K-S null hypothesis for GSEA

Question

[mubangafchama]

Hello Could someone please tell me whether the Kolmogorov-Smirnvo used by GSEA (broad institute ) uses a competitive or self -contained null hypothesis. Also, is there a general consensus on which of the two null hypotheses is best for functional enrichment scoring etc. Thank you

[Brian Bushnell]

I can't answer your question, but I can tell you that I have doubts about Broad's software's statistical analyses. Broad software routinely gives variant calls Phred-scaled scores of over 2000. That means there is a 1 in 10^200 chance of it being incorrect. Considering that there are generally considered to be far fewer than 10^100 atoms in the universe, that seems a bit high to me. I think betting on the outcome of such a call is no different than betting on any other similarly random event - e.g., that you spontaneously disappear via random electron migration, or that the Sun explodes tomorrow.

[mubangafchama]

Thank you Brian. Food for thought

Answer 1

[Jean-Karim Heriche]

GSEA uses the competitive type of null-hypothesis but see this paper which discusses the different statistical approaches to gene set analysis. AEA is another method that is worth considering.

[mubangafchama]

Thank you Jean. I had gotten a mixed up after reading somewhere that another way of looking at a competitive hypothesis is that it uses genes (and not phenotype) as a sampling unit, but GSEA has a parameter where you can select to permute by phenotype.