Multiplatform single-cell and spatial transcriptomics reveal sex-specific malignant states and architecture in male breast cancer

Abstract

Male breast cancer (MBC) is rare and remains largely managed using knowledge derived from female breast cancer. Here, we generated a multiplatform single-cell and spatial transcriptomic atlas integrating scRNA-seq, SeekSpace, Visium HD, Xenium In Situ, and multiplex immunohistochemistry data from male and female breast cancer cohorts, covering more than 660,000 cells. We identified male-specific tumor cells (MSTCs) that were enriched in MBC, rare in female breast cancer, and associated with poorer disease-free survival. MSTCs exhibited neural transcriptional programs, increased transcriptome-inferred copy number variation burden, and spatial coupling with fatty acid metabolism signals. MSTC-enriched regions showed reduced antigen-presentation signatures and spatial association with APOE⁺/CD163⁺ macrophages, with GRN-SORT1 emerging as a candidate macrophage-tumor interaction axis. These findings define an MBC-associated malignant state and its immune-metabolic spatial context.

INTRODUCTION

Male breast cancer (MBC) accounts for only ∼1% of all breast cancers and represents a rare disease with unique clinical characteristic (1, 2). Statistical data demonstrate that compared with female breast cancer (FBC), MBC patients have worse prognosis and higher mortality rate (3), attributable to three key factors: First, the scarcity of MBC samples leads to insufficient understanding of its clinical and molecular characteristics; second, due to the specificity of MBC, most breast cancer clinical or basic research studies exclude it, resulting in relatively limited research data on the cellular biological characteristics of MBC; lastly, clinical understanding for MBC are based on FBC research data, leading to treatment strategy selection for MBC often requiring reference to FBC treatment experience (4, 5). However, MBC and FBC exhibit substantial differences in clinical, pathological, and molecular characteristics (6), particularly in the sex-selective expression of steroid hormone receptors (7). Research indicates that, even after excluding clinical-related factors, the overall prognosis of MBC remains significantly worse than that of FBC, suggesting potential unelucidated biological differences between the two.

In recent years, increasing evidence demonstrates that sex is an important factor affecting tumor development and progression (8, 9). Sex differences not only influence tumor metastatic capacity and metabolic behavior but also participate in regulating tumor cell immune escape mechanisms and tumor immune microenvironment (TIME) remodeling, as confirmed in malignant tumors such as bladder cancer (10) and colorectal cancer (11). Sex-related differences cause significant variations in proliferation, metabolism, and immune responses between males and females in malignant tumors (12, 13), highlighting the urgent need for comprehensive MBC research.

The rapid development of single-cell and spatial transcriptomic technologies has provided unprecedented opportunities to dissect breast tissue architecture and tumor ecosystems. Recent studies have generated single-cell and spatial atlases of the adult human breast, aging breast tissue, metastatic breast cancer, and triple-negative breast cancer (TNBC), revealing epithelial diversity, malignant cell-state heterogeneity and spatially organized tumor-microenvironment interactions (14–17). However, existing studies mainly focus on common tumor types, with research on sex differences remaining scarce. Although previous studies conducted single-cell RNA sequencing (scRNA-seq) on six MBC samples (18), deepening our understanding of sex factors in breast cancer TIME composition and metabolic differences, sample size limitations hindered further in-depth research.

Based on the current research status and challenges described above, this study constructed a single-cell and single-cell spatial sequencing dataset containing 23 FBC samples and 55 MBC samples, covering three different spatial transcriptomics platforms, systematically investigating the differences between MBC and FBC. Through this comprehensive analysis, we identified a male-specific tumor cell (MSTC) subtype highly enriched in MBC, termed MSTCs, and found that their presence is associated with an immunosuppressive microenvironment and poor prognosis. Notably, although MSTCs are rare and scarcely detected in FBC, their presence still correlates with worse survival outcomes. These findings provide critical theoretical insights and a robust data foundation for understanding the unique biological characteristics of MBC and developing targeted therapeutic strategies.

RESULTS

Sex-differentiated single-cell and spatial atlas of human breast cancer

To comprehensively explore the impact of sex on breast cancer cell ecosystems and their molecular characteristics, we adopted a multilevel, multiplatform comprehensive research strategy. First, we performed droplet-based scRNA-seq (10x Genomics) on fresh surgical resection tissue samples from 13 untreated MBC patients (table S1) while integrating published scRNA-seq data, including 6 MBC samples (18) and 17 FBC samples (Fig. 1A) (19). To achieve effective integrative analysis across samples, we used the Harmony (20) method in the Seurat V5 software package for batch correction (fig. S1A). Through rigorous quality control and data integration, we obtained high-quality single-cell transcriptome data for 233,606 cells (Fig. 1B). Based on graph-based uniform manifold approximation and projection (UMAP) analysis combined with canonical gene marker expression patterns, we successfully identified seven major cell types: breast cancer cells (n = 127,414), myeloid cells (n = 21,834), T and natural killer cells (n = 46,114), B and plasma cells (n = 13,232), fibroblasts (n = 12,911), endothelial cells (n = 10,963), and mast cells (n = 1138) (fig. S2, B and C, and table S2A). After quality control, each sample retained abundant major cell types (Fig. 1C and table S2B). Using mLLMCelltype (21) (a large language model-based scRNA-seq cell-type annotation tool), we further clustered and annotated cell subpopulation subsets (fig. S1D), identifying a total of 33 cell subtypes (table S2C).

Fig. 1. Single-cell spatial transcriptomic analysis reveals the cell populations of MBC and FBC.

(A) Workflow of multiplatform single-cell and spatial transcriptomics in MBC, integrating scRNA-seq, spatial profiling, and validation to map cell types and male-specific tumor states. (B) Uniform manifold approximation and projection (UMAP) of the integrated scRNA-seq cohort colored by major lineages: cancer cells, T cells, B cells, myeloid cells (macrophages/monocytes), endothelial cells, fibroblasts, mast cells, and plasmablasts. Cohort size: 36 donors (male, n = 19; and female, n = 17). Tn, naïve T cells; Tmem, memory T cells; Tex, exhausted T cells; Teff, effector T cells; T_reg, regulatory T cells; NK, natural killer cells; DC, dendritic cells; ArtEC, arterial endothelium; CapEC, capillary endothelium; venec, venous endothelium; LEC, lymphatic endothelium; iCAF, inflammatory CAF; myCAF, myofibroblastic CAF; meCAF, metabolic CAF; pvCAF, perivascular CAF. (C) Stacked bar charts of lineage composition by sex and, to the right, per-donor compositions. (D) SeekSpace single-cell spatial maps (patients 1 to 10) colored by cell type as in (A); panel titles report per-section nuclei count (N). (E) Visium HD FFPE spatial maps (female sections H1-649ZJJX, H1-BYFBQ82, and H1-GPWTZY6; male patients 11 to 15) colored as in (A); panel titles report per-section cell counts (N). (F) Composition of major lineages per spatial section across SeekSpace and Visium HD datasets colored as in (A). (G) Sex-stratified frequencies of cancer cell states, including cancer stem cells (CSCs), MSTCs, proliferative cancer cells (Pro-Cells), and stress cancer cells (Stress-Cells) and T cell subsets (NK, Teff, Tex, Tmem, Tn, and T_reg). Only statistically significant P values are annotated above comparisons; nonsignificant comparisons are unlabeled. (H) Sex-stratified proportions of macrophage subpopulations (APOE/CD163, APOE, IL1B, ISG15, MKI67, MMP9, SLC40A1, SPP1, Mono_CD14, Mono_CD16, and DCs). Only statistically significant P values are annotated above comparisons; nonsignificant comparisons are unlabeled.

