Integration of exosome-related genes and differential expression analysis reveals potential biomarkers for prostate cancer

Abstract

Background

Prostate cancer (PCa) is a prevalent malignancy in men, with exosomes playing a key role in tumor microenvironment and disease progression, yet their molecular mechanisms remain unclear.This study aims to identify potential biomarkers and therapeutic targets in PCa by integrating exosome-related genes with differentially expressed genes (DEGs).

Methods

Four GEO datasets (GSE32448, GSE46602, GSE69223, GSE6956) were analyzed. Batch effects were corrected using the ComBat method, followed by DEG analysis and feature selection via machine learning (LASSO regression, random forest, SVM). Functional enrichment and molecular docking validated the findings.

Results

Post-correction, sample clustering improved significantly. Of 49 overlapping DEGs and exosome-related genes, EEF2, LGALS3, and MYO1D emerged as key biomarkers, with EEF2 showing the highest predictive power (AUC = 0.786). A risk score model achieved an AUC of 0.886. Immune analysis linked these genes to immune cell subsets, and docking studies revealed strong interactions with small molecules like cycloheximide.

Conclusion

This study elucidates the molecular role of exosome-related genes in PCa, proposing predictive biomarkers and novel therapeutic targets, warranting further clinical validation.

Introduction

Prostate cancer (PCa) ranks among the most common and deadly cancers affecting men globally. Its slow progression offers a window for early intervention, yet advanced stages—particularly castration-resistant forms—remain stubbornly resistant to current therapies [1]. Despite substantial advances in understanding the genetic and molecular basis of PCa, therapeutic options remain limited and often inadequate for advanced stages of the disease. Recent studies have highlighted the significant role of exosomes—small extracellular vesicles—in facilitating intercellular communication and influencing the tumor microenvironment [2–4]. In PCa, exosomes fuel metastasis, dampen immune defenses, and bolster resistance to drugs, making them a tantalizing target for research [5, 6]. In this study, we focused on 49 exosome-related genes identified from the ExoCarta database. These genes play crucial roles in regulating the exosome-mediated processes that impact cancer progression and immune modulation. Exosomes, as carriers of proteins, RNA, and lipids, facilitate tumor growth and resistance, and understanding these gene interactions offers new therapeutic avenues.

Addressing these challenges, this study integrates multiple PCadatasets to explore the relationship between DEGs and exosome-related genes, with the aim of uncovering novel biomarkers and therapeutic targets. Batch effects, which arise from technical variations between different experimental cohorts, can obscure biological signals and lead to misleading conclusions [7–9]. In this study, we employed a robust batch effect correction approach to harmonize data from four publicly available PCa datasets (GSE32448, GSE46602, GSE69223, and GSE6956), ensuring a more accurate representation of the underlying biological processes. Principal component analysis (PCA) before and after batch correction demonstrated significant improvements in sample clustering, confirming the success of this strategy in mitigating technical biases.

To further investigate the molecular landscape of PCa, we performed comprehensive differential gene expression analysis, revealing a set of DEGs with potential relevance to cancer progression. Volcano plots and heatmaps highlighted key genes associated with differentially regulated pathways, while Gene Ontology (GO) and Gene Set Enrichment Analysis (GSEA) provided insights into functional categories and enriched biological processes. Notably, our GO enrichment analysis revealed significant involvement of DEGs in processes such as regulation of endocytosis, ion transport (particularly potassium ion transport), and immune regulation—pathways closely tied to exosome biology and tumor dynamics. These processes are fundamental to cellular homeostasis and may be dysregulated in cancer, contributing to tumor growth, drug resistance, and metastasis [10–12].

Through rigorous feature selection methods, including LASSO regression and random forest, we identified a set of critical genes, such as EEF2, LGALS3, and MYO1D, which are involved in the regulation of endocytosis, ion transport, and immune modulation [13]. These genes were not only implicated in biological processes but also demonstrated strong predictive value when incorporated into risk score models. The integration of immune cell composition analysis and molecular docking further expanded our understanding of how these genes interact with small molecules and immune cells, offering new avenues for therapeutic intervention [14–16].

