Stromal Cell Subsets Modulate T-cell Infiltration in Early Breast Cancer

Abstract

Recent studies of the tumor microenvironment (TME) have elucidated the heterogeneity of stromal cells, with increasing evidence suggesting that stromal subsets play an important role in regulating antitumor immunity in breast cancer. However, the functional diversity of these cells within the tumor-immune microenvironment and how they interact with immune cells in a spatial and clinical context remain poorly understood. We performed multiplex immunofluorescence on tumor microarrays from two cohorts consisting of 591 patients with breast cancer to assess the abundance and spatial co-localization of stromal and immune subsets and their correlation with clinicopathologic features and patient outcomes. We found that stromal cells were spatially distinct. A perivascular-like subset that was disseminated throughout the stroma rather than restricted to vessel-adjacent regions was enriched in an immune-cold environment and associated with T-cell exclusion. An inflammatory-like cancer-associated fibroblast subset was associated with segregation of T cells from cancer cells. Our findings highlight the differential impact of stromal subsets on immune infiltration and activation within the breast cancer TME with potential implications for patient outcomes.

Significance:

This study characterized spatially defined interactions between stromal and immune subsets in large clinical cohorts of early breast cancer. Findings from this study will fill gaps in current knowledge in how diverse stromal cell subsets, particularly novel subsets, interact with immune cells in a clinically relevant context.

Introduction

Breast cancer is the most commonly diagnosed cancer in women, and despite advancements in treatment over the last few decades, it remains the leading cause of cancer-related death among females worldwide (1). Recent studies into the tumor microenvironment (TME) have revealed complex functional roles of stromal cells in tumorigenesis, progression, and suppression (2). Although there is increasing evidence that stromal cells affect tumor immunity, the diversity of stromal cell subsets and their interactions with immune cells in a clinically relevant context remain poorly understood.

Recent advances in technologies such as single-cell RNA sequencing (scRNA-seq) have allowed the characterization of tumors at tremendous resolution, and the results highlight the heterogeneity of stromal and immune cells within the TME. Different cancer-associated fibroblast (CAF) subpopulations or states with discrete phenotypic, transcriptional, and functional properties have been described in various cancer types (3–7). In breast cancer, our previous scRNA-seq study has revealed four distinct populations of stromal cells: myofibroblast-like CAF (myCAF), inflammatory-like CAF (iCAF), and two perivascular-like cells (PVL) in differentiated and immature states (8, 9). Each of these subpopulations displayed distinct morphologies, spatial relationships, and functional properties. Similarly, scRNA-seq of immune cells showed that a significant heterogeneity exists in the infiltrating T-cell population (10). The effectiveness of CD8⁺ T cells is regulated by immune checkpoint molecules, including PD1 and its ligand PD-L1, and these are the targets of most currently available immunotherapies.

One of the major limitations of single-cell technologies is the lack of information on the structural organization of cells within a tumor, which has increasingly been recognized as a factor that contributes to the heterogeneity, variable tumor behaviors, and differential response to treatments observed in most cancers (11). Spatial profiling has hence been explored to provide a more comprehensive view of the intricate cellular makeup of solid tumors, enabling examination of different cell types and their locations and interactions within different layers of spatial organization. Integration of this with clinical data from large cohorts enables the exploration of spatially defined TME components as predictors of treatment response and patient outcomes, with the potential to discover novel biomarkers and therapeutic targets, thereby improving the precision of cancer diagnosis and management (12).

In this study, we performed multiplex immunofluorescence (mIF) on tumor microarrays (TMA) from two independent primary breast cancer cohorts using a panel of stromal and immune markers defined using our scRNA-seq results. The abundance and spatial co-localization of annotated stromal and immune subsets were linked with clinicopathologic data to assess their clinical significance.

Materials and Methods

Clinical cohorts

This study consists of TMA cores from 591 patients with breast cancer with luminal and triple-negative breast cancer (TNBC) subtypes. All tumors were sampled before treatment. Luminal patients (n = 369) were recruited as part of the Breast Boost trial, in which patients received wide local excision with whole breast irradiation with randomization to cavity boost or no boost (13). The TNBC cases (n = 222) were identified by a review of the oncology database at St George Hospital between 2004 and 2019 (14). Tumors were classified into molecular subtypes according to the following definitions: luminal A: ER⁺ ≥ 1%, PR⁺ > 20%, HER2⁻, Ki67 < 14%; luminal B: ER⁺, PR⁺ ≤ 20%, and/or HER2⁺ and/or Ki67 ≥ 14%; and TNBC: ER/PR⁻ (<1%) and HER2⁻.

TMA slides were constructed with 3 × 1 mm cores sampled from the periphery of tumor blocks marked by a breast pathologist on an hematoxylin and eosin (H&E) slide. Paraffin sections were cut at 4 μm onto Superfrost glass slides (Thermo Fisher Scientific) and stained with H&E on a Leica Bond Rx automated immunostainer machine (Leica Biosystems).

mIF

Opal 9 was used for mIF imaging (Akoya Biosciences), with primary antibody conditions optimized for 3,3′-diaminobenzidine staining applied to both monoplex and multiplex optimizations. Each biomarker antibody was paired with a specific Opal fluorophore based on the coexpression patterns of biomarkers in the tissue and their anticipated protein expression levels (Supplementary Table S1).

Image analysis

Slides were scanned using the Vectra Polaris 3.0 (Akoya Biosciences) using 40× magnification. Cell type identification and classification were performed in accordance with the method described by Wang and colleagues (14). Briefly, TMA images were analyzed using QuPath v0.2.3 (RRID: SCR_018257, https://qupath.github.io). All TMA files and their corresponding TMA maps were imported using the TMA module. Tissue core identification numbers generated by the TMA dearrayer were verified before analysis. Tissue regions were segmented into tumor and stromal compartments using a pixel-based machine learning classifier trained using PanCK staining. Cell segmentation was based on DAPI nuclear staining and implemented using the inbuilt cell detection algorithm. Biomarker phenotyping was conducted through object classification using a random forest algorithm applied to mean nuclear intensity measurements to minimize staining leakage from nearby cells. The final combined classifier was applied to all TMA cores. Cell types were defined by expression or coexpression of markers, including epithelial (PanCK⁺), endothelial (CD31⁺), myCAF (CD140b⁺αSMA⁺), iCAF (CD140b⁺αSMA⁻CD146⁻), differentiated PVLs (dPVL; CD146⁺THY1⁻), immature PVLs (imPVL; CD146⁺THY1⁺), PD1⁺ CD8 T cells (PD1⁺CD8⁺), and other CD8 T cells (PD1⁻CD8⁺; Supplementary Table S2).

