Fibroblast cell clusters were identified using the Louvain algorithm at a resolution of r = 0.5, implemented through the “Clustering” function of BBrowserX (BioTuring Inc., California, USA) (RRID: SCR 025984) (11), which initially yielded four distinct fibroblast clusters. To capture disease-associated gene expression changes, differential expression (DEG) analysis was performed between relevant cell groups within each cluster. The thresholds applied for DEG analysis included an average log₂ fold change (log₂ FC) > 0.5 and a false discovery rate (FDR) < 0.05, both of which are widely accepted criteria for identifying significantly regulated genes. Volcano plots were generated to visualize these results, enabling rapid assessment of genes that were significantly up- or downregulated and providing a basis for selecting candidate genes implicated in pathogenesis.

To refine classification, marker gene prioritization first incorporated coverage-based metrics—Within-cluster Coverage, Outside-cluster Coverage, and weighted log₂ FC (Wlog₂ FC)—to better capture both the prevalence and discriminatory power of candidate genes. We then applied stringent differential marker selection (highlog₂ FC, low FDR), ranked candidates by specificity and magnitude of upregulation, and assessed their biological relevance. This integrative framework enabled the systematic annotation of CAF clusters and highlighted ambiguous populations requiring cautious interpretation.

Each fibroblast cluster was further characterized by applying the “Marker Genes” function, which facilitated the assignment of putative CAF subtypes through comparison with established or novel cell-type markers. To confirm the robustness and specificity of these assignments, fibroblasts were separated from other cell types using the “Sub-Cluster” function. Enrichment analyses were subsequently performed on the DEG sets of each fibroblast cluster under both conditions (NC – Normal Control, BC – Breast Cancer) to refine subtype classification.

After defining fibroblast clusters and assigning putative subtype identities, we next performed molecular and functional profiling to delineate their biological programs and potential interactions within the tumor microenvironment. Two complementary analyses were conducted: pathway enrichment analysis of subtype- specific DEGs and inference of intercellular communication networks.

For pathway enrichment analysis, DEGs of each CAF cluster were submitted to the Enrichr platform (RRID: SCR 001575) (12), with functional annotations derived from multiple curated resources including Reactome (RRID: SCR 003485) (13), Wikipathways (RRID: SCR 002134) (14), and the Gene Ontology (GO) biological processes database (RRID: SCR 002811) (15). Significance was assessed using adjusted p-values with FDR < 0.05. This approach enabled the systematic identification of pathways enriched in each CAF subtype, providing insight into distinct transcriptional programs underlying extracellular matrix remodeling, angiogenesis, metabolism, protein synthesis, and immune modulation.

To evaluate potential cell–cell interactions, ligand–receptor pairing was analyzed using CellPhoneDB (v2.1.7) (RRID: SCR 017054) (16) and cross-validated with BBrowserX’s built-in cell-cell communication inference tool. Only statistically significant interactions (p-values < 0.05, permutation test) were retained. We specifically examined pathways with known relevance to tumor-immune crosstalk. Interaction networks were visualized to highlight both outgoing (CAF-derived ligands) and incoming (CAF-expressed receptors) signaling axes, enabling comparative mapping across CAF subtypes. This integrated molecular and functional profiling strategy provided the foundation for subsequent interpretation of CAF subtype-specific roles in shaping the tumor microenvironment and modulating responses to immune checkpoint blockade.

To examine the spatial organization of CAF programs identified in our scRNA-seq analysis, we analyzed a publicly available breast cancer spatial transcriptomics dataset generated using the 10x Genomics Visium FFPE platform (2021) (RRID: SCR 023571) (17). The dataset was accessed via Talk2Data (18) and processed using the SpatialX platform (BioTuring Inc.) (19). The analyzed tissue corresponds to an FFPE section (Block 738811QB, Section 1) from a grade II breast carcinoma of a 73-year-old Asian female, encompassing regions of ductal carcinoma in situ and invasive carcinoma. As no information on treatment status or clinical response was available, the analysis was restricted to baseline spatial organization. Spatial spots were clustered using the Louvain algorithm (resolution = 5). CAF subtypes were assigned based on marker gene signatures derived from our scRNA-seq analysis and projected onto tissue coordinates to assess their spatial distribution.

