scCDAN: Constraint domain adaptation network for cell type annotation across single cell RNA sequencing data

The development of single cell sequencing technology enables researchers to investigate tissue heterogeneity at the cellular level, with cell type annotation being the core of its analysis. Although various computational methods have been proposed, batch effects remain a major issue affecting annotation accuracy. In this study, a constraint domain adaptation network, named scCDAN, is proposed to address batch effects that lead to poor cell type annotation. Firstly, a domain alignment module is used to align the distributions of source and target domain data through adversarial learning strategies to reduce domain discrepancies. Furthermore, to achieve finer granularity in differentiating cell types, a category boundary constraint module is designed to regulate the positional relationships between cells of the same type and those of different types in the feature space. scCDAN combines domain alignment and category boundary constraint modules to directly and indirectly address batch effects, thereby improving the accuracy of cell type annotation. The effectiveness of scCDAN is validated on simulated, cross-platforms, and cross-species datasets. Experimental results demonstrate that scCDAN outperforms comparative methods in both cell type annotation and batch effects correction.