unreproducible velvetg results?

[jeffgao]

Hello,

When I generated a Roadmaps and a Sequences by velveth, and applied multiple times of velvetg on them, I found the results were not consistent. The resulted node number, n50 length, max length, total length, and used reads would always be greatly different each rime I ran velvetg.

I am sure I used the same Roadmaps and Sequences files for each velvetg run, and I deleted LastGraph, stats.txt, contigs.fa, Graph2, and PreGraph files but only kept Roadmaps and Sequences files before I ran velvetg.

I always used the same velvetg command line as the following,
./velvetg results -exp_cov 8 -cov_cutoff 4

But I always saw different results showing up after the assembly: e.g., n50 length varies from 300k to 500k.

More details:
- The velvet version is 1.1.04;
- My system is ubuntu 11.04;
- I used the MP (shortjump) data as described in http://gage.cbcb.umd.edu/data/Staphy...aureus/README;
- I used reverse complement reads and shuffled them into a single file.
- I used "./velveth results 23 -fastq -shortPaired data/shortjump_shuffled.fastq" to generate Roadmaps and Sequences files.
- I used "./velvetg results -exp_cov 8 -cov_cutoff 4" for assembly.

Would anyone give me some hints that what was going on? Thanks a lot!

Best,
Jeff

[jeffgao]

Am I lucky to get some quick replies? such as whether your velvetg results are reproducible (yes or no) and your guess if your answer is "no". Thanks again!

[aloliveira]

ooh,

I constantly use the velvet and every day I deal with this problem, I do not know where the is the problem, to be honest. I never found any documentation or information as to why this occurs. But I think any kind of help or comments would be welcome.

[jeffgao]

aloliveira,

Here is the answer I got from velvet's author Daniel Zerbino:

this is probably linked to the fact that you are using OpenMP.

The number of reads used is extremely stable (coefficient of variation 6e-5 !) as well as the total length (3e-2). I suspect the large jumps in N50 are mainly due to the fact that the number of contigs rather small (~5200, of which probably many tiny ones) so at this level of granularity, there are only a handful of big gaps which are being bridged. If for random reasons a gap is bridged or not, this is enough to multiply the max contig length by a factor of 2, and the N50 is notoriously unstable on small sets (it behaves essentially like a median).

If you remove the "OPENMP=1" while compiling velvet, the output will be consistent. I have tested this. Anyway, looking at N50 alone will be confusing if there are lots of small contigs existing.

- Jeff

[aloliveira]

Thank you very much jeff.

[rahularjun86]

Dear all,
Can we use same Sequence and Roadmap files generated from velveth by "OPENMP=1" for different cov_cutoff (2 to 15)? Or I need to rerun velveth for different cov_cutoff ??
Many thanks,
Rahul