To investigate the functional heterogeneity of CAFs within the tumor microenvironment, we identified four distinct CAF subtypes based on their gene expression profiles: CAF1 as vascular CAFs (vCAF), CAF2 as myofibroblastic CAFs (myCAF), CAF3 as inflammatory CAFs (iCAF), and CAF4 as antigen-presenting CAF-like (apCAF-like) CAFs. Each subtype exhibited unique molecular signatures and pathway enrichments, suggesting distinct roles in tumor progression and therapy resistance.

CAF1 exhibited a robust gene expression profile associated with vascular functions, with key markers including Notch Receptor 3 (NOTCH3), Melanoma Cell Adhesion Molecule (MCAM), Cytochrome C Oxidase Subunit 4I2 (COX4L2), HIG1 Hypoxia Inducible Domain Family Member 1B (HIGD1B), and Cadherin 6 (CDH6) (Figure 2A). NOTCH3, involved in angiogenesis and endothelial cell signaling, reinforced the vascular phenotype of CAF1 (20). MCAM and CDH6, adhesion molecules critical for cell-cell interactions, suggest CAF1’s role in vessel formation and stabilization (21). The NOTCH Regulated Ankyrin Repeat Protein (NRARP) gene, regulating NOTCH signaling, further supports the angiogenic profile (Supplementary Figure S2A; Supplementary Data 2) (22). Additionally, the expression of Gap Junction Protein Alpha 4 (GJA4) and G Protein-Coupled Receptor 4 (GPR4), along with Potassium Voltage-Gated Channel Subfamily A Member 5 (KCNA5), RAS Like Glutamate Rich (RERGL), and Calsequestrin 2 (CASQ2), highlights CAF1’s potential role in modulating endothelial function and vascular responses to tumor growth (23–25). Pathway analysis further enriched CAF1 in mitochondrial energy metabolism pathways, including the Electron Transport Chain, Mitochondrial ATP Synthesis, and the TCA cycle (Citric Acid cycle), underscoring its energetic support for tumor progression. This, combined with pathways related to RNA processing, splicing, and VEGFA–VEGFR2 signaling, indicates CAF1’s active involvement in angiogenesis, vessel stabilization, and possibly immune exclusion via vascular-mediated mechanisms (Figure 2B; Supplementary Data 3).

Exhibiting a transcriptional signature indicative of myofibroblastic differentiation and ECM remodeling, CAF2 expresses key markers such as Maternally Expressed 3 (MEG3), KCNQ1 Opposite Strand/Antisense Transcript 1 (KCNQ1OT1), Integrin Subunit Alpha 11 (ITGA11), Ankyrin Repeat Domain 36C (ANKRD36C), Myocardial Infarction Associated Transcript (MIAT), and ADAM Metallopeptidase With Thrombospondin Type 1 Motif 6 (ADAMTS6) (Figure 2A). ITGA11, an integrin involved in ECM attachment and fibroblast migration, highlights CAF2’s role in promoting tissue stiffness (26). The long noncoding RNAs Nuclear Paraspeckle Assembly Transcript 1 (NEAT1), MIAT, Xist Ribonucleoprotein (XIST), and KCNQ1OT1 suggest transcriptional reprogramming, typical of fibroblast activation and fibrosis (Supplementary Figure S2B; Supplementary Data 2) (27). ADAMTS6, a metalloproteinase, plays a critical role in ECM turnover, reinforcing CAF2’s function in ECM remodeling (28). Pathway analysis reveals strong enrichment of ECM-related pathways, such as Collagen-Containing Extracellular Matrix, Extracellular Matrix Organization, and Collagen Formation, confirming CAF2’s role in desmoplasia, mechanotransduction, and stromal stiffening (Figure 2B; Supplementary Data 3).

