scRNA-seq doublets with less features and UMIs than singlets? How?

Question

[naughtoncolin]

I'm attempting to remove doublets from one of my samples by removing any cells that express two markers that shouldn't be seen in the same cell e.g. both a hematopoietic marker (CD45) & epithelial marker (EPCAM). Initially, the sample's violin plots for nFeature_RNA & nCount_RNA had bimodal distributions. I thought that the upper portions of these bimodal distributions would be comprised of doublets, but removing cells based on the coexpression of these markers reduces the lower portions of the distributions.

If these cells are in fact doublets, why would they have fewer genes and UMIs detected than singlets?

Answer 1

[jared.andrews07]

There are a few possibilities I can think of off the top of my head.

1. The cells being removed are a legitimate population/cell-type. So not actually doublets.
2. They are actually doublets, which doesn't guarantee they'll have more genes detected than your other cells, though they'd likely have a better chance than average of doing so.
3. There's ambient RNA contamination from one cell type or the other leading to certain cells getting erroneously labeled as doublets.

I'd recommend using one of the doublet finder tools (scDblFinder, DoubletFinder, etc) to do this, as they'll simulate what doublets (or triplets) would look like in your data and identify cells with those profiles. Generally, they seem to work pretty well. You may also be interested in reading the relevant sections on "ambient RNA" in OSCA.

Answer 2

[igor]

Everything @jared.andrews07 mentioned is correct, but I would add one more point.

All the cells are of highly variable quality. You may have a true singlet with 500 genes and another with 5000. You can also have a doublet with two 1000-gene cells, which will fall between those two singlets in the distribution.

There was a nice plot to illustrate this by Stoeckius et al (Fig 1E):

The doublets clearly have more UMIs than singlets on average, but you would not be able to draw a cutoff somewhere to clearly differentiate the two populations. For example, at around 1000 UMIs, you are in the thin part of the singlet distribution, but in absolute numbers, there are still more singlets there than doublets.