MultiAlloDriver: a multi-model method to predict and identify cancer driver mutations

Abstract

A minority of driver mutations in cancer significantly alter protein structure and key functionalities, thereby driving cancer progression. Consequently, the prediction and identification of driver mutations hold critical implications for targeted cancer therapy. This study introduces MultiAlloDriver, a novel multi-modal machine learning model based on an attention mechanism, which for the first time, incorporates protein surface information. By integrating information from three dimensions-protein sequence, structure, and surface-the model achieves high-accuracy driver target prediction with an accuracy exceeding 93%. Notably, we utilized this tool to predict and identify the driver effect of the F90S mutation in the PTEN tumor suppressor gene, uncovering mutations associated with cancer signaling mechanisms. Overall, MultiAlloDriver contributes to elucidating the underlying mechanisms of cancer development and progression, providing a robust framework for the identification of driver targets.