DeepMind's AlphaGenome Atlas catalogues 9 billion DNA variants. Independent labs will test whether it holds up.

Google DeepMind and five academic partners released AlphaGenome Atlas on Tuesday, a one-petabyte public dataset of AI-generated predictions for more than 9 billion possible single-letter changes in the human genome. A single-letter change is a substitution of one of the four DNA bases (A, T, C, or G) at one of roughly 3 billion positions. The catalog is built from a single sequence model and is positioned by the consortium as the first resource to both rank a DNA variant across the genome and predict the specific biological process that variant disrupts, such as altered gene regulation, splicing, or protein binding.

Researchers can query the dataset through a browser interface. The underlying model is described in a peer-reviewed Nature paper and a bioRxiv preprint; open code is on GitHub. The partner roster is the Stowers Institute for Medical Research, Broad Institute, University of Exeter, Memorial Sloan Kettering Cancer Center, and Stanford, with Julia Zeitlinger and Žiga Avsec named in the joint press release.

The intended users are academic and clinical researchers, not patients, and the resource is presented as a prioritization layer for variant interpretation, since wet-lab testing of every single-letter change is impractical. The "first to both rank and predict" framing is a press-release claim rather than an independent benchmark, and the field will test it against the existing toolkit: AlphaMissense (missense variants only), CADD (conservation-based scoring), and gnomAD (observed allele frequencies, not predicted effects). The watch item is independent benchmarks on held-out clinical variants.