Identification of Functional Heterogeneity of Carcinoma-Associated Fibroblasts with Distinct IL6-Mediated Therapy Resistance in Pancreatic Cancer

Fibroblasts accumulate in tumors with a putative capacity to regulate pancreatic ductal adenocarcinoma (PDAC) progression (1–3). Collectively, these cells are referred to as carcinoma-associated fibroblasts (CAF). CAFs may contribute to the emergence and progression of PDAC and response to treatment (3–7). The focus on CAFs in pancreatic cancer has been driven by the fact that, in some patients, stromal cells (including fibroblasts) can outnumber the cancer cells (8–11). CAFs are also referred to in the literature as myofibroblasts due to their expression of alpha smooth muscle actin (αSMA; refs. 12, 13). The biology of CAFs in PDAC is still evolving, with increased recognition of their role in shaping the tumor immune microenvironment (1, 3, 14, 15). Moreover, it is likely that CAFs in mouse and human PDAC are a heterogeneous population and have been tentatively classified into inflammatory, myofibroblastic/extracellular matrix–producing, and antigen-presenting CAFs by several different research groups (16–21). However, their precise function in PDAC progression remains unknown.

A subtype of CAFs has been identified in PDAC, defined by expression of fibroblast activation protein (FAP), with an ability to promote PDAC tumor growth, possibly via CXCL12 (SDF1; ref. 22) and CCL2 signaling (23). Recent studies have suggested that targeting FAP⁺ CAFs may lead to suppression of tumor growth (24–26). Loss of FAP protein in stromal cells delays PDAC disease progression in mice (24). Alternatively, αSMA⁺ CAFs have been shown to function in restraining tumors in genetically engineered mouse models (GEMM) of PDAC and help polarize tumor-infiltrating T cells (27). A debate continues as to whether FAP⁺ CAFs and αSMA⁺ CAFs are the same population with model-specific data interpretations from different laboratories (15, 22, 25, 28–30). Therefore, the question that remains unanswered to date is whether FAP and αSMA identify the same CAF population or subsets of fibroblasts with distinct function(s) in PDAC biology and therapy. We conducted an unbiased study to identify different CAFs, with a specific focus to determine the function of FAP⁺ CAFs and αSMA⁺ CAFs in PDAC, using multiparametric analysis coupled with new genetic mouse models. Our results unravel the functional heterogeneity of CAFs within the tumor microenvironment (TME), with distinct roles in PDAC progression and therapy response.

To explore whether FAP⁺ CAFs and αSMA⁺ CAFs are distinct subsets of CAFs, we performed single-cell RNA sequencing (scRNA-seq) on Pdx1^Cre/+; LSL-Kras^G12D/+; Trp53^R172H/+; LSL-YFP (KPCY; see nomenclature of GEMMs in Supplementary Table S1) tumors and identified different cell types present in the TME, including cancer cells, immune cells, endothelial cells, pericytes, and CAFs (Fig. 1A; Supplementary Fig. S1A; Supplementary Tables S2 and S3). The CAF cluster was characterized by expression of mesenchymal genes (Col1a1, Col1a2, and Dcn) and lack of expression of epithelial genes (Krt8, Krt18, and Krt19), and correlated with previously identified subsets based on a similar transcriptome analysis (Supplementary Fig. S1A and S1B). To further appreciate the precise association of FAP⁺ CAFs and αSMA⁺ (encoded by Acta2) CAFs within the CAF clusters, a deeper computational analysis of the CAF clusters was performed to reveal six subsets with distinct transcriptional profiles, termed aCAF, bCAF, cCAF, dCAF, eCAF, and fCAF (Fig. 1A; Supplementary Fig. S1C; Supplementary Table S4). To further investigate potential changes in CAF clusters during PDAC progression, we compared early-stage PDAC with late-stage PDAC in KPC mice. Early-stage PDAC presented with primarily normal and pancreatic intraepithelial neoplasia (PanIN) histology, whereas late-stage PDAC presented with more advanced disease histology (Supplementary Fig. S2A). The distribution of CAF subsets shifted during disease progression, with aCAFs as the most abundant population in early-stage PDAC and cCAFs as the most abundant population in late-stage PDAC, respectively (Fig. 1A).

We evaluated the expression of well-established fibroblast-associated genes within the six CAF subsets. Several genes (Col1a1, Col1a2, Pdgfra, Vim, and Pdpn) were expressed in all six subsets of CAFs, whereas other genes (Tagln, Fap, S100a4, Acta2, and Pdgfrb) were restricted to specific subsets of CAFs (Supplementary Fig. S2B). To ensure that our analysis was not overtly influenced by the presence of pericytes, we assessed the CAF cluster for common pericyte markers (Mcam and Cspg4) and found minimal expression of pericyte marker genes, as also reported by others (refs. 16, 18, 20; Supplementary Fig. S2C). The CAF cluster was also separated into three subpopulations previously termed as inflammatory CAF (iCAF), myofibroblastic CAF (myCAF), and antigen-presenting CAF (apCAF; refs. 16, 18, 20; Supplementary Fig. S3A and S3B). Clustering of CAFs into iCAF, myCAF, and apCAF was associated with differences in subpopulation composition based on PDAC stage, with iCAF being the most abundant subset in early-stage PDAC and myCAF as the most abundant subset in late-stage PDAC (Supplementary Fig. S3A and S3B). This analysis further confirms that clustering CAFs into six subclusters provides more in-depth appreciation of the changing landscape of CAF subsets with PDAC progression (Fig. 1A).

Next, we specifically determined the localization of Fap (encoding FAP) and Acta2 (encoding αSMA) within the CAF subsets and found that whereas Fap- and Acta2-expressing cells largely localized to the cCAF subset (Fig. 1B), more detailed analysis of this subset revealed minimal overlap of FAP⁺ CAFs and αSMA⁺ CAFs within cCAFs (Fig. 1C–E). The cCAF cluster was further subdivided into cCAF1 (defined by expression of Sfrp1, Serpina3n, Chl1, Mmp2, and Igfbp4), cCAF2 (defined by expression of Igfbp7, Serpine2, Col8a1, Id3, and Cdkn2a), cCAF3 (defined by expression of Acta2, Tagln, Lgals1, Tmsb4x, and Rpl41), and cCAF4 (defined by expression of Ero1l, Egln3, Slc2a1, Pgk1, and Bnip3; Fig. 1C; Supplementary Table S5). Fap was predominantly identified in the cCAF2 subcluster and displayed minimal overlap with Acta2 within the cCAF cluster (Fig. 1D and E). Acta2 was predominantly associated with cCAF3 with some association also observed with cCAF2 and cCAF4 (Fig. 1D and E). The existence of Fap⁺ CAFs and Acta2⁺ CAFs as distinct subsets of cells was further confirmed using a murine PDAC scRNA-seq data set (CAFs enriched by sorting for CD45⁻/EpCAM⁻/CD31⁻ cells; ref. 16; Supplementary Fig. S3C and S3D). This provides the confidence that our findings identifying such separation of CAFs subsets can be validated in multiple data sets. Next, we analyzed the scRNA-seq database with 23 human PDAC tumor CAF samples (31), wherein CAFs were identified based on the expression of the mesenchymal genes Col1a1, Col1a2, Dcn, and Pdpn. Human CAFs from 23 human PDAC tumors also present with FAP⁺ CAFs and ACTA2⁺ CAFs as distinct subsets of cells (Fig. 1F).