A gene signature enriched in inflammatory and immunomodulatory pathways characterizes CAF3, with key markers such as Microfibril Associated Protein 4 (MFAP4), Platelet Derived Growth Factor Receptor Like (PDGFRL), Complement C3 (C3), Insulin Like Growth Factor 1 (IGF1), and MAF BZIP Transcription Factor B (MAFB) (Figure 2A). C3, a complement system component, indicates CAF3’s role in immune modulation through inflammation and immune cell recruitment (29). PDGFRL, a receptor involved in stromal-immune interactions, and IGF1, promoting tumor survival, further support this function (30, 31). Pathway analysis revealed a significant intensification of secretory programs within this subtype (Figure 2B, Supplementary Data 3). The Complement and Coagulation Cascades (WP558) pathway, featuring SERPING1, C1S, C1R, and Complement Factor D (CFD), establishes iCAFs as a primary source of innate immune modulators. Concurrently, the Cytokine-Cytokine Receptor Interaction (WP5473) axis, involving IL6ST, CXCL12, CXCL14, and CCL2, positions iCAFs as a central signaling hub. Additionally, iCAFs exhibit high metabolic plasticity via the NRF2 Pathway (WP2884), characterized by antioxidant genes such as Superoxide Dismutase 3 (SOD3), Glutathione Peroxidase 3 (GPX3), and Hepatocyte Growth Factor (HGF). Beyond immunomodulation, pro-angiogenic factors like Retinoic Acid Receptor Responder 1 (RARRES1) and Vascular Endothelial Growth Factor D (VEGFD) suggest that iCAFs foster metastasis through vascular remodeling (Supplementary Figure S2C; Supplementary Data 2) (32, 33). These findings establish iCAFs as essential orchestrators of an inflammatory, tumor-supportive stroma.

CAF4, designated as an apCAF-like subtype, exhibited a mosaic transcriptional signature with features of stromal, epithelial, and immune-related cells. Key markers such as C-X-C Motif Chemokine Ligand 14 (CXCL14), Keratin 19 (KRT19), Mucin Like 1 (MUCL1), CD74 Molecule (CD74), and Major Histocompatibility Complex, Class II, DP Alpha 1 (HLA-DPA1) suggest a hybrid phenotype, with CXCL14 involved in immune cell recruitment and KRT19 marking epithelial-like features (Figure 2A). CD74, an antigen-presenting molecule, supports CAF4’s potential role in immune modulation, similar to that of antigen-presenting CAFs (apCAF) (34). To ensure cellular identity and exclude potential doublets or technical artifacts, we performed stringent quality control and expression profiling. CAF4 cells consistently displayed gene counts within a normal range (Number of genes < 3,000), arguing against technical artifacts combining multiple cell types (Supplementary Figure S2E). Furthermore, while low-level expression of epithelial-associated markers and other transcripts such as HLA-DPA1, MUCL1, and Trefoil Factor 3 (TFF3) were detectable (Supplementary Figure S2D; Supplementary Data 2) (35, 36), CAF4 cells maintained high expression of core fibroblast markers, including Collagen Type I Alpha 1 Chain (COL1A1), Collagen Type III Alpha 1 Chain (COL3A1), Decorin (DCN), and Lumican (LUM), confirming their lineage as bona fide fibroblasts (Supplementary Figure S2F). These findings, distinct clustering observed in UMAP embedding, confirm that CAF4 represents a specialized fibroblast state adapted to the immune-rich niche rather than epithelial contamination (Figure 2C) (37). Pathway analysis further reveals that CAF4 is enriched in mitochondrial energy metabolism (e.g., Electron Transport Chain and Mitochondrial ATP Synthesis) and translation-related programs (e.g., Cap-Dependent Translation Initiation and Ribosomal Subunit Joining) (Figure 2B), indicating elevated energetic and biosynthetic activity. Importantly, CAF4 also shows significant enrichment of antigen processing and presentation pathways, including cross-presentation of exogenous antigens via endosomal compartments and MHC class I–mediated antigen presentation, consistent with an apCAF-like functional phenotype (Supplementary Data 3).

