The update addresses a longstanding challenge in genetics research. Although cells throughout the body generally contain the same DNA, different cell types activate different genes to perform specialized functions. A genetic variation associated with disease may therefore affect one type of cell while having little impact on others.
Earlier approaches that examined average activity across entire tissues could overlook these differences. AlphaGenome 2 is designed to provide more precise predictions by evaluating how the same DNA sequence behaves under different cellular conditions.
The original AlphaGenome, introduced in 2025, could predict molecular activity across numerous tissues and experimental conditions simultaneously. DeepMind has since expanded that work with AlphaGenome Atlas, which contains predictions covering all 9 billion possible single-letter substitutions in the human genome.
For the latest version, researchers redesigned the underlying architecture to concentrate on one biological condition at a time. The model receives both a DNA sequence and information describing a particular tissue, cell type or cellular state, allowing it to generate predictions tailored to that environment.
DeepMind says the change substantially improves its ability to identify genetic variations that influence gene expression in specific cells. According to the company's evaluations, AlphaGenome 2 detects nearly twice as many expression-altering variants as its previous leading model.
On the TenK10K dataset, which contains information from 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants, compared with 16% for the earlier model. It also correctly predicted the direction of gene activity changes in 43% of cases, up from 23%.
These figures come from DeepMind's own testing and have not been independently verified in the available reporting. Nevertheless, they suggest that the newer architecture can identify genetic effects that broader models may miss.
One example involves inflammatory bowel disease. Previous research identified a genetic variation that increases activity of the ETS2 gene in macrophages, a type of immune cell found in the gut.
DeepMind reports that AlphaGenome 2 reproduced this relationship computationally, identifying the relevant variant and predicting its effect on ETS2 activity specifically within those immune cells. The model also distinguished that response from activity in other nearby cell types.
The ability to isolate such effects could help researchers investigate why certain genetic variations contribute to disease and determine which cells are most relevant for further study.
AlphaGenome 2 also introduces greater flexibility when working with new biological data. Scientists can provide gene activity measurements from a previously unseen cell population, allowing the model to predict how genetic variants may behave without requiring additional training.
According to DeepMind, the system can adapt using single-cell sequencing information from only a few dozen cells. It converts those measurements into a representation of the cells' biological state and uses that information to guide its predictions.
Sarah Teichmann, a University of Cambridge professor and co-founder of the Human Cell Atlas consortium, highlighted this capability as an important improvement for researchers studying newly identified cellular conditions.
The model can also estimate biological properties that were not directly measured in an experiment. For example, researchers who provide gene expression data can receive predictions about chromatin accessibility, which helps describe how readily different regions of DNA can be accessed.
In another demonstration, DeepMind applied AlphaGenome 2 to previously unseen data from a COVID-19 mRNA vaccination study. The model distinguished activated immune T cells from resting cells and predicted that a genetic variant associated with severe COVID-19 had a stronger effect on gene activity in the activated population.
Beyond gene expression, AlphaGenome 2 predicts several other molecular characteristics, including histone modifications and chromatin accessibility. The update also adds DNA methylation predictions, expanding the range of biological processes researchers can examine.
These capabilities are particularly relevant to genome-wide association studies, which identify regions of DNA linked to diseases or other traits but often cannot establish which specific variant, gene or cell type is responsible.
DeepMind has generated predictions for disease-associated variants across hundreds of human cellular conditions, allowing researchers to explore those relationships without running every analysis themselves.
The company says AlphaGenome 2 can help identify connections between regulatory DNA regions and the genes they influence. Such relationships can be difficult to establish experimentally, especially in specialized human cells that are not easily represented by conventional laboratory cell lines.
The model's predictions cover a range of complex traits and conditions, including type 2 diabetes, providing researchers with potential starting points for investigating disease mechanisms and prioritizing experiments.
For drug discovery, the distinction between identifying a genetic association and understanding its biological effects is significant. A model that predicts which cells are affected, which genes are involved and whether their activity increases or decreases could help researchers narrow the questions they investigate experimentally.
However, AlphaGenome 2 remains a research tool rather than a validated clinical system. It has not been approved for diagnostic use, and its predictions still require experimental investigation before they can establish biological mechanisms or support medical decisions.
Access is also limited by how DeepMind is distributing the technology. Non-commercial researchers can use AlphaGenome 2 through the AlphaGenome API, while commercial organizations currently have access through a private preview.
That arrangement gives academic researchers an immediate way to experiment with the model but leaves broader commercial availability uncertain. Startups developing genomics or drug-discovery products may need to obtain preview access before incorporating the technology into their workflows.
The new model also requires substantially more computing resources than its predecessor, according to additional reporting. That could make large-scale analyses across extensive collections of variants and cell types more demanding, particularly for organizations with limited computational budgets.
DeepMind is continuing to support the original AlphaGenome for broader applications while developing additional tools around the new model.
The company plans to introduce AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, and connect specialized scientific models with capabilities from Gemini 4. It is also working with research partners, including the Wellcome Sanger Institute, to expand coverage of human cell types and biological states.
Researchers can currently access interactive tutorials alongside the AlphaGenome 2 API. DeepMind is also inviting submissions describing potential applications, with selected projects eligible for additional usage allowances and credits. The submission deadline is November 30.
AlphaGenome 2 represents a shift from predicting genetic activity across broad biological samples toward examining the effects of DNA variations within specific cellular environments. Its reported improvements could make it useful for investigating disease-associated variants, although independent validation, computing requirements and commercial access will influence how widely the technology is adopted.
This analysis is based on reporting from Google DeepMind & Startup Fortune.
Image courtesy of Google DeepMind.
This article was generated with AI assistance and reviewed for accuracy and quality.
