It looks like DESeq2 has been used efficaciously for MS data in published papers (https://www.ncbi.nlm.nih.gov/pubmed/?term=26193490). However, that is for unlabeled MS data, in which the only information is the spectral count. For labeled MS data, it does not appear DESeq2 is being used.

Thus we want to double check with you that Deseq2 is appropriate for labeled mass spectrometry counts. We will also greatly appreciated it if you could send us some publications in this area.

The only type of MS-based quantitation that is amenable to DESeq2 (or similar approaches) is spectral counting. Other approaches, whether labelled of unlabelled, are continuous data; I would also recommend limma for their analysis.

What's your take on normalization/computing offsets? I usually take the super naive approach of assuming that the assumptions underlying the methods used in DESeq2 or edgeR are +/- applicable to spectral count data, but given the unfortunate level of missingness (here I am talking about data that appear to be missing at random), I have never quite convinced myself that this is really a smart thing to do.

I think the key considerations for use of DESeq2 are: (1) are the data roughly count scale, and so the negative binomial variance modeling is useful, (2) can you assign observations to rows with some certainty (or if not extra modeling is needed to deal with this uncertainty but this is not part of DESeq2 right now)

I don't think I know all the different varieties of MS data, but for many quantitative MS data types, it is appropriate to treat them as continuous measurements on which you can use ordinary statistical models and tests, possibly with empirical Bayes moderation (limma) and/or suitable transformation (e.g. https://www.ncbi.nlm.nih.gov/pubmed/20382981)

What's your take on normalization/computing offsets? I usually take the super naive approach of assuming that the assumptions underlying the methods used in DESeq2 or edgeR are +/- applicable to spectral count data, but given the unfortunate level of missingness (here I am talking about data that appear to be missing at random), I have never quite convinced myself that this is really a smart thing to do.