific crash type. For Illinois datasets, we utilize 16 special tokens: \(T_{k}^{\mathcal{F}}\) , where \(1 \leq k \leq 16\) . The parameters of the input and output embedding layers are set as trainable, enabling the model to align the representations of these special tokens with the existing embedding space.  
+
+#### 5.2.3 Supervised Fine-tuning  
+
+During the fine- tuning phase, the traffic forecasting task is framed as a next- token generation task. This process can be described as:  
+
+\[p_{\theta}(T_{i}) = \prod_{j = 1}^{|T_{i}|} p_{\theta}(t_{j}^{(i)} | t_{1}^{(i)}, \dots , t_{j - 1}^{(i)}), \quad (1)\]  
+
+where \(T_{i}\) is the \(i\) - th item in the training data, \(p_{\theta}\) is the LLM, \(t_{j}^{(i)}\) denotes the \(j\) - th token in \(T_{i}\) . By maximizing the likelihood \(p_{\theta}(T) = \prod_{i = 1}^{N} p_{\theta}(T_{i})\) , the LLM's parameters are learned. Both the system prompt and the user prompt are masked for loss computation during training. We also used uniform data sampling strategy during the training process to facilitate the convergence of TrafficSafe LLM<sup>47</sup>. Through this process, the model learns to make prediction for a traffic crash.
+
+<--- Page Split --->
+
+### 5.2.4 Data Split  
+
+We split the Washington and Illinois dataset into training, validation, and test set in a 7:1.5:1.5 ratio. Since the Washington dataset contains relatively few crash events per year, we utilized as many reports as possible to ensure sufficient training data. However, the data distribution across different classes is highly imbalanced. For example, in the crash severity prediction task in Washington dataset, the ratio of \(\# S_{1} / \# S_{5}\) is nearly 100:1, where \(\# S_{k}\) is the number of data with label \(S_{k}\) . The imbalanced data distribution presents a significant challenge for the model's training and evaluation. In Section 5.2.3, we used uniform sampling strategy to train model on this unbalanced data. Similarly, to facilitate the model's evaluation, for the validation set and test set, we removed most of the data with crash severity category of \(S_{1}\) . Specifically, after processing, the dataset consisted of 16,188 records, with 11,332 used for training, 2,428 for validation, and 2,428 for testing. To balance the validation and test set for better evaluation, we removed 1428 \(S_{1}\) data and used 1000 remaining data for validation set and test set separately. Compared with the Washington state, more crash records can be used in Illinois state to generate dataset. As a result, we were able to balance all subsets, including the training, validation, and test sets. Ultimately, the Illinois dataset comprised 42,715 records, with 29,307 used for training, 6,704 for validation, and 6,704 for testing.  
+
+#### 5.2.5 Evaluation Metrics  
+
+In evaluating the model performance as a classification task, we employ weighted accuracy, precision, and F1- score as metrics. In the context of a classification task, we have four notations, True Positive \((TP)\) , True Negative \((TN)\) , False Positive \((FP)\) , False Negative \((FN)\) . Using these notations, we can represent the metrics as follows:  
+
+- Accuracy is one of the most commonly used measures for the classification performance, and it is defined as a ratio between the correctly classified samples to the total number of samples as follows:  
+
+\[\mathrm{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} \quad (2)\]  
+
+- Precision represents the proportion of positive samples that were correctly classified to the total number of positive predicted samples, which reflect the performance of the prediction:  
+
+\[\mathrm{Precision} = \frac{TP}{TP + FP} \quad (3)\]  
+
+- F1-score combines results on precision and recall. It is the harmonic mean of precision and recall, which can be calculated using formula:  
+
+\[\mathrm{F1 - score} = \frac{2}{\mathrm{Precision}^{-1} + \mathrm{Recall}^{-1}} = 2\cdot \left(\frac{\mathrm{Precision}\cdot\mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}}\right) \quad (4)\]  
+
+where \(\mathrm{Recall} = TP / (TP + FN)\) .  
+
+#### 5.2.6 Adopted Baselines  
+
+We follow the recent literature \(^{48}\) and also adopt XGBoost \(^{38}\) , Random forest (RF) \(^{33}\) , Decision Trees (DT) \(^{36}\) , Adaptive boosting (AdaBoost) \(^{49}\) , LogisticRegression (LR) \(^{37}\) , Categorical boosting (CatBoost) \(^{50}\) , and National Average \(^{39}\) as compared baselines. Building upon these foundational models, we particularly focus on enhancing
+
+<--- Page Split --->
+
+their predictive capabilities through advanced techniques and parameter optimization. The detailed descriptions of these models are listed as follows:  
+
+- XGBoost is a scalable and distributed gradient-boosting framework that constructs an ensemble of decision trees by minimizing a regularized loss function. It uses second-order gradients for optimization and includes features like shrinkage, column subsampling, and tree pruning to improve accuracy and prevent overfitting<sup>38</sup>.  
+
+- AdaBoost<sup>34</sup> is an iterative boosting method that sequentially trains weak classifiers (e.g., decision stumps) and assigns higher weights to misclassified instances in subsequent iterations. The final prediction is determined by a weighted majority vote of all classifiers.  
+
+- Random Forest (RF)<sup>33</sup> builds an ensemble of decision trees by randomly sampling both features and data points (via bootstrap aggregation). The aggregated (voted) output of these diverse trees reduces variance and provides robust performance across a variety of tasks.  
+
+- Decision Trees (DT)<sup>36</sup> recursively split the feature space based on selected thresholds, forming a hierarchical tree structure that is easy to interpret. Although they can capture complex interactions, DTs are prone to overfitting if not properly regularized.  
+
+- Logistic Regression (LR)<sup>37</sup> models the probability of a binary outcome through a linear combination of input features passed through the logistic function. Coefficients are typically estimated via maximum likelihood, providing a simple yet effective approach for classification.  
+
+- CatBoost<sup>50</sup> is a gradient-boosting algorithm that efficiently handles categorical features through techniques such as ordered boosting and gradient-based one-hot encoding. By systematically reducing target leakage in encoding, it achieves high predictive accuracy while mitigating overfitting in heterogeneous datasets.  
+
+- National Average<sup>39</sup> predicts crash severity distributions using calibrated Severity Distribution Functions (SDFs). It incorporates road design, traffic control, and crash data to estimate probabilities for different severity levels via a multinomial logit model.  
+
+For these models, the Bayesian optimization method (BayesSearchCV) is used to facilitate the identification of optimal hyperparameters, such as max_depth and learning_rate. The details of the hyperparameters settings of these models are shown in Supplementary Section 4.  
+
+### 5.3 TrafficSafe Attribution  
+
+To identify the feature contribution of each factor to the prediction results, this paper introduces and adapts the concept of Shapley values<sup>40</sup>. In this section, we first explain the calculation process of Shapley values and subsequently propose a novel sentence-level feature contributions calculation method based on Shapley theory for attributing factors in LLMs.  
+
+#### 5.3.1 Definition of Shapley Value  
+
+Shapley value is a concept from cooperative game theory that has been widely adopted in machine learning to interpret model predictions<sup>51</sup>. It provides a way to fairly allocate the contribution of each feature to the outcome of
+
+<--- Page Split --->
+
+a predictive model. In essence, the Shapley value quantifies how much each feature contributes to a prediction by considering all possible combinations of features. Formally, the Shapley value \(\phi\) of a feature (or player) \(i\) in a cooperative game is defined as:  
+
+\[\phi_{i} = \sum_{S\subseteq N\backslash \{i\}}\frac{|S|!(n - |S| - 1)!}{n!}\Big[v(S\cup \{i\}) - v(S)\Big], \quad (5)\]  
+
+where \(N = \{1,2,\ldots ,n\}\) is the index set of \(n\) features, \(S\) is a subset of \(N\) , and \(v(S)\) is the utility of the subset \(S\) , which represents a measurable value, such as accuracy or prediction score, achieved by the model using only the subset \(S\) of features.  
+
+The Shapley value is utilized in both the training and inference stages in TrafficSafe. During the training stage, it quantifies the contributions of four primary categories of information: general information, infrastructure information, event information, and unit information. During the inference stage, the Shapley value is applied to assess the contributions of individual sentences to the prediction outcomes. The specific methodologies and implementation details are outlined in the subsequent sections.  
+
+#### 5.3.2 Feature Contributions at the Training Stage  
+
+The Shapley value is utilized to assess the influence of different components in the training set on the model during training. As outlined in Section 3.2.1, the \(j\) - th prompt \(p_{j}\) in the dataset \(P\) is divided into five parts: \(c_{0}\) : system prompt (i.e. "You are a helpful assistant designed to predict the severity of a traffic crash ..."), \(c_{1}\) : general information, \(c_{2}\) : infrastructure information, \(c_{3}\) : event information, and \(c_{4}\) : unit information. We denote \(p_{j}(k)\) as the \(c_{k}\) portion of \(p_{j}\) . Given an index set \(S\) , we can construct a variant \(p_{j}(S)\) by concatenating the parts in \(S\) . For example, if \(S = \{0,1,2\}\) , then \(p_{j}(S)\) contains \(c_{0}\) , \(c_{1}\) , and \(c_{2}\) . Formally,  
+
+\[p_{j}(S) = \mathrm{concat}_{k\in S}p_{j}(k), \quad (6)\]  
+
+where concat denotes concatenation. The resulting dataset based on \(S\) is \(P(S) = \{p_{j}(S) | j = 0,1,\ldots ,L\}\) , where \(L\) is the dataset size.  
+
+Referring to Equation (5), the contribution of part \(c_{i}\) at training, \(\phi_{i}^{\mathrm{train}}\) , is  
+
+\[\phi_{i}^{\mathrm{train}} = \sum_{S\subseteq N\backslash \{i\}}\frac{|S|!(n - |S| - 1)!}{n!}\cdot \Big[v\big(P(S\cup \{0,i\})\big) - v\big(P(S\cup \{0\})\big)\Big], \quad (7)\]  
+
+where \(N = \{1,2,3,4\}\) indexes the four content parts, and \(v(P(S))\) is a performance metric (e.g., accuracy, F1- score) obtained after retraining the model only on prompts in \(P(S)\) .  
+
+#### 5.3.3 Sentence-level Feature Contributions at the Inference Stage  
+
+Unlike traditional machine learning models that primarily handle fixed- length feature vectors, LLMs process variable- length text sequences as input. This characteristic makes commonly used Shapley value approximation methods, such as KernelSHAP and DeepSHAP, less applicable to LLMs. Recent approaches like TokenSHAP and TransSHAP have been proposed to address this by decomposing input text into tokens and computing Shapley values at the token level. However, applying token- level Shapley value computation to TrafficSafe LLM
+
+<--- Page Split --->
+
+introduces two primary challenges: 1) Computational limitations. The computational complexity of Shapley values is exponential in the number of players. In our TrafficSafe LLM, with an input size of approximately 500 tokens, large- scale computation of token- level Shapley values for crash data becomes impractical. 2) Limited interpretability. Decomposing the prompt at the token level disregards inter- token dependencies, and the arbitrary masking or replacement of tokens can lead to semantic ambiguity and contextual shifts. These issues hinder a precise understanding of how individual features contribute to predictions. Moreover, paragraph- level analysis is too coarse for detailed attribution, since it can merge distinct features into a single category (e.g., driver and vehicle details under "unit information").  
+
+To overcome these limitations, we propose a sentence- level feature contributions calculation method for inputs of LLMs, which proceeds as follows:  
+
+- Sentence segmentation. The prompts are segmented using delimiters (e.g., commas ", or periods ".") to produce sentence-level units.  
+
+- Feature groups annotation. GPT-4o is used to group and label these sentences (see Figure 4 for the groups' content). Each group is represented as \(c_{k}\) , where \(k \in N' = \{1,2,3,\ldots n\}\) . For the Washington dataset, \(n = 14\) , while for the Illinois dataset \(n = 12\) . Given index set \(S' \subseteq N' \setminus \{i\}\) , we can construct the the prompt \(p_j(S')\) similar to the process Equation (6). The dataset built upon \(S'\) can be written as \(P(S') = \{p_j(S') | j = 0,1,2,\ldots ,L\}\) , where \(L\) is the length of the dataset \(P\) .  
+
+- Feature contributions calculation based on the feature groups. Based on the constructed dataset, the feature contribution for the \(i\) -th sentence-group \(\phi_i^{inf}\) can be calculated as:  
+
+\[\phi_{i}^{inf} = \sum_{S'\subseteq N'\setminus \{i\}}\frac{|S'|!(n - |S'| - 1)!}{n!}\cdot \left[p_{\theta}(P(S'\cup \{0,i\})) - p_{\theta}(P(S'\cup \{0\}))\right] \quad (8)\]  
+
+where \(p_{\theta}\) is the LLM that returns the predicted probability of the target. A higher \(\phi_{i}^{\mathrm{inf}}\) indicates a greater contribution of the \(i\) - th sentence group to the model's confidence. To reduce computational overhead, we adopt a stratified sampling- based Shapley estimation method using complementary contributions<sup>46</sup>.  
+
+## Reference  
+
+47. Du, H., Zhao, J., Zhao, Y., Xu, S., Lin, X., Chen, Y., Gardner, L. M. & Yang, H. F. Advancing Real-time Pandemic Forecasting Using Large Language Models: A COVID-19 Case Study. arXiv preprint arXiv:2404.06962 (2024).  
+
+48. Ahmed, S., Hossain, M. A., Ray, S. K., Bhuiyan, M. M. I. & Sabuj, S. R. A study on road accident prediction and contributing factors using explainable machine learning models: Analysis and performance. Transportation research interdisciplinary perspectives 19, 100814 (2023).  
+
+49. Freund, Y., Schapire, R. & Abe, N. A short introduction to boosting. Journal-Japanese Society For Artificial Intelligence 14, 1612 (1999).  
+
+50. Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V. & Gulin, A. CatBoost: unbiased boosting with categorical features. Advances in neural information processing systems 31 (2018).
+
+<--- Page Split --->
+
+71 Chen, H., Lundberg, S. M. & Lee, S.- I. Explaining a series of models by propagating Shapley values. Nature communications 13, 4512 (2022).  
+
+72 Chen, H., Covert, I. C., Lundberg, S. M. & Lee, S.- I. Algorithms to estimate Shapley value feature attributions. Nature Machine Intelligence 5, 590- 601 (2023).  
+
+73 Lundberg, S. M. & Lee, S.- I. A Unified Approach to Interpreting Model Predictions in Advances in Neural Information Processing Systems (eds Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S. & Garnett, R.) 30 (Curran Associates, Inc., 2017).  
+
+74 Goldsmith, R. & Horovicz, M. TokenSHAP: Interpreting Large Language Models with Monte Carlo Shapley Value Estimation. arXiv preprint arXiv:2407.10114 (2024).  
+
+80 Kokalj, E., Škrlj, B., Lavrač, N., Pollak, S. & Robnik- Šikonja, M. BERT meets Shapley: Extending SHAP Explanations to Transformer- based Classifiers in Proceedings of the EACL Hackashop on News Media Content Analysis and Automated Report Generation (eds Toivonen, H. & Boggia, M.) (Association for Computational Linguistics, Online, Apr. 2021), 16- 21.
+
+<--- Page Split --->
+
+## 6 Data Availability  
+
+Details of each raw data source and data processing are described in the Method Section. The processed data examples are available at https://github.com/Puw242/TrafficSafe. In compliance with HSIS data policy, requests for the complete raw dataset should be made via https://highways.dot.gov/research/ safety/hsis.  
+
+## 7 Code Availability  
+
+Code is publicly accessible at https://github.com/Puw242/TrafficSafe.  
+
+## 8 Author Contributions  
+
+Y.Z., P.W. and H.F.Y conceptualized and designed the study. P.W. and Yibo Z. collected data. P.W. and Yibo Z. processed the data and designed prompts. Y.Z. and P.W. performed experiments. Yibo Z. run the baseline models. Y.Z., P.W., Yibo Z. and H.F.Y prepared the figures. Y.Z., P.W., Yibo Z. and H.F.Y analyzed results. Y.Z., P.W., Yibo Z. and H.F.Y wrote the initial draft. H.D. and H.F.Y provided guidance and feedback for the study. H.D. and H.F.Y revised the manuscript. H.F.Y. acquired the funding. H.F.Y. provided computational resources. All authors prepared the final version of the manuscript.  
+
+## 9 Competing Interests  
+
+The authors declare no competing interests.
+
+<--- Page Split --->
+
+## Example Prompt - #EC22961  
+
+You are a helpful assistant designed to predict the task target of a traffic crash. You need to make prediction based on the information below:  
+
+## General Information  
+
+This incident occurred on February 23, 2022, at 2:00 PM, in the city of Bremerton, Kitsap County, on the 303 route increasing milepost direction at milepost 1.87. The location is an Urban - Principal Arterial, not at an intersection and not related to a driveway. The type of roadway is classified as an Urban Multilane Undivided Non- Freeway. The level of access control is Non Limited Access Least Restrictive, the speed limit is 30, and the average annual daily traffic is 37000.  
+
+## Infrastructure Information  
+
+The road width is 52 feet, the road surface is made of Asphalt, the right and left shoulders width is unknown, and the surface type of the left shoulder is unknown. This road does not have a median- separated area, there is no barrier in the median and the median width is unknown. The condition of the road is unknown regarding work zone status, but it is known that the accident occurred during daylight and the road surface condition was dry.  
+
+## Event Information  
+
+There were no pedestrians involved, 3 vehicles involved. The accident has no influence of alcohol or drugs. There were no objects involved. Vehicle1 was moving North, in the direction of increasing milepost, Vehicle2 was also moving North, in the direction of increasing milepost. The first vehicle was moving straight when the second vehicle was stopped in traffic, legally standing.  
+
+## Unit Information  
+
+The unit 1 is a Vanette Under 10,000 lb, non- commercial vehicle. The airbag was not deployed. The vehicle had no defects. The driver was going straight ahead, was not ejected, and was distracted by an unknown factor. Person 1: Motor Vehicle Driver, Female, 47, Restraint use is unknown. The unit 2 is a Vanette Under 10,000 lb, non- commercial vehicle. The airbag was not deployed. The vehicle had no defects. The driver had stopped for traffic, was not ejected, and no violations or factors contributed to the incident. Person 1: Motor Vehicle Driver, Female, 26, Restraint use is unknown. The unit 3 is a Vanette Under 10,000 lb, non- commercial vehicle. The airbag was not deployed. The vehicle had no defects. The driver was going straight ahead, was not ejected, and was distracted by an unknown factor. A drug recognition expert was not requested. Person 1: Motor Vehicle Driver, Female, 50, Restraint use is unknown.  
+
+## Targets  
+
+Please predict the Injury number of the crash choosing from the following tokens (4 options available).  
+
+Assistant: <ZERO>  
+
+Please predict the Severity of the crash choosing from the following tokens (5 options available).  
+
+Assistant: <NO APPARENT INJURY>  
+
+Please predict the crash Type of the crash choosing from the following tokens (14 options available).  
+
+Assistant: <REAR END COLLISIONS>  
+
+Extended Data Figure 1: A Crash Event Prompt Example from Washington Dataset.
+
+<--- Page Split --->
+
+## Example Prompt - #129094  
+
+You are a helpful assistant designed to predict the task target of a traffic crash. You need to make prediction based on the information below:  
+
+## General Information  
+
+This crash occurred in Cook County on 4/2/2022 at 0:00 o'clock. The crash happened in the city of Chicago, classified as Chicago area, on None at milepost 0.0. The roadway is classified as Unknown, and the location was identified as an Urban 2 Lane Roads. This crash was not related to an intersection.  
+
+## Infrastructure Information  
+
+The road surface was Dry with Darkness, Lighted Road lighting conditions and Clear weather at the time of the crash. The crash occurred on a Not Divided Two- way with Stop Sign in place, and it was confirmed that the crash did not occur in a work zone.  
+
+## Event Information  
+
+The crash involved 2 vehicles. The primary driver behavior in the crash was Unable to Determine, with secondary behavior was (Not Applicable).  
+
+## Unit Information  
+
+Vehicle 0, a 2014 model, was moving South and was traveling straight ahead before the crash. Vehicle 1, a 2016 model, was moving South and was traveling straight ahead before the crash. The driver was a 23- year- old male with no visible distractions, sitting in the Driver. The driver's blood alcohol content was not offered. The driver was a 59- year- old male with no visible distractions, sitting in the Driver. The driver's blood alcoho content was not offered. There was also a passenger, a 24- year- old male, seated in the Third Row Left. There was also a passenger, a 24- year- old male, seated in the third Row Right.  
+
+## Targets  
+
+Please predict the Injury number of the crash choosing from the following tokens (4 options available).  
+
+Assistant: <THREE OR MORE THAN THREE>  
+
+Please predict the Severity of the crash choosing from the following tokens (5 options available).  
+
+Assistant: <POSSIBLE INJURY>  
+
+Please predict the crash Type of the crash choosing from the following tokens (16 options available).  
+
+Assistant: <TURNING>  
+
+Extended Data Figure 2: A Crash Event Prompt Prompt Example from Illinois Dataset.
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>Extended Data Figure 3: The Confusion Matrix for TrafficSafe LLM and the Traditional Methods in (a) Washington Dataset and (b) Illinois Dataset. </center>
+
+<--- Page Split --->
+
+## One Crash Case in Washington - #EC36495  
+
+One Crash Case in Washington - #EC36495Assume the crash occurred on April 5, 2022, at 01:00 AM, in an unknown city, Mason county, on the 101 route increasing milepost direction at milepost 314.78. The location is a Rural- Principal Arterial, not at an intersection and not related to a driveway or other characteristic details are unknown. The roadway classification at the site of the accident is Rural 2 Lane Roads. The level of access control is Non Limited Access More Than Average Restriction, speed limit is 50, average annual daily traffic is 2300. The road width is 22, the road surface is made of Asphalt, the right shoulder width is 3 and there is no left shoulder, with the surface type of the left shoulder being unknown. This road is not median- separated, and there is no barrier or median width information available. The occurrence in the work zone is unknown, in conditions where the road surface was wet and the lighting was dark with no street lights. There were no pedestrians involved, 1 vehicle involved. The accident had no influence of alcohol or drugs. There was an object involved, specifically an Earth Bank or Ledge. Vehicle1 was moving north, in the direction of decreasing milepost. Vehicle2 direction and movement are unknown. The first vehicle was moving straight. The unit 1, is an unknown special vehicle type, not a commercial vehicle. The vehicle had tires punctured or blown. The driver was going straight ahead, totally ejected, and had a contributing factor of operating defective equipment. Person 1: Motor Vehicle Driver, Female, 53, Lap & Shoulder Used.  
+
+![](images/Figure_5.jpg)
+
+<center>TrafficSafe Attribution - #EC36495 </center>  
+
+Severity: <FATAL> (correct) Type: <SINGLE VEHICLE WITH OBJECT> (correct) Injury: <ONE> (correct)  
+
+Crash Severity prediction feature attribution (This crash is a FATAL crash)  
+
+Extended Data Figure 4: One Example of Sentence- based Feature Attribution Results for A Crash Resulting in Fatal in Washington Dataset.
+
+<--- Page Split --->
+![PLACEHOLDER_33_0]
+
+<center>Extended Data Figure 5: One Example of Sentence-based Feature Attribution Results for A Crash Resulting in No Apparent Injury in Illinois Dataset. </center>
+
+<--- Page Split --->
+![PLACEHOLDER_34_0]
+  
+
+Extended Data Figure 6: Data Processing Process. Four raw datasets from HSIS (crash, infrastructure, vehicle, and person data) are used to construct a prompt through four steps. (1) Data mapping and organization: Link the datasets and organize them into four parts: general, infrastructure, event, and unit. (2) Satellite image textualization: Retrieve satellite images via GPS coordinates using the Google Maps API, then employ GPT- 4o to extract text- based information. (3) Dimensionality reduction: Combine targets with similar values using GPT- 4o. (4) Prompt generation: Use the processed data from the previous steps to generate a prompt for each part.
+
+<--- Page Split --->
+![PLACEHOLDER_35_0]
+  
+
+Extended Data Figure 7: AI- expert Textualization Process. An example for the infrastructure information part of an event case in Washington dataset is shown.
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+Supplementary.pdf
+
+<--- Page Split --->

preprint/preprint__09fa1c94df81f8d0d856f1843bb02811e39dec61a630b4897174dcfd9f8b640a/images_list.json

@@ -0,0 +1,77 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1. Aberration-corrected HAADF-STEM images of (a) Ru-1, (b) Ru-2, and (c) Ru-3. (d) Ru K-edge XANES spectra of Ru foil, Ru-1, Ru-2 and \\(\\mathrm{RuO_2}\\) . (e) Ru \\(\\mathrm{k}^{3}\\) -weighted Fourier transform of the EXAFS spectra of Ru foil, Ru-1, Ru-2 and \\(\\mathrm{RuO_2}\\) . (f) CO-probe DRIFTS results of CO-absorbed Ru-1, Ru-2 and Ru-3.",
+    "footnote": [],
+    "bbox": [
+      [
+        152,
+        181,
+        840,
+        424
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Fig. 2. (a) Schematic illustration of Ru-catalysed independent N-formylation and amide hydrogenation and a one-pot sequential \\(\\mathrm{CO_2}\\) hydrogenation reaction. (b) Catalytic yield in N-formylation of morpholine over different catalysts. (c) Catalytic yield of NFM hydrogenation over different catalysts. (d) Intrinsic TOF of Ru-1, Ru-2 and Ru-3 toward morpholine N-formylation and NFM hydrogenation, respectively. (e)",
+    "footnote": [],
+    "bbox": [
+      [
+        149,
+        437,
+        844,
+        703
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3. (a) Schematic illustration of the influence of an excess \\(\\mathrm{CO_2}\\) atmosphere during the N-formylation of morpholine. (b) Catalytic conversion and selectivity of Ru-2 in the routes shown in (a). (c) \\(\\mathrm{H}_2\\) -TPR profiles of three Ru catalysts. (d) Catalytic yield in two reactions with hydrogen and deuterium and the corresponding KIE values.",
+    "footnote": [],
+    "bbox": [
+      [
+        257,
+        348,
+        732,
+        616
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Fig. 4. (a) Energy profile of the reduction process. (b) Structures of the key intermediates in the reaction pathways. The asterisk denotes the adsorption site. Colour code: Ru: green; Al: pink; N: blue; C: grey; O: red; H: white.",
+    "footnote": [],
+    "bbox": [
+      [
+        214,
+        220,
+        780,
+        571
+      ]
+    ],
+    "page_idx": 15
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Fig. 5. Proposed catalytic reaction pathways for Ru-2 in the one-pot two-step catalytic process.",
+    "footnote": [],
+    "bbox": [
+      [
+        230,
+        290,
+        760,
+        537
+      ]
+    ],
+    "page_idx": 17
+  }
+]

preprint/preprint__09fa1c94df81f8d0d856f1843bb02811e39dec61a630b4897174dcfd9f8b640a/preprint__09fa1c94df81f8d0d856f1843bb02811e39dec61a630b4897174dcfd9f8b640a.mmd

@@ -0,0 +1,353 @@
+
+# Efficient amine-assisted hydrogenation of CO2 into methanol collectively catalysed by ruthenium single sites and ensembles in a unified catalyst  
+
+Liang Chen chenliang@nimte.ac.cn  
+
+Ningbo Institute of Materials Technology and Engineering, CAS https://orcid.org/0000- 0002- 0667- 540X  
+
+Qihao Yang  
+
+Ningbo Institute of materials technology & engineering, CAS https://orcid.org/0000- 0002- 0933- 4483  
+
+Yinming Wang  
+
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+Dianhui Pan  
+
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+Desheng Su  
+
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+Hao Liu  
+
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+Qiuju Zhang  
+
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+Sheng Dai  
+
+East China University of Science and Technology https://orcid.org/0000- 0001- 5787- 0179  
+
+Ziqi Tian  
+
+Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences https://orcid.org/0000- 0001- 5667- 597X  
+
+Zhiyi Lu  
+
+Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences https://orcid.org/0000- 0002- 2117- 4101  
+
+## Article  
+
+Keywords: amine- assisted CO2 hydrogenation, supported metal catalyst, N- formylation, amide hydrogenation
+
+<--- Page Split --->
+
+Posted Date: May 2nd, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4185890/v1  
+
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on January 11th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 55837- 7.
+
+<--- Page Split --->
+
+Efficient amine-assisted hydrogenation of \(\mathrm{CO_2}\) into methanol collectively catalysed by ruthenium single sites and ensembles in a unified catalyst  
+
+Qihao Yang \(^{1,2,\ddagger}\) , Yinning Wang \(^{1,2,\ddagger}\) , Dianhui Pan \(^{1,2,\ddagger}\) , Desheng Su \(^{1,3,\ddagger}\) , Hao Liu \(^{1,2}\) , Qiuju Zhang \(^{1,2}\) , Sheng Dai \(^{4}\) , Ziqi Tian \(^{1,2,\ast}\) , Zhiyi Lu \(^{1,2,\ast}\) , Liang Chen \(^{1,2,\ast}\)  
+
+\(^{1}\) Key Laboratory of Advanced Fuel Cells and Electrolyzers Technology and Engineering, Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo, Zhejiang 315201, P. R. China \(^{2}\) University of Chinese Academy of Sciences, 100049 Beijing, P. R. China \(^{3}\) School of Materials Science and Chemical Engineering, Ningbo University, Ningbo, Zhejiang 315211, P. R. China \(^{4}\) Key Laboratory for Advanced Materials and Feringa Nobel Prize Scientist Joint Research Center, Institute of Fine Chemicals, School of Chemistry & Molecular Engineering, East China University of Science and Technology, Shanghai 200237, P. R. China  
+
+\(^{4}\) Key Laboratory for Advanced Materials and Feringa Nobel Prize Scientist Joint Research Center, Institute of Fine Chemicals, School of Chemistry & Molecular Engineering, East China University of Science and Technology, Shanghai 200237, P. R. China  
+
+KEYWORDS: amine- assisted \(\mathrm{CO_2}\) hydrogenation, supported metal catalyst, N- formylation, amide hydrogenation.
