DeepMind's AlphaGenome Atlas Maps All 9 Billion Human DNA Variants
The dataset covers every possible single-letter DNA change in the human genome, including the non-coding majority that conventional tools handle poorly. Whether it produces discoveries beyond its training data depends on how researchers apply it.
Google DeepMind released AlphaGenome Atlas on September 8, 2026, a platform containing AI-predicted molecular effects for all 9 billion possible single-nucleotide variants in the human genome. The dataset is 1 petabyte, which DeepMind states is more than 30 times larger than the AlphaFold Database. The platform's AlphaGenome Variant Impact (AVI) score combines predictions from AlphaGenome and AlphaMissense into a single pathogenicity ranking across both coding and non-coding genomic regions; DeepMind states it provides "best-in-class performance across many variant pathogenicity and rare disease benchmarks." The tool focuses substantially on non-coding DNA, which comprises the vast majority of the human genome and regulates when and where protein-coding genes are expressed. Much non-coding DNA appears to be remnants of viruses and other molecular parasites with no known function.
External collaborators used the AVI score to identify a previously overlooked variant in the DNM1 gene linked to epileptic encephalopathy, with experimental screens validating the prediction. Applying Atlas to UK Biobank whole-genome data from over 54,000 participants found 22% more non-coding genetic associations than conventional methods; by focusing on the 1% of non-coding variants Atlas rated most impactful, researchers identified 19 genetic regions. Academic access is free via a web portal and API, with commercial access through Google Cloud planned. Ars Technica noted that until biologists begin using Atlas heavily, it will not be clear what the tool offers beyond what could have been obtained from its training data alone.
Ars Technica reported September 9 that the practical value of AlphaGenome Atlas, Google DeepMind's September 8 AI database predicting effects for all 9 billion possible human single-nucleotide variants, will not be clear until biologists begin using it heavily.
The outlet added that non-coding DNA, which the tool focuses on, includes stretches with no known function that appear to be remnants of viruses and other molecular parasites.
Google DeepMind launched AlphaGenome Atlas on September 8, a database of AI-predicted molecular effects for all 9 billion possible single-nucleotide variants in the human genome.
The 1-petabyte dataset, more than 30 times larger than the AlphaFold Database, introduces an AVI score combining AlphaGenome and AlphaMissense to rank variants by pathogenicity across coding and non-coding regions. Collaborators used the AVI score to identify a previously overlooked DNM1 gene variant linked to epileptic encephalopathy, confirmed experimentally, and Atlas found 22% more non-coding genetic associations in UK Biobank data from over 54,000 participants.
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free