Google DeepMind has released AlphaGenome Atlas, a searchable artificial-intelligence resource containing predictions for the molecular effects of all 9 billion possible single-letter changes in the human genome. Announced on September 8, the one-petabyte database is available worldwide for non-commercial research through a web portal and is intended to help scientists decide which genetic variants warrant closer investigation.
The human genome contains roughly 3 billion DNA letters, and each position can be replaced by three alternative letters. Only about 2% of the genome directly encodes proteins, while much of the remaining 98% regulates when and where genes operate. DeepMind built the atlas by applying its AlphaGenome model across a reference genome and precomputing effects that previously required researchers to query variants individually.
Each variant links to thousands of predictions spanning gene expression, RNA splicing, chromatin accessibility and other regulatory processes across hundreds of human and mouse cell types and tissues. The resource also includes more than 100 million observed insertions and deletions and maps over 2,500 recurring DNA motifs. A new AlphaGenome Variant Impact score combines regulatory predictions with AlphaMissense protein-impact estimates into one ranking, with feature breakdowns showing what drives the result.
Early collaborators reported concrete research uses. Broad Institute and GREGoR Consortium researchers prioritized a DNM1 variant in a severe epilepsy case; the model predicted an incorrect splice site, and laboratory screens supported that mechanism. In a retrospective set of solved rare-disease cases, the score ranked the known causal variant among the top 50 candidates 29.5% of the time, compared with 12.5% for the established CADD method.
University of Exeter researcher Gareth Hawkes also applied the atlas to whole-genome data from more than 54,000 UK Biobank participants. Grouping rare non-coding variants by predicted molecular effects produced 22% more associations with circulating protein levels than an analysis without the atlas. Nature reported that independent specialists view the removal of computational barriers as valuable, especially for teams unable to run such a model across an entire genome.
The release remains a map of predictions rather than proof about every mutation. DeepMind says AlphaGenome performs better for some variant classes than others, can miss effects in regulatory enhancers and is neither validated nor approved for clinical use. Researchers must confirm important signals experimentally and consider patient evidence before drawing diagnostic conclusions; the company describes the atlas as a baseline that can improve as its models and biological data advance.
Comments