+
+<--- Page Split --->
+
+ABSTRACT: Amine-assisted sequential hydrogenation of carbon dioxide \(\mathrm{CO_2}\) , without the requirement of energy-intensive activation of inert \(\mathrm{CO_2}\) , is an efficient route to generate methanol (i.e., an essential secondary feedstock). However, this one-pot sequential hydrogenation process requires a multifunctional catalyst capable of efficiently catalysing both amine N- formylation and amide hydrogenation. Ruthenium- based catalysts possess substantial catalytic prowess in both steps of the sequential hydrogenation process, while the optimal Ru species required for the specific steps depends on distinct criteria. Herein, to optimize the two- step methanol production process, morpholine was used as a typical amine assistant, and \(\mathrm{Al_2O_3}\) - supported ruthenium catalysts (Ru- 1, Ru- 2, Ru- 3) with isolated Ru sites (Ru1) or/and Ru ensembles (Ru \(\mathrm{e}\) , including Ru clusters, \(\mathrm{Ru_c}\) , and Ru nanoparticles, \(\mathrm{Ru_p}\) ) configurations were rationally fabricated via the judicious annealing strategy. Compared to the energy- intensive \(\mathrm{CO_2}\) hydrogenation process, within the sequential catalytic system, morpholine, \(\mathrm{H_2}\) and \(\mathrm{CO_2}\) underwent an initial conversion to give N- formylmorpholine (NFM), which was subsequently hydrogenated, resulting in the low- temperature formation of methanol (120 °C). Additionally, the excellent regeneration (>99%) of morpholine guarantees a sustainable progression towards subsequent cycles. Experimental analysis and density functional theory (DFT) calculations suggest that the metallic \(\mathrm{Ru_e}\) are superior catalytic sites for the N- formylation step, whereas the isolated \(\mathrm{Ru_1}\) sites exhibit higher amide hydrogenation activity. Thus, the optimal Ru- 2 catalyst, which simultaneously features metallic \(\mathrm{Ru_c}\) and isolated \(\mathrm{Ru_1}\) sites, can efficiently synergize the conversion of \(\mathrm{CO_2}\) into methanol
+
+<--- Page Split --->
+
+in a one- pot two- step process with exceptional selectivity (>95%), highlighting the crucial sensitivity of the process to the structure of active metal species. This work not only presents an advanced catalyst suitable for \(\mathrm{CO_2}\) - based methanol production but also elucidates the requirements for rational design of multiple optimized active sites within the single catalyst for multistep catalytic reactions in the future.  
+
+## 1. Introduction  
+
+The anthropogenic emissions of greenhouse gases, primarily \(\mathrm{CO_2}\) , are widely believed to be responsible for a range of adverse environmental issues<sup>1,2</sup>. The facile catalytic reduction of carbon dioxide \(\mathrm{CO_2}\) into value- added secondary carbon feedstocks (i.e., hydrocarbons, alcohols, and carboxylic acids) exhibits excellent potential for mitigating the excessive accumulation of \(\mathrm{CO_2}^{3 - 6}\) . As one of the most attractive chemical products generated via \(\mathrm{CO_2}\) reduction, methanol \(\mathrm{(CH_3OH)}\) has a variety of attractive potential applications, including serving as a basic industrial feedstock, functioning as a liquid organic hydrogen carrier (LOHC), and being utilized in direct methanol fuel cells (DMFCs)<sup>7- 11</sup>. Therefore, the commercial value of catalytic production of \(\mathrm{CH_3OH}\) from \(\mathrm{CO_2}\) is substantial. In the past decades, remarkable researches have been conducted on the selective production of methanol via \(\mathrm{CO_2}\) hydrogenation<sup>12- 18</sup>, mainly focusing on metal oxides (e.g., \(\mathrm{In_2O_3}\) , \(\mathrm{ZnO - ZrO_2}\) , \(\mathrm{In_2O_3 - ZrO_2}\) )<sup>19- 21</sup> and metal/metal oxides (e.g., \(\mathrm{Cu/ZnO/Al_2O_3}\) , \(\mathrm{Cu/In_2O_3}\) , \(\mathrm{Cu/ZrO_2}\) , \(\mathrm{Pd/ZnO}\) )<sup>22- 28</sup>. However, the traditional \(\mathrm{CO_2}\) hydrogenation based on metal oxides is encumbered by the necessity of high catalytic temperature (>300 °C)<sup>19- 21</sup>, resulting in
+
+<--- Page Split --->
+
+excessive energy consumption. Although the introduction of metal components into metal oxides promotes the activation of \(\mathrm{H}_2\) and thus achieves enhanced catalytic performance at relatively lower temperature ( \(< 250^{\circ}\mathrm{C})^{23 - 25}\) , this modification simultaneously leads to a trade- off with the decrease in \(\mathrm{CH}_3\mathrm{OH}\) selectivity, owing to the excessive hydrogenation of \(\mathrm{CO}_2\) and the exacerbation of the reverse water gas shift (RWGS) reaction<sup>29- 31</sup>.  
+
+Compared to the traditional \(\mathrm{CO}_2\) hydrogenation process, the amine- assisted two- step hydrogenation of \(\mathrm{CO}_2\) to \(\mathrm{CH}_3\mathrm{OH}\) , involving N- formylation of amine with \(\mathrm{CO}_2\) (i.e., first step) and subsequent amide hydrogenation (i.e., second step) in the presence of homogeneous catalysts<sup>32- 35</sup>, especially ruthenium complexes<sup>32- 34</sup>, exhibits superior catalytic activity and selectivity at mild condition ( \(< 180^{\circ}\mathrm{C}\) ), thus making it to be a promising and energy- efficient alternative for methanol production. However, the inherent difficulties in separation and recycling of homogeneous catalysts, coupled with the inferior stability, hinder their widespread application and the scaling up of this amine- assisted \(\mathrm{CO}_2\) - to- MeOH route. Furthermore, for the sequential catalytic reaction systems, the theoretical requirement for achieving optimal overall catalytic performance involves the engagement of at least dual types of active sites<sup>36- 38</sup>. Therefore, the existent homogeneous catalysts with a single designated activity structure may not provide the optimum catalytic activity for both steps of the sequential \(\mathrm{CO}_2\) hydrogenation process. In this context, the immobilization of homogeneous catalysts onto stable heterogeneous matrix (e.g., \(\mathrm{Al}_2\mathrm{O}_3\) , \(\mathrm{In}_2\mathrm{O}_3\) , \(\mathrm{ZrO}_2\) ) can be the most straightforward strategy, as it offers the benefit of easy separation.
+
+<--- Page Split --->
+
+More importantly, the atomic- scale heterogeneity of supported metal catalysts is almost unavoidable, resulting in diverse structural morphologies or/and local coordination environments for the active components<sup>39,40</sup>, which offers the potential to optimize each step of the sequential \(\mathrm{CO_2}\) hydrogenation reaction and thus fine- tune the overall catalytic performance.  
+
+Based on the aforementioned considerations, we rationally synthesized a series of \(\mathrm{Al_2O_3}\) - based heterogeneous catalysts featuring isolated Ru sites (Ru<sub>1</sub>) or/and Ru ensembles (Ru<sub>e</sub>, including Ru clusters, Ru<sub>c</sub>, and Ru nanoparticles, Ru<sub>p</sub>) via a traditional impregnation method coupled with annealing treatment under different atmospheres. The catalysts (i.e., Ru- 1, Ru- 2, Ru- 3), with dominant morphological distributions of atomically dispersed Ru (Ru- 1), Ru clusters (Ru- 2) and Ru nanoparticles (Ru- 3), exhibited distinct activity and selectivity for the sequential \(\mathrm{CO_2}\) hydrogenation to \(\mathrm{CH_3OH}\) with morpholine as the amine assistant. Ru- 2, which simultaneously possesses atomically dispersed Ru<sub>1</sub> sites and metallic Ru<sub>c</sub> species, presented optimal catalytic activity in the N- formylation of morpholine with \(\mathrm{CO_2}\) , as well as the subsequent hydrogenation of the generated amide intermediates (i.e., N- formylmorpholine, NFM) to give methanol with the regeneration of morpholine. Moreover, the superior catalytic performance (selectivity of \(\mathrm{CHO_3H} > 97\%\) , regeneration of morpholine \(>99\%\) ) of Ru- 2 towards one- pot two- step \(\mathrm{CO_2}\) hydrogenation was well maintained in three consecutive cycles. The mechanistic experiments and density functional theory (DFT) results reveal that the deoxygenative hydrogenation (C=O bond cleavage) and deaminative hydrogenation (C- N cleavage)
+
+<--- Page Split --->
+
+are the rate- determining steps of morpholine N- formylation and NFM hydrogenation, which can be accelerated over \(\mathrm{Ru_e}\) and \(\mathrm{Ru_1}\) sites, respectively. This work presents a heterogeneous catalysis protocol for amine- assisted hydrogenation of \(\mathrm{CO_2}\) towards methanol production, highlighting the significance of active species heterogeneity in enhancing the catalytic performance for multistep sequential reactions.  
+
+## 2. Results and Discussion  
+
+2.1 Synthesis and characterizations of \(\mathrm{Al_2O_3}\) -based Ru catalysts  
+
+A series of \(\mathrm{Al_2O_3}\) - supported Ru catalysts (Ru- 1, Ru- 2, Ru- 3) were rationally synthesized via an impregnation- annealing method, with Ru- Macho and \(\alpha\) - \(\mathrm{Al_2O_3}\) being employed as metal precursor and support, respectively (see Section 2 in Supplementary Information (SI) for details). The powder X- ray diffraction (PXRD) patterns exhibited no diffraction peaks related to Ru or \(\mathrm{RuO_2}\) phase in the three catalysts (Supplementary Fig. 1), which might be ascribed to the low contents (Supplementary Table 1) or/and small sizes of the Ru species. Transmission electron microscopy (TEM) images show that the morphologies of the synthesized catalysts with loaded Ru species remained consistent with the original \(\mathrm{Al_2O_3}\) matrix (Supplementary Fig. 2). To identify the differences in atomic- scale structures among the three Ru catalysts, the aberration- corrected high- angle annular dark- field scanning transmission electron microscopy (HAADF- STEM) technique was adopted. The Ru species in Ru- 1 catalysts were atomically dispersed (Fig. 1a and Supplementary Fig.
+
+<--- Page Split --->
+
+3), while Ru clusters emerged in Ru- 2 and these clusters grew into larger nanoparticles in Ru- 3 (Fig. 1b- c and Supplementary Figs. 4- 5).  
+
+![](images/Figure_1.jpg)
+
+<center>Fig. 1. Aberration-corrected HAADF-STEM images of (a) Ru-1, (b) Ru-2, and (c) Ru-3. (d) Ru K-edge XANES spectra of Ru foil, Ru-1, Ru-2 and \(\mathrm{RuO_2}\) . (e) Ru \(\mathrm{k}^{3}\) -weighted Fourier transform of the EXAFS spectra of Ru foil, Ru-1, Ru-2 and \(\mathrm{RuO_2}\) . (f) CO-probe DRIFTS results of CO-absorbed Ru-1, Ru-2 and Ru-3. </center>  
+
+Additional spectroscopic characterizations were conducted to elucidate the electronic structure and coordination environment of the catalysts. The high- resolution X- ray photoelectron spectroscopy (XPS) spectra of Ru 3p in Ru- 1 revealed two prominent peaks at the binding energies of 485.0 eV (Ru 3p \(_{1 / 2}\) ) and 462.2 eV (Ru 3p \(_{3 / 2}\) ), similar to those of Ru oxides \(^{41}\) , indicating the partially oxidized valence state of Ru species in Ru- 1 (Supplementary Fig. 6). In contrast, the binding energies of Ru species in Ru- 3 at 483.7 eV (Ru 3p \(_{1 / 2}\) ) and 461.8 eV (Ru 3p \(_{3 / 2}\) ) showed a striking resemblance to the characteristic signals of metallic Ru \(^{42}\) ,
+
+<--- Page Split --->
+
+consistently supporting the presence of nanoparticles as observed via HAADF- STEM. Similarly, Ru- 2 also demonstrated an average valence state that closely aligns with metallic Ru, but slightly elevated compared to Ru- 2, implying the potential existence of a subset of Ru species in the partially oxidized valence state. In addition, the XPS spectra signals of the P species, originating from the metal precursor (i.e., Ru- Macho), are notably prominent in Ru- 1 and Ru- 2, but significantly diminished in Ru- 3 (Supplementary Fig. 7), suggesting that the P species may contribute to the formation of the Ru species with oxidized valence state.  
+
+To probe the local microstructure of Ru species with enhanced precision, X- ray absorption spectroscopy (XAS) was employed. As shown in Fig. 1d, X- ray adsorption near edge structure (XANES) analysis showed that the energy absorption thresholds of Ru- 1 and Ru- 2 were located between those of \(\mathrm{RuO_2}\) and Ru foil, but Ru- 1 aligned more closely with \(\mathrm{RuO_2}\) , illustrating an increase in the valence state from the Ru foil to Ru- 2, Ru- 1 and \(\mathrm{RuO_2}\) , which is consistent with the XPS observation results. The extended X- ray absorption fine structure (EXAFS) spectrum of Ru- 1 only exhibited one main peak at \(\sim 1.4 \AA\) , and no fingerprinting signal peak of Ru- Ru interactions \((\sim 2.3 \AA)\) cannot be observed, verifying the atomic dispersion of Ru in Ru- 1. The best fitting result of the obtained EXAFS data revealed that each Ru atom was coordinated by \(\sim 3\) O atoms and \(\sim 1\) P atom on average (Fig. 1e, Supplementary Fig. 8a and Supplementary Table 2). For the Ru- 2 catalyst, in addition to the prominent peak at \(\sim 1.4 \AA\) , a relatively weak peak at \(\sim 2.3 \AA\) that corresponds to Ru- Ru first coordination shell could be identified. The low Ru- Ru coordination number (C.N.) of \(3.0 \pm 1.0\)
+
+<--- Page Split --->
+
+determined for Ru- 2 suggested the presence of tiny Ru clusters, which agrees well with the HADDF- STEM results (Fig. 1e, Supplementary Fig. 8b and Supplementary Table 2).  
+
+To gain an in- depth understanding of the Ru species distribution in the catalysts, a CO- probe diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) experiment was conducted to acquire semiquantitative information about the proportions of the surface Ru species. The red characteristic peaks in the DRIFTS spectra (Fig. 1f, Supplementary Fig. 9), corresponding to two linearly adsorbed CO molecules on partially oxidized \(\mathrm{Ru_1}\) species, decreased in intensity with increasing size of \(\mathrm{Ru_e}\) from the Ru- 1 catalyst to the Ru- 3 catalyst, while the characteristic peaks for adsorbed CO on metallic \(\mathrm{Ru_e}\) species (blue) increased in intensity. The statistical percentages of \(\mathrm{Ru_1}\) and \(\mathrm{Ru_e}\) , determined by integrating the characteristic peaks, revealed that the \(\mathrm{Ru_1}\) contents in Ru- 1, Ru- 2 and Ru- 3 are 100%, 47.7%, and 37.2%, respectively (see Section 4 in the Supplementary Information and Supplementary Table 3). The metal dispersion (i.e., available Ru active sites) was further calculated on the basis of the CO adsorption determined from CO- pulse adsorption experiments (see Section 4 in the Supplementary Information and Supplementary Table 4). The metal dispersion of Ru- 1 (>99%), Ru- 2 (69.5%) and Ru- 3 (59.4%) also decreased with increasing proportion of Ru ensembles, which is consistent with the results of the CO- probe DRIFTS experiment.  
+
+2.2 Catalytic performance of \(\mathrm{Al_2O_3}\) -based Ru catalysts towards the morpholine- assisted sequential \(\mathrm{CO_2}\) hydrogenation
+
+<--- Page Split --->
+
+The Ru catalysts, featuring various surface metal species, exhibited exceptional catalytic performance in each reaction step (Fig. 2a). For the N- formylation of morpholine, both Ru- 2 and Ru- 3 demonstrated superior catalytic performance and intrinsic activity, with turnover frequencies (TOFs) twice that of Ru- 1 (Fig. 2b- c). In addition, the catalytic generation of amide reached a state of equilibrium after \(36\mathrm{h}\) for Ru- 2 and Ru- 3, while twice the time was required for Ru- 1 (Fig. 2b). This observation highlights the crucial role of \(\mathrm{Ru_e}\) species in N- formylation. Furthermore, Ru- 2 and Ru- 3 exhibited superior selectivity ( \(>95\%\) ) for the amide compared to Ru- 1 ( \(\sim 83\%\) ), with slight formic acid detected (Supplementary Fig. 10).  
+
+![](images/Figure_2.jpg)
+
+<center>Fig. 2. (a) Schematic illustration of Ru-catalysed independent N-formylation and amide hydrogenation and a one-pot sequential \(\mathrm{CO_2}\) hydrogenation reaction. (b) Catalytic yield in N-formylation of morpholine over different catalysts. (c) Catalytic yield of NFM hydrogenation over different catalysts. (d) Intrinsic TOF of Ru-1, Ru-2 and Ru-3 toward morpholine N-formylation and NFM hydrogenation, respectively. (e) </center>
+
+<--- Page Split --->
+
+Yield- time profile of the one- pot two- step tandem process catalysed by Ru- 2. Reaction conditions: 393 K, 10 mmol substrate, 15 ml 1,4- dioxane as a solvent, 100 mg catalyst, 0.5 mmol \(\mathrm{CsCO_3}\) ; for the N- formylation process: \(\mathrm{P(CO_2):P(H_2) = 1:1}\) with a total pressure of 4 MPa; for amide hydrogenation, \(\mathrm{P(H_2) = 4MPa}\) .  
+
+During the hydrogenation of NFM into amine and methanol, both Ru- 1 and Ru- 2 exhibited superior activity, with TOF values (Ru- 1: \(\sim 320 \mathrm{h}^{- 1}\) , Ru- 2: \(\sim 389 \mathrm{h}^{- 1}\) ) 4- 5 times higher than that of the Ru- 3 catalyst ( \(\sim 87 \mathrm{h}^{- 1}\) ). Notably, Ru- 2 showed a high yield and \(>99\%\) methanol selectivity under identical reaction conditions. Based on the exceptional performance of Ru- 2, we decided to investigate its potential utilization in a one- pot two- step tandem process. Over 144 h, Ru- 2 exhibited superb activity (methanol turnover number, \(\mathrm{TON}_{\mathrm{methanol}} = 3300\) ) and stability in three consecutive reaction cycles (Fig. 2e). The morphology of the Ru clusters was well maintained, as observed in the aberration- corrected HAADF- STEM image (Supplementary Fig. 11). We concluded that the \(\mathrm{Ru_e}\) and \(\mathrm{Ru_1}\) sites played dominant roles in the N- formylation and amide hydrogenation reactions, respectively.  
+
+2.3 Catalytic mechanism of sequential \(\mathrm{CO_2}\) hydrogenation over \(\mathrm{Al_2O_3}\) -based Ru catalysts  
+
+To confirm the catalytic reaction pathways of sequential \(\mathrm{CO_2}\) hydrogenation, a series of mechanistic experiments were performed using the optimal Ru- 2 catalyst. The influence of excess \(\mathrm{CO_2}\) during the N- formylation of morpholine was investigated, in which zwitterionic carbamates were spontaneously produced by
+
+<--- Page Split --->
+
+morpholine in an aprotic solvent (Fig. 3a). In route 2, the reaction was initiated after \(\mathrm{CO_2}\) saturation (without additional \(\mathrm{CO_2}\) input during the catalytic reaction), in contrast to the original route 1, while in route 3, the excess \(\mathrm{CO_2}\) was evacuated after \(\mathrm{CO_2}\) saturation and replaced with 2 MPa \(\mathrm{N}_2\) . The absence of excess \(\mathrm{CO_2}\) led to a higher formate selectivity with similar conversion levels (Fig. 3b), indicating the significant role of \(\mathrm{CO_2}\) in inducing the critical intermediate (i.e., zwitterionic carbamate) during the N- formylation process.  
+
+![](images/Figure_3.jpg)
+
+<center>Fig. 3. (a) Schematic illustration of the influence of an excess \(\mathrm{CO_2}\) atmosphere during the N-formylation of morpholine. (b) Catalytic conversion and selectivity of Ru-2 in the routes shown in (a). (c) \(\mathrm{H}_2\) -TPR profiles of three Ru catalysts. (d) Catalytic yield in two reactions with hydrogen and deuterium and the corresponding KIE values. </center>  
+
+The \(\mathrm{H}_2\) temperature- programmed reduction ( \(\mathrm{H}_2\) - TPR, Fig. 3c) profiles of the three catalysts revealed different \(\mathrm{H}_2\) affinities, where Ru- 2 and Ru- 3, possessing abundant \(\mathrm{Ru_e}\) sites, was prone to reduction at a lower temperature, thus manifesting the stronger
+
+<--- Page Split --->
+
+\(\mathrm{H}_{2}\) activation ability compared with Ru- 1. To investigate the influence of \(\mathrm{H}_{2}\) activation in both steps of sequential \(\mathrm{CO}_{2}\) hydrogenation, hydrogen was replaced with deuterium (Fig. 3d). The yield was controlled at the intrinsic stage (conversion \(< 20\%\) ) to calculate the kinetic isotope effect (KIE) value. Similar to the TOF calculation with different catalysts, the reaction rate of N- formylation sharply decreased when hydrogen was replaced with deuterium, with a primary KIE value of \(2.45 (2 < \mathrm{KIE} < 7)\) , indicating that the rate determining step (RDS) in the N- formylation reaction is a step in which hydrogen is involved \(^{43,44}\) . However, this phenomenon was not observed in the amide reaction process, in which nearly identical reaction rates were obtained with hydrogen and deuterium (KIE=0.98). Thus, for the amide hydrogenation reaction, the RDS was speculated to be cleavage of the C- N bond.  
+
+To further confirm this speculation, the catalytic hydrogenation of diverse carbonyl substrates was evaluated next (Supplementary Fig. 12). Due to substrate limitations, cyclohexylformamide was utilized as an analogue of NFM. The Ru- 2 catalyst exhibited almost no activity towards amides and carboxylic acids but achieved close to \(100\%\) yield for cyclohexylmethanol from aldehydes, with the absence of cyclohexylmethanamine or imine by- products, thereby highlighting the priority of C- N bond cleavage over carbonyl reduction during amide hydrogenation.  
+
+Density functional theory (DFT) calculations were further conducted to gain insights into the reduction mechanism. Two Ru- containing models were constructed to consider the \(\mathrm{Ru}_{\mathrm{e}}\) and \(\mathrm{Ru}_{1}\) sites on the \(\mathrm{Al}_{2}\mathrm{O}_{3}\) substrate, labelled Ru- ensemble and
+
+<--- Page Split --->
+
+Ru- SAC in Supplementary Fig. 13, respectively. The surface of \(\mathrm{Al}_2\mathrm{O}_3\) was passivated with a hydroxyl group, and the single Ru atom was coordinated with a \((\mathrm{CH}_3)_3\mathrm{P}\) ligand to reproduce the chemical environment determined by experiments.  
+
+![](images/Figure_4.jpg)
+
+<center>Fig. 4. (a) Energy profile of the reduction process. (b) Structures of the key intermediates in the reaction pathways. The asterisk denotes the adsorption site. Colour code: Ru: green; Al: pink; N: blue; C: grey; O: red; H: white. </center>  
+
+The free energy profile in Fig. 4a and the relevant structures of intermediates in Fig. 4b indicate that the \(\mathrm{Ru_e}\) provides multiple sites that not only activate the unsaturated carbon of the carboxylic group in carbamates by binding with oxygen but also adsorb the active hydrogen to reduce carbamates. The activation energy barrier of the first reduction step was calculated to be 1.01 eV for the \(\mathrm{C} = \mathrm{O}\) cleavage in the zwitterionic
+
+<--- Page Split --->
+
+carbamates (route 2 in Fig. 3a) under \(\mathrm{H}_2\) . Further protonation leads to elimination of the hydroxyl group, thus forming a formamide intermediate, which requires overcoming an energy barrier of \(1.18\mathrm{eV}\) . Despite the relatively low energy barriers of \(\mathrm{Ru_e}\) in the first few steps, the reduction step to form a hemiaminal (an intermediate of formamide hydrogenation) over \(\mathrm{Ru_e}\) requires overcoming a high energy barrier of \(1.45\mathrm{eV}\) , resulting in a sluggish rate to obtain the final product. However, the formamide hydrogenation may spill over onto the single Ru atom site. The migration of two hydrogen atoms from the Ru site to the intermediate would require overcoming two lower activation barriers of \(1.24\mathrm{eV}\) and \(0.95\mathrm{eV}\) . Thus, the presence of \(\mathrm{Ru_1}\) sites can accelerate the deep reduction of formamide to a hemiaminal. Finally, the hemiaminal desorbs and easily decomposes back into morpholine and formaldehyde.  
+
+Based on the results of our experiments and DFT simulations, we proposed a potential catalytic reaction pathway for Ru- 2 involving several steps (Fig. 5). Initially, morpholine absorbs \(\mathrm{CO_2}\) to yield zwitterionic carbamates. Active hydrogen species generated via metallic \(\mathrm{Ru_e}\) then reduce these carbamates to form intermediate A (Fig. 5). Proton transfer from carbamates to intermediate A leads to the formation of intermediate B, which subsequently undergoes natural intramolecular dehydration to produce amide and \(\mathrm{H}_2\mathrm{O}\) . In addition, the electronegative oxygen in intermediate A can also coordinate to the atomically dispersed \(\mathrm{Ru_1}\) site to form intermediate C, which undergoes hydrogenation to form formate and morpholine. This process is probably the primary source of formic acid by- product formation. In the amide hydrogenation step, the amide is coordinated to the electropositive \(\mathrm{Ru_1}\) site and hydrogenated to form
+
+<--- Page Split --->
+
+intermediate E, which undergoes C- N cleavage with the aid of the \(\mathrm{Ru}_1\) site to generate the adsorbed aldehyde (intermediate F) and regenerate the morpholine. The intermediate F is prone to hydrogenation (Supplementary Fig. 14), thus producing methanol. In contrast, the methylamine by- products via imine (intermediate G) pathway were not detected during amide hydrogenation (Supplementary Fig. 14).  
+
+![](images/Figure_5.jpg)
+
+<center>Fig. 5. Proposed catalytic reaction pathways for Ru-2 in the one-pot two-step catalytic process. </center>  
+
+## 3. Conclusion  
+
+In summary, we prepared a series of active heterogeneous Ru catalysts with multiple surface metal species, including atomically dispersed Ru species and Ru ensembles. Among these catalysts, Ru- 2, which contained both \(\mathrm{Ru}_1\) species and \(\mathrm{Ru}_\mathrm{e}\) sites, demonstrated excellent performance in the N- formylation and amide hydrogenation reactions, enabling efficient one- pot two- step methanol production under relatively mild conditions. The critical roles of the active metal species in the
+
+<--- Page Split --->
+
+reaction process were revealed by combining experiments and theoretical calculations, and a possible reaction pathway was proposed. This study provides a potential candidate catalyst for selective reduction of \(\mathrm{CO_2}\) to methanol and reveals the synergistic effect of different metal species in complex multistep reactions at the atomic scale. The strategy of rationally designing multiple optimized active sites within a single catalyst paves the way for enhancing the catalytic performance in various multistep sequential reactions in the future.  
+
+## ASSOCIATED CONTENT  
+
+Supplementary Information. The Supplementary Information is available free of charge via the Internet at http://pubs.acs.org.  
+
+Chemicals and characterization; Preparation of \(\mathrm{Al_2O_3}\) - based Ru catalysts; Catalyst evaluation; Characterization details; Supplementary Figs. 1- 14; Supplementary Tables 1- 4  
+
+## AUTHOR INFORMATION  
+
+## Corresponding Author  
+
+Liang Chen – Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo 315201, P. R. China; University of Chinese Academy of Sciences, Beijing 100049, P. R. China; E- mail: chenliang@nimte.ac.cn  
+
+Zhiyi Lu – Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo 315201, P. R. China; University of Chinese Academy of
+
+<--- Page Split --->
+
+Sciences, Beijing 100049, P. R. China; E- mail: luzhiyi@nimte.ac.cn  
+
+Ziqi Tian – Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo 315201, P. R. China; University of Chinese Academy of Sciences, Beijing 100049, P. R. China; E- mail: tianziqi@nimte.ac.cn  
+
+## Author Contributions  
+
+These authors contributed equally.  
+
+## Notes  
+
+The authors declare no competing financial interest.  
+
+## ACKNOWLEDGMENT  
+
+This work is supported by the National Natural Science Foundation of China (22101288), the Natural Science Foundation of Zhejiang Province (LQ22B010005 and LD21E020001), the Bellwethers Project of Zhejiang Research and Development Plan (2022C01158), the Ningbo Yongjiang Talent Introduction Programme (2021A- 036- B), the Science and Technology Innovation 2025 Program in Ningbo (2022Z205), Youth Innovation Promotion Association, CAS, the Program for Professor of Special Appointment (Eastern Scholar) at Shanghai Institutions of Higher Learning, Feringa Nobel Prize Scientist Joint Research Center, Transformational Technologies for Clean Energy and Demonstration, Strategic Priority Research Program of the Chinese Academy of Sciences (XDA21000000), DNL Cooperation Fund, CAS (Grant No. DNL202008), and “Transformational Technologies for Clean Energy and Demonstration”.
