Google DeepMind unveiled AlphaGenome 2 on November 11, a DNA prediction model designed to estimate how genetic changes affect particular cell types and states. In the company’s TenK10K benchmark, the model identified 30% of causal variants, compared with 16% for the original AlphaGenome after fine-tuning on single-cell data. The results are from a research benchmark, not a measure of diagnostic accuracy or an estimate of any individual’s disease risk.
For researchers, the model could help connect genetic signals to the cellular settings in which they act, narrowing the field for follow-up experiments. Measurements taken across an entire tissue can show overall changes, but effects may vary between cells. If a variant matters only in a specific cell type, an average across the tissue may not reveal where the effect occurs.
AlphaGenome 2 conditions predictions on cell type and state
The original AlphaGenome generated a fixed set of predictions across multiple tissues and experiments from a DNA sequence of up to about one million letters. AlphaGenome 2 lets researchers specify a target context, such as a tissue, cell type or cell state, and incorporates that information while processing the sequence. DeepMind says this approach is intended to detect cell-specific effects that broader tissue measurements may obscure.
The model was trained on bulk and single-cell data from humans and mice, covering more than 107 million single-cell profiles, the company said. Researchers can also provide single-cell gene-activity data to generate context-specific predictions for newly observed cell states, without retraining the model. It can predict properties that have not been measured directly, giving researchers potential candidates to investigate experimentally.
The TenK10K evaluation covered 28 primary immune cell types. DeepMind reported that AlphaGenome 2 identified 30% of causal variants, compared with 16% for the original model after fine-tuning on single-cell data. The comparison reflects performance on this specific benchmark. It is not clinical validation and does not show that the model can predict an individual’s likelihood of developing a disease.
ETS2 example reproduces a known gut-disease mechanism
DeepMind also presented a case involving intestinal disease. The model computationally reproduced a mechanism described in earlier research: a regulatory variant increases ETS2 activity in intestinal macrophages, a type of immune cell. In the example, AlphaGenome 2 identified the variant and predicted that its effect on ETS2 was concentrated in macrophages, with limited effects on other cell types and genes.
The mechanism was not first discovered by AlphaGenome 2. Work published in 2024 by researchers at Imperial College London, the Francis Crick Institute and University College London had linked a regulatory region active in macrophages to increased ETS2 activity and inflammatory function. The example shows the model reproducing an existing biological finding computationally; it does not establish a new cause of disease.
As of November 11, DeepMind made AlphaGenome 2 available through an API for non-commercial research, while commercial research access remained in private preview. The company says the AlphaGenome models are intended for research and have not been validated or approved for clinical or diagnostic use. For now, researchers can use cell- and state-specific predictions to help select experimental leads, but those predictions require further testing.