AlphaGenome Atlas: precomputed effects for ~9 billion DNA letter changes
DeepMind’s follow-on to AlphaGenome: a 1 PB catalogue of single-nucleotide variant molecular predictions, plus an AVI ranking score that blends regulatory and missense signals for research use.
Primary source: Google DeepMind blog (8 Sep 2026): https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/
What’s new: After the Nature AlphaGenome model paper, Google DeepMind released AlphaGenome Atlas (announced 8 Sep 2026): a platform that precomputes AlphaGenome’s molecular-effect predictions for about 9 billion possible single-nucleotide changes in the human genome (every single-letter substitution), described as a ~1 petabyte resource. The point is scale and access: instead of running the model one variant at a time, researchers get a queryable catalogue plus a web portal, API access, and related tooling.
The Atlas also introduces the AlphaGenome Variant Impact (AVI) score: a single ranking number that combines AlphaGenome’s regulatory predictions with AlphaMissense coding impact, aimed at both coding (~2%) and non-coding (~98%) regions. DeepMind reports AVI feature attributions (which molecular processes drive the score), thousands of per-variant molecular tracks across cell types/tissues, and a motif catalogue. Collaborator vignettes in the announcement include rare-disease variant prioritization (e.g. a DNM1 splicing story with experimental follow-up) and UK Biobank analyses that surface more non-coding associations when variants are grouped by predicted molecular effect.
Why it matters: The model brief answered “can one DNA network read regulatory sequence?” The Atlas answers “can the community use that at genome scale without a TPU farm?” That is the AlphaFold-Database pattern applied to regulation: predictions as infrastructure. For medical genetics and rare disease, a ranked, attributable variant map is closer to a daily research tool than a demo video.
This post is distinct from our earlier AlphaGenome Nature methods note: here the story is the precomputed map and AVI, not the architecture itself.
Caveats: Research / non-commercial portal framing in the announcement; not a clinical diagnostic. Precomputed scores inherit AlphaGenome/AlphaMissense errors and training biases. “Best-in-class” pathogenicity claims in a company blog need the accompanying paper and independent benchmarks before you retire other scorers. Collaborator case studies are promising anecdotes, not a prospective diagnostic trial. Treat Atlas as a hypothesis prioritizer for experimental validation.