To obtain spatially resolved transcriptome information, we collected fresh surgical resection tissue samples from 10 patients with untreated breast cancer (7 males and 3 females) for SeekSpace single-cell nuclear spatial transcriptome sequencing, capturing transcriptome data from 143,821 cell nuclei (Fig. 1D, fig. S1H, and table S3). SeekSpace technology can capture transcriptome information at the single-cell nuclear level while providing spatial localization, enabling us to visualize the spatial distribution characteristics of the tumor microenvironment (22, 23). We similarly used the Harmony method in the Seurat V5 (24) software package for batch correction and used the robust cell-type decomposition (RCTD) algorithm (25) to accurately annotate cell identities based on prior information from single-cell datasets (fig. S1E). To further validate result reliability and exclude the influence of technical platform differences, we also used Visium HD (26) spatial gene expression technology with single-cell resolution to perform spatial transcriptome sequencing on five paraffin-embedded MBC specimens while integrating analysis of three publicly available female Visium spatial transcriptome data (Fig. 1E; fig. S1, I to K; and table S4). By fine-tuning the Cellpose (27) cell segmentation model, we precisely aligned 2-μm-resolution capture areas with hematoxylin and eosin (H&E) images and segmented cell nuclei. Considering the technical challenges of cell boundary identification and transcript assignment, we only summed transcript counts in areas where cell nuclear segmentation masks overlapped with 2-μm bins, creating pseudo-single-cell-level gene expression matrices (28). We performed standard analysis procedures on each single-cell-level gene expression matrix from Visium HD and used the RCTD algorithm to annotate each cell’s identity. UMAP analysis results showed that single-cell matrices obtained through cell segmentation algorithms could clearly confirm cell identities in Visium HD data (fig. S1F). Notably, the cell-type distributions observed in spatial transcriptomics data (Fig. 1F) differed from those in scRNA-seq data (Fig. 1C), with higher proportions of cancer cells and fibroblasts detected in spatial platforms. Because the scRNA-seq and spatial transcriptomic datasets were generated from different individuals and by distinct tissue-processing and analytical workflows, these differences should be interpreted as descriptive cross-platform observations rather than direct estimates of biological differences in cell-type abundance. Therefore, we used the spatial datasets primarily to assess spatial localization and tissue architecture.

Through this comprehensive analysis, we established a comprehensive MBC single-cell spatial transcriptome atlas, providing a robust data foundation for in-depth exploration of unique cell subpopulations, cell states, and intercellular interactions in MBC. Preliminary statistical analysis showed that MSTCs were significantly enriched in MBC, accounting for ∼40% of the malignant cells in MBC compared with only about 5% in FBC (Fig. 1G); compared with FBC, MBC exhibited significantly reduced overall T cell proportions but significantly increased regulatory T (T_reg) cell proportions; among macrophage subpopulations, APOE⁺/CD163⁺ macrophages were enriched in MBC, whereas IL1B⁺ macrophages were enriched in FBC (Fig. 1H). Furthermore, meCAFs were significantly higher in males, whereas pvCAFs were lower and VenECs were enriched in males (fig. S1G).

MSTCs associated with poor breast cancer prognosis

To understand transcriptional difference patterns in cancer cells between MBC and FBC, we performed systematic subclustering analysis of cancer cell subpopulations. Through unsupervised clustering, we identified four major cancer cell subpopulations: cancer stem cells (CSCs; high expression of PROM1 and CD44; low expression of CD24), stress cancer cells (Stress-Cells; high expression of HSPs genes), proliferative cancer cells (Pro-Cells; high expression of MKI67 and other cell proliferation–related genes), and MSTCs (high expression of GRIA2, NEGR1, and GFRA1) (Fig. 2A and table S5). To explore whether breast cancer molecular markers—ER, PR, and HER2—influence the formation of MSTCs distribution differences between MBC and FBC, we analyzed the correlation between these molecular marker expression levels and the MSTC signature score. Results showed that, in both MBC and FBC, the MSTC signature score positively correlated with ER expression levels, with stronger correlation in MBC than in FBC (Fig. 2B and fig. S2A). This phenomenon was validated in independent MBC bulk-level RNA sequencing (RNA-seq) data (29) and TCGA data (fig. S2B) (30). In contrast, at the single-cell level, the correlations between the MSTC signature score and PR or HER2 expression were negligible, with low correlation coefficients (Fig. 2B and fig. S2A). Bulk RNA-seq data showed a positive association between the MSTC signature score and PR expression (fig. S2B), whereas single-cell analysis showed only a very weak correlation, not supporting PR as a meaningful cell-level correlate of the MSTC program. Together, these results suggest that the MSTC signature is more closely associated with ER-related expression than with PR or HER2 expression at single-cell resolution.

Fig. 2. Spatial distribution and prognostic relevance of MSTCs.

(A) UMAP of malignant cells classified as CSCs, MSTCs, Pro-Cells, and Stress-Cells, with marker-gene dot plot. (B) Association of MSTC signature scores with ER, PR, and HER2 expression by sex and normalized ER distribution; significant P values are shown. (C) Representative spatial maps of normalized MSTC signature scores at single-cell/pseudo-single-cell resolution, with adjacent violin plots by RCTD-annotated cell class. (D) Representative multiplex immunohistochemistry (mIHC) images from FBC and MBC sections stained for 4′,6-diamidino-2-phenylindole (DAPI), pan-cytokeratin (panCK), GRIA2, NEGR1, and GFRA1; boxed regions are enlarged. Colors: DAPI, blue; GRIA2, orange; panCK, white; NEGR1, yellow; GFRA1, red. Scale bars, 300 and 20 μm. (E) HALO-based MSTC quantification in the mIHC validation cohort (32 FBC and 43 MBC samples; two-sided t tests). (F) MSTC percentages per sample/section by sex. (G) Bulk RNA-seq MSTC signature scores by sex in TCGA-BRCA and GTEx breast tissues. (H and I) Kaplan-Meier disease-free survival (DFS) analyses stratified by high versus low MSTC signature score in TCGA-BRCA (H) and the retrospective MBC cohort (I); log-rank P values and 95% confidence intervals are shown.

Spatial transcriptomic analysis further showed that high MSTC signature scores were mainly localized to tumor-enriched regions in MBC sections (Fig. 2C and fig. S2, C and D). In these regions, MSTC marker genes, including GRIA2, NEGR1, and GFRA1, showed spatially overlapping expression patterns, supporting the tumor cell localization of the MSTC program at the spatial transcriptomic level. To further examine this finding at the protein level, we performed multiplex immunohistochemistry (mIHC) on 32 FBC and 43 MBC paraffin samples. Pan-cytokeratin (panCK) was used to label epithelial tumor cells, and GRIA2, NEGR1, and GFRA1 were used to label MSTCs. HALO-based quantification showed that the proportion of MSTCs was significantly higher in MBC than in FBC (Fig. 2, D to F, and table S6).

Through consistent validation at multiple levels including single-cell, single-cell nuclear spatial transcriptome, Visium HD, and spatial proteomics, we confirmed the high enrichment characteristics of MSTCs in MBC. To determine whether the MSTC signature score is an early characteristic of male glandular development or formation, we analyzed the MSTC signature score in breast cancer and noncancerous tissues from TCGA and GTEx databases. Results showed that, although the MSTC signature score in MBC was higher than in FBC, in noncancerous tissues, the MSTC signature score in males was lower than in females (Fig. 2G). This important finding suggests that MSTC differentiation may be a characteristic acquired during MBC tumor development rather than an innate property of male glandular tissue.

To assess the clinical relevance of the MSTC program, we first performed survival analysis using the TCGA-BRCA cohort. Patients with high MSTC signature scores showed poorer disease-free survival (DFS) (Fig. 2H and table S7). Consistently, retrospective survival analysis of 36 MBC patients with available DFS data showed that patients with higher MSTC proportions had worse DFS (Fig. 2I and table S8). Together, these transcriptomic and mIHC-based analyses showed that MSTCs were enriched in MBC and were associated with poor prognosis. To further determine whether this prognostic association was a general feature of malignant cell-state programs, we also performed DFS analyses for CSC, Stress-Cell, and Pro-Cell signatures in the TCGA-BRCA cohort. In contrast to the MSTC signature, high CSC, Stress-Cell, and Pro-Cell signature groups were not significantly associated with DFS (fig. S2, E to G). These results suggest that the adverse prognostic association was more pronounced for the MSTC-associated program than for the other malignant cell-state signatures examined.

Neural transcriptional and genomic stress programs of malignant MSTCs

Previous bulk RNA-seq studies identified two MBC subtypes, J_M1 and J_M2, with distinct prognoses. Johansson et al. (29) reported that J_M2 tumors contained transcription factors related to nervous system development, but the cellular origin remained unclear. To determine whether MSTCs represent the source of these features, we performed J_M1/ J_M2 scoring at single-cell resolution using Johansson’s classifier genes. CSCs exhibited the highest J_M1 scores, consistent with their aggressive phenotype, while MSTCs showed the highest J_M2 scores (Fig. 3A). This appeared paradoxical because J_M2 associates with good prognosis, whereas MSTCs correlate with poor outcomes. We decomposed the J_M2 gene set into functional modules: neuro-related, immune activation, metabolic, and other genes. MSTCs selectively expressed neuro-related module genes while showing low immune activation gene expression. In contrast, Stress-Cells and Pro-Cells exhibited higher immune gene expression (Fig. 3B and fig. S3A). These findings suggest that MSTCs represent a major cellular contributor to the neuronal component of the bulk J_M2 signature while lacking the immune-activation component that may underlie the favorable prognosis of classical J_M2 tumors.

