AlphaGenome 2 refines mutation reading cell by cell
DeepMind launches AlphaGenome 2, a model that predicts the effect of DNA variants according to cell type and state. Goal: better link genetics, disease, and therapeutic avenues, without leaving the research framework.
Google DeepMind unveiled AlphaGenome 2 on November 11, 2026, an AI model designed to predict how DNA variations influence specific cells, rather than just tissues on average. This is significant for biomedical research, as the same mutation may have no effect in one cell but be decisive in another. The model is available for non-commercial research, with private preview access for certain commercial uses.
A model that looks at the cell, not just the tissue
AlphaGenome 2 succeeds AlphaGenome, launched last year, but shifts focus. Whereas genomic models at the tissue scale smooth signals, this one conditions each prediction on a given cell type or cell state. DeepMind aims to better identify variants that alter gene expression, RNA splicing, chromatin accessibility, and, compared to the first model, CpG DNA methylation.
Practically, the tool relies on over 100 million individual cell profiles across 68 tissues. This base allows transforming static cell atlases into a prediction engine, according to DeepMind’s formula. A mutation can thus be evaluated in an inflammatory monocyte, an activated T cell, or another state, with different results. It’s less spectacular than a robot breaking the world record for hugs, but far more useful for understanding disease.
What DeepMind says it has improved
The difference is not only in data volume. AlphaGenome 2 handles one condition at a time, allowing it to represent the same DNA segment according to cell identity. The model can also receive single-cell RNA sequencing data from a researcher’s experiment, then infer unmeasured properties like chromatin accessibility. In other words, it can fill experimental gaps and help decide which experiments to run next.
DeepMind also claims the system detects nearly twice as many variants capable of modifying gene expression as its previous reference model. On TenK10K, an immune dataset covering 28 cell types, the technical paper states AlphaGenome 2 recovered 30% of expression-linked variants presumed causal at 90% precision, versus 16% for AlphaGenome 1 refined on single-cell data. The model also explains 34% of cell-specific expression variance within a tissue, compared to 23% for its refined predecessor [to be verified according to the comparative protocol].
Examples in inflammation and Covid-19
DeepMind researchers provide two use cases to demonstrate the gained granularity. In an inflammatory bowel disease analysis, AlphaGenome 2 highlighted rs2836882 among a credible set of seven variants, linked it to the ETS2 gene about 290 kilobases away, and localized the effect to intestinal monocytes and macrophages. The mechanism had already been observed experimentally.
Second example: from single-cell data of a study on an mRNA vaccine against Covid-19, new to the model, it distinguished activated T cells from resting T cells. In this activated state, it predicted a stronger effect for a variant associated with severe Covid-19. The interest is clear: biology is not always read in close-up, but often at the right moment.
A powerful tool, but not yet a verdict
DeepMind remains cautious. The model is still costly to run, as each condition requires a separate prediction. It therefore does not replace AlphaGenome 1, which remains better suited for some large-scale tasks, notably the Atlas published in September 2026 covering all possible single-letter substitutions in human DNA. AlphaGenome 2 is not a direct extension of this precomputed atlas, but a separate, finer, and more computationally demanding version.
The paper also highlights its limits. Even at high precision, the model recovers only a minority of variants that statistical analyses consider probably causal. It may underestimate effects outside accessible chromatin, on very distant regulatory elements, or via post-transcriptional mechanisms. Transient, stimulated, or disease-related states are also underrepresented, as are some neuronal subpopulations.
Most importantly, condition representations come from observational profiles in a stable state. They capture correlations, not causalities. DeepMind therefore specifies that AlphaGenome 2 is neither validated nor authorized for clinical or diagnostic use. For biologists, the interest lies elsewhere: narrowing the hypothesis field before returning to the lab bench.
Key points
- November 11, 2026: launch of AlphaGenome 2 by Google DeepMind.
- Over 100 million cell profiles were used for training.
- 68 tissues cover the training database.
- TenK10K: 28 cell types for the immune benchmark.
- AlphaGenome 2 is accessible for non-commercial research via API.
- Google DeepMind also reserves it for a private commercial preview.
By the numbers
- 100 million+ — individual cell profiles used for training, according to DeepMind, 2026.
- 68 tissues — scope of training data, according to the November 11, 2026 announcement.
- 30% — expression-linked variants recovered at 90% precision on TenK10K.
- 290 kilobases — estimated distance between rs2836882 and ETS2 in the intestinal example.
- 6.8 million — common or trait-associated variants precomputed in the launch resources.