A subset of cancer-associated fibroblasts (FAP+/CAF-S1) mediates immunosuppression in breast cancers, but its heterogeneity and its impact on immunotherapy response remain unknown. Here, we identify 8 CAF-S1 clusters by analyzing more than 19,000 single CAF-S1 fibroblasts from breast cancer. We validate the five most abundant clusters by flow cytometry and in silico analyses in other cancer types, highlighting their relevance. Myofibroblasts from clusters 0 and 3, characterized by extracellular matrix proteins and TGFβ signaling, respectively, are indicative of primary resistance to immunotherapies. Cluster 0/ecm-myCAF upregulates PD-1 and CTLA4 protein levels in regulatory T lymphocytes (Tregs), which, in turn, increases CAF-S1 cluster 3/TGFβ-myCAF cellular content. Thus, our study highlights a positive feedback loop between specific CAF-S1 clusters and Tregs and uncovers their role in immunotherapy resistance.
Our work provides a significant advance in characterizing and understanding FAP+ CAF in cancer. We reached a high resolution at single-cell level, which enabled us to identify specific clusters associated with immunosuppression and immunotherapy resistance. Identification of cluster-specific signatures paves the way for therapeutic options in combination with immunotherapies.
This article is highlighted in the In This Issue feature, p. 1241
Cancer-associated fibroblasts (CAF) represent one of the most abundant components in adenocarcinomas and have key protumorigenic functions (1–6). It is now recognized that CAFs are heterogenous and that distinct CAF subsets can be defined on the basis of expression of specific markers (7–17). Recently, studies based on immunocompetent mouse models reported that CAFs expressing the fibroblast activation protein (FAP) marker are associated with an immunosuppressive environment (9, 18–24). Moreover, the concomitant study of six markers, including FAP, smooth-muscle α actin (SMA), and integrin β1 (CD29), revealed the existence of four CAF subsets, referred to as CAF-S1 to S4, in human breast and ovarian cancers (8, 10). Although CAF-S2 (FAPneg CD29lo SMAneg) and CAF-S3 (FAPneg CD29med SMAneg) fibroblasts are also detected in healthy tissues and could be reminiscent of normal fibroblasts, CAF-S1 (FAPhi CD29med-hi SMAhi) and CAF-S4 (FAPneg SMAhi CD29hi) myofibroblast cells are restricted to cancer and metastatic lymph nodes. CAF-S1 fibroblasts are defined by extracellular matrix and inflammation signatures, and CAF-S4 fibroblasts are characterized by a perivascular signature (8, 10). Both CAF-S1 and CAF-S4 promote metastases through complementary mechanisms (17). In contrast, whereas CAF-S1 promotes immunosuppression in human cancer, CAF-S4 does not (8). Specifically, CAF-S1 fibroblasts stimulate immunosuppression by increasing attraction, survival, and overall content of CD4+ CD25+ FOXP3+ regulatory T cells (Tregs) in the tumor microenvironment (8, 10). In line with these observations, FAPhi CAFs have also been suspected to contribute to primary resistance to immunotherapies (7, 9, 25). However, to our knowledge, the direct role of FAPhi CAFs in immunotherapy resistance has not yet been addressed in human cancer. This observation, coupled with the recent finding that FAPhi CAF-S1 fibroblasts exert their immunosuppressive function through a multistep mechanism (8, 10), prompted us to study their heterogeneity and specific roles in primary resistance to immunotherapy in patients with cancer.
In recent years, single-cell RNA sequencing (scRNA-seq) has been developed to analyze intratumoral heterogeneity. We used this new approach to address the heterogeneity of the FAPhi CAF-S1 immunosuppressive subpopulation. We performed scRNA-seq of more than 19,000 CAF-S1 fibroblasts isolated from eight primary breast cancers and identified eight different CAF-S1 clusters. Among them, three CAF-S1 clusters (1, 2, 5) belong to the inflammatory (“iCAF”) subgroup and five CAF-S1 clusters (0, 3, 4, 6, 7) to the myofibroblastic (“myCAF”) subgroup, iCAF and myCAF having been previously identified in pancreatic cancer (7, 16, 26). The eight CAF-S1 clusters identified here are characterized by high expression of genes coding extracellular matrix (ECM) proteins (cluster 0), detoxification pathway (cluster 1), IL signaling (cluster 2), TGFβ signaling pathway (cluster 3), wound healing (cluster 4), IFNγ (cluster 5), IFNαβ (cluster 6), and acto-myosin pathway (cluster 7). Accordingly, we annotated them as follows: ecm-myCAF (cluster 0), detox-iCAF (cluster 1), IL-iCAF (cluster 2), TGFβ-myCAF (cluster 3), wound-myCAF (cluster 4), IFNγ-iCAF (cluster 5), IFNαβ-myCAF (cluster 6), and acto-myCAF (cluster 7). We confirmed the existence and the relative proportions of the five most abundant CAF-S1 clusters (representing up to 91% of sequenced cells) in breast cancer by multicolor flow cytometry using a combination of specific markers. Moreover, we validated the existence of these five CAF-S1 cellular clusters in head and neck squamous cell carcinoma (HNSCC) and in non–small cell lung cancer (NSCLC) by analyzing publicly available scRNA-seq data, demonstrating the relevance of these CAF-S1 clusters across different cancer types. In addition, we uncovered that the abundance of two CAF-S1 clusters of the myCAF subgroup, namely ecm-myCAF and TGFβ-myCAF, is significantly correlated with an immunosuppressive environment, whereas the content in detox-iCAF and IL-iCAF is not. Indeed, ecm-myCAF and TGFβ-myCAF clusters are enriched in tumors with high levels of PD-1+, CTLA4+, and TIGIT+ CD4+ T lymphocytes (themselves enriched in Tregs), and low fraction of CD8+ T lymphocytes. Importantly, we found that, at diagnosis, these CAF-S1 clusters are associated with primary resistance to immunotherapies in patients with melanoma and NSCLC. In agreement with these findings, ecm-myCAF increases the fraction of FOXP3hi T cells and stimulates both PD-1 and CTLA4 protein levels at the surface of CD4+ CD25+ T lymphocytes, which in turn increases the proportion of TGFβ-myCAF. Thus, our study uncovers a positive feedback loop between the immunosuppressive ecm-myCAF and TGFβ-myCAF CAF-S1 clusters and Tregs that could participate in immunotherapy resistance.
Distinct Cellular Clusters Are Identified within the Immunosuppressive CAF-S1 Subset by Single-Cell Approach
We used scRNA-seq to investigate cellular heterogeneity within the CAF-S1 immunosuppressive subset. We isolated CAF-S1 fibroblasts from human breast cancer (see description of prospective cohort 1 in Supplementary Table S1) by FACS as described previously (8, 10, 17). In brief, from freshly resected tumors, we first excluded debris, dead cells, doublets, and epithelial (EPCAM+), hematopoietic (CD45+), endothelial (CD31+), and red blood (CD235a+) cells (Supplementary Fig. S1A). We considered EPCAM− CD45− CD31− CD235a− as the fraction of cells enriched in fibroblasts and next performed FAP and CD29 staining (Supplementary Fig. S1A), which enabled us to distinguish CAF-S1 (FAPhi CD29med-hi) from the other CAF subpopulations (CAF-S2: FAPneg CD29lo; CAF-S3: FAPneg CD29med; CAF-S4: FAPneg CD29hi; Supplementary Fig. S1A), as previously established in refs. 8 and 10. We then performed scRNA-seq of 18,805 CAF-S1 fibroblasts from 7 patients with breast cancer prior to any treatment (Fig. 1). After quality control, 18,296 CAF-S1 fibroblasts with a median of 2,428 genes detected per cell were conserved for further analyses. Unsupervised graph-based clustering identified eight CAF-S1 clusters, visualized with the Uniform Manifold Approximation and Projection (UMAP) algorithm (Fig. 1A). All clusters were found in most patients from luminal (Lum) and triple-negative (TN) breast cancer subtypes, albeit at varying levels (Fig. 1B and C). No individual cluster was associated with a particular phase of cell cycle or with high proliferation (Supplementary Fig. S1B), as shown using G1–S and G2–M gene signatures (27). We confirmed the detection of these different CAF-S1 cellular clusters using the Label Transfer algorithm (28). Indeed, this algorithm successfully transferred with high prediction scores all the eight cluster labels in an independent CAF-S1 scRNA-seq dataset, newly generated from an eighth patient with breast cancer (Supplementary Fig. S1C).
Differential gene expression analyses revealed that each cluster was characterized by a specific transcriptional profile (Supplementary Data S1). Cluster 0 was associated with ECM remodeling, cell-substrate adhesion, and collagen formation; cluster 1 with detoxification and inflammatory response; cluster 2 with response to growth factor, TNF signaling, and IL pathway; cluster 3 with TGFβ signaling pathway and matrisome; cluster 4 with assembly of collagen fibrils and wound healing; cluster 5 with response to IFNγ and cytokine-mediated signaling pathway; cluster 6 with IFNα/β signaling; and cluster 7 with acto-myosin complex (Supplementary Data S1). As examples, we found high expression of LRRC15, a marker recently identified in CAFs from pancreatic cancer (29), and GBJ2 in cluster 0; ADH1B and GPX3 in cluster 1; RGMA and SCARA5 in cluster 2; CST1 and TGFβ1 in cluster 3; SEMA3C and SFRP4 in cluster 4; CCL19 and CCL5 in cluster 5; IFIT3 and IRF7 in cluster 6; and GGH and PLP2 in cluster 7 (Fig. 1D). Interestingly, in human breast cancer, we were able to distinguish the myofibroblastic (“myCAF”) and inflammatory (“iCAF”) fibroblast subgroups (Fig. 1E and F), previously identified in FAP+ fibroblasts from pancreatic cancer (7, 16, 26). CAF-S1 clusters 1, 2, and 5 were identified as iCAF and clusters 0, 3, 4, 6, and 7 as myCAF (Fig. 1E and F). Consistent with data from pancreatic cancer, iCAF showed high expression of chemokines and proinflammatory molecules such as CXCL12 and SOD2 (Fig. 1E), whereas myCAF expressed myofibroblast markers, including COL1A2 and TAGLN (Fig. 1F). In addition, we observed that the iCAF cluster 5 expressed high levels of CD74, encoding MHC II invariant chain (Fig. 1G). CD74 was recently shown to be specifically expressed in the antigen-presenting CAFs (“apCAF”) in pancreatic cancer (16), suggesting that the CAF-S1 cluster 5 might be reminiscent of such apCAFs (Fig. 1G). To summarize, we identified eight different CAF-S1 clusters in breast cancer. Clusters 1, 2, and 5 belong to the iCAF subgroup, with cluster 5 that might correspond to the apCAF cluster, whereas clusters 0, 3, 4, 6, and 7 belong to the myCAF subgroup. Moreover, iCAF clusters are characterized by detoxification (cluster 1), response to stimuli (cluster 2), IFNγ and cytokines (cluster 5); and myCAF clusters by ECM (cluster 0), TGFβ (cluster 3), wound healing (cluster 4), IFNα/β (cluster 6), and acto-myosin (cluster 7). Accordingly, we proposed the following nomenclature for these different FAPhi CAF-S1 clusters: cluster 0, ecm-myCAF; cluster 1, detox-iCAF; cluster 2, IL-iCAF; cluster 3, TGFβ-myCAF; cluster 4, wound-myCAF; cluster 5, IFNγ-iCAF; cluster 6, IFNαβ-myCAF; and cluster 7, acto-myCAF.
