DeepMind unveils AlphaGenome 2, which predicts DNA mutation effects in specific cell types

On November 11, 2026, Google DeepMind introduced AlphaGenome 2, the successor to last year's AlphaGenome model. The new version is designed to predict how changes in a DNA sequence affect gene activity in specific types of human cells, helping researchers trace diseases back to the cells where they originate.

What AlphaGenome 2 changes

The first AlphaGenome predicted many tissues in parallel. AlphaGenome 2 is a conditional model: it receives a DNA sequence and a target biological state, meaning a tissue, a cell type or a cell state, and returns a prediction for that single combination. The state can be picked from a reference catalog or built from a researcher's own single-cell experiment using a cell encoder.

The model predicts gene expression, histone modifications and chromatin accessibility, among other readouts, and adds DNA methylation as a new one. It can also fill in unmeasured data types from measured ones, for example inferring chromatin accessibility from single-cell gene activity. It was trained on human and mouse samples, both bulk and single-cell.

DeepMind's results

DeepMind says the model finds nearly twice as many expression-altering variants in cells as the previous best model. On TenK10K data covering 28 primary immune cell types, it identified 30 percent of causal variants, while the comparison model reached 16 percent. It got the direction of the effect on a gene right in 43 percent of cases versus 23 percent. The benchmark was the original AlphaGenome fine-tuned on single-cell data, along with other specialized tools.

The company also describes case studies. The model computationally reproduced the result of a recent landmark study: it singled out a causal variant in a distant regulatory switch that, in inflammatory bowel disease, boosts activity of the ETS2 gene selectively in gut macrophages, leaving other cell types and genes largely unaffected. It also showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in T cells activated by a vaccine.

DNA sequence prediction models are an important tool in helping bridge that gap, and AlphaGenome 2 represents an impressive advance in the field. - Peter Donnelly, CEO of Genomics Limited, which plans to incorporate the model into its Mystra platform, quoted on the DeepMind blog

New cell states without retraining

According to DeepMind, the model adapts to previously unseen cell states, such as disease states, using RNA-seq from a few dozen cells and no retraining. The results are said to match or outperform models trained on those specific datasets. The company is also releasing precomputed predictions for trait-associated variants across hundreds of cellular conditions, which it says should make it easier to map enhancer-gene links and prioritize variants from GWAS (genome-wide association studies).

Availability and limits

AlphaGenome 2 is available through the AlphaGenome API for non-commercial research, and a limited private preview runs for commercial research. Colab tutorials for expression conditioning are already available, and an AlphaGenome 2 Skills package is due to arrive in Google Antigravity. Community submissions of use cases are open until November 30, with the possibility of additional quota and credits.

The caveats are clear: the results are for research only and have not been validated or approved for clinical or diagnostic applications. The figures come from DeepMind's own announcement and reflect comparisons with selected reference models, not an independent evaluation.

For Polish genomics labs and biotech companies, the key point is that the model is available through the API for non-commercial research. Commercial use is for now limited to the private preview.