In summary, our comprehensive analysis not only provides a deeper understanding of the molecular interactions between DEGs and exosome-related genes in PCa but also highlights the utility of batch effect correction in ensuring the reliability of data derived from heterogeneous cohorts. These findings open up new possibilities for targeted therapeutic strategies and offer potential biomarkers for improving clinical outcomes in PCa.

Method

Data collection and preprocessing

Gene expression data from four public datasets, namely GSE32448, GSE46602, GSE69223, and GSE6956, were obtained from the Gene Expression Omnibus (GEO) database. The datasets included both experimental and control groups for each cohort, with data formatted as normalized expression values for downstream analysis. Samples were inspected for potential batch effects using PCA prior to correction. Batch effects were identified as significant sample dispersion along the PC1 and PC2 axes, which is consistent with previous findings regarding cross-dataset variability [17, 18].

Batch effect correction

To address the batch effect between datasets, we corrected batch effects using the ComBat method in the sva R package, a widely used tool for adjusting systematic technical variations in high-dimensional data [19]. The PCA plot of the corrected data demonstrated improved clustering, indicating successful adjustment of the batch effects. This correction method has been previously validated for use in gene expression studies and has been shown to maintain the biological variability while removing confounding technical noise [17].

Differential gene expression (DGE) analysis

Differential gene expression (DGE) analysis was performed using the limma package [20, 21]. Genes with |log2FC| >0.5 and P < 0.05 were considered differentially expressed. A volcano plot was used to visualize upregulated (red) and downregulated (blue) genes, while genes with no significant difference in expression were represented in gray. For additional clarity, a heatmap of the DEGs was generated using hierarchical clustering, where red and blue colors indicate high and low gene expression levels, respectively.

Exosome-Related genes and overlap with DEGs

A set of exosome-related genes was extracted from the ExoCarta database [18], and the overlap between DEGs and exosome-related genes was assessed using a Venn diagram. The intersecting genes were identified as potentially involved in both differential expression and exosome-related pathways. We extracted exosome-related gene data from the ExoCarta database, which includes a comprehensive list of proteins and RNA associated with exosomes. A total of 49 exosome-related genes were identified and analyzed for overlap with differentially expressed genes (DEGs) in prostate cancer. These overlapping genes were selected for further analysis in our study.

Functional enrichment analysis

Gene Ontology (GO) enrichment analysis was performed using the clusterProfiler R package [22]. GO terms were categorized into molecular functions, biological processes, and cellular components, and visualized using bar plots, bubble plots, and circular visualization plots. The top significantly enriched GO terms were selected based on their adjusted p-values. Gene Set Enrichment Analysis (GSEA) was performed using the GSEABase package [23].

Feature selection and model development

To identify key predictive features, feature selection was performed using LASSO (Least Absolute Shrinkage and Selection Operator) regression, implemented in the glmnet R package [24]. The LASSO regression coefficient path plot was used to visualize the shrinkage of coefficients as the regularization strength (λ) increased. Cross-validation was applied to select the optimal λ, and 10-fold cross-validation was used to evaluate the model’s accuracy and stability. Model performance was further validated using a random forest (RF) algorithm, implemented through the randomForest R package [21]. The random forest model evaluated the importance of genes in predicting the outcome, with key drivers identified based on their contribution to the predictive performance.

Correlation and chromosomal location analysis

Correlation analysis between selected genes (EEF2, LGALS3, MYO1D) was performed using Pearson correlation, and results were visualized in a correlation heatmap. The chromosomal locations of the genes were retrieved from the UCSC Genome Browser, and their positions on distinct chromosomes were mapped for potential genetic insights.

Model evaluation and risk score calculation

Receiver Operating Characteristic (ROC) curve analysis was used to evaluate the predictive performance of selected genes. Area Under the Curve (AUC) values were calculated to assess the discriminatory power of each gene, with EEF2, LGALS3, and MYO1D analyzed for their individual and combined performance in distinguishing between experimental and control groups. Based on the selected genes, a risk score model was developed using a weighted sum of gene expression values. Calibration and net benefit curves were constructed to assess the accuracy and clinical utility of the risk score.

