Credit: Google DeepMind has released AlphaGenome Atlas, a petabyte of precomputed predictions for all 9 billion possible single-letter changes in the human genome, free for academic use. Its headline validations came from UK Biobank whole genomes at Exeter and from the Broad Institute, and commercial access is heading to Google Cloud. DeepMind has published AlphaGenome Atlas, a precomputed prediction for every one of the 9 billion possible single-letter changes in the human genome. The London lab’s dataset runs to a petabyte and is free for academic use.
Each variant gets an AlphaGenome Variant Impact score, a single number combining AlphaGenome with AlphaMissense. It covers the 2% of the genome that codes for proteins and the 98% that does not, and ships alongside a catalogue of more than 2,500 recurrent DNA motifs. The file is more than 30 times the size of the AlphaFold database, which took protein structures from about 190,000 experimental entries to more than 200 million predictions. The validation is the part that matters from here.
Gareth Hawkes, a Medical Research Council fellow at Exeter, ran it against whole genomes from more than 54,000 UK Biobank participants and found 22% more non-coding associations. Narrowing to the 1% of variants the Atlas rates most impactful, he identified 19 regions linked to body mass index. At the Broad Institute, Laura Covill’s team used the score to surface a variant in the DNM1 gene that creates an incorrect splice site. Experimental screens confirmed it.
DNM1 is strongly linked to epileptic encephalopathy. So a London model, run on British public data by a British fellow, produced a result an American institute then checked in the lab. The licensing is the part worth watching. The Atlas is free for non-commercial use now, and DeepMind says commercial access is coming to Google Cloud, which already sells specialist science models.
No pricing has been published. European law treats those two tracks differently. The AI Act does not apply to models developed and put into service for the sole purpose of scientific research and development, and a paper in npj Digital Medicine argued in January that the test is hard to apply. Its authors note that goals change over time, that original intent is difficult to prove, and that academic and commercial work now routinely overlap inside the same project.
DeepMind has run this sequence before. TNW wrote when AlphaFold arrived that the protein mapping AI was not free, and the genome version is thirty times the size. Also tagged with