We assessed the association between the characterized CAF subtypes and the anti-PD-1 response (using Welch’s t-test) (Figure 3A). The vCAF, iCAF, and apCAF-like subtypes were significantly enriched in non-responders compared to responders during treatment. In contrast, the myCAF subtype remained at low, unchanged levels. At the pre-treatment baseline, only the apCAF-like subtype showed modest enrichment in non-responders, suggesting a potential baseline predictive value. Upon treatment, both the vCAF and iCAF subtypes were elevated in non-responders, reflecting their dynamic roles in sustaining immunosuppressive signaling and driving adaptive therapy resistance. The apCAF-like subtype remained higher in non-responders at both pre-treatment and on-treatment stages, reinforcing its stable contribution to a resistant stromal niche. Conversely, responders consistently showed low levels of these three CAF subtypes, indicating a less suppressive stromal environment that may facilitate T cell infiltration and anti-tumor immunity. Collectively, these results suggest that the apCAF-like/admixture subtype serves as a baseline predictor, while the vCAF and iCAF subtypes are numerically associated with adaptive resistance; the myCAF subtype appears to have minimal impact on the treatment outcome.

To explore context-specific roles, we examined the vCAF, iCAF, and apCAF-like subtypes within different breast cancer types (Figure 3B). The vCAF and apCAF-like subtypes were found to be enriched in TNBC. Crucially, the iCAF subtype showed a selective increase in ER+ tumors during treatment, highlighting a differential response mechanism. HER2+ samples were excluded from comprehensive statistical analysis due to limited sample size (n = 3). However, preliminary cross-subtype comparisons (Figure 3C) indicated that the vCAF subtype appeared higher in HER2+ compared to TNBC, and the iCAF subtype was significantly enriched in ER+ relative to TNBC. The myCAF and apCAF-like subtypes showed no significant differences across these breast cancer subtypes. These findings underscore the heterogeneity of CAF distribution and function across breast cancer subtypes, confirming the apCAF-like subtype as a baseline predictor and the vCAF and iCAF subtypes as adaptive drivers of resistance, with the iCAF subtype showing a distinct, selective enrichment in ER+ tumors during therapy. The observations related to HER2+ tumors remain inconclusive due to sample limitations.

To decode the cellular determinants governing therapeutic efficacy, we quantified the global landscape of intercellular communication within the tumor microenvironment, revealing a profound bifurcation in interaction trajectories contingent upon treatment outcomes (Figure 4A; Supplementary Data 4). In non-responders, disease progression was characterized by Stromal Fortification, where vCAFs and iCAFs intensified direct supportive signaling to cancer cells, effectively shielding the tumor niche. In sharp contrast, responders exhibited a distinctive Stromal Remodeling, marked by the synchronous attenuation of tumor-supportive interactions. Notably, while CAF lineages underwent extensive reconfiguration—specifically with iCAFs redirecting signals toward T cells and myeloid cells—the intrinsic interaction repertoire of T cells remained remarkably stable (Δ ≈ 0 to −1). This signaling stasis in the T-cell compartment serves as a critical baseline, demonstrating that the observed therapeutic shift is not driven by an autonomous expansion of immune signaling repertoires, but is instead orchestrated by CAFs actively rewriting the intercellular script. Analysis of high-intensity interaction networks (≥ 42 interactions; Figure 4B) corroborated that the CAF signaling hierarchy in responders shifted decisively from a tumor-supporting to an immune-promoting profile.

To elucidate the molecular mechanisms driving these macroscopic shifts, we interrogated specific ligand–receptor pairs (Figure 4C; Supplementary Data 5), identifying a sophisticated two-pronged reprogramming strategy induced by anti-PD-1 therapy. First, the phenotypic re-education of the iCAF lineage. Analysis of pre-treatment baselines revealed that iCAFs initially engaged T cells via the Amyloid Precursor Protein (APP)–CD74 axis. In responders, this specific iCAF-derived signal was significantly attenuated, replaced by the effective activation of the CXCL9–CXCR3 axis. This molecular switch transforms iCAFs from a physical barrier into an immune recruitment hub, facilitating the infiltration of CD8⁺ and Th1 T cells. Crucially, while iCAF-derived APP signaling diminishes, the APP pathway itself is not extinguished but rather functionally reallocated to distinct stromal subsets (vCAF and apCAF) to sustain antigen presentation, as elucidated in the pathway analysis (Section 3.5). Second, the functional disarmament of protective vCAF and myCAF populations.

