Google DeepMind launches AlphaGenome Atlas for 9 billion mutations

This release builds directly on AlphaGenome, the genome analysis AI Google DeepMind published in June last year. Rather than asking researchers to run the AI themselves on every mutation of interest, the company precomputed results for all roughly 9 billion possible single-letter changes and made them searchable. The scale is enormous: at 1 petabyte, the database is more than 30 times larger than AlphaFold Database, which holds protein structure predictions. The decision to include non-coding regions is notable because, while the human genome has about 3 billion base pairs, only about 2% of it codes for proteins. The remaining 98% regulates gene activity and is largely unexplained. By covering this territory, AlphaGenome Atlas aims to make causal variant hunting in these regions practical. Early validation by the Broad Institute—finding a DNM1 variant tied to epileptic encephalopathy—and Exeter University's 22% improvement in detecting non-coding associations suggest genuine utility, though both results come from the developers' own cited examples. The real test will be whether independent research groups, working on diverse conditions with different data, can consistently translate these predicted impacts into clinically meaningful findings. The database is made freely available to academics from September 8, with commercial access through Google Cloud to follow, indicating a strategy to establish the tool as a standard reference in genomics research while monetizing broader enterprise use.