Exercise-related genes predicts overall survival and tumor immune microenvironment, and identifies the biological role of SLC52A2 in breast cancer Authors: Caiping Chen, Chao Han, Li Xue, Xiang Lu Published: 2026 ABSTRACT AbstractBackgroundBreast cancer (BRCA) is the most common malignant tumor in women, with prognoses differing markedly by molecular subtype. Exercise is known to be related to tumor occurrence and progression, yet its role in BRCA remains understudied. This research seeks to explore the prognostic and immunomodulatory functions of exercise-related genes (ERGs) in BRCA, and their links to molecular subtypes, so as to inform targeted and precision treatment.MethodsWe obtained BRCA gene expression profiles and clinical data from the TCGA-BRCA project. Using LASSO, random forest, gradient boosting machine (GBM), stepwise Cox regression (Stepcox), and extreme gradient boosting (XGBOOST) algorithms, we screened prognostic signature genes from the intersecting ERGs, determined by differential expression analysis and weighted gene co-expression network analysis (WGCNA), established and validated a novel risk prediction model. We further analyzed biological functions, mutation profiles, immune traits, and molecular subtype associations across different risk groups. Additionally, the prognostic, biological and immunological roles of the core risk biomarker in the exercise-related signature was evaluated.ResultsWGCNA identified 345 differentially expressed ERGs in BRCA, and five machine learning algorithms further screened 6 prognostic genes (SNX22, SLC52A2, TNFRSF18, NFE2, EZR, and EMP1). The low-risk group had significantly higher overall survival rate than the high-risk group, a difference driven by the functional roles of the two groups’ differentially expressed genes, which regulated the cell cycle and epithelial-mesenchymal transition (EMT). Additionally, the high- and low-risk groups showed distinct mutation and immunomodulatory features. Consensus clustering of prognosis between the groups linked these differences to BRCA molecular subtypes. As a core candidate factor, SLC52A2 contributed to BRCA progression and poor prognosis in patients by participating in multiple signaling pathways and regulating immune cells.ConclusionERGs-based prognostic signatures could distinguish clinical outcomes of BRCA patients and were correlated with immune regulation and mutation patterns. The ERGs-constructed risk scoring system and molecular subtypes have the potential to serve as preclinical biological and immunological markers, paving the way for clinical management and treatment of BRCA.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05321-5.Keywords: Breast cancer, Exercise, Prognosis, Immune response, Genetic mutation, Molecular cluster Introduction According to 2022 global statistics, female breast carcinoma (BRCA), with 2,308,897 newly diagnosed cases and 665,684 deaths, remained the second most prevalent malignancy and the fourth leading cause of tumor-related deaths in both sexes worldwide [1]. In addition to histological and clinicopathological staging, molecular subtypes based on the expression levels of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) further contribute to the prognosis management and clinical decision-making of BRCA patients. BRCA is categorized into distinct subtypes: luminal A (ER-positive and HER2-negative), luminal B (ER-positive and HER2-positive), triple-negative BRCA (TNBC, ER-negative, PR-negative and HER2-negative), and HER2-enriched (ER-negative, PR-negative and HER2-positive) [2]. Such molecular typing is conducive to formulating personalized chemotherapy, targeted therapy and endocrine therapy plans and affects the patients prognosis. Furthermore, emerging biomarkers and molecular signaling pathways have been increasingly identified as playing pivotal roles in BRCA onset and progression. Yet, the underlying mechanism of the disease remains incompletely elucidated. It is estimated that by 2050, the age-standardized incidence rate (ASIR) of female BRCA will reach 59.63 cases per 100,000 people, representing a 32.13% increase from 2019. Meanwhile, its age-standardized mortality rate (ASMR) is anticipated to rise to 16.42 cases per 100,000 people, a 4.69% uptick compared with 2019 [3]. Therefore, against the grim medical burden, sustained exploration of key genes driving BRCA progression is crucial for imporving the clinical management and treatment of patients with this malignancy. Exercise has emerged as a pivotal strategy for bolstering cancer prevention, improving treatment outcomes, and optimizing prognosis. Mounting evidence attests to the multifaceted benefits of physical activity for cancer patients, including reducing the incidence and recurrence risk of cancer, as well as augmenting the efficacy of surgery and adjuvant therapy. Notably, in comparison to inactive individuals, regular physical activity is associated with a risk reduction of up to 24% for colorectal cancer [4]. Exercise reduced intratumoral hypoxia, a major limiting factor in radiotherapy, by improving tumor perfusion and vascularization [5]. Additionally, preclinical studies have shown that exercise can modulate angiogenesis, specifically normalize tumor vasculature, and enhance drug delivery efficiency, thereby highlighting the potential of exercise to augment the anticancer efficacy of chemotherapeutic agents such as gemcitabine [6] and doxorubicin [7]. Several multi-omics analyses confirmed the significant role of exercise-related genes (ERGs) in the prognosis, immune regulation, molecular mechanisms and treatment response of patients with various cancers, including hepatocellular carcinoma [8], head and neck squamous cell carcinoma [9], and osteosarcoma [10]. Unfortunately, although the role of ERGs in tumorigenesis and progression has been partially characterized, their specific value, whether biological, immunological or clinical, in BRCA was scarcely reported. This study aimed to screen for key genes related to BRCA prognosis through three complementary analytical approaches: weighted gene co-expression network analysis (WGCNA), differential expression analysis, and univariate Cox regression analysis. Following the identification of these prognostic genes, we constructed a prognostic prediction model by employing five machine learning methods and comprehensively assessed its predictive performance. Based on the established prognostic model, we further conducted in-depth functional annotation and pathway enrichment analyses to investigate its correlations with immune function, tumor