Google DeepMind launches AlphaGenome Atlas, a 1-petabyte catalog predicting the effects of ~9 billion human DNA variants

Google DeepMind released AlphaGenome Atlas (2026-09-08), a free ~1-petabyte catalog of precomputed molecular-effect predictions for roughly 9 billion possible single-nucleotide variants across the human genome, ranked by a new AlphaGenome Variant Impact (AVI)…

On 2026-09-08, Google DeepMind launched AlphaGenome Atlas, a free, roughly 1-petabyte catalog of precomputed predictions for the molecular effects of approximately 9 billion possible single-nucleotide variants across the human genome (about 3 billion base pairs, per Hacker News), spanning hundreds of human and mouse cell types per DeepMind. The Decoder adds that the Atlas holds about 27,000 prediction values per variant, and DeepMind and MarkTechPost report the dataset is more than 30 times larger than the AlphaFold Database. The catalog introduces the AlphaGenome Variant Impact (AVI) score — styled 'Variant Impact Score' by The Verge — to rank variants in both protein-coding and non-coding regulatory regions. DeepMind and MarkTechPost describe AVI as combining AlphaGenome and AlphaMissense predictions; The Decoder calls it a small neural network that also folds in evolutionary conservation features, using 18 inputs versus CADD's 150+. DeepMind and MarkTechPost claim best-in-class AVI performance on variant pathogenicity and rare disease benchmarks, while The Decoder gives head-to-head numbers: AVI ranked causal variants in the top 50 for 29.5% of solved GREGoR cases versus 12.5% for CADD. MarkTechPost adds per-variant feature attributions and over 2,500 recurrent DNA sequence motifs. The Verge reports the Atlas builds on DeepMind's AlphaGenome model released in 2025 and sits in the AlphaFold lineage alongside AlphaMissense. Early users named by MarkTechPost — the Broad Institute, University of Exeter, and Stowers Institute — demonstrated rare-disease variant reprioritization; Broad Institute researchers used the Atlas to prioritize a predicted DNM1 splice variant (linked to epileptic encephalopathy per MarkTechPost) in a previously unsolved rare disease case, and The Decoder reports lab experiments confirmed the splice effect, reclassifying the variant as likely disease-causing. An analysis of 54,000+ UK Biobank participants (54,000 genomes per The Decoder) found 22% more non-coding genetic associations, including 19 regions linked to BMI per Hacker News. Access is free for noncommercial research via a no-code web portal, the AlphaGenome API, and a skill on Google's Antigravity platform; commercial access through Google Cloud is planned, and the catalog is not clinically approved per MarkTechPost. Ars Technica adds scope and limitations: the underlying AlphaGenome system is limited to human and mouse sequences and well-studied cell types, its predictions…

• Launched 2026-09-08 by Google DeepMind as a free catalog of precomputed molecular-effect predictions for DNA variants.
• Covers ~9 billion possible single-nucleotide variants across the roughly 3 billion base pairs of the human genome (Hacker News).
• Dataset is ~1 petabyte, more than 30x larger than the AlphaFold Database (DeepMind, MarkTechPost); The Decoder reports ~27,000 prediction values per variant.
• Spans hundreds of human and mouse cell types (DeepMind).
• Introduces the AlphaGenome Variant Impact (AVI) score ranking both coding and non-coding variants: DeepMind/MarkTechPost say it combines AlphaGenome and AlphaMissense predictions; The Decoder describes a small neural network that also uses…
• Benchmarks: DeepMind/MarkTechPost report best-in-class AVI results on variant pathogenicity and rare disease benchmarks; The Decoder reports AVI ranked causal variants in the top 50 for 29.5% of solved GREGoR cases versus 12.5% for CADD.
• Includes per-variant feature attributions and over 2,500 recurrent DNA sequence motifs (MarkTechPost).
• Broad Institute collaborators validated a DNM1 splice variant linked to epileptic encephalopathy in an unsolved rare disease case; The Decoder reports lab experiments confirmed the splice effect, reclassifying the variant as likely…