Based on our scRNA-seq analyses showing the minimal overlap between FAP and αSMA expression in CAFs of human and mouse PDAC, we next examined treatment-naïve human PDAC tissue microarray (TMA) samples (n = 136) for FAP and αSMA protein expression. The results validated a minimal overlap between FAP and αSMA expression, independent of tumor stage or differentiation status (Fig. 2A–C; Supplementary Table S6). Moreover, αSMA positivity correlated with significantly increased overall survival of patients, in contrast to FAP positivity, which was associated with significantly decreased overall survival (Fig. 2D). Furthermore, the ratio of αSMA/FAP demonstrated a significant prognostic value, showing that patients with high αSMA expression and low FAP expression exhibited significantly better overall survival (Fig. 2D). Consistent results were observed on pancreatic tumors from PDAC mouse models (Fig. 2E). These results further suggest that FAP⁺ CAFs and αSMA⁺ CAFs are distinct CAF subsets with potentially distinct functions in PDAC.

Guided by the nonoverlapping expression and the clinical data associated with FAP and αSMA in PDAC CAFs, we generated FAP-thymidine kinase (TK) transgenic mice and crossed them with KTC PDAC GEMMs to disable the accumulation of FAP⁺ CAFs. These mice were studied alongside KTC;αSMA-TK mice (27), which prevent the accumulation of αSMA⁺ CAFs in PDAC (Supplementary Table S1). Specific depletion of FAP⁺ CAFs resulted in significant suppression of PDAC tumor progression and a significant increase in overall survival of mice (Fig. 3A–C; Supplementary Fig. S4A and S4B). In contrast, depletion of αSMA⁺ CAFs resulted in a more aggressive PDAC tumor with decreased overall survival of mice (ref. 27; Fig. 3A–C; Supplementary Fig. S4A and S4B). Successful depletion of FAP⁺ CAFs or αSMA⁺ CAFs was confirmed by IHC analyses of PDAC tumors (Fig. 3B and C). The FAP-TK and αSMA-TK transgenes disable the accumulation of proliferating FAP-expressing and αSMA-expressing cells with similar efficiency upon ganciclovir (GCV) administration, respectively (Fig. 3B and C). Importantly, depletion of FAP⁺ CAFs did not change the number of αSMA⁺ CAFs, and depletion of αSMA⁺ CAFs did not alter the number of FAP⁺ CAFs in the tumors (Fig. 3C). This further affirms our previous (vide supra) assessment that FAP⁺ CAFs and αSMA⁺ CAFs are distinct cell subsets in the TME with opposing functions.

Moreover, scRNA-seq analysis revealed that Pcna expression levels were consistent across the CAF subsets (Supplementary Fig. S4C). The percentage of Fap⁺ and Acta2⁺ CAFs expressing Pcna was also similar (Supplementary Fig. S4C), indicating that these cell types exhibit comparable proliferative indexes. The percentage of Fap⁺ and Acta2⁺ cancer cells was 2.3% and 0.9%, respectively, suggesting that FAP-TK and αSMA-TK transgenes have minimal impact on cancer cells (Supplementary Fig. S4D). Desmin expression (a dominant pericyte marker) did not significantly change in FAP⁺ CAF–depleted tumors or αSMA⁺ CAF–depleted tumors when compared with their respective controls (Supplementary Fig. S4E). Altogether, these data indicate that pericytes are not depleted in these TK model systems, and the observed phenotypes are due to the function(s) of CAFs and not due to perivascular or cancer cells.

Histopathologic analysis of tumors from FAP⁺ CAF–depleted mice revealed increased prevalence of normal tissue, whereas tumors from αSMA⁺ CAF–depleted mice displayed a more aggressive tumor phenotype (Fig. 3C; Supplementary Fig. S4A and S4B). The phenotypes observed upon depletion of FAP⁺ CAFs and αSMA⁺ CAFs were confirmed using alternative models of PDAC. A similar phenotype, together with decreased overall survival of mice, was observed when αSMA⁺ CAFs were depleted in Ptf1a^cre/+; LSL-Kras^G12D/+; Trp53^R172H/+ (KPC^p48) mice, a different genetic mouse model of PDAC (Supplementary Fig. S4F–S4H). Improved histologic scoring was also observed when FAP⁺ CAFs were depleted in the context of orthotopically implanted KPC-derived cell lines when compared with control mice (Supplementary Fig. S4I). These findings further support the specific targeting of FAP⁺ CAFs without any significant impact on cancer cells. Previous studies implicated FAP depletion in the development of a cachexic phenotype in mice (32, 33), albeit with a distinct genetic strategy. Our genetic targeting strategy, limited to only actively proliferating cells, was not associated with body weight loss or muscle wasting over time in non–tumor-bearing adult mice (Supplementary Fig. S5A and S5B). Loss in FAP⁺ cells in the spleen of healthy (tumor-free) FAP-TK mice was also not observed (Supplementary Fig. S5C).

To decipher the mechanistic underpinning for the opposing functions of FAP⁺ CAFs and αSMA⁺ CAFs in PDAC progression, we performed global transcriptomic analyses of control, FAP⁺ CAF–depleted KTC tumors, and αSMA⁺ CAF–depleted KTC tumors. Comparative analyses of gene expression following the depletion of FAP⁺ CAFs or αSMA⁺ CAFs revealed minimal overlap of both downregulated and upregulated genes (Fig. 4A and B). To ascertain whether such distinct transcriptomic profiles were associated with specific biological processes, we evaluated the enriched pathways in FAP⁺ CAF–depleted tumors and αSMA⁺ CAF–depleted tumors. Interestingly, FAP⁺ CAF depletion was associated with upregulation of largely distinct pathways related to protein processing, proteolysis, fibrinogen and blood coagulation, cell junctions, endopeptidase inhibitor activity, and pancreatic secretion, potentially reflecting the improved histology in FAP⁺ CAF–depleted tumors (Fig. 4C). In αSMA⁺ CAF–depleted tumors, gene expression changes were associated with pathways related to epithelial migration, cell proliferation, cytokine production, inflammatory responses, as well as T and B cell–related immunity (Fig. 4D).

