Google DeepMind · 2026-09-08 · major
AlphaGenome Atlas — a prediction for every possible DNA letter change
AlphaGenome Atlas is a 1-petabyte dataset from Google DeepMind holding predicted molecular effects for all 9 billion single-letter DNA changes in the human genome. Free to use for non-commercial research from today.
AlphaGenome's predictions for every single-letter DNA change, precomputed so researchers can look one up instead of running the model.
Key specs
| Variants | 9 billion |
|---|---|
| Dataset size | 1 petabyte |
Quick facts
| Maker | Google DeepMind |
|---|---|
| Coverage | Hundreds of human and mouse cell types and tissues |
| Access | Free website portal, non-commercial use |
| Commercial use | Google Cloud, soon |
| Also included | Over 2,500 recurrent DNA sequence motifs |
| Vs AlphaFold Database | More than 30× larger |
What is it?
AlphaGenome Atlas is a precomputed lookup table for the human genome. Google DeepMind ran its AlphaGenome model across all 9 billion possible single-letter DNA changes and published the results as a 1-petabyte dataset — more than 30 times the size of the AlphaFold Database. Each variant carries thousands of predicted molecular effects across several aspects of gene regulation.
How does it work?
There is no new model here: the Atlas is AlphaGenome's output frozen in place, which is what turns an expensive inference job into a lookup. Predictions span hundreds of human and mouse cell types and tissues, and Google DeepMind pairs them with a collection of over 2,500 recurrent DNA sequence motifs — the 'words' of the genome — and their locations. Researchers reach the data through the website portal, the AlphaGenome API on GitHub, or a skill inside Google Antigravity.
Why does it matter?
Working out what a non-coding variant actually does is the slow step in rare-disease and population genetics, and precomputation takes the compute cost out of that step. Google DeepMind reports three results from early users: grouping rare variants by their predicted molecular effects across more than 54,000 UK Biobank participants surfaced 22% more non-coding genetic associations that statistical noise had hidden; focusing on the 1% of non-coding variants the Atlas rates most impactful pointed at 19 genetic regions linked to body mass index; and prioritising variants with the AVI score found a DNM1 variant strongly linked to epileptic encephalopathy, where the predictions showed it creates an incorrect splice site.
Who is it for?
genomics, rare-disease and population-genetics researchers
Frequently asked questions
- Is AlphaGenome Atlas free to use?
- AlphaGenome Atlas is free for non-commercial use through Google DeepMind's website portal from launch day, with commercial access coming to Google Cloud later. The underlying AlphaGenome base model is separate: it is already available for academic use on GitHub and through the AlphaGenome API, and for commercial use on Google Cloud via Model Garden.
- What kinds of DNA changes does AlphaGenome Atlas cover?
- AlphaGenome Atlas covers single-nucleotide variants — the 9 billion single-letter substitutions possible in the human genome. For each one it stores thousands of predicted molecular effects across several aspects of gene regulation, spanning hundreds of human and mouse cell types and tissues. Google DeepMind also ships a collection of over 2,500 recurrent DNA sequence motifs alongside the variant predictions.
- How does AlphaGenome Atlas compare to the AlphaFold Database?
- AlphaGenome Atlas is more than 30 times larger than the AlphaFold Database at 1 petabyte. The two do different jobs: the AlphaFold Database grew protein 3D structure coverage from around 190,000 experimental structures to more than 200 million predictions in 2022, while the Atlas predicts what single-letter DNA changes do to gene regulation.
- Do I need to run the AlphaGenome model to use the Atlas?
- No — the point of AlphaGenome Atlas is that Google DeepMind already ran the model and stored the answers, so a researcher looks a variant up instead of paying for inference. Three routes reach the same data: the website portal, the AlphaGenome API on GitHub, and a skill inside Google Antigravity for agent-driven workflows.
Try it
Browse alphagenome.google/atlas, or install the API from github.com/google-deepmind/alphagenome