Introduction

Droplet-based single-cell RNA sequencing (scRNA-seq) facilitates measuring the transcriptomes of thousands of cells in a single run. Pooling cells from different samples or conditions before cell partitioning and library preparation can significantly lower costs and reduce batch effects. The task of assigning each cell of a pooled sample to its sample of origin is called demultiplexing. If genetically diverse samples are pooled, single nucleotide polymorphisms in coding regions can be used for demultiplexing. When working with genetically similar or identical samples, an additional experimental step is required to label the cells with a sample-specific barcode oligonucleotide before pooling. Several techniques have been developed to label cells or nuclei with oligonucleotides based on antibodies (Stoeckius et al. 2018; Gaublomme et al. 2019) or lipids (McGinnis et al. 2019). These oligonucleotides are termed hashtag oligonucleotides (HTOs) and are sequenced together with the RNA molecules of the cells resulting in an (HTOs x droplets) count matrix in addition to the (genes x droplets) matrix with RNA read counts.

The demuxmix package implements a method to demultiplex droplets based on HTO counts using negative binomial regression mixture models. demuxmix can be applied to the HTO counts only, but better results are often achieved if the total number of genes detected per droplet (not the complete transcription profile) is passed to the method along with the HTO counts to leverage the positive association between genes detected and HTO counts. Further, demuxmix provides estimated error rates based on its probabilistic mixture model framework, plots for data quality assessment, and multiplet identification, as outlined in the example workflows in this vignette. Technical details of the methods are described in the man pages.

Installation

The demuxmix package is available at Bioconductor and can be installed via BiocManager::install:

if (!require("BiocManager"))
    install.packages("BiocManager")
BiocManager::install("demuxmix")

The package only needs to be installed once. Load the package into an R session with

library(demuxmix)

Quick start

A matrix of raw HTO counts (HTO x cells) and a vector with the number of detected genes per droplet are needed to run demuxmix with default settings. Empty and low-quality droplets should be removed before running demuxmix. A gene with at least one read is usually considered as detected. Here, we simulate a small example dataset.

library(demuxmix)

set.seed(2642)
class <- rbind(
    c(rep(TRUE,  220), rep(FALSE, 200)),
    c(rep(FALSE, 200), rep(TRUE,  220))
)
simdata <- dmmSimulateHto(class)
hto <- simdata$hto
dim(hto)
#> [1]   2 420
rna <- simdata$rna
length(rna) == ncol(hto)
#> [1] TRUE

The dataset consists of 420 droplets with cells labeled with two different HTOs. The first 200 droplets are singlets labeled with the first HTO, followed by another 200 singlets labeled with the second HTO. The remaining 20 droplets are doublets, which are positive for both HTOs. Next, we run demuxmix to assign droplets to HTOs.

dmm <- demuxmix(hto, rna = rna)
summary(dmm)
#>       Class NumObs     RelFreq   MedProb      ExpFPs          FDR
#> 1     HTO_1    198 0.475961538 0.9999994 0.419487123 0.0021186218
#> 2     HTO_2    197 0.473557692 0.9999985 0.287367473 0.0014587181
#> 3   singlet    395 0.949519231 0.9999992 0.706854596 0.0017895053
#> 4 multiplet     20 0.048076923 0.9999995 0.005837305 0.0002918652
#> 5  negative      1 0.002403846 0.9922442 0.007755814 0.0077558137
#> 6 uncertain      4          NA        NA          NA           NA
classes <- dmmClassify(dmm)
table(classes$HTO)
#> 
#>       HTO_1 HTO_1,HTO_2       HTO_2    negative   uncertain 
#>         198          20         197           1           4

The object dmm contains the mixture models used to classify the droplets. The data frame returned by summary shows that 198 droplets were assigned to HTO_1 and 197 to HTO_2, respectively. Since these results meet our expectations and the estimated error rates are reasonably low, we ran dmmClassify to obtain the classifications for each droplet as a data frame with one row per droplet. The first column HTO of this data frame contains the final classification results.

A histogram of the HTO values overlayed with the components from the mixture model can be plotted for quality control. The following command plots a panel with one histogram per HTO in the dataset.

plotDmmHistogram(dmm)

Density histograms overlayed with mixture probability mass function. The density histograms show the distribution of the HTO counts for the first HTO (upper figure) and the 2nd HTO (lower figure). The negative component of the mixture model representing non-tagged cells is shown in blue, and the positive component is in red.