In non-responders, vCAFs strongly expressed signals associated with cancer stemness (JAG1–NOTCH2) and basement membrane reinforcement (COL4A2–SDC1). In responders, this specific malignant signalling axis targeting cancer cells was effectively dismantled, thereby removing key survival inputs. Similarly, myCAFs in responders displayed an attenuation of chemical defense mechanisms; interactions involving the complement-regulatory protein CD55 (CD55–ADGRE5) and the anti-phagocytic signal (COMP–CD47) were significantly diminished. Furthermore, apCAF-like cells in the responder group exhibited upregulated MHC Class I molecules (HLA-B/C), directly contributing to antigen presentation. Collectively, these data demonstrate that effective anti-PD-1 therapy necessitates a coordinated disruption of stromal defense systems alongside immune recruitment.

The detailed ligand-receptor mapping of CAF-mediated cell–cell interactions and signaling dynamics (Figures 5A, B; Supplementary Data 6) unveils subtype-specific reprogramming with distinct functional consequences. In the case of vCAF, reprogramming is characterized by the upregulation of the Thrombospondin-1–CD47 (THBS1–CD47) signaling pathways. These alterations translate into moderate-strength interactions with endothelial and myeloid populations within the responder group (Figure 5B). This enhancement suggests a functional shift toward immune-modulatory activity and immune cell recruitment (38, 39). Concurrently, tumor-promoting signaling axes, such as Midkine–Nucleolin (MDK–NCL) and Fibronectin 1–Integrin (FN1–integrin), are downregulated, which indicates a reduction in ECM-mediated support for tumor growth (40, 41).

In contrast, myCAF exhibits selective functional remodeling rather than a universal attenuation of activity. The downregulation of structural signals like MDK–NCL and COMP–CD47 (Figure 5A) suggests a dismantling of the physical barriers that typically exclude immune cells from the tumor microenvironment (42). Crucially, myCAFs within responders maintain strong, high-affinity interactions with T cells and endothelial cells (Figure 5B). This stark contrast suggests a phenotypic switch from a barrier-forming phenotype to an immune-permissive state, which actively supports T-cell trafficking and vascular normalization. Similarly, iCAF demonstrates a clear shift towards immune-facing signaling. iCAFs emerge as a central hub of communication in responders, marked by peak interaction strengths with myeloid cells and T cells (Figure 5B). This shift is driven by heightened THBS1–CD47 signaling and a reduction in integrin-mediated support, suggesting that iCAFs orchestrate a permissive microenvironment that facilitates robust immune engagement.

The CXCL12–CXCR4 axis, however, exhibited a profound functional divergence between pre-treatment and on-treatment stages, characterized by opposing dynamics in vCAF and apCAF-like populations. (Figures 5C, D). In non-responders, therapeutic intervention induced a pathological intensification of vCAF-derived signaling toward plasmacytoid dendritic cells, a feature markedly attenuated in responders. Notably, while vCAF-to-T cell communication was elevated at the pre-treatment baseline in the resistance group, anti-PD-1 therapy triggered a secondary, broad-spectrum surge in apCAF-mediated CXCL12 signaling toward T cells, pDCs, and B cells. Although this apCAF-driven axis was detectable across all patients post-treatment, its magnitude was significantly more pronounced in non-responders. These dynamics suggest that in the context of therapeutic failure, the CXCL12–CXCR4 axis does not facilitate productive immune recruitment but instead orchestrates an immunosuppressive niche characterized by pDC sequestration and dysfunctional lymphoid entrapment.

To define the molecular determinants underlying these divergent trajectories, we performed a detailed pathway analysis of ligand–receptor pairs exclusive to each response group (Table 1). This analysis revealed a stark dichotomy in signaling programs that was strictly compartmentalized by CAF subtype. Specifically, Responders were characterized by the subtype-restricted activation of Amyloid Precursor Protein (APP), NOTCH, and Midkine (MK) pathways, whereas the Non-response group was dominated by the Thrombospondin-2 (THBS2) axis (43–45). This indicates that therapeutic distinctness is driven by precise functional modules within specific CAF populations rather than ubiquitous stromal activation.

Specifically, in the responder group, the APP pathway operated as a prominent autocrine and paracrine immune-supportive module, predominantly orchestrated by vCAF and apCAF-like populations. Here, vCAF and apCAF-like cells served as primary ligand sources, interacting with CD74 receptors on a diverse range of targets—including B cells, endothelial cells, mast cells, and myeloid cells. Additionally, vCAF-derived APP engaged SORL1 on γδ T cells, while apCAF-like-derived APP targeted TREM2+TYROBP on myeloid lineages. This specific connectivity network promotes antigen cross-presentation and fosters an immune-permissive microenvironment. Similarly, NOTCH signaling in responders was identified as a specific stromal-vascular crosstalk axis driven exclusively by myCAF-derived JAG1. This ligand engaged NOTCH1/4 on endothelial cells and NOTCH2/3 on stromal and myeloid subsets, potentially supporting vascular normalization and stromal remodeling without triggering immunosuppression. Furthermore, the MK pathway emerged as a critical lymphoid-niche organizing mechanism, where apCAF-like cells specifically targeted B cells via MDK–ITGA4+ITGB1 and MDK–SDC1 interactions. This mechanism likely favors B-cell recruitment and tertiary lymphoid structure (TLS) formation.

Conversely, the Non-response group was dominated by a unique broad-spectrum suppressive broadcast driven by apCAF-like cells via the THBS2 pathway. Unlike the spatially coordinated signaling in responders, this resistance-associated module involved apCAF-like-derived THBS2 engaging CD47 on a broad spectrum of targets (B cells, iCAFs, malignant cells, myeloid cells, pDCs, and T cells), as well as NCL, SDC1/4, and CD36 on stromal and endothelial compartments. This specific apCAF-like–THBS2–CD47 axis likely contributes to therapeutic resistance by sustaining an immunosuppressive and exclusionary tumor microenvironment. Collectively, the identification of these response-specific ligand–receptor pairs highlights APP, NOTCH, and MK as potential biomarkers for effective anti-PD-1 therapy, while positing the apCAF-driven THBS2–CD47 axis as a critical target to overcome therapeutic resistance.

To examine the spatial distribution of CAF subtypes within an intact tumor architecture, we analyzed an independent breast cancer spatial transcriptomics dataset (17). Using a two-step spatial mapping strategy, we first assessed whether marker genes derived from our scRNA-seq analysis delineated distinct CAF populations in situ, and subsequently evaluated their anatomical localization within the tumor microenvironment. Established reference markers, including ANTXR1 for myCAFs and PI16 for iCAFs, together with the pan-fibroblast marker FAP (46), were used to contextualize CAF identity. In parallel, subtype-specific markers identified in our scRNA-seq analysis, ITGA11 for myCAFs and IGF1 for iCAFs (Figure 2A), were spatially mapped alongside these reference markers (Figure 6A). This approach delineated two spatially separable CAF populations, defined as ITGA11⁺ANTXR1⁺FAP⁺ myCAFs and IGF1⁺PI16⁺FAP⁺ iCAFs, consistent with their transcriptomic identities.

We next examined the spatial neighborhoods associated with each CAF subtype. ITGA11⁺ANTXR1⁺ myCAFs were predominantly localized within fibrotic stromal regions and were spatially segregated from immune cell–enriched areas (Figure 6B). This spatial distribution is consistent with an immune-excluded stromal architecture and suggests a structural association between myCAF-enriched stroma and limited immune cell accessibility (26, 47). In contrast, the IGF1⁺PI16⁺ iCAF population exhibited a distinct spatial organization, preferentially localizing to immune-associated regions of the tumor microenvironment (Figure 6C). iCAFs were frequently observed in proximity to macrophages and T cells, indicating a baseline anatomical configuration permissive for close stromal–immune spatial association (31, 48). Although spatial transcriptomics captures a static snapshot, this organization provides the spatial context required for the subtype-specific ligand–receptor interactions identified in our network analyses.