Detection and removal of barcode swapping in single-cell RNA-seq data

Multiplexing is a widely-used procedure that allows multiple DNA libraries to be pooled together for efficient sequencing. However, recent reports suggest that the DNA barcodes that label different libraries can “swap” on patterned flow-cell Illumina sequencing machines, including the HiSeq 4000, HiSeq X, and NovaSeq, thereby mislabelling molecules [1, 2]. This may compromise many types of -omic assays, but it is particularly problematic for single-cell RNA-seq (scRNA-seq), where many libraries are multiplexed together. A number of widely used plate-based scRNA-seq library preparation methods isolate and process individual cells in wells of a microwell plate, before performing library preparation in parallel [3]. A unique combination of sample barcodes labels the library of each cell, typically with one barcode at each end of a cDNA molecule. One barcode provides a row index for each cell on the microwell plate and the other barcode provides a column index. Barcode swapping therefore moves transcripts between cells. We generated a dataset (see Supplementary Files, “Richard data”) where two plates of single-cell libraries were multiplexed for sequencing on the HiSeq 4000 using two mutually exclusive barcode sets. We expect to only observe reads