Hyperbolic heterogeneous graph learning for adverse drug reaction prediction, candidate prioritization, and pathway-based interpretation in pharmacovigilance

Abstract

Adverse drug reaction (ADR) prediction remains challenging because spontaneous reporting data leave many exposure unit–ADR associations unobserved and provide limited support for candidate prioritization and mechanistic interpretation. We developed MEG-HHGN, a mechanistic evidence-guided hyperbolic heterogeneous graph neural network. It represents exposure units as combinations of standardized drug names and routes of administration and integrates spontaneous reports with compounds, drug targets, ADR-related targets, and protein–protein interactions. Using 73,530 spontaneous reports collected in Jiangsu Province from 2010 to 2023, we evaluated MEG-HHGN against eight baseline methods through five-fold cross-validation. MEG-HHGN achieved AUC/AP/MCC values of 0.9550/0.9537/0.7917 in known-exposure-unit completion and 0.9090/0.9064/0.6587 in side-information-supported exposure-unit cold-start evaluation. Global scoring of unobserved exposure unit–ADR pairs produced a ranked candidate pool for pharmacovigilance review. Representative cases involving intravenous vinorelbine tartrate–myelosuppression and oral aspirin–proteinuria illustrated PPI-bridged and shared-target pathways, respectively. MEG-HHGN integrates route-aware prediction, candidate prioritization, and pathway-based interpretation in a unified framework to support signal triage, mechanistic hypothesis generation, and evidence review in post-marketing pharmacovigilance.