Demultiplexing droplets with demuxmix

Example datasets

Two example datasets are introduced in this vignette to illustrate a typical demuxmix workflow. The first dataset is a small simulated dataset used to generate the plots when building this vignette. The alternative second dataset is a real dataset and can be downloaded from the ExperimentHub via the scRNAseq package. Both datasets can be used to go through this vignette by running either the first (simulated data) or the second code block (real data) below. Since the real dataset is much larger, some commands may take up to one minute to complete, which is why this vignette was built with the simulated data.

Simulated dataset

Simulated HTO count data are generated for 650 droplets by the method dmmSimulateHto. The logical matrix class defines for each droplet (column) and HTO (row) whether the droplet is positive or negative for that HTO. Thus, the 3 x 650 matrix class below describes a dataset with 3 hashtags and 650 droplets, of which 50 are doublets (with cells tagged by HTO_1 and HTO_2). The remaining 600 droplets consist of 3 blocks of 200 singlets tagged by one of the three HTOs each.

library(demuxmix)
library(ggplot2)
library(cowplot)

set.seed(5636)
class <- rbind(
    c(rep( TRUE, 200), rep(FALSE, 200), rep(FALSE, 200), rep( TRUE, 50)),
    c(rep(FALSE, 200), rep( TRUE, 200), rep(FALSE, 200), rep( TRUE, 50)),
    c(rep(FALSE, 200), rep(FALSE, 200), rep( TRUE, 200), rep(FALSE, 50))
)
simdata <- dmmSimulateHto(
    class = class,
    mu = c(600, 400, 200),
    theta = c(25, 15, 25),
    muAmbient = c(30, 30, 60),
    thetaAmbient = c(20, 10, 5),
    muRna = 3000,
    thetaRna = 30
)
hto <- simdata$hto
rna <- simdata$rna

htoDf <- data.frame(t(hto), HTO = colSums(hto), NumGenes = rna)
pa <- ggplot(htoDf, aes(x = HTO_1)) +
    geom_histogram(bins = 25)
pb <- ggplot(htoDf, aes(x = HTO_2)) +
    geom_histogram(bins = 25)
pc <- ggplot(htoDf, aes(x = HTO_3)) +
    geom_histogram(bins = 25)
pd <- ggplot(htoDf, aes(x = NumGenes, y = HTO)) +
    geom_point()
plot_grid(pa, pb, pc, pd, labels = c("A", "B", "C", "D"))

Characteristics of the simulated dataset. A) The histogram of the HTO counts from the first HTO (HTO_1) shows a clear separation between positive and negative droplets. B) The histogram of the second HTO (HTO_2) looks similar, although the positive droplets have a smaller mean count and a larger dispersion. C) The histogram of the third HTO reveals a more extensive overlap between the distributions of the positive and negative droplets. D) The scatter plot shows a positive correlation between the number of detected genes and HTO counts in the simulated data.

dmmSimulateHto dmmSimulateHto returns a matrix with the simulated HTO counts and a vector with the simulated number of detected genes for each droplet. Figure @ref(fig:simulate) shows that we simulated two HTOs of excellent quality (HTO_1 and HTO_2) and a third HTO (HTO_3) with more background reads, complicating the demultiplexing. Further, the simulated data shows a positive association between the number of detected genes and the number of HTO counts per droplet, which is often observed in real data.

Cell line mixture dataset

The cell line mixture dataset from Stoeckius et al. (2018) consists of cells from 4 different cell lines. Three samples were taken from each cell line and tagged with a different HTO resulting in a total of 12 different HTOs. The downloaded dataset still contains many potentially empty droplets, which are removed using emptyDrops. Subsequently, the numbers of detected genes are calculated and the HTO matrix is extracted from the SingleCellExperiment object. More information about the preprocessing and data structures for single-cell data in Bioconductor can be found in this excellent online book.

library(demuxmix)
library(ggplot2)
library(cowplot)
library(scRNAseq)
library(DropletUtils)

set.seed(8514)
htoExp <- StoeckiusHashingData(type = "mixed")
eDrops <- emptyDrops(counts(htoExp))
htoExp <- htoExp[, which(eDrops$FDR <= 0.001)]
rna <- colSums(counts(htoExp) > 0)
hto <- counts(altExp(htoExp))
dim(hto)

htoDf <- data.frame(t(hto[c("HEK_A", "KG1_B", "KG1_C"), ]),
    HTO = colSums(hto), NumGenes = rna
)
pa <- ggplot(htoDf, aes(x = HEK_A)) +
    geom_histogram(binwidth = 10) +
    coord_cartesian(xlim = c(0, 600), ylim = c(0, 500))
pb <- ggplot(htoDf, aes(x = KG1_B)) +
    geom_histogram(binwidth = 10) +
    coord_cartesian(xlim = c(0, 600), ylim = c(0, 500))
pc <- ggplot(htoDf, aes(x = KG1_C)) +
    geom_histogram(binwidth = 10) +
    coord_cartesian(xlim = c(0, 600), ylim = c(0, 500))
pd <- ggplot(htoDf, aes(x = NumGenes, y = HTO)) +
    geom_point(size = 0.1) +
    coord_cartesian(ylim = c(0, 750))
plot_grid(pa, pb, pc, pd, labels = c("A", "B", "C", "D"))