Fig. 3. Transcriptional and genomic features of MSTCs.

(A) Violin plots of J_M1 and J_M2 meta-module scores across malignant states. (B) J_M2 submodule scores for neuronal, metabolic, immune, and other genes; significant P values are shown for (A) and (B). (C) Differential transcription factor activity highlighting MSTC-enriched TFs. (D) Regulon connection specificity index (CSI) matrix defining three TF modules, including neural-related hubs—NFIA, TAF1, KLF7, and CLOCK. (E) Pseudo-bulk gene-expression heatmap of malignant subtypes across samples, annotated by subtype, sex, age, tumor-node-metastasis (TNM) stage, ER, PR, and HER2 status; values are z scores. (F) NMF-derived coactivation matrix and per-state activity heatmap for seven meta-modules: M1, stress response; M2, ECM/cell adhesion; M3, rapid stimulus response; M4, EMT/growth-factor signaling; M5 and M6, neural/synaptic programs; and M7, DNA repair, cell cycle, and chromatin modification. (G) Representative inferCNV heatmaps from male and female samples and sample-level inferCNV scores for MSTCs versus other cancer cells; red, gain; blue, loss; white, neutral. (H) Monocle trajectory of pooled, downsampled tumor cells colored by subtype with normalized pseudotime and branch proportions. (I) Pseudotime-ordered gene set heatmap showing three functional stages: respiration, circadian/neural synapse, and fatty acid/glutamatergic/oxytocin signaling.

To further characterize MSTCs at the molecular level, we systematically analyzed transcription factors specifically activated in these cells. Through transcription factor activity analysis, we identified 53 transcription factors specifically activated only in MSTCs (see Materials and Methods; fig. S3, B and C, and table S9), including KLF7 (31, 32), which regulates neural development, and TAF1 (33, 34), which regulates neural differentiation and axon growth (Fig. 3C). Using connection specificity index (CSI) to evaluate correlations between regulon pairs, we identified and enriched three major transcriptional regulatory functional modules, with module three enriched for transcription factors related to nervous system development [NFIA (32), TAF1, KLF7, and CLOCK (35)] (Fig. 3D and table S9).

To understand potential transcriptional expression pattern differences within MSTCs, we performed pseudobulk analysis and unsupervised hierarchical clustering on cancer cell subtypes from different patients. Compared with differences between cell subtypes, transcriptional differences caused by sex factors were more significant. Although some transcriptional heterogeneity still existed between patients, we observed that a subset of MSTCs exhibited highly similar transcriptional characteristics, whereas others showed more heterogeneous expression patterns, suggesting the existence of distinct transcriptome modules within this cell subtype (Fig. 3E). Considering that tumor cell development primarily depends on specific functional program modules that enable tumor cells to adapt to unique TIME and interact with specific cellular components, we used nonnegative matrix factorization (NMF) methods (36) to identify potential functional programs shared among different breast cancer cell subpopulations to mitigate the impact of interpatient heterogeneity on analysis results (37, 38). Through NMF analysis, we defined seven major meta-program (MP) modules (Fig. 3F and fig. S3D) and annotated them on the basis of characteristic genes and functional enrichment of each module: M1, cellular stress response; M2, extracellular matrix organization and cell adhesion; M3, rapid cellular response to stimuli; M4, cell growth and proliferation factors and epithelial-mesenchymal transition signals; M5, neural development synaptic function and neuronal signal transduction; M6, synaptic organization and neural signal transduction; and M7, DNA repair, cell cycle, and chromatin modification (Fig. 3F and fig. S3E). As expected, Stress-Cells and Pro-Cells showed high activity of M1 and M2 modules, CSC highly expressed M3 and M4 modules, whereas MSTCs highly expressed M5 and M6 modules. Consistent with these module-level findings, Gene Ontology (GO) enrichment analysis of differentially expressed genes (DEGs) across malignant cell states showed that MSTCs were preferentially enriched for neural/synaptic programs, including regulation of membrane potential, synapse organization, axonogenesis, and synapse assembly. In contrast, CSCs were enriched for stem cell differentiation, gland development, and extracellular-matrix/cytoskeletal organization, whereas Pro-Cells and Stress-Cells mainly showed translational, ribosome-biogenesis, ribonucleoprotein-biogenesis, and RNA-processing programs (fig. S3F). Notably, compared with other cancer cell subtypes, MSTCs also highly expressed the M7 module, which is closely related to DNA repair and chromatin abnormalities, suggesting that this cell state may harbor a transcriptional program associated with genomic stress and chromatin remodeling. This finding is biologically consistent with previous reports linking BRCA2 alterations in MBC to homologous recombination deficiency and chromosomal instability. BRCA2 mutations are more common than BRCA1 mutations in MBC (12% versus 1% in Italian cohorts) (39) and are associated with defective homologous recombination repair pathways, increased chromosomal instability, and characteristic mutational signatures indicative of homologous recombination deficiency (40, 41). The elevated expression of DNA repair–related genes in MSTCs may represent a compensatory cellular response to underlying genomic stress.

To further examine whether this DNA repair–/chromatin-related transcriptional program was accompanied by transcriptome-inferred copy-number alteration patterns, we performed inferCNV (42) analysis in malignant cells from FBC and MBC samples (Fig. 3G). We further quantified transcriptome-inferred copy number variation (CNV) burden using an inferCNV score calculated from the squared sum of inferCNV values. In MBC samples, MSTCs showed significantly higher median inferCNV scores than other cancer cells, whereas this difference was not significant in FBC samples (Fig. 3G). These results suggest that MSTCs in MBC exhibit increased transcriptome-inferred CNV burden relative to other malignant cells, consistent with the enrichment of DNA repair–/chromatin-related transcriptional programs in this cell state. Through pseudotime trajectory analysis, we inferred five differentiation trajectories related to cancer cell-subtype development. We found differences between MBC and FBC in cancer cell development processes, with trajectories 1, 2, and 3 primarily enriched with MBC cells and with trajectories 4 and 5 primarily enriched with FBC cells (Fig. 3H and fig. S4A). MSTCs were enriched at the terminal ends of tumor cell trajectories and showed lower CytoTRACE2-relative scores (fig. S4B), consistent with reduced developmental potential and a relatively more differentiated tumor cell state in MBC. Differential expression gene analysis along pseudotime trajectories revealed three major biological modules at key turning points of cancer cell lineage trajectories: The first stage enriched aerobic respiration-related biological pathways, the second stage enriched circadian rhythm and neuronal synapse formation-related pathways, and the third stage enriched fatty acid and cholesterol metabolism activity pathways (Fig. 3I and fig. S4C). Together, these analyses indicate that MSTCs represent a malignant cell state characterized by neural/circadian transcriptional regulation, DNA repair–/chromatin-related genomic stress features, and increased transcriptome-inferred CNV burden in MBC.

Spatial neural-lipid coupling and reduced immunogenicity in MSTC-enriched regions

Having defined the neural/circadian and genomic stress features of MSTCs, we next examined whether the MSTC-associated neural program was coupled to metabolic and immune features at single-cell and spatial levels. Previous studies reported fatty acid metabolism as an important feature of MBC, but whether this metabolic program overlaps with the MSTC-associated neural program remained unclear (18). We therefore calculated glutamate metabolism, neural signal, fatty acid metabolism, and angiogenesis scores across malignant cell states. MSTCs showed higher glutamate metabolism, neural signal, and fatty acid metabolism scores than CSCs, Pro-Cells, and Stress-Cells (see Materials and Methods; Fig. 4A). Pairwise correlation analysis further showed that fatty acid metabolism scores were more strongly associated with glutamate metabolism and neural signal scores than with angiogenesis scores (Fig. 4B). These results suggest that the MSTC-associated neural program is preferentially coupled to lipid metabolism rather than angiogenesis at single-cell resolution.

Fig. 4. Spatial neural-lipid coupling and reduced immunogenicity in MSTCs.

