Google DeepMind Unveils AlphaGenome 2 to Map Disease at the Cellular Level
AlphaGenome 2 predicts how genetic mutations impact specific cell types using a dataset of 100 million cell profiles, offering a finer-grained view for drug discovery than its predecessor.
Researchers at Google DeepMind have introduced AlphaGenome 2, a sophisticated AI model designed to predict how changes in DNA affect different types of cells. The model aims to solve a complex biological puzzle: why the same genetic mutation can cause disease in one organ while leaving others unaffected. By providing a high-resolution view of cellular behavior, AlphaGenome 2 could significantly accelerate the development of targeted medical treatments.
Beyond Static DNA Maps
While almost every cell in the human body contains the same DNA, different cells—such as heart, skin, or immune cells—activate different genes depending on their function and state. AlphaGenome 2 is the successor to the original AlphaGenome model released last year. While the first version provided a broader look at the genome, AlphaGenome 2 focuses on the "cellular roots" of disease, capturing the specific patterns of gene expression that vary across cell types.
According to Google DeepMind, the new model identifies nearly twice as many genetic variants that change gene expression in cells compared to the previous state-of-the-art model. This granularity is essential for drug discovery, as researchers need to target specific proteins within specific cell types to minimize side effects and increase efficacy.
Massive Data and "Zero-Shot" Predictions
The breakthrough was made possible by training the model on a massive dataset of more than 100 million individual cell profiles across 68 different tissues. DeepMind says this effectively transforms static "cell atlases"—large repositories of biological data—into a predictive engine for human genetics.
A notable feature of AlphaGenome 2 is its ability to make predictions about entirely new types of cells and biological states without requiring additional training. It can also identify gaps in existing experimental data, helping scientists determine which areas of the lab require more investigation.
"When you develop drugs, you’re targeting specific proteins and specific cell types," explained Žiga Avsec, DeepMind’s genomics lead. He noted that the specificity provided by AlphaGenome 2 is vital for moving from a general understanding of genetics to actionable medical solutions.
The Path to Virtual Cells
Despite its advancements, AlphaGenome 2 is not yet replacing its predecessor. The model’s high level of detail comes at a significant computational cost, making it more "expensive" to run than the original AlphaGenome. For instance, DeepMind’s comprehensive Atlas of every possible DNA letter change, released in September, was produced using the older model because the computational demands of AlphaGenome 2 would have been impractical for such a massive scale.
Looking ahead, DeepMind plans to expand the model's training data to include more cell states and potentially data from multiple species. Long-term, these models are expected to contribute to the creation of "virtual cells"—digital simulations that can mimic complex biological systems.
Google is making AlphaGenome 2 available for non-commercial research via the AlphaGenome API starting today, with a limited "private preview" available for commercial research. This release follows a broader trend of AI labs tackling major scientific hurdles, with competitors like Anthropic exploring drug design and OpenAI focusing on complex mathematical problems.
