scDblFinder identifies doublets in single-cell RNAseq directly by creating artificial doublets and looking at their
prevalence in the neighborhood of each cell. The rough logic is very similar to
DoubletFinder, but it is simpler and more efficient. In a
nutshell, instead of creating doublets from random pairs of cells, scDblFinder first overclusters the cells and
create cross-cluster doublets. It also uses meta-cells from each cluster to create triplets. This strategy avoids
creating homotypic doublets and enables the detection of most heterotypic doublets with much fewer artificial doublets.
We also rely on the expected proportion of doublets to threshold the scores, we include a variability in the estimate
of the doublet proportion (
dbr.sd), and use the error rate of the real/artificial predicition in conjunction with
the deviation in global doublet rate to set the threshold.
The approach described here is complementary to doublets identified via cell hashes and SNPs in multiplexed samples.
The latter can identify doublets formed by cells of the same type from two samples, which are nearly undistinguishable from real cells transcriptionally (and hence unidentifiable through the present package), but cannot identify doublets made by cells of the same sample.
scDblFinder was developed under R 3.6. Install with:
if (!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager") BiocManager::install("scDblFinder") # or, to get that latest developments: BiocManager::install("plger/scDblFinder")
Given an object
sce of class
SingleCellExperiment (which does not contain any empty drops, but hasn’t been further filtered) :
# we create a dummy dataset sce <- scater::mockSCE(ncells=500, ngenes=500) library(scDblFinder) sce <- scDblFinder(sce, verbose=FALSE)
This will add the following columns to the colData of
sce$scDblFinder.ratio: the proportion of artificial doublets among the neighborhood (the higher, the more chances that the cell is a doublet)
sce$scDblFinder.weighted: the proportion of artificial doublets among the neighborhood, weighted by distance
sce$scDblFinder.score: the final doublet score
sce$scDblFinder.class: the classification (doublet or singlet)
If you have multiple samples (understood as different cell captures), then it is
preferable to look for doublets separately for each sample (for multiplexed samples with cell hashes,
this means for each batch). You can do this by simply providing a vector of the sample ids to the
samples parameter of scDblFinder or, if these are stored in a column of
colData, the name of the
column. In this case, you might also consider multithreading it using the
library(BiocParallel) sce <- scDblFinder(sce, samples="sample_id", BPPARAM=MulticoreParam(3)) table(sce$scDblFinder.class)
The important sets of parameters in
scDblFinder refer respectively to the expected proportion of doublets, to the clustering, and to the number of artificial doublets used.
The expected proportion of doublets has no impact on the score (the
ratio above), but a very strong impact on where the threshold will be placed. It is specified through the
dbr parameter and the
dbr.sd parameter (the latter specifies the standard deviation of
dbr, i.e. the uncertainty in the expected doublet rate). For 10x data, the more cells you capture the higher the chance of creating a doublet, and Chromium documentation indicates a doublet rate of roughly 1% per 1000 cells captures (so with 5000 cells, (0.01*5)*5000 = 250 doublets), and the default expected doublet rate will be set to this value (with a default standard deviation of 0.015). Note however that different protocols may create considerably more doublets, and that this should be updated accordingly.
Since doublets are created across clusters, it is important that subpopulations are not misrepresented as belonging to the same cluster. For this reason, we favor over-clustering at this stage. This is for instance implemented by scDblFinder’s
overcluster function, and controlled by specifying minimum and maximum cluster sizes. Alternatively, cluster labels can be directly provided.
scDblFinder itself determines a reasonable number of artificial doublets to create on the basis of the size of the population and the number of clusters, but increasing this number can only increase the accuracy.
If the input SCE already contains a
reducedDim slot named ‘PCA’, scDblFinder will used them for the clustering step. In addition, a clustering can be manually given using the
clusters argument of
scDblFinder(). In this way, seurat clustering could for instance be used (in which case we suggest to increase the
resolution parameter) to create the artifical doublets (see
?Seurat::as.SingleCellExperiment.Seurat for conversion to SCE).
To benchmark scDblFinder against alternatives, we used datasets in which cells from multiple individuals were mixed and their identity deconvoluted using SNPs (via demuxlet), which also enables the identification of doublets from different individuals.
The method is compared to: