Hey conor,

Alignment of reads to the genome and then the determination of raw counts via HTseq, followed by differential expression analysis using DESeq2, is common / typical. Filtering raw counts <30 seems high but this may not affect your results that much. I usually go for <5 or <10.

Another way to do this is to pseudo-align to a reference transcriptome FASTA using Kallisto or Salmon, and then feed the count estimates from these into DESeq2.

I don't see the point in the StringTie / FPKM step, given what you have already performed. Nobody should even be using FPKM anymore. FPKM is derived from a normalisation method that renders your samples incomparable via differential expression. Indeed, no cross-sample normalisation is performed. Feed this back to your colleague who recommended FPKM, too.

Would it be better to use the full annotation rather than lncRNA only annotation for HTSeq to identify differentially expressed lncRNA?

You can use either. General results should be the same where it really matters. The differences may only lie in genes (e.g. lncRNAs) that lie on the fringes of statistical significance, which may otherwise have not even been used in whichever downstream analyses you aim to perform. By only focusing on lncRNAs, you save yourself a harsh FDR threshold because you have less tests.

In any case, your reads have already been aligned to the genome, which helps to mitigate some bias in how you then derive raw counts via HTseq.

Are there more appropriate filters that I should use to identify differentially upregulated lncRNA genes?

From DESeq2, typically start at FDR Q < 0.05 and absolute log (base 2) fold change > 2. In DESeq2, note that you should perform lfc shrinkage, like this:

res <- results(dds, contrast=c("Condition", "Treatment", "Control"),
  independentFiltering = TRUE, alpha = 0.05, pAdjustMethod = 'BH', parallel = TRUE)
res <- lfcShrink(dds, contrast=c("Condition", "Treatment", "Control"), res = res)

At what stage in the RNA-Seq pipeline is filtering supposed to be done (before or after DESeq2)?

Usually on the raw counts, but 30 seems high to me. One can also remove when the counts have been normalised. Ask 2 people and you'll get 2 different answers.

Kevin

If you use different annotation sources, separate quantification can avoid issues with ambiguous reads in overlapping annotations from the different sources (where a similar or related overlapping annotation has a different name in the different annotation files). Ideally, having a merged list prior to quantification is theoretically best, but you'd have to have some system for naming the overlapping annotations.

I agree with Kevin that is is probably most helpful to filter prior to testing.

While I might have a preference on characterizing genes with higher FPKM values, I think 50% (or 20%, etc) greater than 1 is often a bit strict. However, if no expression is above 0.1, I would be a little concerned. Also, I might be concerned about having high expression that isn't consistent among your replicates (which I think can be more of an issue for lncRNAs), but I think DESeq2 replicate concordance is typically pretty good (although I'm sure you'll see exceptions to any general rule, if you make enough comparisons).