Google DeepMind published research Friday describing AlphaGenome, a model that predicts disease risk from non-coding DNA variants with accuracy exceeding prior genome-wide association methods on several benchmarks. The work appears in Nature alongside replication studies from the Broad Institute and deCODE Genetics.
Non-coding regions comprise roughly 98 percent of the human genome and regulate when genes activate. Most clinical genetic tests focus on protein-coding mutations; AlphaGenome targets regulatory variants implicated in type 2 diabetes, rheumatoid arthritis, and inflammatory bowel disease.
Methodology
Researchers trained the model on chromatin accessibility maps, transcription factor binding profiles, and population-scale biobank data. AlphaGenome assigns pathogenicity scores to variants never observed in clinical databases, enabling hypothesis generation for laboratory follow-up.
Clinical Prospects
Physicians cautioned that predictive scores require prospective validation before guiding treatment. Britain's National Health Service said it will fund a pilot comparing AlphaGenome prioritization against standard polygenic risk scores in 50,000 participants.
Equity Concerns
Training data skew toward European-ancestry biobanks. DeepMind released a calibration toolkit and committed funding to expand African and South Asian representation in partner datasets through the Genomics England and All of Us programs.
Commercialization
Isomorphic Labs, DeepMind's drug-discovery affiliate, will use AlphaGenome outputs to triage targets for small-molecule programs. Verily paused a competing internal project and said it will evaluate licensing terms.
Open Science
The model weights remain restricted, but scoring APIs are available to academic researchers under attribution licenses. Open-science advocates criticized the limitation; DeepMind cited misuse risks involving synthetic pathogen enhancement research as justification for controlled access.
