AivexaNewsSearch
AI news for builders and product teamsChecked every hour

AlphaGenome 2: Deciphering how cells read DNA to better understand disease

Collected Nov 11, 2026

Google DeepMind announced AlphaGenome 2 on its research blog, a next-generation DNA sequence model that predicts how DNA sequences and their variations affect specific human cell types. The model is available today for non-commercial research through the AlphaGenome API, and in private preview for commercial research. Interactive Colab tutorials for what the team calls expression conditioning accompany the release. The work is credited to Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec, among a longer list of contributors in the acknowledgements.

The core change is architectural. The first AlphaGenome model took a one-million-letter DNA sequence and produced predictions across thousands of experiments from many tissues in parallel. That was efficient but, in the team's account, spread the model's capacity thin and made subtle differences between closely related cell types hard to capture. AlphaGenome 2 is instead a conditional model: it takes a specific biological condition alongside the DNA sequence, where a condition means a tissue, a cell type or a cell state. The condition is fed in at every stage of the architecture, so the model learns how the same stretch of DNA is read differently depending on the cell.

Conditions can come from a reference catalog of cell types, or be generated on the fly from a scientist's own single-cell experiment through a component the team calls a cell encoder. Training drew on both bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles, spanning 68 tissues. The stated goal is to move past bulk tissue averages and resolve activity inside specialized cells, such as distinct heart muscle cells or blood immune cells, where disease-relevant variants may act only in that one cell type.

On outputs, the model predicts the same core experimental readouts as AlphaGenome, including gene expression, histone modifications and chromatin accessibility, and adds DNA methylation, described as a widely requested readout. Being conditional also enables cross-modality imputation: because the model separates a cell's biological state from its measured outputs, a scientist can supply one measurement, such as single-cell gene activity, and have the model predict unmeasured properties like chromatin accessibility. In practice that is a way to decide which experiments are worth running next, rather than a replacement for them.

The headline benchmark is variant recall. On TenK10K, described as a large single-cell map of 28 primary immune cell types, AlphaGenome 2 identified causal variants at 30% versus 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools, and directional gene effects at 43% versus 23%. The team summarizes this as identifying nearly twice as many variants that change gene expression in cells. It also reports better cell-type specificity on genome regions outside its training data, including tracking continuous changes such as blood stem cells maturing into red blood cells and immune lineages.

A worked example concerns inflammatory bowel disease. A recent study showed that a variant in a distant regulatory switch drives IBD by overactivating the ETS2 gene in gut immune cells called macrophages. Running predictions across several gut immune cell types, AlphaGenome 2 reproduced the finding computationally: it singled out the causal variant and showed it selectively boosts ETS2 in gut macrophages while leaving other cell types and genes largely unaffected. Reproducing a known result is a weaker test than a prospective discovery, so it is best read as a sanity check on the model's cell-type resolution rather than proof of new biology.

Generalization to unseen cell states is the other claim. A scientist can upload a single-cell RNA-seq snapshot from just a few dozen cells; that gene activity is compressed into a cell state representation and used to condition the model on states it has not seen. In the team's tests, AlphaGenome 2 matched or outperformed models trained or fine-tuned directly on those datasets. Given single-cell data from a COVID-19 mRNA vaccine study, it distinguished vaccine-activated T cells from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect that would be missed in unactivated cells. Adapting without retraining is a meaningful practical advantage, since fine-tuning a genomics foundation model is expensive and requires labeled data for each new context.

DeepMind has also precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions. Two uses are highlighted. One is mapping enhancer-to-gene links in primary human cell types, such as brain astrocytes, since experimental CRISPR screens mostly rely on stable lab-grown cell lines and therefore lack wiring diagrams for cells in the body. The other is resolving GWAS signals into a specific causal variant, target gene and cell type across complex traits including height and type 2 diabetes, with the model also predicting whether a variant raises or lowers a gene's activity when it is confident.

The release is explicitly research-only: the models have not been validated or approved for clinical or diagnostic use. DeepMind says AlphaGenome 2 Skills will soon arrive in Google Antigravity, its agentic development platform, pairing specialized models with Gemini 4 in that environment. It is also soliciting community use cases by 30 November, with selected projects possibly eligible for additional quota and credits, and inviting roadmap feedback through its community forum.

Why it matters: For computational biologists and the bioinformatics teams supporting drug discovery, the shift from tissue averages to cell-type-conditional predictions targets the step where GWAS associations are hardest to convert into targets, and the API plus private commercial preview gives teams a way to test that on their own variants now. The caveats are the research-only status and the fact that headline gains come from benchmarks rather than independent replication, so I would treat the reported doubling as a strong signal to evaluate rather than an established result. Product teams in life sciences should also note the Antigravity integration, which suggests DeepMind wants these models used inside agentic workflows rather than as standalone endpoints.

Read at Google DeepMind

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Predicting how DNA sequences and their variations impact specific human cell types and disease.