Abstract
<title>Abstract</title> <p>Exosomes play critical roles in prostate cancer (PRAD) progression. This study aimed to identify and validate exosome-related diagnostic biomarkers for PRAD using integrated bioinformatics and machine learning approaches. Transcriptomic data from TCGA-PRAD and three GEO cohorts were intersected with curated exosome-related gene sets to obtain exosome-related DEGs (Exo-DEGs). Core diagnostic genes were selected through a consensus strategy integrating LASSO, SVM-RFE, and random forest. Diagnostic performance was assessed using ROC analysis, nomogram, and decision curve analysis, while model interpretability was examined using SHAP. Immune infiltration was estimated by ssGSEA. Drug prediction used DSigDB enrichment and molecular docking to AKR1B1. Twelve Exo-DEGs were identified, yielding an eight-gene exosome-related diagnostic signature (CAV1, CLU, FGFR2, APOE, AKR1B1, OLFM4, ALB, and GATA4). The signature demonstrated excellent diagnostic accuracy (AUC = 0.966), outperforming single-gene classifiers. Among ten machine-learning classifiers, random forest achieved the best test-set performance (AUC = 0.953), with SHAP highlighting FGFR2, APOE, CAV1, and CLU as top contributors. Core genes were positively associated with multiple immune cell signatures. Molecular docking suggested stable AKR1B1 binding with cholesterol (Vina score − 8.2 kcal/mol) and warfarin (− 8.0 kcal/mol). Bulk analyses showed decreased AKR1B1 expression in tumors, while single-cell analyses revealed prominent expression in myeloid populations, particularly macrophages, linked to inflammasome signaling, EMT, and immune responses. AKR1B1 represents a promising diagnostic biomarker with functional relevance to tumor immunity and metabolism, offering potential for targeted drug development.</p>