+
+<--- Page Split --->
+
+## References  
+
+1. Artz, J. et al. Sustainable conversion of carbon dioxide: an integrated review of catalysis and life cycle assessment. Chem. Rev. 118, 434-504 (2018).  
+
+2. Aresta, M., Dibenedetto, A. & Quaranta, E. State of the art and perspectives in catalytic processes for \(\mathrm{CO_2}\) conversion into chemicals and fuels: the distinctive contribution of chemical catalysis and biotechnology. J. Catal. 343, 2-45 (2016).  
+
+3. Huang, J. E. et al. \(\mathrm{CO_2}\) electrolysis to multicarbon products in strong acid. Science 372, 1074-1078 (2021).  
+
+4. Velty, A. & Corma, A. Advanced zeolite and ordered mesoporous silica-based catalysts for the conversion of \(\mathrm{CO_2}\) to chemicals and fuels. Chem Soc Rev 52, 1773-1946 (2023).  
+
+5. Yang, Y. et al. Operando studies reveal active Cu nanograins for \(\mathrm{CO_2}\) electroreduction. Nature 614, 262-269 (2023).  
+
+6. Sun, R. et al. Heterogeneous catalysts for \(\mathrm{CO_2}\) hydrogenation to formic acid/formate: from nanoscale to single atom. Energy Environ. Sci. 14, 1247-1285 (2021).  
+
+7. Bai, S. T. et al. Homogeneous and heterogeneous catalysts for hydrogenation of \(\mathrm{CO_2}\) to methanol under mild conditions. Chem. Soc. Rev. 50, 4259-4298 (2021).  
+
+8. Zhong, J. et al. State of the art and perspectives in heterogeneous catalysis of \(\mathrm{CO_2}\) hydrogenation to methanol. Chem. Soc. Rev. 49, 1385-1413 (2020).
+
+<--- Page Split --->
+
+9. Xie, Y., Hu, P., Ben-David, Y. & Milstein, D. A reversible liquid organic hydrogen carrier system based on methanol-ethylenediamine and ethylene urea. Angew. Chem. Int. Ed. 58, 5105-5109 (2019).  
+
+10. Fernández-Alvarez, F. J. & Oro, L. A. Homogeneous catalytic reduction of \(\mathrm{CO_2}\) with silicon-hydrides, state of the art. ChemCatChem, 10, 4783-4796 (2018).  
+
+11. Appel, A. M. et al. Frontiers, opportunities, and challenges in biochemical and chemical catalysis of \(\mathrm{CO_2}\) fixation. Chem. Rev., 113, 6621-6658 (2013).  
+
+12. Navarro-Jaén, S. et al. Highlights and challenges in the selective reduction of carbon dioxide to methanol. Nat. Rev. Chem. 5, 564-579 (2021).  
+
+13. Zhao, H. et al. The role of \(\mathrm{Cu_1-O_3}\) species in single-atom \(\mathrm{Cu/ZrO_2}\) catalyst for \(\mathrm{CO_2}\) hydrogenation. Nat. Catal. 5, 818-831 (2022).  
+
+14. Li, H. et al. Synergetic interaction between neighbouring platinum monomers in \(\mathrm{CO_2}\) hydrogenation. Nat. Nanotechnol. 13, 411-417 (2018).  
+
+15. An, B.; Zhang, J.; Cheng, K.; Ji, P.; Wang, C.; Lin, W., Confinement of ultrasmall \(\mathrm{Cu/ZnO_x}\) nanoparticles in metal-organic frameworks for selective methanol synthesis from catalytic hydrogenation of \(\mathrm{CO_2}\). J. Am. Chem. Soc. 139, 3834-3840 (2017).  
+
+16. Hu, J. et al. Sulfur vacancy-rich \(\mathrm{MoS_2}\) as a catalyst for the hydrogenation of \(\mathrm{CO_2}\) to methanol. Nat. Catal. 4, 242-250 (2021).  
+
+17. Lam, E. et al. \(\mathrm{CO_2}\) Hydrogenation on \(\mathrm{Cu/Al_2O_3}\): Role of the metal/support interface in driving activity and selectivity of a bifunctional catalyst. Angew. Chem. Int. Ed. 58, 13989-13996 (2019).
+
+<--- Page Split --->
+
+18. Prieto, G., Zecevic, J., Friedrich, H., De Jong, K. P. & De Jongh, P. E. Towards stable catalysts by controlling collective properties of supported metal nanoparticles. Nat. Mater. 12, 34-39 (2013).  
+
+19. Martin, O. et al. Indium oxide as a superior catalyst for methanol synthesis by \(\mathrm{CO}_{2}\) hydrogenation. Angew. Chem. Int. Ed. 55, 6261-6265 (2016).  
+
+20. Wang, J. et al. A highly selective and stable \(\mathrm{ZnO-ZrO}_{2}\) solid solution catalyst for \(\mathrm{CO}_{2}\) hydrogenation to methanol. Sci. Adv. 3, e1701290 (2017).  
+
+21. Frei, M. S. et al. Mechanism and microkinetics of methanol synthesis via \(\mathrm{CO}_{2}\) hydrogenation on indium oxide. J. Catal. 361, 313-321 (2018).  
+
+22. Beck, A. et al. Following the structure of copper-zinc-alumina across the pressure gap in carbon dioxide hydrogenation. Nat. Catal. 4, 488-497 (2021).  
+
+23. Bahruji, H. et al. Pd/ZnO catalysts for direct \(\mathrm{CO}_{2}\) hydrogenation to methanol. J. Catal. 343, 133-146 (2016).  
+
+24. Behrens, M. et al. The active site of methanol synthesis over \(\mathrm{Cu/ZnO/Al}_{2}\mathrm{O}_{3}\) industrial catalysts. Science 336, 893-897 (2012).  
+
+25. Wu, C. et al. Inverse \(\mathrm{ZrO}_{2} / \mathrm{Cu}\) as a highly efficient methanol synthesis catalyst from \(\mathrm{CO}_{2}\) hydrogenation. Nat. Commun. 11, 5767 (2020).  
+
+26. Shi, Z. et al. \(\mathrm{CO}_{2}\) hydrogenation to methanol over Cu-In intermetallic catalysts: effect of reduction temperature. J. Catal. 379, 78-89 (2019).  
+
+27. Li, K. & Chen, J. G. \(\mathrm{CO}_{2}\) hydrogenation to methanol over \(\mathrm{ZrO}_{2}\) -containing catalysts: insights into \(\mathrm{ZrO}_{2}\) induced synergy. ACS Catal. 9, 7840-7861 (2019).  
+
+28. Samson, K. et al. Influence of \(\mathrm{ZrO}_{2}\) structure and copper electronic state on
+
+<--- Page Split --->
+
+activity of \(\mathrm{Cu / ZrO_2}\) catalysts in methanol synthesis from \(\mathrm{CO_2}\) . ACS Catal. 4, 3730- 3741 (2014).  
+
+29. Kattel, S., Liu, P. & Chen, J. G. Tuning selectivity of \(\mathrm{CO_2}\) hydrogenation reactions at the metal/oxide interface. J. Am. Chem. Soc. 139, 9739-9754 (2017).  
+
+30. Studt, F. et al. Discovery of a Ni-Ga catalyst for carbon dioxide reduction to methanol. Nat. Chem. 6, 320-324 (2014).  
+
+31. Yin, Y. Z. et al. Pd@zeolitic imidazolate framework-8 derived PdZn alloy catalysts for efficient hydrogenation of \(\mathrm{CO_2}\) to methanol. Appl. Catal. B: Environ. 234, 143-152 (2018).  
+
+32. Bai, S.-T., Zhou, C., Wu, X., Sun, R. & Sels, B. Suppressing dormant Ru states in the presence of conventional metal oxides promotes the Ru-MACHO-BH-catalyzed integration of \(\mathrm{CO_2}\) capture and hydrogenation to methanol. ACS Catal. 11, 12682-12691 (2021).  
+
+33. Kar, S. et al. Mechanistic insights into ruthenium-pincer-catalyzed amine-assisted homogeneous hydrogenation of \(\mathrm{CO_2}\) to Methanol. J. Am. Chem. Soc. 141, 3160-3170 (2019).  
+
+34. Zhang, L., Han, Z., Zhao, X., Wang, Z. & Ding, K. Highly efficient ruthenium-catalyzed N-formylation of amines with \(\mathrm{H_2}\) and \(\mathrm{CO_2}\) . Angew. Chem. Int. Ed. 54, 6186-6189 (2015).  
+
+35. Jayarathne, U., Hazari, N. & Bernskoetter, W. H., Selective iron-catalyzed N-formylation of amines using dihydrogen and carbon dioxide. ACS Catal. 8, 1338-1345 (2018).
+
+<--- Page Split --->
+
+36. Zecevic, J., Vanbutsele, G., De Jong, K. P. & Martens, J. A. Nanoscale intimacy in bifunctional catalysts for selective conversion of hydrocarbons. Nature 528, 245-248 (2015).  
+
+37. Jiao, F. et al. Selective conversion of syngas to light olefins. Science 351, 1065-1068 (2016).  
+
+38. Cheng, K. et al. Direct and highly selective conversion of synthesis gas into lower olefins: design of a bifunctional catalyst combining methanol synthesis and carbon-carbon coupling. Angew. Chem. Int. Ed. 55, 4725-4728 (2016).  
+
+39. Zhang, J. et al. Importance of species heterogeneity in supported metal catalysts. J. Am. Chem. Soc. 144, 5108-5115 (2022).  
+
+40. Liu, L. & Corma, A. Identification of the active sites in supported subnanometric metal catalysts. Nat. Catal. 4, 453-456 (2021).  
+
+41. Qadir, K. Intrinsic relation between catalytic activity of CO oxidation on Ru nanoparticles and Ru oxides uncovered with ambient pressure XPS. Nano Lett. 12, 5761-5768 (2012).  
+
+42. Meng, Z. Electron-rich ruthenium on nitrogen-doped carbons promoting levulinic acid hydrogenation to \(\gamma\) -valerolactone: effect of metal-support interaction. ACS Sustainable Chem. Eng. 7, 16501-16510 (2019).  
+
+43. Li, Z. Covalent triazine framework supported non-noble metal nanoparticles with superior activity for catalytic hydrolysis of ammonia borane: from mechanistic study to catalyst design. Chem. Sci. 8, 781-788 (2017).  
+
+44. Li, L. Accelerating chemo- and regioselective hydrogenation of alkynesover
+
+<--- Page Split --->
+
+bimetallic nanoparticles in a metal- organic framework. ACS Catal. 10, 7753- 7762(2020).
+
+<--- Page Split --->
+
+## Table of Contents  
+
+A heterogeneous supported catalyst featuring atomically dispersed Ru sites and Ru- cluster sites exhibited superior catalytic performance for amine- assisted sequential hydrogenation of CO2 into hydrogen via the synergistic effect of the two types of surface- active Ru species. The rate- determining steps of the two reactions were elucidated and correlated with the intrinsic active species.  
+
+![PLACEHOLDER_26_0]
+
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SupplementaryInformation.pdf
+
+<--- Page Split --->

preprint/preprint__09fa1c94df81f8d0d856f1843bb02811e39dec61a630b4897174dcfd9f8b640a/preprint__09fa1c94df81f8d0d856f1843bb02811e39dec61a630b4897174dcfd9f8b640a_det.mmd

@@ -0,0 +1,461 @@
+<|ref|>title<|/ref|><|det|>[[44, 106, 940, 208]]<|/det|>
+# Efficient amine-assisted hydrogenation of CO2 into methanol collectively catalysed by ruthenium single sites and ensembles in a unified catalyst  
+
+<|ref|>text<|/ref|><|det|>[[44, 229, 291, 275]]<|/det|>
+Liang Chen chenliang@nimte.ac.cn  
+
+<|ref|>text<|/ref|><|det|>[[44, 301, 920, 344]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, CAS https://orcid.org/0000- 0002- 0667- 540X  
+
+<|ref|>text<|/ref|><|det|>[[44, 350, 147, 369]]<|/det|>
+Qihao Yang  
+
+<|ref|>text<|/ref|><|det|>[[44, 371, 930, 392]]<|/det|>
+Ningbo Institute of materials technology & engineering, CAS https://orcid.org/0000- 0002- 0933- 4483  
+
+<|ref|>text<|/ref|><|det|>[[44, 396, 171, 415]]<|/det|>
+Yinming Wang  
+
+<|ref|>text<|/ref|><|det|>[[44, 417, 602, 438]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+<|ref|>text<|/ref|><|det|>[[44, 443, 151, 461]]<|/det|>
+Dianhui Pan  
+
+<|ref|>text<|/ref|><|det|>[[44, 464, 602, 485]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+<|ref|>text<|/ref|><|det|>[[44, 490, 150, 508]]<|/det|>
+Desheng Su  
+
+<|ref|>text<|/ref|><|det|>[[44, 511, 602, 531]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+<|ref|>text<|/ref|><|det|>[[44, 536, 115, 554]]<|/det|>
+Hao Liu  
+
+<|ref|>text<|/ref|><|det|>[[44, 557, 602, 577]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+<|ref|>text<|/ref|><|det|>[[44, 582, 150, 601]]<|/det|>
+Qiuju Zhang  
+
+<|ref|>text<|/ref|><|det|>[[44, 603, 602, 624]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, CAS  
+
+<|ref|>text<|/ref|><|det|>[[44, 629, 135, 647]]<|/det|>
+Sheng Dai  
+
+<|ref|>text<|/ref|><|det|>[[44, 650, 839, 671]]<|/det|>
+East China University of Science and Technology https://orcid.org/0000- 0001- 5787- 0179  
+
+<|ref|>text<|/ref|><|det|>[[44, 676, 121, 694]]<|/det|>
+Ziqi Tian  
+
+<|ref|>text<|/ref|><|det|>[[44, 696, 830, 739]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences https://orcid.org/0000- 0001- 5667- 597X  
+
+<|ref|>text<|/ref|><|det|>[[44, 743, 115, 762]]<|/det|>
+Zhiyi Lu  
+
+<|ref|>text<|/ref|><|det|>[[44, 764, 830, 807]]<|/det|>
+Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences https://orcid.org/0000- 0002- 2117- 4101  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 848, 103, 866]]<|/det|>
+## Article  
+
+<|ref|>text<|/ref|><|det|>[[42, 886, 861, 929]]<|/det|>
+Keywords: amine- assisted CO2 hydrogenation, supported metal catalyst, N- formylation, amide hydrogenation
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[44, 46, 290, 65]]<|/det|>
+Posted Date: May 2nd, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 84, 475, 103]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4185890/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 120, 916, 164]]<|/det|>
+License: © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 181, 535, 201]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 236, 944, 280]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on January 11th, 2025. See the published version at https://doi.org/10.1038/s41467- 025- 55837- 7.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[147, 131, 850, 272]]<|/det|>
+Efficient amine-assisted hydrogenation of \(\mathrm{CO_2}\) into methanol collectively catalysed by ruthenium single sites and ensembles in a unified catalyst  
+
+<|ref|>text<|/ref|><|det|>[[147, 303, 850, 361]]<|/det|>
+Qihao Yang \(^{1,2,\ddagger}\) , Yinning Wang \(^{1,2,\ddagger}\) , Dianhui Pan \(^{1,2,\ddagger}\) , Desheng Su \(^{1,3,\ddagger}\) , Hao Liu \(^{1,2}\) , Qiuju Zhang \(^{1,2}\) , Sheng Dai \(^{4}\) , Ziqi Tian \(^{1,2,\ast}\) , Zhiyi Lu \(^{1,2,\ast}\) , Liang Chen \(^{1,2,\ast}\)  
+
+<|ref|>text<|/ref|><|det|>[[147, 393, 850, 600]]<|/det|>
+\(^{1}\) Key Laboratory of Advanced Fuel Cells and Electrolyzers Technology and Engineering, Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo, Zhejiang 315201, P. R. China \(^{2}\) University of Chinese Academy of Sciences, 100049 Beijing, P. R. China \(^{3}\) School of Materials Science and Chemical Engineering, Ningbo University, Ningbo, Zhejiang 315211, P. R. China \(^{4}\) Key Laboratory for Advanced Materials and Feringa Nobel Prize Scientist Joint Research Center, Institute of Fine Chemicals, School of Chemistry & Molecular Engineering, East China University of Science and Technology, Shanghai 200237, P. R. China  
+
+<|ref|>text<|/ref|><|det|>[[147, 616, 850, 748]]<|/det|>
+\(^{4}\) Key Laboratory for Advanced Materials and Feringa Nobel Prize Scientist Joint Research Center, Institute of Fine Chemicals, School of Chemistry & Molecular Engineering, East China University of Science and Technology, Shanghai 200237, P. R. China  
+
+<|ref|>text<|/ref|><|det|>[[147, 780, 849, 836]]<|/det|>
+KEYWORDS: amine- assisted \(\mathrm{CO_2}\) hydrogenation, supported metal catalyst, N- formylation, amide hydrogenation.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 90, 852, 895]]<|/det|>
+ABSTRACT: Amine-assisted sequential hydrogenation of carbon dioxide \(\mathrm{CO_2}\) , without the requirement of energy-intensive activation of inert \(\mathrm{CO_2}\) , is an efficient route to generate methanol (i.e., an essential secondary feedstock). However, this one-pot sequential hydrogenation process requires a multifunctional catalyst capable of efficiently catalysing both amine N- formylation and amide hydrogenation. Ruthenium- based catalysts possess substantial catalytic prowess in both steps of the sequential hydrogenation process, while the optimal Ru species required for the specific steps depends on distinct criteria. Herein, to optimize the two- step methanol production process, morpholine was used as a typical amine assistant, and \(\mathrm{Al_2O_3}\) - supported ruthenium catalysts (Ru- 1, Ru- 2, Ru- 3) with isolated Ru sites (Ru1) or/and Ru ensembles (Ru \(\mathrm{e}\) , including Ru clusters, \(\mathrm{Ru_c}\) , and Ru nanoparticles, \(\mathrm{Ru_p}\) ) configurations were rationally fabricated via the judicious annealing strategy. Compared to the energy- intensive \(\mathrm{CO_2}\) hydrogenation process, within the sequential catalytic system, morpholine, \(\mathrm{H_2}\) and \(\mathrm{CO_2}\) underwent an initial conversion to give N- formylmorpholine (NFM), which was subsequently hydrogenated, resulting in the low- temperature formation of methanol (120 °C). Additionally, the excellent regeneration (>99%) of morpholine guarantees a sustainable progression towards subsequent cycles. Experimental analysis and density functional theory (DFT) calculations suggest that the metallic \(\mathrm{Ru_e}\) are superior catalytic sites for the N- formylation step, whereas the isolated \(\mathrm{Ru_1}\) sites exhibit higher amide hydrogenation activity. Thus, the optimal Ru- 2 catalyst, which simultaneously features metallic \(\mathrm{Ru_c}\) and isolated \(\mathrm{Ru_1}\) sites, can efficiently synergize the conversion of \(\mathrm{CO_2}\) into methanol
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 262]]<|/det|>
+in a one- pot two- step process with exceptional selectivity (>95%), highlighting the crucial sensitivity of the process to the structure of active metal species. This work not only presents an advanced catalyst suitable for \(\mathrm{CO_2}\) - based methanol production but also elucidates the requirements for rational design of multiple optimized active sites within the single catalyst for multistep catalytic reactions in the future.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 301, 280, 318]]<|/det|>
+## 1. Introduction  
+
+<|ref|>text<|/ref|><|det|>[[147, 350, 852, 892]]<|/det|>
+The anthropogenic emissions of greenhouse gases, primarily \(\mathrm{CO_2}\) , are widely believed to be responsible for a range of adverse environmental issues<sup>1,2</sup>. The facile catalytic reduction of carbon dioxide \(\mathrm{CO_2}\) into value- added secondary carbon feedstocks (i.e., hydrocarbons, alcohols, and carboxylic acids) exhibits excellent potential for mitigating the excessive accumulation of \(\mathrm{CO_2}^{3 - 6}\) . As one of the most attractive chemical products generated via \(\mathrm{CO_2}\) reduction, methanol \(\mathrm{(CH_3OH)}\) has a variety of attractive potential applications, including serving as a basic industrial feedstock, functioning as a liquid organic hydrogen carrier (LOHC), and being utilized in direct methanol fuel cells (DMFCs)<sup>7- 11</sup>. Therefore, the commercial value of catalytic production of \(\mathrm{CH_3OH}\) from \(\mathrm{CO_2}\) is substantial. In the past decades, remarkable researches have been conducted on the selective production of methanol via \(\mathrm{CO_2}\) hydrogenation<sup>12- 18</sup>, mainly focusing on metal oxides (e.g., \(\mathrm{In_2O_3}\) , \(\mathrm{ZnO - ZrO_2}\) , \(\mathrm{In_2O_3 - ZrO_2}\) )<sup>19- 21</sup> and metal/metal oxides (e.g., \(\mathrm{Cu/ZnO/Al_2O_3}\) , \(\mathrm{Cu/In_2O_3}\) , \(\mathrm{Cu/ZrO_2}\) , \(\mathrm{Pd/ZnO}\) )<sup>22- 28</sup>. However, the traditional \(\mathrm{CO_2}\) hydrogenation based on metal oxides is encumbered by the necessity of high catalytic temperature (>300 °C)<sup>19- 21</sup>, resulting in
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 298]]<|/det|>
+excessive energy consumption. Although the introduction of metal components into metal oxides promotes the activation of \(\mathrm{H}_2\) and thus achieves enhanced catalytic performance at relatively lower temperature ( \(< 250^{\circ}\mathrm{C})^{23 - 25}\) , this modification simultaneously leads to a trade- off with the decrease in \(\mathrm{CH}_3\mathrm{OH}\) selectivity, owing to the excessive hydrogenation of \(\mathrm{CO}_2\) and the exacerbation of the reverse water gas shift (RWGS) reaction<sup>29- 31</sup>.  
+
+<|ref|>text<|/ref|><|det|>[[147, 312, 852, 894]]<|/det|>
+Compared to the traditional \(\mathrm{CO}_2\) hydrogenation process, the amine- assisted two- step hydrogenation of \(\mathrm{CO}_2\) to \(\mathrm{CH}_3\mathrm{OH}\) , involving N- formylation of amine with \(\mathrm{CO}_2\) (i.e., first step) and subsequent amide hydrogenation (i.e., second step) in the presence of homogeneous catalysts<sup>32- 35</sup>, especially ruthenium complexes<sup>32- 34</sup>, exhibits superior catalytic activity and selectivity at mild condition ( \(< 180^{\circ}\mathrm{C}\) ), thus making it to be a promising and energy- efficient alternative for methanol production. However, the inherent difficulties in separation and recycling of homogeneous catalysts, coupled with the inferior stability, hinder their widespread application and the scaling up of this amine- assisted \(\mathrm{CO}_2\) - to- MeOH route. Furthermore, for the sequential catalytic reaction systems, the theoretical requirement for achieving optimal overall catalytic performance involves the engagement of at least dual types of active sites<sup>36- 38</sup>. Therefore, the existent homogeneous catalysts with a single designated activity structure may not provide the optimum catalytic activity for both steps of the sequential \(\mathrm{CO}_2\) hydrogenation process. In this context, the immobilization of homogeneous catalysts onto stable heterogeneous matrix (e.g., \(\mathrm{Al}_2\mathrm{O}_3\) , \(\mathrm{In}_2\mathrm{O}_3\) , \(\mathrm{ZrO}_2\) ) can be the most straightforward strategy, as it offers the benefit of easy separation.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 94, 852, 260]]<|/det|>
+More importantly, the atomic- scale heterogeneity of supported metal catalysts is almost unavoidable, resulting in diverse structural morphologies or/and local coordination environments for the active components<sup>39,40</sup>, which offers the potential to optimize each step of the sequential \(\mathrm{CO_2}\) hydrogenation reaction and thus fine- tune the overall catalytic performance.  
+
+<|ref|>text<|/ref|><|det|>[[147, 275, 852, 895]]<|/det|>
+Based on the aforementioned considerations, we rationally synthesized a series of \(\mathrm{Al_2O_3}\) - based heterogeneous catalysts featuring isolated Ru sites (Ru<sub>1</sub>) or/and Ru ensembles (Ru<sub>e</sub>, including Ru clusters, Ru<sub>c</sub>, and Ru nanoparticles, Ru<sub>p</sub>) via a traditional impregnation method coupled with annealing treatment under different atmospheres. The catalysts (i.e., Ru- 1, Ru- 2, Ru- 3), with dominant morphological distributions of atomically dispersed Ru (Ru- 1), Ru clusters (Ru- 2) and Ru nanoparticles (Ru- 3), exhibited distinct activity and selectivity for the sequential \(\mathrm{CO_2}\) hydrogenation to \(\mathrm{CH_3OH}\) with morpholine as the amine assistant. Ru- 2, which simultaneously possesses atomically dispersed Ru<sub>1</sub> sites and metallic Ru<sub>c</sub> species, presented optimal catalytic activity in the N- formylation of morpholine with \(\mathrm{CO_2}\) , as well as the subsequent hydrogenation of the generated amide intermediates (i.e., N- formylmorpholine, NFM) to give methanol with the regeneration of morpholine. Moreover, the superior catalytic performance (selectivity of \(\mathrm{CHO_3H} > 97\%\) , regeneration of morpholine \(>99\%\) ) of Ru- 2 towards one- pot two- step \(\mathrm{CO_2}\) hydrogenation was well maintained in three consecutive cycles. The mechanistic experiments and density functional theory (DFT) results reveal that the deoxygenative hydrogenation (C=O bond cleavage) and deaminative hydrogenation (C- N cleavage)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 262]]<|/det|>
+are the rate- determining steps of morpholine N- formylation and NFM hydrogenation, which can be accelerated over \(\mathrm{Ru_e}\) and \(\mathrm{Ru_1}\) sites, respectively. This work presents a heterogeneous catalysis protocol for amine- assisted hydrogenation of \(\mathrm{CO_2}\) towards methanol production, highlighting the significance of active species heterogeneity in enhancing the catalytic performance for multistep sequential reactions.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 294, 367, 312]]<|/det|>
+## 2. Results and Discussion  
+
+<|ref|>text<|/ref|><|det|>[[148, 344, 664, 363]]<|/det|>
+2.1 Synthesis and characterizations of \(\mathrm{Al_2O_3}\) -based Ru catalysts  
+
+<|ref|>text<|/ref|><|det|>[[146, 395, 852, 862]]<|/det|>
+A series of \(\mathrm{Al_2O_3}\) - supported Ru catalysts (Ru- 1, Ru- 2, Ru- 3) were rationally synthesized via an impregnation- annealing method, with Ru- Macho and \(\alpha\) - \(\mathrm{Al_2O_3}\) being employed as metal precursor and support, respectively (see Section 2 in Supplementary Information (SI) for details). The powder X- ray diffraction (PXRD) patterns exhibited no diffraction peaks related to Ru or \(\mathrm{RuO_2}\) phase in the three catalysts (Supplementary Fig. 1), which might be ascribed to the low contents (Supplementary Table 1) or/and small sizes of the Ru species. Transmission electron microscopy (TEM) images show that the morphologies of the synthesized catalysts with loaded Ru species remained consistent with the original \(\mathrm{Al_2O_3}\) matrix (Supplementary Fig. 2). To identify the differences in atomic- scale structures among the three Ru catalysts, the aberration- corrected high- angle annular dark- field scanning transmission electron microscopy (HAADF- STEM) technique was adopted. The Ru species in Ru- 1 catalysts were atomically dispersed (Fig. 1a and Supplementary Fig.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 94, 850, 150]]<|/det|>
+3), while Ru clusters emerged in Ru- 2 and these clusters grew into larger nanoparticles in Ru- 3 (Fig. 1b- c and Supplementary Figs. 4- 5).  
+
+<|ref|>image<|/ref|><|det|>[[152, 181, 840, 424]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 441, 852, 572]]<|/det|>
+<center>Fig. 1. Aberration-corrected HAADF-STEM images of (a) Ru-1, (b) Ru-2, and (c) Ru-3. (d) Ru K-edge XANES spectra of Ru foil, Ru-1, Ru-2 and \(\mathrm{RuO_2}\) . (e) Ru \(\mathrm{k}^{3}\) -weighted Fourier transform of the EXAFS spectra of Ru foil, Ru-1, Ru-2 and \(\mathrm{RuO_2}\) . (f) CO-probe DRIFTS results of CO-absorbed Ru-1, Ru-2 and Ru-3. </center>  
+
+<|ref|>text<|/ref|><|det|>[[147, 603, 852, 883]]<|/det|>
+Additional spectroscopic characterizations were conducted to elucidate the electronic structure and coordination environment of the catalysts. The high- resolution X- ray photoelectron spectroscopy (XPS) spectra of Ru 3p in Ru- 1 revealed two prominent peaks at the binding energies of 485.0 eV (Ru 3p \(_{1 / 2}\) ) and 462.2 eV (Ru 3p \(_{3 / 2}\) ), similar to those of Ru oxides \(^{41}\) , indicating the partially oxidized valence state of Ru species in Ru- 1 (Supplementary Fig. 6). In contrast, the binding energies of Ru species in Ru- 3 at 483.7 eV (Ru 3p \(_{1 / 2}\) ) and 461.8 eV (Ru 3p \(_{3 / 2}\) ) showed a striking resemblance to the characteristic signals of metallic Ru \(^{42}\) ,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 373]]<|/det|>
+consistently supporting the presence of nanoparticles as observed via HAADF- STEM. Similarly, Ru- 2 also demonstrated an average valence state that closely aligns with metallic Ru, but slightly elevated compared to Ru- 2, implying the potential existence of a subset of Ru species in the partially oxidized valence state. In addition, the XPS spectra signals of the P species, originating from the metal precursor (i.e., Ru- Macho), are notably prominent in Ru- 1 and Ru- 2, but significantly diminished in Ru- 3 (Supplementary Fig. 7), suggesting that the P species may contribute to the formation of the Ru species with oxidized valence state.  