As several pathways related to immune cells and inflammatory signaling were upregulated in αSMA⁺ CAF–depleted tumors (Fig. 4C), we performed CIBERSORT analysis to evaluate the abundance of immune cell subsets. Macrophages and B cells were decreased in FAP⁺ CAF–depleted tumors, whereas an increase in macrophages and a decrease in dendritic cells were observed in αSMA⁺ CAF–depleted tumors (Fig. 4E). Gene set enrichment analysis (GSEA) of differentially expressed pathways between the control and depleted groups also revealed that certain pathways (such as immune response pathway and adaptive immune system pathway) were oppositely regulated by FAP⁺ CAF depletion and αSMA⁺ CAF depletion, with downregulation in FAP⁺ CAF–depleted tumors and upregulation in αSMA⁺ CAF–depleted tumors (Supplementary Fig. S5D).

To further probe the tumor immune microenvironment of FAP⁺ CAF–depleted or αSMA⁺ CAF–depleted KTC tumors, multiplex IHC analysis for immune markers was conducted (Fig. 4F). Depletion of αSMA⁺ CAFs reduced the effector T cell (Teff) to regulatory T cell (Treg) ratio (Fig. 4F). FAP⁺ CAF depletion had no significant influence on T cells (Fig. 4F). FAP⁺ CAF depletion, in contrast to αSMA⁺ CAF depletion, was associated with decreased CD11b⁺ (myeloid) cell infiltration (Supplementary Fig. S5E). Depletion of FAP⁺ or αSMA⁺ CAFs was not associated with changes in CD8⁺ cells (Fig. 4F). These results support the notion that opposing functions of FAP⁺ CAFs or αSMA⁺ CAFs are, at least in part, likely due to differential polarization of the tumor immune microenvironment.

To further identify the role of αSMA⁺ CAFs and FAP⁺ CAFs in PDAC progression and response to therapy, the secretome of the CAFs was evaluated, keeping in mind the sum functional contribution of TP-CAFs and TR-CAFs in the PDAC immune microenvironment (Fig. 4E and F). The immune modulatory cytokine IL6 is a critical mediator of polarization of immune cells (34) and has been implicated in PDAC cancer progression and response to chemotherapy (35–38).

Previous studies identified Il6 as a putative marker gene of iCAFs (16, 18); however, we observed that Il6 was also expressed in myCAFs, at both early-stage PDAC and late-stage PDAC in the KPC mice (Supplementary Fig. S3A and S3B). Il6 transcripts are enriched in both cCAFs and eCAFs of early- and late-stage PDAC tumors (Fig. 5A and B; Supplementary Fig. S5F), with the former composed of both Fap⁺ and Acta2⁺ CAFs. Further analysis of the cCAF cluster revealed that Il6 is enriched in cCAF1, 3, and 4, albeit with some expression also detected in cCAF2 (Fig. 5C). Within total CAF and the cCAF subcluster, both Acta2⁺ and Fap⁺ CAFs express Il6 (Fig. 5B and C; Supplementary Fig. S5F). This observation was confirmed in a published data set of murine PDAC (Supplementary Fig. S5G; ref. 16).

To investigate the functional contribution of CAF-derived IL6 to PDAC progression and therapy response, we generated GEMMs in which two distinct gene recombination systems (Flippase- and Cre-mediated recombinations; Supplementary Table S1) independently drive cancer formation (Pdx1-Flp;FSF-Kras^G12D/+;Trp53^frt/frt; KPPF; Supplementary Fig. S6A and S6B) and allow conditional gene recombination in αSMA⁺ CAFs (αSMA-Cre; floxed-gene of interest) or FAP⁺ CAFs (FAP-Cre; floxed-gene of interest; ref. 39). KPPF mice presented with a similar disease progression to the comparable Cre-driven KPPC model (Pdx1-Cre;LSL-Kras^G12D/+;Trp53^loxP/loxP; Supplementary Fig. S6C and S6D). Potential for recombination events in the CAFs in KPPF;αSMA-Cre;R26^Confetti reporter mice was visualized by capture of GFP, RFP, YFP, and CFP fluorescent cells in the desmoplastic stroma associated with PDAC (Supplementary Fig. S6E).

First, KPPF;αSMA-Cre;R26^Dual reporter mice were bred to a conditional IL6 (Il6) gene knockout allele (KPPF;IL6^smaKO;R26^Dual), effectively abrogating IL6 transcription in αSMA⁺ CAFs (Fig. 5D–F; Supplementary Table S1). Using KPPF;αSMA-Cre;R26^Dual reporter mice, we confirmed enrichment of IL6 transcripts in purified tdTomato⁺ fibroblasts and noted higher IL6 expression in αSMA⁺ CAFs than cancer cells, which was largely ablated in the αSMA⁺ CAFs isolated from the KPPF;IL6^smaKO;R26^Dual mice (Fig. 5F). Genomic recombination events captured by PCR reactions in tumors and control cells and organs also confirmed the specificity of the genetic strategy used to ensure conditional deletion of IL6 in αSMA⁺ cells from the KPPF;IL6^smaKO;R26^Dual mice (Supplementary Fig. S6F).

To examine the production of IL6 by cells other than αSMA⁺ CAFs, the KPPF mice were also bred with systemic (whole body) IL6-deleted mice (KPPF; IL6^−/−). PDAC progression, overall survival, and PDAC histology were similar in both the KPPF;IL6^smaKO and KPPF; IL6^−/− mice when compared with the KPPF control (Fig. 5G and H; Supplementary Fig. S7A–S7C). Il6 expression was significantly decreased in tumors of KPPF;IL6^smaKO mice compared with KPPF control tumors and absent in KPPF;IL6^−/− tumors (Supplementary Fig. S7D). Il1b transcript (encoding IL1β) abundance was unchanged (Supplementary Fig. S7D). Collectively, these findings support that IL6 is likely produced by multiple cell types in PDAC, including αSMA⁺ CAFs and FAP⁺ CAFs, but its impact on cancer progression (treatment-naïve mice harboring p53 loss) is not rate limiting (36).

