Abstract / Summary
Abstract Retrospective recall of approved drugs is the standard validation for computational drug repurposing, but it is exposed to a leakage mode that is rarely checked: drug-target annotations in public reference databases can be added after a drug's approval, so a pipeline appears to recover a target it could not have known at the time. We describe a leakage-aware validation protocol that credits a recovered drug only when the target annotation it is scored through is present in a dated 2020 release of the reference database, and that evaluates graph neural network recall only on held-out drugs. We apply it to IgA nephropathy, the leading cause of kidney failure in East Asia, where five drugs have been approved since 2021 and no computational repurposing study has been published. A four-method pipeline combines network proximity on a 953-protein module built from 30 genome-wide association loci, a relational graph convolutional network on PrimeKG, Open Targets genetic association scoring, and LINCS L1000 transcriptomic reversal. The protocol removed the pipeline's top-ranked apparent success, sparsentan, as post-approval annotation backfill; both annotation-safe known drugs, atrasentan and fostamatinib, ranked in the top 10% of 1,989 scored compounds. The consensus nominates fifteen candidates with no trial record in this disease. Four methods converge on LYN kinase inhibitors, led by nintedanib, and a transforming growth factor beta inhibitor, F351, addresses the fibrosis axis that no approved therapy reverses.