Hi I have downloaded a mass spec dataset from maxqb (http://maxqb.biochem.mpg.de/mxdb/).

I have then imported the data Into R specifying 0 values as NA (as I think this is correct for analysis using the Limma package):

 exprs <- read_excel("exprs.xlsx", na="0")

I then normalise using normalizeBetweenArrays and remove any columns with all NA values:

 y <- normalizeBetweenArrays(log2(exprs), method="cyclicloess", iterations=10)
 y <- exprs[rowSumsis.na(exprs)) != ncol(exprs), ]

Then I carry out DE analysis:

group <- factor(pData$MV_only) 
design <- model.matrix(~0+group)
contrast.matrix <- makeContrasts(
  mv_vs_other = MV-other,levels=design)
fit <- lmFit(y, design)
fit2 <- contrasts.fit(fit, contrast.matrix)
fit2 <- eBayes(fit2, trend=T)

and I get the following warning: Warning message:

Partial NA coefficients for 2885 probe(s)

But the analysis seems to work fine.

I have a couple of questions here.

1) Is this the correct way to deal with 0 values? It should be dealt with in normalizeBetweenArrays?

2) To then visualise any of the data in PCA or coolmap (heat map) I convert the NA's back to 0 in the normalised data:

y[is.na(y)] <- 0

Is this also correct?

Thank you

Reuben

I don't yet have a recommended limma pipeline for MaxQuant mass spec data, although this is something we are working on.

The method you've used seems to me to be as good as anything else at the moment. limma is treating the non-detection NAs as "missing at random". While this assumption is not completely correct, and the NAs do contain some weak information about expression levels, I don't yet have a good general way to recover this information.

Converting NAs back to zeros is however going to play havoc with the heatmap or PCA. I think you would need to use an imputation method for this purpose, for example MaxQuant LQF values.

On a more minor note, setting the cyclic loess iterations as high as 10 seems a bit unnecessary.