A prognostic model for breast cancer based on mitochondria-associated endoplasmic reticulum membrane (MAM) signature genes

Abstract

Background

Mitochondria-associated endoplasmic reticulum membranes (MAMs) have emerged as key regulators of breast cancer biology. Growing evidence indicates that MAM dysfunction affects tumor progression, immune regulation, and treatment response. However, their prognostic value and therapeutic relevance in breast cancer remain insufficiently clarified.

Methods

We curated 83 MAM-related genes from the literature and used TCGA breast cancer cohorts to construct a prognostic model through univariate Cox, LASSO, and multivariate Cox regression analyses. Protein-level validation was performed using the Human Protein Atlas (HPA). Functional enrichment, immune infiltration, immunotherapy response prediction, and drug sensitivity profiling were conducted between risk groups. Single-cell RNA-seq data (GSE255068) were analyzed to map gene expression within the tumor microenvironment. Finally, qRT-PCR and western blotting validated the expression of key genes in breast cancer cell lines.

Results

A four-gene signature (VDAC1, RYR2, PINK1, and TESPA1) robustly stratified breast cancer patients into high- and low-risk groups and independently predicted overall survival across multiple cohorts. A prognostic nomogram integrating the risk score with clinicopathological variables significantly improved survival prediction accuracy. Low-risk tumors were characterized by a more inflamed tumor microenvironment, with increased immune cell infiltration and enhanced immune-related pathway activity, and were predicted to respond more favorably to immune checkpoint blockade. In contrast, high-risk patients exhibited greater predicted sensitivity to multiple cytotoxic and targeted agents, a pattern that remained largely consistent across major molecular subtypes, highlighting the robustness of the signature beyond intrinsic subtype classification. Single-cell transcriptomic analyses further revealed distinct cell-type–specific expression patterns of the four genes within the tumor microenvironment. Finally, molecular experiments confirmed elevated expression of VDAC1, RYR2, and PINK1 in breast cancer cell lines, supporting their biological relevance in tumor progression.

Conclusions

This MAM-based prognostic model provides new insights into breast cancer biology and has potential utility in guiding personalized treatment, particularly regarding immunotherapy and chemotherapy sensitivity. Experimental validation highlights VDAC1 as a promising biomarker.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05287-4.

Introduction

Breast cancer is one of the most common malignancies among women and remains a major threat to global health [1]. Contemporary diagnostic and therapeutic paradigms are guided by molecular subtyping, which stratifies the disease into three principal categories: hormone receptor-positive, HER2-positive, and basal-like. The therapeutic arsenal, encompassing locoregional interventions (surgery, radiotherapy) and a spectrum of systemic treatments (chemotherapy, endocrine, targeted, and immunotherapeutic agents), has collectively contributed to substantial improvements in survival rates[2]. Despite these advances, a considerable proportion (30–40%) of patients ultimately experiences metastatic recurrence,So the prognosis for advanced-stage disease remains particularly unfavorable, with a five-year survival rate approximating only 26% [3, 4]. Given this context, it is imperative to develop and validate novel biomarker signatures that can advance early detection capabilities and optimize prognostic assessment.

MAM is a specialized membrane contact site that structurally bridges and functionally couples the endoplasmic reticulum (ER) with mitochondria.This interface is indispensable for facilitating efficient inter-organellar communication [5–8]. Mechanistically, MAM-mediated calcium cycling activates mTORC1 signaling, promotes the mitochondrial localization of hexokinase 2 (HK2), and enhances the Warburg effect in tumor cells[9, 10]. In breast cancer, the MAM component FUNDC1 imparts resistance to anthracycline chemotherapy, while STIM1, another key MAM protein, shapes the immunosuppressive landscape by promoting CTLA-4⁺ regulatory T-cell infiltration, ultimately influencing immunotherapy success [11, 12].Beyond cancer, MAM dysfunction is also implicated in cardiovascular, metabolic, and neurodegenerative diseases. For instance, the IP3R–GRP75–VDAC1 complex mediates calcium overload–induced cell death in cardiac disease [13]; AMPK inactivation under hyperglycemic conditions enhances Ca²⁺ release and mitochondrial dysfunction in diabetes[14]; And autophagy initiation is regulated by the VAPB-PTPIP51 complex through the modulation of Ca²⁺ signalling. [15]. In mitophagy, FUNDC1 promotes mitochondrial fission by recruiting DRP1 to MAMs, while the PINK1/Parkin pathway enhances autophagosome formation by strengthening ER–mitochondria contacts[16].Given their multifaceted role in calcium regulation, lipid metabolism, and stress response, MAMs are emerging as a critical node in cancer biology and a potential therapeutic target. Yet, the clinical translation of these findings, particularly their prognostic and therapeutic value, remains largely unelucidated.

Despite the biological importance of MAMs, their potential as prognostic biomarkers and therapeutic predictors in breast cancer remains largely unexplored. To address this gap, we integrated multi-omics systems analyses, single-cell transcriptomic profiling, and molecular experiments to develop and validate the first MAM-based prognostic model in breast cancer. This model identifies a four-gene signature (VDAC1, RYR2, PINK1, TESPA1) that reflects tumor biology, immune activity, and treatment sensitivity.

