Preranked GSEA permutation

Question

[as823jk]

Hi,

I have a theoretical question about how preranked GSEA computes its null distribution of enrichment scores. My understanding is that this is done by randomly drawing genes from the background pool of genes to form random gene sets of the same size as the original gene set.

Would there be any meaningful difference if I shuffled the weights (e.g. log2FC) linked to each gene instead? So instead of randomly drawing genes to create an artificial gene set, I would randomly draw weights for each gene in the original gene set.

Thanks!

Answer 1

[ATpoint]

From what I understand, what you describe is how the fGSEA package implements its permutation test, randomly drawing genes to get a background distribution. Since it's entirely rank-based, there is no weights. It's simply a ranking. Each gene has the same weight. I generally find that fGSEA returns inflated, like overly small p-values, and I these days generally prefer the camera gene set enrichment test from the limma package. This either comes as a parametric test, directly starting from the count matrix, or it also has a pre-ranked version that you can use. But it does, from what I understand, not do gene permutation, but sample permutation, thereby giving more realistic p-values. There is a number of posts of the limma author, both here at Biostars and over at the Bioconductor Support site, where he recommends exactly this. And recently, using camera more often in parallel with fGSEA, I tend to agree to this. It gives more, at least in my hands, realistic results. I would always recommend you to actually plot the results, be it with the fGSEA plotting routines or limma, or custom, to confirm that the returned p-values are meaningful.