Integrated Spatial and Bulk Transcriptomic Analysis Identifies a Multicellular Ecosystem Involved in Progression of Young Breast Cancer

ABSTRACT

Young breast cancer (YBC), defined as diagnosis at or before the age of 40 years, is characterized by an aggressive phenotype and poor prognosis. The tumor microenvironment (TME), comprising immune and stromal cells, plays a critical role in tumor initiation and progression and is closely associated with clinical outcomes. However, alterations in cell states and the organization of multicellular ecosystems within the YBC microenvironment remain poorly understood. We analyzed 7379 bulk transcriptome samples and spatial transcriptomic data from eight samples using EcoTyper and complementary spatial analyses. The abundance of CD4+ T cells (S07), fibroblasts (S05), monocytes/macrophages (S09), and epithelial cells (S05) was significantly lower in young than in older patients with breast cancer. CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) formed a spatially organized multicellular ecosystem with putative interactions involving CDH1‐PTPRF/ERBB3. Lower ecosystem abundance was associated with poorer prognosis, including among patients who did not achieve a pathological complete response after neoadjuvant therapy. These findings identify a distinct multicellular ecosystem within the YBC microenvironment, providing insights into the microenvironmental features associated with disease progression.

Abbreviations

BSA

bovine serum albumin

CAF

cancer‐associated fibroblast

CCR8

C–C motif chemokine receptor 8

ce

multicellular ecosystem

CI

confidence interval

ctDNA

circulating tumor DNA

CTLA‐4

cytotoxic T‐lymphocyte‐associated protein 4

DAPI

4′,6‐Diamidino‐2‐phenylindole

DFI

disease free interval

DRFS

distant relapse free survival

EMT

epithelial mesenchymal transition

FAP

fibroblast activation protein

FDR

false discovery rate

FFPE

formalin‐fixed paraffin‐embedded

FPKM

fragments per kilobase of transcript per Million mapped reads

GDC

genomic data commons

GO

gene ontology

GSVA

gene set variation analysis

GTEx

genotype tissue expression

HR

hazard ratio

HR+/HER2−

hormone receptor positive/human epidermal growth factor receptor 2 negative

HRP

Horseradish peroxidase

IHC

immunohistochemistry

KEGG

Kyoto Encyclopedia of Genes and Genomes

MDSC

myeloid‐derived suppressor cell

METABRIC

Molecular Taxonomy of Breast Cancer International Consortium

mIF

multiplex immunofluorescence

NK

natural killer

NMF

non‐negative matrix factorization

OS

overall survival

PCA

principal component analysis

pCR

pathological complete response

PD‐1

programmed cell death 1

PDGFR

platelet‐derived growth factor receptor

PD‐L1

programmed cell death ligand 1

RFS

relapse free survival

RNA‐seq

RNA sequencing

RT

room temperature

SASP

senescence‐associated secretory phenotype

SCAN‐B

Sweden Cancerome Analysis Network‐Breast

ST

spatial transcriptomics

TAM

tumor‐associated macrophage

TCGA

The Cancer Genome Atlas

TLS

tertiary lymphoid structure

TME

tumor microenvironment

TNM

tumor‐node‐metastasis

TPM

transcripts Per Million

Treg

regulatory T cell

TSA

tyramide signal amplification

UCEC

uterine corpus endometrial carcinoma

YBC

young breast cancer

1. Introduction

Young breast cancer (YBC), typically defined as breast cancer diagnosed at or before the age of 40 years [1], represents a clinically distinct subgroup characterized by more aggressive phenotypes and poorer survival than in older patients [2, 3, 4, 5, 6]. At the transcriptomic level, YBC has been associated with increased expression of proliferation and stem cell related signatures [7, 8, 9]. Despite the increasing use of intensive and prolonged systemic treatments, significant improvements in overall survival have not been consistently observed [10], especially in luminal subtypes [11]. This discrepancy suggests that the aggressive clinical behavior of YBC may involve biological mechanisms that current treatments do not adequately address. Although tumor‐intrinsic genomic alterations contribute to cancer progression [12, 13], the overall landscapes of somatic mutations and copy number alterations have been reported to differ little between young and older patients [14]. Another important factor influencing the germination of “seeds” (cancer) is the “soil” (microenvironment). The aging microenvironment is characterized by the accumulation of fibroblasts with a senescence‐associated secretory phenotype (SASP) and increased infiltration of myeloid‐derived suppressor cells (MDSCs) and regulatory T (Treg) cells [15, 16, 17], which may create a tumor‐permissive milieu [18]. Characterizing changes in abundance, spatial colocalization and interactions of cell type‐specific cell states within the tumor microenvironment (TME) is therefore important for understanding YBC progression.

