Identifying the relationship between exosome genes and breast cancer risk using bioinformatics and machine learning methods Authors: Xuwen Wang, Yongyong Ding, Shengyu Shi, Xiaomin Ji, Zekun Jiang, Tian Liu Published: 2026 ABSTRACT AbstractBackgroundBreast cancer (BC) is the most prevalent cancer among women globally, with a high mortality rate. The treatment and prevention of this disease are of utmost importance. Exosomes—nanovesicles derived from cells—play a vital role in intercellular communication and have profound implications for numerous physiological and pathological processes. However, relationships between exosomes and the development and prognosis of BC have not been fully elucidated.MethodsUsing gene expression data obtained from the NCBI GEO database for BC and normal tissues, we analyzed differential gene expression. The intersection of differentially expressed genes and exosome-related genes was obtained, and a gene interaction network was established. Genes associated with BC were identified using the Least Absolute Shrinkage and Selection Operator (LASSO), Support Vector Machine-Recursive Feature Elimination (SVM-RFE), and Random Forest (RF) algorithms. Shared genes identified using the three algorithms were used to construct a nomogram for predicting the risk and prognosis of BC. Gene–drug, gene–RNA binding protein, and gene–transcription factor interaction networks were analyzed. Finally, we evaluated relationships between drugs and proteins using molecular docking analyses.ResultsWe identified 595 differentially expressed genes and obtained 37 exosome-related genes. Five differentially expressed exosome-related genes associated with the risk of disease were identified using three machine learning methods. These genes were involved in the regulation of various biological processes and were associated with immune cell infiltration. Drugs targeting four of the five genes were identified.ConclusionExosome genes are related to the occurrence and prognosis of BC and can be used as targets for drug therapy.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05402-5.Keywords: Exosome, Breast cancer, Bioinformatics, Machine learning, Molecular docking analysis Introduction According to the World Health Organization, breast cancer (BC) is the most prevalent malignant tumor worldwide; it is a leading cause of cancer-related mortality among women globally, and its incidence is rising every year [1, 2]. In the United States, the incidence of BC is notably high. This high incidence may be due to environmental factors, such as air pollution [3] and genetic factors. For instance, individuals who carry mutations in BRCA1 and BRCA2 have a substantially elevated risk of developing BC throughout their lifetime [4]. The treatment and prognosis of BC vary depending on its molecular subtype. HER-2-positive BC is more aggressive with a poorer prognosis, and anti-HER-2 therapy represents a crucial treatment strategy for this subtype [5]. Additionally, the risk of recurrence and metastasis is a significant concern. Patients with BC continue to face a high risk of recurrence, necessitating particular attention during treatment and follow-up care [6]. The continuous emergence of novel technologies and methodologies has transformed the landscape of BC diagnosis and treatment substantially. A notable example is the combination of digital mammography and magnetic resonance imaging (MRI), which demonstrates exceptional sensitivity and specificity in the early detection of BC [7]. Moreover, patient-derived models have emerged as an invaluable research tool, enabling the accurate simulation of tumor characteristics under physiological conditions. These models facilitate the translation of promising laboratory research findings into practical clinical applications, thus accelerating the development of more effective BC treatments [8]. These advancements herald new horizons for early BC diagnosis and personalized treatment. Exosomes—small extracellular vesicles first identified in the 1980s—were discovered during the maturation process of reticulocytes when multivesicular bodies fuse with the plasma membrane [9, 10]. They facilitate intercellular communication and exert a profound influence on the tumor microenvironment [11, 12]. They are intricately involved in the modulation of signaling pathways and gene expression related to tumor growth and metastasis. Specific exosome components alter cellular signaling cascades involved in cancer progression and therapeutic responses. The key pathways implicated in BC are associated with the cell cycle, signal transduction, and gene expression regulation. Dysregulation of these pathways can result in augmented cell proliferation and survival, enabling tumor cells to evade apoptotic signals and contributing to disease progression [13, 14]. For instance, alterations in the PI3K/Akt/mTOR signaling pathway have been implicated in the development of BC, underscoring the significance of targeted therapeutic interventions aimed at these pivotal molecular pathways [15, 16]. Exploring these intricate biological mechanisms is crucial for