Safe-by-redesign guidance for toxic industrial chemicals using explainable artificial intelligence: Introducing the DETOX-QSAR model

Abstract

Toxic industrial chemicals (TICs) are dual-use agents with potential for immediate and lethal inhalation toxicity upon exposure. This study first introduced the Descriptor-Engineered Transition to Non-Toxicity via QSAR (DETOX-QSAR) model, a tool for predicting and designing out the acute inhalation toxicity of TICs. We aimed to proactively prevent adverse effects by addressing safety concerns early in the molecular design phase. Inhalation toxicity was modeled using Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) via consensus feature engineering. SHapley Additive exPlanations (SHAP) findings indicate that nX and minwHBa play a decisive role in prediction. Arsine was identified as the highest-risk pilot compound via Multi-Criteria Decision Making (MCDM) for the modification process. Guided by the local SHAP method, descriptor values of Arsine were gradually modified until toxicity was eliminated. This explainable artificial intelligence (XAI) approach advances green toxicology, enhancing international security against the growing chemical industry.