Skip to content
DraconiticResearch
Search Sources API
226,742 studies31 sourcesUpdated 40 min ago
Peer-reviewedCC BY 4.0Molecular Ecology Resources

Transductive Learning From a Transformer Foundation Model Improves the Taxonomic Classification of eDNA Metabarcodes

6 affiliations
  1. 1MARBEC, Univ Montpellier, CNRS, IFREMER, IRD Montpellier France
  2. 2CEFE, Univ Montpellier, CNRS, EPHE‐ PSL University, IRD Montpellier France
  3. 3Institut Universitaire de France Paris France
  4. 4Inria, Univ Montpellier, LIRMM, CNRS Montpellier France
  5. 5SPYGEN Le Bourget‐du‐Lac France
  6. 6Espace‐Dev, IRD, Univ Montpellier, Univ Réunion Montpellier France

Abstract

Environmental DNA (eDNA) has emerged as a powerful and cost‐effective biodiversity monitoring tool, providing comprehensive and non‐invasive taxonomic inventories. However, eDNA‐based monitoring is strongly dependent on the completeness of genetic reference databases, which varies greatly among regions and taxa. Since completing gaps at the species level is unrealistic in the short term, an alternative is to infer coarser taxonomic assignments for sequences belonging to species not yet in reference databases. To address this challenge, we introduce BouillaBERT, a novel deep learning method tailored for the operational classification of eDNA sequences from unknown genera. BouillaBERT leverages data augmentation, transfer and transductive learning to overcome the small size and imbalance of eDNA reference databases for fishes. We implement and train the model across four commonly used fish eDNA markers and compare its performance with lowest common ancestor (LCA) alignment methods and existing deep learning approaches. BouillaBERT outperforms all tested LCA alignment and deep learning methods in recall for the assignment of unknown sequences across a range of commonly used fish eDNA primers, while achieving comparable or higher precision. BouillaBERT can be rapidly re‐trained to fit a custom sample set or used as‐is, enabling more consistent biodiversity assessments across undersampled regions or taxa.

Related

  1. Advances in 16S Taxonomic Classification: Impacts on Species-Level Resolution of Female Bladder Microbiota
    Lisa Karstens, Benjamin Lamp, Juliette Kudrick +1
    bioRxivPreprintCC BY-NC-NDFull textBioinformatics & computational biologybioRxivPMC
  2. MEG-GPT: A transformer-based foundation model for magnetoencephalography data
    Rukuang Huang, Sungjun Cho, Chetan Gohil +2
    Imaging NeurosciencePeer-reviewedCC BY7 citedNeuroscienceOpenAlex
  3. A lightweight privacy-aware federated vision transformer for tomato leaf disease classification
    Debangana Ram, Manasi Gyanchandani, Akhtar Rasool +1
    Scientific ReportsPeer-reviewedCC BY-NC-ND 4.0Full textPlant biologyCrossref
  4. MangoSwin-attention: a swin transformer with attention pooling for generalizable mango leaf disease classification
    Saksham Khurana, Valarmathi Prahasam, Brindha Subburaj +2
    Scientific ReportsPeer-reviewedCC BY-NC-ND 4.0Full textPlant biologyCrossref
  5. A hybrid EfficientNet-B0-CBAM-Vision Transformer framework for accurate cherry leaf disease classification
    Ateeq Ur Rehman, Jaeyoung Choi, Ahmad Almogren +3
    Scientific ReportsPeer-reviewedCC BY-NC-ND 4.0Plant biologyCrossref
BibTeX
@article{bettinger2026transductive,
  title = {{Transductive Learning From a Transformer Foundation Model Improves the Taxonomic Classification of eDNA Metabarcodes}},
  author = {Simon Bettinger and Auguste Verdier and Morgane Bruno and Stéphanie Manel and Alexis Joly and Maximilien Servajean and Alice Valentini and Rodolphe Devillers and David Mouillot},
  journal = {Molecular Ecology Resources},
  year = {2026},
  month = nov,
  doi = {10.1111/1755-0998.70204},
  url = {https://doi.org/10.1111/1755-0998.70204},
  note = {CC BY 4.0},
}
RIS
TY  - JOUR
TI  - Transductive Learning From a Transformer Foundation Model Improves the Taxonomic Classification of eDNA Metabarcodes
AU  - Simon Bettinger
AU  - Auguste Verdier
AU  - Morgane Bruno
AU  - Stéphanie Manel
AU  - Alexis Joly
AU  - Maximilien Servajean
AU  - Alice Valentini
AU  - Rodolphe Devillers
AU  - David Mouillot
PY  - 2026
DA  - 2026/11/11
JO  - Molecular Ecology Resources
DO  - 10.1111/1755-0998.70204
UR  - https://doi.org/10.1111/1755-0998.70204
AB  - Environmental DNA (eDNA) has emerged as a powerful and cost‐effective biodiversity monitoring tool, providing comprehensive and non‐invasive taxonomic inventories. However, eDNA‐based monitoring is strongly dependent on the completeness of genetic reference databases, which varies greatly among regions and taxa. Since completing gaps at the species level is unrealistic in the short term, an alternative is to infer coarser taxonomic assignments for sequences belonging to species not yet in reference databases. To address this challenge, we introduce BouillaBERT, a novel deep learning method tailored for the operational classification of eDNA sequences from unknown genera. BouillaBERT leverages data augmentation, transfer and transductive learning to overcome the small size and imbalance of eDNA reference databases for fishes. We implement and train the model across four commonly used fish eDNA markers and compare its performance with lowest common ancestor (LCA) alignment methods and existing deep learning approaches. BouillaBERT outperforms all tested LCA alignment and deep learning methods in recall for the assignment of unknown sequences across a range of commonly used fish eDNA primers, while achieving comparable or higher precision. BouillaBERT can be rapidly re‐trained to fit a custom sample set or used as‐is, enabling more consistent biodiversity assessments across undersampled regions or taxa.
SN  - 1755-098X
SN  - 1755-0998
ER  -