EXPLAINABLE ARTIFICIAL NEURAL NETWORKS FOR TABULAR DATA: IDENTIFYING STRUCTURAL, ELECTRONIC AND PHYSICOCHEMICAL DETERMINANTS OF PHARMACEUTICAL COMPOUND-INDUCED RESPIRATORY TOXICITY
GSC Biological and Pharmaceutical Sciences · 01.11.2026 · 37 (2) · 001
DOI: 10.30574/gscbps.2026.37.2.0372 ↗Kurz gefasst
Fachartikel in der Zeitschrift GSC Biological and Pharmaceutical Sciences, online erschienen am 01.11.2026. Verfasst von Désiré Mélèdge und 3 weiteren Autor:innen. Die Arbeit nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Respiratory toxicity induced by pharmaceutical compounds is a major concern in drug discovery and development, emphasizing the need for reliable computational approaches to identify potentially hazardous compounds at an early stage. In this study, an artificial neural network (ANN) integrated with eXplainable Artificial Intelligence (XAI) was developed for the binary classification of pharmaceutical compounds as respiratory toxic or non-toxic using a tabular dataset of molecular descriptors. Model performance was evaluated through ten-fold cross-validation and external validation using accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), and area under the receiver operating characteristic curve (AUC-ROC). During internal validation, the ANN achieved an accuracy of 84.05% and an AUC of 91.36%. On the external test set, the model yielded an accuracy of 83.40%, precision of 85.58%, recall of 85.19%, F1-score of 85.38%, MCC of 0.66, and AUC-ROC of 89.53%, demonstrating robust predictive performance and satisfactory generalization to unseen compounds. SHapley Additive exPlanations (SHAP) analysis was further employed to interpret the ANN predictions and identify the molecular characteristics most strongly associated with model output. The most influential descriptors included nBase, BCUTp-1h, SssO, minssCH2, JGI8, maxHBd, nHeteroRing, GATS2m, ETA_BetaP_ns_d, GATS1e. Collectively, these descriptors encode complementary information related to basicity, heterocyclic structure, local electronic environments, hydrogen-bonding characteristics, atomic polarizability, charge distribution, atomic mass, electronegativity, and molecular topology. The SHAP findings suggest that respiratory toxicity predictions are governed by the combined contribution of multiple structural and physicochemical characteristics rather than by a single molecular property.
Zitieren
APA
Mélèdge, D., N’guessan, A., Kouakou, K. J. L., & Zamblé, D. (2026). EXPLAINABLE ARTIFICIAL NEURAL NETWORKS FOR TABULAR DATA: IDENTIFYING STRUCTURAL, ELECTRONIC AND PHYSICOCHEMICAL DETERMINANTS OF PHARMACEUTICAL COMPOUND-INDUCED RESPIRATORY TOXICITY. GSC Biological and Pharmaceutical Sciences, 37(2), 001. https://doi.org/10.30574/gscbps.2026.37.2.0372
Vancouver
Mélèdge D, N’guessan A, Kouakou KJL, Zamblé D. EXPLAINABLE ARTIFICIAL NEURAL NETWORKS FOR TABULAR DATA: IDENTIFYING STRUCTURAL, ELECTRONIC AND PHYSICOCHEMICAL DETERMINANTS OF PHARMACEUTICAL COMPOUND-INDUCED RESPIRATORY TOXICITY. GSC Biological and Pharmaceutical Sciences. 2026;37(2):001. doi:10.30574/gscbps.2026.37.2.0372
BibTeX
@article{Mldge2026,
author = {Mélèdge, Désiré and N’guessan, Aubin and Kouakou, Kouakou Jean Louis and Zamblé, Donatien},
title = {EXPLAINABLE ARTIFICIAL NEURAL NETWORKS FOR TABULAR DATA: IDENTIFYING STRUCTURAL, ELECTRONIC AND PHYSICOCHEMICAL DETERMINANTS OF PHARMACEUTICAL COMPOUND-INDUCED RESPIRATORY TOXICITY},
journal = {GSC Biological and Pharmaceutical Sciences},
year = {2026},
volume = {37},
number = {2},
pages = {001},
doi = {10.30574/gscbps.2026.37.2.0372}
}
EUVIMED