A unified framework combining linear and 3D molecular features for robust drug-protein interaction prediction

We develop PointDPI to predict drug-protein interactions by simultaneously exploiting their linear and 3D structural features. By aligning the features, PointDPI stereoscopically recognizes molecular properties and reduces reliance on 3D structures. Local topological relationships among molecules are further preserved for avoiding distortion. PointDPI predicts key regulatory sites based on the model's gradient. We demonstrate improved performance over several state-of-the-art (SOTA) methods, including increased accuracy in dealing with unseen molecules. Four predicted drug-protein interactions (DPIs) are experimentally validated at both mRNA and protein levels, highlighting the therapeutic potential of adenosine in inflammatory diseases, ondansetron and etodolac in neurological diseases, and neuroprotective action for dopamine.