mutation burden (TMB), and treatment responses. Methods Transcription data and clinical information We retrieved RNA sequencing data for BRCA samples and corresponding clinical information from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). Transcripts per million (TPM)-normalized expression data from 1118 BRCA tissues and 113 adjacent normal tissues were included as the training cohort. Concurrently, clinical data from 1097 BRCA patients, encompassing age, sex, and pathological stage, were downloaded for survival analyses. Somatic mutation variation (SNV) data for 992 BRCA samples were obtained from the TCGA platform. Additionally, gene expression matrices and clinical information from three datasets (GSE20685, n = 327; GSE20711, n = 91; GSE42568, n = 104), all based on the GPL570 platform, were downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/). Three GEO datasets were integrated to establish an external validation cohort. Batch effect correction was conducted using the “sva” R package, and the datasets were merged into a single cohort using the “limma” R package. All datasets were subjected to consistent normalization before batch correction and integration. Candidate ERGs identified by differentially expressed analysis and WGCNA Currently, there is still a lack of cohort studies on exercise interventions, as well as transcriptomic sequencing data from such interventions that could provide us with ERGs. We obtained 6,908 ERGs from GeneCards (www.genecards.org) and previous reports [8, 9]. Expression differences of ERGs between tumor and normal tissues in the TCGA-BRCA project were compared by “limma” R package. The false discovery rate (FDR) less than 0.05 and the relative expression value (|log2 fold change|) more than 1 were taken as the threshold of significant genetic difference. Volcano map was drawn with “ggplot2” R package to show the difference analysis results of all ERGs, while box plots were used to represent the expression differences of core prognostic ERGs in tumor tissues and normal tissues. We utilized the “WGCNA” R package to identify highly correlated ERG clusters based on the TCGA-BRCA dataset. The function “pickSoftThreshold” was employed to select an appropriate soft threshold, ensuring that the constructed gene co-expression network conformed to scale-free properties. The cutHeight parameter was set to 0.95, and a clustering algorithm was applied to delineate distinct gene modules. A cluster dendrogram was generated based on the calculation of gene adjacency and similarity. The dynamic tree cutting algorithm was then used to further segment modules and merge analogous modules. Each color corresponds to a unique module, with each module containing genes exhibiting similar expression patterns. Pearson correlation analysis was performed to assess the association between each gene module and sample traits, and several modules with higher absolute correlation values were selected for subsequent analyses. Candidate ERGs for further investigation were obtained by taking the intersection of the gene modules most strongly positively or negatively correlated with tumorigenesis and the set of differentially expressed genes (DEGs). Function enrichment analyses The R software package “clusterProfiler” was utilized to carry out enrichment analyses on the candidate ERGs, including Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG). GO analyses were utilized to understand their cellular components (CC), molecular functions (MF), and biological processes (BP), while KEGG was performed to explore the pathways in which these genes may participate. Identification of prognostic ERGs by integrated machine learning methods First, univariate Cox regression analysis was utilized to screen ERGs that were associated with OS using the “survival” R package (P < 0.05). Then, to explore the hub ERGs, LASSO Cox regression, random forest technique (RF), gradient boosting machine (GBM), extreme gradient boosting (XGBoost) and Stepwise Cox (StepCox) regression were used. The ten-fold cross-validated LASSO-Cox regression reduces the number of variables in the model and avoids overfitting and covariance compared to the traditional Cox regression model. The selection of genes corresponding to the min(lamda) is clinically valuable for subsequent model construction. A GBM model that aggregated multiple predictors was trained to predict the likelihood of survival with decision-tree-base-learners using the “gbm” R package. XGBoost, an improved gradient boosting algorithm, can train models faster and more efficiently using the “XGBoost” R package. The basic element of the XGBoost algorithm is a decision tree, and the generation of the latter decision tree considers the prediction results of the previous decision tree. Through the XGBoost and GBM algorithms, we chose the fifteen most influential genes by ranking the genes based on their relevance, respectively. Similarly, the random forest algorithm ranked the genes based on their relevance and subsequently chose the genes corresponding to the number of trees with the smallest error rate for model construction via the “randomforest” R package. StepCox algorithm to improve the accuracy and reliability of prognostic signature using “StepReg” R package. Finally, the prognostic signature, constructed by the intersecting genes among the five machine learning methods, was established as follows: risk score = ∑ni = 1(Coef×Ni), where Coef referred to the coefficient of gene i, derived from multivariate Cox regression and Ni represented the expression value of gene i. We randomly divided the patients in the TCGA-BRCA cohort into the training set (n = 727) and the test set (n = 311) in a 7:3 ratio, while the GSE20685, GSE20711 and GSE42568 cohorts were merged as the external validation set. According to the median critical value of the risk score, the BRCA patients were divided into the high-risk group and the low-risk group. Kaplan-Meier curves were generated to visualize the overall survival (OS) of patients in the three sets stratified by high- and low-risk groups. In addition, disease-free survival(DFS), disease-specific survival (DSS), progression-free survival (PFS), together with OS, were used to evaluate the prognostic role of the core genes in exercise-related signature. Construction and validation of predictive model based on prognostic signature ERGs-based prognostic signature, along with age, gender, T stage, N stage, and M stage, were included in univariate and multivariate Cox regression analyses to determine independent risk factors for the prognosis of