Immune cell composition and molecular enrichment

Immune cell composition differences between treatment and control groups were analyzed using the CIBERSORT algorithm [25] to estimate the relative abundance of immune cell subsets based on gene expression data. The distribution of immune cells was visualized using box plots and heatmaps. In parallel, molecular enrichment analysis of small molecules associated with EEF2, LGALS3, and MYO1D was performed using the Metascape tool [26]. Gene-small molecule interaction networks were generated to visualize the relationship between the selected genes and small molecules, providing insights into potential therapeutic targets.

Molecular Docking and binding affinity analysis

Molecular docking simulations were conducted to evaluate the binding affinity between selected genes (EEF2, LGALS3, MYO1D) and small molecules (cycloheximide, QUINOLINE, and topotecan). Docking studies were performed using AutoDock Vina [27], and binding affinities were reported in kcal/mol. The 3D docking structures were visualized using PyMOL [28], and the interactions between amino acid residues and small molecules were analyzed to identify key binding sites.

Results

Batch effect correction and the relationship between DEGs and exosome-related genes

Prior to batch effect correction, PCA revealed that samples from different cohorts (GSE32448, GSE46602, GSE69223, and GSE6956) were widely dispersed along the PC1 and PC2 axes, indicating significant batch effects between the groups (Fig. 1A). After batch effect correction, the PCA plot demonstrated that the samples clustered more tightly, and the corrected data indicated that variability between groups was effectively controlled (Fig. 1B). A volcano plot was used to visualize the results of differential gene expression analysis, where red points represent upregulated genes, blue points represent downregulated genes, and gray points represent genes with no significant differences (Fig. 1C). A heatmap shows gene expression levels of DEGs between control (yellow) and treatment (green) groups across four datasets (GSE32448, GSE46602, GSE69223, GSE6956). Genes are clustered by similar expression patterns, with red indicating upregulation and blue indicating downregulation. The clustering demonstrates clear differentiation between control and treatment groups, indicating treatment-induced changes in gene expression. Batch effect correction was applied, improving the clustering and ensuring that observed differences reflect biological variation rather than technical biases (Fig. 1D). A Venn diagram revealed the intersection between DEGs and exosome-related genes. Of the total, 503 genes were unique to the DEGs, 815 genes were unique to the exosome dataset, and 49 genes were common to both, representing 36.8% of DEGs and 59.6% of exosome-related genes (Fig. 1E).

Fig. 1.

Batch effect correction and DEGs vs. exosome-related genes. (A) PCA before batch effect correction showing sample dispersion across cohorts. (B) PCA after correction demonstrating improved sample clustering. (C) Volcano plot of DEGs, with red for upregulated, blue for downregulated, and gray for non-significant genes. (D) Heatmap displaying DEGs expression patterns, with red for high and blue for low expression. (E) Venn diagram showing overlap between DEGs and exosome-related genes, with 49 common genes

Functional enrichment analysis

A bubble plot illustrates the GO analysis results, showing the functional terms and their corresponding gene ratios (GeneRatio). Significant functional terms were predominantly focused on “regulation of endocytosis,” “metal ion transport,” and “sodium ion transport” (Fig. 2A). Gene Set Enrichment Analysis (GSEA) curve plots revealed the enrichment results for four key pathways, including “Staphylococcus aureus infection,” “muscle cell cytoskeleton,” “chemical carcinogenesis-DNA adducts,” and “complement and coagulation cascades.” These pathways are involved in critical biological processes that may drive immune modulation, tumor progression, and therapy resistance. The running enrichment score (NES) and corresponding p-values for each pathway are displayed in the table, demonstrating significant enrichment of these pathways in the differential gene set (Fig. 2B).

Fig. 2.

