Similarity-based profiling of pharmacovigilance data for drug safety pattern discovery

Spontaneous adverse event (AE) reporting systems enable large-scale pharmacovigilance but are typically analyzed as isolated drug-event pairs. Here, we constructed a drug-drug similarity network at an ingredient-level, using 968,966 reports from the Korea Adverse Event Reporting System (KAERS; 2020-2024). After preprocessing, 382,530 reports involving 1,058 ingredients and 3,749 AE events in MedDRA preferred terms (PTs) were retained. Significant ingredient-event signals were identified using proportional reporting ratio, reporting odds ratio, χ², and information component. Pairwise similarity between ingredients was quantified using a hypergeometric test based on shared significant PTs with a false discovery rate ingredients and 1,111 PTs, resulting in a network of 150 ingredients and 1,267 edges. Community detection revealed modules that recapitulated known pharmacological classes, including antineoplastic agents and contrast media, and exhibited clinically coherent safety profiles. Notably, cross-class clustering, including statins with anti-infective and anti-inflammatory agents, suggested shared downstream biological effects beyond primary indications. These findings demonstrate that a signal-based drug similarity network derived from spontaneously reported data can capture clinically meaningful safety patterns and reveal latent relationships across therapeutic classes, thereby providing a scalable approach to pharmacovigilance and hypothesis generation.