scSplit tool: genotype-free demultiplexing of pooled single-cell RNA-seq

In brief

A variety of methods have been developed to demultiplex pooled samples in a single cell RNA sequencing (scRNA-seq) experiment which either require hashtag barcodes or sample genotypes prior to pooling. We introduce scSplit which utilizes genetic differences inferred from scRNA-seq data alone to demultiplex pooled samples. scSplit also enables mapping clusters to original samples. Using simulated, merged, and pooled multi-individual datasets, we show that scSplit prediction is highly concordant with demuxlet predictions and is highly consistent with the known truth in cell-hashing dataset.

Highlights

scSplit is an accurate, fast, and computationally efficient method for demultiplexing of individual cells from pooled samples from scRNA-seq.
The new tool can be used without needing sample genotypes prior to mixing.
Mainly designed for droplet-based scRNA-seq but can also be used for scRNA-seq data generated from other types of protocols.
The full article can be found at https://doi.org/10.1186/s13059-019-1852-7.

Introduction

Using single-cell RNA sequencing (scRNA-seq) to cell biology at cellular level provides greater resolution than “bulk” level analyses, thus allowing more refined understanding of cellular heterogeneity. For example, it can be used to cluster cells into sub-populations based on their differential gene expression, so that different fates of cells during development can be discovered. Droplet-based scRNA-seq (for example Drop-Seq or 10X Genomics Systems) allows profiling large numbers of cells for sequencing by dispersing liquid droplets in a continuous oil phase in an automated microfluidics system, and as a result is currently the most popular approach to scRNA-seq despite a high cost per run. Methods that lower the per sample cost of running scRNA-seq are required in order to scale this approach up to a population scale. An effective method for lowering scRNA-seq cost is to pool samples prior to droplet-based barcoding with subsequent demultiplexing of sequence reads.

Materials and data

PBMC dataset [2] can be found under http://support.10xgenomics.com/single-cell/datasets
Hashtagged dataset [4] can be found under the accession number GSE108313
Demuxlet dataset [6] can be found under the accession number GSE96583
Result data are available in https://github.com/jon-xu/scSplit_paper_data
scSplit software is freely available at https://github.com/jon-xu/scSplit/
The software release is archived in zenodo http://doi.org/10.5281/zenodo.3464622
10X Genomics Chromium 3' Reagent Kit
Illumina Nextseq 500

Procedure

scSplit workflow — The overall pipeline of scSplit tool.

Data quality control and filtering: The mixed sample BAM file is first filtered to keep only the reads with a list of valid barcodes to reduce technical noise. Additional filtering is then performed to remove reads that meet any of the following: mapping quality score less than 10, unmapped, failing quality checks, secondary or supplementary alignment, or PCR or optical duplicate. The BAM file is then marked for duplication, sorted and indexed.
SNV calling (A): Freebayes v1.2 is used to call SNVs on the filtered BAM file, set to ignore insertions and deletions (indels), multi-nucleotide polymorphisms (MNPs), and complex events. A minimum base quality score of one and minimum allele count of two is required to call a variant. The output VCF file is further filtered to keep only SNVs with quality scores greater than 30.
Building allele count matrices (B): The “matrices.py” script is run which produces two.csv files, one for each of reference and alternate allele counts as output.
Model initialization (C): find the distinct groups of cells in the scRNA-seq and use them to initialize the Allele Fraction Model (SNVs by samples).
E-M iterations till convergence (D): Initialized allele fraction model and the two allele count matrices are used together to calculate the probability of each cell belonging to the clusters. After each round, allele fraction model is updated based on the probability of cell assignment and this is iterated until overall likelihood of the model reaches convergence.
Alternative presence/absence genotypes (E): matrix indicating cluster genotypes at each SNV is built in this step.
Find distinguishing variants for clusters and use to assign samples to clusters (F): In order to assign each model cluster back to the specific sample, distinguishing variants are identified so that genotyping of the least number of loci using the a suitable platform may be performed. Gram-Schmidt orthogonalization is used to get the minimum set of informative P/A genotypes.

Results

Fig. 1 scSplit

Results on simulated, merged hash-tagged scRNA-seq datasets confirmed scSplit a useful tool to demultiplex pooled single cells. A Confusion matrix showing scSplit demultiplexing results on simulated 2-, 3-, 4- and 8-mix; B TPR and FDR of for singlets and doublets predicted by scSplit and demuxlet compared to known truth before merging; C TPR and FDR of for singlets and doublets predicted by scSplit and demuxlet compared to cell hashing tags.

Acknowledgements

Publication

Genotype-free demultiplexing of pooled single-cell RNA-seq

Jun Xu, Caitlin Falconer, Quan Nguyen, Joanna Crawford, Brett D. McKinnon, Sally Mortlock, Anne Senabouth, Stacey Andersen, Han Sheng Chiu, Longda Jiang, Nathan J. Palpant, Jian Yang, Michael D. Mueller, Alex W. Hewitt, Alice Pébay, Grant W. Montgomery, Joseph E. Powell & Lachlan J.M Coin

Genome Biology volume 20, Article number: 290 (2019)