DeepMind precomputes every possible DNA variant's effect: the AlphaGenome Atlas

What shipped

Google DeepMind released the AlphaGenome Atlas on September 8, 2026: a precomputed catalogue of predicted molecular effects for roughly 9 billion single-nucleotide variants — every possible single-letter substitution in the human genome. It's a 1-petabyte dataset, over 30 times the size of the AlphaFold Database, and it's free to query through a website portal, an API, and a skill in Google Antigravity.

The underlying model, AlphaGenome, already existed. It predicts how a given genetic variant affects things like gene expression, splicing, and chromatin accessibility. What's new is the scale of precomputation: instead of asking researchers to run the model one variant at a time, DeepMind ran it against essentially the entire space of possible single-letter changes and published the results as a lookup table.

Why precomputation is the actual story

The interesting engineering decision here isn't the model — it's the choice to trade compute-now for query-latency-later, at genome scale. AlphaGenome as a standalone model is useful when you already have a candidate variant and want its predicted effect. But most real variant-interpretation work starts from the opposite direction: a researcher has a disease phenotype and a list of thousands of candidate variants, and needs to rank them fast. Running a deep model per candidate, per analysis, doesn't scale to that workflow.

Precomputing everything up front turns variant interpretation into a database lookup. That's the same move DeepMind made with the AlphaFold Database in 2022 — going from roughly 190K experimentally solved protein structures to over 200 million predicted ones, searchable by anyone without running inference themselves. The Atlas applies that pattern to sequence variants instead of structures.

Alongside the raw predictions, DeepMind is releasing the AlphaGenome Variant Impact (AVI) score — a single number per variant that folds together AlphaGenome's regulatory predictions with AlphaMissense (DeepMind's earlier model for protein-coding impact). That fusion matters mechanically: the human genome is about 2% protein-coding and 98% non-coding, and the non-coding fraction is where most trait-associated variants actually sit but is also where causal inference is hardest, since there's no protein sequence to check for damage. A single score that works across both regions, plus feature attributions that show why — splicing disruption versus chromatin accessibility versus conservation — gives researchers a ranked, explainable starting point instead of raw model output they have to interpret from scratch.

Where it's already been used

DeepMind cites two early applications worth noting because they're concrete, not aspirational. In collaboration with the GREGoR Consortium, researchers at the Broad Institute (Laura Covill and Anne O'Donnell-Luria) used the AVI score to re-prioritize variants in unsolved rare-disease cases and found one in the gene DNM1, linked to epileptic encephalopathy. The AlphaGenome predictions underneath the score showed the mechanism directly: the variant created an incorrect splice site that led to an abnormally extended protein — a hypothesis specific enough to validate experimentally, which the team then did.

Separately, Gareth Hawkes at the University of Exeter used the Atlas to look for rare, non-coding variants associated with population-level traits — a search that's normally swamped by statistical noise, since the overwhelming majority of non-coding changes have no effect at all. Having precomputed, ranked scores for every candidate in a region turns that into a filtering problem instead of an open-ended one.

My take

What I'd flag for anyone evaluating this: the value isn't the AVI score's accuracy in isolation — DeepMind reports it as best-in-class on pathogenicity and rare-disease benchmarks, but benchmark performance for variant-effect predictors is notoriously easy to overstate depending on which validated variant sets you test against. The value is that the score exists before you ask the question, for every variant, computed once and shared. That's what makes the DNM1 discovery believable: the researchers weren't waiting on inference time to test a hypothesis, they were scanning precomputed rankings and finding the mechanism already annotated.

For teams building on top of models like this, the AlphaFold-to-AlphaGenome-Atlas pattern is worth internalizing on its own: if your model's value comes from being queried against a large, mostly-static space — structures, variants, molecules — precomputing and serving beats an on-demand API almost every time. It shifts cost from every user's inference call to one upfront batch job, and it turns your product from "a model" into "a resource," which is a much stickier thing to build a research field around.