Varformer: multimodal transformer prioritizes drug targets from population-scale variation
anshulkundaje · x · 2026-10-23
Aaron Wenteler et al. introduce Varformer, a multimodal transformer trained on population-scale missense variation to prioritize drug targets across ancestries while scoring individual variants.
Key points:
- Combines gene-level data (druggability, safety, conservation, tissue expression) with variant features (functional scores, pathogenicity)
- Trained on populations including Admixed American (2.6M variants), Non-Finnish European (10.8M), African (2.7M), and South Asian (2.2M)
- Uses a non-negative risk estimator for positive-unlabeled learning, jointly optimizing gene prioritization and variant prediction losses
- Preprint, GitHub code, and model weights released
More from Apps
- Kubernetes System Design in Small Easy Steps joins ChapterPal's free open textbook collection — burkov · 2026-10-23
- Dev calls out ChatGPT as Claude's cross-device workflow leaves it behind — sytelus · 2026-10-23
- Using one molecule as a personal benchmark: vitrify model predicts 3D structure and vibrational modes — johnowhitaker · 2026-10-23
- Liberal arts grad ships an AI-built e-commerce site selling Taihang apples overnight — yangyi · 2026-10-23
- ChatGPT Dots manages all task threads but still buggy, per Peter Yang review — FinanceYF5 · 2026-10-23
- Ex-judge Bridget Mary McCormack joins Clio to build AI for courts, not just lawyers — EJWalters · 2026-10-23