An Integrated Stress Response Pathway activates Perivascular Cancer-Associated Fibroblasts to Drive Angiogenesis and Tumor Progression

ATF4 is a major effector of the Integrated Stress Response (ISR), a homeostatic mechanism coupling cell growth and survival to bioenergetic demands. Although the pro-tumorigenic role of the ISR in a tumor cell-intrinsic manner has been established, its role in cell-extrinsic processes remains unexplored. Using novel conditional knockout ATF4 mouse models, we show that global, or broblast (FB)-specic loss of host ATF4 results in abnormal tumor vascularization and a pronounced tumor growth delay in syngeneic melanoma and pancreatic tumor models, a phenotype which is largely reversed by co-injection of ATF4 wt/wt FBs. Single-cell tumor transcriptomics uncovered a reduction of markers associated with FB activation in a cluster of perivascular cancer-associated broblasts (CAF) in ATF4 Δ/Δ mice. ATF4 Δ/Δ FBs displayed signicant defects in collagen biosynthesis and deposition and reduced ability to support angiogenesis in vitro. Mechanistically, ATF4 directly regulates the expression of the Col1a1 gene as well as the biosynthesis of glycine and proline, the major amino acids comprising collagen bers. Analysis of human tumor samples revealed a strong correlation between ATF4 and collagen levels and between an ATF4 FB signature and expression of collagen genes. Our ndings uncover a novel role of stromal ATF4 in shaping CAF functionality, a key driver of disease progression and therapy resistance. expression (below 1st quartile) and high COL1A1 expression (above 3rd quartile). Survival analysis using Kaplan–Meier and the log-rank test between COL1A1-low and -high groups, was performed using R package survival and OncoLnc 61 . Correlation analysis and Kaplan-Meier plots were produced using R package ggplot2. for host ATF4’s role in tumor progression and metastasis. ATF4 is essential for the CAF activation via direct regulation of Col1a1 expression and by impacting multiple additional steps in the collagen synthesis pathway, including Glycine (Gly) and Proline (Pro) pools. The resulting abrogation of Collagen I (and potentially additional collagen isoforms) in ATF4-decient FBs leads in dramatic reduction in secreted extracellular matrix collagen, which in turn results in defective CAF activation and reduced levels of angiogenic cytokine signaling to endothelial cells. The resulting defective angiogenesis leads to reduced support for primary and metastatic tumor growth.


Introduction
The tumor microenvironment (TME) is a diverse ecosystem comprised of multiple cell types, including malignant, as well as untransformed stromal and immune cells, with overlapping or sometimes opposing functions that impact tumor growth and progression to metastasis and shape therapeutic responses 1,2 .
Among stromal cells, cancer-associated broblasts (CAFs) constitute a distinct and heterogeneous population, characterized as one of the most active and functionally important components of the TME 3,4 . These intratumoral broblasts are often co-opted to support multiple hallmarks of cancer 5,6 . CAFs are the primary source of extracellular matrix (ECM) components, including collagens, bronectin, matrix metalloproteinases (MMPs) 7 , which have been shown to modulate tumor stiffness and facilitate tumor progression [8][9][10] . In addition to ECM production and organization, CAFs impact tumor metabolism 11,12 and also secrete a plethora of cytokines, chemokines, growth factors, and exosomes to further support tumor progression and modulate responses to treatment [13][14][15] .
Although diverse subtypes of CAFs express multiple markers such as α-smooth muscle actin (αSMA), platelet-derived growth factor receptor β (PDGFRβ) 4,16 , or broblast activation protein (FAP) 17 , none of these is considered to be su ciently unique to distinguish all CAFs within the TME. However, recent advancements in single-cell RNA sequencing (scRNA-seq) and careful analysis of CAFs perspicuous spatial distribution and distinctive functional characteristics have identi ed distinct CAF subtypes in melanoma and pancreatic tumors 18−20 . Interestingly, single-cell expression pro ling on isolated CAFs from a genetically engineered mouse model of breast cancer identi ed a subset of vascular CAFs (termed vCAFs). These vCAFs are localized in the perivascular region and characterized by the expression of genes involved in vascular development 21 . These ndings further support previous reports establishing a melanoma and pancreatic cancer tissue arrays revealed a signi cant positive correlation between collagen I and ATF4 levels and TCGA analysis showed a negative correlation between expression levels of collagen I with overall patient survival in multiple malignancies. These studies uncover a novel, protumorigenic role of the ISR pathway via CAF-dependent mechanisms and suggest new modes for therapeutic intervention.

Results
Deletion of host ATF4 inhibits tumor growth and extends survival in syngeneic melanoma and pancreatic tumor models To test the impact of host ATF4 on tumor growth, we generated a tamoxifen-inducible knockout mouse by crossing Atf4 / mice with Rosa26::CreER T2 (Fig. 1a and Extended Data Fig. 1a) as we previously described 33 . Rosa26::CreER T2 :Atf4 wt/wt and Rosa26::CreER T2 :Atf4 / mice were treated with tamoxifen ( Fig. 1a), resulting in almost complete (90-100%) excision of ATF4 as assayed by qRT-PCR analysis of ATF4 mRNA levels in whole liver, lung and spleen homogenates of Atf4 D/D mice ( Fig. 1b and Extended Data Fig. 1b). Ablation of ATF4 was well tolerated causing only a modest and transient decrease in hematocrit and body weight (Extended Data Fig. 1c and data not shown). We have followed Atf4 wt/wt and Atf4 D/D mice for over a year with no overt toxicities observed. Full necropsy analysis revealed no signi cant pathological aberrations in the Atf4 D/D cohort apart from some mild/moderate toxicity on the small intestine, spleen and liver (data not shown).
