Clustering PacBio CCS reads

Question

[lelle]

I have a mock metabarcoding dataset that was sequence with PacBio. A 4.5kb long stretch of DNA was amplified and sequenced with Circular Consensus Sequencing. This means the error rate is around 1%, but a lot of these are indels.

I want to cluster these sequences to create OTUs. What software would you use?

Clustering software I have tired so far and why they did not work (well):

1. vsearch/usearch: Relies on finding starting points for clusters by sorting reads by abundance. My reads contain enough errors to make almost all of them singletons.
2. cd-hit: Chooses cluster start by longest sequence. This sometimes choses the wrong starting points creating two clusters where it should be one.
3. swarm: only works for low error rates
4. HPC-clust: needs a multiple sequence alignment of all sequences. This is not feasible because of high error rate and high biological variability.

[Rohit]

Vmatch works well for clustering, but not sure about the indel problem though. Since indels are more common, you could reduce the indel penalty and reduce the identity % based on the distribution of indels

[lelle]

Thanks for the suggestion. It seems the clustering would be based on local alignments/matches. I need something that takes into account the full sequences. I would also prefer to use free (as in freedom) or at least open source software.

[Rohit]

Vmatch can cluster to keep only the longest from the subset (-nonredundant option), also the license is free for academic use and easy to obtain

[lelle]

I don't think the -nonredundant option would address my problem. I want the sequences clustered by full length similarity. Unless I misunderstand the documentation (I did not read all of it), vmatch searches for local matches and clusters by that. If I understand correctly it might be possible to force full length matches by setting the psmal parameter of the -dbcluster option to 100. But at that point it just feels like I am forcing the tool to do something it was not designed for.

[Rohit]

dbcluster 100 would be the best option as you said. The local matches are chained in the later steps, global alignments are more forced than local ones which are chained