benjjneb
commented
1 month ago First, "quality trimming" via cutadapt (which is the per-read trimming of tails of reads based on the read-specific quality scores) is NOT recommended for working with DADA2. It is better to handle that with a consistent across-reads truncation length, as per the truncLen approach in filterAndTrim.
That said, given the high level at which reads are passing the denoising step, this is probably not a major problem in your data.
With deep sequencing can come many types of spurious ASVs from different sources, including rare off-target amplification, rare amplification issues, and rare library artefacts among others. But if these things are rare, they are not necessarily much of a concern. The first thing to check is whether that is the case: Are these unexpected ASVs that are outside the expected length range or taxonomic ID typically rare? You've shown a table of all ASVs, but what if you weight this by their abundance? Is it almost all as expected? A simple follow-on solution if that is the case is to filter the data based on the known length distribution of the targeted amplicon to remove much of this.
In short, I don't see any obvious issue with your data from what you've posted. You probably have some spurious off-target stuff in your data at very low abundances. When looking at unique ASV counts that can seem significant, but when accounting for the abundance it usually isn't. But certainly check if my guess is the case.
msalamon2
Hello,
I have a dataset sequenced with Novaseq 6000 with 250PE for 77 libraries and four primers: 12S, COI, ITS and 18S, which gave ~1-6 million paired reads per library, with some lower exceptions with < 1M reads. Before running DADA2, I used cutadapt to remove primers, nextera transposase sequence and polyG tails in sequential steps (this was probably not necessary and the best approach), using the options --discard-untrimmed for the primer removing step, --nextseq-trim=20 for all the steps as it is recommended for Novaseq ,and discarding reads <10 bp with -m 10 at the end. I am worried about the high number of ASVs output by DADA2 for each primer, which is between 33K (ITS, using the ITS pipeline workflow) and 80K (18S). I used the following parameters for the filterAndTrim, changing the truncLen for each primer (except ITS) based on the length at which the quality drop in the quality profiles post-cutadapt, and the length distribution of the reads checked with fastqc (not all the reads seemed to be the same length post-cutadapt):
out <- filterAndTrim(fnFs, filtFs, fnRs, filtRs, maxN = 0, maxEE = c(2,2), truncQ=2, rm.phix = T, truncLen=c(185,185), compress=T, multithread=F)
I also used the custom Loess function (weights, span, and degree) for the error learning to accommodate for the binned quality scores, based on the solution proposed by @hhollandmoritz in issue #1307. I have also used the option pool="pseudo" for the denoising, but not for other steps.
The quality scores post-trimming and filtering looked good, as well as the error profiles and the summary table of the % of reads passing each step. Here is an example below for 12S: BCV001_12SMimammal_trimmedDADA2.pdf plot_errors_DADA2_eDNA_12SMimammal_F.pdf plot_errors_DADA2_eDNA_12SMimammal_R.pdf
Given this, I moved forward with the taxonomic assignment with sintax in vsearch using a cutoff of 0.8, but got relatively low % of assignments at the species/genus/family level (12-18%), even though I was using recent, broad (all Eukaryotes) reference datasets, except for the plants with ITS, which gave 84% of assignments for these taxonomic levels.
Due to the low assignment % and the high number of ASV, we are wondering if something went wrong with the trimming with cutadapt or with DADA2. After reading this issue #1609, it seems that one issue might be that using cutadapt resulted in length variation for the reads, but I am not sure how I would trim the primers + nextera transposase sequence + polyG tails without using it. Indeed, it seems that there is substantial length variation in the ASVs for all my primers (example below is for 12S after chimera removal):
I am still unsure if I should be interpret these results as an indication of spurious ASVs (thus resulting in the low % of assignment) due to introduced noise from the trimming with cutadapt, or if they could also be explained by the deep sequencing and the excepted complexity of the communities sampled.
Your input would be very helpful, thank you. Mathilde Salamon