HieRMVir: Interpretable Viral Classification via Hierarchical Deep Learning.

Abstract

Accurate identification of pathogens-especially those with pandemic potential-remains a significant challenge, particularly when traditional sequence alignment methods fail. While recent genome sequence identification methods have shown promise, most do not account for the hierarchical structure of biological taxonomy or the varying informativeness of genomic features across classification levels. To address these limitations, we propose HieRMVir (Hierarchical Random forest and Mutual information-based Viral genome classifier), a novel hierarchical deep learning framework that integrates random forest (RF)-based feature weighting with mutual information (MI)-guided attention regularization for interpretable and accurate viral sequence classification. HieRMVir performs classification across three levels and leverages feature importance scores from RF to scale input features, while MI scores are used to guide the attention mechanism towards statistically informative k-mer patterns through regularized loss. Experimental results on over one million genome sequences demonstrate that HieRMVir achieves an average accuracy of 95.8% (95% CI: 95.3-96.4%), outperforming existing methods on multiple metrics. Evaluation using hierarchical performance metrics and the analysis of learned attention weights further highlight the biological relevance and interpretability of HieRMVir.