A comprehensive benchmarking study on computational tools for cross-omics label transfer from single-cell RNA to ATAC data

With continuous progress of single-cell chromatin accessibility profiling techniques, scATAC-seq has become more commonly used in investigating regulatory genomic regions and their involvement in developmental, evolutionary, and disease-related processes. At the same time, accurate cell type annotation plays a crucial role in comprehending the cellular makeup of complex tissues and uncovering novel cell types. Unfortunately, the majority of existing methods primarily focus on label transfer within scRNA-seq datasets and only a limited number of approaches have been specifically developed for transferring labels from scRNA-seq to scATAC-seq data. Moreover, many methods have been published for the joint embedding of data from the two modalities, which can be used for label transfer by adding a classifier trained on the latent space. Given these available methods, this study presents a comprehensive benchmarking study evaluating 27 computational tools for scATAC-seq label annotations through tasks involving single-cell RNA and ATAC data from various human and mouse tissues. We found that when high-quality paired data were available to transfer labels across unpaired data, Bridge and GLUE were the best performers; otherwise, bindSC and GLUE achieved the highest prediction accuracy overall. All these methods were able to use peak-level information instead of purely relying on the gene activities from scATAC-seq. Furthermore, we found that data imbalance, cross-omics dissimilarity on common cell types, data binarization, and the introduction of semi-supervised strategy usually had negative impacts on model performance. In terms of scalability, we found that the most time and memory efficient methods were Bridge and deep learning-based algorithms like GLUE. Based on the results of this study, we provide several suggestions for future methodology development.