BRCA patients in the term of OS. To further improve the clinical value, a nomogram was constructed based on the independent risk factors to predict the 1-, 3-, and 5-year OS of BRCA patients in TCGA dataset using the “rms” R package. ROC curves and calibration curves comparing the predicted and observed OS chances were applied to evaluate the prognostic performance of the nomogram. Decision curve analysis (DCA) was conducted to determine the clinical usefulness of the nomogram by quantifying the net benefits at different threshold probabilities. Exploring potential mechanisms of hub ERGs in BRCA The GeneMANIA database (http://genemania.org) was employed to explore the interactions between hub ERGs and their associated genes, as well as the biological functions involved. Gene Set Enrichment Analysis (GSEA) was conducted to compare pathway differences between the high- and low-risk groups using the “clusterProfiler” R package, with the use of the annotated gene sets (“h. all. v2024.1. Hs. symbols.gmt”, “c2.cp.kegg.v7.4.symbols.gmt”, and “c5.go.v7.4.symbols.gmt”) from the MSigDB database (https://www.gsea-msigdb.org). Gene set permutations were performed 1000 times for each analysis to obtain a normalized enrichment score (NES), which was used for sorting pathways enriched in each phenotype. Gene sets with adjusted p-value < 0.05 were considered as significant. Additionally, DEGs between the two groups were identified using the “limma” R package, adhering to the criteria of an adjusted p-value < 0.05 and an absolute log2 fold change > 1. Gene set variation analysis is another enrichment tool to evaluate the correlation between important pathways and the hub ERGs. Immunological analyses Firstly, we integrated multiple algorithms, including MCPCOUNTER, TIMER, CIBERSORT, XCELL, QUANTISEQ, and EPIC, to evaluate the association between exercise-related risk scores and tumor immune infiltration cells (TIICs), and Spearman’s correlation analysis was performed for this purpose. Subsequently, the single-sample GSEA (ssGSEA) algorithm, implemented via the “GSEABase” and “GSVA” R packages, was utilized to assess immune cell infiltration in tumor samples from the TCGA-BRCA dataset. This approach yielded results regarding immune cell presence and immune system function for each sample. Differential box plots were generated using the “ggpubr” and “reshape2” R packages to visualize score differences between the high-risk and low-risk groups. Additionally, the “ESTIMATE” R package was employed to calculate stromal and immune scores within tumor tissues. These scores were then visualized across risk groups using box plots created with the “ggpubr” R package. The expression profiles of various immune checkpoint genes were analyzed across different risk subgroups. Finally, the analyses of TIIC and tumor immune microenvironment (TIME) were also applied to core prognostic ERGs. Somatic mutation analysis Single nucleotide variation (SNV) data were downloaded from TCGA database to investigate whether differences in gene mutation profiles exist between high-risk and low-risk groups. We utilized the “maftools” R package to analyze somatic variant data from TCGA-BRCA samples, which were stored in mutation annotation format (MAF). The tumor mutation burden (TMB) score for each patient was calculated using the formula: (total number of mutations/total number of covered bases) × 10⁶. A waterfall plot was generated to visualize the mutated genes and their variant frequencies. Consensus cluster analysis for BRCA based on ERGs Using the six hub ERGs, we performed consensus clustering on the TCGA-BRCA cohort with the “ConsensusClusterPlus” R package. We tested k = 2 to 10 to find the optimal cluster number, repeating the process 500 times, and selected the best k via the cumulative distribution function. Differences in survival, TIME, and immune checkpoints among subgroups were evaluated using our previous methodology. A heatmap showed the immune landscape of BRCA patients across different clusters. Response to drugs and immunotherapy The Immunophenoscore (IPS) quantifies tumor immunogenicity on a 0–10 scale. A higher IPS suggests improved prognosis and response to immunotherapy. To forecast responsiveness to ICIs therapy, IPS data from TCGA-BRCA patients treated with anti-PD-1/anti-CTLA-4 was retrieved from TCIA (tcia.at/home). IPS levels were then compared between high- and low- risk groups. The sensitivity of various drugs was predicted in patients between two risk groups. The “oncoPredict” R package was utilized for drug prediction. Wilcoxon signed-rank test was utilized to explore the difference in half-maximal inhibitory concentration (IC50) between different risk groups. The results were performed by using the R package “ggplot2”. Cell culture and small RNA interference Human breast cancer cell lines MCF7 and T47D were used in this study. Detailed procedures for cell culture, small RNA interference (siRNA) transfection, and knockdown efficiency verification are described in Supplementary Materials. Proliferation, colony formation and transwell invasion assays Human breast cancer cell lines MCF7 and T47D were used in this study. Detailed procedures for cell culture, small RNA interference (siRNA) transfection, and knockdown efficiency verification are described in Supplementary Materials. Statistical analysis All statistical analyses were performed using R 4.4.2 software. The correlation matrices were conducted using Pearson or Spearman correlation. Wilcoxon test was conducted for the comparisons between the two groups. Survival differences were compared using K-M curves with a Log-rank test. P-value < 0.05 was considered statistical significance. Results Identification of candidate ERGs via WGCNA and differential expression analysis First, 6908 ERGs were included in WGCNA. We constructed a co-expression network and set the soft-thresholding power (β) to 11 when the scale-free topology fitting index reach 0.9 (Fig. 1A). Using the “merge dynamics” algorithm, we then obtained nine modules (Fig. 1B). A heatmap showed the correlation and significance of module genes between tumor and normal tissues (Fig. 1C). Among them, genes in the blue (n = 843, cor = 0.28, p < 0.001), brown (n = 485, cor = 0.49, p < 0.001), and green (n = 360, cor = 0.06, p < 0.001) modules were positively correlated with tumorigenesis (Fig. 1D, F). Conversely, pink (n = 786, cor = − 0.52, p < 0.001) and red (n = 360, cor = − 0.68, p < 0.001) module genes were negatively correlated with tumorigenesis (Fig. 1G, H). Identification of candidate ERGs. A Scale independence and