Google’s AlphaGenome 2 predicts how DNA changes affect specific cells

Google DeepMind released AlphaGenome 2 on November 11, an AI model that predicts how small DNA differences affect gene activity in specific cell types. The release, covered in The Rundown, could help researchers trace disease risk to the cells involved and choose which experiments to run first.

DeepMind’s Adam Gayoso called it “tying human genetics to cell biology” in an interview with The Verge. Access is available through the AlphaGenome API for noncommercial research, with commercial research in private preview.

How the model adds detail

AlphaGenome 2 takes a selected cell type or state as an input, so that context shapes its predictions. This gives scientists a finer view of how a DNA change may affect different cells, or the same type of cell under different conditions.

In a test across 28 primary immune cell types from the TenK10K dataset, DeepMind says AlphaGenome 2 identified 30% of DNA variants that alter gene activity. The earlier model, after further training on data from individual cells, identified 16%.

Labs can also supply gene activity measurements from a few dozen cells without retraining the model, according to DeepMind. In the company’s vaccine example, AlphaGenome 2 predicted a stronger effect from a DNA variant in activated T cells than in resting T cells.

DeepMind also says the model matched an experimental finding on inflammatory bowel disease from a study published in June 2024. It picked out a DNA change that increases activity of ETS2, a gene involved in inflammation, in gut immune cells. The earlier research traced the mechanism to macrophages, a type of immune cell.

Why it matters

Scientists have found plenty of DNA differences tied to diseases such as type 2 diabetes. To turn those links into a treatment, researchers need to work out what the changes do inside cells. As Genomics CEO Peter Donnelly explains, drug teams need to identify the gene involved, the relevant cells and whether raising or lowering the gene’s activity could help.

AlphaGenome 2 could help labs decide which experiments to run first. A prediction linking a DNA variant to a gene in a particular cell type could guide which stretch of DNA researchers alter, which cells they test and which gene’s activity they measure. They would still need to connect the predicted effect to disease risk and confirm it in the lab.

Knowing which cells matter could also help researchers design a more selective treatment. In the 2024 ETS2 study, drugs called MEK inhibitors reduced inflammation in human macrophages and cultured samples from people with inflammatory bowel disease. The researchers warned about effects in other tissues and proposed directing treatment specifically to macrophages as a possible way to reduce toxicity.

The added detail also brings a tradeoff. DeepMind’s Žiga Avsec told The Verge that predictions cost more to compute. Researchers have to decide when a question calls for that extra detail.

Donnelly, whose company plans to evaluate AlphaGenome 2, says systematic tests across diseases are still needed to establish what it adds beyond existing genetic evidence. Recovering the ETS2 finding is encouraging, but the model’s value for drug discovery will depend on whether its predictions point researchers toward targets that hold up in new experiments.