On September 8, 2026, Google DeepMind released AlphaGenome Atlas, a precomputed map of what its AlphaGenome model predicts would happen for every possible single-letter change in the human genome - about 9 billion single-nucleotide variants. For each variant the Atlas holds thousands of molecular effect predictions, covering gene regulation across hundreds of human and mouse cell types and tissues, RNA splicing, gene expression and chromatin accessibility, with protein-level impacts drawn from AlphaMissense. The team also catalogued more than 2,500 DNA sequence motifs.
The scale is the headline. DeepMind describes the dataset as about one petabyte, roughly 30 times larger than the AlphaFold Protein Structure Database. It is available through a web portal at alphagenome.google/atlas, through the AlphaGenome API, and as a skill for Google's Antigravity agent tool. Non-commercial academic use is free; commercial access through Google Cloud is described as forthcoming. A research paper accompanies the announcement.
DeepMind points to early uses by outside groups: researchers at the Broad Institute used it in identifying rare disease variants in the DNM1 gene linked to epileptic encephalopathy, and a University of Exeter study reported 22 percent more non-coding genetic associations in protein-level analyses. The significance is practical. Most disease-associated variation sits in non-coding DNA, where the effect of a mutation is hard to read, and a lookup table that already holds a model's best guess for every position removes the need for each lab to run the model itself - much as the AlphaFold database did for protein structures.
The limits are those of the underlying model. These are predictions, not measurements, and DeepMind says plainly that the Atlas is not a substitute for professional medical advice, diagnosis or treatment. It covers single-letter changes only, not insertions, deletions or combinations of variants, and DeepMind frames the release as "a baseline rather than an endpoint." Whether its predictions hold up across clinical use will be settled by the kind of external validation studies that are only beginning.