(A) Violin plots comparing four pathway scores across malignant states (CSC, MSTC, Pro-Cells, and Stress-Cells): glutamate metabolism, neural signal, fatty acid metabolism, and angiogenesis. Only statistically significant P values are annotated above comparisons; nonsignificant comparisons are unlabeled. (B) Scatter plots showing pairwise associations between neural signal or glutamate metabolism scores (x axes) and fatty acid metabolism or angiogenesis scores (y axes). Points are colored by malignant state as in (A); lines indicate fitted trends per state and overall. Pearson’s r and two-sided P are reported in panels. (C) Representative joint density maps showing the spatial distribution of neural signal and fatty acid metabolism scores in male and female sections. (D) Distribution of cells jointly high for neural signal and fatty acid metabolism scores (co-high%; threshold defined in Materials and Methods) across major cell classes, stratified by sex (male sections, N = 12; and female sections, N = 6). P values were calculated using Kruskal-Wallis tests for multigroup comparisons and two-sided Wilcoxon rank sum tests for two-group comparisons. (E) Sex-stratified associations between fatty acid metabolism score and relative mean distance to MSTC-enriched regions. The x axis shows normalized relative mean distance; Pearson’s r and two-sided P values are shown. (F to H) Gene set enrichment analyses (MSTCs versus other malignant cells) showing negative enrichment scores for MHC class I, MHC class II, hypoxia, and tumor angiogenesis, indicating lower activity of these programs in MSTCs. (I to K) Corresponding z-score heatmaps of gene sets across malignant states (CSC, MSTC, Pro-Cells, and Stress-Cells), supporting the enrichment results.

We next evaluated whether this neural-lipid coupling could be observed in spatial transcriptomic datasets. Representative joint density maps from SeekSpace and Visium HD showed spatial overlap between neural signal and fatty acid metabolism scores in MBC sections (Fig. 4C and fig. S5, A and B). To quantify this pattern, we calculated the proportion of cells jointly high for neural signal and fatty acid metabolism scores (“co-high%”) across spatial sections. This analysis showed that co-high cells were enriched in cancer cell regions and were more frequent in MBC than in FBC sections (Fig. 4D). In addition, distance-based spatial analysis showed that fatty acid metabolism scores were higher in regions closer to MSTC-enriched areas in MBC (Fig. 4E). Together, these analyses support spatial coupling between MSTC-associated neural programs and fatty acid metabolism in MBC tissue architecture.

Given previous observations linking fatty acid metabolism with reduced lymphocyte functional signals in MBC, we next examined whether MSTCs also exhibited altered immunogenicity-related transcriptional programs. Gene set enrichment analysis (GSEA) of DEGs in MSTCs showed down-regulation of immune-recognition and antigen-presentation–related pathways, including major histocompatibility complex (MHC) class I and MHC class II programs (Fig. 4, F and I). Hypoxia- and angiogenesis-related programs were also reduced in MSTCs relative to other malignant states (Fig. 4, G to K). These data suggest that MSTCs combine neural-lipid metabolic coupling with reduced antigen-presentation signatures, consistent with a low-immunogenicity malignant cell state.

MSTCs in MBC are associated with immunosuppressive macrophages

To explore potential explanations for the differences in immune infiltration between MBC and FBC, we systematically evaluated chemokine and cytokine expression patterns of all cell types in the tumor microenvironment of both. Results showed that CXCL9/10/11 (43) related to immune activation were highly expressed in macrophages in FBC, whereas MBC was primarily enriched with myeloid-derived immunosuppressive cytokine-related genes, suggesting that the reason for lower immune cell infiltration in tumor microenvironment of MBC may be due to enrichment of myeloid-derived immunosuppressive cells (Fig. 5A and fig. S6A).

Fig. 5. Coenrichment of MSTCs and macrophage subpopulations in MBC.

(A) Dot plot showing selected cytokine and chemokine genes across major cell classes, grouped by sex. Genes are grouped into immune-activation–related (IA) and immunosuppressive/myeloid-suppressive (ISM) categories. Circle size denotes the fraction of expressing cells; color denotes mean expression per class. (B) Ligand-receptor analysis. Scaled expression heatmap of chemokine genes among cancer cells clusters (left) and TAM clusters (right). Interactions are connected by lines and colored by the dominant TAM clusters. (C) Pearson correlation between MSTC proportions and total T cells (in TIME)/total macrophage (in TIME)/APOE⁺/CD163⁺ macrophage/T_reg cells proportions across patients. (D) Compare the distribution density differences of APOE⁺/CD163⁺ macrophages in males and females using mIHC data, as well as Pearson correlation with MSTC proportions. (E) Representative mIHC images showing coenrichment. Marked by DAPI, blue; panCK, white; GRIA2, orange; NEGR1, yellow; CD163, green; and APOE, light blue. Channel-separated images are shown in fig. S6B, with merged panels from (E) as channel references. (F) BANKSY identified spatially defined tissue domain, color by region. Dot plots displayed the MSTC signature score, group by region. (G) The barplot shows the proportions of APOE⁺/CD163⁺ macrophages in TIME and the proportions of MSTCs colored by region. (H) The spatial heatmap highlights the gene expression levels. (I) Enrichment results of differential gene pathways in R2 and R7. (J) Enrichment of gene signature based on GSEA analysis in R2 and R7. Permutation test, n = 100,000 permutations.

Ligand-receptor inference revealed distinct putative communication patterns between malignant cell states and macrophage subpopulations: Stress-Cells showed broad inferred interactions with ISG15⁺ macrophages, Pro-Cells showed CCR3-related inferred interactions with MMP9⁺ macrophages, whereas MSTCs showed CCR10-related inferred interactions with SPP1⁺ macrophages and CXCR2-related inferred interactions with APOE⁺/CD163⁺ macrophages (Fig. 5B). In single-cell-level correlation analysis, we found that MSTC proportion changes negatively correlated with total T cell proportions [correlation coefficient (r) = −0.4, P = 0.0021] in the TIME and positively correlated with macrophage proportions (r = 0.46, P = 0.0032). Further subpopulation analysis showed that MSTC proportions were positively correlated with both T_reg cell proportions (r = 0.7, P < 0.001) and APOE⁺/CD163⁺ macrophage proportions (r = 0.34, P = 0.03) (Fig. 5C). mIHC validation results showed that APOE⁺/CD163⁺ macrophage proportions in MBC were significantly higher than in FBC, and APOE⁺/CD163⁺ macrophages were positively associated with MSTC proportions in the TIME (Fig. 5D). Spatial analysis revealed that APOE⁺/CD163⁺ macrophages and MSTCs showed spatial proximity in the tumor microenvironment (Fig. 5E and fig. S6, B to D), suggesting potential cellular interactions between these two populations. These results indicate that, in MBC, MSTCs show CXCR2-related inferred signaling with APOE⁺/CD163⁺ macrophages and are enriched in TIME regions with low T cell infiltration, consistent with immune-desert–like microenvironments, although the direction of causality cannot be established from these data.

To further examine this spatial association, we used the BANKSY (44) method to perform fine spatial region analysis on Visium HD datasets that exhibited the highest MSTC signature score. This analysis identified nine spatial regions of interest, among which regions R1 to R4 showed high MSTC signature score, R5 to R7 showed moderate scores, and R8 and R9 showed the lowest scores (Fig. 5F). By calculating the proportions of MSTCs and APOE⁺/CD163⁺ macrophages within each region, we identified regions R2 and R7 as key areas of spatial coenrichment (Fig. 5G). To confirm the reproducibility of these findings, the same spatial analysis was also applied to additional Visium HD datasets with sufficient cell coverage, yielding consistent spatial colocalization patterns between MSTCs and immunosuppressive macrophages (fig. S6E).

Compared with other spatial regions, these key focus areas had the highest neural signal intensity, accompanied by the highest APOE⁺/CD163⁺ macrophage signals and T_reg signals (FOXP3) (Fig. 5H). Compared with other regions, highly expressed genes in focus regions were significantly enriched in fatty acid metabolism, mammary gland development, and RET signaling–related pathways (Fig. 5I). Additionally, GSEA analysis revealed selective down-regulation of multiple immune activation pathways in focus regions. Most notably, tumor necrosis factor–α signaling via nuclear factor κB and allograft rejection pathways showed the most significant suppression, along with complement signaling and interferon-γ response pathways (Fig. 5J and table S10). These results indicate that regions with MSTCs enrichment in MBC coincide with selective down-regulation of key antitumor immune pathways, particularly those involved in macrophage-mediated immunity and T cell responses, a pattern consistent with immunosuppressive microenvironments but not sufficient to determine whether MSTCs are the initiating cause of these states.

