Google DeepMind unveils AlphaGenome 2 AI model for disease research

Model predicts effects in specific cells

The new system offers a finer view than its predecessor by analyzing over 100 million individual cell profiles across 68 different tissues. While nearly every human cell carries the same DNA, different types like heart muscle or immune cells switch genes on and off based on their function. This model captures those differences to predict how genetic variants impact individual cell types and states without needing retraining for new biological conditions.

Data training enables broader predictions

AlphaGenome 2 identifies nearly twice as many variants that change gene expression compared to the previous state-of-the-art model released last year. The technology allows static cell atlases to function as a predictive engine for human genetics by mapping vast datasets needed for high-resolution analysis. Researchers noted this specificity is vital when developing drugs that target specific proteins and specific cell types to treat particular diseases.

Future work expands data sources

Google DeepMind intends to enhance its genetics models through additional data generation, covering cell states currently underrepresented in existing repositories. Lead researcher Žiga Avsec stated future versions could be trained on data across multiple species to provide a broader view of biology. The team is also working to identify gaps in experimental data so scientists can decide what to investigate next in the lab.