Building on our coverage 11 days ago, DeepMind has published an AI-built atlas that predicts the effect of every single-letter change in human DNA.
On 2026-09-08 DeepMind announced AlphaGenome Atlas: a precomputed, searchable dataset of predictions for all 9 billion possible single-base mutations across the human reference genome. The Atlas is roughly one petabyte of model outputs and is available for non-commercial use through a web browser, so researchers can query AI-based consequences without running large genomics models themselves.
That matters because researchers often spend weeks rerunning huge models or running lab experiments to interpret a single genetic variant. With the Atlas, they can look up AI predictions instantly, which should speed up variant interpretation and help teams prioritise which changes to test experimentally.
DeepMind built the resource by running its AlphaGenome model across the reference genome and comparing each base to the three alternate letters it could be, then storing the predicted effects on things like gene activity, RNA splicing and chromatin interactions. Think of it as a map you can look up instead of redrawing every time.
Practically, this lowers the technical barrier for many labs and academic groups. You no longer need pet-scale compute just to see what an AI predicts for a specific mutation. But it's not a substitute for experiments: the Atlas offers model-based hypotheses, not clinical proof, and the full dataset's size means most users will query it in the browser rather than download it.
The real test will be whether independent researchers start trusting these predictions enough to change what they test in the lab, and whether regulators and clinicians accept AI-derived evidence. Expect rapid follow-up in the form of benchmarks, independent validations and debate over how these predictions should influence real-world genetics.