Cell-cell communication between MSTCs and macrophage subpopulations in MBC

To study gene expression programs and spatial location distributions of MSTCs and explore the impact of MSTC-enriched regions on tumor microenvironment remodeling, we referenced the method established by Reina-Campos et al. (28) and developed a spatial distance–based gene expression gradient analysis method. By calculating the average Euclidean distance from immune cells and stromal cells in the tumor microenvironment to the nearest five MSTCs, we observed spatial distribution relationships between gene expression gradients and relative distances (Fig. 6A). Results showed that, as relative distances between cells increased, gene expression gradients changed significantly. Expression levels of genes related to immune cell and stromal cell functions in the tumor microenvironment all gradually increased with increasing relative distance from MSTCs, including immune function–related CD74, IGHG1, and IGKC and angiogenesis-related CCN1, CCN2, and VIM genes. This finding indicates that MSTC-enriched regions coincide with functionally suppressive tumor microenvironment features. Spatial domain coenrichment analysis showed that APOE⁺ macrophages had high colocalization relationships with ISG15⁺ macrophages, MMP9⁺ macrophages, and MSTCs, which was highly consistent with single-cell-level analysis results. These patterns suggest close spatial and transcriptional coordination between MSTCs and multiple macrophage subpopulations in MBC, raising the possibility that these populations influence each other (Fig. 6, B and C).

Fig. 6. Analysis of cell-cell communications in MBC.

(A) The convolved gene expression of immune or stromal cells along the average relative spatial distance to MSTCs. (B) Heatmap showing the spatial colocalization patterns between different cell types. Rows represent target cell types and columns represent neighboring cell types. The enrichment score indicates the likelihood of spatial proximity between cell-type pairs, with blue indicating low enrichment and red indicating high enrichment. (C) Visualization of spatial enrichment networks using graph. (D) CellChat identifies the interaction networks of macrophages and tumor cells. Nodes represent cell types, and edges represent the number of significant ligand-receptor (LR) pairs. (E) Xenium In Situ analysis of GRN and SORT1 expression in macrophage and tumor cell populations. GRN and SORT1 expression was compared between APOE⁺ TAMs and other macrophages and between MSTCs and other tumor cells. Each point represents one Xenium region; paired Wilcoxon P values are indicated. (F) Spatial association analysis in Xenium data. Top: GRN expression in APOE⁺ TAMs or other macrophages along the standardized distance to the nearest MSTC. Bottom: SORT1 expression in MSTCs or other tumor cells along the standardized distance to the nearest APOE⁺ TAM. Cell density is shown as log₁₀ cell count. (G) Cell-cell communication analysis based on 24 MBC Xenium In Situ data.

Based on this hypothesis, we systematically analyzed intercellular communication between different cell types in MBC. Cell communication network analysis results showed that communication signals from APOE⁺ macrophages, SPP1⁺ macrophages, and ISG15⁺ macrophages were most active in MBC (Fig. 6D and fig. S6F). These macrophage subpopulations were predicted to communicate with MSTCs through two candidate signaling routes: first, through cholesterol and testosterone signaling pathways linked to lipid metabolism and local hormone metabolism in MBC; and second, GRN-SORT1 and glutamate-receptor [Glu–(SLC1A3 + GLS)–GRIA2] signaling networks that are candidate routes of interaction with MSTCs and may be relevant to their tumor-associated transcriptional programs (Fig. 6D). Particularly interesting, the GRN-SORT1 axis plays important roles in neuronal survival and functional maintenance and is closely related to neurodegenerative diseases (45, 46). In this signaling pathway, almost all macrophage subpopulations and cancer cell subpopulations have rich interactions with MSTCs. We compared the TCGA FBC patient cohort and 74 MBC patients from GSE31259. Results consistently showed that, in MBC patients, GRN-SORT1 expression levels were positively correlated, whereas a negligible inverse correlation existed in females (fig. S6G).

Last, we collected 24 independent paraffin-embedded MBC specimens and used 10x Genomics’ Xenium In Situ platform to perform single-cell resolution spatial transcriptomics analysis on a 5001-gene panel. Due to the limited gene panel detected by the Xenium In Situ platform, we performed independent analysis and annotation procedures for Xenium data. After quality filtering, we retained 286,651 cells from 24 independent samples. We used the Harmony method in Scanpy to integrate samples and annotated cell types on the basis of major gene marker expression (fig. S7, A to C).

To comprehensively validate our key findings across an independent platform and larger patient cohort, we performed multiple spatial analyses on the Xenium dataset. First, we validated the spatial colocalization patterns between MSTCs and APOE⁺ macrophages observed in our Visium HD analysis. Spatial proximity analysis revealed that APOE⁺ macrophages were significantly enriched in regions with high MSTCs density compared with regions with other tumor cell subtypes across all 24 samples (fig. S7D). The mean nearest neighbor distance between MSTCs and APOE⁺ macrophages was significantly shorter than that between other tumor cells and APOE⁺ macrophages, confirming the preferential spatial association between these two cell types. Second, we examined the spatial coexpression patterns of the MSTC signature score and fatty acid metabolism pathways. In the independent Xenium dataset, cell-level analysis further showed a positive association between MSTC signature scores and fatty acid metabolism scores (fig. S7E). This result provided additional spatial-transcriptomic support for the coupling between MSTC-associated programs and fatty acid metabolism, although the effect size was modest.

In the Xenium In Situ platform, we found that all 24 MBC samples exhibited high GRN-SORT1 signal activity levels (Fig. 6E). Spatial analysis showed SORT1 expression in tumor cells, including MSTCs, whereas GRN expression was enriched in TIME/macrophage compartments adjacent to MSTC-rich regions. Additionally, ligand-receptor interaction analysis of Xenium In Situ data showed that tumor-associated macrophages in MBC communicate with tumor cells through GRN-SORT1 signals and express CTLA4-related ligands that are associated with T cell receptor expression, suggesting potential macrophage–T cell communication through CTLA4-related interactions (Fig. 6F).

DISCUSSION

Recent research increasingly clearly demonstrates that sex characteristics are key factors causing tumor heterogeneity and affecting tumor development and progression, particularly playing important roles in TIME remodeling and metabolic pathway regulation (47, 48). Despite evidence that fatty acid metabolism influences the immunosuppressive state of MBC, the rarity of MBC has limited in-depth investigation.

Recent single-cell and spatial studies have mapped adult and aging breast tissue, metastatic breast cancer, and TNBC, defining epithelial diversity, spatial tumor architecture, and immune-stromal organization (14–17). However, MBC remains underrepresented in these atlases. Here, we provide a comprehensive single-cell and spatial transcriptomic analysis comparing MBC and FBC. MSTCs are enriched in male disease and rare in female disease, and their enrichment associates with poorer DFS. Neuron-like differentiation phenomena in cancer cells have also been reported in other tumor types, such as neuroendocrine differentiation in prostate cancer associated with treatment resistance and poor prognosis (49) and neuroendocrine characteristics in lung cancer linked to aggressive phenotypes (50). However, neuron-like differentiation in breast cancer has rarely been reported previously. This study provides a comprehensive single-cell and spatial characterization of an MBC-enriched malignant cell state. These cells exhibit a highly malignant epithelial phenotype; however, the precise underlying mechanisms remain to be elucidated. Although the causal relationship is not yet fully understood, we speculate that this phenomenon may be analogous to neuroendocrine differentiation in prostate cancer, representing a specific adaptive alteration during the evolutionary progression of MBC.

Relative to prior bulk classifications, our data refine interpretation of the J_M1/ J_M2 framework. Bulk studies, including work by Johansson and by Severson, defined J_M1/ J_M2 subtypes in MBC, with the J_M2 label often mixing neuronal and immune-activation signals at the aggregate level. At single-cell resolution, these components decouple: MSTCs capture the neuronal component, whereas immune-activation signals arise mainly from other cancer cell states and immune cells. This decoupling reconciles why tumors can score high on bulk J_M2 yet show unfavorable DFS when MSTCs dominate, and it supports interpreting bulk subtypes through a cell-state and spatial lens. Lipid metabolism provides a second axis of context. Bulk studies frequently report that fatty acid metabolism correlates with angiogenesis. Our cell-level data refine this view: The neuronal-metabolic program that defines MSTCs is more strongly coupled to lipid metabolism than to angiogenesis. Thus, lipid and angiogenic signals can covary in bulk-level analyses, whereas MSTCs show a preferential neural-lipid metabolic association with weak angiogenesis-related coupling at single-cell resolution. This resolves an apparent discrepancy between bulk and single-cell results.

