Google DeepMind launches AlphaGenome Atlas for variant research

Google DeepMind launched AlphaGenome Atlas on September 8, giving researchers an online tool for searching precomputed predictions about how single-letter DNA changes could affect molecular biology. The company says the one-petabyte resource covers 9 billion possible single-nucleotide variants in a reference human genome.

These entries are model-generated predictions, not experimental measurements of all 9 billion variants. The figure represents roughly 3 billion positions in the genome, with three alternative DNA letters possible at each one. By precomputing the results, DeepMind aims to help researchers rank candidate variants before choosing which ones to test in the laboratory.

AlphaGenome Atlas is a specialized genomics resource, distinct from Google DeepMind’s open Gemma model family.

The release also introduces the AlphaGenome Variant Impact, or AVI, score. It brings together outputs from AlphaGenome, which predicts molecular and regulatory effects, and AlphaMissense, a model focused on protein-altering variants. The result is a single number designed to rank changes across protein-coding and non-coding regions.

Each AVI score includes additive feature contributions intended to show which predicted processes drove it, including splicing, gene expression or chromatin accessibility. The score is a prioritization aid, not experimental proof that a variant causes a particular effect.

DeepMind is offering the Atlas through a no-code web portal for non-commercial research and also directs researchers to its AlphaGenome API. Commercial access through Google Cloud is planned, the company says, rather than generally available through the Atlas at launch.

DeepMind cited two collaborator projects as early case studies. The company said Broad Institute researchers used AVI to prioritize a variant in DNM1. AlphaGenome predicted that the change would create an incorrect splice site and an abnormal protein extension, and DeepMind said experimental screens validated that specific prediction. That reported result covers one prediction; it does not establish the accuracy of Atlas predictions more broadly.

DeepMind also said Gareth Hawkes of the University of Exeter applied Atlas predictions to whole-genome data from more than 54,000 UK Biobank participants. According to the company, the work found 22% more non-coding genetic associations and identified 19 regions linked to body mass index after the analysis was restricted to the 1% of variants that Atlas predicted would have the greatest impact. Those figures are company-reported collaborator results, not broad experimental validation of the Atlas.

The Atlas is intended to help researchers prioritize and interpret variants, but its predictions still require targeted experiments. DeepMind states that AlphaGenome has not been validated or approved for clinical use and that Atlas information is not medical advice, diagnosis or treatment.