Materials and methods

Data collection

The initial gene set comprising 83 MAM-related genes was curated through a comprehensive review of existing literature[17–19]. For model construction and validation, we procured RNA sequencing data alongside detailed clinical annotations for a cohort encompassing 941 breast tumor specimens and 95 adjacent normal tissues. These data were sourced from the TCGA-BRCA project using the UCSC Xena platform (https://xenabrowser.net, accessed on April 23, 2025).The GSE20711 dataset was downloaded from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo) for external validation of the prognostic model. Single-cell RNA sequencing data from the GEO repository (accession GSE255068) was leveraged to interrogate the expression patterns of the signature genes at single-cell resolution.

Construction and validation of the prognostic model

The construction of our prognostic model followed a stepwise procedure. The pipeline commenced with a univariate Cox regression analysis to isolate MAM-related genes (MaMs) exhibiting potential prognostic value.To enhance model generalizability and mitigate overfitting, we employed Least Absolute Shrinkage and Selection Operator (LASSO) regression with its penalty parameter (λ) optimized by tenfold cross-validation.[20]. In conclusion, the genetic elements identified by LASSO technique were integrated into a multivariate Cox regression model, thereby constructing a reliable and autonomous prognostic signature. For each patient, a continuous risk value was calculated by combining the standardised levels of the final prognostic genes with their respective regression values calculated from the multivariate Cox model. The calculation followed the standard formula: risk score = Σ (Gene_Expression_i × Coefficient_i).Using the cohort-wide median risk score as a cutoff, we stratified patients into two distinct prognostic categories: a high-risk group and a low-risk group, for all subsequent comparative analyses. The accuracy of MaRGs in different cohorts was evaluated by comparing the survival curves and ROC curve analysis. We also performed stratified analysis based on clinical characteristics to determine the importance of MAPGS in different groups.Protein expression of the model genes and its impact on overall survival (OS) in breast cancer patients were further assessed utilizing the publicly available Human Protein Atlas (HPA) database.

Construction and evaluation of the nomogram

A prognostic nomogram, which integrated the molecular risk score with clinical parameters (age, sex, tumor stage), was developed from the Cox regression results to estimate 1-, 3-, and 5-year OS. The model's calibration, discrimination, and clinical utility were respectively assessed using calibration plots, ROC curves, and DCA, while its time-dependent predictive performance was quantified with the "timeROC" package.

Functional enrichment analysis of differentially expressed genes

In order to ascertain the existence of any potential biological basis for risk stratification, differentially expressed genes (DEGs) between the groups were initially screened using the 'limma' package, with thresholds of |log2FC|> 0.585 and adjusted p-value < 0.05 applied. The functional profile of these significant differentially expressed genes (DEGs) was then interrogated through Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses implemented in the “clusterProfiler” package.Furthermore, gene set enrichment analysis (GSEA) was conducted utilising the c2.all.v2024.1.Hs.symbols dataset as a reference, in order to identify the pathways that distinguish the two risk groups[21].

Immune landscape characterization

The cellular composition of the tumour microenvironment (TME) was initially estimated by way of the 'ESTIMATE' protocol, a computational method that uses bulk transcriptomic data to calculate stromal, immune, and combined (ESTIMATE) scores[22].To obtain a more granular and robust view of immune cell infiltration, we deployed a consensus approach utilizing seven established deconvolution algorithms… thereby providing a comprehensive portrait of the immune landscape stratified by risk [23].Activity levels of immune-related functional pathways were quantified using single-sample gene set enrichment analysis (ssGSEA) via the "GSVA" and "GSEABase" R packages.

Immunotherapy response and drug sensitivity analyses

To estimate the response likelihood to immune checkpoint inhibition (ICI), the Immunophenoscore (IPS) data are obtained directly from The Cancer Immunome Atlas (TCIA, https://tcia.at/home).The SubMap procedure (http://cloud.genepattern.org/gp) was used to calculate the resemblance between our sample subgroups and known recipients of anti–PD-1 and anti–CTLA-4 treatments[24].The Wilcoxon test was used to compare the expression levels of common immune checkpoint genes between high- and low-risk cohorts[25].Chemotherapeutic and targeted agent susceptibility was profiled in silico using the 'pRRophetic' package. This tool predicts the half-maximal inhibitory concentration (IC50) for a broad panel of 198 anticancer compounds by leveraging baseline gene expression data from our cohort, as described previously[26, 27].

Single-cell RNA sequencing analysis

The scRNA-seq data analysis pipeline was implemented primarily with the “Seurat” package. Raw data from dataset GSE255068 were subjected to quality control, retaining cells with mitochondrial gene percentages below 20% and unique gene counts between 200 and 7500, while filtering out genes detected in fewer than three cells. Technical batch effects were mitigated with the “Harmony” package. We then identified the top 3,000 highly variable genes for dimensionality reduction and cell clustering. The resulting clusters were visualized using UMAP and annotated into distinct cell types with the “SingleR” package, which references canonical marker genes.