Recent studies have provided preliminary insights into the cellular landscape of the YBC microenvironment but have largely focused on immune cell types [19, 20], reporting relative enrichment of M0 and M1 over M2 macrophages in subsets of YBC [21]. YBC has also been associated with inflammatory and immunosuppressive features, including upregulation of interferon‐stimulated genes, increased expression of programmed cell death ligand 1 (PD‐L1) and cytotoxic T‐lymphocyte‐associated protein 4 (CTLA‐4), and enrichment of Treg cells [22, 23]. Advances in single‐cell RNA sequencing have enabled detailed characterization of cell states and the multicellular ecosystem within the TME [24, 25, 26]. However, the widespread use of formalin‐fixed and paraffin‐embedded specimens and limited sample sizes of many single‐cell studies can constrain analyses of clinical outcomes. Increasing evidence indicates that cells interact within the multicellular ecosystem (CE) in the microenvironment [27, 28, 29]. These ecosystems, exemplified by tertiary lymphoid structure (TLS) and oncofetal ecosystem, may have greater effects on tumor initiation and progression than isolated cell types [30, 31, 32]. Nevertheless, the presence and clinical relevance of such multicellular ecosystems in the YBC microenvironment remain unclear.

Here, we leveraged EcoTyper, a machine learning framework [33], to characterize cell states and multicellular ecosystem in the YBC microenvironment. By integrating large‐scale bulk and spatial transcriptomic data, we aimed to delineate YBC‐associated cellular alterations and multicellular organization, identify biological features related to aggressive disease and poor prognosis, and provide a basis for investigating potential therapeutic targets (Figure 1A).

FIGURE 1.

Changes in cell state abundance in young breast cancer. (A) Overview of the study design. The icons were obtained from Servier Medical Art and used under the Creative Commons Attribution 4.0 Unported License (https://creativecommons.org/licenses/by/4.0/). (B) Heatmap of differences in cell state abundance between young (≤ 40 years) and older (> 40 years) individuals in breast cancer and normal breast tissue datasets. ***p < 0.001; **p < 0.01; *p < 0.05. (C) Correlations between natural age and the abundances of CD4+ T cells (S07), epithelial cells (S05), fibroblasts (S05), and monocytes/macrophages (S09) in tumor tissues. (D) Differences in the abundance of CD4+ T cells (S07), epithelial cells (S05), fibroblasts (S05), and monocytes/macrophages (S09) between young (≤ 40 years) and older (> 40 years) in TCGA pan‐cancer dataset. ***p < 0.001; **p < 0.01; *p < 0.05. (E) Kaplan–Meier survival curves for patients with high and low abundances of CD4+ T cells (S07), epithelial cells (S05), fibroblasts (S05), and monocytes/macrophages (S09) in TCGA‐BRCA, METABRIC, and SCAN‐B datasets. TCGA‐BRCA, The Cancer Genome Atlas breast cancer cohort; METABRIC, Molecular Taxonomy of Breast Cancer International Consortium; SCAN‐B, Sweden Cancerome Analysis Network‐Breast.

2. Methods

2.1. Data Acquisition and Deconvolution

Transcriptomic profiles and corresponding clinical data for The Cancer Genome Atlas (TCGA) breast cancer cohort (1091 tumors and 113 adjacent normal tissues) and pan‐cancer datasets were obtained from UCSC Xena. Data from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC; n = 1904), Genotype‐Tissue Expression (GTEx; n = 204), and Sweden Cancerome Analysis Network‐Breast (SCAN‐B; n = 3409) were retrieved from cBioPortal, the GTEx portal, and GEO (GSE96058), respectively. RNA‐seq datasets (TCGA, GTEx and SCAN‐B) were analyzed in transcripts per million (TPM) or fragments per kilobase of transcript per million mapped reads (FPKM) format, and microarray datasets (METABRIC, GSE25066 and GSE22226) were analyzed as normalized expression intensities in non‐logarithmic space, as required by EcoTyper. The GSE25066 and GSE22226 were used to evaluate outcomes after neoadjuvant chemotherapy. Somatic mutation data for TCGA‐BRCA and METABRIC were obtained from GDC‐TCGA and cBioPortal, respectively [34]. In all datasets, YBC was defined as breast cancer diagnosed at or before the age of 40 years (≤ 40 years).

2.2. Identification of Cell States

We used EcoTyper, an established machine learning framework, to identify cell type‐specific transcriptional states across 12 cell types [33], including B cells, CD4+ T cells, CD8+ T cells, dendritic cells, endothelial cells, epithelial cells, fibroblasts, mast cells, monocytes/macrophages, natural killer (NK) cells, plasma cells, and neutrophils. Downloaded bulk RNA sequencing (RNA‐seq) datasets were input into EcoTyper in TPM or FPKM format. Cell states within the TME were recovered using the pretrained non‐negative matrix factorization (NMF) model following EcoTyper recovery workflow.

2.3. Analysis of Spatial Transcriptomic Data

2.3.1. Integration of Spatial Transcriptomic Data

Spatial transcriptomic (ST) datasets were obtained from 10x Genomics (https://www.10xgenomics.com/), following the preprocessing procedures provided with the datasets. The data were analyzed using the Seurat package. Raw counts were normalized using the SCTransform function, and the datasets were integrated for subsequent analyses.

