Google DeepMind introduced AlphaGenome 2 on November 11, extending its DNA prediction work to specific cell types and biological states. The model is available for noncommercial research through the AlphaGenome API, while commercial access is in private preview, according to the company.
The distinction matters because a genetic change can have different consequences in different cells. DeepMind says the model is designed to investigate that context rather than relying only on measurements averaged across a tissue. It is a research tool, not an approved diagnostic system.
Moving from tissue averages to cellular context
DeepMind’s announcement says training incorporated more than 107 million individual cell profiles. Researchers can supply a DNA sequence together with a target biological condition. The company also describes support for incorporating information from a new single-cell experiment without retraining the entire model.
The accompanying AlphaGenome 2 research paper explains why this is a difficult inference problem. Many disease-associated genetic variants sit outside protein-coding regions. Cell atlases provide detailed observations, but individual datasets may not include enough donors to establish how a particular variant changes gene activity.
What the research paper reports
The paper describes predictions conditioned on a DNA segment of roughly one million letters and a target cell context. Its evaluations examine regulatory effects, including gene expression and chromatin accessibility. These are measurements relevant to biological regulation, not direct predictions of whether an individual patient will develop a disease.
The authors report improved recovery of causal single-cell expression-associated variants at a specified precision threshold. They also describe an atlas of enhancer-to-gene relationships and scores for seven million common variants. These results come from the research team’s evaluation and should not be confused with independent clinical validation.
Access has practical limits
The official AlphaGenome documentation says noncommercial API access is free, but query rates can vary with demand. It distinguishes looking up precomputed atlas results from running live predictions. The latter is intended for workloads involving thousands of queries rather than unrestricted scanning of millions of variants.
That distinction affects how a research group might plan an experiment. Before committing to a large analysis, a team should check whether its intended measurements are already represented, which model or resource supplies them, and how much live computation is required. That is a practical planning recommendation, not a claim that every use case is supported.
Prediction still needs experimental follow-up
DeepMind explicitly says AlphaGenome 2 has not been validated or approved for clinical or diagnostic purposes. For readers outside genomics, the useful framing is hypothesis prioritization: a computational result can suggest where researchers might look next, without establishing a treatment decision or a causal explanation on its own.
The immediate test is therefore research usefulness. Can a laboratory choose better follow-up experiments, reproduce results in its own data and understand where the model fails? Those questions are more informative than treating a benchmark gain as proof that the underlying biology has been solved. The new release expands the questions researchers can ask, while leaving verification firmly in the experimental workflow.