Molecular biology assays

Cell culture

Human breast cancer cell lines (MCF-7, T-47D, MDA-MB-231, MDA-MB-468, AU565, and SK-BR-3) were obtained from the American Type Culture Collection (ATCC). These cell lines were deliberately selected to represent the major intrinsic molecular subtypes of breast cancer, including luminal A (MCF-7, T-47D), basal-like/triple-negative (MDA-MB-231, MDA-MB-468), and HER2-enriched tumors (AU565, SK-BR-3), thereby enabling subtype-spanning biological validation of the MAM signature. Cells were cultured at 37 °C in a humidified atmosphere containing 5% CO₂ using DMEM or RPMI-1640 supplemented with 10% fetal bovine serum and 1% penicillin–streptomycin. Cell line authenticity and mycoplasma-free status were routinely confirmed by STR profiling and mycoplasma testing.

Quantitative real-time PCR

Total RNA was extracted from cultured cells using the FastPure Cell/Tissue Total RNA Isolation Kit (Vazyme), followed by quantification with a NanoDrop spectrophotometer. Subsequently, 500 ng of total RNA was reverse-transcribed into cDNA using the HiScript III RT SuperMix (Vazyme). Quantitative real-time PCR (qPCR) was then performed with ChamQ Universal SYBR qPCR Master Mix on a [ABI Q6; Accurate Biology] real-time PCR system, with all primers synthesized by Sangon Biotech. Gene expression levels were normalized to GAPDH and calculated using the comparative 2^–ΔΔCt method.

Western blot analysis

Cells were lysed on ice using a strong RIPA lysis buffer supplemented with PMSF and phosphatase inhibitors. The lysates were centrifuged at 12,000 × g for 10 min at 4 °C to collect the supernatant. Protein concentration was quantified using a BCA assay kit (CWBIO). Total proteins (30 μg per lane) were denatured, separated by SDS-PAGE (Yeasen prefabricated gels), and transferred to PVDF membranes via wet transfer. The membranes were blocked with 5% non-fat milk in TBST for 1 h at room temperature, followed by incubation with primary antibodies (Proteintech) overnight at 4 °C. After washing with TBST, the membranes were incubated with HRP-conjugated secondary antibodies for 1 h at room temperature. Protein bands were visualized using an ECL substrate (Vazyme Biotech) and imaged with a chemiluminescence imaging system. Band intensities were analyzed with ImageJ software.

Statistical analysis

This study employed the R programming platform (v4.2.2) for all data processing, statistical analysis, and figure creation. Survival probabilities between the dichotomized risk cohorts were plotted via Kaplan–Meier curves, and the homogeneity of survival functions was formally tested using the log-rank method. Inter-group comparisons for continuous metrics involved the Wilcoxon rank-sum test, whereas differences in distribution of categories was conducted through the utilisation of the chi-square test. Linear dependencies between two continuous parameters were examined by calculating Pearson's correlation coefficients. A probability value (p-value) of less than 0.05 in a two-tailed test was defined as the benchmark for a statistically significant finding.

Results

Development and validation of the MAM-related prognostic model

Our study cohort comprised 1,036 breast cancer cases from TCGA-BRCA, for which transcriptomic profiles and corresponding clinical data were integrated. To ensure unbiased model validation, the dataset was randomly partitioned into training (70%) and testing (30%) subsets using the createDataPartition function from the R caret package.To mitigate overfitting and pinpoint the most prognostically relevant MaRGs, we applied LASSO regression (Fig. 1A-B). The genes selected by LASSO were subsequently entered into a multivariate Cox regression analysis, which culminated in an optimized prognostic signature based on four genes: PINK1, RYR2, TESPA1, and VDAC1. The risk score was calculated as follows: h0 exp[(0.359 × PINKE) + (0.455 × RYR2)—(0.254 × TESPA1) + (0.5 × VDAC1)].Stratification of patients into high- and low-risk categories was performed using the median risk score from the training set as the cutoff. A clear demarcation between these two prognostic groups was visually apparent in the ordered risk heatmaps, score distribution plots, and survival status plots, collectively affirming the model's discriminative capacity (Fig. 1C-D).Survival trajectories diverged significantly between the groups, with Kaplan–Meier analysis demonstrating markedly inferior overall survival for high-risk patients relative to their low-risk counterparts (P < 0.001, Fig. 1D). The model's predictive accuracy was consistently maintained over time, as evidenced by time-dependent receiver operating characteristic (ROC) analysis. The area under the curve (AUC) for predicting 1-, 3-, and 5-year overall survival reached 0.669, 0.646, and 0.633, respectively, indicating robust prognostic performance (Fig. 1E).When compared with previously reported breast cancer biomarker models, our four-gene MaRG model achieved higher AUC and concordance index (C-index) values, demonstrating superior prognostic accuracy [28–30] (Fig. 1F–I). The generalizability of the prognostic model was assessed by applying it to an independent external cohort, the GEO dataset GSE20711 (n = 98).The analysis successfully reproduced the key finding: low-risk patients had significantly superior survival (P = 0.045, Fig. 1J). Furthermore, the AUC values for 1-year OS prediction were 0.816(Fig. 1K), thereby supporting the model's robustness and transferability to independent data.

