Knowledge-graph-enhanced multi-scale modeling for drug-drug interaction prediction

Drug-drug interaction (DDI) prediction is crucial for understanding combined medication effects and preventing adverse reactions. Traditional machine learning methods rely on handcrafted features and lack generalization, while existing deep learning approaches often fail to capture global and multi-scale drug relationships. To overcome these limitations, we propose ALG-DDI, a multi-scale feature fusion model that integrates three types of drug information: attribute (intrinsic drug structure), local correlations (with proteins and diseases), and global semantic information from the medical knowledge graph PrimeKG. We encode these using attribute masking, the idea of RGCN and GraphSAGE, and ComplEx, respectively. A transformer encoder with attention mechanism then fuses these multi-scale representations. The resulting drug pair vector is fed into a fully connected network for DDI prediction, which we also extend to DDI event prediction. Extensive evaluations on three datasets-including comparative experiments, cross-validation, retrospective analysis, and case studies-demonstrate that ALG-DDI outperforms existing state-of-the-art methods.