Explainable AI-driven prediction of influenza neuraminidase inhibitors using a stacked ensemble-learning framework
Issued Date
2025-12-01
Resource Type
ISSN
00104825
eISSN
18790534
Scopus ID
2-s2.0-105022780590
Pubmed ID
41274115
Journal Title
Computers in Biology and Medicine
Volume
199
Rights Holder(s)
SCOPUS
Bibliographic Citation
Computers in Biology and Medicine Vol.199 (2025)
Suggested Citation
Meewan I., Schaduangrat N., Mookdarsanit L., Mookdarsanit P., Shoombuatong W. Explainable AI-driven prediction of influenza neuraminidase inhibitors using a stacked ensemble-learning framework. Computers in Biology and Medicine Vol.199 (2025). doi:10.1016/j.compbiomed.2025.111313 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/113357
Title
Explainable AI-driven prediction of influenza neuraminidase inhibitors using a stacked ensemble-learning framework
Corresponding Author(s)
Other Contributor(s)
Abstract
Influenza is a widespread respiratory infection that poses a persistent global health threat. While neuraminidase inhibitors (NAIs) are key antiviral agents, their effectiveness is increasingly challenged by emerging resistant strains and viral variants. This highlights the urgent need for novel NAIs with improved efficacy. However, conventional drug discovery methods are time-consuming, costly, and not well-suited for high-throughput screening. Consequently, computational approaches offer a promising alternative for accelerating the discovery of next-generation NAIs in a more efficient and cost-effective manner. To address these issues, we propose a highly precise and interpretable computational methodology, called XAI-NAI, that can be utilized to perform rapid and accurate prediction of NAI activity. In XAI-NAI, we employed 21 molecular descriptors and embeddings to extract sufficient information from NAIs. Subsequently, each molecular representation was input into six powerful machine learning (ML) methods for constructing 126 base-regressors. Next, all base-regressors were applied to generate 126-D new feature vectors containing multi-aspect information. Finally, a two-step feature selection method was used to optimize the 126-D new feature vectors, and the optimal subset was fed into the final meta-regressor using a stacking strategy. Experimental results on the independent test showed that XAI-NAI was effective and attained superior predictive performance than several conventional ML-based prediction models and related published prediction models for NAI activity prediction, achieving an R<sup>2</sup> of 0.750, an RMSE of 0.831, and a MAE of 0.576. In addition, XAI-NAI, along with molecular docking and molecular dynamics simulations, was applied to perform in silico drug repurposing by identifying potential NAIs among FDA-approved drugs. We anticipate that XAI-NAI will serve as a useful computational tool for supporting the community-wide effort to provide new insights for accurately screening and predicting promising NAIs.