At a glance
Access: AlphaGenome 2 is available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research.
Official announcement: AlphaGenome 2
What shipped
Google DeepMind announced AlphaGenome 2 on Nov 11 2026, a next-generation model that predicts how DNA sequences and their variations affect specific human cell types. DeepMind describes it as a step-change improvement in variant predictive performance over the first AlphaGenome.
The first AlphaGenome launched last year to predict how DNA variants affect molecular activity across broad tissues. DeepMind later added AlphaGenome Atlas, a platform with predictions for all 9 billion possible single-letter changes in the human genome.
What changed in the design. The original model produced fixed parallel outputs for bulk tissues and cell lines from a 1-million-letter DNA sequence. DeepMind says that approach spread the model's capacity thin and made it hard to capture subtle differences between closely related cell types.
AlphaGenome 2 is a conditional model instead. It predicts DNA effects for one condition at a time, defined as a tissue, cell type or state. It takes that target condition as input at every stage of its architecture, alongside the million-letter sequence.
Training and outputs. DeepMind trained the model on bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles. It predicts the same core readouts as AlphaGenome, such as gene expression, histone modifications and chromatin accessibility. It adds DNA methylation, which DeepMind calls a widely requested readout.
Key capabilities
Cell-type resolution. DeepMind says the model can resolve activity inside specialized cells, such as distinct heart muscle or blood immune cells. The aim is to uncover DNA variants that act only in specific cell types.
New cell states without retraining. DeepMind says the model can generalize to entirely new single-cell experiments. A scientist uploads a single-cell RNA-seq snapshot from just a few dozen cells, and the model conditions on the resulting cell state. DeepMind says it can match or outperform models trained or fine-tuned directly on those datasets.
Cross-modality imputation. The model separates a cell's biological state from its outputs. If a scientist supplies one measurement, such as single-cell gene activity, it can predict unmeasured properties such as chromatin accessibility. DeepMind says this can help researchers prioritize which experiments to do next.
Worked examples from DeepMind.
- In a landmark study of inflammatory bowel disease, a variant was found to overactivate the ETS2 gene in gut macrophages. DeepMind says AlphaGenome 2 singled out the causal variant and showed it selectively boosts ETS2 in those cells.
- Given unseen single-cell data from a COVID-19 mRNA vaccine study, the model distinguished vaccine-activated T cells from resting ones.
- It also found that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state.
Precomputed scores. DeepMind precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions. It says they extend enhancer-gene mapping to primary human cell types such as brain astrocytes, and help pinpoint likely variants and cell types across traits from height to type 2 diabetes.
What it means for agent work
A specialist model like this is a tool an agent calls, not the agent itself. AlphaGenome 2 takes a sequence plus a condition and returns predictions. That fits a workflow in which a coordinating agent selects the condition, calls the model and passes the results to a human reviewer.
The condition input is the design point that matters for pipelines. The condition can come from a reference catalog of cell types or be generated from a scientist's own single-cell experiment through DeepMind's cell encoder. Log which condition produced each prediction, because the same stretch of DNA is read differently across cell types.
Imputation and experiment prioritization are natural agent tasks: flag which readouts are missing, rank candidate variants and propose the next experiment for a person to approve.
The research-only boundary shapes the design. DeepMind says AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. Keep outputs behind human review and out of automated clinical decisions.
DeepMind says it will soon release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, and expects specialized models to combine with agentic science tools such as Gemini 4 in Antigravity. Those Skills are not out yet, so agent integration today goes through the API. Genomics PLC's CEO Peter Donnelly said in the announcement that the company looks forward to incorporating the model in its Mystra platform, which is a stated intent rather than a shipped integration.
If you are planning how agents should call specialist models under approval and audit controls, read Governed AI agents need approval queues and audit trails, or Book a Consultation.
How to try it
AlphaGenome 2 and interactive Colab tutorials for expression conditioning are available today for non-commercial research through the AlphaGenome API, with commercial research in private preview. DeepMind has not published pricing for the commercial preview.
DeepMind also invites researchers to share interesting use cases through its form by Nov 30 2026. Selected projects may be eligible for additional quota and credits. A community forum is open for feedback on the future roadmap, and the announcement links to the paper and the API access page.