Fig. 1.

Construction and validation of the MaRGS. A Cross validation method to select optimal genes. B The Lasso coefficient profiles. C The OS KM curves between high- and low-risk groups in the TCGA-BRCA cohort. D The model genes, risk score and clinical outcomes in the two risk groups. E The time-dependent ROC curves of the MaRGS. F–H The time-dependent ROC curves of Wang’s, Yang’s and Zhang’s gene signature. I C-idex of our signature, Wang’s, Yang’s and Zhang’s signature. J The OS KM curves between high- and low-risk groups in the GEOcohort. K The time dependent ROC curves of the MaRGS in the GEO cohort

Stratified analysis and protein expression of MaRGs

To further ascertain the clinical translatability of the four-gene signature, we conducted stratified survival analyses within predefined clinical subsets.Risk score distributions varied significantly with patient age and tumor stage. Older patients exhibited higher risk scores (P = 0.00026), and patients with T4-stage disease showed higher risk scores than those with T1–T3 tumors (P = 0.0083, 0.0056, and 0.0083, respectively) (Fig. 2A–B).Stratified survival analyses across various clinical subsets (age, sex, TNM stage) confirmed the robustness of our model, with the high-risk group exhibiting inferior survival in all but the M1 stage subgroup. The non-significance in M1 stage is noted, potentially attributable to its limited sample size (Fig. 2C-L).To further evaluate whether the prognostic value of the signature was consistent across intrinsic molecular backgrounds, additional survival analyses were performed within PAM50-defined breast cancer subtypes. Notably, the four-gene signature retained its prognostic discrimination within the major molecular subtypes, with high-risk patients consistently exhibiting poorer overall survival compared to their low-risk counterparts. While the magnitude of survival differences varied among subtypes, the overall trend remained concordant, supporting the subtype-independent prognostic robustness of the signature (Supplementary Figure S1).Protein-level validation was pursued using immunohistochemical images from the HPA. This investigation revealed elevated protein levels of VDAC1, RYR2, and PINK1 in breast carcinoma specimens compared to normal glandular tissues. In contrast, TESPA1 protein levels did not exhibit a significant differential expression (Fig. 3A-D).Kaplan–Meier analysis revealed that elevated expression of VDAC1, RYR2, and PINK1 was significantly associated with poorer patient prognosis (P < 0.05). Conversely, low expression levels of TESPA1 conferred a substantial survival disadvantage (P < 0.001, Fig. 3E).

Fig. 2.

Investigating the association of the signature with the clinicopathological characteristics of patients in the TCGA-BRCA cohort.The distribution characteristics of different clinicopathological factors in the two risk groups (A-B). The OS KM curves of the two risk groups stratified by the clinicalpathological factors. AJCC M stage, (C-D), E–F Age, AJCC N stage, (G-H),(I-J) Stage, (K-L) AJCC T stage, (G. H)

Fig. 3.

Expression and Prognosis of Key Genes in Normal Tissue and Breast Cancer Tissue A Expression of VDAC1 in normal breast tissue and tumor tissue; B Expression of TESPA1 in normal breast tissue and tumor tissue; C Expression of RYR2 in normal breast tissue and tumor tissue; D Expression of PINK1 in normal breast tissue and tumor tissue; E: Kaplan–Meier survival curve of the four marker genes

Construction and evaluation of the nomogram

Both univariate and multivariate Cox regression analyses confirmed the risk score as an independent prognostic factor, retaining significance after adjustment for key clinical variables (Fig. 4A-B). Consequently, a prognostic nomogram was developed that integrated this molecular risk score with key clinical features, thereby providing individualised predictions of 1-, 3-, and 5-year survival (Fig. 4C). The practical utility of the nomogram was confirmed through calibration curves which demonstrated a significant correlation between predicted probabilities and observed outcomes (Fig. 4D). Its discriminative power was further established by an AUC of 0.763 in ROC analysis, exceeding the performance of models based solely on clinical parameters (Fig. 4E). The clinical utility of the integrated nomogram was further validated by decision curve analysis, which showed it provided a greater net benefit than any single prognostic factor across a spectrum of risk thresholds (Fig. 4F).

Fig. 4.

