The Pulse
DeepMind Maps 9 Billion Human DNA Variants With AlphaGenome
Google DeepMind has released AlphaGenome Atlas, a searchable dataset predicting the molecular effects of every possible single-letter change in the human genome. The 1-petabyte resource adds a scoring system designed to help researchers pri
AI.info Team ·
“The human genome is a massive search space. Finding a genetic variant that matters for a specific disease or trait can be a difficult process, like finding a needle in a haystack.”
Dr. Gareth Hawkes, Medical Research Council fellow, University of Exeter
Google DeepMind has released a searchable map of the predicted effects of every possible single-letter change in the human genome, turning a calculation that would be impractical to repeat variant by variant into a precomputed research resource.
Called AlphaGenome Atlas, the platform contains predictions for roughly 9 billion single-nucleotide variants. DeepMind says the dataset is available for academic research through a free website portal, the AlphaGenome API and an integration with Google Antigravity, the company’s AI-powered scientific workbench.
AlphaGenome Atlas puts the genome’s 98% in reach
The atlas builds on AlphaGenome, a model DeepMind introduced in June 2025 to predict how DNA variants affect gene regulation. AlphaGenome can analyze long DNA sequences and estimate changes in processes including gene expression, RNA splicing, chromatin accessibility and interactions between distant sections of DNA.
DeepMind says the new resource applies those predictions across the complete set of possible single-letter substitutions in the human genome. Researchers can examine variants in protein-coding regions as well as the much larger non-coding portion of DNA, which does not directly produce proteins but helps control when, where and how genes operate.
Only about 2% of the human genome directly codes for proteins, according to DeepMind. The remaining 98% contains regulatory sequences that influence gene activity and includes many variants associated with traits and disease. That imbalance has made non-coding variants difficult to interpret, because researchers often face huge candidate lists with little experimental guidance.
A one-petabyte catalogue with a single impact score
AlphaGenome Atlas stores thousands of predicted molecular effects for each variant across hundreds of human and mouse cell types and tissues. DeepMind describes the result as a 1-petabyte dataset, more than 30 times larger than the AlphaFold Database was after its 2022 expansion.
The platform also introduces the AlphaGenome Variant Impact, or AVI, score. AVI combines AlphaGenome’s predictions with AlphaMissense, DeepMind’s model for estimating the effects of protein-altering variants, and condenses the results into a single value that researchers can use to rank candidate mutations.
DeepMind also provides feature attributions for AVI scores. Those breakdowns indicate which predicted processes contribute to a variant’s ranking, such as altered splicing, gene expression or chromatin accessibility. The atlas includes more than 2,500 recurring DNA sequence motifs and their locations, giving researchers another way to connect a variant with the regulatory sequences it may disrupt.
Rare disease research supplies an early test
DeepMind says external collaborators have already used AlphaGenome Atlas in rare disease studies and population genetics. In work with the GREGoR Consortium, researchers used AVI scores to prioritize variants in cases where the genetic cause of a disease remained unknown.
One of those predictions pointed to a variant that created an incorrect splice site, leading to an abnormal extension in the resulting protein. Experimental screens validated the predicted effect and identified nearby variants with similar behavior, according to DeepMind.
The resource is also intended to help with variants that influence common traits rather than causing a single rare disorder. At the University of Exeter, Hawkes analyzed whole-genome data from more than 54,000 UK Biobank participants. Grouping rare variants according to their predicted molecular effects allowed the study to identify 22% more non-coding genetic associations than an approach that did not use the atlas, DeepMind says.
In a separate analysis of body-mass-index data, Hawkes focused on the 1% of non-coding variants that AlphaGenome Atlas predicted to have the greatest impact. That process identified 19 genetic regions for follow-up research.
Predictions narrow the search; experiments still decide
AlphaGenome Atlas does not replace laboratory testing or establish that a mutation causes a disease. Its entries are model predictions derived from patterns in experimental genomic data, and the atlas is most useful when it helps researchers decide which candidates to test first.
That distinction matters because a high predicted impact can indicate that a variant changes a molecular process without proving that the change produces a particular clinical outcome. DeepMind presents the atlas as a baseline that can improve as AlphaGenome and related models are updated.
The company is making the atlas available for non-commercial use through its website. DeepMind says commercial access will become available on Google Cloud, while the AlphaGenome base model is already offered for academic use through its API and GitHub repository and for commercial use through Google Cloud’s Model Garden.
From individual variant analysis to genome-scale search
AlphaGenome originally gave researchers a way to submit a DNA sequence and inspect predictions for a particular variant. Atlas changes the workflow by precomputing the results for the entire set of possible single-letter changes, allowing users to search and rank candidates without running every analysis from scratch.
That shift is most significant for non-coding DNA, where the number of plausible candidates can overwhelm traditional statistical and experimental methods. The atlas turns those candidates into a ranked research queue, but the concrete value will depend on how often its highest-scoring variants survive laboratory validation and improve disease diagnosis or biological understanding.