DeepMind unveils AlphaGenome 2 for cell-specific DNA predictions

Google DeepMind announced AlphaGenome 2 on November 11, a research model designed to predict how DNA variants affect particular cell types and biological states. Unlike its predecessor, which produced predictions across many tissues and experiments at once, the new system takes a target condition alongside a DNA sequence and evaluates that setting specifically, the company said.

DeepMind reported that the model identified 30% of causal variants in the TenK10K benchmark, which covers 28 primary immune-cell types. The company compared that result with 16% for its earlier AlphaGenome model fine-tuned on single-cell data. The figures are company-reported benchmark results, not measures of diagnostic accuracy or a person’s disease risk.

The model was trained on more than 107 million individual cell profiles from human and mouse samples, according to DeepMind. Researchers can also provide gene-activity data from a newly observed cell state, which the company says lets AlphaGenome 2 generate predictions without retraining. It can infer other properties from supplied measurements, including how DNA is packaged and whether it is accessible. Those predictions could help researchers decide...