Mouse melanoma B16F10 cells were injected orthotopically (subcutaneously) in the right ank of Atf4 wt/wt and Atf4 D/D mice (Extended Data Fig. 1d). A pronounced delay in tumor growth was observed in Atf4 D/D mice accompanied by a signi cant increase in their survival, compared to Atf4 wt/wt littermates (Fig. 1c,d and Extended Data Fig. 1e). These effects were not gender dependent (Extended Data Fig. 1f). In addition, a signi cant inhibitory phenotype on tumor growth was observed when host ATF4 was excised following the establishment of palpable B16F10 tumors, indicating that host ATF4 expression contributes to both tumor initiation and progression (Fig. 1e). To test if these ndings extend to other tumor types, we injected subcutaneously mice with syngeneic MH6419 pancreatic tumor cells, originated from the Kras LSL−G12D/wt ;Trp53 / ;Pdx1-Cre (KPC) model of spontaneous pancreatic cancer 34 . Similar to the melanoma growth results, global host ATF4 ablation resulted in a signi cant delay in tumor growth and extension of overall survival (Fig. 1f,g and Extended Data Fig. 1g). In total, these results suggest that host ATF4 signi cantly contributes to the establishment and growth of syngeneic tumors.

ATF4 is essential for CAF activation in the tumor microenvironment
To delineate the role of host ATF4 in the sequence of events leading to tumor growth, we performed transcriptional pro ling at the single-cell level (scRNA-seq) in smaller (150 mm 3 ) and larger (300 mm 3 ) size B16F10 tumors grown in Atf4 wt/wt and Atf4 D/D mice (Extended Data Fig. 2a). We acquired single-cell transcriptomes from a total of 7,414 cells from small and 28,166 cells from larger B16F10 tumors, respectively, for downstream analysis. Graph-based clustering of cells following Uniform Manifold Approximation and Projection (UMAP), identi ed 7 distinct cell types in small B16F10 tumors (Fig. 2a), with CAFs accounting for 6.12% of the total cells. To identify the different cell types represented by the clusters, we cross-referenced the gene signature of each cluster with known markers of cell populations described in the literature (Fig. 2b) 35,36 . We then performed differential gene expression (DE) analysis for each cell type to identify potential transcriptome changes between the Atf4 wt/wt and Atf4 D/D cohorts. We con rmed host ATF4 deletion by the absence of Atf4 mRNA expression in Atf4 D/D mice across all the "host" clusters (i.e. CAFs, dendritic cells, endothelial cells, etc.), while Atf4 levels remained unchanged in melanoma clusters compared to the Atf4 wt/wt mice (Extended Data Fig. 2b). Interestingly, we observed a substantial decrease in the total number of endothelial cells in the Atf4 Δ/Δ grown tumors (Extended Data Fig. 2c), which implies a potential reduction in angiogenesis. Although there was a decrease in the total number of T cells/NK cells in Atf4 D/D grown tumors, anti-CD8 treatment caused a small increase in the rate of tumor growth in both Atf4 wt/wt and Atf4 D/D mice, suggesting that other mechanisms must account for the dramatic differences seen between these cohorts (data not shown). We did, however, observe striking differences in gene expression in a cluster that corresponds to CAFs. In this cluster, we identi ed 148 DE genes (Supplementary Table 1), including, a signi cant downregulation of Col1a1 and Col1a2 in the Atf4 D/D mice-grown tumors (Fig. 2c). These two genes encode the pro-alpha1(I) and pro-alpha2(I) chains, respectively, essential components of type I collagen, which is the most abundant collagen (~ 90%) in the body and in the ECM 37 . Furthermore, several additional collagen genes were downregulated in tumors grown in Atf4 D/D mice (data not shown). Notably, the expression levels of Acta2 (encoding αSMA) and Pdgfrβ (platelet-derived growth factor receptor beta), which are broadly reported as markers of CAFs 4,6 , were nearly absent in the tumors grown in Atf4 D/D mice (Fig. 2c). Geneset enrichment analysis using Reactome pathways revealed that genes with higher levels of expression in Atf4 wt/wt compared to Atf4 D/D CAFs were mainly enriched for CAF-related functions, such as extracellular matrix organization, collagen formation and biosynthesis (Fig. 2d, left). In contrast, genes expressed at higher levels in Atf4 D/D CAFs were associated mainly with ISR activation, including response of GCN2 to amino acid de ciency and eukaryotic translation initiation (Fig. 2d, right), a nding consistent with loss of ATF4 in these cells. Several studies have de ned distinct CAF subtypes with distinguishable functional programs based on speci c gene expression signatures. Here, using the gene signatures described in Bartoschek et al. 21 we identi ed 3 distinct cell subclusters, vascular CAFs (vCAFs), matrix CAFs (mCAFs) and cyclin/melanoma CAFs (cCAFs/melCAFs) (Fig. 2e). Interestingly, within the cCAFs we identi ed the melCAFs, a newly discovered sub-population which retains traits from CAFs but also melanoma cells.
The vCAFs (Extended Data Fig. 2d) are considered essential for vascular development and angiogenesis and are also characterized by the highest levels of expression of Acta2 and Pdgfrβ among all CAF subclusters (Fig. 2f). Notably, the vCAFs were reduced substantially in the Atf4 D/D mice (Fig. 2g) and the Col1a1 and Col1a2 were signi cantly downregulated only in the vCAFs subcluster in the Atf4 D/D mice (Fig. 2h).
