Bridging scRNA-seq and scATAC-seq With Graph Domain Adaptation for Cell Type Annotation

Accurate cell type annotation of single-cell ATAC-seq (scATAC-seq) data is crucial for downstream analysis but remains challenging due to the modality gap between scATAC-seq and scRNA-seq. Cross-omics methods typically transfer annotations from scRNA-seq References via label propagation, but often suffer from overcorrection, which can obscure biological variability. To address the challenges associated with cross-omics alignment in scATAC-seq cell type annotation, we propose scGDA, a task-specific graph domain adaptation (GDA) framework that integrates modality-specific graph modeling, structural representation learning, and cross-modality alignment for scRNA-seq to scATAC-seq label transfer. In particular, scGDA independently constructs k-nearest neighbor graphs for both scRNA-seq and scATAC-seq datasets and employs dual graph neural network modules to capture both global and local structural features within each omics. These omics-specific representations are subsequently aligned within a shared embedding space, enabling effective knowledge transfer for cell type annotation. Extensive experiments across multiple benchmark datasets demonstrate that scGDA achieves competitive or superior accuracy compared with existing methods. The source code is available at https://anonymous.4open.science/r/scGDA-4B05.