GSEA analysis in R

Question

[rajdeepboral00]

I have used DESeq2 to perform differential gene expression study and saved the result. For using GSEA of clusterProfiler package, it needs a rank list, now the result of DESeq2 should be ranked on the basis of log2foldChange or test statistic? ANd if log2FC then should i shrunken the l2FC?

Answer 1

[ATpoint]

It doesn't matter really. Both shrunken logFCs and test statistics are fine and there is no real justification of one over the other. This was asked many times before.

[rajdeepboral00]

So, basically when i ranking them on the basis of shrunken log fold change , the analysis is ceoming non-significant and if i ranked them on the basis of test statistics, it is becoming significant. So, what to trust now?

[ATpoint]

None I guess, these metrics are similar so it's concerning and indicates suprious calls. clusterProfiler uses fGSEA, and I remember posts from the limma author mentioning that (I hope I quote correctly now) that in his opinion PValues/FDRs from fGSEA (gene-permutation based tests) are strongly exaggerated, so you should probably consider cutoffs way smaller than usual 0.05. I personally use something in the realm of 10^-5 for fGSEA, but lately only use the camera geneset test from limma as it feels more robustish (not that I have data to show that, but concept-wise I like it more than gene-permutations).

As for clusterProfiler/fGSEA, look at the data, plot the enrichments (see https://bioconductor.org/packages/devel/bioc/vignettes/fgsea/inst/doc/fgsea-tutorial.html under "One can make an enrichment plot for a pathway") and check whether the geneset really shows good shifting to either the left or righthand-side of the plot, rather than relatively even coverage across the ranking range, regardless of the pvalue.