very low percent uniquely mapped for only one read out of the pair

[hello_world]

Hello!
I am baffled by this issue. I am having trouble getting any significant percentage of reads to map. I don't know if it is a library prep issue or a mapping parameters issue. I used the Illumina kit to make a paired end library and size-selected for an insert of 200bp. Reads are 101bp. The core facility which did the sequencing also mapped the reads using BWA.
Here is their report:

Total Reads 163,135,475
Post-Filter Reads 81,274,386 - 49.8% of Total Reads
Failed Reads 81,861,089 - 50.2% of Total Reads
Mapped Reads (Read 1) 36,101,462 - 44.4% of Post-Filter Reads
Mapped Reads (Read 2) 31,427,539 - 38.7% of Post-Filter Reads
Uniquely-Mapped Reads (Read 1) 32,297,769 - 39.7% of Post-Filter Reads
Uniquely-Mapped Reads (Read 2) 6,464,605 - 8.0% of Post-Filter Reads
Insert Size 189

So all my libraries have this issue. The number of uniquely mapped reads for the second real is very, very low. What could be causing this to happen? Is it because the second read is overlapping into the first? As insert size is 189 and each read is 101bp, I expect some overlap. How will this be treated by the aligner?

I also tried doing my own mapping using Tophat2. I tried to map with a default -r value (distance between mate pairs) and with -r set to 0. I am getting 3% mapped reads using Tophat. Should I try with a negative -r value?

What could be going on? Is this a problem with my library prep? Or can it be solved with the correct alignment settings?

Thanks!

[grakocevic]

Overlap shouldn't be the issue.

I'm not sure what the filter the core is referring to (probably Illumina's chastity), but whatever it is, 50% failure is not normal. Try running FastQC on your data for some clues as to what may have gone wrong.

[Brian Bushnell]

I assume that "uniquely-mapped reads" in this case means "unambiguously mapped" rather than "mapped to a location that had not been previously mapped to". Is that correct?

Anyway, ambiguous alignments can be caused by low quality reads. The lower the quality, the more ambiguous the alignments will be - and read 2 typically has lower quality than read 1. Posting the base quality and base content histograms might be helpful. This sounds like a library failure - possibly, the sequencing machine was highly overloaded (clusters too dense).

If you want to try a more sensitive aligner, try BBMap. It outputs a nice set of statistics concerning the mapping rates, pairing rates, and ambiguous mapping rates for read 1 and read 2, as well as error rates, which might help diagnose the problem.

[hello_world]

Ambiguous alignments can be caused by low quality reads. The lower the quality, the more ambiguous the alignments will be - and read 2 typically has lower quality than read 1. Posting the base quality and base content histograms might be helpful. This sounds like a library failure - possibly, the sequencing machine was highly overloaded (clusters too dense).

Good suggestion. I had no idea there could be such dramatic quality difference between read 1 and read 2.
I am not familiar with FastQC but I'll go figure it out. A colleague also suggested I use Picard.

[Brian Bushnell]

I had no idea there could be such dramatic quality difference between read 1 and read 2.

Normally, there isn't anything so drastic...

[hello_world]

I have generated gone through the fastqc files for my paired read libraries.
Strangely enough, the problem is not low quality reads. The problem is over-representation of 10-15 sequences. For one of my over-represented sequences, there are 30K reads. Also, most of the over-represented sequences blat to the mitochondrial genome.
This library was prepared from a rare stem cell population. I am not shocked that, starting from picograms of RNA, I have some bottlenecks.
My question now is if I can save some of the data or get some information out of this sequencing run despite this issue.
For example, can I treat the library as single end read instead of paired? If I pool my four biological replicates, I might still get enough real reads to do differential expression analysis.
Is there a way to proceed given the read 2 bottleneck?
Thanks!