AlphaGenome 2 succeeds AlphaGenome, released last year. Nearly every cell in the body carries the same DNA, but different cell types switch different genes on and off, so the same variant can affect a heart muscle cell, a skin cell or an immune cell differently. The new model is built to predict those differences by cell type and state.
According to DeepMind, the model was trained on more than 100 million single-cell profiles across 68 tissues, and it identifies nearly twice as many variants that change gene expression as its previous state-of-the-art model. The researchers found it can also make predictions for new cell types and biological states without retraining, and can point to gaps in experimental data.
DeepMind's genomics lead Žiga Avsec told The Verge that this specificity matters in drug development, which targets particular proteins and cell types. He said AlphaGenome 2's predictions are more expensive to compute, so it does not replace its predecessor, and that producing the genome-wide Atlas DeepMind released in September with it would have been impractical.
Commercial terms and broader availability have not been announced. DeepMind says it is generating more data to improve its models and that future versions could be trained on data across multiple species.