Finally, we wondered whether these CAF-S1 clusters accumulate differentially in the different breast cancer subtypes. Because we performed analysis on patients prior to any treatment, the fresh samples collected for scRNA-seq were mostly collected from Lum patients, the HER2 and TN breast cancer patients being preferentially treated in neoadjuvant settings. Consequently, there was no HER2 patient and only 2 patients with TN breast cancer in our prospective cohort 1 (Supplementary Table S1). Still, in this dataset, we could detect that patients with TN breast cancer exhibited higher proportions of iCAF clusters than patients with LumA breast cancer, which accumulated more myCAF clusters (in LumA: iCAF = 43.4%, myCAF = 56.6%; in TN: iCAF = 57.1%, myCAF = 42.9%; P = 1.29e-64 from Fisher exact test). Because of the low number of TN breast cancers in our dataset, we sought to address this question by taking advantage of The Cancer Genome Atlas (TCGA) database, which contains RNA-sequencing (RNA-seq) data from a high number of patients with LumA and TN breast cancer. To this end, we first defined specific signatures of the five most abundant CAF-S1 clusters (that represented up to 91% of sequenced CAF-S1 cells) by identifying differentially expressed genes in each cluster compared with the others (Supplementary Fig. S1D). As we also used these signatures for detecting these clusters in melanoma, NSCLC, and HNSCC data (see below), we next discarded any gene of these signatures that was also expressed by melanoma, NSCLC, and HNSCC cancer cells to avoid any signal from cancer cells and to guarantee a signal strictly specific of CAF-S1 clusters (Supplementary Data S2 and Supplementary Fig. S1D for specificity of CAF-S1 cluster signatures). We thus assessed the differential of expression of CAF-S1 cluster–specific signatures between LumA and TN breast cancer subtypes from the TCGA RNA-seq database (https://portal.gdc.cancer.gov/). We confirmed the accumulation of iCAF clusters in TN and myCAF clusters in LumA breast cancer (Supplementary Fig. S1E). Specifically, detox-iCAF and IL-iCAF showed higher expression in TN compared with Lum breast cancer, whereas ecm-myCAF, TGFβ-myCAF, and wound-myCAF expression was higher in LumA breast cancer than in other breast cancer subtypes (Supplementary Fig. S1E); this increase in iCAF content in TN breast cancer was consistent with the reported presence of numerous tumor-infiltrating lymphocytes (TIL) in some TN breast cancers (8). In summary, using scRNA-seq from a large number of FAP+ CAF-S1 fibroblasts isolated from breast cancer, we detected eight clusters that exhibit distinct signatures and accumulated differentially in breast cancer subtypes.
CAF-S1 Cellular Clusters Are Validated by Multicolor Flow Cytometry in Breast Cancer
We next aimed at validating CAF-S1 clusters by using multicolor flow cytometry (FACS) on fresh breast cancer samples. By analyzing the percentage of each cluster among CAF-S1 defined by scRNA-seq, we first observed that the five first clusters represented up to 91% of total sequenced cells (Fig. 2A). We thus decided to focus our FACS analysis on these five most abundant clusters and sought to identify surface markers for each cluster. Using pairwise comparisons of CAF-S1 cluster expression profiles, we identified six surface markers with commercially available antibodies and designed a gating strategy to identify the five most abundant clusters (Fig. 2B; Supplementary Fig. S2A). We sought to validate the specificity of these six markers in an independent CAF-S1 dataset. To do so, we studied the CAF-S1 scRNA-seq data corresponding to the eighth patient, in which the cluster labels were transferred successfully by the Label Transfer algorithm (as shown Supplementary Fig. S1C). Indeed, the gating strategy relying on these six markers and based on 7 patients with breast cancer efficiently delineated the five most abundant CAF-S1 clusters in the independent dataset (Supplementary Fig. S2B). This way, we confirmed that these markers were specific of each CAF-S1 cluster. We thus analyzed fresh breast cancer samples by FACS by applying the following gating strategy: CAF-S1 fibroblasts (isolated as CD45−, EPCAM−, CD31−, CD235a−, FAPhi CD29med) were first separated on the basis of ANTXR1 protein level that distinguished myCAF (ANTXR1+) from iCAF (ANTXR1−) fibroblasts in breast cancer. ANTXR1+ (myCAF) CAF-S1 clusters 0 (ecm-myCAF), 3 (TGFβ-myCAF), and 4 (wound-myCAF) were next distinguished according to SDC1, LAMP5, and CD9 protein levels. ANTXR1+ SDC1+ LAMP5− were defined as cluster 0 (ecm-myCAF), ANTXR1+ LAMP5+ SDC1+/− as cluster 3 (TGFβ-myCAF), and ANTXR1+ SDC1− LAMP5− CD9+ as cluster 4 (wound-myCAF; Fig. 2C). ANTXR1− (iCAF) CAF-S1 clusters 1 (detox-iCAF), and 2 (IL-iCAF) were separated using GPC3 and DLK1 markers. ANTXR1− GPC3+ DLK1+/− were defined as cluster 1 (detox-iCAF); ANTXR1− GPC3− DLK1+ as cluster 2 (IL-iCAF; Fig. 2C). ANTXR1+ CAF-S1 cells that were negative for LAMP5, SDC1, and CD9 and ANTXR1− GPC3− DLK1− cells were pooled and referred to as “other cluster.” By applying this gating strategy on 44 fresh samples (prospective cohort 2, Supplementary Table S1), we validated the existence of these five most abundant clusters in breast cancer (Fig. 2D). The percentage of each cluster among CAF-S1 cells defined by FACS confirmed the single-cell results, including clear heterogeneity among CAF-S1 fibroblasts and ecm-myCAF as the most abundant population in the majority of patients (Fig. 2D). We next analyzed whether there was any correlation between the respective proportions of these five CAF-S1 clusters across patients (see correlation matrix Fig. 2E, and detailed correlation plots on the right). We detected that the relative abundances of ecm-myCAF and TGFβ-myCAF (both myCAF) were correlated together, as well as those of detox-iCAF and IL-iCAF (both iCAF; Fig. 2E, detailed correlation plots on the right). Conversely, the proportions of ecm-myCAF and TGFβ-myCAF were anticorrelated with the ones of detox-iCAF and IL-iCAF (Fig. 2E). In addition, the wound-myCAF was negatively correlated with detox-iCAF and IL-iCAF but also with ecm-myCAF (Fig. 2E, detailed correlation plots at bottom), suggesting that these different CAF-S1 clusters could accumulate differentially in breast cancer, but in a coordinated manner.
CAF-S1 Cellular Clusters Are Identified across Cancer Types
We next sought to test the existence of CAF-S1 cellular clusters in other cancer types. To do so, we analyzed publicly available scRNA-seq data from HNSCC (30) and NSCLC (31), because these two studies had isolated enough CAFs to investigate clusters. These published studies included 18 patients with HNSCC, of which 5 matched pairs of primary tumors and lymph nodes metastases, for a total of 5,902 total cells analyzed (30). Moreover, in the NSCLC cohort, more than 52,000 total cells were collected from 5 different patients (31). In these two studies, 1,422 cells and 1,465 cells were annotated as CAFs in HNSCC and NSCLC cohorts, respectively (30, 31). To pursue the analysis strictly on CAF-S1 fibroblasts, CAF-S1 were distinguished from CAF-S4 based on the expression of FAP and MCAM, two markers regulated at RNA levels in CAF-S1 (FAPhi MCAMlo) and CAF-S4 (FAPlo MCAMhi), respectively (ref. 8; Supplementary Fig. S2C). As a result, 603 CAF-S1 cells in HNSCC and 959 in NSCLC were further analyzed. We compared the similarity between CAF-S1 cells from distinct cancer types by mixing the referent (breast cancer) and target (HNSCC or NSCLC) datasets. Data integration was done using “anchor” correspondences across single cells from different datasets on the basis of the similarity of their expression profiles, as described in ref. 28 (Fig. 3). We used CAF-S1 cluster–specific signatures defined by differentially expressed genes in each cluster compared with the others (Supplementary Data S2; Supplementary Fig. S1D). Remarkably, we found systematic correspondences between CAF-S1 clusters from breast cancer and CAF-S1 clusters from either HNSCC (Fig. 3A) or NSCLC (Fig. 3B). Visualization of the clusters using specific signatures confirmed that we could detect the five most abundant clusters in HNSCC and NSCLC (Fig. 3A and B). Hence, we could confirm the existence of the five most abundant CAF-S1 clusters in HNSCC and NSCLC, highlighting their relevance in other cancers.