TMA core measurements were linked with corresponding clinical data using patient ID. For patients with more than one core, the median values of these cores were used for analysis. The abundance of cell types was measured for each core as the percentage of cells per total number of detections. Cell proportions were computed within the stromal and tumor regions separately. To account for differences in locations, the proportion of immune cells was taken from the entire core, whereas for stromal cells, proportions were taken from stromal regions only to reduce the impact of misclassified stromal cells in tumor regions.

Spatial analysis

To investigate the spatial proximity between cell types, we measured the proportion of cell type A with nearby cell type B within a radius of 30 to 100 μm, with an interval of 10 μm. Cell type A with at least one cell type B within 30 μm were determined as B-adjacent cells. Type A cells without type B cells within 100 μm were defined as B-distal cells. We also measured the minimal distance between a pair of cell types using the calculate_minimum_distances_between_celltypes function from the SPIAT package (15). The patient-level metrics were summarized as the median across TMA cores.

Immune phenotypes

To characterize patterns of immune distribution, we classified cores into three categories: immune-cold, immune-intermixing, and immune-segregated. Immune-cold was determined by a CD8⁺ T-cell infiltration (sum of PD1⁺CD8⁺ and PD1⁻CD8⁺ proportions) below the median. We then examined the co-localization of CD8⁺ T and epithelial cells using the area under the curve (AUC) of the cross-K function (15). A positive AUC score of the cross-K function indicates co-localisztion of these cells, whereas a negative AUC score indicates a separation of CD8⁺ T cells from epithelial cells. We classified cores with CD8⁺ T-cell proportion over the median and a negative cross-K AUC as immune-segregated. These cores displayed immune aggregates forming outside the epithelial nest. An immune intermixing group was defined with CD8⁺ T-cell proportion over the median and a positive cross-K AUC.

Association of spatial features with clinical variables

We compared the enrichment of stromal subset–related features in groups of patients stratified by clinical characteristics. Features include fibroblast proportions in the stromal region, minimal distance between pairs of cells, proportion of cell type A with cell type B present within a 30-μm neighborhood, and the normalized mixing score defined by the interactions between cell types A and B divided by the sum of interactions between A–A and B–B (Supplementary Table S3; ref. 15). The Wilcoxon rank-sum test was performed using the compare_means function from ggpubr (0.6.0, RRID: SCR_021139) for each feature to determine its difference between groups classified by clinical characteristics. P value was corrected using the Benjamini–Hochberg (BH) method.

Statistical analysis

Statistical analysis was completed using R 4.4.2. Univariate and multivariate survival analyses were conducted using Cox proportional hazards modeling, with a P value < 0.05 considered significant. Clinical variables with known prognostic values, including age, lymph node metastasis, grade, tumor size, and molecular subtypes (luminal A vs. luminal B) were included in the multivariate survival analyses. Survival predictions were represented using Kaplan–Meier (KM) plots (survival 3.7.0, RRID: SCR_021137). Association between specific stromal and immune phenotypes and clinicopathologic features were examined using Kruskal–Wallis tests and Pearson correlations (stats 4.4.2, RRID: SCR_025678). Statistical significance for comparing patient groups was determined by a P value < 0.05 using the Wilcoxon rank-sum test.

Xenium data generation

Formalin-fixed, paraffin-embedded (FFPE) blocks from three TNBC tumors were sectioned at 5 μm and placed onto Xenium slides following the Xenium In Situ FFPE Tissue Preparation Guide (10x Genomics, CG000578). Slides were then deparaffinized and decrosslinked (10x Genomics, CG000580). A predesigned and add-on gene panel of 5,101 genes (100 custom add-on genes additional to the default 10× 5k panel; Supplementary Table S4) was added to the tissue following instructions in Xenium Prime (5K) In situ Gene Expression (10x Genomics, CG000760). Cell segmentation reagents were applied to assist segmentation of nuclei and cell bodies (10x Genomics, CG000760). Xenium slides were processed on the Xenium Analyzer (serial number XETG00531, software version 3.4.1.0, software analysis version Xenium-3.3.0.1) for imaging and analysis (Decoding and Imaging user guide, 10x Genomics, CG000584).

Xenium cell type annotation and spatial analysis

Raw Xenium data were converted into AnnData objects using anndata (v5.2.1; ref. 16). Cells with fewer than 10 detected transcripts or fewer than 1 expressed gene were excluded. Cell annotation was performed based on a combined positivity of genes corresponding to the protein markers in the mIF panel (Supplementary Table S5). To identify epithelial cells, EPCAM was used as a surrogate of keratin genes encoding PanCK. Adjacent and disseminated PVLs were identified as PVLs (dPVL or imPVL) localized within 30 μm or beyond 100 μm to the nearest endothelial cell, respectively. Distance from PD1⁺CD8⁺ and PD1⁻CD8⁺ T cells to the nearest adjacent or disseminated PVLs was computed using var_by_distance function in Squidpy (RRID: SCR_026157) to assess the spatial proximity of T cells to PVL phenotypes (17).

Differential gene expression and pathway enrichment

The differential gene expression (DEG) analysis was performed using DESeq2 (1.46.0, RRID: SCR_000154; ref. 18). To reduce the false discovery of DEGs arising from transcript contamination from nearby cells, we constrained DEGs to genes expressed by at least 5% of PVLs in a breast cancer scRNA-seq dataset (8). Gene Ontology enrichment analysis was computed separately for upregulated and downregulated genes comparing between adjacent and disseminated PVLs. Genes were first filtered by significance [absolute log₂ fold change (FC) >0.5 and adjusted P value < 0.05], and then the top 100 genes from each subset (ranked by log₂ FC) were analyzed with clusterProfiler (4.14.6, RRID: SCR_016884; ref. 19). Enrichment was performed for Biological Process terms with BH adjustment.

