HVIface: sequence-based deep learning for decoding human-virus protein-protein interfaces.

Introduction: Human-virus protein-protein interactions (PPIs) govern critical steps in viral entry, replication, and immune modulation. Despite their importance, the physicochemical determinants underlying human-virus interface recognition remain incompletely understood. A detailed characterization of these determinants is essential for uncovering viral hijacking mechanisms and enabling therapeutic targeting. To address this, we developed a sequence-based deep learning framework, HVIface, for residue-level prediction of human-virus PPI interfaces. Methods: HVIface was trained on a curated dataset of 73 structurally resolved human-virus complexes obtained from the Protein Data Bank. The model integrates 18 sequence-derived features, including residue contact potentials, physicochemical compatibility measures, and co-evolutionary signals. Performance was evaluated on an independent test set and compared against existing methods such as PIPENN and PeSTo. Additionally, feature importance analysis was conducted to identify key determinants contributing to interface prediction. Results: HVIface achieved an accuracy of 0.85 on the independent test set, outperforming existing approaches in identifying human-virus interface residues. Feature importance analysis highlighted environmentally weighted charge compatibility and co-evolution-derived contact potentials as dominant contributors. Case studies involving HIV-1 Integrase-Nup62, Influenza PB1-Nup85, and HCV NS5A-LASP1 complexes demonstrated that viral proteins preferentially target solvent-accessible, electrostatically favorable host regions through compact hotspot-driven architectures. Discussion: The findings reveal that electrostatic complementarity and localized residue clustering are central to viral-host interface formation, suggesting conserved physicochemical principles underlying viral targeting strategies. HVIface not only provides a robust predictive framework but also offers mechanistic insights into human-virus interface chemistry. These insights can facilitate systematic exploration of viral interactions and support the development of targeted therapeutic interventions. The source code is available at: https://github.com/krishnakantgupta-ai/HVIface_MMAD.