+
+<|ref|>text<|/ref|><|det|>[[147, 388, 852, 892]]<|/det|>
+To probe the local microstructure of Ru species with enhanced precision, X- ray absorption spectroscopy (XAS) was employed. As shown in Fig. 1d, X- ray adsorption near edge structure (XANES) analysis showed that the energy absorption thresholds of Ru- 1 and Ru- 2 were located between those of \(\mathrm{RuO_2}\) and Ru foil, but Ru- 1 aligned more closely with \(\mathrm{RuO_2}\) , illustrating an increase in the valence state from the Ru foil to Ru- 2, Ru- 1 and \(\mathrm{RuO_2}\) , which is consistent with the XPS observation results. The extended X- ray absorption fine structure (EXAFS) spectrum of Ru- 1 only exhibited one main peak at \(\sim 1.4 \AA\) , and no fingerprinting signal peak of Ru- Ru interactions \((\sim 2.3 \AA)\) cannot be observed, verifying the atomic dispersion of Ru in Ru- 1. The best fitting result of the obtained EXAFS data revealed that each Ru atom was coordinated by \(\sim 3\) O atoms and \(\sim 1\) P atom on average (Fig. 1e, Supplementary Fig. 8a and Supplementary Table 2). For the Ru- 2 catalyst, in addition to the prominent peak at \(\sim 1.4 \AA\) , a relatively weak peak at \(\sim 2.3 \AA\) that corresponds to Ru- Ru first coordination shell could be identified. The low Ru- Ru coordination number (C.N.) of \(3.0 \pm 1.0\)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 94, 850, 186]]<|/det|>
+determined for Ru- 2 suggested the presence of tiny Ru clusters, which agrees well with the HADDF- STEM results (Fig. 1e, Supplementary Fig. 8b and Supplementary Table 2).  
+
+<|ref|>text<|/ref|><|det|>[[147, 202, 852, 821]]<|/det|>
+To gain an in- depth understanding of the Ru species distribution in the catalysts, a CO- probe diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) experiment was conducted to acquire semiquantitative information about the proportions of the surface Ru species. The red characteristic peaks in the DRIFTS spectra (Fig. 1f, Supplementary Fig. 9), corresponding to two linearly adsorbed CO molecules on partially oxidized \(\mathrm{Ru_1}\) species, decreased in intensity with increasing size of \(\mathrm{Ru_e}\) from the Ru- 1 catalyst to the Ru- 3 catalyst, while the characteristic peaks for adsorbed CO on metallic \(\mathrm{Ru_e}\) species (blue) increased in intensity. The statistical percentages of \(\mathrm{Ru_1}\) and \(\mathrm{Ru_e}\) , determined by integrating the characteristic peaks, revealed that the \(\mathrm{Ru_1}\) contents in Ru- 1, Ru- 2 and Ru- 3 are 100%, 47.7%, and 37.2%, respectively (see Section 4 in the Supplementary Information and Supplementary Table 3). The metal dispersion (i.e., available Ru active sites) was further calculated on the basis of the CO adsorption determined from CO- pulse adsorption experiments (see Section 4 in the Supplementary Information and Supplementary Table 4). The metal dispersion of Ru- 1 (>99%), Ru- 2 (69.5%) and Ru- 3 (59.4%) also decreased with increasing proportion of Ru ensembles, which is consistent with the results of the CO- probe DRIFTS experiment.  
+
+<|ref|>text<|/ref|><|det|>[[148, 848, 849, 904]]<|/det|>
+2.2 Catalytic performance of \(\mathrm{Al_2O_3}\) -based Ru catalysts towards the morpholine- assisted sequential \(\mathrm{CO_2}\) hydrogenation
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 94, 853, 410]]<|/det|>
+The Ru catalysts, featuring various surface metal species, exhibited exceptional catalytic performance in each reaction step (Fig. 2a). For the N- formylation of morpholine, both Ru- 2 and Ru- 3 demonstrated superior catalytic performance and intrinsic activity, with turnover frequencies (TOFs) twice that of Ru- 1 (Fig. 2b- c). In addition, the catalytic generation of amide reached a state of equilibrium after \(36\mathrm{h}\) for Ru- 2 and Ru- 3, while twice the time was required for Ru- 1 (Fig. 2b). This observation highlights the crucial role of \(\mathrm{Ru_e}\) species in N- formylation. Furthermore, Ru- 2 and Ru- 3 exhibited superior selectivity ( \(>95\%\) ) for the amide compared to Ru- 1 ( \(\sim 83\%\) ), with slight formic acid detected (Supplementary Fig. 10).  
+
+<|ref|>image<|/ref|><|det|>[[149, 437, 844, 703]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 720, 853, 887]]<|/det|>
+<center>Fig. 2. (a) Schematic illustration of Ru-catalysed independent N-formylation and amide hydrogenation and a one-pot sequential \(\mathrm{CO_2}\) hydrogenation reaction. (b) Catalytic yield in N-formylation of morpholine over different catalysts. (c) Catalytic yield of NFM hydrogenation over different catalysts. (d) Intrinsic TOF of Ru-1, Ru-2 and Ru-3 toward morpholine N-formylation and NFM hydrogenation, respectively. (e) </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 225]]<|/det|>
+Yield- time profile of the one- pot two- step tandem process catalysed by Ru- 2. Reaction conditions: 393 K, 10 mmol substrate, 15 ml 1,4- dioxane as a solvent, 100 mg catalyst, 0.5 mmol \(\mathrm{CsCO_3}\) ; for the N- formylation process: \(\mathrm{P(CO_2):P(H_2) = 1:1}\) with a total pressure of 4 MPa; for amide hydrogenation, \(\mathrm{P(H_2) = 4MPa}\) .  
+
+<|ref|>text<|/ref|><|det|>[[146, 255, 853, 646]]<|/det|>
+During the hydrogenation of NFM into amine and methanol, both Ru- 1 and Ru- 2 exhibited superior activity, with TOF values (Ru- 1: \(\sim 320 \mathrm{h}^{- 1}\) , Ru- 2: \(\sim 389 \mathrm{h}^{- 1}\) ) 4- 5 times higher than that of the Ru- 3 catalyst ( \(\sim 87 \mathrm{h}^{- 1}\) ). Notably, Ru- 2 showed a high yield and \(>99\%\) methanol selectivity under identical reaction conditions. Based on the exceptional performance of Ru- 2, we decided to investigate its potential utilization in a one- pot two- step tandem process. Over 144 h, Ru- 2 exhibited superb activity (methanol turnover number, \(\mathrm{TON}_{\mathrm{methanol}} = 3300\) ) and stability in three consecutive reaction cycles (Fig. 2e). The morphology of the Ru clusters was well maintained, as observed in the aberration- corrected HAADF- STEM image (Supplementary Fig. 11). We concluded that the \(\mathrm{Ru_e}\) and \(\mathrm{Ru_1}\) sites played dominant roles in the N- formylation and amide hydrogenation reactions, respectively.  
+
+<|ref|>text<|/ref|><|det|>[[147, 678, 850, 734]]<|/det|>
+2.3 Catalytic mechanism of sequential \(\mathrm{CO_2}\) hydrogenation over \(\mathrm{Al_2O_3}\) -based Ru catalysts  
+
+<|ref|>text<|/ref|><|det|>[[147, 766, 851, 896]]<|/det|>
+To confirm the catalytic reaction pathways of sequential \(\mathrm{CO_2}\) hydrogenation, a series of mechanistic experiments were performed using the optimal Ru- 2 catalyst. The influence of excess \(\mathrm{CO_2}\) during the N- formylation of morpholine was investigated, in which zwitterionic carbamates were spontaneously produced by
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[146, 92, 852, 336]]<|/det|>
+morpholine in an aprotic solvent (Fig. 3a). In route 2, the reaction was initiated after \(\mathrm{CO_2}\) saturation (without additional \(\mathrm{CO_2}\) input during the catalytic reaction), in contrast to the original route 1, while in route 3, the excess \(\mathrm{CO_2}\) was evacuated after \(\mathrm{CO_2}\) saturation and replaced with 2 MPa \(\mathrm{N}_2\) . The absence of excess \(\mathrm{CO_2}\) led to a higher formate selectivity with similar conversion levels (Fig. 3b), indicating the significant role of \(\mathrm{CO_2}\) in inducing the critical intermediate (i.e., zwitterionic carbamate) during the N- formylation process.  
+
+<|ref|>image<|/ref|><|det|>[[257, 348, 732, 616]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[146, 631, 852, 761]]<|/det|>
+<center>Fig. 3. (a) Schematic illustration of the influence of an excess \(\mathrm{CO_2}\) atmosphere during the N-formylation of morpholine. (b) Catalytic conversion and selectivity of Ru-2 in the routes shown in (a). (c) \(\mathrm{H}_2\) -TPR profiles of three Ru catalysts. (d) Catalytic yield in two reactions with hydrogen and deuterium and the corresponding KIE values. </center>  
+
+<|ref|>text<|/ref|><|det|>[[147, 792, 852, 887]]<|/det|>
+The \(\mathrm{H}_2\) temperature- programmed reduction ( \(\mathrm{H}_2\) - TPR, Fig. 3c) profiles of the three catalysts revealed different \(\mathrm{H}_2\) affinities, where Ru- 2 and Ru- 3, possessing abundant \(\mathrm{Ru_e}\) sites, was prone to reduction at a lower temperature, thus manifesting the stronger
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 855, 485]]<|/det|>
+\(\mathrm{H}_{2}\) activation ability compared with Ru- 1. To investigate the influence of \(\mathrm{H}_{2}\) activation in both steps of sequential \(\mathrm{CO}_{2}\) hydrogenation, hydrogen was replaced with deuterium (Fig. 3d). The yield was controlled at the intrinsic stage (conversion \(< 20\%\) ) to calculate the kinetic isotope effect (KIE) value. Similar to the TOF calculation with different catalysts, the reaction rate of N- formylation sharply decreased when hydrogen was replaced with deuterium, with a primary KIE value of \(2.45 (2 < \mathrm{KIE} < 7)\) , indicating that the rate determining step (RDS) in the N- formylation reaction is a step in which hydrogen is involved \(^{43,44}\) . However, this phenomenon was not observed in the amide reaction process, in which nearly identical reaction rates were obtained with hydrogen and deuterium (KIE=0.98). Thus, for the amide hydrogenation reaction, the RDS was speculated to be cleavage of the C- N bond.  
+
+<|ref|>text<|/ref|><|det|>[[147, 515, 852, 757]]<|/det|>
+To further confirm this speculation, the catalytic hydrogenation of diverse carbonyl substrates was evaluated next (Supplementary Fig. 12). Due to substrate limitations, cyclohexylformamide was utilized as an analogue of NFM. The Ru- 2 catalyst exhibited almost no activity towards amides and carboxylic acids but achieved close to \(100\%\) yield for cyclohexylmethanol from aldehydes, with the absence of cyclohexylmethanamine or imine by- products, thereby highlighting the priority of C- N bond cleavage over carbonyl reduction during amide hydrogenation.  
+
+<|ref|>text<|/ref|><|det|>[[147, 789, 851, 882]]<|/det|>
+Density functional theory (DFT) calculations were further conducted to gain insights into the reduction mechanism. Two Ru- containing models were constructed to consider the \(\mathrm{Ru}_{\mathrm{e}}\) and \(\mathrm{Ru}_{1}\) sites on the \(\mathrm{Al}_{2}\mathrm{O}_{3}\) substrate, labelled Ru- ensemble and
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 187]]<|/det|>
+Ru- SAC in Supplementary Fig. 13, respectively. The surface of \(\mathrm{Al}_2\mathrm{O}_3\) was passivated with a hydroxyl group, and the single Ru atom was coordinated with a \((\mathrm{CH}_3)_3\mathrm{P}\) ligand to reproduce the chemical environment determined by experiments.  
+
+<|ref|>image<|/ref|><|det|>[[214, 220, 780, 571]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 589, 850, 682]]<|/det|>
+<center>Fig. 4. (a) Energy profile of the reduction process. (b) Structures of the key intermediates in the reaction pathways. The asterisk denotes the adsorption site. Colour code: Ru: green; Al: pink; N: blue; C: grey; O: red; H: white. </center>  
+
+<|ref|>text<|/ref|><|det|>[[147, 714, 852, 882]]<|/det|>
+The free energy profile in Fig. 4a and the relevant structures of intermediates in Fig. 4b indicate that the \(\mathrm{Ru_e}\) provides multiple sites that not only activate the unsaturated carbon of the carboxylic group in carbamates by binding with oxygen but also adsorb the active hydrogen to reduce carbamates. The activation energy barrier of the first reduction step was calculated to be 1.01 eV for the \(\mathrm{C} = \mathrm{O}\) cleavage in the zwitterionic
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 92, 853, 486]]<|/det|>
+carbamates (route 2 in Fig. 3a) under \(\mathrm{H}_2\) . Further protonation leads to elimination of the hydroxyl group, thus forming a formamide intermediate, which requires overcoming an energy barrier of \(1.18\mathrm{eV}\) . Despite the relatively low energy barriers of \(\mathrm{Ru_e}\) in the first few steps, the reduction step to form a hemiaminal (an intermediate of formamide hydrogenation) over \(\mathrm{Ru_e}\) requires overcoming a high energy barrier of \(1.45\mathrm{eV}\) , resulting in a sluggish rate to obtain the final product. However, the formamide hydrogenation may spill over onto the single Ru atom site. The migration of two hydrogen atoms from the Ru site to the intermediate would require overcoming two lower activation barriers of \(1.24\mathrm{eV}\) and \(0.95\mathrm{eV}\) . Thus, the presence of \(\mathrm{Ru_1}\) sites can accelerate the deep reduction of formamide to a hemiaminal. Finally, the hemiaminal desorbs and easily decomposes back into morpholine and formaldehyde.  
+
+<|ref|>text<|/ref|><|det|>[[147, 515, 852, 907]]<|/det|>
+Based on the results of our experiments and DFT simulations, we proposed a potential catalytic reaction pathway for Ru- 2 involving several steps (Fig. 5). Initially, morpholine absorbs \(\mathrm{CO_2}\) to yield zwitterionic carbamates. Active hydrogen species generated via metallic \(\mathrm{Ru_e}\) then reduce these carbamates to form intermediate A (Fig. 5). Proton transfer from carbamates to intermediate A leads to the formation of intermediate B, which subsequently undergoes natural intramolecular dehydration to produce amide and \(\mathrm{H}_2\mathrm{O}\) . In addition, the electronegative oxygen in intermediate A can also coordinate to the atomically dispersed \(\mathrm{Ru_1}\) site to form intermediate C, which undergoes hydrogenation to form formate and morpholine. This process is probably the primary source of formic acid by- product formation. In the amide hydrogenation step, the amide is coordinated to the electropositive \(\mathrm{Ru_1}\) site and hydrogenated to form
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 852, 262]]<|/det|>
+intermediate E, which undergoes C- N cleavage with the aid of the \(\mathrm{Ru}_1\) site to generate the adsorbed aldehyde (intermediate F) and regenerate the morpholine. The intermediate F is prone to hydrogenation (Supplementary Fig. 14), thus producing methanol. In contrast, the methylamine by- products via imine (intermediate G) pathway were not detected during amide hydrogenation (Supplementary Fig. 14).  
+
+<|ref|>image<|/ref|><|det|>[[230, 290, 760, 537]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[147, 553, 850, 608]]<|/det|>
+<center>Fig. 5. Proposed catalytic reaction pathways for Ru-2 in the one-pot two-step catalytic process. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 643, 267, 660]]<|/det|>
+## 3. Conclusion  
+
+<|ref|>text<|/ref|><|det|>[[147, 692, 852, 896]]<|/det|>
+In summary, we prepared a series of active heterogeneous Ru catalysts with multiple surface metal species, including atomically dispersed Ru species and Ru ensembles. Among these catalysts, Ru- 2, which contained both \(\mathrm{Ru}_1\) species and \(\mathrm{Ru}_\mathrm{e}\) sites, demonstrated excellent performance in the N- formylation and amide hydrogenation reactions, enabling efficient one- pot two- step methanol production under relatively mild conditions. The critical roles of the active metal species in the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 853, 336]]<|/det|>
+reaction process were revealed by combining experiments and theoretical calculations, and a possible reaction pathway was proposed. This study provides a potential candidate catalyst for selective reduction of \(\mathrm{CO_2}\) to methanol and reveals the synergistic effect of different metal species in complex multistep reactions at the atomic scale. The strategy of rationally designing multiple optimized active sites within a single catalyst paves the way for enhancing the catalytic performance in various multistep sequential reactions in the future.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 367, 386, 386]]<|/det|>
+## ASSOCIATED CONTENT  
+
+<|ref|>text<|/ref|><|det|>[[148, 418, 850, 475]]<|/det|>
+Supplementary Information. The Supplementary Information is available free of charge via the Internet at http://pubs.acs.org.  
+
+<|ref|>text<|/ref|><|det|>[[147, 506, 851, 600]]<|/det|>
+Chemicals and characterization; Preparation of \(\mathrm{Al_2O_3}\) - based Ru catalysts; Catalyst evaluation; Characterization details; Supplementary Figs. 1- 14; Supplementary Tables 1- 4  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 632, 395, 650]]<|/det|>
+## AUTHOR INFORMATION  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 678, 350, 695]]<|/det|>
+## Corresponding Author  
+
+<|ref|>text<|/ref|><|det|>[[147, 716, 853, 810]]<|/det|>
+Liang Chen – Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo 315201, P. R. China; University of Chinese Academy of Sciences, Beijing 100049, P. R. China; E- mail: chenliang@nimte.ac.cn  
+
+<|ref|>text<|/ref|><|det|>[[147, 842, 850, 898]]<|/det|>
+Zhiyi Lu – Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo 315201, P. R. China; University of Chinese Academy of
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 94, 704, 112]]<|/det|>
+Sciences, Beijing 100049, P. R. China; E- mail: luzhiyi@nimte.ac.cn  
+
+<|ref|>text<|/ref|><|det|>[[147, 145, 852, 240]]<|/det|>
+Ziqi Tian – Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo 315201, P. R. China; University of Chinese Academy of Sciences, Beijing 100049, P. R. China; E- mail: tianziqi@nimte.ac.cn  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 265, 340, 282]]<|/det|>
+## Author Contributions  
+
+<|ref|>text<|/ref|><|det|>[[149, 305, 439, 323]]<|/det|>
+These authors contributed equally.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 357, 200, 373]]<|/det|>
+## Notes  
+
+<|ref|>text<|/ref|><|det|>[[148, 397, 570, 415]]<|/det|>
+The authors declare no competing financial interest.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 448, 364, 465]]<|/det|>
+## ACKNOWLEDGMENT  
+
+<|ref|>text<|/ref|><|det|>[[147, 483, 852, 911]]<|/det|>
+This work is supported by the National Natural Science Foundation of China (22101288), the Natural Science Foundation of Zhejiang Province (LQ22B010005 and LD21E020001), the Bellwethers Project of Zhejiang Research and Development Plan (2022C01158), the Ningbo Yongjiang Talent Introduction Programme (2021A- 036- B), the Science and Technology Innovation 2025 Program in Ningbo (2022Z205), Youth Innovation Promotion Association, CAS, the Program for Professor of Special Appointment (Eastern Scholar) at Shanghai Institutions of Higher Learning, Feringa Nobel Prize Scientist Joint Research Center, Transformational Technologies for Clean Energy and Demonstration, Strategic Priority Research Program of the Chinese Academy of Sciences (XDA21000000), DNL Cooperation Fund, CAS (Grant No. DNL202008), and “Transformational Technologies for Clean Energy and Demonstration”.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[147, 95, 245, 111]]<|/det|>
+## References  
+
+<|ref|>text<|/ref|><|det|>[[145, 130, 852, 196]]<|/det|>
+1. Artz, J. et al. Sustainable conversion of carbon dioxide: an integrated review of catalysis and life cycle assessment. Chem. Rev. 118, 434-504 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[147, 206, 851, 299]]<|/det|>
+2. Aresta, M., Dibenedetto, A. & Quaranta, E. State of the art and perspectives in catalytic processes for \(\mathrm{CO_2}\) conversion into chemicals and fuels: the distinctive contribution of chemical catalysis and biotechnology. J. Catal. 343, 2-45 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[147, 316, 849, 372]]<|/det|>
+3. Huang, J. E. et al. \(\mathrm{CO_2}\) electrolysis to multicarbon products in strong acid. Science 372, 1074-1078 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 389, 851, 483]]<|/det|>
+4. Velty, A. & Corma, A. Advanced zeolite and ordered mesoporous silica-based catalysts for the conversion of \(\mathrm{CO_2}\) to chemicals and fuels. Chem Soc Rev 52, 1773-1946 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[147, 500, 851, 556]]<|/det|>
+5. Yang, Y. et al. Operando studies reveal active Cu nanograins for \(\mathrm{CO_2}\) electroreduction. Nature 614, 262-269 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[147, 574, 851, 667]]<|/det|>
+6. Sun, R. et al. Heterogeneous catalysts for \(\mathrm{CO_2}\) hydrogenation to formic acid/formate: from nanoscale to single atom. Energy Environ. Sci. 14, 1247-1285 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 686, 851, 741]]<|/det|>
+7. Bai, S. T. et al. Homogeneous and heterogeneous catalysts for hydrogenation of \(\mathrm{CO_2}\) to methanol under mild conditions. Chem. Soc. Rev. 50, 4259-4298 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 760, 851, 816]]<|/det|>
+8. Zhong, J. et al. State of the art and perspectives in heterogeneous catalysis of \(\mathrm{CO_2}\) hydrogenation to methanol. Chem. Soc. Rev. 49, 1385-1413 (2020).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 850, 187]]<|/det|>
+9. Xie, Y., Hu, P., Ben-David, Y. & Milstein, D. A reversible liquid organic hydrogen carrier system based on methanol-ethylenediamine and ethylene urea. Angew. Chem. Int. Ed. 58, 5105-5109 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[148, 205, 850, 260]]<|/det|>
+10. Fernández-Alvarez, F. J. & Oro, L. A. Homogeneous catalytic reduction of \(\mathrm{CO_2}\) with silicon-hydrides, state of the art. ChemCatChem, 10, 4783-4796 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[148, 279, 850, 334]]<|/det|>
+11. Appel, A. M. et al. Frontiers, opportunities, and challenges in biochemical and chemical catalysis of \(\mathrm{CO_2}\) fixation. Chem. Rev., 113, 6621-6658 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[148, 353, 850, 408]]<|/det|>
+12. Navarro-Jaén, S. et al. Highlights and challenges in the selective reduction of carbon dioxide to methanol. Nat. Rev. Chem. 5, 564-579 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[148, 427, 850, 482]]<|/det|>
+13. Zhao, H. et al. The role of \(\mathrm{Cu_1-O_3}\) species in single-atom \(\mathrm{Cu/ZrO_2}\) catalyst for \(\mathrm{CO_2}\) hydrogenation. Nat. Catal. 5, 818-831 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[148, 501, 850, 556]]<|/det|>
+14. Li, H. et al. Synergetic interaction between neighbouring platinum monomers in \(\mathrm{CO_2}\) hydrogenation. Nat. Nanotechnol. 13, 411-417 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[148, 575, 851, 705]]<|/det|>
+15. An, B.; Zhang, J.; Cheng, K.; Ji, P.; Wang, C.; Lin, W., Confinement of ultrasmall \(\mathrm{Cu/ZnO_x}\) nanoparticles in metal-organic frameworks for selective methanol synthesis from catalytic hydrogenation of \(\mathrm{CO_2}\). J. Am. Chem. Soc. 139, 3834-3840 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[148, 724, 850, 779]]<|/det|>
+16. Hu, J. et al. Sulfur vacancy-rich \(\mathrm{MoS_2}\) as a catalyst for the hydrogenation of \(\mathrm{CO_2}\) to methanol. Nat. Catal. 4, 242-250 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[148, 798, 850, 890]]<|/det|>
+17. Lam, E. et al. \(\mathrm{CO_2}\) Hydrogenation on \(\mathrm{Cu/Al_2O_3}\): Role of the metal/support interface in driving activity and selectivity of a bifunctional catalyst. Angew. Chem. Int. Ed. 58, 13989-13996 (2019).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 93, 850, 187]]<|/det|>
+18. Prieto, G., Zecevic, J., Friedrich, H., De Jong, K. P. & De Jongh, P. E. Towards stable catalysts by controlling collective properties of supported metal nanoparticles. Nat. Mater. 12, 34-39 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[148, 205, 850, 260]]<|/det|>
+19. Martin, O. et al. Indium oxide as a superior catalyst for methanol synthesis by \(\mathrm{CO}_{2}\) hydrogenation. Angew. Chem. Int. Ed. 55, 6261-6265 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[147, 279, 850, 335]]<|/det|>
+20. Wang, J. et al. A highly selective and stable \(\mathrm{ZnO-ZrO}_{2}\) solid solution catalyst for \(\mathrm{CO}_{2}\) hydrogenation to methanol. Sci. Adv. 3, e1701290 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[147, 354, 850, 409]]<|/det|>
+21. Frei, M. S. et al. Mechanism and microkinetics of methanol synthesis via \(\mathrm{CO}_{2}\) hydrogenation on indium oxide. J. Catal. 361, 313-321 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[147, 427, 850, 483]]<|/det|>
+22. Beck, A. et al. Following the structure of copper-zinc-alumina across the pressure gap in carbon dioxide hydrogenation. Nat. Catal. 4, 488-497 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 501, 849, 556]]<|/det|>
+23. Bahruji, H. et al. Pd/ZnO catalysts for direct \(\mathrm{CO}_{2}\) hydrogenation to methanol. J. Catal. 343, 133-146 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[147, 575, 850, 630]]<|/det|>
+24. Behrens, M. et al. The active site of methanol synthesis over \(\mathrm{Cu/ZnO/Al}_{2}\mathrm{O}_{3}\) industrial catalysts. Science 336, 893-897 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[147, 649, 850, 704]]<|/det|>
+25. Wu, C. et al. Inverse \(\mathrm{ZrO}_{2} / \mathrm{Cu}\) as a highly efficient methanol synthesis catalyst from \(\mathrm{CO}_{2}\) hydrogenation. Nat. Commun. 11, 5767 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[147, 723, 850, 778]]<|/det|>
+26. Shi, Z. et al. \(\mathrm{CO}_{2}\) hydrogenation to methanol over Cu-In intermetallic catalysts: effect of reduction temperature. J. Catal. 379, 78-89 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[147, 797, 850, 852]]<|/det|>
+27. Li, K. & Chen, J. G. \(\mathrm{CO}_{2}\) hydrogenation to methanol over \(\mathrm{ZrO}_{2}\) -containing catalysts: insights into \(\mathrm{ZrO}_{2}\) induced synergy. ACS Catal. 9, 7840-7861 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[147, 871, 850, 889]]<|/det|>
+28. Samson, K. et al. Influence of \(\mathrm{ZrO}_{2}\) structure and copper electronic state on
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[181, 93, 849, 149]]<|/det|>
+activity of \(\mathrm{Cu / ZrO_2}\) catalysts in methanol synthesis from \(\mathrm{CO_2}\) . ACS Catal. 4, 3730- 3741 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[147, 167, 850, 223]]<|/det|>
+29. Kattel, S., Liu, P. & Chen, J. G. Tuning selectivity of \(\mathrm{CO_2}\) hydrogenation reactions at the metal/oxide interface. J. Am. Chem. Soc. 139, 9739-9754 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[147, 242, 850, 297]]<|/det|>
+30. Studt, F. et al. Discovery of a Ni-Ga catalyst for carbon dioxide reduction to methanol. Nat. Chem. 6, 320-324 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[147, 315, 850, 409]]<|/det|>
+31. Yin, Y. Z. et al. Pd@zeolitic imidazolate framework-8 derived PdZn alloy catalysts for efficient hydrogenation of \(\mathrm{CO_2}\) to methanol. Appl. Catal. B: Environ. 234, 143-152 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[147, 427, 851, 558]]<|/det|>
+32. Bai, S.-T., Zhou, C., Wu, X., Sun, R. & Sels, B. Suppressing dormant Ru states in the presence of conventional metal oxides promotes the Ru-MACHO-BH-catalyzed integration of \(\mathrm{CO_2}\) capture and hydrogenation to methanol. ACS Catal. 11, 12682-12691 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 575, 850, 668]]<|/det|>
+33. Kar, S. et al. Mechanistic insights into ruthenium-pincer-catalyzed amine-assisted homogeneous hydrogenation of \(\mathrm{CO_2}\) to Methanol. J. Am. Chem. Soc. 141, 3160-3170 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[147, 686, 850, 780]]<|/det|>
+34. Zhang, L., Han, Z., Zhao, X., Wang, Z. & Ding, K. Highly efficient ruthenium-catalyzed N-formylation of amines with \(\mathrm{H_2}\) and \(\mathrm{CO_2}\) . Angew. Chem. Int. Ed. 54, 6186-6189 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[147, 798, 850, 890]]<|/det|>
+35. Jayarathne, U., Hazari, N. & Bernskoetter, W. H., Selective iron-catalyzed N-formylation of amines using dihydrogen and carbon dioxide. ACS Catal. 8, 1338-1345 (2018).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 93, 850, 186]]<|/det|>
+36. Zecevic, J., Vanbutsele, G., De Jong, K. P. & Martens, J. A. Nanoscale intimacy in bifunctional catalysts for selective conversion of hydrocarbons. Nature 528, 245-248 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[147, 205, 849, 260]]<|/det|>
+37. Jiao, F. et al. Selective conversion of syngas to light olefins. Science 351, 1065-1068 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[147, 279, 850, 373]]<|/det|>
+38. Cheng, K. et al. Direct and highly selective conversion of synthesis gas into lower olefins: design of a bifunctional catalyst combining methanol synthesis and carbon-carbon coupling. Angew. Chem. Int. Ed. 55, 4725-4728 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[147, 391, 848, 446]]<|/det|>
+39. Zhang, J. et al. Importance of species heterogeneity in supported metal catalysts. J. Am. Chem. Soc. 144, 5108-5115 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[147, 465, 850, 520]]<|/det|>
+40. Liu, L. & Corma, A. Identification of the active sites in supported subnanometric metal catalysts. Nat. Catal. 4, 453-456 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[147, 539, 850, 630]]<|/det|>
+41. Qadir, K. Intrinsic relation between catalytic activity of CO oxidation on Ru nanoparticles and Ru oxides uncovered with ambient pressure XPS. Nano Lett. 12, 5761-5768 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[147, 650, 850, 742]]<|/det|>
+42. Meng, Z. Electron-rich ruthenium on nitrogen-doped carbons promoting levulinic acid hydrogenation to \(\gamma\) -valerolactone: effect of metal-support interaction. ACS Sustainable Chem. Eng. 7, 16501-16510 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[147, 761, 850, 854]]<|/det|>
+43. Li, Z. Covalent triazine framework supported non-noble metal nanoparticles with superior activity for catalytic hydrolysis of ammonia borane: from mechanistic study to catalyst design. Chem. Sci. 8, 781-788 (2017).  
