Paired-end questions

[scubachris]

Greetings to all!
I was recently given some sequencing data for a not yet sequenced organism, so I am trying a de-novo assembly with Velvet. This is the first time I've been working with sequencing data, so I have one important question (and lots of others, but we'll have time for those):
The data is FASTQ Illumina 1.3+ paired-end. I have shuffled the two files according to the Velvet manual, but I have a lot of low quality positions, for example:

@HWUSI-EAS1785_0222:5:1:16937:1092#NCCAAT/1
GATCGGAAGAGCACACGTCTGAACTCCAGTCACGCCAATATCTCGTATGCCGTCTTCTGCTTGAAAAAANNANNAAAAAA
AAGACANAAAANNAGA
+
hhhhhhhhhhhhhhhhhhhhhghhhhhhhhhhhhhhhhhhhahhhhhhhhhhfhhgfggdhhgadBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBB
@HWUSI-EAS1785_0222:5:1:16937:1092#NCCAAT/2
NNNNNNNNNNGCGTCGTGTAGGGAAAGAGTGTAGATCTCGGTGGTCGCCGTATCATTAAAAAAAAAAAAAAAAACAGGAA
GAGGAAATATGAGAAG
+
BBBBBBBBBBLJJHK__________YYTYYYYYYYb____b_bb_YYYYY_______Z____BBBBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBB

I am thinking of trimming the low quality positions, but then the mate-pairs would have different sizes. Would this have an effect on the Velvet output? You know, garbage in, garbage out...
Thanks a lot, and really hoping for an answer..

[scubachris]

OK, so far I am replacing all low quality bases with "N", so I don't have to trim anything. Still, any other ideas are welcome....

[seb567]

OK, so far I am replacing all low quality bases with "N", so I don't have to trim anything. Still, any other ideas are welcome....

Velvet will replace every N by a A. Therefore, replacing Ns by As will have no impact on the memory usage, at least I believe.

If you have way too much data, you can discard any pair of reads if one of the two sequences contains low-quality bits.

You can also correct sequencing errors using Quake.

Website: http://www.cbcb.umd.edu/software/quake/
Paper in Genome Biology: http://genomebiology.com/2010/11/11/R116/abstract

The software coming from the team of Professor Steven Salzberg are those with the highest quality and are usually very easy to use.

I, for example, use hawkeye, AMOS, MUMmer and bowtie, just to name a few.


-seb

p.s.: You can also try Ray, a de novo assembler I created.

Website: http://denovoassembler.sourceforge.net/
paper in Journal of Computational Biology http://dx.doi.org/doi:10.1089/cmb.2009.0238

[scubachris]

Hello, thanks a lot for the reply! I'll have a look at both programs!
As for the low quality reads, since my coverage wasn't much (around 15), I was trying to be somewhat parsimonious, so I did this:
1. I trimmed the 'B' characters from the ends of the mate pairs, and then trimmed the longest one to match the other.
2. If one of the mate pairs was completely useless, I discarded it, stored the other elsewhere, and treated it as a single read.

[scubachris]

Hello again,
I managed to do a decent filtering by trimming trailing 'B's and excluding reads with Phred < 20.
After running a De novo assembly with CLC Genomics Workbench, I got some contigs with extremely high coverage. (see here http://img683.imageshack.us/f/67898655.png/). (My predicted coverage by taking into account number of reads and genome size, is around 15-20x).
The parts of these contigs that have the most coverage are repetitive sequences. Also, a couple of them are like this: http://img198.imageshack.us/f/86423716.png/
Is this some sort of contamination? For this particular sequence I could not get any BLAST results.
What should I do with these? Perhaps run the assembly again with other parameters?
Any help appreciated!

[Thorondor]

velvet does fine with paired ends with different length. If you trim first and then shuffle the reads together you need to check if the trimming did not discard any reads! If so, you need to use the supported select_paired.pl script and then shuffle the sorted reads.

Your high coverage reads are obvious repeats. Getting better results with repetitive sequences depends on your insert length, read quality, coverage and read length.

I am not really into CLC but what I heard so far is that it does some weird things. ;-) And you do not know how it exactly works. :-/

[scubachris]

Hi, thanks a lot for the reply! I'm afraid the problem is with the coverage, hopefully we'll be doing some more sequencing soon! I didn't see much difference in the results of Velvet, ABySS, or CLC, so I'll stick with CLC for now, since it has quite a useful toolbox as well.