2.3.2. Cellular Colocalization Analysis

We selected the top 10 marker genes for each of four cell states: CD4+ T cells (S07) (LRRC69, OPRM1, CCDC65, CACNG8, TMC2, PLGLB1, MYLK3, SLC22A9, ANKRD36C, and DSC2), epithelial cells (S05) (AGR2, SLC44A4, AGR3, RAB11FIP1, TSPAN8, GOLM1, BCAS1, ERBB3, LLGL2, and PLS1), fibroblasts (S05) (PPP1R32, ST6GALNAC2, RSPH9, ARHGAP39, LRRC8B, IFT122, TTC32, CLASP1, ZNF273, and DPY19L2), and monocytes/macrophages (S09) (DLEC1, LMNTD1, RIBC1, CACNB1, C2orf80, C4orf47, AVIL, SERPINF2, DHX40, and MIR194‐2HG). These marker sets were used to estimate the cell state enrichment by AddModuleScore function with default parameter [33]. We assessed spatial colocalization among CD4+ T cells (S07), epithelial cells (S05), fibroblasts (S05), and monocytes/macrophages (S09) using the Jaccard similarity index, calculated from presence‐absence profiles of each cell state across barcoded spots within each tissue section. The cell states were hierarchically clustered, and the Jaccard coefficients were visualized using pheatmap. Spatial distributions were visualized using SpatialFeaturePlot.

2.3.3. Analysis of Putative Ligand–Receptor Interactions

To identify putative ligand–receptor interactions, we utilized the SpaGene package to infer them from a predefined list of ligand–receptor pairs [35, 36]. The SpaGene_LR function was used to assess spatial colocalization and obtain false discovery rate (FDR) values. Pairs with FDR < 0.05 were considered to show significant spatial colocalization, and those consistently significant across all slices were selected for further analysis. A spatial CE enrichment score was calculated from the marker genes of CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) using the AddModuleScore function, and the ligand–receptor pairs abundance was derived similarly.

We used MISTy, a multiview machine learning framework, to assess the relationships between ligand–receptor pair scores and CE enrichment by estimating feature importance in different spatial context [37]. Then, we defined three spatial milieus: (1) an intrinsic view within a single spot; (2) a juxta view with a radius of two spots representing the local neighborhood; (3) a para view with a radius of five spots representing the broader neighborhood. Each dataset was analyzed independently, followed by analysis of the combined datasets.

2.4. Multiplex Immunofluorescence Staining

Co‐expression of CD4 and OPRM1, CD68 and DLEC1, and PDGFR (α + β) and ZDHHC1 was used to identify CD4+ T cells (S07), monocytes/macrophages (S09), and fibroblasts (S05), respectively. Multiplex immunofluorescence (mIF) staining was performed on formalin‐fixed paraffin‐embedded (FFPE) tissue sections using a PANO 7‐plex IHC kit (Panovue, #10004100100). Sections were cut at a thickness of 4 μm, deparaffinized in xylene and rehydrated through a graded ethanol series. Heat‐induced antigen retrieval was performed in Tris‐EDTA buffer (pH 9.0, Proteintech) using a microwave oven. Sections were blocked with a 2.5% bovine serum albumin (BSA) buffer for 15 min at room temperature (RT). Multiplex staining was performed through sequential rounds of immunostaining. Each cycle used a primary antibody against one of the following targets: OPRM1 (Abclonal, #A16939), DLEC1 (Proteintech, #20027‐1‐AP), CD4 (Starter, #SDT‐R163), ZDHHC1 (Abmart, #PH7848), PDGFR (α + β) (Starter, #S‐393‐11), or CD68 (Proteintech, #66231‐2‐Ig). Primary antibody incubation was followed by incubation with a horseradish peroxidase (HRP)‐conjugated anti‐rabbit/mouse secondary antibody (Panovue). Signals were detected by tyramide signal amplification (TSA) using Opal fluorophores PPD 480, PPD 520, PPD 570, PPD 620, PPD 690, and PPD 780 (Panovue). After each TSA reaction, the antibody‐fluorophore complex was stripped by microwave‐mediated antigen retrieval to allow subsequent staining cycles. The staining sequence and fluorophore assignments were optimized to minimize spectral interference. After all staining cycles, nuclei were counterstained with DAPI (Panovue), and slides were mounted with antifade mounting medium (Panovue). Images were acquired using a KFBIO Digital Pathological Slide Scanner (KF‐FL‐400). Marker expression and spatial distribution were quantified in each field of view using HALO software (Indica Labs).

2.5. Construction of Multicellular Ecosystem Score

To derive a CE score from bulk transcriptomic data, we performed principal component analysis (PCA) on the abundance of CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09), using the prcomp function. We assessed the correlations between the first principal component (PC1) score and the abundances of the three constituent cell states. The PC1 score was used to represent CE abundance, with its sign reversed when these correlations were negative.

