Google AlphaGenome 2 zooms into cellular roots of disease with finer predictions

A closer look at cell differences

AlphaGenome 2 succeeds DeepMind’s AlphaGenome, released last year. Although most cells carry the same DNA, different cell types activate different genes, and those patterns can shift with a cell’s state. The same genetic variant may therefore affect cells in different ways.

The new model predicts variant effects for individual cell types and states. Google says this finer resolution may reveal disease-related effects in particular cells that broader models could miss.

Training across tissues

DeepMind trained AlphaGenome 2 using more than 100 million individual cell profiles from 68 tissues. The company says it found nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model. The model can also predict outcomes for new cell types and biological states without retraining, and flag gaps in experimental data that researchers might study. Google DeepMind research scientist Adam Gayoso described the advance as “tying human genetics to cell biology.”

Potential value for treatment research

The model could help scientists examine genetic variants found in large populations and investigate their effects at the cellular level, Gayoso told The Verge. That may help connect genetic changes with disease.

Google DeepMind genomics lead Žiga Avsec said cell-level detail matters for drug development because treatments target particular proteins and cell types. He said new technologies for measuring individual cells and their states have made the large training datasets possible.

Costs and future directions

DeepMind plans to improve its models by gathering more data, including for cell states that are poorly represented. Avsec said future versions might draw on data from multiple species, and that such models could eventually contribute to “virtual cells,” digital simulations of biological systems.

AlphaGenome 2 is not replacing its predecessor soon. Avsec said its more detailed predictions require more computing resources and cost more to produce, making it impractical for building the genome-wide Atlas covering every possible DNA variant.