Immunosuppressive Environment Correlates with Specific CAF-S1 Clusters
As we showed that CAF-S1 fibroblasts exert immunosuppression in breast and ovarian cancers (8, 10), we next investigated whether this function could be exerted by all CAF-S1 clusters or restricted to specific ones (Fig. 4). We first tested whether we could detect correlations between the content of CAF-S1 clusters and immune cells. To this end, fresh breast cancer samples (prospective cohort 2, Supplementary Table S1) were characterized both in terms of CAF-S1 cluster content (as shown Fig. 2) and immune cell infiltration, including CD4+, CD8+ T lymphocytes and natural killer (NK) cells (Supplementary Fig. S3A–S3C). We tested associations between stromal and immune cells and analyzed variables exhibiting at least one significant correlation with another variable (Fig. 4A). The correlation matrix obtained by unsupervised hierarchical clustering highlighted that ecm-myCAF and TGFβ-myCAF are clustered together on the one side and detox-iCAF and IL-iCAF are clustered on the other, while the wound-myCAF cluster was quite isolated (Fig. 4A), suggesting that these different clusters interact differentially with T cells. Interestingly, we found that ecm-myCAF, TGFβ-myCAF and wound-myCAF showed specific associations with T lymphocytes (Fig. 4A). Indeed, we first observed that the proportion of ecm-myCAF among CAF-S1 fibroblasts was significantly correlated with CD45+ hematopoietic cell infiltration (Supplementary Fig. S3D). More specifically, the abundance of ecm-myCAF was correlated with infiltration of PD-1+, CTLA4+, and TIGIT+ CD4+ T lymphocytes, but anticorrelated with CD8+ T lymphocytes (Fig. 4A, right, red square). Likewise, although the content in TGFβ-myCAF did not show any global association with CD45+ hematopoietic cells, its abundance was positively correlated with infiltration by CTLA4+ CD4+ T lymphocytes and negatively correlated with CD8+ T lymphocytes (Fig. 4A, right, green square). Thus, these data indicate that the abundance of ecm-myCAF and TGFβ-myCAF is associated with an immunosuppressive environment enriched in Tregs. In contrast to ecm-myCAF and TGFβ-myCAF clusters, the abundance of detox-iCAF and IL-iCAF was correlated with CD8+ T-cell infiltration (Fig. 4A, right, yellow and green squares). The wound-myCAF cluster was globally correlated with T lymphocytes among CD45+ cells (Supplementary Fig. S3D) and anticorrelated with CTLA4+, TIGIT+, PD-1+, and NKG2A+ CD4+ T lymphocytes (Fig. 4A, bottom, blue square, and Supplementary Fig. S3D). The enrichment in wound-myCAF was also anticorrelated with CTLA4+ CD8+, TIGIT CD8+, CD244+ CD8+, CD244+ NK (Fig. 4A, bottom, blue square) markers of exhaustion, suggesting a global association of this cluster with high T-lymphocyte infiltration and immunoprotective environment.
To validate these data in an independent and large cohort of patients with breast cancer, we next tested the association between CAF-S1 clusters and T-cell signatures in the publicly available TCGA database. RNA-seq data from the TCGA database enabled us to confirm that the expression of ecm-myCAF and TGFβ-myCAF clusters was positively correlated with the one of FOXP3 (Fig. 4B), one of the main Treg markers. In addition, although wound-myCAF showed no real association with FOXP3, detox-iCAF and IL-iCAF clusters were negatively correlated with FOXP3 (Fig. 4B). In agreement with these data, we observed a positive correlation between T-cell cytolytic index, as defined in ref. 32, and detox-iCAF and IL-iCAF clusters, but not with ecm-myCAF, TGFβ-myCAF or wound-myCAF (Fig. 4C). To summarize, although detox-iCAF and IL-iCAF are correlated with an immunocompetent environment, both ecm-myCAF and TGFβ-myCAF are associated with an immunosuppressive environment, poor in CD8+ T lymphocytes and enriched in CD4+ T lymphocytes expressing high levels of immune checkpoints, including PD-1 and CTLA4.
Positive Feedback Loop between ecm-myCAF and TGFβ-myCAF with PD-1+ and CTLA4+ Tregs
As described above, we observed that abundance of ecm-myCAF and TGFβ-myCAF, but not detox-iCAF and IL-iCAF, is correlated with that of PD-1+ and/or CTLA4+ CD4+ T lymphocytes in breast cancer. We investigated the role of CAF-S1 clusters in generating a CD4+ CD25+–enriched immunosuppressive environment. We therefore established primary cultures of CAF-S1 clusters to perform in vitro functional assays. Although we did not achieve establishing every CAF-S1 cluster in culture, we succeeded at isolating ecm-myCAF and iCAF clusters by applying two distinct methods, that is, (i) by leaving CAF-S1 fibroblasts directly escaping and spreading from breast cancer samples seeded in plastic dishes, and (ii) by sorting FAPhi CD29med cells by FACS and expanding them in culture in plastic dishes. After expansion during a few weeks, we compared the identity of these different cells in the same culture conditions, that is, those compatible with coculture with CD4+ CD25+ T lymphocytes for performing functional assays (see below). We observed that CAF-S1 obtained by spreading expressed high levels of myCAF genes, whereas CAF-S1 isolated by sorting exhibited high expression of iCAF genes (Fig. 5A). We thus applied the cluster-specific signatures established from scRNA-seq data (Supplementary Data S2 and Supplementary Fig. S1D) and found that spread CAF-S1 fibroblasts were enriched in ecm-myCAF, whereas sorted CAF-S1 were enriched in detox-iCAF, IL-iCAF, and IFNγ-iCAF clusters (Fig. 5A; see also the Methods section “RNA Sequencing of CAF-S1 Primary Cell Lines Isolated from Breast Cancer”). Thus, although we did not obtain all pure CAF-S1 clusters in vitro, we succeeded in generating ecm-myCAF and iCAF using these two methods that enabled us to compare the respective properties of these clusters. We have previously demonstrated that the global CAF-S1 subpopulation has no direct effect on CD4+ CD25− T cells but increases the proportion of FOXP3+ Tregs among CD4+ CD25+ T lymphocytes (8, 10). As the content in ecm-myCAF and TGFβ-myCAF in breast cancer was associated with a CD4+-enriched immunosuppressive microenvironment whereas iCAF clusters were not, we compared the function of these myCAF and iCAF clusters on CD4+ CD25+ T cells using functional assays, as previously performed in refs. 8 and 10.
First, we tested the impact of myCAF and iCAF clusters on the content of CD4+ CD25+ FOXP3+ T lymphocytes in vitro (Fig. 5B). ecm-myCAF increased the percentage of FOXP3+ T cells among the CD4+ CD25+ population and enhanced the FOXP3 protein level in these T cells (Fig. 5B). In contrast, iCAF clusters had no impact on either the percentage of CD4+ CD25+ FOXP3+ T lymphocytes or on FOXP3 protein levels (Fig. 5B). We also tested the impact of culture on the identity and immunosuppressive activity of normal fibroblasts. Fibroblasts isolated upon spreading from healthy tissues were FAPneg-lo and devoid of immunosuppressive activity at early time point, but became FAPpos-hi and immunosuppressive at later passages (Supplementary Fig. S4A and S4B), suggesting that the long-term maintenance of CAFs on plastic dishes may activate them. On the basis of the ability of ecm-myCAF to increase CD4+ CD25+ FOXP3+ T lymphocytes, we next compared the capacity of CAF-S1 clusters to modulate the proportion of PD-1+, CTLA4+, TIGIT+, TIM3+, and LAG3+ on CD4+ CD25+ FOXP3+ T lymphocytes, considering both the percentage of positive cells and surface protein levels of these immune checkpoints (Fig. 5C–G). ecm-myCAF significantly increased both the percentage of PD-1+ and CTLA4+ CD4+ CD25+ FOXP3+ T lymphocytes and immune checkpoint levels at their surface (Fig. 5C and D). In contrast to ecm-myCAF, iCAF clusters affected neither the percentage of PD-1+ and CTLA4+ T cells nor CTLA4 protein levels (Fig. 5C and D). Furthermore, although iCAF clusters increased PD-1 protein levels, this effect was at lower efficiency than ecm-myCAF. Both myCAF and iCAF clusters increased the proportion of CD4+ CD25+ FOXP3+ TIGIT+ cells (Fig. 5E), whereas they had no effect on TIM3+ and LAG3+ T cells (Fig. 5F and G). Hence, in agreement with the correlations observed in breast cancer between ecm-myCAF and PD-1+ and CTLA4+ CD4+ T lymphocytes, we show here that CAF-S1 from ecm-myCAF have a direct function on Tregs by enhancing PD-1 and CTLA4 immune checkpoint levels at the surface of CD4+ CD25+ FOXP3+ T lymphocytes, whereas iCAF clusters have no or minimal effect on these cells. Finally, we observed that the upregulation of immune checkpoints at the surface of CD4+ CD25+ T cells upon coculture with ecm-myCAF was also detected intracellularly (Supplementary Fig. S4C), suggesting that ecm-myCAF increased the total protein levels in T cells. Moreover, we found that expression of FOXP3, CTLA4, and TIGIT in CD4+ CD25+ T cells was also upregulated at mRNA levels following coculture with ecm-myCAF (Supplementary Fig. S4D), PD-1 RNA being almost not detected in T cellsin vitro. Furthermore, we observed that mRNA levels of NFAT and STAT family members were also elevated in CD4+ CD25+ T lymphocytes upon coculture with ecm-myCAF (Supplementary Fig. S4E). As NFAT and STAT are well-known transcriptional regulators of immune checkpoints in T cells, these data indicate that ecm-myCAF promotes upregulation of immune checkpoints at RNA levels in CD4+ CD25+ T lymphocytes, potentially through the activation of NFAT/STAT signaling pathways.
Considering the impact of CAF-S1 clusters on Tregs, we next wondered whether T lymphocytes could in turn modulate CAF-S1 cluster identity. We thus evaluated whether coculturing CD4+ CD25+ T lymphocytes had any impact on the marker cluster levels at the surface of CAF-S1 fibroblasts (Fig. 5H). Upon coculture, to avoid any contamination by T cells, we isolated CAF-S1 by FACS and analyzed cluster markers expressed at their surface. We observed that the coculture of CD4+ CD25+ T cells significantly increased the expression of the TGFβ-myCAF–specific marker LAMP5 at the surface of ecm-myCAF fibroblasts (Fig. 5H), thereby suggesting that the content in TGFβ-myCAF increases upon coculture with CD4+ CD25+ T lymphocytes. This effect was detected only in myCAF and not in iCAF fibroblasts (Fig. 5H), as expected, considering that TGFβ-myCAF and ecm-myCAF CAF-S1 fibroblasts belong to the myCAF subgroup (Fig. 1E and F). Consistent with this observation, ANTXR1 protein level also showed a trend to increase (although without reaching significance) in ecm-myCAF fibroblasts upon coculture with CD4+ CD25+ T cells, whereas it remained strictly unchanged and low in iCAF cells (Fig. 5H). Quite surprisingly, DLK1, marker of IL-iCAF, also increased upon coculture, suggesting a potential plasticity between ecm-myCAF and IL-iCAF. In contrast, the other markers did not show any significant variation upon coculture and remained either high (SDC1) or low (GPC3 and CD9), as expected on the basis of respective cluster identity (Fig. 5H). These observations suggest that CD4+ CD25+ T lymphocytes might promote conversion of ecm-myCAF (ANTXR1+ SDC1+ LAMP5− CD9+/−) into TGFβ-myCAF (ANTXR1+ LAMP5+ SDC1+/− CD9+/−), both clusters being myCAF. Taken as a whole, we found that ecm-myCAF can have a direct impact on PD-1 and CTLA4 protein levels at the surface of FOXP3+ T lymphocytes. Reciprocally, Tregs can promote the conversion of ecm-myCAF into TGFβ-myCAF, thereby underlying a positive feedback loop between these two clusters and CD4+ CD25+ PD-1+ or CTLA4+ T cells that could account for the positive correlations we observed in breast cancer.
