Network-Based Drug Repositioning

Abstract

Drug repositioning -- identifying new therapeutic indications for approved drugs -- offers a faster and cheaper path to treatment than de novo drug discovery, reducing development timelines from 12-15 years to 3-5 years and costs from $2.6 billion to $300-500 million per indication. Network-based approaches that model drugs, targets, diseases, and genes as interconnected nodes in biological knowledge graphs can systematically predict repositioning candidates by identifying drugs proximal to disease modules in the interactome. However, existing network methods achieve hit rates of only 5-15% in experimental validation, limited by incomplete knowledge graphs, simplistic proximity metrics, and inability to capture polypharmacology (multi-target drug effects). This study developed and benchmarked six drug repositioning approaches -- literature-based similarity (DrugBank indication overlap), transcriptomic signature matching (CMap/L1000), network proximity (shortest path in PPI), graph embedding (node2vec on drug-target-disease graph), heterogeneous GNN link prediction (RGCN), and a proposed Multi-Evidence Network Repositioning Engine (MENRE) combining heterogeneous knowledge graph embedding (drug-target-disease-pathway-side effect, 248,000 nodes, 2.4 million edges), transcriptomic perturbation signatures, molecular structural similarity, and clinical evidence from EHR-derived drug-disease co-occurrence via a multi-relational graph transformer -- for predicting novel drug-disease associations across 1,248 approved drugs and 862 diseases. MENRE achieved the highest prediction accuracy: AUROC 0.942 on held-out known indications (vs network proximity 0.768, RGCN 0.884, CMap 0.812; p < 0.001), with precision at top-50 predictions of 38.4% (vs RGCN 24.6%, CMap 18.2%). Experimental validation of MENRE's top 20 novel predictions by cell-based phenotypic assays confirmed 7 of 20 (35%) as biologically active in the predicted disease context. Clinical validation using retrospective EHR analysis (n = 4.2 million patients) identified 3 repositioning candidates with significant protective associations (HR 0.62-0.78; p < 0.01). These results establish multi-evidence network repositioning as the most effective computational approach for systematic drug repurposing