Hi Mark and all,

I use calcNormFactors with default (TMM) method usually without problems for RNA-Seq before giving the counts to limma-voom.

In one experiment, though, I have noticed that the MA plot for my AvsB comparison has given me a cloud of points not centered around 0.

I have checked and used DESeq2 standard pipeline, which we have used often, too,  and its MAplot looks nicely centered. Then I have used calcNormFactors with 'RLE' method, the one suggested by Huber/Anders and limma/voom again, and it worked fine with the MA plot nicely centered. 

Why is this? How can I know apriori when TMM or RLE is going to work better? is this due to replicates being very disperssed? I can say that TMM worked nicely for a AAA,BBB,CCC comparison in AvsB and AvsC with a very good MDS plot, and in this case where TMM did not work and RLE did, the design was AAA vs BBB and the MDS was not that nice with clear separation more in PC2 than in PC1. Might be that?

Thanks a lot

Jose

My group normalizes a lot of RNA-seq data, and I have not yet seen a real dataset for which it can be definitely said that RLE "works" and TMM doesn't. You would need to provide more evidence to convince me that your data is such an example.

It has nothing to do with whether groups are well separated or replicates are consistent. Normalization does not use the design matrix.