Skip to content- Sources: primary
- Summary: AlphaGenome 2 takes a million-letter DNA sequence plus a target condition, either a cataloged cell type or a state encoded from a user's single-cell RNA-seq sample, and predicts expression, chromatin accessibility, histone marks and DNA methylation. It was trained on more than 107 million cell profiles, and on the TenK10K immune dataset it recovered 30% of causal variants against 16% for a fine-tuned AlphaGenome. It is available through the AlphaGenome API for non-commercial research and in private preview for commercial use.
- Why it matters: Drug discovery and genetics teams can score variants in disease-relevant cell types through an API instead of tissue averages.
- Background: Genome-wide association studies flag regions of DNA linked to a trait, but the causal variant, its target gene and the cell type where it acts are often unclear. The first AlphaGenome predicted many bulk tissue readouts in parallel. The new model conditions on one cell state at a time, which also lets it impute unmeasured readouts from a single measured one.