Snords and MiRs in RNA-Seq data

Question

[lhawthorn]

I have RNA-Seq data from a large number of samples analyzed using CuffDiff. They are 50bp pe reads and libraries prepared using Truseq (so size selected). The most highly up and down regulated transcripts are snords and miRNAs. They have unusually high fold changes but the q-values are high (~1). But in many cases the q values are high because the mir/snord is absent in one of the samples. Are these artifacts of alignment and would you suggest ignoring them? Or should they be included in my analyses? Thanks

[Devon Ryan]

I'd recommend including them. You might be able to use the GTF file from cufflinks (or stringTie or whatever), use that with featureCounts to get counts and then put them in edgeR. The results from that tend to be a little more reliable (you could also try DESeq2, but it seems edgeR performs a bit better when you randomly have a sample with a dropout). Note that it's good to confirm results with an orthogonal technology regardless of what's used.