2.6. Enrichment Analysis

Gene expression matrices from the TCGA‐BRCA, METABRIC, and SCAN‐B cohorts were used to assess the correlations between gene expression and CE abundance. Genes with a correlation coefficient < 0 and an adjusted p‐value < 0.05 were selected for downstream analyses. Their potential biological functions were investigated using the Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome pathway, and gene ontology (GO) resources [38, 39, 40]. Pathway activity was assessed using GSVA with the Hallmark gene sets. Differences in pathway activity between CE groups were evaluated using limma. Pathways with p < 0.05 were considered significantly enriched.

2.7. Survival Analysis

The prognostic value of the CE score was evaluated using Kaplan–Meier survival curves and univariate and multivariate Cox proportional hazards regression models. Survival endpoints were defined as in the original studies. Optimal cutoff values were determined using surv_cutpoint to classify patients into CE‐high and CE‐low status. Survival distributions were estimated using the Kaplan–Meier method and compared using the log‐rank test.

2.8. Statistical Analysis

All analyses were performed in R software version 4.3.0. Continuous variables were compared using Student's t‐test or the Mann–Whitney U test, and categorical variables were compared using the chi‐square test or Fisher's exact test. Correlations were assessed using Pearson's correlation or Spearman's rank correlation coefficient. Survival distributions were estimated using the Kaplan–Meier method and compared using log‐rank tests. Cox proportional hazards regression models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs). Multiple testing correction was performed using the Benjamini–Hochberg method. p < 0.05 was considered statistically significant. The specific statistical methods are described in the corresponding sections.

3. Results

3.1. Identification of Cell States Specific to YBC

We first applied EcoTyper to characterize the breast tissue microenvironment using five bulk transcriptomic data sources initially, including tumor tissues from TCGA, METABRIC, and SCAN‐B, and normal tissues from TCGA and GTEx (Table S1; Figure S1). We identified 69 cell states across 12 cell types (Figure 1B). CD4+ T cells (S07), fibroblasts (S05), monocytes/macrophages (S09), and epithelial cells (S05) were significantly less abundant in YBC (≤ 40 years) than in old breast cancer (OBC) patients (> 40 years). This pattern was not observed in tumor‐adjacent or normal breast tissues from young (≤ 40 years) and older (> 40 years) individuals (Figure 1B,C; Figure S2; Table S2). This pattern of age‐related reductions in abundance across all four cell states was not observed in any of the other solid tumors examined (Figure 1D), suggesting that the changes might be specific to YBC. These differences in cell state abundance were also not associated with BRCA1 or BRCA2 mutations (Figure S3). We next examined associations with clinicopathological features. Among young patients with breast cancer, these four cell states were less abundant in Grade 3 and basal‐like tumors, but their abundances did not differ significantly across tumor stages (Figure S4). Lower cell state abundance was also associated with poorer prognosis (Figure 1E). Together, these findings identify reductions in distinct immune, stromal, and epithelial cell states in YBC and link these alterations to aggressive clinicopathological features and unfavorable outcomes.

3.2. YBC‐Associated Cell States Form a Spatially Organized Multicellular Ecosystem

Correlation analysis of bulk transcriptomic data revealed strong associations among the abundances of CD4+ T cells (S07), fibroblasts (S05), monocytes/macrophages (S09), and epithelial cells (S05) (Figure 2A). Enrichment analysis of genes associated with lower cell state abundance identified the PD‐L1 expression and programmed cell death 1 (PD‐1) checkpoint pathway in cancer as a shared pathway for CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) (Figure S5; Table S3). These findings prompted us to investigate whether the cell states were spatially associated. Enrichment of the three cell states in ST data was quantified using AddModuleScore. Jaccard similarity analysis revealed substantial spatial overlap among these states (Figure 2B). Consistently, mIF imaging and visualization of ST sections showed that the three cell states tended to cluster together and colocalize (Figure 2C,D; Figures S6 and S7). We combined their marker‐gene signatures to calculate a spatial CE enrichment score. Analysis of putative ligand–receptor interactions identified 45 pairs with significant spatial colocalization across all sections with FDR < 0.05 (Figure 2E). Among these pairs, CDH1‐ERBB3 (r = 0.36) and CDH1‐PTPRF (r = 0.46) showed the strongest associations with the CE score (Figure 2E). Spatial mapping showed that these pairs were enriched in the regions occupied by the ecosystem comprising CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) (Figure 2F). Analyses of within‐spot and neighborhood relationships also identified the strongest association between CDH1‐PTPRF and CE enrichment (Figure S8). Gene signatures related to lower CE abundance were enriched in pathways involved in tumor progression (Figure S9; Table S3). Taken together, these findings indicate that CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) colocalize and may form a distinct multicellular ecosystem within the TME.

FIGURE 2.

