Comprehensive evaluation of noise reduction methods for single-cell RNA sequencing data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional data requiring normalization and batch correction.
• Technical variations and biases obscure true biological signals in scRNA-seq datasets.
• Existing computational methods lack clear guidance for selecting appropriate procedures across diverse experimental scenarios.

Purpose of the Study:

• To comprehensively assess the performance of 28 scRNA-seq noise reduction methods.
• To evaluate method performance across 55 diverse simulated and real-world scenarios.
• To provide evidence-based recommendations for selecting appropriate scRNA-seq data processing techniques.

Main Methods:

• Performance evaluation of 28 scRNA-seq noise reduction procedures.
• Utilized simulated and real datasets encompassing 55 varied scenarios.
• Assessed factors including batch effects, cell population imbalance, data complexity, dropout rates, and library sizes.
• Employed quantitative metrics and low-dimensional embeddings for performance assessment.

Main Results:

• Identified technical and biological factors influencing the performance of each evaluated method.
• Recommended specific noise reduction methods tailored to different experimental scenarios.
• Highlighted a challenging scenario leading to overcorrection by most methods.
• Demonstrated variability in batch mixing and preservation of biological structures across methods.

Conclusions:

• A comprehensive guideline for selecting scRNA-seq noise reduction procedures is provided.
• The study identifies limitations and unsolved issues in current batch correction methodologies.
• There is an urgent need for novel metrics to assess batch correction effectiveness, particularly for subtle cell-type mixing.