EdgeR and Deseq Normalization problems

[moriah]

Hi all,

I am trying to run DE analysis on miRNAs samples (from two conditions).

In a recently published article it was mentioned that "TMM" normalization method is the least suitable to miRNAs and they suggested to use quantile normalization.

I tried to change the method parameter to "upperquantile"
y <- calcNormFactors(y,method="upperquartile")

What I was getting when running the next step is:
y <- estimateCommonDisp(y) # Estimates common dispersion
Error in if (mx < tol) { : missing value where TRUE/FALSE needed

When looking at the y variable I have noticed that all of the norm.factors are NaN.
Worth mentioning that I didn't have this problem when running the "TMM" and "RLE" normalization methods.

My questions:
1. What might cause the problem when running "upperquantile" normalization in EdgeR?
2. How can I use the "quantile" normalization and not the "upperquantile"? Change the Percentile some how? (p parameter)
3.From the manual I noticed that the "RLE" normalization is similar to the implemented normalization in DEseq , but when running EdgeR with "RLE" vs DEseq I have been getting very different results ? Why is that?

I will appreciate your answers,

Thanks,
Moriah

[markrobinsonca]

Hi Moriah,

Some quick thoughts:

1. I'm afraid that I can't tell from the information you've provided why method="upperquantile" gives NaNs. If you send me a count table, I can have a look, there should be a simple explanation.

2. No, changing the percentile will not allow quantile normalization. It is possible to do quantile normalization in edgeR through offsets (keeping the count data as is), but this is not what the authors of the miRNA paper are suggesting (I've only skimmed that paper and it wasn't clear how they used the normalized data in the differential testing and it appears that their testing strategy does not account for biological variation - binomial or Fisher's test).

3. Yes, basically with method="RLE", you will get the same normalization factors in edgeR (lib.size * norm.factor) as sizeFactors() in DESeq. But, edgeR and DESeq differ slightly in how the dispersions are estimated and how the exact test is performed. You mentioned "very different results", what do you mean by this?

If this doesn't answer your question, can I suggest you post to the Bioconductor mailing list? See the section "how to get help" in the edgeR User's Guide.

Best,
Mark

[moriah]

Hi Mark,

Thank you for your answer!

1. How can I send you my table? Thanks for the suggestion!
2. What method would you suggest when dealing with DE of micorRNA?
3. In DESeq I got one differentially expressed microRNA, and in EDgeR I got around 100.

Thanks,
Moriah

[gpcr]

@moriah : Any luck with your new analysis with upper-quartile... as you were following methods in this article :
"Evaluation of normalization methods in mammalian microRNA-Seq data."

[am@i]

Hi
I am getting all the p- values and FDR values as non-significant while performing DE analysis of miRNA's in DEseq. Whereas almost all are significant with chi-square test.What is the reason for such discrepancy? How should I use DEseq to get meaningful result of DE analysis for miRNA?