Functional enrichment analysis of DEGs. (A) Bubble plot showing GO analysis results with significant terms like “regulation of endocytosis,” “metal ion transport,” and “sodium ion transport.” (B) GSEA curve plots for key pathways, including “Staphylococcus aureus infection” and “muscle cell cytoskeleton,” with enrichment scores (NES) and p-values

Feature selection and model performance evaluation

The LASSO regression coefficient path plot demonstrates the stepwise shrinkage of coefficients as the L1 norm increases. The results show that the coefficients of several features rapidly approach zero as the regularization strength increases, indicating that only a few features are retained in the model for further analysis (Fig. 3A).The LASSO regression cross-validation results illustrate the change in the binomial distance of the model across different Log(λ) values. By selecting an optimal λ value, the model’s predictive accuracy was maximized, showing the best performance at this λ value (Fig. 3B). The 10-fold cross-validation accuracy analysis revealed that as the number of features increased, the model’s accuracy fluctuated. The optimal model performed best with a moderate number of features, reaching an accuracy of approximately 0.81 (Fig. 3C).The 10-fold cross-validation error results indicated that as the number of features increased, the model’s error gradually decreased, suggesting that feature selection positively influenced the stability and accuracy of the model (Fig. 3D).As the number of trees in the random forest model increased, the training, validation, and testing errors stabilized, indicating that the model’s generalization ability significantly improved after 200 trees (Fig. 3E). The random forest model evaluation results revealed the importance ranking of features, highlighting genes such as EEF2, LGALS3, and MYO1D as key drivers in the model. These genes are involved in critical pathways such as regulation of endocytosis, ion transport, and immune modulation, which are integral to tumor growth, immune evasion, and drug resistance. These features contributed the most to the model’s predictive performance and may be critical factors in the associated biological processes (Fig. 3F).

Fig. 3.

Feature selection and model performance evaluation. (A) LASSO regression coefficient path plot showing stepwise shrinkage of coefficients. (B) LASSO regression cross-validation results illustrating model performance across different Log(λ) values. (C) 10-fold cross-validation accuracy analysis revealing optimal performance with a moderate number of features. (D) 10-fold cross-validation error analysis indicating improved model stability with more features. (E) Random forest model performance showing stabilization of errors after 200 trees. (F) Feature importance ranking from random forest, highlighting key genes such as ANXA2, TSPAN1, and EEF2

Feature selection and model evaluation results

The Venn diagram illustrates the overlap of features selected by three methods: LASSO regression, random forest (RF), and support vector machine (SVM). The diagram shows that each method selected a different number of features, with three common features identified by all three methods (Fig. 4A).Gene expression analysis revealed significant differences in the expression levels of EEF2, LGALS3, and MYO1D between the experimental group (Treat) and the control group (Control) (Fig. 4B).Correlation analysis between genes showed a negative correlation between EEF2 and LGALS3 expression (r = -0.04), a weak positive correlation between MYO1D and EEF2 expression (r = 0.20), and a positive correlation between MYO1D and LGALS3 (r = 0.04). These correlations may provide insights into the potential mechanisms underlying their interactions in related biological processes (Fig. 4C).The chromosomal location map of the genes displayed the positions of EEF2, LGALS3, and MYO1D on different chromosomes, indicating that these genes are located in distinct chromosomal regions, suggesting they may have roles in different genetic contexts (Fig. 4D).Receiver operating characteristic (ROC) curve analysis was used to assess the performance of EEF2, LGALS3, and MYO1D in the predictive model. The results showed that EEF2 had an AUC of 0.786, LGALS3 had an AUC of 0.751, and MYO1D had an AUC of 0.716, indicating good performance, with EEF2 demonstrating the best predictive ability (Fig. 4E). The ROC curve for the integrated model demonstrated its performance, with an AUC of 0.886 and a 95% confidence interval (CI) of 0.844–0.923, indicating high sensitivity and specificity in distinguishing the experimental and control groups (Fig. 4F).

Fig. 4.

