What happened
On 11 November 2026, Google DeepMind published AlphaGenome 2: Deciphering how cells read DNA to better understand disease on its blog, with an accompanying technical paper. AlphaGenome 2 is a DNA sequence model that predicts how a DNA sequence, and variants within it, affect a specific human cell type or state. The first AlphaGenome was announced in June 2025 and described in a Nature paper in January 2026.
Key details
- Conditional design. According to DeepMind, the original model predicted many tissues in parallel. AlphaGenome 2 instead takes a biological condition (a tissue, cell type or state) together with a sequence of one million DNA letters and predicts effects for that condition. Conditions can also be generated from a user’s own single-cell data, and the model can predict some unmeasured readouts from a single measurement, for example chromatin accessibility from gene activity.
- Outputs. Gene expression, histone modifications, chromatin accessibility and DNA methylation.
- Training data. DeepMind says the model was trained on more than 100 million individual cell profiles across 68 tissues, from human and mouse samples.
- Headline benchmark. On the TenK10K dataset (28 primary immune cell types), DeepMind reports that AlphaGenome 2 identified causal variants 30% of the time versus 16% for the comparison, and directional gene effects 43% versus 23%. The comparison was against the original AlphaGenome fine-tuned on single-cell data and other specialised tools. These are developer-reported figures.
- Case example. DeepMind says the model reproduced a computational finding linking a variant to ETS2 overactivation in gut macrophages in inflammatory bowel disease.
- Access. Available now for non-commercial research through the AlphaGenome API. Commercial research access is in private preview. Colab tutorials for expression conditioning are provided. DeepMind lists “AlphaGenome 2 Skills” in Google Antigravity as planned for a later release. The announcement does not mention open model weights.
- Stated limits. The models are “intended for research use only and have not been validated or approved for clinical or diagnostic purposes.” The GWAS-linked predictions are precomputed rather than run on demand.
What it means for researchers and research offices
For genomics groups, the practical change is that cell-type-specific variant interpretation, for example in immune cells, can be queried through an existing API rather than requiring a lab to train a model. The ability to condition on a lab’s own single-cell data is the feature most likely to matter for study-specific work, though its performance outside the published benchmarks has not been independently assessed.
For research offices and technology-transfer teams, the access split is the immediate issue. Non-commercial use is open; commercial use is by private preview only, so projects with industry funding, sponsored collaborations or spin-out intent should check terms before building workflows on API outputs. Because access is through a hosted API and not downloadable weights, data governance questions (what sequence or single-cell data may be sent to a third-party service, and under what consent and data-use agreements) fall to the research data management and compliance function. See also our guidance on research data management.
Benchmarks reported by the developer on a single dataset are a starting point, not validation. Results used in publications should state the model version, access date and conditioning inputs so they can be reproduced.
What to watch
- Independent evaluations and peer-reviewed publication of the AlphaGenome 2 paper.
- Terms and pricing of commercial access when it leaves private preview.
- Whether weights or on-demand variant scoring become available.
- DeepMind’s use-case submission form (deadline 30 November 2026), which signals the kinds of applications the team is soliciting.
Sources
- Google DeepMind, “AlphaGenome 2: Deciphering how cells read DNA to better understand disease”, 11 November 2026 (read 13 November 2026)
- AlphaGenome 2 technical paper (PDF) (read 13 November 2026)
- AlphaGenome API documentation (read 13 November 2026)
- STAT, coverage of the original AlphaGenome launch, 25 June 2025 (read 13 November 2026)
