Issue with DADA2: Low ASV numbers in PacBio full-length ITS data

[cherry52HZ]

Dear Dr. Benjamin Callahan,
I’ve been following your 16S PacBio tutorial (https://benjjneb.github.io/LRASManuscript/LRASms_fecal.html) to analyze my full-length ITS amplicon data (obtained for marine phytoplankton) with DADA2. However, I encountered an issue that I would like your help:

   The number of ASVs generated is unexpectedly low—only 10–30 ASVs per sample on average. In contrast, when we analyze shorter amplicons (ITS1, ITS2, and 18S rDNA V4, ~300 bp) using the standard DADA2 pipeline for short reads, each sample usually yield hundreds of ASVs.

   I’m wondering if there might be oversights in our adapted pipeline (see below) or if specific parameters need adjustment for PacBio full-length ITS data. We have checked the DADA2 discussions at GitHub but didn’t find similar issues reported by others. I would greatly appreciate your insights on potential problem steps and optimization suggestions.

Thanks in advance
Regards,
Yingchao Li.

Our adapted script (key steps) is attached for your reference.

Issue with DADA2 Low ASV numbers in PacBio full length ITS data.docx

[benjjneb]

The key diagnostic is the verbose output from the derepFastq command on your primer-free/filtered fastq files. Example:

Encountered 28197 unique sequences from 29663 sequenced read.

This indicates that there is very low replication (multiple reads with the same sequence) in the data. DADA2 fundamentally relies on repeated observations of the same sequence to identify ASVs. Thus, it is not an appropriate tool for data with <10% replication for typical applications, as in that regime DADA2 is very insensitive to any but the most abundant ASVs in the sample.

The follow-up question would be why replication is so low here. Is it poor quality sequencing? Is it a hyper-diverse community? Off-target amplification? Technical issues with library construction that introduce spurious variation in the reads?

[nchen2022]

We also observed relatively low replication rates in the data, and I was wondering whether this might be influenced by the inherent error rate of PacBio HiFi sequencing. Specifically, could sequencing errors—even at reduced frequencies—create artificial sequence diversity, thereby making true biological replication appear lower than it actually is?

If so, I’d appreciate your insights on how best to address this. How should we evaluate the error rates of PacBio HiFi results? Would additional error correction steps or adjusting the bioinformatics pipeline help mitigate such effects?

[benjjneb]

If sequencing error rates are high enough such that the fraction of error-free reads is small, then there will be very little replication in the data. The fraction of error-free reads is approximately exp(-expected_errors_per_read). Expected errors per read is a product of the per-base error rate and the length of the reads. So to increase the error-free fraction, one can either decrease the per-base error rate by, for example, using standard HiFi sequencing instead of Kinnex 16S sequencing, and sequencing a shorter amplicon. It is also important to be using modern PacBio instruments/chemistries. For example PacBio RSII sequencing wasn't accurate enough to make ASV inference from full-length 16S effective.