Google DeepMind AlphaGenome Atlas Maps All 9 Billion Possible DNA Variants
Google DeepMind's AlphaGenome Atlas maps 9 billion DNA variants with AVI scores, accelerating rare disease and population genetics research to aid scientists.
Summary
Google DeepMind launched AlphaGenome Atlas on September 8, 2026, precomputing molecular effects for all 9 billion possible single nucleotide variants in the human genome. The one petabyte dataset, more than 30 times larger than the AlphaFold Database, provides thousands of predictions per variant across hundreds of human and mouse cell types and tissues. Its AlphaGenome Variant Impact score combines AlphaGenome with protein focused AlphaMissense, ranking changes across the protein coding 2% and noncoding 98% of the genome. Scores link to biological drivers including expression, splicing, chromatin accessibility and conservation, plus more than 2,500 recurring DNA motifs. Google DeepMind says its testing shows leading performance across pathogenicity and rare disease benchmarks.
Laura Covill and Anne O’Donnell-Luria of the Broad Institute, working with the GREGoR Consortium, used AVI to identify a previously overlooked DNM1 variant linked to epileptic encephalopathy. AlphaGenome predicted an incorrect splice site and abnormal protein extension, which experiments validated alongside similar nearby variants. University of Exeter researcher Gareth Hawkes analyzed more than 54,000 UK Biobank participants and found 22% more noncoding associations by grouping rare variants through predicted effects, including regulators of aging linked PLA2G7 and oxygen sensor EGLN1. Restricting hundreds of millions of noncoding BMI variants to the predicted top 1% identified 19 genetic regions. Julia Zeitlinger and Melanie Weilert at the Stowers Institute used Atlas motifs to distinguish transcription factors affecting DNA accessibility from those also controlling gene activity.
Atlas is free for noncommercial use through its website, AlphaGenome API and Google Antigravity, with commercial Google Cloud access coming soon. The AlphaGenome base model remains available academically through GitHub and the API, and commercially through Model Garden. Atlas is not clinically validated or approved and cannot replace professional medical advice, diagnosis or treatment.
Positives
- 9 billion precomputed variant predictions create a genome wide map that would be impractical to reproduce through laboratory testing.
- AVI ranks changes across both the protein coding 2% and noncoding 98% of the human genome.
- Experimental screens validated the predicted splicing defect and abnormal protein extension caused by a DNM1 variant.
- More than 54,000 UK Biobank genomes yielded 22% more noncoding associations when variants were grouped by predicted molecular effects.
- The predicted top 1% of noncoding BMI variants narrowed hundreds of millions of candidates to 19 genetic regions.
Risks & concerns
- AlphaGenome has not been clinically validated or approved, and Atlas cannot provide medical advice, diagnosis or treatment.
- Computational predictions still require targeted experimental validation before researchers can confirm biological effects.
- Hundreds of millions of harmless noncoding changes create statistical noise that can conceal trait and disease associations.
- Commercial Atlas access is not yet available and is planned for Google Cloud at an unspecified future date.
- Google DeepMind describes Atlas as a baseline, indicating its genome maps remain subject to improvement as the underlying models advance.