Google DeepMind releases AlphaGenome Atlas covering 9 billion single-letter DNA variants
Precomputing a petabyte of predictions turns a model into a lookup table, and shows how scientific AI is shifting from papers to platforms.
Google DeepMind on 8 September 2026 released AlphaGenome Atlas, a resource containing predicted molecular effects for roughly 9 billion single-nucleotide variants, which the company describes as every possible single-letter change in the human genome. The dataset runs to about a petabyte, more than 30 times the size of the AlphaFold Database, and holds thousands of predictions per variant across many cell types and tissues.
The predictions cover gene regulation across hundreds of human and mouse cell types, RNA splicing, gene expression, chromatin accessibility, transcription factor binding and protein impact. A new AlphaGenome Variant Impact score, or AVI, combines the regulatory predictions of the AlphaGenome model with the protein-level predictions of AlphaMissense into a single ranking that applies both to the roughly 2 percent of the genome that codes for proteins and to the 98 percent that does not. Each score comes with feature attributions showing which biological processes drive it, and the Atlas also catalogues more than 2,500 recurrent DNA sequence motifs with their genomic locations.
Access is through a free website at alphagenome.google/atlas for academic and non-commercial use, the AlphaGenome API on GitHub, a Google Antigravity skill, and Google Cloud's Model Garden for commercial use, which DeepMind says is coming soon.
DeepMind cited early results from partners. Researchers at the Broad Institute working with the GREGoR Consortium, including Laura Covill and Anne O'Donnell-Luria, used the predictions to identify a DNM1 variant linked to epileptic encephalopathy. Gareth Hawkes at the University of Exeter reported finding 22 percent more non-coding genetic associations in an analysis of more than 54,000 UK Biobank participants, including variants affecting the genes PLA2G7 and EGLN1. Julia Zeitlinger and Melanie Weilert at the Stowers Institute used the Atlas to categorise transcription factor effects. Other collaborators named include Boston Children's Hospital, Harvard University, Memorial Sloan Kettering Cancer Center, Massachusetts General Hospital and the University of Kansas Medical Center.
Why it matters
Precomputing predictions for every possible variant turns a model into a lookup table, which changes who can use it: a clinician or geneticist no longer needs to run inference, only to search. It also signals a pattern for scientific AI in which the atlas built from a model becomes more useful than the model itself, as happened with protein structures. The focus on the non-coding 98 percent of the genome targets the region where most disease-associated variants sit and where interpretation has been hardest. The commercial route through Google Cloud shows how these resources are being positioned as platforms rather than papers.
One-letter DNA change
Any of roughly 9 billion possible single-nucleotide variants in the human genome
AlphaGenome regulatory predictions
Splicing, expression, chromatin accessibility and transcription factor binding across hundreds of cell types
AlphaMissense protein predictions
Estimated impact on the protein for variants inside coding regions
AVI score with attributions
Single impact ranking for coding and non-coding variants, with the processes that drive it
Served via web, API and cloud
Free portal for academic use, GitHub API, Antigravity skill, Google Cloud for commercial use
What you can learn from this
A single-nucleotide variant is a one-letter change in DNA. The human genome is about three billion base pairs long, and each position could hold one of three alternative letters, which is where a figure of roughly nine billion possible variants comes from. Most of those variants never occur in a living person, but a resource that covers all of them means any newly sequenced variant already has a prediction waiting. That is the practical difference between a model and an atlas. The trade-off is storage, hence a petabyte of data.
Coding and non-coding variants need different reasoning. A change inside a gene's protein-coding region can alter an amino acid, and tools such as AlphaMissense predict whether that substitution is likely to damage the protein. A change outside coding regions cannot be assessed that way; it may instead alter when and where a gene is switched on by changing a regulatory sequence. Models such as AlphaGenome predict those regulatory effects from the surrounding DNA context. Combining the two into one score gives a comparable number for either type.
Sequence-to-function models predict experimental readouts from DNA. Rather than being trained on disease labels, models of this type learn to predict measurable properties such as chromatin accessibility, splicing patterns or gene expression directly from a stretch of sequence. Because those readouts vary by cell type, predictions are produced per tissue. A variant's effect is then estimated by comparing the prediction for the reference sequence with the prediction for the altered sequence. This is why the Atlas reports effects across hundreds of cell types rather than a single yes-or-no label.
Feature attribution makes a score inspectable. A single number is hard to trust in a clinical or research setting, so the Atlas pairs each AVI score with attributions indicating which predicted processes contributed most. Attribution methods estimate how much each input feature moved the output, giving a researcher a lead to test in the laboratory. It does not prove causation, but it turns an opaque prediction into a hypothesis. Score plus explanation is becoming the standard shape for scientific AI tools.
Precomputation shifts cost from inference to storage and serving. Running a large model over billions of inputs once and storing the outputs means users pay nothing in compute at query time. The cost moves to keeping a petabyte online and indexed for fast lookup by genomic coordinate. That design favours organisations with cloud infrastructure and explains the multiple access routes: a browser for individuals, an API for pipelines, and a cloud listing for commercial users.
We teach this
How to use this in practice
Look up a known variant in the Atlas. Visit alphagenome.google/atlas, use the site's search to find a well-documented variant, for example one recorded in a public clinical database, and read its AVI score and attributions. Write down which predicted process dominates and which cell types show the largest effect, then compare that with whatever the public database says about the variant's known clinical significance. Done looks like a paragraph summarising one variant's predicted mechanism in plain language, with a note on whether the prediction and the database agree.
Sketch the pipeline from sequence to score. Draw a flow: reference DNA window, altered DNA window, model predictions for each, difference calculation, protein-impact prediction, combination into AVI, attribution. Note at each step what the input and the output are, and mark the step at which a cell type is chosen, since that is where tissue-specific differences enter. Done looks like a one-page diagram you could use to explain to a non-specialist why the score differs between tissues.
Explore the API with a small script. Clone the AlphaGenome repository from GitHub, follow its setup instructions including any access key it requires, and write a short script that requests predictions for one variant and prints the few outputs with the largest effect. Keep it under 50 lines, and add a comment above each call explaining what it requests and what comes back. Done looks like a script that runs end to end and prints structured output you can read without the documentation open.
Compare a coding and a non-coding example. Pick one missense variant and one intronic or intergenic variant, retrieve both from the Atlas, and compare how their scores are explained. Note which relies on protein-impact features and which on regulatory features, and whether the non-coding one shows a cell-type pattern that would be invisible to a protein-only tool. Done looks like a two-column table contrasting the two explanations.
Sources
- AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome — Google DeepMind
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report