Hello!

I have to analyze data obtained from Affymetrix miRNA 4.0 microarray. It is the first time I have to analyze such kind of data and I wanted to be sure that I was doing it right. My main concerns are about the normalization of the probe's intensities and the filtering in limma.

# Normalization
library(oligo)
library(pd.mirna.4.0)
celFiles <- list.celfiles("~/Desktop/Affymetrix_miRNA/RawData_miRNA", full.names=TRUE)
rawData <- read.celfiles(celFiles, pkgname="pd.mirna.4.0")
eset <- rma(rawData)

# I did some plots and everything look really great after normalization. 

# Limma
# I directly use the eset data to calculte the miRNAs differentially expressed. 

Question:

1. Normalization: Should I do something else for the normalization or apply rma is usually enough? Should I do something with the information of the spike-in probes? 

2. Filtering: I have seen that among the probes on this micro-array some are unrelevant for my analysis:

snoRNA, spike-in,...

should I remove such probes before computing all the statistics from limma?

3. More globally, I found many tutorials talking about normalization of microarrays for genes but not for miRNAs. Are they differences in the processing of those two types of microarrays that I should know?

4. From what I have seen, some miRNAs are represented by several probes on the chip. They seem to be clustered during the normalization step and the creation of the expression dataset. Can someone explain me how it is done or point me to article/post that explain such thing?

Thanks in advance!

1. No, that's about it. You could hypothetically use the spike-ins for something, but I have yet to find a reasonable use for them.
2. Probably. I generally remove all the extra cruft Affy puts on the arrays before fitting the model.
3. Well, the miRNA arrays are sort of different from most other Affy arrays. First, most miRNA transcripts are shorter than the Affy probes (21-23 nt vs 25 nt), so Affy tiles down some number (four, I believe) of the same exact probe on different regions of the array. You then quantile normalize, which is reasonable, but then you fit a medianpolish model, which is probably not necessary - you could probably just take the average of the probes and go from there. But I don't think the medianpolish is going to do anything wrong per se, and it's easier to do than trying to figure out how to compute the averages.
4. Yeah, especially with the miRNA 4.0 array. If a given miRNA is conserved across various species, instead of putting down individual probes for each species, Affy just puts down one probeset and re-labels it a bunch of times. It's like more efficient and stuff. Or are you talking about the fact that multiple probes go into a single probeset? In that case, then yeah, multiple probes are summarized to make a single probeset (and hence a single measure of expression for the thing being measured). If you google 'Irizarry affymetrix medianpolish' you will get a bunch of links that will explain things further.