Identifying breast cancer subtypes and exploring spatial expression patterns of the key subtype-specific gene based on pathological images, single-cell sequencing, and spatial transcriptomic data

Abstract

Background

Breast cancer (BRCA) is a heterogeneous disease. Accurate prognosis and molecular subtypes are critical for personalized treatment in BRCA. This study aimed to identify prognostic subtypes of BRCA based on features derived from pathological images and to explore their potential molecular and cellular mechanisms using spatial transcriptomics (ST) and single-cell RNA sequencing (scRNA-seq).

Methods

Hematoxylin and eosin (H&E)-stained images were obtained from the TCGA-BRCA dataset using the OTSU algorithm. Pathological features were extracted using PyRadiomics. Feature selection was performed via Cox regression and machine learning. Subtypes were defined through consensus clustering, followed by differential gene expression. Key genes were identified using WGCNA and Lasso regression analysis. The spatial distribution and cellular interactions of key genes were analyzed by integrating ST and scRNA-seq datasets.

Results

Among 1,406 extracted pathological features, 25 core features were identified, based on which two BRCA subtypes were defined. The C1 subtype was associated with advanced disease and poor prognosis. ESRP1 was identified as a key gene upregulated in the C1 subtype. ST analysis revealed that ESRP1 was highly expressed in tumor lesion areas and co-localized with epithelial and fibroblast cells. Cell-cell communication analysis revealed that fibroblasts served as central hubs in the TGF-β signaling network, particularly in regions enriched with ESRP1.

Conclusions

This study identified pathology-derived prognostic BRCA subtypes that complement existing molecular classifications. The core gene, ESRP1, exhibited distinct spatial expression patterns in BRCA tissues, primarily localized to tumor core regions enriched with fibroblast cells. These findings provide exploratory insights into BRCA heterogeneity and potential therapeutic targets.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s12282-026-01883-y.

Introduction

Breast cancer (BRCA) is one of the most common malignant tumors among women worldwide, accounting for approximately 31% of all female cancers [1]. Despite significant advancements in early screening and treatment over the past decades, BRCA remains a leading cause of cancer-related deaths in women [2, 3]. BRCA is characterized by a high degree of heterogeneity, which is a critical factor influencing clinical behavior and outcomes [4]. This heterogeneity is reflected in histological types, molecular characteristics, treatment responses, and prognoses, posing significant challenges to clinical diagnosis and personalized treatment strategies [5]. Traditionally, BRCA subtyping has relied on markers such as estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), which classify tumors into luminal A-like, luminal B-like, HER2⁺, and triple-negative breast cancer (TNBC) subtypes [6, 7]. While these classifications are clinically practical, the marked differences in response and outcomes among patients suggest that the traditional subtyping system may not be universally applicable and often fails to capture the full complexity of tumor heterogeneity [8, 9]. Therefore, accurately identifying additional BRCA subtypes with biological and clinical significance is essential for optimizing prognostic evaluation and therapeutic strategies.

In recent years, advancements in computational pathology, high-throughput omics technologies, and machine learning algorithms have provided new avenues for tumor subtyping and biomarker discovery [10–12]. Histological features of tumors not only convey phenotypic information but are also closely associated with tumor progression and patient prognosis. Hematoxylin and eosin (H&E) staining is a routine technique in biopsies and is commonly used for cancer detection and subtype diagnosis. Machine learning has overcome the limitations of visual inspection of H&E images, enabling the extraction of hidden information imperceptible to the human eye [13]. In particular, the PyRadiomics platform enables high-throughput extraction of quantitative features from medical images, including shape, texture, and intensity-based descriptors, which can objectively characterize tumor morphological heterogeneity [14]. Studies have demonstrated that molecular biomarker expression in BRCA can be predicted using machine learning models based on H&E images [15]. Furthermore, the emergence of spatial transcriptomics (ST) and single-cell RNA sequencing (scRNA-seq) technologies has provided crucial insights into cellular heterogeneity within tissues, tumors, and immune cells. These technologies allow for the spatial mapping of individual cells within tissue sections, facilitating functional interpretation in complex multicellular systems [16, 17]. They also help reveal cellular heterogeneity in BRCA [18], offering new perspectives on the mechanisms underlying tumor initiation and progression.

This study aimed to identify prognostic subtypes of BRCA based on features derived from H&E pathological images and to explore the expression patterns and potential cell–cell communication networks of key subtype-specific genes by integrating ST and scRNA-seq data, thereby providing new insights into BRCA heterogeneity and potential therapeutic targets.

