Remove contamination across cell clusters

Question

[i.am.filippov]

I've got an scRNA-seq dataset filtered for immune cells. After applying the Seurat workflow I've discovered that some non-immune markers are expressed by cell subsets in almost all the clusters. See heat map:

I'm fairly certain that those cells are contamination. How would you filter those out based on expression values?

Thanks in advance!

[rpolicastro]

Why do you think those are contaminating cells?

[EagleEye]

How can you say that these are contaminated? For which cell types the selections are done in scRNA-seq and which non-immune cell types you are getting?

[i.am.filippov]

rpolicastro @EagleEye that's a good question! The cells were sorted by CD45+ before sequencing. Some epithelial cells were detected nonetheless, based on EPCAM+. Now, these cells expressing salivary markers - I'm not 100% sure they're contaminating. It might as well be that the immune cells in the tissue are expressing some tissue-specific genes. But my original question is more about the technical process of removing these cells if they indeed are contamination.

Answer 1

[firestar]

There are many ways to do this. Here is a working example.

library(Seurat)
pbmc_small <- pbmc_small %>%
              NormalizeData() %>%
              FindVariableFeatures() %>%
              ScaleData()

VlnPlot(pbmc_small,features="PPBP")

Let's say we want to remove cells expressing PPBP above 5.5. Here is one way:

# get positions of genes that match condition
x <- which(GetAssayData(pbmc_small)["PPBP",]<5.5)
# subset to new object
pbmc_small1 <- subset(pbmc_small,cells=x)

VlnPlot(pbmc_small1,features="PPBP")

Here is another way. This way, you can plot your good and bad cells on a UMAP for example before deciding to remove it.

# save good and bad cells as a new metadata variable
pbmc_small$good <- GetAssayData(pbmc_small)["PPBP",]<5.5
# UMAPPlot(pbmc_small,group.by="good")
pbmc_small1 <- subset(pbmc_small,subset=good==TRUE)

[i.am.filippov]

Thanks, this makes sense! I was also wondering if there's some semi-automated threshold detection method that is used in practice.