Datasets:

---

DataSeer

/

das-extraction-md-labels

like 0

Follow

DataSeer Research Data Services Ltd 3

ERROR: type should be string, got "https://doi.org/10.1038/s41590-023-01504-2\n\nConserved transcriptional connectivity of\nregulatory T cells in the tumor microenvi-\nronment informs new combination cancer\ntherapy strategies\n\nReceived: 25 March 2022\n\nAccepted: 5 April 2023\n\nPublished online: 1 May 2023\n\n Check for updates\n\n  2, Jesse A. Green1, Sham Rampersaud\n\nAriella Glasner1,10, Samuel A. Rose2,10, Roshan Sharma2,10, Herman Gudjonson2,\n  1, Izabella K. Valdez1,\nTinyi Chu\nEmma S. Andretta1, Bahawar S. Dhillon1, Michail Schizas1, Stanislav Dikiy\nAlejandra Mendoza1, Wei Hu\n  1, Ojasvi Chaudhary\nTianhao Xu2, Linas Mazutis3, Gabrielle Rizzuto\nAlvaro Quintanal-Villalonga\nCharles M. Rudin\n\n  4,5,\n  6, Parvathy Manoj6, Elisa de Stanchina7,8,\n & Alexander Y. Rudensky\n\n  1, Zhong-Min Wang\n\n  6, Dana Pe’er\n\n  1,\n  2,\n\n  2,9\n\n  1,9\n\nWhile regulatory T (Treg) cells are traditionally viewed as professional\nsuppressors of antigen presenting cells and effector T cells in both\nautoimmunity and cancer, recent findings of distinct Treg cell functions in\ntissue maintenance suggest that their regulatory purview extends to a wider\nrange of cells and is broader than previously assumed. To elucidate tumoral\nTreg cell ‘connectivity’ to diverse tumor-supporting accessory cell types, we\nexplored immediate early changes in their single-cell transcriptomes upon\npunctual Treg cell depletion in experimental lung cancer and injury-induced\ninflammation. Before any notable T cell activation and inflammation,\nfibroblasts, endothelial and myeloid cells exhibited pronounced changes\nin their gene expression in both cancer and injury settings. Factor analysis\nrevealed shared Treg cell-dependent gene programs, foremost, prominent\nupregulation of VEGF and CCR2 signaling-related genes upon Treg cell\ndeprivation in either setting, as well as in Treg cell-poor versus Treg cell-rich\nhuman lung adenocarcinomas. Accordingly, punctual Treg cell depletion\ncombined with short-term VEGF blockade showed markedly improved\ncontrol of PD-1 blockade-resistant lung adenocarcinoma progression\nin mice compared to the corresponding monotherapies, highlighting a\npromising factor-based querying approach to elucidating new rational\ncombination treatments of solid organ cancers.\n\nDiverse stromal cell types found within the tumor microenvironment\n(TME) can support cancer initiation and progression by acting as acces-\nsory cells, yet their relationships and interdependencies remain poorly\nunderstood. Cells of the innate and adaptive immune system, when\n\nmobilized by immunotherapeutic agents, have been implicated in\nlimiting cancer progression, yet some of the very same cell types can\nsupport tumor growth either directly or indirectly by facilitating\ntumor-promoting functions of other accessory cell types. Treg cells,\n\nA full list of affiliations appears at the end of the paper.\n\n e-mail: peerd@mskcc.org; rudenska@mskcc.org\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1020\n\nnature immunologyArticle\fa\n\nKrasLSL-G12D/WTTrp53fl/flFoxp3GFP-DTR\n\nAd-Cre\n\nDT/\nPBS\n\nAnalysis\n\n0 wk\n\n18–20 wk\n\nb\n\ns\nl\nl\ne\nc\n+\n4\nD\nC\n\nf\no\ng\ne\nr\nT\n%\n\n40\n\n30\n\n20\n\n10\n\n0\n\n****\n\nCtrl\n\nDT\n\nc\n\ns\nl\nl\ne\nc\n+\n5\n4\nD\nC\n\nf\no\n+\n4\nD\nC\n%\n\n8\n\n6\n\n4\n\n2\n\n0\n\nNS\n\nCtrl\n\nDT\n\ne\nCtrl DAPI Foxp3 (GFP)\n\nDT DAPI Foxp3 (GFP)\n\nNS\n\nNS\n\nNS\n\nd\n\n)\ng\n(\n\nt\nh\ng\ne\nW\n\ni\n\n2.0\n\n1.5\n\n1.0\n\n0.5\n\n0\n\nCtrl\n\nDT\n\nTumor-free PBS\nTumor PBS\nTumor-free DT\nTumor DT\n\n20\n\n15\n\n10\n\n5\n\n0\n\ns\nl\nl\ne\nc\n+\n5\n4\nD\nC\n\nf\no\n+\n8\nD\nC\n%\n\nf\n\n200 µm\n\ng\nLYVE1 Foxp3 (GFP) GP38 DAPI\n\nCD11c Foxp3 (GFP) F4/80 Draq7\n\n200 µm\n\n)\n\n0\n0\n0\n,\n1\n×\n(\n\nr\ne\nb\nm\nu\nn\nG\nE\nD\n\n4\n\n3\n\n2\n\n1\n\n0\n\nh\n\n)\n\nm\nµ\n(\ne\nc\nn\na\nt\ns\ni\nD\n\n200\n\n150\n\n100\n\n50\n\n0\n\nUp\nDown\n\nVECFib\n\nNeu\nMac/DC\nLEC\n\nCD4\n\nCD8\n\n****\n\n****\n\n****\n\n-fib\nTreg\n\n-mac\n-LEC\nTreg\n\n-fib\nTreg\n\n-mac\n-LEC\nTreg\n\nTreg\n\nTreg\n\nTumor-free\nzone\n\nTumor\nnodule\n\nFig. 1 | Early transcriptional responses of principal accessory cell populations\nin the lung adenocarcinoma TME to Treg cell depletion. a, Schematic of the\nexperimental design. b,c, Quantification of Treg (CD4+Foxp3+) one-tailed unpaired\nt-test P = 12.87, d.f. = 7 ****P < 0.0001 and Tcon (TCRβ+CD4+ and TCRβ+CD8+)\ncell populations; left, one-tailed t-test P = 0.3799, d.f. = 7, not significant (NS)\nP = 0.3576; right, one-tailed t-test P= 0.1925, d.f. = 7, NS P = 0.4264, in tumor-\nbearing lungs 48 h after diphtheria toxin (DT) or PBS (Ctrl) administration.\nd, Quantification of lung weight in tumor-free and tumor-bearing mice 48 h after\nDT-induced Treg cell depletion. One-way analysis of variance (ANOVA) followed\nby Sidak’s multiple-comparisons test. Tumor-free PBS versus tumor-free DT,\nP = 0.004037, d.f. = 10 NS P > 0.9999; tumor PBS versus tumor DT, P = 0.7450,\nd.f. = 10, NS P = 0.9787. e, Representative IF staining of Foxp3+ cells in tumor-\nbearing lungs of Ctrl and DT-treated mice. f, Numbers of upregulated (red)\nor downregulated (blue) DEGs (P < 0.05) 48 h after DT or PBS administration\n\nidentified by bulk RNA-seq analysis of the indicated cell subsets. Fib, fibroblasts;\nNeu, neutrophils; Mac, macrophages; CD4 and CD8, effector CD4+ and CD8+\nT cells. g, Representative IF staining of the indicated cell types. h, Quantification\nof distances between Treg cells and the indicated cell types. One-way ANOVA,\nalpha = 0.05, followed by Tukey’s multiple-comparison test Treg-Fib tumor-free\nzone versus tumor nodule, q = 8.041, d.f. = 2544 ****P < 0.0001. Treg-LEC tumor-\nfree zone versus tumor nodule q = 10.08, d.f. = 2544, ****P < 0.0001, Treg-Mac\nversus tumor-free zone versus tumor nodule q = 17.79, d.f. = 2544, ****P < 0.0001.\nAt least 200 cells were counted in each comparison. Three independent sections\nper mouse were analyzed. Three and four mice were used in each group in two\nindependent experiments. Data are presented as the mean ± s.e.m. (b–d)\n(b and c) N = Ctrl-5, DT-4, (d) N = 3 tumor-free PBS, 3 tumor-free DT, 4 tumor PBS,\n4 tumor DT. Data are presented as the mean ± s.e.m.\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1021\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\n\n\fexpressing the transcription factor Foxp3, are highly enriched in\nhuman solid organ cancers and their experimental animal models, and\nat sites of inflammation and injury, where they exert both their essen-\ntial immunosuppressive function and distinct tissue repair-promoting\nmodalities1–4. Depletion of Treg cells results in restraint of tumor growth\nin numerous experimental cancer models5–9. Nevertheless, some\ntumors eventually progress after an initial response to Treg depletion5.\nThe latter can be due to waning functionality of effector T cells due to\nnegative regulation by co-receptors, foremost PD-1, expected to occur\nprimarily in PD-1 blockade-responsive tumors expressing PD-L1. An\nalternative, yet not mutually exclusive, explanation, is that Treg cell\ndepletion induces compensatory modulation of key accessory cell\ntypes in the TME, which may affect predominantly PD-1 nonresponsive\ncancers. Thus, early changes in diverse cellular components of the\nTME upon short-term Treg cell depletion may directly and indirectly\nimpact its overall effect on tumor growth. Thus, we sought to eluci-\ndate the interplay between Treg cells and other cellular components\nof the TME by investigating early changes in their features upon Treg\ndepletion in experimental cancer settings. Specifically, we wished to\nuse a genetically engineered mouse model that is characterized by\nnatural evolution of the TME, pronounced Treg cell presence, resist-\nance to PD-1 blockade and close resemblance to human disease.\nTherefore, we used KrasLSL-G12D/WTTrp53fl/fl mice harboring a Foxp3GFP-DTR\nallele (KP-DTR), in which intratracheal infection with a Cre-expressing\nreplication-deficient adenovirus induces lung adenocarcinoma (LuAd)\nformation. These mice offer a well-established model of non-small cell\nlung cancer (NSCLC) in humans, a disease where only some respond\nto PD-1/PD-L1 blockade-based therapies7,10–12. Our studies revealed\nthat Treg cells profoundly affect the transcriptional programs of key\naccessory cells including endothelial cells, fibroblasts, monocytes\nand macrophages in the TME. Moreover, these Treg cell dependencies\nof the transcriptional states of accessory cells are largely conserved\nin human lung cancer.\n\nResults\nEarly responses of tumor microenvironment cells to Treg cell\ndepletion\nTo enable temporally controlled Treg cell depletion in KP adenocarci-\nnomas, we generated KrasLSL-G12D/WTTrp53fl/flFoxp3GFP-DTR mice, in which\nall Treg cells express the diphtheria toxin receptor (DTR)13. We reasoned\nthat since Treg cells are typically found in the tumor margins, early com-\npensatory responses of key accessory cell types—tumor-associated\nfibroblasts, vascular endothelial cells (VECs) and lymphatic endothe-\nlial cells (LECs), and macrophages (Mac)—to Treg cell depletion likely\nprecede effects on the tumor growth. Because the expansion of acti-\nvated self-reactive T cells, observed 72–96 h after DT-mediated Treg\ncell depletion in Foxp3GFP-DTR mice, induces pronounced inflamma-\ntory responses13, we sought to minimize these confounding factors\nby analyzing early transcriptional responses of KP tumor cells, lung\n\nepithelial cells (ECs), VECs, LECs, macrophages and T cells 48 h follow-\ning DT administration to tumor-bearing KP-DTR mice (Fig. 1a,b and\nExtended Data Fig. 1a,e). As expected, pronounced local and systemic\nT cell activation and inflammation, typically elicited by an extended\nTreg cell depletion regimen, were not observed (Fig. 1c and Supplemen-\ntary Fig. 1c), and tumor volume was unaffected at this early time point\n(Fig. 1d) even though neutrophils were moderately increased (Extended\nData Fig. 1c,k). Highly efficient tumoral Treg cell depletion in situ was\nconfirmed by immunofluorescence (IF) microscopy of DT-treated as\ncompared to control (Ctrl) mice, in which Treg cells were found mainly at\nthe boundaries of tumor foci (Fig. 1e). Bulk RNA-sequencing (RNA-seq)\nanalyses of cell subsets purified by fluorescence-activated cell sort-\ning (FACS) from the lungs of DT-treated KP adenocarcinoma-bearing\nKP-DTR mice showed pronounced changes in gene expression in LECs,\nmacrophages and fibroblasts, while T cells, which are considered the\nmain targets of Treg cell suppression, changed the least (Fig. 1f and\nSupplementary Table 1). Among accessory cells, the most pronounced\ntranscriptional responses were observed in fibroblasts, endothelial\ncells and CD11c+ myeloid cells, highlighting Treg cell ‘connectivity’ to\nthese cell types in tumor-bearing lungs (Extended Data Fig. 1f–h and\nSupplementary Table 1). Importantly, DT-induced Treg cell ablation in\ntumor-free control KP-DTR mice resulted in minor, if any, changes in\ngene expression in all lung cell populations analyzed with the excep-\ntion of VECs (Extended Data Fig. 1j,k). This was consistent with the\npredominantly intravascular localization of Treg cells in the lung of\nunchallenged mice in contrast to their heavy presence in the cancerous\nlung parenchyma (Extended Data Fig. 1f,g)14. These results suggest that\nthe observed transcriptional changes in accessory cells in cancerous\nlungs are not due to a systemic response to Treg cell depletion. Next, we\ninvestigated whether shared groups of genes underwent modulation in\ndifferent accessory cell types and observed correlated gene expression\nchanges in endothelial cells and fibroblasts (Extended Data Fig. 1f, g).\nThese included programs related to endothelial-to-mesenchymal\ntransition (EndMT)-related genes (Id2, Itgav and Cxcl12), which were\npreviously shown to be modulated by Treg cells in the hair follicles15, and\ninflammation-related genes (Il6, Ccl5, Acacb, Ccl22, Arg1 and Tnfrsf18),\nwhose expression is affected by Treg cells in adipose tissue in the con-\ntext of metabolic inflammation and muscle injury16,17 (Extended Data\nFig. 1g). Cell-type-specific gene expression changes confirmed the\nshared gene expression changes were not due to sample impurities\n(Extended Data Fig. 1h). Considering the early transcriptional response\nof several TME cell types to Treg depletion, we assessed whether Treg\ncells were found in the proximity of these ‘first responders’ using IF\nanalysis of tumor-bearing lungs. Indeed, GFP-DTR+ Treg cells were found\nin markedly closer proximity to Lyve-1+ LECs, GP38+ fibroblasts and\nF4/80+ macrophages within and near tumor nodules than in areas\nfurther away from tumor nodules in the same tumor-bearing lung\n(Fig. 1g,h). Collectively, we have shown that Treg cells are highly con-\nnected in the KP TME.\n\nFig. 2 | Single-cell transcriptomic analysis of ‘Treg cell dependencies’\nof accessory cell states in mouse lung adenocarcinoma tumor\nmicroenvironment. a,b, t-distributed stochastic neighbor embedding (t-SNE)\nplots (27,000 cells) representing cell populations from major cell lineages\nisolated from 48 h DT-treated or PBS-treated (Ctrl) tumor-bearing lungs\n(three mice per group) colored by cell type (a) and condition (b). c, A density\nplot showing the distribution of cells between experimental conditions.\nd,e, t-SNE plots (2,815 cells) representing distribution of the VEC populations\ncolored by subtype (d) and condition (e). f, A density plot of endothelial cells\nshowing the distribution of cells between experimental conditions. g, Graph of\nneighborhoods of endothelial cells computed using MiloR and t-SNE embedding.\nEach dot represents a neighborhood and is color coded by the false discovery rate\n(FDR)-corrected P value (alpha = 1) quantifying the significance of enrichment\nof DT cells compared to control in each neighborhood. The size of the dot\nrepresents the number of cells in the neighborhood. h, Swarm plot depicting the\n\nlog fold change of differential cell-type abundance in DT-treated versus\ncontrol samples in each neighborhood across different endothelial cell types.\nEach dot represents a neighborhood and is color coded by the FDR-corrected\nP value (alpha = 1) quantifying the significance of enrichment of DT cells\ncompared to control in each neighborhood. A neighborhood is classified as a cell\ntype if it comprises at least 80% of cells in the neighborhood, otherwise it is called\n‘mixed’. i, Heat map showing average factor cell score in each cluster for each\nexperimental condition in the VEC population. The scores were row normalized\nbetween 0 and 1. Each row represents a factor, and each column represents a\ncluster in a specific experimental condition. The clusters are grouped based on\ntheir phenotype. j, Gene expression heat maps showing the top 200 genes that\ncorrelated the most with the imputed activated VEC factor indicated (Methods).\nEach column represents a cell; cells are ordered based on their factor score (in\nascending order from left to right) indicated by the green bar. Select genes of\ninterest are noted on the right. b, e, h and i; Ctrl, PBS, gray; DT, red.\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1022\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fa\n\nd\n\ng\n\ni\n\nj\n\n4\n\n0\n\n–4\n\nb\n\ne\n\nCtrl\nDT\n\nEpithelial\nLymphatic\nendothelial\nVascular\nendothelial\nMyeloid\nNeutrophil\nFibroblast\nT/NK cell\nB cell\n\nArtery vein\naCap\ngCap\nInflammatory\ncapillary\nLymphatic\nendothelial\n\n50\n100\n150\n\n200\n\nFDR\n\n0.75\n\n0.50\n\n0.25\n\nArtery vein\naCap\ngCap\nInflammatory capillary\nLymphatic endothelial\n\n1.0\n\n0.5\n\n0\n\nc\n\n2\nE\nN\nS\n-\nt\n\nf\n\n2\nE\nN\nS\n-\nt\n\nCtrl\nDT\n\nh\n\ngCap\n\nMixed\n\naCap\n\nLymphatic\nendothelial\n\nArtery vein\n\nInflammatory\ncapillary\n\nCtrl\n\nDT\n\nt-SNE1\n\nCtrl\n\nDT\n\nt-SNE1\n\nHigh\n\nLow\n\nHigh\n\nLow\n\n3.0\n\n2.5\n\n2.0\n\n1.5\n\n1.0\n\n0.5\n\n–\nl\no\ng\n1\n0\n(\nF\nD\nR\n)\n\nCtrl\n\nDT\n\nlogFC\n\n–4.0\n\n–2.0\n\n0\n\n2.0\n\n4.0\n\n6.0\n\n8 - angiogenesis\n15 - inflammation hypoxia\n17 - inflammation EndMT\n3 - activated VEC\n19 – IFN response\n\nActivated VEC (3)\n\nAngiogenesis (8)\n\nInflammation, hypoxia (15)\n\nInflammation, EndMT (17)\n\nCell scores\n\nCtrl\n\nDT\n\n4\n\n0\n\n–4\n\nBcl3\nNoct\nRelb\nTnf\nCcrl2\nCc40\nIrf5\nCsf1\nNfκb2\nIcosl\nEgr2\nDll1\nPim1\nIrf1\nIcam1\nFgf2\nTank\nIl6\nTgif1\nNinj1\nTnip1\n\n4\n\n0\n\n–4\n\nLpl\nCd36\nMiga2\nTap1\nWars1\nCd74\nLy6e\nGbp6\nIdo1\nCiita\nOas2\nVegfa\nThbd\nSlco2a1\nJup\nIcam2\nLima1\nCldn5\nPard6g\nCd47\nFmo1\nAlas1\nBmpr2\nSptbn1\nSmad6\nSema3c\n\n4\n\n0\n\n–4\n\nKlf6\nNfil3\nBhlhe40\nMaff\nSerpine1\nPlaur\nTnfaip3\nIcam1\nNfkbia\nJunb\nHbegf\nRel\nRelb\nFosl2\nHmox1\nTimp3\nIrf8\nBatf3\nNfkbiz\nPvr\nCcr7\nStat3\n\nEmp3\nSerpina3\nPsmg4\nCd63\nIl1r1\nLgmn\nCsrp2\nLcn2\nCfb\nLgals4\nNpm3\nTraf4\nKpnb1\nTimp1\nGda\nCh25\nTgm2\nPrkca\nCsrp2\nNgf\nAmmecr1\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1023\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2Nhood size\fSingle-cell analysis of tumoral Treg cell ‘connectivity’\nTo explore the impact of Treg cells on the diverse cell states in the TME,\nwe performed single-cell RNA sequencing (scRNA-seq) of sorted CD45−\nand CD45+ cell populations using the 10X platform (Extended Data\nFig. 2a). These populations were isolated from tumor-bearing lungs of\nKP-DTR mice treated for 48 h with DT or vehicle control 3 months after\nadenoviral Cre-driven tumor initiation (Fig. 1a). After pre-processing,\nwe clustered cells using PhenoGraph18 and annotated clusters using\nexpression of known markers into major cell types (Extended Data\nFig. 2b–d). To ensure our inferences were robust, we focused on the\nmajor hematopoietic and non-hematopoietic cell types in the TME\nthat had substantial numbers of cells. The final processed datasets\nincluded LECs, VECs, LECs, fibroblasts, lymphoid cells and myeloid\ncells (macrophages, monocytes, dendritic cells (DCs) and neutrophils;\nFig. 2a). Similarly to population-level assessments, scRNA-seq showed\nthat short-term Treg cell depletion had profound effects on transcrip-\ntional features of fibroblasts, myeloid and endothelial cells compared\nto lymphocytes (Extended Data Fig. 2e and Fig. 2a–c). To gain deeper\ninsight into the phenotypic response of accessory cells whose transcrip-\ntomes were most affected by Treg removal—endothelial cells, fibroblasts\nand myeloid cells, we separately clustered and embedded each subtype\nto ascribe finer-grain identities (Fig. 2d and Extended Data Fig. 3; for\nannotation strategy see Methods). Furthermore, we used Milo19 to\nquantify changes in abundance of subpopulations and cell states after\nTreg cell depletion (Methods). We found several cell states affected by\nTreg cell depletion, with the most pronounced phenotypic shifts in capil-\nlary VECs, mesenchymal stem cells (MSCs), Col14a1 matrix fibroblasts,\nmonocytes and macrophages (Fig. 2d–h and Extended Data Fig. 4a–d).\nTherefore, Treg cell depletion markedly affected the distribution and\nabundancies of several cell states and subsets in the TME.\n\nShared and distinct Treg cell-dependent gene programs\nWe then sought to characterize genes that respond to Treg cell depletion\nin these key accessory cell subsets. We used factor analysis to char-\nacterize gene expression programs—sets of genes whose expression\nchanges in a coordinated way in a specific set of cells and assessed\ntheir differential usage in cell populations from control or DT mice to\nelucidate the response to Treg cell depletion. Specifically, factor analysis\nmethods are well suited to decompose data into factors, which rep-\nresent coordinated expression programs across cells and reduce the\nimpact of noise on analysis, which can be dominant at an individual\ngene level20. We used single-cell hierarchical Poisson factorization\n(scHPF), designed specifically for scRNA-seq21,22 and applied it to each\ncell lineage separately to dissect the observed gene expression changes.\nEach cell and gene present in the expression matrix was assigned a\nscore for each factor, enabling biological interpretation of that factor\n(see Supplementary Table 2 for factor gene and cell matrices). Factors\nwere robust to random initializations of the model and robust to slight\nchanges in parameters (Methods and Supplementary Fig. 1).