Importance of Proper Pairs vs Aligned Reads for RNASeq data

Question

I have stranded, paired end RNASeq reads that I have aligned using STAR. I plan do conduct a differential expression analysis with DESeq2. After running quality control checks, a good portion of my reads are aligned but I'm a little concerned about the proportion of properly paired reads

Here is the inner distance plot for the samples

For human RNA sequencing, does anyone have a rule of thumb that they'd consider is an appropriate proportion of properly paired reads? Are there any good papers that talk about this issue? As MultiQC marks the column as yellow, I'm concerned the proper pair percentage may not be satisfactory.

How important are properly paired reads for differential expression analyses?

Answer 1

Differential gene analysis sees only the read counts. So the proportion of not properly paired reads will have an effect on the counts, if the programme / software used, counts fragments / read pairs instead of single reads. For example HTSeq will count

From https://htseq.readthedocs.io/en/release_0.11.1/count.html

For paired-end data, does htseq-count count reads or read pairs? Read pairs. The script is designed to count “units of evidence” for gene expression. If both mates map to the same gene, this still only shows that one cDNA fragment originated from that gene. Hence, it should be counted only once.

In Rsubread, featureCounts you set requireBothEndsMapped=TRUE to count fragments.

Short anaswer is if the sequencing is deep enough, you should still have enough reads. In your situation, I estimated that you lose about 40% of your reads if you use only properly paired reads, which means you lose quite a bit of counts.

The larger question is what reads are you missing?

I see that you have about 20-30% of reads that are non-uniquely mapped, I suppose part of this will contribute towards the not properly paired reads. I would check first whether most of the not properly paired comes from multiple aligned.

Then if not, you can check where these not properly paired reads are falling, if they are some repetitive regions or say rRNA etc, then most likely thats a problem with the library prep..

Answer 2

If you check the RSeQC docs you can get an explanation of how the inner distance stats work. By the look of your inner-distance MultiQC plot I guess that your sequencing is 2x100bp reads. This plot shows that your insert sizes on the pairs is pretty short, meaning that a lot of the paired reads are overlapping significantly. This will be affecting the alignments a little, resulting in the low proportion of properly paired reads.

The good news is that for differential expression, this probably doesn't really matter too much. What is important is getting fragment counts, so as long as at least one of the pair mates is mapping properly, then you will will be getting sensible counts - even if read 2 is not mapping properly. You should be able to have confidence in your results from multiple replicates if in doubt - you have more than just -rep1 right? ;)

Note that the yellow background in the MultiQC column is not passing judgement with categories like pass/warn/fail in FastQC - it just happens that the colour scheme ended up there for your samples. It's a continuous gradient from red to yellow to green for those values (0-100%).