GSEA on single-cell cluster data

Question

[roy.granit]

Hello Biostars, I have 10x SC expression data which I have processed using Seurat, and now I wish to test the association of several gene-sets with certain clusters. What approach/R package would you recommend?

I suppose I can run GSEA on the expression of individual cells, but next how do I summarize a cluster?

Thanks!

[Martombo]

GSVA tries to estimate the pathway expression levels in every individual sample and then allows you to perform comparisons or unsupervised analysis like clustering on the transformed matrix.

Answer 1

[igor]

Some options:

• Take the top X cluster markers (genes upregulated in that cluster) and run GSEA (or any other pathway analysis) on those.
• Calculate the fold changes between that cluster and other clusters and use that as a basis for a pre-ranked list for GSEA.

[roy.granit]

Thanks, not sure about these suggestions, I believe they are correct for pathway over-representation but not for hyper-geometric/GSEA analysis. I was seeking some tool that would take the individual GSEA scores for each cell in a certain cluster and conduct a statistical comparison of these to the scores by cells in rest of cluster/other cluster. Something like GSVA which was suggested here above..

[igor]

Due to the dropouts, the result for individual cells will be extremely noisy. Cluster-level analysis would be much cleaner.

GSEA is not meant for over-representation, but I have seen it used that way many times. Regardless, as I mentioned, there are many other pathway analysis tools that are appropriate for that approach.

Single-sample enrichment approaches like GSVA are fine. However, unless the differences between your sub-populations are dramatic, they may not be able to effectively capture them. This is why I suggested to focus on the genes that are actually differentially expressed.

Answer 2

[marcovicari3]

Hi Roy. What do you think about pagoda?

http://hms-dbmi.github.io/scde/pagoda.html

Best, Marco