Warnings, parameters tuning and precaution regarding overcorrection #419
Description
Hello!
Thank you for the tool and all the work you do explaining details of its functioning to the community.
I have rather typical sample that should be well suited for the CellBender workflow but It would be nice if you could explain a few things to me in general about CellBender behavior.
Experimenting with the available settings I was able to find optimal values I believe good for this sample. Training curve always have a little spike in the beginning but in the end it looks pretty normal.
Q1: Should we be worried about convergence rate? It is said in the documentation that optimal values are 0.25-0.35 but is there a threshold I could definitely say that it is the case when we should experiment with learning rate ?
Q2: Is it more beneficial to add more surely empty droplets? Will it improve the performance if one set total-droplets-included to the value in the end of the plateau?
I have tried this but on different samples it seems to have different effects but overall I noticed that learning curve is getting more deviations from the expected line.
In every run for all my samples and all the parameters I see A LOT OF warnings about changed gene counts.
WARNING: The expression of the highly-expressed gene X decreases quite markedly after CellBender. Check to ensure this makes sense!
It is a little bit confusing. So many genes is not possible to check by hand, among ribosomal and mitochondrial there are other meaningful genes.
Q3: How can I be sure that expression is not overcorrected? There are genes that a dramatically changed after CellBender with 99% counts being removed.
Thank you! Appreciate if you could find time to explain this little things to me.
Evgeniia