The plots generated by the code above reveal that the quality of the different HTOs varies in the cell line mixture dataset. Most HTOs demonstrate a nicely separated bimodal distribution, as exemplarily shown for HEK_A, but HG1_C and, to a lesser extent, HG1_B, demonstrate a larger overlap between the distributions of the positive and negative droplets. As with the simulated data, there is a positive association between HTO counts and the detected number of genes. However, the association appears noisier because four different cell lines with highly distinct RNA profiles and different cell surface characteristics (likely influencing HTO labeling) were pooled. Consequently, the HTO counts and the number of detected genes are very different between cells from different samples but very similar between cells from the same sample. This uncommon experimental design makes it difficult to leverage the association between the number of genes detected and HTO counts during demultiplexing. With default parameters, demuxmix automatically selects the most appropriate model, and, for this dataset, naive instead of regression mixture models are used for most HTOs. In addition, the large number of 12 pooled samples further complicates the demultiplexing.

Running demuxmix

demuxmix takes a matrix of HTO counts and a vector with the numbers of detected genes per droplet as input and returns an object of class Demuxmix containing a mixture model for each HTO. Several additional parameters can be passed to demuxmix, but all these parameters have default values that work well across many datasets. With the default settings, demuxmix automatically selects either naive mixture models or regression mixture models for each HTO, depending on which model provides the best separation between positive and negative droplets.

dmm <- demuxmix(hto, rna = rna)
dmm
#> Demuxmix object with 3 HTOs and 650 cells.
#>   HTO_1: RegMixModel; converged
#>          mu=(29.012, 604.96)
#>          RNA coef. negative comp: 1.075, p=3.11e-35
#>                    positive comp: 0.9322, p=1.97e-35
#>   HTO_2: RegMixModel; converged
#>          mu=(30.364, 383.49)
#>          RNA coef. negative comp: 0.809, p=2.68e-14
#>                    positive comp: 0.8115, p=3.42e-21
#>   HTO_3: RegMixModel; converged
#>          mu=(62.133, 199.78)
#>          RNA coef. negative comp: 1.143, p=2.07e-21
#>                    positive comp: 1.08, p=8.43e-43
classLabels <- dmmClassify(dmm)
head(classLabels)
#>     HTO      Prob    Type
#> 1 HTO_1 0.9995790 singlet
#> 2 HTO_1 1.0000000 singlet
#> 3 HTO_1 0.9983449 singlet
#> 4 HTO_1 0.9988615 singlet
#> 5 HTO_1 0.9999997 singlet
#> 6 HTO_1 0.9999999 singlet

summary(dmm)
#>       Class NumObs    RelFreq   MedProb    ExpFPs         FDR
#> 1     HTO_1    198 0.31083203 0.9999958 1.3099257 0.006615786
#> 2     HTO_2    197 0.30926217 0.9999975 1.3625168 0.006916329
#> 3     HTO_3    184 0.28885400 0.9923134 5.5638155 0.030238128
#> 4   singlet    579 0.90894819 0.9998231 8.2362580 0.014224971
#> 5 multiplet     51 0.08006279 0.9999863 0.6500497 0.012746072
#> 6  negative      7 0.01098901 0.8841169 0.8390094 0.119858491
#> 7 uncertain     13         NA        NA        NA          NA

# Compare demultiplexing results to ground truth from simulation
table(classLabels$HTO, simdata$groundTruth)
#>                    
#>                     HTO_1 HTO_1,HTO_2 HTO_2 HTO_3
#>   HTO_1               198           0     0     0
#>   HTO_1,HTO_2           0          48     0     0
#>   HTO_1,HTO_2,HTO_3     0           2     0     0
#>   HTO_2                 0           0   197     0
#>   HTO_2,HTO_3           0           0     1     0
#>   HTO_3                 0           0     0   184
#>   negative              0           0     0     7
#>   uncertain             2           0     2     9

For the simulated data, the object dmm contains three regression mixture models. We then apply dmmClassify to obtain a data frame with one row for each droplet. The first column contains the classification result. The second column contains the posterior probability that the assigned HTO is correct. The last column contains the type of the assignment, which is either “singlet”, “multiplet”, “negative” (not tagged by any HTO), or “uncertain” (posterior probability too small to classify the droplet with confidence). Only droplets of type singlet should be kept in the dataset. Multiplets of two or more cells from the same sample cannot be detected at the demultiplexing step. The comparison with the true labels from the simulation shows that most droplets were classified correctly.

