Google DeepMind launched AlphaGenome Atlas, a roughly one-petabyte database holding a machine-generated guess for what happens at each of the nearly 9 billion possible single-letter edits to human DNA.
Only about 2 percent of the genome codes directly for proteins. The rest acts as regulatory instructions, controlling when and how strongly other genes switch on. Working out what a single letter swap does to that regulatory layer has historically meant slow, one-variant-at-a-time lab work. The Atlas instead ships a precomputed guess for every possible swap, generated by the existing AlphaGenome model.
These are predictions, not measured lab results. Google DeepMind frames each entry as a lead worth testing, citing two research teams (one at the Broad Institute, another using UK Biobank data) that used the Atlas to prioritize candidates rather than confirm answers. The launch post does not publish independent accuracy figures. Access is free through a web portal the company says needs no coding.
Precomputing every possible answer flips the usual order of biology research: the scarce resource stops being a hypothesis to generate and becomes the judgment call of which candidate is worth a wet-lab experiment.
Google DeepMind announced AlphaGenome Atlas on the Google blog on September 8, 2026.
