Sample demultiplexing, multiplet detection, experiment planning and novel cell type verification in single cell sequencing

Abstract

Identifying and removing multiplets from downstream analysis is essential to improve the scalability and reliability of single cell RNA sequencing (scRNA-seq). High multiplet rates create artificial cell types in the dataset. Sample barcoding, including the cell hashing technology and the MULTI-seq technology, enables analytical identification of a fraction of multiplets in a scRNA-seq dataset.We propose a Gaussian-mixture-model-based multiplet identification method, GMM-Demux. GMM-Demux accurately identifies and removes the sample-barcoding-detectable multiplets and estimates the percentage of sample-barcodingundetectable multiplets in the remaining dataset. GMM-Demux describes the droplet formation process with an augmented binomial probabilistic model, and uses the model to authenticate cell types discovered from a scRNA-seq dataset.We conducted two cell-hashing experiments, collected a public cell-hashing dataset, and generated a simulated cellhashing dataset. We compared the classification result of GMM-Demux against a state-of-the-art heuristic-based classifier. We show that GMM-Demux is more accurate, more stable, reduces the error rate by up to 69×, and is capable of reliably recognizing 9 multiplet-induced fake cell types and 8 real cell types in a PBMC scRNA-seq dataset.

Artificial-Cell-Type Aware Cell Type Classification in CITE-seq

“…Each ACT droplet contains multiple cells of different types. Since the ADT counts of an ACT droplet equals to the sum of the ADT counts of its member cells, also because the logarithm dampens small-scale changes in the ADT counts (such as log(2 • ) = log( ) + log (2), with ≫ 2), when individual cells merge into an ACT droplet, in the surface marker space, after CLR, the ACT droplet is positioned as slightly exceeding the highest coordinate in each surface marker dimension, across all of its member cells (Xin, et al, 2019). Mathematically, assuming 8 is the set of cells included in an ACT droplet , then the CLR-transformed surface marker count vector 8 of approximately equals…”

“…Additionally, ACT clusters encompass much fewer cells than BCT clusters; and there could be many more ACT clusters than BCT clusters. According to Xin, et al (Xin, et al, 2019), with a total of BCT clusters, there could be as many as 2 C − − 1 ACT clusters. Altogether, a CITE-seq dataset is likely to share the following properties in the surface marker space: (1) it may contain a large number of clusters;…”

“…In a cell hashed scRNA-seq dataset, droplets are categorized into either multi-sample multiplets (MSM) or single-sample droplets (SSDs). Prior work (Xin, et al, 2019) has established that ACT clusters have MSM percentages approaching and exceeding 1 − 1/ , where is the number of hashtags used in cell hashing (assuming even cell distributions among samples in cell hashing). For the PBMC_16k dataset, we used = 4 hashtags.…”

“…If a putative homogeneous droplet cluster in PBMC_16k has a MSM ratio approaching or exceeding 75%, then it is highly likely to be an ACT cluster. We used GMM-Demux (Xin, et al, 2019) to classify droplets into MSMs and SSDs.…”

“…For instance, in the above example (Figure 2B), there exists a sizable CD3 + CD19 + droplet cluster which is rarely observed in PBMCs. Prior work has shown that such droplets are all multiplets and recommends subsequent removal of such droplets from downstream analysis (Xin, et al, 2019). In this work, we call cell types that truly exist as biological cell types (BCT) and cell types created by CITE-seq multiplets as artificial cell types (ACT).…”