I don't think it makes sense to do this. Just because a gene is up-regulated or down-regulated does not mean that the process, function or component that the gene is annotated to is also respectively directionally regulated. There is no directionality associated to how a gene impacts a GO term (aside from some GO terms having directionality like "negative regulation of cell proliferation"). This is one of the limitations to GO analysis. It is specifically for this reason that one should consider all DEGs when performing a GO analysis. Then use advanced pruning algorithms like Elim or Weight as proposed by Alexa et al (2006) to identify truly enriched GO terms.

You can read more about this here: http://eepurl.com/bubBe9

Check this: "Separate enrichment analysis of pathways for up- and downregulated genes"

Yes, when the enrichment software you're using doesn't consider the fold change, you should run up and down genes separately. I think this is because GO enrichment usually looks for significant enrichment with respect to the whole set of GO annotations for that species.

You should take care to check for similar/same terms between groups. Otherwise, look for GO/pathway software that considers fold change.

I'd like to restart this discussion because I didn't quite understand the result (contradictory answers).

Is there consensus now, 5 years after this discussion? What should we do?

Since I'm doing GO annotation just now, I came up with the same question and landed here. Here's my 2p, without too much thinking...

In favour of keeping up- and down-regulated genes together: other things being equal, you should have more power because you rely on larger sets of DE genes. Also, you test fewer hypotheses so fewer chances of false-positive calls.

In favour of separate analysis: imaging you have 5 genes up-regulated and they are all in the same pathway; you also have 50 genes downregulated scattered across pathways. I think that testing the 55 genes together will obscure the pathway enriched for the 5 genes. Is this a far-fetched scenario? Dunno, but I don't think so... You can have a treatment that specifically affects a pathway and many genes move in the same direction in that pathway (e.g., genes associated with DNA replication). This in turn causes havoc in a number of other pathways in a seemingly random manner.

I think I'm leaning towards separate analyses but I can easily be persuaded otherwise. In practice, try both and try to pick the most sensible results without fooling yourself.