Google DeepMind's AlphaGenome Atlas Maps 9 Billion Human DNA Mutations Instantly

Google DeepMind has released AlphaGenome Atlas, a searchable database that predicts the molecular consequences of every possible single-letter mutation in the human genome. Rather than making researchers run an ML model each time they want to score a variant, DeepMind precomputed the entire space and packaged it as an interactive resource with a unified impact score.

The platform covers predictions for 9 billion single-nucleotide variants, meaning every possible single-letter change across the roughly 3 billion base-pair reference sequence, with each position multiplied by the three alternative letters that could appear there.

From model to atlas

AlphaGenome, the underlying model released earlier this year, predicts how a given DNA change affects gene regulation, splicing, chromatin accessibility, and other molecular readouts. Useful in isolation, but querying it variant by variant was a bottleneck. Precomputing AlphaGenome's predictions at scale removes that constraint and turns a research model into a browsable resource.

The resulting dataset weighs in at 1 petabyte, more than 30 times the size of the AlphaFold Database, which catalogued 3D structures for over 200 million proteins.

One score across coding and non-coding DNA

The central addition is the AlphaGenome Variant Impact (AVI) score, a single number that combines predictions from AlphaGenome and AlphaMissense, DeepMind's model for protein-altering variants. Researchers get one ranking metric plus a mechanistic breakdown of what drives it.

The score covers both coding regions, the 2% of the genome that encodes proteins, and non-coding regions, the remaining 98% that orchestrates gene activity and houses most trait-associated variants. Non-coding variants are where most genome-wide association signals fall, and where interpretation has historically been hardest.

The Atlas exposes four layers:

• Molecular effect predictions: thousands of predictions per variant across gene regulation readouts, spanning hundreds of human and mouse cell types and tissues.
• AVI score: a single impact number per variant.
• AVI feature attributions: which biological processes, such as RNA splicing or gene expression, drive that score.
• DNA sequence motifs: over 2,500 recurrent DNA sequences and their genomic locations.

Benchmark performance and early wins

DeepMind's testing shows the AVI score reaches best-in-class performance across variant pathogenicity and rare disease benchmarks. Two collaborations illustrate what that means in practice.

In rare-disease work with the GREGoR Consortium, researchers used AVI scores to surface a previously overlooked variant in a gene called DNM1, which is strongly linked to epileptic encephalopathy. The predictions showed the variant creating an incorrect splice site that extended the resulting protein abnormally, and experimental screens confirmed it.

On the population-genetics side, a Medical Research Council fellow at Exeter applied Atlas to UK Biobank whole-genome data from over 54,000 participants. Grouping rare variants by predicted molecular effects uncovered 22% more non-coding genetic associations that would otherwise have dissolved into statistical noise. Filtering to the top 1% of non-coding variants by predicted impact, he identified 19 genomic regions potentially linked to body mass index.

How to access it

Access is free for non-commercial research through several routes:

• The Atlas web portal, aimed at researchers without a coding background.
• The AlphaGenome API on GitHub for programmatic queries.
• A skill inside Google Antigravity, DeepMind's agentic development platform, for composing Atlas into automated scientific workflows.
• Commercial availability on Google Cloud is coming; the base AlphaGenome model is already on Cloud Model Garden.

One firm caveat: AlphaGenome has not been validated for or approved for clinical use. It is a research tool for hypothesis generation and prioritization, not diagnostics.

Why precomputing changes the calculus

The shift from on-demand inference to a fully materialized lookup table over the entire mutation space has practical downstream effects. Any variant lookup becomes instant, changing how researchers can filter cohorts of tens of thousands of genomes. Rare-variant burden tests can be reweighted by predicted molecular impact rather than treating all variants equally, which is what unlocked the extra associations in the UK Biobank analysis. Comparing variants across the genome becomes trivial because every score was computed with the same model against the same reference.

AlphaFold turned protein structure prediction from a per-query modeling problem into a browsable resource that reshaped how biologists approached structural questions. Applying the same treatment to variant effects, especially across the non-coding 98%, gives labs a shared prior over which mutations are worth pursuing experimentally.

What to update in your mental model

For anyone working in genomics ML, variant effect prediction across the human genome is no longer bottlenecked by inference cost. The bottleneck has shifted to integrating a dense score field into downstream analyses. Non-coding variant prioritization now performs competitively with coding-variant prioritization on standard benchmarks, a gap that did not close until recently. Model interpretability, via AVI feature attributions that decompose scores into splicing, chromatin, expression, and conservation contributions, ships as a first-class output.

For teams building on top of Atlas, the interesting design space is downstream: combining AVI scores with cohort data, feeding them into agentic pipelines that plan follow-up wet-lab experiments, and fusing them with protein-level predictors for end-to-end variant-to-phenotype reasoning.