ecm-myCAF and TGFβ-myCAF Are Associated with Primary Resistance to Immunotherapy
Given the direct effect of specific CAF-S1 clusters on PD-1 and CTLA4 protein levels on Tregs, we next wondered whether some CAF-S1 clusters could be associated with immunotherapy resistance. As we did not have access to data from breast cancer treated by immunotherapy, we took advantage of publicly available data from patients with metastatic melanoma treated with anti–PD-1 (pembrolizumab) therapy (33), which has revolutionized melanoma treatment. As defined in the aforementioned study, we considered patients as “nonresponders” to anti–PD-1 if they showed progressive disease, and as “responders” in the case of complete or partial response. By performing gene set enrichment analysis, we first observed that, at time of diagnosis, expression of CAF-S1–specific genes, but not of normal fibroblast content, was significantly enriched in tumors from nonresponder patients (Fig. 6A). Using CAF-S1 cluster–specific signatures (Supplementary Data S2), we observed that ecm-myCAF, TGFβ-myCAF, and wound-myCAF gene expression was enriched in nonresponders compared with responders, whereas detox-iCAF, IL-iCAF, and IFNγ-iCAF clusters were not (Fig. 6B). We next compared the content in each CAF-S1 cluster in responders versus nonresponders. We confirmed that ecm-myCAF, TGFβ-myCAF, and wound-myCAF expression was significantly higher in nonresponders than in responders, whereas detox-iCAF and IFNγ-iCAF expression was similar between the two subgroups of patients (Fig. 6C). Moreover, neither the normal fibroblast content nor the cytolytic index, defined in ref. 32, was different between responders and nonresponders (Fig. 6D and E). In agreement with these observations, the reciprocal analysis (i.e., determining the number of responders and nonresponders according to low or high CAF-S1 cluster expression) confirmed that the number of nonresponder patients was significantly associated with tumors showing high expression of ecm-myCAF, TGFβ-myCAF, or wound-myCAF at diagnosis, whereas the other CAF-S1 clusters, the general CAF content, or the cytolytic index were not informative on patient response to immunotherapy (Fig. 6F–H). Together, these data show that three specific CAF-S1 clusters (ecm-myCAF, TGFβ-myCAF, and wound-myCAF) are indicative at diagnosis of anti–PD-1 response in patients with metastatic melanoma, whereas the other CAF-S1 clusters (detox-iCAF, IL-iCAF, and IFNγ-iCAF), the total CAF content, or the cytolytic index are not. Finally, we sought to validate the impact of CAF-S1 clusters, in particular ecm-myCAF, TGFβ-myCAF, and wound-myCAF, on primary immunotherapy resistance in a series of patients with metastatic NSCLC, another recently established clinical indication for immunotherapy (here treated in second- or third-line setting with nivolumab; see Supplementary Table S2 for detailed description of the NSCLC cohort 4). Similar to melanoma, we validated that CAF-S1 signature, evaluated in tumor specimens sampled at diagnosis, was significantly enriched in nonresponder patients (Fig. 6I). In contrast, the normal fibroblast content was higher in responders than in nonresponders, suggesting that CAF-S1 was significantly enriched in nonresponders (Fig. 6I). Importantly, we confirmed that ecm-myCAF, TGFβ-myCAF, and wound-myCAF were associated with nonresponder patients with NSCLC, as opposed to detox-iCAF, IL-iCAF, and IFNγ-iCAF clusters (Fig. 6J). In conclusion, in contrast to detox-iCAF, IL-iCAF, and IFNγ-iCAF clusters, the abundance of ecm-myCAF and TGFβ-myCAF at diagnosis is associated with resistance to immunotherapy in both melanoma and NSCLC, consistent with their capacity to increase PD-1+ and CTLA4+ protein levels in Tregs.
By performing scRNA-seq on more than 19,000 CAF-S1 fibroblasts from patients with breast cancer, we identified eight cellular clusters within the CAF-S1 immunosuppressive subset in human breast cancer. We validated the existence of the five most abundant clusters by FACS using specific surface markers. We dissected the most prominent pathways and gene-specific signatures, characterizing each CAF-S1 cluster as follows: ECM (cluster 0/ecm-myCAF), detoxification (cluster 1/detox-iCAF), response to stimuli (cluster 2/IL-iCAF), TGFβ (cluster 3/TGFβ-myCAF), wound healing (cluster 4/wound-myCAF), IFNγ and cytokines (cluster 5/IFNγ-iCAF), IFNα/β (cluster 6/IFNαβ-iCAF), and acto-myosin (cluster 7/acto-myCAF). We took advantage of these specific signatures to show that the different CAF-S1 clusters exhibit a distinct accumulation in breast cancer subtypes and to confirm their existence in publicly available datasets for HNSCC and NSCLC, thereby underscoring the relevance of our findings. Among the five most abundant clusters, ecm-myCAF and TGFβ-myCAF are specifically associated with an immunosuppressive environment, as their abundance correlates with that of PD-1+ and/or CTLA4+ CD4+ T cells. Importantly, ecm-myCAF and TGFβ-myCAF are enriched at time of diagnosis in samples from patients with melanoma and NSCLC who did not respond to immunotherapies. Consistent with these observations, we demonstrated that ecm-myCAFs are able to increase the expression of PD-1 and CTLA4 at the surface of FOXP3+ Tregs. Reciprocally, CD4+ CD25+ T lymphocytes promote the conversion of ecm-myCAF into TGFβ-myCAF fibroblasts. These data therefore demonstrate an interesting reciprocal cross-talk between specific ecm-myCAF and TGFβ-myCAF with CD4+ CD25+ T lymphocytes that could promote immunosuppression and be involved in resistance to immunotherapies.
Heterogeneity in cellular composition represents a major challenge in modern oncology. In recent years, scRNA-seq appeared as a revolutionary technique for addressing intratumor cellular complexity (27, 30, 34–44). In addition to cancer cells, previous studies have provided the first comprehensive catalogs of normal cells that compose the tumor microenvironment, most of these studies being focused on immune cell heterogeneity (31, 45–51). Recently, we and others have identified different CAF subsets in various adenocarcinomas, including two subpopulations defined by either adhesion/wound healing (CAF-S1) or perivascular/contractile (CAF-S4) signatures (7–16). The first single-cell data on CAFs from human cancers and mouse models confirmed the existence of ECM-rich (CAF-S1) and contractile (CAF-S4) subpopulations (16, 24, 26, 29, 38), indicating that CAF-S1 and CAF-S4 can be detected in distinct cancer types and across species. Here, we go a step further by analyzing a large number of immunosuppressive CAF-S1 cells, reaching, to our knowledge, an unprecedented resolution of this subpopulation. Unbiased methods enabled us to identify specific gene signatures that distinguish different CAF-S1 cellular clusters, which highlights the diversity of this particular CAF subpopulation within cancers. We confirmed the identification of two CAF-S1 myCAF and iCAF fibroblast subtypes previously identified in pancreatic cancers (7, 16, 26, 29). Our analysis increases CAF-S1 cellular resolution by showing that both myCAF and iCAF fibroblasts can be further subdivided into different cellular clusters. CAF-S1 from clusters 1 and 2 (detox-iCAF and IL-iCAF) are characterized by detoxification and inflammatory signaling and by response to IL and TNF pathways, respectively. CAF-S1 from cluster 5 (IFNγ-iCAF) are enriched in IFNγ signature and exhibit high expression of CD74, a marker of the apCAF subset recently identified in pancreatic cancers (16), suggesting cluster 5/IFNγ-iCAF might be the “apCAF” counterpart in breast cancer. Among myCAF, CAF-S1 cluster 0 is enriched in genes encoding ECM and ECM remodeling proteins and was referred to as “ecm-myCAF.” Interestingly, ecm-myCAF–specific signature contains the LRRC15 gene that has been recently identified in pancreatic cancer (29). In addition, myCAFs from cluster 3 are specified by TGFβ-signaling pathway and were defined as “TGFβ-myCAF.” Importantly, we confirmed the existence of the five most abundant clusters by using flow cytometry in breast cancer. We also detected these distinct CAF-S1 cellular clusters in HNSCC and NSCLC, demonstrating the relevance of our findings across cancer types. In addition, we unraveled that the CAF-S1 clusters exhibit a distinct accumulation in the different breast cancer subtypes, with the detox-iCAF and IL-iCAF being predominant in TN breast cancer, consistent with the possible presence of numerous TILs in this specific breast cancer subtype (8).
Cancer immunotherapy has emerged as an effective therapy in oncology. However, despite encouraging results, many patients with advanced cancer do not respond to immune checkpoint inhibitors, and little is known about the mechanisms of primary resistance. Identification of biomarkers that may reliably discriminate between responder and nonresponder patients before initiating therapy is hence needed to select patients who are likely to benefit from immuno-oncology drugs. Here, we show that several CAF-S1 clusters could actually contribute to primary resistance to immunotherapy. In addition to well-known factors involved in immunotherapy resistance, such as lack of antigen presentation, tumor immunogenicity deficiency, T-cell exclusion, or defective tumor cell response to IFNγ (52–55), we identify here specific CAF-S1 cellular clusters as new key players in primary resistance to immunotherapy. We demonstrate that the abundance of ecm-myCAF and TGFβ-myCAF in tumors is anticorrelated with CD8+ T-cell infiltration but correlated with PD-1+ and CTLA4+ CD4+ T-cell content in breast cancer. Moreover, our results support the existence of a reciprocal cross-talk between the ecm-myCAF and TGFβ-myCAF clusters, on the one hand, and Tregs on the other, which could reinforce the role of these clusters in immunotherapy resistance. Indeed, ecm-myCAF enhance the expression of PD-1 and CTLA4 at the surface of CD4+ CD25+ FOXP3+ T lymphocytes, thereby leading to a global increase in PD-1+ and CTLA4+ Tregs. Reciprocally, Tregs tend to promote the conversion of CAF-S1 fibroblasts from ecm-myCAF to TGFβ-myCAF. This reciprocal cross-talk explains, at least in part, the enrichment in ecm-myCAF and TGFβ-myCAF in tumors that are highly infiltrated by Tregs. It also provides some clues for the link of ecm-myCAF and TGFβ-myCAF with immunotherapy resistance, in agreement with recent findings on LRCC15 (29), which is one of the genes of the ecm-myCAF signature. In contrast to ecm-myCAF and TGFβ-myCAF, wound-myCAFs are not associated with an immunosuppressive environment, but correlated with a high global infiltration by T lymphocytes. Hence, as the wound-myCAF cluster is enriched in tumors from patients who do not respond to immunotherapy, it might serve as a new surrogate marker of primary resistance to immunotherapies in highly infiltrated tumors, which are usually sensitive to this type of treatment. Still, the mechanistic explanation for the role of wound-myCAF in this paradoxical observation remains elusive. As expected, if we consider that IFNγ-iCAFs are reminiscent of the apCAF subset recently identified (16), this cluster is not indicative of immunotherapy resistance. Taken as a whole, these findings suggest that, although important, infiltration by CD8+ and Tregs might not be sufficient on its own for defining nonresponder patients. Thus, assessing the content in specific CAF-S1 clusters in tumors at time of diagnosis could provide an additive value for predicting primary resistance to immune checkpoint inhibitors.