\n\nWe reasoned that gene programs most affected by Treg cell pres-\nence would have differential factor cell scores between the control\nand DT conditions. To evaluate this systematically, we computed the\naverage cell score of every factor in each cluster for each condition\n(Fig. 2i) and identified those that have higher averages in DT compared\nto control. In the endothelial lineage, we identified four major gene\nprograms that were robust to random initializations (Supplementary\nFig. 1), were biologically relevant and had significantly differential cell\nscores (Mann–Whitney test; Methods) following Treg cell depletion\ncompared to control in at least one of the endothelial cell subtypes\n(Fig. 2i). We then visualized expression of the genes with the highest\nfactor loadings in the relevant cell subtype (Fig. 2j). We observed several\nnotable patterns, including the inflammatory or activated capillary\nVEC factor (factor 3), a highly Treg cell-dependent factor character-\nized by cytokine/chemokine-, Notch and nuclear factor-κB (NF-κB)\nsignaling-, and co-stimulation pathway-related gene expression\n(Fig. 2j; see Supplementary Table 3 for endothelial factors of inter-\nest). Other highlighted factors enriched following Treg cell depletion\nin the endothelial cell population included genes related to the NF-κB\nsignaling pathway (Nfkbia, Rel, Hbegf), inflammation/hypoxia (Klf6,\nSerpine1, Plaur; factor 15) and vascularization (Vegfa, Thbd and Slco2a1;\nfactor 8), and genes linked to transforming growth factor-beta-induced\nEndMT (Emp3, Timp1 and Tgm2; factor 17). Besides cancer, the latter\nprocess is induced in aberrant tissue remodeling and fibrosis23,24. These\nobservations indicate that Treg cells impact specific features of certain\nendothelial cell subsets in the TME.\n\nNotably, the observed transcriptomic perturbations were not\nunique to endothelial cells. The Treg cell depletion-induced gene pro-\ngrams related to interferon (IFN) response, inflammatory cytokines\n(ICs) and chemokines, STAT3 and interleukin (IL)-6 signaling appeared\nto be shared across accessory cell populations. The three most differen-\ntially expressed gene (DEG) programs observed in fibroblasts following\nTreg cell depletion included an inflammatory secretory phenotype (Ccl2,\nHif1a, Rel, Cxcl1; factor 22), IFN response (Irf7, Ifit3, Isg15; factor 9) and\nECM-related genes (Fbn1, Fn1, Lamc2, Notch2; factor 14; Extended Data\nFig. 5a,b and Supplementary Table 4). On the other hand, several fac-\ntors in monocytes (factors 2, 5, 7, 13, 17, 21 and 22) and macrophages\n(factors 15, 17 and 23) including IFN and hypoxia response emerged as\ndifferentially abundant (Extended Data Fig. 5c,d and Supplementary\nTable 5; for all significant factors across cell subsets, see Supplementary\nTable 6). These results suggested that Treg cell communication with vari-\nous cells in the TME imparted both shared and distinct transcriptional\nfeatures across and within specific cell populations in either a direct\nor indirect manner.\n\nTreg cell dependency of accessory cell states in lung injury\nTo test whether the Treg cell ‘connectivity’ to key accessory cell types\nobserved in lung cancer represents a generalizable facet of tissue organ-\nization, we examined perturbations of their transcriptional states upon\n\nFig. 3 | Shared early transcriptional responses induced upon Treg cell\ndepletion in mouse lung adenocarcinoma TME and bleomycin-induced\nlung inflammation. a, t-SNE plots (24,592 cells) representing cell populations\nisolated from the lungs of mice administered with diphtheria toxin (DT) or PBS\n(Ctrl) for 48 h. Lung injury-induced inflammation was induced in both groups\nof mice upon bleomycin treatment 21 d before DT/PBS administration. The\ndata represent analysis of three mice per group colored by cell type (left) and\ncondition (middle), and a density of the distribution of cells between conditions\n(right). b, t-SNE embedding of endothelial cells isolated from Ctrl and DT\nafter bleomycin administration color coded by cell type (left) or experimental\ncondition (middle), and density plots of the distribution of endothelial cells\nbetween conditions (right). c, Heat map showing average factor cell score in\neach cell type for each experimental condition for endothelial cell subsets. The\nscores were row normalized between 0 and 1. Each row represents a factor, and\neach column represents an endothelial cell subset in a specific experimental\ncondition. Factors of interest are highlighted by a red box. d, Heat map showing\n\nthe 72 shared genes specific to activated VEC factor in both lung challenge\nmodels (Methods and Supplementary Table 9). Each column represents a\ncell; cells are ordered based on their factor score (in ascending order from\nleft to right), indicated by the green bar. e, Heat map showing factor cell score\nacross experimental conditions averaged over each myeloid cluster in each\nexperimental condition for bleomycin-administered cells. The rows are factors\nand columns are clusters for each experimental condition. The clusters are\ngrouped based on the cell type they are associated with. The heat map shows\nrow-normalized scores from 0 to 1. The left color bar shows the average factor\ncell score. f, Heat maps showing the 54 shared genes between mouse lung tumor\nand injury-induced inflammation in the Arg1+ macrophage factor (tumor factor\n23 corresponding to injury-induced inflammation factor 0; Supplementary\nTable 10). Each column is a cell; cells are ordered based on their factor score\n(in ascending order from left to right) indicated by the green bar. The treatment\ncondition for each cell is indicated by gray for PBS and red for DT bars. Select\ngenes of interest are shown.\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1024\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fa\n\nb\n\nc\n\n1.0\n\n0.5\n\n0\n\ne\n\n1.0\n\n0.5\n\n0\n\nEpithelial\nEndothelial\nMyeloid\nNeutrophil\nFibroblast\nT/NK cell\nB cell\n\nCtrl\nDT\n\naCap\ngCap\nInflammatory capillary\nLymphatic endothelial\n\nCtrl\nDT\n\n2\nE\nN\nS\n-\nt\n\n2\nE\nN\nS\n-\nt\n\nCtrl\n\nDT\n\nHigh\n\nt-SNE1\n\nLow\n\nCtrl\n\nDT\n\nHigh\n\nt-SNE1\n\nLow\n\nCtrl\n\nDT\n\n15 inflammation,\nEndMT, angiogenesis\n\n12 -activated VEC\n\n13 -inflammation,\nhypoxia\n\naCap\ngCap\nInflammatory capillary\nLymphatic endothelial\n\nArg1 mac\nAlveolar mac\nC1qa mac\nCcr7 cDC2\n\nCsf3r mono\nMono\nRetnla mac\n\ncDc1\n\ncDc2\npDc\n\nCtrl\n\nDT\n\n0\n\n15\n\nd\n\n4\n\n0\n\n–4\n\nf\n\n4\n\n0\n\n–4\n\nKP activated VEC 3\n\nInjury activated VEC 15\n\nCell scores\nCtrl DT\n\nKP (23)\n\nInjury (0)\n\nCell scores\nDT\nCtrl\n\nMaff\nBhlhe40\nTnfaip3\nFosl2\nCsf1\nIcoslg\nBcl3\nBirc3\nEgr2\nRelb\nPim1\nNoct\nWsb1\nSoca2\nCasp4\nIrf5\nNrid2\nIcam4\nLat2\nSh2b2\nCtps1\nDll1\n\nVegfa\nSpp1\nVcan\nPlaur\nTgm2\nInhba\nCol4a1\nOlr1\nNdrg1\nSlc2a1\nErrfi1\nSod2\nSphk1\nGsr\nArg1\nUpp1\nAlas1\nSmox\nArg2\nPrdx6\nUck2\nBlcap\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1025\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fa\n\n1 mm\n\nb\n\no\nd\nn\nE\n\nb\nF\n\ni\n\nl\ny\nM\n\nFactor\n\n(17) inflammation, EndMT\n(15) inflammation, hypoxia\n(8) angiogenesis\n(19) IFN\n(3) activated VEC\n(22) inflammatory cytokine\n(21) inflammatory cytokine\n(14) ECM\n(9) IFN\n(23) Arg1+ macrophage\n(2) C1Q+ macrophage\n(15) proliferation\n(21) monocyte, hypoxia\n(17) IFN\n(13) Csf3r monocyte\n(5) monocyte, coagulation\n\nTumor\nRNA %\n0.4\n\n0\n\nTumor\nspot\n\n+\n–\n\np\na\nC\na\n\np\na\nC\ng\n\ne\nt\ny\nc\ni\nr\ne\nP\n\ne\ng\na\nh\np\no\nr\nc\na\nM\n\nt\ns\na\nl\nb\no\nr\nb\nfi\n+\n\n1\na\n4\n1\nl\no\nC\n\nt\ns\na\nl\nb\no\nr\nb\nfi\no\ny\nM\n\nt\ns\na\nl\nb\no\nr\nb\nfi\n+\n\n1\na\n3\n1\nl\no\nC\n\ne\nt\ny\nc\no\nn\no\nM\n\ne\ng\na\nh\np\no\nr\nc\na\nm\n\nr\na\nl\no\ne\nv\nl\nA\n\nl\na\ni\nl\ne\nh\nt\no\nd\nn\ne\nc\ni\nt\na\nh\np\nm\ny\nL\n\nc\n\nInflammatory cytokine\n\nd\n\nIFN response\n\n1 mm\n\n0\n\n0.6\n\n0\n\n0.4\n\ne\n\nTreg depleted - Ctrl\nmean factor score\n\n0.04\n\n0\n\n–0.04\n\n–log10(P.adj)\n\n0\n50\n100\n150\n200\n\nSignaling\nniche\n\nIFN\n\nIC\n\nIFN + IC\n\nother\n\nCtrl\n\nTreg depleted\n\nCtrl\n\nTreg depleted\n\nTreg depleted\n\nA\n\nIA\n\nIA\n\nIA\n\nLV\n\nA\n\nLV\n\nIA\n\nV\n\nV\n\nV\n\nBr\n\nA\n\nIC\n\nIFN\n\nf\n\ne\np\ny\nt\n\nl\nl\ne\nC\n\nTumor\nT cell/ILC2\nNK\nNeutrophil\nMSC\nMonocyte\ncDC\nBasophil/mast\nAT2\nAlveolar macrophage\n\n0\n\n1\n\n2\n\n0\n\n0.2\n\n0.4\n\n0.6\n\nCell-type RNA fraction log2 fold change\n\nAdjusted empirical P\n\n0.1\n\n0\n\nFig. 4 | Spatial transcriptomics identifies distinct inflammatory cytokine and\nIFN signaling niches in lung adenocarcinoma following Treg cell depletion.\na, Tumor region identification in KP LuAd sections using Visium ST. The\nfraction of tumor cell RNA in each Visium spot (top right) was determined by\nBayesPrism deconvolution, binarized (bottom right; Methods), and compared\nto histological H&E images (left). b, Factor scores and Bonferroni-adjusted\ntwo-sided t-test P values differentially expressed factors between control and Treg\ncell-depleted conditions in ST. c,d, Representative tissue sections from control\n(left) or Treg cell-depleted (right) conditions. Tumor regions are outlined, and\nspots are colored by factor score. Scores represent IC (c; 18 genes) or IFN (d;\n\n103 genes) gene programs shared across all lineages (Br, bronchi; A/V, artery/\nvein; LV, lymphatic vessel; IAs, immune cell aggregates). e, ST analysis revealed\ndistinct signaling niches. Spots were assigned to niches based on thresholding\na gamma distribution fitted to IC or IFN signaling module scores across all spots\n(Methods). f, Enrichment of cell-type RNA fractions in signaling niches. Adjusted\nempirical P value corresponds to the probability of obtaining the mean observed\nRNA fraction for that cell type (Methods). Fractions with adjusted P > 0.01 are\nnot shown. In a and c–e, images are representative of, and analysis performed\non (b and f), one of two serial sections for each of four samples (DT and Ctrl, two\nbiological replicates each).\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1026\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\n\n\fLung progenitor-like\n\nEMT\n\nHigh plasticity\n\nTumor\n\nSpot\n\n+\n\n–\n\nTumor state deconvolution\n\nTumor subtype RNA %\n\n200 µm\n\nTumor state\n\nGastric\n\nEndoderm.like\n\nAT2.like\n\nAT1.like\n\nEMT\n\nLung.progenitor.like\n\nHigh.plasticity\n\nc\n\ns\nn\no\ni\ns\ne\nl\n\nf\no\nr\ne\nb\nm\nu\nN\n\n10\n\n5\n\n0\n\n1 mm\n\nF3\n\nAnxa1\n\nPtgs2\n\ne\n\nEMT\nHigh plasticity\n\nLung pro\n\nCondition\n\nImmune response\n\nCtrl\nTreg\ndepletion\n\n10\n\n5\n\n0\n\n–\n\n+\n\nGastric\nHigh plasticity\n\nGastric\nHigh plasticity\n\nPf4\n\nGkn2\n\n90\n\nDlk1\n\nCtsh\n\nId2\n\nApoc1\n\nCol18a1\n\nScd2\n\nMgst1\n\nItgav\nPhlda1\n\nPlin2\n\nEgr1\nFosl1\nPtprn\n\nIl6 Sprr1a\nCxcl3\n\nCxcl1\n\nLgals3\nKrt7\n\nErrfi1\n\nThbs1\n\nJunb\nKrt18\n\nPmvk\n\nIfrd1\n\nNedd4\n\nCxcl16\n\nIl4ra\n\nS100a11\n\nIer5\nDusp1\n\nCxcl2\n\nEreg\n\nMeg3\n\nSox9\n\nLgi3\n\nBpifa1\n\nBex2\n\nCxcl10\n\nSignificant\nFalse\nTrue\n\nImmune\nresponse\n\n–\n+\n\na\n\nb\n\nd\n\nj\n\nP\nd\ne\nt\ns\nu\nd\na\n0\n1\ng\no\nl\n–\n\n60\n\n30\n\n0\n\n−3\n\n−2\n\n−1\n\n0\n\n1\n\n2\n\nlog2 fold change\nTreg depletion response – nonresponse\n\nSox9\n\n0\n\n2\n\nPf4\n\n0\n\n4\n\nFig. 5 | High-plasticity state and heterogeneity revealed by lung\nadenocarcinoma responses to Treg cell depletion. a, ST analysis of tumor states.\nBayesPrism deconvolution using additional labeled tumor cells from Yang et al.28\nwas performed to assign tumor-state-specific RNA fractions. Correspondence\nof regions with highlighted differential tumor states (middle) to H&E section\nis shown (right). Dashed lines denote regions with the indicated dominant\ntumor states (red, high plasticity; yellow, EMT; black, lung progenitor-like).\nb, Spots labeled by tumor-state cluster. In a and b, images are representative\nof, and analysis performed on (c and d), one of two serial sections for each of\n\nfour samples (DT and Ctrl, two biological replicates each). c, Quantification of\ntumor lesion area types across Treg cell depletion and control conditions (left)\nor between tumors with or without detectable immune response in Treg cell-\ndepleted condition (right; N = 85 lesion areas). d, Differential gene expression\n(two-sided Wilcoxon test Benjamini–Hochberg adjusted) of tumor spots\nin lesions with and without immune response to Treg cell depletion. e, Log-\nnormalized expression of Sox9 and Pf4 (Cxcl4) in a representative tumor-bearing\nlung section after Treg cell depletion. Inset at top left indicates immune response\nstatus of tumor lesion areas.\n\nidentical short-term Treg cell depletion in a setting of bleomycin-induced\nfibrotic lung inflammation using scRNA-seq analysis (Fig. 3a,b and\nExtended Data Fig. 6a–d). Not only were all cell populations detected\nin tumor-bearing lungs also present in inflamed lungs, Treg deple-\ntion in this setting also generated similar transcriptional responses\n(Fig. 3a,b and Extended Data Fig. 6c,d). Independent analysis of the\ngene programs in the inflamed lung using scHPF (see Supplementary\nTable 7 for factor matrices) identified Treg cell depletion-associated\nendothelial factors (Fig. 3c and Supplementary Table 8). We correlated\ngene scores associated with each factor from lung tumors to lung injury\nto identify similarities. We found that the activated VEC factor in the\nlung injury (factor 15) correlated strongly (Pearson correlation > 0.70)\n\nwith its counterpart in the tumor setting (factor 3), indicating that\nthe same set of genes responded to the loss of Treg cells in both chal-\nlenges. In fact, 72 of the top 200 genes associated with factor 3 spe-\ncific to the tumor endothelial cell inflammatory capillary subset were\nshared with the top 200 genes associated with factor 15 specific to the\nsame subset of cells in the injury model (Fig. 3d and Supplementary\nTable 9). Other endothelial cell factors, namely inflammation/hypoxia\n(factor 13), NF-κB signaling and EndMT (factor 12), that were observed\nin the inflamed lung upon short-term Treg cell depletion also correlated\npositively, even if weakly, with related tumor factors 15 and 8, respec-\ntively (Extended Data Fig. 6e). Consistently, factor analyses of other\nlineages revealed overlapping differential gene programs between\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1027\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\n\n\fa\n\nb\n\nGastric\n\nIC score\n\n200 µm\n\n0\n\n0.6\n\n0\n\n0.6\n\nHigh plasticity\n\nSox9\n\n0\n\n0.5\n\n0\n\n2\n\nFig. 6 | Local histological and immune response heterogeneity following Treg\ncell depletion. a, H&E staining of representative tumor section characterized by\nhistological and immune response state heterogeneity after Treg cell depletion.\nInsets at bottom represent a zoomed-in view of gastric (left) and high-plasticity\n(right) areas. Black arrows highlight neutrophil infiltration in a high-plasticity\narea. b, Tumor RNA fraction within highlighted high-plasticity and gastric\nepithelial states (left) and gene expression modules (right) of tumor lesion\nshown in a. Images are representative of one of two serial sections for each of four\nsamples (DT and Ctrl, two biological replicates each).\n\ntumor and injury models, including Treg cell depletion-induced gene\nprograms in Arg1+ macrophages (Fig. 3e,f and Supplementary Table 10)\nand IC signatures in Col14a1 matrix fibroblasts. These findings sug-\ngested that Treg cell-dependent transcriptional programs are not limited\nto the TME and can be shared across pathological conditions.\n\nSpatial distribution of Treg cell-dependent tumor\nmicroenvironment gene programs\nTo gain insights into the spatial organization of the identified accessory\ncell populations, gene programs and their relationship to transcrip-\ntional states of tumor cells, we profiled four tissue sections (two control,\ntwo Treg cell depleted) using the 10X Visium platform. We used Bayes-\nPrism25,26, a Bayesian framework that jointly estimates cell-type frac-\ntions and cell-type-specific gene expression using a labeled scRNA-seq\nreference, to deconvolve each spatial transcriptomics (ST) spot into\nconstituent cell populations. Deconvolution was performed using our\nscRNA-seq datasets labeled with 26 distinct cell populations selected\n\nto optimize granularity, robustness and concordance with underly-\ning histological features in paired H&E-stained sections (Methods,\nFig. 4a, Extended Data Fig. 7a–e and Supplementary Table 11). Next, we\nassessed whether the gene factors that changed upon Treg cell deple-\ntion in scRNA-seq were also identified by ST analysis. Consistently,\nwe observed upregulation of endothelial and fibroblast IC and IFN\nsignaling-related gene signatures after Treg cell depletion within spots\nassigned to the corresponding cell type (Fig. 4b). We also observed\nincreased use of genes associated with the activated VEC factor in capil-\nlary aerocyte (aCap) endothelium assigned spots, as well as increased\nIFN and proliferation related gene signatures in myeloid spots. IC and\nIFN factors shared many genes across all three analyzed accessory\nlineages (18 for IC, 103 for IFN), which suggested that similar gene pro-\ngrams were induced across colocalized cell types by common stimuli,\nindicative of a signaling niche. The spatial behavior of shared genes\nin these two programs showed localization to two distinct signaling\nniches in the tissue, with the IC gene program (Cxcl2, Ier3, Fosl1, Il6)\nlocalized to the tumor core and the IFN response gene program (Ifit1,\nStat1, Isg15, Irf7) localized to the periphery of, or distal to tumor lesions.\nInspection of the same H&E-stained sections confirmed dense tumor\ncell presence with potential hypoxia and neutrophil infiltration at IC\nfoci, and immune cell aggregates at sites with strong IFN response\nsignal (Fig. 4c–e and Supplementary Table 12). Further, ST analysis\nrevealed concordant differential distribution of tumor cells and acces-\nsory cell types within these territories with higher frequency of tumor\ncells, basophils/mast cells, neutrophils and MSCs in IC territories and\na high frequency of T cells/type 2 innate lymphoid cells (ILC2s), natu-\nral killer (NK) cells, conventional dendritic cells (cDCs), monocytes\nand alveolar macrophages in IFN territories (Fig. 4f). Taken together,\nthese results point to two primary inflammatory and spatially distinct\nmodes of lung TME response to Treg cell depletion within tumor mass\nand tumor margin.\n\nTumor states associated with response to Treg cell depletion\nKP LuAds adopt a range of recurrent transcriptional states with features\nof differentiated lung ECs, their progenitors or epithelial progenitors\nfrom other tissues including the gastrointestinal tract and liver and\nEMT (epithelial to mesenchymal transition)-associated ones27–30. We\nnext sought to identify potential associations between tumor states\nand the identified TME niches, that is, IC-positive, IFN-positive and cold\n(negative) ones. We first identified tumor cells within our ECs by call-\ning KRAS p.Gly12Asp mutations. Because optimized dissociative TME\nsingle-cell analysis protocols are suboptimal for capturing tumor cells,\nwe identified only 239 tumor cells within our mouse scRNA-seq dataset.\nTo enable robust deconvolution of tumor cell states, we substituted\ntumor cells from our scRNA-seq dataset with those from a published\ndataset that had better capture of KP LuAd cells (KP-tracer dataset;\nN = 18,083)28. With this updated reference, we performed an additional\nspot deconvolution to more accurately capture tumor states in the tis-\nsue. TME fractions for other cell types remained relatively unchanged\nbetween deconvolutions (Extended Data Fig. 7c).