Hi David,

This is a really interesting workflow. My comments:

(1) I encourage you to start a discussion about this idea on seqanswers.com; you'll reach more people and may have a better discussion there. Ideally, you'll get a Tophat developer to chime in on what I perceive to be the main issue, which is:

This may seem similar to setting tophat to ignore non-unique reads. However, it is not. This approach gives you 10-15% more reads. I think it is because if tophat finds (for example) that the forward read maps to one site but the reverse read maps to two sites it throws away the whole read.

Remember that Tophat uses Bowtie to map reads, so it would make sense to look carefully at the Bowtie documentation to see how it handles paired-end reads. I can't find anything that directly addresses your issue. The other thing to consider is how Tophat maps reads -- it breaks them up in order to find splice junctions -- and so I'm not sure that Tophat/Bowtie is really mapping paired reads; it may be doing some hybrid single/paired-end mapping. Also, at one time, you could specify Bowtie parameters when running Tophat, but I don't see that option anymore.

(2) It would be interesting to know whether you get qualitatively different results via Cufflinks (or another transcriptome analysis software package) using your method vs. just using Tophat w/ and w/o ignoring non-unique reads. A skeptical view of your workflow would note that (a) multi-mapping reads may be legitimate and should not be filtered out and (b) Cufflinks/compare/diff assembly and quantitation may smooth out stray reads enough so that your method isn't necessary.

Thanks for the interesting post,
J.

On Feb 23, 2011, at 9:41 AM, David Matthews wrote:

Hi Jeremy,

I thought I'd write to get a discussion of a workflow for people doing RNA seq that I have found very useful and addresses some issues in mapping mRNA derived RNA-seq paired end data to the genome using tophat. Here is the approach I use (I have a human mRNA sample deep sequenced with a 56bp paired end read on an illumina generating 29 million reads):

1. Align to hg19 (in my case) using tophat and allowing up to 40 hits for each sequence read
2. In samtools filter for "read is unmapped", "mate is mapped" and "mate is mapped in a proper pair"
3. Use "group" to group the filtered sam file on c1 (which is the "bio-sequencer" read number) and set an operation to count on c1 as well. This provides a list of the reads and how many times they map to the human genome, because you have filtered the set for reads that have a mate pair there will be an even number for each read. For most of the reads the number will be 2 (indicating the forward read maps once and the reverse read maps once and in a proper pair) but for reads that map ambiguously the number will be multiples of 2. If you count these up I find that 18 million reads map once, 1.3 million map twice, 400,000 reads map 3 times and so on until you get down to 1 read mapping 30 times, 1 read mapping 31 times and so on...
4. Filter the reads to remove any reads that map more than 2 times.
5. Use "compare two datasets" to compare your new list of reads that map only twice to pull out all the reads in your sam file that only map twice (i.e. the mate pairs).
6. You'll need to sort the sam file before you can use it with other applications like IGV.

What you end up with is a sam file where all the reads map to one site only and all the reads map as a proper pair. This may seem similar to setting tophat to ignore non-unique reads. However, it is not. This approach gives you 10-15% more reads. I think it is because if tophat finds (for example) that the forward read maps to one site but the reverse read maps to two sites it throws away the whole read. By filtering the sam file to restrict it to only those mappings that make sense you increase the number of unique reads by getting rid of irrational mappings.

Has anyone else found this? Does this make sense to anyone else? Am I making a huge mistake somewhere?

A nice aspect of this (or at least I think so!) is that by filtering in this manner you can also create a sam file of non-unique mappings which you can monitor. This can be useful if one or more genes has a problem of generating a lot of non-unique maps which may give problems accurately estimating its expression. Also, you also get a list of how many multi hits you have in your data so you know the scale of the problem.

Best Wishes,
David.

__________________________________
Dr David A. Matthews

Senior Lecturer in Virology
Room E49
Department of Cellular and Molecular Medicine,
School of Medical Sciences
University Walk,
University of Bristol
Bristol.
BS8 1TD
U.K.

Tel. +44 117 3312058
Fax. +44 117 3312091