To track the regulation and fate of CAFs during B16F10 tumor progression and reorganization of the TME, we also performed scRNA-seq on larger sized tumors (Extended Data Fig. 2a). Graph-based clustering was used to identify 5 distinct cell clusters (Fig. 2i), which have been validated by a unique gene signature (Extended Data Fig. 2e). Atf4 levels remained extremely low in Atf4 D/D mice across all the "host" clusters (Extended Data Fig. 2f). Similar to the smaller tumor data, the expression levels of Acta2 in CAFs cluster were signi cantly reduced in the tumors grown in Atf4 D/D mice (Fig. 2j) and the number of endothelial cells was higher in the Atf4 wt/wt mice indicating an important role of ATF4 in tumor vascularization (Extended Data Fig. 2g). In contrast to the smaller sized tumors, no difference was observed in the expression levels of Col1a1 and Col1a2 between the Atf4 wt/wt and Atf4 D/D cohorts (Fig. 2j), as observed in smaller sized tumors, suggesting the presence or activation of an alternative mechanism of Col1 gene expression regulation to compensate for the ATF4 loss during tumor progression. Geneset enrichment analysis revealed that genes upregulated in Atf4 wt/wt CAFs were related mainly to ISR/UPR pathways as well as to smooth muscle contraction (Extended Data Fig. 2h). By following the aforementioned approach, we identi ed 4 distinct CAFs subclusters (Fig. 2k). The newly discovered sub-population melCAFs, was separated from the cCAFs sub-cluster and was found to form a distinct sub-cluster (Extended Data Fig. 2i,j). vCAFs cluster (Extended Data Fig. 2k), featured the highest expression of Acta2 and Pdgfrβ among all CAF subclusters again (Extended Data Fig. 2l) and critically, was diminished in tumors grown in Atf4 D/D mice (Fig. 2l). Interestingly, the vCAFs remained as a distinct subcluster during the transition from smaller to larger sized melanoma tumors suggesting the importance of this subcluster in shaping the TME ( Fig. 2m and Extended Data Fig. 2m,n). Collectively, these results suggest that host ATF4 deletion impairs CAF functionality at different stages of tumor development resulting in a tumor-inhibiting phenotype.

Host ATF4 loss results in abnormal tumor vascularization and reduced ECM component deposition
The results from the scRNA-seq data suggested speci c defects in tumors grown in Atf4 Δ/Δ compared to Atf4 wt/wt mice. To further explore these differences, we stained larger size B16F10 and MH6419 tumors of equal volume (approx. 300 mm 3 (Fig. 3c,d). Abnormal blood vessels provide reduced levels of nutrients and oxygen to tumor tissue resulting in intratumoral necrotic areas. Indeed, B16F10 tumors from Atf4 Δ/Δ mice presented a higher percentage of necrotic areas compared to the Atf4 wt/wt tumors (Extended Data Fig. 3f-h). It is wellestablished that activated CAFs in the TME contribute to tumor neo-angiogenesis by providing structural support and local secretion of chemokines and angiogenic factors 6 . Staining of B16F10 and MH6419 tumors for αSMA, showed dramatic reductions in tumors derived from Atf4 D/D mice compared to levels observed in Atf4 wt/wt mice, in which was primarily restricted in the blood vessels (Fig. 3e,f and Extended Data Fig. 3i-m). Levels of additional CAF markers such as broblast activation protein (FAP) in MH6419 tumors and PDGFRβ in B16F10 tumors, were also found to be signi cantly reduced in Atf4 D/D mice (Fig. 3g,h and Extended Data Fig. 3n-p). These ndings further corroborate the results from the scRNAseq analysis, where the expression levels of Acta2 and Pdgfrβ in CAFs cluster were signi cantly reduced in the tumors grown in Atf4 D/D mice. The expression of markers of other cell types which also contribute to blood vessel functionality such as neural/glial antigen 2 (NG2; pericytes) was not appreciably altered in the melanoma tumors from Atf4 D/D mice (data not shown). A primary function of all broblasts is the synthesis and maintenance of ECM. Interestingly, collagen levels were signi cantly reduced in tumors grown in Atf4 D/D mice (Fig. 3i,j and Extended Data Fig. 3q,r), further supporting a critically important role of ATF4 in CAF activation and stromagenesis in the solid tumors. Finally, immuno uorescence staining of human melanoma tissues revealed that ATF4 is highly expressed (apart from the tumor cells) on CAFs (aSMA) that localized on the perivascular area (CD34) compared to other CAFs reside away from the blood vessels (Fig. 3k).
We then sought to investigate if the absence of activated broblasts is related to the tumor-inhibitory effects observed in Atf4 D/D mice. Since the syngeneic B16F10 tumors are grown subcutaneously, we isolated dermal broblasts from tumor-naïve Atf4 wt/wt (DFB wt/w ) and Atf4 D/D (DFB D/D ) mice (Extended Data Fig. 4a). These were then co-injected with B16F10 cells (3:1 ratio) into the anks of Atf4 wt/wt or Atf4 D/D mice (Fig. 3l). DFB wt/wt injected into Atf4 D/D mice nearly completely reversed the tumor growth inhibition observed in Atf4 D/D + DFB D/D group, while DFB D/D injected into Atf4 wt/wt mice, caused a delay in tumor growth compared to the Atf4 wt/wt + DFB wt/w group (Fig. 3m). Although our results strongly implicated the broblast compartment in the tumor growth de ciency in Atf4 D/D mice, we could not exclude a contribution of other cellular compartments. To further investigate the contribution of activated broblasts, we genetically excised ATF4 in a more tissue-speci c manner by crossing Atf4 / with Col1a1::CreER T2 mice (Extended Data Fig. 4b). The Col1a1 promoter has shown activity speci cally in broblasts and osteoblast compartments. Deletion of ATF4 in these compartments following tamoxifen treatment did not result in any weight loss, reduction of hematocrit, or any other overt phenotypes (data not shown). Col1a1 driven-speci c ATF4 deletion caused a signi cant B16F10 tumor growth delay, similar to that observed in Atf4 D/D mice (Fig. 3n). This phenotype was accompanied by reduced MVD (Fig. 3o,p). Thus, these results lend strong support to the notion that ATF4 de ciency in broblasts creates an inhibitory TME through abnormal angiogenesis and reduced collagen deposition.
