Google DeepMind announced AlphaGenome 2 on November 11, 2026, giving genetics researchers a way to predict how a DNA change may affect a particular cell type or state—including cells represented by their own experimental data. The practical aim is to help choose which genetic changes and cells deserve further investigation.
Why the same DNA can have different effects
Most cells in a person share essentially the same DNA, but use different parts of it. Gene expression is a measure of how actively a gene is being used. Pooling measurements from a tissue can blur effects confined to one cell population, as DeepMind explains in its announcement.
AlphaGenome 2 pairs a DNA sequence with a chosen cell type or state, then predicts molecular effects in that setting. Researchers can select a reference cell type or supply their own human single-cell gene-activity measurements without retraining the model on that dataset. The getting-started guide describes both routes.
The original AlphaGenome already produced outputs for tissues and cell types, but from a fixed catalogue. The new workflow makes the biological setting an input. Comparing cells requires a prediction for each condition, rather than one universal score.
A bowel-disease example shows why cell type matters
DeepMind’s ETS2 tutorial revisits a DNA region linked to inflammatory bowel disease. Across seven candidate changes and seven intestinal immune-cell types, one candidate stands out: its protective version is predicted to reduce activity of the ETS2 gene in macrophages and monocytes, two kinds of immune cell, with much smaller effects in the other illustrated cells.
That gives researchers a focused relationship to investigate: a DNA change, a target gene and particular cells. It is not a newly discovered cause of bowel disease. The tutorial revisits a mechanism investigated experimentally in a 2024 Nature study; agreement with that earlier work illustrates prioritisation, not a replacement for experimental evidence.
What the benchmark does—and does not—show
In their release manuscript, DeepMind’s researchers report that AlphaGenome 2 recovered 30% of benchmark-positive associations between DNA variants and gene activity at 90% precision, versus 16% for the original model fine-tuned on single-cell data. The TenK10K evaluation covered 28 blood immune-cell types.
Here, precision describes how often the associations flagged by the model matched the benchmark’s positive labels; recall describes how many of those positives it recovered.
An illustrative calculation using those rounded results: out of 100 benchmark-positive associations, the newer model would recover about 30 rather than 16—a gain of 14. About 70 would still be missed at that threshold.
These are developer-reported results against statistically selected associations and matched controls, not a patient-screening trial. The 90% precision figure is not 90% diagnostic accuracy, and a low predicted effect does not establish that a variant is harmless.
Check the evidence behind each prediction
AlphaGenome 2 can also estimate biological measurements that were not collected. The documentation distinguishes three evidence labels for individual outputs:
Observed: the particular measurement for that biological condition was included in training. The new prediction itself was not observed in a lab.
Supported imputation: an unmeasured output inferred from informative measurements along a route that has been benchmarked. It remains a hypothesis.
Exploratory imputation: an output without adequate experimental support for that inference; performance is unbenchmarked and can be unreliable.
One practical example is splicing—the way cells cut and join RNA. Single-cell gene-activity measurements alone do not validate predictions of how that processing changes. A plausible-looking output is not enough reason to treat a proposed mechanism as established.
The FAQ also warns that cell states far from the training data may not be captured reliably. Supplying a rare or disease-related state does not itself establish that the model understands it.
How researchers can use it now
The current documentation offers free non-commercial API access, with query limits that vary with demand. At the November 11 launch, commercial research access was a private preview; the announcement did not specify commercial prices or a general-availability date.
A sensible starting point is to choose a relevant reference cell type, compare a small number of biologically justified alternatives, and check each output’s evidence label before choosing a follow-up experiment. For custom cells, the input tutorial requires raw gene-expression counts and warns that the encoder does not explicitly correct batch effects—differences introduced by how samples were processed.
For researchers who instead need precomputed variant scores without running new predictions, the official guide lists AlphaGenome Atlas separately. RohitAI’s guide to AlphaGenome Atlas explains that lookup-based approach.
DeepMind says AlphaGenome 2 has not been validated or approved for clinical or diagnostic use, and the API documentation prohibits clinical decision-making. Its immediate role is to sharpen a research question: which DNA change, which gene and which cell should an experiment investigate?