AlphaGenome Atlas: a high-resolution map of human DNA

gist

Google DeepMind's Pushmeet Kohli and Žiga Avsec introduce AlphaGenome Atlas, a queryable database of predicted effects for every possible single-nucleotide variant in the human genome. The AlphaGenome model pre-computed regulatory impact across about 9 billion single-letter changes into a 1-petabyte atlas, surfaced through a no-code portal and a single AlphaGenome Variant Impact (AVI) score spanning coding and non-coding regions. Early users at the Broad Institute and on UK Biobank data report faster rare-disease triage and more non-coding associations for complex traits such as BMI.

ideas

• Map every possible letter change. Atlas stores pre-calculated AlphaGenome predictions for all ~9 billion single-nucleotide variants as a 1-petabyte regulatory-impact database researchers can query without re-running the model.
• AVI collapses thousands of channels. One Variant Impact score combines coding and non-coding predictions so labs can prioritize candidates without scanning every molecular readout.
• Rare disease gets a splice-site lead. At the Broad Institute, Laura Covill's team used AVI to highlight a DNM1 variant predicted to create an incorrect splice site and help close an unsolved case.
• Complex traits gain non-coding signal. On 54,000+ UK Biobank participants, Gareth Hawkes grouped variants by predicted molecular effects, found 22% more non-coding associations, and linked 19 BMI regions from the top 1% of impactful variants.
• Zero-code access is the distribution bet. A public website portal targets clinical researchers and biologists who will not write API clients, with a longer write-up on the Google DeepMind blog.

quotes