Feature selection and gene expression analysis. (A) Venn diagram illustrating overlap of features selected by LASSO regression, random forest, and SVM. (B) Gene expression analysis showing significant differences in EEF2, LGALS3, and MYO1D between the treatment and control groups. (C) Correlation analysis between genes showing various correlations in their expression. (D) Chromosomal location map of EEF2, LGALS3, and MYO1D. (E) ROC curve analysis showing predictive performance of EEF2 (AUC = 0.786), LGALS3 (AUC = 0.751), and MYO1D (AUC = 0.716).(F) ROC curve of the integrated model showing high predictive performance (AUC = 0.886, 95% CI = 0.844–0.923)

Evaluation of the risk score model

Based on the genes EEF2, LGALS3, and MYO1D, we developed a risk score system. The score range for each gene is labeled below. The total score ranges from 0 to 220, with the risk level increasing from low to high, providing a quantitative basis for disease risk prediction (Fig. 5A). Figure 5B shows the calibration curve, which evaluates the agreement between the predicted and actual probabilities. The dotted line represents the ideal perfect prediction, the solid line represents the bias-corrected curve, and the dashed line shows the original (uncorrected) model’s predictions. By comparing these three curves, we can assess the accuracy and reliability of the model’s predictions. Figure 5C presents the net benefit curve under different cost-benefit ratios. The red curve represents the net benefit when considering only the genes, the gray curve represents the net benefit without considering the genes, and the black curve shows the net benefit when all risk factors are considered. At various threshold probabilities, the red curve outperforms the others, indicating that the gene risk score enhances the net benefit of the predictive model.

Fig. 5.

Risk score model evaluation. (A) Risk score system based on EEF2, LGALS3, and MYO1D, with score ranges indicating risk levels. (B) Calibration curve comparing predicted and actual probabilities. (C) Net benefit curve showing the enhanced performance of the gene risk score in the predictive model

Analysis of immune cell composition and molecular enrichment in the treatment and control groups

Figure 6A displays the distribution differences of immune cell compositions between the treatment group (red) and the control group (green). The box plot shows variations in immune cell components between the two groups, with significant differences observed in multiple cell types, such as CD8 + T cells, B cells, and regulatory T cells.The heatmap demonstrates the correlation between MYO1D, LGALS3, and EEF2 with immune cell components. This plot reveals the correlations between these three genes and various immune cell subsets. Red indicates a positive correlation, while blue indicates a negative correlation, with the strength of the correlation indicated by the color intensity (Fig. 6B). The bar plot shows the number of small molecules associated with LGALS3, MYO1D, and EEF2 and their adjusted p-values. This plot illustrates the abundance of several chemical compounds related to these genes, including QUINOLINE, cycloheximide, and TDZD-8 (Fig. 6C). The bubble plot further illustrates the gene ratios (GeneRatio) of small molecules associated with LGALS3, MYO1D, and EEF2. It can be observed that compounds like QUINOLINE and cycloheximide occupy significant positions within these gene networks (Fig. 6D). The gene-small molecule network map demonstrates the interactions between LGALS3, MYO1D, and EEF2 with various small molecules. Red hexagons represent the key genes studied, and blue nodes represent the small molecules associated with these genes. The network shows multiple small molecules, such as cycloheximide, QUINOLINE, and topotecan, interacting with these genes through complex relationships, revealing potential biological processes in which these genes might be involved (Fig. 6E).

Fig. 6.

Immune cell composition and molecular enrichment. (A) Box plot showing differences in immune cell composition between treatment and control groups. (B) Heatmap displaying correlations between MYO1D, LGALS3, EEF2, and immune cell components. (C) Bar plot showing small molecules associated with the genes and their adjusted p-values. (D) Bubble plot illustrating gene ratios of small molecules related to the genes. (E) Gene-small molecule network map showing interactions between the genes and associated small molecules