The parameter model can be used to select a specific mixture model. The naive mixture model selected in the code below does not use any information from the RNA data. As shown in the following output, the naive mixture model performs slightly worse than the regression mixture model mainly because more droplets are assigned to the class “uncertain”.

dmmNaive <- demuxmix(hto, model = "naive")
dmmNaive
#> Demuxmix object with 3 HTOs and 650 cells.
#>   HTO_1: NaiveMixModel; converged
#>          mu=(29.707, 614.7)
#>   HTO_2: NaiveMixModel; converged
#>          mu=(30.825, 390.29)
#>   HTO_3: NaiveMixModel; converged
#>          mu=(58.938, 191.29)
classLabelsNaive <- dmmClassify(dmmNaive)
summary(dmmNaive)
#>       Class NumObs     RelFreq   MedProb    ExpFPs        FDR
#> 1     HTO_1    175 0.293132328 0.9960974  5.400830 0.03086188
#> 2     HTO_2    179 0.299832496 0.9969071  4.060804 0.02268606
#> 3     HTO_3    180 0.301507538 0.9949591  4.711928 0.02617738
#> 4   singlet    534 0.894472362 0.9960974 14.173562 0.02654225
#> 5 multiplet     61 0.102177554 0.9811527  4.066596 0.06666551
#> 6  negative      2 0.003350084 0.9421215  0.115757 0.05787849
#> 7 uncertain     53          NA        NA        NA         NA

# Compare results of the naive model to ground truth from simulation
table(classLabelsNaive$HTO, simdata$groundTruth)
#>                    
#>                     HTO_1 HTO_1,HTO_2 HTO_2 HTO_3
#>   HTO_1               175           0     0     0
#>   HTO_1,HTO_2           0          42     0     0
#>   HTO_1,HTO_2,HTO_3     0           3     0     0
#>   HTO_1,HTO_3           9           0     0     0
#>   HTO_2                 0           0   179     0
#>   HTO_2,HTO_3           0           0     7     0
#>   HTO_3                 0           0     0   180
#>   negative              0           0     0     2
#>   uncertain            16           5    14    18

Another useful parameter is the acceptance probability pAcpt, which can be passed to the demuxmix method to overwrite the default value, or directly set in the object dmm as shown in the code block below. The parameter is used at the classification step and specifies the minimum posterior probability required to classify a droplet. If the posterior probability of the most likely class is smaller than pAcpt, the droplet is classified as “uncertain”. Setting pAcpt to 0 forces the classification of all droplets. For HTO datasets of moderate quality, the default value can be lowered to recover more droplet in the dataset, if a larger error rate is acceptable. The summary method estimates the FDR depending on the current setting of pAcpt.

pAcpt(dmm)
#> [1] 0.729
pAcpt(dmm) <- 0.95
summary(dmm)
#>       Class NumObs     RelFreq   MedProb     ExpFPs         FDR
#> 1     HTO_1    191 0.327615780 0.9999978 0.39912417 0.002089655
#> 2     HTO_2    190 0.325900515 0.9999986 0.24119804 0.001269463
#> 3     HTO_3    155 0.265866209 0.9943603 1.60287265 0.010341114
#> 4   singlet    536 0.919382504 0.9999483 2.24319486 0.004185065
#> 5 multiplet     45 0.077186964 0.9999972 0.07140102 0.001586689
#> 6  negative      2 0.003430532 0.9732852 0.05342954 0.026714771
#> 7 uncertain     67          NA        NA         NA          NA

pAcpt(dmm) <- 0
summary(dmm)
#>       Class NumObs    RelFreq   MedProb    ExpFPs         FDR
#> 1     HTO_1    200 0.30769231 0.9999952  1.943649 0.009718245
#> 2     HTO_2    198 0.30461538 0.9999974  1.683204 0.008501031
#> 3     HTO_3    191 0.29384615 0.9918053  8.372435 0.043834737
#> 4   singlet    589 0.90615385 0.9997875 11.999288 0.020372305
#> 5 multiplet     52 0.08000000 0.9999849  1.092788 0.021015155
#> 6  negative      9 0.01384615 0.8036236  1.588729 0.176525464
#> 7 uncertain      0         NA        NA        NA          NA