\n\nIn spots with tumors, the tumor-state fractions exhibited regional\nvariation in gene expression programs, sometimes within seemingly\nthe same tumor nodule (Fig. 5a). Tumor spots were clustered by their\ntumor-state fractions and typically showed a dominant tumor state\nin each spot (Extended Data Fig. 8a) manifested in the expression of\ncorresponding marker genes (Extended Data Fig. 8b), forming continu-\nous spatial patches of similar phenotypes (Fig. 5b and Extended Data\nFig. 8c). Spots were grouped into tumor lesion areas, or contiguous\npatches of tumor cells belonging to the same cluster and quanti-\nfied across control and Treg cell-depleted conditions. Tumor states\nwere also compared across tumors that had a detectable immune\nresponse (>10% of spots in IC or IFN signaling niches) or not in Treg\ncell-depleted sections. Treg cell depletion resulted in a pronounced\nincrease in tumor lesion areas that corresponded to a high-plasticity\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1028\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fcell state, specifically among tumor nodules associated with an immune\nresponse (Fig. 5c and Extended Data Fig. 8d,e)29. This was consistent\nwith a significant enrichment of high-plasticity cell-state genes upregu-\nlated by tumor cells after Treg cell depletion in scRNA-seq (Extended\nData Fig. 8f,g and Supplementary Tables 13 and 14). Therefore, TME\nresponse to the removal of Treg cells may elicit a gene program in\nLuAds that represents an unstable transitional state, which can give\nrise to other tumor states28,29. While IC and IFN niches were observed\nin the majority of tumor nodules after Treg cell depletion, there were\nsome nodules, and even areas within individual nodules, that did not\n(Fig. 4c). In particular, those with a gastric epithelial lineage gene\nexpression program were selectively devoid of IC or IFN responses\n(Fig. 5c and Extended Data Fig. 8e). We assessed differential gene\nexpression between immune response ‘rich’ and ‘poor’ lesion areas\nand found increased expression of Gkn2 (gastrokine), Pf4 (platelet fac-\ntor 4/Cxcl4) and Sox9 among other genes (Fig. 5d,e and Supplementary\nTable 15). Interestingly, Sox9 expression in lung tumor cells was shown\nto enable their escape from NK cell-mediated killing in certain cases27,\nsuggesting one potential mechanism of immune evasion. Similarly to\na recent analysis of CRISPR-edited tumors31, the observed response\nto Treg cell depletion was spatially restricted, as even ‘nonresponsive’\nareas that were directly adjacent to responsive ones were deprived of\nimmune cell or IC signals (Fig.6a,b). Therefore, regional variation in\ntumor state appears to define the TME response to Treg cell depletion.\n\nConserved Treg cell-dependent features of human and mouse\ntumor microenvironment\nNext, we sought to test whether Treg cell-dependent TME features\nobserved in mice are conserved in human cancer (Fig. 7a) by leveraging\nvariation in Treg cell abundance in 25 primary or local metastatic LuAd\nsamples from 23 individuals, analyzed using scRNA-seq (Supplemen-\ntary Tables 16 and 17). Despite differences in composition and propor-\ntion of accessory cell types in these datasets, we were able to detect all\ncell populations corresponding to those observed in mice (Fig. 7b and\nExtended Data Fig. 9a,b). To determine whether the factors induced\nafter Treg cell depletion in mice are present in human LuAd samples\nwith a low abundance of Treg cells, we first determined the frequency\nof Treg cells among all hematopoietic cells in each sample (Fig. 7c and\nExtended Data Figs. 9c and 10a,b). Next, we performed scHPF analy-\nsis for each of the cell lineages under investigation (Supplementary\nTable 18) and looked for orthologous genes shared between human\nand mouse factors to align gene programs between species (Fig. 7d).\nThen, we assessed the correlation of mean factor usage in single cells to\nTreg cell frequency across human samples. This identified three factors\nnegatively correlated with Treg cell proportion that corresponded to\naspects of the compensatory endothelial response to Treg cell depletion\nin the KP mouse model (Extended Data Fig. 10c). The latter included\nfactors whose most associated genes were related to activated aCap\n(CAR4, CD36, IFNGR1, FAS, CX3CL1, TNFRSF11b, EDN1; factors 3 and\n5; Fig. 7e), inflammation and hypoxia (VEGFB, PLAUR, SERPINE1, IL6,\nCXCL1, BCL3, PVR, IRF4, BATF3, TFP12; factors 4 and 5) and angiogenesis\n\nfactors (factor 3). We used the sum of these three factors as a general\nTreg cell-responsive endothelial gene program to account for potential\nsample-specific, cell-type-specific or condition-specific effects that\nwould separate a shared underlying biological program into separate\nfactors during factorization (Extended Data Fig. 10d). Comparing this\nscore to Treg cell proportion, we observed a clear negative correlation—\nstronger than any factor individually—across tumor samples (Fig. 7f),\nwhich suggested conserved Treg cell influence on this gene expression\nprogram. To further identify specific components of this shared Treg\ncell-responsive endothelial gene expression program, we compared the\nloadings of genes in the factors related to inflammation and hypoxia\nacross species (factors 4 + 5 in human LuAd, factor 15 in KP mouse;\nFig. 7g). This identified genes encoding key inflammatory mediators\n(IL6, CSF3, VCAM1, SELE, PTGS2) and a host of VEGF-induced genes in\nendothelial cells (RND1, ADAMTS1, ADAMTS4, ADAMTS9, AKAP12) as\nconserved members of expression programs induced in endothelial\ncells in Treg cell-poor TMEs across species.\n\nSimilar analyses of fibroblasts and myeloid cells also revealed\ncorresponding Treg cell-dependent mouse and human factors. For\nexample, human fibroblast factors 3, 5 and 22 corresponded to IC\nmouse fibroblast factors 21 and 22, with overlapping genes including\nIL6, IL1RL1, NFKB1, CCL2 and LIF (Extended Data Fig. 10e,f), while factor\n9 (AP1 TF family members, KLF2/4, SOX9, HES1, IRF1) was negatively\nassociated with Treg cell proportion. Additionally, high usage of con-\nserved CSF3R monocyte factor 16 (CSF3R, PROK2, VCAN) in human ‘Treg\ncell-poor’ LuAd samples was consistent with the hypoxia, angiogenesis\nand NF-κB signaling related features (VEGFA, HIF1A, CEACAM1, NOTCH1,\nBCL3, BCL6) of this population in Treg cell-depleted mice (Extended\nData Fig. 10g,h and Extended Data Fig. 5d). Notably, several human\nmyeloid factors and corresponding mouse factors showed positive\ncorrelation with Treg cell presence, such as an SPP1/FOLR2 macrophage\nfactor, a cell cycle factor and a C1Q+ macrophage factor (C1Q, antigen\npresentation-related genes), which included genes encoding known\nnegative regulators of innate and adaptive immunity (CFH, CR1L, LAG3,\nPDCD1LG2, LILRB4, IL18BP; Extended Data Fig. 10i). Interestingly, we\nobserved similarly pronounced downregulation of this gene program\nupon Treg cell depletion in both lung tumors and bleomycin-induced\ninjury, suggesting that Treg cells within both niches sustain certain\nimmunomodulatory myeloid cell states. Further analysis of correla-\ntion between conserved Treg cell-dependent human and mouse factors\nrevealed a set of opposing TME programs (Extended Data Fig. 10j). One\nfactor group in Treg cell-poor or Treg cell-deprived tumors included IL-1β/\nIL-18 signaling-related genes (IL18RAP, IL1RAP) expressed in angiogenic\nmonocytes and tumor necrosis factor (TNF)/IL-1β-induced genes in\nfibroblast and endothelial cells involved in monocyte and neutrophil\nrecruitment (CSF3, CXCL1, CXCL2, CXCL8, CCL2). The other, positively\nassociated with Treg cell presence, featured immunomodulatory genes\nthat inhibit IL-1β/IL-18 signaling (TMEM176B, IL18BP; see Supplemen-\ntary Table 19 for KP/injury/LuAd factors). These findings suggest a\nconserved role of Treg cells in tuning transcriptional states of principal\naccessory cell types in the TME.\n\nFig. 7 | Factor analysis of Treg cell ‘dependencies’ of accessory cell\ntranscriptional states in human and mouse lung adenocarcinomas.\na, Schematic of the experimental design. b, t-SNE plot of all cells (82,991 total\ncells) from 25 primary human LuAd or local metastases labeled by lineage.\nc, t-SNE of T/NK cell lineage colored by unique molecular identifier (UMI) counts\nof Treg cell marker genes (maximum of two). d, Jaccard similarity between genes\nassociated with mouse and human factors in tumor endothelial cells. Factors of\ninterest with high correlation are highlighted by a green box. e, Conservation\nof activated VEC signature genes. Normalized gene loading (fraction of gene\nscore across all factors) of genes within the mouse activated VEC signature\nacross all human endothelial factors. Upper and lower notches of the box plot\ncorrespond to the 75th and 25th quartiles, respectively, and the middle notch\ncorresponds to the median. Whiskers extend to the farthest data point no more\n\nthan 1.5 times the interquartile range from the hinge, with outliers beyond that\ndisplayed as individual points. Select genes with high loadings of factors 3 and 5\nare highlighted (N = 45 genes). f, Mean log2 sum of inflammation/angiogenesis\nassociated human endothelial factor (3, 4 and 5) cell loadings plotted against\nlog2 Treg cell proportion in each human sample. Spearman correlation estimate\n(R) and P value are listed. Trend line represents a linear model fit between\nthe two and shading indicating the 95% confidence interval (N = 19 human\nsamples). g, Normalized gene scores (fraction of gene scores across all factors)\nin orthologous genes between mouse and human inflammation/hypoxia factors.\nGenes significantly attributed to both human factors and mouse factors are\nhighlighted as conserved. VEGF-induced genes in endothelial cells were derived\nfrom the CytoSig database.\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1029\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fa\n\nb\n\ne\n\ns\ne\nn\ne\ng\nC\nE\nV\ng\nn\nd\na\no\n\ni\n\nl\ne\nn\ne\ng\nd\ne\nz\ni\nl\na\nm\nr\no\nN\n\n0\n\n1.00\n\n0.75\n\n0.50\n\n0.25\n\n0\n\ng\n\n5\n+\n4\nn\na\nm\nu\nh\ng\nn\nd\na\no\n\ni\n\nl\ne\nn\ne\ng\nd\ne\nz\ni\nl\na\nm\nr\no\nN\n\nTreg cell depletion\nin mice\n\nCompute associations\nbetween altered genes\n\nIdentify compensatory\nprograms\n\nLevel of Treg cell\npresence in humans\n\nCandidates for\ncombination therapy\n\nMouse KP factors\n\nc\n\nCD3E\n\nCD4\n\nB cell\nBlood endothelial\nEpithelial\nFibroblast\nLymphatic endothelial\nMyeloid\nNeutrophil\nT/NK\n\nFOXP3\n\nIL2RA\n\nd\n\ns\nr\no\nt\nc\na\nf\nd\nA\nu\nL\nn\na\nm\nu\nH\n\n2\n\n1\n\n0\n\n0.15\n\n0.1\n\n0.05\n\n0\n\n16\n7\n5\n10\n19\n21\n22\n3\n20\n14\n18\n13\n9\n6\n15\n17\n8\n1\n2\n4\n12\n111\n\nICAM4\n\nBIRC3\n\nTNFAIP3\n\nTIFA\n\n0.6\n\n0.4\n\nRAB20\nBCOR\n\nELMSAN1\n\n0.2\n\nCDC42EP4\n\nNOCT\n\nBCL3\nPIM1\n\nRELB\n\nSLC25A25\nSHB\nTGIF1\nFOXP4\n\nCSF1\n\n10\n\n1\n\n4\n\n15\n\n0\n\n13 7 12 8 6\n\n2\n\n5 18\n\n14\n\n113\n\n9\n\n19 17 16\n\nf\n\ne\ng\na\ns\nu\n5\n\n,\n\n4\n\n,\n\n3\nr\no\nt\nc\na\nf\n\n2\ng\no\n\nl\n\n2\n\n0\n\n–2\n\n–4\n\n3 = inflammatory capillary\n4, 5 = inflammation\n\nR = –0.41, P = 0.082\n\nCells\n100\n200\n300\n400\n\n1\n\n2\n\n3\n\n4\n\n5\n\n6\n\n7\n\n8\n\n9\n\n10\n\n11\n\n12\n\n13\n\n14\n\n15\n\n16\n\n17\n\n18\n\n19\n\n20 21\n\n22\n\nFactor\n\n–6\n\n–5\n\n–4\n\n-3\n\nlog2 Treg/CD45+\n\nIL6\n\nSELE\n\nTNFAIP3\n\nBIRC3\n\nCSF3\n\nVCAM1\n\nRCAN1\n\nMB21D2\nIER3\n\nKDM6B\n\nRND1\n\nARID5A\n\nADAMTS1\nMAP3K8\n\nSDC4\n\nADAMTS4\n\nADAMTS9\n\nPTGS2\n\nSOX7\n\nTM4SF1\n\nTRIB1\n\nNFKBID\n\nNFKBIA\n\nTGIF2\n\nREL\n\nAKAP12\n\nTNFSF9\n\nSLC25A25\nMTF1\n\nPLAUR\n\nPVR\n\nSAT1\nZBTB10\n\nNOCT\n\nFOSL1\n\nEMP1\n\nARL5B\n\nBMP2\n\nPMAIP1\n\nSLC20A1\n\nPTPRE\n\nPFKFB3\n\nINHBA\n\nConserved\n\nConserved &\nVEGF induced\n\n0\n\n0.25\n\n0.50\n\n0.75\n\nNormalized gene loading mouse 15\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1030\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\n\n\fCombinatorial Treg cell depletion therapy\nThese results highlighted candidate compensatory pathways, whose\ntargeting in combination with current clinical-stage intratumoral Treg\ncell depletion strategies32,33 could improve therapeutic efficacy. In this\nregard, the increased expression of VEGF pathway-related genes upon\nTreg cell deprivation was of particular interest suggesting that height-\nened VEGF signaling may ‘buffer’ the negative impact of Treg cell deple-\ntion on the tumor-supporting TME function and facilitate a rebound in\nthe tumor progression. We tested the above possibility by investigating\nwhether combining short-term Treg cell depletion with VEGF blockade\ncan lead to an improved control of KP tumor progression. We trans-\nplanted KP adenocarcinomas into Foxp3GFP-DTR mice and administered\nthem with DT and mouse VEGF-A neutralizing antibody (aVEGF) after\ntumors became macroscopically detectable (Fig. 8a). While Treg cell\ndepletion and VEGF blockade alone could slow tumor progression,\ntheir combination had a markedly more pronounced therapeutic\neffect (Fig. 8b). Assessment of the overall survival rate, when mice\nwere left untreated after the initial response and killed after rebound\n(tumor volume reached 1 cm3, maximum allowed by the institutional\nguidelines) showed that combination therapy improved survival in\ncomparison to either monotherapy or untreated groups (Fig. 8b). While\nsimilarly increased numbers and activation level of tumoral T cells were\nobserved in ‘DT + aVEGF’ and ‘DT-only’ in comparison to ‘aVEGF-only’\ntumor samples on day 20 of transplantation (Fig. 8c), IFN-γ-producing\nCD4+ T cells and IFN-γ-producing and TNFα-producing CD8+ T cells\nwere markedly increased in the combination treatment group as were\nmonocyte numbers (Fig. 8d). Moreover, we observed further increases\nin tumor hypoxia and apoptosis upon combined Treg cell depletion and\nVEGF blockade in comparison to both monotherapeutic modalities\nand untreated control groups (Fig. 8e,f). Notably, KP tumors failed to\nrespond to PD-1 blockade, which did not offer additional therapeutic\n\nbenefits when combined with VEGF blockade in full agreement with a\nrecent study of antiangiogenic, anti-PD-1 and chemotherapy in a KP lung\ncancer model34. Recent studies revealed high amounts of chemokine\nreceptor CCR8 displayed by Treg cells in human cancers35,36 highlighting\ntheir depletion as a therapeutic strategy33,37,38. Thus, we examined the\ntherapeutic potential of short-term VEGF blockade combined with\nantibody-mediated depletion of CCR8-expressing Treg cells, which\nrepresented only a fraction of intratumoral Treg cells in KP tumors\n(Fig. 8g). While CCR8 antibody treatment alone diminished tumor\ngrowth, a markedly more pronounced effect was observed when it\nwas combined with VEGF blockade (Fig. 8h). Notably, this regimen\nwas associated with a mere 15% decrease in overall population of\ntumor-associated Treg cells in the absence of their noticeable changes\nin the secondary lymphoid organs (Fig. 8i). Besides VEGF-A, whose\nneutralization was conducted as a proof-of-concept approach for the\ndiscovery of orthogonal combination therapy, we noted additional\ncandidate compensatory pathways enriched in the Treg cell-poor or\ncell-depleted TME including the CCR2–CCL2 axis, inhibitors of which\nare currently tested as monotherapies or combination therapies of\nhuman cancers. To further test the utility of assessment of early TME\nresponses to Treg cell depletion for identifying combinatory therapeu-\ntic modalities, we subjected KP tumor transplanted mice to a similar\nshort-term treatment with CCR8 antibody and a selective CCR2 antago-\nnist RS-504393 (CCR2i). The latter combination provided minimal\nadditional therapeutic benefit in comparison to anti-CCR8 and CCR2i\nmonotherapies contrary to aVEGF/CCR8 combination (Fig. 8j,k). These\nresults suggest that CCR2 blockade and Treg cell depletion may converge\non shared or partially overlapping TME states, whereas VEGF blockade\noffers an orthogonal intervention and highlights potential for discovery\nof orthogonal cancer therapies through single-cell and spatial analyses\nof early TME responses to acute perturbation.\n\nFig. 8 | Systemic or intratumoral CCR8+ Treg cell depletion combined with\nVEGF blockade restrains KP adenocarcinoma progression. a, Schematic of\nthe experimental design; s.c., subcutaneous. b, Tumor growth dynamics upon\nthe indicated therapeutic interventions. The data represent mean values of\ntumor volume measurements (left). Adjusted P values for day 20 measurements:\nPBS-IgG versus DT-IgG P < 0.0001; PBS-IgG versus PBS-αVEGF P = 0.0004;\nPBS-IgG versus DT-αVEGF P < 0.0001; DT-IgG versus PBS-αVEGF P = 0.0328;\nDT-IgG versus DT-αVEGF P = 0.0109; PBS-αVEGF versus DT-αVEGF; P = 0.0005.\nRepresentative image of tumor volumes at day 20 (center). Kaplan–Meyer\nsurvival curves followed by log rank (Mantel–Cox) of KP tumor-bearing mice\n(right). The ‘survival’ time reflects the end point of the experiment when tumor\nvolume in individual mice reached 1 cm3; adjusted P values: PBS-IgG versus\nDT-IgG P = 0.0012; PBS-IgG versus PBS-αVEGF P > 0.05 (NS), PBS-IgG versus\nDT-αVEGF P = 0.0078; DT-IgG versus PBS-αVEGF P > 0.05 (NS); DT-IgG versus\nDT-αVEGF P = 0.05; PBS-αVEGF versus DT-αVEGF P = 0.0186. c,d, Quantification\nof the indicated immune cell subsets and frequencies of activated (CD44hi\nCD62lo), proliferating (Ki67+) and IFN-γ-producing TCRβ+ CD4+ and TCRβ+ CD8+\ncells in tumor samples shown in Fig. 8b in the indicated experimental groups\nof mice analyzed on day 20. e, Representative HIF1α and TUNEL staining of KP\ntumor sections. f, Quantification of HIF1α expression and apoptosis (TUNEL\nstaining) in KP tumor sections; staining areas and signal intensity normalized\nby the total area and mean background intensity, respectively. 3–5 tumors from\neach experimental group were analyzed. (PBS-IgG N = 5; DT-IgG N = 4; PBS αVegf\nN = 3; DT αVegf N = 3) with four sections per individual tumor sample. Data\nrepresent the mean ± s.e.m. g, Proportion of intratumoral Treg cells on day 20\nafter KP tumor transplantation. Data represent the mean ± s.e.m. of one of two\nindependent experiments; N = 8. h, Tumor growth dynamics upon the indicated\ntherapeutic interventions. Gray arrows indicate days of neutralizing antibody\nadministration. The data represent mean values of tumor volume measurements\n(left). Adjusted P values for day 20 measurements: IgG versus αCCR8 P < 0.0001;\nIgG versus αVEGF P < 0.0001; IgG versus αCCR8-αVEGF P < 0.0001; αCCR8\nversus αVEGF P = 0.0434; αCCR8 versus αCCR8-αVEGF P = 0.0044; αVEGF\nversus αCCR8-αVEGF P < 0.0001. i, Quantification of proportion and absolute\nnumbers of intratumoral and splenic Treg cells following treatment (left) and the\ncorresponding Treg cell numbers in spleens in the treated animals (right). Data in\n\nh and i represent the mean ± s.e.m. of one of two independent experiments, IgG N\n= 10, CCR8 N = 10, αVegf N = 8, CCR8-αVegf N = 8. j, Tumor growth dynamics upon\nthe indicated therapeutic interventions (left). Gray and black arrows indicate\ntiming of neutralizing antibody and CCR2 inhibitor (CCR2i) administration,\nrespectively. The data represent the mean ± s.e.m. values of tumor volume\nmeasurements. Adjusted P values of day 20 measurements: IgG versus αCCR8\nP = 0.0009; IgG versus αVEGF P < 0.0001; IgG versus CCR2i P < 0.0001; IgG\nversus αCCR8-αVEGF P < 0.0001; IgG versus αCCR8-CCR2i P < 0.0001; αCCR8\nversus αVEGF P = 0.9982; αCCR8 versus CCR2i P = 0.6138; αCCR8 versus αCCR8-\nαVEGF P < 0.0001; αCCR8 versus αCCR8-CCR2i P = 0.0041; αVEGF versus CCR2i\nP = 0.9551; αVEGF versus αCCR8-αVEGF P = 0.0003; αVEGF versus αCCR8-CCR2i\nP = 0.0363; CCR2i versus αCCR8-αVEGF P = 0.0018; CCR2i versus αCCR8-\nCCR2i P = 0.2271; αCCR8-αVEGF versus αCCR8-CCR2i P = 0.4530. Plots include\ndata from two independent experiments combined with nine animals in each\ngroup in experiment 1 (IgG N = 9, αCCR8 N = 9, αVEGF N = 9, CCR2i N = 9, αCCR8\nN = αVEGF-9, αCCR8-CCR2i N = 9) and 4–6 animals per group in experiment 2\n(IgG N = 4; CCR2i N = 6; CCR8-CCR2i N = 6). k, Kaplan–Meyer survival curves\nfollowed by Log-rank (Mantel–Cox) of KP tumor-bearing mice. The ‘survival’ time\nreflects the end point of the experiment when tumor volume in individual mice\nreached 1 cm3. Adjusted P values: IgG versus αCCR8 ***P < 0.0001; IgG versus\nαVEGF ***P < 0.0001; IgG versus CCR2i ***P < 0.0001; IgG versus αCCR8-αVEGF\n***P < 0.0001; IgG versus αCCR8-CCR2i ***P < 0.0001; αCCR8 versus αVEGF\nP = 0.5687 (NS); αCCR8 versus CCR2i P = 0.7411 (NS); αCCR8 versus αCCR8-\nαVEGF ***P = 0.0002; αCCR8 versus αCCR8-CCR2i P = 0.0342; αVEGF versus\nCCR2i P = 0.8054 (NS); αVEGF versus αCCR8-αVEGF ***P = 0.0006; αVEGF versus\nαCCR8-CCR2i P = 0.0666 (NS); CCR2i versus αCCR8-αVEGF ***P = 0.0003;\nCCR2i versus αCCR8-CCR2i P = 0.6749 (NS); αCCR8-αVEGF versus αCCR8-CCR2i\n*P = 0.0489. Plots include data from two independent experiments combined\nwith 5–11 animals in each group in experiment 1 (IgG N = 9, αCCR8 N = 9; αVEGF\nN = 9; CCR2i N = 11; αCCR8-αVEGF N = 9; αCCR8-CCR2i N = 5) and 4–10 animals\nper group in experiment 2 (IgG N = 7; CCR2i N = 10; CCR8-CCR2i N = 4). In b–d,\nh and i, plots are representative of one of two experiments with 8–10 mice per\ngroup each, at day 20 after transplantation. Number of mice per group in b and c:\nPBS-IgG N = 10; DT-IgG N = 10; PBS-αVEGF N = 9; DT-αVEGF N = 9; number of mice\nper group in h and i: IgG N = 10; αCCR8 N = 10; αVEGF N = 8; αCCR8-αVEGF N = 8.