Google DeepMind has released AlphaGenome 2, a model that predicts how a DNA variant changes gene regulation in a specific cell type or cell state, rather than averaging across whole tissues. Announced on November 11, 2026, the model targets a long-standing problem in genomics: many variants linked to disease sit in non-coding DNA, and their effects depend heavily on which cells carry them.
What the Model Does Differently
The original AlphaGenome could predict regulatory effects of genetic variants, but its predictions were not tied closely to individual cell types. AlphaGenome 2 conditions each prediction on a chosen tissue, cell type or cell state. Researchers can also supply their own single-cell RNA-seq data, which a cell encoder turns into a condition without further fine-tuning, according to Unite.AI's review of the release.
The model predicts gene expression, RNA splicing and chromatin accessibility, and it adds CpG DNA methylation relative to the first version. DeepMind says it was trained on more than 100 million single-cell profiles drawn from human and mouse data.
The Benchmark Numbers
The headline result comes from TenK10K, a single-cell immune dataset covering 28 cell types. There, the model recovered 30% of putatively causal expression-linked variants at 90% precision. AlphaGenome 1, fine-tuned on single-cell data, recovered 16% under the same conditions. On the direction of effect, AlphaGenome 2 recovered 43% of variant effect directions at 90% accuracy, up from 23% for the predecessor.
Those are DeepMind's own figures, drawn from a preprint that has not been peer reviewed. Independent evaluations of the second version had not been reported at the time of the announcement, so the numbers should be read as promising rather than settled. The model also explained a median 34% of within-tissue, cell-type-specific expression variance on held-out regions, compared with 23% for the earlier model.
Research Examples
In one inflammatory bowel disease analysis, the model flagged the variant rs2836882 and linked it to the gene ETS2, roughly 290 kilobases away. It localized the effect to gut monocytes and macrophages, which matches a mechanism that had previously been established experimentally.
In a second example, the model used COVID-19 vaccine single-cell data to distinguish activated from resting T cells. It predicted a stronger effect in the activated state for a variant associated with severe COVID-19. DeepMind also released precomputed resources, including enhancer-gene links across 640 reference conditions and 39,323 genes, and scores for more than 6.8 million common and trait-associated variants across 510 conditions.
Access, Limits and Cost
Non-commercial use of the API is free, although query rates vary with demand. Commercial research is in private preview. Neither of the two reports reviewed for this article gives a timeline for releasing the model weights, so it is unclear how easily outside groups will be able to run the model on their own infrastructure.
Compute is another constraint. Implicator reports that the more detailed predictions cost more to run than the original AlphaGenome, and that this makes the newer model impractical for the genome-wide atlas DeepMind released in September. The same coverage cites a Charité-led preprint posted on November 9, which tested the original model rather than the second version and found it often failed to carry promoter disruption through to the rest of a gene's transcript.
The model also has known blind spots. It can underestimate effects outside accessible chromatin, at very distant regulatory elements, or through mechanisms that act after transcription. Transient, stimulated and disease states are underrepresented in the training data, and the model struggles to separate some fine-grained neuronal subpopulations. Because its cell representations come from observational profiles, the predictions capture correlations rather than causation, and its ability to predict knockout or overexpression experiments remains uncertain. Both versions are restricted to research use and have not been validated or approved for clinical or diagnostic purposes.
Voices From the Field
Žiga Avsec, DeepMind's genomics lead, explained the motivation in terms drug developers will recognize. "When you develop drugs, you're targeting specific proteins and specific cell types," he said.
James Lee, an immunologist at the Francis Crick Institute, was more pointed about the gap the model addresses. Of the new model, he said: "They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments." He also said the original model's lack of cell-type specificity meant it "hasn't been hugely useful for us."
Why It Matters
Pinpointing which variants matter in which cells is one of the bottlenecks between genome-wide association studies and drug targets. A model that narrows the search from thousands of candidate variants to a short list of cell-specific mechanisms could shorten the path from a genetic signal to a biological hypothesis. That is the promise DeepMind is pitching to researchers.
The caveats are equally important. The benchmarks are self-reported, the weights are not yet public, and the best results so far come from a limited set of immune and tissue contexts. The next signals will come from independent groups testing the model on their own data, and from whether the cell-specific predictions hold up in wet-lab experiments.
