An AI map for narrowing variants that cannot all be tested
Google DeepMind released AlphaGenome Atlas on September 8. The one-petabyte resource precomputes AlphaGenome predictions for roughly nine billion possible single-nucleotide variants in the human genome. Researchers can use a free non-commercial portal and an API, with commercial access planned through Google Cloud.
The Atlas links thousands of predictions for gene expression, splicing, chromatin accessibility and other molecular processes across cell and tissue contexts. Its AVI score combines AlphaGenome and AlphaMissense to rank candidates in both protein-coding regions and the roughly 98% of the genome that is non-coding. Google says collaborators prioritized a previously overlooked DNM1-related rare-disease variant and then validated the predicted mechanism experimentally.
A prediction atlas is not a diagnostic report
The project did not experimentally test all nine billion changes; it precomputed model outputs. Benchmarks and case studies are reported by Google and collaborators, and performance can shift across populations, diseases and assays. A predicted molecular association also does not by itself establish disease causality or a therapeutic target.
DeepMind explicitly says AlphaGenome has not been validated or approved for clinical use. Researchers need independent cohorts, functional experiments and clinical-genetics review, and the score should not be used alone for diagnosis or treatment.