RxPairEvid and GraphFusion: Multimodal pharmacovigilance and mechanistic knowledge-graph evidence for drug-drug interaction prediction.

Drug-drug interactions (DDIs) are a major cause of preventable adverse drug reactions, yet experimental screening cannot keep pace with the growing number of possible drug pairs. We present RxPairEvid, a large-scale multimodal evidence bank containing 1.07 million labelled drug pairs. The resource harmonizes chemical, target, pathway, perturbational, safety, and mechanistic knowledge-graph evidence in a unified relational framework. On top of this resource, we develop RxPairEvid-GraphFusion, a regularized neural model that combines pair-level tabular evidence with concatenated mechanistic graph embeddings for the two drugs. We evaluate the framework under both a conventional random drug-pair split and a strict drug-disjoint cold-start split, and we explicitly distinguish an early-stage mechanistic screening setting from a broader post-market signal-prioritization setting. On the random split, GraphFusion achieves AUROC 0.988, AUPRC 0.968 and F1 0.900; on the drug-disjoint split it attains AUROC 0.761, AUPRC 0.590 and F1 0.453, reflecting the expected difficulty of emerging-drug prediction. Calibration and threshold analyses, together with SHAP attributions, identify the most influential mechanistic evidence channels. RxPairEvid-GraphFusion provides a scalable and more interpretable foundation for DDI screening and medication-safety decision support, particularly under realistic cold-start settings.