Google DeepMind Releases AlphaGenome 2 for Single-Cell Variant Prediction
Google DeepMind has released AlphaGenome 2, a next-generation DNA sequence model that uses a conditional architecture to predict gene regulation and variant effects at the single-cell level. Trained on over 100 million single-cell profiles from humans and mice, it nearly doubles the accuracy of identifying disease-causing variants compared to previous models. This breakthrough bridges the gap between genomics and single-cell biology, allowing researchers to pinpoint disease-causing genetic variants within specific cell types and states. Such precision dramatically enhances our understanding of complex human traits, rare diseases, and immune responses. AlphaGenome 2 shifts from a parallel prediction approach to a conditional architecture where the biological condition serves as input at every stage. It currently supports non-commercial research via the AlphaGenome API and adds brand-new DNA methylation prediction capabilities alongside standard gene expression and chromatin accessibility metrics.
## BACKGROUND
Genomics models like the original AlphaGenome analyze DNA sequences to predict how regulatory elements control gene expression across broad tissue types. However, predicting how subtle single-nucleotide DNA variations alter gene expression within distinct, fine-grained cell types has long remained a complex computational hurdle.