Removing low complexity reads from RNA-seq

Question

[Asaf]

I'm mapping bacterial RNA-seq to the genome and found something very weird. A very abundant RNA has a AATAATAAT repeat somewhere in the middle and I have a lot of reads which map to the gene (it's paired end sequencing and the second mate maps) but the number of repeats is larger (up to 8 AAT repeats). Since I have several millions reads of this gene and it's a small-RNA I get thousands such reads. I'm trying to figure out their source, whether it's biological or just artifact of the RT/PCR/sequencing etc.

To overcome this issue I started screening the reads for low-complexity reads and removing them (using dust filter) which seems to work. My questions are:

1. Is it common to remove low-complexity reads from the data?
2. Why should they be removed? Is it because the mapping will be difficult or wrong or the reads are probably a result of an error?

Thanks

[mark.ziemann]

Are you using a mapping quality filter before generating counts? using a mapQ threshold >10 will remove most low-complexity reads because they map to many places in the genome. HTSeq, Bedtools and FeatureCounts all have facility to do this.

[Asaf]

My analysis is a bit different, I don't remove multiple mapped reads. In addition I work on bacteria so the genome is much smaller.

[mark.ziemann]

In that case, "Dusting" or "RepeatMasking" reads would seem appropriate.