Materials and methods

Acquisition and feature extraction of pathological images

This study included 1,030 BRCA samples from the TCGA-BRCA database (https://portal.gdc.cancer.gov/), with both clinical data and H&E-stained whole slide images available. The whole-slide images (WSIs), high-resolution digital scans of entire histopathology slides, were preprocessed using the OTSU algorithm to segment the tissue regions from the background. OTSU is a widely used image processing method that automatically distinguishes foreground tissue from background based on differences in pixel intensity distribution. To standardize image analysis, WSIs at 40× magnification were divided into sub-images of 1024 × 1024 pixels, while those at 20× magnification were first divided into 512 × 512 pixel sub-images and then upsampled to 1024 × 1024 pixels [19]. From the segmented tissue regions of each WSI, representative image patches were randomly sampled to ensure adequate coverage of tumor morphology as well as the surrounding stromal microenvironment. Sampled patches were visually inspected to exclude low-quality regions when necessary. Furthermore, a subset of sampled patches was reviewed by an experienced pathologist to ensure that the selected regions adequately represented both tumor and stromal components. A total of 1,488 quantitative pathological features were extracted using PyRadiomics (version 3.0.1), an open-source Python package that enables standardized extraction of quantitative radiomic features. The extracted features include first-order statistics describing pixel intensity distributions, shape-based descriptors characterizing geometric properties, and multiple texture features derived from gray-level matrices such as gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighboring gray tone difference matrix (NGTDM) features. The extracted features were subsequently normalized. After quality control, features with constant or zero values were excluded, resulting in 1,406 features retained for subsequent analysis.

Selection of prognosis-related pathological features using machine learning

Univariate Cox regression analysis was first performed on the extracted pathological features to identify those significantly associated with overall survival (OS). Features with a p-value < 0.05 were retained for further analysis. Three machine learning algorithms—AdaBoost, XGBoost, and Random Forest (RF)—were then applied using the adabag, xgboost, and randomForest R packages, respectively, to evaluate feature importance. The top 50 features from each model were selected based on weight importance, and the intersection of these features across all three models was identified using a Venn diagram. The overlapping features were considered key candidates for further analysis.

Consensus clustering analysis

Based on the intersecting pathological features identified above, unsupervised consensus clustering was conducted using the ConsensusClusterPlus R package. The optimal number of clusters (k) was determined by evaluating the cumulative distribution function (CDF) curves, with k = 2 showing the highest clustering stability. Accordingly, two BRCA subtypes were defined: C1 and C2. The Wilcoxon rank-sum test was used to compare the expression levels of the intersecting pathological features between the two subtypes.

Survival and clinical correlation analysis

To assess the robustness of the subtype classification, the TCGA-BRCA cohort was randomly divided into a training set (70%) and a validation set (30%). Kaplan–Meier (KM) survival analysis and log-rank tests were performed independently in total, training, or validation sets to confirm the prognostic significance of the C1/C2 subtypes. KM survival analysis was performed using the survival and survminer R packages. Chi-square tests were used to compare the distribution of clinical variables (such as age, tumor stage, hormone receptor status, and histological subtype) between the C1 and C2 subtypes. To evaluate the relationship between the identified subtypes and established molecular classifications, samples with complete ER, PR, and HER2 annotation were classified into four immunohistochemistry-based surrogate subtypes: luminal A-like, luminal B-like, HER2⁺, and TNBC. Classification was based on ER, PR, and HER2 status following commonly accepted guidelines: luminal A-like tumors were ER and/or PR-positive, HER2-negative; luminal B-like tumors were ER and/or PR-positive, HER2-positive; HER2-positive tumors were HER2-positive, ER and PR-negative; TNBC tumors were negative for ER, PR, and HER2. Ki-67 data were not available in the TCGA clinical annotations; therefore, Ki-67 was not incorporated in the subtype definitions. Missing data for ER, PR, and HER2 status were inherent to the TCGA-BRCA clinical annotation (28–30% missing rate), and analyses involving molecular subtypes were restricted to samples with complete receptor status information. The distribution of C1 and C2 across these molecular subtypes was assessed using Chi-square tests. Stratified KM survival analyses were performed within each molecular subtype to evaluate whether C1/C2 subtypes retain prognostic value independent of conventional classifications. Hazard ratios (HRs) and 95% confidence intervals (CIs) were visualized using forest plots. In addition, both univariate and multivariate Cox regression analyses were conducted to evaluate whether the identified subtypes serve as independent prognostic factors in BRCA.

Differential gene expression and pathway enrichment analysis

RNA-seq transcriptomic data from the TCGA-BRCA cohort were analyzed using the limma R package to identify differentially expressed genes (DEGs) between the C1 and C2 subtypes. The criteria were set as fold change = 1.5 and p < 0.05. These DEGs were further subjected to Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis using the clusterProfiler R package, aiming to elucidate their potential associated signaling pathways.

Immune cell infiltration analysis

To evaluate the tumor immune microenvironment (TIME), the CIBERSORT algorithm was applied to estimate the relative abundances of 22 immune cell types in TCGA-BRCA samples. The differences in immune cell infiltration between the C1 and C2 subtypes were assessed using the Wilcoxon rank-sum test. Immune cells with significant differences were visualized using box plots.

Weighted gene co-expression network analysis (WGCNA)

WGCNA was performed to identify gene modules significantly associated with the newly defined BRCA subtypes. An adjacency matrix was constructed using a soft-thresholding power determined according to the scale-free topology criterion. Gene modules were identified via dynamic tree cutting, and their correlations with subtype phenotypes were evaluated. The module most strongly associated with the identified subtypes was selected for further analysis. A Venn diagram was used to intersect the genes in this key module with the DEGs identified earlier, yielding a set of potential subtype-specific candidate genes.

