So I usually run all my samples with

strandedness

set to

auto

and usually have no issues until I tried this on total RNA. The confusing part of this error message is that I didn't set the strandedness to

reverse

in the samplesheet, so shouldn't the inferred strandedness be the same (even though

unstranded

should be wrong in this case). The infer_experiment output below:

This is PairEnd Data
Fraction of reads failed to determine: 0.0719
Fraction of reads explained by "1++,1--,2+-,2-+": 0.2513
Fraction of reads explained by "1+-,1-+,2++,2--": 0.6768

Does this warning merely mean that while

reverse

is the best guess, the fraction of reads that failed determination is too high? or that the dominant fraction not significantly different from the other fractions?

The issue here has been that the method we used to predict strandedness (Salmon) made its assessments by a different formula to the tool used in the QC step at the end of the workflow. Your message is timely, because we've tried to address this in the latest release- try it out and see if it helps.

Yea there are a few other samples that are suspicious, will look into it. Also odd looking GC content plots. I had my scripts use the auto inference in the past as I was processing some older batches of rnaseq where some were unstranded (but not clearly annotated as such...), so it was convenient to let it infer. Though I'd imagine for future batches if I do know the expected strandedness I shouldn't let it infer (to prevent it from erroneous guesses?). Or is the inference generally quite reliable?

If there's nothing wrong it should be quite straightforward and the auto inference should mostly get it right. The new version allows you to control how stringent it is as wll.

Interesting, will give it a go.

It might be sensible to set the strandedness once you're sure of it on subsequent runs, but just to save yourself the compute resource used in the inference.