Protein-Small molecule binding affinity analysis

Figure 7 presents the molecular docking and binding affinity analysis between EEF2, LGALS3, MYO1D, and three small molecules (cycloheximide, QUINOLINE, topotecan). Figure 7A shows the binding model of EEF2 with cycloheximide. The left panel displays the 3D docking structure of EEF2 with cycloheximide, while the right panel shows the surface representation, highlighting the interactions of amino acid residues. The binding affinity is -7.7 kcal/mol, indicating a strong interaction. Figure 7B presents the binding model of EEF2 with QUINOLINE. The left panel shows the 3D docking structure of EEF2 with QUINOLINE, with the right panel providing the surface representation, emphasizing the interactions between amino acids and the small molecule. The binding affinity of -6.4 kcal/mol suggests a stable interaction. Figure 7C illustrates the binding model of EEF2 with topotecan. The left panel displays the 3D docking structure of EEF2 with topotecan, and the right panel presents the surface view. The binding affinity is -7.6 kcal/mol, demonstrating a strong binding interaction. Figure 7D shows the binding model of LGALS3 with cycloheximide. The left panel presents the 3D docking structure, and the right panel shows the surface view, revealing key amino acid interactions. The binding affinity is -6.6 kcal/mol, indicating a relatively strong interaction. Figure 7E displays the binding model of LGALS3 with QUINOLINE. The left panel shows the 3D docking structure, and the right panel presents the surface representation. The binding affinity of -5.1 kcal/mol reflects a weaker interaction. Figure 7F presents the binding model of LGALS3 with topotecan. The left panel shows the 3D docking structure, and the right panel illustrates the surface view. The binding affinity is -7.7 kcal/mol, indicating a strong interaction. Figure 7G depicts the binding model of MYO1D with cycloheximide. The left panel shows the 3D docking structure of MYO1D with cycloheximide, and the right panel provides the surface view. The binding affinity of -7.6 kcal/mol demonstrates a strong binding interaction. Figure 7H shows the binding model of MYO1D with QUINOLINE. The binding affinity is -5.7 kcal/mol, indicating a weaker interaction. Figure 7I presents the binding model of MYO1D with topotecan. The binding affinity of -8.4 kcal/mol indicates a very strong interaction, suggesting potential application value in drug design.

Fig. 7.

Protein-small molecule binding affinity analysis. (A-I) Molecular docking analysis showing binding models and affinities between EEF2, LGALS3, MYO1D, and small molecules (cycloheximide, QUINOLINE, topotecan)

Discussion

PCa is one of the most common cancers among men worldwide, particularly in regions such as North America, Europe, and parts of Asia. According to global cancer statistics, PCa is the second leading cause of cancer-related death in men, following lung cancer. Although PCa typically progresses slowly, early detection and treatment can significantly improve survival rates. However, for advanced and metastatic PCa, particularly castration-resistant PCa (CRPC), the therapeutic options are limited, and the prognosis for patients is poor. New biomarkers and therapeutic targets are urgently needed to improve treatment outcomes and prolong survival [29, 30].

In recent years, with the advancement of molecular biology techniques, increasing attention has been paid to the molecular mechanisms underlying PCa, particularly those related to cell signaling, tumor microenvironment, and immune evasion. Among these studies, exosomes have emerged as crucial mediators of intercellular communication, playing a significant role in the progression, metastasis, and immune modulation of PCa. Exosomes, by carrying molecules such as miRNA, mRNA, and proteins, regulate the interactions between tumor cells and the surrounding microenvironment, thereby promoting tumor growth and metastasis. Therefore, investigating exosome-related genes and pathways not only helps to reveal the potential pathogenesis of prostate cancer but also provides new directions for early diagnosis and targeted therapy [31, 32].

This study presents a comprehensive analysis of batch effect correction, DEG identification, functional enrichment, and model evaluation, with a particular focus on the relationship between differentially expressed genes and exosome-related genes. The results demonstrated a systematic approach to control batch effects, elucidate potential biomarkers, and develop predictive models for disease outcomes.

PCArevealed significant batch effects prior to correction, as evidenced by the wide dispersion of samples along the PC1 and PC2 axes. These batch effects, often observed in multi-cohort studies, can obscure true biological signals and compromise downstream analyses. After batch effect correction, the PCA plot indicated improved clustering, suggesting that the variability between cohorts had been effectively minimized. Such corrections are crucial for ensuring that observed gene expression changes reflect biological rather than technical variation, a concern that is increasingly important in genomics research where datasets from different platforms or batches are commonly integrated [33].

