DeepMind’s AlphaGenome 2 predicts how the same DNA variant affects different cells

Google DeepMind released AlphaGenome 2 on November 11. The model predicts how a change in DNA affects gene activity in a particular human cell type or state. A research team can compare predictions for the same variant across cell types before choosing where to look for an effect in the lab. DeepMind opened its API for non-commercial research at launch and made commercial research access available in private preview.

The first AlphaGenome, introduced in June 2025, primarily learned from measurements taken across mixtures of cells. A change in a relatively rare cell type can be difficult to pick out when its activity is averaged together with that of other cells in a tissue. The new model can take a cell type or state into account when predicting gene expression, a measure of how actively a cell uses a gene’s instructions. Its training included measurements from individual cells alongside tissue-level data from humans and mice. DeepMind also added predictions for DNA methylation, a chemical marking of DNA associated with gene regulation.

What 30 percent means in the immune-cell test

DeepMind evaluated the model using TenK10K, a study of immune cells in blood. The group behind that dataset paired genome sequences from 1,925 people with gene activity measurements from more than five million cells across 28 immune cell types. That gave DeepMind a way to test predictions against estimates of where variants affect gene activity.

In one test, DeepMind set the bar so that 90 percent of the calls counted as correct against the benchmark’s labels. At that threshold, AlphaGenome 2 recovered 30 percent of the variants the benchmark treated as likely causes of a change in expression. A version of the first model adapted to single-cell data recovered 16 percent. Most candidate variants still went undetected at that threshold. The figures describe performance against this benchmark, which is based on research data, and give no estimate of accuracy in diagnosing a patient.

A second look at a known bowel disease finding

A 2024 study in Nature offers a concrete example of the question the model tries to answer. Researchers from the Francis Crick Institute and other institutions studied a stretch of DNA associated with inflammatory bowel disease. Their experiments linked a change in a regulatory region to greater activity of a gene called ETS2 in human inflammatory macrophages, immune cells that help regulate inflammation in tissue. The work connected a statistical association with a gene and a cell type.

DeepMind revisited that known result with its model. Of seven candidates it examined, only the variant identified in the earlier study showed a substantial predicted effect. The model predicted higher ETS2 activity in macrophages and monocytes from the gut. This was a retrospective check against an experimental finding. For a newly proposed variant, the prediction gives a research team a cell type and an effect to test.

Access for research

The model can also use previously unseen measurements of gene activity from a human cell sample to make predictions for that cell state without another round of model training. The documentation says accuracy can suffer when the cells differ substantially from the training data. It labels some estimates for unmeasured properties as exploratory, so those outputs can be treated as hypotheses.

DeepMind also calculated variant scores in advance for 510 combinations of human tissues and cell types. The data cover variants linked to traits as well as variants that are common in the population. The downloadable release is a filtered set of those predictions. It gives researchers a way to consult existing results without running a fresh prediction for each candidate.

The public API is free for non-commercial use. Request limits vary with demand, and lookups of precomputed results allow higher request rates than fresh predictions, according to the documentation. DeepMind’s release places commercial research access in private preview. The terms for the public API exclude use for clinical decisions.