ecm-myCAFs and TGFβ-myCAFs are characterized by signatures of genes involved in collagen synthesis and ECM organization and response to TGFβ stimulus, respectively. TGFβ was shown to attenuate anti–PD-L1 response by contributing to T-cell exclusion, whereas TGFβ inhibition was shown to unleash potent cytotoxic T-cell response against tumor cells (11, 19, 56–58). Consistent with these observations, TGFβ-myCAF expresses high levels of TGFβ1 and TGFβ3. In addition, stromal and mesenchymal gene expression has been previously related to resistance to anti–PD-1/PD-L1 blockade in melanoma and urothelial and colorectal cancers (29, 38, 58–60). The dense networks of CAF-secreted collagen fibers in tumor nests have been shown to constitute a physical barrier that prevents T cells from reaching the tumor bed (61). CAF-S1 fibroblasts also exert an active function in immunosuppression by increasing attraction, survival, and activation of FOXP3+ T lymphocytes (8, 10). Here, we go a step further by showing that ecm-myCAF are able to stimulate PD-1 and CTLA4 protein levels at the surface of FOXP3+ T cells, which might influence anti–PD-1 immunotherapy effectiveness. The increase in CTLA4 expression at the surface of Tregs could also be a mechanism to bypass PD-1 blockade and to contribute to immune exhaustion persistence. Such increased CTLA4 expression by Tregs has been suspected to be the biological basis for the additive effect of anti-CTLA4 antibodies when used in combination with anti–PD-1 checkpoint inhibitors. This combination has recently been shown to improve overall survival, as compared with chemotherapy, in the first-line setting of patients with metastatic NSCLC (62). These data thus support the development of strategies combining PD-1 and/or CTLA4 blockade with therapies targeting CAF-S1 cluster components to overcome primary resistance to immune checkpoint blockade.
Cohorts of Patients
Patients with Breast Cancer.
The study developed here is based on samples taken from surgical residues available after histopathologic analyses and not required for diagnosis. There is no interference with clinical practice. Analysis of primary tumor samples was performed in accordance with the relevant national law and with recognized ethical guidelines (Declaration of Helsinki) on the protection of people taking part in biomedical research. All patients hospitalized at Institut Curie (patients with breast cancer) received a welcome booklet explaining that their samples may be used for research purposes. All patients included in our study were thus informed by their referring oncologist that biological samples collected through standard clinical practice could be used for research purposes and they gave their verbal informed consent. In the case of patient refusal, which could be either orally expressed or written, residual tumor samples were not included in our study. Human experimental procedures for analyses of tumor microenvironment by F. Mechta-Grigoriou's lab were approved by the Institutional Review Board and Ethics committee of the Institut Curie Hospital group (approval February 12, 2014) and CNIL (Commission Nationale de l'informatique et des Libertésapproval no.: 1674356 delivered March 30, 2013). The Biological Resource Centre (BRC) is part of the Pathology Department in the Diagnostic and Theragnostic Medicine Department headed by Dr. A. Vincent-Salomon. BRC is authorized to store and manage human biological samples according to French legislation. The BRC has declared defined sample collections that are continuously incremented as and when patient consent forms are obtained (declaration number: DC-2008-57). The BRC follows all currently required national and international ethical rules, including the Declaration of Helsinki. The BRC has also been accredited with the AFNOR NFS-96-900 quality label (renewed and currently valid until 2021). All samples are pseudo-anonymous when they arrive from the BRC in the lab. In addition, the BRC collections have been declared to the CNIL (approval no. 1487390 delivered February 28, 2011). Lum tumors were defined by positive immunostaining for estrogen receptor (ER) and/or progesterone receptor (PR). The cutoff used to define hormone receptor positivity was 10% of stained cells. Ki-67 (proliferation) score further distinguishes LumA from LumB tumors (below 15%: Lum A, above: Lum B). HER2-amplified carcinomas have been defined according to ERBB2 immunostaining using American Society of Clinical Oncology guidelines. TN immunophenotype was defined as follows: ER−PR− ERBB2− with the expression of at least one of the following markers: KRT5/6+ or EGFR+.
Patients with NSCLC.
NSCLC samples were from the routine diagnostic samples stored in the Pathology Department of Bichat Hospital, from patients treated with immuno-oncology drugs in the Thoracic Oncology Department of Bichat hospital headed by G. Zalcman. The deidentified clinical data were part of the thoracic oncology database of patients with lung cancer from the Clinical Investigation Centre (coheaded by G. Zalcman) CIC-1425/CLIP2 of Bichat Hospital (Regional Health Agency authorization no.: 17-1381), in accordance with French regulatory rules for observational clinical research. Patients received checkpoint inhibitors after progression upon chemotherapy-based first or second line, according to the registration of immuno-oncology drugs. During the current study period, the anti–PD-1 nivolumab mAb represented the most frequent drug used in such setting. Efficacy of immuno-oncology treatment was assessed every 8 to 12 weeks by whole-body CT scan, by a weekly multidisciplinary tumor board, including thoracic specialized radiologists, thoracic oncologists, and pulmonologists, according to RECIST v.1.1 criteria, defining objective responders (OR), patients with stable disease (SD), and patients showing tumor progression (Progr) with 20% or more increase in their tumor volume without any clinical benefit. The best response status observed at 4 months was used in this study. SD patients, for whom immuno-oncology drug was given for more than 6 months because of a clinical benefit at 4-month evaluation (n = 3), without any criteria for progressive disease (then long-lasting SD) were included in the group of responder patients in this study. Date of the CT scan–assessed progression was recorded. Some patients showed an early clinical progression that required an early (before 8 weeks) CT scan assessment. In these series, according to the above criteria, there were 22 responder patients and 48 progressive patients, who received less than 4 months of treatment. Date of progression was retained as the date of CT scan showing RECIST progression. Date and cause of death or date of last news with vital status was systematically recorded. Postprogression second- or third-line treatments were registered. There was no unbalance according to postprogression treatments. PD-L1 staining was performed and interpreted by A. Guyard on 4-μm paraffin-embedded sections from diagnosis pretreatment biopsy samples containing at least 200 tumor cells, using Cell Signaling Technology E1L3N commercially available clone on the Leica Bond platform. All but 5 patients (having less than 200 tumor cells in the remaining pathologic block) had PD-L1 IHC analysis.
CAF-S1 RNA-seq at Single-Cell Level
Isolation of CAF-S1 from Breast Cancer.
CAF-S1 fibroblasts were isolated from a total of 8 primary breast cancers (surgical residues prior any treatment; see Supplementary Table S1 for details on the prospective cohort). Seven breast cancers were initially studied. In addition, another breast cancer sample was added for validating the CAF-S1 clusters by using the Label Transfer algorithm from Seurat R package. CAF-S1 fibroblasts were isolated from breast cancer by using BDFACS ARIA III sorter (BD Biosciences). Fresh human breast cancer primary tumors were collected directly from the operating room after surgical specimen macroscopic examination and selection of areas of interest by a pathologist. Samples were cut into small pieces (around 1 mm3) and digested in CO2-independent medium (Gibco #18045-054) supplemented with 150 μg/mL liberase (Roche #05401020001) and DNase I (Roche #11284932001) for 40 minutes at 37°C with shaking (180 rpm). Cells were then filtered through a 40-μm cell strainer (Thermo Fisher Scientific #223635447) and resuspended in PBS+ solution (PBS, Gibco #14190; EDTA 2 mmol/L, Gibco #15575; Human Serum 1 %, BioWest #S4190-100) at a final concentration between 5 × 105 and 106 cells in 50 μL.To isolate CAF-S1 fibroblasts, we first applied a selection to exclude epithelial (EPCAM+), hematopoietic (CD45+), endothelial (CD31+), and CD235a+ (red blood) cells and next used CAF-S1 markers (FAP and CD29). To do so, cells in suspension were then stained with an antibody mix containing anti-EPCAM-BV605 (BioLegend, #324224), anti-CD31-PECy7 (BioLegend, #303118), anti-CD45-APC-Cy7 (BD Biosciences, #BD-557833), anti-CD235a-PerCP/Cy5.5 (BioLegend, #349109), anti-CD29-Alexa Fluor 700 (BioLegend, #303020), anti-FAP-APC (primary antibody, R&D Systems, #MAB3715) for flow cytometry cell sorting to perform scRNA-seq. All antibodies except FAP were purchased already conjugated with fluorescent dyes. Anti-FAP antibody was conjugated with fluorescent dye Zenon APC Mouse IgG1 Labeling Kit (Thermo Fisher Scientific, #Z25051). The following isotype control antibodies for each CAF marker were used: iso-anti-CD29 (BioLegend, #400144) and iso-anti-FAP (primary antibody, R&D Systems, #MAB002).
Cell suspensions were stained immediately after dissociation of breast cancer tumor samples during 15 minutes at room temperature with the antibody mix in PBS+ solution. DAPI (2.5 μg/mL; Thermo Fisher Scientific, #D1306) was added just before flow cytometry sorting. Signals were acquired on the BDFACS ARIA III sorter (BD Biosciences) for cell sorting. At least 5 × 105 events were recorded. Compensations were performed using single staining on anti-mouse IgG and negative control beads (BD Biosciences, #552843) for each antibody. Data analysis was performed using FlowJo version X 10.0.7r2. Cells were first gated on the basis of forward (FSC-A) and side (SSC-A) scatters (measuring cell size and granularity, respectively) to exclude debris. Dead cells were excluded on the basis of their positive staining for DAPI. Single cells were next selected based on SSC-A versus SSC-W parameters. Gating included EPCAM−, CD45−, CD31−, CD235a− cells, to remove epithelial (EPCAM+), hematopoietic (CD45+), endothelial (CD31+), and red blood cells (CD235a+).
Single-Cell CAF-S1 RNA-seq.
