ATHENA: A deep learning–based AI for functional prediction of genomic mutations and synergistic vulnerabilities in prostate cancer

Abstract Identifying functional mutations that drive therapy resistance remains a major challenge in prostate cancer. Large-scale sequencing often produces extensive lists of mutations but provides limited insight into which alterations are functionally relevant. To overcome this gap, we developed ATHENA (Attention-based Therapeutic Network Analyzer), a deep learning–based AI framework that predicts the functional impact of genomic mutations and reveals their synergistic vulnerabilities. Integrated with our RNA/DNA-informed variant discovery pipeline OncoVar, ATHENA models nonlinear dependencies among mutations to distinguish driver events from passenger variants. Trained on large multi-cohort datasets and interpreted using SHAP analysis, ATHENA not only stratifies patients by clinical outcomes but also predicts which specific mutations alter tumor behavior and therapy response, enabling direct validation through base editing experiments. Applied to prostate cancer progression models, the OncoVar–ATHENA framework identified stage-specific driver signatures across castration-resistant, AR-variant–driven, and metastatic disease, and uncovered cooperative interactions such as SYVN1–STC2 that promote tumor proliferation. By moving beyond simple mutation identification, ATHENA enables functional prediction of genomic interactions. This approach accelerates the discovery of actionable targets and provides a foundation for rational design of next-generation combination therapies in advanced prostate cancer. Competing Interest Statement P.M. served as a scientific consultant to Accutar Biotechnology, Inc. No other authors have COI to disclose.