Timeline

DeepMind releases AlphaGenome 2, a DNA model that predicts variant effects by cell type

Conditioned on a single cell type, the model adds DNA methylation outputs; DeepMind says it finds nearly twice as many expression-altering variants as the previous state of the art.

  • Ideas & essays
  • Models & capabilities
  • Notable

Google DeepMind released AlphaGenome 2, a successor to its AlphaGenome DNA-sequence model. It predicts how a DNA sequence and its variants affect specific human cell types, tissues and cell states, with the aim of linking genetic variants to gene regulation in the cells where disease arises.

The main change is architectural. The original model predicted many tissues and cell lines in parallel from a one-million-letter sequence. AlphaGenome 2 instead takes a single biological condition as an input alongside the sequence. A cell encoder lets researchers build that condition from their own single-cell data, including cell states absent from training, without retraining. The model can also use one measurement, such as single-cell gene activity, to impute unmeasured ones, such as chromatin accessibility, and adds DNA methylation to its outputs. DeepMind said training used more than 100 million individual cell profiles from human and mouse samples.

DeepMind reported that the model identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model. On TenK10K, a benchmark of 28 primary immune cell types, it said AlphaGenome 2 identified causal variants for 30% of cases against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools, and 43% against 23% for directional effects. It also said the model reproduced a published finding that an inflammatory-bowel-disease-linked variant raises ETS2 expression specifically in gut macrophages. These are the developers’ own evaluations, and DeepMind’s post describes several of its examples as computational predictions rather than experimentally validated results.

Access follows the earlier pattern: the model is available now through the AlphaGenome API for non-commercial research, with a private preview for commercial research. The post does not say whether weights are released. DeepMind states that AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic use. It also released precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions, building on the AlphaGenome Atlas published in September.