RNA-seq normalization

[Golsheed]

Hello,

I have a raw count data matrix (with genes in rows and samples in columns) that needs to be normalized. I can't use DESeq2 or Limma directly for this, since the design matrix is not fixed for my data. Specifically, one of the columns of the design matrix changes from gene to gene. So I need to do the normalization myself. Since I'm new to this, I'm not sure if I'm on the right track, so any help, comment, or reference would be much appreciated. Here's what I have in mind to do:

Suppose that the raw counts matrix is called "reads". Then using edgeR, I can do the following:

dge <- DGEList(counts=reads)
dge <- calcNormFactors(dge)

I have a couple of questions on this:
(1) is there a function in edgeR that directly outputs the matrix of normalized counts?
(2) If not, can I calculate the normalized count for each gene-sample combination with
dge$counts/(dge$samples$lib.size*dge$samples$norm.factors)?

I'd appreciate your help,
Golsheed

[illinu]

How did you get your count data matrix? I use the trinity pipeline for this analysis because it does all, align the reads to the transcriptome, calculate abundance, normalize and output the matrix with FPKM values, it also does DGE

[Gordon Smyth]

You can't output normalized counts from edgeR, because the normalization factors are applied to the fitted model, not to the counts themselves. Of course you can output FPKM or counts-per-million, but this isn't the same thing.

The edgeR authors have long argued that "normalized count" is not a meaningful concept, because any standardization of the counts would cause them to be no longer counts.

Anway, you don't need normalized counts to fit models with gene-specific design matrices. This is done for example in the spliceVariants() function in edgeR and also in the DEXSeq package. In both cases, they work with the raw counts.

[bastianwur]

oeh...but there's the dge$pseudo.counts matrix, which I thought were the normalized counts.

[Gordon Smyth]

oeh...but there's the dge$pseudo.counts matrix, which I thought were the normalized counts.

Read the section on normalization in the edgeR User's Guide. Amongst other things, it explains why the pseudo counts should not be used as normalized counts.

[bastianwur]

Guess I should've taken a look at your name ^^ -> sorry *ffttt*.