Google DeepMind announced AlphaGenome 2, a successor to its previously released genome-predicting model designed to show how genetic variants affect individual cell types and biological states. While most human cells contain identical DNA, different cell types activate and deactivate specific genes depending on their function and condition, meaning the same genetic variant can produce distinct effects across different cells.
The new model was trained on more than 100 million individual cell profiles across 68 tissues to provide enhanced precision compared to broader genome prediction tools. According to the company, AlphaGenome 2 identifies nearly twice as many variants capable of altering gene expression in cells compared to the previous state-of-the-art approach. The researchers also found that the model can make predictions about previously unseen cell types and biological states without requiring retraining, and it can identify gaps in experimental datasets that might guide future laboratory investigations.
The increased specificity offered by AlphaGenome 2 addresses a practical need in medical research and drug development. As DeepMind’s genomics lead explained, targeted treatments require understanding effects on specific proteins and cell types, making this granular level of analysis essential for disease treatment strategies. Genetic variants identified in large population studies can now be examined at the cellular level to understand potential disease connections.
The development of such models has only recently become feasible through technological advances that enable high-resolution capture of individual cells and their states, creating the large datasets necessary for training. DeepMind plans to enhance its genetics models by generating additional data, particularly for underrepresented cell states, and future versions may incorporate data from multiple species. Long-term applications could include digital simulations designed to replicate biological systems.
Google is making AlphaGenome 2 available for non-commercial research through the AlphaGenome API, with limited private preview access for commercial applications. The more detailed predictions come at a higher computational cost than its predecessor, making it impractical for generating comprehensive genome-wide atlases like the one DeepMind released earlier this year.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI