Summary: Google DeepMind’s AlphaGenome Atlas, published September 8, 2026, precomputes regulatory-effect predictions for 9 billion possible single-letter DNA variants in a roughly 1-petabyte queryable resource. The AVI score spans coding and non-coding regions with feature attributions, and early collaborators used it to rank a DNM1 variant and find 22% more non-coding associations in UK Biobank data.
Where most genomics tools have you hand over a variant and then wait for a prediction, AlphaGenome Atlas works the other way around. Molecular effect predictions for all 9 billion possible single-letter DNA changes in the human genome are computed ahead of time, and researchers reach them through a portal, an API, or a Google Antigravity skill — querying, ranking, and exploring rather than inferring on demand.
That move, from on-demand inference to a precomputed atlas, is the product story here. It repeats the pattern from the moment AlphaFold stopped predicting one protein structure at a time and published a database of 200 million predictions. The database became the interface, and large-scale analysis opened up to researchers with no coding experience.
What’s actually in the atlas
Google DeepMind published AlphaGenome Atlas on September 8, 2026. The dataset runs to roughly 1 petabyte, more than 30 times the size of the AlphaFold Database. Inside, every variant carries thousands of molecular effect predictions tied to gene regulation, spanning hundreds of human and mouse cell types and tissues.
The headline addition is the AlphaGenome Variant Impact (AVI) score, which folds AlphaGenome’s regulatory predictions and AlphaMissense’s protein impact predictions into a single number. Its reach is the point: it applies to the 2% of the genome that codes for proteins and the 98% that does not. Trait-associated variants cluster in the non-coding portion, where one interpretable score has been hard to come by.
Each AVI score also ships with feature attributions that break the number down into interpretable categories such as chromatin accessibility, splicing, and conservation. A researcher therefore sees more than the fact that a variant is high-impact — the biological mechanism driving that impact is visible as well.
Why precomputation changes the workflow
For product builders, the design choice worth studying is what Google DeepMind did not do: they didn’t just ship a faster model. Everything was precomputed and handed over as a queryable resource.
That has real consequences for how teams build on top of it. Instead of integrating a model and managing inference costs, a genomics tool can treat the atlas as a data layer. The AVI score turns into a filter to apply before deeper analysis begins. The linked DNA sequence motifs — over 2,500 recurrent sequences and their locations — let you connect variants to the functional elements they disrupt.
The same logic sits behind choosing a structured data extraction tool for AI agents: do the expensive computation once, do it well, and expose it through a clean interface rather than repeating it on every request.
Early validation from rare disease and population genetics
Collaborators at Google DeepMind have already put the atlas to work in two concrete settings.
In unsolved rare disease research, the GREGoR Consortium applied the AVI score to rank candidate variants. Researchers at the Broad Institute turned up a variant affecting the DNM1 gene, linked to epileptic encephalopathy, that earlier work had overlooked. The AlphaGenome predictions exposed the mechanism: the variant created an incorrect splice site, which produced an abnormal protein extension. Experimental screens confirmed the prediction.
In population genetics, Gareth Hawkes at the University of Exeter ran the atlas over whole-genome data from more than 54,000 UK Biobank participants. Grouping rare variants by predicted molecular effects uncovered 22% more non-coding genetic associations than the statistical noise would have allowed anyone to detect. He also pinpointed 19 genetic regions tied to body mass index by concentrating on the 1% of non-coding variants the atlas predicts to be most impactful.
What this means for your roadmap
If you build tools for genetic analysis, the atlas moves the default architecture. A precomputed, rankable dataset becomes the starting point instead of a model call. That lowers the barrier to non-coding variant interpretation, historically the hard part of the problem.
Limits exist. The atlas is available for non-commercial use through the website portal today, with commercial access coming soon on Google Cloud. The AlphaGenome base model already sits on GitHub and is reachable through the AlphaGenome API for academic use, while Cloud Model Garden covers commercial use.
The bigger question is whether a precomputed atlas ages well. Google DeepMind frames it as a baseline rather than an endpoint — as the underlying models improve, the maps get regenerated. For builders, that means designing integrations that treat the atlas as a versioned data source rather than a static reference.
That’s the lesson from small language models in the enterprise too: the value isn’t only the model, it’s the packaging and the interface that make the model’s output usable at scale.