Development and assessment of the nomogram. A Univariate regression. B Multivariate regression of the clinicopathological indicators and gene signature. C A comprehensive nomogram for predicting BC patients’ survival probability. D Calibration curves of the nomogram at 1-, 3-, and 5-year intervals. E The time-dependent ROC curves of the nomograma and clinicopathological indicators. F DCA curves of the clinicopathological indicators and this nomogram.* p < 0.05, **p < 0.01,*** p < 0.001

Functional enrichment and pathway analysis

To functionally characterize the prognostic groups defined by our model, we applied GSEA. This revealed that the favorable prognosis in the low-risk group was hallmarked by the activation of immune regulatory processes and a transcriptional profile akin to the normal-like subtype, including specific gene sets governing immuno-regulatory cellular crosstalk. Conversely, the group with a high risk was associated with pathways linked to estrogen receptor (ESR1) signalling, luminal B subtype features, and metastasis-related signatures such as bone metastasis and brain metastasis (Fig. 5A-B).A systematic profiling of differentially expressed genes (DEGs) underscored a pervasive theme of immune activation. Gene Ontology (GO) enrichment highlighted key roles in immune response regulation, lymphocyte differentiation, and intercellular communication (Biological Process); the T cell receptor complex and immunoglobulin complex (Cellular Component); as well as antigen binding and cytokine receptor activity (Molecular Function) (Fig. 5C). Consistent with this, KEGG pathway interrogation confirmed that the DEGs were predominantly engaged in immune and inflammatory signaling, tumor immunity, and cell differentiation cascades (Fig. 5D).

Fig. 5.

Enrichment analysis of differentially expressed genes between the low-risk and high-risk groups. A-B GSEA results for the gene sets “c2.cp.kegg_legacy.v2024.1.Hs.symbols”; C GO functional enrichment analysis of high and low risk groups in Breast cancer; D KEGG pathway enrichment analysis of high and low risk groups in Breast cancer. BP, biological process; CC, cellular component; MF, molecular function.

Immune landscape analysis

The interrogation of the tumour environment using the ESTIMATE algorithm indicated a markedly more active immune and stromal landscape in the low-risk cohort, which was reflected in significantly elevated immune, stromal, and composite ESTIMATE scores (Fig. 6A).The low-risk group was defined by a significantly more inflamed tumor immune microenvironment, as quantified by elevated infiltration levels of multiple immune cell types (B cells, CD4⁺/CD8⁺ T cells, Tregs, macrophages, dendritic cells, neutrophils, NK cells) across seven computational algorithms (Fig. 6B).The findings of the study indicate that patients with low-risk tumors demonstrate a more active immune status within their respective tumour microenvironments.In addition, correlation analyses further demonstrated that the risk score was negatively correlated with the infiltration levels of the majority of immunological cell subsets (r range: –0.4 to 0.4), including T cells, NK cells, macrophages, dendritic cells, neutrophils and B cells (Fig. 6C).Further ssGSEA robustly enriched key immune-effector pathways—such as antigen presentation, cytokine signaling, and T-cell activation/co-stimulation—within the low-risk cohort (Fig. 6D). Considering the pronounced molecular heterogeneity of breast cancer, we further assessed whether the immune infiltration patterns associated with the risk signature were preserved across PAM50 molecular subtypes. Notably, within each major subtype, low-risk tumors consistently displayed higher immune and stromal scores, as well as increased infiltration of effector immune cell populations. Although the absolute magnitude of immune infiltration varied among molecular subtypes, the directionality and overall immune trends associated with risk stratification remained concordant, indicating that the immune landscape defined by the signature is largely independent of intrinsic molecular subtype (Supplementary Figure S2).Collectively, these results suggest that a potentiated state of immune surveillance and effector activation underlies the favorable prognosis observed in low-risk patients.

Fig. 6.

Immune landscape diversity across MaRGS groups. A Comparison of tumor microenvironment composition between risk groups in the TCGA-GC cohort. B The distribution of distinctive infiltrated immune cells between two risk groups calculated using multiple algorithms.C The correlation between the riskscore and tumor-infiltrating immune cells estimated by multiple algorithms. D The difference of immune function scores between risk groups.* p < 0.05, **p < 0.01,*** p < 0.001

Prediction of immunotherapy response

Interrogation of Immunophenoscore (IPS) data from The Cancer Immunome Atlas (TCIA) indicated that elevated IPS in the low-risk group was associated with a greater likelihood of positive response to immune checkpoint inhibition, underscoring its predictive potential (Fig. 7A).The SubMap algorithm was next applied to predict immunotherapy response. It predicted a heightened likelihood of benefit from anti-PD-1 blockade in the low-risk group, whereas the association with anti-CTLA-4 response was less pronounced (Fig. 7B).Interrogation of a panel of 43 canonical immune checkpoint molecules revealed a distinct expression pattern, with transcripts for PDCD1 (PD-1), CD274 (PD-L1), PDCD1LG2 (PD-L2), and CTLA-4 being significantly more abundant in tumors from the low-risk group (P < 0.001). Conversely, emerging checkpoint molecules such as LAG3, TIGIT, and HAVCR2 (TIM-3) were overexpressed in the high-risk group, suggesting potential roles in immune escape(Fig. 7C–E) [31–33].

Fig. 7.

