Google DeepMind introduced AlphaGenome Atlas on Tuesday, publishing molecular impact predictions for 9 billion single-nucleotide variants across the human genome. The 1-petabyte dataset catalogues the effects of every possible single-letter DNA mutation across both coding and non-coding sequences, spanning hundreds of human and mouse cell types and tissues.
The platform opened for non-commercial academic research through a free web portal, an application programming interface, and as a skill inside Google Antigravity. DeepMind plans to release commercial access on Google Cloud soon. The dataset is more than 30 times larger than the AlphaFold Database, which expanded in 2022 to over 200 million protein structure predictions.
Scoring variant impact
To summarize individual mutations, DeepMind released the AlphaGenome Variant Impact score. The metric combines outputs from AlphaGenome with AlphaMissense, a model focused on protein-altering variants, into a single numerical ranking. The system assesses both the 2% of the genome that codes for proteins and the remaining 98% that forms non-coding regulatory sequences.
The database also links each score to specific feature attributions, identifying whether molecular mechanisms such as RNA splicing or gene expression drive a mutation's impact. In addition, the release includes a catalogue of more than 2,500 recurring DNA sequence motifs to help locate binding sites for transcription factors.
Research findings
External research teams tested the atlas on rare diseases and population data prior to release. Scientists Laura Covill and Anne O'Donnell-Luria at the Broad Institute, working with the GREGoR Consortium, used the scoring tool to identify a previously overlooked mutation in the DNM1 gene linked to epileptic encephalopathy. Experimental testing confirmed that the variant created an incorrect splice site, leading to an abnormal protein extension.
At the University of Exeter, Medical Research Council fellow Gareth Hawkes applied the atlas to whole-genome data from more than 54,000 UK Biobank participants. Grouping rare variants by predicted molecular effect revealed 22% more non-coding associations than statistical analysis alone, pinpointing regulatory variants tied to the proteins PLA2G7 and EGLN1. Hawkes also evaluated non-coding variants linked to body mass index, isolating 19 genetic regions from the top 1% of impactful predictions. At the Stowers Institute for Medical Research, Julia Zeitlinger and Melanie Weilert used the sequence motifs to classify transcription factors that alter DNA accessibility versus those that regulate gene expression.
DeepMind stated that AlphaGenome Atlas is not approved or validated for clinical use and does not constitute medical advice. The underlying AlphaGenome base model remains available on GitHub and the AlphaGenome API for academic use, and on Google Cloud Model Garden for commercial applications.
