When should a pathway enrichment result be treated as exploratory rather than mechanistic evidence?

Why

An over-representation test asks whether more genes from a pathway sit in your hit list than chance would predict. Chance depends on the background, the set of genes that could have been called. A survey of 186 articles found that 95 percent of over-representation analyses did not use an appropriate background or did not describe it, and 43 percent did not correct for testing many gene sets.

Using the whole genome as background for a tissue makes tissue-typical pathways look enriched, because only expressed genes can ever be called. Timmons and colleagues show liver GO terms appearing for what was simply a list of liver-expressed genes. RNA-seq adds length bias: longer genes have more power to be called, and goseq corrects the test for it.

Even a carefully done enrichment only says the list resembles a pathway's membership. It does not show the pathway's activity changed, which direction it changed, or that it causes the phenotype. Treat it as a ranked set of hypotheses, then test the strongest with perturbation, protein or activity measurements.