Ethics

The Breast Boost study was approved by the Human Research Ethics Committee of St George Hospital, Sydney, Australia (ref No: 96/84), and written informed consent was obtained from all participants. Ethics approval for the TNBC cohort was granted by the South Eastern Sydney Local Health District Human Research Ethics Committee at the Prince of Wales Hospital, Sydney (Boost: HREC 96/16 and TNBC: HREC 2018/ETH00138), who granted a waiver of consent to analyze the tissue blocks. The TNBC samples profiled using Xenium Prime in this study were collected following protocols x13-0133, x16-018, and x19-0496. Ethical approval for this study was acquired by the Sydney Local Health Districts Ethics committee (Royal Prince Alfred Hospital zone) and St Vincent’s hospital Ethics Committee. Consent for the use of samples in this study was obtained from all patients prior to collection of tissue, and data were deidentified as per approved protocol. All studies were conducted in accordance with the Declaration of Helsinki.

Results

Patient demographics

A total of 1,356 TMA cores were processed, with a median of three cores per patient (range, 1–6). 868 TMA cores were from luminal tumors and 488 cores from TNBC tumors (Fig. 1A). In the luminal cohort, the majority of tumors were luminal A (n = 275, 75%) and invasive ductal carcinoma (n = 317, 85.9%). Other histology subtypes include lobular carcinoma (n = 34, 9.2%), micropapillary (n = 7, 1.9%), mucinous (n = 10, 2.7%), and tubulolobular (n = 1, 0.3%). Fifty percent (n = 184) of patients in this cohort received adjuvant endocrine therapy, 16% (n = 59) received chemotherapy, and 10% (n = 37) received both. The median length of follow-up for this cohort is 16 years for survival. Analysis of clinicopathologic features revealed that age, nodal status, and molecular subtypes (luminal A vs. luminal B) were independent predictors of patient survival when adjusted for other clinical factors, consistent with known results from the literature (Supplementary Table S6). The interventional endpoint of this trial, boost versus no boost, had no impact on patient outcome (13). For the TNBC cohort, the majority of patients had ductal histology (n = 204, 91.9%) and node-negative disease (n = 139, 62.6%). Of the total patients, 70.3% received adjuvant chemotherapy, including both anthracycline-based and anthracycline-free regimens. The median length of follow-up was 4.5 years for survival. Older age, node positivity, no chemotherapy, and large tumor size were associated with poorer survival, whereas a high tumor-infiltrating lymphocyte (TIL) score was associated with better survival in the univariate analyses. Multivariate analysis showed a significant correlation between age and nodal status and overall survival (OS) after adjusting for other significant clinical characteristics as confounders (Supplementary Table S7).

Figure 1.

Cohort description and cell type annotations. A, Cohort overview. B, Cell types of interest defined by the expression of markers. C, mIF images were processed by tissue segmentation and cell segmentation to map the distribution of individual cells in stromal and tumor regions. D, mIF images showing marker staining and cell type annotation. E, Heatmap showing z-score–normalized median intensities of markers across cell types. Stars indicate the key markers of cell types used for classification. F, Median percentages of cell types within cells with assigned classes across the entire TMA core, stromal and tumor regions. G, Proportions of iCAFs, myCAFs, endothelial cells, and PVLs within all detected cells in the stromal region in luminal and TNBC cohorts. H, Proportions of PD1⁻CD8⁺ T cells and PD1⁺CD8⁺ T cells within all detected cells in the stromal and tumor regions in luminal and TNBC cohorts. I, Compositions of classified cell types in the entire TMA cores per patient, with unclassifiable cells excluded. In the luminal cohort, the vertical dotted line separates patients with CD8 T% above and below the median CD8 T% of the TNBC cohort. Color bars below the bar plot denote the clinical variables per patient. [A, Created in BioRender (RRID: SCR_018361). Zhang, H. (2026) https://BioRender.com/y2inm3s.]

Distinct stromal and immune landscapes in luminal cancer and TNBC

mIF data were processed using tissue and cell segmentation to locate individual cells in the tumor or stromal regions. Cells were classified into one of the eight types based on the expression of eight markers, including epithelial (PanCK⁺), endothelial (CD31⁺), myCAF (CD140b⁺αSMA⁺), iCAF (CD140b⁺αSMA⁻CD146⁻), dPVLs (CD146⁺THY1⁻), imPVLs (CD146⁺THY1⁺), PD1⁺ CD8 T cells (PD1⁺CD8⁺), and other CD8 T cells (PD1⁻CD8⁺; Fig. 1B–D). The marker intensity was consistent with the cell type definitions (Fig. 1E).

Cell proportions were quantified as the number of cells normalized by the total number of detected cells within the entire TMA core, as well as within tumor and stromal regions separately. Compositions of classified cells were shown in Fig. 1F. Proportions per patient were summarized as the median across cores. Due to the sparsity of imPVLs, we merged dPVLs and imPVLs into a general PVL group for downstream analyses. TNBC tumors were enriched in endothelial cells compared with luminal tumors, whereas the proportion of iCAFs, myCAFs, and PVLs were significantly elevated in luminal tumors (P < 0.05; Fig. 1G).

Within the stromal compartment, iCAFs represented the most predominant cell type in both luminal and TNBC cases (median 20.59% in luminal and 7.89% in TNBC; Fig. 1G). The least abundant cell type is imPVLs, accounting for 0.036% in luminal and 0.041% in TNBC. PVLs account for 7.66% [standard deviation (SD) 6.45%] and 6.28% (SD 8.38%) of total detected cells in luminal and TNBC respectively, with a SD comparable with that of iCAFs and myCAFs (iCAF, SD 10.75% in luminal, 7.37% in TNBC; myCAF, SD 8.59% in luminal, 7.36% in TNBC; Fig. 1G). In contrast, endothelial cells showed relatively stable abundance across patients (SD 1.73% in luminal, 3.12% in TNBC), despite their presumed spatial association with PVL cells (Fig. 1G).

