Batch correction methods used in single-cell RNA sequencing analyses are often poorly calibrated.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables large-scale biological studies.
• Combining scRNA-seq data across experiments or runs is crucial but challenged by batch effects.
• Existing batch correction methods for scRNA-seq data vary in effectiveness.

Purpose of the Study:

• To evaluate and compare the performance of eight widely used batch correction methods for scRNA-seq data.
• To introduce a novel approach for quantifying batch correction artifacts at both fine and cluster scales.
• To identify the most reliable method for scRNA-seq batch correction.

Main Methods:

• Comparative analysis of eight batch correction algorithms: MNN, SCVI, LIGER, Combat, ComBat-seq, BBKNN, Seurat, and Harmony.
• Development of a novel quantitative framework to assess data alteration during batch correction.
• Evaluation of method performance based on cell-to-cell distances and cluster-level effects.

Main Results:

• Several methods (MNN, SCVI, LIGER, Combat, ComBat-seq, BBKNN, Seurat) introduce detectable artifacts and alter data significantly.
• Poor calibration was observed in many tested batch correction methods.
• Harmony demonstrated consistent high performance across all evaluation metrics and methodologies.

Conclusions:

• Harmony is the only recommended method for batch correction in scRNA-seq data due to its minimal introduction of artifacts.
• Users should exercise caution with other popular batch correction tools, as they may compromise data integrity.
• The novel evaluation framework provides a robust way to assess batch correction performance.