Spatial colocalization of cell states and putative ligand–receptor interactions. (A) Correlations among the abundances of CD4+ T cells, epithelial cells, fibroblasts, and monocytes/macrophages. Colors indicate correlation coefficients (r), with blue representing negative correlations (r < 0) and red representing positive correlations (r > 0). Crosses indicate non‐significant correlations (p > 0.05). (B) Heatmap of spatial overlap among CD4+ T cells, epithelial cells, fibroblasts, and monocytes/macrophages in spatial transcriptome (n = 8), assessed using the Jaccard similarity index. Darker red indicates greater similarity. (C) Representative multiplex immunofluorescence images showing the expression and colocalization of the CD4 (orange) and OPRM1 (yellow), representing CD4+ T cells (S07); PDGFR (α + β) (red) and ZDHHC1 (cyan), representing fibroblasts (S05); and CD68 (white) and DLEC1 (green) representing monocytes/macrophages (S09) in young and old breast cancer (OBC) patients. Nuclei were counterstained with DAPI (blue). (40× objective lens, Scale bar: 100 μm). PDGFR, platelet‐derived growth factor receptor; DAPI, 4′,6‐diamidino‐2‐phenylindole. (D) Spatial distributions of CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) in breast cancer samples. (E) Heatmap of associations between ligand–receptor pair scores and the multicellular ecosystem comprising CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09). Merged represents combined spatial transcriptomic data (n = 8), analyzed using Pearson's correlation coefficient. (F) Spatial distributions of the multicellular ecosystem (left) and ligand–receptor pair scores (right) in breast cancer samples.

3.3. Prognostic and Therapeutic Implications of the Multicellular Ecosystem

We derived the CE score from bulk transcriptomic data using PCA and examined its associations with clinical outcome (Figure S10). CE abundance was significantly lower in YBC patients (Figure 3A). A similar pattern was observed in young patients with uterine corpus endometrial carcinoma (UCEC) (Figure 3B). Patients were classified into CE‐low and CE‐high status using optimal cutoff value (Figures S11 and S12). CE‐low status was associated with higher histological grade and more aggressive molecular subtypes (Figure S13). CE‐low status was associated with poorer outcomes in univariable analyses (Figure 3C) and in multivariable Cox models for disease‐free interval (DFI) in TCGA (HR, 2.98; 95% CI, 1.00–8.90), relapse‐free survival (RFS) in METABRIC (HR, 1.26; 95% CI, 1.01–1.56), and overall survival (OS) in SCAN‐B (HR, 1.34; 95% CI, 1.02–1.74) after adjustment for age, TNM stage, and molecular subtype, etc. (Figure 3D). Analyses stratified by molecular subtype consistently showed poorer outcomes, in CE‐low patients with hormone receptor‐positive/human epidermal growth factor receptor 2‐negative (HR+/HER2−) breast cancer across TCGA, METABRIC, and SCAN‐B cohorts (Figure S14). We also evaluated the prognostic relevance of CE status after neoadjuvant chemotherapy. In GSE25066 and GSE22226, CE‐low status was also associated with poorer prognosis among patients who failed to achieve pathological complete response (pCR) (Figure 3E), especially in HR+/HER2− subtype (Figure S15). In GSE25066, CE‐low status remained associated with poorer distant relapse‐free survival (DRFS) in multivariable analysis (HR, 2.05; 95% CI, 1.04–4.06) (Figure 3F). CE abundance and status yielded concordance indices (C‐indices) comparable to those of conventional clinicopathological factors, including tumor stage and molecular subtype, supporting their potential prognostic value (Table S4).

FIGURE 3.

Prognostic value of the multicellular ecosystem. (A) Comparisons of multicellular ecosystem abundance between young (≤ 40 years) and older (> 40 years) breast cancer patients in the TCGA‐BRCA, METABRIC and SCAN‐B datasets. ****p < 0.0001; ***p < 0.001; **p < 0.01; *p < 0.05. (B) Comparisons of multicellular ecosystem abundance between young (≤ 40 years) and older (> 40 years) cancer patients in pan‐cancer datasets of TCGA. ***p < 0.001; **p < 0.01; *p < 0.05. (C) The Kaplan–Meier survival curves for CE‐high and CE‐low patients in TCGA‐BRCA, METABRIC, and SCAN‐B datasets. CE, multicellular ecosystem; TCGA‐BRCA, The Cancer Genome Atlas breast cancer cohort; METABRIC, Molecular Taxonomy of Breast Cancer International Consortium; SCAN‐B, Sweden Cancerome Analysis Network‐Breast. (D) Forest plots of univariable and multivariable Cox regression models including CE status and clinicopathological covariates. CI, confidence interval; HR, hazard ratio. ^aAdjusted for age (≤ 40 or > 40 years), T stage, N stage, M stage, molecular subtype, and adjuvant therapy; ^bAdjusted for age (≤ 40 or > 40 years), Tumor stage, N stage, molecular subtype, histological grade, endocrine therapy, and chemotherapy; ^cAdjusted for age (≤ 40 or > 40 years), tumor size, nodal status, molecular subtype, histological grade, endocrine therapy, and chemotherapy. (E) The Kaplan–Meier survival curves for patients in GSE25066 and GSE22226 datasets, stratified by CE status and achievement of a pathological complete response (pCR). (F) Forest plots of multivariate Cox regression models including CE status and clinicopathological covariates. CI, confidence interval; HR, hazard ratio. ^dAdjusted for age (≤ 40 or > 40 years), T stage, N stage, molecular subtype, and histological grade; ^eAdjusted for age (≤ 40 or > 40 years), T stage, molecular subtype, and histological grade.