Selection of key genes

Least Absolute Shrinkage and Selection Operator (Lasso) Cox regression analysis was performed using the glmnet R package to further narrow down the candidate subtype-specific genes. A 10-fold cross-validation was applied to determine the optimal value of the regularization parameter lambda. The expression levels of the selected genes were then evaluated between the identified subtypes as well as between tumor and normal samples. Gene expression levels in TCGA-BRCA were log2-normalized. Furthermore, the expression of the key gene was compared across established molecular subtypes (luminal A-like, luminal B-like, HER2⁺, TNBC) and between C1 and C2 within each molecular subtype to evaluate whether its prognostic relevance is independent of conventional molecular classifications.

ST and scRNA-seq analysis

ST data (GSE190811) and scRNA-seq data (GSE161529) were obtained from the Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/). The Seurat R package was used for data processing. For ST data, raw count matrices were normalized, and spatial parameters were corrected using the SCTransform function. The spatial expression distribution of key genes was visualized using the SpatialFeaturePlot function [20]. To define regions of interest for spatial analysis, tissue regions were manually annotated based on the corresponding H&E-stained images. Specifically, regions were delineated based on histopathological features, including cellular density, tissue architecture, and tumor–stroma boundaries, thereby enabling distinction between tumor-enriched and stromal or adjacent non-tumor areas. The annotation was independently performed by two investigators, and discrepancies were resolved through consensus. These annotated regions were subsequently used to guide downstream spatial analysis and interpretation.

For the scRNA-seq data, low-quality cells were removed based on the following criteria: fewer than 1,000 or more than 20,000 detected genes, or mitochondrial RNA content exceeding 20%. Cell type annotation was performed using the RCTD algorithm, and the annotated cell types were then mapped onto ST slices to visualize their spatial distributions.

Cell–cell communication analysis

Based on the cell types annotated from the scRNA-seq data, CellChat was employed to analyze intercellular communication networks. Interaction networks among cell types were constructed, and both the number and weight of cell–cell communications were quantified. Special attention was given to the communication network of the TGF-β signaling pathway. Network graphs were generated to visualize the overall communication landscape and mapped onto ST slices.

Statistical analysis

All statistical analyses in this study were performed using R software (version 4.2.2). All tests were two-sided, and p < 0.05 was considered statistically significant.

Results

Identification of potential BRCA subtypes based on key pathological features

To explore potential subtypes of BRCA, a comprehensive evaluation was conducted on 1,406 pathological features. Through univariate Cox regression analysis, a total of 104 features were identified to be significantly associated with BRCA prognosis. The names and expression patterns of these features are provided in Table S1. To further assess the prognostic importance of these features, three machine learning algorithms—AdaBoost, XGBoost, and RF—were employed to rank the 104 features based on feature weight importance. The top 50 features identified by each algorithm are presented in Table S2. To identify robust and BRCA-relevant features, the intersection of the top 50 features from all three algorithms was taken, yielding 25 overlapping features (Fig. 1A), with their names listed in Table S3. Subsequently, consensus clustering analysis was performed on the TCGA-BRCA samples using these 25 key pathological features. When the number of clusters (k) was set to 2, the clustering showed optimal stability and separation (Fig. 1B–C). Based on this, two BRCA subtypes were identified and designated as C1 and C2. Further analysis compared the expression levels of these 25 features between subtypes C1 and C2. As shown in Fig. 1D–E, compared with the C2 subtype, 9 features were significantly upregulated in the C1 subtype (p < 0.0001, Fig. 1D), while 15 features were significantly downregulated (p < 0.0001, Fig. 1E). These subtype-specific expression patterns of pathological features highlight the underlying biological heterogeneity within BRCA.

Fig. 1.

Identification of pathology-derived BRCA subtypes based on key pathological features. A Overlapping pathological features among the top 50 selected by AdaBoost, XGBoost, and Random Forest algorithms. B The cumulative distribution function curve in consensus clustering analysis. C Consensus clustering divides TCGA-BRCA samples into two subtypes: C1 and C2. D Overlapping pathological features highly expressed in the C1 subtype. E Overlapping pathological features highly expressed in the C2 subtype. ****p < 0.0001; ns: not significant. BRCA breast cancer

