Which HTSeq mode is suitable for RNAseq data to be used for differential gene expression?

Question

[Arindam Ghosh]

Which of the three modes -- union, intersection-strict, intersection-nonempty-- is suitable for RNAseq data to be used for differential gene expression?

[Matteo Schiavinato]

There is no "more suitable" one, all of them will do depending on what you want to achieve! I usually use union but you should check, biologically speaking, which one fits more your system.

[GouthamAtla]

Go with default, but I would suggest to use FeatureCounts instead of HTSeq. or Salmon.

[Arindam Ghosh]

Why? Any explanation for not prefering HTSeq.

[GenoMax]

featureCounts is:

1. Much faster (multi-threaded).
2. No need to sort BAM files but can take sorted ones saving time.
3. Will create a matrix of reads counts if you feed it multiple BAM files that can be directly used for DE.

[GouthamAtla]

Also better at assigning PE reads compared to HTSeq.

[Kevin Blighe]

Any room for the pseudo-aligners here, i.e., to avoid the necessity to produce a BAM in the first place? - Kallisto, Salmon, et al.

[grant.hovhannisyan]

If you use STAR as mapper, you can specify the option --quantMode GeneCounts and STAR will output read counts equivalent to htseq --union option. Thus you can skip counting step in your pipeline and save time.

Answer 1

[h.mon]

As stated above, featureCounts is faster and probably equivalent or better than HTSeq. But to answer directly your question, I will quote the author of HTSeq:

I haven't used anything else than "union" since long, and, admittedly, I would find it hard to now come up with good examples where the intersection modes would be preferable in practice. (When I wrote htseq-count two years ago, that was not so clear yet, of course.)