Benchmarking methods for detecting differential states between conditions from multi-subject single-cell RNA-seq...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) allows simultaneous transcriptome quantification of thousands of cells.
• Multisubject, multicondition experiments are crucial for identifying cell-type-specific differential states (DS).
• Naive scRNA-seq analysis methods can yield false positives due to biological replicate variation (pseudoreplicate bias).

Purpose of the Study:

• To comprehensively compare 18 existing methods for identifying DS changes in multisubject scRNA-seq data.
• To evaluate the performance of different statistical approaches in handling biological variability.
• To identify the most reliable methods for accurate differential state analysis.

Main Methods:

• Comparison of 18 distinct statistical methods for differential state analysis.
• Evaluation of pseudobulk methods, mixed models, and naive single-cell approaches.
• Assessment of methods accounting for subjects as random effects versus latent variables.

Main Results:

• Pseudobulk methods generally demonstrated superior performance in multisubject scRNA-seq analysis.
• Pseudobulk and mixed models significantly outperformed naive methods by effectively modeling subjects.
• Naive methods, while sensitive, produced a high number of false positives; latent variable modeling did not improve their performance.

Conclusions:

• Pseudobulk methods are recommended for robust differential state analysis in multisubject scRNA-seq studies.
• Accurate modeling of biological subjects is critical to avoid pseudoreplicate bias and reduce false positives.
• The study provides a benchmark for selecting appropriate statistical tools for complex scRNA-seq experimental designs.