Robust annotation and discovery of novel cell types in single-cell ATAC-seq data through cross-modal reference alignment.

CARA accurately transfers cell-type knowledge from scRNA-seq to scATAC-seq, outperforming baseline methods and robustly detecting novel cell types across diverse datasets.

• Why it matters: Precise cell-type annotation in scATAC-seq is hindered by data sparsity, high dimensionality, limited reference labels, and batch effects, impeding understanding of regulatory element accessibility and cell differentiation.
• What they did: The study introduces CARA, a cross-modal Bayesian framework that combines pretraining on scRNA-seq with semi-supervised learning on scATAC-seq, enabling annotation, batch correction, and novel cell-type detection across multiple modalities.
• The result: CARA consistently outperforms existing methods, preserves lineage structures such as hematopoietic trajectories, and identifies rare or aberrant populations, facilitating discovery and mechanistic insights in single-cell analyses.