ATF4-dependent, multilayered regulation of the collagen biosynthetic pathway contributes to protumorigenic broblast functionality The single-cell transcriptomic analysis revealed a profound impact of ATF4 on broblasts activation status and collagen mRNA levels. To cross-validate some of the scRNA-seq ndings, we performed genome-wide microarray analysis on isolated lung broblasts from Atf4 wt/wt (LFB wt/wt ) and Atf4 D/D (LFB D/D ) mice. We identi ed over 3000 genes to be differentially expressed, with a profound reduction in expression of collagen-associated (i.e. Col1a1, Co1a2, etc.) and broblast activation (i.e. Pdgfrβ) genes in LFB D/D (Fig. 4a), results which were con rmed by qRT-PCR analysis ( Fig. 4b and Extended Data Fig. 4c).
Similarly, Col1a1 levels were signi cantly reduced in DFB D/D (Extended Data Fig. 4d). Not surprisingly, ECM organization/degradation and collagen biosynthesis pathways were the most impaired in LFB D/D , validated by Geneset enrichment analysis on the 100 most downregulated genes in Atf4 D/D mice (Fig. 4c). The biosynthesis of collagen is a highly coordinated process, involving mRNA synthesis and translation into pro-collagen, hydroxylation, glycosylation, and crosslink formation (Fig. 4d). Because both in vitro and in vivo RNA sequencing analysis showed downregulation of Col1a1 expression in the absence of ATF4, we reasoned that ATF4 directly regulates its expression. Analysis of mouse chromatin immunoprecipitation sequencing (ChIP-seq) data 38 , revealed potential binding sites of ATF4 inside intron 5 of Col1a1, as alternative transcription start sites (TSS) (Fig. 4e, left). ChIP-qRT-PCR validated the ChIPseq results (Fig. 4e, right). We hypothesized that the severe phenotype of reduced collagen levels in tumors grown in ATF4 de cient mice could also involve additional steps in the pathway. Translation of Col1a1 mRNA requires adequate levels of glycine, proline and/or hydroxy-proline, which account for 70-100% of its polypeptide chain. Since ATF4 drives the transcription of genes involved in amino acid biosynthesis and transport 29 , we examined the levels of glycine and proline in ATF4 pro cient and de cient LFB. Using NMR spectroscopy, we observed that intracellular levels of both amino acids were signi cantly reduced in ATF4 de cient cells (Fig. 4f). To further corroborate these ndings, we measured the metabolic ux from serine to glycine and glutamine to proline by labeling the cells with serine- 13  and Extended Data Fig. 4j-m). Notably, re-expression of a mouse ATF4 homolog in LFB D/D , resulted in the detection of intracellular procollagen levels similar to the levels found in LFB wt/wt (Fig. 4m). Collectively, these results demonstrate that ATF4 is required to maintain a functional phenotype in broblasts, through the regulation of multiple steps of the collagen biosynthetic pathway.
ATF4 de cient broblasts display signi cantly attenuated pro-angiogenic activity To further investigate the abnormal vascularization phenotype observed in both B16F10 and MH6419 tumors, primary lung endothelial cells were isolated from healthy Atf4 wt/wt (EC wt/wt ) and Atf4 D/D (EC Δ/Δ ) mice ( Fig. 5a and Extended Data Fig. 5a), and tested for their ability to form endothelial tubes on Matrigel-coated plates. ATF4 was con rmed to be successfully excised in ECs (Fig. 5b,c and Extended Data Fig. 5b). After 4 and 8 hours of plating, EC Δ/Δ did not show any differences on sprouting or tube formation parameters compared to the EC wt/wt . In contrast, stimulation of endothelial cells with conditioned medium (CM) derived from LFB WT and LFB D/D revealed a signi cant defect in the response of EC Δ/Δ (Fig. 5d,e and Extended Data Fig. 5c). Interestingly, the CM from LFB D/D caused a signi cant reduction in both sprouting and tube formation of EC wt/wt compared to those treated with CM from the LFB wt/wt , indicating a possible de ciency in the LFB D/D secretome (Fig. 5e). To further probe the reason for this de ciency, CM from ATF4 pro cient and de cient LFB as well as from LFB D/D expressing the mouse ATF4 homolog, were analyzed for multiple secreted angiogenesis-related proteins. Intriguingly, VEGF, SDF-1, IGFBP-2 and IGFBP-9 levels were signi cantly reduced in the CM from LFBs D/D , and these levels were re-established in LFB D/D with restored ATF4 levels ( Fig. 5f,g). It is well established that TGFβ/Smad3 pathway is active in CAFs, which in turn secrete VEGF, among the other cytokines, to boost angiogenesis 39 . Interestingly, here we show that expression of both p-SMAD3 and T-SMAD3 were downregulated in LFB D/D post TGFb treatment (Fig. 5h), in agreement with already published data 40 . Taken together, these results suggest that ATF4 loss in broblasts impairs their pro-angiogenic activity through a defective secretome.

Host ATF4 ablation inhibits lung metastasis of melanoma tumors
Since acute deletion of host ATF4 caused a signi cant tumor-inhibitory phenotype, and since activated CAFs also play a critical role in the establishment of the metastatic niche 41-43 , we speculated that ATF4 de ciency could also result in an inhibitory effect on lung metastasis. B16F10 melanoma tumors, similar to human melanoma metastasize to multiple sites, but primarily to the lung 44 . We rst examined the impact of host ATF4 deletion in the pre-metastatic niche by analyzing gene expression changes on lungs from Atf4 wt/wt and Atf4 D/D mice at 4 weeks post tamoxifen treatment. Genome-wide microarray analysis identi ed more than 170 genes as being differentially expressed, with 21 genes to be signi cantly downregulated in Atf4 D/D lungs, including Col1a1, validated by qRT-PCR analysis (Fig. 6a,b). Importantly, pathway analysis on the most dysregulated genes revealed defects on collagen formation, extracellular matrix organization and integrin cell surface interactions pathways (Fig. 6c), which is consistent with the gene expression data on LFBs. Mass spectrometry of lung tissue extracts revealed pronounced reductions in glycine and proline levels on Atf4 D/D mice compared to their Atf4 wt/wt littermates (Fig. 6d).
These results indicate that loss of host ATF4 might cause an unfavorable metastatic niche, possibly through the regulation of broblast functionality. Tail vein injection of B16F10 cells resulted in e cient lung colonization in Atf4 wt/wt mice at 3 weeks. In contrast, a signi cantly reduced number of lung metastases (evaluated macroscopically), as well as area covered by metastatic melanoma cells (determined by H&E stained serial lung sections), were observed in Atf4 D/D compared to Atf4 wt/wt mice ( Fig. 6e-h). Critically, in a more physiologically relevant model of metastasis, where equal volume (appr. 300 mm 3 ) of B16F10 tumors in both genotypes were surgically excised and lungs were examined at 4 weeks post excision (Fig. 6i), the results were even more striking: 6 out of 9 lungs from Atf4 D/D mice lacked any detectable metastases, while the other 3 presented with only a small single metastatic nodule.
In contrast, a signi cantly higher number of metastases was observed in all the lungs of Atf4 wt/wt littermates (Fig. 6j,k). Together, these results indicate that host ATF4 acts as a driving factor in the development of the pre-metastatic niche and e cient metastatic process in B16F10 melanoma tumors.
Association between ATF4 levels, stromagenesis and poor prognosis in melanoma patients To investigate the relevance of our ndings in human malignancies, we analyzed the expression of Col1a1, Acta2 and multiple other genes in relation to ATF4 activity in different cohorts of melanoma (SKCM) and pancreatic cancer patients (PAAD) from the TCGA database. Since ATF4 is primarily regulated at the translational level 29,33,38 , we used ATF4 transcriptional targets as a surrogate for ATF4 activation. However, none of the well-characterized ATF4 targets in epithelial cells that we and others have previously reported 33 were expressed at signi cant levels in broblasts or showed any correlation with the collagen genes or Acta2 (data not shown). This is likely due to the tissue-speci c and transformation-speci c repertoire of genes regulated by ATF4. Therefore, we decided to use a " broblast ATF4 target signature" consisting of the 30 most downregulated genes in Atf4 D/D compared to Atf4 wt/wt broblasts as a surrogate for ATF4 activation in this cell type (Fig. 7a). Interestingly, we found that Col1a1, Acta2, Pdgfrβ and FAP displayed a signi cant positive correlation with this dataset primarily in melanoma (SKCM) and pancreatic tumors (PAAD) but also in several other tumor types, including lung adenocarcinoma (LUAD), kidney renal clear cell carcinoma (KIRC) etc. (Fig. 7b and Extended Data Fig. 6a). In contrast, no signi cant correlation was found using a list of 30 randomly chosen genes ( Fig. 7c). To further probe this relationship, human malignant melanoma and high-density pancreatic cancer tissue arrays were stained for collagen (COL1) and ATF4 by immunohistochemistry (Extended Data Fig. 6b,c). Indeed, a signi cant positive correlation was found in melanoma tissue array (Fig. 7d,e and Extended Data Fig. 6d), which was stronger in the metastatic group compared to the primary tumor group (Extended Data Fig. 6e). These ndings further corroborate the results from the mouse melanoma metastasis model. A signi cant positive correlation was observed in the pancreatic tissue array (Extended Data Fig. 7a,b). Interestingly, we noticed that this correlation was stronger in grade 2 compared to grade 3 pancreatic cancer patients, suggesting that ATF4 may exert a stronger regulatory role on collagen deposition at earlier disease stages (Extended Data Fig. 7c). These ndings are consistent with the scRNA-seq results from B16F10 tumors, where we observed a gradual loss of Col1a1 regulation by ATF4 moving from smaller to larger size tumors. Notably, high expression of Col1a1 also correlated with poor prognosis in melanoma and many other tumor patients (Fig. 7f and Extended Data Fig. 7d). Together, these ndings suggest that ATF4-dependent activation of CAFs dictates early ECM organization and CAFs-instructed angiogenesis to support the growth of primary tumors as well as the metastatic phenotype (Fig. 7g).

Discussion
The non-malignant ecosystem of the tumor microenvironment is increasingly being appreciated as a key driver of tumor progression, aggressiveness, and resistance to therapy. Given the cardinal features of CAFs in the tumor microenvironment 4,41,45 , a better understanding of their transitory roles during tumor evolution and the mechanisms underlying these genotypic and phenotypic changes is crucial for developing novel therapeutic approaches. We have uncovered an essential, novel role for the master ISR effector ATF4 in shaping CAF functionality to dictate ECM organization and angiogenesis to support a tumor-promoting phenotype in experimental models of melanoma and pancreatic cancer (Fig. 7g).
Employing unbiased single-cell transcriptomic analysis of small and large-sized melanoma tumors, we surprisingly detected impaired CAFs activation in Atf4 D/D mice, based on the expression levels of Acta2 and Pdgfrβ, as one of the most commonly used CAF markers 4,46 . Interestingly, the levels of aSMA, PDGFRβ and FAP detected by immuno uorescent staining on melanoma and pancreatic tumors were substantially reduced to nearly undetectable in Atf4 D/D mice, indicating that ATF4 is essential for CAFs activation within the TME. A more constrained deletion of ATF4 in the broblast/osteoblast compartment in a similar tumor growth pro le as in the global ATF4 KO mice. Notably, co-injection studies of broblasts from Atf4 wt/wt mice led to signi cant recovery of tumor growth rates in Atf4 D/D mice.
