Application of an AMF Bioinformatics Pipeline Across Five Datasets Using QIIME 2, VSEARCH and MaarjAM

Community Contributions Tutorials
, ,
1

Hello QIIME 2 Community,

We would like to share the application of an Arbuscular Mycorrhizal Fungi (AMF) bioinformatics pipeline across five sequencing datasets using QIIME 2, VSEARCH, and a MaarjAM-based Naive Bayes taxonomic classifier.

The collection includes two AMF sequencing datasets, Alerce and Cottonwood, along with three mock-community datasets: AMV45/AMDGR, WANDA, and RICE.

The objective of this contribution is to demonstrate how reference-based sequence screening and subsequent AMF taxonomic classification can be applied across different datasets using a reproducible bioinformatics workflow.

1. Datasets included in the workflow

The five datasets included in this contribution are:

A. Sequencing datasets

  1. Alerce

  2. Cottonwood

B. Mock-community datasets

  1. AMV45/AMDGR

  2. WANDA

  3. RICE

Each dataset has an associated workflow for sequence screening and subsequent taxonomic classification.

2. Overview of the bioinformatics pipeline

The general workflow consists of the following stages:

  1. Loading representative or amplicon sequences.

  2. Preparing or loading the corresponding non-AMF reference sequences.

  3. Screening sequences against the non-AMF reference using VSEARCH.

  4. Separating matched and non-matched sequences.

  5. Retaining non-matched sequences as candidate AMF sequences.

  6. Training a Naive Bayes classifier using MaarjAM reference sequences and taxonomy.

  7. Performing taxonomic assignment using QIIME 2.

  8. Generating taxonomic assignment artifacts and interactive visualizations.

3. Non-AMF reference-based sequence screening

VSEARCH is used to compare input sequences against the selected non-AMF reference database.

The workflows use the following principal parameters:

  • Search method: --usearch_global

  • Sequence identity threshold: 90%

  • Strand: plus

  • Number of threads: 22

The principal command structure is:

vsearch \
  --usearch_global input_sequences.fasta \
  --db nonAMF_reference.fasta \
  --id 0.90 \
  --strand plus \
  --matched results/nonAMF_hits.fasta \
  --notmatched results/AMF_retained.fasta \
  --blast6out filter_results.txt \
  --threads 22

The command generates two main sequence files:

Matched sequences:

nonAMF_hits.fasta

Contains sequences matching the selected non-AMF reference database.

Non-matched sequences:

AMF_retained.fasta

Contains sequences that did not match the selected non-AMF reference database. These are retained as candidate AMF sequences for subsequent taxonomic investigation.

The alignment information is saved in filter_results.txt.

4. Reference databases used across the five datasets

The notebooks use the following non-AMF reference artifacts:

Dataset Input sequences Non-AMF reference artifact
Alerce denoising/rep_seqs.qza nonAMF_WANDA_AML2.qza
Cottonwood denoise/rep-seqs-dada2.qza nonAMF_WANDA_AML2.qza
AMV45/AMDGR AMV45_AMDGR_amplicons.fasta noAMF_AMV45NF-AMGDR_trimmed.qza
WANDA WANDA_AML2_amplicons.fasta nonAMF_WANDA_AML2.qza
RICE denoising/rep-seqs.qza noAMF_AMV45NF-AMGDR_trimmed.qza

For the AMV45/AMDGR workflow, the non-AMF reference is prepared using the QIIME 2 feature-classifier extract_reads method.

The primer sequences used are:

  • Forward primer: AAGCTCGTAGTTGGATTTCTG

  • Reverse primer: CCCAACTATCCCTATTAATCAT

The extraction uses a minimum sequence length of 300 bp and a maximum sequence length of 700 bp.

The resulting reference artifact is subsequently used for VSEARCH screening in both the AMV45/AMDGR and RICE workflows.

5. Examination of screening outputs

The notebooks include a Python-based FASTA record-counting procedure to examine the filtering outputs.

The recorded validation results are:

Dataset Matched non-AMF records Non-matched retained records
Alerce 1,307 4,135
Cottonwood 369 1,434
AMV45/AMDGR 0 61
WANDA 1 68
RICE 1 673

These values represent FASTA records counted in the corresponding notebook outputs.

Non-matched records are considered candidate AMF sequences for subsequent analysis; they are not automatically confirmed AMF identifications.

6. MaarjAM-based taxonomic classification

Following sequence screening, the retained candidate sequences are subjected to taxonomic assignment using the MaarjAM reference database.

The workflow uses two reference artifacts:

  • maarjam-ref-seq.qza

  • ref-taxonomy-maarjam.qza

A Naive Bayes classifier is trained using the QIIME 2 feature-classifier plugin and the fit_classifier_naive_bayes method.

The resulting classifier is then applied to the retained candidate sequences using classify_sklearn.

The same general classification approach is implemented across the five dataset workflows.

7. Generation of taxonomic assignment outputs

The taxonomic assignment results are saved as QIIME 2 artifacts (.qza).

The corresponding results are also converted into interactive visualizations (.qzv) using the QIIME 2 metadata tabulate visualizer.

These outputs allow researchers to inspect the assigned taxonomy associated with the candidate AMF sequences.

The workflow produces separate taxonomic assignment artifacts and visualizations for the individual datasets.

8. Software environment

The associated resource collection documents the following software information:

  • QIIME 2 framework: 2025.4.0

  • Python: 3.10.14

  • scikit-learn: 1.4.2 (version reported during classifier training)

A QIIME 2 warning was generated during classifier training indicating that the trained classifier may produce unreliable results with other scikit-learn versions.

Users should therefore maintain a compatible software environment when reproducing the workflow.

9. Resource availability

The sequencing datasets, mock-community resources, non-AMF reference artifacts, MaarjAM reference resources, trained classifier, and associated QIIME 2 outputs are available through our Zenodo preview record:

INNOVATIVE PIPELINE FOR ACCURATE AMF QUANTIFICATION

Access the Zenodo preview record

10. Potential applications

This collection of workflows may be useful for:

  • Reference-based screening of AMF sequencing data.

  • Investigation of candidate AMF sequences.

  • MaarjAM-based taxonomic classification.

  • Application of AMF bioinformatics workflows to different sequencing datasets.

  • Examination of mock-community taxonomic assignments.

  • Development and evaluation of reproducible fungal metabarcoding workflows.

11. Community discussion

We welcome feedback and suggestions from the QIIME 2 community regarding:

  • Reference database selection for AMF sequence screening.

  • VSEARCH identity thresholds and sequence-filtering strategies.

  • Application of MaarjAM-based taxonomic classifiers.

  • Reproducibility of workflows across different datasets.

  • Evaluation of AMF taxonomic assignments using mock communities.

We hope that sharing these five workflows and their associated resources will be useful to researchers working on AMF metabarcoding, fungal biodiversity, and microbial community analysis.

Thank you to the QIIME 2 community for providing a platform for sharing reproducible bioinformatics resources.