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

Area of Science:

• Genomics
• Computational Biology
• Single-cell Analysis

Background:

• Single-cell chromatin accessibility (scATAC-seq) is vital for understanding gene regulation in development and disease.
• Accurate cell type annotation is essential for interpreting complex tissue cellularity.
• Limited methods exist for transferring labels from single-cell RNA sequencing (scRNA-seq) to scATAC-seq data.

Purpose of the Study:

• To comprehensively benchmark 27 computational tools for scATAC-seq cell type annotation.
• To evaluate tool performance using both paired and unpaired scRNA-seq and scATAC-seq data from human and mouse tissues.
• To identify optimal methods for cross-modal label transfer and assess factors influencing performance.

Main Methods:

• Benchmarking of 27 computational tools for scATAC-seq label annotation.
• Utilized paired and unpaired single-cell RNA and ATAC sequencing data from diverse human and mouse tissues.
• Evaluated performance based on prediction accuracy, scalability (time and memory efficiency), and impact of data characteristics.

Main Results:

• Bridge and GLUE were top performers when high-quality paired data were available for label transfer.
• bindSC and GLUE showed the highest prediction accuracy overall for unpaired data.
• Factors like data imbalance, cross-omics dissimilarity, and semi-supervised strategies negatively impacted performance.
• Bridge and deep-learning methods like GLUE demonstrated superior scalability.

Conclusions:

• Bridge, GLUE, and bindSC are recommended for scATAC-seq cell type annotation, depending on data availability.
• Peak-level information is crucial for accurate annotation, beyond gene activity.
• Future scATAC-seq methodology development should consider data characteristics and scalability.