Hi all, 

I'm using voom and limma to analyse RNA-seq data. In general, I have followed the basic instructions regarding RNA-seq analysis presented in the limma manual. 

The matrix of read counts is filtered so that I keep genes where cpm>=1 in at least 3 samples. This is done using the following code:

keep=rowSums(cpm(matrix)>=1)>=3
matrix=rnaseqMatrix_parents[keep,]

Next, I apply the TMM normalization and use the results as input for voom

DGE=DGEList(matrix)
DGE=calcNormFactors(DGE,method =c("TMM"))
v=voom(DGE,design,plot=T)

However, with one of my datasets, the plot produced by voom looks generally ok, but with a gap midway

Looking back at the manual for limma, I noticed the following note: 

If the data are very noisy, one can apply the same between-array normalization methods as would be used for microarrays, for example: v <- voom(counts,design,plot=TRUE,normalize="quantile")

Although I am not certain that my dataset qualifies as "very noisy" (happy to read what people have to say about it or point me to examples) I decided to give it a whirl. 

v2=voom(matrix,design,plot=T, normalize="quantile")

The plot looks better than before:

I should finish by saying that I went on and perform the differential expression analysis using the two outputs from voom. Things change marginally (1570 vs 1758 DEG out of 15729 genes at 5% FDR when using the "TMM" vs "quantile" strategy). Not surprisingly, the genes differentially expressed in one configuration only are associated with Q-values close to the 0.05 cut-off.

I am a bit uncertain about using the quantile normalization route (did not find a reference where people had used it), so I am eager to have inputs/guidelines about when to decide using TMM vs quantile norm.

Input on this most welcome!

Thanks!

Jean-Marc

You ask a good question. Personally I reserve quantile normalization for really extreme situations. On the other hand, one of my close colleagues, Wei Shi, likes to use quantile normalization almost routinely, see for example:

http://www.ncbi.nlm.nih.gov/pubmed/25894659

As you have seen, it doesn't make a spectacular difference in most cases; scale normalization is simpler but quantile normalization tends to give a few more DE genes.

To me, your voom plot with TMM looks fine. The plots shows that you have some clusters of transcripts with special characteristics, evidenced by the gap and by the blob of points in the middle left of the plot. These transcripts are probably outliers in a subset of samples and this is something to explore as part of the downstream analysis, but to me is not something for the normalization to solve. It doesn't cause any serious problems to voom.

I believe the original voom paper (Law et al.) actually uses quantile normalization. although I've never used it, it would be interesting to hear if the gurus have any pointers here.

Also, TMM and quantile normalization are separate and you should always ensure that you are using TMM instead of "simple" normalization by read depth. An easy way to ensure this works as expect is to create an edgeR::DGEList with your data and call calcNormFactors on it. Use that data object to pass through voom, and not just a count matrix 

Update

In the event that someone skimming this thread doesn't read through the comments, I wanted to provide an update to this answer in light of Gordon's replies to this answer:

Take note that (contrary to what I said in my answer) Gordon suggests that you actually wouldn't want to do both TMM followed by quantile normalization, but rather choose one or the other.