\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1031\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fa\n\nGFP-DTR\n\nFoxp3\n\nb\n\n3\n\n)\n\nm\nm\n\ns.c. KP\n\nDT/\nPBS\n\nαVEGF/\nIgG\n\nAnalysis\n\n(\n\nl\n\no\nv\nr\no\nm\nu\nT\n\nPBS-IgG\n\nDT-IgG\n\nPBS αVEGF\n\nDT αVEGF\n\n800\n\n600\n\n400\n\n200\n\n0\n\n*\n*\n*\n\n*\n*\n*\n\n*\n*\n*\n\n*\n\n*\n*\n*\n\n*\n\nDay 0\n\nDay 8,9\n\nDay 9, 12, 15\n\nDay 20\n\n5 7 8 9 1\n\n1\n\n2\n1\n\n3\n1\n\n4\n1\n\n5\n1\n\n6\n1\n\n9\n1\n\n0\n2\n\nDays\n\n***\n\nNS\n\nNS\n\n****\n\n**** ****\n\n****\n\nNS\n\nNS\n\n120\n\n****\n\n**** ****\n\n90\n\n60\n\n30\n\n0\n\n4\nD\nC\n\nf\no\n7\n6\nK\n%\n\nI\n\n100\n\n80\n\n60\n\n40\n\n20\n\n0\n\nPBS-IgG\n\nDT-IgG\n\nPBS-αVEGF\n\nDT-αVEGF\n\nd\n\nr\ne\nb\nm\nu\nn\nγ\n-\nN\nF\nI\n\n4\nD\nC\n\n250,000\n\n200,000\n\n150,000\n\n100,000\n\n50,000\n\n0\n\ni\n\nl\na\nv\nv\nr\nu\nS\n\n100\n\n80\n\n60\n\n40\n\n20\n\n0\n\n0\n\n20\n\n40\n\nDays\n\n****\n\n***\n\nNS\n\n****\n\n*** ****\n\nr\ne\nb\nm\nu\nn\nγ\n-\nN\nF\nI\n\n8\nD\nC\n\n100,000\n\n80,000\n\n60,000\n\n40,000\n\n20,000\n\n0\n\nPBS-IgG\nDT-IgG\nPBS αVEGF\nDT αVEGF\n\n****\n*\n\nNS\n\n**\n\n**\n\n****\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\n****\n\nNS\n\nNS\n\n*\n\nNS\n\nNS\n\n***\n\n**** ****\n\n100\n\n***\n\n**** ***\n\n8\nD\nC\n\nf\no\n7\n6\nK\n%\n\nI\n\n80\n\n60\n\n40\n\n20\n\n0\n\n120\n\n90\n\n60\n\n30\n\n0\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\n**\n\n*\n\nNS\n\nNS\n\nNS *\n\n150,000\n\n100,000\n\n50,000\n\nr\ne\nb\nm\nu\nn\no\nn\no\nM\n\n0\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\n****\n\nNS\n\nNS\n\n1,500,000\n\n****\n\n**** **\n\n1,000,000\n\n500,000\n\n0\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\n**\n\nNS\n\nNS\n\n***\n\n**\n\n**\n\n200,000\n\n150,000\n\n100,000\n\n50,000\n\n0\n\nPBS-IgG\n\nDT-IgG\nDT-αVEGF\nPBS-αVEGF\n\n4\nD\nC\n\nf\no\n\no\n\nl\n\nL\n2\n6\nD\nC\n\ni\n\nh\n\n4\n4\nD\nC\n%\n\n8\nD\nC\n\nf\no\n\no\n\nl\n\n2\n6\nD\nC\n\ni\n\nh\n\n4\n4\nD\nC\n%\n\nPBS -IgG\n\nDT-IgG\n\nPBS-αEGF\n\nDT-αEGF\n\n500 µm\n\nPBS-IgG\n\nDT-IgG\n\nPBS-αEGF\n\nDT-αEGF\n\nc\n\nr\ne\nb\nm\nu\nn\n4\nD\nC\n\nr\ne\nb\nm\nu\nn\n8\nD\nC\n\ne\n\nL\nE\nN\nU\nT\n\na\n1\nF\nI\nH\n\ng\n\n****\n\n*\n\nNS\n\nNS NS **\n\n****\n\n*\n\nNS\n\n**\n\n*\n\n****\n\nf\n\na\ne\nr\na\nl\na\nt\no\nt\n/\nL\nE\nN\nU\nT\n\n100\n\n80\n\n60\n\n40\n\n20\n\n0\n\ny\nt\ni\ns\nn\ne\nt\nn\n\ni\n\nL\nE\nN\nU\nT\n\n120\n\n80\n\n40\n\n0\n\nPBS-IgG\n\nPBS-αVEGF\nDT-IgG\nDT-αVEGF\n\nPBS-IgG\n\nPBS-αVEGF\nDT-IgG\nDT-αVEGF\n\n****\n\nNS\n\nNS\n\n**\n\nNS\n\n**\n\n**\n\n***\n\nNS\n\n100\n\nNS\n\nNS\n\n*\n\na\ne\nr\na\nl\na\nt\no\nt\n/\nα\n1\nF\nI\nH\n\n80\n\n60\n\n40\n\n20\n\n0\n\ny\nt\ni\ns\nn\ne\nt\nn\n\ni\n\nα\n1\nF\nI\nH\n\n120\n\n80\n\n40\n\n0\n\nPBS-IgG\n\nPBS-αVEGF\nDT-IgG\nDT-αVEGF\n\nPBS-IgG\n\nPBS-αVEGF\nDT-IgG\nDT-αVEGF\n\nNS\n\nNS\n\nNS\n\nNS NS NS\n\n***\n\nNS\n\nNS\n\n**\n\n*\n\n**\n\ni\n\n4\nD\nC\nP\nK\nf\no\ng\ne\nr\nT\n%\n\n80\n\n60\n\n40\n\n20\n\n0\n\nn\ne\ne\nl\np\ns\nn\n\ni\n\nr\ne\nb\nm\nu\nn\ng\ne\nr\nT\n\n100,000\n\n80,000\n\n60,000\n\n40,000\n\n20,000\n\n0\n\nIgG\nαCCR8\nαVEGF\nαCCR8-αVEGF\n\nIgG\nαCCR8\nαVEGF\nαCCR8-αVEGF\n\nIgG\n\nαCCR8\n\nαVEGF\n\nCCR2i\n\nαCCR8-αVEGF\n\nαCCR8-CCR2i\n\nIgG /αCCR8/αVEGF\n\nh\n\n)\n\n3\n\nm\nm\n\n(\n\nl\n\no\nv\nr\no\nm\nu\nT\n\n1,000\n\n800\n\n600\n\n400\n\n200\n\n0\n\nIgG\n\nαCCR8\n\nαVEGF\nαCCR8-αVEGF\n\n*\n*\n*\n\n*\n*\n**\n*\n*\n\n*\n\n*\n*\n**\n*\n\nIgG\n\n10864\n\n12 14 16 18 19 20\n\nDays\n\nIgG /αCCR8,/αVEGF,\nCCR2i\n\nNS\nNS\nNS\n\nNS\n\n*\n*\n*\n\n*\n*\n*\n\nNS\n\n*\n\n*\n*\n\n*\n*\n*\n\n*\n*\n*\n\n*\n*\n*\n\n*\n*\n*\n\n*\n*\n\nk\n\ni\n\nl\na\nv\nv\nr\nu\ns\n\nf\no\ny\nt\ni\nl\ni\n\nb\na\nb\no\nr\nP\n\n100\n\n80\n\n60\n\n40\n\n20\n\n0\n\nIgG\n\nαCCR8\n\nαVEGF\n\nCCR2i\n\nαCCR8-αVEGF\n\nαCCR8-CCR2i\n\ng\ne\nr\nT\nP\nK\nf\no\n\n+\n\n8\nR\nC\nC\n%\n\n100\n\n80\n\n60\n\n40\n\n20\n\n0\n\nj\n\n)\n\n3\n\nm\nm\n\n(\n\nl\n\no\nv\nr\no\nm\nu\nT\n\n1,000\n\n800\n\n600\n\n400\n\n200\n\n0\n\n5\n\n7\n\n9\n\n11\n\n13\n\n15\n\n18\n\n20\n\n0\n\n10\n\n20\n\n30\n\n40\n\nTime (days)\n\nDays elapsed\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1032\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\n\n\fDiscussion\nSuccesses in therapeutic targeting of PD-1 and CTLA-4 pathways in T\nlymphocytes are viewed as clinical evidence supporting the notion\nof cancer surveillance by cells of the adaptive immune system akin\nto that of pathogens. On the other hand, a growing realization of\nthe important roles immune cells play in supporting normal tissue\nfunction, maintenance and repair suggests an alternative, even if not\nmutually exclusive view of tumor–immune interactions. Within the\nlatter framework, the TME can be considered as a tissue-supporting\nmulticellular network, which in response to cues emanating from can-\ncerous cells supports their growth. In this regard, cancer represents\na special state of a parenchymal cell, whose support by both immune\nand non-immune cells is guided by common, yet poorly understood\nprinciples of tissue organization. Treg cells suppress immune responses\ndirected against self-antigens and foreign antigens to protect tissues\nfrom inflammation-associated loss of function1. Besides this indirect\ntissue-supporting functionality, Treg cells were also implicated in direct\nresponses to injury and other forms of tissue damage through produc-\ntion of tissue repair factors16,39–42 suggesting that these functions of Treg\ncells are conserved. Furthermore, Treg cells were shown to support skin\nand hematopoietic stem cell niches15,43,44. Therefore, it is reasonable to\nassume that in human solid organ malignancies and in experimental\nmouse cancers Treg cells likely play similar dual roles supporting tumor\ngrowth-promoting accessory cell states.\n\nHere, we showed that Treg cells have a profound impact on states\nof key accessory cells in a genetic autochthonous mouse model of\nNSCLC, in an experimental model of lung injury and in human LuAds.\nUsing robust unsupervised data-driven computational analyses, we\nfound that Treg cells support conserved gene programs—factors—across\nexperimental models of lung cancer and injury, suggesting their role\nin coordinating broad, shared accessory cell functions that extend\nto various conditions and tissue states. The latter included human\nimmunomodulatory C1Q+ (CFH, CR1L, LAG3, PDCD1LG2, LILRB4, IL18BP)\nand SPP1/FOLR2 macrophage factors and their mouse counterparts,\nwhich were positively associated with the Treg cell presence. A similar\nmacrophage gene program is also reported to be enriched in NSCLC\nlesions45 and sustained by Treg cells in mouse models of melanoma and\nbreast cancer46.\n\nOur analysis of the distribution of activated cell types and gene\nexpression programs with respect to their localization within and\naround tumor nodules showed high concordance of characteristic\ngene programs that were identified by scRNA-seq and ST analyses\nin situ. Notably, Treg cell depletion induced the IC response program\nlocalized to tumor nodule cores, while the IFN response program\nwas most notable in the margins of tumor foci. The display of these\nprograms by multiple cell types present within the same local niche\nsuggests that they are elicited by common signals (‘signal niche’), for\nexample, hypoxia response in the tumor nodule cores and a transient\nburst of IFN-γ produced by CD8+ T cells and NK cells concentrated in\nthe tumor margins47. We also observed heterogeneity between Treg cell\ndepletion responsive and nonresponsive tumor foci distinguished by\nthe presence or paucity of the IC gene program. Interestingly, tumor\nnodules that failed to induce the IC gene program in response to Treg\ncell depletion expressed Sox9 in agreement with a recent study where\nupregulation of Sox9 in human LuAd conferred resistance to NK cells27.\nAmong the conserved gene programs negatively associated\nwith Treg cell presence in mouse and human lung cancer, we noted the\nVEGF signaling pathway. This included increased expression of VEGF\nsignaling-related genes in endothelial cells and increased expres-\nsion of VEGF-A in myeloid and other cell types. This most likely rein-\nforces the immunosuppressive TME providing support for tumor\ngrowth consistent with a recent report of tumor ischemia caused by\nthe transient spike in intratumoral IFN-γ following CD25 antibody\nphotoimmunotherapy-induced Treg cell depletion47. In addition,\nlung EC-derived VEGF was shown to specify development of CAR4hi\n\nendothelial cells and promote vascularization and tissue regenera-\ntion following injury48,49. VEGF has also been suggested to exert an\nimmunomodulatory effect on cells of the innate and adaptive immune\nsystem50. Considering VEGF targeting being an approved therapy for\nsome human cancers, combined VEGF-A and Treg cell targeting serves\nas a proof-of-concept for a rational combination therapy instructed\nby the new knowledge of TME transcriptional connectivity. While\nnear complete loss of the Treg cell pool in KP-DTR mice led to systemic\nautoimmunity and inflammation, VEGF blockade coupled with CCR8\nantibody-mediated depletion of intratumoral Treg cells showed impres-\nsive therapeutic efficacy with no adverse effects. The latter owes to the\nfact that CCR8 expression is selectively enriched in highly activated\nintratumoral Treg cell subsets in human and mouse malignancies35,36,38.\nOur observation that a combination of CCR8 antibody-mediated intra-\ntumoral Treg cell depletion with CCR2 blockade did not yield additional\nbenefit in comparison to the corresponding monotherapies suggests\nthat the latter either directly or indirectly converge on a shared regula-\ntory node and highlights the utility of preclinical selection of combina-\ntorial therapeutic strategies informed by scRNA-seq and ST analyses\nof early TME responses.\n\nOur study highlights a generalizable approach where perturba-\ntion of a given cell population in an engineered genetic cancer model\nenabled computational learning of its ‘connectivity’ and influence on\nthe TME and other diseased tissue states, which could then be com-\npared to the human clinical settings. A surfeit of secreted and cell\nsurface molecules has been implicated in Treg cell-mediated immuno-\nsuppressive and tissue-supporting functions. However, none of these\nindividual modalities can predominantly account for the bulk of these\nfunctionalities. Combinatorial targeting of these putative mediators\nwill enable elucidation of the molecular mechanisms of the observed\nTreg cell dependencies in the TME. Our results suggest that Treg cells\nserve as an essential component of a complex network of accessory\ncells of both hematopoietic and non-hematopoietic origin. Shared\nperturbations in their transcriptional states observed across the three\ndifferent settings imply that the identified interdependencies of Treg\ncells and other components of tissue-supporting cellular networks are\nconserved and can be exploited to develop new strategies for rational\ntherapies of cancer and other diseases.\n\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information,\nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41590-023-01504-2.\n\nReferences\n1.\n\nJosefowicz, S. Z., Lu, L.-F. & Rudensky, A. Y. Regulatory T cells:\nmechanisms of differentiation and function. Annu. Rev. Immunol.\n30, 531–564 (2012).\n\n2. Sakaguchi, S. et al. Regulatory T cells and human disease. Annu.\n\nRev. Immunol. 38, 541–566 (2020).\n\n3. Glasner, A. & Plitas, G. Tumor resident regulatory T cells. Semin.\n\nImmunol. 52, 101476 (2021).\n\n4. Bos, P. D. Treg cells in cancer: beyond classical immunological\n\ncontrol. Immunol. Invest. 45, 721–728 (2016).\n\n5. Bos, P. D., Plitas, G., Rudra, D., Lee, S. Y. & Rudensky, A. Y.\n\nTransient regulatory T cell ablation deters oncogene-driven\nbreast cancer and enhances radiotherapy. J. Exp. Med. 210,\n2435–2466 (2013).\n\n7.\n\n6. Grinberg-Bleyer, Y. et al. NF-κB c-Rel is crucial for the regulatory\nT cell Immune checkpoint in cancer. Cell 170, 1096–1108 (2017).\nJoshi, N. S. et al. Regulatory T cells in tumor-associated tertiary\nlymphoid structures suppress anti-tumor T cell responses.\nImmunity 43, 579–590 (2015).\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1033\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\f8. Fujimura, T., Ring, S., Umansky, V., Mahnke, K. & Enk, A. H.\n\nRegulatory T cells stimulate B7-H1 expression in myeloid-derived\nsuppressor cells in ret melanomas. J. Invest. Dermatol. 132,\n1239–1246 (2012).\n\n9. Gyori, D. et al. Compensation between CSF1R+ macrophages and\n\nFoxp3+ Treg cells drives resistance to tumor immunotherapy. JCI\nInsight 3, 120631 (2018).\n\n10. DuPage, M., Dooley, A. L. & Jacks, T. Conditional mouse lung\ncancer models using adenoviral or lentiviral delivery of Cre\nrecombinase. Nat. Protoc. 4, 1064–1072 (2009).\n\n11. Herter-Sprie, G. S. et al. Synergy of radiotherapy and PD-1\n\nblockade in Kras-mutant lung cancer. JCI Insight 1, e87415 (2016).\n\n12. Skoulidis, F. et al. STK11/LKB1 mutations and PD-1 inhibitor\n\nresistance in KRAS-mutant lung adenocarcinoma. Cancer Discov.\n8, 822–835 (2018).\n\n13. Kim, J. M., Rasmussen, J. P. & Rudensky, A. Y. Regulatory T cells\nprevent catastrophic autoimmunity throughout the lifespan of\nmice. Nat. Immunol. 8, 191–197 (2007).\n\n31. Dhainaut, M. et al. Spatial CRISPR genomics identifies regulators\nof the tumor microenvironment. Cell 185, 1223–1239 (2022).\n32. Campbell, J. R. et al. Fc-Optimized anti-CCR8 antibody depletes\nregulatory T cells in human tumor models. Cancer Res. 81,\n2983–2994 (2021).\n\n33. Onda, M., Kobayashi, K. & Pastan, I. Depletion of regulatory\n\nT cells in tumors with an anti-CD25 immunotoxin induces CD8\nT cell-mediated systemic antitumor immunity. Proc. Natl Acad.\nSci. USA 116, 4575–4582 (2019).\n\n34. Martinez-Usatorre, A. et al. Overcoming microenvironmental\n\nresistance to PD-1 blockade in genetically engineered lung cancer\nmodels. Sci. Transl. Med. 13, eabd1616 (2021).\n\n35. Plitas, G. et al. Regulatory T cells exhibit distinct features in\nhuman breast cancer. Immunity 45, 1122–1134 (2016).\n\n36. De Simone, M. et al. Transcriptional landscape of human tissue\n\nlymphocytes unveils uniqueness of tumor-infiltrating T regulatory\ncells. Immunity 45, 1135–1147 (2016).\n\n37. Van Damme, H. et al. Therapeutic depletion of CCR8+\n\n14. Fan, X. et al. CD49b defines functionally mature Treg cells that\n\nsurvey skin and vascular tissues. J. Exp. Med. 215, 2796–2814 (2018).\n15. Ali, N. et al. Regulatory T cells in skin facilitate epithelial stem cell\n\ntumor-infiltrating regulatory T cells elicits antitumor immunity\nand synergizes with anti-PD-1 therapy. J. Immunother. Cancer 9,\ne001749 (2021).\n\ndifferentiation. Cell 169, 1119–1129 (2017).\n\n38. Villarreal, D. O. et al. Targeting CCR8 induces protective\n\n16. Burzyn, D. et al. A special population of regulatory T cells\npotentiates muscle repair. Cell 155, 1282–1295 (2013).\n\n17. Feuerer, M. et al. Lean, but not obese, fat is enriched for a unique\npopulation of regulatory T cells that affect metabolic parameters.\nNat. Med. 15, 930–939 (2009).\n\n18. Levine, J. H. et al. Data-driven phenotypic dissection of AML\n\nreveals progenitor-like cell that correlate with prognosis. Cell 162,\n184–197 (2015).\n\nantitumor immunity and enhances vaccine-induced responses in\ncolon cancer. Cancer Res. 78, 5340–5348 (2018).\n\n39. Arpaia, N. et al. A distinct function of regulatory T cells in tissue\n\nprotection. Cell 162, 1078–1089 (2015).\n\n40. Dombrowski, Y. et al. Regulatory T cells promote myelin\n\nregeneration in the central nervous system. Nat. Neurosci. 20,\n674–680 (2017).\n\n41. Hui, S. P. et al. Zebrafish regulatory T cells mediate organ-specific\n\n19. Dann, E., Henderson, N. C., Teichmann, S. A., Morgan, M. D. &\nMarioni, J. C. Differential abundance testing on single-cell\ndata using k-nearest neighbor graphs. Nat. Biotechnol. 40,\n245–253 (2022).\n\n20. van Dijk, D. et al. Recovering gene interactions from single-cell\n\ndata using diffusion. Cell 174, 716–729 (2018).\n\nregenerative programs. Dev. Cell 43, 659–672 (2017).\n\n42. Ito, M. et al. Brain regulatory T cells suppress astrogliosis and\npotentiate neurological recovery. Nature 565, 246–250 (2019).\n43. Fujisaki, J. et al. In vivo imaging of Treg cells providing immune\nprivilege to the haematopoietic stem-cell niche. Nature 474,\n216–219 (2011).\n\n21. Levitin, H. M. et al. De novo gene signature identification from\n\n44. Hirata, Y. et al. CD150high bone marrow Tregs maintain\n\nsingle‐cell RNA‐seq with hierarchical Poisson factorization. Mol.\nSyst. Biol. 22, e8557 (2019).\n\nhematopoietic stem cell quiescence and immune privilege via\nadenosine. Cell Stem Cell 22, 445–453 (2018).\n\n22. Szabo, P. A. et al. Single-cell transcriptomics of human T cells\nreveals tissue and activation signatures in health and disease.\nNat. Commun. 10, 4706 (2019).\n\n23. Ma, J., Sanchez-Duffhues, G., Goumans, M.-J. & Ten Dijke, P.\n\nTGF-β-induced endothelial-to-mesenchymal transition in disease\nand tissue engineering. Front. Cell Dev. Biol. 8, 260 (2020).\n\n24. Potenta, S., Zeisberg, E. & Kalluri, R. The role of\n\nendothelial-to-mesenchymal transition in cancer progression.\nBr. J. Cancer 99, 1375–1379 (2008).\n\n25. Chu, T., Wang, Z., Pe’er, D. & Danko, C. G. Cell type and gene\n\nexpression deconvolution with BayesPrism enables Bayesian\nintegrative analysis across bulk and single-cell RNA sequencing\nin oncology. Nat. Cancer 3, 505–517 (2022).\n\n45. Casanova-Acebes, M. et al. Tissue-resident macrophages provide\n\na pro-tumorigenic niche to early NSCLC cells. Nature 595,\n578–584 (2021).\n\n46. Liu, C. et al. Treg cells promote the SREBP1-dependent metabolic\nfitness of tumor-promoting macrophages via repression of CD8+\nT cell-derived interferon-γ. Immunity 51, 381–397 (2019).\n\n47. Kurebayashi, Y. et al. Rapid depletion of intratumoral regulatory\nT cells induces synchronized CD8 T- and NK-cell activation\nand IFNγ-dependent tumor vessel regression. Cancer Res. 81,\n3092–3104 (2021).\n\n48. Niethamer, T. K. et al. Defining the role of pulmonary endothelial\n\ncell heterogeneity in the response to acute lung injury. Elife 9,\ne53072 (2020).\n\n26. Niec, R. E. et al. Lymphatics act as a signaling hub to regulate\n\nintestinal stem cell activity. Cell Stem Cell 29, 1067–1082 (2022).\n\n27. Laughney, A. M. et al. Regenerative lineages and\n\n49. Vila Ellis, L. et al. Epithelial Vegfa specifies a distinct endothelial\npopulation in the mouse lung. Dev. Cell 52, 617–630 (2020).\n50. Lapeyre-Prost, A. et al. Immunomodulatory activity of VEGF in\n\nimmune-mediated pruning in lung cancer metastasis. Nat. Med.\n26, 259–269 (2020).\n\n28. Yang, D. et al. Lineage tracing reveals the phylodynamics,\n\nplasticity, and paths of tumor evolution. Cell 185,\n1905–1923 (2022).\n\ncancer. Int. Rev. Cell Mol. Biol. 330, 295–342 (2017).\n\nPublisher’s note Springer Nature remains neutral with regard to\njurisdictional claims in published maps and institutional affiliations.\n\n29. Marjanovic, N. D. et al. Emergence of a high-plasticity cell state\nduring lung cancer evolution. Cancer Cell 38, 229–246 (2020).\n\n30. LaFave, L. M. et al. Epigenomic state transitions characterize\n\ntumor progression in mouse lung adenocarcinoma. Cancer Cell\n38, 212–228 (2020).\n\nOpen Access This article is licensed under a Creative Commons\nAttribution 4.0 International License, which permits use, sharing,\nadaptation, distribution and reproduction in any medium or format,\nas long as you give appropriate credit to the original author(s) and the\nsource, provide a link to the Creative Commons license, and indicate\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1034\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fif changes were made. The images or other third party material in this\narticle are included in the article’s Creative Commons license, unless\nindicated otherwise in a credit line to the material. If material is not\nincluded in the article’s Creative Commons license and your intended\nuse is not permitted by statutory regulation or exceeds the permitted\n\nuse, you will need to obtain permission directly from the copyright\nholder. To view a copy of this license, visit http://creativecommons.\norg/licenses/by/4.0/.\n\n© The Author(s) 2023, corrected publication 2023\n\n1Immunology Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 2Program for Computational and Systems\nBiology, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 3Institute of Biotechnology, Life Sciences Centre,\nVilnius University, Vilnius, Lithuania. 4Human Oncology & Pathogenesis Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center,\nNew York, NY, USA. 5Department of Pathology & Laboratory Medicine, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY,\nUSA. 6Department of Medicine, Thoracic Oncology Service, New York, NY, USA. 7Antitumor Assessment Core Facility, New York, NY, USA. 8Molecular\nPharmacology Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 9Howard Hughes Medical Institute, Sloan\nKettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 10These authors contributed equally: Ariella Glasner, Samuel A. Rose,\nRoshan Sharma.