I had a question regarding interpreting some of the multiqc reports, I noticed that i was getting roughly ~ 40% Salmon mapping rate for all my samples, and some of my samples had high Picard percent (90%) duplication. I was wondering if the 40% is too low to proceed with downstream analysis, and should I exclude the samples with high percent duplication?

Mapping rates at that level are not all that unusual with e.g. ribo-depleted libraries where there's a high level of non-exonic content from immature transcripts. I'd maybe question a bit if it was a poly-A library. Duplication is tricky with RNA-seq because it can be highly duplicated by its nature. If you have a handful of extremely highly expressed transcripts they can dominate the sequencing and you end up with a lot of duplication. Looking at your RSeQC duplication plot I'd say you have one oddball sample (day0_cc030) you might want to take a look at. It has more reads with more duplication. Unless there's anything significantly different about its biology you might want to go back and check sample QC (RIN) to see if you might have had unusually low levels of starting material.

The dupRadar results will give you a better idea about the nature of the duplicates, but 90% is definitely concerning (even if you have one or a few very highly expressed loci, it doesn't leave many reads for the rest). If you have a high rate of unmapped reads (>5%), definitely find out what they are. Sometimes it's rRNA (which get posttranscriptionally modified and often does not sequence well), which is annoying but not concerning, but I'd be worried about contamination. A simple, but comprehensive way is to take a few thousand unmapped reads and just BLAST them against the nt database.

Regarding the day0_cc030, all the day 0s were collected from the same species at the same time, so theoretically they should be the same, there should be nothing different about its biology. I believe the RIN was fine, so should I just scrap this sample for downstream analysis? dupRadar keeps failing for this particular group of samples, i don't skip the step, but i tried to at least skip problematic samples with:

process {
    withName:'DUPRADAR' {
        errorStrategy = 'ignore'
    }
}

but it ends up failing for all the samples. I was thinking of just running dupradar outside the pipeline to double check? I'm just not sure if I should include the high duplicate samples into my diffferential gene analysis.

The problem with super-high duplicates (assuming they are not all "real") is that your actual information content sinks dramatically -- with >95% dups you only have <5% useful information (that is only less than one fifth of what you get with 75% dups, which is already pretty bad). If you look at the PCA, on the other hand, the day0_cc030 is somewhere in the middle, close to other day0 sample, so it does not look like an outlier. The conservative thing would probably be to drop that sample, but you can run your stats both ways and see if it makes a difference. What error are you getting from dupRadar? You could run it manually to see if you can work around that.

even with the config file: dupRadar just shows this:

[18/09a1de] process > NFCORE_RNASEQ:RNASEQ:DUPRAD... [100%] 25 of 25, failed: 25

But is there a set duplication threshold where I should generally scrap a sample, like 75%, I would rather take a conservative approach as I plan on collecting more replicates, so data avaliability isn't really an issue.

Any error messages would be in the

.nextflow.log

or in the

.command.*

files in the work directories of the failed processes. Any thresholds for what is acceptable QC would have to be set by you, there are no generally accepted "standards". As I said you can do it both ways. Given the PCA, I would expect that no huge differences, but the extra sample should lower the p-values bit. That's why the "conservative" thing would be to leave it out, whereas in cases of real outliers (with very different expression) it would be "conservative" to retain them (since the noise introduced by such a sample would likely increase p-values).