Upon isolation, CAF-S1 cells were directly collected into RNase-free tubes (Thermo Fisher Scientific, #AM12450) precoated with DMEM (GE Life Sciences, #SH30243.01) supplemented with 10% FBS (Biosera, #1003/500). At least 6,000 cells were collected per sample. In these conditions, cell concentration was checked in control samples and was of 200,000 cells/mL. Single-cell capture, lysis, and cDNA library construction were performed using Chromium system from 10X Genomics, with the following kits: Chromium Single Cell 3′ Library & Gel Bead Kit v2 kit (10X Genomics, #120237) and Chromium Single Cell A Chip Kits (10X Genomics, #1000009). Generation of gel beads in Emulsion (GEM), barcoding, post GEM-reverse transcription cleanup and cDNA amplification were performed according to the manufacturer's instructions. Targeted cell recovery was 3,000 cells per sample to retrieve enough cells, while preserving a low multiplet rate. Cells were loaded accordingly on the Chromium Single cell A chips, and 12 cycles were performed for cDNA amplification. cDNA quality and quantity were checked on Agilent 2100 Bioanalyzer using Agilent High Sensitivity DNA Kit (Agilent, #5067-4626) and library construction followed according to 10X Genomics protocol. Libraries were next run on the Illumina HiSeq (for patients P5–7) and NovaSeq (for patients P1–4) with a depth of sequencing of 50,000 reads per cell. Processing of raw data, including demultiplexing of raw base call (BCL) files into FASTQ files, alignment, filtering, barcode, and Unique Molecular Identifiers (UMI) counting, were performed using 10X Cell Ranger pipeline version 2.1.1. Reads were aligned to Homo sapiens (human) genome assembly GRCh38 (hg38).
scRNA-seq Data Processing
Preprocessing of raw data was initially performed using Cell Ranger software pipeline (version 2.1.1). This step included demultiplexing of raw base call (BCL) files into FASTQ files, reads alignment on human genome assembly GRCh38 using STAR, and counting of unique molecular identifier (UMI). A first set of 18,805 CAF-S1 cells from 7 patients with breast cancer (corresponding to seven sequencing runs, P1–7) were analyzed using Seurat R package (version 3.0.0; ref. 63). A second set of 1,646 CAF-S1 cells from 1 patient with breast cancer (patient 8) was used for validation (see the “Label Transfer” section below) and analyzed using the same methodology.
As a quality-control step, we first filtered out low-quality cells, empty droplets, and multiplet captures based on the distribution of the unique genes detected (nonzero count) in each cell for each patient. Cells with less than 200 genes detected and more than 6,000 genes detected (for patient 1), more than 5,000 genes detected (for patients 3, 5, and 6), more than 4,500 genes detected (for patients 2, 7, and 8) or more than 4,000 genes detected (for patient 4) were excluded. Distribution of cells based on the fraction of expressed mitochondrial genes was also plotted. Cells with a fraction of mitochondrial genes higher than 5% were discarded to eliminate dying cells or low-quality cells with extensive mitochondrial contamination. For each patient, the mitochondrial fraction was computed using the Percentage FeatureSet function from Seurat with argument pattern = “λMT-”. Following these quality-control criteria, 18,296 CAF-S1 cells (patient 1 = 1,825 cells; patient 2 = 3,300 cells; patient 3 = 2,810 cells; patient 4 = 3,153 cells; patient 5 = 2,486 cells; patient 6 = 3,179 cells; and patient 7 = 1,543 cells), and 1,582 CAF-S1 cells (patient 8) were finally conserved in the first and second datasets, respectively, for downstream analyses.
Normalization and Data Integration.
Integration of the 7 breast cancer scRNA-seq from the first dataset was done using Seurat functions FindIntegrationAnchors and IntegrateData after library-size normalization of each cell using NormalizeData function with default parameters. Thirty dimensions were used for canonical correlation analysis (CCA); 30 principal components (PC) were used in the weighting procedure of IntegrateData function. Data were scaled using the ScaleData function and the variables “nUMI” and “percent.mt” were used for regression. The same parameters were used for the normalization of the second dataset.
Clustering and Data Visualization.
Principal component analysis (PCA) dimensionality reduction was run using default parameters. Number of the included components (PCs) was assessed using the JackStraw procedure implemented in JackStraw and ScoreJackStraw functions. Thirty PCs were conserved. Graph-based clustering approach was used to cluster the cells from the first dataset using FindNeighbours (k = 20) and FindClusters functions (res = 0.35). Ten CAF-S1 clusters were obtained at this resolution. For visualization of the data, the nonlinear dimensional reduction technique UMAP was applied using the RunUMAP function from Seurat.
Analysis of Differential Gene Expression and Signaling Pathways.
Genes specifically upregulated in each of the 10 clusters of the first dataset were identified using pairwise differential analysis. Although the median number of differentially expressed genes in each pairwise combination was 126 genes, two combinations gave a very limited number of differentially expressed genes, with only 9 genes between clusters 0 and 5 and 22 genes between clusters 3 and 6. Biological meaning of each cluster was also determined using the Metascape tool (http://metascape.org) using all genes significantly upregulated in each of the 10 initial clusters (one cluster vs. all other clusters; function FindAllMarkers with following parameters: logfc.threshold = 0.25, test = wilcox for Wilcoxon rank sum test). Consistent with pairwise analysis, biological pathways identified for clusters 0/5 on the one hand, and 3/6 on the other hand, were redundant and thus combined. Clusters 0 and 5 were then defined as cluster 0/ecm-myCAF and clusters 3 and 6 as cluster 3/TGFβ-myCAF to finally identify 8 biologically distinct CAF-S1 clusters.
Gene Signatures of CAF-S1, CAF-S1 Clusters, and Normal Fibroblasts (Supplementary Data S2).
Specific gene signatures of CAF-S1 clusters 0 to 5 were defined by performing a differential analysis (Wilcoxon rank-sum test) between clusters 0 and 5. Differentially expressed genes between clusters (one cluster vs. all other clusters) with an Padj < 0.05 were selected. As these signatures were used for detecting CAF-S1 clusters in RNA-seq data from single cells and bulk of different cancer types including melanoma, NSCLC, and HNSCC, we excluded genes expressed in tumor cells by using scRNA-seq data from tumor cells of melanoma (27), NSCLC (31), and HNSCC (30). We defined genes expressed in tumor cells (and thus excluded from CAF-S1 cluster signatures) if more than 10% of tumor cells showed an expression level higher than 1 in any of the aforementioned scRNA-seq data. Details of the data are given in the first tab of Supplementary Data S2. CAF-S1 global signature was initially published in ref. 8 and submitted to the same type of analysis, excluding genes detected in tumor cells, and thus adapted for bulk analysis. The first 100 most significant genes were considered for the CAF-S1–specific signature. The normal fibroblast signature was defined by the genes significantly upregulated in normal fibroblasts (FAPneg CD29med SMAneg) isolated from healthy juxta-tumor tissues, compared with CAF-S1 fibroblasts isolated from breast cancer. Genes that are expressed in tumor cells were excluded from the signature, following the same strategy as this one described above. Cytolytic index was defined as the geometric mean of granzyme A (GZMA) and perforin (PRF1) gene expression, as described in ref. 32. All these signatures are given in the second tab of Supplementary Data S2.
To validate the CAF-S1 clusters identified in the first dataset, another set of data corresponding to 1,582 CAF-S1 fibroblasts after quality control and collected from an additive breast cancer sample was analyzed using the Seurat pipeline. The Label Transfer algorithm, described in ref. 28 and implemented in the Seurat V3.0 R package, was applied using functions FindTransferAnchors and TransferData. The first dataset of 18,296 CAF-S1 cells was used as reference, and the second dataset of 1582 CAF-S1 cells was used as query. When finding anchors, dimensional reduction was performed by projecting the PCA from the reference onto the query. Thirty dimensions were used.
Single-Cell Data Integration from Breast Cancer, HNSCC, and NSCLC
Integration between breast cancer and HNSCC or breast cancer and NSCLC single-cell data was done using the method described in ref. 28 and implemented in Seurat V3.0 R package. In brief, the identification of cell pairwise correspondences between single cells across datasets (called “anchors”) allows transforming datasets into a shared space. Dimensionality reduction of both datasets was performed using diagonalized CCA, and L2-normalization was applied to the canonical correlation vectors prior to the identification of anchors. Default parameters were used for FindIntegrationAnchors and IntegrateData function in Seurat V3.0 package.
Flow Cytometry Analysis of the Five Most Abundant CAF-S1 Clusters and Immune Cells
Forty-four breast cancer samples were cut into small fragments and digested in CO2-independent medium (Gibco, #18045-054) supplemented with 5% FBS (PAA, #A11-151), 2 mg/mL collagenase I (Sigma-Aldrich, #C0130), 2 mg/mL hyaluronidase (Sigma-Aldrich, #H3506), and 25 mg/mL DNase I (Roche, #11284932001) for 45 minutes at 37°C with shaking (180 rpm). After tissue digestion, cells were filtered using a cell strainer (40 mm, Thermo Fisher Scientific, #223635447) and washed using PBS solution (Gibco, #14190) supplemented with 2 mmol/L EDTA (Gibco, #15575) and 1% human serum (BioWest, #S4190-100). Cells were then separated into two groups for analyzing CAF-S1 cluster panel and immune cell panel, respectively.
CAF-S1 Cluster Panel.
