Google DeepMind launched AlphaGenome Atlas on Sept. 10, the largest-ever map for predicting genomic effects. It covers 9 billion possible single-nucleotide variants (SNVs), has a total database capacity of 1 PB—30 times larger than the AlphaFold Database launched in 2022—and is freely available to academic researchers starting today.
The Scale and Architecture of AlphaGenome Atlas
AlphaGenome Atlas has a total capacity of 1 PB, making it 30 times larger than the AlphaFold Database launched by Google DeepMind in 2022. At the time, the AlphaFold Database expanded protein structure data from 190,000 experimentally determined structures to more than 200 million predicted structures and provided a visualization interface to lower the barrier to entry. AlphaGenome Atlas uses the same strategy: AI performs calculations at a scale that would be impossible to achieve in the laboratory, then packages the results into an interface that researchers can query directly.
The core metric, the AVI score, is calculated by combining the prediction results of the AlphaGenome and AlphaMissense models. It condenses complex molecular biology predictions into a single value, so researchers do not need to compare dozens of metrics individually and can instead use the score to determine which variants merit further study.
AVI Functional Attribution and Research Tools for 2,500 DNA Sequence Motifs
Each AVI score comes with an “AVI functional attribution” report identifying the molecular mechanism primarily affected by the variant, such as RNA splicing, gene expression levels, or chromatin compactness. This allows researchers to determine not only the severity of the effect but also where in the process the problem occurs.
AlphaGenome Atlas also organizes more than 2,500 recurring DNA sequence motifs, which can be used to locate transcription factor binding sites. Transcription factors are proteins that determine whether genes are switched on or off, and the ability to locate these motifs directly determines whether researchers can interpret the function of a non-coding sequence.
Julia Zeitlinger and Melanie Weilert of the Stowers Institute for Medical Research in the United States used the sequence motif resources to distinguish which transcription factors are responsible only for opening up DNA and which can actually activate or silence genes—details that previously required repeated experiments to piece together gradually.
The GREGoR Consortium’s DNM1 Epilepsy Case
The GREGoR Consortium approached Laura Covill and Anne O'Donnell-Luria of the Broad Institute, who used AVI scores to rescreen thousands of candidate variants and identify suspected pathogenic variants that had previously been overlooked. They focused on the DNM1 gene, which is strongly associated with epileptic encephalopathy. AlphaGenome’s prediction indicated that one variant created an incorrect splice site, causing cells to misread the gene’s instructions and resulting in an extra segment being added to the protein downstream.
Subsequent experimental screening confirmed this conclusion and also identified nearby variants with similar effects. Google DeepMind specifically noted in its database documentation that this case is credible because it was followed by experimental validation rather than being concluded solely on the basis of the AVI score.
Gareth Hawkes Identifies 22% More Non-Coding Region Associations
Gareth Hawkes, a researcher at the University of Exeter in the United Kingdom, applied AlphaGenome Atlas to whole-genome data from more than 54,000 UK Biobank participants. After grouping variants according to their predicted effects, he identified 22% more associations in non-coding regions that had previously been buried in statistical noise. He successfully located variants regulating PLA2G7, which is associated with aging, and EGLN1, a key gene involved in cellular oxygen sensing.
Hawkes then focused on the top 1% of non-coding variants that the system determined to have the greatest impact. After comparing hundreds of millions of data points, he identified 19 gene regions associated with BMI, which could guide the next stage of research. This case shows that AlphaGenome Atlas can not only identify pathogenic variants within individual families but also recover signals previously buried in population data from hundreds of thousands of people.
Conditions for Free Academic Access, Clinical Validation Limitations, and Antigravity Integration
The current access conditions and usage restrictions for AlphaGenome Atlas are as follows:
Free academic version: Limited to non-commercial academic use; available for queries starting today
Commercial version: Coming soon to Google Cloud for enterprise users
AlphaGenome base model: Already available on GitHub for free academic use
Enterprise commercial model: Available through Model Garden
Clinical limitation: The website clearly states that the predictions have not been clinically validated and cannot be used as medical advice
Google DeepMind acknowledges that AlphaGenome Atlas is a foundation rather than an endpoint. As the model continues to improve, the map will become increasingly detailed, and the current version is not the final answer. In addition, AlphaGenome Atlas has been integrated into Google Antigravity as a directly callable function in scientific workflows, advancing the creation of a unified scientific research platform.
Frequently Asked Questions
Which models are used to calculate AlphaGenome Atlas’s AVI score, and what is the rationale behind its design?
The AVI score is calculated by combining the prediction results of the AlphaGenome and AlphaMissense models, condensing complex molecular biology predictions into a single value. Each score includes an “AVI functional attribution” report identifying whether the variant affects RNA splicing, gene expression levels, or chromatin compactness. Google DeepMind’s internal tests indicate that the AVI score achieved best-in-class performance across multiple pathogenicity and rare disease benchmarks.
What specific findings did Gareth Hawkes achieve in the UK Biobank study?
Hawkes applied AlphaGenome Atlas to whole-genome data from more than 54,000 UK Biobank participants and identified 22% more associations in non-coding regions that had previously been buried in statistical noise. After focusing on the top 1% of the most influential non-coding variants, he further identified 19 gene regions associated with BMI.
What are AlphaGenome Atlas’s current access conditions for academia and its clinical-use limitations?
The free academic version is limited to non-commercial use and is available for queries starting today. The AlphaGenome base model is already available on GitHub. The enterprise commercial version is coming soon to Google Cloud and can also be accessed through Model Garden. The website clearly states that the predictions have not been clinically validated and cannot be used as medical advice.