Quality control

The demuxmix package implements methods for assessing data quality and model fit. All plotting methods plot a panel with one graph for each HTO in the dataset as default. Specific HTOs can be selected via the parameter hto. The most informative plot is probably the histogram of the HTO data overlaid with the mixture probability mass function, as this plot shows both the raw data and the model fit.

plotDmmHistogram(dmm)

Density histograms overlayed with mixture probability mass functions. The density histogram is shown for each HTO in the simulated dataset. The negative component of the respective mixture model representing non-tagged cells (blue) and the positive component (red) are plotted on top of the histogram. The black curve is the mixture pmf.

dmmOverlap(dmm)
#>        HTO_1        HTO_2        HTO_3 
#> 6.612645e-10 1.392127e-05 4.829136e-02

First, the histograms should be used to verify the model fit. The model fit is good if the mixture pmf closely follows the shape of the histogram. The model fit in the simulated data is adequate for all three HTOs. Second, the histogram should be bimodal, and the blue and red components should show little overlap, which is the case for the first two HTOs (HTO_1 and HTO_2). The third HTO (HTO_3) was simulated to harbor more background reads (mean of 60 reads), and, consequently, the blue component is shifted towards the right. The method dmmOverlap calculates the area intersected by the two components. The area is close to zero for the first two HTOs but 0.048 for HTO_3. As a rule of thumb, an overlap less than 0.03 can be considered excellent. As seen in this example dataset, a value around 0.05 will lead to some droplets being classified as “uncertain” but is still sufficient to accurately demultiplex the majority of droplets.

The histogram plots can look very different across different real datasets depending on (i) the HTO sequencing depth, (ii) the number of pooled samples, (iii) the number of droplets, and (iv) the quality of the HTO experiment (cross-staining, background HTOs, cell debris). Only the last factor (iv) relates to the actual quality of the data. Specifically, (i) and (ii) can both lead to a histogram that looks like a vertical blue line with an x-offset of 0 and a flat horizontal red line with a y-offset of 0. The reason is that a large sequencing depth results in large HTO counts in positive droplets so that the probability mass of the red component spreads over an extensive range of values on the x-axis and appears flat compared to the sharp location of the blue background component. Similarly, a large number of pooled samples, e.g., the 12 samples in the cell line mixture dataset, causes the red component to appear flat as it only covers an area of about 1/12 compared to 11/12 covered by the blue component. In such cases, it is helpful to zoom into the critical part of the histogram where the red and blue components overlap. For the cell line mixture dataset, the following command generates a suitable histogram for the first HTO (panel B).

pa <- plotDmmHistogram(dmm, hto=1)
pb <- plotDmmHistogram(dmm, hto=1) +
    coord_cartesian(xlim = c(0, 200), ylim = c(0, 0.01))
plot_grid(pa, pb, labels = c("A", "B"))

Another useful quality plot is the histogram of the posterior probability that a droplet is positive for the respective HTO.

plotDmmPosteriorP(dmm)

Histograms of posterior probabilities. Each histogram shows the distribution of the posterior probabilities that a droplet contains a tagged cell. Posterior probabilities were obtained from the mixture model fitted to the respective HTO data.

As seen in the first two histograms for HTO_1 and HTO_2, most droplets should have a posterior probability close to 0 (negative droplet) or close to 1 (positive droplet). Consistent with the previous quality plots, the third HTO (HTO_3) demonstrates some droplets with posterior probabilities between 0 and 1, reflecting droplets with cells of uncertain origin.

Finally, if regression mixture models were used, the decision boundary in relation to the number of detected genes and the HTO counts can be plotted.

plotDmmScatter(dmm)

Decision boundary. The scatter plots show the relation between the number of detected genes and HTO counts for each of the three HTOs. The color indicates the posterior probability. The black dashed line depicts the decision boundary where the posterior probability is 0.5.

These plots are only available if regression mixture models were used. A droplet with many detected genes is required to have more HTO reads in order to be classified as positive. If naive mixture models were used, the dashed decision boundaries between blue (negative) and red (positive) droplets would be vertical lines.