Cells were stained with Live/Dead NIR (1:1000, BD Bioscience #565388) for 20 minutes in PBS. Cells were then washed and stained for 45 minutes with an antibody cocktail containing anti-CD235a-APC-Cy7 (1:20, BioLegend, #349115), anti-EPCAM-BV605 (1:25, BioLegend, #324224), anti-CD31-PECy7 (1:50, BioLegend, #303118), anti-CD45-BUV395 (1:25, BD Biosciences, #BD-563792), anti-CD29-Alexa Fluor 700 (1:50, BioLegend, #303020), anti-FAP (1:100, R&D Systems, #MAB3715) coupled using fluorescent dye Zenon APC Mouse IgG1 Labeling Kit (Thermo Fisher Scientific, #Z-25051), anti-ANTXR1-AF405 (1:33, Novus Biologicals, #NB-100-56585), anti-LAMP5-PE (1:10, Miltenyi Biotec, #130-109-156), anti-SDC1-BUV737 (1:25, BD Biosciences, #BD-564393), anti-GPC3-AF594 (1:20, R&D Systems, #FAB2119T, 100UG), anti-DLK1-AF488 (1:25, R&D Systems, #FAB1144G-100), and anti-CD9-BV711 (1:200, BD Biosciences, #BD-743050). Isotype control antibodies for each CAF cluster marker used were: mouse IgG1 isotype control - BV711 (1:200, BD Biosciences, #563044), mouse IgG1 isotype control - AF405 (1:3, Novus Biologicals, #IC002V), mouse IgG1 isotype control - BUV737 (1:25, BD Biosciences, #564299), mouse IgG2B isotype control - AF488 (1:12,5, R&D Systems, #IC0041G), mouse IgG2A isotype control - AF594 (1:5, R&D Systems, #IC003T), REA Control-PE (1:10, Miltenyi Biotec, #130-113-462). Cells were then washed and acquired using LSR FORTESSA analyzer (BD Biosciences) the same day or fixed in 4% paraformaldehyde (PFA; Electron Microscopy Sciences, #15710) for 20 minutes then washed and kept in PBS+ solution overnight and acquired the next day. At least 5 × 105 events were recorded. Compensations were performed using single staining on anti-mouse IgG and negative control beads (BD Biosciences #552843) for each antibody and on cells for Live/Dead staining. Data analysis was performed using FlowJo version 10.4.2 (LLC). Cells were first gated on the basis of forward (FSC-A) and side (SSC-A) scatters (measuring cell size and granulosity, respectively) to exclude debris. Dead cells and red blood cells were excluded on the basis of their positive staining for Live/Dead NIR and CD235a, respectively. Single cells were next selected on the basis of SSC-H versus SSC-A parameters. Cells were then gated on EPCAM−, CD45−, CD31− cells, for excluding epithelial cells (EPCAM+), hematopoietic cells (CD45+), and endothelial cells (CD31+). DAPI−, EPCAM−, CD45−, CD31− cells were separated on four subsets (CAF-S1 to CAF-S4) according to FAP and CD29. CAF-S1 subset was first gated on ANTXR1. ANTXR1+ cells were next gated according to SDC1 and LAMP5. ANTXR1+ SDC1+ LAMP5− were defined as cluster 0/ecm-myCAF and ANTXR1+ SDC1− LAMP5+ as cluster 3/TGFβ-myCAF. ANTXR1+ SDC1− LAMP5− were gated on CD9, and ANTXR1+ SDC1− LAMP5− CD9+ were defined as cluster 4/wound-myCAF. ANTXR1−/lo cells were gated on DLK1 and GPC3. Cluster 1/detox-iCAF was defined as ANTXR1− GPC3+ DLK1−/+ and cluster 2 as ANTXR1− GPC3− DLK1+. ANTXR1− GPC3− DLK1−, and ANTXR1+ SDC1− LAMP5− CD9− were designated as other clusters.
Among the 44 breast cancer samples analyzed for CAF-S1 clusters, 37 were characterized at the meantime for immune content. Cell types were analyzed on the Live/Dead negative fraction and defined as hematopoietic cell (CD45+), CD4+/CD8+ T lymphocytes; B lymphocytes (CD45+ CD14−CD3− CD19+); NK (CD45+ CD14− CD3− CD56+), cytotoxic NK (CD56+ CD16+), and noncytotoxic NK (CD56+ CD16−); CD45+ CD14− CD3+ CD4+/CD8+ and myeloid cells (CD45+ CD14+). On each identified population, the percentage of positive cells for the following checkpoints was also evaluated: PD-1, CTLA4, NKG2A, TIGIT, CD244, CD158K, CD69, CD161. Cells were stained with Live/Dead (1:1,000, Thermo Fisher Scientific, #L34955) for 20 minutes in PBS. Cells were then washed and stained for 45 minutes with an antibody cocktail containing anti-CD45-APC-cy7 (1:20, BD Biosciences, #BD-557833), anti-CD14-BV510 (1:50, BD Biosciences, #563079), anti-CD56-BUV395 (1:25, BD Biosciences, #563554), anti-CD16-BV650 (1:25, BD Biosciences, #563692), anti-PD-1-BUV 737 (1:20, BD Biosciences, #565299), anti-CD3-AF700 (1:25, BD Biosciences, #557943), anti-NKG2A-BV786 (1:20, BD Biosciences, #747917), anti-TIGIT-BV605 (1:20, BD Biosciences, #747841), anti-CD158K-PE (1:10, Miltenyi Biotec, #130-095-205), anti-CD244-FITC (1:10, BD Biosciences, #550815), anti-CTLA4-Pe-cy5 (1:10, BD Biosciences, #555854), anti-CD19-Percp-cy5.5 (1:20, BD Biosciences, #561295), anti-CD4 APC (1:25, Miltenyi Biotec, #130-092-374), anti-CD8-PE-TexasRed (1:100, Life Technologies, #MHCD0817), anti-CD69-BV710 (1:25, BD Biosciences, #563836), anti-CD161-PE-VIO770 (1:100, Miltenyi Biotec, #130-113-597). Cells were then washed and acquired using LSR FORTESSA analyzer (BD Biosciences) the same day or fixed in 4% PFA (Electron Microscopy Sciences, #15710) for 20 minutes, then washed and kept in PBS+ solution overnight and acquired the next day. At least 5 × 105 events were recorded. Compensations were performed using single staining on anti-mouse IgG and negative control beads (BD Biosciences, #552843) for each antibody and on cells for Live/Dead staining. Data analysis was performed using FlowJo version 10.4.2 (LLC).
RNA-seq of CAF-S1 Primary Cell Lines Isolated from Breast Cancer
RNAs were extracted from CAF-S1 fibroblasts with Qiagen miRNeasy Kit (Qiagen, #217004) according to the manufacturer's instructions. Among the 7 CAF-S1 primary cell lines studied here, 3 were isolated by sorting and 4 by spreading. These 7 CAF-S1 primary cell lines were generated from 7 different patients with breast cancer. RNA integrity and quality were analyzed using the Agilent RNA 6000 Pico Kit (Agilent Technologies, #5067-1513). cDNA libraries were prepared using the TruSeq Stranded mRNA Kit (Illumina, #20020594) followed by sequencing on NovaSeq (Illumina). Reads were mapped on the human reference genome (hg38; Gencode release 26) and quantified using STAR (version 2.5.3a) with parameters “outFilterMultimapNmax = 20; alignSJoverhangMin = 8; alignSJDBoverhangMin = 1; outFilterMismatchNmax = 999; outFilterMismatchNoverLmax = 0.04; alignIntronMin = 20; alignIntronMax = 1000000; alignMatesGapMax = 1000000; outMultimapperOrder = Random.” Only genes with one read in at least 5% of all samples were kept for further analyses. Normalization was conducted with DESeq2 R package and raw read matrix was log2 transformed. For identifying the identity of CAF-S1 primary cell lines, a score was computed by the mean of expression of the genes that compose the iCAF/myCAF signatures (as defined in ref. 7) and for each CAF-S1 cluster signature (as defined in this article, and shown in Supplementary Data S2). P values are from DESeq2 analysis.
Isolation of CD4+ CD25+ T Lymphocytes.
CD4+ CD25+ T lymphocytes were isolated from peripheral blood of healthy donors obtained from the Etablissement Français du Sang, Paris, Saint-Antoine Crozatier blood bank through a convention with the Institut Curie. Briefly, peripheral blood mononuclear cells (PBMC) were isolated using Lymphoprep (Stemcell, #07861), as described previously in ref. 8. CD4+ CD25+ cells were purified from 5 × 108 PBMCs by using magnetic cell separation with the human CD4+ CD25+ Tregs Isolation Kit (Miltenyi Biotec, #130-091-301), according to the manufacturer's instructions. The purity of CD4+ CD25+ T lymphocytes was determined by flow cytometry, as described in ref. 8.
Isolation of CAF-S1 Clusters in Culture.
To isolate the different CAF-S1 clusters, we first started by sorting the cells according to their specific markers, but we failed to keep them alive with their different identities. We then tested two distinct methods of isolation by spreading and sorting. For the “spreading” method, tumors were cut into small pieces and incubated in plastic dishes (Falcon, #353003) in DMEM (HyClone, #SH30243.01) supplemented with 10% FBS (Biosera, #FB-1003/500), streptomycin (100 μg/mL), and penicillin (100 U/mL; Gibco #15140-122) in a humidified in 1.5% O2 and 5% CO2 incubator, to let fibroblasts spread and expand for at least 2 to 3 weeks at 37°C. For isolating fibroblasts by the “sorting” method, tumors were digested using enzymatic cocktails described above in “Isolation of CAF-S1 from Breast Cancer” and sorted by BDFACS ARIA III using the gating strategy detailed above in “CAF-S1 RNA-seq at Single Cell” in 48-well plastic dishes (TPP plates, #192048) precoated with FBS for 2 hours. CAF-S1 sorted cells were next expanded for 3 to 4 weeks at 37°C in plastic dishes (TPP plates, #192048) in pericyte medium (ScienCell, #1201) supplemented with 2% FBS (ScienCell, #0010) in a humidified in 1.5% O2 and 5% CO2 incubator. For comparing cellular identity of sorted and spread CAF-S1 fibroblasts in the exact same conditions used in functional assays, both types of fibroblasts (spread and sorted) were transferred into plastic dishes (Falcon, #353047) in DMEM (HyClone, #SH30243.01) in 20% O2, as these culture conditions are compatible with coculture with CD4+ CD25+ T lymphocytes that are applied for in vitro functional assays. Using these protocols, seven different CAF-S1 cell lines from 7 different patients have been isolated, 3 by sorting and 4 by spreading. To avoid any in vitro activation, these CAF-S1 primary cell lines isolated by sorting and spreading were used no later than passage 5. Moreover, in each experiment, properties of spread and sorted cells were compared at the same passages.
Treg-CAF-S1 Cluster Functional Assays.
5 × 104 CAF-S1 cells (spread and sorted) were plated on 24-well plates (Falcon, #353047) in DMEM (HyClone, #SH30243.01) with 10% FBS (Biosera, #FB-1003/500) at 1.5% O2 overnight for complete adherence. The medium was then removed and 5 × 105 CD4+ CD25+ T lymphocytes were added in 500 μL of DMEM 1% FBS (ratio 1–10) and incubated overnight (for RNA analysis) or for 24 hours (for FACS) at 37°C, 20% O2. For FACS analysis, nonadherent cells (CD4+ CD25+) were then harvested, washed, and stained for 30 minutes at room temperature using the following markers: Live/Dead (1:1,000, BD Biosciences, #562247), anti-CD45-BUV395 (1:50, BD Biosciences, #BD-563792), anti-CD4-APC (1:50, Miltenyi Biotec, #130-092-374), anti-CD25 PE-cy7 (1:33, BD Biosciences, #557741), anti-FOXP3-FITC (1:33, eBioscience, #53-4776-42), anti-CTLA4-Pe-cy5 (1:20, BD Biosciences, #555854), anti-PD-1-BUV737 (BD Biosciences, #565299), anti-TIGIT-BV605 (1:50, BD Biosciences, #747841), anti-LAG3-BV510 (1:50, BD Biosciences, #744985), anti-TIM3-BV711 (1:50, BD Biosciences, #565566). Adherent cells were trypsinized, washed, and stained for CAF-S1 clusters markers (ANTXR1, CD9, SDC1, LAMP5, GPC3, DLK1) in addition to Live/Dead NIR to remove dead cells and CD45 to remove remaining Tregs. Cells (Treg panel and CAF-S1 panel) were acquired using ZE5 cell analyzer (Bio-Rad) and analyzed by FlowJo v10.4.2. For RNA analysis, nonadherent cells (CD4+ CD25+) were harvested by pipetting, and spun down. RNA was then extracted using Single Cell RNA Purification Kit (Norgen Biotek Corp., #51800) according to the manufacturer's recommendations. RNA integrity and quality were analyzed using the Agilent RNA 6000 Pico Kit (Agilent Technologies, #5067-1513). cDNA libraries were prepared using the TruSeq RNA Exome Kit (Illumina, #20020189) followed by sequencing on NovaSeq (Illumina). Reads were mapped on the human reference genome (hg38; Gencode release 29) and quantified using STAR (version 2.6.1a) with parameters “outFilterMultimapNmax = 20; alignSJoverhangMin = 8; alignSJDBoverhangMin = 1; outFilterMismatchNmax = 999; outFilterMismatchNoverLmax = 0.04; alignIntronMin = 20; alignIntronMax = 1000000; alignMatesGapMax = 1000000; outMultimapperOrder = Random.” Only genes with one read in at least 5% of all samples were kept for further analyses. Normalization was conducted with DESeq2 R package.
