Google DeepMind has released the AlphaGenome Atlas, a precomputed database of over 9 billion possible single-letter DNA changes in humans. The dataset exceeds 1 petabyte in size, making it approximately 30 times larger than the AlphaFold database. The tool is currently free for noncommercial researchers via a web portal, the AlphaGenome API, and Google Antigravity.
The Atlas is based on the AlphaGenome model launched in January, which can analyze up to 1 million DNA letters to predict molecular measurements like RNA splicing and gene expression. According to Nature, the model outperformed other non-specialized models in 25 out of 26 benchmarks. DeepMind plans to introduce commercial access to the database through Google Cloud in the future.
To simplify data access, DeepMind introduced the AlphaGenome Variant Impact (AVI) score. This metric combines predictions from AlphaGenome and AlphaMissense to provide a single numerical value for prioritizing genetic variants. This approach removes the need for coding knowledge, which previously limited the API's usage to around 9,000 researchers.
Collaborators have already demonstrated the utility of the Atlas. Researchers at the University of Exeter found 22% more associations regarding rare noncoding variants affecting blood proteins by using the Atlas. In one instance, they reduced a pool of 526 candidates to just four. Meanwhile, the Stowers Institute for Medical Research used the tool to categorize transcription factors across different cell types.