\n\n e-mail: peerd@mskcc.org; rudenska@mskcc.org\n\nNature Immunology | Volume 24 | June 2023 | 1020–1035\n\n1035\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fMethods\nExperimental model and mouse details\nMice. Animals were housed at the Memorial Sloan Kettering Cancer\nCenter (MSKCC) animal facility under specific pathogen-free condi-\ntions according to institutional guidelines. All studies were performed\nunder protocol 08-10-023 and approved by the MSKCC Institutional\nAnimal Care and Use Committee. Mice used in this study had no previ-\nous history of experimentation or exposure to drugs. Foxp3GFP-DTR and\nKrasLSL-G12D Trp53fl/fl mice were previously described10,13. Adult male and\nfemale mice (6 weeks or older) were used for all experiments.\n\nLung adenocarcinoma and bleomycin-induced fibrotic injury induc-\ntion. Cre recombinase-mediated induction of KP LuAds was previously\ndescribed10. Briefly, mice were anesthetized with a 160–180 μl keta-\nmine–xylazine mixture and infected with Cre recombinase-expressing\nadenovirus (1 × 108 plaque-forming units) via intratracheal administra-\ntion. Tumors developed within approximately 3 months. For the induc-\ntion of fibrotic injury, pharmaceutical-grade bleomycin (Fresenius\nKabi) was administered intranasally to anesthetized mice (0.06 U per\nmouse). For the s.c. KP tumor growth model, KP cells were resuspended\nin sterile PBS and injected to the flank subcutaneous space (1 × 106 KP\ncells in 200 μl per mouse).\n\nDiphtheria toxin, VEGF, PD-1, CCR8 antibody and CCR2i treatments.\nDT (List Biological Laboratories) was administered to mice (1 μg per\nmouse in PBS) via retro-orbital injection twice on two consecutive\ndays. For tumor transplantation experiments, DT was injected on days\n8 and 9 after tumor s.c. transplantation. Mouse polyclonal neutral-\nizing VEGF-A antibody (R&D clone AF-493-M) or control IgG (BioXcell\nclone BE0130) were injected on days 9, 12 and 15 (20 μg per mouse\nper injection) with or without DT, or on days 8, 10, 12, 14 and 17 with\nor without CCR8 antibody (BioLegend clone SA214G2; 240 μg per\nmouse per injection). PD-1 antibody (BioXcell clone BE0146) alone\nor in combination with VEGF-A antibody was administered on days\n8, 10, 12, 14 and 17 (250 μg per mouse per injection). RS-504393 CCR2\ninhibitor (CCR2i; 2517, Tocris) was administered (50 mg per kg body\nweight) daily in combination with CCR8 antibody, VEGF-A antibody or\ntheir combination. In these experiments, CCR8 and VEGF-A antibodies\nwere administered on days 8, 10 and 12, and CCR2i was administered\ndaily starting on day 8 and ending on day 12.\n\nHuman lung adenocarcinoma samples. Individuals with LuAd under-\ngoing a surgical resection or tissue biopsy at MSKCC were identified\nand biospecimens collected prospectively from 2017 to 2020. All par-\nticipants from whom biospecimens were obtained provided informed\nconsent for an MSKCC-wide biospecimen collection and analysis pro-\ntocol. Recruitment was designed to capture a wide, unbiased swath\nof heterogeneous disease, with a slight emphasis on EGFR-mutated\ntumors with a high propensity to transform to more aggressive sub-\ntypes. Biases may be present related to this recruitment design, the\nrace, sex, smoking status and the general population of MSKCC. Use of\nall participant material and data described in this paper was performed\nunder ethical approval obtained from the MSKCC Institutional Review\nBoard (study nos. 06-107 and 12-245). Only continuous trends between\ncell proportion and factor use were assessed across all participants\nand therefore controls based on sample groupings are not relevant.\n\nCell isolation and flow cytometry. For isolation of immune and stro-\nmal cells, lungs were perfused, placed into 5 ml microcentrifuge tubes\ncontaining 400 μl of cold serum-free RPMI and chopped with scis-\nsors (1–2 mm). Lung fragments were placed in 2–3 ml of pre-warmed\ndigestion medium (RPMI 1640, 10 mM HEPES buffer pH 7.2 to 7.6, 1%\npenicillin–streptomycin, 1% l-glutamine, liberase (Sigma-Aldrich,\n05401020001) and 1 U ml−1 DNase I (Sigma-Aldrich, 10104159001;\n2–3 ml)) and incubated for 30 min at 37 °C. After digestion, supernatant\n\nNature Immunology\n\nwas collected and cells were resuspended in ice-cold RPMI 1640 contain-\ning 5% FCS (Thermo Fisher, 35010CV), 1 mM HEPES pH 7.2 to 7.6 (Corn-\ning, MT25060CI), 1% penicillin–streptomycin (Corning, MT30002CI)\nand 200 mM l-glutamine (Corning, MT25005CI). After additional\ndigestion for 1 h of the remaining tissue, both digested cell fractions\npassed through a 100-μm strainer (Corning, 07-201-432), washed and\nFACS sorted. For cell isolation from transplanted KP tumor-bearing\nmice, tumors were placed into 5 ml microcentrifuge tubes containing\n400 μl of cold serum-free RPMI 1640, chopped with scissors and incu-\nbated in digestion medium containing 1 mg ml−1 collagenase (Sigma,\n11088793001) and 1 U ml−1 DNase I (Sigma-Aldrich, 10104159001) and\nbeads on a shaker at 37 °C for 1 h. For cytokine production measure-\nments, cells were incubated at 37 °C, 5% CO2 for 3 h in the presence of\n50 ng ml−1 phorbol-12-myristate-13-acetate (Sigma-Aldrich, P8139),\n500 ng ml−1 ionomycin (Sigma-Aldrich, I0634), 1 μg ml−1 brefeldin A\n(Sigma-Aldrich, B6542) and 2 μM monensin (Sigma-Aldrich, M5273).\nCells were stained with Ghost Dye Red 780 (Tonbo Bioscience, 13-0865)\nor Zombie NIR Flexible Viability Kit (BioLegend, 423106) and a mixture\nof fluorophore-conjugated antibodies for 30 min at 4 °C cells, washed\nand fixed with 1% paraformaldehyde (Electron Microscopy Sciences,\n15710). For intracellular staining, cells were fixed and permeabilized\nwith the BD Cytofix/Cytoperm Kit or with the Thermo Fisher Transcrip-\ntion Factor Fix/Perm Kit according to the manufacturer’s instructions\nand analyzed on a BD LSR II flow cytometer or sorted on a BD Aria II flow\ncytometer. Post-sort cell purity was routinely higher than 95%. Flow\ncytometry data were collected on an LSR II using FACS Diva v8.0 (BD),\nor on Aurora using SpectroFlo v2.2.0.3 (Cytek). Flow cytometry data\nwere analyzed using FlowJo v 10.6.1 (BD).\n\nImmunofluorescence microscopy, histological and spatial tran-\nscriptomic analyses. Perfused lungs were fixed for 1 h at 22 °C in 4%\nparaformaldehyde and dehydrated at 4 °C in 30% sucrose, snap-frozen\nin OCT compound (Sakura Tissue-Tek, 4583). For ST, samples were\nflash frozen without fixation. All samples were sectioned with a Leica\nCM1950 Cryostat at −2 °C, to a thickness of 10 μm. Sections were fixed\nin acetone for 20 min at −20 °C, rehydrated in PBS, blocked with 10%\nnormal donkey serum ( Jackson ImmunoResearch, 017-000-121) in PBS,\n0.3% Triton X-100, and stained overnight with fluorophore-conjugated\nantibodies at 4 °C in a humidified chamber. Thereafter, nuclei were\nstained with DAPI (5 mg ml−1; Abcam, 28718-90-3) or Draq7 (5 μM;\nAbcam, 109202) for 20 min at 22 °C. Sections were imaged in Slow-\nFade mounting medium (Life Technologies, S36938) using a confocal\nLeica SP8 microscope. For histology, tissues were fixed in 10% neu-\ntral buffered formalin, embedded in paraffin, and sectioned. For the\nTUNEL assay, sections were processed under standardized conditions\nusing the DeadEnd Fluorometric Detection System (Promega, G3250),\nand subsequent immunohistochemistry was carried out using BOND\nPolymer Refine Detection Kit (Leica, DS9800), according to the manu-\nfacturer’s instructions. All Images were processed and analyzed using\nImageJ package v2.0.0-rc-69/1.52p. Distances between cells of interest\nwere quantified following the same strategy and using similar code as\ndescribed elsewhere51.\n\nAntibodies. See Supplementary Table 20 for all antibodies used in\nthis study.\n\nRNA-seq library preparation and sequencing. Cell populations\nwere sorted straight into TRIzol (Thermo Fisher, 15596018), RNA was\nprecipitated with isopropanol and linear acrylamide, washed with 75%\nethanol, and resuspended in RNase-free water. After RiboGreen quan-\ntification and quality control by Agilent BioAnalyzer, 0.4–2.0 ng total\nRNA with RNA integrity numbers ranging from 1.0 to 9.9 underwent\namplification using the SMART-Seq v4 Ultra Low Input RNA Kit (Clo-\nnetech, 63488), with 12 cycles of amplification. Subsequently, 1.5–10 ng\nof amplified cDNA was used to prepare libraries with the KAPA Hyper\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fPrep Kit (Kapa Biosystems, KK8504) using 8 cycles of PCR. Samples\nwere barcoded and run on a HiSeq 4000 or HiSeq 2500 in rapid mode\nin a 50 bp/50 bp paired-end run, using the HiSeq 3000/4000 SBS Kit\nor HiSeq Rapid SBS Kit v2 (Illumina). An average of 32 million paired\nreads were generated per sample, and the percentage of mRNA bases\nper sample ranged from 62% to 88%.\n\nRNA-seq analysis. Paired-end RNA-seq reads were mapped to the\ngenome using STAR52 v2.7.3a. Gene annotations were downloaded\nfrom Ensembl release 83, which is based on mouse genome assembly\nGRCm38. R v3.6.0 was used for generating count matrices and DESeq2\n(ref. 53) was used for principal-component analysis, to identify DEGs\nand for Spearman correlations calculations and for hierarchical clus-\ntering and generation of k-means heat maps.\n\nSingle-cell RNA sequencing. Single-cell RNA-seq was performed on\nFACS-sorted mouse lung KP cells or human LuAd samples, on the Chro-\nmium instrument (10X Genomics) following the user guide manual\n(CG00052 Rev E) for 3′ v2 and v3 as previously described54. Briefly,\nsorted cells were washed once with PBS containing 0.04% BSA and\nresuspended in PBS containing 0.04% BSA to a final concentration\nof 700–1,200 cells per μl. Viability of cells was confirmed to be above\n80%, as confirmed with 0.2% (wt/vol) Trypan Blue staining (Countess\nII). Then samples were encapsulated in microfluidic droplets at a dilu-\ntion of ∼70 cells per ml (doublet rate ∼3.9%). Encapsulated cells were\nsubjected to a reverse transcription (RT) reaction at 53 °C for 60 min.\nAfter RT, the emulsion droplets were broken and barcoded cDNA was\npurified with DynaBeads MyOne SILANE, followed by 14 cycles of PCR\namplification (98 °C for 180 s; (98 °C for 15 s, 67 °C for 20 s, 72 °C for\n60 s) × 12 cycles; 72 °C for 60 s). Then, 50 ng of PCR-amplified barcoded\ncDNA was fragmented with the reagents provided in the kit and puri-\nfied with SPRI beads to obtain an average fragment size of 600 bp.\nNext, the DNA library was ligated to the sequencing adaptor followed\nby indexing PCR (98 °C for 45 s; (98 °C for 20 s, 54 °C for 30 s, 72 °C for\n20 s) × 10 cycles; 72 °C for 60 s). An average of 5,000 cells were targeted\nfor each tumor sample. The resulting DNA library was double-size\npurified (0.6–0.8×) with SPRI beads and sequenced on an Illumina\nNovaSeq platform (R1: 26 cycles (KP), 28 cycles (LuAd); i7: 8 cycles; R2:\n96 cycles (KP), 90 cycles (LuAd)) resulting in 184.5–186.1 million reads\nper sample (average reads per single cell, 42,000; average reads per\ntranscript, 4.40–7.14; KP).\n\nVisium spatial gene expression slides were permeabilized at 37 °C\nfor 12–18 min and polyadenylated. mRNA was captured by primers\nbound to the slides. RT, second-strand synthesis, cDNA amplification\nand library preparation proceeded using the Visium Spatial Gene\nExpression Slide & Reagent Kit (10X Genomics PN 1000184) accord-\ning to the manufacturer’s protocol. After evaluation by real-time PCR,\ncDNA amplification included 11–12 cycles; sequencing libraries were\nprepared with 8 cycles of PCR. Indexed libraries were pooled equimolar\nand sequenced on a NovaSeq 6000 in a PE28/120 run using the NovaSeq\n6000 S1 Reagent Kit (200 cycles; Illumina). An average of 219 million\npaired reads were generated per sample.\n\nComputational analysis of scRNA-seq data. For basic pre-processing\nand lineage identification see Supplementary Methods. We performed\ndimensionality reduction using principal-component analysis (specify-\ning 50 principal components (PCs); nPC = 50), then visualized the data\nin two dimensions 2D using t-SNE on the PCs (perplexity parameter set\nto 50 (KP) or 100 (injury)). The cells were grouped into clusters using\nPhenoGraph18 on the PC space, with k = 30 (Extended Data Fig. 2b). We\nestablished that clustering was robust to slight changes in k, by reclus-\ntering the cells under varying k (k ϵ (20, 25, 30, 35, 40, 45)) and measur-\ning consistency using the adjusted Rand index (using the sklearn\npackage in Python), obtaining an average Rand index > 0.85. To anno-\ntate each cluster as a specific lineage, we computed the average\n\nNature Immunology\n\nexpression of known lineage markers (Extended Data Fig. 2c,d). All the\ngenes used for annotation are listed in the heat map48,55–60.\n\nFor human LuAd samples, non-empty droplets were defined using\nCellBender on a per-sample basis61. The expected number of cells was\ndefined by SEQC output after the initial quality filters described above,\nplus 25,000, to ensure an adequate number of empty droplets in each\nhuman sample. A learning rate of 0.0001 (modified to 0.00005 for\nsamples needing a slower learning rate) was used with 300 epochs.\nViable cells were identified with a library size greater than 500 UMIs,\ngene number greater than 250, log10 genes per UMI greater than 0.8\n(complexity), and less than 20% mitochondrial transcripts.\n\nUMI counts from non-empty droplets with doublets removed\nwere normalized by first dividing by the library size (UMI counts per\ndroplet), multiplying by a scale factor of 10,000, and then taking the\nnatural logarithm of 1 + the normalized counts. Before dimensionality\nreduction and clustering, genes were filtered out if they were detected\nin less than 10 cells, had low transcript annotation quality (transcript\nsupport level 4 or 5 in Ensembl 85), or belonged to categories includ-\ning mitochondrial transcripts, highly expressed ncRNAs, ribosomal\nRNAs, immunoglobulin transcripts, hemoglobin genes or T cell antigen\nreceptor variable regions. This resulted in 18,597 retained genes and\n84,909 cells. The median total counts and number of cells per sample\nare listed in Supplementary Table 11.\n\nDoublet detection. For all mouse model samples, we performed\ndoublet detection using Scrublet62 with default parameters (that is,\nexpected_doublet_rate = 0.06, min_counts = 2, min_cells = 3, min_gene_\nvariability_pctl = 85, log_transform = true, n_prin_comps = 30) on each\nsample individually. Since we were more interested in analyzing specific\nlineages, we removed doublets when processing each lineage individu-\nally (as described below).\n\nIn human samples, doublets were identified on non-empty drop-\nlets for each sample individually using DoubleDetection (https://doi.\norg/10.5281/zenodo.2678041) with a P-value threshold of 1 × 10−7 and\na voter threshold of 0.8. This algorithm was used because of its higher\nrelative accuracy among doublet detection methods63, important for\nconsistency across heterogeneous sample mixtures. Doublets were\nremoved before lineage identification.\n\nDensity plots. For analysis of individual lineages (mouse samples), see\nSupplementary Methods. t-SNE plots are valuable to build a hypoth-\nesis but it can be difficult to glean the density of cells from different\nconditions due to cells (dots) overlapping on top of each other. To\ncomplement the t-SNE plots (colored by conditions such as Fig. 2b,e)\nand further highlight the finding that Treg cell depletion has different\neffects in different cell populations, we chose to represent the distribu-\ntion of the cells in the t-SNE plot using a density plot. We used the kde-\nplot implementation in Seaborn package in Python (with non-default\nparameter thres = 0).\n\nscRNA-seq differential expression. Differential expression test-\ning between tumor cells of control or DT conditions was performed\nusing MAST64 on log-normalized values. Only genes in at least 10%\nof cells in either condition and a minimum log fold change of 0.25\n(2,654 genes) were used as input. Significant genes were defined as\nadjusted P value < 0.05 and log fold change > 0.5. Gene-set enrichment\nanalysis was performed using fgsea65 with log fold-change values of\nsignificant genes. Gene sets derived from tumor clustering in work by\nMarjanovic et al.29 were used to assess enrichment (Extended Data Fig.\n8g; scrna_tumor_de_fgsea).\n\nMilo analysis. Milo incorporates information from biological rep-\nlicates to assign a P value for fold changes in neighborhood cellular\nabundance between experimental conditions, where neighborhoods\nare defined as regions of transcriptionally similar cells in a k-nearest\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fneighbor (kNN) graph generated. This method provided us with rigor-\nous statistics to compare the frequency of different transcriptional\nstates between conditions.\n\nFor each lineage, we sought to quantify the changes in density\nof control and DT cells in each neighborhood in the kNN graph using\nMilo19. Conceptually, Milo is analogous to differential gene expression\nanalysis, but instead of identifying genes that are differential between\ntwo groups of cells, Milo tests for differential cell density in (possibly\noverlapping) neighborhoods in the kNN graph, across different condi-\ntions. Milo also considers the originating sample of each cell and treats\nany batch effect as a covariate. However, since we did not observe sig-\nnificant batch effects in our data, the design matrix we supplied only\nincluded the sample identity and experimental condition of each cell.\nTo perform the analysis, we first constructed a kNN graph (k = 30)\non PCs using the buildGraph function in Milo. For each lineage, we used\nthe same number of PCs (nPCs = 50) as for clustering and cell-type\nannotation above. We constructed neighborhoods on top of the kNN\ngraph using the Milo makeNhoods function with default parameters\n(prop = 0.1, refined = true), then counted cells in each neighborhood\nusing the countCells function and assessed statistical significance\nusing testNhoods and calcNhoodDistance for spatial FDR correction.\nWe used default parameters in all these cases. Results were then visu-\nalized using the plotNhoodGraphDA function with alpha set to 1 in all\ncases (implying that neighborhoods with spatial FDR < 1 are colored in\nall visualizations). We further assigned each neighborhood to a cell type\nif more than 80% of the cells in it belonged to that cell type; otherwise,\nthe neighborhood was termed ‘mixed’.\n\nFactor analysis. To identify gene programs and their usage across\ncells, we used the scHPF package21. scHPF is a Bayesian factorization\nmethod that explicitly models sparsity in scRNA-seq count data, using\nhierarchical Poisson factorization to achieve positive-valued loadings\nacross a selected number of factors for individual cells and genes. The\nmethod provides gene scores, which assign each gene a score for gene\nmembership in a factor, and cell scores, which quantify the usage of\neach factor by a cell. Cells with high cell scores for a factor will use the\ngene program represented by that factor at higher levels; the gene pro-\ngram, in turn, consists of genes with high gene scores for that factor. In\nthe context of response to Treg cell perturbation in cells from different\nlineages, scHPF provided an ideal unsupervised and data-driven way\nto extract gene programs (factors) that are systematically altered by\nthe perturbation.\n\nIn the mouse tumor samples, scHPF was run using default hyperpa-\nrameters in the endothelial, fibroblast and myeloid lineages to obtain\n20 endothelial-specific factors, 25 fibroblast-specific factors and 25\nmyeloid-specific factors.\n\nDifferential factor usage between diphtheria toxin and control. We\nexpect that the coordinated gene program response to the impact of\nTreg cell depletion should reflect as factor cell scores being differential\nbetween control and DT conditions. To quantify this, we computed the\naverage cell score of every factor in each cluster of cells (grouped by the\ncell type they belong to) for each condition. This result is presented as a\nheat map in Fig. 2i for endothelial cells, Extended Data Fig. 5a for fibro-\nblasts, Extended Data Fig. 5c for myeloid cells in the tumor model and\nFig. 3g for endothelial cells in the bleomycin injury model. Investigating\naverages at the cluster level ensures that any factors that reflect subtle\nshifts in cell states within a cell type will be identified. We then studied\nthose factors that have higher averages in DT compared to control.