In both stromal and tumor regions, TNBC tumors harbored a higher abundance of PD1⁺CD8⁺ T cells than luminal cases (P < 0.05; Fig. 1H), aligning with the notion that TNBC is a more immune-inflamed subtype and can benefit from immune checkpoint therapy (20, 21). Surprisingly, luminal cases showed a higher level of PD1⁻CD8⁺ T in the tumor region than TNBC (P < 0.001; Fig. 1H). Despite luminal tumors containing a lower proportion of CD8⁺ T cells in general than TNBC (1.58% vs. 2.64%), 116 of 369 (31.4%) luminal tumors showed a CD8⁺ T-cell percentage higher than the median of TNBC tumors (Fig. 1I), highlighting heterogeneity in immune infiltration within luminal tumors. Additionally, the CD8⁺ T-cell percentage in luminal tumors increased with younger age at diagnosis (P = 0.025; Supplementary Fig. S1A), no lymph node metastasis (P = 0.013; Supplementary Fig. S1B), and low grade compared with median grade (P = 0.045; Supplementary Fig. S1C). No significant difference in CD8⁺ T-cell percentage was observed between luminal A and luminal B subtypes (Supplementary Fig. S1D) or between small (<20 mm) and large (≥20 mm) tumors (Supplementary Fig. S1E).

We calculated the ratio of PD1⁺CD8⁺ to total CD8⁺ T cells to quantify the prevalence of CD8⁺ T cells expressing inhibitory receptors. This ratio varied across patients, ranging from 0 to 0.55 in luminal (median = 0.05) and 0 to 1 (median = 0.26) in TNBC. Patients with higher overall CD8⁺ T-cell proportions tended to have a higher PD1⁺CD8⁺ ratio (luminal, Pearson’s R = 0.15, P = 0.005; TNBC, Pearson’s R = 0.22, P = 0.00086), suggesting that tumors more permissive of T-cell infiltration also exhibit increased frequency of inhibitory T-cell phenotypes.

Spatially distinct distribution of stromal subsets and correlation with T-cell infiltration

The three stromal subsets exhibited distinct spatial distributions. myCAFs were most closely associated with cancer cells, as indicated by their shortest median distance to tumor cells and their highest representation in cancer-adjacent compartments (Fig. 2A). In contrast, iCAFs and PVL cells were predominantly located at greater distances from tumor regions.

Figure 2.

Spatial distribution of stromal cell types and their correlation with CD8⁺ T-cell infiltration. A, Median distance from iCAFs, myCAFs, and PVLs to their nearest epithelial cells and the proportion of each stromal cell type with epithelial cells present within a 30 μm radius. B, Correlation between the log-transformed proportions of CD8⁺ T cells and total stromal cells. C, PD1⁻CD8⁺ T cell percentages in TMA cores. Cores are stratified into low and high groups based on the median proportions of myCAFs, PVLs, and iCAFs within the total stromal population in the stromal region.

We observed no significant association between overall CAF abundance within the stromal compartment and T-cell infiltration in luminal cases (Pearson’s R = 0.026, P = 0.62; Fig. 2B). In TNBC, however, there was a modest negative correlation (Pearson’s R = −0.14, P = 0.039; Fig. 2B). When stratifying by stromal subset, divergent associations with CD8⁺ T-cell infiltration emerged. Cores enriched in PVL cells exhibited reduced CD8⁺ T cell infiltration (P = 0.002 for luminal; P = 0.0026 for TNBC; Fig. 2C). Conversely, iCAF-enriched cores demonstrated increased CD8⁺ T-cell infiltration (P < 0.001 for luminal; P < 0.001 for TNBC; Fig. 2C), indicating a possible role for iCAFs in supporting or permitting T-cell entry into the TME.

Immune compositions and spatial distribution reveal three immune infiltration phenotypes

To further characterize immune infiltration patterns in breast cancer, we classified TMA cores into three immune phenotypes based on T-cell enrichment and tumor–T cell co-localization (Fig. 3A and B). The immune-cold phenotype was defined by total CD8⁺ T-cell proportions below the median across all TMA cores (1.44% in luminal and 1.36% in TNBC). Immune-segregated cores displayed spatial separation between tumor and T-cell aggregates (cross-K AUC < 0, CD8⁺ T% ≥ 1.44% in luminal, CD8⁺ T% ≥ 1.36% in TNBC). Immune-intermixing cores exhibited co-localization of tumor cells and CD8⁺ T cells (cross-K AUC ≥ 0, CD8⁺ T% ≥ 1.44% in luminal, CD8⁺ T% ≥ 1.36% in TNBC).

Figure 3.

Immune phenotypes defined by T-cell proportions and the co-localization between T cells and epithelial cells. A, Cell type maps of represented cores from the luminal cohort with immune-cold, immune-segregated, and immune-intermixing phenotypes. B, mIF images of three immune phenotypes in luminal and TNBC cohorts. Image on the left shows the entire TMA core; image on the right shows a zoom-in region. The white box indicates the location of the zoom-in region in the TMA core. C, Percentages of immune phenotypes in luminal and TNBC cohorts. D, KM curves for OS stratified by the presence of immune-cold cores in the luminal cohort. E, Multivariable Cox proportional hazard regression analysis of OS for the presence of immune-cold cores in the luminal cohort, considering age, lymph node metastasis, grade, tumor size, and molecular subtypes. *, P < 0.05; ***, P < 0.001. F, KM curves for OS stratified by the presence of immune intermixing cores in the TNBC cohort. G, The ratio of PD1⁺CD8⁺ T cells to all CD8⁺ T cells within immune-cold, -segregated, and -intermixing cores. H, Percentages of iCAFs, myCAFs, and PVLs in total stromal cells within TMA cores of different immune phenotypes.

The most prevalent phenotype was immune-cold in both cohorts, with a slightly higher proportion in TNBC (48.62% in luminal, 50.20% in TNBC). This was followed by immune-segregated (32.83% in luminal, 35.04% in TNBC) and immune-intermixing (18.55% in luminal, 14.75% in TNBC; Fig. 3C). The presence of immune-cold cores was associated with reduced survival in luminal tumors but not in TNBC (luminal: univariate P = 0.0026; multivariate P = 0.018; Fig. 3D and E), whereas in TNBC, the presence of intermixing cores showed a trend to be associated with better survival (P = 0.097; Fig. 3F). It is worth noting that an increased CD8⁺ T-cell percentage was associated with better survival in the luminal cohort (univariate P = 0.027) but not significant when accounting for clinical variables (multivariate P = 0.211). This suggests that the presence of an immune-cold region better predicts the patient outcome than a summarized proportion of PD1⁺CD8⁺ or PD1⁻CD8⁺ T cells. PD1⁺CD8⁺ T-cell ratios display a similar level between segregated and intermixing cores in both subtypes, which were higher than that in the immune-cold cores (Fig. 3G). The composition of CAF subsets varied by immune phenotype. In both cohorts, iCAFs were most abundant in immune-segregated and -intermixing phenotypes and least abundant in immune-cold cores (Fig. 3H). In contrast, PVL cells were most abundant in immune-cold cores (Fig. 3H). myCAFs were more uniformly distributed, with slightly higher levels in intermixing phenotypes in luminal cases (Fig. 3H). These findings suggest that distinct stromal subsets influence T-cell infiltration in different manners. PVLs may play a role in excluding T cells from infiltrating the tumor, whereas iCAFs may relate to separation of cancer and T cells.

