Google DeepMind unveiled AlphaGenome 2 on November 11, 2026 – a model that predicts how DNA variants change gene activity in specific human cell types. According to the developers, on the TenK10K immune cell benchmark it identified 30% of causal variants – DNA changes that affect gene activity. The best result among the comparison methods, which included the first-generation AlphaGenome fine-tuned on single-cell data and other specialized tools, was nearly half that – 16%.
Predictions like these help narrow the search for the causes of disease: they point to which DNA change disrupts a particular gene and in which cells it happens. Researchers get specific variants and genes to test in the lab – in the editors’ view, a first step toward finding drug targets.
The key difference in the new version is that it accounts not only for cell type but also for cell state – for example, whether an immune cell is resting or responding to an infection. To handle a new state, the model can use a researcher’s own data without additional training.
At a glance
- AlphaGenome 2 analyzes a DNA sequence of about 1 million base pairs, conditioned on a chosen tissue, cell type or cell state.
- It was trained on human and mouse data, including more than 107 million individual cell profiles.
- The model can predict the effects of DNA variants in new cell states from gene activity data for just a few dozen cells.
- Non-commercial research access is open via the AlphaGenome API; commercial access is in private preview.
What changed in the AlphaGenome 2 architecture
The first AlphaGenome took a DNA sequence and predicted its molecular properties across thousands of experiments from different tissues at once. DeepMind explains that this approach made it hard to capture subtle differences between closely related cell types.
In the second version, the researcher specifies a single condition – a tissue, cell type or cell state. That condition is fed into the model at every stage of DNA processing, and the model makes predictions for one condition at a time. This helps it capture the fact that the same stretch of the genome can affect different cells in different ways.
The new architecture also lets the model predict properties that scientists did not measure directly. For example, from single-cell gene activity data it can estimate chromatin accessibility – how easily the proteins that regulate genes can reach particular stretches of DNA. On top of the model’s existing readouts, DeepMind added prediction of DNA methylation – chemical marks that help regulate gene activity without changing the DNA sequence itself.
AlphaGenome 2 uses single-cell gene activity measurements (scRNA-seq) to make predictions for new cell states without fine-tuning. On data from a COVID-19 mRNA vaccine study, it predicted that a DNA variant linked to severe disease disrupts gene activity more strongly in activated T cells than in resting ones.
How AlphaGenome 2 differs from the first version
| Feature | AlphaGenome | AlphaGenome 2 |
|---|---|---|
| Length of the analyzed DNA sequence | 1 million base pairs | 1 million base pairs |
| How it works | Predicts properties for thousands of experiments across different tissues at once | Conditions each prediction on a specific tissue, cell type or cell state |
| New cell states from a researcher’s data | Not supported | Uses single-cell gene activity data without fine-tuning |
| DNA methylation prediction | No | Yes |
| Access at launch | AlphaGenome API preview for non-commercial research | AlphaGenome API for non-commercial research plus a private commercial preview |
Results on TenK10K
On TenK10K, a dataset covering 28 primary immune cell types, DeepMind compared the new model with the original AlphaGenome fine-tuned on single-cell data and with specialized tools.
| Metric | Best comparison method | AlphaGenome 2 |
|---|---|---|
| Identifying causal DNA variants | 16% | 30% |
| Predicting the direction of the effect on gene activity | 23% | 43% |
On both metrics the result improved by roughly 1.9x (editors’ calculation: 30 / 16 = 1.875; 43 / 23 ≈ 1.87). These figures describe performance on this particular benchmark, not the model’s overall accuracy.
In an inflammatory bowel disease (IBD) case study, the model computationally reproduced a finding from a study published in Nature. It singled out a DNA change that boosts the activity of the ETS2 gene in macrophages – immune cells in the gut. The change sits in a region that regulates ETS2 from a distance. According to the model’s predictions, other cell types and genes are largely unaffected.
DeepMind also precomputed the effects of DNA variants associated with a range of traits and diseases, from height to type 2 diabetes. These variants were found in genome-wide association studies (GWAS), and AlphaGenome 2 scored their impact on gene activity across hundreds of cell and tissue types and states.
Access and limitations
AlphaGenome 2 is available free of charge for non-commercial research via the AlphaGenome API. The Python client installs with pip install alphagenome, and an API key is required. The documentation includes interactive Google Colab notebooks – a service for running Python code in the browser without installing anything locally – among them tutorials on analyzing your own single-cell data. DeepMind plans to release AlphaGenome 2 Skills later for Google Antigravity, its agentic development platform. Until November 30, 2026, the company is accepting submissions of AlphaGenome 2 use cases; selected projects may receive additional quota and credits.
The announcement does not disclose the model’s parameter count, detailed terms or pricing for commercial access, exact rate limits for the free API, or a release date for Skills in Antigravity. The company also has not said whether it plans to release the AlphaGenome 2 weights and code.
What this means for research
AlphaGenome 2 helps scientists decide which DNA changes are worth studying further by predicting how they affect gene activity in different cells. Researchers can feed it their own gene activity measurements without retraining the network. And for DNA changes associated with various traits and diseases, DeepMind has already prepared scores that can be explored without large-scale computation.
The model’s predictions still need to be verified experimentally. It cannot be used in medicine for diagnosis or treatment decisions: DeepMind notes that AlphaGenome models are “intended for research use only and have not been validated or approved for clinical or diagnostic purposes.” For now, the practical value of AlphaGenome 2 is helping scientists choose which experiments to run next.
