We did 16s metagenome sequecning using illumina GA 2x.10% of the reads were compllete after assembly (paired end merge). Same 10% of the reads were retained for downstream processing. We had a good representation of the sample with in that 10%.

Most of the time, > 90% should be possible but this will depend on the theoretical overlap bp length and quality of the read at 3' end.

10% could be dangerous because if the failure to overlap is due to the nature of the 90% of the sequences (having higer GC/AT or composition which may compromise quality), bias can be a huge issue.

Further, difference in the length of PCR amplicon and the pair-end merging parameter can play a role too. For example, the v3 region is 170-190 bp. If somehow 10% are from the shorter v3 fragments i.e. the 170ish bp. then you're missing out on the bigger v3 fragment

Best to check the root cause for failure. Do you have the sequence QC report?

Han Ming Gan..here the 10% mentioned here was the reads rettained for further analysis after sequence assembly..100% completely merged sequence were considered for the analysis..from the raw reads after each preprocessing steps like quality filtering, dereplication, paired end merge, non atgc removal and chimera removal their is a reduction in read volume. This 10% refers to the reads remaing for OTU analysis after preprocessing steps.

Indeed 10% is really on the low end . Could this be a result of poor sequencing run? That too should be a concern. Using the miseq, nearly 90% or more of the merged reads are retained. Ga IOC are known to be poor in seq quality particularly at 3 end compared to other illumina platform. Diff chemistry perhaps

The reads were obtained from Hiseq Illumina sequencer platform. GC contents is around 50-50%, subsequently were filtrated by quality control. I would discard the poor sequecing run as a possible hypothesis. We are using CLC workbench changing kmer, the best assembly was around 10%. However using DNAstar, we reach around 40%. I keep you informed..

In my case, for enriched simple metagenome for example reactor sludge adopt to specific pollutant, we had get around 80% with both velvet, IDBA-UD and CLC; For more complex systems like digestion sludge or activated sludge from wastewater treatment plant the assembled reads usually count to the best 40%.

personally, i will suggest to take a look at IDBA-UD cause it automatically combines different kmer sizes for assembly, which is more convenient to use.

Plus combing libraries with different insert size is especially helpful to improve the assembly in term of N50.

The ideal assembly percentage may vary according to the software you use. Look within the program you are using the assembly that returns the best amount of N50. For example I use idba_ud software, this program the best N50 is always found in the file contig100.

http://bioinformatics.oxfordjournals.org/content/early/2012/04/06/bioinformatics.bts174

http://i.cs.hku.hk/~alse/idba_ud/

Assembly of metagenomic data is different from the assembly of data obtained from isolated sample as there are multiple species in metagenomics sample and so genomes. The regular assemblers are not applicable in case of metagenome because they take whole data as a single genome and generate a large de-bruinjn graph (Velvet) and need to identify various genomes from this large graph (MetaVelvet) 

So I recommand Metavelvet, which uses output of velvet itself.

Alternatively you can also use IDBA-UD and Omega.   

http://metavelvet.dna.bio.keio.ac.jp/

I think, it all depends upon coverage. If you have good coverage the assembly would be better, generally 30x coverage is considered as minimum to obtain a good quality assembly of targeted genomes.

Further, use metaspades for assembly that is best assembler till now for metagenomic investigation.