We are performing trial shotgun assemblies of 80 millions paired-end sequences illumina reads obtained from a bioreactor microbial community. After removing contigs shorter than 1kbp, the assembly using Velvet has total length of 26 Mbases. Starting from this result we run Metavelvet following the procedure described in the manual and the final assembly has a total length of 20 Mbp. Not only the total size of the assembly was lower using metavelvet, but also the sum of the contigs larger than 10 kbp decreased. We have performed various trials changing parameters but the metavelvet assembly is still worse that the Velvet assembly. Does someone have suggestions?
Is it really worse ?
Metavelvet account for varying abundance, that is the main difference with velvet. In an environment with high diversity. Velvet (and all assemblers) assume clonality since the whole algorithm is based on this null-hypothesis: "should I not stick these two reads together based on their similarity" and an associated probability.
In this respect, velvet expects assumes an even coverage while metavelvet first bin reads in respect to contigs found in velvet coverage of with the same reads. A similar assembly as in velvet is then performed within each binning: many reads will not be assembled together, resulting in smaller contigs.
This is at the detriment of size, but assembled contigs are more reliable. To test an assembly quality, one could use: GC skew, codon use bias, etc over the whole contig-length.
I suggest this reading, while I'm confident there is more recent publications on the subject : http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0031386
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics
- Bioinformatic Tools