Deconvolution: Visium data

Once we have chosen a deconvolution method from #99, we can then move ahead and generate spot deconvolution results on the Visium data (n = 30) with that chosen method. So it'll be similar to #128 but instead of using the number of nuclei derived from cellpose on the DAPI channel from IF, we'll use the VistoSeg cell counts.

Remember that we decided not to run cellpose on the histology images (aka #95). It might be worth revisiting a bit the conclusions from @heenadivecha at #91. From what I recall, we know that tissue wrinkles affect the counts. I know that @kmaynard12 has manually annotated some of these wrinkles. But one option we had discussed was to cap the number of cells per spot at some maximum, like 20 or even 10. That would be the "precision" we would be willing to consider. For example with a cap of 20, we would be looking at max a 5% change (1 / 20 * 100 = 5). The other option I think we had considered was dropping spots with a very high VistoSeg cell count. But well, would the result be the same for a spot under the cap regardless of whether it's neighbor spot (above the cap) is included or not? I would imagine that the deconvolution results are independent for each spot, but I don't know, hm...

[ ] Update on number of cells cap vs dropping spots (either from Kristen's manual wrinkle annotations or due to a threshold on the VistoSeg cell counts).
[ ] Run the selected method from #99 on the n = 30 Visium samples. Unlike #128, I think that we'll likely focus on just the layer-level resolution here. But we might also run the broad cell type resolution too if at #99 we found that the results were quite different between the resolutions (I wouldn't expect this).

LieberInstitute / spatialDLPFC

Deconvolution: Visium data #129