[mira_talk] Re: Lots of contigs, then segmentation fault
- From: Bastien Chevreux <bach@xxxxxxxxxxxx>
- To: mira_talk@xxxxxxxxxxxxx
- Date: Tue, 3 May 2011 21:16:31 +0200
On Tuesday 19 April 2011 18:14:47 Egon Ozer wrote:
> I'd be happy to provide my data to you for testing. Do you want the sff
> files or my extracted fasta, qual, and xml files for the 454 data?
Hello Egon,
your data set made MIRA (and me) sweat, actually, quite a lot. It's not that
much that version 3.2.1 crashed on it, but that my newer development version,
while not crashing, performed ... really not good: way too many contigs for my
liking.
I've been busy the week-end over to understand what happened that MIRA
absolutely did not like that data set and found the reason: it looks like that
this paired-end FLX data contains a lot more false duplicates than I have ever
seen up to now. These false duplicates contain, I think, PCR artefacts ... and
these "sequencing errors" let MIRA believe that there are repeats and/or
ploidy differences.
I had to develop a couple of new algorithms to deal with these kind of things.
Not everything I thought of has been implemented yes, but already I think the
improvements are good enough to test. E.g., here are the results of 3.2.1.15:
Number of contigs: 116
Largest contig: 893586
N50 contig size: 172613
N90 contig size: 34046
N95 contig size: 21118
and here for my current development version:
Number of contigs: 75
Largest contig: 901116
N50 contig size: 397873
N90 contig size: 108334
N95 contig size: 52586
Almost halved the number of contigs and N50 doubled. Taking then a hybrid
assembly with your 454 and Solexa data, I get this:
Number of contigs: 55
Largest contig: 894849
N50 contig size: 588120
N90 contig size: 139889
N95 contig size: 62263
The number of contigs was more than halved and the N50/90/95 numbers trippled.
The next release on SourceForge will contain those enhancements (but can take
a week or two). Contact me if you want to test the current head of the
development tree before that :-)
B.
Other related posts:
