Influence of single-cell RNA sequencing data integration on the performance of differential gene expression analysis

Large-scale comprehensive single-cell experiments are often resource-intensive and require the involvement of many laboratories and/or taking measurements at various times. This inevitably leads to batch effects, and systematic variations in the data that might occur due to different technology platforms, reagent lots, or handling personnel. Such technical differences confound biological variations of interest and need to be corrected during the data integration process. Data integration is a challenging task due to the overlapping of biological and technical factors, which makes it difficult to distinguish their individual contribution to the overall observed effect. Moreover, the choice of integration method may impact the downstream analyses, including searching for differentially expressed genes. From the existing data integration methods, we selected only those that return the full expression matrix. We evaluated six methods in terms of their influence on the performance of differential gene expression analysis in two single-cell datasets with the same biological study design that differ only in the way the measurement was done: one dataset manifests strong batch effects due to the measurements of each sample at a different time. Integrated data were visualized using the UMAP method. The evaluation was done both on individual gene level using parametric and non-parametric approaches for finding differentially expressed genes and on gene set level using gene set enrichment analysis. As an evaluation metric, we used two correlation coefficients, Pearson and Spearman, of the obtained test statistics between reference, test, and corrected studies. Visual comparison of UMAP plots highlighted ComBat-seq, limma, and MNN, which reduced batch effects and preserved differences between biological conditions. Most of the tested methods changed the data distribution after integration, which negatively impacts the use of parametric methods for the analysis. Two algorithms, MNN and Scanorama, gave very poor results in terms of differential analysis on gene and gene set levels. Finally, we highlight ComBat-seq as it led to the highest correlation of test statistics between reference and corrected dataset among others. Moreover, it does not distort the original distribution of gene expression data, so it can be used in all types of downstream analyses.