AlphaGenome 2 shrinks predictions to single cells, flagging nearly twice as many expression-changing variants as its predecessor
Synopsis
AlphaGenome 2 is Google's next-generation DNA sequence model, moving prediction from tissue averages down to individual cell types and cell states while letting researchers name the target condition. The earlier AlphaGenome predicted thousands of experiments in parallel from one sequence, spreading model capacity thin and blurring closely related cell types. The new conditional model receives the target tissue, cell type or state at every stage of its architecture, was trained on more than 107 million single-cell profiles across 68 tissues, and identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model.
Interpretation
The effect of a DNA variant on gene expression can now be predicted for a specific cell type and cell state, not only as a cross-tissue average. The earlier AlphaGenome started from a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues all at once. According to Google, AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model.
The model treats cell type as an input condition instead of letting all conditions share one parallel output. A conditional model means the target tissue, cell type or state is fed into the network alongside the DNA sequence at every stage of the architecture. Training used bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles across 68 tissues.
The same conditional design lets the model fill in experimental readouts that were not measured. Cross-modality imputation means that from one measurement, such as single-cell gene activity, the model predicts unmeasured properties such as chromatin accessibility. Google says this can help researchers decide which experiments to run next, but the summary gives no comparison numbers for imputation accuracy.
Perspective
The demonstrated scope is variant-effect prediction on human and mouse bulk and single-cell data, 68 tissues and more than 107 million single-cell profiles, plus cross-modality imputation of unmeasured properties from one measurement; the intended readers are genetics and single-cell researchers doing variant interpretation and experiment prioritization.
The nearly twofold figure is an overall comparison from the releasing organization, and the summary does not break it down by cell type, tissue or evaluation set; prediction for unseen cell and disease states is described only qualitatively.
