Unveiling Novel Potential Biomarkers in Prostate Cancer: A Comprehensive Bioinformatics and Machine Learning Approach

Prostate cancer (PCa) exhibits marked biological heterogeneity, resulting in substantial differences in disease recurrence and clinical outcomes and highlighting the need for improved molecular risk stratification. In this study, we developed a machine learning–driven prognostic framework to identify transcriptomic features associated with disease-free survival (DFS) and stratify patients according to recurrence risk. Differentially expressed genes between prostate cancer and normal prostate tissue were initially screened, and their prognostic relevance was evaluated in the TCGA-PRAD cohort. A prognostic signature was subsequently developed using LASSO-Cox and multivariate Cox regression, generating a patient-level risk score for recurrence stratification. The final prognostic model comprised 33 genes and effectively separated patients into high- and low-risk groups, with significantly different DFS outcomes. The model demonstrated good prognostic discrimination, as indicated by the concordance index and time-dependent area under the receiver operating characteristic curve (AUC), and significantly separated patients into high- and low-risk groups. Random Survival Forest analysis further supported the prognostic relevance of the selected features, while SHAP analysis identified the relative contribution of individual genes to patient-level risk prediction. Drug–gene interaction analysis further identified a subset of the 33 signature genes targeted by FDA-approved drugs, providing hypothesis-generating information regarding potential therapeutic relevance. The results support the potential of the 33-gene signature for molecular risk stratification and recurrence prediction in PCa. However, the model was developed and evaluated within the TCGA-PRAD cohort and requires independent external and prospective validation before clinical translation.