average connectivity. B Cluster dendrogram. C Heatmap of the correlation between BRCA risk and ERGs. D–H Scatter plot of brown (D), green (E), blue (F), pink (G), and red (H) modules. Each point in the scatter plot represents a gene. I volcano map of DEGs in the TCGA-BRCA cohort. J Venn diagram of DEGs and modules genes. J, K GO (J) and KEGG (K) enrichment analysis of the green module. ERG, exercise-related gene. BRCA, breast cancer. DEG, differentially expressed gene. TCGA, The Cancer Genome Atlas. GO, Gene Ontology. KEGG, Kyoto Encyclopedia of Genes and Genomes As the volcano map shown (Fig. 1I), differential expression analysis was performed on 6,908 ERGs using criteria of | log2 fold change | > 1 and FDR < 0.05, yielding 365 upregulated and 834 downregulated genes in BRCA tissues. Intersection of these upregulated genes with those associated with the WGCNA positive module identified 171 candidate ERGs, while intersection of the downregulated genes with those related to the negative module identified 174 candidate ERGs (Fig. 1J). To further understand the biological roles of these genes in BRCA, we conducted a functional enrichment analysis. GO enrichment analysis revealed that these genes were involved in biological processes such as nuclear division, regulation of cell cycle phase transition, chromosome segregation, DNA replication, etc. (Fig. 1K). KEGG enrichment analysis not only confirmed their close association with the cell cycle but also identified enrichment in pathways such as the P53 signaling pathway, cellular senescence, microRNAs in cancer, cellular senescence, etc. (Fig. 1L). These results suggested that the candidate genes screened via WGCNA and differential expression analysis played critical roles in the carcinogenesis and progression of BRCA. Prognostic ERGs screened via multiple machine learning methods Univariate Cox regression analysis was used to identify the risk and protective genes among candidate ERGs. Through the analysis of the OS for the TCGA-BRCA cohort, 16 protective genes (SOX10, SNX22, SLC16A11, TSLP, APOD, S100B, TP63, TNFRSF18, KCNH8, NRG1, FBLN5, THSD7B, NFE2, HSD11B1, CCND2, and EGR2) and 9 risk genes (PLK1, SLC7A5, SLC52A2, GGCT, ULBP1, ALPK3, ACSL1, EZR, and EMP1) were confirmed their prognostic potentials (Fig. 2A). Furthermore, five machine learning algorithms were integrated to obtain the hub prognostic ERGs. RF identified 14 prognostic genes as the key markers with importance (Fig. 2B, C). LASSO analysis narrowed down the candidate ERGs and 20 genes with optimal λ values were screened (Fig. 2D, E). GBM and XGBoost algorithms each screened the top 15 characteristic genes as prognostic markers (Fig. 2F, G). StepCox regression analysis defined five protective genes and four risk factors (Fig. 2H). To improve the stability and robustness of feature selection, we integrated the results from five machine learning algorithms and adopted an intersection strategy to identify consensus prognostic genes. This approach can effectively reduce algorithm-specific bias and improve the reliability of the final gene signature. Accordingly, six overlapping prognostic genes were identified from the intersection of the five methods(Fig. 2I), including SNX22, SLC52A2, TNFRSF18, NFE2, EZR, and EMP1(survival coefficient, − 0.567, 2.620, − 0.937, − 0.582, 5.309, 3.974, respectively). We obtained the formula for the survival risk of prognosis in patients with BRCA: risk score= (− 0.567 × SNX22) + (2.620 × SLC52A2) + (− 0.937 × TNFRSF18) + (− 0.582 × NFE2) + (5.309 × EZR) + (3.974 × EMP1). According to the median risk score, the patients in TCGA-BRCA cohort were divided into low- and high-risk subgroups. Kaplan-Meier survival curves were plotted, revealing that high-risk individuals had a significantly unfavorable prognosis than low-risk counterparts (Fig. 2J). The heat map showed the expression distribution of six genes across the high and low risk groups of the TCGA-BRCA samples (Fig. 2K). In this prognostic map, the risk-associated genes (SLC52A2, EZR, and EMP1) were upregulated in the high-risk group, whereas the protective genes (SNX22, TNFRSF18, and NFE2) exhibited downregulation therein. Identification of the prognostic biomarkers from the integrated machine learning methods. A Univariate forest plot showing the prognostic biomarkers among candidate ERGs. B Correlations between error rate and classification trees in the random forest algorithm. C The relative variable importance of prognosis-associated genes. D Selection of the λ in LASSO regression analysis via 10-fold cross-validation. E The coefficient profiles of the expression of 17 variables in LASSO method. F Barplot showing the top 15 important genes in GBM algorithm. G Barplot showing the four risk genes and five protective genes via Stepwise Cox regression analysis. H Lollipop chart showing the top 15 important genes in XGBoost algorithm. I Venn diagram of five machine learning methods. J The Kaplan-Meier survival curves for the high- and low- risk groups in the TCGA-BRCA cohort. K Heatmap showing the expression of six prognostic biomarkers. LASSO, least absolute shrinkage and selection operator Exercise-related prognostic signature was the independent risk factor for OS of BRCA patients To further validate the prognostic value of exercise-related signature, the TCGA-BRCA cohort was randomly split into a training set and a test set, while GSE20685, GSE20711, and GSE42568 cohorts were incorporated as an external validation set. Comparison of OS between the high- and low- risk groups across these three sets revealed that patients in the high-risk group had significantly poorer prognosis than their low-risk counterparts (Fig. 3A–C). Setting the median risk score as the threshold and plotting the survival status across three sets revealed that nearly all high-risk BRCA patients died, further demonstrating the stability of our exercise-related prognostic signature (Fig. 3D–F). Moreover, the results of univariate and multivariate Cox regression analyses of the age, gender, clinicopathological stage, and this signature risk score, are shown in forest plots. Age > 60 years (HR = 1.035, 95%CI 1.019–1.050, p < 0.001), T4 stage (HR = 2.242, 95% CI 1.049–4.792 p = 0.037), N3 stage (HR = 2.789, 95% CI 1.319–5.896, p = 0.007), M1 stage (HR = 2.893, 95% CI 1.462–5.723, p = 0.002), along with high risk score of this newly identified signature (HR = 4.472, 95% CI 1.580–14.125, p = 0.005) were included into multivariate Cox analysis and were identified as independent prognostic risk factors (Fig. 3G, H). These results confirmed the prognostic