the development of effective therapeutic strategies and for improving patient outcomes. Exosome gene expression has emerged as a pivotal area of research in the context of BC and its metastasis. Recent studies have revealed the pivotal role of exosomes in facilitating intercellular communication within the tumor microenvironment, thereby influencing cancer progression and metastasis [17, 18]. Expression patterns of exosome genes are correlated with prognosis in BC, supporting their value as predictive biomarkers. Changes in exosome gene profiles can mirror the dynamic alterations occurring within the tumor microenvironment as cancer progresses. For instance, specific exosome microRNAs regulate the expression of target genes implicated in vital signaling pathways associated with tumorigenesis and metastasis [19, 20]. Furthermore, the cargo carried by exosomes can influence the behavior of neighboring cells, fostering an environment conducive to cancer cell survival and dissemination [21, 22]. Exosome-mediated intercellular communication is a fundamental component of the regulatory networks that govern gene expression patterns in BC. The exosome-mediated interactions between tumor cells and their microenvironment contribute to the establishment of an intricate gene regulatory network that promotes tumor progression [23, 24]. Understanding these mechanisms can offer profound insights into the roles of exosome gene expression in BC, underscoring its potential as a therapeutic target and prognostic tool. The significance of exosome genes in the progression and development of BC, which is a multifaceted and heterogeneous condition characterized by the uncontrolled proliferation of breast cells, has garnered considerable attention. Exosome genes have the potential to serve as biomarkers for early diagnosis and prognostic evaluations. In this study, the detailed role of exosome-related genes in BC was evaluated through analyses of gene expression profiles. By leveraging a multifaceted approach that integrates gene expression analyses, functional enrichment analyses, and advanced machine learning techniques, we aimed to identify and validate promising biomarkers. This study reveals the crucial role of exosome genes in BC and supports their potential as biomarkers for diagnosis and prognosis. Delving deeper into this domain could pave the way for the development of refined diagnostic and treatment strategies. Methods Data acquisition Four sets of gene expression profile data pertinent to BC were retrieved from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/): GSE50428 (platform GPL13648, microarray), GSE71053 (platform GPL570, microarray), GSE139038 (platform GPL27630, microarray), and GSE205185 (platform GPL21185, microarray). In total, 90 BC samples (experimental or treatment group) and 46 control samples (control group) were included in this study. For data preprocessing, the R software package “limma” was employed to normalize each set of raw data obtained from the GEO. Subsequently, the “limma” and “sva” packages were utilized to concatenate the four datasets and perform batch correction, ensuring data consistency and comparability. The “ggplot2” and “ggpubr” packages were used to conduct a principal component analysis (PCA), facilitating the visualization of the combined data prior to correction and subsequent to batch correction. The R packages “limma,” “ggplot2,” and “pheatmap” were used to visually represent differentially expressed genes (DEG) between normal and cancer samples within the corrected and combined dataset. Values of P < 0.05 were considered statistically significant. To identify significant differences in gene expression between groups, we employed a threshold of LogFC (logarithmic fold change) > 1 and adjusted P-value < 0.05. The adjusted P-values were calculated using the Benjamini–Hochberg false discovery rate (FDR) correction method. We obtained exosome-related genes from the GeneCards database (https://www.genecards.org/) [25]. Using “exosome” as the keyword, we applied stringent screening criteria, setting “score > 1.00” to pinpoint relevant, high-quality exosome-related genes. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses We obtained the intersection of DEGs and exosome-related genes for comprehensive GO and KEGG pathway enrichment analyses. Intersecting genes were ranked by -log10(adjusted-P) × log₂(fold-change) for GO/KEGG over-representation analysis. GO enrichment analyses provide a broad spectrum of information, allowing for the systematic categorization and annotation of genes. In particular, functional enrichment was evaluated in the biological processes (BP), molecular functions (MF), and cellular components (CC) categories [26–28]. The KEGG database includes data for complex signaling pathways involving numerous genes. The intersecting genes were evaluated through a KEGG enrichment analysis to identify key cellular processes and signaling pathways related to BC. We used multiple R packages, including “clusterProfiler,” “org.Hs.eg.db,” “ggplot2,” and “pathview,” for the KEGG pathway enrichment analysis of intersecting