Google DeepMind Ships 1PB AlphaGenome Atlas of 9B DNA Variants
TL;DR
- Atlas is 30x larger than the AlphaFold Database; an 80x computation speedup via model distillation and GPU kernel optimization made it feasible.
- The AVI score placed known causal variants in the top 50 candidates 29.5% of the time, versus 12.5% for existing methods, a documented performance gap.
- Beta testers at the Broad Institute used AVI scores to identify a DNM1 splice-site variant as a probable cause of severe epileptic encephalopathy.
Google DeepMind pre-computed the regulatory impact of every possible single-letter change to the human genome. All 9 billion of them. In a post on the Google blog, the team calls the result AlphaGenome Atlas: "a database that predicts the effects of every possible single nucleotide variant in the human genome." The finished file runs to 1 petabyte.
The reason it exists is that only about 2% of the genome codes for protein. The other 98% is regulatory and, until now, thinly annotated. AlphaGenome Atlas assigns each of those 9 billion positions a summary number the team calls the AVI score. "This single, easy-to-use score combines predictions for both coding and non-coding regions," the post says.
Two field trials are cited. At the Broad Institute, Laura Covill's team ran a rare-disease case through the atlas: "The tool highlighted a critical variant in the DNM1 gene, predicting that it created an incorrect splice site." Dr. Gareth Hawkes then applied it to 54,000+ UK Biobank participants and, by grouping variants by their predicted molecular effect, "uncovered 22% more non-coding genetic associations," landing on 19 genetic regions linked to body mass index.
Access is a browser tab. "AlphaGenome Atlas is available today through an intuitive website portal that requires zero coding skills," the announcement says.
The post gives no false-positive rate for AVI, no ancestry breakdown of the UK Biobank cohort, and no license terms for commercial reuse. It follows Google's male fruit fly connectome release from four days earlier, part of a busy run on our Healthcare AI beat.
What others are reporting
-
Google DeepMind Read →
First-party technical post detailing three access channels: free browser portal, GitHub API, and a skill in Google Antigravity. Establishes the commercial-vs-academic licensing split and the coming Google Cloud tier.
AlphaGenome Atlas charts the molecular effects of DNA variants across the genome.
-
Scientific American Read →
Highlights the commercial licensing gap versus AlphaFold's fully open model and frames Atlas as a workflow shift: researchers now look up predictions rather than running computation.
This represents the first time that any researcher in the world can access a comprehensive map of the human genome and its variations by simply opening a browser.
-
Fortune Read →
Reports three beta-testing case studies with quantitative benchmarks, including AVI placing causal variants in top 50 candidates 29.5% of the time versus 12.5% for existing tools.
We bought the book, but we did not understand how to read it.
-
IEEE Spectrum Read →
Details the 80x computation speedup via model distillation and GPU kernel optimization, and flags the 1M base-pair field-of-view cap as a ceiling for long-range enhancer predictions.
Understanding this language of life can unlock so many things.
-
The Register Read →
Places Atlas within the competitive AI product landscape and connects the DNM1 epilepsy variant identification to the AVI score's mechanism, not just its outcome.
By precomputing AlphaGenome's predictions at scale, we have created an easily accessible resource that vastly expands the model's reach.
-
The Next Web Read →
Unique EU AI Act regulatory framing and traces the cross-continental validation chain from UK Biobank discovery at Exeter to Broad Institute experimental confirmation.
The file is more than 30 times the size of the AlphaFold database, which took protein structures from about 190,000 experimental entries to more than 200 million predictions.
-
Unite.AI Read →
Frames Atlas explicitly as a 'first-generation' resource with acknowledged training data gaps, while reporting UK Biobank found 22% more non-coding associations than previously detectable.
The most comprehensive catalogue of how genetic mutations affect molecular biology.
Shared on Bluesky by 1 AI expert
Originally reported by blog.google
Read the original article →Original headline: Google DeepMind Ships AlphaGenome Atlas With Predictions for All 9B Single-Letter DNA Variants