\n\nTo ensure that our factors are significantly differential between\ncontrol and DT, we considered cell scores for each factor in each cluster\nand computed P values between the two conditions using a Mann–Whit-\nney U test as implemented in the scipy.stats.mannwhitneyu package\nin Python. The P values are reported in Supplementary Table 6. We\nthen considered factors that were robust to random initialization of\n\nNature Immunology\n\nscHPF (Supplementary Fig. 1 and ‘Robustness analysis of factors’),\nwere biologically relevant and had P values < 0.01 for further analysis.\n\nRobustness analysis of factors. We assessed the robustness of the\nobtained factors in two ways. First, we sought to ensure that the\nobtained factors were robust to random initializations. For this, we\nfixed the number of factors computed and reran the model for 20\niterations. To quantify the similarity across iterations, we computed\nPearson correlation (for both gene and cell scores) between best match-\ning factors between iterations. The best matching factors between\nany two iterations were identified using an implementation of the\nHungarian66 matching algorithm. The algorithm matches each factor\nfrom one iteration to the best matching factor from a second itera-\ntion such that the total cost is minimized, where the cost is defined as\n(1 − pairwise correlation score between two iterations). We used the\nPython (v3.8) implementation of the linear_sum_assignment function\nin the optimize module of SciPy package (v1.7.1). After matching, we\nreported the median correlation score between pairs of iterations\n(Supplementary Fig. 1).\n\nSecond, we sought to ensure that the factors we identified in our\nanalysis (highlighted in red in Figs. 2i and 3g and Extended Data Fig. 5a,\nc) as being different between control and DT conditions (‘Differential\nfactor usage between diphtheria toxin and control’) were robust to\nchanges in parameters, mainly the choice of number of factors. This\ntest ensures that the obtained factors were not identified by chance\nand that they constitute robust signal in the data. For this, we fixed the\nnumber of factors computed above (that is, 20 factors for endothelial,\n25 factors for fibroblasts and 25 factors for myeloid) as the baseline.\nThen, we reran scHPF for a range of number of factors (around the\nchosen value) and computed correlations with the specific factors\nof interest. To compute the correlation to the best matching factor,\nwe used the same strategy of the Hungarian matching algorithm as\ndescribed above. The average correlation over 20 such iterations was\nthen reported (Supplementary Fig. 1).\n\nWe repeated the same computation to assess the robustness of\n\nchosen factors in the bleomycin injury model.\n\nComparison of human and mouse factors. In human samples, scHPF\nwas run with default hyperparameters and ten random initializations in\nthe endothelial, fibroblast and myeloid lineages, using raw UMI counts\nfor genes expressed in at least 1% of cells within the lineage. This left\n12,533 genes in the endothelial lineage, 13,216 genes in the fibroblast\nlineage and 12,253 genes in the myeloid lineage for factor analysis. To\nselect the number of factors for downstream analysis, scHPF was first\nrun with two more factors than the number of PhenoGraph clusters\nwithin the lineage, then subsequently increased nine times by steps\nof one, for a total of nine separate factorizations (that is, k = (17, 18, 19\n… 250). To achieve consistent granularity across lineages, we chose\nthe factorization in which ~90% of the variance in a cells’ expression\n(on average) was explained by the top 7 factors, given by 22 factors\nfor endothelial and fibroblasts lineages and 27 factors for myeloid\nlineage cells.\n\nAfter matrix factorization in human samples, we identified gene\nprograms associated with Treg cell presence in LuAd tumors by calcu-\nlating the Spearman correlation between the log2 average factor cell\nscore and log2 Treg cell proportion of CD45+ cells in each sample. This\ncalculation was also performed using the Treg cell proportion of CD3+\ncells in each sample to ensure consistency; however, the Treg cell propor-\ntion of CD45+ cells are referenced in the primary results (Extended Data\nFig. 9a). We assessed the stability of gene programs using a similar strat-\negy to that used for mouse above, and robustness of factor associations\nto Treg cell presence was assessed by the same correlation calculation\nusing matched factors in a separate run of scHPF factorized using a\ndifferent value of k. The relationship of factors across lineages was\nassessed by the pairwise Spearman correlation of log2 average factor\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fcell scores in each sample from one factor to all other factors. Only\nsamples with enough representative cells were used for correlation\nanalysis in each lineage (>5 cells in endothelial and fibroblast, >20\ncells in myeloid). Sample 16 was removed from all factor correlations\ndue to outlier values driven by high IFN signatures, and sample 17 was\nremoved from endothelial correlations due to outlier values driven\nby low cell numbers.\n\nTo identify conserved gene programs (factors) in endothelial,\nfibroblast and myeloid cells between human and mouse tumors, we\ncompared the gene scores of orthologous genes. First, the genes used\nfor factorization were filtered for orthologs that had a one-to-one\ncorrespondence between species (Ensembl 85 annotations) and were\nexpressed in both species. A gene was assigned to a factor if its gene\nscore was two standard deviations greater than the mean of gene\nscores for all genes in that factor. Then, a Jaccard similarity score was\ncalculated between all mouse and human factors of a given lineage\nby dividing the number of shared assigned genes by the number of\nunique assigned genes in each pair of factors. The z-score of Jaccard\nvalues for all human factors against each mouse factor was used to\nidentify human factors with greater homology to a mouse factor than\nbackground. Typically, a Jaccard similarity score greater than 0.06 in\nthe endothelial lineage and 0.07 in the fibroblast and myeloid lineages\nand would define one (and no more than three) factors in human with\nhomology to a mouse factor.\n\nThe validity of factor mappings across species was assessed by\nexamining the genes shared between conserved factors to ensure they\nbelonged to coherent biological programs (inflammation, angiogen-\nesis, and so on). To find genes with similarly high scores across con-\nserved factors, we normalized the gene score for each gene by the sum\nof its scores across all factors (fraction of total gene score), which also\nenabled comparison across factorizations. This was used to compare\nTreg cell-associated inflammation and hypoxia programs in Fig. 4g; we\ncompared normalized gene scores from the sum of human factors 4 and\n5 to those in mouse factor 15, as these corresponded to the same under-\nlying biological process across species (see below). In Fig. 4g, genes\nwere listed as conserved if they were assigned (as described above) to\nboth the human and mouse factors being compared. VEGF-regulated\ngenes in endothelial cells were identified using gene sets derived by\nDhainaut et al.31 with data from the CytoSig database, which houses\npublic cytokine response datasets for many cell types and treatment\npairs (https://cytosig.ccr.cancer.gov/).\n\nIn certain cases, gene or cell scores for several factors were\nsummed to relate an underlying biological process to similar gene\nexpression programs in mouse (as above) or Treg cell proportion across\nhuman participants. An underlying biological process (for example,\ninflammation) could be split across several factors due to similar but\nnonoverlapping expression programs (for example, cell-type-specific\nsignaling) or very similar expression programs with sample-specific or\ncondition-specific effects. Comparisons including only partial signal in\nthese cases, when only a single factor was compared to another entity,\ncould mask associations to the broader biological program. In Fig. 4f,\nwe summed cell loadings for human endothelial factors 3, 4 and 5 to\nrelate the conserved Treg cell-responsive endothelial expression pro-\ngram to Treg cell proportion across tumor samples. We reasoned that\nthese factors were related to a shared underlying biological process\nbecause they were each individually negatively associated with Treg\ncell proportion across samples to various degrees (Extended Data\nFig. 9c), and their genes aligned with different components of Treg\ncell depletion-induced expression program in mouse tumors: factor\n3, aCap; factor 4/5, inflammation and hypoxia with features of the\nmouse activated VEC (Fig. 4e). Additionally, factors 4 and 5 shared\ninflammation-relevant genes (IL6, CSF3) but with different sample\nspecificities, which indicated that sample-specific effects rather than\nthe underlying biology could have separated this gene program across\ntwo human factors (Extended Data Fig. 9d). Therefore, a summed factor\n\nNature Immunology\n\nscore was found to be more appropriate in capturing certain endothe-\nlial gene program relationships to Treg cell proportion.\n\nExpression heat maps\nOnce we identified the factors of interest in each of the cell types,\nbased on our definition of higher average cell score in DT compared\nto control conditions, we zoomed into the genes that contributed the\nmost to those factors. We were particularly interested in understand-\ning the genes that drive the factor score in a specific subpopulation of\ncells. In our analysis, we sought to focus on specific subtypes with the\nhighest average cell score for the factor. As such, we isolated the cell\ntypes of interest and correlated the factor usage (cell scores) with gene\nexpression. Details of the subsetting are provided in Supplementary\nTable 19. To elaborate, we provide an example: we identified factors 9,\n14 and 22 to be enriched in DT-treated cells compared to control in the\nfibroblast subpopulation in the mouse tumor model. These factors had\nthe highest cell usage scores among the COL14A1 subtype. Therefore,\nto identify genes that are driving these factors and ensure that we focus\non gene programs specific to the COL14A1 subtype, we subset this cell\ntype of interest and correlate factor usage with gene expression. In\ncases where the cell type of interest was small (for example, the inflam-\nmatory capillary subset in endothelial cells in the mouse tumor model),\nwe subsetted the cell type of interest combined with the phenotypically\nmost similar cell type (for example, we grouped the inflammatory cap-\nillary subset with aCap in the mouse tumor model endothelial cells).\nThis ensured we had sufficient cell numbers to compute the correlation\nand allowed us to identify genes specific to the cell type of interest in\ncontrast to its nearest phenotypically similar subtype.\n\nTo this end, we correlated gene expression against the cell scores\nin the isolated set of cells and identified the top 200 most correlated\ngenes as being relevant to that factor for that specific subpopulation.\nTo ensure that the correlation scores were not influenced by any poten-\ntial outliers (cells with deviant cell scores), we compared our results\nagainst correlation computed between the imputed gene expression\nand imputed factor cell scores (using MAGIC57, nPCs = 20, k = 30, ka = 10,\nt = 4). In both scenarios, we obtained highly similar results. The expres-\nsion heat maps (Figs. 2i and 3i and Extended Data Figs. 5a,c, 6b and 7a,d)\ndisplay the result from imputed data.\n\nWe followed the same procedure for the bleomycin injury model\n\ndata.\n\nSpatial transcriptomics\nRead mapping and quantification. We processed Visium ST data with\nthe SpaceRanger pipeline from 10X Genomics (v1.3.1). The mkfastq\nfunction was used to generate FASTQ files from raw base calls and the\ncount function was used in combination with a matched brightfield\nH&E-stained image to align to a modified mm10 genome, perform\ntissue detection and count UMIs for each spot. The modified genome\nconsisted of Ensembl 100 annotations with an added transcript to\ndetect DTR-GFP expressed from the Foxp3 promoter (sDTR-eGFP).\nUMI counts were summed by gene symbol and sDTR-eGFP reads were\nsummed together with Foxp3. All analyses of differential cell-type abun-\ndance or gene expression were performed in the first serial section of\neach biological sample to preserve the independence of observations.\nGene expression counts were log normalized using SCTransform67 with\nSeurat (v4.1.1)68 to compare between spots. Spots with fewer than 1,000\nUMIs were excluded from analysis.\n\nDeconvolution of Visium spots to cell-type RNA fractions. Visium\ncaptures transcripts from sectioned tissue placed over 55-μm-diameter\nspots, such that each spot sums gene expression from multiple cells. We\nused the BayesPrism algorithm25,26 to deconvolve cell types present in\neach spot and thereby improve the effective resolution of the technol-\nogy. As input, BayesPrism accepts a spot-by-gene count matrix and a\nscRNA-seq reference dataset labeled by cell type; it utilizes a Bayesian\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fapproach to jointly model cell-type fractions and cell-type-specific\ngene expression within each Visium spot.\n\nIn our ST analysis, we used two separate scRNA-seq references\nfor deconvolution—one containing the small number of tumor cells\n(N = 239) captured in our study (‘non-merged reference’) and another\ncontaining cells from a separate study that sampled more tumor\ncells (N = 18,083) from specific tumor sub-states (‘merged reference’,\ndetailed below). The merged reference was used to assess the presence\nof granular transcriptional states within tumors, while the non-merged\nreference was used to study accessory cell populations without the\ninfluence of batch effects (data from two separate studies) or con-\nfounding during deconvolution (limited resolution between normal\nepithelial and certain tumor states, that is, AT2 versus AT2-like tumors).\nThe non-merged reference includes scRNA-seq data solely consisting\nof cells from identical experimental conditions to the ST data (KP\ntumor-bearing lungs treated with PBS or DT; data from Fig. 2) and\nwas used for all analyses in Fig. 4 and Extended Data Fig. 7c,e,f. Cell\nfraction estimates from the non-merged reference were used to distin-\nguish tumor spots from normal spots because of the better-matched\nexperimental characteristics, the capture of tumor cells from the Treg\ncell-depleted state and the lower chance of similarity to normal EC\ntypes by using all tumor cells as a single reference population. The\ncell-type fraction estimates of tumor sub-states from the merged\nreference were used for analysis in Fig. 5 and Extended Data Fig. 8 only\nin tumor spots defined using the non-merged reference. Additional\ndetails of scRNA-seq reference construction and applications of the\ncell-type fraction estimates are mentioned below.\n\nSelection of cell types and marker genes. The accuracy and reli-\nability of cell fraction estimates depends on the presence of features\nin Visium data that are specific to labeled populations (highly specific\ncell-type markers give better deconvolution), the transcriptional dis-\ntance between populations (better separated populations give better\ndeconvolution) and how closely matched the scRNA-seq reference is\nwith populations profiled in situ by ST. We thus optimized both gene\nselection and cell-type label granularity in our scRNA-seq reference\nand leveraged the ability of BayesPrism to encode separate cell states\nwithin a population to better match the reference in specific conditions\n(that is, control versus Treg cell depleted).\n\nFeature selection before deconvolution can improve the\nsignal-to-noise ratio by removing genes that are irrelevant to cell type but\nbehave similarly to relevant genes, and it can also mitigate the influence\nof genes that change due to batch effects. We therefore chose to focus on\ncell-type marker genes in our deconvolution, which is a recommended\noption in BayesPrism. Marker genes were computed by conducting\npairwise t-tests across cell types (findMarker function in SCRAN) using\nlog-normalized data. We defined marker genes by a minimum P value of\n0.05 and minimum log fold-change value of 0.25 across all comparisons.\nGenes with fewer than ten counts across all Visium sections, or those\ndetected in fewer than five cells in scRNA-seq data, were removed in\naddition to ribosomal genes, mitochondrial genes and genes associated\nwith the cell cycle (https://github.com/dpeerlab/spectra/).\n\nWe merged highly similar cell types to avoid confounding deconvo-\nlution. To ensure adequate resolution between cell types, we computed\nmarker genes as described above, starting at the most granular level of\nannotation and iteratively merging cell populations with their closest\nneighbor (by transcriptional distance), until each cell population had\nat least 30 marker genes. This included collapsing Artery/Vein with\ngCap cells (labeled as gCap); CD8+ T cells, effector T cells, exhausted\nCD8+ T cells, MAIT, gdT, TH2, naïve T cell, activated T cell, Treg and ILC2\npopulations (T cell/ILC2); B and plasma cells (B cells); monocyte and\nCsf3r+ monocytes (monocyte); cDC1 and cDC2 (cDC); and Csf3r+ neu-\ntrophil, Ccl3+ neutrophil, and Siglecf+ neutrophil (neutrophil). We\nfurther merged the inflammatory capillary population with aCap cells\n(aCap), and Arg1+ with C1q+ macrophage populations (macrophages),\n\nNature Immunology\n\nas these are arguably specialized cell states of the same overarching cell\ntype. Cycling T cells were also removed to prevent misassignment to\ntissue regions with increased expression of cell cycle-related genes. The\nresulting filtered scRNA-seq reference comprised 4,219 marker genes\nand 23,178 cells labeled as 26 cell populations (see Extended Data Fig.\n7a for full list of cell populations included).\n\nWhile our feature selection strategy ensured adequate resolution\nbetween cell types, transcriptional heterogeneity within cell types can\nalso influence deconvolution. BayesPrism initially performs inference\nat the cell-state level, which can account for condition-specific hetero-\ngeneity in transcriptional states within cell types during deconvolution.\nCell states can be captured by the algorithm through cell-type-specific\nexpression estimates but can also be included as labels in the reference\ndata. We observed substantial transcriptional shifts in accessory cell\npopulations between control and Treg cell-depleted conditions by\nscRNA-seq (Fig. 2 and Extended Data Figs. 4 and 5), and thus labeled\ncells from these accessory populations to help capture heterogeneity\nwithin cell types. Control and Treg cell-depleted states were assigned\nfor aCap, gCap, LECs, Col13a1+ and Col14a1+ fibroblasts, pericytes,\nmyofibroblasts, AT1, AT2, cDC, macrophage, alveolar macrophage,\nneutrophil and monocyte populations in the scRNA-seq reference.\nBayesPrism sums cell-state fractions at the cell-type level before the\nfinal update step and downstream analysis.\n\nFollowing cell-state definition, BayesPrism was run jointly on serial\nsections, to allow sharing of information across more spots during the\nfinal update step and filtering out of genes whose expression fraction\n(reads/total reads) was greater than 0.01 in 10% of Visium spots. For the\nrobustness and reproducibility analysis, each section was deconvolved\nindependently.\n\nKP tumor cells are known to adopt a range of recurrent cell states\nas they progress27–30. To assess tumor transcriptional states and their\nrelation to Treg cell depletion, we performed a second deconvolution\nusing BayesPrism across Visium spots with a scRNA-seq reference\ncontaining more granular tumor-state labels. Given the limited number\nof tumor cells in our reference (239 cells), we decided to incorporate\nscRNA-seq data from Yang et al.28, which contains ~50,000 KP tumor\ncells (referred to as KP-Tracer data), to more accurately assign general\nstates within tumor regions. A key advantage of BayesPrism is that it\ncan incorporate single-cell data from multiple sources, which do not\nneed to be matched with our data. The algorithm uses cell types in the\nscRNA-seq reference as a prior for possible cell states in the Visium data,\nwhile disregarding cell-type fractions in the reference.