Abstract
Drug repurposing can accelerate therapeutic development, yet existing computational approaches typically rely on a single evidence type and lack validation across biological scales. Here we present LinkD, an integrated framework combining structure-informed latent diffusion modeling for drug–target interaction prediction with entropy-aware selectivity scoring, large-scale cellular phenotype validation, and real-world clinical evidence from population-scale electronic health records (EHRs). LinkD comprises four modules. LinkD-DTI, a diffusion-based affinity prediction model, ranks first on 8 of 9 benchmarks spanning BindingDB, Davis, and KIBA (RMSE 0.447–0.699) under random and cold-start settings. LinkD-Select computes entropy-based selectivity scores that distinguish on-target from off-target interactions, recovering 95.3% of known selective drug–target pairs with concordant docking support. LinkD-Pheno validates predicted molecular effects through concordance of drug sensitivity and CRISPR gene-dependency profiles across 960 cancer cell lines. LinkD-Agent ( https://linkd-agent.onrender.com/ ) is an interactive AI system enabling transparent, multi-evidence hypothesis generation without programming expertise. Population-level analyses of 11.5 million individuals from Mount Sinai and UK Biobank EHR cohorts show that LinkD-prioritized drug–disease associations are enriched for protective clinical signals, including the beta-blockers carvedilol (OR: 0.66, 95% CI: 0.59–0.73) and propranolol (OR: 0.77, 95% CI: 0.61–0.97) associated with reduced prostate cancer risk—corroborated by ADRB2 docking and in vitro growth inhibition. These results establish LinkD as a scalable, multi-scale framework for systematic drug repurposing.
Full Text
The Full Text of this preprint is available as a PDF (19.9 MB). The Web version will be available soon.