Google DeepMind has released AlphaGenome 2, a DNA model that predicts how a change in someone’s genome plays out in a specific type of human cell, rather than in a tissue average. The company announced the model on Tuesday, 11 November.
The model takes a stretch of DNA about a million letters long, together with a chosen tissue, cell type or cell state, and predicts how that DNA behaves in that condition: which genes are active, and several measures of how the DNA is marked and packaged, including chromatin accessibility, histone marks and DNA methylation.
That shift matters because every cell in the body carries the same DNA, yet a heart muscle cell reads it very differently from an immune cell. The first AlphaGenome, previewed in June 2025 and published in Nature in January 2026, produced fixed parallel outputs for bulk tissues and cell lines. AlphaGenome 2 is a conditional model: it predicts for one tissue, cell type or cell state at a time.
DeepMind trained the model on more than 107 million individual cell profiles from human and mouse bulk and single-cell data, covering 68 tissues. It can also adapt to a scientist’s own single-cell experiment, with as few as a few dozen cells, without retraining.
On the TenK10K benchmark, a large single-cell map of 28 primary immune cell types, DeepMind says AlphaGenome 2 identified 30 per cent of causal variants, against 16 per cent for the original AlphaGenome fine-tuned on single-cell data, and called the direction of a variant’s effect on its gene correctly 43 per cent of the time, against 23 per cent. The company says the model finds nearly twice as many gene-changing variants as its predecessor on one test.
The launch included an independent validation that DeepMind highlighted: researchers at the Francis Crick Institute replicated a laboratory result on inflammatory bowel disease that had taken a year and a half to produce, in 30 minutes using AlphaGenome 2. Sarah Teichmann of the University of Cambridge, who works with the Human Cell Atlas, said the model adapts directly to new single-cell data without retraining.
AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API, with commercial research in private preview. DeepMind has also precomputed scores for trait-linked variants across hundreds of human cell and tissue types, covering seven million common variants linked to traits from height to type 2 diabetes.
The model is for research use only. DeepMind says it has not been validated or approved for clinical or diagnostic purposes, and the same applies to the original AlphaGenome. The company is working with the Wellcome Sanger Institute to profile more cell types.
There are real limits to the advance. The model still misses 70 per cent of causal variants on the TenK10K test, and its predictions need considerably more compute than the original, which DeepMind’s genomics lead Ziga Avsec said makes a genome-wide atlas with this version impractical. The original AlphaGenome will stay in service.
The release follows DeepMind’s AlphaGenome Atlas, launched on 8 September 2026, which holds predictions for all nine billion possible single-letter changes in the human genome. Genomics Limited, led by Peter Donnelly, is already folding AlphaGenome 2 into its Mystra platform. For researchers trying to trace a disease back to the cells where it begins, the new model offers a far sharper lens than anything before it.
