HieRMVir: Interpretable Viral Classification via Hierarchical Deep Learning.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate pathogen identification is crucial, especially for pandemic threats, but traditional methods struggle with complex genomic data.
• Existing genome identification tools often overlook biological taxonomy hierarchy and feature informativeness across classification levels.

Purpose of the Study:

• To develop a novel hierarchical deep learning framework for accurate and interpretable viral genome classification.
• To address limitations in existing methods by incorporating taxonomic structure and feature informativeness.

Main Methods:

• Proposed HieRMVir (Hierarchical Random forest and Mutual information-based Viral genome classifier), a deep learning framework.
• Integrated Random Forest (RF) for feature weighting and Mutual Information (MI) for attention regularization.
• Employed a three-level classification system guided by feature importance and MI scores for informative k-mer patterns.

Main Results:

• HieRMVir achieved an average accuracy of 95.8% (95% CI: 95.3-96.4%) on over one million genome sequences.
• The framework outperformed existing methods across multiple performance metrics.
• Hierarchical metrics and attention weight analysis confirmed biological relevance and interpretability.

Conclusions:

• HieRMVir offers a significant advancement in viral genome classification, enhancing accuracy and interpretability.
• The hierarchical approach effectively leverages taxonomic structure and genomic feature informativeness.
• This method holds promise for improved pathogen detection and surveillance.