value of exercise-related signature in BRCA patients. Construction and validation of an six-gene prognostic signature. A–C The Kaplan-Meier survival curves for the LS and HS groups in the TCGA training set (A), test set (B) and external validation set (C). D–F Risk score (bottom) and survival time (top) distribution of the TCGA training set (D), test set (E) and external validation set (F). Patients from the three sets were stratified into two groups according to the median value of risk scores. G, H Forest plot of univariate (G) and multivariate (H) Cox regression analysis of risk score and clinical characteristics using the TCGA-BRCA cohort. HR and P-values were displayed The performance and stability of the predictive model based on exercise-related signature Based on the above-mentioned independent risk factors, we developed a novel prognostic model, providing a reliable and quantifiable method for predicting the progression of BRCA. The final nomogram scores of each patient obtained by combining all items can be used to predict 1-, 3-, and 5-year OS rates (Fig. 4A). The performance of the model was verified across training set (AUC = 0.808), test set (AUC = 0.791), and external validation set (AUC = 0.784) cohorts via Receiver Operating Characteristic (ROC) curve analysis for 5 years OS, demonstrating that this model predicted patients’ prognosis with considerable accuracy (Fig. 4B–D). The calibration curves of the three sets all indicated that the predictive model had a good fit with the real cases (Fig. 4E–G), while DCA elicited that the nomogram provided a significant net benefit (Fig. 4H–J). Overall, these results confirm that the nomogram holds marked predictive value. Establishment of a novel predictive model for OS of BRCA patients. A Nomogram constructed based on exercise-related risk score, age, T stage, N stage and M stage as predictive factors to predict 1-, 3-, and 5-year survival probability. B–D ROC curves for the training set (B), test set (C) and external validation set (D). E–G Calibration curves for the survival probability at 1, 3, and 5 years of patients in the training set (E), test set (F) and external validation set (G). I–K DCA curves the training set (I), test set (J) and external validation set (K). OS, overall survival. ROC, receiver operating characteristic. DCA, decision curev analysis Biological function and potential mechanism underlying the hub ERGs To explore the regulatory links of the six hub ERGs with other genes in BRCA, we analyzed PPI networks from the GeneMANIA database. These networks showed co-localization, co-expression, and physical interactions between the hub ERGs and other genes, including ATG3, WNT3A, and WNT8B (Fig. 5A). Next, gene sets downloaded from the MSigDB database, including “h.all.v2024.1.Hs.symbols.gmt”, “c2.cp.kegg.v7.4. symbols.gmt”, and “c5.go.v7.4.symbols.gmt”, were used in GSEA. The Hallmarks-GSEA results showed that epithelial mesenchymal transition (EMT), myogenesis, UV response down, TNFalpha signaling via NFkb and hypoxia were mainly enriched in the high-risk group (Fig. 5B). The KEGG-GSEA analysis revealed that the upregulated DEGs between the high- and low-risk groups are enriched in pathways including focal adhesion, regulation of actin cytoskeleton, pathways in cancer, tight junction, ECM receptor interaction (Fig. 5C). GO-GSEA analysis showed that the top five functions were contractile fiber, cell substrate junction, external encapsulating structure organization, collagen containing extracellular matrix, and extracellular matrix structural constituent (Fig. 5D). These results suggested that the gene expression patterns in high-risk group samples were associated with the construction of extracellular matrix and EMT, which might be the biological basis leading to disease progression in high-risk group patients. Biological function of the six-gene prognostic signature. A Gene-gene interaction network from GeneMania database. B GSEA for enriched pathways between the high- and low- risk group. C Cluster plot for KEGG enrichment analysis. D Bar plot showing GO enrichment analysis. GSEA, gene set enrichment analysis Difference of tumor microenvironment and mutation landscape between risk groups Next, the study proceeded to assess the connection between exercise-related signature and the Tumor Immune Microenvironment (TIME) in the TCGA-BRCA cohort. Using the MCPCOUNTER, TIMER, CIBERSORT, XCELL, QUANTISEQ, and EPIC methods, a positive correlation was identified between most immune cells, including key effector cells, and exercise-related risk scores (Fig. 6A). Furthermore, ssGSEA results revealed heightened immune-related activities in the high-risk category, such as APC co-stimulation, CCR, and type II IFN response. However, the low-risk group had relatively stronger functions in cytolytic activity, human leukocyte antigens, and inflammation-promotion(Fig. 6B). The samples in high-risk group showed a notable presence of various immune cells, such as macrophages, mast cells, neutrophils, and regulatory T cells (Fig. 6C). The examination of the ESTIMATE algorithm confirmed the existence of increased ESTIMATE and stromal scores in the high-risk group (Fig. 6D–F). Finally, analysis indicated that immune checkpoints were more differently expressed between the high- and low-risk groups (Fig. 6G), suggesting potential responsiveness to immune checkpoint blockade (ICB) treatments in these patients. Association of the exercise-related with the immune microenvironment in the TCGA-BRCA cohort. A Bubble plot of correlation coefficients between immune cells and risk scores. B Box plot showing differences in immune-related functions between the high- and low-risk groups using the ssGSEA algorithm. C Box plot showing differences in immune cells between the high- and low-risk groups using the ssGSEA algorithm. D–F Immune score (D), stromal score (E) and ESTIMATE score (F) in the two risk subgroups. G Differences in the expression of immune checkpoints across the risk subgroups. ssGSEA, single sample gene set enrichment analysis. *P < 0.05, **P < 0.01 and ***P < 0.001 Somatic mutation data for 992 BRCA samples from TCGA database were extracted by “maftools” method. We utilized waterfall plot to exhibit high-frequency mutated genes in BRCA, such as TP53 (34%), PIK3CA (34%), CDH1(13%), TTN (17%), and MUC16 (10%) (Fig. S1A). We annotated 15,413 protein-coding genes with somatic mutations that mostly consist of missense mutations (Fig. S1B, C), and single-nucleotide polymorphism (SNP) was more common than insertion, deletion, and oligonucleotide polymorphisms (ONPs) (Fig. S1D). In addition, among the six combinations of mutation types (i.e., T > G, T > C, T > A, C > T, C > G and C > A), C > T was the most common (Fig. S1E, F). The somatic mutation rates varied considerably among the samples, though an average of 34 mutations and maximum of 3783 mutations occurred in BRCA samples and variant types were also shown via box plot (Fig. S1G). The co-occurence and mutual exclusion among top 25 mutated genes were also displayed (Fig. S1H, I, and J). To further explore the possible mechanism of exercise-related prognostic signature, we divided the somatic mutation data of TCGA-BRCA into two groups according to the median risk score (Fig. S1J and K). The waterfall plots of the top 20 mutated genes in the two risk groups indicated that the profile of highly mutated genes was generally similar between the two groups, which was demonstrated by no obvious difference of TMB between two risk groups (Fig. S1L, M). These results suggested that There was no evidence to confirm the correlation between the predictive model and tumor mutations in BRCA. Response prediction of risk signature to immunotherapy and drug Subsequently, to investigate the immunotherapy responses of BRCA patients in different risk groups, we downloaded the IPS scores of the TCGA-BRCA cohort from the TCIA database. The results showed that among the CTLA4(−)/PD1(−), CTLA4(−)/PD1(+), CTLA4(+)/PD1(−), and CTLA4(+)/PD1(+) subgroups, patients with low-risk BRCA exhibited higher IPS, indicating a better response to immunotherapy (Fig. 7A–D). On the other hand, to predict the sensitivity of several common drugs, drug analysis was conducted for high- and low-risk BRCA (Fig. 7E). ABT737, Dabrafenib, Elephantin, Leflunomide, Navitoclax, Pevonedistat, Topotecan, Sabutoclax, VE821, Venetoclax, XAV939 showed higher sensitivity in low-risk BRCA. Moreover, Acetalax, Alpelisib, AZD1332, Cediranib, Dasatinib, Ibrutinib, JAK8517, LCL161, Osimertinib, NU7441, Luminespib, Sapitinib, Staurosporine, Taselisib, and UMI-77 exhibited lower sensitivity in low-risk patients than in high-risk patients. Immune therapy response and drug sensitivity. A–D Group comparison of IPS between groups of different BRCA disease subtypes: CTLA4(−)/PD1(−) (A), CTLA4(−)/PD1(+) (B), CTLA4(+)/PD1(−) (C), CTLA4(+)/PD1(+) (D). E Boxplots showing the difference of IC50 between risk groups. CTLA4, cytotoxic T lymphocyte-associated antigen-4. PD1, programmed death-1. *P < 0.05, **P < 0.01 and ***P < 0.001 Identification of three senescence clusters Unsupervised consensus clustering was conducted to explore molecular subtypes based on the expression of ERGs comprising the risk signature. Consensus clustering results showed that there was a significant difference when k = 3 with a curve of a gentle slope (Fig. 8A, B, and C) and patients in TCGA-BRCA cohort were grouped into three clusters (Fig. 8D). The PCA plots (Fig. 8E), t-SNE maps (Fig. 8F), and Sankey diagram (Fig. 8G) revealed a clear association between clusters and risk groups. Cluster 2 (C2) and Cluster 3 (C3) are the main components of the low-risk group, whereas Cluster 1 (C1) accounts for a large proportion of the high-risk group. Subsequent survival analysis illustrated those patients in the C2 group had the most favorable prognosis, while patients in the C1 group had the worst clinical outcomes in the term of OS and PFS (Fig. 8H, I). The TME scores of the different clusters were subsequently calculated, revealing that cluster C1 had the highest stromal score among all clusters, while cluster C3 exhibited the highest immune score (Fig. 8J–L). Using the MCPcounter algorithm, we calculated the scores of 10 immune cell types in each sample. We found that in C3, which is associated with improved prognosis, the infiltration levels of T cells, neutrophils, myeloid dendritic cells, CD8 + T cells, B cells, and cytotoxic lymphocytes were significantly higher than those in the other two clusters (Fig. 8M–R). In addition, applying the immune checkpoints inhibitors analysis, it was identified that C3 cluster was significantly correlated with LAG3, CTLA4, PDCD1, and so on, suggesting that patients involved in the C3 group might benefit from immunotherapy, especially anti-PD-1 receptor blockers (Fig. 8S). Consensus Clustering Analysis for six-gene prognostic signature. A–C CDF plot (A), tracking plot (B) and Delta plot (C) of concordance Clustering analysis. D Plot of consensus Clustering results for BRCA. E–G PCA plot (E), tSNE (F) and sankey diagram of the connection between clusters and high-and low-risk group. G, H The Kaplan-Meier survival curves for OS (G) and PFS (H) based on the three clusters. J–L Immune score (J), stromal score (K) and ESTIMATE score (L) in the three clusters. M Heatmap showing the immune infiltrating cells in the three clusters. N Differences in the expression of immune checkpoints across the three clusters. CDF, cumulative distribution function. PCA, principal component analysis. *P < 0.05, **P < 0.01 and ***P < 0.001 Prognostic, biological and immunological analyses of SLC52A2 In this study, three protective ERGs and three risk ERGs collectively constituted the exercise-related signature, with the expression levels of the three risk genes accounting for a relatively high weight in the risk score calculation formula. While EZR and EMP1 were fully validated to exert oncogenic effects in previous studies [11, 12], the role of SLC52A2, a core prognostic marker, remains poorly understood. By analyzing the expression pattern and prognostic value of SLC52A2 in the TCGA-BRCA cohort, we observed that SLC52A2 was significantly upregulated in BRCA tissues (Fig. 9A, B). Using its optimal cut-off value as the threshold for grouping, patients in the SLC52A2 high-expression group exhibited less favorable clinical outcomes compared to those in the low-expression group, as evidenced by the comparisons of DFS (Fig. 9C), DSS (Fig. 9D), OS (Fig. 9E), and PFS (Fig. 9F). This finding suggests that SLC52A2 exerts an oncogenic effect in BRCA. Further GSVA revealed that the upregulation of SLC52A2 expression was associated with the activation of multiple signaling pathways, including UV response down, androgen response, protein secretion, TGF-βsignaling, and EMT (Fig. 9G). Additionally, ESTIMATE analysis demonstrated that samples from the SLC52A2 high-expression group had elevated stromal scores and ESTIMATE scores (Fig. 9H). Moreover, SLC52A2 expression showed strong correlations with the infiltration levels of follicular helper T cells, regulatory T cells, and M0/M1 macrophages (Fig. 9I). These results suggested