cellbender failing on sample with ~150 cells

[ktpolanski]

Hello,

Sometimes you're stuck with a sample which doesn't have a lot going on. This is one of those. This is what the knee plot looks like for it:

I fed it into a relatively recent cellbender (installed on 08.06.2021, so only a few commits behind master), and got the following strange error about some tensor being empty:

cbk.5277079.txt

Is there anything that can be done to salvage this sample? Thanks and sorry for the trouble.

[sjfleming]

@ktpolanski you found a bug! Hmmm, I will have to keep that one in mind. The issue is that I try to compute an "average size" for cells on the first pass through the encoder neural network... but in this sample, by chance, there were actually no "cells" in the first minibatch. So I will need to fix this.

But for now, I bet there's something you can do.

Try this:

cellbender remove-background \
    --input TransTHY8814436/raw.h5 \
    --output TransTHY8814436/TransTHY8814436_knee.h5 \
    --expected-cells 100 \
    --total-droplets-included 1000 \
    --fpr 0.01 0.05 0.1 \
    --cuda

(The smaller --total-droplets-included will hopefully (1) let CellBender know the prior on empty droplet counts should be larger than it was estimating before, and (2) it will increase the ratio of cells to empties in each minibatch, hopefully side-stepping the bug.)

[ktpolanski]

Thanks for the suggestion! Our GPU queues are super congested and the job is still stuck as pending, I'll let you know once it goes through.

[sjfleming]

Hopefully these kinds of things are solved in v0.3.0 and much smoother for users.

Closed by #238