Use marker genes from defined cell types as reference for CellAssign

[allyhawkins]

Stacked on #221

This PR is a continuation of the analysis started in #221. Here I broke out the comparison of using the top marker genes as a reference for cell assignment as a separate notebook. Before filing this, I addressed some of the concerns brought up in the initial review:

For the "circular" bit: how did you decide on 10 markers for each cell type? I also wonder if we might want to use "any" for these markers, or something in between: I kind of assume that most marker lists are some kind of a mix, where genes that mark a set of related types might be included in all of their marker lists. The fact that CellAssign uses a binary matrix certainly implies that it would expect some of that kind of data. (But does CellAssign require upregulation for a marker? I kind of assume so, but it is worth checking).

• Honestly, I don't really have a rationale for picking 10 markers. I just liked the number and was trying to grab the most highly expressed marker genes for each cell type without getting too many genes that may limit the ability to differentiate cell types.
• I didn't change the number of marker genes used, but I did slightly change how I grab them. Previously I wasn't doing any sorting based on log fold change. Here before creating the matrix, I only grab the top 10 markers with the highest positive fold change.
• I did read a bit more about CellAssign to confirm that the expectation of the marker genes is that they are highly expressed in a given cell type in comparison to all other cell types. See some of the text taken directly from the paper https://www.nature.com/articles/s41592-019-0529-1.
  • "CellAssign allows for flexible expression of marker genes, assuming that marker genes are more highly expressed in the cell types they define relative to others."
  • "Marker genes are assumed to be overexpressed in cell types where they are markers—not necessarily at similar levels—compared to those where they are not."
• I looked at the top markers identified for the Tumor_myoblast population and of the top 10 (I also looked at the top 20 on my own) none of them were genes that were mentioned in the paper. Also, none of them were Myoblast specific genes that I recognized (MYOG, MYOD1, MYF5). I looked a few up and most of them seemed more ubiquitously expressed.
• Despite none of them being myoblast-specific genes, using them to define Myoblasts was able to recapitulate cell type assignments fairly well. Using only the markers with the highest fold change meant the assignments were more consistent with the manual annotations, which is what I expect.

Here's a copy of the completed notebook: cell-assign-sarcoma-marker-genes.nb.html.zip

[allyhawkins]

Thanks for taking a look at this @jashapiro. I added some explanatory text around the analysis that was already here, but I didn't update any plots or any of the actual analysis. I will note that I did try to use any rather than all for the FindMarkers strategy and there was not a noticeable difference which is why I just kept using all. I do think the idea of doing a logFC cutoff is good, but I'm not sure we need it at this stage.