Pathway enrichment analysis showed greater activity in CE‐low tumors for pathways related to cell proliferation (mTORC1 signaling and MYC targets V1/V2), cancer‐related mechanisms (KRAS signaling up, p53 pathway, IL6‐JAK‐STAT3 signaling, TNF‐α signaling via NF‐κB, TGF‐β signaling, epithelial mesenchymal transition (EMT), adipogenesis, and angiogenesis), and treatment response of breast cancer (estrogen response early/late) (Figure S16). These findings were consistent with the more aggressive clinicopathological features of CE‐low tumors.

4. Discussion

Host cells take on dual roles in their interactions with cancer cells, making it hard to discern whether they are friends or foes [27]. Understanding the diversity of cell states and the multicellular ecosystems formed by their interactions is therefore important for elucidating cancer initiation and progression. Previous studies of the YBC microenvironment have primarily examined changes in cellular composition [19, 20], particularly immune cell populations, potentially overlooking cell state heterogeneity and multicellular organization. We used EcoTyper to extend this characterization to cell type specific transcriptional states [33]. The reduced abundances of CD4+ T cells (S07), fibroblasts (S05), monocytes/macrophages (S09), and epithelial cells (S05) support the view that YBC has distinct microenvironmental features [19]. Moreover, current theory emphasizes that cells act cooperatively rather than in isolation [27, 29, 41]. Consistent with this perspective, CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) were spatially colocalized and appeared to form a multicellular ecosystem associated with breast cancer outcomes.

The aggressive behavior of YBC has often been attributed to tumor‐intrinsic features, including increased proliferative activity, unfavorable molecular subtypes [2, 42, 43], and genetic susceptibility, such as a high frequency of TP53 and BRCA1/2 mutations in this population [44, 45]. However, these factors may not fully explain its distinct biology and poor outcomes. Our findings provide a complementary microenvironmental perspective. Tumorigenesis is a dynamic process involving both gene–environment interactions and a transition from a tumor‐suppressive to a tumor‐permissive microenvironment [27, 46]. Consistent with this framework, lower CE abundance was associated with increased activity of cancer‐related pathways, including KRAS signaling up, p53 pathway, and IL6‐JAK‐STAT3 signaling. Disruption of a spatially organized immune‐stromal ecosystem may therefore be an additional feature associated with tumor progression.

Tumor biology is strongly influenced by the cellular composition and spatial organization of the TME [47, 48, 49, 50]. Reduced immune cell infiltration may indicate feeble ability to combat tumors [29, 51]. Previous studies have reported lower abundance of M2 macrophages and Th17 cells in YBC [19, 20], although associations between microenvironmental features and clinical outcomes have been inconsistent [52, 53]. One explanation is that analyses of isolated cell states may underestimate the importance of spatially coordinated multicellular structures. In this study, CE‐low status was related to poorer survival in both univariable and multivariable Cox analyses, suggesting potential prognostic value, especially in HR+/HER2− breast cancer. This is the most common subtype and generally has a favorable prognosis with endocrine therapy [54]. Nevertheless, nearly 30% of high‐risk HR+/HER2− patients may experience recurrence within 5 years post‐surgery [55], and this risk can persist for up to 20 years [56]. CE‐low status may help identify a subset of patients with high‐risk microenvironmental features for further evaluation. The KEYNOTE‐756 and CheckMate‐7FL trials have supported the potential value of adding PD‐1 blockade to neoadjuvant treatment in high‐risk HR+/HER2− patients [57, 58]. However, identifying patients who may benefit from the addition of immunotherapy remains challenging. Our findings suggest that CE‐low status, which is associated with high‐risk phenotypes, may serve as a candidate biomarker for tailoring neoadjuvant treatment strategies, including consideration of anti‐PD‐1 therapy. At the patient level, achieving pCR after neoadjuvant therapy is associated with better survival [59, 60, 61]. However, the efficacy of neoadjuvant chemotherapy varies significantly, with pCR rates ranging from 7.5% to 50.3% across different subtypes [61], and the prognosis of patients who failed to achieve pCR remains heterogeneous. Circulating tumor DNA (ctDNA) is a potential prognostic biomarker [62], but one study reported undetectable baseline ctDNA in 16% of patients [63]. Its applicability in HR+/HER2− disease also warrants further evaluation because ctDNA levels are much lower in this subtype [64]. The CE identified here may provide an additional candidate biomarker for identifying patients with HR+/HER2− who remain at high risk of recurrence or metastasis after neoadjuvant therapy.

