Speaker
Description
Influenza remains a persistent global health threat due to its high mutability, rapid transmission, and significant socioeconomic impact. This study addresses the critical research question: How can computational frameworks overcome current limitations to accurately and robustly predict human-influenza protein-protein interactions (PPIs)? Understanding these interactions is crucial for uncovering the mechanisms of viral replication and pathogenesis and for developing targeted antiviral strategies. However, existing methods rely heavily on high-resolution but scarce protein structures, and encounter difficulties in generating high-quality negative samples, which could lead to suboptimal predictive performance with limited generalizability capability, and restricted practical applications in infectious disease research. To address these challenges, we propose SEMVS, an innovative end-to-end framework for PPI prediction. SEMVS introduces an adaptive negative sampling strategy to enhance training by generating high-quality negative samples. We develop a double-view deep learning approach to extract global and local sequence features, enabling a more comprehensive understanding of protein interactions. Experimental evaluations demonstrate that SEMVS outperforms traditional classifiers, state-of-the-art deep learning models, and large language models (LLMs), achieving an accuracy of 0.987, sensitivity of 0.986, specificity of 0.988, and MCC of 0.974. We also demonstrate our designed adaptive negative sampling improves accuracy, F1-score, and AUROC by 17.1%, 26.4%, and 19.1%, respectively, compared to random sampling. The ablation studies highlight the importance of combining global and local feature extraction. Additionally, to test the generalizability of our proposed framework, we apply it with fine-tuning to PPI prediction across other human-virus pairs, which suggests robust and comparable performance. Further clustering analysis of predicted interactions uncovers conserved viral mechanisms targeting specific human proteins, validated by structural alignments based on AlphaFold3 and supporting literature. In conclusion, we provide a scalable and robust framework based on sampling-enhanced double-view learning for influenza-human PPI prediction, with broad implications for understanding viral pathogenesis and facilitating targeted antiviral therapies development.
| Expedited Notification | No thanks, I do not require Expedited Notification |
|---|