Endothelial–distal PVLs are associated with T-cell exclusion

PVL is defined as a perivascular-like stromal cell population marked by CD146 positivity; these include pericytes, vascular smooth muscle cells, and other mural cells. In previous studies, we observed that a subset of PVLs is spatially dissociated from endothelial cells, suggesting that these cells may have functions distinct from their conventional role in vascular contractility and blood flow regulation (9). In the luminal cases, we found that about 97.12% of PVLs were distant (>30 μm) from endothelial cells. Even at a 100 μm threshold, a large majority of PVLs remained distant from endothelial cells (Fig. 4A and B). TNBC exhibited a lower proportion of disseminated PVLs (70.12% at 30 μm, 22.11% at 100 μm; Supplementary Fig. S2A), likely due to the higher endothelial abundance in TNBC compared with luminal subtype (Fig. 1G). There was no prominent disseminated PVL in normal breast samples, suggesting the dissociation of PVLs from endothelial cells is specific to tumors (Fig. 4B).

Figure 4.

PVLs distal from endothelial cells are associated with T-cell exclusion. PVLs were classified as adjacent if an endothelial cell was present within 30 μm and as disseminated if no endothelial cell was present within 100 μm. A, Percentage of PVLs with endothelial cells present within a 10 to 100 μm radius in the luminal cohort. B, mIF images of TMA cores exhibiting disseminated PVLs in luminal and TNBC tumors, and the co-localization of PVLs and blood vessels in a normal breast sample. C, Percentages of CD8⁺ T cells in TMA cores with high and low disseminated PVL percentage stratified by the median. D, Mean percentages of PD1⁺CD8⁺ and PD1⁻CD8⁺ T cells in the neighborhood of adjacent and disseminated PVLs. Neighborhood is defined as the 10 nearest cells to each PVL. Error bar indicates the standard error. E and F, Percentages of disseminated and adjacent PVLs in the stromal region of cores classified as immune-cold, -segregated, or -intermixing phenotypes. G, Gene expression heatmap of canonical marker genes across cell types identified in Xenium data. Cell types were defined using the same combination of markers derived from the mIF panel. Expression values represent mean counts scaled per gene. H, Bar plots showing the proportions of adjacent, disseminated, and other PVLs across Xenium sections from three TNBC tumors. I, Representative regions from Xenium sections illustrating the spatial distribution of PVLs, endothelial cells, epithelial cells, and CD8⁺ T cells. Both PD1⁺CD8⁺ and PD1⁻CD8⁺ T cells were found near adjacent PVLs, whereas they were distant from disseminated PVLs. J, Density plot of PD1⁺CD8⁺ and PD1⁻CD8⁺ T cells across a distance (0–300 μm) to the nearest adjacent or disseminated PVL. K, Volcano plot of DEGs between adjacent and disseminated PVLs. DEGs with an absolute log₂ FC >0.5 and adjusted P value < 0.05 were classified as significantly upregulated or downregulated. The top 10 DEGs by log₂ FC in each direction were labeled. L, Bar plots showing the top 10 enriched GO Biological Process terms for the top 100 genes upregulated in disseminated PVLs (left, blue) and adjacent PVLs (right, yellow). Bar length represents the −log₁₀ adjusted P value, with greater values indicating stronger enrichment.

To investigate whether disseminated PVLs form at the expense of PVL-bounded vessels, we assessed the correlation between disseminated PVL (PVLs without an endothelial cell present within 100 μm) and the proportion of endothelial cells with PVL present within 30 μm. No significant association was observed (luminal, Pearson’s R = −0.037, P = 0.51; TNBC, Pearson’s R = 0.077, P = 0.26; Supplementary Fig. S2B and S2C). This suggests the existence of a spatially distinct PVL population that is not vessel-dependent.

We have observed a reduced CD8⁺ T-cell abundance in PVL-enriched tumors (Fig. 2C). To explore the relationship between PVL spatial distribution and T-cell infiltration, we examined the association between disseminated PVL abundance and CD8⁺ T-cell proportions across cores. In luminal and especially in TNBC cohorts, CD8⁺ T-cell infiltration was reduced in cores with a higher percentage of disseminated PVLs, suggesting that this PVL subpopulation may negatively regulate T-cell entry (luminal: P = 0.018; TNBC: P < 0.001; Fig. 4C). To better understand this, we compared the cellular neighborhoods surrounding PVLs distal versus adjacent to endothelial cells. In both luminal and TNBC, disseminated PVLs had lower proportions of PD1⁺CD8⁺ T cells in their vicinity compared with adjacent PVLs (P < 0.01; Fig. 4D). This supports the observation that disseminated PVLs are associated with T-cell exclusion. As further evidence, the proportion of disseminated PVLs in the stromal region was highest in immune-cold cores across both subtypes (Fig. 4E). In contrast, adjacent PVLs were more abundant in segregated and intermixing subtypes (Fig. 4F). Together, these findings suggest that the influence of PVLs on T-cell exclusion was primarily associated with vessel independent PVLs.

