Should I remove these cells? (scRNA 10x 5')

Question

[msn]

Hello all, quick question today.

I did some clean up on some cells , nFeature_Genes > 200 & nFeature_Genes < 4500 & percent.mt < 20 , nothing too extreme , there is this one sample with a chuck of cells outside the linear norm and I am not sure if they biologically relevant as interesting and should be kept or if their ratio of molecules detected to genes detected should be ringing red flags for me.

I have circled the the cells of interest in the plot. note the axis labels as this is already trimmed data. I have also shuffled the printing of the dots and am certain it is sample specific-ish. the ratio doesnt fit for doublets (scran detectdoublets confirms) , mito counts are 0, which doesn't mean they aren't dead cells per say, could be part of the library prep artifact killing cells with a different phenotype of necrosis vs apoptosis (mito graph subset on the single sample).

Other hypothesis is that its empty GEMs, and its ambient RNA from a massive death somewhere in tissue processing/library prep. BUT then you would expect all other good gems to have the same molecule counts. So dont think that's it either.

Thanks for your help and insight

[LChart]

There seem to be very few genes expressed but at a high count. What genes are actually being expressed in this cluster? Do these cells show up on one of the top 10 or 15 PCs? Maybe the loadings from those will help.

[msn]

I'll check the loadings, thanks!

Answer 1

[ATpoint]

I personally try to combine QC with celltype information. You usually have an idea which "crude" celltypes you expect in your data, so I typically run something like UCell to score the cells against a selection of canonical markers and the do these sorts of plots per celltype. For example, if this population of cells you circle would be something like a macrophage that expresses genes in the thousands then it would probably be a chunk of damaged cells. If this was erythrocyte contamination it could be "normal" as these cells have no nucleus and only limited RNA left in the cell. Number of genes is vastly different per celltype. We work a lot on neutrophils, there we often only see about 1000 genes on average per cell, while something like vascular endothelium, macrophages or fibroblasts often express many more genes. That having said, using a single cutoff over heterogeneous cells can remove interesting biology, try to do it per-celltype or at least run a crude clustering and do it per-cluster. If a cluster looks odd in QC metrics then see if you can make sense of it via markers, and if not consider removing it.

[msn]

Thanks I will look into this!