Prediction of the immunotherapy response in the two risk groups. A-D Comparisons of the IPS in the two risk groups. E Submap analysis between the two risk groups. F The differential expression profiles of immune-checkpoint genes between two risk groups. * p < 0.05, **p < 0.01,*** p < 0.001

Drug sensitivity analysis

Computational drug sensitivity screening via the pRRophetic algorithm positioned the high-risk cohort as potential beneficiaries of multiple therapeutic agents. This group exhibited significantly lower predicted IC50 values for a suite of drugs (all P < 0.001, Fig. 8A-J), encompassing 5-fluorouracil, cisplatin, cyclophosphamide, docetaxel, pirarubicin, gefitinib, gemcitabine, apatinib, lapatinib, oxaliplatin, palbociclib, ribociclib, vinblastine, vinorelbine, and zoledronic acid. These findings suggest distinct pharmacogenomic profiles between risk groups, potentially guiding individualized treatment strategies.To further determine whether the observed drug sensitivity patterns were influenced by intrinsic molecular heterogeneity, stratified analyses were performed within PAM50 molecular subtypes. Notably, the differential drug response between high- and low-risk groups was largely preserved across major subtypes, including Luminal A, Luminal B, HER2-enriched, and Basal-like tumors. Although the magnitude of sensitivity varied among subtypes, the directionality of drug response remained concordant, underscoring the robustness of the risk signature in capturing pharmacogenomic vulnerability beyond molecular subtype classification (Supplementary Fig. 3).

Fig. 8.

Drug sensitivity analysis between the low- and high-risk groups. A 5-fluorouracil. B Cisplatin. C Cyclophosphamide. D Docetaxel. E Epirubicin. F Gefitinib. G Gemcitabine. H Lapatinib. I Olaparib. J Oxaliplatin. K Palbociclib, L Ribociclib, M Vincristine, N Vinorelbine, O Zoledronate

Validation of model genes in single-cell RNA sequencing data

We leveraged single-cell RNA sequencing data from the GSE255068 dataset to deconvolute the breast tumor microenvironment. UMAP visualization resolved the transcriptomic landscape into 27 discrete cell clusters (Fig. 9A). Subsequent annotation based on canonical marker genes classified these clusters into nine principal lineages, such as common myeloid progenitors (CMPs), endothelial cells, and fibroblasts, among others (Fig. 9B). The defining markers of each cell cluster were visualized via a heatmap, complemented by a bar plot that quantified the relative abundance of each cell type within the tumor microenvironment (Fig. 9C-D).Among the four model genes, VDAC1 showed broad expression across multiple cell types; PINK1 was enriched in endothelial cells; TESPA1 was predominantly expressed in T cells; and RYR2 exhibited high expression in tissue stem cells (Fig. 9E). These findings confirm that the model genes display distinct, cell-type–specific expression patterns, reinforcing their biological relevance and robustness.

Fig. 9.

Validation of the gene signature in the scRNA-seq data. A-B Distribution of cell clusters and annotated cell types. C Annotation and Validation Map of Cells in Breast Tumor Tissue. D Percentage Stacked Bar Chart Showing the Proportion of Cells in Different Samples. E The expression level of the model genes in each cellular subtype

Validation of key LL-genes by qRT-PCR and western blotting

Validation of four key mitochondria-associated endoplasmic reticulum membrane (MAM) genes revealed distinct but heterogeneous expression patterns across breast cancer cell lines. Quantitative real-time PCR analysis showed that VDAC1, RYR2, and PINK1 transcripts were significantly upregulated in multiple breast cancer cell lines compared with the normal breast epithelial cell line MCF-10A, whereas TESPA1 expression did not exhibit a consistent or statistically significant difference (Fig. 10A). These expression patterns were broadly concordant with the immunohistochemical observations derived from the Human Protein Atlas.

Fig. 10.

Validation of key mitochondria-associated endoplasmic reticulum (MAM) genes in breast cancer cells. A The mRNA expression levels of VDAC1, RYR2, PINK1, and TESPA1 were measured by qRT-PCR in breast cancer cell lines compared to the normal breast epithelial cell line MCF-10A. Data are presented as mean ± SD (n = 3). *p < 0.05, **p < 0.01, ***p < 0.001 vs. MCF-10A group; ns, not significant. B Protein expression of VDAC1 was further confirmed by Western blot analysis in a panel of breast cancer cell lines

To further assess protein-level expression, western blot analysis was performed in representative breast cancer cell lines spanning major molecular subtypes. VDAC1 protein was readily detectable and showed elevated expression across several breast cancer cell lines, consistent with its transcript-level upregulation (Fig. 10B). This experiment was conducted to confirm the protein-level detectability and cross-subtype expression of MMM signature genes in commonly used experimental models, rather than to establish functional causality.

Discussion

Our study substantiates the emerging paradigm of MAMs as a central signaling hub governing inter-organellar crosstalk. The dysregulation of genes associated with the MAM complex has been demonstrated to significantly contribute to the phenotypic diversity and clinical progression of breast cancer. This process primarily occurs through the disruption of core physiological pathways, including but not limited to calcium metabolism, lipid processing, and the management and regulation of oxidative stress[34, 35]. Established clinicopathological parameters, including age, tumor stage, grade, and nodal status, constitute the cornerstone of breast cancer prognosis and treatment planning. However, their predictive accuracy is inherently constrained by the substantial heterogeneity of the disease[36].The present study establishes and robustly validates a novel, MAM-derived four-gene prognostic signature (VDAC1, RYR2, PINK1, TESPA1) for breast cancer. The signature demonstrated stable and superior prognostic performance across independent TCGA and GEO cohorts, outperforming conventional TNM staging and previously published gene signatures. Importantly, subtype-stratified survival analyses based on PAM50 classification further confirmed that the prognostic discrimination of the risk score was preserved within major molecular subtypes, including Luminal A, Luminal B, and HER2-enriched breast cancers. Although statistical significance was not consistently observed in all subgroups—most notably within the Basal-like/TNBC cohort, likely reflecting limited sample size and extreme biological heterogeneity—the overall trend toward poorer survival in high-risk patients was maintained. These findings indicate that the prognostic value of the MAM-related signature is not driven by a single molecular subtype, but rather reflects shared MAM-associated mechanisms operating across heterogeneous breast cancer backgrounds.The integrated nomogram combining risk score and clinical parameters provided robust predictive accuracy and superior clinical utility, suggesting its potential for individualized prognostic assessment and clinical decision-making.

Breast tumours are embedded within a complex and ever-changing microenvironment, which can be considered a sophisticated ecosystem composed of neoplastic cells coexisting alongside a variety of resident as well as infiltrating immune cells, stromal elements, vascular networks, and an extracellular matrix scaffold[37]. Tumor cells can reshape the TME to promote immune evasion by suppressing cytotoxic immune cell function and facilitating immunosuppressive cell infiltration. For example, in colorectal cancer, Tregs and mature regulatory dendritic cells (mregDCs) form a “Treg–mregDC–lymphatic niche,” which suppresses immune effector activity and supports tumor immune escape [38]. Furthermore, metabolic reprogramming within the TME, characterised by hypoxia, acidity, and glucose deprivation, has been demonstrated to impair the survival and functionality of effector T cells and NK cells[39].Our comprehensive dissection of the tumor immune landscape unveiled a markedly more immunologically active state in low-risk patients. This was characterized by significantly elevated stromal, immune, and composite ESTIMATE scores, concomitant with a broad enrichment of multiple antitumor immune cell populations.Notably, these immune features were largely preserved across molecular subtypes, with low-risk tumors within Luminal, HER2-enriched, and Basal-like subgroups consistently exhibiting higher immune infiltration compared to their high-risk counterparts. Although baseline immune composition differed quantitatively among subtypes—consistent with established subtype biology—the qualitative association between low-risk status and immune activation remained concordant.The application of functional enrichment analysis revealed the presence of enhanced antigen presentation, cytokine–cytokine receptor interaction, and T-cell activation in individuals classified as being at low risk. Furthermore, a robust negative correlation between risk score and immune infiltration was observed across the majority of immune subsets, thereby emphasising the immunologically active landscape of the low-risk group.

While immune checkpoint blockade has fundamentally transformed the therapeutic landscape for triple-negative breast cancer, its clinical efficacy remains constrained by low objective response rates and a pressing lack of predictive biomarkers [40]. Our findings suggest that the MAM-derived gene signature may serve as a complementary biomarker for immunotherapeutic stratification across breast cancer subtypes.Patients classified as low-risk exhibited higher Immunophenoscore (IPS) values, indicating enhanced predicted sensitivity to PD-1/PD-L1 and CTLA-4 inhibitors. The expression of key immune checkpoint molecules was found to be markedly increased in low-risk tumours, indicating potential susceptibility to immune checkpoint blockade.The key immune checkpoint molecules in question include PD-1, PD-L1,PD-L2, and CTLA-4.

The findings of the present study are supported by the results of previous studies, which lends weight to the mechanistic plausibility of the present findings. The present study demonstrates that high VDAC1 expression in cancer cells is associated with the Warburg effect, which leads to lactate accumulation and extracellular acidification. This, in turn, has been shown to suppress T-cell and NK-cell function, thereby promoting immune evasion[41]. Conversely, low TESPA1 expression impairs CD8⁺ T-cell proliferation and cytotoxic function, leading to enhanced tumor growth and immune suppression [42]. RYR2 mutations have been linked to increased infiltration of CD8⁺ T cells, activated memory CD4⁺ T cells, and M1 macrophages, suggesting an immunoregulatory role in the breast cancer microenvironment [43]. Likewise, loss of PINK1 promotes tumor cell proliferation and invasion while amplifying metabolic reprogramming that weakens immune effector activity [44].

Together, these observations suggest that the MAM-related gene signature not only predicts survival outcomes but also reflects the immune and metabolic state of the tumor microenvironment.