To validate the immune-cold association of disseminated PVLs, we generated an independent spatial transcriptomics dataset for three TNBC tumors using the Xenium Prime platform (5,001 plus 100 genes add-on panel; Supplementary Table S4). Stromal and immune phenotypes were annotated based on the combined expression of genes corresponding to the markers used in the mIF panel (Supplementary Table S5). The classified cell types exhibited expression of canonical lineage markers, confirming the fidelity of our cell type annotation (Fig. 4G; ref. 22). The overlap of VWF, ENG, and CD34 for endothelial cells and imPVL can be attributed to the spillover of transcripts from endothelial cells. Within this dataset, disseminated PVLs accounted for an average of 15% of the total PVLs (range, 8.1%–20.4%), comparable with the prevalence observed in the mIF TNBC cohort (Fig. 4H). Spatial analysis revealed that both PD1⁻CD8⁺ and PD1⁺CD8⁺ T cells were distributed further away from disseminated PVLs than from adjacent PVLs (Fig. 4I and J), consistent with the CD8⁺ T-cell exclusion associated with disseminated PVLs in the mIF TNBC cohort. DEG analysis revealed upregulation of genes related to extracellular matrix (ECM) organization and vascular stability in adjacent PVLs, including COL15A1, COL4A1, POSTN, IGFBP4, and PDGFRB (Fig. 4K and L). In contrast, disseminated PVLs exhibited upregulation of genes related to vascular permeability, hypoxia signaling and apoptosis, such as NDRG1, VEGFA, LDHA, MFGE8, and YWHAZ (Fig. 4K and L; refs. 23–25). These divergent expression profiles suggest that disseminated PVLs transition away from a vessel maintenance role toward a stress-response state, potentially linked to aberrant vasculature and reduced immune cell infiltration.

iCAFs constrain spatial contacts between T cells and cancer cells in luminal breast cancer

In luminal tumors, iCAFs were enriched in immune segregated and intermixing cores, indicating that their presence is not directly associated with an immune-cold phenotype (Figs. 3H and 5A). However, we observed that the minimal distance between CD8⁺ T cells and epithelial cells in luminal cancers was greater in the presence of iCAFs (Fig. 5B), suggesting that iCAFs may spatially restrict T cells within the stromal compartment.

Figure 5.

iCAF are associated with the segregation of CD8⁺ T cells from tumor in the luminal cohort. A, Representative TMA core showing the spatial distribution of CD8⁺ T cells in iCAF-dense (blue) and epithelial-dense (red) regions. B, Minimal distance (min dist) from CD8⁺ T cells to the nearest epithelial cell in TMA cores with low vs. high iCAF proportions. C, KM curves for OS stratified by the percentage of PD1⁻CD8⁺ T cells in iCAF-high vs. iCAF-low groups. D, Multivariable Cox proportional hazard regression analysis of OS for the CD8⁺ T cells% in iCAF-low group, considering age, lymph node metastasis, tumor grade, tumor size, and molecular subtype. *, P < 0.05; ***, P < 0.001.

Remarkably, CD8⁺ T-cell infiltration was associated with improved patient survival for luminal tumors with low iCAF abundance in both univariate and multivariate analyses (Fig. 5C and D), but this prognostic value was lost in iCAF-high tumors. This suggests that high iCAF was associated with T-cell dysfunction. Similarly, CD8⁺ T-cell infiltration was associated with improved survival in myCAF-low tumors but not in myCAF-high tumors (Supplementary Fig. S3A). However, the significance was lost after accounting for clinical variables, especially age at diagnosis (Supplementary Fig. S3B). One possible explanation is that myCAFs may also be associated with T-cell dysfunction, although we were unable to explore this further due to limitations of our dataset.

Clinical correlation of spatially resolved cell compositions

To assess the correlation of spatial context of stromal subsets with clinical characteristics, we evaluated the enrichment of six spatial features per stromal subset in patients stratified by prognostic variables from survival analysis: age at diagnosis, nodal metastasis, tumor grade, histologic subtypes, molecular subtypes, and TILs (Supplementary Tables S6 and S7). The spatial features include the proportion of PVLs, myCAFs, and iCAFs within the stromal compartment, together with their spatial proximity to endothelial, epithelial, and CD8 T cells measured by minimal distance, normalized mixing score, and proportion of adjacent cells (“Materials and Methods”; Supplementary Table S3). Strong associations with clinical variables were defined by significant elevation of at least two spatial features within the same clinical group.

In luminal tumors, increased proximity between PVL and cancer epithelial cells, as reflected by a higher proportion of cancer-adjacent PVLs and shorter PVL–epithelial minimal distance, was linked to lymph node metastasis (Fig. 6A). Enhanced iCAF–cancer proximity showed similar associations with lymph node metastasis (Fig. 6A), whereas myCAF–cancer adjacency did not show significant relationship with assessed clinical variables (Supplementary Fig. S4A). On the other hand, closer proximity between iCAF and CD8⁺ T cells was observed for tumors diagnosed at a younger age (age < 55; Fig. 6A). There was a similar trend for PVL–CD8 T-cell proximity but reached significant for the minimal distance only. Additionally, increased stromal iCAF proportions distinguished luminal A subtype and low-grade tumors (Fig. 6A). For all three stromal cell types, their adjacency to endothelial cells increased in high-grade tumors (Fig. 6A; Supplementary Fig. S3A). This is due to the enrichment of endothelial cells in high-grade luminal cases (Supplementary Fig. S4B).

Figure 6.

A, Association of iCAF- and PVL-related spatial features with clinical variables in the luminal cohort. Patients were stratified by clinical variables. Dot colors indicate patient groups in which the feature is increased. Dot outlines denote statistical significance of feature enrichment (red, significant; gray, not significant), assessed by the Wilcoxon rank-sum test with P values corrected using the BH method. NMS, normalized mixing score. B, Same as A for the TNBC cohort. C, Univariate and multivariate analyses of the association between cell type proportions and OS in the luminal cohort. D, Same as C for the TNBC cohort.

In TNBC, PVL–endothelial proximity correlated with nodal metastasis (Fig. 6B). Across all the three stromal cell types, high proportions in the stroma and increased spatial separation from CD8⁺ T cells and endothelial cells were associated with reduced immune infiltration (Fig. 6B; Supplementary Fig. S4C). Additionally, enhanced iCAF and myCAF adjacency to cancer cells corresponded with reduced TILs (Fig. 6B; Supplementary Fig. S4C). These findings indicate that in TNBC, CD8⁺ T cell and endothelial cell enrichment strongly correlate with clinical TIL assessment, whereas stromal cell abundance generally inversely correlates with immune infiltration.