genes and subsequent visualization of the results. Gene Set Enrichment Analysis (GSEA) and single-sample Gene Set Enrichment Analysis (ssGSEA) of DEGs between experimental and control groups We conducted enrichment analyses of all genes in the experimental (treatment) and control groups. Intersecting genes were ranked by log₂(fold-change) for GSEA. A standardized enrichment score for each analysis was computed based on 1000 simulated counts. We established an the Benjamini-Hochberg FDR (False Discovery Rate) threshold of less than 0.05 to identify significant KEGG pathways. The results were visualized using the R packages “ggplot2,” “fgsea,” “clusterProfiler,” and “GseaVis.” Using the R package “GSVA,” we performed an ssGSEA to obtain immune cell scores [29]. We analyzed differences in immune cell infiltration between the experimental (treatment) and control groups using the R packages “reshape2,” “ggpubr,” and “corrplot.” The results were then visualized. Then, we examined the intersecting genes and immune cell scores generated using three machine learning methods (LASSO, SVM-RFE, and RF algorithms) to elucidate the relationships between these genes and immune cells and visualized the findings. Screening of exosome genes using machine learning methods We utilized three machine learning methods, Least Absolute Shrinkage and Selection Operator (LASSO), Support Vector Machine-Recursive Feature Elimination (SVM-RFE), and Random Forest (RF) algorithms, to identify risk biomarkers for BC. These three algorithms are widely used for the discovery of diagnostic and prognostic factors [30]. Initially, we conducted univariate Cox regression analysis of the corrected combined gene data and differentially expressed exosome genes using the R package “glmnet.“. The purpose of this analysis was to filter out genes with significant differential expression and associations with exosomes. We then performed a LASSO regression analysis using the R package “glmnet.” This package enables the establishment of steering/penalty parameters and ten-fold cross-validation to select the most relevant variables and reduce overfitting [26]. We conducted a LASSO logistic regression analysis to screen out characteristic genes associated with BC. We also employed SVM-RFE to screen for characteristic BC genes. The SVM-RFE algorithm was implemented in R software using the package “e1071.” This method is effective for identifying genes with high discriminative power [31, 32]. Lastly, we employed the R package “randomForest” to use the RF model to screen for BC genes. RF involves a collection of individual decision trees that are iteratively constructed from random subsets of predictors (features) and dependent variables (target variables) [33]. RF prevents overfitting of a single decision tree and improves model performance based on a large number of related decision trees derived from the same data set [26]. Using the RF algorithm, genes with an importance score of > 2 were screened for further analyses. We identified the common characteristic genes obtained using the three machine learning methods and visualized the results using the R package “ggvenn.” Next, we analyzed the expression differences of these genes between the experimental (treatment) and control groups to explore their correlation with BC. Finally, we utilized the R package “circlize” to create a circular plot of the chromosomal locations of the intersection-feature genes. Construction and validation of a risk nomogram The intersecting signature genes identified through three distinct machine learning methodologies were used to develop a comprehensive BC risk nomogram. This model was validated using receiver operating characteristic curves (ROC) and calibration curves. A decision curves analysis (DCA) was used evaluate accuracy and reliability. The R packages “rms,” “rmda,” “glmnet,” “pROC,” and “ggsci” were used for analyses and the visualization of results. Drug enrichment analysis and network interaction analysis We obtained drug–gene relationships from DSigDB (https://dsigdb.tanlab.org/). The R packages “clusterProfiler,” “org.Hs.eg.db,” “tidytable,” and “ggplot2” were employed to examine the relationships between the intersection-feature genes and drugs and to visualize the results. Drug Regulatory Networks: Cytoscape serves as a comprehensive modeling environment for the integration and analysis of biomolecular interaction networks and their associated states [34]. It enables the visualization of various types of attribute data related to the nodes and edges in biological networks [35]. We utilized Cytoscape (https://cytoscape.org/) to analyze the interactions between the intersecting characteristic genes and drugs and subsequently visualized the results. We obtained RNA-binding protein (RBP) data for all intersecting genes from ENCORI (https://rnasysu.com/encori/) and employed Strawberry Perl software (version 5.3.0, https://www.perl.org) to extract the intersecting gene network relationship file and node attribute file. Subsequently, we constructed a RBP regulatory network using Cytoscape. We obtained transcription