Barcode identification for single cell genomics

Single-cell sequencing experiments use short DNA barcode ‘tags’ to identify reads that originate from the same cell. In order to recover single-cell information from such experiments, reads must be grouped based on their barcode tag, a crucial processing step that precedes other computations. However, this step can be difficult due to high rates of mismatch and deletion errors that can afflict barcodes. Here we present an approach to identify and error-correct barcodes by traversing the de Bruijn graph of circularized barcode k-mers. This allows for assignment of reads to consensus fingerprints constructed from k-mers, and we show that for single-cell RNA-Seq this improves the recovery of accurate single-cell transcriptome estimates. Availability and implementation Freely available source code is available at Github: https://github.com/pachterlab/Sircel This Github repository also contains iPython notebooks to reproduce all analysis presented in this paper. Introduction Tagging of sequencing reads with short DNA barcodes is a common experimental practice that enables a pooled sequencing library to be separated into biologically meaningful partitions. This technique is in the cornerstone of many single-cell sequencing experiments, where reads originating from individual cells are tagged with cell-specific barcodes; as such, the first step in any single-cell sequencing experiment involves separating reads by barcode to recover single-cell profiles (Svensson et al., 2017; Trapnell, 2015); (Klein et al., 2015). For example, in the Drop-Seq protocol, which is a popular microfluidic-based single-cell experimental platform, DNA barcodes are synthesized on a solid bead support, using split-and-pool DNA synthesis (Macosko et al., 2015). Similar split-and-pool barcoding strategies are used in other single-cell sequencing assays such as Seq-Well (Gierahn et al., 2017) and Split-seq (Rosenberg et al., 2017). One consequence of this synthetic technique is that deletion errors are extremely prevalent; by some estimates 25% of all barcode sequences observed contain at least one deletion (Macosko et al., 2015). Ignoring such errors can therefore dramatically lower the number of usable reads in a dataset, while incorrectly grouping reads together can confound single cell analysis. Current approach to “barcode calling”, the process of grouping reads together by barcode, use simple heuristics to first identify barcodes that are likely to be uncorrupted, and then “error correct” remaining barcodes to increase yields. However the complex . CC-BY-NC 4.0 International license peer-reviewed) is the author/funder. It is made available under a The copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/136242 doi: bioRxiv preprint first posted online May. 9, 2017;