Overview
AlphaGenome 2 is Google DeepMind’s DNA-sequence model for predicting how sequences and genetic variants affect gene regulation in specific human cell types and states. It was announced on November 11, 2026. (deepmind.google)
Capabilities
- A conditional model takes a DNA sequence of up to approximately one million letters and a target tissue, cell type, or cell state. Conditions can come from a reference catalog or a researcher’s single-cell experiment.
- Predicts molecular readouts including gene expression, histone modifications, chromatin accessibility, and DNA methylation. It can also infer missing experimental measurements across modalities.
- Can use single-cell RNA-seq data from a few dozen cells to make predictions for previously unseen cell states without retraining.
- Training used bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. (deepmind.google)
Research uses and reported performance
The model can help researchers prioritize variants and experiments, link enhancers to target genes, and investigate which cells may mediate associations found in genome-wide association studies. Google also reports precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions.
On the TenK10K immune-cell benchmark, Google reports identifying 30% of causal variants versus 16% for the earlier comparison models, and 43% versus 23% for directional gene effects. These figures are benchmark-specific, not a guarantee of performance in every setting. (deepmind.google)
Access and pricing
AlphaGenome 2 is available through the AlphaGenome API for non-commercial research; commercial research access was announced as a private preview. The official API repository says non-commercial API use is free, with query rates varying by demand. Currency: not applicable to the free non-commercial API. Input and output charges: no separate charges are specified for that free use; commercial input and output pricing and currency remain unverified. (deepmind.google)
Limitations
Predictions have varying confidence and should be treated as research hypotheses rather than established clinical findings. The model and its outputs are intended for research only and have not been validated or approved for clinical or diagnostic use. The technical paper says model source code and weights will be provided upon final publication. (storage.googleapis.com)