Separate runs for up/down

[christophista]

Hello

I have recently played around with clusterProfiler::gseGO. The function needs a ranked list of genes (value is rank, name is gene) to perform the GSEA analysis. As I understand it, per default it uses all genes of an experiment at a time, hence those that have higher or lower expression relative to a control (and hence in my case a positive or negative rank as I am using the log2FC as the rank criteria).

I also separate the genes into those that have a positive and those that have a negative log2FC and ran the GSEA separately. It worked, but that is no quality indicator whether it should be done like this or if the generated results are valid and meaningful. Also, all tutorials and guides never suggested this strategy. The only reference in that respect that I found was
https://royalsocietypublishing.org/doi/full/10.1098/rsif.2013.0950?rss=1
however, the authors only separated for GO terms, which is quite different than GSEA.

What I observed was:

• The overlap of doing all genes at once and doing up/down separately was not too great.
• Some terms were present in the up and down likewise

Both observations are probably fine, i.e., separate lists increase the power of the GSEA to a certain extend, and genes of the same term can have up and down directions, so it is not a surprise.

Thus I would like to kindly ask for your opinion on this.

• Is the strategy valid?
• What are your thoughts on doing the analysis separately?
• Would you recommend it?

Thank you in advance!