DeepMind Launches AlphaGenome Atlas and AVI Score

DeepMind has launched a pre-calculated atlas covering approximately nine billion DNA substitutions and introduced a unique score, AVI, to prioritize variants. A case of epilepsy and cohort analyses illustrate its relevance, while the organization acknowledges limitations and proposes non-commercial access, with an offering via Google Cloud to come.

A GREGoR Case and Numbers Against Established Methods

The GREGoR consortium used a record dedicated to unresolved rare diseases as a real-world testing ground. In a child suffering from severe epilepsy with no diagnosis despite sequencing, the AVI elevated a variant of the DNM1 gene, previously classified as uncertain, to the top of the candidates. Predictions from AlphaGenome suggested the creation of an aberrant splice site extending the protein by 13 amino acids, an effect specific to a version of the gene expressed exclusively in the brain, explaining the lack of signal in blood analyses. A laboratory experiment confirmed this hypothesis, and researchers recommend reclassifying the variant as likely disease-causing. In cases already elucidated by GREGoR, the AVI positioned the causal variant among the top 50 choices in 29.5% of cases, while CADD achieved this in 12.5% of cases. DeepMind indicates that the AVI outperformed existing tools on clinically annotated variants, particularly in non-coding regions, although some competing tools remain superior for certain tasks.

What the Atlas Covers and How It Produces Its Predictions

The AlphaGenome Atlas provides predictions for each of the approximately nine billion single-letter changes, estimating their impact on molecular processes across hundreds of cell types and tissues. The dataset reaches a petabyte, more than 30 times the AlphaFold database. It relies on AlphaGenome, introduced in 2025, which reads sequences of one million letters of DNA and predicts gene expression intensity, binding of regulatory proteins, and transcript splicing. The model, which previously had to be queried variant by variant, now benefits from pre-calculated responses, averaging about 27,000 prediction values per variant. The atlas specifically targets the approximately 98% of the non-coding genome, where most disease-associated variants reside and whose effects were the hardest to deduce. More broadly, it aims to help sort through the numerous human variations at the individual level, while distinguishing rare deleterious variants remains beyond the reach of comprehensive experimental screening.

AVI, a Unique Score Based on 18 Characteristics for Each Variant

To make thousands of output values per variant actionable, the team designed the AlphaGenome Variant Impact Score (AVI), a single number summarizing the probability of impact. A small neural network aggregates predictions from AlphaGenome with those from the AlphaMissense model and two measures of evolutionary stability of a DNA site. The AVI is based on 18 input characteristics, significantly fewer than CADD, which utilizes over 150. Training was conducted by considering very rare variants as likely harmful and common variants as likely benign, due to uncertainty about the harmfulness of most individual mutations. For each variant, the atlas also details the processes contributing to the score, such as splicing or the action of a regulatory switch.

Population Studies: 22% Additional Links Detected

The atlas is designed to enhance cohort analyses, where the effect of rare variants must be evaluated collectively. When benign and effective variants mix, the signal gets drowned out. Gareth Hawkes (University of Exeter) used the atlas to group only variants predicted to have similar effects in data from over 54,000 participants in the UK Biobank. This strategy uncovered 22% additional associations between non-coding variants and blood protein levels compared to standard filters. From the predictions, the team also derived 2,601 recurring short DNA motifs, preferred binding sites for regulatory proteins.

Known Limitations and Access Modalities for the Atlas

DeepMind highlights limitations: not all cell types are modeled, and some effects mediated by the quantity of other regulatory proteins still elude the model. The atlas and the AVI are described as tools intended for research and represent only one possible piece of evidence in a diagnostic process. Regarding access, the atlas can be used for free for non-commercial purposes via a web portal, an API, and as an integrated skill in Google Antigravity. Commercial access will be offered through Google Cloud.