ORF prediction in assembled NGS metagenomic/metatranscriptomic data

[someperson]

Hi, Sorry I'm relatively new to this and am just looking for some qualified input.
I have sequenced several metatranscriptomes using an Illumina GA system and I have assembled them using meta-velvet. Now I would like to do a reliable gene-prediction on the assembled contigs.

Basically I would use Orphelia or FragGeneScan for this, however from the respective descriptions I gather that they have been specifically designed for the Identification of coding regions in short single-reads not in assembled data. Whereas classical ORF-prediction tools like GLIMMER suppose that you have homogenic and more-or-less complete sequence data.
My data however is very heterogenic and my largest contigs are just over 2kb.

Therefore, what would be the best approach for ORF-prediction in assembled metatranscripome/metagenome-data?

Can I use Orphelia or FragGeneScan for this, or are they unreliable for datasets of highly varying sequence lengths? Do you know of any better suited tools?

EDIT:
I've tested FragGeneScan on my data and do get peptides of varying lengths, which is hopeful. But since its metatranscriptomic data, I can't really validate how much false positives i get or how much of the genetic potential is missed.

Any experience or opinions on wether I should optimize my data for such gene-predictions (For example to sort the contigs into size and so seperate predictions for each contig-sizerange (e.g. <100bp, <500bp and >500bp)?

[GenoMax]

I am not answering your question but wanted to mention MG-RAST, just in case http://metagenomics.anl.gov/.

[someperson]

Thanks for the reply. I am aware of MG-RAST, which also does gene-prediction based on FragGeneScan. But here also, as far as I know, this is basically done on unassembled data (MG-RAST does not assemble, does it?).

Also I don't want to have an overview of the functions present (as MG-RAST gives) but I want to get to the actual peptide sequences to do comparative further in-depth analyses on amino-acid sequence level.

[JackieBadger]

Most people use this http://www.bioinformatics.org/sms2/orf_find.html

[rhinoceros]

FGS works just fine on assemblies, it's what I've using on Illumina assemblies. Be-aware though that the latest version 1.16 has a bug which leads to memory allocation failures on random sets. However, some guy from the mg-rast team has fixed it. It's available here. Just remember to use the correct parameters, I believe '-complete=1 -train=complete' for predicting from contigs..

[someperson]

Thanks, that's good to know!