Google DeepMind maps 9 billion human DNA variants
AlphaGenome Atlas uses AI to predict the molecular impact of every possible single-letter change in the human genome. The platform brings those predictions together in a one-petabyte catalogue and has already been used to identify variants associated with rare diseases and complex traits.
Summary
Google DeepMind has introduced AlphaGenome Atlas, a platform containing predictions about the molecular effects of roughly 9 billion single-letter variants in the human genome. The system uses the AlphaGenome model to precompute the impact of every possible change and turn the results into a searchable map for researchers.
The Atlas contains one petabyte of data and is designed to make a poorly understood part of the genome easier to explore. Around 2% of human DNA codes for proteins; the remaining 98% is non-coding and regulates gene activity, but is much harder to interpret.
In practice
The platform’s main tool is the AlphaGenome Variant Impact (AVI) score, which combines predictions from AlphaGenome with AlphaMissense, DeepMind’s model for protein-altering variants. AVI lets researchers rank variants by predicted impact and examine which biological processes contribute most to each score.
The Atlas contains thousands of molecular-effect predictions for each variant, covering different aspects of gene regulation across hundreds of human and mouse cell types and tissues. It also links each score to specific biological features, including gene expression, RNA splicing and chromatin accessibility, and includes a catalogue of more than 2,500 recurring DNA motifs.
Context
In collaboration with the Broad Institute and the GREGoR Consortium, researchers used the Atlas to prioritise variants in an unsolved rare-disease case. A variant in the DNM1 gene was linked to the creation of an incorrect splice site, and laboratory experiments validated the prediction.
In another study, Gareth Hawkes applied the Atlas to data from more than 54,000 UK Biobank participants. By grouping non-coding variants according to their predicted molecular effects, he found 22% more genetic associations and identified 19 regions linked to body-mass index among the 1% of variants predicted to have the greatest impact.
Why it matters
- The Atlas turns billions of scattered predictions into a resource for searching variants and comparing their effects quickly.
- Interpreting non-coding regions could help identify disease mechanisms that do not appear in protein-coding parts of DNA.
- The platform is already available for academic research through a no-code portal; DeepMind says commercial access through Google Cloud will follow.