no empty droplet

[jessicaliu70]

cellranger:

code used:
cellbender remove-background
--input $paths
--output "./cellbender/$sample.h5"
--cuda
--total-droplets-included 40000
--expected-cells 15000
--fpr 0.01
--epochs 300
--z-dim 100
--z-layers 500
--empty-drop-training-fraction 0.3
--low-count-threshold 5;

log:
cellbender:remove-background: CellRanger v3 format
cellbender:remove-background: Trimming dataset for inference.
cellbender:remove-background: Including 32067 genes that have nonzero counts.
cellbender:remove-background: Prior on counts in empty droplets is 5
cellbender:remove-background: Prior on counts for cells is 2290
cellbender:remove-background: Excluding barcodes with counts below 5
cellbender:remove-background: Using 14667 probable cell barcodes, plus an additional 24667 barcodes, and 58825 empty droplets.
cellbender:remove-background: Largest surely-empty droplet has 121 UMI counts.
cellbender:remove-background: Running inference...

[sjfleming]

This one is pretty similar to #187 I would say. I think that the main problem is a failure to identify the prior on counts in empty droplets. This should be able to be fixed with --low-count-threshold 50 for this sample as well. (I might also try --expected-cells 10000, which can help cellbender identify the "really good" cells in order to form its prior on cell counts... but it might not matter, since it seems cellbender got that prior pretty close to correct anyway.)

Also, if you keep seeing the test ELBO drop down and deviate from the train ELBO like what is shown above, I might suggest trying to reduce --z-dim to something maybe closer to --z-dim 20.

[sjfleming]

These kinds of issues will hopefully be solved in v0.3.0

Potentially closed by #238