U.S.-based Google DeepMind announced on November 11 (local time) an AI model, “AlphaGenome 2,” that predicts how DNA sequences and their mutations affect specific types of human cells. It aims to help elucidate cellular-level mechanisms that cause disease. For noncommercial research purposes, it is available through the “AlphaGenome API.” For commercial research, it will be offered as a limited preview.
The first-generation “AlphaGenome,” announced last June, was a model capable of reading DNA sequences of up to one million base pairs and predicting gene activity at the tissue level and elsewhere. Human cells fundamentally possess the same DNA, but the genes that are active differ by cell type, including immune cells and brain cells. AlphaGenome 2 allows users to specify the tissue or cell type they want to predict and the state of the cells as “conditions,” then makes predictions in accordance with those conditions. Training used data from 68 human and mouse tissues, along with data from more than 100 million cells. The model can also now predict DNA methylation—the chemical markings on DNA involved in switching genes on and off.
Overview of AlphaGenome 2. In addition to the DNA sequence, tissue and cell types are entered as “conditions,” allowing gene activity and other properties to be predicted for each set of conditions (Image: from Google DeepMind’s paper)
Another feature is that if researchers input data from several dozen cells obtained in their own experiments, the model can tailor its predictions to the state of those cells without being retrained. According to DeepMind, this produces performance equal to or better than that of models trained directly on such data. As an example, DeepMind says the model distinguished between T cells activated by a COVID-19 vaccine and resting T cells, and showed that mutations associated with severe disease disrupt gene activity more strongly in activated T cells.
In terms of performance, AlphaGenome 2 can detect about twice as many mutations that alter gene activity within cells as the previously highest-performing model. In an evaluation covering 28 types of immune cells, it accurately identified 30 percent of mutations that cause changes in gene activity, surpassing the original AlphaGenome further trained on cell-level data (16 percent). It also reproduced the results of an existing study that identified a mutation causing excessive activity of a specific gene in intestinal immune cells in inflammatory bowel disease (IBD).
Data containing precomputed prediction results across hundreds of cell and tissue types will also be provided for mutations associated with genetic predispositions and diseases such as type 2 diabetes. This will allow researchers to use the results without conducting large-scale computations. UK-based Genomics PLC plans to integrate AlphaGenome 2 into its “Mystra” platform. It also plans to provide an AlphaGenome 2 skill for Google’s development environment, “Google Antigravity.”
AlphaGenome 2 is intended for research purposes only and has not undergone validation or approval for use in clinical practice or diagnosis. The paper identifies as challenges the fact that only some causal mutations can still be detected with high accuracy and that data remain limited for stimulated cells and cells in diseased states.
