I would like to understand the effects of the end-to-end read mapping of bowtie1

[Rachel Hillmer]

Hello bioinformatics pioneers!

Upon digging into the bowtie/tophat documentation (I am using bowtie1 since I have very short (29bp after barcode trimming), single end reads), I realize that the algorithm works (I think) differently from what I would a priori assume. I am using bowtie 1.4.1, not called directly, but called from within tophat. I am using the --transcriptome-index flag to guide mapping to annotated cDNAs first, since I am working with a very well-annotated genome (Arabidopsis thaliana).

My intuition would say, for a well-annotated genome (such as mine), map to the annotated cDNAs first, transform these into chromosome-mapped reads, and then map the leftover reads onto the chromosomes. (This is not perfectly satisfying, I should probably want to map in a more unbiased way, but I am assuming that what I care about is quantifying the expression of a known gene, rather than discovering new spliceforms. Perhaps this is a big assumption.)

But, as far as I can tell, bowtie has another filter: it only accepts end-to-end reads (I know bowtie2 has made such mapping optional, making unnconnected local alignments possible). Most notably, this eliminates reads that happen to fall at low abundance in introns and at the edges of UTRs. I presume this is a noise filtering function. Does anyone know of any quantification that has been done (from your own analysis, or in a paper) on the effects of this filtering?

Cheers,
~Rachel

[chadn737]

The local alignments does soft-clipping to enhance the alignment. When it does this it will ignore some of the base pairs at the ends of the reads. I think the primary reason for choosing a local alignment is when you have significant drop-offs in read quality at the ends of the reads and would thus expect more mismatches.

I would not however, use such an option when you have very short reads, such as yours. Its one thing to soft-clip a few base pairs from longer reads. But at very short read lengths, you could actually end up with more inaccurate mapping.

As a side, I would not worry about. I have been working with Arabidopsis RNA-seq for a few years now. If you have quality data, then you will have no problem mapping the vast majority of your data in end-to-end mode. I have had samples with varying amounts of adapter contamination and other problems and can still get an average of 97% of reads mapping, of which ~87% are unique. At your short read lengths I would suspect a greater amount of non-unique mapped reads....all the more reason to map end-to-end.

[Rachel Hillmer]

Oh! I understand now. I somehow thought "end-to-end" meant that in order for a read to be aligned, its tail had to exactly pick up where the head of another read began, for example:
--------->----------->
but not
---------> ------------>

I now understand. End to end means keep the entire read, from one end of it to the other.
Thanks for the clarification.