Benchmarking automated cell type annotation tools for single-cell ATAC-seq data.

Area of Science:

• Genomics and Bioinformatics
• Single-cell analysis
• Computational biology

Background:

• Single-cell chromatin accessibility (scATAC-seq) is crucial for studying regulatory genomics in development, evolution, and disease.
• Accurate cell type annotation is essential for understanding tissue composition and identifying novel cell types.
• Existing automated annotation methods are primarily designed for scRNA-seq and their adaptability to scATAC-seq is unclear.

Purpose of the Study:

• To benchmark the performance of five automated cell type annotation methods for scATAC-seq data.
• To evaluate classification accuracy and scalability of these methods across diverse human and mouse tissues.
• To assess method robustness under varying data sizes, mislabeling rates, sequencing depths, and unique cell types.

Main Methods:

• Evaluation of five scATAC-seq annotation methods: Bridge integration, Conos, scJoint, scGCN, and Seurat v3.
• Benchmarking using publicly available scATAC-seq datasets from mouse and human tissues (brain, lung, kidney, PBMC, BMMC).
• Analysis of performance across different data sizes, mislabeling rates, sequencing depths, and unique cell type proportions.

Main Results:

• Bridge integration showed the best overall accuracy and robustness, performing well across various data conditions.
• Conos was the most computationally efficient but yielded the lowest prediction accuracy.
• scJoint performed poorly on complex datasets but better on simpler annotations; scGCN and Seurat v3 had moderate performance, with scGCN being time-consuming and poor for unique cell types.

Conclusions:

• Bridge integration is a highly accurate and robust method for scATAC-seq cell type annotation.
• Method selection should consider trade-offs between accuracy, scalability, and computational efficiency.
• Further development is needed for methods to effectively annotate cell types unique to scATAC-seq data.