Help with Seurat QC ambiguity

Question

I have four PBMC samples from 10X scRNA-seq

> cancer
An object of class Seurat 
36601 features across 18338 samples within 1 assay 
Active assay: RNA (36601 features, 0 variable features)
>

And this is Seurat QC plot

If I am not wrong, samples were of relatively low quality, with gene expression data revealing the presence of mitochondrial genes, as well as MALAT-1, which are suggestive of poor sample quality (dead/dying cells).

Any way by this plot I filtered cells to remove data that had more than 20 "percent" mitochondrial expression , > 2000 features or < 100 features

cancer <- subset(cancer, subset = nFeature_RNA > 100 & nFeature_RNA < 2000 & percent.mt < 20)

By this I lost most of cells

> cancer
    An object of class Seurat 
    36601 features across 6,883 samples within 1 assay 
    Active assay: RNA (36601 features, 0 variable features)
    >

Please, can somebody experienced in scRNA-seq inspect this plot and my thresholds and tell me if I am wrong, any suggestion, if losing this number of cells is normal

Thank you

Answer 1

Answer from @haci, converted from comment:

I think you should not be using a cut-off higher than 20 or 25 for the reasons I explained above.

I would advise against using cells with more than 20-25% percent.mito. This metric is used as a proxy for cell death and there is not much point keeping these cells if the cell death is so severe to the point where let's say half the reads map to the mitochondrial genome (with a 50% cut-off). There will be many caveats one being apoptotic cells from different lineages clustering together, not interesting unless you are studying apoptosis related phenomena. In my opinion 6000 good quality cells are better than 10000 apoptotic cells.

Answer 2

From the graphs, you're losing most of the cells through the percent.mitochondrial threshold. The distribution is so broad on that mitochondrial plot that I probably wouldn't filter for mitochondrial expression (except possibly at ~95%). If I did have a lower threshold, I'd put it at around 50% rather than 20%, because that's where there's a change in density on the mitochondrial plot.

This suggests that there's a very large mitochondria-rich population in the PBMC dataset. I don't have a good understanding of biology and cell populations, but if I were presented with these data I'd probably try clipping out those cells with 65-95% mtDNA and see if there is a strong enrichment for markers that correspond to any particular cell population (e.g. platelets).