Population-scale drug target and variant prioritization with Varformer

Abstract

Human genetic variation is a powerful driver of therapeutic target prioritization, yet current models struggle to aggregate variable numbers of variants per gene and are bottlenecked by the scarcity of validated targets. We present Varformer, a multimodal transformer that uses cross-modal attention to integrate contextual representations of missense variants with gene-level features. Trained via positive-unlabeled learning on clinically validated targets across four genetic ancestry groups, Varformer achieves a precision@20 of 0.95, 0.86, and 0.48 across three benchmarks of increasing difficulty. It surpasses a logistic regression ablation by 20% and DrugnomeAI by 23% on the hardest benchmark. Cross-population rank normalization isolates targets with consistent genetic support across ancestry groups. The variant encoder also scores individual variants. Across 1,164 carrier-variant pairs in Genes & Health, Varformer ranked the known pathogenic variant in the carrier's top 20 in 57% of pairs. Varformer is available as an open-source framework for unified target and variant prioritization.