Multi-View Transformers for Structure-Aware HA-NA Drift Risk Scoring and Mutation Hotspot Mapping.

Area of Science:

• Virology
• Computational Biology
• Genomics

Background:

• Influenza A virus rapidly evolves via mutations in haemagglutinin (HA) and neuraminidase (NA), impacting vaccine efficacy.
• Existing sequence-only models lack integration of 3D protein structure and long-term evolutionary context for mutation risk assessment.

Purpose of the Study:

• To develop a structure-aware computational framework, TRIAD-Influenza, for prioritizing emerging influenza A HA and NA variants.
• To link codon-level mutations to 3D protein structure and evolutionary dynamics for enhanced risk prediction.

Main Methods:

• A multi-view transformer architecture (TRIAD) integrating codon/residue sequences, predicted HA/NA 3D structures, and phylogenetic context.
• Curated over 3x10^5 HA/NA sequences (2010-2024) with codon-aware alignment and predicted 3D protein structures.
• Developed mutation hotspot mapping using gradient saliency and a contact-weighted mutation risk index (CMRI).

Main Results:

• TRIAD-Influenza achieved strong predictive performance (AUROC ≈0.89) on internal validation and maintained discrimination (AUROC ≈0.85-0.86) on external cohorts.
• Predicted risk scores correlated with experimental antigenic distances (Spearman ρ ≈0.82).
• CMRI hotspots identified known epitopes and escape residues, validating the model's biological relevance.

Conclusions:

• TRIAD-Influenza enables rapid, structure-aware prioritization of influenza A HA/NA sequences.
• The model provides interpretable mutation hotspot maps for targeted experimental validation and surveillance.
• Integrating sequence, structure, and phylogeny enhances the prediction of viral evolution and potential pandemic threats.