Our analysis identified 49 common genes between DEGs and exosome-related genes, which represent 36.8% of DEGs and 59.6% of exosome-related genes. This overlap is intriguing, as it suggests a functional relationship between differential expression and exosome involvement. In the context of prostate cancer, these genes could be key drivers in the tumor microenvironment, where exosomes are known to influence cancer progression and metastasis. Notably, genes such as EEF2, LGALS3, and MYO1D, identified in our study, have been previously implicated in prostate cancer development. Exosomes, known for their roles in intercellular communication and molecular transport, have been implicated in numerous biological processes, including immune response modulation, cancer progression, and metastasis. The identification of these shared genes lays a foundation for further investigations into the role of exosomes in disease pathophysiology, particularly in the context of the differential gene expression profiles observed between experimental and control groups [34]. Notably, genes such as ANXA2, TSPAN1, andEEF2, which were identified as key contributors to model performance, have been previously linked to cellular signaling and exosome biogenesis [35, 36].

Gene Ontology (GO) and Gene Set Enrichment Analysis (GSEA) revealed significant functional enrichment in terms related to “regulation of endocytosis,” “metal ion transport,” and “sodium ion transport,” which may contribute to the observed differential gene expression. These processes are fundamental to cellular homeostasis and have important implications in cancer biology, where dysregulated ion transport and endocytosis often contribute to tumor growth, drug resistance, and metastasis [37]. The pathways enriched in “Staphylococcus aureus infection” and “complement and coagulation cascades” further highlight the immune-related functions of these DEGs, suggesting that immune modulation via exosome-mediated transfer of molecular signals could be an underlying mechanism.

GSEA also uncovered key enriched pathways related to chemical carcinogenesis and DNA damage, which are consistent with the known role of exosomes in cancer progression. The identification of these pathways underscores the potential for exosome-related genes to serve as biomarkers for cancer diagnosis and prognosis, as well as potential therapeutic targets [38, 39].

The use of LASSO regression, random forest, and support vector machine (SVM) methods for feature selection highlighted the importance of genes such as EEF2, LGALS3, and MYO1D in the predictive models. The combination of these three genes resulted in a robust model, with an AUC of 0.886, demonstrating excellent discriminatory power between experimental and control groups. This high performance is promising for future clinical applications, where these genes may serve as biomarkers for disease stratification, prognosis, or treatment response. The identification of these genes aligns with previous studies that have implicated them in tumorigenesis, immune modulation, and cellular signaling [40, 41].

The performance of the risk score system, developed from the selected genes, is particularly noteworthy. With an accuracy of 0.81 and a clear calibration curve, the model offers a reliable method for disease risk prediction. The inclusion of these exosome-related genes could potentially improve the precision of existing predictive models in oncology, especially in personalized medicine where gene expression profiling is used to guide therapeutic decisions [42].

Our analysis of immune cell composition revealed significant differences between the treatment and control groups, with alterations observed in key immune cell types, including CD8 + T cells, B cells, and regulatory T cells. These changes are consistent with the growing body of evidence suggesting that exosomes can modulate immune responses, influencing the tumor microenvironment in prostate cancer and potentially impacting therapeutic outcomes [43]. In prostate cancer, exosomes have been shown to play a pivotal role in immune evasion, making them a critical factor in immune modulation [44]. The positive and negative correlations observed between MYO1D, LGALS3, and EEF2 with various immune cell subsets further suggest that these genes could be integral to immune regulation in cancer.

The gene-small molecule interaction network provided insights into potential therapeutic compounds associated with the selected genes. Small molecules like QUINOLINE and cycloheximide, identified as significant interactors with MYO1D, LGALS3, and EEF2, have been studied for their roles in cancer treatment and immune modulation. Particularly, these small molecules may have a profound impact on prostate cancer by disrupting exosome-mediated signaling pathways, thereby enhancing treatment response and immune cell infiltration [45]. The binding affinities calculated through molecular docking further strengthen the hypothesis that these compounds could potentially be repurposed for cancer therapy, particularly in combination with exosome-targeting strategies.

