I came across an article benchmarking different tissue-specific gene expression identifying methods.
I am wondering why is it not a common practice to use differential expression analysis methods, like DESeq2, to identify tissue specific genes? Practically, it can be really a time-consuming task, if you have multiple tissues - identifying DE genes using one-vs-one for each tissue and then find the common DEs for the tissue of interest. Apart from practicality, is there any reason not to do it?
Thank you very much for your question and the linked paper. It helps me immensely. I did what you mention in DESeq2 - the problem is that your results are horrible to interpret (using pairwise tests, expression clustering and GO-Term analysis).
Philosophically, tissue specific suggests a gene is _only_ expressed in one tissue. Just because a gene is DE between tissue type A and B, with positive logFoldChange (i.e. higher in tissue A), that doesn't mean it is not expressed in B, it can easily be expressed in both, while being higher in tissue A.
Practically, most DGE tools assume that:
That changes between the two conditions are limited, or at least the mean change is 0.
That the compositional changes more or less reflect the expression level changes. The normalisation methods in DESeq2 and edgeR are designed to cope with a certain amount of compositional change between samples (i.e. changes in the total amount of RNA in a cell between samples/changes in expression levels of highly expressed genes changing the reads available for other genes), but there is probably a limit to how big the changes can be.
That the "gene" is the same sequence in both samples - i.e. there haven't really been any big splicing changes (although using tximport and weighting data be effect gene length can offset this somewhat).
All three of these assumptions are less likely to be true when comparing tissues than when comparing two samples from the same tissue.
Thank you very much for your question and the linked paper. It helps me immensely. I did what you mention in DESeq2 - the problem is that your results are horrible to interpret (using pairwise tests, expression clustering and GO-Term analysis).