The relationship between MSTCs, hormone-receptor expression, and local steroid–metabolic programs also requires careful interpretation. The positive association between MSTC signature scores and ER expression may reflect the ER-positive/luminal epithelial context in which most MBCs arise, consistent with the known predominance of hormone-receptor–positive disease in MBC (4, 7). However, PR expression should not be interpreted as a simple linear surrogate of ER activity at single-cell resolution. Although PR has long been used as a biomarker of ERα function, PR is not merely an ER-induced target gene but can also modulate ERα chromatin binding and transcriptional output (51). Moreover, loss or reduction of PR expression in ER-positive breast cancer may reflect altered cross-talk between ER and growth-factor signaling pathways rather than complete loss of ER activity (52, 53). Therefore, the positive association between MSTC scores and ER, together with the weak association with PR, suggests that MSTCs retain an ER/luminal context but are not defined by canonical ER-PR transcriptional output alone. In this regard, the cholesterol- and testosterone-related ligand-receptor signals identified in our cell-cell communication analysis should be interpreted as candidate local lipid/steroid-metabolic interactions in the MBC microenvironment, rather than as direct evidence that MSTCs are driven by classical ER signaling. This interpretation is consistent with prior work showing broad steroid hormone-receptor interplay in MBC and the role of cholesterol as a precursor for steroid hormone synthesis (7, 54). Future studies integrating ER/PR/AR chromatin occupancy, local steroid–metabolic profiling, and functional perturbation experiments will be required to determine how hormone-related signaling contributes to MSTC biology.

We identified potential links between MSTCs and immunosuppressive microenvironmental features in MBC, including reduced MHC I/II expression and CXCR2-related inferred interactions with APOE⁺/CD163⁺ macrophages. APOE is a gene widely expressed in tumor and immune cells, primarily participating in lipoprotein metabolism (55, 56). Reports indicate that APOE⁺ macrophages are associated with early recurrence of clear cell renal carcinoma (57) and dynamic evolution and immunosuppression of aggressive acral melanoma (58). We identified inferred communication patterns and spatial associations between MSTCs and APOE⁺/CD163⁺ macrophage populations in MBC. Together with spatial transcriptomic and mIHC evidence, these findings support the presence of macrophage-enriched immunosuppressive niches around MSTC-enriched regions, although functional interactions remain to be experimentally validated.

Although this study provides a comprehensive single-cell and spatial transcriptomic resource for MBC, several limitations should be noted. First, due to the scarcity of MBC samples, our cohort was predominantly Asian and all collected MBC specimens were histologically confirmed as invasive ductal carcinoma, which limits generalization to other populations and histological subtypes. In addition, BRCA2 and other key germline or somatic variables were not analyzed because this retrospective, multicenter study lacked uniform germline testing and matched tumor DNA sequencing. To avoid bias, we did not impute BRCA2 status from expression surrogates or scRNA-derived copy-number proxies; future cohorts should incorporate harmonized WES/WGS or targeted DNA panels with matched normal for integrated modeling. Second, although we identified putative pathways such as GRN-SORT1, functional validation still requires dedicated perturbation experiments. Third, direct comparison of cell-type proportions across scRNA-seq and spatial platforms should be interpreted cautiously as these datasets were generated from different individuals rather than matched tumor samples. Therefore, the observed composition differences may reflect both platform-specific effects and interpatient or regional heterogeneity. We accordingly used the spatial datasets primarily to validate spatial localization and tissue organization, rather than as evidence for absolute cross-platform differences in cell-type abundance. Fourth, although SeekSpace and Visium HD enabled spatial mapping of major cell lineages, we did not perform a comprehensive cross-platform analysis of all immune cell subsets relative to MSTCs. This limitation reflects platform differences in immune cell capture and annotation, as well as the use of nonmatched tissue sections from different individuals. These factors limit robust comparison of sparse immune populations such as B cells and some T cell subsets. We therefore focused on reproducible spatial patterns, including MSTC-rich regions, APOE⁺/CD163⁺ macrophage proximity, immune/stromal gene gradients and Xenium validation. Larger matched spatial cohorts will be required to fully characterize T cell, B cell, and myeloid distributions relative to MSTCs. Fifth, given the current maturity of spatial transcriptomics, we could not perform deeper functional analysis of the spatial organization of the MBC microenvironment. Last, cell segmentation in Visium HD remains challenging: The highly irregular morphology of tumor and microenvironmental cells increases boundary uncertainty and may affect analysis and interpretation. Although we used 10x Genomics Xenium In Situ and SeekSpace nuclei sequencing for cross-validation, the limited gene panel of Xenium and the lower immune cell capture of SeekSpace constrain a comprehensive view of the MBC TIME. Continued advances in spatial technologies and computation should enable deeper interrogation and broader validation.

In summary, this study provides a comprehensive single-cell and spatial transcriptomic resource for MBC and identifies an MBC-associated malignant cell state with distinct neural-lipid, genomic stress, and immune-metabolic spatial features. These findings offer a hypothesis-generating framework for future mechanistic and therapeutic studies aimed at improving the biological understanding and clinical management of MBC. However, because this study is based on observational single-cell and spatial transcriptomic data without functional perturbation experiments, the mechanistic relevance and causal direction of the associations identified here remain to be established.

MATERIALS AND METHODS

Patient sample collection

This study was reviewed and approved by the Medical Ethics Committee of Affiliated Hospital of Qingdao University (ethics code: QYFY-WZLL-28676). The study was conducted in accordance with the principles of the Declaration of Helsinki and the International Conference on Harmonization Guidelines for Good Clinical Practice. All patients whose human specimens were collected provided written informed consent. All authors approved the manuscript for submission for publication and vouch for the accuracy and completeness of the data reported. From January 2021 to December 2024, we collected surgically resected primary breast cancer tissues from MBC and FBC patients treated in hospitals across China for single-cell sequencing (10x Genomics), SeekSpace spatial transcriptome sequencing, and Visium HD spatial transcriptome sequencing, totaling 28 cases. Paraffin-embedded tissue samples from 2008 to 2019 were collected for mIHC staining and Xenium In Situ analysis.

10x Chromium scRNA-seq: Tissue processing, library preparation, and sequencing

Fresh surgical tissues were transported on ice in sterile dishes containing 10 ml of 1× Dulbecco’s phosphate-buffered saline (DPBS; Thermo Fisher Scientific, catalog no. 14190144) to remove residual storage solution. Tissues were minced on ice and dissociated at 37°C for ∼40 min at 50 rpm in 0.25% trypsin (Thermo Fisher Scientific, catalog no. 25200-072) and deoxyribonuclease I (10 μg/ml; Sigma-Aldrich, catalog no. 11284932001) prepared in phosphate-buffered saline with 5% fetal bovine serum (FBS; Thermo Fisher Scientific, catalog no. SV30087.02). Cell suspensions were collected every 20 min, filtered through 40-μm strainers, subjected to RBC lysis (Thermo Fisher Scientific, catalog no. 00-4333-57), and washed with 1× DPBS containing 2% FBS. Cell viability was assessed by 0.4% Trypan blue staining on a Countess II Automated Cell Counter (Thermo Fisher Scientific).

Single cells were loaded onto a Chromium Controller to generate Gel Bead-in-Emulsions. Following lysis, poly(A) and RNA hybridized to barcoded gel beads. Reverse transcription appended cell barcodes and unique molecular identifiers (UMIs) to cDNA, followed by second-strand synthesis, adaptor ligation, and amplification. Libraries were prepared according to the manufacturer’s protocol (10x Genomics, CG000206 Rev. D), quality checked using a Bioanalyzer High Sensitivity DNA kit (Agilent) and Qubit High Sensitivity DNA assay (Thermo Fisher Scientific), and sequenced on an Illumina NovaSeq 6000 using paired-end chemistry (2 × 150).

10x Genomics Visium HD FFPE spatial transcriptomics: Library preparation and sequencing

Formalin-fixed, paraffin-embedded (FFPE) samples passing RNA QC (DV200 > 30%) were sectioned at 5 μm onto Visium Gene Expression slides (10x Genomics), baked at 42°C for 3 hours, and dried overnight. Slides were deparaffinized at 60°C, immersed in xylene, rehydrated through an ethanol gradient, and H&E stained (Mayer’s hematoxylin; Dako bluing reagent; alcoholic eosin). After whole-slide imaging, probe hybridization, probe release, and library construction were performed using the Visium HD Spatial Gene Expression Reagent Kit per User Guide CG000685 (Human, 6.5 mm). Libraries were sequenced in paired-end 150–base pair mode on Illumina platforms.

