Google DeepMind releases AlphaGenome Atlas: 1-petabyte catalogue of effect predictions for ~9 billion human DNA variants, with new AVI score outperforming CADD

DeepMind launched AlphaGenome Atlas, a 1-PB precomputed catalogue covering ~9 billion possible human single-nucleotide variants, plus an AVI impact score that beat CADD on rare-disease benchmarks and helped reclassify a DNM1 epilepsy splice variant as likely…

Google DeepMind released AlphaGenome Atlas, a one-petabyte dataset of precomputed molecular effect predictions for roughly 9 billion possible single-nucleotide variants in the human genome, with about 27,000 prediction values per variant — more than 30 times the size of the AlphaFold Database. It also introduces the AlphaGenome Variant Impact (AVI) score, a small neural network combining AlphaGenome regulatory predictions, AlphaMissense, and evolutionary conservation features, using 18 inputs versus CADD's 150+. DeepMind reports best-in-class AVI performance on variant pathogenicity and rare-disease benchmarks: it ranked causal variants among the top 50 candidates in 29.5% of solved GREGoR cases versus 12.5% for CADD, and performed better on noncoding variant classification. The atlas includes per-variant feature attributions and over 2,500 recurrent DNA sequence motifs. Early users at the Broad Institute, University of Exeter, and Stowers Institute demonstrated rare-disease variant reprioritization; in a GREGoR epilepsy case, AVI elevated a previously unclear DNM1 splice variant that lab experiments confirmed as likely disease-causing (linked to epileptic encephalopathy per Broad collaborators). A UK Biobank analysis of 54,000+ genomes surfaced 22% more noncoding-variant associations using the atlas. Access is currently free/noncommercial via a web portal, API, and a Google Antigravity skill, with commercial access (via Google Cloud per one report) coming soon; the tool is not clinically approved.

• Release: Google DeepMind's AlphaGenome Atlas, announced/reported 2026-09-08 to 2026-09-09
• Scale: 1-petabyte dataset of precomputed molecular effect predictions for ~9 billion possible human single-nucleotide variants, ~27,000 prediction values per variant
• Size comparison: more than 30x larger than the AlphaFold Database
• AVI score: small neural network combining AlphaGenome, AlphaMissense, and evolutionary conservation features; 18 input features versus CADD's 150+
• Benchmark: ranked causal variants in the top 50 for 29.5% of solved GREGoR cases versus 12.5% for CADD; also outperformed existing tools on clinically classified and noncoding variants
• Case study: a GREGoR epilepsy case — a DNM1 splice variant, previously unclear, was elevated by AVI and confirmed by lab experiments as likely disease-causing; Broad Institute collaborators linked it to epileptic encephalopathy
• UK Biobank: analysis of 54,000+ genomes found 22% more noncoding-variant associations using the atlas
• Dataset extras: per-variant feature attributions and over 2,500 recurrent DNA sequence motifs