C1 subtype is associated with poor prognosis in BRCA patients

To further investigate the association between the newly identified BRCA subtypes and clinical outcomes, we first analyzed the distribution of clinical features between the C1 and C2 subtypes. The results revealed that a significantly higher proportion of patients in the C1 subtype were diagnosed at advanced stages (stage III & IV) compared to the C2 subtype (27.616% vs. 20.476%, p < 0.05, Table 1), suggesting a potential link between the C1 subtype and disease progression. Survival analysis further supported this observation. The KM curve demonstrated that patients classified as C2 had significantly longer OS than those in the C1 group (p < 0.01, Fig. 2A). To evaluate the robustness of these findings, the TCGA-BRCA cohort was randomly divided into a training set and a validation set. KM analysis confirmed that the prognostic difference between C1 and C2 was significant in both the training cohort (p = 0.0026) and the independent validation cohort (log-rank p = 0.035; Fig. 2B), demonstrating the reproducibility of the subtype classification. Moreover, subgroup analysis revealed that the C2 subtype was consistently associated with better prognosis across various clinical settings, including patients who were ER⁺, PR⁻, HER2⁻, diagnosed at stage I & II, with invasive ductal carcinoma (IDC), or were < 60 years (p < 0.05, Figure S1A–O). A forest plot was used to further visualize the prognostic differences between C1 and C2 across different subgroups, showing that patients in the C2 subtype had a generally better prognosis than those in C1 across multiple clinical categories (Fig. 2C). These findings suggest that the C1 subtype is not only associated with more advanced disease stages but also with significantly poorer survival outcomes.

Table 1.

Comparison of clinical characteristics between C1 and C2

Variables		Total (n = 1003)	C1 (n = 583)	C2 (n = 420)	p
Age (years), n(%)	≥ 60	564 (56.231)	327 (56.089)	237 (56.429)	0.915
	< 60	439 (43.769)	256 (43.911)	183 (43.571)
ER status, n(%)	Negative	165 (16.451)	88 (15.094)	77 (18.333)	0.270
	Positive	550 (54.835)	319 (54.717)	231 (55.000)
	Missing	288 (28.714)	176 (30.189)	112 (26.667)
PR status, n(%)	Negative	230 (22.931)	121 (20.755)	109 (25.952)	0.116
	Positive	482 (48.056)	283 (48.542)	199 (47.381)
	Missing	291 (29.013)	179 (30.703)	112 (26.667)
HER2 status, n(%)	Negative	599 (59.721)	342 (58.662)	257 (61.190)	0.261
	Positive	102 (10.169)	67 (11.492)	35 (8.333)
	Missing	302 (30.110)	174 (29.846)	128 (30.476)
Stage, n(%)	I & II	756 (75.374)	422 (72.384)	334 (79.524)	0.010
	III & IV	247 (24.626)	161 (27.616)	86 (20.476)
Histological, n(%)	IDC	279 (27.817)	157 (26.930)	122 (29.048)	0.460
	Others	724 (72.183)	426 (73.070)	298 (70.952)

Abbreviations: ER, estrogen receptor; PR, progesterone receptor; HER2, human epidermal growth factor receptor 2; IDC, invasive ductal carcinoma

Fig. 2.

C1 subtype is associated with poor prognosis in BRCA patients. A Kaplan-Meier curve for overall survival time between C1 and C2 in the total TCGA-BRCA cohort. B Kaplan-Meier curve for overall survival time between C1 and C2 in the training and validation cohort; the TCGA-BRCA cohort was randomly divided into a training set (70%) and a validation set (30%). C The forest plot shows the prognostic differences of subtypes across different subgroups. BRCA breast cancer, ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, HR hazard ratio, CI confidence interval

To address the potential confounding effect of conventional molecular subtypes, we systematically evaluated the distribution of C1 and C2 subtypes across established molecular classifications. Among 665 samples with complete ER, PR, and HER2 annotation, we classified patients into luminal A-like (n = 453), luminal B-like (n = 71), HER2⁺ (n = 27), and TNBC (n = 114). Chi-square analysis revealed that the C1 and C2 subtypes were distributed across all molecular subtypes without statistically significant enrichment in any single group (p = 0.089; Table S5). Stratified KM survival analyses within each molecular subtype showed that C1 consistently exhibited a trend toward poorer prognosis compared to C2 across all molecular subtypes (Figure S2A-D), with a statistically significant difference observed in the luminal A-like subgroup (p = 0.0051). These results indicate that the C1/C2 classification captures prognostic information that is largely independent of, and complementary to, the conventional ER/PR/HER2-based molecular subtyping system.

Identified subtype classification is an independent prognostic factor for BRCA patients

To determine whether the identified subtype serves as an independent prognostic factor in BRCA, univariate and multivariate Cox regression analyses were performed. The univariate Cox analysis indicated that several variables were significantly associated with OS of BRCA patients, including tumor stage (OR = 2.608, 95% CI: 1.847–3.682, p < 0.001), age (OR = 2.190, 95% CI: 1.550–3.095, p < 0.001), and the identified subtype (OR = 0.585, 95% CI: 0.408–0.838, p = 0.003). In the multivariate Cox regression model, the subtype remained a significant independent prognostic factor (OR = 0.568, 95% CI: 0.396–0.816, p = 0.002). Additionally, age (OR = 2.379, 95% CI: 1.680–3.368, p < 0.001) and tumor stage (OR = 2.761, 95% CI: 1.952–3.904, p < 0.001) also emerged as significant predictors of prognosis (Table S4). These findings suggest that the identified BRCA subtype classification serves as an independent predictor of patient prognosis and holds substantial potential for clinical application in risk stratification and personalized treatment planning.