IL6 signaling is reported to confer prosurvival signals to cancer cells via the JAK/STAT signaling pathway in the context of chemotherapy (35, 36). We next investigated whether CAF-derived IL6 might have a functional contribution in the context of chemotherapy response. IL6 was genetically deleted in FAP⁺ CAFs also for this assessment (Supplementary Table S1; Supplementary Fig. S7E). In contrast with loss of IL6 in FAP⁺ CAFs (KPPF;IL6^fapKO mice), KPPF;IL6^smaKO mice (with loss of IL6 in αSMA⁺ CAFs) were associated with a significant increase in the overall survival of mice upon treatment with gemcitabine (Fig. 5I) and identical to KPPF;IL6^−/− mice (Fig. 5I). Loss of IL6 from αSMA⁺ CAFs resulted in improved histopathology and reduced tumor burden in the context of gemcitabine treatment (Fig. 5J and K). Analysis of the TCGA human pancreatic cancer cohort further revealed that the ACTA2/IL6-low expression subgroup exhibited less progressive disease in response to primary therapy than the ACTA2/IL6-high expression subgroup (Supplementary Fig. S7F). Minimal impact on lung and liver metastasis was observed in KPPF;IL6^smaKO mice compared with KPPF mice treated with gemcitabine (Supplementary Fig. S8A). Gemcitabine treatment increased the number of putative IL6-producing αSMA⁺ CAFs in KPPF mice (Fig. 6A), further supporting their role in therapy resistance. Gemcitabine-enhanced αSMA⁺ CAFs abundance was also reversed in KPPF;IL6^smaKO mice (Fig. 6A).

Cancer cells in the tumors of KPPF mice treated with gemcitabine showed elevated levels of phosphorylated Stat3, ERK1/2, and Akt, which was attenuated in KPPF;IL6^smaKO mice treated with gemcitabine (Fig. 6B–D). Gemcitabine treatment in KPPF;IL6^smaKO mice did not significantly affect tumor collagen deposition or vasculature when compared with controls (Supplementary Fig. S8B and S8C). Gemcitabine treatment or IL6 genetic depletion did not significantly alter the presence of FSP1⁺ or FAP⁺ CAFs (Supplementary Fig. S8D and S8E). In contrast, cleaved caspase-3, indicative of apoptosis, increased upon gemcitabine treatment in KPPF;IL6^smaKO mice treated with gemcitabine (Fig. 6E). Collectively, these results suggest that αSMA⁺ CAF–derived IL6 confers tumor resistance to gemcitabine by promoting cancer cell survival, whereas FAP⁺ CAF–derived IL6 exerts an insignificant impact in the response to gemcitabine (Fig. 5I).

Given the well-documented role of IL6 on the tumor immune microenvironment (34), we evaluated the immune composition of tumor, spleen, and peripheral blood of KPPF, KPPF;IL6^−/−, and KPPF;IL6^smaKO mice, with and without gemcitabine therapy (Supplementary Fig. S9A and S9B). Intratumoral immune cell frequencies were affected by loss of IL6, whereas immune frequencies in the spleen and peripheral blood were largely unchanged (Fig. 6F; Supplementary Fig. S10A–S10C). The number of Tregs and Teffs significantly changed in the tumors of both KPPF;IL6^−/− and KPPF;IL6^smaKO mice (Fig. 6F). Further, the frequency of CD11b⁺PD-L1⁺ cells was significantly reduced in KPPF;IL6^−/− and KPPF;IL6^smaKO mice compared with control KPPF mice (Fig. 6F). Gemcitabine treatment with IL6 loss resulted in increased frequency of CD11c⁺ cells (Fig. 6F). These observed alterations in the immune microenvironment suggested that tumors with loss of IL6 may be sensitive to immune-checkpoint blockade in combination with gemcitabine. Therefore, we tested the therapeutic benefit of dual immune-checkpoint blockade using anti-CTLA4 and anti–PD-1 antibodies [anti-CTLA4 and anti–PD-1 (αCP)]. Although αCP did not reveal efficacy in KPPF or KPPF;IL6^−/− mice, it synergized with gemcitabine to reveal efficacy with a significant increase in the overall survival of KPPF;IL6^−/− mice (Fig. 6G). The immune profiling of the tumor suggested that gemcitabine treatment results in an increase in putative CD11c⁺ dendritic cells (Fig. 6F) with the likely potential for enhanced tumor antigen presentation for an improved Teff response upon αCP and depletion of IL6 (Fig. 6G). These results further emphasize the contribution of IL6 in PDAC therapy response.

Several elegant studies have identified the potential diversity of CAFs in pancreatic cancer (15–17, 19, 20, 31). Although specific markers for different classes of CAFs have not been identified, multiple studies using scRNA-seq have shown that one can classify CAFs into different transcriptionally defined clusters (16–19, 31). IHC studies and mouse models fluorescently labeling mesenchymal cell populations have also suggested that CAFs can express diverse markers that do not always overlap (40–43). In this regard, FAP⁺ CAFs and αSMA⁺ CAFs have been identified as being part of the same scRNA-seq cluster and purported to exhibit similar functions; however, other studies showed that an additional subset of CAFs (Meflin⁺ CAFs) gives rise to increasing numbers of αSMA⁺ CAFs but not FAP⁺ CAFs during tumor progression (44).

Although recent studies in mice and clinical trial results have demonstrated that suppression of CAFs can lead to acceleration of pancreatic cancer in some contexts (27, 44–47), the question remains whether all CAFs are targeted by such strategies or a subset of CAFs are manipulated, resulting in more aggressive PDAC. This becomes an important unaddressed question because other studies have suggested that targeting FAP⁺ CAFs results in control of PDAC (24), although therapeutic efficacy data from an ongoing PDAC clinical trial are currently unavailable (ClinicalTrials.gov identifier: NCT03932565). Therefore, using scRNA-seq, multiplex immunostaining, and new genetic mouse models, our studies focused on identifying whether FAP⁺ CAFs and αSMA⁺ CAFs possess similar biological and functional relevance in PDAC or display diversity with respect to their biology and actions.

We identify that FAP⁺ CAFs and αSMA⁺ CAFs are distinct populations of fibroblasts in both human PDAC samples and transgenic mouse models. The αSMA⁺ CAFs predominantly act to restrain PDAC (TR-CAFs), and the net function of FAP⁺ CAFs is to promote PDAC (TP-CAFs; Fig. 6H). These data were supported by computational analyses of scRNA-seq data. FAP⁺ CAFs and αSMA⁺ CAFs present in the same scRNA-seq subclusters of CAFs (cCAF) due to some commonalities in their transcriptomes; however, they are distinct populations of cells with largely distinct, nonoverlapping transcriptomes. We also show that FAP⁺ CAFs and αSMA⁺ CAFs regulate different cancer-associated transcriptomic networks and define the tumor immune composition in distinct manners. Suppression of αSMA⁺ CAFs leads to a decreased Teff/Treg ratio, potentially contributing to accelerated tumor growth. Deletion of type I collagen in αSMA⁺ CAFs was associated with decreased abundance as well as impaired activation of T cells (21), suggesting that the observed changes in T cells upon depletion of αSMA⁺ CAFs is in part mediated by reduced collagen I. In contrast, the FAP⁺ CAFs have an impact on inflammation, and their suppression leads to a decrease in CD11b⁺ myeloid cells and an inhibition of PDAC with increased overall survival of mice. Macrophages have been reported to be tumor-promoting in PDAC (48–50), suggesting that alterations in macrophages may contribute to increased survival of FAP⁺ CAF–depleted mice.