SeekSpace single-cell spatial transcriptomics: Library preparation and sequencing

Fresh-frozen tissues were cryosectioned at 10 to 20 μm on a cryostat (Leica, −20°C). Regions of interest were placed onto SeekSpace Chips without folds and briefly warmed to ensure adhesion. Single-nucleus suspensions with spatial barcodes were prepared using the SeekSpace Single Cell Spatial Transcriptome-seq Kit (K02501-08). Nuclei were counted with AO/PI on a Countstar Rigel S2. Nuclei were divided into eight polymerase chain reaction (PCR) tubes for reverse transcription (600 to 30,000 nuclei per tube) using distinct barcoded primers with 15 annealing cycles (8°C → 42°C). After washing, nuclei were pooled and mixed with ligation reagents and then loaded with barcoded hydrogel beads and partitioning oil into SeekOne DD Chip S3. Emulsion droplets were generated on the SeekOne Digital Droplet System and incubated for 60 min at 20°C followed by 10 min at 65°C to obtain barcoded cDNA and spatial barcodes. Droplets were decrosslinked, and a second reverse transcription and PCR preamplification were performed. The preamplified product was used to construct both spatial barcode and cDNA libraries with sample indexing. Libraries were purified with VAHTS DNA Clean Beads (Vazyme N411-01), QC-checked by Qubit (Thermo Fisher Scientific, Q33226), and Bio-Fragment Analyzer (Bioptic Qsep400) and sequenced on an Illumina NovaSeq X Plus (PE150).

10x Genomics Xenium In Situ transcriptomics: Assay and imaging

FFPE sections (5 μm) were processed following 10x Genomics User Guides CG000578 and CG000580. Xenium Prime 5K Human Pan Tissue and Pathways Assay (PN-1000671) was performed according to CG000760: Sections were hybridized with probe sets at 50°C for 16 to 24 hours, washed, ligated, and amplified. Cell segmentation staining used the Xenium Cell Segmentation Staining Reagents (PN-1000661). After autofluorescence quenching and 4′,6-diamidino-2-phenylindole (DAPI) staining, sections were imaged and decoded on a Xenium Analyzer (PN-1000481). Postrun H&E staining followed CG000613.

Multiplex immunohistochemistry

mIHC was performed on 4-μm-thick FFPE whole-tissue sections using sequential primary-antibody staining paired with the TSA 7-color kit (abs50037, Absinbio, Shanghai), followed by DAPI staining. For example, deparaffinized slides were incubated with anti-panCK antibody (ab7753, Abcam), for 30 min and then treated with anti-rabbit/mouse horseradish peroxidase–conjugated secondary antibody (abs50015-02, Absinbio, Shanghai) for 10 min. Labeling was then developed for 10 min using TSA 520 according to the manufacturer’s instructions. Slides were washed in TBST buffer and then transferred to preheated citrate solution (90°C) before being heat treated using a microwave set at 20% of maximum power for 15 min. Slides were cooled in the same solution to room temperature. Between all steps, the slides were washed with tris buffer. The same process was repeated for the following antibodies/fluorescent dyes, in order: anti-GRIA2 (ZENBIO, R22839; 1:100; TSA 620), anti-panCK (Abcam, ab7753; 1:200; TSA 780), anti-NEGR1 (ZENBIO, 163125; 1:100; TSA 570), anti-APOE (Proteintech, 66830-1; 1:100; TSA 480), anti-GFRA1 (Affbiotech, DF7309; 1:250; TSA 690), and anti-CD163 (Abcam, ab182422; 1:100; TSA 520). Each slide was then treated with two drops of DAPI (abs47047616, Absinbio, Shanghai), washed in distilled water, and manually coverslipped. Slides were air dried and imaged with Pannoramic MIDI II (3DHISTECH). Images were analyzed using Indica Halo software.

scRNA-seq data preprocessing

The raw data were analyzed using the Cell Ranger pipeline (version 7.1.0), with gene-count data generated using the default and recommended parameters. Align the FASTQ output obtained from sequencing data with the GRCh38 reference genome using the STAR (59) algorithm. We used Seurat (version 5.0.1) R package to do all subsequent analyses. To eliminate the influence of low-quality cells such as empty droplets and multiplets, cells with expressed genes <200 or >6000 were excluded. The percentage of UMIs mapped to mitochondria was set to less than 25%.

We use in Seurat:: NormalizeData, Seurat:: FindVariableFeatures and Seurat:: ScaleData functions to normalize all data. Seurat:: IntegrateLayers function was used to remove technical variability (method = RPCAIntegration). The top 50 principal components (PCs) were calculated using the Seurat::RunUMAP function. A UMAP dimensional reduction was performed on the scaled matrix using the top 50 PC analysis components to obtain a two-dimensional representation.

Spatial data processing

For 10x Visium HD spatial transcriptomics data, nuclei were segmented on maximum projection DAPI-stained images using a fine-tuned Cellpose model. Baysor was used to predict cell boundary segmentation based on transcript identity and location, using prior Cellpose nuclear segmentation for 10x Visium HD. The prior-segmentation-confidence parameter was set to 0.95, with min-molecules-per-cell set to the median nuclear transcript count. Baysor segmentations lacking nuclei were filtered out, whereas those containing multiple nuclei were split by assigning transcripts to the nearest nuclear centroid within segmentation boundaries. Cell boundaries were visualized as polygons using the alphashape Python package. Before downstream processing, cells with nuclear transcript count n < 8, total transcript count n < 20, or total transcript count n > 800 were filtered out.

Cell cluster and annotation

We conducted an initial clustering analysis on the integrated data using Seurat::FindNeighbors and Seurat::FindClusters (resolution of 0.8). The cell clusters obtained from the unsupervised clustering were manually annotated on the basis of well-known marker genes. These marker genes include T cells (CD3D, CD3E, and CD3G), B cells (CD19, CD79A, and CD79B), plasma cells (IGHM, CD38, and PRDM1), mast cells (MS4A2, TPSB2, and TPSAB1), myeloid cells (CD68, LYZ, CD14, FCGR3A, S100A8, and S100A9), endothelial cells (ENG, KDR, and PECAM1), fibroblasts (COL1A1, COL1A2, FN1, and DCN), and epithelial cells (EPCAM, KRT8, KRT18, KRT19, and MUC1). Visualization was performed using the Seurat::DotPlot function, and cell identities were annotated on the basis of the expression levels and proportions of marker genes within the clusters.

Next, we performed a second round of clustering to further characterize the subtypes of major cell types within the tumor microenvironment. To mitigate potential data heterogeneity caused by cell types, cell numbers, or intersample variability, we first filtered out features with an expression proportion below 50 within the cell type in the second round of clustering and used the Harmony method for batch correction. In the second clustering, we used 50 PCs for dimensionality reduction and unsupervised clustering (resolution of 1).

To enhance annotation accuracy, we used an automated large language model (LLM)–based consensus annotation approach. We first identified DEGs for each cluster using Seurat::FindAllMarkers with parameters: only.pos = TRUE, min.pct = 0.25, and logfc.threshold = 0.5. The DEGs were then processed through an interactive consensus annotation system that leveraged multiple AI models (Claude-3.5-Sonnet, Gemini-2.0-Flash, Qwen-Max, and GPT-4o) to provide tissue-specific cell-type annotations for MBC. The system analyzed the top 15 marker genes per cluster across a maximum of five discussion rounds with a controversy threshold of 0.8 for consensus determination. Because mLLMCelltype only broadly classified macrophages as M1 and M2 subtypes, we performed additional manual annotation of macrophage subpopulations based on highly expressed genes specific to each macrophage cluster.

Spatial deconvolution (RCTD)

According to the scRNA-seq dataset, spacexr (version 2.2.0) was applied to infer the cell-type composition of each bin. Specifically, the default parameters were used in the creat.RCTD function, except for minimum number of cell of >1 for each cell types and minimum of UMI counts of >1 per pixel, and mode parameter was set to “doublet” in the run.RCTD function.

Cell-cell communication analysis

Cell-cell communication analysis was conducted with the scRNA-seq data by using the CellChat software (60) (version 2.1.2). Specifically, only receptors and ligands expressed in >1% of cells were further evaluated, while a cell-cell communication was considered nonexistent if the ligand or the receptor was unmeasurable.