Treg-CAF-S1 Cluster Intracellular Staining.
5 × 104 CAF-S1 cells (spread) were plated on 24-well plates (Falcon, #353047) in DMEM (HyClone, #SH30243.01) with 10% FBS (Biosera, #FB-1003/500) at 1.5% O2 overnight for complete adherence. The medium was then removed and 5 × 105 CD4+ CD25+ T lymphocytes were added in 500 μL of DMEM 1% FBS (ratio 1–10) and incubated overnight at 37°C, 20% O2. Nonadherent cells (CD4+ CD25+) were then harvested, washed, and stained for 30 minutes at room temperature with Live/Dead (1:1,000, BD Biosciences, #562247); after washing, the cells were divided into two groups (one fixed and permeabilized and the second kept without fixation/permeabilization except for FOXP3 staining) and stained using the following markers: anti-CD45-BUV395 (1:50, BD Biosciences, #BD-563792), anti-CD4-APC (1:50, Miltenyi Biotec, #130-092-374), anti-CD25 PE-cy7 (1:33, BD Biosciences, #557741), anti-CTLA4-Pe-cy5 (1:20, BD Biosciences, #555854), anti-PD-1-BUV737 (BD Biosciences, #565299), anti-TIGIT-BV605 (1:50, BD Biosciences, #747841), anti-LAG3-BV510 (1:50, BD Biosciences, #744985), anti-TIM3-BV711 (1:50, BD Biosciences, #565566) anti-FOXP3-FITC (1:33, eBioscience, #53-4776-42).
Comparison of Fibroblasts from Normal Healthy Tissue with CAF-S1 from Breast Cancer
Primary fibroblasts were collected by using the spreading method (see above) from juxta-tumors, that is, tissues defined as heathy by referent pathologists. Juxta-tumors were cut into small pieces, put in plastic dishes (Falcon, #353003), and cultured in DMEM (HyClone, #SH30243.01) supplemented with 10% FBS (Biosera, #FB-1003/500), streptomycin (100 μg/mL), and penicillin (100 U/mL; Gibco #15140122) for 2–3 weeks at 37°C. Spread fibroblasts were next analyzed at early and late passages (passages 2 and 5, respectively) for verifying expression of CAF-S1 markers. Primary cells were trypsinized, resuspended in PBS and stained with Live/Dead Fixable Aqua Dead Cell Stain Dye (Thermo Fisher Scientific, #L34957) diluted in PBS for 20 minutes at room temperature and fixed in PFA 4% for 20 minutes at room temperature. After a rapid wash in PBS+, the cells were stained with an anti-FAP antibody (1:100, R&D Systems, #MAB3715) or isotype control (1:100, R&D Systems, #MAB002) in PBS+ for 40 minutes at room temperature. Both antibody and isotype control were coupled using fluorescent dye Zenon APC Mouse IgG1 Labeling Kit (Thermo Fisher Scientific, #Z-25051). Cells were acquired using LSR FORTESSA analyzer (BD Biosciences). Fifty thousand events per sample were recorded.
RNA-seq from NSCLC Samples
FFPE biopsies (n = 120) from treatment-naïve NSCLC were processed for RNA extraction using high FFPET RNA Isolation Kit (Roche, #06650775001) following the manufacturer's instructions. RNA integrity and quality were analyzed using the Agilent RNA 6000 Pico Kit (Agilent Technologies, #5067-1513). Samples with a DV200 higher than 40% were chosen for RNA-seq (n = 70). cDNA library was prepared using the Nextera XT Sample Preparation Kit (Illumina, #FC-131-10) followed by sequencing on NovaSeq (Illumina). Reads were mapped on the human reference genome (release hg19/GRCh37) and quantified using STAR (version 2.5.3a) with parameters “outFilterMultimapNmax = 20; alignSJoverhangMin = 8; alignSJDBoverhangMin = 1; outFilterMismatchNmax = 999; outFilterMismatchNoverLmax = 0.04; alignIntronMin = 20; alignIntronMax = 1000000; alignMatesGapMax = 1000000; outMultimapperOrder = Random.” Only genes with one read in at least 5% of all samples were kept for further analyses. Normalization, unsupervised analysis (PCA), and differential analysis between responder and nonresponder patients were conducted with DESeq2 R package.
All statistical analyses and graphical representation of data were performed in the R environment (https://cran.r-project.org, version 3.5.3) or using GraphPad Prism software (version 8.1.1). Statistical tests used are in agreement with data distribution: Normality was first checked using the Shapiro–Wilk test, and parametric or nonparametric two-tailed tests were applied according to normality, as indicated in each figure legend. scRNA-seq data presented in Figs. 1–3 and Supplementary Figs. S1 and S2 were analyzed using Seurat R package (version 3.0). Correlation matrix shown in Figs. 2 and 4 was computed using cor function from stats R package with method = “pearson” and use = “pairwise.complete.obs”. Corrplot R function was used for the clustering and the visualization of the correlation matrix with the following parameters: order = “hclust” and hclust.method = “ward.D2”. Quantifications from FACS analysis shown in Fig. 5 are shown using mean ± SEM. Differential analysis between CD4+ CD25+ T cells cultured alone or in the presence of ecm-myCAF in Supplementary Fig. S4D and S4E was conducted using the DESeq2 R package. Gene Set Enrichment Analysis (GSEA) software version 3.0 (Broad Institute) was used in Fig. 6. For melanoma RNA-seq data, the following parameters were applied: Enrichment statistic = “weighted”, Metric for ranking genes = “Signal2Noise.” For NSCLC RNA-seq data, GSEAPreranked was used with log2 fold change from DESeq2 differential analysis as metric for ranking genes and “classic” mode for enrichment score.
scRNA-seq data sorted from breast cancer samples are available on the European Genome-Phenome Archive (EGA) platform (https://ega-archive.org) under accession number EGAS00001004030. RNA-seq data from iCAF (sorted) and ecm-myCAF (spread) CAF-S1 fibroblasts maintained in culture are available on the EGA platform under accession number EGAS00001004031. RNA-seq data from NSCLC samples are available on the EGA platform under accession number EGAS00001004032.
Disclosure of Potential Conflicts of Interest
G. Zalcman has received speakers bureau honoraria from BMS and AstraZeneca and is a consultant/advisory board member for BMS, AstraZeneca, MSD, and Roche. F. Mechta-Grigoriou received research support from Innate-Pharma, Roche, and BMS. No potential conflicts of interest were disclosed by the other authors.
Conception and design: G. Zalcman, A. Vincent-Salomon, F. Mechta-Grigoriou
Development of methodology: Y. Kieffer, H.R. Hocine, G. Gentric, L. Albergante
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): H.R. Hocine, G. Gentric, F. Pelon, B. Bourachot, S. Lameiras, C. Bonneau, A. Guyard, S. Baulande, G. Zalcman, A. Vincent-Salomon
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y. Kieffer, C. Bernard, L. Albergante, C. Bonneau, K. Tarte, A. Zinovyev, F. Mechta-Grigoriou
Writing, review, and/or revision of the manuscript: Y. Kieffer, K. Tarte, A. Zinovyev, G. Zalcman, A. Vincent-Salomon, F. Mechta-Grigoriou
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. Guyard
Study supervision: F. Mechta-Grigoriou
We thank Drs. Virginie Mieulet and Josh Waterfall for critical reading of the manuscript, as well as Dr. Emanuela Romano, Dr. Itay Tirosh, and the GDR 3697 Micronit for fruitful discussions. We are grateful to Charlotte Martinat from the Diagnostic and Theragnostic Department, Arnaud Meng from the “Stress and Cancer” lab (U830), as well as Renaud Leclere and André Nicolas from the experimental pathology platform at the Institut Curie for help and advice. We are also grateful to Celine Namour and Zohra Brouk for clinical data management at the Early Phase Clinical Research Unit, Thoracic Oncology Department, CIC1425/CLIP2 Paris-Nord, from Bichat Hospital (Assistance Publique-Hôpitaux de Paris). Y. Kieffer, C. Bonneau, G. Gentric, and H.R. Hocine were supported by the Institut National du Cancer, INCa (INCa-DGOS-9963; INCa-11692), SIRIC (INCa-DGOS-4654), the Medical Research Foundation (FRM) and the Foundation ARC (AAP SIGN'IT 2019), and the foundation “Chercher et Trouver.” A. Zinovyev is partly supported by the Agence Nationale de la recherche (ANR-19-P3IA-0001). The NGS platform was supported by a grant from France Génomique Consortium (ANR-10-INBS-09-08). The experimental work was supported by grants from the Ligue Nationale Contre le Cancer (Labelisation), Inserm (PC201317), Institut Curie (Incentive and Cooperative Program Tumor Micro-environment PIC TME/T-MEGA, PIC3i CAFi), Bristol-Myers Squibb Foundation (call 2018 for Research in Immuno-oncology), ICGex (ANR-10-EQPX-03), SIRIC (INCa-DGOS-4654), INCa (STROMAE INCa-DGOS-9963, CaLYS INCa-11692, and INCa-DGOS-Inserm-12554). F. Mechta-Grigoriou acknowledges the French Pink Ribbon Association “Le cancer du sein, Parlons-en” and the Simone and Cino del Duca Foundation for attribution of their “Grand Prix.” F. Mechta-Grigoriou is very grateful to all her funders for providing support throughout the years.