AlphaGenome Atlas Maps 9 Billion DNA Variants

Google DeepMind has released AlphaGenome Atlas, a precomputed catalogue of the predicted molecular consequences of all 9 billion possible single-letter changes to human DNA. Announced on September 8, the dataset runs to roughly one petabyte — more than thirty times the size of the AlphaFold Database — and carries thousands of molecular predictions for each variant across hundreds of human and mouse cell types and tissues.

The Atlas is not a new model. It is the AlphaGenome model, which DeepMind released in June 2025, run exhaustively in advance so that researchers no longer have to submit variants one at a time and wait. That shift from on-demand inference to a queryable resource is the whole point: it turns a tool you have to know how to drive into something closer to a reference work you can look things up in.

One number for coding and non-coding DNA

Alongside the data, DeepMind introduced the AlphaGenome Variant Impact score, or AVI, which folds predictions from AlphaGenome and the earlier AlphaMissense protein model into a single ranking. The useful part is its coverage. AlphaMissense addressed variants in protein-coding DNA, which is about 2% of the genome; most variants that turn up in association studies sit in the other 98%, where their effects are regulatory and much harder to interpret. AVI puts both on one scale, and ships with feature attributions indicating which process — RNA splicing, gene expression, chromatin accessibility — a given change is predicted to disrupt.

Access is split along familiar lines. There is a free web portal for non-commercial and academic research, the AlphaGenome API for programmatic use, and a skill for Google Antigravity so the Atlas can be pulled into agent workflows directly. Commercial licensing is routed through Google Cloud.

What early users found

DeepMind published results from three groups given advance access. At the Broad Institute, Laura Covill and Anne O'Donnell-Luria used the Atlas on unsolved rare disease cases and flagged a variant in the DNM1 gene tied to epileptic encephalopathy; the predictions indicated it created an incorrect splice site that abnormally extended the resulting protein, which the team then confirmed experimentally. At the University of Exeter, Gareth Hawkes applied it to 54,000 UK Biobank participants and reported 22% more non-coding genetic associations than conventional analysis surfaced, including regulatory variants affecting PLA2G7, linked to aging, and EGLN1, involved in oxygen sensing. A group at the Stowers Institute led by Julia Zeitlinger with Melanie Weilert used it to sort transcription factors by whether they primarily open up DNA or drive gene activation.

Predictions, not measurements

Every number in the Atlas is a model output. Nine billion predictions is nine billion hypotheses, and the value of a hypothesis generator at this scale depends entirely on how well its false positive rate is understood — a question the resource itself cannot answer. The Broad's DNM1 finding is instructive in the right way: the model pointed at a mechanism, and a laboratory confirmed it. That is the intended loop, and it is a different thing from treating a high AVI score as a finding.

DeepMind is explicit about the boundary, stating that the Atlas has not been validated or approved for any clinical use and is not a substitute for professional medical advice, diagnosis or treatment. The caution is warranted, and it will be tested. A free, comprehensive, searchable map of variant effects is precisely the kind of artifact that leaks out of research settings, and the gap between a predicted splice disruption and a diagnosis is one that a well-designed portal makes very easy to forget.

Still, the scale is the argument. Rare disease diagnosis fails most often not because a patient's causal variant is missing from their sequence data but because nobody can say which of the thousands of variants they carry matters. Handing that triage problem a ranked, mechanism-annotated prior across the entire genome is a real change in what a small lab can attempt without a bioinformatics team behind it.