Functional characterization of DEGs associated with BRCA subtypes and their correlation with immune cell infiltration

To further elucidate the potential molecular mechanisms underlying the identified BRCA subtypes, DEGs between the C1 and C2 subtypes were analyzed. A total of 746 significant DEGs were identified, among which 429 genes were downregulated and 317 genes were upregulated in the C1 subtype compared to C2 (Fig. 3A). Given that the C1 subtype is associated with poorer prognosis, we focused on the upregulated genes in C1 to explore potential pathogenic mechanisms. The KEGG results revealed 10 significantly enriched pathways, primarily involved in metabolic processes, including cytokine–cytokine receptor interaction, glutathione (GSH) metabolism, and tryptophan metabolism (Fig. 3B). Considering the close interplay between metabolic reprogramming and the TIME, we further evaluated differences in immune cell infiltration between the two subtypes. Compared with the C2 subtype, the C1 subtype had significantly lower infiltration levels of B cells, T follicular helper cells, and resting dendritic cells, while infiltration levels of M2 macrophages and neutrophils were markedly higher (p < 0.05; Fig. 3C). These findings suggest that immunosuppressive cells are relatively enriched in the C1 subtype and may play a critical role in driving its poor clinical outcomes.

Fig. 3.

Functions of DEGs between C1 and C2 and immune cell infiltration. A Volcano plot reveals the DEGs between C1 and C2. B KEGG pathways of DEGs upregulated in C1. C Immune cell infiltration shows significant differences between C1 and C2. *p < 0.05, **p < 0.01. DEGs differentially expressed genes

ESRP1 is a key DEG in the potential BRCA subtype

To further identify gene modules significantly associated with the potential BRCA subtypes, WGCNA was performed. Based on the scale independence criterion, a total of 14 co-expression modules were constructed (Fig. 4A-B). The module–trait correlation heatmap revealed that the purple module showed the strongest positive correlation with C1 and was therefore selected as the key module, which contained 145 genes (Fig. 4C). By intersecting these 145 genes with the previously identified 746 DEGs, 18 key DEGs closely associated with the subtypes were obtained (Fig. 4D).

Fig. 4.

Identification of subtype-related DEGs. A, B WGCNA reveals 14 co-expression modules based on DEGs between C1 and C2. C Correlation of 14 co-expression modules with the identified subtype. D Intersections between DEGs in the identified subtypes and genes in the purple module. DEGs differentially expressed genes, WGCNA Weighted gene co-expression network analysis

Next, Lasso regression analysis was performed using OS as the outcome to further narrow down these 18 genes, resulting in 7 candidate genes: OAF, DCAF13, GSN, ESRP1, MYLK, CLEC14A, and DUSP1 (Fig. 5A). Differential expression analysis showed that DCAF13 and ESRP1 were significantly upregulated in tumor tissues compared to normal tissues, while the other five genes were downregulated (p < 0.0001, Fig. 5B). In addition, both DCAF13 and ESRP1 exhibited significantly higher expression levels in the C1 subtype compared to the C2 subtype (p < 0.05, Fig. 5C).

Fig. 5.

ESRP1 is a key DEG in the BRCA subtype. A Lasso regression analysis was performed on the identified subtype-related DEGs and identified 7 key genes. B Key gene expression between normal and tumor samples in TCGA-BRCA. (C) Key gene expression between C1 and C2. *p < 0.05, **p < 0.01, ****p < 0.0001, ns not significant, DEG differentially expressed gene, BRCA breast cancer

Taken together with previous findings, the C1 subtype was associated with multiple pathological features related to metabolically active regions and stromal infiltration (the 9 pathological features highly expressed in C1). ESRP1, as an epithelial-specific splicing regulator, plays a critical role in controlling epithelial–mesenchymal transition (EMT) [21]. Therefore, we selected ESRP1 as a representative gene for subsequent ST analysis to further explore its role in tumor tissue heterogeneity.

Additionally, to investigate whether the association between ESRP1 and the C1 subtype is confounded by molecular subtypes, we evaluated ESRP1 expression across established molecular classifications and within each molecular subtype. ESRP1 expression varied across four molecular subtypes, with the highest levels in HER2⁺ samples, followed by TNBC (p < 0.001, Figure S3A). Moreover, ESRP1 expression showed a higher level in C1 than C2 across four molecular subtypes, with significance in the luminal A-like subgroup (p < 0.05; Figure S3B), suggesting that the elevated ESRP1 expression in C1 is not merely a reflection of molecular subtype bias.

