Periodical illumina read length distribution after trimming of low-quality bases

[luxmare]

In my NGS data analysis, before mapping, I trimmed low-quality bases (<Q20) from 3' ends until a high quality (≥Q20) base appears. After that, I plotted the distribution of read length and obtained the weird periodical read length distribution. Please see attached.

In the graph, length distributions from different lanes or tiles were drawn in different colors. Frequencies of reads were oscillated with 5bp intervals.

I also saw this kind of weird length distribution for other our RNA-seq and genome sequence dataset, and RNA-seq data from SRA as well.

Does anyone know the reason why such periodical length distribution was appeared after trimming?

Thanks in advance.

[kmcarr]

I can't remember any details but I do recall hearing once that there is something about the Illumina quality scoring algorithms which creates these 5bp cycles.

[luxmare]

Thank you, kmcarr.
Do you mean that such weird distribution is caused by the base calling algorithm in the illumina pipeline?

Can we just ignore the length distribution after trimming of low-quality bases? We would not worry about it?

[fkrueger]

Funny that you mention this, I have done something quite similar recently.

I wanted to find out whether the increase in sequencing errors towards later sequencing cycles (which is equivalent to a drop in Phred quality) can be described by some kind of mathematical formula. I used a couple of sequence files to determine the starting position of poor qualities. Poor qualities were defined as reads which exceeded a certain number of low quality basecalls in total (in the attached figure there had to be at least 8 quality values below 30). I tried various different thresholds (qualities 10, 15, 20, 30) but the graph does not change much.

Interestingly the pattern I got did not increase steadily towards later cycles (as I expected), and I also saw a periodicity of - you might have guessed - 5 bp for poor quality starting positions. This seems to be indeed a feature of the Illumina pipeline algorithms used. Even though it looks artefactual and I found this slightly worrying I don't think one can do much about it, as it is present in all samples irrespective of their origin.

This led me to the conclusion that the increased error rate one sees towards the end of longer reads is not chemistry or run-time related but seems to be largely the cumulative effect of these spikes of low quality basecalls which are introduced into the reads with a periodicity of 5 bp. Quite odd, isn't it?

[luxmare]

Thank you fkrueger.
I really think so, it's weird. I hope this hidden bias will be improved in the near future.