Google DeepMind released AlphaGenome 2 on Nov. 11, a DNA sequence model that predicts how genetic variants affect gene regulation in specific human cell types and cell states — the difference between knowing that a variant matters and knowing where in the body it matters.

The first AlphaGenome scored variants against the genome as a whole. The second version bridges single-cell biology and human genetics: it predicts effects across hundreds of human cell types and, according to DeepMind, can generalise to cell and disease states it never saw in training. DeepMind says the model recovers about 30% of causal single-cell eQTLs — the statistical links between variants and gene expression measured cell by cell — compared with roughly 16% for the earlier approach.

The significance is practical as much as technical. Most disease-associated variants found by genome-wide studies sit in non-coding regions, where their effect depends on which cell is reading the DNA at the time. A model that assigns a variant to the right cell type, and to a state such as activation or stress, gives researchers a sharper hypothesis to test in the lab — and a way to rank drug targets that may only matter in one tissue.

Nature covered the release, noting the model's ability to predict the cell types where a variant is likely to act. DeepMind researcher Pushmeet Kohli highlighted the design that allows prediction across hundreds of cell types and generalisation to unseen states.

As always with genomics models, validation is the hard part: prediction is cheap, experimental confirmation is not, and the field has learned to be careful about benchmarks built from the same data used to train the model. AlphaGenome 2's value will be settled by how many of its cell-type-specific predictions survive a wet lab.