Multi-mapped reads in RNAseq experiment

[nmerienn]

Dear all,

I am currently working with RNAseq data obtained from different strains of Plasmodium parasite. We are mainly interested in differential expression between the strains. After quality trimming of the reads, I planned to align reads on Plasmodium genome with TopHat2 or HISAT2 (still not yet decided, I am comparing both), and to use HTseq-count for counting.

Alignment works quite correctly, but I have a lot of reads with multiple alignments (> 40% with MAPQ < 10) and I am not sure to really understand how to handle them for the downstream analysis. Based on what I read, I first understood that for differential expression analysis, it would be better to keep multi-mapped reads for counting, but I found that multireads are not counted with HTseq-count. Do I have to force HTseq-count to count them (remove NH optional flag in .bam file)? Is there any way to reduce the levels of multi-mapped reads?

Thank you very much for any help.

[GenoMax]

Were these libraries ribo-depleted? If not the multi-mapping reads could be from rRNA.

[nmerienn]

Thank you for your answer.

Libraries were subjected to polyA selection, so there should not be so much remaining rRNA inside (normally).

[GenoMax]

Thank you for your answer.

Libraries were subjected to polyA selection, so there should not be so much remaining rRNA inside (normally).

One would think so but that may always not be the case. You should check what the multi-mapping reads are? Ideally aligning against rDNA repeat (from your species) would give you an exact idea of how much rRNA had remained.

[nmerienn]

Thank you for your advice. I will align them on rRNA sequences to check if this is remaining rRNA.

But in general, I think there is always some multi-mapped reads in RNAseq data. What is the consensus concerning them for counting, is it better to keep them or remove before counting?

Thanks

[GenoMax]

See this article for some pointers.

You could be strict and drop them altogether, allow the aligner to randomly pick one spot (out of many where read aligns well) or allow the reads to map in every location that they map equally well to. All those options have some consequences which you would need to weigh. There are newer methods like Salmon which consider read distribution to assign the reads, which may be important if you are looking for transcripts.