Previous studies showed that the overall functional contribution of αSMA⁺ CAFs is tumor-restraining (20, 27, 47). The tumor-restraining contribution of αSMA⁺ CAFs is in part mediated by their production of type I collagen (21). Our studies identify that IL6 produced by αSMA⁺ CAFs does not contribute to progression of PDAC despite depletion of about 50% of the total IL6 present in the TME. To address the impact of total IL6 in the TME on PDAC progression, we also crossed IL6 whole body knockout mice (IL6^−/−) with KPPF mice. Complete lack of IL6 in the TME of PDAC also has no impact on PDAC progression. These data clearly indicate that IL6 does not play a role in PDAC progression; therefore, it is not surprising that IL6 deleted from the subset of cells in the TME also does not reveal a functional importance. The total loss of IL6 from the TME or just from the αSMA⁺ CAFs has an impact only when gemcitabine is provided to suppress cancer cell proliferation and tumor growth. This shows that the optimal response to chemotherapy is compromised by the presence of IL6 in the TME, likely by providing cancer cell survival benefit. This aspect of drug resistance is an applied role of IL6 but not its natural role in the context of PDAC progression. Therefore, the sum totality of the actions of αSMA⁺ CAFs is “tumor-restraining.” When we begin to audit one by one the specific proteins produced by αSMA⁺ CAFs using dual-recombinase genetic mouse models, we begin to unravel their specific role in PDAC progression. Specific deletion of type I collagen from αSMA⁺ CAFs revealed its tumor-restraining action (21) but without an impact on chemotherapy resistance (27). Driven by many published reports and scRNA-seq data, we show that deletion of IL6 from αSMA⁺ CAFs does not alter the progression of PDAC and therefore does not contribute to the “tumor-restraining” properties of αSMA⁺ CAFs. In contrast, IL6 produced by αSMA⁺ CAFs has an “applied” function in providing protection against chemotherapy-induced inhibition of tumor growth, resulting in perceived therapy resistance. In addition, the abundance of αSMA⁺ CAFs was significantly increased by gemcitabine treatment, presumably leading to further increased IL6 production by αSMA⁺ CAFs. Our study reveals the complexity of CAF biology in PDAC, specifically the different CAF roles in natural progression of the disease and how their function affects therapeutic intervention.

Collectively, our studies show that targeting FAP⁺ CAFs emerges as a viable strategy to achieve inhibition of PDAC. Such a strategy must contemplate sparing the αSMA⁺ CAFs, and this study suggests that this may be feasible, since they are minimally overlapping in their respective protein biomarker presentation. We show that depletion of FAP⁺ CAFs does not affect αSMA⁺ CAF content in PDAC. In this regard, FAP-directed chimeric antigen receptor T-cell therapy might be an effective strategy, as previously demonstrated (51). Targeting chemokines that are selectively secreted by FAP⁺ CAFs might also be a viable strategy (22). Importantly, the strategy to inhibit FAP⁺ CAFs must consider prevention of bone marrow toxicities and cachexia (32, 33, 51). Strategies to increase the number of αSMA⁺ CAFs can represent another strategy to control PDAC (20). In this regard, activating the vitamin D receptor on αSMA⁺ CAFs could program them to further control PDAC (52).

Our studies indicate that αSMA⁺ CAF–derived IL6 confers chemoresistance and negatively regulates T cells in the TME. In this regard, organoid-based studies revealed a subset of CAFs displaying an immunomodulatory phenotype that included IL6 production (18). Although αSMA⁺ CAFs as a whole emerged as tumor-restraining, in the context of gemcitabine therapy stress, cancer cells likely utilize the IL6 produced by αSMA⁺ CAFs to promote their survival through activation of STAT3 signaling. Previous studies reported that high CAF abundance was associated with increased MAPK and STAT3 cosignaling in proliferative and invasive cancer cells (53). In mice treated with gemcitabine, deletion of IL6 in αSMA⁺ CAFs led to a reduction in phospho-STAT3 and phospho-ERK1/2 levels in cancer cells, indicating potential paracrine signaling between αSMA⁺ CAFs and cancer cells to promote cancer cell survival. Alternatively, Ly6C^hi monocytes have been reported to express IL6R (54); thus, it is possible that αSMA⁺ CAF–derived IL6 acts indirectly on cancer cells to promote chemoresistance.

It is conceivable that IL6 production by αSMA⁺ CAFs is for self-preservation purposes in nontreatment conditions but is also utilized by cancer cells for induction of survival signaling pathways in the context of gemcitabine resistance. Previous studies have reported that IL6 signaling confers prosurvival signals to cancer cells via the JAK/STAT signaling pathway in the context of chemotherapy (35, 36). We also observed a significant polarization of the PDAC immune microenvironment upon deletion of αSMA⁺ CAF–produced IL6; however, such changes in Teff and Treg frequencies did not affect PDAC progression. The immune-checkpoint blockade therapy was ineffective when combined with gemcitabine in KPPF mice; however, a combinatorial benefit was observed with inhibition of IL6 signaling, indicating that αSMA⁺ CAF–produced IL6 may be a critical suppressor of immune-checkpoint blockade therapy in PDAC. The cell death, generation of neoantigens, and increased CD11c⁺ cells induced by gemcitabine and the suppression of IL6 likely favor the emergence of Teffs to augment the efficacy of immune-checkpoint blockade. A trial targeting IL6R with tocilizumab in conjunction with nab-paclitaxel and gemcitabine is ongoing (NCT02767557); however, our data suggest that addition of checkpoint blockade to this treatment regimen has the potential to improve therapeutic response.

In summary, this study demonstrates that CAFs are not uniform in their biology and exhibit functional diversity with therapeutic implications for pancreatic cancer.

