A genetic variant linked to a disease rarely acts everywhere in the body at once; it usually matters in one kind of cell. Google DeepMind's AlphaGenome 2 is built around that fact. Where the first AlphaGenome predicted how DNA changes affect gene activity at the level of broad tissues, the new version makes predictions for a specific cell type or state, which is closer to the question geneticists actually need answered.
The design change is the main story. The original model read about a million letters of DNA and produced outputs for thousands of experiments at once, which DeepMind says spread its capacity thin. AlphaGenome 2 is conditional: it predicts for one condition at a time, and that condition, a tissue, cell type or state, is fed into every stage of the network. A condition can come from a reference catalogue or be built from a researcher's own single-cell data, and DeepMind says a snapshot of just a few dozen cells is enough to describe a new cell state without retraining. Training drew on more than 100 million individual cell profiles across 68 tissues, and the model adds DNA methylation to the readouts it predicts.
The reported gains are large. On the TenK10K benchmark, which covers 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants against 16% for the best comparison, and got the direction of gene effects right in 43% of cases against 23%. DeepMind's examples include reproducing a published finding that a variant linked to inflammatory bowel disease raises ETS2 activity in gut macrophages while leaving other cells largely untouched, and showing that a variant tied to severe COVID-19 has a stronger effect in vaccine-activated T cells than in resting ones.
For researchers, access is the practical point. Non-commercial users can call it now through the AlphaGenome API with Colab tutorials, commercial research is in private preview, and DeepMind plans AlphaGenome 2 skills in Google Antigravity alongside Gemini 4. Genomics Limited says it will build it into its Mystra platform, and the Wellcome Sanger Institute is collaborating on profiling more cell types.
The limits are stated plainly. The model is for research only and has not been validated for clinical or diagnostic use. Direction-of-effect predictions are described as reliable only when the model is confident, and the cell-type specificity improvement is given without a figure. Weights, licence terms and pricing are not mentioned, and the benchmark results come from DeepMind's own announcement rather than an independent evaluation.