DeepMind Releases AlphaGenome 2 to Predict DNA Variant Effects by Cell Type

Google DeepMind has released AlphaGenome 2, a DNA-sequence model designed to predict how genetic variants affect gene activity in specific cell types and cell states, a step toward linking disease-associated mutations to the cells through which they act. The system builds on AlphaGenome, which can process long DNA sequences, and AlphaGenome Atlas, a precomputed map of predicted effects for billions of single-letter variants in the human genome. AlphaGenome 2 adds conditionality: users can specify a tissue, cell type or cellular state, allowing the model to focus predictions on a particular biological context rather than averaging signals across a tissue. DeepMind trained the model on human and mouse single-cell datasets totaling more than 107 million cells across 68 tissues. It predicts genomic readouts including gene expression, chromatin accessibility, histone modifications and DNA methylation. In testing on TenK10K, a single-cell dataset covering 28 primary immune-cell types, AlphaGenome 2 identified about 30 percent of relevant causal genetic variants, compared with 16 percent for a first-generation model fine-tuned on single-cell data. It correctly predicted the direction of expression effects for 43 percent of tested variants versus 23 percent. The model can also accept user-provided single-cell expression data to predict variant effects in previously unseen cell states without retraining. DeepMind highlighted a retrospective analysis of an inflammatory bowel disease study from Francis Crick Institute researcher James Lee and colleagues. Their experiments had linked a noncoding variant to overactivity of the ETS2 gene in intestinal macrophages. AlphaGenome 2 independently recovered the key variant and predicted its effect mainly in that cell type, a result Lee said matched laboratory findings. DeepMind noted the AI analysis took about 30 minutes, while the original work involved broader experimental validation over more than a year. The company also provided precomputed predictions for variants associated with many traits across hundreds of cell and tissue conditions, helping researchers prioritize candidate genes and cell types. AlphaGenome 2 is open to noncommercial research through the AlphaGenome API and is in private preview for commercial use. DeepMind says predictions remain hypotheses and are not validated for clinical or diagnostic use.