factor (TF) data for all intersecting genes from TRRUST (https://www.grnpedia.org/trrust/) and utilized Strawberry Perl to generate an intersecting gene TF network relationship file and node attribute file. We developed a transcription factor regulation network using Cytoscape. Molecular docking analysis We selected several intriguing or previously reported drugs from the drug enrichment data for a molecular docking analysis. We retrieved the 3D structures of the selected drugs from the Explore Chemistry database (https://pubchem.ncbi.nlm.nih.gov/). We obtained the corresponding protein structures from the Protein Data Bank (PDB) (https://www.rcsb.org/). A molecular docking analysis of the drugs and protein structures was conducted using CB-DOCK2 (https://cadd.labshare.cn/cb-dock2/index.php). The docking molecule with the most favorable (lowest) free energy, indicating the highest stability, was selected for visualization. Results DEG analysis To obtain a large sample size, we acquired four datasets from the NCBI GEO database, including data for 90 BC samples and 46 normal tissue samples. These four datasets were concatenated, and the combined results are depicted in Fig. 1A. To account for batch effects in the data, we subjected the combined dataset to normalization procedures, with the outcomes presented in Fig. 1B. Data batch correction and analysis of genetic differences. A, B Analysis plots before and after batch correction; Heatmap (C) and volcano plot (D) of gene difference analysis To delve deeper into the extensive genetic information contained within tissue samples, we evaluated differential gene expression between experimental (treatment) and control samples. These DEGs may contribute to the tumor microenvironment in BC. We further evaluated the DEGs using a hierarchical cluster analysis. An exhaustive hierarchical clustering heatmap revealed distinct gene expression clusters, elucidating the fundamental mechanisms underlying the dysregulation of gene expression in BC. The top 100 DEGs are shown in Fig. 1C. Between the experimental and control groups, 595 genes were differentially expressed, including 416 that were downregulated and 179 that were upregulated in the experimental (treatment) group (Supplementary Material 1), as shown in Fig. 1D. Functional enrichment analysis of differentially expressed exosome genes We performed a joint analysis of the 595 DEGs and 878 exosome gene, and identified 37 differentially expressed exosome genes, as depicted in Fig. 2A. The detailed information of 37 exosomal genes is presented in Supplementary Material 2. We performed a GO functional analysis of these 37 exosome genes. GO terms are assigned to three primary categories: BP, CC, and MF. Notably, the exosome genes were predominantly enriched for terms in the CC category, as illustrated in Fig. 2B. In the BP category, the exosome genes were enriched for the regulation of monoatomic ion transmembrane transport. This process may be vital for differentiating the control group from the experimental (treatment) group. Within the CC category, the collagen-containing extracellular matrix was the most abundant term. In the MF category, calcium-dependent protein binding and modified amino acid binding were the most significant terms. These results are presented in Fig. 2C. As expected, single genes possessed multiple functions, and multiple genes were enriched within a single component. The complex network is illustrated in Fig. 2D. Functional enrichment analysis based on intersection genes. A Intersection gene analysis of differential genes and exosome genes; Circle plot (B) and bubble plot (C) of GO enrichment analysis of intersection genes; D Analysis of the network relationship between exosome genes and GO functional components; Bubble diagram (E) and network diagram (F) of KEGG pathway enrichment analysis of intersection genes; G Proteoglycans pathway in cancer. (GO: gene ontology; KEGG: Kyoto encyclopedia of genes and genomes) We performed a KEGG enrichment analysis of the intersecting exosome genes. These genes were predominantly involved in fluid shear stress and atherosclerosis, proteoglycans in cancer, and the renin-angiotensin system, as illustrated in Fig. 2E. These findings underscore the intricate interplay between signaling pathways and exosome genes, which play a pivotal role in the initiation and progression of BC. Figure 2F–G offers a comprehensive view of the intersecting exosome genes and enriched pathways, providing a detailed breakdown of these intricate interactions. GSEA To uncover the biological pathways associated with BC, we performed a GSEA of the differentially expressed exosome genes. The top five pathways related to the intersecting exosome genes are summarized in Fig. 3A. Notably, four of the top five pathways were up-regulated, while one pathway was down-regulated. Results for the three most significantly enriched pathways are displayed in Fig. 3B–D. The upregulated pathways in BC primarily included the cell cycle, Epstein-Barr virus infection, antigen processing and presentation, and the regulation of lipolysis in adipocytes. The