PacBio full-length 16S CCS data: almost all reads unique during dereplication and very few reads retained after denoising

[alitcd]

Hello,

I am analysing PacBio Sequel II full-length 16S rRNA CCS data (~1450 bp) using the DADA2 long-read workflow, but I am observing an unusually high number of unique sequences.

Dataset

Platform: PacBio Sequel II

Data type: CCS reads

Target: full-length 16S (~1450 bp)

~10,000–13,000 reads per sample

Primers trimmed using cutadapt

CCS minimum passes ≥5

Primers:

F: AGRGTTTGATYNTGGCTCAG
R: TASGGHTACCTTGTTASGACTT

During dereplication:

derepFastq()

almost all reads appear to be unique (for example ~11,200 unique sequences from ~12,300 reads).

After running:

learnErrors()
dada()
removeBimeraDenovo()

only a very small number of reads remain per sample.

This seems similar to issue #1975 where it was noted that when almost every read is unique, DADA2 cannot infer ASVs because repeated observations of the same sequence are required.

Before concluding that the data is unsuitable for DADA2, I wanted to ask:

Is such a high unique/read ratio expected for PacBio full-length 16S CCS data?

Could this behaviour be caused by mixed read orientation or incomplete primer trimming?

Are there recommended filtering parameters for PacBio full-length 16S data prior to denoising?

I can provide example FASTQ files or the full script if helpful.

Thank you for any suggestions.

[alitcd]

Yes, I have read that discussion, and I believe this is likely the same issue.

1. Yes, this dataset is PacBio full-length 16S CCS data.
2. We checked the read orientation and it is correct. We also confirmed that the primers were properly trimmed. Based on these checks, the problem does not appear to be related to mixed orientation or incomplete primer removal.
   Please let us know if there are any additional filtering recommendations or steps you would advise for this type of dataset.

[benjjneb]

The typical error rates for PacBio HiFi data using modern machines and chemistries are low enough for DADA2 to work with long-read 16S, but there is run-to-run and machine-to-machine variability. If the error rates in your run are higher than usual, then there may be too little replication to effectively use DADA2.

Another possiblity is that you are working in a hyperdiverse environment. In this case, one can still use DADA2, but sensitivity will be constrained by the requirement of at least 2 repeated error-free reads of an ASV. Given your read depth of about 10k/sample, and assuming error-free-read rates are 50% based on other HiFI 16S datasets, you will be constrained to detecting ASVs at 4/10k0.04% or higher. It is possible that a large amount of your sampled community could be below that limit of detection.

almost all reads appear to be unique (for example ~11,200 unique sequences from ~12,300 reads).

This diagnostic suggests that DADA2 will be limited in its ability to detect rare vatiants, but there is still enough replication there that it will likely work to detect the more abundant parts of the community.

[cjfields]

We have also seen this issue pop when we're given data from the sequencing center that hasn't been prefiltered to 99.9% accuracy.