Google Deepmind releases AlphaGenome 2 to predict gene effects in single cells

From tissues to single cells

Every cell in the human body contains the same DNA blueprint, but different cell types interpret those instructions differently. For example, a heart muscle cell and an immune cell will activate different genes despite sharing the same genetic code. Understanding these specific differences remains one of the most significant challenges in modern genomics.

Google Deepmind's previous model, AlphaGenome, attempted to predict genetic properties across thousands of experiments involving many tissues simultaneously. However, this broad approach spread the model's capacity too thin. It made it difficult for the system to capture the subtle nuances that distinguish closely related cell types from one another.

AlphaGenome 2 addresses this by shifting to a conditional model. Instead of looking at a tissue as a whole, it takes a specific biological state as an extra input. This allows the system to calculate the effects of DNA sequences and variations on one specific cell type at a time.

How the model processes DNA

The new system was trained on a massive dataset of more than 107 million individual cell profiles from 68 different tissues in humans and mice. It does more than just predict gene expression, which is how strongly a gene is turned on. It also calculates how DNA is physically packaged in the nucleus to show which parts of the genome are accessible.

Additionally, AlphaGenome 2 can predict DNA methylation. These are chemical tags that function as molecular switches to silence or activate specific genes. By modeling these various layers of biology, the AI provides a more complete picture of how a cell functions based on its genetic makeup.

One of the model's most useful features is its ability to infer missing data from a single measurement. If a researcher only knows a cell's gene activity, the model can provide predictions for other properties, such as how tightly the DNA is packaged. This helps scientists decide which physical lab experiments are worth the time and resources to perform next.

Testing and disease insights

In testing on the TenK10K benchmark dataset, which includes 28 primary immune cell types, AlphaGenome 2 identified about 30 percent of causal variants. It also correctly predicted whether a variant would upregulate or downregulate a gene in 43 percent of cases. This is a significant improvement over the fine-tuned version of its predecessor, which managed only 16 and 23 percent respectively.

The model has already demonstrated its ability to reproduce known biological mechanisms. It successfully isolated a DNA variant in a regulatory switch that overactivates a specific gene in gut macrophages, which is a known mechanism for inflammatory bowel disease. Notably, the model showed that this effect occurred only in those specific cells while other cell types remained unaffected.

The AI also showed promise in analyzing vaccine responses. When applied to COVID-19 mRNA vaccine data, it distinguished between activated and resting T cells. It revealed that a specific genetic change associated with severe COVID-19 cases interferes with cell activity mostly during the activated phase, a detail that would be hidden if researchers only looked at resting cells.

Research and drug development

Deepmind has already precomputed predictions for hundreds of cell and tissue conditions to support AI-driven drug development. These predictions help researchers analyze genome-wide association studies, which link DNA regions to traits like height or type 2 diabetes. While these studies often only locate signals in broad stretches of DNA, AlphaGenome 2 can narrow down the likely causal variants.

The model can also identify the specific cell types affected by a mutation, such as astrocytes in the brain. When the model is confident enough, it can predict whether a mutation upregulates or downregulates a gene. This capability is particularly important for drug development, where understanding the direction of a genetic change is vital for creating effective treatments.

A major advantage of this new architecture is its ability to adapt to new experiments without needing to be retrained. Researchers can provide data on gene activity from a small group of cells, and the model can calculate effects for that previously unseen cell state. In many cases, it outperforms specialized models that were trained specifically on that small amount of data.

Availability and next steps

AlphaGenome 2 is currently available for noncommercial research through an API. For commercial users, access is provided through a closed preview. It is important to note that the system is not approved for use in clinical diagnostics at this time.

Google Deepmind plans to roll out dedicated features for AlphaGenome 2 on its agentic development platform, Google Antigravity. This platform is designed to combine the model with other tools, such as Gemini 4, to create automated research workflows. These workflows aim to streamline how scientists move from data collection to biological discovery.

By providing a tool that can predict complex cellular behaviors from DNA sequences, Deepmind aims to bridge the gap between human genetics and cell biology. The goal is to provide a more granular understanding of how individual cells contribute to human health and disease.

Try it today

Check the Google Deepmind website for information on accessing the AlphaGenome 2 API for noncommercial research purposes.

Watch

How the model performs when integrated into automated workflows on the Google Antigravity platform.

Further applications of the model in identifying specific cell types for other complex diseases.