Assessment of batch-correction methods for scRNA-seq data with a new test metric

Single-cell transcriptomics is a versatile tool for exploring heterogeneous cell populations. As with all genomics experiments, batch effects can hamper data integration and interpretation. The success of batch effect correction is often evaluated by visual inspection of dimension-reduced representations such as principal component analysis. This is inherently imprecise due to the high number of genes and non-normal distribution of gene expression. Here, we present a k-nearest neighbour batch effect test (kBET, https://github.com/theislab/kBET ) to quantitatively measure batch effects. kBET is easier to interpret, more sensitive and more robust than visual evaluation and other measures of batch effects. We use kBET to assess commonly used batch regression and normalisation approaches, and quantify the extent to which they remove batch effects while preserving biological variability. Our results illustrate that batch correction based on log-transformation or scran pooling followed by ComBat reduced the batch effect while preserving structure across data sets. Finally we show that kBET can pinpoint successful data integration methods across multiple data sets, in this case from different publications all charting mouse embryonic development. This has important implications for future data integration efforts, which will be central to projects such as the Human Cell Atlas where data for the same tissue may be generated in multiple locations around the world. [Before final publication, we will upload the R package to Bioconductor]