DeepMind releases AlphaGenome Atlas, mapping predicted effects of 9 billion DNA variants

Google DeepMind has released AlphaGenome Atlas, a free dataset that predicts the molecular effect of nearly every possible single-letter change to human DNA. The company described it in a blog post on Tuesday as a “predictive map of every possible DNA letter change in the human genome.” Researchers can reach the predictions through a web portal, through DeepMind’s AlphaGenome interface, and as a skill inside Antigravity, Google’s platform for building AI agents.

DNA is written with four chemical letters, A, C, G, and T, and the human genome, the body’s full DNA code, holds about three billion letter pairs. Those letters set how genes switch on and off and by how much. A single changed letter can be harmless, mark an ordinary difference between people, or play a part in disease, and telling those cases apart is a long-standing problem in biology. DeepMind puts the number of possible single-letter substitutions at roughly nine billion.

Atlas contains a prediction for each of those nine billion variants, estimating effects such as how much of a given protein a cell makes. DeepMind’s team calls it “the most comprehensive catalogue of how genetic mutations affect molecular biology.” To help researchers narrow the field, Google is also releasing a Variant Impact Score, shortened to AVI, that draws on its other DNA-effect models to rank variants. The company says researchers can now “rapidly rank variants and interpret their molecular effects at the same time.”

The project builds on AlphaGenome, a model DeepMind released last year to find genetic drivers of disease, and on AlphaMissense, an earlier tool that predicted which small mutations change proteins. Atlas extends the predictions to most of the genome, including the long stretches that do not code for proteins but control when and where genes are active. Ziga Avsec, DeepMind’s genomics lead, said in a press briefing that the model itself was already public and that the catalog took time to build. “Basically it took us some time to really precompute and also analyze this many variants because the space is so big,” he said.

The predictions come from a model trained on public human and mouse genome databases, then applied to billions of variants to make a dataset of about one petabyte, or a million gigabytes. The Verge’s report gives no independent check of the full nine-billion-variant catalogue, and the suggestion that Atlas could speed research and new treatments is attributed to DeepMind’s scientists. The release also lands as DeepMind cofounder Demis Hassabis steps back from running the lab to focus on scientific research, including the drug-discovery spinoff Isomorphic Labs.

Google says the data is available for noncommercial research through its website starting today, with commercial access on Google Cloud to follow “soon” and no firm date attached. Worth watching is whether outside labs report that AVI scores line up with results from bench experiments, and what terms Google sets for the paid Google Cloud version once it opens.