I was a little surprised that the P-value histograms are, if anything, 'worse' after gene filtering. You can clearly see the P-value distribution is not uniform for C120_batch13_1 without gene-filtering but looks reasonable-ish for the other 4 captures. In these plots, the top row is without gene-filtering and the bottom row is with gene-filtering. The same is true without and with gene-filtering. Nothing seems amiss in the barcode plots, with the knee being well-identified in all 5 captures. In these plots, the top row is without gene filtering and the bottom row is after filtering out mitochondrial, ribosomal protein, TCR, and BCR genes.

MA plot comparing the retained low-count barcodes to the discarded barcodes.A histogram of P-values for low-total barcodes.

Captureįollowing the advice in OSCA Advanced 'Testing for empty droplets' and 'emptyDrops behaviour when excluding mitochondrial/ribosomal genes', I made: Interestingly, Cell Ranger (v7) reports ~20,000 cells for this capture, so presumably they do some gene-filtering before applying their implementation of the Empty Drops algorithm. The other captures within this batch have similar and expected numbers of non-empty droplets.Īfter excluding mitochondrial, ribosomal protein, TCR, and BCR genes (hereafter referred to as gene-filtering), emptyDrops() returns a number of non-empty droplets for C120_batch13_1 that is more like what I expect (~20,000). The number of non-empty droplets in this capture is 3x larger than is reasonable/expected (~65,000 rather than ~20,000 non-empty droplets from 1 lane (capture) of 10x 5' scRNA-seq data). The problematic capture is capture 1 of batch 13 ( C120_batch13_1).

Within a batch, each capture is roughly a technical replicate because the cells come from the same population (a tube containing a mixture of PBMCs from different donors labelled with HTOs) which are then distributed across the 5 captures.

12 batches (named 3-14 for reasons), with each batch containing 5 captures, and targeting ~20,000 cells/capture.

Human PBMCs profiled with 10x Genomics 5' chemistry (v2), including gene expression (GEX), a large panel (BioLegend's TotalSeq™-C Human Universal Cocktail) of antibody derived tags (ADTs), and VDJ sequencing of T-cell receptors (TCRs) and B-cell receptors (BCRs).

In my case, the knee point is being properly identified, but there are still too many non-empty droplets.Īnother point of difference to the earlier report is that the genes driving the differences between the retained low-count barcodes and discarded barcodes in this capture are T-cell receptor (TCR) genes, which I suspect may suggest the need for assay-specific or sample-specific gene filtering. I found similarities to a previously reported issue ( 'emptyDrops behaviour when excluding mitochondrial/ribosomal genes') whereby the performance of emptyDrops(), and in particular the identification of the knee point in the barcode rank plot, was improved by excluding mitochondrial and ribosomal protein genes (this is now documented in the 'Non-empty droplets versus cells' section of the man page for emptyDrops()). Initially, I noticed this affecting 1/60 captures in a large scRNA-seq experiment I'm analysing.īut upon investigating the issue further I'm wondering whether I should also revise how non-empty droplets are being identified in the other 59/60 captures (and more generally in other similar-ish experiments).

I've encountered a situation where DropletUtils::emptyDrops() was calling 'too many' non-empty droplets and would appreciate some advice.

Codelink v2 violations