Our studies also provide for a putative mechanism underlying the effect of ATF4. These data suggest that transcriptional changes in CAFs elicited by ATF4 loss are associated with primary CAFs functions, such as extracellular matrix organization, collagen formation and biosynthesis. We have shown reduced levels of deposited collagen in melanoma and pancreatic tumors post-global ATF4 excision which has been validated by ex vivo broblast-derived matrix assays. Mechanistically, we identi ed and validated an ATF4 binding site at the intron 5 of the Col1a1 gene. Intriguingly, in humans, this site has been identi ed as the second most common binding site region, with more than 20 transcription factors binding sites, deeming these loci regulatory hotspots 47,48 . Therefore, our data suggest a new multifactorial role of ATF4 on the regulation of the collagen synthesis pathway which further contributes to the protumorigenic broblast functionality.
There is overwhelming evidence related to the heterogeneity and plasticity of CAFs 6,45,46 . The combination of scRNA sequencing with functional assays has uncovered unprecedented diverse and spatially distinct CAF subpopulations based on speci c gene expression signatures 21,49,50 . Among our key ndings, we identi ed the vascular CAFs (vCAFs), a spatially distinct CAF subcluster featured by the highest levels of aSMA and PDGFRβ, that were reduced in Atf4 D/D mice. CAFs have been ascribed wellestablished roles in supporting angiogenesis through the release of pro-angiogenic factors, like VEGFA, FGF2 and SDF-1/CXCL-12 (CAFs secretome) 15,51,52 or via exertion of mechanical forces 23 within the tumor milieu. Our data indicate that conditioned medium (CM) from LFB wt/wt but not LFB D/D could support vascular tube formation and sprouting in in vitro assays of angiogenic activity. Interestingly, CM from ATF4-replete LFBs exhibited signi cantly higher levels of VEGF and SDF-1/CXCL-12 which have been shown to drive angiogenesis in malignancy 15,53 . However, we cannot exclude a possible proangiogenic role of vCAFs through altering mechanical forces, as they appear to be predominantly localized in proximity to the vasculature, and future studies will be required to address this. Kojima and colleagues have demonstrated that TGFβ and SDF-1 autocrine signaling facilitates the differentiation of stromal broblasts into CAFs in invasive human breast carcinomas 54 . Interestingly, we uncovered an impaired TGFβ/Smad3 signaling in TGFβ-treated LFBs lacking ATF4, suggesting a potential involvement of this pathway on the defective secretome of the LFB D/D .
Collectively, our work highlights the vital importance of the ATF4 on regulating the differentiation of stromal broblasts to CAFs and the activation of the latter through the collagen I synthesis and TGFβ/Smad3 pathways, respectively. Overall, the lack of overt toxicities following transient ATF4 deletion in mice, coupled with the demonstrated pro-tumorigenic role of ATF4 in a tumor-intrinsic as well as tumorextrinsic manner, further supports the notion that a clinically useful therapeutic window may exist for ISR and ATF4 inhibition as an attractive antitumor strategy.

Methods
Plasmids and other reagents. The list of reagents, assays and adenovirus used in this study is provided in the Supplementary Tables 2,3. Table 4. Cell culture. All cell lines are listed in the Supplementary Table 5. B16F10 and MH6419 cells were cultured in RPMI-1640 supplemented with 10% FBS in the presence of 5% CO2 at 37 °C. Isolated EC wt/wt and EC Δ/Δ cells were cultured in Endothelial Cell (EC) medium. Isolated LFB wt/wt , LFB Δ/Δ , DFB wt/wt and DFB Δ/Δ cells were cultured in phenol-free DMEM/F12. All cell lines were determined to be free of mycoplasma with repeated testing.
Lung endothelial cell isolation. Lungs were isolated from 8-10 weeks old Atf4 wt/wt and Atf4 D/D mice (males and females). Tissues from 8 were minced and digested in 10 ml of collagenase lysis buffer (5 mg/ml of collagenase type II (Worthington, Cat#LS004176)) dissolved in phenol red-free EC medium (ScienCell, Cat#1001-prf) without FBS and Pen/Strep under continues rotation on a rocker at 37 °C for 35-45 min. An equal volume of phenol-free EC medium supplemented with 10% FBS, 100 U/ml penicillin and 100 mg/ml streptomycin were added to the lysed tissues to quench collagenase and then passed through 70 µµ and 40 µµ cell strainer. Cells were spun at 300 g for 5 min, the pellet was resuspended in 10 ml of complete EC medium (phenol-free EC medium supplemented with 10% FBS, 100 U/ml penicillin, 100 mg/ml streptomycin, 1 µΜ of 4-hydroxytamoxifen (4-HT), 1x non-essential amino acids (NEAA) and 55 µΜ β-mercaptoethanol (β-ME)) and cells were plated in 10 cm plates and incubated at 37 °C for 1 h.
Immuno uorescence. Tumor tissues were cut in 8-10 µµ (or 100 µµ for confocal microscopy) thick sections, coded and stored at -80 °C. Slides thawed at room temperature and subsequently xed with 2% paraformaldehyde for 20 min. After 3x washes with TBS, tissues were blocked with 8% BSA and 1% Immunohistochemistry. For H&E staining, 5 µm thick para n sections were mounted on Superfrost Plus™ slides and stained using a Gemini AS Automated Slide Stainer. Slides were nally mounted with a resinous mounting medium (Thermo Scienti c ClearVue™ coverslipper).
EdU proliferation assay. Fibroblast cell proliferation was detected using the Click-iT EdU Alexa Fluor 488 Imaging kit according to the manufacturer's instructions.
qRT-PCR. Total RNA was isolated using the Macherey-Nagel kit and cDNA was synthesized using the High Capacity RNA-to-cDNA kit, according to manufacturer's instructions. qRT-PCR was performed with Power SYBR green PCR master mix. For data analysis, the QuantStudio 6 Flex Real-Time PCR System (Applied Biosystems) was used. Relative gene expression levels were de ned using the DDCt method and normalization was performed to 18S rRNA. All primers used in this study are described in the Supplementary Table 6.
