Google DeepMind's AlphaGenome Atlas maps roughly 9 billion DNA variants

Google DeepMind's AlphaGenome Atlas maps roughly 9 billion DNA variants

Google DeepMind has released AlphaGenome Atlas, a platform containing predictions for the molecular effects of roughly 9 billion single-nucleotide variants, essentially every possible single-letter change in the human genome. It has been available since 8 September 2026 through a free website portal for academic and non-commercial use, through the AlphaGenome API, and as a skill inside Google Antigravity. DeepMind calls it the most comprehensive catalogue yet of how genetic mutations affect molecular biology. The underlying dataset runs to 1 petabyte, more than 30 times the size of DeepMind's AlphaFold Database of protein structures.

The Atlas builds on AlphaGenome, DeepMind's existing AI model for predicting how a genetic variant affects biological processes such as gene regulation. DeepMind says AlphaGenome has already found wide research use for analyzing specific variants; the Atlas instead precomputes its predictions across the whole genome up front, turning them into a resource researchers can browse without running the model themselves. Alongside it, DeepMind introduces the AlphaGenome Variant Impact score, called AVI, a single number that combines AlphaGenome's predictions with those of AlphaMissense, DeepMind's separate model for scoring protein-altering variants. The AVI score works across both the 2% of the genome that codes for proteins and the remaining 98% that does not, letting researchers rank and interpret variants from either region at once. Each score comes with feature attributions pointing to which biological process, such as RNA splicing or gene expression, is predicted to be most disrupted. The Atlas also catalogues more than 2,500 recurring short DNA sequences, along with their locations, that DeepMind describes as the genome's regulatory vocabulary.

DeepMind highlights three research collaborations that used the Atlas ahead of its 8 September 2026 public release. Working with the GREGoR Consortium, researchers at the Broad Institute, including Laura Covill and Anne O'Donnell-Luria, used the AVI score to prioritize candidate variants in unsolved rare disease cases that earlier analysis had missed. That search turned up a variant in a gene called DNM1, which is strongly linked to epileptic encephalopathy. AlphaGenome's predictions showed how the variant worked: it created an incorrect splice site, a defect in the cell's genetic instructions, leading to an abnormal extension of the resulting protein. Experimental screens then confirmed that prediction and turned up nearby variants with similar effects.

At the University of Exeter, Medical Research Council fellow Gareth Hawkes applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. Grouping rare, non-coding variants by their predicted molecular effect surfaced signals that would otherwise have been lost in statistical noise, turning up 22% more non-coding genetic associations than without that grouping. The approach pinpointed regulatory variants controlling the levels of proteins circulating in the body, including PLA2G7, linked to aging, and EGLN1, a cellular oxygen sensor. Hawkes also used the Atlas to examine hundreds of millions of non-coding UK Biobank variants for links to body mass index; by focusing on just the 1% of those variants the Atlas predicts to be most impactful, Hawkes identified 19 genetic regions that could help direct further research into the trait. At the Stowers Institute for Medical Research, Julia Zeitlinger and Melanie Weilert used the Atlas's motif catalogue for a different purpose: working out which transcription factors, the proteins that switch genes on or off, merely make DNA more accessible versus which ones actively toggle gene activity.

DeepMind frames the Atlas as a successor to its AlphaFold Database, for which it grew the 3D structure information available in 2022 from around 190,000 experimentally determined protein structures to more than 200 million predicted ones, covering nearly all catalogued proteins known to science, and which the company says became a resource that drove discoveries across the life sciences. On its own testing, DeepMind says the AVI score delivers best-in-class performance across many variant pathogenicity and rare disease benchmarks, though the announcement does not publish the actual scores behind that claim, nor does it disclose the formula or weighting behind how AVI combines AlphaGenome and AlphaMissense. It is also unclear from the announcement whether the DNM1 finding has been separately published or peer reviewed. AlphaGenome Atlas is free for academic and non-commercial use starting 8 September 2026; commercial access is planned for Google Cloud soon, with no date given. The underlying AlphaGenome base model was already available separately for academic use on GitHub and via its own API, and for commercial use on Google Cloud through Model Garden. DeepMind also says Atlas resources can feed into its broader agentic systems, such as Google Antigravity, to support end-to-end scientific workflows, and describes the Atlas itself as a baseline rather than an endpoint, expecting its genome maps to grow more comprehensive as models like AlphaGenome keep improving.

