Google DeepMind Releases AlphaGenome Atlas, a Catalog of Molecular Effects for 9 Billion Human DNA Variants

Human DNA is written in a four-letter chemical alphabet — A, C, G and T — and the human genome contains roughly 3 billion letter pairs. Single-letter changes can be harmless, account for ordinary differences between people, or contribute to disease. Because about 9 billion single-letter substitutions are possible, researchers have long faced the challenge of working out which changes matter and how. Previously, interpreting a variant required running an AI model one variant at a time or performing laboratory experiments, a painstakingly slow process.

The Atlas turns that into a lookup table. For each variant, it provides thousands of predictions about molecular effects, including gene regulation across hundreds of human and mouse cell types and tissues. It also assigns every variant an AlphaGenome Variant Impact (AVI) score, which combines AlphaGenome's regulatory predictions with AlphaMissense, DeepMind's model for protein-altering variants. This lets the score cover both the protein-coding portion of the genome (about 2%) and the remaining 98% of non-coding DNA. Each AVI score is decomposed into additive attributions for categories such as chromatin accessibility, splicing, conservation and protein effects, so researchers can see which biological process a variant may disrupt. The resource also includes a compendium of more than 2,500 recurrent DNA motifs, including transcription factor binding sites.

The project builds on AlphaGenome, a model DeepMind released in 2025, and scales it across every possible single-base change. Applying the model to all 9 billion variants produced a dataset of roughly 1 petabyte, which DeepMind says is more than 30 times the size of the AlphaFold Database. “Basically it took us some time to really precompute and also analyze this many variants because the space is so big,” Ziga Avsec, DeepMind's genomics lead, said in a press briefing. The AlphaGenome model was trained using public databases of human and mouse genomes.

The Atlas is available today for non-commercial research through a web portal, the AlphaGenome API, and as a skill in Google's agentic development platform Antigravity. DeepMind says commercial access will come through Google Cloud “soon.” The underlying AlphaGenome model is already available for academic use on GitHub and for commercial use on Model Garden. Kohli said that Isomorphic Labs, DeepMind's sister company devoted to drug discovery, would have access to Atlas but would still require a commercial license; terms were not specified. DeepMind underlined that the resource is not for clinical use.

Early results from external researchers accompany the release. Within the GREGoR Consortium, Laura Covill and Anne O'Donnell-Luria of the Broad Institute used AVI scores to re-prioritize variants that earlier analyses had overlooked. The scores flagged a variant in DNM1, a gene strongly linked to epileptic encephalopathy. The underlying predictions showed the variant creates an incorrect splice site that abnormally extends the protein, and experimental screens validated the mechanism while finding nearby variants with similar effects. In population genetics, Gareth Hawkes, a Medical Research Council fellow at the University of Exeter, applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants; grouping rare variants by predicted molecular effect uncovered 22% more non-coding associations than previously detectable.

Kohli framed the release as carrying forward the unfinished business of the Human Genome Project, which mapped the full human DNA sequence in 2003. “As the saying goes, we bought the book,” he said, “but we did not understand how to read it.” A paper describing the Atlas and its construction is being released on bioRxiv. The launch continues DeepMind's push into science after AlphaFold, the protein-structure prediction work that won Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry, and comes as Hassabis steps back from daily AI-lab management to concentrate on research efforts including Isomorphic Labs. DeepMind says the AVI score shows best-in-class performance on variant-pathogenicity and rare-disease benchmarks, with details in the technical report.