DeepCOI: a large language model-driven framework for fast and accurate taxonomic assignment in animal metabarcoding

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Metabarcoding faces challenges with incomplete taxonomic data and high computational demands.
• Accurate and efficient taxonomic assignment is crucial for biodiversity assessment.

Purpose of the Study:

• To introduce DeepCOI, a novel large language model-based classifier for DNA metabarcoding.
• To enhance the accuracy and efficiency of taxonomic assignment using cytochrome c oxidase I (COI) gene sequences.

Main Methods:

• Developed DeepCOI, a classifier pre-trained on seven million COI gene sequences using self-supervised learning.
• Evaluated DeepCOI's performance across eight major phyla, comparing it against existing metabarcoding methods.

Main Results:

• DeepCOI achieved high accuracy with an AU-ROC of 0.958 and AU-PR of 0.897.
• Demonstrated significantly reduced inference time compared to current methods.
• Showcased interpretability by identifying key taxonomically informative sequence positions.

Conclusions:

• DeepCOI offers a scalable and efficient solution for accurate taxonomic assignment in metabarcoding.
• The model enhances biodiversity assessment and environmental monitoring capabilities.
• DeepCOI represents a significant advancement in computational approaches for DNA-based biodiversity studies.