AlphaGenome 2: DeepMind AI reads DNA effects cell type by cell type
By Anika SharmaAI & Innovations Editor, Peepals Tech
Representative image: chaiyananuwatmongkolchai / Pixabay
Google DeepMind released AlphaGenome 2 on 11 November 2026, an AI model that predicts how DNA changes act in specific cell types, free for non-commercial research.
What Happened
Google DeepMind, the artificial intelligence research lab owned by Google, released AlphaGenome 2 on Wednesday, 11 November 2026. It is a new AI model that predicts how a small change in a person's DNA alters gene activity inside one specific type of cell, such as a particular immune cell, rather than across a whole tissue. The company says the model is free for non-commercial research through its AlphaGenome API and is in a private preview for commercial research.
Key Facts
- AlphaGenome 2 follows the first AlphaGenome model, which DeepMind released in 2025 and described in a paper in the journal Nature in 2026.
- DeepMind says the new model was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles from 68 tissues.
- On a single-cell immune benchmark called TenK10K, which covers 28 cell types, DeepMind reports that AlphaGenome 2 found 30% of likely causal variants at 90% precision, against 16% for the first model fine-tuned on single-cell data.
- On predicting the direction of a variant's effect (whether it raises or lowers a gene's activity), the reported score rose from 23% to 43%.
- Researchers can feed in their own single-cell RNA data to describe a cell condition without retraining the model; the documentation allows up to 256 cells per request.
- Unlike the first version, whose code and weights were opened for non-commercial use in January 2026, the new model's weights have not been released.
- DeepMind says AlphaGenome 2 is a research tool and has not been validated or approved for clinical or diagnostic use.
Why It Matters
Every cell in the body carries the same DNA, yet a heart cell, a brain cell and an immune cell behave very differently. The difference lies in which genes each cell switches on and off. A variant is a single changed letter in the DNA code. Many variants linked to diseases sit not inside genes but in the stretches of DNA that control them, so the same change can matter in one cell type and do nothing in another. Finding which variant actually drives a disease, and in which cells, is one of the slowest steps in genetics research.
The first AlphaGenome predicted variant effects for bulk tissues and cell lines, which blur together the many cell types inside an organ. AlphaGenome 2 moves to cell-type resolution, which is closer to how diseases such as autoimmune conditions actually work. If its predictions hold up, scientists could narrow a long list of suspect DNA changes to a short list worth testing in the laboratory, saving time and money.
The release also shows how large AI labs are turning towards biology. DeepMind says the first model's API was used by more than 3,000 scientists and handled about one million requests a day by January 2026, when it opened the model's code. Later reports put the number of researchers who had used the API at about 9,000. Keeping the new model's weights closed while offering free API access for non-commercial work gives DeepMind control over how it is used and leaves a door open for paid commercial use.
There are limits. The benchmark results come from DeepMind itself, and no independent group has yet repeated them. One report notes that the new model needs considerably more computing power than the first, which makes running it across every possible variant in the genome impractical for now.
Peepals calculation: On the TenK10K benchmark, the share of likely causal variants found rose from 16% to 30%, a gain of 14 percentage points, or about 1.9 times as many (30 ÷ 16 ≈ 1.88). For the direction of effect, the score rose from 23% to 43%, a gain of 20 points, again about 1.9 times (43 ÷ 23 ≈ 1.87). Both measures nearly doubled. Our calculation from the reported figures.
| Measure (TenK10K benchmark) | First model | AlphaGenome 2 |
|---|---|---|
| Likely causal variants found (at 90% precision) | 16% | 30% |
| Direction of effect correct | 23% | 43% |
| Weights released | Yes, non-commercial | No |
| Access | API and code | API, commercial preview |
Impact
Short-term: Academic genetics groups, including those in India working on population-scale genome projects, can test AlphaGenome 2 through the free API for non-commercial research. Companies have to apply for the private preview.
Long-term: If the cell-level predictions prove reliable in independent tests, the model could speed up the search for disease-causing variants and for drug targets. Because the weights are closed, researchers will depend on DeepMind's servers and terms of use.
Who is affected: Geneticists, computational biologists, drug developers and research hospitals that study the genetic roots of disease. Patients are not directly affected yet, since the model is not approved for clinical use.
Key Takeaway
AlphaGenome 2 lets researchers ask how a single DNA change behaves in a specific cell type, and DeepMind reports it nearly doubles the first model's benchmark scores, though the results still await independent checks.
Questions and Answers
What is AlphaGenome 2?
AlphaGenome 2 is an AI model from Google DeepMind, released on 11 November 2026, that predicts how changes in DNA affect gene activity in specific cell types, to help scientists find which genetic variants drive disease.
How is AlphaGenome 2 different from the first AlphaGenome?
The first model made predictions for whole tissues and cell lines. The second works at the level of individual cell types, was trained on more than 107 million single-cell profiles, and, according to DeepMind, finds nearly twice as many likely causal variants on an immune-cell benchmark.
Can anyone use AlphaGenome 2?
Researchers can use it free for non-commercial work through the AlphaGenome API. Commercial research is offered only in a private preview, and the model's weights have not been released.
Can doctors use AlphaGenome 2 to diagnose patients?
No. DeepMind says the model is a research tool that has not been validated or approved for clinical or diagnostic use.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
By Anika Sharma, AI & Innovations EditorCovers AI & Innovation and more for Peepals Tech. All stories by Anika Sharma
Get the day's stories in your inbox
Peepals Weekly: the week decoded, free. One email a week, unsubscribe in one click.
Anika Sharma is a pen name of the Peepals Global Editorial Team, which researches, fact-checks and publishes every story. The portrait is an illustration.
