Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects

Why it matters

Google DeepMind has released AlphaGenome 2, a conditional DNA sequence-to-function model that predicts how genetic variants affect molecular activity in a chosen cell type rather than averaging over bulk tissue. The model reads a roughly one-megabase DNA window along with a target cell state, using a convolutional encoder, transformer blocks, and a decoder, with the cell-state embedding modulating activations throughout. Unlike its predecessor, which predicted a fixed catalog of tracks in parallel, AlphaGenome 2 handles one condition per call and can adapt to single-cell datasets it never saw during training, without fine-tuning. It outperformed fine-tuned AlphaGenome 1, Borzoi, and Scooby on single-cell eQTL and enhancer-gene benchmarks, including CRISPRi validation data and a Parkinson's disease cohort. Access is currently through a hosted API for non-commercial research, with weights promised after final publication. DeepMind also published precomputed atlases covering enhancer-gene links and variant scores across many cell conditions, and commercial access is in private preview.