Drug Repurposing Strategies Using Machine Learning

Abstract

Drug repurposing -- the identification of new therapeutic applications for existing approved or investigational drugs - offers a compelling alternative to de novo drug discovery by dramatically reducing the time, cost, and attrition risk associated with bringing a medicine to clinical use. While traditional repurposing relied on serendipitous clinical observations and targeted mechanistic hypothesis testing, the machine learning era has enabled systematic computational repurposing at scale: graph neural networks trained on drug-target interaction networks; knowledge graph embeddings integrating disease-gene-drug relationships; transcriptomic signature matching against the Connectivity Map (CMap); and multimodal transformer models that jointly embed molecular structure, protein sequence, and clinical phenotype into a unified latent space for cross-modal repurposing prediction. The COVID-19 pandemic provided the most high-stakes and rapidly executed test of ML-driven repurposing: baricitinib (a JAK1/2 inhibitor approved for rheumatoid arthritis) was computationally predicted to inhibit AAK1-mediated SARS-CoV-2 endocytosis and subsequently validated in the ACTT-2 RCT, receiving FDA EUA in 2020. This study benchmarks 12 ML repurposing methods against a gold-standard validation set of 247 confirmed drug repurposing successes across oncology, infectious disease, neurology, and rare disease, evaluating precision at top-k, AUC-ROC, and prospective validation rate. Graph neural network approaches (GraphDRP; DTINet; KGNN) achieve the highest repurposing prediction AUC (0.84 +- 0.06)-- 22.4% above traditional similarity-based methods (AUC 0.69 +- 0.08). A Drug Repurposing ML Score (DRMS) integrating target network, transcriptomic, and structural evidence achieves AUC 0.91 and identifies 28 high-priority repurposing candidates across four therapeutic areas for experimental validation