Google DeepMind released AlphaGenome 2 on November 11, a model that predicts how a DNA sequence affects gene activity in a specific tissue, cell type or cell state. The announcement was authored by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec, and is accompanied by a technical paper.
The original AlphaGenome produced a fixed set of parallel predictions for bulk tissues and cell lines from a 1-million-letter DNA sequence. AlphaGenome 2 is a conditional model: the target condition is fed in at every stage of the architecture, so the same sequence can be read differently depending on the cell type requested. A cell encoder lets scientists turn their own single-cell data into a condition, and the model adds DNA methylation to the readouts the first version already covered, such as gene expression, histone modifications and chromatin accessibility.
DeepMind says the model was trained on bulk and single-cell data from human and mouse, including more than 107 million individual cell profiles. On the TenK10K dataset, which covers 28 primary immune cell types, it identified causal variants at a rate of 30% versus 16% and directional gene effects at 43% versus 23%, compared with the original AlphaGenome fine-tuned on single-cell data and other specialised tools. Overall, DeepMind reports that it identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model.
The company also says AlphaGenome 2 reproduced a published finding that a genetic variant drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages, and that it matched or beat models trained directly on new datasets using only a few dozen cells as input, without retraining.
AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with commercial research access in private preview and Colab tutorials published. DeepMind states the model is for research use only and has not been validated for clinical or diagnostic use.