Google DeepMind announced AlphaGenome 2 on November 11, 2026, moving its DNA sequence model from broad tissue-level predictions toward specific human cell types and states. Researchers can access it now through the AlphaGenome API for non-commercial research. Commercial research access is a separate private preview, according to DeepMind’s announcement.
The release gives genetics researchers two useful capabilities: predicting a variant’s molecular effects in a chosen biological context, and supplying their own single-cell RNA sequencing data to define contexts outside the model’s reference catalog. DeepMind also reports stronger variant identification results, although those figures are company-reported benchmarks, not independent clinical validation.
The consequential change is how researchers can frame a question. A DNA change might have little apparent effect across a whole tissue but matter in one specialized cell population. AlphaGenome 2 makes that population part of the prediction request.
How Cell-Type Conditioning Changes the Model
AlphaGenome 1 took a DNA sequence roughly one million letters long and predicted molecular properties across thousands of tissue and cell-line experiments in parallel. Its outputs were fixed: researchers received predictions for the contexts the model had been designed to cover.
AlphaGenome 2 instead predicts for one biological condition at a time. That condition can describe a tissue, cell type or cell state. DeepMind says the model receives the condition alongside the DNA sequence at every stage of its architecture, allowing the same sequence to produce different predictions in different cellular contexts. Both versions use million-letter sequence windows. The announcement explains this architectural change.
A tissue contains multiple kinds of cells. Averaging their activity can obscure a regulatory effect that appears strongly in one population and weakly elsewhere. A conditional model gives researchers a way to ask about that population directly, such as a particular immune cell type in blood.
My view is that this change matters more for experimental planning than the headline training-data count. A researcher deciding which cells to study needs a context-specific hypothesis. A large aggregate score alone provides less guidance about where to look.
The architecture also supports cross-modality imputation: predicting an unmeasured molecular property from an available measurement. DeepMind gives the example of supplying single-cell gene activity and predicting chromatin accessibility, which describes how accessible regions of DNA are to regulatory machinery.
That can help prioritize experiments. However, an accessibility prediction inferred from gene activity is not a second experimental measurement. Researchers should treat it as another model-generated hypothesis when deciding what to measure next.
What the Training Data and New Outputs Cover
DeepMind says AlphaGenome 2 was trained on more than 107 million individual cell profiles, using both bulk and single-cell data from human and mouse samples. Its headline description places the training coverage at 68 tissues.
Those figures describe profiles and tissue coverage, not 107 million distinct people or independently sampled patients. They indicate the scale of the biological measurements used for training.
The model predicts several molecular readouts:
| Predicted readout | What it helps researchers investigate |
|---|---|
| Gene expression | How active particular genes are in a biological condition |
| Chromatin accessibility | Which DNA regions are accessible to regulatory machinery |
| Histone modifications | Chemical marks on DNA-associated proteins involved in gene regulation |
| DNA methylation | Chemical marks on DNA that can relate to regulatory activity |
DNA methylation is a newly added readout compared with the previous AlphaGenome model, according to DeepMind.
Together, these outputs offer several views of a possible regulatory effect. A useful research workflow could examine whether a predicted expression change coincides with predicted changes in accessibility or other regulatory features. Agreement between predictions can help formulate an experiment, but it remains agreement within a modeling workflow.
Adapting Predictions to a Lab’s Own Single-Cell Data
Researchers do not have to limit themselves to predefined cell types. DeepMind’s AlphaGenome page says scientists can upload their own single-cell RNA-seq data to adapt AlphaGenome 2 to unseen biological states without fine-tuning.
Single-cell RNA-seq measures gene activity in individual cells. In this workflow, that information describes the biological condition in which the model should interpret a DNA sequence. DeepMind describes a cell encoder that generates the condition from a researcher’s single-cell experiment.
That opens up questions involving states that a broad tissue label cannot capture well:
- Disease progression: comparing variant predictions in differently profiled disease states.
- Drug treatment: examining predictions in a state represented by treated-cell data.
- Early development: investigating changing cellular programs over developmental stages.
These are application areas identified by DeepMind, not demonstrations that the model will perform equally well in every submitted dataset.
The distinction between conditioning and retraining is important. The advertised workflow uses a researcher’s expression data as context for the existing model; it does not require training a new model for each experiment. The Python repository describes this as expression conditioning with a researcher’s own human single-cell RNA-seq data.
The opened documentation does not provide enough detail to specify upload formats, preprocessing requirements or minimum cell counts here. Researchers should use the linked expression-conditioning tutorials before preparing a dataset.
A practical implication is that a lab can investigate its experimentally observed state instead of relying entirely on a generic cell label. The quality and relevance of that state description become part of interpreting the resulting predictions.
What the Benchmark Improvement Actually Says
DeepMind reports results on TenK10K, a single-cell dataset covering 28 primary immune cell types. Its announcement gives these comparisons with the original AlphaGenome fine-tuned on single-cell data:
| Evaluation outcome | Earlier model | AlphaGenome 2 |
|---|---|---|
| Identification of variants labeled causal in the benchmark | 16% | 30% |
| Identification of directional gene effects | 23% | 43% |
The announcement also discusses comparisons with specialized tools. These are DeepMind-reported results, and the opened sources provide no independent replication of those measurements. The benchmark figures appear in the release announcement.
The arithmetic supports the “nearly twice” description for these particular outcomes: 30 divided by 16 is about 1.88, while 43 divided by 23 is about 1.87. In absolute terms, the differences are 14 and 20 percentage points.
That is narrower than saying the model is twice as accurate across genetics research. The reported gains concern named evaluation outcomes in an immune-cell dataset.
The announcement’s use of “causal variants” also needs context. It describes variants identified within a benchmark evaluation. A model prediction about a variant’s molecular effect does not, by itself, prove that the variant causes a patient’s disease. Establishing that connection requires additional genetic, experimental and clinical evidence.
Frequently Asked Questions
3 questions
1Do I Need My Own Single-Cell RNA-Seq Data?
No. AlphaGenome 2 supports conditions selected from a reference catalog. Supplying your own data is an additional route for investigating newly profiled biological states.
