Dual-Channel Dual-Scale Interactive Learning for the Prediction of Compound-Protein Interaction

Compound-Protein Interaction (CPI) serves as an essential indicator for efficiently screening potential candidate drugs. Previous studies have typically focused on modeling CPIs either from intramolecular or intermolecular interactions, disregarding the diversity of interactions and the fine dependencies between these two types of interactions, thereby limiting the accuracy of CPI predictions. We argue that properly considering both intramolecular and intermolecular interactions allows for a more comprehensive understanding of the interactions between compounds and proteins. To this end, we propose a novel approach called Dual-channel Dual-scale Interactive learning (DDI) for CPI predictions. DDI simultaneously captures various CPI information from both intramolecular and intermolecular interactions using a dual-channel encoding structure as backbone. Furthermore, to fetch the complicated relationships between the two types of interactions, we design a dual-scale interactive learning paradigm to facilitate interactive transmission of collaborative information between intramolecular and intermolecular interactions in both local and global hyperspace (atom-level and molecule-level), enhancing the learning of each other. Finally, we predict CPIs based on the rich interaction information from dual channels. Exhaustive experimental studies on two benchmarks verify the superiority of DDI in CPI predictions.