Turning off downsampling in SpaCET.deconvolution.matched.scRNAseq

[ahwanpandey]

Hello,

I already have a downsampled dataset. How do I turn off downsampling in "SpaCET.deconvolution.matched.scRNAseq" to save some time and memory?

I thought using a very large value of sc_nCellEachLineage would skip it, but seems it is still doing something in the background?

Thanks!
Ahwan

[beibeiru]

Hi, @ahwanpandey ,

Please reinstall our package and run it like this.

SpaCET_obj <- SpaCET.deconvolution.matched.scRNAseq(
    SpaCET_obj, 
    sc_counts=sc_counts, 
    sc_annotation=sc_annotation, 
    sc_lineageTree=sc_lineageTree, 
    sc_downsampling=FALSE, # this is a new parameter
    coreNo=6
)

Best,
Beibei

[ahwanpandey]

This is great, thanks! I just tested and works great.

So testing with 1,000 cells per lineage is very fast and smooth. However, using 10,000 seems to be very slow. Is this expected behaviour?

One reason I wanted to use more cells is because that is what I had used for RCTD and wanted to compare at similar cell numbers.

Thanks again.

[beibeiru]

Which step is very slow? Down-sampling or Create the reference?

Best,
Beibei

[ahwanpandey]

This part, creating the reference. It has been at 0% for the last 15 minutes and it is still at the first lineage. This only happens with 10k cells each lineage. 1k cells each lineage works perfectly fine. I have already downsampled the data beforehand.

[ahwanpandey]

Oh and I am running with 8 threads.

[beibeiru]

The SpaCET algorithm compares one cell type against all others individually to identify cell-type-specific markers.
As a result, the computation time increases with the number of cell types.

To address this, I enabled the downsampling parameter. Based on my evaluation, downsampling has a limited effect on the results while significantly improving efficiency.

Best,
Beibei

[ahwanpandey]

Totally understand what you are saying.

The original reference is really huge and I had already downsampled it to two subsets; 1k cells per cell type and 10k cells per cell type. So having the downsampling parameter to be optional in SpaCET is very useful to save time. Thanks for that.

So the reference I am using has 9 cell types. When using the reference that I downsampled to 1k cells per cell type (so total of 9k cells in reference) works great and is very fast.

But when using the same reference downsampled to 10k cells per cell type (total of 90k cells in reference), it is stuck at the section in the screenshot above. Yesterday it was stuck at that part for more than and hour which is why I cancelled it. Let me see if I can let the job run for longer and see how long it takes.