Chromatin Immunoprecipitation (ChIP). The ChIP assay was performed as previously described 55 . IgG was used for pulldown as an experimental control. The primer sequences are listed in Supplementary  Table 6.
Conditioned medium. Fibroblasts were plated and cultured until 80% con uency. After 3x washes with PBS, phenol-free DMEM/F12, without any supplement, was added and kept for 24 h. The medium was spun at 300 g for 5 min, ltered through 0.22 µm lter and stored at -80 °C until further use.
Fibroblast-derived matrices (FDMs). Decellularized FDMs were generated as described previously 56 , using lung and dermal broblasts cultured at con uence in the presence of 75 µg/ml ascorbic acid for 8 days.
Tube formation assay. Each well of the 8-well slide chamber was coated with matrigel. 10 5 cells endothelial cells in 300 µl of complete EC medium were plated on matrigel on each well. Cells were allowed to attach for 4 h, and phase-contrast images were taken at 8, 12 and 16 h post-plating, using the Zeiss Observer.Z1 (Zeiss).
Protein angiogenesis array. Conditioned medium was added to the membranes of the proteome pro ler mouse angiogenesis array and processed according to the manufacturer's instructions.
Flow cytometry. Endothelial and broblast cells were stained with Live/Dead Fixable Aqua Dead Cell Stain Kit (Invitrogen, Cat#L34957) for live/dead cell discrimination. Cell proliferation was detected using the Click-iT EdU Alexa Fluor 488 Imaging Kit (Invitrogen, Cat#C10632) according to the manufacturer's instructions. For endothelial cells, cell surface staining against CD31 (Biolegend, Cat#102508) was performed for 30 min at 4 °C. All data acquisition was done using a FACSCanto II (BD Biosciences).
Derivatized with 60 ul 1:1 mix pyridine: MTBSTFA, 60 min 80C. After derivatization, samples were transferred to GC-MS vials and left at room temp for 4 days before running GC-MS (this was to stabilize the norvaline signal). NMR spectrometry. Cell or tissue samples were extracted using a biphasic extraction protocol. LC-MS grade Methanol and chloroform were purchased from Thermo Fisher. 0.22 micron ltered milli-Q water was used for extraction purpose. Fifty mg of tissue and/or 106 cells were extracted using 500 µλ 2:2:1 methanol/chloroform/water. The cell samples were sonicated using a sonicator bath and the tissue samples were homogenized using steal bids in a Tissuelyzer II system (Eppendorf). The samples were further centrifuged (13300 rpm, 4 o C). The upper fraction containing the polar metabolites were carefully collected and dried using a vacuum centrifuge (Eppendorf).
NMR spectra were acquired using Avance III HD 700 MHz NMR spectrometer (Bruker Biospin, Billerica, MA) tted with a 3 mm NMR triple resonance inverse probe and SampleJet system for automated high throughput spectral acquisition. All spectra were acquired at 298K. The pulse program of the acquired NMR spectra took the shape of the rst transient of a 2-dimensional NOESY and generally of the form RD-90-t-90-tm-90-ACQ. Where RD = relaxation delay, t = small time delay between pulses, tm = mixing time and ACQ = acquisition 57 . Continuous irradiation of water during RD and tm was used to suppress the water signal. The spectra were acquired using 1-second interscan delay, 0.1 second mixing time, 76K data points and 14 ppm spectral width with a variable number of scans depending on starting sample mass. The FIDs were zero-lled to 128K; 0.1 Hz of linear broadening was applied followed by Fourier transformation.
NMR spectra were imported into Chenomx v 8.0. (Edmonton, Canada) for quantitative targeted pro ling 58 . The processor module was used to phase and baseline correct the spectra followed by internal standard calibration and deletion of the water region. The processed spectra were then imported to the pro ler module for the targeted pro ling of selected metabolites. Quanti ed data from this process were exported for further analysis.
Metabolic tracing study. For serine-13C3 labeling experiments, cells were cultured in RPMI medium lacking glucose, serine, and glycine (TEKnova) supplemented with 2 g/L glucose and 0.03 g/L serine-13C3 (Sigma Aldrich) for 1 and 3 hours before harvesting. Cells were washed twice with ice-cold PBS prior to extraction with 600 µΛ of 80:20 acetonitrile:water over ice for 15 min. Cells were scraped off plates to be collected with supernatants, sonicated for 30 sec, then spun down at 15,000 RPM for 15 min. 7-8.5 min, 40%B; 8.5-8.7 min, 40% → 80%B; 8.7-10 min, 80%B. The overall runtime was 10 min and the injection volume was 6 µΛ. Agilent Q-TOF was operated in negative mode and the relevant parameters were as listed: ion spray voltage, 3500 V; nozzle voltage, 1000 V; fragmentor voltage, 125 V; drying gas ow, 11 L/min; capillary temperature, 300 °C, drying gas temperature, 320 °C; and nebulizer pressure, 40 psi. A full scan range was set at 50 to 1200 (m/z). The reference masses were 119.0363 and 980.0164.
The acquisition rate was 2 spectra/s. Data processing was performed with Agilent Pro nder B.08.00 (Agilent technologies). The mass tolerance was set to +/-15 ppm and RT tolerance was +/-0.2 min.
