ViBE: a hierarchical BERT model to identify eukaryotic viruses using metagenome sequencing data

Area of Science:

• Virology
• Bioinformatics
• Genomics

Background:

• Viruses are widespread and mutate rapidly, making their identification challenging, especially novel strains.
• Current methods like homology-based sequence alignment struggle to detect distantly related viruses.
• Accurate viral detection from metagenomic data requires models that understand genomic patterns.

Purpose of the Study:

• To develop a novel computational model for detecting eukaryotic viruses in metagenomic samples.
• To classify detected viruses at the order level using deep learning.
• To improve the identification of novel and diverse viral sequences.

Main Methods:

• Development of a hierarchical BERT model named ViBE.
• Pre-training ViBE on read-like sequences from reference viral genomes.
• Fine-tuning ViBE for classifying paired-end reads of DNA and RNA viruses into taxonomic orders.

Main Results:

• ViBE demonstrated higher recall than alignment-based methods while maintaining comparable precision.
• ViBE significantly outperformed existing alignment-free methods across all tested scenarios.
• The model's effectiveness was validated on real-world sequencing datasets, including the vaginal virome.

Conclusions:

• ViBE offers a powerful, accurate, and efficient approach for eukaryotic virus detection and classification from metagenomic data.
• The model enhances the ability to discover novel viruses and expand our understanding of the viral ecosystem.
• ViBE represents a significant advancement in computational virology for analyzing complex biological samples.