In terms of patient outcomes, we stratified cohorts into “high” or “low” groups based on median cell-type proportions across patients. In the luminal cohort, reduced PD1⁻CD8⁺ T cells and iCAFs proportions, and high PVLs and endothelial cell proportions were associated with inferior OS (Fig. 6C). In multivariate analysis adjusting for age, node metastasis, tumor grade, tumor size, and molecular subtype, endothelial cell abundance remained significantly associated with worse survival (Fig. 6C). In TNBC, similar trends were demonstrated for PD1⁻CD8⁺ T cells, PVLs, and endothelial cells, but none of the cell types achieved statistical significance in univariate or multivariate analysis (Fig. 6D).

Discussion

In this study, we have characterized the spatially distinct features of stromal subsets in relation to immune cells (Fig. 7). This was done in two large breast cancer cohorts consisting of more than 1,300 TMA cores from 591 patients. Both cohorts have well-annotated clinicopathologic data as well as robust long-term outcome data with follow-up of up to 16 years. This gives our study sufficient power to identify clinically meaningful correlations, whereas addressing limitations of many prior spatial profiling studies that relied on only a small number of samples or patients and were therefore constrained by tissue selection biases and the challenges of both intratumoral and intertumoral heterogeneity.

Figure 7.

Diagrammatic summary of key findings. A, Disseminated PVLs are associated with T-cell exclusion and are enriched in immune cold phenotype. B, myCAFs are located in close proximity to cancer cells, whereas iCAFs and PVLs are disseminated throughout the stroma. C, iCAFs are enriched in the immune-segregated phenotype in which T cells are present but restricted away from cancer cells. [Created in BioRender. Chen, J. (2026) https://BioRender.com/d28tsrm.]

We make the novel observation that the large majority of PVLs are located distant from endothelial cells, despite their perivascular lineage (Fig. 7A). In breast cancers, but not normal breast, the majority of PVLs were disseminated throughout the stroma rather than restricted to vessel-adjacent regions. Morphologically, these cells closely resemble CAFs and share expression of canonical CAF markers such as ACTA2 and PDGFRB and most likely overlap with cells annotated as vascular CAFs (vCAF) in prior studies (3, 22). In Bartoschek and colleagues (3), vCAFs were defined as cells enriched for mural genes such as NOTCH3, EPAS1, COL18A1, and NR2F2 and perivascular markers such as RGS5, MCAM, DES, KCNJ8, and HIGD1B. They do not uniformly exhibit classical fibroblast markers such as DCN, FAP, and PDGFRα. It was also observed that these cells lay along vessels early and become increasingly detached with tumor progression (3), consistent with perivascular origin and delamination. At the protein level, vCAFs were defined as CD146^high CD31^neg with FAP and PDPN used as discriminators for “activated CAFs” and pericytes required extra markers such as RGS5 (22). vCAFs and pericytes were hard to distinguish in some datasets in this study, underscoring phenotypic overlap between these two subsets. Our previously published scRNA-seq data demonstrated distinct transcriptional differences between PVLs and CAFs, supporting their unique identity within the stromal compartment. Gene expression from our Xenium data demonstrated the consistency between the CAF and PVL cell types identified by mIF and their corresponding transcriptomic profiles (Fig. 4G). Together, our studies and the literature support recasting “vCAFs” as perivascular-like stromal cells that frequently lack classic fibroblast–ECM features and may delaminate from vessels into the tumor stroma. In normal breast tissue, this PVL subset resembles vascular smooth muscle cells, transcriptionally and spatially distinct from pericytes (26).

Our study is the first to link increased PVL proportion with CD8⁺ T-cell exclusion and a cold immunophenotype (Fig. 7A), consistent across both luminal and TNBC cohorts. This discovery was uniquely enabled by the application of a marker panel targeting stromal and lymphocyte subsets to large cohorts of disease. In the luminal cohort, higher PVL proportion was prognostic of poorer OS. Although this was not statistically significant in the TNBC cohort, potentially limited by its smaller sample size, a similar trend was observed. PVL-associated T-cell exclusion was primarily associated with the subpopulation of disseminated PVLs, in which increased endothelial–distal PVLs correlated with decreased CD8⁺ T-cell infiltration. Spatial transcriptomic analysis revealed that both PD1⁺CD8⁺ and PD1⁻CD8⁺ T cells were localized further away from disseminated PVLs while closer to adjacent PVLs, confirming the association of disseminated PVLs with immune-cold phenotype and T-cell exclusion (Fig. 4J). Additionally, disseminated PVLs displayed an upregulation of genes associated with vascular permeability and hypoxia response, which is a new insight into these cells (Fig. 4L). Previous studies examining endothelial–pericyte interactions have observed that pericytes detach and migrate away from blood vessels as a response to aberrant PDGFβ signaling (27, 28). These detached pericytes can then differentiate into myofibroblasts, contributing to fibrosis (29, 30), and also lead to vessel instability and vulnerability (31). Our disseminated PVLs may represent these detached pericytes; however, our data suggest these cells do not differentiate into myCAF and maintain CD146 expression. More work is warranted to investigate signaling pathways driving these phenotypes, which may potentially be targeted to reverse the associated effect on T-cell infiltration.

myCAFs and iCAFs have been extensively reported in the literature. Consistent with previous studies (8, 32, 33), we found that myCAFs were located in close proximity to epithelial cells, whereas iCAFs were more broadly distributed throughout the stroma (Fig. 7B). Our study further suggests that iCAFs may spatially restrict T cells within the stromal compartment, as evidenced by an association of iCAFs with a segregated immunophenotype (Fig. 3H) and increased distance between CD8⁺ T cells and epithelial cells in the presence of iCAFs (Fig. 7C). This is further supported by the diminished survival benefit conferred by T-cell infiltration in iCAF-high tumors (Fig. 5C). We propose that this may be mediated by chemokine-directed migration of T cells toward iCAFs, which robustly express the T-cell chemokines CXCL12 and CCL2 (8).

