Pushmeet Kohli, vice-president for research at DeepMind, announced the launch of the AlphaGenome Atlas, a pre-computed database that predicts the biological impact of every possible single-letter change in the human genome. The resource, hosted by Google, is now openly accessible to academic researchers worldwide.
What the Atlas contains
The AlphaGenome Atlas maps each of the 9 billion potential single-base substitutions to likely effects on gene regulation, expression and protein synthesis. Using the AlphaGenome model introduced last year, DeepMind evaluated every reference base against the three alternative nucleotides, generating around 27 000 predictions per variant across hundreds of human and mouse cell types.
To aid interpretation, the team also provides an AlphaGenome Variant Impact (AVI) score, which combines regulatory predictions with those from the earlier AlphaMissense model that focuses on protein-altering changes. An AVI score of 10 places a variant in the top 10 percent of impact, while a score of 30 ranks it among the strongest one in a thousand.
Why the Atlas matters
Only about two percent of the genome encodes proteins; the remaining 98 percent governs when and where genes are switched on. Until now, scientists have struggled to assess mutations in these non-coding regions. By delivering a comprehensive, instantly searchable map, the Atlas promises to speed up the identification of disease-causing variants and guide the development of new therapies.
"We bought the book of the Human Genome Project, but we did not understand how to read it," Kohli said, echoing the sentiment that the original sequencing effort left many functional questions unanswered.
Early testing and practical applications
Researchers at the Broad Institute, working with the GREGoR Consortium, used the AVI score to re-evaluate unsolved rare-disease cases. In one instance, the Atlas highlighted a splice-altering mutation in the DNM1 gene linked to epileptic encephalopathy, a finding later confirmed in the laboratory.
In a retrospective analysis of previously solved cases, the AVI score ranked the true causal variant within the top 50 candidates 29.5 percent of the time, compared with 12.5 percent for the established CADD method.
University of Exeter Medical Research Council fellow Gareth Hawkes applied the Atlas to over 54 000 UK Biobank genomes, uncovering rare non-coding variants that influence blood protein levels and increasing association yields by 22 percent.
At the Stowers Institute, investigator Julia Zeitlinger used the Atlas motif catalogue to classify transcription factors by their regulatory roles across cell types, a task that would have been impractical without the computational resource.
Integration and future steps
Ewan Birney, director of the European Bioinformatics Institute (EMBL-EBI), confirmed that the AVI score will be incorporated into Ensembl's Variant Effect Predictor, expanding its reach within the bioinformatics community.
While the Atlas offers a powerful starting point, DeepMind cautions that its predictions do not replace experimental validation. According to genomics lead Žiga Avsec, the model performs well for splicing and promoter variants but may miss effects in distal enhancers, and its overall accuracy is lower than that of AlphaFold for protein structures.
DeepMind plans to make the Atlas available for commercial use through a licensing arrangement on Google Cloud in the near future, while the free academic version remains online.