Comparison to hashedDrops

The method hashedDrops in the package DropletUtils provides an alternative approach for demultiplexing HTO-labeled single cell data. The following code block runs hashedDrops and compares the results to demuxmix.

suppressPackageStartupMessages(library(DropletUtils))
suppressPackageStartupMessages(library(reshape2))
hd <- hashedDrops(hto)
hdrops <- rownames(hto)[hd$Best]
hdrops[!hd$Confident] <- "uncertain"
hdrops[hd$Doublet] <- "multiplet"

dmux <- classLabels$HTO
dmux[classLabels$Type == "multiplet"] <- "multiplet"

comp <- melt(as.matrix(table(dmux, hdrops)))
colnames(comp) <- c("demuxmix", "hashedDrops", "Count")
comp$color <- ifelse(comp$Count > 100, "black", "white")
ggplot(comp, aes(x = demuxmix, y = hashedDrops, fill = Count)) +
    geom_tile() +
    scale_fill_viridis_c() +
    geom_text(aes(label = Count), col = comp$color, size = 5)

Comparison between demuxmix and hashedDrops. The heatmap depicts the classification results for the simulated dataset obtained from demuxmix on the x-axis and hashedDrops on the y-axis.

The classifications of both methods are highly concordant. No droplet has been assigned to different singlet classes by the two approaches. However, hashedDrops assigned more droplets tagged by HTO_3 to the category “uncertain”, which is expected when looking at the lower panel of Figure @ref(fig:qualityScatterplot). In contrast to demuxmix, hashedDrops does not utilize the positive association between the number of genes and HTO counts explicitly simulated in this dataset.

Special usecase: pooling non-labeled with labeled cells

If precious rare cells are pooled with highly abundant cells, labeling the highly abundant cells only but not the rare cells avoids additional losses of the rare cells during the labeling process. However, such a design results in a more challenging demultiplexing task. The real dataset used as an example in this section consists of rare cerebrospinal fluid cells (non-labeled) and peripheral blood mononuclear cells (PBMCs) stained with oligonucleotide-conjugated antibodies. In this design, the “negative” cells identified by demuxmix correspond to the CSF cells. The dataset is included in the demuxmix package as a data frame.

data(csf)
head(csf)
#>                      HTO NumGenes freemuxlet freemuxlet.prob
#> AAACCCATCAATCGGT-1 21087     2640        0,0         0.00000
#> AAACCCATCATTTACC-1  3616      414        0,0        -1.47424
#> AAACGAATCGGTCTAA-1  5387     1841        0,0         0.00000
#> AAACGCTAGGGAGTGG-1  3293     1743        0,0         0.00000
#> AAACGCTAGTGATTCC-1  3876     3296        0,0         0.00000
#> AAAGAACCACACTGGC-1  1150     3540        1,1         0.00000

csf <- csf[csf$NumGenes >= 200, ]
nrow(csf)
#> [1] 2510
hto <- t(matrix(csf$HTO, dimnames = list(rownames(csf), "HTO")))

The data frame contains the number of HTO reads and the number of detected genes per droplet in the first two columns. We remove all droplets with less than 200 detected genes since these droplets are unlikely to contain intact cells. The HTO counts are then converted into a matrix as required by demuxmix. The matrix has only one row since only the PBMCs were stained. For this example dataset, CSF cells and PBMCs from two genetically unrelated donors were pooled so that genetic demultiplexing could be used to benchmark the HTO-based demultiplexing. The third column contains the result from the genetic demultiplexing using freemuxlet (Kang et al. 2018). The fourth column contains freemuxlet’s logarithmized posterior probability.

dmm <- demuxmix(hto, rna = csf$NumGenes)
dmm
#> Demuxmix object with 1 HTO and 2510 cells.
#>   HTO: RegMixModel; converged
#>        mu=(897.39, 4977.8)
#>        RNA coef. negative comp: 0.02205, p=0.448
#>                  positive comp: 0.3756, p=1.16e-73

summary(dmm)
#>       Class NumObs   RelFreq   MedProb   ExpFPs         FDR
#> 1       HTO   1099 0.4625421 0.9999986 3.708186 0.003374146
#> 2   singlet   1099 0.4625421 0.9999986 3.708186 0.003374146
#> 3 multiplet      0 0.0000000        NA 0.000000         NaN
#> 4  negative   1277 0.5374579 0.9999490 4.608103 0.003608538
#> 5 uncertain    134        NA        NA       NA          NA

dmmOverlap(dmm)
#>        HTO 
#> 0.02682714

demuxmix selected a regression mixture model with a significant regression coefficient in the positive component, indicating that the number of detected genes in the stained PBMCs is predictive of the HTO counts observed in these cells. Although demuxmix selected a regression model for the negative component as well, the smaller coefficient and the larger p-value suggest that the association is much weaker in the CSF cells. This is common since a larger amount of background HTOs is required in order to detect the association in the negative droplets.

