SIAgentNews.com
← All news

Google DeepMind launches AlphaGenome 2 to predict DNA effects in specific cells

The new model helps scientists understand how tiny genetic variations impact different human cell types to better uncover the drivers of disease.

· First reported by Google DeepMind

107 millionindividual cell profiles used for training

The short version

  • Trained on over 107 million individual cell profiles across 68 different human and mouse tissues.
  • Identifies nearly twice as many variants that change gene expression in cells compared to previous state-of-the-art models.
  • Uses a conditional architecture to predict DNA effects for one specific tissue, cell type, or state at a time.
  • Can predict missing experimental data, such as chromatin accessibility, from a single measurement like gene activity.
  • Available now for non-commercial research via an API and in private preview for commercial use.

Why it matters

This tool helps bridge the gap between broad genetic studies and specific cellular biology. It allows researchers to pinpoint exactly which cells are affected by a genetic mutation, which is essential for understanding complex diseases.

Moving beyond tissue averages

Every cell in the human body contains the same DNA blueprint, but different cells perform different jobs by switching specific genes on or off. For example, a heart muscle cell and an immune cell react differently to the same genetic instructions. While global projects like the Human Cell Atlas are mapping these expressions, predicting how small DNA variations change these behaviors remains a major challenge.

Google DeepMind released AlphaGenome 2 to address this specific gap in human genetics. The previous version of the model looked at broad tissue averages, which can sometimes mask the unique behaviors of individual cell types. AlphaGenome 2 is designed to zoom in on specialized cells, such as specific blood immune cells or distinct heart muscle types, to find the cellular drivers of disease.

By focusing on these specific cellular states, the model helps scientists prioritize which genetic variants to study first. It can identify variants that only cause problems in certain cells, which might be overlooked by traditional methods. This level of detail is critical for understanding rare diseases and complex human traits that do not affect the entire body equally.

A new conditional architecture

The technical core of AlphaGenome 2 is a fundamental redesign of how the model processes biological information. Instead of trying to predict many different tissues at the same time, it functions as a conditional model. This means it predicts the effects of DNA sequences for one specific condition—such as a specific cell type or a disease state—at a time.

This architecture allows the model to learn how the same stretch of DNA is read differently depending on the environment. Because it treats the biological condition as a primary input at every stage, it can distinguish between closely related cell types. This precision helped the model identify nearly twice as many variants that change gene expression compared to previous state-of-the-art tools.

One unique feature of this design is called cross-modality imputation. This allows the model to predict missing experimental data based on what is already known. For instance, if a scientist provides a measurement of single-cell gene activity, the model can predict other unmeasured properties like chromatin accessibility. This helps researchers plan more efficient experiments by identifying which data points are most worth collecting next.

Predicting unseen disease states

AlphaGenome 2 is designed to generalize to new experiments without needing to be retrained every time. Scientists can upload a small snapshot of gene activity from just a few dozen cells to

Expanding the research toolkit

Google DeepMind plans to integrate AlphaGenome 2 into its agentic development platform, Google Antigravity. This upcoming release will introduce AlphaGenome 2 Skills, which are designed to work alongside other specialized models like Gemini 4. By combining these tools, the platform aims to help scientists accelerate their research and make new discoveries more efficiently.

The team is also looking for ways to highlight the practical applications of this technology. They are encouraging the scientific community to share interesting use cases that others can learn from. Researchers who submit their projects by November 30 may be eligible for additional credits and quota to support their work.

Important safety and community notes

It is important to note that AlphaGenome 2 is currently intended for research purposes only. The model has not been validated or approved for any clinical or diagnostic uses. This distinction ensures that the tool remains a resource for fundamental biology and disease discovery rather than a direct medical tool.

The development of this model involved significant collaboration and communication expertise. The team expressed gratitude to several individuals, including Evie Gray, Adriana Fernandez Lara, and Alex Wilkins, who helped communicate the findings of this work. They also continue to seek community input to help shape the future roadmap of the project.

Sources and links

First reported by Google DeepMind. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news