\n\nTo minimize computational burden and sample-specific biases,\nwe processed the KP-Tracer data by removing mutation-specific and\nmesenchymal cell states (these were largely sample-specific), and\ndownsampling the remaining tumor states to a maximum of 2,000 cells\n(for balanced sampling), leaving 18,083 cells. The original tumor-state\nlabels of EMT-1, EMT-2 and pre-EMT were combined (labeled as EMT),\nas were early gastric, late gastric and gastric-like populations (gastric)\nto limit deconvolution to the most representative overarching tumor\nstates. These cells were combined with our accessory cell data in the\nsame count matrix, with tumor cells removed. Marker gene selec-\ntion and gene filtering were performed as above but adding 24 genes\nupregulated in AT2 cells relative to the AT2-like tumor state (t-test on\nlog-normalized scRNA-seq data with adjusted P value < 0.001, log fold\nchange > 1, expressed in 15% more cells relative to AT2-like) to better\ndiscriminate tumor from normal states. The final merged reference\ncontained 40,787 cells and 4,546 genes after cell and feature selection,\nand we ran BayesPrism deconvolution with it using identical settings\nto the non-merged reference.\n\nVisium data are very noisy. To discriminate robust evidence for\ncell type, we only included cell-type fractions above background at\nparticular spots, using the same mixed-model strategy as the compute.\nbackground function from SpaceFold26, with modifications detailed\nbelow. Specifically, for each cell type in each tissue section, a gamma\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fmixture model with two components was fit for cell-type fraction\n(gammamixEM from mixtools69), and a Gaussian mixture model with\ntwo components was fit for the summed deconvolved gene expression\nvalues (Mclust from Mclust70) across all spots. Mixture model distribu-\ntions were checked for agreement with data structure by histogram and\noverlay of the fitted distribution. After determining parameters for the\nmixture components, we identified spots with >70% posterior prob-\nability of being assigned to the mixture component having the higher\nmean and used the minimum value of these as a threshold for calling a\ncell type ‘present’. Cell-type thresholds below 0.001 were reset to 0.001,\nand summed deconvolved gene expression thresholds below 50 were\nreset to 50. To enable comparison across tissue sections and prevent\nerroneous cutoffs due to tissue-specific composition, the median of\ncell fraction and summed deconvolved gene expression cutoffs across\nall eight tissue sections was used for each cell type. These values were\nsubsequently refined with the guidance of H&E staining (see below for\ndetails). Illustrations of spot binarization for the presence of specific\ncell types are shown in Fig. 4a and Extended Data Fig. 7d,e.\n\nAssessing robustness and accuracy. We assessed the robustness\nand accuracy of cell-type RNA fraction estimates before proceeding\nfurther with analysis downstream of our deconvolution. We performed\nbootstrap analysis to determine robustness to the sparse capture of\nVisium. Specifically, we ran BayesPrism on one tissue section with\neach spot randomly downsampled to 90% of its reads, repeated this\n20 times, and calculated the Spearman correlation of cell-type fraction\nestimates between the original and each downsampled deconvolu-\ntion. Cell fraction estimates across spots were highly consistent, with\nSpearman correlations ≥ 0.87 for all trials (Extended Data Fig. 7a). We\nnext compared average cell-type fraction between individually decon-\nvolved serial sections across all samples, validating the expectation\nthat cell-type fractions captured by serial sections are highly similar\n(Spearman R = 0.99; Extended Data Fig. 7b). To ensure consistency\nbetween our two deconvolution approaches, we compared cell-type\nfractions of non-tumor accessory cells with and without additional\ntumor states from the KP-Tracer study in our scRNA-seq reference.\nAverage log(cell-type RNA fraction) values from each tissue section\nwere highly correlated (Spearman R = 0.97), suggesting that decon-\nvolved accessory populations were generally not influenced by tumor\nRNA, and that the inclusion of tumor cells from a separate study did\nnot impact accessory cell deconvolution (Extended Data Fig. 7c). One\nexception was in resolving AT1-like and AT2-like epithelial states, which\nare highly similar to several of the added tumor states; the added states\nlikely improved their resolution in tumor regions, but not in non-tumor\nregions, due to transcriptional similarity with normal epithelial states.\nCell-type assignment was cross-referenced with the underlying\ntissue histology from matched H&E-stained brightfield images to con-\nfirm accurate positioning of cell types where possible (Extended Data\nFig. 7d,e). For example, spots deemed to possess different capillary\ntypes, pericytes and alveolar macrophages were consistent with the\nliterature and anatomical features (Extended Data Fig. 7d). gCap and\nartery/vein cells were localized around blood vessels and alveoli, with\nsome penetration into tumor areas, whereas aCap cells were mainly\ndistributed over alveoli and surrounding tumor areas, consistent with\ntheir propensity to surround areas of injury48. Pericytes were localized\naround blood vessels and bronchi, consistent with published annota-\ntions60, and alveolar macrophages were concentrated in areas sur-\nrounding tumor regions, as previously shown45. LECs, DCs and B cells\nare all expected in areas containing lymphoid aggregates emanating\nfrom a lymphatic vessel and were indeed detected in these regions by\nour ST analysis (Extended Data Fig. 7e). Moreover, regions represent-\ning part of an IC signaling niche were found to have higher neutrophil\ncell-type fraction (Fig. 4f), which was readily apparent in the aligned\ntissue section due to the unique appearance of neutrophils in H&E\nstaining (Fig. 5f).\n\nNature Immunology\n\nUpon assessing marker gene expression and inspecting histology,\nwe noted that several cell types including AT2, gCap, MSCs, monocytes\nand LECs had more modes in the distribution of their cell-type fractions\nand summed deconvolved gene expression values across spots, likely\ndue to regional variation in cell-type composition and read density. To\naccount for this variation, we reset the presence/absence thresholds for\nthese cell types as above but using mixture models with three mixture\ncomponents instead of two. As a result, the minimum value from spots\nassigned to the mixture component with the second highest mean (one\nabove background mixture component) with >70 posterior probability\nwas used as a threshold. The median cell-type fraction and summed\ndeconvolved gene expression threshold values of the three-component\nmixture models across all tissue sections was applied to all spots (as\nfor two-component models above).\n\nAnalysis of gene program usage across conditions. To assess the\ndifferential use of gene programs between control and Treg cell-depleted\nconditions identified by factor analysis in scRNA-seq data, we used\nthe AddModuleScore function in Seurat to compute the relative\nlog-normalized expression of each factor’s genes relative to a ran-\ndom set of background genes with similar average expression in the\ntissue. Specifically, all genes were split into 24 expression bins and 100\ncontrol features were randomly selected for each feature in the input\ngene program from a corresponding bin. The average log-normalized\nexpression of control features was then subtracted from the average\nlog-normalized expression of the features of interest to derive a mod-\nule score. Module scores were computed across spots from all four\nsamples at the same time. A t-test was performed to compare gene\nprogram module scores in control and Treg cell-depleted conditions\nfor gene programs of interest, and P values were adjusted by Benja-\nmini–Hochberg correction. To measure the difference in relevant\ncellular contexts, comparisons were restricted to spots with cell-type\nfractions above background for cell types in which the gene program\nof interest was found to be differential by scRNA-seq, creating a table\ncomparing cell type by gene program of interest across conditions\n(Fig. 4b). The visualization of specific module scores was performed\nin Figs. 4c,d and 5g.\n\nDefinition of signaling niches. Certain gene programs that increased\ntheir abundance in both our scRNA-seq and ST analysis following\nTreg cell depletion shared many genes across endothelial, fibroblast\nand myeloid lineages. This included factors that contained many\nIFN-stimulated genes (IFN factors) and factors that contained genes\nrelated to IC and hypoxia signaling (IC factors). To determine shared\ngenes between IFN and IC factors, genes were assigned to each rel-\nevant factor from the mouse scRNA-seq in the same way as detailed\nin ‘Comparison of human and mouse factors’ and the intersection\nof genes across all three lineages for IFN or IC factors was taken. The\nIFN factors were defined as fibroblast factor 9, endothelial factor 19\nand myeloid factor 17. IC factors were defined as fibroblast factor\n22, endothelial factor 15 and myeloid factor 21. The module score\nof shared genes for IFN (N = 103 genes) or IC (N = 18 genes)-related\ngene programs (See Supplementary Table 12 for gene lists) was then\nused to define ‘signaling niches’ or Visium spots where a common\nsignaling pathway may drive downstream gene expression in several\ncolocalized cell types.\n\nTo assign spots to a signaling niche, we took advantage of the fact\nthat most spots across all tissue sections did not show signal for IFN or\nIC gene programs. Therefore, we modeled the background rate of these\ngene programs by fitting their module scores plus a pseudocount of\none to a gamma distribution using maximum likelihood estimation\n(fitdistr from MASS package71) across all spots on the four biologically\nindependent sections being analyzed. Alignment with the gamma dis-\ntribution was checked by a histogram of the gene scores and density\noverlay of the fit distribution. The module score corresponding to an\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fupper tail probability of 0.01 in the fit distributions was then used as\na threshold above which spots were assigned to that signaling niche.\nAssignment of spots to the IFN or IC signaling niche was not mutually\nexclusive and gave 397 spots assigned to IC niches, 330 spots assigned\nto IFN niches and 21 spots assigned to both. An illustration of spot\nassignment to signaling niches is shown in Fig. 4e.\n\nCell-type enrichment in signaling niches. To assess the presence\nof different cell types within signaling niches relative to background\ncell-type fractions across the tissue (Fig. 4f), we took a random sample\nof 100 spots across all tissue sections and averaged the fraction of each\ncell type, then repeated for 10,000 iterations to form an empirical prob-\nability distribution of mean cell-type fractions of randomly selected\nspots. The empirical P value was calculated as the fraction of iterations\nin our empirical distribution with an average cell-type fraction above\nthe average for all spots in a given signaling niche (IFN or IC). Empirical\nP values were adjusted by Benjamini–Hochberg correction to account\nfor multiple hypothesis testing. To measure the magnitude of enrich-\nment for each cell type, the log2 average cell-type fractions from the\ntotal empirical distribution were subtracted from average cell-type\nfraction values from either signaling niche.\n\nDefinition of tumor-state regions. To classify spots within tumor\nlesions into areas of consistent transcriptional phenotypic state (‘tumor\nlesion areas’), we first selected spots with tumor RNA above back-\nground (detailed above) and used cell-type fractions from the decon-\nvolution with the merged reference. To visualize the co-occurrence\nof tumor states, we z-scored fractions of tumor states in tumor spots\nand hierarchically clustered the spots into seven groups (R cutree with\nk = 7) using Pearson correlation distance and average agglomeration\n(Extended Data Fig. 8a). This analysis revealed that tumor spots were\ntypically dominated by a single tumor state. When plotted in their tissue\ncontext, we found that they often aggregated spatially (Fig. 5a), pro-\nviding further support for the presence of consistent transcriptional\nphenotypes within lesional areas.\n\nEach of the seven clusters was then labeled based on the tumor\nstate with the highest cell-type fraction in the cluster. The validity of\ncluster labels was assessed by the expression of tumor-state marker\ngenes defined by previous studies28. We found clearly higher expres-\nsion of marker genes in their corresponding cluster relative to other\ntumor clusters (Extended Data Fig. 8b). The classification of tumor\nspots in their tissue location is shown in Fig. 5b and Extended Data\nFig. 8c. In H&E staining, the location of different tumor states often cor-\nresponded to a noticeable change in histology, further supporting our\nclassifications. For instance, neighboring gastric and high-plasticity\nregions also exhibited a more differentiated morphology in the gastric\ntumor area and less structure in the high-plasticity area (Fig. 5f,g).\nWhile our strategy increased the resolution of tumor transcriptional\nstates in our deconvolution, there may be additional tumor states\nin situ that are not contained within our merged scRNA-seq refer-\nence due to heterogeneity of the model. In these cases, the cell-type\nfractions from missing tumor states would be assigned to the closest\ntranscriptional neighbor.\n\nTumor lesion areas were defined by separating connected com-\nponents (contiguous spots in tissue) of the same tumor-state cluster.\nAT1-like and AT2-like tumor-state clusters were merged for tumor lesion\narea definition because of the higher degree of mixing between these\ntumor states observed previously29 and in our analysis (Extended Data\nFig. 8a). Only lesion areas greater than six spots were kept for subse-\nquent analysis, to avoid micrometastases or regions dominated by\ntumor edges due to sectioning. This resulted in 47 and 38 tumor lesion\nareas in control and Treg cell-depleted tissue sections, respectively.\nTumor lesion areas were deemed to have an immune response in Treg\ncell-depleted tissue sections if >10% of constituent spots were part of\nan IC or IFN signaling niche (Fig. 5e).\n\nNature Immunology\n\nDifferential expression of tumor areas\nWe were interested in detecting differential gene expression between\ntumor lesion areas. We first collected all spots in tumor lesion areas (1)\nexhibiting an immune response and (2) exhibiting no response after\nTreg cell depletion (defined in the paragraph above), then performed a\nWilcoxon rank-sum test between the two groups of spots. SCTransform\nlog-normalized values were used as input and only genes detected\nin at least 10% of spots in either condition and with an average log\nfold-change value > 0.25 between conditions were tested. We detected\n259 genes that were differentially higher in responding tumors and\n142 genes that were higher in non-responding tumors at Benjamini–\nHochberg-adjusted P value < 0.01 and log fold-change > 0.5 (Fig. 5d and\nSupplementary Table 15). The SCTransform log-normalized expression\nlevels of specific genes associated with non-responsiveness to Treg cell\ndepletion are shown in Fig. 5e.\n\nStatistics\nFor all mouse experiments, statistical analyses were performed using\nGraphPad Prism 9 and are detailed in the figure legends. Mice were allo-\ncated randomly to experimental groups. No statistical methods were\nused to predetermine sample sizes but our sample sizes are similar to\nthose reported in previous publications5,13. Data collection and analysis\nwere not performed blind to the conditions of the experiments.\n\nStatistical tests used for analysis of RNA-seq and ST data are\ndescribed. For scRNA-seq and ST, count data were assumed to be distrib-\nuted according to a negative binomial distribution and log-transformed\ndata according to a normal distribution. In other analyses, data dis-\ntribution was assumed to be normal but this was not formally tested.\nscRNA-seq data analysis was performed using custom code rely-\ning primarily on Python v3.8.11 using Scanpy v1.8.1 package for basic\npre-processing and analysis. Visualization of the data was done using\nMulticoreTSNE v0.1 implementation of t-SNE in Python, and clustering\nwas done using PhenoGraph v1.5.7 package in Python. Factor analysis\nwas done using scHPF v0.5.0 implementation in Python v3.7.11. Dif-\nferential abundance testing between scRNA-seq conditions was per-\nformed using Milo v1.3.4. Identification of factors (Hungarian matching\nalgorithm) was implemented using the linear_sum_assignment module\nin optimize submodule of SciPy (v1.7.1) in Python (v3.8). For human\nfactor analysis, Spearman correlation coefficients and P values were\ncalculated in R using ggpubr (0.4.0) and results were visualized using\nggplot2 v3.3.5.\n\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\n\nData availability\nRaw and processed bulk, scRNA-seq and Visium data from mouse\nare available from the Gene Expression Omnibus under super series\naccession GSE202159. Human tumor scRNA-seq data are available at\nthe Human Tumor Atlas Network (HTAN) data coordinating center web\nplatform (https://humantumoratlas.org/). Source data are provided\nwith this paper.\n\nCode availability\nNo new algorithms were developed for this paper. All analy-\nsis code is available at https://github.com/dpeerlab/Treg_\ndepletion_reproducibility/.\n\nReferences\n51. Levine, A. G. et al. Stability and function of regulatory\n\nT cells expressing the transcription factor T-bet. Nature 546,\n421–425 (2017).\n\n52. Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner.\n\nBioinformatics 29, 15–21 (2013).\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\f53. Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold\nchange and dispersion for RNA-seq data with DESeq2. Genome\nBiol. 15, 550 (2014).\n\n54. Chan, J. M. et al. Signatures of plasticity, metastasis, and\n\nimmunosuppression in an atlas of human small cell lung cancer.\nCancer Cell 39, 1479–1496 (2021).\n\n55. Azizi, E. et al. Single-cell map of diverse immune phenotypes in\nthe breast tumor microenvironment. Cell 174, 1293–1308 (2018).\n56. Buechler, M. B. et al. Cross-tissue organization of the fibroblast\n\nlineage. Nature 593, 575–579 (2021).\n\n57. Gillich, A. et al. Capillary cell-type specialization in the alveolus.\n\nNature 586, 785–789 (2020).\n\n58. Zepp, J. A. et al. Distinct mesenchymal lineages and niches\n\npromote epithelial self-renewal and myofibrogenesis in the lung.\nCell 170, 1134–1148 (2017).\n\n59. Zilionis, R. et al. Single-cell transcriptomics of human and mouse\nlung cancers reveals conserved myeloid populations across\nindividuals and species. Immunity 50, 1317–1334 (2019).\n60. Travaglini, K. J. et al. A molecular cell atlas of the human lung\nfrom single-cell RNA sequencing. Nature 587, 619–625 (2020).\n\n61. Fleming, S. J. et al. Unsupervised removal of systematic\n\nbackground noise from droplet-based single-cell experiments\nusing CellBender. Preprint at bioRxiv https://doi.org/10.1101/\n791699 (2019).\n\n62. Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: computational\n\nidentification of cell doublets in single-cell transcriptomic data.\nCell Syst. 8, 281–291 (2019).\n\n63. Xi, N. M. & Li, J. J. Benchmarking computational doublet-detection\n\nmethods for single-cell RNA sequencing data. Cell Syst. 12,\n176–194 (2021).\n\n64. Finak, G. et al. MAST: a flexible statistical framework for assessing\ntranscriptional changes and characterizing heterogeneity in\nsingle-cell RNA-sequencing data. Genome Biol. 16, 278 (2015).\n65. Korotkevich, G. et al. Fast gene-set enrichment analysis. Preprint\n\nat bioRxiv https://doi.org/10.1101/060012 (2021).\n\n66. Munkres, J. Algorithms for the assignment and transportation\n\nproblems. J. Soc. Ind. Appl. Math. 5, 32–38 (1957).\n67. Hafemeister, C. & Satija, R. Normalization and variance\n\nstabilization of single-cell RNA-seq data using regularized\nnegative binomial regression. Genome Biol. 20, 296 (2019).\n68. Hao, Y. et al. Integrated analysis of multimodal single-cell data.\n\nCell 184, 3573–3587 (2021).\n\n69. Benaglia, T., Chauveau, D., Hunter, D. R. & Young, D. S. mixtools:\n\nCycle for Survival and the Marie-Josée and Henry R. Kravis Center\nfor sequencing. This work was supported by the Irvington Cancer\nResearch Institute Postdoctoral Fellowship (to A.G.), NCI Cancer\nCenter Support grant P30 CA08748, NCI grant U54 CA209975 (to\nA.Y.R., D.P. and C.L.), NIAID grant R01AI034206 (to A.Y.R.), NCI Human\nTumor Atlas Network U2C CA233284 (to D.P.), Immunology T32\ntraining grant 5T32CA009149-44 (to S.A.R), Wrobel Family Foundation\n(to S.A.R) Alan and Sandra Gerry Metastasis and Tumor Ecosystems\nCenter at MSKCC (to R.S. and D.P.), NCI grant R35CA263816 (to\nC.M.R.), the Robert J. and Helen C. Kleberg Foundation (to C.M.R. and\nD.P.) and Ludwig Center for Cancer Immunotherapy at MSKCC. A.Y.R.