Despite its foundational role in breast cancer management, the enduring efficacy of chemotherapy is frequently undermined by the emergence of drug resistance, which represents a principal barrier to achieving long-term disease control [45]. Evidence from prior research strongly implicates several MAM-associated proteins in mediating resistance to chemotherapeutic agents. This is exemplified by GRP75, whose inhibition has been shown to synergize with cisplatin to promote apoptosis. Furthermore, BCL-2, a known resident of the MAM, exerts its influence on cisplatin sensitivity by fine-tuning the calcium cross-talk between the ER and mitochondria, as observed in ovarian carcinoma cells [46, 47]. PERK, a key effector of the unfolded protein response, drives therapeutic resistance by integrating the regulation of ER stress, ROS production, and calcium signaling. These interconnected processes collectively underpin the development of a tumor phenotype that is resilient to therapy[48].Consistent with the aforementioned mechanisms, our drug sensitivity analysis revealed that high-risk patients exhibited increased susceptibility to multiple commonly employed chemotherapeutic and targeted agents, including but not limited to, cisplatin, docetaxel, and lapatinib. Importantly, subtype-stratified drug sensitivity analyses demonstrated that these associations were largely consistent across molecular subtypes, with most agents overlapping with those identified in the unstratified analysis. This suggests that MAM-associated vulnerabilities may transcend subtype-specific therapeutic paradigms.VDAC1, a central channel protein in the outer mitochondrial membrane, regulates metabolite transport and facilitates drug accumulation within mitochondria, potentially enhancing the cytotoxic effects of platinum-based compounds [49–52]. In contrast, PINK1 can phosphorylate PTEN at Ser179, preventing its nuclear translocation and tumor-suppressive activity while sustaining AKT signaling and resistance to apoptosis induced by chemotherapy [53]. Moreover, in TNBC, paclitaxel has been shown to inhibit PINK1-mediated mitophagy, preventing PD-L1 degradation and leading to immune evasion and chemoresistance [54].These results emphasize that MAM integrity and its related signaling pathways play key roles in modulating drug sensitivity, metabolism, and apoptosis, making them promising targets for reversing chemoresistance.

This study presents several innovations.First, it is the first to integrate MAM-related genes into a prognostic model for breast cancer, bridging the gap between organelle communication and clinical prediction.Second, the model’s biological validity was supported by multi-omics, multi-algorithm, and single-cell–level validation, ensuring robustness.Third, by combining molecular and clinical parameters into a nomogram, we developed a clinically applicable risk assessment tool for precision oncology.Nevertheless, this study is not without its limitations. In light of the retrospective analysis, it can be concluded that the utilisation of external multicentre prospective validation is imperative for confirming the model's clinical applicability. Although subtype-stratified analyses mitigate concerns regarding molecular heterogeneity, limited sample size within certain subgroups—particularly Basal-like tumors—may constrain statistical power.Additionally, mechanistic studies exploring how MAM-related genes regulate immune checkpoint expression and chemotherapeutic response were not performed.Elucidating the spatiotemporal landscape of MAMs in the tumor microenvironment will demand a multimodal integration of cutting-edge techniques, particularly spatial transcriptomics, metabolomics, and functional studies spanning in vitro and in vivo models.

Conclusion

Our establishes a concise four-gene prognostic model derived from MAM-related genes, which enables reliable prediction of survival outcomes in breast cancer patients.Beyond prognostic value, the model provides mechanistic insights into how MAMs regulate tumor metabolism, immune modulation, and therapeutic resistance. Our study provides a hypothesis-generating biological framework suggesting that MAM-associated pathways may influence therapeutic vulnerability, thereby supporting future exploration of combination strategies integrating MAM-related targets with conventional or immune-based treatments in breast cancer.

Author contributions

Jintao Cui: Conceptualization, Methodology, Formal Analysis, Investigation,Writing—Original Draft, Visualization,Validation. Bin Xu: Software, Data Curation, Writing—Review &amp; Editing,Visualization. Songfu Han:Writing—Review,Data Curation,Project Administration. Chaoyang Guo:Supervision, Project Administration, Writing—Review &amp; Editing. Limin Wei and Xinshuai Wang: Conceptualization, Resources, Supervision, Project Administration, Writing—Review &amp; Editing.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The data supporting the findings of this study are publicly available from multiple open-access databases. Bulk RNA sequencing data and corresponding clinical information were obtained from The Cancer Genome Atlas Breast Cancer cohort (TCGA-BRCA) via the UCSC Xena platform (https://xenabrowser.net). External validation and single-cell RNA sequencing datasets were retrieved from the Gene Expression Omnibus (GEO) under accession numbers GSE20711 and GSE255068 (https://www.ncbi.nlm.nih.gov/geo). Protein expression and immunohistochemistry data were obtained from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org). Immunophenoscore (IPS) data used for immunotherapy response prediction were downloaded from The Cancer Immunome Atlas (TCIA) (https://tcia.at). All data analyzed in this study are publicly available without restrictions.

Declarations

Ethics approval and consent to participate

Not applicable. This study was conducted using publicly available datasets, and no ethical approval was required. Not applicable. This study did not involve direct participation of human subjects or the collection of identifiable personal data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.