Statistical methods suitable for DE analysis of non coding RNAs

Question

I am currently working on DE analysis of coding as well as non-coding features of A. thaliana using the edgeR package.

Is the negative binomial method that is normally used for the DE analysis of genes suitable for non-coding features such as miRNA, lncRNA, snRNA, snoRNA, pseudogenes, transposons, uORFs and 5' and 3' UTRs? Or are there special statistical methods that are more suitable for modelling each of these, like I have read that the zero-inflated model used by lncDIFF can better model lncRNA counts? I see that my miRNA counts matrix is abundant in zero counts and very low counts of genes. Should I use lncDIFF for that?

Also, among the various DE test methods in edgeR such as exactTest(), glmFit() and glmLRT(), glmQLFit() and glmQLFTest(), which is/are the one(s) most suitable for running DE tests on such non-coding features?

Answer 1

There is no question that lncDIFF is the approach to use, especially if the alternative is performing edgeR directly, and thats demonstrable (below). I don't think anyone would question the legitimacy of your approach. If you look at Fig 1 for the paper describing the test, here there is no question about it edgeR used directly does not perform for this task at all. The paper describing lncDIFF is here,

https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-019-5926-4

You present lncDIFF as an option within edgeR. This I don't know, but lncDIFF can be directly downloaded from GitHub, https://github.com/qianli10000/lncDIFF.

---

The one thing you have to be careful about is you are using plants, whilst these tests are designed around humans, see below for details.

---

How it's tested: GLM methods versus an Exact test, e.g. Fisher's. Both are all valid tests. However, GLM methods are preferred over an exact test, in my statistical experience.

First, you need to obtain a fit before you can perform a test, which ever GLM you choose.

x <- model.matrix(...)
y <- estimateDisp(...)
fit <- glmFit(y, x, robust=TRUE)
lrt <- glmLRT(fit, contrast=c(-1,1,0,0,0)) 

The basic issue is GLM likelihood ratio test versus quasi likelihood test, which the right approach? The clear answer is QLT because the authors of lncDIFF developed the test around QLT (see the paper). That shouldn't be general advice outside lncDIFF

The correct approach is to try both, if there major differences thats an issue: I would say in that scenario is consider QLT - but only for lncDIFF, as described. If there are no major differences you are good to go.