Google DeepMind’s AlphaGenome 2 predicts how tiny DNA differences change genes in specific kinds of cells, a tool for tracking down causes of disease
DeepMind says it finds nearly twice as many gene-changing DNA variants as its first version. It is free for noncommercial research and not approved for medical use.
Google DeepMind, Google’s AI research lab, released AlphaGenome 2 on Nov. 11, an AI model that predicts how small differences in a person’s DNA change the way genes work inside specific kinds of cells. DeepMind says the tool can help scientists find the DNA changes and cells behind diseases, a first step toward new drug targets.
Every cell in the body carries the same DNA, but a heart muscle cell and an immune cell switch on different genes. A DNA variant, a spot where one person’s genetic code differs from another’s, may matter in only one kind of cell. Large genetic studies have linked thousands of variants to diseases, but finding which variant does the damage, in which cells, is hard.
What DeepMind says it does
- It identified nearly twice as many variants that change gene activity in immune cells as DeepMind’s original model: 30% versus 16% on a test set covering 28 types of immune cells.
- It reproduced a known finding about inflammatory bowel disease: a variant in a distant genetic switch that overactivates a gene called ETS2 only in certain gut immune cells.
- Scientists can upload gene-activity data from a few dozen of their own cells, and the model makes predictions for those cell states without being retrained. In data from a COVID-19 vaccine study, it showed a variant linked to severe COVID-19 disrupted genes more strongly in activated immune cells.
- It adds predictions for DNA methylation, chemical tags that help turn genes on and off.
AlphaGenome 2 is available free for noncommercial research through DeepMind’s AlphaGenome online service and in private preview for companies. Genomics Ltd., a company whose chief executive, Peter Donnelly, called human genetics one of the most powerful tools in drug discovery, said it plans to add the model to its Mystra platform.
How it works
The first AlphaGenome, released last year, read a stretch of DNA up to 1 million letters long and predicted its effects across broad tissues all at once. The new version predicts for one tissue, cell type or cell state at a time, which DeepMind says lets it see how the same DNA is read differently in closely related cells. It was trained on data from more than 107 million individual human and mouse cells, including maps from the Human Cell Atlas, a global project charting the body’s cell types.
Limits
The performance figures are DeepMind’s own, and its announcement did not cite a peer-reviewed paper. The predictions are meant to guide lab experiments, not replace them. DeepMind says AlphaGenome models are for research only and have not been validated or approved for clinical or diagnostic use.
Sources: Google DeepMind blog post, Nov. 11, 2026; Google DeepMind AlphaGenome page.