It would help if you could explain your specific scenario in more detail. What exactly are you trying to do?

I've worked with pathway analysis (a little), and I gather the important tests are about random graphs. Given some gene co-expression graph, and the set of known genes, we could construct random gene expression graphs (null hypothesis of random association), and see how unlikely the experimentally derived graph is.

If you found Ubiquitin connects to Ubiquitin, that should have happened by chance quite easily, and youll find p>0.50. If you find a complex mesh of a dozen otherwise poorly connected things, that will be extremely unlikely by chance, and has a p<0.01.

Of course these general examples are susceptible to all sorts of errors, random graphs clearly are not biologically relevant, and are subject to sampling and selection bias. It's just the best tool we have at the moment.

I'm no expert, but I don't think this is a solved problem, and these pathway based p-values could use a lot more research.