Key facts

  • Google DeepMind released AlphaGenome Atlas on 8 September 2026: a free website portal, an API and a Google Antigravity skill containing predictions for the effects of roughly 9 billion possible single-letter DNA variants, in a 1-petabyte dataset more than 30 times the size of DeepMind's AlphaFold Database.
  • The release introduces the AlphaGenome Variant Impact (AVI) score, a single number combining AlphaGenome and AlphaMissense predictions to rank variants across both the 2% protein-coding and 98% non-coding genome, alongside a catalogue of more than 2,500 recurring DNA sequence motifs.
  • At the Broad Institute, Laura Covill and Anne O'Donnell-Luria used the AVI score, working with the GREGoR Consortium, to find a previously overlooked variant in the gene DNM1 that is strongly linked to epileptic encephalopathy; AlphaGenome showed it creates a faulty splice site, and experimental screens confirmed the prediction.
  • University of Exeter fellow Gareth Hawkes applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants, uncovering 22% more non-coding genetic associations and identifying 19 genetic regions that could help direct further body-mass-index research by focusing on the top 1% of predicted most-impactful non-coding variants.
  • DeepMind says its own testing found best-in-class performance for the AVI score across variant pathogenicity and rare disease benchmarks, but the announcement does not publish those scores; commercial access via Google Cloud is planned soon with no date given, while academic and non-commercial use is free starting 8 September 2026.

Why it matters

Testing how each of the genome's roughly 9 billion possible single-letter variants affects biology in the lab, one at a time, is not practical, and that gap has limited how fast genetic findings turn into an understanding of disease. AlphaGenome Atlas precomputes those effects across the whole genome up front, including the 98% that does not code for protein but still houses most trait-associated variants and has historically been the harder region to interpret. DeepMind explicitly compares the release to its own AlphaFold Database, for which the 3D structure information available grew from about 190,000 experimental protein structures to more than 200 million predicted ones when it was expanded in 2022, and which went on to reshape structural biology research; the new Atlas is more than 30 times larger by data volume, and DeepMind is betting it can do for genetic variants what AlphaFold did for protein shapes.

Who it affects

The immediate audience is academic and non-commercial researchers in rare disease genetics, population genetics and molecular biology; DeepMind names specific users at the Broad Institute, working through the GREGoR Consortium, the University of Exeter and the Stowers Institute for Medical Research. Commercial users, including biotech and pharmaceutical companies that might want the Atlas on Google Cloud, are not yet served: that access is only planned, with no date given. Patients are affected only indirectly, through whatever new disease-gene links and drug targets this research produces downstream; the Atlas itself is a research resource, not a clinical tool.

How to use it

AlphaGenome Atlas's website portal has been free to use since 8 September 2026 for academic and non-commercial purposes; the Atlas is also available through the AlphaGenome API and as a skill inside Google Antigravity. Commercial use is planned for Google Cloud soon, but DeepMind gives no date or pricing. Separately, the underlying AlphaGenome base model was already available for academic use on GitHub and via its own API, and for commercial use on Google Cloud through Model Garden, before this Atlas release.

How solid is it

DeepMind says its own testing found the AVI score gives best-in-class performance across many variant pathogenicity and rare disease benchmarks, but the announcement does not publish the scores behind that claim, and it does not explain the formula or weighting behind how AVI combines AlphaGenome and AlphaMissense predictions. The strongest evidence offered is the Broad Institute's DNM1 case, where AlphaGenome's predicted mechanism, an incorrect splice site, was confirmed by wet-lab experimental screens; the announcement does not say whether that finding has been separately published or peer reviewed. The other two case studies, from the University of Exeter and the Stowers Institute, are statistical and exploratory rather than experimentally confirmed.

Risks and caveats

Most of what the Atlas produces are computational predictions, not experimentally settled facts: only the DNM1 case is described as validated in the lab. The 19 genetic regions linked to body mass index and the regulatory variants tied to PLA2G7 and EGLN1 are presented as leads for further research, not final conclusions. DeepMind's headline performance claim for the AVI score rests on its own unpublished benchmark scores, and the underlying paper is linked but not named or dated in the announcement. Commercial terms and timing on Google Cloud remain undecided, so any organization planning to build on the Atlas commercially has no date or price to plan around yet.