+
+<|ref|>text<|/ref|><|det|>[[147, 873, 850, 891]]<|/det|>
+44. Li, L. Accelerating chemo- and regioselective hydrogenation of alkynesover
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[181, 94, 850, 150]]<|/det|>
+bimetallic nanoparticles in a metal- organic framework. ACS Catal. 10, 7753- 7762(2020).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[420, 94, 576, 111]]<|/det|>
+## Table of Contents  
+
+<|ref|>text<|/ref|><|det|>[[147, 144, 852, 312]]<|/det|>
+A heterogeneous supported catalyst featuring atomically dispersed Ru sites and Ru- cluster sites exhibited superior catalytic performance for amine- assisted sequential hydrogenation of CO2 into hydrogen via the synergistic effect of the two types of surface- active Ru species. The rate- determining steps of the two reactions were elucidated and correlated with the intrinsic active species.  
+
+<|ref|>image<|/ref|><|det|>[[330, 336, 666, 444]]<|/det|>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[42, 93, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 353, 150]]<|/det|>
+SupplementaryInformation.pdf
+
+<--- Page Split --->

preprint/preprint__0a029bedc77ee23f09bec94b1dd73e6a81b91ac30d9e51c5375983d20db515bb/images_list.json

@@ -0,0 +1,62 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "FIG. 1. a) Non-primitive hexagonal unit cell of the \\(\\mathrm{V_2O_3}\\) high-temperature rhombohedral metallic phase. b) Schematic of the rhombohedral-to-monoclinic distortion along each of the three equivalent hexagonal axes. c) PEEM experimental setup. X-ray radiation, with tunable energy resonant with the vanadium \\(\\mathrm{L_{2,3}}\\) edge, impinges on the sample surface and the emitted electrons are collected and imaged through electrostatic and magnetic lenses. The \\(\\mathrm{V_2O_3}\\) film is coated with gold metal electrodes, allowing to drive a current through the device (see sketch of a typical resistive switching current-voltage curve in the bottom panel) while simultaneously acquiring XLD-PEEM images.",
+    "footnote": [],
+    "bbox": [
+      [
+        102,
+        102,
+        888,
+        430
+      ]
+    ],
+    "page_idx": 3
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "FIG. 2. XLD-PEEM images before (a) and during (b-j) the application of an electric current at \\(T = 120 \\mathrm{K}\\) . The homogeneous regions at the top and bottom of each image are the gold electrodes. The area in between is the exposed \\(\\mathrm{V}_2\\mathrm{O}_3\\) antiferromagnetic monoclinic phase, exhibiting a striped domain nanotexture. For currents larger than \\(1.5 \\mathrm{mA}\\) , the striped domains disappear in the region delimited by the white dashed lines, demonstrating the appearance of a rhombohedral metallic filament, which widens as the current is increased.",
+    "footnote": [],
+    "bbox": [
+      [
+        520,
+        105,
+        900,
+        627
+      ]
+    ],
+    "page_idx": 4
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "FIG. 3. a) Detail of the XLD-PEEM image shown in Fig. 2 in the region where the metallic filament is formed upon the application of a current above the threshold \\(I_{th}\\) . b) Schematic of the monoclinic domains crossing at \\(60^{\\circ}\\) and forming a topological defect. Blue, red and yellow areas identify the three possible monoclinic domains corresponding to the three equivalent order parameter directions \\(\\hat{\\epsilon}_{n}\\) . The order parameter at the boundaries between different domains is oriented along \\(\\hat{\\epsilon}_{1} + \\hat{\\epsilon}_{2}\\) ( \\(2\\pi /3\\) ) for the red-blue interface and along \\(\\hat{\\epsilon}_{2} + \\hat{\\epsilon}_{3}\\) ( \\(4\\pi /3\\) ) for the blue-yellow interface. The mixed red-yellow triangular region indicates the local suppression of the strain at the topological defect. The energy functionals shown on the left and right, illustrate how a topological defect (green plot, solid line) decreases the insulator-metal energy difference, \\(\\Delta\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        113,
+        102,
+        460,
+        325
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "FIG. 4. a) I-V curve measured during the XLD-PEEM imaging. The sudden drop in the voltage measured at \\(I_{th} = 1.05 \\mathrm{mA}\\) indicates the first resistive switch. b) Line profiles of the XLD-PEEM images in Fig. 2. The grey shaded area indicates the progressive widening of the metallic rhombohedral filament. The direction of the line profiles is shown by the white dashed line in the XLD-PEEM image on top, where we report a detail of Fig. 2j. c) Width, \\(d\\) , of the metallic filament as a function of current. The blue/red markers represent the values of \\(d\\) obtained from the XLD-PEEM images below/above \\(I_{th}\\) . The green solid line shows an estimate of \\(d\\) , derived from a parallel resistors model predicting a sudden jump of \\(d\\) to \\(200 \\mathrm{mA}\\) in \\(I_{th}\\) (see Supplementary Information Section S5).",
+    "footnote": [],
+    "bbox": [
+      [
+        220,
+        100,
+        797,
+        573
+      ]
+    ],
+    "page_idx": 6
+  }
+]

preprint/preprint__0a029bedc77ee23f09bec94b1dd73e6a81b91ac30d9e51c5375983d20db515bb/preprint__0a029bedc77ee23f09bec94b1dd73e6a81b91ac30d9e51c5375983d20db515bb.mmd

@@ -0,0 +1,181 @@
+
+# Mott resistive switching initiated by topological defects  
+
+Claudio Giannetti  
+
+claudio.giannetti@unicatt.it  
+
+Università Cattolica del Sacro Cuore https://orcid.org/0000- 0003- 2664- 9492  
+
+Alessandra Milloch Università Cattolica del Sacro Cuore  
+
+Ignacio Figueruelo- Campanero IMDEA Nanociencia  
+
+Wei- Fan Hsu KU Leuven  
+
+Selene Mor Università Cattolica del Sacro Cuore  
+
+Simon Mellaerts KU Leuven https://orcid.org/0000- 0002- 6715- 3066  
+
+Francesco Maccherozzi  
+
+Diamond Light Source, Chilton, Didcot, Oxfordshire, OX11 0DE, UK. https://orcid.org/0000- 0003- 4074- 2319  
+
+Larissa Ishibe Veiga Diamond Light Source  
+
+Sarnjeet Dhesi Diamond Light Source https://orcid.org/0000- 0003- 4966- 0002  
+
+Mauro Spera Università Cattolica del Sacro Cuore https://orcid.org/0000- 0001- 9041- 364X  
+
+Jin Seo KU Leuven https://orcid.org/0000- 0003- 4937- 0769  
+
+Jean- Pierre Locquet https://orcid.org/0000- 0002- 4214- 7081  
+
+Michele Fabrizio International School for Advanced Studies https://orcid.org/0000- 0002- 2943- 3278  
+
+Mariela Menghini IMDEA Nanoscience https://orcid.org/0000- 0002- 1744- 798X
+
+<--- Page Split --->
+
+## Article  
+
+## Keywords:  
+
+Posted Date: June 6th, 2024  
+
+DOI: https://doi.org/10.21203/rs.3.rs- 4019377/v1  
+
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+Additional Declarations: There is NO Competing Interest.  
+
+Version of Record: A version of this preprint was published at Nature Communications on October 31st, 2024. See the published version at https://doi.org/10.1038/s41467- 024- 53726- z.
+
+<--- Page Split --->
+
+# Mott resistive switching initiated by topological defects  
+
+Alessandra Milloch, \(^{1,2,3,*}\) Ignacio Figueroelo- Campanero, \(^{4,5,\dagger}\) Wei- Fan Hsu, \(^{3}\) Selene Mor, \(^{1,2}\) Simon Mellaerts, \(^{3}\) Francesco Maccherozzi, \(^{6}\) Larissa Ishibe Veiga, \(^{6}\) Sarnjeet S. Dhesi, \(^{6}\) Mauro Spera, \(^{1}\) Jin Won Seo, \(^{7}\) Jean- Pierre Locquet, \(^{3}\) Michele Fabrizio, \(^{8}\) Mariela Menghini, \(^{4}\) and Claudio Giannetti \(^{1,2,9,\ddagger}\)  
+
+\(^{1}\) Department of Mathematics and Physics, Università Cattolica del Sacro Cuore, Brescia I- 25133, Italy \(^{2}\) ILAMP (Interdisciplinary Laboratories for Advanced Materials Physics), Università Cattolica del Sacro Cuore, Brescia I- 25133, Italy \(^{3}\) Department of Physics and Astronomy, KU Leuven, B- 3001 Leuven, Belgium \(^{4}\) IMDEA Nanociencia, Cantoblanco, 28049 Madrid, Spain \(^{5}\) Facultad Ciencias Físicas, Universidad Complutense, 28040 Madrid, Spain \(^{6}\) Diamond Light Source, Didcot, Oxfordshire OX11 0DE, UK \(^{7}\) Department of Materials Engineering, KU Leuven, 3001 Leuven, Belgium \(^{8}\) Scuola Internazionale Superiore di Studi Avanzati (SISSA), Via Bonomea 265, 34136 Trieste, Italy \(^{9}\) CNR- INO (National Institute of Optics), via Branze 45, 25123 Brescia, Italy  
+
+Avalanche resistive switching is the fundamental process that triggers the sudden change of the electrical properties in solid- state devices under the action of intense electric fields [1]. Despite its relevance for information processing, ultrafast electronics, neuromorphic devices, resistive memories and brain- inspired computation [1- 14], the nature of the local stochastic fluctuations that drive the formation of metallic regions within the insulating state has remained hidden.  
+
+Here, using operando X- ray nano- imaging, we have captured the origin of resistive switching in a \(\mathrm{V}_2\mathrm{O}_3\) - based device under working conditions. \(\mathrm{V}_2\mathrm{O}_3\) is a paradigmatic Mott material [3], which undergoes a first- order metal- to- insulator phase transition together with a lattice transformation that breaks the threefold rotational symmetry of the rhombohedral metallic phase [2, 5, 6, 8- 11, 15]. We reveal a new class of volatile electronic switching triggered by nanoscale topological defects appearing in the shear- strain based order parameter that describes the insulating phase. Our results pave the way to the use of strain engineering approaches to manipulate such topological defects and achieve the full dynamical control of the electronic Mott switching. Topology- driven, reversible electronic transitions are relevant across a broad range of quantum materials, comprising transition metal oxides, chalcogenides and kagome metals.  
+
+The insulator- to- metal transition (IMT) in Mott materials is a key mechanism for the development of next generation Mottronic devices [3, 13]. The intrinsic correlated nature of the Mott insulating state makes these systems fragile to external stimuli [16, 17], such as the application of an electric field, which can drive the collapse of the electronic band structure and the sudden release of a large number of free carriers [18, 19]. At the macroscopic level, this phenomenon manifests in the resistive switching process, i.e., a sharp increase of the current flow when the applied voltage exceeds a threshold value [6, 20- 27]. This strong non- linearity triggered many efforts to develop neuromorphic building blocks for the hardware implementation of neural networks [14] or for ultrafast volatile and non- volatile memories or processors [12, 28, 29]. The state- of- the- art macroscopic models [30] are based on resistor networks that consider interconnected nodes transforming from the insulating to metallic state in the presence of an electric field. Above a certain threshold, a percolative, avalanche transition takes place,  
+
+thus leading to the formation of conductive filaments and the consequent sudden drop in resistivity [22, 31].  
+
+The full control and exploitation of this process is currently prevented by a limited knowledge of the early- stage firing dynamics. Microscopically, little is known about the nature of the nanoscale regions that trigger the avalanche process. Also the relation between the electronic and structural properties of the switched regions and those of the pristine insulating template is a matter of debate. Pioneering optical microscopy experiments captured the real- time formation of macroscopic metallic channels [4, 32- 35], but lacked the resolution and sensitivity to address the microscopic origin of the switching process.  
+
+Here, we adopt resonant X- ray microscopy to record nanoscale snapshots of the switching dynamics in a \(\mathrm{V}_2\mathrm{O}_3\) - based nanodevice during the application of an electric field. The results unveil the fundamental role played by the order parameter topology of the underlying lattice nanotexture. The breaking of the \(C_3\) symmetry upon transition to the insulating monoclinic phase leads to the formation of three twin shear- strain domains with boundaries oriented along the three hexagonal directions [36, 37]. The geometrical constraints then produce shear- strain topological defects at the corners of monoclinic domains crossing with an angle of \(60^{\circ}\) . These nanoscale
+
+<--- Page Split --->
+![](images/Figure_1.jpg)
+
+<center>FIG. 1. a) Non-primitive hexagonal unit cell of the \(\mathrm{V_2O_3}\) high-temperature rhombohedral metallic phase. b) Schematic of the rhombohedral-to-monoclinic distortion along each of the three equivalent hexagonal axes. c) PEEM experimental setup. X-ray radiation, with tunable energy resonant with the vanadium \(\mathrm{L_{2,3}}\) edge, impinges on the sample surface and the emitted electrons are collected and imaged through electrostatic and magnetic lenses. The \(\mathrm{V_2O_3}\) film is coated with gold metal electrodes, allowing to drive a current through the device (see sketch of a typical resistive switching current-voltage curve in the bottom panel) while simultaneously acquiring XLD-PEEM images. </center>  
+
+topological defects act as seeds for the formation of the metallic phase, thus triggering the macroscopic volatile resistive switching.  
+
+\(\mathrm{V_2O_3}\) is a prototypical Mott insulator that undergoes a thermally- driven transition from a high- temperature paramagnetic, rhombohedral metal to a low- temperature antiferromagnetic monoclinic insulator [38- 40]. The lattice transformation at the critical temperature \(T_{IMT}\) implies the breaking of the \(C_3\) symmetry of the non- primitive hexagonal unit cell of the high- temperature metallic phase (see Fig. 1a). The structural transition can thus be described [37] by a vector order parameter:  
+
+\[\vec{\epsilon} = (\epsilon_{31},\epsilon_{23}) = \epsilon \left(\cos \phi_n,\sin \phi_n\right) \quad (1)\]  
+
+associated to the shear strain components \(\epsilon_{31}\) and \(\epsilon_{23}\) that characterize the monoclinic distortion. Below \(T_{IMT}\) , the amplitude of the order parameter, \(\epsilon\) , becomes non- zero, while the phase can assume three different values:  
+
+\[\phi_{n} = (2n - 1)\frac{\pi}{3} \quad (2)\]  
+
+corresponding to the distortion along the three equivalent hexagonal axes of the rhombohedral phase, indicated in the following by the versors \(\hat{\epsilon}_n\) , \(n = 1,2,3\) (see Fig. 1b)).  
+
+Resistive switching can be induced by applying an electric field across a patterned micro- gap at temperatures close to \(T_{IMT}\) [4, 33, 41]. The resistive switching device investigated here is formed by a \(20 \mathrm{nm} \mathrm{V}_2\mathrm{O}_3\) film coated with gold electrodes. \(\mathrm{V}_2\mathrm{O}_3\) is grown by oxygen- assisted Molecular Beam Epitaxy on a \((0001)\) - \(\mathrm{Al}_2\mathrm{O}_3\) substrate with a \(40 \mathrm{nm} \mathrm{Cr}_2\mathrm{O}_3\) buffer layer to reduce any interfacial residual strain [42]. The resulting \(\mathrm{V}_2\mathrm{O}_3\) film has the \(c\) axis oriented parallel to the surface normal, with \(T_{IMT} = 145 \mathrm{K}\) (see Supplementary Information Fig. S1). Two gold electrodes allow the application of an electric bias across the gap of width \(w = 2 \mu \mathrm{m}\) and length \(l = 30 \mu \mathrm{m}\) (figure 1c)). The gap region between the electrodes is imaged using PhotoEmission Electron Microscopy (PEEM), combined with X- ray Linear Dichroism (XLD) at the \(\mathrm{L_{2,3}}\) vanadium edge (513- 530 eV, see Figure S2) [36, 37, 43]. The XLD- PEEM images are obtained from the normalized difference between images recorded with the light electric field vector, \(\vec{E}\) , perpendicular and \(16^{\circ}\) to the surface normal at a photon energy \(520.6 \mathrm{eV}\) . Since the XLD signal depends on the angle between the in- plane component of \(\vec{E}\) and the position dependent order parameter, \(\vec{\epsilon} (\mathbf{r})\) [37] (see Fig. 1b) and c)), this technique provides a map - with \(\sim 30 \mathrm{nm}\) spatial resolution - of the
+
+<--- Page Split --->
+
+three different monoclinic domains during the resistive switching process.  
+
+Figure 2a) shows an XLD- PEEM image obtained in the monoclinic insulating phase at \(T = 120 \mathrm{K}\) . The \(\mathrm{V}_2\mathrm{O}_3\) nanotexture exhibits features typical of the monoclinic insulating phase [36, 37]. Monoclinic domains with different \(\phi_n\) give rise to different XLD contrast, which can be identified as different color intensities within the XLD- PEEM image. The minimization of the total strain leads to the formation of stripe- like domains, with symmetry- constrained directions [37]. Each monoclinic insulating domain extends over a few micrometers, thus connecting the two electrodes, and it is characterized by a width \(w_{dom} \sim 200 \mathrm{nm}\) [37].  
+
+The XLD- PEEM imaging is then repeated while driving a current, \(I\) , through the device and measuring the voltage drop, \(V\) , across the gold contacts. Figures 2b)- j) show the XLD- PEEM images acquired at increasing values of \(I\) , following the upward branch of the hysteresis cycle. The presence of an in- plane electric field across the electrodes introduces a weak image blurring that becomes significant for \(V \geq 6 - 8 \mathrm{V}\) . Despite this, the nanodomains are well resolved during the resistive switching process, which first manifests itself at the voltage drop observed between \(0.08 \mathrm{mA}\) and \(1.1 \mathrm{mA}\) (Fig. 2 c) and d) respectively). As the current is further increased, the melting of the monoclinic nanotexture in the region delimited by white dashed lines (Fig. 2 e)- j)) progresses with a widening channel with a homogenous intensity. The XLD contrast measured in the region between the white dashed lines corresponds to the signal of the high- temperature rhombohedral phase. This is also confirmed from the angle dependence of the XLD signal [37]. As shown in the Supplementary Information Fig. S3, images collected with two different X- rays polarization angles, with respect to the in- plane \(\mathrm{V}_2\mathrm{O}_3\) axes, show no intensity variation upon sample rotation in the metallic channel, as opposed to the lateral monoclinic domains, for which the XLD signal depends on the angle between the light polarization and \(\vec{\epsilon} (\mathbf{r})\) . The constant XLD contrast region in the middle of the gap therefore appears due to the formation of a metallic channel with rhombohedral lattice structure ( \(\epsilon = 0\) ). XLD- PEEM images obtained under the same conditions, but with a larger field of view, capture the whole gap of the device (see Supplementary Information Fig. S4). The metallic channel consistently forms in the same location within the gap with no additional metallic paths observed. Furthermore, when the applied current is removed, the metallic channel disappears and the monoclinic domains reappear with the same pre- switching configuration, indicating a volatile process.  
+
+The formation of the metallic channel is pinned by a specific topology of the lattice nanotexture, characterized by V- shaped domains, i.e. at the crossing point of domains with the same \(\phi_n\) with directions that differ by \(\pi /3\) . Fig. 3a) shows a detail of the switching region, using a colorscale that highlights the three differ  
+
+![](images/Figure_2.jpg)
+
+<center>FIG. 2. XLD-PEEM images before (a) and during (b-j) the application of an electric current at \(T = 120 \mathrm{K}\) . The homogeneous regions at the top and bottom of each image are the gold electrodes. The area in between is the exposed \(\mathrm{V}_2\mathrm{O}_3\) antiferromagnetic monoclinic phase, exhibiting a striped domain nanotexture. For currents larger than \(1.5 \mathrm{mA}\) , the striped domains disappear in the region delimited by the white dashed lines, demonstrating the appearance of a rhombohedral metallic filament, which widens as the current is increased. </center>  
+
+ent domains with a monoclinic distortion along \(\hat{\epsilon}_1\) (red, \(\phi_1 = \pi /3\) ), \(\hat{\epsilon}_2\) (blue, \(\phi_2 = \pi\) ) and \(\hat{\epsilon}_3\) (yellow, \(\phi_3 = 5\pi /3\) ). The stabilization of the monoclinic nanotexture is driven by the Saint- Venant compatibility condition [37], which ensures a continuity of the medium during a deformation
+
+<--- Page Split --->
+![](images/Figure_3.jpg)
+
+<center>FIG. 3. a) Detail of the XLD-PEEM image shown in Fig. 2 in the region where the metallic filament is formed upon the application of a current above the threshold \(I_{th}\) . b) Schematic of the monoclinic domains crossing at \(60^{\circ}\) and forming a topological defect. Blue, red and yellow areas identify the three possible monoclinic domains corresponding to the three equivalent order parameter directions \(\hat{\epsilon}_{n}\) . The order parameter at the boundaries between different domains is oriented along \(\hat{\epsilon}_{1} + \hat{\epsilon}_{2}\) ( \(2\pi /3\) ) for the red-blue interface and along \(\hat{\epsilon}_{2} + \hat{\epsilon}_{3}\) ( \(4\pi /3\) ) for the blue-yellow interface. The mixed red-yellow triangular region indicates the local suppression of the strain at the topological defect. The energy functionals shown on the left and right, illustrate how a topological defect (green plot, solid line) decreases the insulator-metal energy difference, \(\Delta\) . </center>  
+
+via the curl- free condition:  
+
+\[\vec{\nabla}\times \vec{\epsilon} (\mathbf{r}) = 0. \quad (3)\]  
+
+The conservation of the parallel component of \(\vec{\epsilon} (\mathbf{r})\) across an interface between two different domains has two important implications:  
+
+i) the interface between two different monoclinic domains is oriented along \(\hat{\epsilon}_{n}\) of the third domain; ii) the interface between a monoclinic and a rhombohedral metallic domain is oriented perpendicularly to \(\hat{\epsilon}_{n}\) of the monoclinic domain.  
+
+If we consider, for example, a domain with order parameter along \(\hat{\epsilon}_{2}\) (blue in Figure 3b)), its interface is oriented along \(\hat{\epsilon}_{1}\) , i.e. at \(\pi /3\) angle, when it neighbours an \(\hat{\epsilon}_{3}\) domain (yellow), whereas it is oriented along \(\hat{\epsilon}_{3}\) , i.e. at \(2\pi /3\) angle, when it neighbours an \(\hat{\epsilon}_{1}\) domain (red), in agreement with the nanotexture reported in Fig. 3. The Saint- Venant condition corresponds to a fixed phase jump \(\delta \phi = 2\pi /3\) of \(\vec{\epsilon} (\mathbf{r})\) across any interface between two monoclinic domains. We note that this condition is satisfied throughout the nanotextured domain structure, except at the V- shaped vertex structure formed by two \(\hat{\epsilon}_{2}\)  
+
+domains with boundaries oriented along \(\hat{\epsilon}_{1}\) and \(\hat{\epsilon}_{3}\) . If we consider a circuit \(\Gamma_{1}\) across the boundary between two striped domains, the total phase shift is given by \(\delta \phi = +2\pi /3 - 2\pi /3 = 0\) thus adhering to the curl- free condition in Eq. 3. In contrast, the topology of the V- shaped structure is such that, if we move around the internal apex \((\Gamma_{2})\) , the total phase- shift is constrained to \(\delta \phi = +2\pi /3 + 2\pi /3 = 4\pi /3\) , thus breaking the curl- free condition. The consequence is that the vertex of the V- shaped domains acts as a topological defect with a fractional Hopf index (see Supplementary Information Section S6). These topological defects are inherently characterized by the strong frustration of the local value of the order parameter \(\vec{\epsilon} (\mathbf{r})\) and fluctuations on spatial and temporal scales that cannot be captured by the present experiment. We further note that the formation of the topological defect is a direct and unavoidable consequence of the quasi- 1D confined geometry of the system. Whereas the component of the order parameter parallel to the electrodes \((\epsilon_{||})\) , see Fig. 3b) can be compensated outside the gap, the perpendicular component \((\epsilon_{\perp})\) has to be minimized to avoid the accumulation of excessive strain energy within the gap region. Thus, considering the directions of \(\vec{\epsilon} (\mathbf{r})\) at the boundaries between different monoclinic domains (see Fig. 3b), the formation of V- shaped domains is a unique configuration that fulfils the requirement \(\epsilon_{\perp} = 0\) .  
+
+The suppression of the symmetry- breaking order parameter, \(\vec{\epsilon} (\mathbf{r})\) , at topological defects has far- reaching implications related to the nature of the resistive switching process. The electronic IMT can be described by a scalar order parameter \(\eta (\mathbf{r})\) [37], which depends on the position \(\mathbf{r}\) and is such that \(\eta = - 1\) in the metallic state and \(\eta = +1\) in the insulating state. The coupling between the electronic and structural transitions can be described by the energy functional [37]:  
+
+\[F[\epsilon ,\eta ]\propto \int d\mathbf{r}\left\{\left(\eta^{2}(\mathbf{r}) - 1\right)^{2} - g(\epsilon^{2}(\mathbf{r}) - \epsilon_{t}^{2}(V))\eta (\mathbf{r})\right\} ,\]  
+
+where \(g\) is the coupling between the electronic order parameter and the strain and \(\epsilon_{t}(V)\) is a threshold parameter that controls the first- order IMT and can depend on the applied voltage \(V\) . When \(\epsilon^{2}(\mathbf{r}) > \epsilon_{t}^{2}(V)\) , the insulating phase with \(\eta = +1\) is locally favoured, whereas for strain smaller than the threshold value, i.e. \(\epsilon^{2}(\mathbf{r}) < \epsilon_{t}^{2}(V)\) , the metallic solution is stabilized. \(\epsilon_{t}^{2}(V)\) thus represents the threshold above which the insulating monoclinic state \((\eta = +1\) , \(\epsilon \neq 0\) ) becomes stable. The description of the electric- field induced transition is based on the observation [18] that the electric field directly couples to the electronic bandstructure of a Mott insulator, making the metallic phase more stable. The transition can thus be described assuming that \(\epsilon_{t}^{2}(V)\) increases with increasing \(V\) . The energy difference between the insulating and metallic phase can be expressed as \(\Delta (\mathbf{r}, V) = F[- 1] - F[+1] \simeq g \left[ \epsilon^{2}(\mathbf{r}) - \epsilon_{t}^{2}(V) \right]\) . If we start from the insulating phase with \(\epsilon^{2}(\mathbf{r}) > \epsilon_{t}^{2}(V) = 0\) , the IMT takes place when \(V\) is increased up to the point
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>FIG. 4. a) I-V curve measured during the XLD-PEEM imaging. The sudden drop in the voltage measured at \(I_{th} = 1.05 \mathrm{mA}\) indicates the first resistive switch. b) Line profiles of the XLD-PEEM images in Fig. 2. The grey shaded area indicates the progressive widening of the metallic rhombohedral filament. The direction of the line profiles is shown by the white dashed line in the XLD-PEEM image on top, where we report a detail of Fig. 2j. c) Width, \(d\) , of the metallic filament as a function of current. The blue/red markers represent the values of \(d\) obtained from the XLD-PEEM images below/above \(I_{th}\) . The green solid line shows an estimate of \(d\) , derived from a parallel resistors model predicting a sudden jump of \(d\) to \(200 \mathrm{mA}\) in \(I_{th}\) (see Supplementary Information Section S5). </center>  
+
+that \(\Delta (\mathbf{r}, V) = 0\) . A topological defect, which locally suppresses \(\epsilon^2 (\mathbf{r})\) , thus acts as a seed with a lower threshold compared to the rest of the system.  
+
+Intriguingly, we also note from Eq. 4 that the IMT can take place at a non- zero value of \(\epsilon^2 (\mathbf{r})\) , which allows the formation of a non- thermal metallic state \((\eta = - 1)\) with a finite monoclinic distortion \((\epsilon^2 (\mathbf{r}) \lesssim \epsilon_t^2 (V))\) , as already observed in non- equilibrium optical experiments [37, 44]. The nature of the early- stage switching process can be inferred by a direct comparison between the electrical state of the device and the melting of the monoclinic domains. The \(I - V\) curve of the device, as measured in- situ during the PEEM imaging, is plotted in Fig. 4a).  