down-regulated pathways were involved in DNA replication, focal adhesion, the Wnt signaling pathway, and fatty acid degradation. These discoveries offer additional insights into the deregulated molecular cascade and abnormal cellular processes that contribute to the pathogenesis of BC. Functional enrichment analysis of intersection exosome genes. A GSEA examines the top five ranked pathways; B-D Generalized spectral estimation study of the most significantly enriched 2 to 4 downregulated pathways. (GSEA: gene set enrichment analysis) Construction and validation of three machine learning models Utilizing three machine learning algorithms, we screened for potential BC genes from a pool of 37 intersecting exosome genes with prognostic value. Initially, through the Lasso regression model, we pinpointed 15 exosome genes linked to BC: SCRN2, TGFBR3, PKD2, DAAM2, MME, FN1, ZFP36, ANXA1, PIGR, CXCL12, GPRC5A, LRRC15, CA4, MMP9, and ANPEP (Fig. 4A–B). Subsequently, we leveraged SVM, a highly effective machine learning algorithm, to identify intersecting exosome genes associated with BC. Through stringent screening processes, we identified 27 exosome genes (Fig. 4C–D) closely related to BC. Lastly, in the RF model, setting the relative correlation score to > 2, we identified 12 genes associated with BC (Fig. 4E–F). The construction and validation of the three machine learning models are summarized in Fig. 4. Construction and Verification of the LASSO model, the SVM-RFE model, and the RF model. A, B 37 different iterations of the LASSO model’s cross-validation process to adjust the choice of parameters. Each curve represents a gene. LASSO analysis of these coefficients. The best lambda is presented at the vertical dashed line; C, D An SVM-RFE method for feature gene selection; E, F Correlation between the total number of trees in a RF and the error rate. An order based on the relative significance of genes. (LASSO: Least absolute shrinkage and selection operator; SVM-RFE: Support Vector Machine-Recursive Feature Elimination; RF: random forest) Gene correlation analysis Using the three machine learning methods, we identified five common genes: TGFBR3, MME, PIGR, GPRC5A, and CA4 (depicted in Fig. 5A). Subsequently, we evaluated the expression of these five genes in the experimental (treatment) and control groups, as illustrated in Fig. 5B. Four genes, namely TGFBR3, MME, PIGR, and CA4, showed lower expression in the experimental (treatment) group than in the control group, whereas GPRC5A was over-expressed in the experimental group. Next, we analyzed the correlations among these five genes. TGFBR3, MME, and PIGR were positively correlated with each other. Conversely, TGFBR3 and MME displayed negative correlations with GPRC5A (Fig. 5C). Lastly, we explored the chromosomal distribution of these five genes. As shown in Fig. 5D, TGFBR3 and PIGR were located on the same chromosome(chr1). Identifying the associated signature gene. A Venn diagram for identifying common feature genes of LASSO, SVM-RFE and RF model; B Differential expression of five common genes between experimental (treat) and control groups; C Association analysis of common genes. D Map of the distribution of common genes on chromosomes. (ns: no significance; *P < 0.05; **P < 0.01; ***P < 0.001.) Construction and validation of a nomogram To evaluate the diagnostic significance of the five exosome genes in BC, we constructed a nomogram (Fig. 6A). A higher total score was associated with an increased risk of BC. Additionally, we evaluated the performance of a single gene versus the risk model. The area under the curve (AUC) of the risk model was 0.987 (Fig. 6B), and the model significantly outperformed any single gene for predicting the risk of BC (Fig. 6C). These findings suggest that the combined use of these five genes in a risk model can provide a more accurate and reliable diagnostic tool for BC. Calibration curves further demonstrated that the predicted model fitted well with actual gene expression data (Fig. 6D). Additionally, a DCA (Fig. 6E) showed that the model provides a high net benefit. These findings validate the robustness and clinical applicability of our risk model. Constructing and validating the nomogram. A Nomogram of five common exosome genes associated with BC risk; B ROC curve of risk model (AUC = 0.987); C ROC curve of five genes (AUC: TGFBR3 = 0.935, MME = 0.889, PIGR = 0.85, GPRC5A = 0.811, CA4 = 0.814); Calibration curve (D) and DCA (E) of the prediction model. BC: breast cancer; ROC: receiver operating characteristic curves; AUC: area under the curve; DCA: decision curves analysis Immunocorrelations and molecular interaction analysis We utilized the R package “ssGSEA” to obtain immune scores for the corrected gene expression data in the experimental (treatment) and control groups. Through an analysis of these immune scores, we detected differentially expressed immune cells between the two groups. The immune cells with high expression levels in the experimental (treatment) group included Myeloid-derived suppressor cells (MDSC), CD56dim natural killer (NK) cells, NK cells, regulatory T cells, activated CD8 T cells, activated dendritic cells, CD56 bright NK cells, and again activated CD8 T cells. In the experimental (treatment) group, Plasmacytoid dendritic cells, Type 2 T helper cells, Mast cells, Eosinophils, Type 1 T helper cells, and Memory B cells exhibited low expression levels (Fig. 7A). Subsequently, we examined the relationship between model exosome genes and immune cell infiltration. CA4 did not have a significant regulatory relationship with immune cells. The other four genes demonstrated positive or negative regulatory relationships with immune cells (Fig. 7B). Analysis of immune correlates and molecular interactions. A Differentially infiltrated immune cells in the control and experimental (treat) groups; B Analysis of model exosome genes and immune cell infiltration regulation; Bar graph of enrichment analysis of model exosome genes and therapeutic drugs (C) and bubble graph (D). Network diagram of drug (E), RNA-binding protein (F), and TF (G) interactions with model exosome genes. (ns: no significance; *P < 0.05; **P < 0.01; ***P < 0.001. TF: transcription factor.) We performed a therapeutic drug enrichment analysis of the model exosome genes. Figure 7C–D shows the 30 drugs with the most significant enrichment, among which dobutamine exhibited the highest level of enrichment. Then, we constructed a network diagram depicting the regulation of genes by drugs, as presented in Fig. 7E. Finally, we analyzed the transcriptional regulation of RBPs and TFs on model exosome genes. The results, as shown in Fig. 7F, indicated that a single RBP can regulate the expression of multiple genes, and a gene can be regulated by multiple RBPs. In the TF regulatory network (Fig. 7G), we observed that GPRC5A and CA4 were not regulated by any TFs. Each of the remaining three genes was regulated by more than one TF. Molecular docking analysis We evaluated therapeutic drugs targeting the model exosome genes through a molecular docking analysis. We obtained the tertiary structure of the protein corresponding to each gene and the 3D structure of the drug for the molecular docking analysis. Figure 8A displays the docking results for bicalutamide with the MME gen(PDB ID: 6GID), and Fig. 8B shows the docking results for praziquantel with the TGFBR3 gene (PDB ID: 8DC0). Figure 8C–F presents the docking results for ferulic acid, p-coumaric acid, coumarin, and acetazolamide with the CA4 gene(PDB ID: 3F7U), respectively. Docking analysis of drug and protein molecules. A 3D analysis diagram of bicalutamide and MME molecular docking; B 3D analysis diagram of molecular docking between praziquantel and TGFBR3; 3D analysis diagram of molecular docking of CA4 with ferulic acid (C), p-coumaric acid (D), coumarin (E) and acetazolamide (F) Discussion BC remains one of the leading causes of cancer-related morbidity and mortality worldwide. The heterogeneity of the disease and complexity of its underlying mechanisms pose significant challenges to treatment and decision-making processes. Accordingly, novel approaches are needed to deepen our understanding of BC biology and ultimately improve patient outcomes. Our results shed light on the crucial role of exosome-related genes in the initiation and progression of BC. By integrating phenotypic data with cutting-edge machine learning techniques, we identified potential biomarkers that could prove invaluable for clinical applications. The application of machine learning methods streamlined the selection of significant exosome-related genes and revealed their diagnostic value in BC. Our DEG analysis, functional enrichment studies, and machine learning models collectively emphasize the potential of these biomarkers to guide therapeutic decisions and personalize treatment strategies for patients with BC. The DEG analysis provided significant insights into the molecular mechanisms underlying BC. Using data from multiple cohorts, we identified 37 differentially expressed exosome genes between BC samples and normal tissues. These genes were involved in several critical biological pathways related to BC (Fig. 2C–F). A KEGG enrichment analysis revealed that the intersecting exosome genes were involved in signaling pathways associated with BC, supporting their critical role in tumorigenesis and progression. The enriched pathways included fluid shear stress and atherosclerosis, proteoglycans in cancer, and the renin-angiotensin system, indicating that complex interactions between these biological processes and exosome gene expression patterns could contribute to BC. These pathways may mediate the effects of exosome genes on the tumor microenvironment and the metastatic propensity of BC cells. For example, the modulation of fluid shear stress can bolster the survival of circulating tumor cells, which play a crucial role in the metastatic process [36]. Fluid shear stress can promote the release of exosomes and the adhesion of blood, and they also link chronic inflammatory signaling (shared with atherosclerotic pathways) to BC‑associated vascular remodeling and metastatic dissemination [37, 38]. Acconcia et al. reported that angiotensin‑mediated