Back to news
researchGoogleDeepMind2026-11-11

Google DeepMind releases AlphaGenome 2, identifying nearly twice as many variants that change gene expression

Google DeepMind released AlphaGenome 2 on November 11, moving from parallel prediction to a conditional architecture that predicts sequence effects for a single cell type and identifies close to twice as many expression-altering variants as the previous best model.

Google DeepMind released AlphaGenome 2 on November 11, a new DNA sequence model that predicts how DNA sequences and their variants affect specific human cell types, giving researchers a new tool for mining cellular drivers of disease. Every human cell carries the same DNA blueprint, but different cell types switch different genes on to perform their functions. Google previously shipped the original AlphaGenome, which predicts variant effects on broad tissue-level molecular activity, and AlphaGenome Atlas covering predictions for all nine billion single-letter variants across the human genome.

AlphaGenome 2 narrows from broad tissue averages down to individual cell types and states, and rebuilds the model architecture, producing a step change in variant prediction performance and enabling predictions for cell and disease states the model has not seen. Compared with the previous best model, AlphaGenome 2 identifies close to twice as many variants capable of changing gene expression within cells. The model is trained on more than 100 million single-cell atlases across 68 tissue types and draws on over 107 million single-cell profiles, and is open to non-commercial research through the AlphaGenome API with commercial research in private preview.

On architecture, the original model could read DNA sequences up to a million letters at once and synchronously predict thousands of experimental properties across many tissues, but that parallel approach spreads compute and struggles to capture fine differences between closely related cell types. AlphaGenome 2 drops parallel prediction for a conditional model that predicts DNA effects for one condition at a time, feeding the target biological condition as input at every stage of the architecture so it can learn how the same DNA segment reads differently across cell types. The conditional architecture also supports cross-modal data imputation, predicting missing experimental measurements by separating cellular state from outputs, so inputting only single-cell gene activity lets it predict unmeasured properties such as chromatin accessibility. The model better explains gene expression differences between cell types within the same tissue in genomic regions not covered by training data, and adds DNA methylation prediction.

谷歌DeepMindAlphaGenome 2DNA单细胞生物模型