CoFEND Predicts Drug Interactions for New Drugs

Predicting drug-drug interactions (DDIs) for new drugs, known as the cold-start problem, is vital for patient safety and minimizing adverse reactions. Existing methods often fall short by considering only partial relationships, leading to incomplete similarity modeling, and by separating similarity computation from DDI prediction. This research proposes CoFEND (Cross-Modal Fusion End-to-End Learning Network) to overcome these limitations. CoFEND integrates diverse multimodal information, including molecular structures, enzymes, and transporters, to build comprehensive drug-centered knowledge graphs. It then employs a four-channel graph autoencoder to seamlessly fuse cross-modal similarities within a single, end-to-end learning framework, ensuring better alignment between similarity computation and DDI prediction. Additionally, CoFEND offers a novel two-stage interpretability scheme that precisely identifies key factors for both the drug causing the interaction (perpetrator) and the drug affected by it (victim), providing deeper mechanistic insights. Experiments on real datasets confirm CoFEND's superior prediction accuracy and interpretability.