inflammation, angiogenesis, and extracellular matrix remodeling are well‑established drivers of breast cancer progression, metastasis, and epithelial‑mesenchymal transition [39]. Furthermore, proteoglycans play a pivotal role in the physical structure of the extracellular matrix, which is essential for cancer cell proliferation and migration [40]. The identification of these pathways emphasizes the significance of exosome genes in BC biology and suggests that they are potential therapeutic targets for intervention. Further analyses of these signaling pathways may lead to the development of novel therapeutic approaches for BC, especially therapies targeting the molecular mechanisms that promote tumor progression and metastasis [41, 42]. This research contributes to the growing body of evidence that emphasizes the significance of exosome communication in cancer biology, providing valuable insights that could lead to advancements in diagnostic and therapeutic methodologies for BC. We conducted a GSEA to identify key pathways in BC (Fig. 3). The cell cycle pathway emerged as a predominant upregulated pathway, whereas the DNA replication pathway was downregulated. These two pathways are essential for regulating cell proliferation, which is often dysregulated in cancer, leading to uncontrolled cell division [43–45]. The GSEA results suggest that exosome genes promote tumor growth and progression, which aligns with previous findings and further underscores their significance in cancer biology [46]. Of note, Focal adhesion, WNT Signaling pathway, and Fatty Acid Degradation pathways were significantly downregulated in BC. We identified potential biomarkers for BC diagnosis using three machine learning algorithms. In particular, five exosome-related genes (TGFBR3, MME, PIGR, GPRC5A, and CA4) were associated with diagnosis. Among these five genes, four genes (TGFBR3, MME, PIGR, and CA4) exhibited low expression levels in tumor tissues, whereas GPRC5A showed high expression. We developed a nomogram based on these five key exosome-related genes with significant predictive value (AUC: 0.987, 95% CI: 0.971–0.998). There was good consistency between the calibration curves and DCA curves. Given these promising results, further research should evaluate the mechanisms by which these five exosome-related genes contribute to BC. ssGSEA revealed significant correlations between exosome-related genes and immune cell infiltration in the tumor microenvironment, suggesting that these genes are crucial for regulating immune responses in BC. Frequences of specific immune cell types, such as MDSCs and activated CD8 T cells, differ between BC samples and normal tissues, and exosome-related genes may influence immune cell dynamics. This is particularly important as the tumor microenvironment is a critical factor in cancer progression and the response to therapy. The five differentially expressed exosome genes correlated with immune cell infiltration may shape the immune landscape in BC. Furthermore, the observed positive and negative regulatory relationships between exosome genes and various immune cell types indicate that these genes are potential biomarkers for predicting patient outcomes and responses to immunotherapy. Understanding the interplay between exosome-related genes and immune cell infiltration deepens our understanding of BC biology and creates opportunities for the development of targeted therapies aimed at modulating the immune microenvironment. These insights align with recent literature that underscores the role of exosomes in cancer immunology, reinforcing the idea that exosome-related gene signatures could be crucial in advancing personalized medicine approaches in BC [47–49]. Despite extensive research focused on the pharmacological treatment of BC, effective drugs that significantly improve survival rates and mitigate complications are still limited. In the current study, potential drug molecules targeting exosome genes were predicted using the Drug Signature Database (DSigDB). Notably, dobutamine was pinpointed as a candidate therapeutic agent for BC. Dobutamine enhances myocardial contractility and is commonly used in the treatment of heart failure and shock. However, few studies have evaluated its application to tumor treatment. Based on our findings, future research should evaluate dobutamine as a potential treatment for BC and its utility in clinical applications. “Drug–target–disease” networks provide a basis for discovering new drugs and therapeutic targets [50]. Out of the five differentially expressed exosome genes in this study, four were identified as potential drug targets. Drugs targeting the PIGR gene have not yet been discovered, and this should be a focus of research in the future. Bicalutamide, a non-steroidal antiandrogen medication, is frequently employed in prostate cancer treatment. However, few studies have evaluated the application of bicalutamide in BC treatment. Bicalutamide inhibits the proliferation and invasion of MDA-MB-231 BC cells in a dose-dependent manner [51]. Our results suggest that bicalutamide exerts its effects by targeting