DeepMind

AI · Biotech · Genomics

Launch

DeepMind's AlphaGenome 2 predicts DNA effects across hundreds of cell types

A single genetic variant can behave differently in every cell type, and this model finally accounts for that variation.

• Google DeepMind launched AlphaGenome 2 on November 11, 2026, a DNA sequence model that predicts how genetic variants affect gene activity across hundreds of specific human and mouse cell types.
• It succeeds AlphaGenome, released in 2025, which predicted variant effects across broad tissues but couldn't pinpoint which specific cell types were affected.
• The model uses a 'conditional' architecture that takes a specific cell type and state as input, rather than predicting all tissues in parallel like its predecessor.
• Trained on more than 100 million individual cell profiles across 68 tissues, AlphaGenome 2 nearly doubled its detection rate of causal variants, from about 16% to 30%.
• In a validation test, the model reproduced in 30 minutes a finding about an inflammatory bowel disease gene that took Francis Crick Institute researchers a year and a half of lab experiments to confirm.
• AlphaGenome 2 can also analyze new single-cell experimental data it has never seen before, such as distinguishing vaccine-activated immune T cells from resting ones without retraining.
• The model is available now via the AlphaGenome API for non-commercial research, with a private preview for commercial use; genomics firm Genomics.com plans to integrate it into its Mystra drug-discovery platform.
• Since 98% of the human genome is non-coding and historically hard to interpret, cell-type-specific predictions could help researchers pinpoint disease-driving variants and validate drug targets far faster than lab experiments allow.