Overview
AlphaGenome 2 is the next-generation DNA sequence model from Google DeepMind, released on November 11, 2026 and succeeding AlphaGenome.
The architectural change is that it is a conditional model. Beyond a one-million-letter DNA sequence, the input includes a biological condition, meaning a tissue, cell type or cell state, and that condition participates at every stage of the architecture. Conditions come from two places: a reference catalogue, or generated on the fly by a cell encoder from the scientist's own single-cell experiment, which Google says needs only a few dozen cells. That second route matters, because it means you no longer wait for the catalogue to cover your system.
Outputs keep the core experimental readouts from the previous generation, including gene expression, histone modifications and chromatin accessibility, and add DNA methylation, the most requested addition. Cross-modality imputation is supported too: input single-cell gene activity and infer chromatin accessibility you never measured.
Training covers 107 million single-cell profiles across 68 tissues, drawing on human and mouse bulk and single-cell data. Google frames the effect as turning static cell atlases into a predictive engine for human genetics.
The performance numbers are stated plainly. On TenK10K, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 reliably identifies nearly twice as many causal variants, 30% against 16%, and directional gene effects at 43% against 23%. Those comparisons are same dataset and same terms, measured against the original AlphaGenome fine-tuned on single-cell data alongside other specialised tools, so they hold up reasonably well.
Generalisation to unseen cell states is the other capability worth noting. On genomic regions outside its training data the model explains more of how expression differs across cell types within a tissue, and it predicts unseen states without retraining. Google's example is telling vaccine-activated T cells apart from resting ones.
Availability splits in two. Non-commercial research can use it through the AlphaGenome API on day one, with Colab tutorials for expression conditioning; commercial research runs through a private preview. Google also says AlphaGenome 2 Skills will arrive in Google Antigravity.
The usage boundary belongs in plain text. Google states that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. That is not boilerplate; it keeps the model out of any diagnostic pipeline.
Key Features
- Conditional architecture: Takes a one-million-letter DNA sequence plus a tissue, cell type or cell state, with the condition participating at every layer to tighten predictions from tissue averages down to specific cells.
- Conditions from your own single-cell data: A cell encoder generates the condition from your own single-cell experiment, needing only a few dozen cells according to Google, so you are not blocked waiting for the catalogue.
- DNA methylation added: Adds DNA methylation to gene expression, histone modifications and chromatin accessibility, the readout users asked for most.
- Cross-modality imputation: Infers unmeasured properties from one measurement, for example predicting chromatin accessibility from single-cell gene activity.
- Trained on 107 million profiles: Covers 68 tissues across human and mouse bulk and single-cell data, which Google describes as turning static atlases into a predictive engine.
- Nearly double the causal variants: 30% detection on TenK10K against 16% for the fine-tuned previous model, and 43% against 23% on directional gene effects, same dataset and terms.
- Generalises to unseen states: Predicts cell states absent from training, with Google's example being vaccine-activated versus resting T cells, without retraining.
- API and Colab from day one: Non-commercial research gets the AlphaGenome API immediately plus interactive Colab tutorials for expression conditioning, with Antigravity Skills to follow.
Use Cases
- Mapping GWAS loci to specific cell types
- Locating mechanisms behind rare and complex disease
- Forecasting effects before designing single-cell experiments
- Prioritising non-coding variants by pathogenicity
- Imputing missing modalities
- Attributing drug targets to cell types
- Validating known mechanisms such as ETS2 regulation in intestinal macrophages in IBD
Pros
- Prediction granularity drops from tissue averages to individual cell types and states, surfacing variants that act in one cell only
- 30% against 16% on causal variants and 43% against 23% on directional effects, same dataset and terms
- Adds DNA methylation and supports cross-modality imputation from a single measurement
- Conditions can come from your own single-cell experiment with a few dozen cells, without catalogue coverage
- Generalises to unseen cell states without retraining per state
- Available to non-commercial research through API and Colab on release day
Pricing
Free for non-commercial research through the AlphaGenome API and Colab tutorials; commercial research runs through a private preview requiring an application. The community use-case call runs to November 30, and selected projects may receive additional quota and credits. Intended for research only, not validated or approved for clinical or diagnostic use.
Summary
Genome models used to tell you a sequence looked regulatory in liver tissue. AlphaGenome 2 narrows that to which cell type, in which state. The same DNA reads differently in heart and immune cells, and disease usually strikes only one of them, which is where the real value of this upgrade sits. Causal variant detection moves from 16% to 30% on the same dataset and the same terms. Non-commercial research can use it today; commercial use remains in private preview.
Version History
- AlphaGenome 2 released (2026-11-11): Google DeepMind released AlphaGenome 2, a conditional DNA sequence model taking a one-million-letter sequence plus a tissue, cell type or state and predicting gene expression, histone modifications, chromatin accessibility and newly added DNA methylation. Trained on 107 million single-cell profiles across 68 tissues from human and mouse bulk and single-cell data. On TenK10K it identifies causal variants at 30% against 16% for the previous model fine-tuned on single-cell data, and directional gene effects at 43% against 23%, and generalises to unseen cell states. Open to non-commercial research via the AlphaGenome API and Colab, commercial use in private preview, with Google Antigravity Skills to follow. Google states it is for research only and not validated or approved for clinical or diagnostic use.