that SLC52A2 exhibited prognostic and immunomodulatory roles in BRCA, and might have potential pro-tumor effects. Elevated SLC52A2 served as an oncogene in BRCA. A, B Box plots showing the increased expression of SLC52A2 in BRCA compared to the normal tissues. C–F The Kaplan-Meier survival curves showed that high-SLC52A2 patients with BRCA had poor prognosis in term of DFS (C), DSS (D), OS (E), and PFS (F). G Barplor showing the SLC52A2-related biological function via GSVA. H violin plots showing the immune score, stromal score and ESTIMATE score between the high- and low- SLC53A2 group. I The mantel test showing the correlation between SLC52A2 expression and tumor immune infiltration cells in BRCA. DFS, disease free survival. DSS, disease specific survival. PFS, progression free survival. GSVA, gene set variation analysis. ***P < 0.001 Silencing SLC52A2 surpressed the proliferation, invasion, and migration of breast cancer cells in vitro To investigate the effect of SLC52A2 on the biological behaviors of breast cancer cells, MCF7 and T47D cells were transfected with SLC52A2-targeting siRNAs or negative control siRNA (NC). The siRNA with the optimal transfection efficiency was selected for subsequent detection of cell proliferation, colony formation, invasion and migration abilities. The CCK-8 assay results indicated that the OD values at 72 h in the siSLC52A2 groups were significantly higher than those in the NC group(Fig. 10A), suggesting that SLC52A2 inhibition slow down the growth rate of breast cancer cells. In the colony formation assay, the number of colonies formed in the siSLC52A2 groups was significantly increased compared with the NC group (P < 0.001, Fig. 10B), and the colonies were larger in size, indicating that suppressing SLC52A2 expression impaired the clonogenic potential of breast cancer cells. Additionally, the Transwell invasion assay demonstrated that the number of invaded cells in the siSLC52A2 groups was significantly more than that in the NC group(P < 0.001, Fig. 10C). Collectively, these results confirm that inhibition of SLC52A2 supressed the proliferation, invasion, and migration of MCF7 and T47D breast cancer cells. Decreased SLC52A2 expression inhibited proliferation, clony formation, invasion and migration of breast cancer cells. A CCK-8 assays. B Representative images of clony formation assays. C Representative images of Transwell assays. ***P < 0.001 Discussion BRCA is a malignant disease characterized by significant tumor heterogeneity. Traditional prognostic indicators, such as clinicopathological stage, tumor size, and histological grade, are no longer sufficient to precisely guide the treatment and management of patients, particularly those diagnosed at an early stage [13]. Currently, ER, PR, and HER2 status are used for molecular classification, serving as supplementary bases for risk stratification and treatment selection. The complex molecular landscape of BRCA forms the core of clinical decision-making, enabling tailored treatment strategies and improved oncological outcomes. Building on this framework, additional biomarkers, including Ki67 [14], TP53 [15], BRCA1/2 [16], PTEN [17], and CHEK2 [18], have been gradually integrated into the clinical decision-making process. While these markers provide valuable evidence to support personalized therapy, further research is needed to deepen our understanding of BRCA’s intricate molecular characteristics. Exercise, as a non-pharmacological intervention and preventive measure, helps to reduce the risk of cancer. The risk of cancer could be reduced by exercising at least 3–5 h per week, with women benefiting from appropriate high-intensity activities, such as heavy housework and dancing, which are considered to be effective in lowering the risk of BRCA [19]. During cancer treatment, physical exercise can alleviate the toxic side effects related to the treatment and enhance the efficacy of the supplementary treatment [20]. Exercise exerts a notable influence on the survival and prognosis of cancer patients. A prospective study of 6295 BRCA patients demonstrated that those who survived for more than five years and engaged in 4.9 to 17.4 metabolic equivalent tasks (METs) per week, particularly those with an exercise volume of ≥ 17.4 METs, exhibited a reduction in all-cause mortality as physical activity levels increased [21]. Multiple meta-analyses have consistently demonstrated an inverse relationship between post-diagnosis physical activity levels in BRCA patients and both cancer-specific and all-cause mortality [22, 23]. These compelling findings directed our attention toward elucidating the molecular basis through which exercise mitigates the risk of BRCA and enhances patient prognosis. Therefore, identifying ERGs with prognostic value was of great importance to explore the molecular basis of exercise in BRCA. In this study, we initially employed WGCNA and differential expression profiling on ERGs to pinpoint module genes exhibiting the strongest association with BRCA pathogenesis. Next, leveraging univariate Cox regression in conjunction with an integrative framework of multiple machine learning algorithms, we further delineated exercise-related prognostic signature genes for BRCA patients, culminating in the establishment and validation of a prognostic model with robust high performance. Although this procedure might exclude some potentially critical genes, it help improve the stability of the model and reduced bias introduced by individual methods. These results enabled us to establish a connection between exercise and the molecular characteristics of the prognosis of BRCA patients. We analyzed the differentially expressed ERGs in tumor samples and found that they were significantly enriched in biological processes associated with cell growth. Furthermore, enrichment analysis of the DEGs between the high-risk and low-risk groups uncovered notable differences in EMT potential between the two cohorts. Orange et al. treated human colon cancer(CC) cells with serum collected immediately after acute exercise and observed a significant reduction in cell proliferation compared with the control group [24]. Another randomized controlled trial confirmed that a single session of resistance training (RT) or high-intensity interval training (HIIT) can inhibit the in vitro growth of CC cells, potentially mitigating recurrence risk [25]. A study on prostate cancer indicates that tumor cells harbor exercise-sensitive genes, which may be involved in exercise-mediated regulation of malignant cell growth and apoptosis [26]. Moderate-intensity