Google DeepMind Releases AlphaGenome Atlas, Predictions for All 9 Billion Possible Human DNA Mutations

Google DeepMind has released AlphaGenome Atlas, a free platform containing precomputed predictions for the molecular effects of all 9 billion possible single-nucleotide variants in the human genome — every single-letter DNA change that could theoretically occur.

The dataset totals 1 petabyte, which DeepMind says is more than 30 times larger than the AlphaFold Database. It is available through a public website portal, the AlphaGenome API, and as a skill in Google Antigravity, DeepMind's agentic development environment.

What's in the Atlas

AlphaGenome Atlas builds on AlphaGenome, DeepMind's existing model for predicting how genetic variants affect biological processes such as gene expression and RNA splicing. Rather than requiring researchers to query the model variant-by-variant, DeepMind has precomputed AlphaGenome's outputs at genome-wide scale, covering thousands of molecular effect predictions per variant across hundreds of human and mouse cell types and tissues.

The release includes several new components:

• AlphaGenome Variant Impact (AVI) score — a single number per variant that combines predictions from AlphaGenome and AlphaMissense (DeepMind's protein-variant-impact model), covering both coding regions (roughly 2% of the genome) and non-coding regions (the remaining 98%, where most trait-associated variants reside).
• AVI feature attributions — a breakdown of which biological processes, such as chromatin accessibility or splicing, drive each variant's AVI score.
• DNA sequence motifs — a catalogue of more than 2,500 recurrent DNA sequences and their genome locations.

DeepMind claims the AVI score delivers "best-in-class performance" across multiple variant pathogenicity and rare disease benchmarks, though it did not publish specific benchmark figures or comparison scores in the announcement.

Reported research applications

According to DeepMind, external collaborators have already used the Atlas in active research. Working with the GREGoR Consortium, researchers at the Broad Institute — led by Laura Covill and Anne O'Donnell-Luria — say they used the AVI score to identify a previously overlooked variant in the DNM1 gene linked to epileptic encephalopathy. DeepMind states the underlying AlphaGenome predictions showed the variant created an incorrect splice site, leading to an abnormal protein extension, and that experimental screens validated this prediction along with nearby variants showing similar effects.

DeepMind also says the Atlas is being applied to identify rare non-coding variants associated with protein levels and complex traits in population genetics research, though it did not disclose specific published results for this application in the announcement.

What this means

This is an infrastructure release, not a new model: the underlying AlphaGenome and AlphaMissense models already existed, and this launch is about precomputing and packaging their outputs at genome-wide scale for accessibility, similar to how the 2022 AlphaFold Database expansion turned a research model into a widely used reference tool covering more than 200 million structure predictions.

The practical value hinges on adoption and validation beyond DeepMind's own collaborators. The DNM1 case is a genuine example of a novel finding traced through the platform, but it is one data point from a partner relationship, not independent third-party confirmation of the AVI score's general accuracy. Free access and a no-code portal lower the barrier for rare disease researchers who lack computational resources to run AlphaGenome themselves, which is likely the more significant near-term impact than any single benchmark claim.