On September 8, 2026, Google DeepMind released AlphaGenome Atlas. The freely accessible, 1-petabyte database contains AI-generated predictions for every possible single-letter change across the human genome, roughly 9 billion variants, plus more than 100 million short insertions and deletions.
Until now, accessing those predictions meant writing code against an API, a barrier that limited usage to around 9,000 researchers globally. The atlas converts that process into a browser search, which sounds like a minor convenience until you consider that the difference between calling an API and Googling something is often the difference between who does science and who doesn’t. Tools like AI-powered websites have already begun reshaping who can access complex capabilities without specialized technical skills.
From Model to Map
DeepMind precomputed AI predictions for every DNA position in a reference human genome and packaged them into a searchable database with a single impact score per variant.
DeepMind built the atlas by running its AlphaGenome model across every position in a reference human genome, comparing each DNA base against its three nucleotide alternatives and precomputing the predicted molecular effects: chromatin state changes, transcription factor binding shifts, splicing alterations, and gene expression impacts.
DeepMind developed the AlphaGenome Variant Impact (AVI) score to make those thousands of predictions per variant usable. It combines AlphaGenome’s regulatory predictions with AlphaMissense’s protein-pathogenicity estimates, covering both coding and non-coding regions in one unified framework.
An AVI of 10 places a variant in the top 10% for predicted biological disruption; AVI 30 puts it in the top 0.1%.
The score is also decomposable. Researchers can see exactly which processes, whether splicing, gene expression, or protein alteration, drive a given variant’s number, giving them mechanism alongside ranking.
If you remove the friction you also increase the curiosity for people to dive in.
Žiga Avsec, AlphaGenome team lead, Google DeepMind, via Nature
Real-World Use, Real-World Limits
Demonstrated use cases show genuine progress in rare-disease research, though leading experts urge caution about clinical application.
Early results suggest the AVI score moves the needle in rare-disease research. On solved cases from the GREGoR Consortium, AVI placed the known causal variant in the top 50 candidates 29.5% of the time, compared to 12.5% for the previous standard tool.
Broad Institute researchers used those predictions to identify a non-coding variant as a likely cause of severe epilepsy, a finding later supported by experimental validation.
At the University of Exeter, statistical geneticist Gareth Hawkes applied atlas predictions across 54,000 UK Biobank whole genomes. The analysis surfaced rare non-coding variants associated with blood protein levels across thousands of traits.
Preprint co-author Julia Zeitlinger describes the parallel motif-mapping work as building “a searchable dictionary for non-coding DNA.”
Expert reaction mixes genuine appreciation with pointed caution. Martin Kircher, a Berlin-based bioinformatician, calls the atlas “a useful and generous way to scale up access to a strong model,” while stressing it will not replace experiments or individual case-level clinical judgment.
The sharpest note comes from Ben Lehner of the Wellcome Sanger Institute.
This isn’t an ‘AlphaFold moment’. The predictive performances of genomic models like AlphaGenome are still far from brilliant and these models should not yet be used alone to make clinical decisions.
Ben Lehner, molecular biologist, Wellcome Sanger Institute, via Nature
Worth noting: the supporting research is currently a bioRxiv preprint, meaning peer review is ongoing.
Access and What Comes Next
The atlas is free for non-commercial research now, with commercial access through Google Cloud planned for the near term.
The atlas is free for non-commercial research via web portal and API, with commercial access planned through Google Cloud licensing. DeepMind’s drug-discovery sister company Isomorphic Labs will operate under that commercial framework.
Avsec expects AI agents, systems that autonomously navigate biological databases, to become the atlas’s most active power users as the tooling matures.
The atlas doesn’t hand you a diagnosis or a drug target. What it does is remove a significant computational wall that kept high-capacity genomic AI out of reach for most researchers. Whether it eventually earns its AlphaFold comparison depends on experimental data that, at the required scale, simply doesn’t exist yet.
