New deep learning-based methods for visualizing ecosystem properties using environmental DNA metabarcoding data.

Area of Science:

• Ecology
• Bioinformatics
• Computational Biology

Background:

• Environmental DNA (eDNA) metabarcoding is a powerful tool for biodiversity assessment.
• Current eDNA data analysis methods struggle to extract all ecological information due to data complexity.
• Improved dimensionality reduction and clustering are needed for eDNA data.

Purpose of the Study:

• To develop and evaluate novel deep learning methods for analyzing eDNA metabarcoding data.
• To enhance the extraction and visualization of ecological information from eDNA samples.
• To improve the accuracy and efficiency of biodiversity monitoring using eDNA.

Main Methods:

• Developed two deep learning (DL) methods using neural networks (NNs): variational autoencoders and deep metric learning.
• Combined sequence counts per molecular operational taxonomic unit (MOTU) and nucleotide sequences as input.
• Applied methods to three diverse eDNA datasets.

Main Results:

• DL methods accurately represent biodiversity indicators (MOTU richness, α- and β-diversity) in a 2D latent space.
• Nonlinear DL methods effectively extract features and avoid common eDNA biases.
• Outperformed traditional dimensionality reduction techniques (PCA, t-SNE, NMDS, UMAP).

Conclusions:

• Neural networks offer a more efficient approach for structuring eDNA metabarcoding data.
• The developed DL methods improve ecological interpretation and biodiversity monitoring.
• This work advances the application of machine learning in ecological research.