ESRP1 is highly expressed in the BRCA core lesion region and closely associated with the spatial distribution of epithelial and fibroblast cells

To further investigate the spatial expression pattern of ESRP1 in BRCA tissues, we analyzed four tumor samples from the ST dataset GSE190811. Taking sample 1 as an example, ESRP1 expression was significantly elevated in the tumor lesion area compared to the surrounding areas (Fig. 6A). To determine the cellular composition of ESRP1-enriched regions, we integrated scRNA-seq data from GSE161529, annotated the cell types present in the ST samples, and projected these annotations onto the ST map. A total of eight major cell types were identified: epithelial cells, CD8⁺ T cells, myoepithelial cells, fibroblasts, endothelial cells, neutrophils, B cells, and M2 macrophages (Fig. 6B). Spatial mapping of these cell types revealed that ESRP1-enriched areas were predominantly composed of epithelial cells and fibroblasts, both of which were densely distributed in the tumor lesion region (Fig. 6C). In contrast, CD8⁺ T cells were mainly localized at the tumor periphery (Fig. 6C). Additionally, B cells exhibited a relatively high expression along the tumor edge, whereas endothelial cells and neutrophils were more diffusely distributed throughout the tumor regions. This spatial expression pattern was also consistently observed in the other three samples: ESRP1 was uniformly and highly expressed in the tumor core area, and its spatial distribution closely overlapped with the enrichment of epithelial and fibroblast cells (Figure S4–S6 A–C).

Fig. 6.

ESRP1 is highly expressed in the BRCA lesion region and closely associated with the spatial distribution of epithelial and fibroblast cells. A Tumor lesion location and spatial expression of ESRP1 in ST sample 1. B Cell type annotations based on the single-cell RNA-seq dataset GSE161529 and their corresponding spatial distribution within the ST sample 1. C Spatial localization of eight annotated cell types (epithelial cells, CD8⁺ T cells, endothelial cells, M2 macrophages, myoepithelial cells, fibroblasts, neutrophils, and B cells) in the tumor lesion region. BRCA breast cancer, ST spatial transcriptomics

Fibroblasts play a central role in TGF-β-mediated intercellular communication and are closely associated with ESRP1-enriched regions in BRCA

To further elucidate the intercellular communication landscape within BRCA, we conducted cell communication analysis based on ST samples. As shown in Figure S7A, fibroblasts exhibited the highest degree of interaction and communication intensity with other cell types, demonstrating strong centrality within the network both in terms of the interaction number (left) and interaction weight (right). While epithelial and endothelial cells also displayed certain levels of interaction, their overall influence was notably weaker than that of fibroblasts. Given that the TGF-β signaling pathway is frequently activated in tumors and has been reported to regulate EMT, we further explored the TGF-β-mediated intercellular communication network. As shown in Figure S7B, fibroblasts served not only as prominent targets but also as key sources of TGF-β signaling, engaging in extensive bidirectional signaling with multiple cell types, including epithelial cells, neutrophils, myoepithelial cells, and B cells, thereby forming a fibroblast-centered TGF-β communication network. The spatial communication map further supports these findings (Figure S7C), where fibroblasts in the tumor core region were in close spatial contact with various cell types, most notably showing strong interactions with B cells and epithelial cells. Strikingly, these regions with dense fibroblast interaction coincided with areas of high ESRP1 expression, suggesting that ESRP1 may contribute to BRCA tumor progression and microenvironment remodeling by modulating fibroblast-driven TGF-β signaling. Consistent spatial interaction patterns were also observed in the other three ST samples, where fibroblasts consistently emerged as central hubs within the cell-cell communication network, reinforcing their pivotal regulatory role in the BRCA tumor microenvironment (TME) (Figure S4–S6 D–F).

Discussion

In this study, two prognostically distinct subtypes of BRCA were identified and validated based on quantitative features extracted from H&E-stained pathological images. The results revealed significant differences between the two subtypes in terms of clinical outcomes, molecular characteristics, and immune microenvironment profiles. Specifically, the C1 subtype exhibited worse OS, more advanced tumor staging, and markedly elevated expression of the ESRP1. Further integration of ST and scRNA-seq analyses demonstrated that ESRP1 was highly expressed in tumor lesion regions and was co-localized with epithelial and fibroblast cells. Notably, fibroblasts played a central role in the TGF-β signaling communication network within the TME.

Unlike traditional BRCA subtyping approaches based on transcriptomic profiles or genomic mutations, this study employed a machine learning-driven strategy using pathological image features, offering improved interpretability and clinical relevance. In recent years, the convergence of digital pathology and artificial intelligence has gradually transformed the landscape of tumor subtyping [22–25]. Based on 1,406 pathological features and three machine learning algorithms, we successfully classified BRCA into two new subtypes: C1 and C2. Clinically, the C1 subtype consistently exhibited significantly poorer survival across multiple stratified subgroups. Even within traditionally low-risk groups such as ER⁺ and early-stage (I&II) patients, C1 subtype individuals had markedly worse prognoses than those in the C2 group, suggesting that C1 may have more aggressive biological characteristics. Multivariate Cox regression analysis further confirmed that the identified subtype was an independent prognostic factor, highlighting its potential utility for guiding personalized treatment strategies. Importantly, our analysis demonstrated that the C1/C2 classification is largely independent of the conventional ER/PR/HER2-based molecular subtyping system. Cross-tabulation analysis showed that C1 and C2 subtypes were distributed across all four molecular subtypes (luminal A-like, luminal B-like, HER2⁺, TNBC) without significant enrichment in any single subgroup. Stratified survival analyses within each molecular subtype further supported this notion, as C1 consistently showed a trend toward worse prognosis compared to C2 across all molecular subgroups. These findings suggest that the C1/C2 classification may serve as a complementary tool to existing molecular subtyping, rather than a replacement, potentially refining risk stratification within molecularly defined subgroups. However, given that external validation in independent cohorts has not been performed, the clinical applicability of this classification remains exploratory and hypothesis-generating at this stage.

