4 weeks ago by
There are at least 2 things that I would try:
1) Another method (DESeq2, limma-voom), etc. There are also multiple ways to calculate a p-value within edgeR (such as with dispersions estimated for "edgeR-robust")
You may not be able to lock-down the exact right one to use for a project ahead of time. Sometimes the results are similar, sometimes they are very different.
For example, it is a work in progress, but (when I have some free time) I have been trying to show this with public data, and I have accordingly added a "Update" to my GitHub acknowledgement. For example, for 2 out of the 3 E-MTAB-2682 comparisons, I could identify the gene being altered with DESeq2 or limma-voom but not edgeR (but the point is that you will find a project where the method doesn't work if you test enough projects, not that edgeR is worse than DESeq2 or limma-voom). You can see this in the Target_Recovery_Status.xlsx file on the SourceForge page (which should continually change over time, but probably slowly).
2) You can try increasing the FDR to 0.25 (or I have even seen the FDR increased to 0.50 to try and decrease false negatives). However, if you use a method for RNA-Seq analysis, I think it is rare to find no differences with FDR < 0.50 with any method. So, it is probably good to think of those results like a hypothesis.