The dark side of enrichment analysis: Confusing term names, overlapping pathways, and study bias

[@kanefoster8780]

love the idea of this series! really important to consider the downsides of these approaches

[@BDBL9]

Exactly,!! I have tried real hard to explain this understudied part on my defense. It was specifically noticeable when used wide range of GO hierarchy, but i tried level 2 or 3 only it made perfect sense. Thank you for the nice explanation.

[@larsjuhljensen]

Yes you can use levels or GO Slim terms. The problem is just that different levels don't really mean the same in different parts of GO; level 3 can be very broad terms in one part of GO and very specific terms in another part. I have in the past sometimes cut GO based on number of genes annotated with the term instead of the level in the ontology.

[@nirajshah2306]

Hi Lars, Thanks for this video. I have two questions: 1) In the differentially expressed gene list, there are many genes that does not p-adj values but they do have p-values. I do not include these genes as universe list for over-representation analysis. Am I doing the right thing by excluding these genes ? 2) On importing GO terms for the genes, one gene corresponds to many GO terms. Should I use all these GO terms for each gene or something else can be done ? Thanks..

[@larsjuhljensen]

1) All the genes that you have measured should in principle have p-values, most of which are not significant, some of which are only initially significant, and a smaller number of which are still significant after correction for multiple testing. You generally want to use the ones that are significant after correction for multiple testing as your list (foreground) and all the ones measured as the universe (background). So no, it does not sound like you are doing the right thing. 2) Regarding GO terms, it is not really something you should need to worry about. Generally speaking, an enrichment tool will use all of the GO terms for performing the enrichment analysis without you needing to do anything for that to happen.

[@Sara-gc9gu]

Would love to hear your thoughts on an ORA pathway analysis using Clusterprofiler, in particular the compareCluster function!

[@larsjuhljensen]

Over-representation analysis (ORA) is for all practical purposes a synonym for enrichment analysis. The terminology is not used consistently in the literature, with GSEA both being the name of a specific tool and the general concept of doing enrichment analysis with gene sets. I personally use it in the broad way, in which case GSEA and ORA refer to the same kind of analysis. I am not an expert on clusterProfiler specifically, but the caveats that I mentioned about pathway enrichment analysis should apply to all such tools, since it stems from the pathway databases that they use as input (i.e. overlapping pathways containing genes that are not pathway-specific). It is not immediately clear to me from the documentation what the compareCluster function does; it would appear to do an enrichment/over-representation analysis of each cluster, but I cannot tell what it uses as background. If it uses the complete gene list as background and thus looks at what each cluster is enriched for compared to the list as a whole, that can be a very good way to avoid the study bias issues (since the foreground and background sets would come from the same list and thus presumably have identical biases). That type of analysis can be very useful to functionally characterize clusters as I briefly talked about in my presentation about stringApp enrichment analysis (kzbin.info/www/bejne/d4aoqo2tYs59fsk).

[@Sara-gc9gu]

@@larsjuhljensen Thank you for your detailed reply! Yes, there is the option of using your own background in the function.

[@larsjuhljensen]

In that case I would suggest that you try using all clusters (i.e. your gene list) as background and each individual cluster as a separate foreground.

[@joelsteele1786]

Sara have you considered BiNGO? It works well for showing you what terms are really significant and their hierarchy.

[@larsjuhljensen]

@@joelsteele1786 Another option is to do enrichment analysis in Cytoscape stringApp and then send the enrichment results to EnrichmentMap (stringApp has a button for that).