Residue-Level Multiview\nDeep Learning for ATP Binding\nSite Prediction and Applications in Kinase Inhibitors

Accurate identification of adenosine triphosphate (ATP)\nbinding\nsites is crucial for understanding cellular functions and advancing\ndrug discovery, particularly in targeting kinases for cancer treatment.\nExisting methods face significant challenges due to their reliance\non time-consuming precomputed features and the heavily imbalanced\nnature of binding site data without further investigations on their\nutility in drug discovery. To address these limitations, we introduced\nMultiview-ATPBind and ResiBoost. Multiview-ATPBind is an end-to-end\ndeep learning model that integrates one-dimensional (1D) sequence\nand three-dimensional (3D) structural information for rapid and precise\nresidue-level pocket-ligand interaction predictions. Additionally,\nResiBoost is a novel residue-level boosting algorithm designed to\nmitigate data imbalance by enhancing the prediction of rare positive\nbinding residues. Our approach outperforms state-of-the-art models\non benchmark data sets, showing significant improvements in balanced\nmetrics with both experimental and AI-predicted structures. Furthermore,\nour model seamlessly transfers to predicting binding sites and enhancing\ndocking simulations for kinase inhibitors, including imatinib and\ndasatinib, underscoring the potential of our method in drug discovery\napplications.