Targeting non‐malignant components of the TME, rather than cancer cells alone, has emerged as a therapeutic strategy [65]. Several approaches have been developed to modulate specific TME components, including inhibition of tumor‐resident Tregs by targeting CCR8 [66], depletion of FAP+ cancer‐associated fibroblasts (CAFs) [67], and reprogramming of tumor‐associated macrophages (TAM) via CD40 agonists [68]. However, the therapeutic effects of most approaches remain uncertain. The challenge stems from the state‐dependent functions of individual cell types and the complex interactions among them, which may limit the efficacy of therapies targeting individual TME components. This complexity highlights the rationale for targeting the multicellular ecosystem within TME for cancer prevention and treatment. Growing evidence supports the feasibility of therapeutically remodeling these ecosystems, as exemplified by the induction of tertiary lymphoid structures (TLSs) in immunotherapy scenario [69]. The CE identified in this study may provide a conceptual basis for future ecosystem‐level therapeutic strategies in YBC, although whether its restoration or remodeling improves treatment response requires experimental validation.

This study has several limitations as well. First, most of our findings are computational in nature. Although mIF imaging and ST analyses supported the spatial colocalization of CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09), and computational analyses identified putative ligand–receptor interactions, validation using perturbation experiments and animal or patient‐derived models is needed. Second, the ST data lacked single‐cell resolution, and each spot could contain signals from multiple cell types. Although mIF provided imaging‐based evidence of spatial proximity, high‐resolution spatial profiling is needed to delineate cellular organization more precisely. Finally, the datasets used were derived primarily from populations of European ancestry, which may limit generalizability. Validation in large prospective, multicenter cohorts is needed to determine whether this CE can serve as a robust biomarker for risk stratification or treatment guidance.

5. Conclusions

In summary, our study comprehensively characterizes the tumor microenvironment in YBC by integrating large‐scale bulk and spatial transcriptomic data. We identified a spatially organized ecosystem comprising CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09). Lower abundance of this multicellular ecosystem was associated with poorer outcomes among patients who did not achieve pCR after neoadjuvant treatment. These findings provide new insights into the microenvironmental features of YBC and a basis for further investigation of prognostic biomarkers and ecosystem‐directed therapeutic intervention.

Author Contributions

Xiaolin Chen: conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, software, visualization, writing – original draft. Zhuozhi Liang: conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, software, supervision, validation, visualization, writing – original draft. Xinming Chen: formal analysis, project administration, validation, writing – original draft. Qian Ouyang: funding acquisition, methodology, project administration. Qianru Chen: data curation, investigation, resources. Liling Zhu: conceptualization, supervision, writing – review and editing. Shunrong Li: conceptualization, funding acquisition, investigation, supervision, writing – review and editing. Kai Chen: conceptualization, funding acquisition, methodology, project administration, supervision, writing – review and editing.

Funding

This work was supported by grants from the National Natural Science Foundation of China (#82503417, #82605434 and #82303552), Guangdong Provincial Clinical Research Center for Breast Diseases (#2023B110005), Guangdong Science and Technology Department (#2024B1212030002), and Shanwei Science Technology Commission (#250424138940494), Postdoctoral start‐up program supported by Zenith Institute of Medical Sciences, Guangzhou (#KQ05220250020), Yat‐sen Postdoctoral start‐up program supported by China Guangfa Talent Cultivation Project (#9390025005), and by the grants from the Youth project supported by Guangdong medical research foundation (#B2026204). This work was also supported by grants from Guangdong Provincial Fund Committee Regional Joint Fund—Youth Fund Project (#2023A1515110943).

Ethics Statement

This study was conducted in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committee of Sun Yat‐Sen Memorial Hospital, Sun Yat‐Sen University (SYSKY‐2024‐928‐01). For the publicly available datasets, ethics approval and informed consent were obtained by the original investigators. Archived tissue sections used for multiplex immunofluorescence were obtained from the Department of Pathology. The requirement for informed consent for the retrospective use of these specimens was waived by the ethics committee.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Recovery of CD4+ T cells, epithelial cells, fibroblasts, and monocytes/macrophages cell states. (A–D) Heatmap depicts CD4+ T cells, epithelial cells, fibroblasts, and monocytes/macrophages cell states deconvoluted from bulk transcriptome data. (A) for CD4+ T cells; (B) for epithelial cells; (C) for fibroblasts; and (D) for monocytes/macrophages cell states.

Figure S2: Correlation analysis of CD4+ T cells (S07), epithelial cells (S05), fibroblasts (S05), and monocytes/macrophages (S09) abundance with natural age in normal tissues. (A) In TCGA dataset; (B) In GTEx dataset.

