Google DeepMind has released the AlphaGenome Atlas, a dataset predicting the effects of approximately nine billion single-letter DNA changes in the human genome. The dataset is one petabyte in size, which is more than 30 times larger than the AlphaFold protein structure database. It is based on the AlphaGenome AI model introduced in 2025, which analyzes DNA sequences up to one million letters long to predict gene expression and protein binding.
The system includes the AlphaGenome Variant Impact Score (AVI), which simplifies thousands of prediction values into a single number. AVI uses a small neural network combining AlphaGenome predictions, the AlphaMissense protein model, and evolutionary conservation data. It utilizes 18 input features, whereas the benchmark tool CADD uses over 150.
In a clinical application, AVI identified a disease-causing variant in the DNM1 gene for a child with severe epilepsy who previously lacked a diagnosis. The variant creates an incorrect splice site in a brain-specific gene version, lengthening the protein by 13 building blocks. In a review of solved cases from the GREGoR consortium, AVI ranked the causal variant among the top 50 candidates in 29.5% of cases, compared to 12.5% for CADD.
The atlas is also used to enhance population studies. Research using data from over 54,000 UK Biobank participants showed that grouping variants by predicted effect increased the discovery of links between noncoding variants and blood protein levels by 22%. Additionally, the team derived 2,601 recurring short DNA patterns that serve as regulatory protein binding sites.
DeepMind states that the AlphaGenome Atlas and AVI are research tools and not diagnostic tools. The atlas is currently available for noncommercial use via a web portal, an API, and as a skill in Google Antigravity, with a commercial version planned for Google Cloud.