\nand D.P. are HHMI investigators.\n\nAuthor contributions\nA.G., D.P. and A.Y.R. conceived the study and designed the\nexperiments. A.G., S.A.R., R.S., D.P. and A.Y.R. interpreted the data and\nwrote the manuscript. A.G. and J.A.G. performed the experiments.\nS.R., I.K.V., E.S.A., B.S.D. and Z.-M.W. assisted with cell isolation and\nin vivo tumor experiments. A.G., M.S. and S.D. performed bulk RNA-seq\nanalysis. O.C., T.X. and L.M. prepared scRNA-seq samples. S.A.R., R.S.\nand H.G. performed analysis of the scRNA-seq data. W.H. and A.M.\nassisted with imaging and analysis. S.A.R. and T.C. performed analysis\nof Visium data. G.R. performed pathological analysis. A.Q.-V., P.M.,\nJ.E., E.D.S. and C.M.R. performed scRNA-seq of human LuAd samples.\nD.P. and A.Y.R. supervised the study. Correspondence and requests for\nmaterials should be addressed to the corresponding authors.\n\nCompeting interests\nA.Y.R. is a member of SAB, and has equity in Surface Oncology, RAPT\nTherapeutics, Sonoma Biotherapeutics, Santa Ana Bio and Vedanta\nBiosciences and is an SAB member of BioInvent and Amgen; A.Y.R.\nholds a therapeutic Treg cell depletion IP licensed to Takeda. C.M.R.\nhas consulted regarding oncology drug development with AbbVie,\nAmgen, Astra Zeneca, D2G, Daiichi Sankyo, Epizyme, Genentech/\nRoche, Ipsen, Jazz, Kowa, and Merck, and is a member of the SAB of\nAuron, Bridge Medicines, Earli, and Harpoon Therapeutics. D.P is a\nmember of the SAB and has equity in Insitro. The remaining authors\ndeclare no competing interests.\n\nAdditional information\nExtended data is available for this paper at\nhttps://doi.org/10.1038/s41590-023-01504-2.\n\nan R package for analyzing mixture models. J. Stat. Softw. 32,\n1–29 (2009).\n\n70. Scrucca, L., Fop, M., Brendan, M. T. & Raftery, A. E. mclust 5:\n\nSupplementary information The online version\ncontains supplementary material available at\nhttps://doi.org/10.1038/s41590-023-01504-2.\n\nclustering, classification and density estimation using Gaussian\nfinite mixture models. R J. 8, 289–317 (2016).\n\n71. Venables, W. N. & Ripley, B. D. Modern Applied Statistics with S\n(Springer New York, 2002). https://doi.org/10.1007/978-0-387-\n21706-2\n\nAcknowledgements\nWe thank members of the A.Y.R. and D.P. laboratories for discussions,\nT. Tammela for the KP cell line, MSKCC Single Cell Analytics Innovation\nLab and Integrated Genomics Operation Core facility funded by\nthe NCI Cancer Center Support Grant (CCSG, P30 CA08748),\n\nCorrespondence and requests for materials should be addressed to\nDana Pe’er or Alexander Y. Rudensky.\n\nPeer review information Nature Immunology thanks Shannon Turley\nand the other, anonymous, reviewer(s) for their contribution to the\npeer review of this work. Primary Handling Editor: L. A. Dempsey, in\ncollaboration with the Nature Immunology team.\n\nReprints and permissions information is available at\nwww.nature.com/reprints.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 1 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 1 | Short-term Treg depletion in lung KP adenocarcinoma\nbearing mice. (a, b, d) Representative gating of normal and tumor (EpCAM)\ncells, (D) TCRβ+CD4+ (CD4), TCRβ+CD8+ (CD8) T cells, myeloid cells (MHCII+GR1-\nCD11b+), neutrophils (Neu) (MHCII-GR1+CD11b+), vascular endothelial cells\n(VEC) (CD45-CD31+GP38−), fibroblasts (Fib) (CD45-CD31-GP38+), and lymphatic\nendothelial cells (LEC) (CD45−CD31+GP38+) (A, B) in KP-lungs in diphtheria toxin\n(DT, N = 3) and PBS (Ctrl, N = 4) mice and (D) in tumor-free lungs (DT, N = 4), PBS\n(Ctrl, N = 4). (c, e) Cell frequencies in KP tumors from (A, B, D). Data represent\nmean ± SEM of one of two independent experiments. (c) Two-way ANOVA\nalpha = 0.05, Šídák’s multiple comparisons CD4 t = 2.254, df = 35 ns P = 0.1953,\nCD8 t = 1.235, df = 35 ns P = 0.8320, MHCII+/Gr1- CD11b+ (MAC/DC) t = 0.5098\ndf = 35 ns P = 0.9987, MHCII-/Gr1+ CD11b+ (Neu) t = 2.985, df = 35, * P = 0.0355,\nVEC, t = 0.2030, df = 35 ns P>0.9999, Fib t = 0.09821, df = 35 ns P>0.9999, Lec t =\n0.08549, df = 35 ns P>0.9999. (e) Two-way ANOVA, Alpha = 0.05, Tukey’s multiple\n\ncomparisons EC, t = 1.056. df = 12, ns P = 0.5261 Tumor t = 1.217, df = 12, ns P =\n0.4332. (f) Fold change (FC) deferentially expressed genes (DEG). (g) k-means\nclustering of FC DEG between DT and Ctrl. Columns - log2 FC (DT/Ctrl) for cells,\neach row is a gene. Select genes are labeled. (h) Z-score-normalized counts for\nselected genes in (G). (i) Cell frequencies in tumor-free DT and Ctrl lungs. Two-\nway ANOVA, alpha = 0.05, Šídák’s multiple comparisons. Epcam t = 0.3437, df =\n37, ns P>0.9999, CD4 t = 1.413, df = 32 ns 0.769, CD8 t = 1.434, df = 32 ns P = 0.7550,\nMHCII+/Gr1- CD11b+ (MAC/DC) t = 0.8971, df = 32, ns P = 0.9771, MHCII-/Gr1+\nCD11b+ (Neu) t = 3.664, df = 32 ** P = 0.0071, VEC t = 0.3854, df = 32 ns P>0.9999,\nFib t = 0.008845, df = 32 ns P>0.9999, LEC t = 0.01251, df = 32 ns P>0.9999. Data\nrepresent mean ± SEM of one of two independent experiments N = DT-3, PBS-3.\n(j) DEG Numbers. (k) FC DEG in cells isolated from tumor-free lungs of DT vs Ctrl\nmice. (F, J, K) DEG - differentially expressed genes (p<0.05). Red-upregulated,\nblue-downregulated.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 2 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 2 | scRNA-seq analysis and annotation of major TME cell\ntypes in lung KP adenocarcinomas. (a) Sorting strategy. CD45+ and CD45− cells\nwere sorted from lungs of PBS (Ctrl) and DT treated (48 hr) mice (3 mice per\ngroup) harboring KP lung tumors. (b) t-SNE plots embedding (27,606 cells)\nrepresenting distribution of all the cells isolated in (A), colored by PhenoGraph\nclusters (k = 30) (left), or sample (right) (related to Fig. 2a, which shows major\ncell lineages). (c) Heatmap showing the average expression of cell type specific\n\nmarkers in each cluster. The rows are genes and columns are clusters. Shown\nexpression is row normalized between 0–1 and genes are grouped to indicate\nthe subtype they typically are associated with. All the genes used for annotation\nare shown. (d) t-SNE embedding (same as B) colored by lineages inferred using\nthe average expression of each cluster shown in the heatmap in (C). (e) t-SNE\nembedding reflecting experimental conditions (Ctrl: PBS, gray; DT: diphtheria\ntoxin, red).\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 3 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 3 | Heatmaps of Treg depletion-induced gene\nexpression changes in fibroblasts, endothelial and myeloid cells in lung\nKP adenocarcinomas. (a) Heatmap showing the average expression of known\nendothelial markers in each endothelial cell cluster. Rows indicate cluster and\ncolumns indicate genes. The heatmap is column normalized between 0–1 and\nthe genes are grouped to indicate the subtype they typically are associated\n\nwith (top). t-SNE embeddings (2815 cells) (bottom) representing distribution\nof endothelial cells color coded by their cluster identity inferred using the\ngene expression pattern for each cluster shown in the heatmap (left) and cell\ntype annotation derived from the heatmap above (right). All genes used for\nannotation are shown. (b) Same as (A) for fibroblasts. (c) Same as (A) for myeloid\ncells.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 4 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 4 | Neighborhood analysis of Treg depletion-induced\ngene expression changes in endothelial cells, fibroblast and myeloid cells\nin lung KP adenocarcinomas. (a) t-SNE embedding of fibroblasts (3,791 cells)\n(top) color coded by cell subtype (left) or experimental condition (right). A\ndensity plot of the distribution of fibroblasts between conditions (bottom).\nCtrl – PBS, gray; DT - diphtheria toxin, red. (b) Graph of neighborhoods of\nfibroblast cells computed using MiloR and embedded on t-SNE (top). Each dot\nrepresents a cellular neighborhood and is color coded by the FDR corrected\np-value (alpha = 1) quantifying the significance of enrichment of DT cells\n\ncompared to control in each neighborhood. The size of the dot represents the\nnumber of cells in the neighborhood. (bottom). Swarm plot depicting the log-\nfold change in differential abundance of DT treated cells against control cells in\neach neighborhood across different fibroblast cell types. Each dot represents\na neighborhood and is color coded by the FDR corrected p-value (alpha = 1)\nquantifying the significance of enrichment of DT cells compared to control in\neach neighborhood. A neighborhood is classified as a cell type if it comprises at\nleast 80% of cells in the neighborhood, or called ‘mixed’ otherwise. (c) Same as\n(A) for myeloid cells. (d) Same as (B) for myeloid cells.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 5 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 5 | Factor analysis of Treg-dependent gene expression\nby fibroblasts and myeloid cells in lung KP adenocarcinomas. (a) Heatmap\nshowing factor cell score across experimental conditions averaged over each\nfibroblast cluster in each experimental condition. The rows are factors and\ncolumns are clusters for each experimental condition. The clusters are grouped\nbased on the cell type they are associated with. The heatmap is row normalized\nfrom 0–1. (b) Heatmaps showing the top 200 genes that correlate the most with\n\nimputed cell scores of the indicated factors (see Methods) for fibroblast subsets.\nEach column is a cell; cells are ordered based on their factor score in ascending\norder from left to right indicated by the green bar. The experimental condition\nfor each cell is indicated by the grey for PBS (Ctrl) and red for diphtheria toxin-\ntreated conditions (DT) bar. Select examples of genes of interest are noted. (c)\nHeatmap as in A for myeloid cells. (d) Heatmaps as in B for myeloid cells.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 6 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 6 | Treg-dependent gene expression changes in\nendothelial and myeloid cells in bleomycin-induced lung inflammation vs\nKP adenocarcinomas. (a) Schematic of the experimental design. (b) Numbers\nof Treg and effector T cells in Ctrl (PBS) and DT (Diphtheria Toxin) treated lungs,\nat day 21 after bleomycin administration. (Left) Two-way ANOVA, Alpha = 0.05,\nfollowed by Tukey’s multiple comparisons test was performed. PBS Ctrl vs. PBS\nBL, q = 11.66 df = 8 ***P = 0.0002, PBS Ctrl vs. DT Ctrl q = 0.9285, DF = 8, ns P =\n0.9103. PBS Ctrl vs. DT BL q = 0.1986, df = 8 ns P = 0.9989. DT Ctrl vs. DT BL q =\n0.7299 df = 8, ns P = 0.9529. Center Two-way ANOVA, Alpha = 0.05, followed by\nŠídák’s multiple comparisons test was performed. PBS Ctrl vs DT Ctrl t = 0.3479\ndf = 8 ns P = 0.9997, PBS BL vs DT BL t = 1.575 df = 8 ns P = 0.633. (Right) Two-\nway ANOVA, Alpha = 0.05, followed by Tukey’s multiple comparisons test was\n\nperformed. PBS Ctrl Vs DT Ctrl q = 00.7223 df = 8 ns P = 0.9542 PBS BL vs DT BL t\n= 0.1102 df = 8 ns P = 0.9998. A representative of two independent experiments\nwith 3 mice per group in each is shown. (c, d) t-SNE embedding of endothelial\ncells isolated from lungs of DT treated and Ctrl mice color coded by cell type\n(left) or experimental condition (middle) and density plots of the distribution\nof endothelial cells between conditions (right). (c) fibroblast, (D) myeloid cells.\n(e) Heatmaps showing the top 200 genes that correlate the most with imputed\ncell scores of the indicated factors for endothelial cells. Each column is a cell;\ncells are ordered based on their factor score in ascending order from left to right\nindicated by the green bar. The treatment condition for each cell is indicated by\ngrey (Ctrl) and red (DT) bars. Select genes of interest are shown.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 7 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 7 | Robustness and validation of BayesPrism\ndeconvolution. (a) For each cell type, Spearman’s correlation of cell fraction\nacross all spots was calculated between deconvolution using all available reads\nand 1 of 20 separate deconvolutions using the available reads downsampled to\n90%. Points represent the mean of the 20 Spearman’s correlation calculations\nand error bars are the minimum and maximum correlation values. (b)\nComparison of cell fractions across separately deconvolved serial sections. For\nall four biological samples, the average cell fraction for each cell type is plotted\nin the first serial section relative to the second. Trend line indicates a slope of 1.\nSpearman’s correlation is shown. (c) Comparison of average log cell fractions in\n\neach of 8 tissue sections using the standard scRNA-seq reference or the reference\nwith tumor RNA substituted for KP-Tracer tumor cells. Trend line indicates a\nslope of 1. Spearman’s correlation is shown. (d, e). Examples of positive spots for\ncertain populations of interest are associated with histological features. Images\nare from representative areas of control and Treg depleted tissue sections. Plots\nwith positive spots display the same example areas in the top of each panel\narrangement with the H&E stained image at lower resolution. (Br = bronchi; A/V\n= artery / vein; LV = lymphatic vessel). Analysis performed on (A-C) and images\nare representative of (D-E), one of two serial sections for each of four samples (DT\nand Ctrl two biological replicates each). One experiment was performed.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fA\n\nT\nM\nE\n\nr\ne\nt\ns\nu\nc\n\nl\n\nn\no\ni\nt\nc\na\nr\nf\n\nl\nl\n\ne\nC\n\nCtrl 1\n\nC\n\nCell fraction cluster\n\n1\n2\n3\n4\n5\n6\n7\n\nRNA %\nZ- score\n4\n2\n0\n- 2\n- 4\n\nB\n\ni\n\nn\no\ns\ns\ne\nr\np\nx\ne\n\nd\ne\nz\n\ni\nl\n\na\nm\nr\no\nn\n\ng\no\nL\n\n3\n\n2\n\n1\n\n0\n\n4\n\n2\n\n0\n\n5\n\n4\n\n3\n\n2\n\n1\n\n0\n\n1.5\n\n1.0\n\n0.5\n\n0.0\n\ne\nk\n\ni\nl\n.\n2\nT\nA\n\ne\nk\n\ni\nl\n.\n1\nT\nA\n\nc\ni\nr\nt\ns\na\nG\n\ne\nk\n\ni\nl\n.\n\nm\nr\ne\nd\no\nd\nn\nE\n\ne\nk\n\ni\nl\n.\nr\no\nt\ni\nn\ne\ng\no\nr\np\n.\ng\nn\nu\nL\n\ny\nt\ni\nc\ni\nt\ns\na\np\n.\nh\ng\nH\n\ni\n\nl\n\nHopx\n\nSftpc\n\n8\n\n6\n\n4\n\n3\n\n2\n\n1\n\n0\n\n3\n\n2\n\n1\n\n0\n\nFn1\n\nGkn2\n\nSox2\n\nGc\n\nItga2\n\nAT1.like AT2.like EMTEndoderm.likeGastricHigh.plasticity\n\nT\nM\nE\n\ne\nk\n\ni\nl\n-\n1\nT\nA\n\ne\nk\n\ni\nl\n-\n2\nT\nA\n\nc\ni\nr\nt\ns\na\nG\n\ne\nk\n\ni\nl\n-\n\nm\nr\ne\nd\no\nd\nn\nE\n\ny\nt\ni\nc\ni\nt\ns\na\np\n-\nh\ng\nH\n\ni\n\nl\n\ne\nk\n\ni\nl\n-\nr\no\nt\ni\nn\ne\ng\no\nr\np\n\ng\nn\nu\nL\n\nCtrl 2\n\nTreg depleted 2\n\nGastric\n\nEndoderm.like\n\nAT2.like\n\nAT1.like\n\nEMT\n\nLung.progenitor.like\n\nHigh.plasticity\n\nGastric\n\nEndoderm.like\n\nAT2.like\n\nAT1.like\n\nEMT\n\nLung.progenitor.like\n\nHigh.plasticity\n\n1mm\n\n1mm\n\n1mm\n\nImmune response\n\nTumor state\n\nGastric\n\nEndoderm.like\n\nTumor state\n\nAT2.like\n\nGastric\n\nAT1.like\nEndoderm.like\n\nAT2.like\nEMT\nAT1.like\n\nEMT\n\nLung.progenitor.like\n\nLung.progenitor.like\n\nHigh.plasticity\n\n-\n\n+\nTRUE\n\nCondition\n\nD\n\ns\na\ne\nr\na\nn\no\ns\ne\n\ni\n\nl\n\nf\no\nr\ne\nb\nm\nu\nN\n\n20\n\n10\n\n0\n\nAT1-like AT2-like\n\nEMT Endoderm Gastric\n\nF\n\nGm10076\n\n30\n\nPtma\n\nCtrl\nControl\n\nTreg\nTreg depletion\ndepletion\n\nHigh\nplasticity\n\nLung\nprogenitor\n\nTnfrsf12a\n\nH3f3b\n\nF3\n\nIfrd1\n\nNr1d1\n\nGadd45b\n\nAtf4\n\nCcnl1\n\nClk4\n\nCoq10b\n\nSrsf11\n\nWsb1\n\nCxcl16\n\n20\n\nP\nd\ne\nt\ns\nu\nd\na\n\nj\n\n0\n1\ng\no\nl\n-\n\n10\n\nHbb- bs\n\nHspa8\n\nTle5\n\nGm10073\n\nGm9843\n\nTubb5\n\nSec61b\n\nCcl21a\n\nFtl1- ps1\n\nFkbp1a\n\nTmsb4x\n\nFkbp4\nCrip2\n\nSt3gal4\n\nTrpt1\n\nEif3f\n\nEno1\n\nHsp90ab1\n\nGm10250\n\nStmn1 Gm2000\nGm9493\n\nGm11808\n\nLrrc58\nSh3bgrl3\n\nGstp1\n\nPtms\nRnf5\n\nPrdx2\n\nRtraf\n\nFkbp2\n\nAldoa\n\nTpi1\n\nKrtcap2\n\nMgst3\n\nHsp90b1\n\nTuba1b\nGsta4\nMettl7a1\n\nLdha\n\nAtpif1\n\nDnaja1\n\nSfn\n\nHspb1\n\n0\n-2.5\n\nMapk3\n\nSnrpb\n\nSlc25a25 Ppp1r15a\n\nH4c8\n\n5430416N02Rik\n\nU2af1\n\nGigyf2\n\nZkscan5\n\nEwsr1\nSnhg16\n\nSrsf3\nChtop\n\nRsrc2\n\nUbl3\nNop58\n\nEif5\nSlc6a14\nXpc\n\nKlf5\n\nActn1\n\nNsrp1\n\nTra2b\nMafk\n\nNfkbia\nKlf6\n\nRcan1\nMaff\n\nCcl2\n\nSntb2\n\nFabp3\nPlk2\n\nPtpn14\n\nNrg1\n\nMab21l4\n\nDazl\n\nOser1\n\nPhlda1\n\nEgr1\n\nGc\n\nSprr1a\n\nLog2 fold-change Treg depleted - Control\n\n0\n\n2.5\n\ns\na\ne\nr\na\nn\no\ns\ne\n\ni\n\nl\n\nf\no\nr\ne\nb\nm\nu\nN\n\n10\n\n5\n\n0\n\nE\n\nG\n\nt\ne\ns\n\ne\nn\ne\ng\n\nr\ne\nt\ns\nu\nc\n\nl\n\nr\no\nm\nu\nt\n\nl\n\na\nt\ne\n\ni\n\nc\nv\no\nn\na\nj\nr\na\nM\n\nAT1-like\n\nAT2-like\n\nGastric\n\nHigh\nplasticity\n\nLung\nprogenitor\n\nHPCS\n\n0\n\n5\n\n10\n\n-log10 adjusted P\n\nExtended Data Fig. 8 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\n\n\fExtended Data Fig. 8 | Spatial transcriptomic analysis of tumor cell states\nperturbed in response to Treg cell depletion. (a) Hierarchical clustering of\ntumor spots by tumor state RNA fractions. (b) Log normalized expression of\ntumor state marker genes in assigned spot clusters from A. (c) H&E staining of\n3 independent KP tumor sections (in addition to those shown in Fig. 5b) with\ntumor spots denoted by their assigned cluster in A. (d) Number of tumor lesion\nareas identified across all lung tumor states in control or Treg depleted mice\n(85 tumor lesions total). (e) Number of tumor lesion areas identified in Treg\ndepleted sections across all tumor states colored by immune response status\nin Treg depleted mice (N = 38 tumor lesion areas). (f) Differentially expressed\n\ngenes (Wilcoxon test BH adjusted) between tumor cells between control and Treg\ndepleted conditions. (N = 239 cells total). (g) GSEA of differentially expressed\ngenes in F within gene sets defined by different tumor clusters identified\nin Marjanovic et al.29 which partially align with tumor states identified by\ndeconvolution. Dashed line indicates adjusted p-values <0.05. (NES = normalized\nenrichment score; HPCS = high plasticity cell state). Analysis performed on (A-D,\nB-G) and images are representative of (C), one of two serial sections for each of\nfour samples (DT and Ctrl two biological replicates each). One experiment was\nperformed.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 9 | Clustering and cell linage annotation of scRNA-seq\ndatasets of human lung adenocarcinomas. (a) Heatmap displaying genes used\nto determine lineage assignments for single cell PhenoGraph clusters in human\nLuAd samples. Color bar represents the mean log normalized gene expression\n\nin each PhenoGraph cluster scaled from 0 to 1 for each gene. (b) Global t-SNE\nembedding across of all human lineages as in Fig. 4b colored by sample. ID.\nAll genes used for annotation are shown. (c, d) t-SNE embeddings of the T/NK\nlineage colored by PhenoGraph cluster (C) or sample ID (D).\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 10 | See next page for caption.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\fExtended Data Fig. 10 | Association of Treg abundance with transcriptional\nfeatures of endothelial cells in human LuAd and loadings of human and\nmouse fibroblast and endothelial cell factors. (a) Treg proportion of\nhematopoietic cells (CD45+) calculated from scRNA-seq data across all samples.\n(b) Treg proportion of hematopoietic cells compared to the Treg proportion of\nCD3+ cells across all human samples. (c) Mean log2 cell loading of CAR4+ capillary\n(factor 3) and other inflammation/hypoxia associated human endothelial factors\n(4,5) plotted against log2 Treg proportion in each patient sample. Spearman\ncorrelation estimate (R) and p value are listed. Trend line represents a linear\nmodel fit between the two and shading indicating the 95% confidence interval.\n(d) t-SNE of human endothelial cells colored by factor 3, 4, or 5 cell loading\n\n(max 2.5) or sample ID. (N = 19 patient samples). (e,g) Heatmap showing Jaccard\nsimilarity of genes associated with human and mouse fibroblast (E) or myeloid\n(G) factors. (f,h,i) Mean log2 cell loading of factors negatively associated with\nTreg frequency in fibroblasts (F) and myeloid cells (H), or positively associated\nin myeloid cells (I) plotted against log2 Treg proportion in each patient sample.\nSpearman correlation estimate (R) and p value are listed. Trend line represents\na linear model fit between the two and shading indicating the 95% confidence\ninterval. (fibroblast N = 20; myeloid N = 23). (j) Heatmap showing the Spearman’s\ncorrelation between Treg cell frequency associated human factors with\nconserved trends in mouse Treg-depletion.\n\nNature Immunology\n\nArticlehttps://doi.org/10.1038/s41590-023-01504-2\f\fβ\n\n\fβ\n\n\f  μ\n\n"