+
+XLD- PEEM images were recorded at specific values of \(I\) . The \(I - V\) plot shows that the first resistive switching event occurs at the threshold current \(I_{th} = 1.05 \mathrm{mA}\) . In Figure 4b) we report a linecut of the XLD- PEEM image acquired at specific values of \(I\) ; the image profile is taken along a line crossing the monoclinic domains in the middle of the device gap (see white solid line in Fig. 4, top panel). For large currents running through the device, the line profile in Fig. 4b) displays a flat region, which indicates the melting of the monoclinic nanodomains due to the formation of the rhombohedral metallic channel. As highlighted by the grey area in Fig. 4b), the width \(d\) of the metallic filament increases with the current, from
+
+<--- Page Split --->
+
+\(d = 0.23\pm 0.05\mu \mathrm{m}\) at \(I = 1.5\mathrm{mA}\) to \(d = 3.7\pm 0.2\mu \mathrm{m}\) at \(I = 10\mathrm{mA}\)  
+
+Modelling the device as a circuit with two parallel resistors (see Supplementary Information Section S5) allows an estimation of \(d\) of the rhombohedral filament corresponding to the observed voltage drop. For large currents running through the device, the experimentally determined values of \(d\) match well with those predicted for a metallic channel forming in the gap, which has the resistivity of the high- temperature rhombohedral phase, as shown in Fig. 4. However, in correspondence of the first resistive switching event at \(I_{th} = 1.05\mathrm{mA}\) , the model predicts the sudden formation of a \(\sim 200\mathrm{nm}\) wide metallic rhombohedral filament, which is not visible in the XLD- PEEM images (see Fig. 4c and Supplementary Information Fig. S5), despite being well above the experimental resolution of the microscope. To explain this discrepancy, one might suspect that a rhombohedral metallic filament forms below the surface of the \(\mathrm{V}_2\mathrm{O}_3\) film, where it is not detected by PEEM which has a surface sensitivity limited to the first few nanometers. In fact, two arguments act against this possibility: i) the presence of the \(\mathrm{Cr}_2\mathrm{O}_3\) buffer layer reduces the substrate- film lattice mismatch from \(4.2\%\) to \(0.1\%\) , thus almost entirely removing the residual epitaxial strain in the film [42], which is known to suppress the monoclinic phase and favour interfacial metallicity [42, 45]. In contrast to highly- strained films, in which the metal to insulator resistivity jump is strongly suppressed [45], the films in the present study display the 5- order of magnitude resistivity change typical of the unstrained metal- to- insulator transition (see Fig. S1); ii) the curl- free conditions force the interface between monoclinic and rhombohedral metallic regions to be oriented perpendicularly to the order parameter of the monoclinic domain. The formation of a sub- surface metallic layer would lead to a sharp \((\ll 20\mathrm{nm})\) monoclinic- rhombohedral interface parallel to \(\vec{\epsilon}\) , thus leading to a dramatic increase of the strain energy of the system. Our results are compatible with a complex scenario in which the topology- driven resistive switching likely occurs via the sudden transformation of a single \(200\mathrm{nm}\) wide insulating monoclinic domain into a metallic channel with a non- thermal monoclinic lattice structure. At a second stage, the Joule heating leads to the thermally driven monoclinic- to- rhombohedral structural transition and the formation of rhombohedral metallic channels perpendicular to both the metallic electrodes and the \(\hat{\epsilon}_2\) order parameter direction, as observed in Fig. 2.  
+
+The X- ray- based nanoimaging of a Mott device under operating conditions allowed us to simultaneously capture the formation of nanoscale conductive paths and the topology of the underlying symmetry- broken nanotexture. The present results expand our knowledge of the resistive switching process in Mott materials by demonstrating the leading role of inherent topological defects in initiating the avalanche process. The  
+
+methodologies used in this work imply that nanoscale strain engineering approaches could unlock a gate to manipulating topological defects and controlling the electronic switching dynamics in real devices, such as Mott- transition- based RRAM [46, 47], Mott memristor [48- 50] and artificial neurons [51, 52]. The concept of topology- driven resistive switching will be key to assessing the possible non- thermal nature of the early stage electronic phase [37] as well as the microscopic origin of memory and non- volatile effects recently observed in Mott devices [6]. We note that the relation between topological defects and electronic phase transitions established here is a general concept, potentially extendable to other systems that undergo first- order phase transitions accompanied by a symmetry breaking, as described by the energy functional (4). Relevant examples embrace transition- metal oxides [3, 53], such as vanadates, nickelates and manganites, and layered materials, such as \(1T\) - TaS2 [54- 57], in which the IMT is accompanied by charge-, lattice- and orbital- ordered states with reduced symmetry. Further platforms include cuprate superconductors [58] and kagome metals [59] in which light- or magnetic- induced discontinuous electronic transitions coexist with charge- order. Topological defects in the order parameter therefore provide a framework for understanding non- equilibrium electronic phase transitions, allowing all- optical control of hidden states of matter in a broad class of quantum materials [57, 60- 64].  
+
+We thank Diamond Lights Source for the provision of beamtime under proposal numbers MM- 27218, MM- 31711 and MM- 34455. We thank Manuel R. Osorio and Fernando J. Urbanos for the fabrication of sample electrodes at the Centre for Micro and Nanofabrication of IMDEA Nanociencia. A.M., S.M. and C.G. acknowledge financial support from MIUR through the PRIN 2015 (Prot. 2015C5SEJJ001) and PRIN 2017 (Prot. 20172H2SC4.005) programs and from the European Union - Next Generation EU through the MUR- PRIN2022 (Prot. 20228YCYY7) program. C.G. acknowledges support from Università Cattolica del Sacro Cuore through D.I, D.2.2 and D.3.1 grants. S.M. acknowledges partial financial support through the grant "Finanziamenti ponte per bandi esterni" from Università Cattolica del Sacro Cuore. I.F.C. and M.M. acknowledge support from the "Severo Ochoa" Programme for Centres of Excellence in R&D (CEX2020- 001039- S) and the Spanish AEI- MCIN PID2021- 122980OB- C52 (ECoSOC- ECLIPSE). I.F.C holds a FPI fellowship from the Spanish AEI- MCIN (PRE2020- 092625). W.- F.H., S.M., J.W.S. and J.- P.L. acknowledge financial support by the KU Leuven Research Funds Project No. C14/21/083, iBOF/21/084, KAC24/18/056 and C14/17/080, as well as the FWO AKUL/13/19 and AKUL/19/023, and the Research Funds of the INTERREG- E- TEST Project (EMR113) and INTERREG- VL- VL- PATHFINDER Project (0559).
+
+<--- Page Split --->
+
+[1] Z. Wang, H. Wu, G. W. Burr, C. S. Hwang, K. L. Wang, Q. Xia, and J. J. Yang, Resistive switching materials for information processing, Nature Reviews Materials 5, 173 (2020).[2] Y. Zhou and S. Ramanathan, Mott Memory and Neuromorphic Devices, Proceedings of the IEEE 103, 1289 (2015).[3] Y. Tokura, M. Kawasaki, and N. Nagaosa, Emergent functions of quantum materials, Nature Physics 13, 1056 (2017).[4] J. del Valle, J. G. Ramirez, M. J. Rozenberg, and I. K. Schuller, Challenges in materials and devices for resistive- switching- based neuromorphic computing, Journal of Applied Physics 124, 211101 (2018).[5] C. Adda, B. Corraze, P. Stoliar, P. Diener, J. Tranchant, A. Filatre- Furcate, M. Fourmigue, D. Lorcy, M.- P. Besland, E. Janod, and L. Cario, Mott insulators: A large class of materials for Leaky Integrate and Fire (LIF) artificial neuron, Journal of Applied Physics 124, 152124 (2018).[6] J. del Valle, J. G. Ramirez, M. J. Rozenberg, and I. K. Schuller, Subthreshold firing in Mott nanodevices, Nature 569, 388 (2019).[7] A. Pérez- Tomás, Functional Oxides for Photoneuromorphic Engineering: Toward a Solar Brain, Advanced Materials Interfaces 6, 1900471 (2019).[8] J. del Valle, P. Salev, Y. Kalcheim, and I. K. Schuller, A caloritronics- based Mott neuristor, Scientific Reports 10, 4292 (2020).[9] Z. Zhang, S. Mondal, S. Mandal, J. M. Allred, N. A. Aghamiri, A. Fali, Z. Zhang, H. Zhou, H. Cao, F. Rodolakis, J. L. McChesney, Q. Wang, Y. Sun, Y. Abate, K. Roy, K. M. Rabe, and S. Ramanathan, Neuromorphic learning with Mott insulator NiO, Proceedings of the National Academy of Sciences 118, e2017239118 (2021).[10] X. Deng, S.- Q. Wang, Y.- X. Liu, N. Zhong, Y.- H. He, H. Peng, P.- H. Xiang, and C.- G. Duan, A Flexible Mott Synaptic Transistor for Nociceptor Simulation and Neuromorphic Computing, Advanced Functional Materials 31, 2101099 (2021).[11] S. Deng, H. Yu, T. J. Park, A. N. Islam, S. Manna, A. Pofelski, Q. Wang, Y. Zhu, S. K. Sankaranarayanan, A. Sengupta, and S. Ramanathan, Selective area doping for Mott neuromorphic electronics, Science Advances 9, eade4838 (2023).[12] Y. Ran, Y. Pei, Z. Zhou, H. Wang, Y. Sun, Z. Wang, M. Hao, J. Zhao, J. Chen, and X. Yan, A review of Mott insulator in memristors: The materials, characteristics, applications for future computing systems and neuromorphic computing, Nano Research 16, 1165 (2023).[13] Z. Yang, C. Ko, and S. Ramanathan, Oxide electronics utilizing ultrafast metal- insulator transitions, Annual Review of Materials Research 41, 337 (2011).[14] A. Mehonic and A. J. Kenyon, Brain- inspired computing needs a master plan, Nature 604, 255 (2022).[15] W.- F. Hsu, S. Mellaerts, C. Bellani, P. Homm, N. Uchida, M. Menghini, M. Houssa, J. W. Seo, and J.- P. Locquet, Raman spectroscopy and phonon dynamics in strained \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review Materials 7, 074606 (2023).[16] J. Zhang and R. D. Averitt, Dynamics and control in complex transition metal oxides, Annual Review of Materials Research 44, 19 (2014).[17] D. Basov, R. Averitt, and D. Hsieh, Towards properties on demand in quantum materials, Nature materials 16, 1077 (2017).[18] G. Mazza, A. Amaricci, M. Capone, and M. Fabrizio, Field- driven mott gap collapse and resistive switch in correlated insulators, Physical review letters 117, 176401 (2016).[19] C. T. Suen, I. Markovic, M. Zonno, S. Zhdanovich, N.- H. Jo, M. Schmid, P. Hansmann, P. Pupuhal, K. Fursich, V. Zimmerman, et al., Nature of the current- induced insulator- to- metal transition in \(\mathrm{Ca}_2\mathrm{RuO}_4\) as revealed by transport- ARPES, arXiv:2308.05803 (2023).[20] D. J. Wouters, S. Menzel, J. A. J. Rupp, T. Hennen, and R. Waser, On the universality of the I- V switching characteristics in non- volatile and volatile resistive switching oxides, Faraday Discuss. 213, 183 (2019).[21] Y. Kalcheim, A. Camjayi, J. del Valle, P. Salev, M. Rozenberg, and I. K. Schuller, Non- thermal resistive switching in Mott insulator nanowires, Nature Communications 11, 2985 (2020).[22] P. Stoliar, L. Cario, E. Janod, B. Corraze, C. Guillot- Deudon, S. Salmon- Bourmand, V. Guiot, J. Tranchant, and M. Rozenberg, Universal Electric- Field- Driven Resistive Transition in Narrow- Gap Mott Insulators, Advanced Materials 25, 3222 (2013).[23] V. Guiot, L. Cario, E. Janod, B. Corraze, V. Ta Phuoc, M. Rozenberg, P. Stoliar, T. Cren, and D. Roditchev, Avalanche breakdown in \(\mathrm{CaTa}_4\mathrm{Se}_8\mathrm{- }x\mathrm{Te}_x\) narrow- gap Mott insulators, Nature communications 4, 1722 (2013).[24] F. Nakamura, M. Sakaki, Y. Yamanaka, S. Tamaru, T. Suzuki, and Y. Maeno, Electric- field- induced metal maintained by current of the Mott insulator \(\mathrm{Ca}_2\mathrm{RuO}_4\) , Scientific reports 3, 2536 (2013).[25] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Origin of hysteresis in resistive switching in magnetite is Joule heating, Physical Review B 79, 245131 (2009).[26] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Statistical distribution of the electric field- driven switching of the Verwey state in \(\mathrm{Fe}_3\mathrm{O}_4\) , New Journal of Physics 14, 013019 (2012).[27] E. Janod, J. Tranchant, B. Corraze, M. Querre, P. Stoliar, M. Rozenberg, T. Cren, D. Roditchev, V. T. Phuoc, M.- P. Besland, et al., Resistive switching in Mott insulators and correlated systems, Advanced Functional Materials 25, 6287 (2015).[28] R. Waser and M. Aono, Nanoionics- based resistive switching memories, Nature materials 6, 833 (2007).[29] D. Ielmini, Resistive switching memories based on metal oxides: mechanisms, reliability and scaling, Semiconductor Science and Technology 31, 063002 (2016).[30] J. S. Lee, S. Lee, and T. W. Noh, Resistive switching phenomena: A review of statistical physics approaches, Applied Physics Reviews 2, 031303 (2015).[31] P. Stoliar, J. Tranchant, B. Corraze, E. Janod, M.- P. Besland, F. Tesler, M. Rozenberg, and L. Cario, A Leaky- Integrate- and- Fire Neuron Analog Realized with a Mott Insulator, Advanced Functional Materials 27, 1604740 (2017).[32] M. Lange, S. Guénon, Y. Kalcheim, T. Luibrand, N. M.  
+
+complex transition metal oxides, Annual Review of Materials Research 44, 19 (2014).[17] D. Basov, R. Averitt, and D. Hsieh, Towards properties on demand in quantum materials, Nature materials 16, 1077 (2017).[18] G. Mazza, A. Amaricci, M. Capone, and M. Fabrizio, Field- driven mott gap collapse and resistive switch in correlated insulators, Physical review letters 117, 176401 (2016).[19] C. T. Suen, I. Markovic, M. Zonno, S. Zhdanovich, N.- H. Jo, M. Schmid, P. Hansmann, P. Pupuhal, K. Fursich, V. Zimmerman, et al., Nature of the current- induced insulator- to- metal transition in \(\mathrm{Ca}_2\mathrm{RuO}_4\) as revealed by transport- ARPES, arXiv:2308.05803 (2023).[20] D. J. Wouters, S. Menzel, J. A. J. Rupp, T. Hennen, and R. Waser, On the universality of the I- V switching characteristics in non- volatile and volatile resistive switching oxides, Faraday Discuss. 213, 183 (2019).[21] Y. Kalcheim, A. Camjayi, J. del Valle, P. Salev, M. Rozenberg, and I. K. Schuller, Non- thermal resistive switching in Mott insulator nanowires, Nature Communications 11, 2985 (2020).[22] P. Stoliar, L. Cario, E. Janod, B. Corraze, C. Guillot- Deudon, S. Salmon- Bourmand, V. Guiot, J. Tranchant, and M. Rozenberg, Universal Electric- Field- Driven Resistive Transition in Narrow- Gap Mott Insulators, Advanced Materials 25, 3222 (2013).[23] V. Guiot, L. Cario, E. Janod, B. Corraze, V. Ta Phuoc, M. Rozenberg, P. Stoliar, T. Cren, and D. Roditchev, Avalanche breakdown in \(\mathrm{CaTa}_4\mathrm{Se}_8\mathrm{- }x\mathrm{Te}_x\) narrow- gap Mott insulators, Nature communications 4, 1722 (2013).[24] F. Nakamura, M. Sakaki, Y. Yamanaka, S. Tamaru, T. Suzuki, and Y. Maeno, Electric- field- induced metal maintained by current of the Mott insulator \(\mathrm{Ca}_2\mathrm{RuO}_4\) , Scientific reports 3, 2536 (2013).[25] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Origin of hysteresis in resistive switching in magnetite is Joule heating, Physical Review B 79, 245131 (2009).[26] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Statistical distribution of the electric field- driven switching of the Verwey state in \(\mathrm{Fe}_3\mathrm{O}_4\) , New Journal of Physics 14, 013019 (2012).[27] E. Janod, J. Tranchant, B. Corraze, M. Querre, P. Stoliar, M. Rozenberg, T. Cren, D. Roditchev, V. T. Phuoc, M.- P. Besland, et al., Resistive switching in Mott insulators and correlated systems, Advanced Functional Materials 25, 6287 (2015).[28] R. Waser and M. Aono, Nanoionics- based resistive switching memories, Nature materials 6, 833 (2007).[29] D. Ielmini, Resistive switching memories based on metal oxides: mechanisms, reliability and scaling, Semiconductor Science and Technology 31, 063002 (2016).[30] J. S. Lee, S. Lee, and T. W. Noh, Resistive switching phenomena: A review of statistical physics approaches, Applied Physics Reviews 2, 031303 (2015).[31] P. Stoliar, J. Tranchant, B. Corraze, E. Janod, M.- P. Besland, F. Tesler, M. Rozenberg, and L. Cario, A Leaky- Integrate- and- Fire Neuron Analog Realized with a Mott Insulator, Advanced Functional Materials 27, 1604740 (2017).[32] M. Lange, S. Guénon, Y. Kalcheim, T. Luibrand, N. M.
+
+<--- Page Split --->
+
+Vargas, D. Schwebius, R. Kleiner, I. K. Schuller, and D. Koelle, Imaging of electrothermal filament formation in a mott insulator, Physical Review Applied 16, 054027 (2021). [33] J. Del Valle, N. M. Vargas, R. Rocco, P. Salev, Y. Kalcheim, P. N. Lapa, C. Adda, M.- H. Lee, P. Y. Wang, L. Fratino, et al., Spatiotemporal characterization of the field- induced insulator- to- metal transition, Science 373, 907 (2021). [34] D. Babich, J. Tranchant, C. Adda, B. Corraze, M.- P. Besland, P. Warnicke, D. Bedau, B. Bertoncini, J.- Y. Mevellec, B. Humbert, J. Rupp, T. Hennen, D. Wouters, R. Llopis, L. Cario, and E. Janod, Lattice contraction induced by resistive switching in chromium- doped \(\mathrm{V}_2\mathrm{O}_3\) : a hallmark of Mott physics, arXiv preprint arXiv:2105.05093 (2021). [35] T. Luibrand, A. Bercher, R. Rocco, Tahouni- Bonab, L. Varbaro, C. Rischau, and e. a. Dominguez, Characteristic lengthscales of the electrically- induced insulator- to- metal transition, arXiv preprint arXiv:2301.00456 (2023). [36] A. Ronchi, P. Homm, M. Menghini, P. Franceschini, F. Maccherozzi, F. Banfi, G. Ferrini, F. Cilento, F. Parmigiani, S. S. Dhesi, et al., Early- stage dynamics of metallic droplets embedded in the nanotextured Mott insulating phase of \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review B 100, 075111 (2019). [37] A. Ronchi, P. Franceschini, A. De Poli, P. Homm, A. Fitzpatrick, F. Maccherozzi, G. Ferrini, F. Banfi, S. S. Dhesi, M. Menghini, et al., Nanoscale self- organization and metastable non- thermal metallicity in Mott insulators, Nature Communications 13, 3730 (2022). [38] D. McWhan, T. Rice, and J. Remeika, Mott transition in Cr- doped \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review Letters 23, 1384 (1969). [39] D. McWhan and J. Remeika, Metal- insulator transition in \((\mathrm{V}_{1 - x}\mathrm{Cr}_x)_2\mathrm{O}_3\) , Physical Review B 2, 3734 (1970). [40] D. McWhan, A. Menth, J. Remeika, W. Brinkman, and T. Rice, Metal- insulator transitions in pure and doped \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review B 7, 1920 (1973). [41] S. Guénon, S. Scharinger, S. Wang, J. Ramírez, D. Koelle, R. Kleiner, and I. K. Schuller, Electrical breakdown in a \(\mathrm{V}_2\mathrm{O}_3\) device at the insulator- to- metal transition, Europhysics Letters 101, 57003 (2013). [42] L. Dillemans, T. Smets, R. R. Lieten, M. Menghini, C.- Y. Su, and J.- P. Locquet, Evidence of the metal- insulator transition in ultrathin unstrained \(\mathrm{V}_2\mathrm{O}_3\) thin films, Applied Physics Letters 104, 071902 (2014). [43] J.- H. Park, L. Tjeng, A. Tanaka, J. Allen, C. Chen, P. Metcalf, J. Honig, F. De Groot, and G. Sawatzky, Spin and orbital occupation and phase transitions in \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review B 61, 11506 (2000). [44] A. Ronchi, P. Franceschini, P. Homm, M. Gandolfi, G. Ferrini, S. Pagliara, F. Banfi, M. Menghini, C. Giannetti, et al., Light- assisted resistance collapse in a \(\mathrm{V}_2\mathrm{O}_3\) - based mott- insulator device, Physical Review Applied 15, 044023 (2021). [45] A. Pofelski, S. Valencia, Y. Kalcheim, P. Salev, A. Rivera, C. Huang, M. A. Mawass, F. Kronast, I. K. Schuller, Y. Zhu, et al., Domain nucleation across the metal- insulator transition of self- strained \(\mathrm{V}_2\mathrm{O}_3\) films, arXiv preprint arXiv:2312.09051 (2023). [46] A. Sawa, Resistive switching in transition metal oxides, Materials Today 11, 28 (2008). [47] Y. Wang, K.- M. Kang, M. Kim, H.- S. Lee, R. Waser,  
+
+D. Wouters, R. Dittmann, J. J. Yang, and H.- H. Park, Mott-transition-based RRAM, Materials Today 28, 63 (2019). [48] M. D. Pickett, G. Medeiros-Ribeiro, and R. S. Williams, A scalable neuristor built with Mott memristors, Nature Materials 12, 114 (2013). [49] M. Yoshida, R. Suzuki, Y. Zhang, M. Nakano, and Y. Iwasa, Memristive phase switching in two-dimensional \(1\mathrm{T} - \mathrm{TaS}_2\) crystals, Science Advances 1, e1500606 (2015). [50] S. Kumar, J. P. Strachan, and R. S. Williams, Chaotic dynamics in nanoscale \(\mathrm{NbO}_2\) Mott memristors for analogue computing, Nature 548, 318 (2017). [51] F. Tesler, C. Adda, J. Tranchant, B. Corraze, E. Janod, L. Cario, P. Stoliar, and M. Rozenberg, Relaxation of a spiking Mott artificial neuron, Phys. Rev. Appl. 10, 054001 (2018). [52] X. Zhang, Y. Zhuo, Q. Luo, Z. Wu, R. Midya, Z. Wang, W. Song, R. Wang, N. K. Upadhyay, Y. Fang, F. Kiani, M. Rao, Y. Yang, Q. Xia, Q. Liu, M. Liu, and J. J. Yang, An artificial spiking afferent nerve based on Mott memristors for neurorobotics, Nature Communications 11, 51 (2020). [53] M. Imada, A. Fujimori, and Y. Tokura, Metal-insulator transitions, Rev. Mod. Phys. 70, 1039 (1998). [54] I. Vaskivskyi, J. Gospodaric, S. Brazovskii, D. Svetin, P. Sutar, E. Goreshnik, I. A. Mihailovic, T. Mertelj, and D. Mihailovic, Controlling the metal- to- insulator relaxation of the metastable hidden quantum state in \(1\mathrm{T} - \mathrm{TaS}_2\) , Science Advances 1, e1500168 (2015). [55] M. J. Hollander, Y. Liu, W.- J. Lu, L.- J. Li, Y.- P. Sun, J. A. Robinson, and S. Datta, Electrically driven reversible insulator- metal phase transition in \(1\mathrm{T} - \mathrm{TaS}_2\) , Nano Letters 15, 1861 (2015). [56] S.- H. Lee, J. S. Goh, and D. Cho, Origin of the insulating phase and first- order metal- insulator transition in \(1\mathrm{T} - \mathrm{TaS}_2\) , Phys. Rev. Lett. 122, 106404 (2019). [57] F. Y. Gao, Z. Zhang, Z. Sun, L. Ye, Y.- H. Cheng, Z.- J. Liu, J. G. Checkelsky, E. Baldini, and K. A. Nelson, Snapshots of a light- induced metastable hidden phase driven by the collapse of charge order, Science Advances 8, eabp9076 (2022). [58] B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, and J. Zaanen, From quantum matter to high- temperature superconductivity in copper oxides, Nature 518, 179 (2015). [59] T. Asaba, A. Onishi, Y. Kageyama, T. Kiyosue, K. Ohtsuka, S. Suetsugu, Y. Kohsaka, T. Gaggl, Y. Kasahara, H. Murayama, K. Hashimoto, R. Tazai, H. Kontani, B. R. Ortiz, S. D. Wilson, Q. Li, H. H. Wen, T. Shibauchi, and Y. Matsuda, Evidence for an odd- parity nematic phase above the charge- density- wave transition in a kagome metal, Nature Physics 10.1038/s41567- 023- 02272- 4 (2024). [60] L. Stojchevska, I. Vaskivskyi, T. Mertelj, P. Kusar, D. Svetin, S. Brazovskii, and D. Mihailovic, Ultrafast switching to a stable hidden quantum state in an electronic crystal, Science 344, 177 (2014). [61] J. Zhang, X. Tan, M. Liu, S. W. Teitelbaum, K. W. Post, F. Jin, K. Nelson, D. N. Basov, W. Wu, and R. D. Averitt, Cooperative photoinduced metastable phase control in strained manganite films, Nature Materials 15, 956 (2016). [62] S. Wandel, F. Boschini, E. H. da Silva Neto, L. Shen, M. X. Na, S. Zohar, Y. Wang, S. B. Welch, M. H.
+
+<--- Page Split --->
+
+Seaberg, J. D. Koralek, G. L. Dakovski, W. Hettel, M.- F. Lin, S. P. Moeller, W. F. Schlotter, A. H. Reid, M. P. Minitti, T. Boyle, F. He, R. Sutarto, R. Liang, D. Bonn, W. Hardy, R. A. Kaindl, D. G. Hawthorn, J.- S. Lee, A. F. Kemper, A. Damascelli, C. Giannetti, J. J. Turner, and G. Coslovich, Enhanced charge density wave coherence in a light- quenched, high- temperature superconductor, Science 376, 860 (2022). [63] Y. Cheng, A. Zong, L. Wu, Q. Meng, W. Xia, F. Qi, P. Zhu, X. Zou, T. Jiang, Y. Guo, J. van Wezel, A. Kogar, M. W. Zuerch, J. Zhang, Y. Zhu, and D. Xiang, Ultrafast formation of topological defects in a two- dimensional  
+
+[64] A. Verma, D. Golež, O. Y. Gorobtsov, K. Kaj, R. Russell, J. Z. Kaaret, E. Lamb, G. Khalsa, H. P. Nair, Y. Sun, R. Bouck, N. Schreiber, J. P. Ruf, V. Ramaprasad, Y. Kubota, T. Togashi, V. A. Stoica, H. Padmanabhan, J. W. Freeland, N. A. Benedek, O. G. Shpyrko, J. W. Harter, R. D. Averitt, D. G. Schlom, K. M. Shen, A. J. Millis, and A. Singer, Picosecond volume expansion drives a later- time insulator- metal transition in a nano- textured Mott insulator, Nature Physics 10.1038/s41567- 024- 02396- 1 (2024).
+
+<--- Page Split --->
+
+## Supplementary Files  
+
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+SIV203ResistiveSwitching.pdf
+
+<--- Page Split --->

preprint/preprint__0a029bedc77ee23f09bec94b1dd73e6a81b91ac30d9e51c5375983d20db515bb/preprint__0a029bedc77ee23f09bec94b1dd73e6a81b91ac30d9e51c5375983d20db515bb_det.mmd

@@ -0,0 +1,244 @@
+<|ref|>title<|/ref|><|det|>[[44, 108, 864, 175]]<|/det|>
+# Mott resistive switching initiated by topological defects  
+
+<|ref|>text<|/ref|><|det|>[[44, 196, 199, 214]]<|/det|>
+Claudio Giannetti  
+
+<|ref|>text<|/ref|><|det|>[[55, 223, 365, 240]]<|/det|>
+claudio.giannetti@unicatt.it  
+
+<|ref|>text<|/ref|><|det|>[[44, 269, 728, 289]]<|/det|>
+Università Cattolica del Sacro Cuore https://orcid.org/0000- 0003- 2664- 9492  
+
+<|ref|>text<|/ref|><|det|>[[44, 294, 365, 333]]<|/det|>
+Alessandra Milloch Università Cattolica del Sacro Cuore  
+
+<|ref|>text<|/ref|><|det|>[[44, 339, 313, 378]]<|/det|>
+Ignacio Figueruelo- Campanero IMDEA Nanociencia  
+
+<|ref|>text<|/ref|><|det|>[[44, 385, 155, 424]]<|/det|>
+Wei- Fan Hsu KU Leuven  
+
+<|ref|>text<|/ref|><|det|>[[44, 431, 370, 472]]<|/det|>
+Selene Mor Università Cattolica del Sacro Cuore  
+
+<|ref|>text<|/ref|><|det|>[[44, 478, 510, 518]]<|/det|>
+Simon Mellaerts KU Leuven https://orcid.org/0000- 0002- 6715- 3066  
+
+<|ref|>text<|/ref|><|det|>[[44, 524, 255, 542]]<|/det|>
+Francesco Maccherozzi  
+
+<|ref|>text<|/ref|><|det|>[[44, 546, 945, 587]]<|/det|>
+Diamond Light Source, Chilton, Didcot, Oxfordshire, OX11 0DE, UK. https://orcid.org/0000- 0003- 4074- 2319  
+
+<|ref|>text<|/ref|><|det|>[[44, 593, 250, 632]]<|/det|>
+Larissa Ishibe Veiga Diamond Light Source  
+
+<|ref|>text<|/ref|><|det|>[[44, 639, 610, 679]]<|/det|>
+Sarnjeet Dhesi Diamond Light Source https://orcid.org/0000- 0003- 4966- 0002  
+
+<|ref|>text<|/ref|><|det|>[[44, 685, 730, 725]]<|/det|>
+Mauro Spera Università Cattolica del Sacro Cuore https://orcid.org/0000- 0001- 9041- 364X  
+
+<|ref|>text<|/ref|><|det|>[[44, 732, 510, 772]]<|/det|>
+Jin Seo KU Leuven https://orcid.org/0000- 0003- 4937- 0769  
+
+<|ref|>text<|/ref|><|det|>[[44, 778, 397, 818]]<|/det|>
+Jean- Pierre Locquet https://orcid.org/0000- 0002- 4214- 7081  
+
+<|ref|>text<|/ref|><|det|>[[44, 824, 784, 865]]<|/det|>
+Michele Fabrizio International School for Advanced Studies https://orcid.org/0000- 0002- 2943- 3278  
+
+<|ref|>text<|/ref|><|det|>[[44, 870, 598, 911]]<|/det|>
+Mariela Menghini IMDEA Nanoscience https://orcid.org/0000- 0002- 1744- 798X
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[44, 46, 103, 63]]<|/det|>
+## Article  
+
+<|ref|>sub_title<|/ref|><|det|>[[44, 84, 135, 102]]<|/det|>
+## Keywords:  
+
+<|ref|>text<|/ref|><|det|>[[44, 121, 291, 140]]<|/det|>
+Posted Date: June 6th, 2024  
+
+<|ref|>text<|/ref|><|det|>[[44, 159, 474, 179]]<|/det|>
+DOI: https://doi.org/10.21203/rs.3.rs- 4019377/v1  
+
+<|ref|>text<|/ref|><|det|>[[42, 197, 914, 240]]<|/det|>
+License: © © This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License  
+
+<|ref|>text<|/ref|><|det|>[[42, 258, 534, 277]]<|/det|>
+Additional Declarations: There is NO Competing Interest.  