Moreover, the C1 subtype was enriched in several metabolic pathways, particularly GSH metabolism and tryptophan metabolism, both of which are closely linked to redox homeostasis and immune evasion. GSH is a key intracellular antioxidant that neutralizes reactive oxygen species and facilitates detoxification, thereby maintaining cellular redox balance [26]. Elevated GSH levels in BRCA have been associated with chemotherapy resistance. Tryptophan metabolism is a crucial modulator of tumor immune escape [27]; the accumulation of its metabolites promotes regulatory T cell differentiation while suppressing effector T cell activity, resulting in an immunosuppressive TME that facilitates cancer cell survival and immune evasion [28]. Our immune infiltration analysis further supports this observation, as C1 tumors showed significantly higher infiltration of M2 macrophages and neutrophils, indicative of an immunosuppressive microenvironment. The distinct metabolic and immune features of the C1 subtype may collectively drive its unfavorable clinical outcomes, suggesting that patients with this subtype may benefit from therapies targeting metabolic pathways or immune checkpoints.

By integrating WGCNA and Lasso Cox regression analyses, we identified ESRP1 as a key DEG most closely associated with the C1 subtype. ESRP1 is an epithelial-specific splicing regulatory factor and is considered a critical molecule in promoting the acquisition of invasive and migratory capabilities in tumor cells [29–31]. Previous studies have shown that ESRP1 contributes to tumor progression by regulating the alternative splicing of multiple genes involved in cell–cell adhesion, cytoskeletal remodeling, and cell migration [31, 32]. Liu et al. reported that ESRP1 is highly expressed in BRCA tissues and may serve as a potential predictor of poor prognosis [29]. Consistently, our study also demonstrated that ESRP1 was significantly upregulated in tumor samples from the TCGA-BRCA cohort. ST analysis further revealed that ESRP1 was highly enriched in the tumor lesion region and exhibited strong co-localization with both epithelial cells and fibroblasts, suggesting its potential role in epithelial–stromal interactions within BRCA. Moreover, ESRP1 showed subtype-specific high expression in the C1 group, reinforcing its close association with this newly identified aggressive subtype. Furthermore, our stratified analysis demonstrated that ESRP1 expression varied significantly across molecular subtypes, with the highest levels observed in HER2⁺ and TNBC groups. Importantly, within the luminal A-like subgroup, the largest molecular subtype, ESRP1 expression was still significantly higher in C1 than in C2, suggesting that the association between ESRP1 and the C1 subtype is not merely a surrogate of receptor status.

Notably, we identified nine key pathological features that were significantly upregulated in the C1 subtype. Among them, log-sigma-2-mm-3D_glszm_SmallAreaLowGrayLevelEmphasis, log-sigma-3-mm-3D_glrlm_ShortRunLowGrayLevelEmphasis, log-sigma-5-mm-3D_glszm Zone%, and square_glszm SmallAreaLowGrayLevelEmphasis originated from metabolically active regions, corresponding to the glycolytic core, the hypoxia-metabolic stress interface, metabolic compartmentalization, and the diffusion edge of metabolic products, respectively. In addition, square_glcm ClusterShade, log-sigma-5-mm-3D_ngtdm Busyness, log-sigma-5-mm-3D_glrm RunLengthNonUniformityNormalized, wavelet-LH_glcm_ClusterShade, and log-sigma-3-mm-3D_glrlm RunLengthNonUniformityNormalized were associated with stromal fibrosis/collagen disorder, cellular density gradient heterogeneity, ECM remodeling, neovascular distortion within the stroma, and tumor–stroma interface disruption. These pathological features are closely related to cancer progression. For example, log-sigma-2-mm-3D_glszm_SmallAreaLowGrayLevelEmphasis has been reported to predict progression-free survival in head and neck squamous cell carcinoma [33], and log-sigma-5-mm-3D_glszm Zone% has been linked to the growth rate of renal tumors [34]. Furthermore, these regions align well with ESRP1’s known functional roles. Gökmen-Polar et al. demonstrated that ESRP1 can drive ER⁺ BRCA progression by modulating fatty acid/lipid metabolism and redox pathways [35]. Another study found that ESRP1 regulates the alternative splicing of MENA and promotes EMT, facilitating tumor cell invasiveness in BRCA [36]. These findings suggest that ESRP1 not only exhibits marked expression advantages in the C1 subtype but also shows strong consistency in spatial structure, pathological features, and functional associations with metabolism and invasiveness. Collectively, this indicates that ESRP1 may serve as a core driver gene in the C1 subtype, playing a pivotal role in tumor progression and remodeling of TME.

