Adding ONT read support for ampliseq

[ShareiFae]

Description of feature

Would it be possible to add support for Oxford Nanopore Technology reads? I've found two projects (the second integrating the first) that might allow a quick implementation.

https://github.com/treangenlab/emu
https://github.com/genomic-medicine-sweden/gms_16S

[d4straub]

Hi @ShareiFae ,
emu seems indeed interesting, its also available as container https://bioconda.github.io/recipes/emu/README.html.
However, if I understood correctly, it compares sequences to a database and makes count tables from that. It is not a reference-free (de-novo) representative sequence (OTU/ASV/whatever) calculation like DADA2 or Deblur. Therefore I have the impression the tool is misplaced in this pipeline. Let me know if you have a different opinion.

[ShareiFae]

A shame, the downstream analysis and the report of this pipeline is really really good and I would like it to have it with nanopore data even it they cannot be done exactly the same way.

With EMU, as demonstrated by this other pipeline, you can create a phyloseq object that functions as Pseudo OTU tables:
https://github.com/josephpetrone/RESCUE/blob/master/detailed_steps.md#rstudio-and-phyloseq-handoff
https://github.com/josephpetrone/RESCUE

I've also found a pipeline that utilizes Quiime:
https://github.com/MaestSi/MetONTIIME

[samuell]

Hi @d4straub , I'm interested in hearing more about this. Is it the fact that it brings a database at all that you see as a problem here, or the size of the database?

EMU ships with a bundled database as an 84MB fasta file as part of the repo, but allows customizing this database, if that clarifies anything.

[d4straub]

Hi there,
databases are no problem. The "problem" is that many parts of the pipeline use ASV (/OTU) sequences. And EMU doesnt support that, is that correct?
There was/is also a discussion of nanopore analysis in nf-core slack: https://nfcore.slack.com/archives/CEA7TBJGJ/p1707383491473489
Essentially, newer nanopore data (from recent flowcells with high quality) might be possible to process with the pipeline and DADA2 with minor modifications (see benjjneb/dada2#759). Also possible is clustering with VSEARCH (that would be relatively easy to add into the pipeline).
Generally it is not only necessary to identify a solution, but also implementing it. I have neither the need nor time to do so. But if someone want to tackle that, I'm supportive.

[d4straub]

In the meanwhile a new pipeline was proposed for nanopore amplicon sequencing. This new pipeline is not aiming at producing ASVs but rather directly taxonomically classify sequences. See nf-core/proposals#104

[ShareiFae]

Following up on ONT support: I think Savont could be a promising route specifically for integrating full-length ONT 16S data into ampliseq.

Repo:
https://github.com/bluenote-1577/savont

Benchmark page:
https://github.com/bluenote-1577/savont/wiki/Savont-Preliminary-Results

What seems relevant for ampliseq is that Savont is an ASV-based workflow for high-accuracy long-read amplicons, and for ONT data it could potentially take over the ONT 16S ASV inference step: from full-length ONT 16S reads to ASVs / feature-table-like outputs, instead of relying on methods that were originally designed around short-read data.

From the repo/docs, the relevant pieces are:

• ASV generation from ONT R10.4 SUP full-length 16S reads
• read-to-ASV clustering output (final_asvs.fasta, final_clusters.tsv)
• optional 16S taxonomic profiling/classification against EMU or SILVA

What makes it interesting is the preliminary benchmark page:

• On the Zymo D6300 ONT R10.4 dataset, Savont is reported to be much more sensitive than unoise3 at low coverage.
• The benchmark states that traditional ASV methods like unoise need >1000x depth for reliable ASV generation, whereas Savont can recover ASVs reliably even above ~30x depth.
• For species detection, the same benchmark reports Savont as 100-1000x more sensitive than traditional ASV methods.
• On the Zymo D6331 dataset, Savont and Emu show comparable abundance estimates, but the benchmark notes potential false-positive species calls in Emu.
• In a small comparison against PacBio HiFi + unoise3, ONT + Savont ASVs were reported to be very similar to the “gold standard” ASVs.

So this looks like a realistic candidate for adding ONT 16S support in ampliseq, with Savont handling the ONT-specific ASV generation step and ampliseq using those results downstream.

Given the benchmark results, Savont looks like one of the most concrete current options for bringing ONT full-length 16S into ampliseq.

[d4straub]

Thanks, savont seems indeed very interesting. It appears like it could be added neatly into the pipeline as alternative to DADA2, adding the possibility to analyse ONT data. Currently the documentation isnt complete, a pre-print seems to be in preparation. I asked for recommended preprocessing and exact commands of the benchmarking in bluenote-1577/savont#5.

[d4straub]

https://doi.org/10.64898/2026.02.26.708165 claims that unoise3 but not DADA2 can generate ASVs from (most recent) ONT reads. However, only for shorter amplicons (up to near-full length 16S) and higher sequencing depth than PacBio sequencing is required, especially the longer the amplicons get or the more divers the sample is. I dont think that a substantial number of research is directed from Illumina sequencing to ONT for amplicon sequencing analysis based on these results.
Regarding preprocessing, the methods state

Primer trimming was carried out with Cutadapt v.5.1 [32] using the following parameters -g {Fw…RCRev} --revcomp --discard-untrimmed, where Fw and RCRev are the forward primer and reverse complement reverse primer without overhangs, respectively, found in Table 1. Empty sequences generated during primer trimming were removed using a custom script.

that is essentially done in ampliseq. The last step (rm empty sequences) is solved in ampliseq with cutadapt's --minimum-length 1.
If savont proves to solve the sequencing depth problem, as indicated in their wiki, it continues to appear to me as feasible approach to integrate to the pipeline.

[d4straub]

Small update:

• preprint of Savont in https://doi.org/10.64898/2026.05.26.727271
• I am testing Savont currently with ONT Zymo samples sequenced locally, it seems great, but best practice for multi-sample handling is unclear as of the time of writing

Preprocessing could be: cutadapt (if primers are there), Porechop_ABI, chopper (maybe unneccessary)

[d4straub]

I started an implementation of Savont. Still in a draft stage, feedback welcome: #1026

[d4straub]

Savont & preprocessing tools are now in dev branch, meaning Nanopore data analysis works now. However, to put that into a release, some fiddling with params is still necessary to integrate it better.
You can try it out with

NXF_VER=25.10.4 nextflow pull nf-core/ampliseq -r dev
NXF_VER=25.10.4 nextflow run nf-core/ampliseq -r dev -profile test_savont,singularity --ref_taxonomy_storage ref_taxonomy_storage --outdir results_test_nanopore_2.19.0dev -resume