AlphaGenome 2 narrows variant prediction to single cells: nearly twice as many causal variants recovered on an immune cell atlas
Synopsis
AlphaGenome 2 moves prediction of DNA variant effects from bulk-tissue averages down to specific cell types and cell states, recasting the model as a conditional one that takes a target biological condition as input. On the TenK10K single-cell map of 28 primary immune cell types, it identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data, and 43% versus 23% for directional gene effects. Trained on more than 107 million individual cell profiles across 68 tissues, it is available through the AlphaGenome API for non-commercial research.
Interpretation
The unit of DNA variant effect prediction shrinks from bulk-tissue averages to individual cell types and cell states, so researchers can see regulatory variants that act only in particular cells. The previous AlphaGenome took a one-million-letter DNA sequence and predicted thousands of experiments across many tissues in parallel all at once; that parallel approach spread the model's capacity thin and made subtle differences between closely related cell types hard to capture. On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data, and 43% versus 23% for directional gene effects.
The model treats cell type as an input condition rather than sharing one parallel output across all cells, so how the same stretch of DNA is read differently across cell types can be learned separately. The conditional model receives a one-million-letter DNA sequence together with a target biological condition at every stage of its architecture; the condition is a tissue, cell type or cell state, drawn from a reference catalog or generated on the fly from a scientist's own single-cell experiment by a cell encoder. Training used human and mouse bulk and single-cell data totaling more than 107 million individual cell profiles across 68 tissues; the model predicts gene expression, histone modifications and chromatin accessibility, and adds DNA methylation as a readout.
The model adapts to new single-cell experiments without retraining, by compressing a gene-activity snapshot from a few dozen cells into a cell state representation used as the condition. On unseen single-cell data from a COVID-19 mRNA vaccine study, the model distinguished vaccine-activated immune T cells from resting ones and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly when cells are in the activated state, an effect that would be missed in unactivated cells. On unseen cell states, AlphaGenome 2 can match or outperform models trained or fine-tuned directly on those datasets.
Perspective
The demonstrated scope is human and mouse bulk and single-cell data across 68 tissues and more than 107 million individual cell profiles, with readouts including gene expression, histone modifications, chromatin accessibility and DNA methylation; AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.
Performance on unseen cell states is described as matching or outperforming models trained or fine-tuned directly on those datasets, not as better in every case; precomputed variant scores give the direction of gene up- or down-regulation only when the model is confident.