Graph Neural Networks with Heterogeneous Message Passing for Multi-Scale Drug-Drug Interaction Prediction

Abstract

Adverse drug-drug interactions (DDIs) cause approximately 195,000 hospitalizations annually in the US alone. Existing computational DDI prediction methods operate at a single biological scale — either molecular fingerprints or protein targets — missing the complex multi-scale mechanisms underlying polypharmacy risks. We present HetDDI-GNN, a heterogeneous graph neural network operating on a unified knowledge graph integrating molecular structures (1.2M atoms), protein-protein interactions (18K nodes), metabolic pathways (2.1K reactions), and clinical co-prescription data (3.4M records). HetDDI-GNN achieves AUROC of 0.952 on DrugBank DDI prediction and 0.918 on an external clinical validation set from the FDA Adverse Event Reporting System, outperforming single-scale baselines by 4-8%.