A comprehensive comparison of supervised and unsupervised methods for cell type identification in single-cell RNA-seq

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cellular heterogeneity.
• Accurate cell type identification is a fundamental challenge in scRNA-seq data analysis.
• Existing computational methods are broadly classified into supervised and unsupervised approaches.

Purpose of the Study:

• To comprehensively evaluate the performance of supervised and unsupervised cell type identification methods.
• To investigate the impact of various factors (e.g., dataset size, sequencing depth, batch effects, reference quality) on method performance.
• To provide guidelines for selecting appropriate cell typing methods based on scientific objectives and data characteristics.

Main Methods:

• Performance assessment of 8 supervised and 10 unsupervised methods.
• Utilized 14 diverse public scRNA-seq datasets across different tissues, species, and protocols.
• Analyzed the influence of factors like cell count, cell type number, sequencing depth, batch effects, reference bias, population imbalance, novel cell types, and computational efficiency.

Main Results:

• Supervised methods generally showed superior performance compared to unsupervised methods across most scenarios.
• Unsupervised methods demonstrated comparable or better performance in identifying unknown or novel cell types.
• The efficacy of supervised methods is highly dependent on the quality and relevance of the reference dataset, with suboptimal references diminishing their advantage.

Conclusions:

• The choice between supervised and unsupervised methods depends on the specific research question and the properties of the scRNA-seq dataset.
• Reference dataset quality is a critical determinant of supervised method success.
• The study provides a framework for method selection and introduces a reusable R pipeline for future evaluations.