Everyone carries the same set of DNA, yet the genes switched on and off in cardiomyocytes and anti-infection immune cells are worlds apart. The same mutation can also lead to opposite outcomes depending on which cell it lands in—and it is this complexity that has long stalled disease gene mapping. On November 11 local time, Google DeepMind launched AlphaGenome 2, taking the answer to "what will a genetic variant do" down to individual cell types and states.
Fed 107 million single cells, its hit rate nearly doubled
The previous generation of AlphaGenome focused only on overall molecular activity across different tissues and struggled with subtle differences within specific cells. The new model adopts a conditional architecture, taking target tissue, cell type, or state as input. It ingested more than 107 million single-cell atlases from humans and mice, covering 68 tissues, and its ability to identify relevant variants is nearly twice that of the previous generation. In the TenK10K test across 28 types of primary immune cells, it detected disease-causing variants at a rate of 30%, compared with only 16% for the fine-tuned previous generation; its accuracy in predicting the direction of expression changes also rose from 23% to 43%.
It can also make predictions for cells it was not trained on
Even more practical is its generalization: after uploading a small amount of single-cell RNA sequencing data, the model can analyze cell types and states it has never seen. In COVID-19 mRNA vaccine data, it distinguished activated and resting immune T cells, and showed that a variant linked to severe disease had a stronger effect when cells were activated. It also adds predictions for histone modifications, chromatin accessibility, and DNA methylation, which the previous generation lacked, and supports cross-modal inference: given only gene activity, it can infer unmeasured metrics. The model is open to non-commercial research through the AlphaGenome API, while commercial use goes through a private preview. Google stresses that it is for research only and has not been approved for clinical or diagnostic use.
