Statistical modeling, estimation, and remediation of sample index hopping in multiplexed droplet-based single-cell RNA-seq data

Abstract

Sample index hopping can substantially confound the analysis of multiplexed sequencing data due to the resulting erroneous assignment of some, or even all, of the sequencing reads generated by a cDNA fragment in a given sample to other samples. In those target samples, the data cross-contamination artifact takes the form of "phantom molecules", molecules that exist only in the data by virtue of read misassignment. The presence of phantom molecules in droplet-based single-cell RNA-seq data should be a cause of great concern since they can introduce both phantom cells and artifactual differentially-expressed genes in downstream analyses. More importantly, even when the index hopping rate is very small, the fraction of phantom molecules in the entire dataset can be high due to the distributional properties of sequencing reads across samples. To our knowledge, current computational methods are unable to accurately estimate the underlying rate of index hopping nor adequately correct for the resultant misassignment in droplet-based single cell RNA-seq data. Here, we introduce a probabilistic model that formalizes the phenomenon of index hopping and allows the accurate estimation of its rate. Application of the proposed model to several multiplexed datasets suggests that the sample index hopping probability for a given read ranges between 0.003 to 0.009, arguable low numbers, even though, counter-intuitively, they can give rise to a large fraction of phantom molecules -as high as 85% -in any given sample. We also present a model-based approach for inferring the true sample of origin of the reads that are affected by index hopping, thus allowing the purging of the majority of phantom molecules in the data. Using empirical and simulated data, we show that we can reassign reads to their true sample of origin and remove predicted phantom molecules through a principled probabilistic procedure that optimally minimizes the false positive rate. Thus, even though sample index hopping often substantially compromises single-cell RNA-seq data, it is possible to accurately quantify, detect, and reassign the affected reads and remove the phantom molecules generated by index hopping. Code and reproducible analysis notebooks are available at https://github.com/csglab/phantom_purge.

Computational Methods for Single-Cell RNA Sequencing

“…scRNA-seq experiments of thousands of cells experience additional demultiplexing challenges including sequencing errors and hopping of barcodes/indices, which can prevent the collapse of reads to their appropriate UMI, cell, and library barcode. In large-scale, complex experiments, these issues require attention, and several methods are available to correct for UMI and cellular barcode errors (23) and index hopping (24).…”