Descriptor-Based Explainable QSAR Modeling Approaches for the Prediction of Respiratory Toxicity of Volatile and Gas-Phase Chemicals.

Abstract

Toxic gases pose significant risks to human health and ecosystems because of their capacity to deplete oxygen levels and disrupt critical physiological functions. Effective management of these hazards is crucial, given their widespread use in various industrial sectors, including agriculture, metallurgy, pharmaceuticals, and plastics. A dataset of 229 gaseous-phase chemicals from authoritative toxicological sources was used to develop robust binary quantitative structure-activity relationship (QSAR) models for predicting toxicity. Employing molecular descriptor selection, correlation analyses, and outlier detection, random forest and XGBoost emerged as the most accurate models, each achieving a high accuracy of 0.915, with ROC areas of 0.992 and 0.986, respectively. SHapley Additive exPlanations (SHAP) were applied to enhance the explainability of the best model, providing global and local insights into the importance and direction of each descriptor. Key features influencing toxicity were identified globally, while locally, descriptors were ranked on a molecule-by-molecule basis, offering guidance for mitigating toxic effects through structural modifications. Although numerous studies have focused on predicting the toxicity of chemical gases, few have systematically analyzed the specific molecular descriptors driving toxicity. This study addresses this gap by providing a descriptor-based analysis, offering deeper mechanistic insights, and enabling more accurate predictions. The QSAR approach aligns with the 4R principles, promoting ethical, cost-effective, and efficient alternatives to animal testing, thus contributing to safer chemical design and environmental monitoring.