An interpretable molecular framework for predicting cancer driver missense mutations

Abstract

Missense mutations play a critical role in human disease, contributing to both inherited disorders and cancer. However, accurately predicting their functional impact--particularly for cancer driver mutations--remains a major challenge due to limited validated labels and the complex molecular basis of oncogenesis. Here, we systematically characterized over 120,000 missense variants across pathogenic, benign, driver, passenger, recurrent somatic, and common population classes, using a comprehensive set of mechanistically grounded molecular features. By assessing the statistical burden of variations, we demonstrated that these features effectively discriminate among diverse variant classes and reveal a consistent enrichment of functional sites, structural integrity, and biophysical changes in pathogenic and driver mutations. Building on these insights, we developed MutaPheno, an interpretable framework for predicting the functional consequences of missense mutations. The model integrates 34 molecular-level features, encompassing structural, functional, physicochemical, and contextual descriptors, using a random forest algorithm. Trained exclusively on pathogenic and benign variants, MutaPheno achieved strong accuracy in predicting cancer driver mutations, outperforming both cancer-specific and general pathogenicity tools, while also demonstrating superior robustness when tested on unseen proteins. Our findings highlight the shared mechanisms between pathogenic and driver mutations and emphasize the role of molecular features in improving variant interpretation. MutaPheno provides a transparent and generalizable tool that can facilitate driver discovery and the development of targeted therapies.