MiSeq - 16S - what to do with unassembled reads?

[Elsie]

Brian you are a superstar!, thanks so much for that. I've been hunting through references and everything says "merge reads, discard unmerged", fine if everything does merge. Thanks so much, I really appreciate the advice.

[Brian Bushnell]

I recommend against following guidance of the form "X must be done in situation Y". PE data doesn't have to be merged in any situation; you can always use the ends individually if you want. Even after you merge them, you don't get a whole 16s (though with 250bp paired reads, you CAN get a whole V4 region). Whether or not it is wise to merge them depends on the insert size distribution, the accuracy of the merging tool, and the capabilities/limitations of the downstream analysis tools. If you have a low merge rate, it's likely that a lot of your reads don't overlap.

Merging causes various biases, and can turn good data into bad data, as can any preprocessing step. For really good 16s analysis I recommend PacBio reads of insert, which can produce full-length 16s sequences at around 99.5% accuracy by creating a consensus from sequencing the same molecule multiple times.

If you are analyzing merged reads, you may simply wish to discard all unmerged reads. This will cause bias (against low-complexity and repetitive sequences) but when you split reads into merged and unmerged, bias is inevitable no matter how you proceed. It's probably fine as long as you are not doing anything quantitative, but unacceptable if you are. If you want to quantify abundance based on 16s, just use the raw reads. Or switch to a smaller region like V4 in which you can ensure all pairs are overlapping.

Also, I suggest reading this thread, which compares the accuracy of various pair-merging tools.

Edit: With sufficient coverage, you may be able to improve your merging rate using error correction. But again, that risks increasing bias.

[Elsie]

Hi,
I'm working with 16S data, and have been using PEAR (http://bioinformatics.oxfordjournals...cs.btt593.full) for merging my reads (thanks PEAR people, great tool!). This has been going great but my most recent set of data has lots of unassembled reads (> 50%) - can I include these into a 16S analysis pipeline and if so, how? Every reference I've read says PE data must be merged before 16S analysis (QIIME, mothur and UPARSE).

thanks.