All acronyms designating specific GEMMs are listed in Supplementary Table S1. LSL-Kras^G12D/+ (55), Pdx1-Cre (55), Ptf1a-Cre (55), αSMA-TK (56), αSMA-RFP (57), FAP-TK (58), Rosa26-loxP-Stop-loxP-YFP (59), Rosa26-CAG-Brainbow 2.1 (60), LSL-Trp53^R172H/+ (55), FSF-Kras^G12D/+ (61), Pdx1-Flp (61), Trp53^frt/+ (62), αSMA-Cre (56), CMV-Cre (63), IL6^loxP/loxP (64), Trp53^loxP/loxP (65), and Tgfbr2^loxP/loxP (66) mouse strains were previously documented. The FAP-Cre transgenic strain was newly generated. A 5-kb sequence flanking the Fap promoter and partial Exon 1 was cloned into the pORF-HSV1-TK vector (InvivoGen) using Not I and Age I. The sequence-confirmed FAP-TK construct was released from the vector using Not I and Swa I before purification and injection into fertilized eggs. FAP-Cre was generated similarly, where the Cre sequence was inserted into the pORF-FAP-TK plasmid with Not I and BstE II digestion. The FAP-Cre plasmid was digested with Not I and Swa I, excised, and injected into fertilized eggs as described for FAP-TK. The transgenic mice were generated in the MD Anderson Genetically Engineered Mouse Facility on the C57Bl/6 genetic background. FAP-TK mice have been deposited at The Jackson Laboratory (stock 034655). These mice were bred onto PDAC GEMMs or implanted orthotopically with 689KPC cancer cells, as previously described (67). The KPPF and R26^Dual mice of the dual-recombinase system were kindly provided by D. Saur (Technische Universität München). Mice were maintained on a mixed genetic background, and both male and female mice were evaluated.

GCV (sud-gcv, InvivoGen) was administered i.p. daily at 50 mg/kg of body weight (approximately 1.5 mg per 25 g mouse). GCV was administered to KTC mice at 21 to 37 days of age, and to KPC^p48 mice at 50 to 51 days of age. Control groups were TK-negative mice that received GCV or phosphate-buffered saline (PBS), or were not injected. In the orthotopic tumor model (689KPC), GCV was administered 15 days following tumor implantation, and mice were euthanized at 40 days following tumor implantation. All mice were housed under standard housing conditions at the MD Anderson Cancer Center (MDACC) animal facility, and all animal procedures were approved by the MDACC Institutional Animal Care and Use Committee. Investigators were not blinded to group allocation but were blinded for the histologic assessment of phenotypic outcome with no randomization method used. The experimental endpoint was defined when the animals developed significant signs of illness leading to their death or requiring euthanasia. Survival curves for KTC mice were plotted based on the number of days after start of GCV or PBS treatment, and KPC^p48 survival curves were plotted based on the entire lifespan of the mice with GCV or PBS treatment start time indicated. For therapeutic treatments, mice were given gemcitabine (G-4177, LC Laboratories) i.p. twice per week at 40 mg/kg of body weight. Anti-CTLA4 (9H10; Bio X Cell) and anti–PD-1 (RMP1-14; Bio X Cell) antibodies were i.p. administered at 200 μg/mouse twice per week.

For analysis of immune cell populations in spontaneous tumors of indicated transgenic mouse strains, the staining and flow cytometry procedures were conducted as previously described (21). For analysis of cells from KPC tumors, the tumors were minced and digested in collagenase IV (4 mg/mL) and dispase (4 mg/mL) in DMEM for 1 hour at 37°C. Digestion was stopped by the addition of FBS to neutralize the protease activity, followed by washing with FACS buffer 3 times. Digested tissues were then filtered through a 70-μm mesh followed by a 40-μm mesh, centrifuged, and incubated in ACK lysis buffer for 3 minutes at room temperature. Prior to staining, spleen was minced and filtered through a 40-μm mesh and was incubated in ACK lysis buffer for 3 minutes at room temperature. FAP and its corresponding isotype antibody were conjugated with the Zenon Alexa Fluor 647 Rabbit IgG Labeling Kit according to the manufacturer's instructions. Samples were stained with antibody and fixable viability dye eFluor 780 in FACS buffer for 30 minutes on ice, followed by washing prior to analysis on a BD LSRFortessa X20. For sorting experiments, samples were analyzed and sorted on a BD FACSAria. Unstained and single-stained samples were used for compensation controls. Data analysis was performed in FlowJo software (TreeStar, Inc). Details on the antibodies, sources, and dilution are listed in Supplementary Table S7.

PDAC samples with less than 10% pancreatic adenocarcinoma areas were defined as early-stage PDAC, whereas PDAC samples with greater than 50% pancreatic adenocarcinoma areas were defined as late-stage PDAC. The tumors of KPC mice were processed to obtain single-cell suspensions (see “Flow Cytometry” section). scRNA-seq on these samples was conducted at the MDACC Advanced Technology Genomics Core. Single-cell Gel Bead-In-Emulsions (GEM) generation and barcoding, post–GEM-RT cleanup and cDNA amplification, library construction and Illumina-ready sequencing library generation were prepared by following the manufacturer's guidelines. A High-Sensitivity dsDNA Qubit Kit was used to estimate the cDNA and library concentration. An HS DNA Bioanalyzer was used for the quantification of cDNA. A DNA 1000 Bioanalyzer was used for the quantification of libraries. The “c-loupe” files were generated by using Cell Ranger software pipelines following the manufacturer's guidelines. Cells from unfractionated tumor were encapsulated using 10X Genomics’ Chromium Controller and Single-Cell 3′ Reagent Kits v2. Following capture and lysis, cDNA was synthesized and amplified to construct Illumina sequencing libraries. The libraries from about 1,000 cells per sample were sequenced with Illumina NextSeq 500. The run format was 26 cycles for read 1, 8 cycles index 1, and 124 cycles for read 2. scRNA-seq data were processed by the Advanced Technology Genomics Core at MDACC.

Cell clustering was performed using the Seurat R software as previously described (68, 69). Specifically, cell clustering was conducted using the nonlinear dimensional reduction technique by the Uniform Manifold Approximation and Projection (UMAP) algorithm. To identify marker genes of cell clusters, we compared each of the cell clusters using pairwise differential expression analysis with settings recommended for data with batch effect, with average fold change (FC) expression compared with other included clusters >2. The clustering of CAFs was performed using an analogous gene signature used in recently published data sets (16).

