Google DeepMind researchers have introduced a new artificial intelligence model designed to determine how changes in DNA affect different types of cells. Called AlphaGenome 2, the model offers a more detailed way to analyse the effects of genetic variations at the cellular level. This could provide important clues about the causes of diseases and help guide the development of new treatments.
AlphaGenome 2 is the successor to AlphaGenome, DeepMind’s genome-prediction model launched last year. Although almost every cell in the human body carries the same DNA, different cell types — from heart muscle and skin cells to immune cells fighting infections — activate and deactivate different genes. These patterns can also change depending on a cell’s type and state. As a result, the same genetic variant can have very different effects in different cells.
The new model is designed to capture these differences by predicting how genetic variants affect individual cell types and biological states. In a blog post published on Wednesday, Google said AlphaGenome 2 provides a more fine-grained view than broader models, potentially revealing disease-related effects in specific cells that might otherwise go unnoticed.
To achieve this level of detail, AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 tissues. DeepMind says the model effectively turns static cell atlases — large databases that map and catalogue individual cells — into a predictive engine for human genetics. According to the company, AlphaGenome 2 identifies variants capable of changing gene expression in cells nearly twice as often as its previous leading model.
Researchers also found that AlphaGenome 2 can make predictions about new cell types and biological states without requiring retraining. The model can also identify gaps in experimental data, potentially helping scientists decide what to investigate next in the laboratory.
Adam Gayoso, a Google DeepMind research scientist involved in the project, told The Verge that the biggest breakthrough was connecting human genetics with cell biology. He said the model gives researchers a way to examine genetic variants identified in large populations, understand their potential effects at the cellular level and potentially link those effects to diseases.
Žiga Avsec, Google DeepMind’s head of genomics research, told The Verge that this level of detail is particularly important when trying to understand how to treat a specific disease. “When you’re developing drugs, you’re targeting specific proteins and specific cell types,” he said, explaining why such detailed analysis matters.
According to Avsec, advances in technologies capable of mapping individual cells and their states at high resolution, alongside the creation of the large datasets needed to train the model, have only recently made a system like AlphaGenome 2 possible. DeepMind is now working to improve its genetic models by generating more data, including information about cell states that are currently underrepresented.
Avsec also said future versions could be trained on data from multiple species, giving the models a broader understanding of biology. In the longer term, systems such as AlphaGenome 2 could contribute to efforts to create “virtual cells” — digital simulations designed to imitate biological systems.
Despite describing AlphaGenome 2 as the successor to AlphaGenome, Avsec said the new model is unlikely to replace the earlier version in the near future. More detailed predictions require greater computing power and are therefore more expensive to generate. These requirements would make it impractical to use AlphaGenome 2 to create the genome-wide Atlas published by DeepMind in September, which covers every possible single-letter DNA change in the human genome.
Google announced that AlphaGenome 2 is available from today through the AlphaGenome API for non-commercial research. The company has also launched a limited “private preview” for commercial research.
The launch is the latest example of Google and competing AI laboratories increasingly turning to artificial intelligence to tackle major challenges in science and medicine. Pushmeet Kohli, who leads DeepMind’s AI-for-science research, hinted that more significant scientific results are on the way. The organisation is also expected to open an automated materials science laboratory in the United Kingdom.
Meanwhile, Anthropic is experimenting with drug design and has given Claude access to a wet laboratory where physical experiments are conducted. OpenAI, for its part, has been solving a notable number of long-standing problems in mathematics and related fields.
