DeepMind maps all 9 billion possible typos in your DNA

Google DeepMind has worked out in advance what happens when you change any single letter of the human genome. That's all 9 billion possible substitutions: three alternatives at each of about 3 billion positions. The result, AlphaGenome Atlas, is a 1-petabyte dataset, about 30x the size of the AlphaFold Database. Academic and non-commercial researchers can use it free through a no-code web portal and the AlphaGenome API. Commercial access is coming through Google Cloud.

Each variant gets thousands of predictions about gene regulation, RNA splicing and chromatin accessibility, across hundreds of human and mouse cell types. These roll up into one AlphaGenome Variant Impact (AVI) score, so labs can decide which mutations to study first. The real prize is the non-coding 98% of the genome: most trait-linked variants sit there, and older tools have been weakest there. When researchers grouped UK Biobank variants by predicted effect, they found 22% more non-coding genetic associations.

The caveats are real. The model looks at a 1-million-base-pair window and can't reliably predict very long-range enhancer effects. Most diseases also involve many variants, not one. And a single score can oversimplify a very complex system.

Why it matters: it's the AlphaFold playbook applied to DNA. Compute everything once, and every rare-disease lab gets a lookup table instead of a GPU bill.