Google DeepMind announced AlphaGenome 2 on Nov. 11, 2026, a DNA prediction model designed to show how genetic changes act in specific cell types and states. The approach aims to expose disease-linked effects that broad tissue measurements can obscure, giving scientists a more focused starting point for experiments.
The original AlphaGenome produced fixed parallel outputs across tissues and experiments from a million-letter DNA sequence. AlphaGenome 2 instead takes a target condition, such as a tissue, cell type or cell state, alongside the sequence and uses that condition throughout its architecture, according to DeepMind.
Its announcement pairs a reported improvement on an immune-cell benchmark with a computational reproduction of an established bowel-disease mechanism. The examples show how the model could focus laboratory work on particular variants and cells.
That allows predictions to ask how the same stretch of DNA behaves in different cellular settings. DeepMind says this can help identify effects confined to particular cells that broader tissue models miss. A tissue-level result can provide useful context while still leaving the cell responsible for a disease-linked effect unresolved.
The company reports training the model on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles. The profiles supply training material for learning how DNA activity varies across biological settings.
DeepMind also says researchers can supply single-cell gene-activity data to condition predictions on newly observed states without retraining the model. The system can then predict properties that were not measured, helping scientists decide which experiments to prioritize.
On TenK10K, a dataset covering 28 primary immune-cell types, DeepMind reports that AlphaGenome 2 identified 30% of causal variants in the evaluation, compared with 16% for the original AlphaGenome fine-tuned on single-cell data.
These developer-reported benchmark results measure variant identification, not diagnostic accuracy or a person’s disease risk.
For researchers, the practical attraction is a larger set of candidate variants to investigate in the cells where they may act. Connecting a disease-associated genetic signal to a particular variant, target gene and cell type can make the next experiment more focused.
Reproducing an existing disease clue
The bowel-disease example makes the distinction concrete. In its Nov. 11 announcement, DeepMind says AlphaGenome 2 computationally reproduced a finding involving a regulatory variant that increases activity of the ETS2 gene in gut macrophages, a type of immune cell.
The model singled out the variant and predicted its selective effect on ETS2 in those macrophages, with other cell types and genes largely unaffected. The useful clue is the combination of variant, gene and cell type: it points to a specific cellular setting for investigating the effect.
The underlying biology predates this model. In a June 7, 2024 report, Imperial College London described research conducted with the Francis Crick Institute and UCL that linked a macrophage-active regulatory region to increased ETS2 activity and inflammatory functions. The biological mechanism was therefore already established; AlphaGenome 2's contribution in this example is reproducing it computationally.
As of the Nov. 11 announcement, DeepMind offered AlphaGenome 2 through its API for non-commercial research, with commercial research in private preview. The company states that both AlphaGenome models are intended for research and have not been validated or approved for clinical or diagnostic purposes.
The immediate opportunity is to choose better-targeted experiments: investigate a genetic change in the cell type or state where its effect is predicted to appear.
