From Pathways to Prediction: A Comparative Machine Learning Framework for Prostate Cancer Survival

Background Prostate cancer shows substantial clinical and molecular heterogeneity, limiting the prognostic accuracy of conventional clinicopathologic models. Single-gene alterations and tumor mutational burden provide limited prognostic discrimination. Pathway-level genomic abstraction may better capture cumulative oncogenic disruption. Methods Genomic and clinical data from 2,231 prostate adenocarcinoma patients were analyzed. Somatic mutations were mapped to 11 cancer-related signaling pathways. A composite pathway-based risk score integrating pathway burden, p53 pathway status, and high-risk co-alterations was developed. Prognostic performance was evaluated using survival analysis, Cox regression, time-dependent receiver operating characteristic curves, and machine-learning models. Genomic generalizability was assessed in an independent external cohort. Results The proposed score stratified patients into distinct risk groups with significantly different overall survival (log-rank p < 0.0001). Each one-point increase was associated with a 31% higher mortality risk (hazard ratio 1.31, 95% confidence interval 1.21–1.42). The model showed moderate discrimination (concordance index 0.5897) and more stable predictive performance than tumor mutational burden alone. Machine-learning models achieved similar performance, and feature importance analysis identified p53 pathway disruption and pathway burden as key predictors. Conclusions The proposed framework is a mutation-based genomic risk stratification tool derived from targeted sequencing data that provides interpretable prognostic stratification with performance comparable to machine-learning models.