Genome-wide gene expression microarray analysis. Microarray services were provided by the UPENN Molecular Pro ling Facility, including quality control tests of the total RNA samples by Agilent Bioanalyzer and Nanodrop spectrophotometry. All protocols were conducted as described in the Affymetrix WT Plus Reagent Kit Manual and the Affymetrix GeneChip Expression Analysis Technical Manual. Brie y, 250 ng of total RNA was converted to the rst-strand cDNA using reverse transcriptase primed by poly(T) and random oligomers that incorporated the T7 promoter sequence. Second-strand cDNA synthesis was followed by in vitro transcription with T7 RNA polymerase for linear ampli cation of each transcript, and the resulting cRNA was converted to cDNA, fragmented, assessed by Bioanalyzer, and biotinylated by terminal transferase end labeling. Five and a half micrograms of labeled cDNA were added to Affymetrix hybridization cocktails, heated at 99ºC for 5 min and hybridized for 16 h at 45ºC to Clariom D Mouse Arrays using the GeneChip Hybridization oven 645. The microarrays were then washed at low (6X SSPE) and high (100 mM MES, 0.1M NaCl) stringency and stained with streptavidinphycoerythrin. Fluorescence was ampli ed by adding biotinylated anti-streptavidin and an additional aliquot of streptavidin-phycoerythrin stain. A GeneChip 3000 7G scanner was used to collect uorescence signal. Affymetrix Command Console and Expression Console were used to quantitate expression levels for targeted genes; default values provided by Affymetrix were applied to all analysis parameters. The GEO accession number is GSE159020.
Single-cell RNA-sequencing. Cells were loaded into a 10X Genomics Chromium Single-Cell controller following manufacturer's instructions using the 10 × 3' RNA-Seq V2 kit. Illumina sequencing libraries were prepared then sequenced either on three lanes of a HiSeq 4000 (28 bp x 98 bp) or a NovaSeq 6000 (28 bp x 91 bp). Samples were sequenced to a median depth of 14,188 +/-932.6 reads per cell with a median 2,291 +/-334 median gene count detected per cell. The fraction of reads mapping con dently to the transcriptome was 60% +/-4.6%. The percent of reads from mitochondrial genes had a median of 6 +/-0.41%. Initial data processing was performed with Cell Ranger V3.0.1. The GEO accession number is GSE159996.
Mouse necropsy. Mouse necropsy was performed according to the Comparative Pathology Core's standardized approach for rodent studies. Subsequently, patients in each cancer type were divided into two groups according to COL1A1 expression: low COL1A1 expression (below 1st quartile) and high COL1A1 expression (above 3rd quartile). Survival analysis using Kaplan-Meier and the log-rank test between COL1A1-low and -high groups, was performed using R package survival and OncoLnc 61 . Correlation analysis and Kaplan-Meier plots were produced using R package ggplot2.
In vivo mouse studies. All animal experiments have been approved by the University Laboratory Animal Resources and Institutional Animal Care and Use Committee of the University of Pennsylvania regulations. Both males and females were used in in vivo experimental procedures. Mice were housed in pathogen-free conditions. ATF4 excision was achieved by oral gavage of tamoxifen (200 mg/kg BW) for 5 consecutive days. For tumor growth studies, 5 × 10 5 B16F10 or MH6419 cells were injected into the anks of 9-10-week-old Atf4 wt/wt and Atf4 Δ/Δ mice. For co-injection studies, 5 × 10 4 B16F10 mixed with 1.5 × 10 5 broblast (Atf4 wt/wt or Atf4 Δ/Δ ) cells were injected into the anks of 9-10-week-old Atf4 wt/wt and Atf4 Δ/Δ mice. For sc-RNA-seq studies, 5 × 10 5 B16F10 or MH6419 cells were injected into the anks of 9-10-week-old Atf4 wt/wt and Atf4 Δ/Δ mice. For lung colonization studies, 1.5 × 10 5 B16F10 cells were injected in the tail vein of Atf4 wt/wt and Atf4 Δ/Δ mice. For, lung metastasis studies, 5 × 10 5 B16F10 were injected into the anks of 9-10-week-old Atf4 wt/wt and Atf4 Δ/Δ mice. Tumors reached appr. 300 mm 3 were surgically removed and 4 weeks later mice were euthanized the lungs were harvested and stored in 10% formalin. The primers for genotyping used in this study are described in Supplementary Table 6. Also, the mouse models used in this study are described in Supplementary Table 7. Patient samples. Malignant melanoma with normal skin tissue array and pancreas cancer tissue array with adjacent normal pancreas tissue were purchased from US Biomax, Inc. Formalin-xed para nembedded human melanoma tumors were obtained from patients resected at the Hospital of the University of Pennsylvania upon signing the informed consent in accordance with the IRB protocol No. Table 8     High ATF4 levels or ATF4-dependent gene expression correlate with increased COL1 expression or deposition in human tumors. a, "Fibroblast ATF4 target signature" comprised of the 30 most downregulated genes in LFBD/D compared to LFBwt/wt cells. b, Pearson correlation between the broblast ATF4 target signature and COL1A1, ACTA2, PDGFRb and FAP in Skin Cutaneous Melanoma (SKCM) and Pancreatic adenocarcinoma (PAAD). The linear regression lines along with 95% con dence intervals (shaded regions) are shown. c, Pearson correlation between 30 randomly chosen genes and COL1A1 and PDGFRb in SKCM. The linear regression lines along with 95% con dence intervals (shaded regions) are shown. d, Representative images from human melanoma tissue arrays stained for COL1 and ATF4 proteins (Up: high expression; Down: low expression). Scale bars, 100mm. e, Pearson correlation between the % ATF4 area and % COL1 area. f, Kaplan-Meier plot of survival time of SKCM patients with high (n=151) or low (n=151) COL1A1 expression. Log-rank (Mantel-Cox) test. g, Working model for host ATF4's role in tumor progression and metastasis. ATF4 is essential for the CAF activation via direct regulation of Col1a1 expression and by impacting multiple additional steps in the collagen synthesis pathway, including Glycine (Gly) and Proline (Pro) pools. The resulting abrogation of Collagen I (and potentially additional collagen isoforms) in ATF4-de cient FBs leads in dramatic reduction in secreted extracellular matrix collagen, which in turn results in defective CAF activation and reduced levels of angiogenic cytokine signaling to endothelial cells. The resulting defective angiogenesis leads to reduced support for primary and metastatic tumor growth.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. ExtendedDatalegends.docx ExtendedDataFigures17.pdf SupplementaryTables1232020.docx