How to compare pathway significances after treatment?

Question

[Cancan]

Suppose there is a data set which records protein information at time point 1 and 2. At time point 1, all subjects protein information are recorded. Right after time point 1, some subjects receive treatment or placebo. Then at time point 2, subject protein information are recorded again. For individual time point 1 and 2, I can perform over representation analysis at each time point. However, I do not know how to compare whether the treatment significantly affect pathways.

How do I compare pathways over representation analysis affected by treatment? I would like to conclude some pathways are inhibited or promoted by treatment.

[Cancan]

The protein information are protein level at time point 1 and at time point 2.

Answer 1

[jared.andrews07]

For RNA-seq, you do this by performing a DE analysis with the timepoint accounted for in your design (as shown/described in the DESeq2 vignette, design would be ~timepoint + treatment), then use the resulting DE list for GSEA via fgsea or similar.

I don't have good knowledge of proteomics, but I assume there must be similar packages and methods used.

[Cancan]

Maybe this is irrelevant. A lot of pathways do overlap. So technically one would use mixed effect model in conjunction with DE analysis?

[jared.andrews07]

I don't really understand how your comments relate to each other here, but will try to address them anyway.

Yes, lots of pathways overlap. There are ways to "collapse" pathways have majority overlap in their members to a single term, see the collapsePathways function from fgsea. I find this a little messy and tend to prefer to return all results and just cherry pick representative ones in figures while providing the full list in a supplemental table.

Both DESeq2 and edgeR are fixed effect models, though you may be able to handle one factor as a random effect with limma. In this case though, I don't believe you have any random effects, as well-described in this Stats.SE post.

[Cancan]

I was thinking that your dependent variable is set to be protein levels. Then dependent columns are becoming correlated due to overlap of pathways. This correlation requires usage of mixed effect model. If there is no correlation, I would have not objection to the statement. For ANOVA, one needs to check residual independence.