I recently performed RNA-seq analysis on auditory brainstem of mouse.The experiment involves wildtype mouse brains vs a transgenic mouse where two microRNA,s are knocked out. The idea was to see the difference in expression of genes between wildtype vs KO as microRNAs regulate gene expression. Problems: 1.The data is messy where the RPM values between 3 replicates are too different. The PCA plot is all over the place where samples from same area and same mouse don't group together. 2.The differential gene expression by edgeR reveals a total of 5 genes and none of those are interesting to me as future research candidates. The genes I identified from literature as targets of these microRNAs and would be interesting candidates for my research question are in the total list produced by edgeR but they are not significantly different.

My question is if we perform a gene ontology /enrichment analysis ,would it reveal the genes(as differentialy expressed) I am interested in as compared to an over all result that edgeR produces? For example now I have genes related to blood or calcium uptake as differentially regulated but I am interested in genes specifically related to inhibitory synapses and see if their expression is changed in KO samples compared to wildtype ones.

I hope the question is not too confusing since I don't know much about R and these analysis .The analysis was done by the company who did the RNA-seq. My apologies!

```

The analysis using edgeR is a by-gene analysis that specifically tests for evidence that a given gene is differentially expressed in one group versus another. All of the various gene set tests test for evidence that a set of genes is differentially expressed. A conventional over-representation analysis like a GO hypergeometric test won't be useful in this context, as it needs significant by-gene results.

You might be able to use something like mroast or fry to do self-contained gene set tests and achieve significant results for some interesting sets of genes even if none of the genes are themselves significant, particularly if you choose a more powerful method than the default.

That said, if the incoming data are quite noisy and you don't have sufficient replication, any results you get may be suspect. There is no substitute for good data.