In my RNAseq dataset of differentiated stem-cell lines, some samples have far fewer significantly differentially expressed genes than others. QC shows that this is because there are way fewer reads for one sample than the others.
Can gene co-expression networks be used to infer differential expression? For example, if certain genes just miss the cutoff q-value cutoff If so, what papers have done so?
While you can use networks to find differentially expressed genes (see the WGCNA package, which does this) in my experience this ends up largely matching what you'd get using a traditional package with a looser threshold for significance. Given the time savings of traditional packages, there's rarely any gain to using networks (it won't hurt to try, just note that it'll take some time). If some very interesting genes are just above your significance threshold then change your threshold. You need to validate your findings in some way anyway, so your p-value threshold is partly just a way to protect you from wasting your time (but if background knowledge suggests that your time wouldn't be wasted...).
Co-expression network will give you an idea about the genes having similar expression patterns and the nodes will be decided on the basis of correlation scores and nothing much you will get from differential gene expression perspective.
I would suggest you to try some other models for differential expression analysis like baySeq , DESeq, edgeR, NOIseq. Depending upon the model you use for analysis you might get some extra significant DEG. Also, try different significance cutoff and see how it's changing the analysis result to get the threshold that suits the data best.
