DeepMind releases AlphaGenome 2 to predict how DNA variants act inside specific cells

Google DeepMind on Wednesday released AlphaGenome 2, a new artificial-intelligence model designed to predict how changes in human DNA affect gene activity inside specific types of cells, rather than across broad tissue averages. The company said in a blog post announcing the model that it is intended to help researchers pinpoint which genetic variants drive disease and in which cells they act, a step the company frames as central to turning genome-wide association studies into usable leads for drug discovery.

The model is a successor to the original AlphaGenome, which DeepMind introduced in 2025 and formally described in a peer-reviewed paper in Nature. That version predicted molecular activity across many tissues at once. AlphaGenome 2 instead takes a single biological condition, such as a tissue, a cell type or a disease state, as an input alongside the DNA sequence, and produces predictions tailored to that condition. DeepMind said the new model was trained on more than 107 million individual cell profiles spanning 68 tissues in humans and mice, and that it still accepts DNA sequences of up to one million letters at a time, the same scale as its predecessor.

Finer resolution, fewer missed variants

According to DeepMind, the practical gain is accuracy at a finer grain. The company said AlphaGenome 2 identifies nearly twice as many expression-altering genetic variants as the previous state-of-the-art approach. On a public dataset called TenK10K, which covers 28 types of immune cells, DeepMind said the model correctly flagged the causal variant in 30 percent of test cases, compared with 16 percent for a version of the original AlphaGenome that had been separately fine-tuned on single-cell data, and that it identified the correct direction of a gene's response 43 percent of the time, against 23 percent previously.

The model also adds a fourth prediction category, DNA methylation, to the gene-expression, chromatin-accessibility and histone-modification outputs the original AlphaGenome produced, and it can use one type of experimental measurement to infer others it was not directly given. DeepMind said a notable new feature is that the model can take a small sample of new single-cell data, on the order of a few dozen cells, and generalize its variant predictions to that previously unseen cell state without being retrained, a capability the company used in its validation tests.

Testing against known disease biology

DeepMind illustrated the model's behavior with two worked examples. In one, AlphaGenome 2 was shown to reproduce a finding, originally published in Nature in 2024, that a distant regulatory variant overactivates the ETS2 gene specifically in gut macrophages, a mechanism linked to inflammatory bowel disease; DeepMind said the model singled out the same causal variant from unlabeled sequence data. In the second, given single-cell data from a COVID-19 vaccine study the model had not previously seen, it distinguished vaccine-activated T cells from resting ones and showed that a variant already associated with severe COVID-19 disrupted gene activity more strongly once those cells were activated.

Both examples are drawn from DeepMind's own announcement rather than independent replication, and the company reiterated that AlphaGenome 2, like its predecessor, is for research purposes only and has not been validated or approved for clinical or diagnostic use. The underlying AlphaGenome Atlas, a separately released, petabyte-scale index of predicted effects for nearly all 9 billion possible single-letter DNA changes in the human genome, which DeepMind published in September, carries the same caveat.

Who gets access, and reaction from outside researchers

DeepMind is making AlphaGenome 2 available immediately through its API for non-commercial research, with commercial research access limited to a private preview for now; the company also said a version will be built into its Google Antigravity coding-agent tools. Researchers can apply to a community program, with submissions open until November 30, for extra computing quota and credits to run their own projects, according to the announcement, which was also reported by The Verge. Code for the client library is published on GitHub, and usage guides are hosted on a separate documentation site.

DeepMind quoted two outside scientists in its release. Sarah Teichmann, a co-founder of the international Human Cell Atlas consortium who now leads a research group at the University of Cambridge's Stem Cell Institute, said of the model's ability to work from small amounts of new single-cell data:

"What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining."

Peter Donnelly, chief executive of the Oxford-spinout drug-discovery company Genomics, said sequence-prediction models of this kind help close the gap between the raw statistical associations produced by genome-wide association studies and the specific, mechanistic evidence that drug programs need before committing to a target. Donnelly's company said it plans to incorporate AlphaGenome 2 into Mystra, its existing AI platform for identifying drug targets from human genetics data.

What's next

The people most directly affected by the release are academic genomics researchers, computational biologists and drug-target-discovery teams at pharmaceutical and biotech companies, who make up the model's intended non-commercial and, eventually, commercial user base; DeepMind's framing is explicitly aimed at helping them triage the thousands of DNA variants that large genetic studies turn up each year, most of which have no known function. The company has not published a timeline for moving commercial access out of private preview, and it has not said whether or when any descendant of the model might be positioned for clinical or diagnostic settings, a use it continues to rule out for the current version. For now, DeepMind says the near-term signal of adoption will come from the community research projects it selects from its November 30 call for proposals, and from whether companies such as Genomics follow through on plans to fold the model into their own drug-discovery pipelines.