Next, we look at some QC plots.

histo <- plotDmmHistogram(dmm)
scatter <- plotDmmScatter(dmm) + coord_cartesian(xlim = c(2, 4))
plot_grid(histo, scatter, labels = c("A", "B"), nrow = 2)

Demultiplexing a pool of labeled PBMCs and non-labeled CSF cells. A) The density histogram overlaid with the mixture pmf shows a good separation between the positive red component (PMBCs) and the negative blue component (CSF cells). B) The scatter plot shows the number of HTO reads (x-axis) versus the number of detected genes (y-axis) on the logarithmic scale. The color indicates the posterior probability of the droplet containing a tagged cell.

Overall, the histogram reveals a good model fit but also shows some background staining of the CSF cells. The mean of the negative component is 897.4 reads. Still, the overlap with the positive component (mean of 4977.8) is reasonably small, and the larger mean of the negative component is probably driven partly by the large sequencing depth of the HTO library. Moreover, the scatter plot shows that a smaller set of cells has a lower RNA content (y-axis) and that those cells require less HTO counts (x-axis) in order to be classified as positive (red color).

Finally, we use genetic demultiplexing to assess demuxmix’s performance. Multi-sample multiplets detected by freemuxlet are removed since multiplets cannot be detected when just one of two samples is stained. We also remove cells that were not classified with high confidence by freemuxlet.

class <- dmmClassify(dmm)
highConf <- csf$freemuxlet %in% c("0,0", "1,1") &
    exp(csf$freemuxlet.prob) >= 0.99
table(class$HTO[highConf], csf$freemuxlet[highConf])
#>            
#>              0,0  1,1
#>   HTO       1036   22
#>   negative    39 1143
#>   uncertain   52   55

# Sensitivity "P(class=PBMC | PBMC)"
sum(csf$freemuxlet[highConf] == "0,0" & class$HTO[highConf] == "HTO") /
    sum(csf$freemuxlet[highConf] == "0,0" & class$HTO[highConf] != "uncertain")
#> [1] 0.9637209

# Specificity "P(class=CSF | CSF)"
sum(csf$freemuxlet[highConf] == "1,1" & class$HTO[highConf] == "negative") /
    sum(csf$freemuxlet[highConf] == "1,1" & class$HTO[highConf] != "uncertain")
#> [1] 0.9811159

With the default acceptance probability of 0.9, demuxmix achieved a sensitivity and specificity above 95%. Only 107 cells were classified as “uncertain” and have to be discarded.

For comparison, we run demuxmix again and, this time, manually select the naive mixture model.

dmmNaive <- demuxmix(hto, model = "naive")
class <- dmmClassify(dmmNaive)
table(class$HTO[highConf], csf$freemuxlet[highConf])
#>            
#>              0,0  1,1
#>   HTO       1027   28
#>   negative    45 1120
#>   uncertain   55   72

# Sensitivity "P(class=PBMC | PBMC)"
sum(csf$freemuxlet[highConf] == "0,0" & class$HTO[highConf] == "HTO") /
    sum(csf$freemuxlet[highConf] == "0,0" & class$HTO[highConf] != "uncertain")
#> [1] 0.9580224

# Specificity "P(class=CSF | CSF)"
sum(csf$freemuxlet[highConf] == "1,1" & class$HTO[highConf] == "negative") /
    sum(csf$freemuxlet[highConf] == "1,1" & class$HTO[highConf] != "uncertain")
#> [1] 0.9756098

The naive model achieved a slightly lower sensitivity and specificity than the regression mixture model. In addition, more cells were classified as “uncertain”, demonstrating the benefit of modeling the relationship between the number of detected genes and HTO counts.