Library Seurat (version 3.6.1), dplyr, and cowplot were loaded into R to explore quality control metrics, filter cells, normalize data, cluster cells, and identify cluster biomarkers. To filter out low-quality cells, a threshold with a minimum of 200 and a maximum of 4,000 to 7,000 genes per cell was used. Cells with more than 10% of the mitochondrial genome were also removed for further analysis. To remove the influence of technical characteristics from downstream analyses, the “sctransform” package was used for normalization. The “RunUMAP” function was used for clustering the cells. The “FindAllMarkers” function was used to identify the specific markers for each cell cluster. The “DoHeatmap” function was used to show the top genes in each cluster. The “VlnPlot” function was used to show expression probability distributions across cell clusters of the genes we selected to assign the cell-type identity and the genes that we were interested in. The Markov affinity-based graph imputation of cells (MAGIC) algorithm, which utilizes the nearest-neighbor graphing and a diffusion operator to restore or “smooth” missing transcripts from the single-cell expression data based on the expression of similar cells, was used (70). MAGIC smoothing of the cancer cell cluster was performed using the library Matrix and library Rmagic (70) based on a pooled gene signature of CAFs.

Mouse tissues were fixed in 10% neutral buffered formalin, embedded in paraffin, and sectioned at 5-μm thickness. Sections were processed for hematoxylin and eosin (H&E) staining, Masson's trichrome staining using Gomori's Trichome Stain Kit (38016SS2, Leica Biosystems), or picrosirius red (Direct Red 80, Sigma-Aldrich) according to the manufacturer's instructions. Histopathologic assessments were conducted in a blinded fashion by scoring H&E-stained sections for relative percentages of the listed histopathologic phenotypes. A weighted histology score was then applied to the percentages as follows: for tumors with less than 5% normal tissue 2 points, else 0 points; greater than 30% PanIN or ADM tissue 2 points, else 0 points; cancer area greater than 30% 4 points, else 0 points; poorly differentiated PDAC area greater than 30% 5 points, else 0 points; and necrosis area greater than 5% 6 points, else 0 points. The weighted scores were then summed for each animal to be interpreted as a higher value meaning worse histopathology. Images were obtained with a Leica DM 1000 LED microscope using a 20× objective and an MC120 HD Microscope Camera with Las V4.4 Software (Leica). Tumor scores for orthotopic tumors were evaluated based on H&E sections of the entire pancreas and attributed a score on a scale from 1 (minor involvement) to 4 (extensive involvement).

All antibodies, sources, and dilutions are listed in Supplementary Table S7. Formalin-fixed, paraffin-embedded (FFPE) sections were processed for IHC staining as previously described (71) after citrate-based antigen retrieval (pH = 6). Staining for αSMA was performed with the M.O.M. Kit (Vector Laboratories) following the manufacturer's instructions. For all stainings, counterstaining with hematoxylin was performed and DAB positivity was examined in 10 visual fields at 200× magnification. For αSMA IHC on mouse tumor sections, immunoreactive scores were obtained from the sum of distribution and intensity scores for each section, established on a scale of 1 to 4 (72). FAP IHC was performed and imaged identically as the FAP staining in the multiplex staining panel (see below). Pseudocolored images with the FAP-520 channel being colored brown and the DAPI channel colored blue on a white background were used for the scoring. The stromal region was scored on a scale of 0 to 3 for the density of FAP⁺ fibroblasts in each image. The scores were averaged for each mouse and presented.

All human pancreatic tumor sections (n = 136) were fixed on TMA slides that contain three representative 1-mm cores for each patient (two representative cores of tumor and one core of matched benign pancreatic tissue). The TMAs were constructed from FFPE blocks of archived PDAC specimens as previously described (73). This study was approved by the Institutional Review Board of MDACC (IRB LAB05-0854) and in accordance with the U.S. Common Rule. Written informed consent was obtained from all patients. The patients included in this cohort received no neoadjuvant therapy. Cases and clinical information were retrieved from the surgical pathology files of the Department of Pathology, MDACC (Supplementary Table S6). Immunofluorescence images were obtained using the LSM800 confocal laser scanning microscope under 100× magnification and analyzed by ZEN software (Zeiss). The quantification of indicated staining was based on the average reading of the two representative tumor cores for each patient.

The multiplex staining procedures, spectral unmixing, and cell segmentation using the Nuance and inForm imaging softwares were previously published (74). Antibody concentrations used for the multiplex staining can be found in Supplementary Table S7. Multiplex stained slides were imaged with the Vectra Multispectral Imaging System, using Vectra software version 3.0.3 (PerkinElmer). Each tissue section was scanned in its entirety using a 4× objective, and up to 50 cancer regions (at 20×) were selected for multispectral imaging using the Phenochart software (PerkinElmer). Each multiplex field was scanned every 20 nm of the emission light spectrum across the range of each emission filter cube. Filter cubes used for multispectral imaging were DAPI (440–600 nm), FITC (520–680 nm), Cy3 (570–690 nm), Texas Red (580–700 nm), and Cy5 (680–720 nm). Multispectral images from single-marker stained slides with the corresponding fluorophores were used to generate a spectral library using the Nuance Image Analysis software (PerkinElmer). The library contained the emitting spectral peaks of all fluorophores and was used to unmix each multispectral image (spectral unmixing) to its individual eight components using the inForm 2.4 image analysis software. All spectrally unmixed image cubes were subsequently segmented into individual cells based on the nuclear DAPI counterstain. For the immune cell population analysis, all spectrally unmixed and segmented images were analyzed using the inForm phenotyping algorithm. This allows for the individual identification of each DAPI-stained cell according to the pattern of fluorophore expression and nuclear/cellular morphologic features. Cells were phenotyped into eight different classes according to the markers they expressed: CD3⁺ T cells (CD3⁺), Teffs (CD3⁺CD4⁺FoxP3⁻), cytotoxic T cells (CD3⁺CD8⁺), Tregs (CD3⁺CD4⁺FoxP3⁺), myeloid cells (CD11b⁺), cancer cells (CK8⁺), and other cells (negative for all markers). For fibroblast cell population analysis, FAP staining was performed after antigen retrieval with TE buffer (pH = 9), and αSMA and CK8 staining was performed after antigen retrieval with citrate buffer (pH = 6.0). Slides were imaged with a Zeiss Axio Scan.Z1 and LSM800 confocal microscope. Percent positive area for a given fibroblast marker and negative for CK8 was quantified in ImageJ.

Tissues from the R26^Dual lineage tracing mouse strain, expressing Pdx1-Flp–driven intrinsic EGFP and αSMA-Cre–driven tdTomato, were fixed in 4% paraformaldehyde overnight at 4°C and equilibrated in 30% sucrose overnight at 4°C. Tissues were then embedded in O.C.T. compound (TissueTek) and processed for 5-μm-thick frozen sections. Sections were blocked for 1 hour with 4% cold water fish gelatin (Aurion) and immunostained overnight at 4°C with anti-αSMA antibody (followed by goat anti-mouse Alexa Fluor 647 secondary antibody). Slides were then mounted with Vectashield Mounting Medium (Vector Laboratories) to a glass coverslip and visualized under the LSM800 confocal laser scanning microscope and ZEN software (Zeiss).

