Google DeepMind on Tuesday unveiled AlphaGenome Atlas, a catalogue of predicted molecular effects for about 9 billion single-letter DNA changes across the human genome, Nature reported. The company also published a same-day blog describing the atlas as a 1-petabyte resource built on its earlier AlphaGenome model.
According to Nature and DeepMind, the atlas precomputes AlphaGenome predictions for every possible single-nucleotide substitution, so researchers do not have to score each candidate through the API one by one. DeepMind says roughly 9,000 researchers already used the model via API after its earlier release; the atlas is meant to remove that friction with a free, non-commercial website portal, plus API access and a skill hook in Google Antigravity. Commercial Cloud access is described as coming later.
A centerpiece is the AlphaGenome Variant Impact score, or AVI, which collapses AlphaGenome and AlphaMissense signals into one ranking number for both coding and non-coding regions. Nature notes that collaborators used AVI to prioritize rare-disease candidates, including a DNM1 splice-disrupting variant tied to epileptic encephalopathy that experimental screens later backed. DeepMind also cites UK Biobank work that surfaced more non-coding associations for circulating proteins and body-mass index when analysts focused on high-AVI variants.
The atlas ships with feature attributions that point to mechanisms such as splicing, chromatin accessibility, and gene expression, plus a library of more than 2,500 recurrent DNA motifs. DeepMind and Nature both stress the resource is for research, not clinical diagnosis. The model paper behind AlphaGenome already ran in Nature earlier; Tuesday’s drop is the genome-wide precomputed map, not another model launch.
This is the AlphaFold-database move for regulatory DNA: ship the bulk predictions so wet-lab and clinical genetics teams can query first and experiment second.