Spatial immunophenotypes have been shown to correlate with survival and response to immune checkpoint inhibitors in breast cancer (34–36). By classifying our cores into immune-cold, -segregated, and -intermixing phenotypes, we demonstrated a significantly poorer survival in patients with the presence of immune-cold cores in luminal tumors, adding to existing evidence that spatial immune contexture is prognostic in breast cancer. Furthermore, we examined stromal phenotypes in these immunophenotypes and found an enrichment of PVLs in immune-cold cores, consistent with our finding that PVLs correlate with T-cell exclusion (Fig. 7A). This was primarily related to endothelial–distal PVLs, whereas endothelial-adjacent PVLs were not enriched in immune-cold cores. This also suggests that this effect was not attributable to blood vessel density. We found that iCAFs were enriched in immune-segregated cores in which T cells were present but restricted to the tumor periphery (Fig. 7C). T-cell exclusion is one of the leading causes of resistance to immune checkpoint inhibition (37), and recent studies suggest a role of CAFs in mediating this (38, 39). Our result provides further evidence to support this and suggests that PVLs and iCAFs contribute to this in different ways. Whereas PVLs exclude T cells from migrating into the tumor, iCAFs were related to the segregation of cancer and T cells. Although our findings provide novel observations on how stromal heterogeneity, particularly PVLs and iCAFs, shape CD8⁺ T-cell infiltration within a spatially relevant context and demonstrate its clinical relevance, further work is required to elucidate the functional and mechanistic basis underlying these observations.

For decades, conventional methods such as immunohistochemistry have been the primary tools for tumor characterization. However, these methods are limited in their capacity to capture the complexity of cellular diversity within tumors. The application of mIF and digital pathology enables the use of expanded protein panels, providing a more comprehensive, spatially resolved view of tumors and their TME. However, these technologies require greater standardization and automation of staining and imaging protocols to enhance reproducibility and clinical applicability (12). One of the limitations of our findings is the potential overcalling of iCAFs due to the lack of a positive marker for iCAFs in our mIF panel. iCAFs were annotated as PDGFRβ⁺ cells negative for other stromal markers (αSMA, CD146, and THY1) and hence may include non-iCAF phenotypes belonging to subsets not annotated by these markers, such as antigen-presenting CAFs and matrix CAFs (2).

This study is based on TMA cores rather than whole-slide sections to improve the feasibility to analyze large clinical cohorts to strengthen potential clinical applicability and statistical power. To account for the well-documented intratumoral heterogeneity of breast cancer, we utilized a sampling strategy of three representative 1-mm cores. This approach is supported by validation studies demonstrating that three-core sampling provides high concordance with whole-tissue sections for standard biomarkers and clinical outcomes, capturing the diversity of the TME while maximizing experimental throughput for spatial transcriptomic analysis (40, 41). The results of our Xenium profiling of surgical resections demonstrated similar findings in the prevalence of disseminated PVLs and their association with CD8⁺ T-cell exclusion, providing further reassurance that our findings are not substantially affected by the use of TMA cores. Although the QuPath algorithm worked well to allow objective quantitative cell annotation suitable for application to large clinical cohorts with efficient high throughput of data, the training of slides was semiquantitative, subject to interobserver variability. We addressed this by manually cross-checking cores across different slides, supervised by a pathologist. Advancements in deep learning Artificial Intelligence algorithms will help to overcome these technical challenges to allow automated deep interrogation of vast quantities of data generated from whole tumors of large cohorts.

Conclusion

Using mIF to map the stromal–immune dynamics in breast cancer in two large clinical cohorts, our study highlights that stromal subsets exhibit distinct spatial distribution, interaction with immune cells and clinical correlations. The clinical correlations identified in our study warrant further investigation as potential biomarkers, ideally through incorporation into prospective clinical trials. Specific stromal–immune interactions may also be explored as novel therapeutic targets to modulate the tumor-immune microenvironment and enhance responses to immunotherapy.

Supplementary Material

Supplementary Figure 1

Association of CD8 T cell percentages with clinical characteristics in the luminal cohort.

Supplementary Figure 2

Correlation between PVLs and endothelial cells.

Supplementary Figure 3

Survival analyses of PD1⁻CD8⁺ T cells in myCAF-high vs. myCAF-low groups

Supplementary Figure 4

Association of myCAF-related spatial features with clinical variables.

Supplementary Table 1

Antibodies used in mIF panel.

Supplementary Table 2

Marker combinations for defining cell types in the mIF panel.

Supplementary Table 3

Description of features selected for assessing the association with clinicopathological variables.

Supplementary Table 4

Xenium gene list for custom-designed add-ons

Supplementary Table 5

Marker gene combinations for defining cell types in Xenium.

Supplementary Table 6

Univariate and multivariate Cox regression of clinicopathological predictors of overall survival in in the luminal cohort.

Supplementary Table 7

Univariate and multivariate Cox regression of clinicopathological predictors of overall survival in in the TNBC cohort.

Data Availability

Raw mIF images and Xenium Prime data are available at https://doi.org/10.6019/S-BSST3074. Processed data can be obtained from https://doi.org/10.5281/zenodo.20199307. Other data generated in this study are available from the corresponding author upon reasonable request.

Code Availability

Code related to the analyses in this study can be found on GitHub at https://github.com/Swarbricklab-code/Stromal_Tcell_Infiltration_mIF.

Authors’ Disclosures

J. Chen reports grants from National Breast Cancer Foundation Australia during the conduct of the study. H. Zhang reports grants from National Breast Cancer Foundation Australia during the conduct of the study. P.H. Graham reports grants from New South Wales Cancer Institute Ontario Clinical Oncology Group outside the submitted work. A. Swarbrick reports grants from Cancer Council New South Wales Cancer Institute, National Breast Cancer Foundation, and National Health and Medical Research Council during the conduct of the study, grants from 10x Genomics outside the submitted work, and research support from John and Deborah McMurtrie A. Swarbrick holds a conjoint professorial appointment at UNSW Sydney. No disclosures were reported by the other authors.

Authors’ Contributions

J. Chen: Conceptualization, formal analysis, methodology, writing–original draft, writing–review and editing. H. Zhang: Formal analysis, writing–original draft, writing–review and editing. T. Ruan: Supervision, writing–review and editing. S.Z. Wu: Conceptualization, methodology. I. Slapetova: Data curation, formal analysis, methodology. E. Millar: Supervision, methodology, writing–review and editing, sample acquisition. P.H. Graham: Writing–review and editing, sample acquisition, clinical data collection. J. Lynch: Writing–review and editing, sample acquisition, clinical data collection. L.H. Browne: Writing–review and editing, sample acquisition, clinical data collection. E. Lim: Resources, supervision, funding acquisition, writing–review and editing. A. Swarbrick: Conceptualization, resources, supervision, funding acquisition, writing–review and editing.