Further cell-cell communication analysis revealed that fibroblasts play a central role in the communication network within the BRCA TME. As the predominant stromal cells in the TME, fibroblasts can influence cancer cell metastasis and therapeutic resistance by secreting extracellular matrix proteins and signaling molecules such as TGF-β and IL-6 [37, 38]. TGF-β serves as a critical regulatory axis for fibroblast function during malignant tumor progression. TGF-β signaling promotes EMT, facilitating tumor cells to lose their epithelial characteristics and acquire enhanced migratory and metastatic abilities [39]. Notably, TGF-β1 is not only a key cytokine responsible for the activation of cancer-associated fibroblasts [40] but also an essential factor for sustaining their pro-tumorigenic functions. Our analysis showed that fibroblasts act as both major sources and important targets of TGF-β signaling, establishing extensive bidirectional communication networks with multiple cell types, including epithelial cells, B cells, and neutrophils. Importantly, previous studies have indicated that in hypoxic conditions, the TGF-β–ESRP1 axis synergistically induces EMT and promotes BRCA metastasis [36]. ESRP1 exhibits strong functional coordination with the TGF-β pathway: TGF-β can regulate ESRP1 expression to promote epithelial cell transition into a mesenchymal phenotype [41], while ESRP1-mediated splicing events—such as the CD44v/CD44s and FGFR2-IIIb/IIIc isoform switches—directly influence tumor cell adhesion and motility. In our study, we observed strong interactions between fibroblasts and multiple cell types in regions of high ESRP1 expression. These findings further support a potential synergistic relationship between ESRP1 and fibroblast-mediated TGF-β signaling. We speculate that in the C1 subtype, ESRP1 may modulate alternative splicing events in coordination with TGF-β signals released by fibroblasts, thereby enhancing EMT and immunosuppressive potential in BRCA. This synergy likely contributes to TME remodeling and facilitates BRCA metastasis.

From a clinical perspective, the C1/C2 subtype classification based on pathological image features may provide exploratory insights into BRCA heterogeneity, but its clinical utility remains to be established. Because H&E-stained slides are routinely available, this classification could potentially be implemented as a computational pathology tool, though its performance in real-world decision-making is unproven. The C1 subtype is associated with elevated ESRP1 expression, metabolic reprogramming, and an immunosuppressive microenvironment, which generate hypotheses for future mechanistic studies.

However, some limitations should be noted. First, these observations are based on internal validation within the TCGA-BRCA cohort, and treatment information was not incorporated, limiting the ability to rule out treatment-related confounding. The proposed subtypes and their prognostic value were established and internally validated using a random training/validation split, but the C1/C2 classification should currently be considered exploratory and hypothesis-generating rather than ready for clinical application. Second, while ST and scRNA-seq data were utilized to reveal the spatial distribution of key genes and potential intercellular communication, direct experimental evidence from in vitro or in vivo models is still lacking. Future studies should explore the mechanistic role of the ESRP1–TGF-β axis in BRCA and evaluate the potential efficacy of anti-TGF-β or anti-fibrotic therapies in patients with the C1 subtype. Third, approximately 28–30% of TCGA-BRCA samples lacked complete ER, PR, or HER2 annotation, and these missing cases were excluded from subtype-stratified analyses. This introduces a potential selection bias, which should be considered when interpreting the apparent independence of C1/C2 classification from conventional molecular subtypes. The reduced sample size may also partly explain why survival differences between C1 and C2 were significant only in the luminal A-like subgroup. Further validation in larger, well-annotated, and independent cohorts is warranted to confirm these findings and assess their clinical relevance.

Conclusion

In summary, this study identified two prognostically distinct subtypes of BRCA based on pathological image features and validated them through transcriptomic, single-cell, and ST data. Two subtypes with distinct prognostic outcomes were defined, among which the C1 subtype exhibited high ESRP1 expression, fibroblast enrichment, and a TGF-β-centered intercellular communication profile, suggesting a potentially more aggressive phenotype and poorer clinical prognosis. Importantly, the C1/C2 classification was shown to be largely independent of conventional ER/PR/HER2-based molecular subtypes, suggesting its potential as a complementary prognostic tool. ESRP1 was identified as a key regulatory factor that may promote tumor progression by mediating EMT and stromal remodeling, though these mechanisms remain to be experimentally validated. Overall, our findings provide exploratory insights into BRCA heterogeneity and highlight potential avenues for subtype-specific investigation.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(1.5MB, xlsx)}

Supplementary Material 2 ^{(13.6KB, xlsx)}

Supplementary Material 3 ^{(10.4MB, docx)}

Author contributions

LGX contributed to the conception and design. LGX, XFT, XPZ and RNG contributed to the collection and assembly of data. XHZ, XJL, MZQ and RKL analyzed and interpreted the data. All authors wrote and approved the final manuscript.

Funding

Not applicable.

Data availability

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Conflict of interest

The authors report no conflict of interest.

Ethical approval and consent to participate

In accordance with the Declaration of Helsinki, The Ethics Committee of Longyan First Affiliated Hospital of Fujian Medical University deemed that this research is based on open-source data, so the need for ethics approval was waived.

Consent for publication

Not applicable.