Fine-Grained Assignment of Unknown Marine eDNA Sequences Using Neural Networks

Area of Science:

• Ecology
• Bioinformatics
• Genomics

Background:

• Environmental DNA (eDNA) metabarcoding enables simultaneous species detection across diverse environments.
• Current bioinformatics tools struggle with accurate taxonomic assignments when species are missing from reference databases.
• Existing methods often overlook crucial nucleotide positional information.

Purpose of the Study:

• To develop a novel deep neural network architecture for enhanced eDNA metabarcoding analysis.
• To improve the accuracy of taxonomic assignments, particularly for underrepresented species.
• To address limitations in current bioinformatics tools for eDNA data.

Main Methods:

• Proposed a deep neural architecture leveraging nucleotide identity and positional patterns in short sequences.
• Conducted in-silico validation using NCBI GenBank sequences.
• Compared the new approach against state-of-the-art tools (Obitools, Kraken2, Lolo) and embedding methods.

Main Results:

• Achieved high classification accuracy: 94.7% at genus level and 86.5% at family level.
• Significantly outperformed existing reference-based pipelines.
• Demonstrated robustness with limited training data and improved performance with sequence alignment.

Conclusions:

• AI-powered eDNA metabarcoding offers a powerful complement to existing taxonomic assignment tools.
• The method is particularly valuable for incomplete reference databases and non-species-level resolution.
• Enhances capabilities for biodiversity monitoring and ecosystem management.