Summary
Google DeepMind says AlphaGenome 2 predicts how DNA variants affect gene activity in specific cell types and states. The research model uses single-cell data and reported higher performance on an immune-cell benchmark than earlier approaches.
Google DeepMind has announced AlphaGenome 2, an AI model designed to predict how DNA sequences and genetic variants affect activity in particular human cell types and states. The company says the model was trained on more than 107 million individual cell profiles across 68 tissues and, on one immune-cell benchmark, identified nearly twice as many variants with relevant effects as earlier approaches.
The model is available through the AlphaGenome API for non-commercial research; commercial research access is in private preview. Google DeepMind says it is intended for research and has not been validated or approved for clinical or diagnostic use.
From tissue averages to individual cell types
DNA is largely shared across the body, but different cells use different parts of it. A genetic variant can therefore have a measurable effect in one cell type while making little difference in another. Regulatory elements such as enhancers can influence genes from a distance, making it difficult to determine which gene a variant affects and where that effect occurs.
The first AlphaGenome predicted properties across many tissues and experiments in parallel from a DNA sequence. AlphaGenome 2 changes that approach: it takes a target tissue, cell type or cell state as an input alongside a DNA sequence of up to one million letters. The target condition can come from a reference catalogue or be represented from a researcher’s own single-cell experiment using the model’s cell encoder.
Google DeepMind says this design helps the model learn how the same DNA sequence is read differently in different cells. Its predicted readouts include gene expression, histone modifications and chromatin accessibility, as well as DNA methylation. It can also use one measured property—for example, gene activity—to predict another that was not measured, such as chromatin accessibility. Such predictions could help researchers decide which experiments to pursue.
Reported results and research uses
On the TenK10K dataset, a single-cell map covering 28 primary immune cell types, Google DeepMind reports that AlphaGenome 2 identified causal variants at a rate of 30%, compared with 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools. For directional gene effects—whether a variant is predicted to raise or lower a gene’s activity—the reported figures were 43% and 23%, respectively.
The company also describes tests on cell types and states not seen during training. In one example, the model used data from a COVID-19 mRNA vaccine study to distinguish activated from resting immune T cells. It predicted that a variant associated with severe COVID-19 had a stronger effect on gene activity in activated cells, an effect that could be missed when looking only at resting cells.
Another example concerns inflammatory bowel disease (IBD). Applying the model to gut immune cells, Google DeepMind says it computationally reproduced a finding in which a variant in a distant regulatory element boosts the ETS2 gene in macrophages, while having little effect on other cell types and genes.
These examples show how cell-specific predictions could help connect genetic associations to candidate genes and the cells in which they act. The results are computational predictions for research: the model can help prioritise hypotheses, but its intended use is not clinical diagnosis or treatment.