DeepMind releases AlphaGenome Atlas

Google DeepMind released the AlphaGenome Atlas, a precomputed database of molecular-effect predictions covering roughly nine billion possible single-nucleotide variants — every point mutation that could occur across the human genome. Each variant carries thousands of predicted effects on gene regulation across hundreds of human and mouse cell types and tissues, drawn from the company’s earlier AlphaGenome model. The resulting dataset runs to about one petabyte, which Google DeepMind said is more than 30 times the size of the AlphaFold Protein Structure Database.

The release also introduced an AlphaGenome Variant Impact (AVI) score, a single summary metric intended to help researchers triage which of the billions of catalogued variants are worth investigating further, alongside feature attributions explaining what drove a given prediction. The atlas is browsable through a web portal, queryable via an API, and available as a skill inside Google’s Antigravity coding platform; access for academic and non-commercial research is free, with commercial use planned via Google Cloud’s Model Garden.

The underlying predictions are computational, not experimentally validated variant-by-variant — the atlas is a hypothesis-generation tool rather than a clinical resource, and its value depends on independent labs testing the predictions it flags as significant. Its scale is nonetheless a marked change from prior practice: variant-effect prediction has typically been run on demand for a handful of genes or a research group’s own dataset, and the atlas instead exposes the results for the entire genome up front, positioning routine variant screening as a lookup rather than a modelling task for individual research groups.