Session Info

sessionInfo()
#> R version 4.4.2 (2024-10-31)
#> Platform: x86_64-pc-linux-gnu
#> Running under: Ubuntu 24.04.1 LTS
#> 
#> Matrix products: default
#> BLAS:   /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3 
#> LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so;  LAPACK version 3.12.0
#> 
#> locale:
#>  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
#>  [3] LC_TIME=en_US.UTF-8        LC_COLLATE=C              
#>  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
#>  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
#>  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
#> [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
#> 
#> time zone: Etc/UTC
#> tzcode source: system (glibc)
#> 
#> attached base packages:
#> [1] stats4    stats     graphics  grDevices utils     datasets  methods  
#> [8] base     
#> 
#> other attached packages:
#>  [1] reshape2_1.4.4              DropletUtils_1.27.0        
#>  [3] SingleCellExperiment_1.29.1 SummarizedExperiment_1.37.0
#>  [5] Biobase_2.67.0              GenomicRanges_1.59.1       
#>  [7] GenomeInfoDb_1.43.2         IRanges_2.41.1             
#>  [9] S4Vectors_0.45.2            BiocGenerics_0.53.3        
#> [11] generics_0.1.3              MatrixGenerics_1.19.0      
#> [13] matrixStats_1.4.1           cowplot_1.1.3              
#> [15] ggplot2_3.5.1               demuxmix_1.9.0             
#> [17] BiocStyle_2.35.0           
#> 
#> loaded via a namespace (and not attached):
#>  [1] viridisLite_0.4.2         farver_2.1.2             
#>  [3] R.utils_2.12.3            fastmap_1.2.0            
#>  [5] digest_0.6.37             lifecycle_1.0.4          
#>  [7] statmod_1.5.0             magrittr_2.0.3           
#>  [9] compiler_4.4.2            rlang_1.1.4              
#> [11] sass_0.4.9                tools_4.4.2              
#> [13] utf8_1.2.4                yaml_2.3.10              
#> [15] knitr_1.49                S4Arrays_1.7.1           
#> [17] labeling_0.4.3            dqrng_0.4.1              
#> [19] DelayedArray_0.33.2       plyr_1.8.9               
#> [21] abind_1.4-8               BiocParallel_1.41.0      
#> [23] HDF5Array_1.35.2          withr_3.0.2              
#> [25] sys_3.4.3                 R.oo_1.27.0              
#> [27] grid_4.4.2                fansi_1.0.6              
#> [29] beachmat_2.23.2           colorspace_2.1-1         
#> [31] Rhdf5lib_1.29.0           edgeR_4.5.0              
#> [33] scales_1.3.0              MASS_7.3-61              
#> [35] cli_3.6.3                 rmarkdown_2.29           
#> [37] crayon_1.5.3              httr_1.4.7               
#> [39] DelayedMatrixStats_1.29.0 scuttle_1.17.0           
#> [41] cachem_1.1.0              rhdf5_2.51.0             
#> [43] stringr_1.5.1             zlibbioc_1.52.0          
#> [45] parallel_4.4.2            BiocManager_1.30.25      
#> [47] XVector_0.47.0            vctrs_0.6.5              
#> [49] Matrix_1.7-1              jsonlite_1.8.9           
#> [51] maketools_1.3.1           locfit_1.5-9.10          
#> [53] limma_3.63.2              jquerylib_0.1.4          
#> [55] glue_1.8.0                codetools_0.2-20         
#> [57] stringi_1.8.4             gtable_0.3.6             
#> [59] UCSC.utils_1.3.0          munsell_0.5.1            
#> [61] tibble_3.2.1              pillar_1.9.0             
#> [63] htmltools_0.5.8.1         rhdf5filters_1.19.0      
#> [65] GenomeInfoDbData_1.2.13   R6_2.5.1                 
#> [67] sparseMatrixStats_1.19.0  evaluate_1.0.1           
#> [69] lattice_0.22-6            R.methodsS3_1.8.2        
#> [71] bslib_0.8.0               Rcpp_1.0.13-1            
#> [73] gridExtra_2.3             SparseArray_1.7.2        
#> [75] xfun_0.49                 buildtools_1.0.0         
#> [77] pkgconfig_2.0.3

References

Gaublomme, JT, B Li, C McCabe, A Knecht, Y Yang, E Drokhlyansky, N Van Wittenberghe, et al. 2019. “Nuclei Multiplexing with Barcoded Antibodies for Single-Nucleus Genomics.” Nature Communications 10 (1): 2907.

Kang, HM, M Subramaniam, S Targ, M Nguyen, L Maliskova, E McCarthy, E Wan, et al. 2018. “Multiplexed Droplet Single-Cell RNA-Sequencing Using Natural Genetic Variation.” Nature Biotechnology 36 (1): 89–94.

McGinnis, CS, DM Patterson, J Winkler, DN Conrad, MY Hein, V Srivastava, JL Hu, et al. 2019. “MULTI-Seq: Sample Multiplexing for Single-Cell RNA Sequencing Using Lipid-Tagged Indices.” Nature Methods 16 (7): 619–26.

Stoeckius, M, S Zheng, B Houck-Loomis, S Hao, BZ Yeung, WM Mauck, P Smibert, and R Satija. 2018. “Cell Hashing with Barcoded Antibodies Enables Multiplexing and Doublet Detection for Single Cell Genomics.” Genome Biology 19 (1): 224.

Demultiplexing oligonucleotide-labeled scRNA-seq data with demuxmix