Abstract
Annotating cell types is a critical step in single-cell RNA sequencing (scRNA-seq) data analysis. Some supervised or semi-supervised classification methods have recently emerged to enable automated cell type identification. However, comprehensive evaluations of these methods are lacking. Moreover, it is not clear whether some classification methods originally designed for analyzing other bulk omics data are adaptable to scRNA-seq analysis. In this study, we evaluated ten cell type annotation methods publicly available as R packages. Eight of them are popular methods developed specifically for single-cell research, including Seurat, scmap, SingleR, CHETAH, SingleCellNet, scID, Garnett, and SCINA. The other two methods were repurposed from deconvoluting DNA methylation data, i.e., linear constrained projection (CP) and robust partial correlations (RPC). We conducted systematic comparisons on a wide variety of public scRNA-seq datasets as well as simulation data. We assessed the accuracy through intra-dataset and inter-dataset predictions; the robustness over practical challenges such as gene filtering, high similarity among cell types, and increased cell type classes; as well as the detection of rare and unknown cell types. Overall, methods such as Seurat, SingleR, CP, RPC, and SingleCellNet performed well, with Seurat being the best at annotating major cell types. Additionally, Seurat, SingleR, CP, and RPC were more robust against downsampling. However, Seurat did have a major drawback at predicting rare cell populations, and it was suboptimal at differentiating cell types highly similar to each other, compared to SingleR and RPC. All the code and data are available from https://github.com/qianhuiSenn/scRNA_cell_deconv_benchmark.