Google DeepMind Releases AlphaGenome Atlas With Precomputed Molecular Effect Predictions and AVI Scores for 9 Billion Human DNA Variants
MarkTechPost Asif Razzaq ● Covered by 4 sources
DeepMind put 9 billion human DNA changes into a searchable atlas. It turns variant hunting from slow guesswork into a precomputed lookup table.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has turned AlphaGenome into something closer to a genome map than a model demo. The company says AlphaGenome Atlas contains precomputed predictions for every possible single-nucleotide variant in the human genome, which works out to roughly 9 billion one-letter changes.
That matters because the old workflow was clunky. AlphaGenome, released in June 2025, could already predict how a DNA variant affects things like gene expression and RNA splicing, but only one variant or one region at a time. DeepMind’s team ran it across the whole human genome, stored the outputs, and ended up with a 1-petabyte dataset. They describe that as more than 30 times the size of the AlphaFold Database, which holds over 200 million protein structure predictions.
The Atlas comes with four linked pieces. There are the molecular effect predictions themselves, with thousands of predictions per variant across hundreds of human and mouse cell types and tissues. There is the AVI score, a single impact number that combines AlphaGenome’s regulatory predictions with AlphaMissense so it can cover coding and non-coding DNA. There are feature attributions, which break the score into interpretable chunks like chromatin accessibility, splicing, and conservation. And there is a catalogue of over 2,500 recurrent DNA motifs, including transcription factor binding sites.
The access story is mixed, which is very Google Cloud in 2026. The Atlas is available as a free web portal for academic use, through the AlphaGenome API, and as a skill in Google Antigravity. Non-commercial research users can query it now. Commercial access on Google Cloud is listed as coming soon. Meanwhile, the underlying AlphaGenome model is already on GitHub for academic use and in Model Garden on Google Cloud for commercial use.
DeepMind says AVI performs best in class across many variant pathogenicity and rare disease benchmarks, with the technical report carrying the details. The bigger point is simpler: a lookup table beats rerunning a giant model every time someone wants to check a candidate variant. That is a practical win, not a flashy one, and science usually needs more of those.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI release that sounds useful instead of merely grand. Precomputing the boring part and leaving the interpretation exposed is exactly the sort of unglamorous move genomics needed. The industry keeps worshipping models; sometimes the real product is a table.
Read more about this at: MarkTechPost