Google DeepMind has announced AlphaGenome 2, a new version of its DNA model that predicts how small differences in a person's genetic code change the way genes behave inside particular human cell types. Announced on 11 November 2026, the model was trained on more than 100 million individual cell profiles from 68 tissues, and on one large immune-cell test it picked out 30% of causal variants, compared with 16% for the earlier approach.
The finding
Every cell in the body carries the same DNA, but a heart muscle cell and an immune cell switch on different genes. A tiny change in the DNA, called a variant, can matter in one kind of cell and not in another. The first AlphaGenome, released last year, made its predictions for whole tissues, which are averages over many cell types. AlphaGenome 2 works at the level of single cell types and cell states.
DeepMind says this lets researchers spot variants that act only in a small group of cells, effects that a tissue-wide average would hide. It also says the new model can make predictions for cell states it was never trained on.
How it was done
The team changed how the model handles cell types. The original version took a stretch of DNA one million letters long and produced outputs for thousands of experiments at once. AlphaGenome 2 is a "conditional" model: it takes the DNA sequence together with one chosen condition (a tissue, a cell type or a cell state) and predicts the effect for that condition only. The condition is fed in at every stage of the model.
- Data: more than 107 million cell profiles, from both human and mouse, mixing single-cell and bulk tissue data.
- What it predicts: gene expression, histone modifications and chromatin accessibility, as before, plus DNA methylation, which is new.
- Filling gaps: given one kind of measurement, the model can estimate others that were not measured, which DeepMind says can help scientists choose their next experiments.
- New samples: a scientist can upload gene-activity data from a few dozen cells, and the model adapts to that cell state without retraining.
Results
| Test (TenK10K, 28 immune cell types) | AlphaGenome 2 | Earlier model |
|---|---|---|
| Causal variants identified | 30% | 16% |
| Direction of gene effect identified | 43% | 23% |
The comparison is with the original AlphaGenome fine-tuned on single-cell data and other specialised tools. DeepMind gives two worked examples:
- Bowel disease: a recent study linked a variant to inflammatory bowel disease (IBD) through over-activation of the ETS2 gene in gut immune cells called macrophages. AlphaGenome 2 reproduced this on a computer, singling out the variant and showing that the effect was limited to those cells.
- COVID-19 vaccine data: on single-cell data it had not seen, from a study of an mRNA vaccine, the model told activated T cells apart from resting ones. It also showed that a variant linked to severe COVID-19 has a stronger effect in the activated cells.
DeepMind has also precomputed predictions for variants linked to traits and diseases, from height to type 2 diabetes, across hundreds of cell and tissue conditions, so researchers do not have to run the calculations themselves.
"What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining"Sarah Teichmann, co-founder of the Human Cell Atlas consortium, University of Cambridge
"AlphaGenome 2 represents an impressive advance in the field."Peter Donnelly, CEO, Genomics Limited
Who can use it
AlphaGenome 2 is available now through the AlphaGenome API for non-commercial research, with Colab tutorials, and in private preview for commercial research. DeepMind says AlphaGenome 2 Skills will come soon to Google Antigravity, its agentic development platform. Genomics Limited says it plans to add the model to its Mystra platform.
Limits stated by the authors
DeepMind says both AlphaGenome models are meant only for research. Neither has been validated or approved for diagnosing patients or for any clinical use. It is working with researchers, including at the Wellcome Sanger Institute, to cover a wider range of cell types and states.
Announced by Google DeepMind on its official blog: read the announcement




