How to deal with "The expression of the highly-expressed gene _ decreases quite markedly after CellBender. Check to ensure this makes sense!" warning?

[Ratoncito]

I just ran 69 samples through cellbender with default settings and noticed that in about 1/10th of my runs I get a warning like this:

WARNING: The expression of the highly-expressed gene Avp decreases quite markedly after CellBender. Check to ensure this makes sense!

In some cases it even includes biologically interesting genes such as Oxt, but I digress...

Just wondering if I should take this warning seriously or how would I "ensure this makes sense". Any feedback welcome.

[sjfleming]

Hi @Ratoncito , yes we were a bit vague about that "ensure this makes sense" statement. And it does show up a lot! Including in cases where it's totally fine. The noisier the dataset, the more likely this will show up.

It is not meant to be an error, just a warning.

As for how to follow the advice and "ensure this makes sense", I'd recommend looking at some version of a plot like this:

where you label your gene of interest and sure it's not way off the diagonal. The diagonal here is kind of like our naive expectation about how much each gene should be removed.

• x-axis: take some droplets you're sure are empty (from the empty droplet plateau) and take the mean expression over cells
• y-axis: take the mean (over cells) counts removed for each gene np.array((adata.layers['raw'] - adata.layers['cellbender']).mean(axis=0)).squeeze()
• (see here for some more detailed code)

Also look at UMAPs for those genes... UMAPs like this:

Make sure it doesn't look like cellbender is unfairly just destroying all counts for a gene. (As of v0.3.0, it should be protected against doing that... but I still like to tell people to check.)

Do I always "check to make sure"? No, if I'm running v0.3.0+, I do not always check. But if I see a gene that is "biologically interesting such as Oxt", then for sure I will check on that one.