+
+<|ref|>text<|/ref|><|det|>[[42, 313, 940, 356]]<|/det|>
+Version of Record: A version of this preprint was published at Nature Communications on October 31st, 2024. See the published version at https://doi.org/10.1038/s41467- 024- 53726- z.
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[231, 100, 770, 117]]<|/det|>
+# Mott resistive switching initiated by topological defects  
+
+<|ref|>text<|/ref|><|det|>[[132, 128, 872, 170]]<|/det|>
+Alessandra Milloch, \(^{1,2,3,*}\) Ignacio Figueroelo- Campanero, \(^{4,5,\dagger}\) Wei- Fan Hsu, \(^{3}\) Selene Mor, \(^{1,2}\) Simon Mellaerts, \(^{3}\) Francesco Maccherozzi, \(^{6}\) Larissa Ishibe Veiga, \(^{6}\) Sarnjeet S. Dhesi, \(^{6}\) Mauro Spera, \(^{1}\) Jin Won Seo, \(^{7}\) Jean- Pierre Locquet, \(^{3}\) Michele Fabrizio, \(^{8}\) Mariela Menghini, \(^{4}\) and Claudio Giannetti \(^{1,2,9,\ddagger}\)  
+
+<|ref|>text<|/ref|><|det|>[[160, 173, 844, 300]]<|/det|>
+\(^{1}\) Department of Mathematics and Physics, Università Cattolica del Sacro Cuore, Brescia I- 25133, Italy \(^{2}\) ILAMP (Interdisciplinary Laboratories for Advanced Materials Physics), Università Cattolica del Sacro Cuore, Brescia I- 25133, Italy \(^{3}\) Department of Physics and Astronomy, KU Leuven, B- 3001 Leuven, Belgium \(^{4}\) IMDEA Nanociencia, Cantoblanco, 28049 Madrid, Spain \(^{5}\) Facultad Ciencias Físicas, Universidad Complutense, 28040 Madrid, Spain \(^{6}\) Diamond Light Source, Didcot, Oxfordshire OX11 0DE, UK \(^{7}\) Department of Materials Engineering, KU Leuven, 3001 Leuven, Belgium \(^{8}\) Scuola Internazionale Superiore di Studi Avanzati (SISSA), Via Bonomea 265, 34136 Trieste, Italy \(^{9}\) CNR- INO (National Institute of Optics), via Branze 45, 25123 Brescia, Italy  
+
+<|ref|>text<|/ref|><|det|>[[165, 304, 830, 377]]<|/det|>
+Avalanche resistive switching is the fundamental process that triggers the sudden change of the electrical properties in solid- state devices under the action of intense electric fields [1]. Despite its relevance for information processing, ultrafast electronics, neuromorphic devices, resistive memories and brain- inspired computation [1- 14], the nature of the local stochastic fluctuations that drive the formation of metallic regions within the insulating state has remained hidden.  
+
+<|ref|>text<|/ref|><|det|>[[165, 376, 830, 510]]<|/det|>
+Here, using operando X- ray nano- imaging, we have captured the origin of resistive switching in a \(\mathrm{V}_2\mathrm{O}_3\) - based device under working conditions. \(\mathrm{V}_2\mathrm{O}_3\) is a paradigmatic Mott material [3], which undergoes a first- order metal- to- insulator phase transition together with a lattice transformation that breaks the threefold rotational symmetry of the rhombohedral metallic phase [2, 5, 6, 8- 11, 15]. We reveal a new class of volatile electronic switching triggered by nanoscale topological defects appearing in the shear- strain based order parameter that describes the insulating phase. Our results pave the way to the use of strain engineering approaches to manipulate such topological defects and achieve the full dynamical control of the electronic Mott switching. Topology- driven, reversible electronic transitions are relevant across a broad range of quantum materials, comprising transition metal oxides, chalcogenides and kagome metals.  
+
+<|ref|>text<|/ref|><|det|>[[87, 532, 486, 797]]<|/det|>
+The insulator- to- metal transition (IMT) in Mott materials is a key mechanism for the development of next generation Mottronic devices [3, 13]. The intrinsic correlated nature of the Mott insulating state makes these systems fragile to external stimuli [16, 17], such as the application of an electric field, which can drive the collapse of the electronic band structure and the sudden release of a large number of free carriers [18, 19]. At the macroscopic level, this phenomenon manifests in the resistive switching process, i.e., a sharp increase of the current flow when the applied voltage exceeds a threshold value [6, 20- 27]. This strong non- linearity triggered many efforts to develop neuromorphic building blocks for the hardware implementation of neural networks [14] or for ultrafast volatile and non- volatile memories or processors [12, 28, 29]. The state- of- the- art macroscopic models [30] are based on resistor networks that consider interconnected nodes transforming from the insulating to metallic state in the presence of an electric field. Above a certain threshold, a percolative, avalanche transition takes place,  
+
+<|ref|>text<|/ref|><|det|>[[515, 532, 916, 559]]<|/det|>
+thus leading to the formation of conductive filaments and the consequent sudden drop in resistivity [22, 31].  
+
+<|ref|>text<|/ref|><|det|>[[515, 560, 916, 718]]<|/det|>
+The full control and exploitation of this process is currently prevented by a limited knowledge of the early- stage firing dynamics. Microscopically, little is known about the nature of the nanoscale regions that trigger the avalanche process. Also the relation between the electronic and structural properties of the switched regions and those of the pristine insulating template is a matter of debate. Pioneering optical microscopy experiments captured the real- time formation of macroscopic metallic channels [4, 32- 35], but lacked the resolution and sensitivity to address the microscopic origin of the switching process.  
+
+<|ref|>text<|/ref|><|det|>[[515, 720, 916, 878]]<|/det|>
+Here, we adopt resonant X- ray microscopy to record nanoscale snapshots of the switching dynamics in a \(\mathrm{V}_2\mathrm{O}_3\) - based nanodevice during the application of an electric field. The results unveil the fundamental role played by the order parameter topology of the underlying lattice nanotexture. The breaking of the \(C_3\) symmetry upon transition to the insulating monoclinic phase leads to the formation of three twin shear- strain domains with boundaries oriented along the three hexagonal directions [36, 37]. The geometrical constraints then produce shear- strain topological defects at the corners of monoclinic domains crossing with an angle of \(60^{\circ}\) . These nanoscale
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[102, 102, 888, 430]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[84, 444, 919, 520]]<|/det|>
+<center>FIG. 1. a) Non-primitive hexagonal unit cell of the \(\mathrm{V_2O_3}\) high-temperature rhombohedral metallic phase. b) Schematic of the rhombohedral-to-monoclinic distortion along each of the three equivalent hexagonal axes. c) PEEM experimental setup. X-ray radiation, with tunable energy resonant with the vanadium \(\mathrm{L_{2,3}}\) edge, impinges on the sample surface and the emitted electrons are collected and imaged through electrostatic and magnetic lenses. The \(\mathrm{V_2O_3}\) film is coated with gold metal electrodes, allowing to drive a current through the device (see sketch of a typical resistive switching current-voltage curve in the bottom panel) while simultaneously acquiring XLD-PEEM images. </center>  
+
+<|ref|>text<|/ref|><|det|>[[85, 544, 487, 584]]<|/det|>
+topological defects act as seeds for the formation of the metallic phase, thus triggering the macroscopic volatile resistive switching.  
+
+<|ref|>text<|/ref|><|det|>[[85, 585, 487, 704]]<|/det|>
+\(\mathrm{V_2O_3}\) is a prototypical Mott insulator that undergoes a thermally- driven transition from a high- temperature paramagnetic, rhombohedral metal to a low- temperature antiferromagnetic monoclinic insulator [38- 40]. The lattice transformation at the critical temperature \(T_{IMT}\) implies the breaking of the \(C_3\) symmetry of the non- primitive hexagonal unit cell of the high- temperature metallic phase (see Fig. 1a). The structural transition can thus be described [37] by a vector order parameter:  
+
+<|ref|>equation<|/ref|><|det|>[[170, 711, 485, 729]]<|/det|>
+\[\vec{\epsilon} = (\epsilon_{31},\epsilon_{23}) = \epsilon \left(\cos \phi_n,\sin \phi_n\right) \quad (1)\]  
+
+<|ref|>text<|/ref|><|det|>[[85, 740, 487, 794]]<|/det|>
+associated to the shear strain components \(\epsilon_{31}\) and \(\epsilon_{23}\) that characterize the monoclinic distortion. Below \(T_{IMT}\) , the amplitude of the order parameter, \(\epsilon\) , becomes non- zero, while the phase can assume three different values:  
+
+<|ref|>equation<|/ref|><|det|>[[228, 802, 485, 828]]<|/det|>
+\[\phi_{n} = (2n - 1)\frac{\pi}{3} \quad (2)\]  
+
+<|ref|>text<|/ref|><|det|>[[85, 837, 487, 878]]<|/det|>
+corresponding to the distortion along the three equivalent hexagonal axes of the rhombohedral phase, indicated in the following by the versors \(\hat{\epsilon}_n\) , \(n = 1,2,3\) (see Fig. 1b)).  
+
+<|ref|>text<|/ref|><|det|>[[515, 544, 917, 867]]<|/det|>
+Resistive switching can be induced by applying an electric field across a patterned micro- gap at temperatures close to \(T_{IMT}\) [4, 33, 41]. The resistive switching device investigated here is formed by a \(20 \mathrm{nm} \mathrm{V}_2\mathrm{O}_3\) film coated with gold electrodes. \(\mathrm{V}_2\mathrm{O}_3\) is grown by oxygen- assisted Molecular Beam Epitaxy on a \((0001)\) - \(\mathrm{Al}_2\mathrm{O}_3\) substrate with a \(40 \mathrm{nm} \mathrm{Cr}_2\mathrm{O}_3\) buffer layer to reduce any interfacial residual strain [42]. The resulting \(\mathrm{V}_2\mathrm{O}_3\) film has the \(c\) axis oriented parallel to the surface normal, with \(T_{IMT} = 145 \mathrm{K}\) (see Supplementary Information Fig. S1). Two gold electrodes allow the application of an electric bias across the gap of width \(w = 2 \mu \mathrm{m}\) and length \(l = 30 \mu \mathrm{m}\) (figure 1c)). The gap region between the electrodes is imaged using PhotoEmission Electron Microscopy (PEEM), combined with X- ray Linear Dichroism (XLD) at the \(\mathrm{L_{2,3}}\) vanadium edge (513- 530 eV, see Figure S2) [36, 37, 43]. The XLD- PEEM images are obtained from the normalized difference between images recorded with the light electric field vector, \(\vec{E}\) , perpendicular and \(16^{\circ}\) to the surface normal at a photon energy \(520.6 \mathrm{eV}\) . Since the XLD signal depends on the angle between the in- plane component of \(\vec{E}\) and the position dependent order parameter, \(\vec{\epsilon} (\mathbf{r})\) [37] (see Fig. 1b) and c)), this technique provides a map - with \(\sim 30 \mathrm{nm}\) spatial resolution - of the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 102, 488, 130]]<|/det|>
+three different monoclinic domains during the resistive switching process.  
+
+<|ref|>text<|/ref|><|det|>[[85, 132, 488, 291]]<|/det|>
+Figure 2a) shows an XLD- PEEM image obtained in the monoclinic insulating phase at \(T = 120 \mathrm{K}\) . The \(\mathrm{V}_2\mathrm{O}_3\) nanotexture exhibits features typical of the monoclinic insulating phase [36, 37]. Monoclinic domains with different \(\phi_n\) give rise to different XLD contrast, which can be identified as different color intensities within the XLD- PEEM image. The minimization of the total strain leads to the formation of stripe- like domains, with symmetry- constrained directions [37]. Each monoclinic insulating domain extends over a few micrometers, thus connecting the two electrodes, and it is characterized by a width \(w_{dom} \sim 200 \mathrm{nm}\) [37].  
+
+<|ref|>text<|/ref|><|det|>[[85, 293, 488, 793]]<|/det|>
+The XLD- PEEM imaging is then repeated while driving a current, \(I\) , through the device and measuring the voltage drop, \(V\) , across the gold contacts. Figures 2b)- j) show the XLD- PEEM images acquired at increasing values of \(I\) , following the upward branch of the hysteresis cycle. The presence of an in- plane electric field across the electrodes introduces a weak image blurring that becomes significant for \(V \geq 6 - 8 \mathrm{V}\) . Despite this, the nanodomains are well resolved during the resistive switching process, which first manifests itself at the voltage drop observed between \(0.08 \mathrm{mA}\) and \(1.1 \mathrm{mA}\) (Fig. 2 c) and d) respectively). As the current is further increased, the melting of the monoclinic nanotexture in the region delimited by white dashed lines (Fig. 2 e)- j)) progresses with a widening channel with a homogenous intensity. The XLD contrast measured in the region between the white dashed lines corresponds to the signal of the high- temperature rhombohedral phase. This is also confirmed from the angle dependence of the XLD signal [37]. As shown in the Supplementary Information Fig. S3, images collected with two different X- rays polarization angles, with respect to the in- plane \(\mathrm{V}_2\mathrm{O}_3\) axes, show no intensity variation upon sample rotation in the metallic channel, as opposed to the lateral monoclinic domains, for which the XLD signal depends on the angle between the light polarization and \(\vec{\epsilon} (\mathbf{r})\) . The constant XLD contrast region in the middle of the gap therefore appears due to the formation of a metallic channel with rhombohedral lattice structure ( \(\epsilon = 0\) ). XLD- PEEM images obtained under the same conditions, but with a larger field of view, capture the whole gap of the device (see Supplementary Information Fig. S4). The metallic channel consistently forms in the same location within the gap with no additional metallic paths observed. Furthermore, when the applied current is removed, the metallic channel disappears and the monoclinic domains reappear with the same pre- switching configuration, indicating a volatile process.  
+
+<|ref|>text<|/ref|><|det|>[[85, 796, 488, 876]]<|/det|>
+The formation of the metallic channel is pinned by a specific topology of the lattice nanotexture, characterized by V- shaped domains, i.e. at the crossing point of domains with the same \(\phi_n\) with directions that differ by \(\pi /3\) . Fig. 3a) shows a detail of the switching region, using a colorscale that highlights the three differ  
+
+<|ref|>image<|/ref|><|det|>[[520, 105, 900, 627]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[515, 643, 917, 752]]<|/det|>
+<center>FIG. 2. XLD-PEEM images before (a) and during (b-j) the application of an electric current at \(T = 120 \mathrm{K}\) . The homogeneous regions at the top and bottom of each image are the gold electrodes. The area in between is the exposed \(\mathrm{V}_2\mathrm{O}_3\) antiferromagnetic monoclinic phase, exhibiting a striped domain nanotexture. For currents larger than \(1.5 \mathrm{mA}\) , the striped domains disappear in the region delimited by the white dashed lines, demonstrating the appearance of a rhombohedral metallic filament, which widens as the current is increased. </center>  
+
+<|ref|>text<|/ref|><|det|>[[515, 810, 916, 876]]<|/det|>
+ent domains with a monoclinic distortion along \(\hat{\epsilon}_1\) (red, \(\phi_1 = \pi /3\) ), \(\hat{\epsilon}_2\) (blue, \(\phi_2 = \pi\) ) and \(\hat{\epsilon}_3\) (yellow, \(\phi_3 = 5\pi /3\) ). The stabilization of the monoclinic nanotexture is driven by the Saint- Venant compatibility condition [37], which ensures a continuity of the medium during a deformation
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[113, 102, 460, 325]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[85, 341, 488, 525]]<|/det|>
+<center>FIG. 3. a) Detail of the XLD-PEEM image shown in Fig. 2 in the region where the metallic filament is formed upon the application of a current above the threshold \(I_{th}\) . b) Schematic of the monoclinic domains crossing at \(60^{\circ}\) and forming a topological defect. Blue, red and yellow areas identify the three possible monoclinic domains corresponding to the three equivalent order parameter directions \(\hat{\epsilon}_{n}\) . The order parameter at the boundaries between different domains is oriented along \(\hat{\epsilon}_{1} + \hat{\epsilon}_{2}\) ( \(2\pi /3\) ) for the red-blue interface and along \(\hat{\epsilon}_{2} + \hat{\epsilon}_{3}\) ( \(4\pi /3\) ) for the blue-yellow interface. The mixed red-yellow triangular region indicates the local suppression of the strain at the topological defect. The energy functionals shown on the left and right, illustrate how a topological defect (green plot, solid line) decreases the insulator-metal energy difference, \(\Delta\) . </center>  
+
+<|ref|>text<|/ref|><|det|>[[86, 550, 277, 563]]<|/det|>
+via the curl- free condition:  
+
+<|ref|>equation<|/ref|><|det|>[[230, 568, 485, 588]]<|/det|>
+\[\vec{\nabla}\times \vec{\epsilon} (\mathbf{r}) = 0. \quad (3)\]  
+
+<|ref|>text<|/ref|><|det|>[[85, 596, 488, 638]]<|/det|>
+The conservation of the parallel component of \(\vec{\epsilon} (\mathbf{r})\) across an interface between two different domains has two important implications:  
+
+<|ref|>text<|/ref|><|det|>[[102, 646, 488, 725]]<|/det|>
+i) the interface between two different monoclinic domains is oriented along \(\hat{\epsilon}_{n}\) of the third domain; ii) the interface between a monoclinic and a rhombohedral metallic domain is oriented perpendicularly to \(\hat{\epsilon}_{n}\) of the monoclinic domain.  
+
+<|ref|>text<|/ref|><|det|>[[85, 732, 488, 877]]<|/det|>
+If we consider, for example, a domain with order parameter along \(\hat{\epsilon}_{2}\) (blue in Figure 3b)), its interface is oriented along \(\hat{\epsilon}_{1}\) , i.e. at \(\pi /3\) angle, when it neighbours an \(\hat{\epsilon}_{3}\) domain (yellow), whereas it is oriented along \(\hat{\epsilon}_{3}\) , i.e. at \(2\pi /3\) angle, when it neighbours an \(\hat{\epsilon}_{1}\) domain (red), in agreement with the nanotexture reported in Fig. 3. The Saint- Venant condition corresponds to a fixed phase jump \(\delta \phi = 2\pi /3\) of \(\vec{\epsilon} (\mathbf{r})\) across any interface between two monoclinic domains. We note that this condition is satisfied throughout the nanotextured domain structure, except at the V- shaped vertex structure formed by two \(\hat{\epsilon}_{2}\)  
+
+<|ref|>text<|/ref|><|det|>[[515, 103, 918, 460]]<|/det|>
+domains with boundaries oriented along \(\hat{\epsilon}_{1}\) and \(\hat{\epsilon}_{3}\) . If we consider a circuit \(\Gamma_{1}\) across the boundary between two striped domains, the total phase shift is given by \(\delta \phi = +2\pi /3 - 2\pi /3 = 0\) thus adhering to the curl- free condition in Eq. 3. In contrast, the topology of the V- shaped structure is such that, if we move around the internal apex \((\Gamma_{2})\) , the total phase- shift is constrained to \(\delta \phi = +2\pi /3 + 2\pi /3 = 4\pi /3\) , thus breaking the curl- free condition. The consequence is that the vertex of the V- shaped domains acts as a topological defect with a fractional Hopf index (see Supplementary Information Section S6). These topological defects are inherently characterized by the strong frustration of the local value of the order parameter \(\vec{\epsilon} (\mathbf{r})\) and fluctuations on spatial and temporal scales that cannot be captured by the present experiment. We further note that the formation of the topological defect is a direct and unavoidable consequence of the quasi- 1D confined geometry of the system. Whereas the component of the order parameter parallel to the electrodes \((\epsilon_{||})\) , see Fig. 3b) can be compensated outside the gap, the perpendicular component \((\epsilon_{\perp})\) has to be minimized to avoid the accumulation of excessive strain energy within the gap region. Thus, considering the directions of \(\vec{\epsilon} (\mathbf{r})\) at the boundaries between different monoclinic domains (see Fig. 3b), the formation of V- shaped domains is a unique configuration that fulfils the requirement \(\epsilon_{\perp} = 0\) .  
+
+<|ref|>text<|/ref|><|det|>[[515, 460, 917, 580]]<|/det|>
+The suppression of the symmetry- breaking order parameter, \(\vec{\epsilon} (\mathbf{r})\) , at topological defects has far- reaching implications related to the nature of the resistive switching process. The electronic IMT can be described by a scalar order parameter \(\eta (\mathbf{r})\) [37], which depends on the position \(\mathbf{r}\) and is such that \(\eta = - 1\) in the metallic state and \(\eta = +1\) in the insulating state. The coupling between the electronic and structural transitions can be described by the energy functional [37]:  
+
+<|ref|>equation<|/ref|><|det|>[[520, 584, 912, 615]]<|/det|>
+\[F[\epsilon ,\eta ]\propto \int d\mathbf{r}\left\{\left(\eta^{2}(\mathbf{r}) - 1\right)^{2} - g(\epsilon^{2}(\mathbf{r}) - \epsilon_{t}^{2}(V))\eta (\mathbf{r})\right\} ,\]  
+
+<|ref|>text<|/ref|><|det|>[[515, 625, 917, 877]]<|/det|>
+where \(g\) is the coupling between the electronic order parameter and the strain and \(\epsilon_{t}(V)\) is a threshold parameter that controls the first- order IMT and can depend on the applied voltage \(V\) . When \(\epsilon^{2}(\mathbf{r}) > \epsilon_{t}^{2}(V)\) , the insulating phase with \(\eta = +1\) is locally favoured, whereas for strain smaller than the threshold value, i.e. \(\epsilon^{2}(\mathbf{r}) < \epsilon_{t}^{2}(V)\) , the metallic solution is stabilized. \(\epsilon_{t}^{2}(V)\) thus represents the threshold above which the insulating monoclinic state \((\eta = +1\) , \(\epsilon \neq 0\) ) becomes stable. The description of the electric- field induced transition is based on the observation [18] that the electric field directly couples to the electronic bandstructure of a Mott insulator, making the metallic phase more stable. The transition can thus be described assuming that \(\epsilon_{t}^{2}(V)\) increases with increasing \(V\) . The energy difference between the insulating and metallic phase can be expressed as \(\Delta (\mathbf{r}, V) = F[- 1] - F[+1] \simeq g \left[ \epsilon^{2}(\mathbf{r}) - \epsilon_{t}^{2}(V) \right]\) . If we start from the insulating phase with \(\epsilon^{2}(\mathbf{r}) > \epsilon_{t}^{2}(V) = 0\) , the IMT takes place when \(V\) is increased up to the point
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[220, 100, 797, 573]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[84, 585, 919, 671]]<|/det|>
+<center>FIG. 4. a) I-V curve measured during the XLD-PEEM imaging. The sudden drop in the voltage measured at \(I_{th} = 1.05 \mathrm{mA}\) indicates the first resistive switch. b) Line profiles of the XLD-PEEM images in Fig. 2. The grey shaded area indicates the progressive widening of the metallic rhombohedral filament. The direction of the line profiles is shown by the white dashed line in the XLD-PEEM image on top, where we report a detail of Fig. 2j. c) Width, \(d\) , of the metallic filament as a function of current. The blue/red markers represent the values of \(d\) obtained from the XLD-PEEM images below/above \(I_{th}\) . The green solid line shows an estimate of \(d\) , derived from a parallel resistors model predicting a sudden jump of \(d\) to \(200 \mathrm{mA}\) in \(I_{th}\) (see Supplementary Information Section S5). </center>  
+
+<|ref|>text<|/ref|><|det|>[[85, 696, 487, 737]]<|/det|>
+that \(\Delta (\mathbf{r}, V) = 0\) . A topological defect, which locally suppresses \(\epsilon^2 (\mathbf{r})\) , thus acts as a seed with a lower threshold compared to the rest of the system.  
+
+<|ref|>text<|/ref|><|det|>[[85, 745, 487, 877]]<|/det|>
+Intriguingly, we also note from Eq. 4 that the IMT can take place at a non- zero value of \(\epsilon^2 (\mathbf{r})\) , which allows the formation of a non- thermal metallic state \((\eta = - 1)\) with a finite monoclinic distortion \((\epsilon^2 (\mathbf{r}) \lesssim \epsilon_t^2 (V))\) , as already observed in non- equilibrium optical experiments [37, 44]. The nature of the early- stage switching process can be inferred by a direct comparison between the electrical state of the device and the melting of the monoclinic domains. The \(I - V\) curve of the device, as measured in- situ during the PEEM imaging, is plotted in Fig. 4a).  
+
+<|ref|>text<|/ref|><|det|>[[515, 697, 917, 867]]<|/det|>
+XLD- PEEM images were recorded at specific values of \(I\) . The \(I - V\) plot shows that the first resistive switching event occurs at the threshold current \(I_{th} = 1.05 \mathrm{mA}\) . In Figure 4b) we report a linecut of the XLD- PEEM image acquired at specific values of \(I\) ; the image profile is taken along a line crossing the monoclinic domains in the middle of the device gap (see white solid line in Fig. 4, top panel). For large currents running through the device, the line profile in Fig. 4b) displays a flat region, which indicates the melting of the monoclinic nanodomains due to the formation of the rhombohedral metallic channel. As highlighted by the grey area in Fig. 4b), the width \(d\) of the metallic filament increases with the current, from
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 102, 488, 128]]<|/det|>
+\(d = 0.23\pm 0.05\mu \mathrm{m}\) at \(I = 1.5\mathrm{mA}\) to \(d = 3.7\pm 0.2\mu \mathrm{m}\) at \(I = 10\mathrm{mA}\)  
+
+<|ref|>text<|/ref|><|det|>[[85, 135, 488, 757]]<|/det|>
+Modelling the device as a circuit with two parallel resistors (see Supplementary Information Section S5) allows an estimation of \(d\) of the rhombohedral filament corresponding to the observed voltage drop. For large currents running through the device, the experimentally determined values of \(d\) match well with those predicted for a metallic channel forming in the gap, which has the resistivity of the high- temperature rhombohedral phase, as shown in Fig. 4. However, in correspondence of the first resistive switching event at \(I_{th} = 1.05\mathrm{mA}\) , the model predicts the sudden formation of a \(\sim 200\mathrm{nm}\) wide metallic rhombohedral filament, which is not visible in the XLD- PEEM images (see Fig. 4c and Supplementary Information Fig. S5), despite being well above the experimental resolution of the microscope. To explain this discrepancy, one might suspect that a rhombohedral metallic filament forms below the surface of the \(\mathrm{V}_2\mathrm{O}_3\) film, where it is not detected by PEEM which has a surface sensitivity limited to the first few nanometers. In fact, two arguments act against this possibility: i) the presence of the \(\mathrm{Cr}_2\mathrm{O}_3\) buffer layer reduces the substrate- film lattice mismatch from \(4.2\%\) to \(0.1\%\) , thus almost entirely removing the residual epitaxial strain in the film [42], which is known to suppress the monoclinic phase and favour interfacial metallicity [42, 45]. In contrast to highly- strained films, in which the metal to insulator resistivity jump is strongly suppressed [45], the films in the present study display the 5- order of magnitude resistivity change typical of the unstrained metal- to- insulator transition (see Fig. S1); ii) the curl- free conditions force the interface between monoclinic and rhombohedral metallic regions to be oriented perpendicularly to the order parameter of the monoclinic domain. The formation of a sub- surface metallic layer would lead to a sharp \((\ll 20\mathrm{nm})\) monoclinic- rhombohedral interface parallel to \(\vec{\epsilon}\) , thus leading to a dramatic increase of the strain energy of the system. Our results are compatible with a complex scenario in which the topology- driven resistive switching likely occurs via the sudden transformation of a single \(200\mathrm{nm}\) wide insulating monoclinic domain into a metallic channel with a non- thermal monoclinic lattice structure. At a second stage, the Joule heating leads to the thermally driven monoclinic- to- rhombohedral structural transition and the formation of rhombohedral metallic channels perpendicular to both the metallic electrodes and the \(\hat{\epsilon}_2\) order parameter direction, as observed in Fig. 2.  
+
+<|ref|>text<|/ref|><|det|>[[86, 763, 488, 870]]<|/det|>
+The X- ray- based nanoimaging of a Mott device under operating conditions allowed us to simultaneously capture the formation of nanoscale conductive paths and the topology of the underlying symmetry- broken nanotexture. The present results expand our knowledge of the resistive switching process in Mott materials by demonstrating the leading role of inherent topological defects in initiating the avalanche process. The  
+
+<|ref|>text<|/ref|><|det|>[[515, 103, 917, 487]]<|/det|>
+methodologies used in this work imply that nanoscale strain engineering approaches could unlock a gate to manipulating topological defects and controlling the electronic switching dynamics in real devices, such as Mott- transition- based RRAM [46, 47], Mott memristor [48- 50] and artificial neurons [51, 52]. The concept of topology- driven resistive switching will be key to assessing the possible non- thermal nature of the early stage electronic phase [37] as well as the microscopic origin of memory and non- volatile effects recently observed in Mott devices [6]. We note that the relation between topological defects and electronic phase transitions established here is a general concept, potentially extendable to other systems that undergo first- order phase transitions accompanied by a symmetry breaking, as described by the energy functional (4). Relevant examples embrace transition- metal oxides [3, 53], such as vanadates, nickelates and manganites, and layered materials, such as \(1T\) - TaS2 [54- 57], in which the IMT is accompanied by charge-, lattice- and orbital- ordered states with reduced symmetry. Further platforms include cuprate superconductors [58] and kagome metals [59] in which light- or magnetic- induced discontinuous electronic transitions coexist with charge- order. Topological defects in the order parameter therefore provide a framework for understanding non- equilibrium electronic phase transitions, allowing all- optical control of hidden states of matter in a broad class of quantum materials [57, 60- 64].  
+
+<|ref|>text<|/ref|><|det|>[[515, 500, 917, 870]]<|/det|>
+We thank Diamond Lights Source for the provision of beamtime under proposal numbers MM- 27218, MM- 31711 and MM- 34455. We thank Manuel R. Osorio and Fernando J. Urbanos for the fabrication of sample electrodes at the Centre for Micro and Nanofabrication of IMDEA Nanociencia. A.M., S.M. and C.G. acknowledge financial support from MIUR through the PRIN 2015 (Prot. 2015C5SEJJ001) and PRIN 2017 (Prot. 20172H2SC4.005) programs and from the European Union - Next Generation EU through the MUR- PRIN2022 (Prot. 20228YCYY7) program. C.G. acknowledges support from Università Cattolica del Sacro Cuore through D.I, D.2.2 and D.3.1 grants. S.M. acknowledges partial financial support through the grant "Finanziamenti ponte per bandi esterni" from Università Cattolica del Sacro Cuore. I.F.C. and M.M. acknowledge support from the "Severo Ochoa" Programme for Centres of Excellence in R&D (CEX2020- 001039- S) and the Spanish AEI- MCIN PID2021- 122980OB- C52 (ECoSOC- ECLIPSE). I.F.C holds a FPI fellowship from the Spanish AEI- MCIN (PRE2020- 092625). W.- F.H., S.M., J.W.S. and J.- P.L. acknowledge financial support by the KU Leuven Research Funds Project No. C14/21/083, iBOF/21/084, KAC24/18/056 and C14/17/080, as well as the FWO AKUL/13/19 and AKUL/19/023, and the Research Funds of the INTERREG- E- TEST Project (EMR113) and INTERREG- VL- VL- PATHFINDER Project (0559).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 135, 490, 875]]<|/det|>
+[1] Z. Wang, H. Wu, G. W. Burr, C. S. Hwang, K. L. Wang, Q. Xia, and J. J. Yang, Resistive switching materials for information processing, Nature Reviews Materials 5, 173 (2020).[2] Y. Zhou and S. Ramanathan, Mott Memory and Neuromorphic Devices, Proceedings of the IEEE 103, 1289 (2015).[3] Y. Tokura, M. Kawasaki, and N. Nagaosa, Emergent functions of quantum materials, Nature Physics 13, 1056 (2017).[4] J. del Valle, J. G. Ramirez, M. J. Rozenberg, and I. K. Schuller, Challenges in materials and devices for resistive- switching- based neuromorphic computing, Journal of Applied Physics 124, 211101 (2018).[5] C. Adda, B. Corraze, P. Stoliar, P. Diener, J. Tranchant, A. Filatre- Furcate, M. Fourmigue, D. Lorcy, M.- P. Besland, E. Janod, and L. Cario, Mott insulators: A large class of materials for Leaky Integrate and Fire (LIF) artificial neuron, Journal of Applied Physics 124, 152124 (2018).[6] J. del Valle, J. G. Ramirez, M. J. Rozenberg, and I. K. Schuller, Subthreshold firing in Mott nanodevices, Nature 569, 388 (2019).[7] A. Pérez- Tomás, Functional Oxides for Photoneuromorphic Engineering: Toward a Solar Brain, Advanced Materials Interfaces 6, 1900471 (2019).[8] J. del Valle, P. Salev, Y. Kalcheim, and I. K. Schuller, A caloritronics- based Mott neuristor, Scientific Reports 10, 4292 (2020).[9] Z. Zhang, S. Mondal, S. Mandal, J. M. Allred, N. A. Aghamiri, A. Fali, Z. Zhang, H. Zhou, H. Cao, F. Rodolakis, J. L. McChesney, Q. Wang, Y. Sun, Y. Abate, K. Roy, K. M. Rabe, and S. Ramanathan, Neuromorphic learning with Mott insulator NiO, Proceedings of the National Academy of Sciences 118, e2017239118 (2021).[10] X. Deng, S.- Q. Wang, Y.- X. Liu, N. Zhong, Y.- H. He, H. Peng, P.- H. Xiang, and C.- G. Duan, A Flexible Mott Synaptic Transistor for Nociceptor Simulation and Neuromorphic Computing, Advanced Functional Materials 31, 2101099 (2021).[11] S. Deng, H. Yu, T. J. Park, A. N. Islam, S. Manna, A. Pofelski, Q. Wang, Y. Zhu, S. K. Sankaranarayanan, A. Sengupta, and S. Ramanathan, Selective area doping for Mott neuromorphic electronics, Science Advances 9, eade4838 (2023).[12] Y. Ran, Y. Pei, Z. Zhou, H. Wang, Y. Sun, Z. Wang, M. Hao, J. Zhao, J. Chen, and X. Yan, A review of Mott insulator in memristors: The materials, characteristics, applications for future computing systems and neuromorphic computing, Nano Research 16, 1165 (2023).[13] Z. Yang, C. Ko, and S. Ramanathan, Oxide electronics utilizing ultrafast metal- insulator transitions, Annual Review of Materials Research 41, 337 (2011).[14] A. Mehonic and A. J. Kenyon, Brain- inspired computing needs a master plan, Nature 604, 255 (2022).[15] W.- F. Hsu, S. Mellaerts, C. Bellani, P. Homm, N. Uchida, M. Menghini, M. Houssa, J. W. Seo, and J.- P. Locquet, Raman spectroscopy and phonon dynamics in strained \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review Materials 7, 074606 (2023).[16] J. Zhang and R. D. Averitt, Dynamics and control in complex transition metal oxides, Annual Review of Materials Research 44, 19 (2014).[17] D. Basov, R. Averitt, and D. Hsieh, Towards properties on demand in quantum materials, Nature materials 16, 1077 (2017).[18] G. Mazza, A. Amaricci, M. Capone, and M. Fabrizio, Field- driven mott gap collapse and resistive switch in correlated insulators, Physical review letters 117, 176401 (2016).[19] C. T. Suen, I. Markovic, M. Zonno, S. Zhdanovich, N.- H. Jo, M. Schmid, P. Hansmann, P. Pupuhal, K. Fursich, V. Zimmerman, et al., Nature of the current- induced insulator- to- metal transition in \(\mathrm{Ca}_2\mathrm{RuO}_4\) as revealed by transport- ARPES, arXiv:2308.05803 (2023).[20] D. J. Wouters, S. Menzel, J. A. J. Rupp, T. Hennen, and R. Waser, On the universality of the I- V switching characteristics in non- volatile and volatile resistive switching oxides, Faraday Discuss. 213, 183 (2019).[21] Y. Kalcheim, A. Camjayi, J. del Valle, P. Salev, M. Rozenberg, and I. K. Schuller, Non- thermal resistive switching in Mott insulator nanowires, Nature Communications 11, 2985 (2020).[22] P. Stoliar, L. Cario, E. Janod, B. Corraze, C. Guillot- Deudon, S. Salmon- Bourmand, V. Guiot, J. Tranchant, and M. Rozenberg, Universal Electric- Field- Driven Resistive Transition in Narrow- Gap Mott Insulators, Advanced Materials 25, 3222 (2013).[23] V. Guiot, L. Cario, E. Janod, B. Corraze, V. Ta Phuoc, M. Rozenberg, P. Stoliar, T. Cren, and D. Roditchev, Avalanche breakdown in \(\mathrm{CaTa}_4\mathrm{Se}_8\mathrm{- }x\mathrm{Te}_x\) narrow- gap Mott insulators, Nature communications 4, 1722 (2013).[24] F. Nakamura, M. Sakaki, Y. Yamanaka, S. Tamaru, T. Suzuki, and Y. Maeno, Electric- field- induced metal maintained by current of the Mott insulator \(\mathrm{Ca}_2\mathrm{RuO}_4\) , Scientific reports 3, 2536 (2013).[25] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Origin of hysteresis in resistive switching in magnetite is Joule heating, Physical Review B 79, 245131 (2009).[26] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Statistical distribution of the electric field- driven switching of the Verwey state in \(\mathrm{Fe}_3\mathrm{O}_4\) , New Journal of Physics 14, 013019 (2012).[27] E. Janod, J. Tranchant, B. Corraze, M. Querre, P. Stoliar, M. Rozenberg, T. Cren, D. Roditchev, V. T. Phuoc, M.- P. Besland, et al., Resistive switching in Mott insulators and correlated systems, Advanced Functional Materials 25, 6287 (2015).[28] R. Waser and M. Aono, Nanoionics- based resistive switching memories, Nature materials 6, 833 (2007).[29] D. Ielmini, Resistive switching memories based on metal oxides: mechanisms, reliability and scaling, Semiconductor Science and Technology 31, 063002 (2016).[30] J. S. Lee, S. Lee, and T. W. Noh, Resistive switching phenomena: A review of statistical physics approaches, Applied Physics Reviews 2, 031303 (2015).[31] P. Stoliar, J. Tranchant, B. Corraze, E. Janod, M.- P. Besland, F. Tesler, M. Rozenberg, and L. Cario, A Leaky- Integrate- and- Fire Neuron Analog Realized with a Mott Insulator, Advanced Functional Materials 27, 1604740 (2017).[32] M. Lange, S. Guénon, Y. Kalcheim, T. Luibrand, N. M.  
+
+<|ref|>text<|/ref|><|det|>[[513, 135, 920, 875]]<|/det|>
+complex transition metal oxides, Annual Review of Materials Research 44, 19 (2014).[17] D. Basov, R. Averitt, and D. Hsieh, Towards properties on demand in quantum materials, Nature materials 16, 1077 (2017).[18] G. Mazza, A. Amaricci, M. Capone, and M. Fabrizio, Field- driven mott gap collapse and resistive switch in correlated insulators, Physical review letters 117, 176401 (2016).[19] C. T. Suen, I. Markovic, M. Zonno, S. Zhdanovich, N.- H. Jo, M. Schmid, P. Hansmann, P. Pupuhal, K. Fursich, V. Zimmerman, et al., Nature of the current- induced insulator- to- metal transition in \(\mathrm{Ca}_2\mathrm{RuO}_4\) as revealed by transport- ARPES, arXiv:2308.05803 (2023).[20] D. J. Wouters, S. Menzel, J. A. J. Rupp, T. Hennen, and R. Waser, On the universality of the I- V switching characteristics in non- volatile and volatile resistive switching oxides, Faraday Discuss. 213, 183 (2019).[21] Y. Kalcheim, A. Camjayi, J. del Valle, P. Salev, M. Rozenberg, and I. K. Schuller, Non- thermal resistive switching in Mott insulator nanowires, Nature Communications 11, 2985 (2020).[22] P. Stoliar, L. Cario, E. Janod, B. Corraze, C. Guillot- Deudon, S. Salmon- Bourmand, V. Guiot, J. Tranchant, and M. Rozenberg, Universal Electric- Field- Driven Resistive Transition in Narrow- Gap Mott Insulators, Advanced Materials 25, 3222 (2013).[23] V. Guiot, L. Cario, E. Janod, B. Corraze, V. Ta Phuoc, M. Rozenberg, P. Stoliar, T. Cren, and D. Roditchev, Avalanche breakdown in \(\mathrm{CaTa}_4\mathrm{Se}_8\mathrm{- }x\mathrm{Te}_x\) narrow- gap Mott insulators, Nature communications 4, 1722 (2013).[24] F. Nakamura, M. Sakaki, Y. Yamanaka, S. Tamaru, T. Suzuki, and Y. Maeno, Electric- field- induced metal maintained by current of the Mott insulator \(\mathrm{Ca}_2\mathrm{RuO}_4\) , Scientific reports 3, 2536 (2013).[25] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Origin of hysteresis in resistive switching in magnetite is Joule heating, Physical Review B 79, 245131 (2009).[26] A. Fursina, R. Sofin, I. Shvets, and D. Natelson, Statistical distribution of the electric field- driven switching of the Verwey state in \(\mathrm{Fe}_3\mathrm{O}_4\) , New Journal of Physics 14, 013019 (2012).[27] E. Janod, J. Tranchant, B. Corraze, M. Querre, P. Stoliar, M. Rozenberg, T. Cren, D. Roditchev, V. T. Phuoc, M.- P. Besland, et al., Resistive switching in Mott insulators and correlated systems, Advanced Functional Materials 25, 6287 (2015).[28] R. Waser and M. Aono, Nanoionics- based resistive switching memories, Nature materials 6, 833 (2007).[29] D. Ielmini, Resistive switching memories based on metal oxides: mechanisms, reliability and scaling, Semiconductor Science and Technology 31, 063002 (2016).[30] J. S. Lee, S. Lee, and T. W. Noh, Resistive switching phenomena: A review of statistical physics approaches, Applied Physics Reviews 2, 031303 (2015).[31] P. Stoliar, J. Tranchant, B. Corraze, E. Janod, M.- P. Besland, F. Tesler, M. Rozenberg, and L. Cario, A Leaky- Integrate- and- Fire Neuron Analog Realized with a Mott Insulator, Advanced Functional Materials 27, 1604740 (2017).[32] M. Lange, S. Guénon, Y. Kalcheim, T. Luibrand, N. M.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[85, 100, 490, 878]]<|/det|>
+Vargas, D. Schwebius, R. Kleiner, I. K. Schuller, and D. Koelle, Imaging of electrothermal filament formation in a mott insulator, Physical Review Applied 16, 054027 (2021). [33] J. Del Valle, N. M. Vargas, R. Rocco, P. Salev, Y. Kalcheim, P. N. Lapa, C. Adda, M.- H. Lee, P. Y. Wang, L. Fratino, et al., Spatiotemporal characterization of the field- induced insulator- to- metal transition, Science 373, 907 (2021). [34] D. Babich, J. Tranchant, C. Adda, B. Corraze, M.- P. Besland, P. Warnicke, D. Bedau, B. Bertoncini, J.- Y. Mevellec, B. Humbert, J. Rupp, T. Hennen, D. Wouters, R. Llopis, L. Cario, and E. Janod, Lattice contraction induced by resistive switching in chromium- doped \(\mathrm{V}_2\mathrm{O}_3\) : a hallmark of Mott physics, arXiv preprint arXiv:2105.05093 (2021). [35] T. Luibrand, A. Bercher, R. Rocco, Tahouni- Bonab, L. Varbaro, C. Rischau, and e. a. Dominguez, Characteristic lengthscales of the electrically- induced insulator- to- metal transition, arXiv preprint arXiv:2301.00456 (2023). [36] A. Ronchi, P. Homm, M. Menghini, P. Franceschini, F. Maccherozzi, F. Banfi, G. Ferrini, F. Cilento, F. Parmigiani, S. S. Dhesi, et al., Early- stage dynamics of metallic droplets embedded in the nanotextured Mott insulating phase of \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review B 100, 075111 (2019). [37] A. Ronchi, P. Franceschini, A. De Poli, P. Homm, A. Fitzpatrick, F. Maccherozzi, G. Ferrini, F. Banfi, S. S. Dhesi, M. Menghini, et al., Nanoscale self- organization and metastable non- thermal metallicity in Mott insulators, Nature Communications 13, 3730 (2022). [38] D. McWhan, T. Rice, and J. Remeika, Mott transition in Cr- doped \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review Letters 23, 1384 (1969). [39] D. McWhan and J. Remeika, Metal- insulator transition in \((\mathrm{V}_{1 - x}\mathrm{Cr}_x)_2\mathrm{O}_3\) , Physical Review B 2, 3734 (1970). [40] D. McWhan, A. Menth, J. Remeika, W. Brinkman, and T. Rice, Metal- insulator transitions in pure and doped \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review B 7, 1920 (1973). [41] S. Guénon, S. Scharinger, S. Wang, J. Ramírez, D. Koelle, R. Kleiner, and I. K. Schuller, Electrical breakdown in a \(\mathrm{V}_2\mathrm{O}_3\) device at the insulator- to- metal transition, Europhysics Letters 101, 57003 (2013). [42] L. Dillemans, T. Smets, R. R. Lieten, M. Menghini, C.- Y. Su, and J.- P. Locquet, Evidence of the metal- insulator transition in ultrathin unstrained \(\mathrm{V}_2\mathrm{O}_3\) thin films, Applied Physics Letters 104, 071902 (2014). [43] J.- H. Park, L. Tjeng, A. Tanaka, J. Allen, C. Chen, P. Metcalf, J. Honig, F. De Groot, and G. Sawatzky, Spin and orbital occupation and phase transitions in \(\mathrm{V}_2\mathrm{O}_3\) , Physical Review B 61, 11506 (2000). [44] A. Ronchi, P. Franceschini, P. Homm, M. Gandolfi, G. Ferrini, S. Pagliara, F. Banfi, M. Menghini, C. Giannetti, et al., Light- assisted resistance collapse in a \(\mathrm{V}_2\mathrm{O}_3\) - based mott- insulator device, Physical Review Applied 15, 044023 (2021). [45] A. Pofelski, S. Valencia, Y. Kalcheim, P. Salev, A. Rivera, C. Huang, M. A. Mawass, F. Kronast, I. K. Schuller, Y. Zhu, et al., Domain nucleation across the metal- insulator transition of self- strained \(\mathrm{V}_2\mathrm{O}_3\) films, arXiv preprint arXiv:2312.09051 (2023). [46] A. Sawa, Resistive switching in transition metal oxides, Materials Today 11, 28 (2008). [47] Y. Wang, K.- M. Kang, M. Kim, H.- S. Lee, R. Waser,  
+
+<|ref|>text<|/ref|><|det|>[[515, 102, 920, 878]]<|/det|>
+D. Wouters, R. Dittmann, J. J. Yang, and H.- H. Park, Mott-transition-based RRAM, Materials Today 28, 63 (2019). [48] M. D. Pickett, G. Medeiros-Ribeiro, and R. S. Williams, A scalable neuristor built with Mott memristors, Nature Materials 12, 114 (2013). [49] M. Yoshida, R. Suzuki, Y. Zhang, M. Nakano, and Y. Iwasa, Memristive phase switching in two-dimensional \(1\mathrm{T} - \mathrm{TaS}_2\) crystals, Science Advances 1, e1500606 (2015). [50] S. Kumar, J. P. Strachan, and R. S. Williams, Chaotic dynamics in nanoscale \(\mathrm{NbO}_2\) Mott memristors for analogue computing, Nature 548, 318 (2017). [51] F. Tesler, C. Adda, J. Tranchant, B. Corraze, E. Janod, L. Cario, P. Stoliar, and M. Rozenberg, Relaxation of a spiking Mott artificial neuron, Phys. Rev. Appl. 10, 054001 (2018). [52] X. Zhang, Y. Zhuo, Q. Luo, Z. Wu, R. Midya, Z. Wang, W. Song, R. Wang, N. K. Upadhyay, Y. Fang, F. Kiani, M. Rao, Y. Yang, Q. Xia, Q. Liu, M. Liu, and J. J. Yang, An artificial spiking afferent nerve based on Mott memristors for neurorobotics, Nature Communications 11, 51 (2020). [53] M. Imada, A. Fujimori, and Y. Tokura, Metal-insulator transitions, Rev. Mod. Phys. 70, 1039 (1998). [54] I. Vaskivskyi, J. Gospodaric, S. Brazovskii, D. Svetin, P. Sutar, E. Goreshnik, I. A. Mihailovic, T. Mertelj, and D. Mihailovic, Controlling the metal- to- insulator relaxation of the metastable hidden quantum state in \(1\mathrm{T} - \mathrm{TaS}_2\) , Science Advances 1, e1500168 (2015). [55] M. J. Hollander, Y. Liu, W.- J. Lu, L.- J. Li, Y.- P. Sun, J. A. Robinson, and S. Datta, Electrically driven reversible insulator- metal phase transition in \(1\mathrm{T} - \mathrm{TaS}_2\) , Nano Letters 15, 1861 (2015). [56] S.- H. Lee, J. S. Goh, and D. Cho, Origin of the insulating phase and first- order metal- insulator transition in \(1\mathrm{T} - \mathrm{TaS}_2\) , Phys. Rev. Lett. 122, 106404 (2019). [57] F. Y. Gao, Z. Zhang, Z. Sun, L. Ye, Y.- H. Cheng, Z.- J. Liu, J. G. Checkelsky, E. Baldini, and K. A. Nelson, Snapshots of a light- induced metastable hidden phase driven by the collapse of charge order, Science Advances 8, eabp9076 (2022). [58] B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, and J. Zaanen, From quantum matter to high- temperature superconductivity in copper oxides, Nature 518, 179 (2015). [59] T. Asaba, A. Onishi, Y. Kageyama, T. Kiyosue, K. Ohtsuka, S. Suetsugu, Y. Kohsaka, T. Gaggl, Y. Kasahara, H. Murayama, K. Hashimoto, R. Tazai, H. Kontani, B. R. Ortiz, S. D. Wilson, Q. Li, H. H. Wen, T. Shibauchi, and Y. Matsuda, Evidence for an odd- parity nematic phase above the charge- density- wave transition in a kagome metal, Nature Physics 10.1038/s41567- 023- 02272- 4 (2024). [60] L. Stojchevska, I. Vaskivskyi, T. Mertelj, P. Kusar, D. Svetin, S. Brazovskii, and D. Mihailovic, Ultrafast switching to a stable hidden quantum state in an electronic crystal, Science 344, 177 (2014). [61] J. Zhang, X. Tan, M. Liu, S. W. Teitelbaum, K. W. Post, F. Jin, K. Nelson, D. N. Basov, W. Wu, and R. D. Averitt, Cooperative photoinduced metastable phase control in strained manganite films, Nature Materials 15, 956 (2016). [62] S. Wandel, F. Boschini, E. H. da Silva Neto, L. Shen, M. X. Na, S. Zohar, Y. Wang, S. B. Welch, M. H.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[90, 102, 488, 247]]<|/det|>
+Seaberg, J. D. Koralek, G. L. Dakovski, W. Hettel, M.- F. Lin, S. P. Moeller, W. F. Schlotter, A. H. Reid, M. P. Minitti, T. Boyle, F. He, R. Sutarto, R. Liang, D. Bonn, W. Hardy, R. A. Kaindl, D. G. Hawthorn, J.- S. Lee, A. F. Kemper, A. Damascelli, C. Giannetti, J. J. Turner, and G. Coslovich, Enhanced charge density wave coherence in a light- quenched, high- temperature superconductor, Science 376, 860 (2022). [63] Y. Cheng, A. Zong, L. Wu, Q. Meng, W. Xia, F. Qi, P. Zhu, X. Zou, T. Jiang, Y. Guo, J. van Wezel, A. Kogar, M. W. Zuerch, J. Zhang, Y. Zhu, and D. Xiang, Ultrafast formation of topological defects in a two- dimensional  
+
+<|ref|>text<|/ref|><|det|>[[515, 103, 919, 234]]<|/det|>
+[64] A. Verma, D. Golež, O. Y. Gorobtsov, K. Kaj, R. Russell, J. Z. Kaaret, E. Lamb, G. Khalsa, H. P. Nair, Y. Sun, R. Bouck, N. Schreiber, J. P. Ruf, V. Ramaprasad, Y. Kubota, T. Togashi, V. A. Stoica, H. Padmanabhan, J. W. Freeland, N. A. Benedek, O. G. Shpyrko, J. W. Harter, R. D. Averitt, D. G. Schlom, K. M. Shen, A. J. Millis, and A. Singer, Picosecond volume expansion drives a later- time insulator- metal transition in a nano- textured Mott insulator, Nature Physics 10.1038/s41567- 024- 02396- 1 (2024).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[43, 42, 312, 70]]<|/det|>
+## Supplementary Files  
+
+<|ref|>text<|/ref|><|det|>[[43, 92, 768, 113]]<|/det|>
+This is a list of supplementary files associated with this preprint. Click to download.  
+
+<|ref|>text<|/ref|><|det|>[[60, 130, 348, 150]]<|/det|>
+SIV203ResistiveSwitching.pdf
+
+<--- Page Split --->

preprint/preprint__0a0dcde223c0f68e72f0ff32ff701bc4d28087fec6db01b2248e07b20064711f/images_list.json

@@ -0,0 +1,25 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 | Experimental approach. a, Saccade countermanding task. Monkeys initiated trials by fixating a central point. After a variable time, the center of the fixation point was extinguished, and a peripheral target was presented at one of two possible locations. On no stop-signal trials monkeys were required to shift gaze to the target, whereupon after \\(600\\pm 0\\) ms a high-pitch auditory feedback tone was delivered, and \\(600\\pm 0\\) ms later fluid reward was provided. On stop-signal trials ( \\(\\sim 40\\%\\) of trials) after the target appeared, the center of the fixation point was re-illuminated after a variable stop-signal delay, which instructed the monkey to cancel the saccade in which case the same high-pitch tone was presented \\(1,500\\pm 0\\) ms after target presentation followed \\(600\\pm 0\\) ms later by fluid reward. Stop-signal delay was adjusted such that monkeys successfully canceled the saccade in \\(\\sim 50\\%\\) of trials. In the remaining trials, monkeys made non-canceled errors, which were followed after \\(600\\pm 0\\) ms by a low-pitch tone, and no reward was delivered. Monkeys could not initiate trials earlier after errors. b, Grand average cumulative distributions of all RT for both monkeys on trials with no stop-signal (solid) and non-canceled errors (dashed). c, Grand average probability of non-canceled errors (P(error)) as a function of stop-signal delay. Inset shows the distribution of SSRT across all",
+    "footnote": [],
+    "bbox": [
+      [
+        327,
+        94,
+        720,
+        664
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Fig 6 | Extrinsic and intrinsic circuitry for executive control. The laminar distribution observed for Conflict (orange), Event Timing (dark blue), and Goal Maintenance (dark red) are summarized with selected anatomical connections based on published studies. Sampled neurons were likely broad spike pyramidal and narrow spike, possibly inhibitory neurons. The laminar densities of calretinin (CR), calbindin (CB), and parvalbumin (PV) neurons observed and of D1 and D2 receptors are indicated on the far right. Left, Conflict signal can arise in SEF through afferents from frontal eye field (FEF). SEF can receive coincident inputs from Fixation neurons (STOP) and Movement neurons (GO) in FEF, directly, or in SC, indirectly, via thalamus, terminating in L2/3. These inputs are integrated within the synapses of L2/3 and L5 Conflict neurons. Intracortical processing produces later activation of Conflict neurons in L6 which can relay this signal to the Thalamus. Right, Top: Schematic of the activity profile for Goal Maintenance and Event Timing neurons in distinct phases indicated by the number. We conjecture that Goal Maintenance neurons, mainly located in L2/3, suppress unwanted movement through push-pull basal ganglia circuitry with pyramidal neurons directly projecting to the indirect pathway (D2) and inhibitory neurons, inhibiting pyramidal neurons that can project to the direct (D1) pathway. The gray symbol indicates that these neurons are distinct from those reported in this study. Input from dorsolateral prefrontal cortex (DLPFC) and anterior cingulate cortex (ACC), terminating in L2/3 can inform SEF of the anticipated reward association based on the experienced stop-signal delay contingent on successful stopping. Dopamine (DA) neuron projections in L2/3 from the SNpc and VTA can also relay this information. These inputs can result in the phasic response in Goal Maintenance neurons (phase 1, red). Following the phasic response, activity can remain elevated via recurrent connections and balance of excitation and inhibition (phase 2, red). The auditory feedback tone, integrated with the task rule from DLPFC cues the termination of operant control on behavior, resulting in the inhibition of pyramidal and inter-neurons by CR and CB neurons. This results in the termination of the sustained activity (phase 3). Event Timing neurons can receive input from DLPFC and ACC terminating in L2/3 informing neurons in L2/3 and L5 about an upcoming event. Ramping results from recurrent connections (1, dark blue). SEF can receive information about stop-signal appearance and successful stopping from ventrolateral prefrontal cortex (VLPFC) and DLPFC and Conflict neurons within the microcircuitry. This information can suppress the ramping activity via inhibitory",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 9
+  }
+]