How to calculate the absolute p-value for GSEA which returned 0.0 as the p-value?

Question

[Riq]

I am trying to perform a GSEA for ten biochemical properties to check which properties are enriched for my set of genes. I am using the Windows application (v4.2.2) with GSEAPreranked package. However, when I perform the analysis some of them returns a p-value of "0.0". When I checked the user guide it states:

"In the GSEA report, a p value of zero (0.0) indicates an actual p value of less than 1/number-of-permutations. For example, if the analysis performed 100 permutations, a reported p value of 0.0 indicates an actual p value of less than 0.01. For a more accurate p value, increase the number of permutations performed by the analysis. Typically, you will want to perform 1000 permutations (phenotype or gene_set). (If you attempt to perform significantly more than 1000 permutations, GSEA may run out of memory.)"

I tried increasing the permutation to 10,000 and 100,000 which still yielded a "0.0" value. Is there a way to calculate the absolute p-value which I can use to later calculate FDR q-value?

Answer 1

[LauferVA]

Hello Mahasish,

The answer to your question is implicit in the question itself.

GSEA is a form of permutation based testing. The p-value is calculated empirically by permuting the gene labels at random, then seeing how many times out of the total number of permutations a random permutation is more extreme than the actual result you generated.

example: if you run 10,000 perms, and the actual data has enrichment for a pathway more extreme than all but 13 of the permutations, then the p-value would be 13/10000 = 0.0013.

If you are generating a p-value of 0.0, this means that in 0 out of 10,000 cases was the random permutation more enriched than your actual data for that pathway.

There are two solutions I'd consider.

1) Record the p-value as p < 0.0001 2) Re-run the pathway analysis with 10,000,000 permutations. This will likely give you at least a few perms that are more enriched than your data.

If you choose 2), then I do agree with rpolicastro's comment, above: fgsea is the way to go.

Finally, are you only testing one pathway? If not, you will need to control for multiple testing.

VAL

[Riq]

Thank you Vincent for your answer. When I increase the permutation to 1,000,000 the software cannot handle that many permutations and return an error "Java Out of Memory". I am testing for one pathway only. I have used "fgsea" package to calculate the really small p-values.

Answer 2

[rpolicastro]

You may want to consider switching to fgsea which has algorithmic improvements to GSEA for better p-value estimation [PREPRINT]. It helps to resolve most cases where the p-value is really small.

[Riq]

Thank you so much. "fgsea" was able to calculate the really small p-value for me.