DeepMind releases AlphaGenome 2 to predict cell-type-specific effects of DNA variants

Google DeepMind said Wednesday it has developed AlphaGenome 2, a new DNA-sequence model that it says can predict how genetic variants affect gene regulation in specific human cell types, a step aimed at helping researchers sort through which DNA changes may matter in disease biology.

The advance, announced Nov. 11 in a DeepMind blog post by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec, centers on making predictions conditional on biological context. Unlike the earlier AlphaGenome, DeepMind says AlphaGenome 2 takes not just a DNA sequence but also a condition such as tissue, cell type or cell state, reflecting the idea that the same DNA can be read differently in different cells. The model is available now for noncommercial research through the AlphaGenome API, while commercial research access is in private preview. DeepMind said both AlphaGenome and AlphaGenome 2 are for research use only and “have not been validated or approved for clinical or diagnostic purposes.”

DeepMind describes AlphaGenome 2 as a “next-generation model” for predicting how DNA sequences and their variations affect specific human cell types. The company said it trained the model on more than 107 million individual cell profiles spanning 68 tissues, using both bulk and single-cell data from human and mouse samples.

Like the original AlphaGenome, the new model predicts several broad signs of gene regulation, including gene expression, histone modifications and chromatin accessibility — measures researchers use to understand whether genes are active and how tightly DNA is packaged. DeepMind said AlphaGenome 2 also adds DNA methylation, a chemical marking of DNA linked to gene control. The company said the model can also perform cross-modality imputation, meaning it can infer missing experimental measurements from limited inputs.

The company’s strongest performance claims came from a benchmark on the TenK10K single-cell immune dataset, which covers 28 primary immune cell types. DeepMind said AlphaGenome 2 identified 30% of causal variants, compared with 16% for the earlier AlphaGenome fine-tuned on single-cell data and other specialized tools. It also said the model captured 43% of directional gene effects, compared with 23% for those earlier approaches. Those results come from DeepMind’s own benchmarking and have not yet been broadly validated by outside groups.

In plain terms, the problem DeepMind is trying to address is a common one in human genetics. Large genetic studies can point to stretches of DNA associated with disease risk, but researchers often still struggle to pin down which variant in that region is actually causal, which gene it changes, and in which cell type that effect happens. A model that can narrow those possibilities by cell type could make follow-up biology faster and more targeted.

DeepMind pointed to inflammatory bowel disease as one example of how the system might be used. The company said AlphaGenome 2 computationally reproduced a published 2024 Nature finding that highlighted a regulatory variant increasing ETS2 activity specifically in gut macrophages, a type of immune cell in the intestine. That kind of result is important because it aims to connect three pieces at once: the variant, the gene it affects and the cell type where the effect shows up.

Professor Sarah Teichmann, a genomics researcher quoted in DeepMind’s announcement, said the model’s appeal is that it can adapt directly to new single-cell data without retraining. “What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states from even a small sample of cells,” she said in the blog post.

The new system builds on AlphaGenome, the earlier model in the same research line, which was published in Nature on Jan. 28, 2026. That publication gave the broader program a peer-reviewed foundation as a tool for predicting regulatory effects of DNA variants. AlphaGenome 2 extends that work by making those predictions dependent on tissue, cell type or cell state, which is central to many disease questions.

That could make the model useful in research areas where the biology depends heavily on cell context, including immune disease, cancer and other conditions in which the same genetic variant may matter in one cell population but not another. Peter Donnelly, CEO of Genomics PLC, said in the DeepMind blog that “DNA sequence prediction models are an important tool in helping bridge that gap,” though such comments in the announcement are not independent validation.

For now, the announcement is best understood as a research tool update rather than a clinical milestone. The AlphaGenome line already has a peer-reviewed paper behind it, but AlphaGenome 2’s latest claims — including its benchmark gains and disease-related examples — still await broader independent evaluation in the hands of outside researchers.