constant-load exercise markedly suppresses lung tumor growth, and lowered MMP9 levels in lung cancer tissues, potentially inhibiting EMT and controlling tumor metastasis to a certain degree [27]. Further investigations revealed that patients stratified with high-risk scores exhibited compromised immune regulatory capacity. Consensus clustering analyses unveiled potential interrelationships between molecular subtypes and the prognostic model in the context of survival trajectories and immune regulatory circuitry. Exercise could, to a certain extent, enhance the mobilization of white blood cells, particularly natural killer (NK) cells and gamma delta (γδ) T cells [28]. The heightened immune surveillance induced by exercise further improves immune function. Research by Pedersen et al. demonstrated that, in five distinct mouse tumor models, voluntary wheel-running exercise reduced tumor incidence and growth rate by approximately 60% via a mechanism involving catecholamine-induced natural killer cell mobilization and redistribution [29]. Specifically, blood T cells induced by acute exercise exhibit a stronger in vitro stimulatory response to autologous antigen-presenting cells pulsed with tumor-associated antigens [30]. Furthermore, exercise facilitates the retention of functional T cells, and an increase in maximal oxygen uptake is negatively correlated with T cell senescence [31]. Exercise-induced mobilization of immune cells and activation of immune surveillance partly underpin the correlation between ERGs and immune function modulation in BRCA, yet this mechanistic link awaits further experimental validation. Additionally, our study investigated the differential drug sensitivity. The beneficial effects of exercise on treatment response have been validated in several cohort studies. A subgroup analysis of 87 neoadjuvant chemotherapy patients from a randomized controlled trial by Sanft et al. showed that the pathological complete response (pCR) rate was significantly higher in the exercise intervention group than in the usual care group (53% vs. 28%), especially among patients with hormone receptor-positive/HER2-negative and triple-negative BRCA [32]. Among BRCA patients receiving neoadjuvant chemotherapy, supervised aerobic or resistance training promoted greater tumor shrinkage and pathological pCR, and helped maintain relative dose intensity in hormone receptor-negative cases [33]. Such clinical benefits may be attributed to improved drug delivery efficiency resulting from enhanced hemodynamics and vascular remodeling induced by exercise intervention administered concurrently with antineoplastic therapy [34]. Nevertheless, studies on the relationship between ERGs expression differences and treatment response in BRCA remain insufficient. Our findings may thus open a new avenue for identifying eligible populations and advancing precision medicine in the combined application of exercise intervention and antineoplastic therapy. SLC52A2, a key protein regulating riboflavin transmembrane transport, was previously thought to be associated with an increased risk of cardiovascular and neurodegenerative diseases due to its abnormal expression [35]. Research focusing on gastric cancer revealed that SLC52A2 was upregulated in tumor tissues, thereby offering a new perspective on its role in tumor progression [36]. Subsequent study further confirmed that abnormal SLC52A2 expression in colorectal cancer was linked to disease advancement [37]. The correlation of SLC52A2 expression with an increased proportion of stromal cells and the activation of EMT in BRCA provided potential clues for understanding how it might contribute to tumor progression and poor patient prognosis. Nevertheless, the specific mechanisms underlying these processes remain to be further elucidated through experimental investigations. Several inherent limitations in the present study demand attention. First, relying on public datasets, our findings’ generalizability hinged on validation in real population. Second, the functions and mechanisms of ERGs in BRCA progression were only supported by in silico analysis and in vitro experiments, lacking direct evidence from animal experiments. Further experimental studies are warranted to validate the expression patterns, biological functions, and clinical value of these genes in BRCA. Additionally, the mechanism by which ERGs affect BRCA patients’ treatment responses remains unclear. Conclusion In summary, we developed a molecular clustering system and a novel ERGs-based prognostic signature model for BRCA. This system demonstrated significant utility in enhancing the prognostic prediction of BRCA. It also facilitated risk stratification for BRCA patients, characterization of the TIME, and prediction of gene mutations. Additionally, we performed comprehensive functional and immunomodulatory analyses of ERGs, and specifically identified the potential dual role of SLC52A2 in tumor progression and immune regulation. These findings not only hold the potential to improve patients’ prognostic assessment but also deepen our understanding of the potential mechanisms by which ERGs are involved in BRCA pathogenesis and provide novel clues for the development of targeted therapeutic strategies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Author contributions All authors contributed to the study conception and design. Data curation was performed by Xiang Lu and Li Xue. Funding acquisition was performed by Xiang Lu and Caiping Chen. Writing – original draft was performed by Caiping Chen. Methodology, investigation, formal analysis was performed by Li Xue. Validation was performed by Caiping Chen. Visualization was performed by Chao Han. Conceptualization, project administration, resources, software, writing – review & editing was performed by Xiang Lu. All authors reviewed the manuscript and approved thefinal submitted version. Funding The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The Jiaxing Key Discipline of Medicine (Mastropathy, Innovation Subject, 2023-FC-001), and the Jiaxing Provinces and Cities Jointly Cultivate Discipline—General Surgery (Minimally Invasive, 2023-PYXK-001). Data availability Data is available upon reasonable request from the corresponding author. Declarations Ethics approval and consent to participate The cohort data used in this article were all from online databases and therefore did not involve ethics. Consent to participate is not applicable to this study. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement Data is available upon reasonable request from the corresponding author.