MME and GPRC5A, presenting a novel basis for the application of bicalutamide in the treatment of BC. Bicalutamide binds to the catalytic active domain of MME (neprilysin), inhibits its metalloprotease activity, thereby regulating tumor-related peptide degradation and suppressing breast cancer cell proliferation. RBPs and TFs are crucial for RNA transcription and synthesis. RBPs facilitate the maturation, transport, localization, and translation of RNA, while TFs are essential in the process of gene transcription. Notably, TFs for GPRC5A and CA4 were not detected within the TF regulatory network, highlighting the need for further analyses of their regulatory mechanisms. Molecular docking is a key approach employed in the field of drug discovery and screening [52]. We used a molecular docking analysis to construct a structurally stable drug–protein interaction model. We found that few drugs target the MME and TGFBR3 proteins. Conversely, a relatively large number of drugs acted on the CA4 protein. Furthermore, we observed that diverse drugs exhibit specificity for different targets within the same protein. However, drugs specifically targeting the GPRC5A and PIGR proteins have not been reported to date, emphasizing the need for further research efforts. Limitations and conclusions It is crucial to recognize several limitations of this study. The most significant limitation was the relatively small sample size, which might affect the generalizability of our findings. Extensive clinical validation studies are needed to verify the reliability and reproducibility of our findings. Furthermore, the functions of the exosome genes identified in this study should be evaluated using wet lab experiments. Additionally, the integration of multiple datasets may introduce batch effects, which could potentially bias the results. Despite these limitations, we successfully identified 37 differentially expressed exosome genes that are significantly linked to BC. By applying machine learning algorithms, we identified candidate biomarkers for risk and prognosis. These findings deepen our comprehension of the molecular mechanisms driving BC and provide new directions for research focused on uncovering the therapeutic potential of exosome genes in clinical contexts. Ultimately, continued research on these genes holds promise for the advancement of targeted interventions and precision medicine strategies for BC, potentially resulting in the development of more effective and personalized therapies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Differentially expressed genes between breast cancer and normal tissues. Supplementary Material 2. Acknowledgements We thank Liwen Bianji (Edanz) (www.liwenbianji.cn) for editing the English text of a draft of this manuscript. Abbreviations Area under the curve Breast cancer Biological processes Cellular components Decision curves analysis Differential expression gene Exosome gene expression False Discovery Rate Gene Expression Omnibus Gene ontology Gene set enrichment analysis Kyoto encyclopedia of genes and genomes Least absolute shrinkage and selection operator Myeloid-derived suppressor cells Molecular functions Magnetic resonance imaging National Center for Biotechnology Information Principal component analysis Protein Data Bank RNA-binding Protein Random forest Receiver operating characteristic curves Single-sample gene set enrichment analysis Support Vector Machine-Recursive Feature Elimination Transcription factor Author contributions WXW and LT: Writing–original draft, Data curation, Conceptualization, Supervision. DYY and SSY: Methodology, Investigation, Formal analysis, Software, Resources, Project administration. JXM: data acquisition, bioinformatics analysis, figure processing. JZK and LT: Writing–review & editing. Funding This work was not supported by any funding. Data availability The dataset analyzed during this study is available in the NCBI Gene Expression Omnibus (GEO) repository under accession number GSE50428 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE50428), GSE71053 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE71053), GSE139038 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE139038) and GSE205185 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE205185). Declarations Ethics approval and consent to participate Not applicable. 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. Contributor Information Zekun Jiang, Email: JZK_missef95@163.com. Tian Liu, Email: 568523745@qq.com. References Associated Data Supplementary Materials Supplementary Material 1: Differentially expressed genes between breast cancer and normal tissues. Supplementary Material 2. Data Availability Statement The dataset analyzed during this study is available in the NCBI Gene Expression Omnibus (GEO) repository under accession number GSE50428 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE50428), GSE71053 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE71053), GSE139038 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE139038) and GSE205185 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE205185).