View Single Post
Old 03-10-2015, 07:45 PM   #18
Brian Bushnell
Super Moderator
 
Location: Walnut Creek, CA

Join Date: Jan 2014
Posts: 2,695
Default

BBNorm cannot be used for raw PacBio data, as the error rate is too high; it is throwing everything away as the apparent depth is always 1, since all the kmers are unique. Normalization uses 31-mers (by default) which requires that, on average, the error rate is below around 1/40 (so that a large number of kmers will be error-free). However, raw PacBio data has on average a 1/7 error rate, so most programs that use long kmers will not work on it at all. BBMap, on the other hand, uses short kmers (typically 9-13) and it can process PacBio data, but does not do normalization - a longer kmer is needed.

PacBio CCS or "Reads of Insert" that are self-corrected, with multiple passes to drop the error rate below 3% or so, could be normalized by BBNorm. So, if you intentionally fragment your reads to around 3kbp or less, but run long movies, then self-correct, normalization should work.

PacBio data has a very flat coverage distribution, which is great, and means that typically it does not need normalization. But MDA'd single cells have highly variable coverage regardless of the platform, and approaches like HGAP to correct by consensus of multiple reads covering the same locus will not work anywhere that has very low coverage. I think your best bet is really to shear to a smaller fragment size, self-correct to generate "Reads of Insert", and use those to assemble. I doubt normalization will give you a better assembly with error-corrected single-cell PacBio data, but if it did, you would have to use custom parameters to not throw away low-coverage data (namely, "mindepth=0 lowthresh=0"), since a lot of the single-cell contigs have very low coverage. BBNorm (and, I imagine, all other normalizers) have defaults set for Illumina reads.
Brian Bushnell is offline   Reply With Quote