Hi,

I am trying to perform benjamini-hochberg adjustment for the p-values obtained from running goana(). I am wondering which approach is appropriate: 1) get all GO terms with p-values < 1 and adjust them altogether. 2) get all GO terms with p-values < 0.1 (or 0.05) and adjust them altogether. 

Going with one or the other would mean that I would be getting FDR values that are different. Ideally I would like to filter the GO terms below a threshold of FDR < 0.1 and because there are some GO terms that are close to the threshold, I'm trying to decide whether I should consider them insignificant and chuck them away. What would be appropriate? Thank you.

The whole point of adjusting for multiple testing is to correct for the total number of tests that were performed. You can't just throw away tests with large p-values, that would defeat the entire purpose of the correction. (The exception is with independent filtering, but that is not applicable here - the p-value is hardly independent of itself.) In short, you should adjust based on all the tests that were performed, i.e., all GO terms that were used.

P.S. It's debatable whether the term-level FDR has much meaning for GO enrichment analyses. This is due to the fact that GO terms lie in a hierarchy, with more general terms at the root and specific terms at the leaves. If one node is a true positive, then it could be argued that all ancestors of that node are true positives as well. Thus, if you have enough detection power, you might end up with many significant GO terms, all of which are actually driven by a single biological process. This leads to an inconsistency between the term-level FDR (which the BH method controls) and the FDR across distinct biological processes (which is of more interest).