Figure S3: Abundance of CD4+ T cells state (S07), epithelial cells state (S05), fibroblasts state (S05), and monocytes/macrophages state (S09) between BRCA1/BRCA2 mutation carriers and non‐carriers. (A) In Young breast cancer (YBC) patients from TCGA‐BRCA dataset; (B) In Young breast cancer (YBC) patients from METABRIC dataset; (C) In old breast cancer (OBC) patients from TCGA‐BRCA dataset; (D) In old breast cancer (OBC) patients from METABRIC dataset. ns, not significant; ***p < 0.001; **p < 0.01; *p < 0.05.

Figure S4: Abundance of CD4+ T cells state (S07), epithelial cells state (S05), fibroblasts state (S05), and monocytes/macrophages state (S09) of breast cancer patients. (A) The abundance differences among different molecular subtypes in young breast cancer; (B) The abundance differences among different histological grade in young breast cancer; (C) The abundance differences among different tumor stage in young breast cancer;(D) The abundance differences among different molecular subtypes in old breast cancer; (E) The abundance differences among different histological grade in old breast cancer; (F) The abundance differences among different tumor stage in old breast cancer. ns, not significant; ****p < 0.0001; ***p < 0.001; **p < 0.01; *p < 0.05.

Figure S5: Venn diagram of the shared pathways enriched among genes negatively associated with CD4+ T cells (S07), epithelial cells (S05), fibroblasts (S05), and monocytes/macrophages (S09) abundance.

Figure S6: The percentage of CD4_S07, Macrophage_S09, and Fibroblast_S05 positive cells in young vs. old breast cancer (OBC) patients.

Figure S7: Spatial relationships among CD4_S07, Macrophage_S09, and Fibroblast_S05 cells. (A) Pairwise analyses of the proportions of CD4_S07, Macrophage_S09, and Fibroblast_S05 cells within 100 μm of each other. (B) Pairwise analyses of the nearest‐neighbor distance among CD4_S07, Macrophage_S09, and Fibroblast_S05 cells.

Figure S8: Spatial transcriptome analysis of multicellular ecosystem and ligand–receptor pairs. (A) Importance of ligand–receptor pairs in the prediction of multicellular ecosystem within a spot; (B) in the local neighborhood (within a 2‐spot radius); (C) and in the extended neighborhood (within a 5‐spot radius).

Figure S9: Enriched pathways of genes negatively associated with multicellular ecosystem.

Figure S10: Correlations between the principal component 1 (PC1) score and the abundance of the constituent cell states. Heatmap of Pearson correlation coefficients between the PC1 score and the abundance of CD4+ T cells (S07), fibroblasts (S05), and monocytes/macrophages (S09) in breast cancer (TCGA‐BRCA, METABRIC, SCAN‐B, GSE25066, GSE22226), and pan‐cancer datasets of TCGA.

Figure S11: The optimal cutoff point to dichotomize multicellular ecosystem (CE) abundance into CE‐high and CE‐low subgroups using maximally selected rank statistics. (A) In TCGA‐BRCA; (B) In METABRIC; (C) In SCAN‐B; (D) In GSE25066; (E) In GSE22226.

Figure S12: Principal component analysis identified two distinct multicellular ecosystem status. (A) In TCGA‐BRCA; (B) In METABRIC; (C) In SCAN‐B; (D) In GSE25066; (E) In GSE22226.

Figure S13: The distribution of clinical phenotypes in different CE status. (A) In TCGA‐BRCA; (B) In METABRIC; (C) In SCAN‐B.

Figure S14: Kaplan–Meier curves for clinical outcomes according to multicellular ecosystem categories in different molecular subtypes of breast cancer. (A) For disease free interval in TCGA‐BRCA; (B) For relapse free survival in METABRIC; (C) For overall survival in SCAN‐B.

Figure S15: Kaplan–Meier curves for clinical outcomes according to pathological complete response status and multicellular ecosystem categories in different molecular subtypes of breast cancer. (A) For DRFS in GSE25066; (B) For OS in GSE22226.

Figure S16: Pathway enrichment in different CE status.

Table S1: Description: The details of datasets involved in this study.

Table S2: Description: Deconvoluted cell states abundance of CD4+ T cells (S07), fibroblasts (S05), monocytes/macrophages (S09), and epithelial cells (S05).

Table S3: Description: Pathway enrichment with genes associated with CD4+ T cells (S07), fibroblasts (S05), monocytes/macrophages (S09), and epithelial cells (S05), and multicellular ecosystem.

Table S4: Description: The C‐index comparison of CE abundance and status with clinicopathological factors.

Data Availability Statement

Publicly available transcriptomic and clinical datasets analyzed in this study were obtained from UCSC Xena, cBioPortal, the GTEx Portal, and GEO (GSE96058, GSE25066, and GSE22226), as described in the Methods. Spatial transcriptomic datasets were obtained from 10x Genomics.