Google DeepMind Opens AlphaGenome Atlas to 9 Billion DNA Variants

Google DeepMind has released AlphaGenome Atlas, a public catalogue of predicted molecular effects for roughly 9 billion possible single-nucleotide variants in the human genome. The portal is free for academic research and pairs the precomputed predictions with an AlphaGenome Variant Impact (AVI) score, giving researchers a way to rank variants without running the base model one mutation at a time.

What AlphaGenome Atlas adds

AlphaGenome Atlas packages genome-wide predictions across gene regulation, including gene expression, RNA splicing, chromatin accessibility, histone modifications, transcription-factor binding and chromatin contacts. It covers coding and non-coding DNA. That distinction matters because most of the genome does not directly encode proteins, yet regulatory variants can influence when and where genes are active.

The scale is the headline constraint. DeepMind describes the dataset as about 1 petabyte, more than 30 times the size of the AlphaFold Database. A web portal makes the data searchable and visual, while the AlphaGenome API and a Google Antigravity skill expose programmatic or agent-based access. DeepMind says commercial cloud access is planned; the announcement currently describes the Atlas website as available for non-commercial use.

How the AVI score is meant to be used

The AVI score combines AlphaGenome predictions with AlphaMissense, DeepMind’s model for protein-altering variants. It reduces multiple model outputs to a single ranking signal, then links that score back to feature attributions such as predicted changes in splicing or gene expression.

This makes the Atlas a prioritization resource rather than a diagnostic service. A high score can tell a researcher which variants deserve closer examination and which molecular process may be affected. It cannot, by itself, establish that a variant causes disease in a patient.

Early research examples and the evidence boundary

DeepMind reports that collaborators used the score in rare-disease research and in analyses of common traits. In one example involving the GREGoR Consortium, predictions helped prioritize a previously overlooked DNM1 variant associated with epileptic encephalopathy; experimental screens then supported the predicted splice-site mechanism.

Another analysis used whole-genome data from more than 54,000 UK Biobank participants. DeepMind says grouping rare variants by predicted molecular effect exposed 22% more non-coding genetic associations and identified 19 regions linked to body-mass-index analysis after focusing on the 1% of variants the Atlas ranked as most impactful.

Those are reported collaborator results, not a guarantee that every high-scoring variant will validate experimentally. The independent Nature paper on the underlying AlphaGenome model reports strong benchmark performance—matching or exceeding external models in 25 of 26 variant-effect evaluations—but the Atlas adds large-scale precomputation and should be assessed on its own use cases.

Access, limits and the next step

Researchers can start with the Atlas website and its academic-use terms. Teams planning commercial work should wait for the announced Google Cloud route and verify licensing, quotas and API availability at the point of use. The underlying AlphaGenome model is also available for academic use through GitHub and the AlphaGenome API.

The most important caveat is clinical status: DeepMind explicitly says AlphaGenome has not been validated for, and is not approved for, clinical use. The immediate value is research triage—turning an otherwise impractical search across billions of possible changes into a ranked, interpretable set of hypotheses. Experimental validation remains the next milestone for any biological or medical conclusion.

Sources

• Google DeepMind: AlphaGenome Atlas announcement
• Nature: Advancing regulatory variant effect prediction with AlphaGenome