By Google DeepMindLearning Objectives
- Understand what a variant-effect model predicts and why non-coding DNA is the hard part
- Understand that the Atlas is a map of predictions rather than a record of measurements
- Evaluate the access terms before planning research or a product around it
What Is AlphaGenome?
AlphaGenome is a variant-effect model from Google DeepMind. Given a change to a single letter of human DNA, it predicts the downstream regulatory consequences, such as whether a gene is read more or less often and how the surrounding machinery responds. That matters because most of the human genome does not code for proteins, and most variants linked to disease fall in those non-coding stretches, where the effect of a change is far harder to reason about than it is inside a gene.
AlphaGenome Atlas, released on September 8, 2026, is the model run exhaustively rather than on demand. DeepMind scored all nine billion possible single-nucleotide variants in the human genome and published the results as a one-petabyte dataset, which it describes as more than 30 times the size of the AlphaFold Database.
⚠️Warning
The Atlas is a map of predictions, not a record of measurements. Every one of those nine billion entries is a model output, not an experimental result. Its value is in narrowing where laboratory work should look, and DeepMind's own terms state the predictions are for theoretical modelling and research only and must not be used for clinical decision-making or relied on as medical advice. Reading the Atlas as though nine billion variants have been characterised would overstate it substantially.
💡Key Concept
Coding versus non-coding: a variant inside a gene can change the protein it builds, which is comparatively tractable to predict. A variant outside one changes how and when genes are switched on. Most disease-associated variants sit in that second category, which is the gap AlphaGenome targets.
📝Note
Access is split by purpose, and it is not open source. The software in the repository is Apache 2.0, but the API is offered free for non-commercial use only, and DeepMind's terms state that outputs should not be used for training other machine learning models. Commercial use runs through Google Cloud instead. Check which side of that line your work falls on before you build.
🎯Tip
Explore AlphaGenome: the Atlas portal is at alphagenome.google/atlas and the API code is on GitHub.
Pricing
Free for non-commercial research through the web portal and the API. Commercial use is a separate arrangement through Google Cloud rather than a published price.
- Browse predictions for any variant
- Non-commercial research only
- No account-level commercial rights
- Programmatic access, academic use
- Outputs may not train other models
- Apache 2.0 code, restricted outputs
- Commercial deployment path
- Standard cloud terms
- Contact Google for terms
Core Features
Exhaustive Variant Scoring
The Atlas covers all nine billion possible single-nucleotide changes rather than only variants someone has already observed.
Non-Coding Regulatory Prediction
Targets the regulatory stretches outside genes, where most disease-associated variants sit and interpretation is hardest.
Programmatic Access
A public API and open-source client code let a research group query predictions at scale instead of through the browser.
Published Collaborator Findings
Launch partners reported concrete results, which is the evidence that the predictions narrow real searches.
Strengths
- Exhaustive coverage — every possible single-letter change, not a curated subset
- Aimed at the hard case — non-coding regulatory effects
- Free for research — no cost barrier for academic work
- Open client code — the repository is Apache 2.0
- Named collaborators — the Broad Institute, Boston Children's Hospital, Memorial Sloan Kettering, Mass General's Center for Genomic Medicine, Harvard, Exeter, Stowers and the University of Kansas Medical Center
Limitations and Considerations
- Predictions, not measurements — laboratory validation still decides
- Explicitly not for clinical use — DeepMind's own terms rule that out
- Outputs may not train other models — a real constraint on downstream research
- Commercial use is separate — through Google Cloud, with no published price
- One petabyte is not a casual download — serious use needs cloud-side querying
Best Use Cases
| Use Case | Why AlphaGenome Matters | Caveat |
|---|---|---|
| Prioritising rare-disease variants | Scores variants no one has studied | Predictions need laboratory follow-up |
| Interpreting non-coding associations | Built for the regulatory genome | Not valid for clinical decisions |
| Large-scale cohort analysis | API access across many variants | Outputs cannot train other models |
| Teaching genomics and AI | Free portal, concrete published findings | Commercial work needs Google Cloud |
Key Takeaways
- AlphaGenome is Google DeepMind's variant-effect model, predicting how a single-letter DNA change affects gene regulation
- Its Atlas release scores all nine billion possible single-nucleotide variants as a one-petabyte dataset, described as more than 30 times the size of the AlphaFold Database
- Every entry is a prediction rather than a measurement, and DeepMind's terms exclude clinical decision-making explicitly
- Access is free for non-commercial research, outputs may not be used to train other models, and commercial use runs through Google Cloud
- Reported collaborator findings include 22 percent more non-coding genetic associations across more than 54,000 UK Biobank participants and a DNM1 variant tied to epileptic encephalopathy