Establishment of primary PDAC cancer cell and myofibroblast lines was conducted as previously described with minor modifications (39, 71). Fresh PDAC tissues from KPPF;IL6^smaKO;R26^Dual mice and KPPF;αSMA-Cre;R26^Dual mice were minced, digested with collagenase IV (17104019, Gibco, 4 mg/mL) and dispase II (17105041, Gibco, 4 mg/mL) in RPMI at 37°C for 1 hour, filtered by 70-μm cell strainers, resuspended in RPMI with 20% FBS, and then seeded on type I collagen–coated dishes (354401, Corning). Cells were cultured in RPMI medium containing 20% FBS and 1% penicillin, streptomycin, and amphotericin B (PSA) antibiotic mixture. Pdx1 lineage–traced cancer cells and αSMA lineage–traced myofibroblasts were further purified by FACS (BD FACSAria II) based on EGFP and tdTomato signals, respectively. The sorted cells were subsequently maintained in vitro. All studies were performed on cells cultivated fewer than 25 passages. DNA from these primary cell lines was extracted using the DNA Mini Kit (51304, Qiagen). Total RNA was extracted from the indicated cells using the Direct-zol RNA Kit (Zymo Research), processed for cDNA synthesis using the Reverse Transcription Kit (Applied Biosystems), and subjected to qRT-PCR using SYBR Green Master Mix (Applied Biosystems). Primer sequences used were as follows: Il6 Forward: 5′-GCTTAATTACACATGTTCTCTGGGAAA-3′; Il6 Reverse: 5′-CAAGTGCATCATCGTTGTTCATAC-3′; Il1b Forward: 5′-GGGCTGCTTCCAAACCTTTG-3′; Il1b Reverse: 5′-TGATACTGCCTGCCTGAAGCTC-3′; Gapdh Forward: 5′-AGGTCGGTGTGAACGGATTTG-3′; Gapdh Reverse: 5′-TGTAGACCATGTAGTTGAGGTCA-3′.

Total RNA was also isolated from tumors of age-matched KTC;αSMA-TK, and KTC;FAP-TK mice (n = 3 mice per in each group), which were administrated with GCV or PBS. RNA extraction was carried out using the Qiagen RNeasy Mini Kit and submitted to the Microarray Core Facility at MDACC. Gene expression analysis was performed using Affymetrix MTA 1.0 Genechip. The Limma package (75) from R Bioconductor was used for quantile normalization of expression arrays and to analyze differential gene expression between the TK groups (KTC;αSMA-TK and the KTC-FAP-TK groups) and their respective control groups (KTC-αSMA-TK^control and KTC;FAP-TK^control; P ≤ 0.05 and FC ≥1.2). Analyses of differentially expressed pathways between the TK and control groups were performed using GSEA (76). For CIBERSORT analysis, R package biomaRt (version 2.50.3) was used to convert mouse genes into human gene symbols and CIBERSORT with built-in LM22 gene signatures was used for deconvolution analysis (77).

The statistical tests used for the comparative analyses presented are listed in the figure legends. Statistical analyses were carried out using GraphPad Prism (GraphPad Software version 8). Kaplan–Meier plots were used for survival analysis, and the log-rank Mantel–Cox test was used to evaluate statistical differences with GraphPad Prism. Error bars represent standard error of the mean unless specified in the figure legends. Statistical significance was defined as P < 0.05.

Source data for each figure are included. Microarray data from KTC;αSMA-TK GEMMs were previously deposited at Gene Expression Omnibus (GEO) under accession number GSE52812 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE52812; ref. 27). Gene expression microarray data from KTC;FAP-TK GEMMs were also deposited in GEO (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE120577). scRNA-seq data were deposited in GEO under accession number GSE198815 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE198815).

K.M. McAndrews reports other support from Stellanova Therapeutics outside the submitted work. Y. Chen reports personal fees from Stellanova Therapeutics outside the submitted work. J.K. Darpolor reports grants from the NIH, the American Legion Auxiliary, and the NCI and a John J. Kopchick Fellowship during the conduct of the study. J.L. Carstens reports grants from the NIH and the Cancer Prevention & Research Institute of Texas during the conduct of the study. H. Wang reports grants from the NIH during the conduct of the study. P. Correa de Sampaio reports personal fees from Deep Science Ventures and Triumvira Immunologics outside the submitted work. V.S. LeBleu reports personal fees and other support from Stellanova Therapeutics during the conduct of the study. No disclosures were reported by the other authors.

K.M. McAndrews: Data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. Y. Chen: Data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. J.K. Darpolor: Data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. X. Zheng: Data curation, investigation. S. Yang: Data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. J.L. Carstens: Data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. B. Li: Software, formal analysis, investigation, methodology, writing–review and editing. H. Wang: Supervision, investigation. T. Miyake: Resources. P. Correa de Sampaio: Data curation, investigation, writing–review and editing. M.L. Kirtley: Resources. M. Natale: Investigation. C.-C. Wu: Data curation, formal analysis. H. Sugimoto: Investigation. V.S. LeBleu: Data curation, formal analysis, supervision, investigation, visualization, writing–original draft. R. Kalluri: Conceptualization, resources, supervision, funding acquisition, writing–original draft.

We thank Erika J. Thompson, David P. Pollock, and Yunxin Chen for their assistance with scRNA-seq, Tapsi Kumar and Nicholas Navin for their assistance with analysis of human scRNA-seq data, Judith Kaye for mouse husbandry support, and Ehsan Ehsanipour for assistance with generation and validation of the FAP-Cre mouse. This work was primarily supported by the Cancer Prevention & Research Institute of Texas (CPRIT). Research in the Kalluri laboratory is also supported by NCI P01CA117969 and CPRIT Award RP150231. R. Kalluri is a Distinguished University Chair supported by the Sid W. Richardson Foundation. J.K. Darpolor was supported by the American Legion Auxiliary Fellowship and the Center for Clinical and Translational Sciences TL1 training grant (NIH UL1TR000371). K.M. McAndrews and Y. Chen were supported by Ergon Foundation Postdoctoral Fellowships. Other support includes The University of Texas MD Anderson Cancer Center (UT MDACC) Flow Cytometry Core (NIH P30CA016672), the UT MDACC Advanced Technology Genomics Core (NIH P30CA016672), and the UT MDACC Genetically Engineered Mouse Facility (NIH P30CA016672).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.