The molecular docking results also highlight the potential of small molecules, such as cycloheximide, QUINOLINE, and topotecan, to bind strongly with the selected genes, especially MYO1D and EEF2. These interactions may provide a basis for the development of targeted therapies that aim to modulate gene expression or interfere with exosome-mediated intercellular communication in cancer. In prostate cancer, this approach could offer a new strategy for overcoming therapeutic resistance and improving patient outcomes, particularly by combining these small molecules with exosome-targeting therapies. This approach is in line with the growing interest in using exosome-based therapies and exosome-inhibitors in oncology [46].

Despite these promising findings, several limitations should be acknowledged. First, the datasets used in this study were of limited sample size, which may affect the generalizability of the results. Second, no independent validation cohort or experimental verification was performed, and the immune cell composition analysis relied on computational inference, which may not fully capture the in vivo complexity. Finally, the therapeutic implications from molecular docking remain preliminary and require further in vitro and in vivo validation.

While this study provides valuable insights into the relationship between exosome-related genes and differentially expressed genes in cancer, focusing on their potential as biomarkers and therapeutic targets, there are several limitations that should be addressed in future work.

First, this study relies entirely on computational analysis, and the sample sizes of the datasets used were limited. This may impact the generalizability and reproducibility of the results. Since no independent validation cohort was used, the findings remain preliminary hypotheses, lacking experimental validation. Therefore, the clinical applicability and robustness of the results need to be further assessed. To improve the generalizability of the findings, future studies should incorporate larger sample sizes and experimental validation to evaluate the identified genes’ performance across different populations, particularly in various clinical stages and treatment conditions.

Furthermore, although this study used a bioinformatic approach to analyze the relationship between exosome-related and differentially expressed genes, there is no experimental wet-lab validation (such as cell-based assays or animal models) of the specific roles these genes play in cancer progression. The immune cell composition analysis, which was computationally inferred, may not fully reflect the in vivo immune responses, particularly in the prostate cancer immune microenvironment. Thus, future research should use experimental methods to validate the role of these genes in immune regulation and the tumor microenvironment.

Additionally, while the molecular docking analysis in this study identified potential small-molecule interventions, these results remain speculative. Molecular docking predictions require further experimental validation, such as surface plasmon resonance (SPR) or isothermal titration calorimetry (ITC), to confirm the binding characteristics of small molecules to the identified genes and their potential therapeutic implications.

Despite these limitations, this study lays the foundation for future investigations into exosome-related genes in early diagnosis, prognosis, and targeted therapy. Future studies should incorporate clinical cohort data, patient-derived xenograft (PDX) models, and clinical trials to further validate the therapeutic efficacy and safety of these small molecules and genes in targeting exosome-related pathways in prostate cancer.

Conclusion

This study provides preliminary insights into the relationship between exosome-related genes and differentially expressed genes in cancer and demonstrates their potential as biomarkers and therapeutic targets. However, these findings must be validated through wet-lab experiments and clinical validation to ensure their clinical relevance.

Author contributions

Shengnan Huang and Chunyan Hu contributed to the conception and design of the study, data analysis, and manuscript writing. Zhongqi Guo, Xiaojing Zhang, Tiantian Wu, and Yuqing Zeng participated in data collection and analysis. Chen Qing and Shulin Cheng provided guidance on the study design and data interpretation. Shulin Cheng also contributed to the acquisition of funding for the project. All authors read and approved the final manuscript.

Funding

This study was supported by the Sichuan Provincial Health and Family Planning Commission (Project No. 18PJ456) and the Nanchong Municipal Science and Technology Bureau (Project No. 16YFZJ0046).

Data availability

The gene expression datasets analyzed in this study were obtained from the Gene Expression Omnibus (GEO) database, specifically from the datasets GSE32448, GSE46602, GSE69223, and GSE6956. These datasets are publicly available and can be accessed through the GEO website ( [https://www.ncbi.nlm.nih.gov/geo/](https:/www.ncbi.nlm.nih.gov/geo) ). The processed data generated during this study, including the batch-corrected gene expression values and differential gene expression results, can be obtained from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study does not involve human participants, animal experiments, or other procedures requiring ethical approval.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.