AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
Google ● Covered by 4 sources
Google DeepMind released AlphaGenome Atlas, a free portal with predictions for 9 billion DNA letter changes. It could save researchers from having to test each mutation in the lab, one by one.
Based on reporting by Google — 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 its AlphaGenome model into something far bigger and easier to poke at: AlphaGenome Atlas, a public website and API that contains predictions for the effects of 9 billion single-letter DNA changes across the human genome. The company says it is the most comprehensive catalogue yet of how mutations affect molecular biology, and that academics can use it for free through the portal.
The basic problem is simple enough. Human DNA has roughly 9 billion possible single-nucleotide variants, and checking them one at a time in a lab is not realistic. Atlas is the workaround: Google has precomputed AlphaGenome’s predictions at scale, then wrapped them in a searchable interface that behaves more like a map than a model. Researchers can look up a variant and see predicted molecular effects without running the computation themselves.
The new piece of glue is the AlphaGenome Variant Impact score, or AVI. It folds together AlphaGenome and AlphaMissense, Google’s model for protein-altering variants, into a single number meant to help scientists rank variants quickly. The system is not just a score, though. Each result is linked to feature attributions, molecular effect predictions, and a library of more than 2,500 DNA motifs, so users can trace a prediction back to the biology behind it.
Google says outside collaborators have already used the atlas to find and validate variants in rare disease work and to spot rare variants linked to common traits. One example involved researchers at the Broad Institute, working with the GREGoR Consortium, who used the AVI score to prioritize overlooked variants and identify a change in DNM1 tied to epileptic encephalopathy. The prediction pointed to an incorrect splice site, and experimental screens backed it up.
The atlas is also being pitched as useful for population genetics. At the University of Exeter, Gareth Hawkes applied it to whole-genome data from more than 54,000 UK Biobank participants and found 22% more non-coding genetic associations by grouping variants according to predicted molecular effects. He also used the most impactful 1% of non-coding variants to identify 19 genetic regions linked to body mass index. Google says the whole dataset is 1 petabyte, more than 30 times larger than the AlphaFold Database, and it is available now through the website, the AlphaGenome API, and as a skill in Google Antigravity.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI release: less demo theatre, more giant pile of useful evidence. The real test, as always, is whether researchers can turn Google’s polished access into discoveries that survive contact with experiments, not just another shiny portal with a good press image.
Read more about this at: Google