Google DeepMind Launches AlphaGenome 2 to Predict DNA Mutation Effects Cell by Cell

Every cell in the human body carries the exact same DNA, yet a neuron, a cardiac muscle cell, and a white blood cell read entirely different parts of it. Capturing this distinction is precisely what Google DeepMind aims to achieve with AlphaGenome 2, unveiled on November 11 on its blog. The model predicts how DNA variations alter gene activity, no longer at the scale of whole tissues, but for a specific cell type or cellular state.

A Model That Reasons Cell by Cell

The original AlphaGenome, launched last year, processed a 1-million-letter sequence to output thousands of predictions at once across tissues and cell lines. While effective, DeepMind notes that this approach diluted the model’s capacity and struggled to distinguish closely related cell types. AlphaGenome 2 introduces an architectural shift: it is fully conditional. Researchers input the DNA sequence alongside the target biological condition—selected from a catalog of cell types or derived from experimental data—and the model predicts solely for that specific state.

The model was trained on human and mouse datasets, encompassing more than 107 million single-cell profiles. It predicts gene expression, histone modifications, and chromatin accessibility, mirroring its predecessor, while adding DNA methylation—a heavily requested feature, according to the developer. It is also capable of imputing missing data: from a single measurement, such as gene expression, it can estimate the others.

Adapting to New Cells Without Retraining

The feature drawing the most praise from researchers cited by DeepMind is its generalization capability. A lab can input expression profiles from just a few dozen cells—such as immune cells following vaccination—and the model leverages this data to predict variant effects in that exact state without requiring retraining. DeepMind highlights an example from an mRNA Covid-19 vaccine study: a variant associated with severe disease disrupted gene activity more prominently in activated T cells than in resting cells.

In terms of benchmarks, the developer reports that on the TenK10K dataset (spanning 28 immune cell types), AlphaGenome 2 identifies nearly twice as many causal variants as its predecessor: 30% compared to 16%, and 43% compared to 23% for the direction of gene expression changes. DeepMind also showed that the model computationally pinpointed a distant genetic “switch” variant that overactivates the ETS2 gene in gut macrophages—previously identified by a published study as a key driver of inflammatory bowel disease.

Availability and Access

AlphaGenome 2 is available starting today for non-commercial research via the AlphaGenome API, alongside interactive Colab tutorials, and in private preview for commercial applications. DeepMind has also precomputed predictions for disease-associated variants across hundreds of cell types and tissues, and announced that “AlphaGenome 2 Skills” will soon arrive in Antigravity, its agentic development platform. The company emphasizes that these models are neither validated nor approved for diagnostic or clinical use.