Genetically supported drug target prioritization for rare diseases

RareGPS is a machine-learning framework prioritizing drug targets for rare and uncommon diseases, integrating 11 genetic, clinical, and experimental evidence sources. It uses the full distribution of genetic associations across allele-frequency bins in an allelic-series model. Across 161 phenotypes, RareGPS outperforms existing resources for predicting drug indications and clinical trial progression; top 1% targets show 58-fold higher likelihood of advancing from nonindicated to phase IV and 8-fold from phase I to IV versus the middle 50%. We validated RareGPS using prescriptome analyses in two million patients and an independent literature evaluation tool (AMELIE). We publish predictions for 3,021,965 gene-phenotype pairs.