AlphaGenome Atlas: DeepMind's 9-Billion-Variant Genome Map
DeepMind precomputed the molecular effects of all ~9 billion single-letter DNA substitutions — a 1-petabyte atlas free for non-commercial research. What it does well, and where researchers say it stops.
From Coding Job to Lookup
On September 8, 2026, Google DeepMind released AlphaGenome Atlas: a platform containing precomputed effect predictions for roughly 9 billion single-nucleotide variants — every possible single-letter change — in the human reference genome. The dataset weighs in at 1 petabyte, more than 30 times the size of the AlphaFold Database, and is available free for non-commercial research through a web portal, the AlphaGenome API, and a skill in Google Antigravity, with commercial access via Google Cloud planned.
The Atlas is built on AlphaGenome, the regulatory-prediction model DeepMind shipped in 2025. According to the company, about 9,000 researchers had queried that model through a programming interface — a path that required writing code and could be slow, which kept many biologists out. The Atlas removes that wall: a researcher can now look up a variant on a web page and read a single number instead of running the model themselves.
What You Actually Get
Four linked resources ship with the Atlas. First, molecular effect predictions: thousands of predictions per variant across gene regulation, spanning hundreds of human and mouse cell types and tissues. Second, the AlphaGenome Variant Impact (AVI) score, which combines AlphaGenome with AlphaMissense — DeepMind's model for protein-altering variants — into one number that works for both coding regions (about 2 percent of the genome) and the non-coding 98 percent. Third, AVI feature attributions that decompose each score into interpretable contributions such as splicing or gene expression. Fourth, a compendium of more than 2,500 recurrent DNA motifs — the "words" of the genome — and their locations.
Early Results from Real Research
The strongest evidence comes from external collaborators. At the Broad Institute, researchers used the AVI score to prioritize variants overlooked in previous work on an unsolved rare disease, identifying a variant affecting the gene DNM1, strongly linked to epileptic encephalopathy — and the underlying predictions showed exactly how it created an incorrect splice site, later validated by experimental screens. At the University of Exeter, Gareth Hawkes applied the Atlas to whole-genome data from over 54,000 UK Biobank participants, uncovering 22 percent more non-coding genetic associations and pinpointing regulatory variants driving proteins including PLA2G7 and EGLN1. At the Stowers Institute, researchers used the motif resource to categorize transcription factors by whether they only open DNA or also activate genes — work they said was experimentally impractical before.
The Limits, Stated Plainly
DeepMind's own genomicists are candid about boundaries. The Atlas predicts; it does not replace the experiment that confirms, and it cannot account for an individual patient's particulars. Performance is stronger for promoter and splicing variants than for enhancers. Regulatory elements acting from far away can fall outside the million-letter window the model reads. Ziga Avsec, the team lead, says the predictions are accurate enough to point downstream research in the right direction but "should not be taken as universal truth." Ben Lehner of the Wellcome Sanger Institute put it more bluntly: this is not an "AlphaFold moment," and clinical decisions should not rest on these models alone.
Verdict
As a research instrument, the Atlas hands a small lab the reach that used to require a large one's compute budget — a searchable, genome-wide dictionary of variant effects, free for non-commercial use. As a clinical tool, it is explicitly a reference layer, not a diagnostic. For anyone working on rare disease or non-coding variation, it is now one of the first places to look; for everyone else, it is a preview of how precomputed foundation-model output is becoming standard research infrastructure.
Sources: Google DeepMind Blog | 澎湃新闻