Google DeepMind has announced AlphaGenome 2, a new artificial intelligence model that predicts how changes in DNA affect gene activity in specific human cell types. The announcement came on November 11, 2026, and the model is aimed squarely at researchers trying to understand how genetic variants contribute to disease.
The human genome is the same in nearly every cell of the body, yet a skin cell, a brain cell and an immune cell behave very differently. That difference comes down to which genes are switched on or off in each cell type — and a DNA variant that is harmless in one cell type can change gene activity in another. AlphaGenome 2 is built to predict exactly that: given a DNA sequence and a specific cell type or cell state, it estimates how a variant would change gene activity there.
Key Takeaways
- The news: Google DeepMind announced AlphaGenome 2 on November 11, 2026 — a next-generation DNA sequence AI model.
- What it does: It predicts how DNA variants change gene activity in specific human cell types and cell states, one condition at a time.
- The training data: More than 100 million individual cell profiles (107M+) across 68 tissues, from human and mouse.
- The benchmark result: On the TenK10K benchmark it identified 30% of putatively causal expression-linked variants at 90% precision, versus 16% for AlphaGenome 1 — nearly twice as many.
- Availability: Free for non-commercial research now via the AlphaGenome API; commercial research is in private preview.
- The limit: Research use only — it is not validated or approved for clinical or diagnostic use.
A Model That Reads DNA in Context
AlphaGenome 2 is a conditional model. Instead of giving one general prediction for a DNA variant, it takes a condition — a tissue, a cell type or a cell state — and predicts the effect of the variant under that condition, one at a time. That matters because most of the DNA variants linked to disease in genome-wide studies sit outside protein-coding genes, in regions that regulate when and where genes are switched on. Their effects are often visible only in the cell type where that regulation is active.
The model was trained on more than 100 million individual cell profiles across 68 tissues, covering both human and mouse data. That scale of single-cell training data is what lets it make predictions at the level of individual cell types rather than broad tissues. It can also adapt to a researcher’s own single-cell experiment through a cell encoder, which lets scientists tune the model to the cell types they are studying rather than relying only on the conditions it was trained on.
This is a separate, new model — not an update to the precomputed AlphaGenome Atlas. Its predecessor, AlphaGenome 1, was published in Nature in January 2026 and predicted molecular activity across broad tissues. AlphaGenome 2 moves that prediction down to the level of specific cell types and states, which is where many disease-linked variants are expected to act.
What the Numbers Show
DeepMind tested AlphaGenome 2 on TenK10K, a single-cell map of 28 primary immune cell types, and the results show a clear step up from the previous model. At 90% precision, AlphaGenome 2 identified 30% of putatively causal expression-linked variants — the variants believed to actually drive changes in gene expression — compared with 16% for AlphaGenome 1 fine-tuned on single-cell data. In other words, it found nearly twice as many variants that change gene expression, at the same high precision.
The model also did better at predicting the direction of a variant’s effect on a gene. It correctly called the direction of effect for 43% of genes, versus 23% for the fine-tuned AlphaGenome 1. Getting the direction right — whether a variant turns a gene’s activity up or down — is important for researchers trying to work out how a variant might contribute to a disease, and it is a harder test than simply flagging that a variant does something.
These are benchmark results on one dataset, and they describe how well the model performs at prediction tasks rather than how well it understands disease itself. But the gap over the previous model is large, and it comes on the exact task — variant effects in specific cell types — that the new model was designed for.
Who Can Use It — and How
AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API, free of charge. Commercial research access is in private preview, so companies will need to wait for wider availability or apply for the preview programme. DeepMind also said AlphaGenome 2 Skills are coming to Google Antigravity, its agentic development platform, which would let researchers build the model into their own analysis workflows.
The free research access is a deliberate choice. Single-cell datasets and the computing power to train on them are out of reach for many academic labs, so an API that any researcher can query lowers the barrier to studying variant effects in specific cell types. It also means the model is likely to be tested quickly across many different diseases and cell types by the wider research community, which is how its real strengths and weaknesses will show up.
What It Can and Can’t Do
AlphaGenome 2 can predict how a DNA variant is likely to change gene activity in a given cell type, and it can do that for cell types and states a researcher supplies. That makes it a tool for prioritising variants — narrowing down the long list of candidates from a genetic study to the ones most likely to matter in a particular cell type, so lab experiments can focus on the strongest leads.
What it cannot do is diagnose disease or guide treatment. The model is for research use only: it has not been validated or approved for clinical or diagnostic use, and its predictions are not medical advice. A predicted effect on gene activity in one cell type is a hypothesis to test in the lab, not a finding about a person’s health.
There is another caveat worth stating plainly. The technical paper describing AlphaGenome 2 is a preprint — it has not yet been peer-reviewed. The benchmark numbers come from DeepMind’s own testing, and independent researchers will need time to verify them and to find where the model falls short. That is normal for a new research model, but it means the results should be read as promising early evidence rather than settled science.
For now, AlphaGenome 2 looks like a practical step forward for genetics research: a model that reads DNA in the context of the cell it acts in, available free to researchers, with honest limits about what it can and cannot do.
Related Guides
- Claude Fable 5.2 and Mythos 5.2 Release
- Apple Siri Agentic AI Agent Ternus 2027
- Nvidia Jensen Huang Shareholder Lawsuit Spy Allegations
Frequently Asked Questions
Q: What is AlphaGenome 2?
A: AlphaGenome 2 is a next-generation DNA sequence AI model announced by Google DeepMind on November 11, 2026. It predicts how DNA variants change gene activity in specific human cell types and cell states.
Q: How is AlphaGenome 2 different from AlphaGenome 1?
A: AlphaGenome 1, published in Nature in January 2026, predicted molecular activity across broad tissues. AlphaGenome 2 is a separate new model that predicts variant effects in specific cell types and states, and it identified nearly twice as many expression-changing variants on the TenK10K benchmark.
Q: What did AlphaGenome 2 score on the TenK10K benchmark?
A: On the TenK10K single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of putatively causal expression-linked variants at 90% precision, versus 16% for AlphaGenome 1 fine-tuned on single-cell data. It also called the direction of gene effects for 43% of genes, versus 23%.
Q: Is AlphaGenome 2 free to use?
A: Yes, for non-commercial research it is available now for free via the AlphaGenome API. Commercial research access is in private preview, and AlphaGenome 2 Skills are coming to Google Antigravity.
Q: Can AlphaGenome 2 be used to diagnose disease?
A: No. AlphaGenome 2 is for research use only and is not validated or approved for clinical or diagnostic use. Its technical paper is also a preprint, not yet peer-reviewed.
Sources: Google DeepMind blog (Nov 11, 2026); Unite.ai (Nov 11, 2026); Nature (Nov 11, 2026).
