Variant calling from paired/unpaired reads

[dcard]

I just happened upon a couple question threads in various places that warn against using both paired and non-paired reads for SNP calling.

https://www.biostars.org/p/140318/
http://gatkforums.broadinstitute.org...ngle-end-reads

It appears that paired and unpaired reads are on two different playing fields as far as their mapping quality, which makes sense, and this makes a difference in inferring SNPs downstream. I've tried looking for more information about this and potential ways to overcome it, but I'm not really seeing anything.

I thought I would pose the question to this forum and see what more knowledgeable people can explain. Namely, a more detailed explanation of this problem would be nice for those of us who don't know a ton about what is going on under the hood of various variant calling pipelines (e.g., samtools and GATK). Is it possible to overcome this issue and take advantage of both paired and unpaired (i.e., single-end sequencing or pairs broken upstream)? Are there tools that can account for this?

Any thoughts are appreciated. I have some datasets where using broken pairs would be great, rather than just ignoring them.

Cheers,
Daren

[Brian Bushnell]

Hi Daren,

When I wrote a variant caller, I used both paired and unpaired reads; properly paired reads were simply given greater weight. The result was better than when ignoring unpaired reads. I suspect that it made little difference except in places with low coverage or near structural variations, but I never analyzed exactly where it made a difference.

I don't know how other programs handle variants in paired vs unpaired reads, but I believe the best method is to include all available information, and just weight it appropriately if some pieces of information have lower confidence than others. The difference in value between paired and unpaired reads varies greatly depending on insert size, genome complexity, ploidy, and read length; it's possible that these guidelines were developed for 2x36bp human sequencing, which would be utterly irrelevant to 1x250bp bacterial sequencing, for example. Or 2x150bp human resequencing, for that matter. Guidelines that Broad publishes are generally going to be human-centric (and thus deal with large, repetitive, diploid genomes), and guidelines that anyone publishes are going to relate to the technology common at the time.