Pathway enrichment analyses

GO enrichment analysis (61): To identify enriched GO biological processes based on DEGs, analysis was performed on DEGs from each cell type using clusterProfiler::gseGO (version 4.8.1) R package. Using org.Hs.eg.db annotation package (version 3.17.0). GSEA: We used GseaVis::gseaNb (version 0.0.8) R package (https://github.com/junjunlab/). For input, we used preranked gene lists by avg_log₂FC calculated from Seurat::FindAllMarkers. Kyoto Encyclopedia of Genes and Genomes (62): We used ClusterGVis::enrichCluster (version 0.1.0) R package. Using org.Hs.eg.db annotation package (version 3.17.0): The ssGSEA analysis was conducted using the R singleseqgset package (version 0.1.2.9000). We used the human hallmark gene sets from the msigdbr package (63) (version 7.5.1) as the gene sets of interest for enrichment analysis. Differences were considered statistically significant at P < 0.05.

Developmental trajectory

We inferred the developmental trajectory of human tumor cells using CytoTRACE (version 0.3.3) and Monocle (version 2.12). Tumor cells from male and female scRNA-seq samples were pooled. To reduce computational burden while preserving the original cell-subtype composition, cells were stratified by the annotated cell subtype and proportionally downsampled to ∼50,000 cells. The number of cells sampled from each subtype was calculated according to its proportion in the pooled tumor cell dataset. CytoTRACE is based on the concept of transcriptional diversity (i.e., the number of genes expressed in cells decreases during differentiation). A log₂-normalized expression matrix was used. For Monocle 2, we constructed a CellDataSet object from the cluster-annotated Seurat object using the newCellDataSet function. We derived DEGs from each cluster using the DifferentialGeneTest function and ordered cells in pseudotime using genes with q < 1 × 10⁻⁵. Dimensionality reduction was performed using the DDRTree algorithm, followed by ordering cells along the trajectory. CytoTRACE2 scores were calculated from the same downsampled tumor cell object using raw counts. The CytoTRACE2-relative score was used as an inferred measure of developmental potential within the analyzed tumor cell population, with lower scores indicating lower inferred developmental potential. CytoTRACE2-relative scores were then projected onto the Monocle trajectory.

CNV inference

To identify malignant and nonmalignant cells, we used two complementary methods to confidently distinguish between malignant and nonmalignant cells in each sample. To more accurately validate the identified cancer cells, we also assessed CNV levels using the R package inferCNV, with immune cells (T cells, B cells, macrophages, and mast cells) and stromal cells (fibroblasts and endothelial cells) as control groups and epithelial cells as the test group. To quantify transcriptome-inferred CNV burden, we calculated an inferCNV score for each malignant cell from the inferCNV output matrix. The score was defined as the squared sum of inferCNV values across chromosomal segments. For each sample and malignant cell group, the median inferCNV score was calculated and used for sample-level comparison. P values were calculated using two-sided t tests.

TCGA data and survival analysis

We used the R package easyTCGA (64) (version 0.0.4.200.0) to download the TCGA pan-cancer gene expression matrix and sample clinical information. Kaplan-Meier survival curves were plotted using ggsurvplot (version 0.4.9) function in the R package Survminer (version 0.4.9).

Functional module scoring

Feature scoring in scRNA-seq, SeekSpace spatial transcriptomics, and Visium HD spatial transcriptomics was calculated using the AddModuleScore function in Seurat, with the following gene sets: neurotrophic meta module (GRIA2, NEGR1, GFRA1,ROBO1,ROBO2, SHANK2, and CHGB), angiogenesis gene set (msigdb, TUMOR_ANGIOGENESIS), fatty acid gene set (msigdb, FATTY_ACID_METABOLISM), and glutamate signal gene set (msigdb, GLUTAMATE_RECEPTOR_ACTIVITY).

Regulon network

The regulon network was explored using the R package SCENIC (version 1.1.3), which analyzed the coexpression of transcriptional factors and their putative target genes. We built and scored gene regulatory network using the default parameters. Raw count matrix was used to build coexpression network using the runCorrelation, and runGENIE3 functions and active gene networks were identified by AUCell. Regulon activity for each cell was calculated as the average normalized expression of putative target genes. To identify differentially active transcription factors between MSTCs and other cell types, we extracted regulonAUC data from SCENIC results and calculated mean activity scores for each group. Statistical comparison was performed to identify transcription factors with significantly higher activity in MSTCs using fold change and difference thresholds (fold_change > 75th percentile, difference > 75th percentile, and mean activity > 0.005). The CSI was calculated to assess transcriptional regulatory relationships using a custom R function. CSI measures the specificity of connections between regulons based on correlation matrices, where CSI values range from 0 to 1, with higher values indicating more specific connections. We applied a CSI threshold of 0.5 and performed hierarchical clustering using Ward.D2 method to identify regulatory modules. The resulting CSI matrix and module assignments were visualized using ComplexHeatmap with custom color schemes and annotations.

NMF analysis

To investigate expression programs underlying tumor heterogeneity in malignant cells, we performed NMF analysis using the cNMF Python package. Patients with at least 150 malignant tumor cells were selected for analysis. For each tumor, UMI matrices were converted to counts per million by normalizing each gene by the total UMI count per cell, followed by log₂ transformation. Data were then mean centered by subtracting the average expression of each gene across all cells. We filtered genes expressed in at least 2% of cells, set all negative values to zero, and performed NMF analysis with factor numbers (K) ranging from 5 to 12, using 1000 iterations and a threshold of 0.4. Expression programs were initially defined as the top 150 genes per K (ranked by NMF scores). Factors appearing in fewer than 75 cells were excluded. Robust NMF programs were hierarchically clustered based on Jaccard similarity using Euclidean (1 − Jaccard similarity) distance as the distance metric. MPs were defined on the basis of representative genes and their enriched biological functions.

Spatial distance analysis for neurotrophic signal distribution

To quantify the spatial relationship between cells and neurotrophic signal centers, we developed a distance-based analysis approach. First, pathway scores for neurotrophic, angiogenesis, glycolysis, and fatty acid metabolism were calculated, and outliers were capped at the 95th percentile to reduce the influence of extreme values. Spatial coordinates from the tissue sections were extracted and normalized to a range of 0 to 1 using min-max normalization to standardize measurements across different tissue dimensions. Cell spatial distribution was analyzed using K-means clustering to identify density centers representing regions of high cell concentration. The optimal number of clusters was determined using the elbow method, evaluating within-cluster sum of squares for k values ranging from 1 to 10. Based on this analysis, k = 3 was selected as the optimal cluster number. K-means clustering was performed with 25 random initializations to ensure robust cluster identification. For each cell $i$, the minimum Euclidean distance to the K centers $d_i$ was computed and normalized as

$$d_i^{\prime }=\frac{d_i-d_{\text{min}}}{d_{\text{max}}-d_{\text{min}}},d_i^{\prime }\in [0,1]$$

Cells were sorted by $d_i^{\prime }$ and binned in groups of 100 cells; within each bin, medians of $d_i^{\prime }$ and pathway scores were computed. Pearson correlations quantified associations between spatial proximity and pathway activity. Pearson correlation coefficients were computed to assess the relationship between spatial proximity to cancer cell centers and various pathway activities, including neurotrophic, angiogenesis, glycolysis, and fatty acid metabolism scores.

Spatial proximity and gene expression gradient analysis

To investigate the spatial relationship between different cell types and MSTCs, we calculated the distance from each cell to its nearest reference cells. For each non-MSTC, we computed the median distance to the top-k nearest MSTCs (k = 3) using Euclidean distance in the spatial coordinate system. The analysis was performed using custom Python functions that used scipy.spatial.distance.cdist for efficient distance matrix calculation. We performed gene expression gradient analysis to identify genes that showed correlation with spatial distance to MSTCs. Using a modified scVelo heatmap approach, we calculated Pearson correlation coefficients between each gene’s expression and the distance to MSTCs across all immune or stromal cells. Genes were filtered to include only those expressed in more than 5% of cells. The heatmap visualization used convolution smoothing with user-defined bins (n_bins = 10) to reveal expression trends along the spatial gradient.

Statistics

Two-sided tests were used. Nonparametric comparisons used Wilcoxon rank sum unless stated. Correlations used Pearson unless normality failed, in which case Spearman was reported. Multiple testing was adjusted by Benjamini-Hochberg false discovery rate. Exact tests, n, and effect sizes appear in the figure legends or the supplementary tables.

Supplementary Materials

The PDF file includes:

Figs. S1 to S7

Legends for tables S1 to S10

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S10