There are roughly 9 billion possible single-letter mutations in the human genome. Testing each one in a laboratory is not just slow — it’s practically impossible. That basic constraint has quietly bottlenecked genetic research for decades. Now, Google DeepMind is attempting to bypass it entirely.
The company has released AlphaGenome Atlas, a platform containing precomputed predictions for the molecular effects of all 9 billion single-nucleotide variants in the human genome. It is, by any reasonable measure, the most comprehensive catalogue of its kind ever assembled. And it’s available free of charge through a web portal aimed at academic researchers, including those with no coding experience.
What AlphaGenome Atlas actually contains
The dataset runs to 1 petabyte — more than 30 times larger than the AlphaFold Database, which itself reshaped how researchers study protein structure when it expanded in 2022. The comparison is intentional. DeepMind built AlphaFold Database to give researchers a single accessible portal into a vast prediction space. AlphaGenome Atlas follows the same model, applied to genetic variation rather than protein folding.
At its core, the platform is built on AlphaGenome, an AI model DeepMind previously released for predicting how genetic variants affect biological processes. Atlas takes that model’s outputs and applies them at genome-wide scale, making predictions browsable and searchable rather than something researchers must generate one variant at a time.
The platform contains four interconnected resources:
- Molecular effect predictions — thousands of predictions per variant, spanning gene regulation across hundreds of human and mouse cell types and tissues
- AlphaGenome Variant Impact (AVI) score — a single number summarising each variant’s predicted biological impact, covering both protein-coding regions and the non-coding 98% of the genome where most trait-associated variants are found
- AVI feature attributions — breakdowns showing which molecular processes, such as RNA splicing or gene expression, are predicted to be most disrupted by a given variant
- DNA sequence motifs — a catalogue of over 2,500 recurring DNA sequences and their locations across the genome
The AVI score is a notable addition. It draws on both AlphaGenome and AlphaMissense, DeepMind’s earlier model for protein-altering variants, condensing their outputs into a single ranking metric. That matters because researchers sifting through thousands of candidate variants need a fast way to prioritise. A single interpretable score, backed by mechanistic detail, is genuinely useful in that context.
Early results from rare disease research
The platform was tested with external collaborators before launch. One result stands out. Researchers from the Broad Institute, working with the GREGoR Consortium on unsolved rare disease cases, used the AVI score to prioritise candidate variants that had been missed in earlier analyses. The approach flagged a variant in a gene called DNM1, which is associated with epileptic encephalopathy.
The AlphaGenome predictions didn’t just identify the variant — they showed how it worked. The mutation created a faulty splice site, causing an abnormal extension in the resulting protein. Experimental validation confirmed the prediction and also identified nearby variants with similar effects. That combination of prioritisation and mechanistic explanation is exactly what rare disease research needs, where the signal is often buried under thousands of benign variants.
Non-coding variants and population genetics
Rare disease is one application. But AlphaGenome Atlas is also being used to study common traits across larger populations, where non-coding variants are notoriously hard to identify against a background of harmless genetic variation. By using AVI scores to filter candidates before statistical analysis, researchers can cut through that noise more efficiently. Work by collaborators testing the approach on protein levels and complex traits suggests the tool improves detection of meaningful non-coding variants in population-scale data.
Why this matters beyond the headline number
The 9 billion variants figure is striking, but the more important question is whether the predictions hold up in practice. DeepMind reports best-in-class performance across multiple variant pathogenicity and rare disease benchmarks. Independent validation will matter here, and the research community will stress-test these predictions thoroughly now that access is open.
Still, the structural contribution is significant regardless. For years, non-coding genomics has suffered from a shortage of usable, interpretable prediction tools. Most variant effect predictors focus on protein-coding regions and struggle with the vast non-coding majority. A platform that covers both, links predictions to specific molecular mechanisms, and makes that information accessible through a free portal fills a genuine gap. The AlphaFold Database precedent suggests this kind of precomputed resource, once available, tends to become infrastructure. AlphaGenome Atlas looks built for the same role.
The platform is available now through the web portal, the AlphaGenome API, and as a skill in Google Antigravity.