Interpretable machine learning for predicting key hazardous properties of chemicals

Abstract

Accurate prediction of hazardous chemical properties such as toxicity, flammability, reactivity, and reactivity with water (RW) is essential for the safe handling, storage, and transport of chemicals in industrial and environmental contexts. Traditional experimental methods are time-consuming, costly, and struggle to capture the complex, dynamic relationships between molecular structure and hazardous properties. Moreover, many conventional models lack interpretability, making it difficult to understand the underlying molecular interactions driving these properties. In this study, we developed 8 machine learning models to predict these four hazardous properties based on molecular descriptors. The models were trained on self-curated datasets, incorporating advanced techniques for feature selection and interpretability. The most optimal model was further applied to predict the hazardous properties of the hazardous chemicals list. Results show that XGBoost achieved superior performance in predicting toxicity (0.768) and reactivity (0.917), while RF excelled in flammability (0.952) and RW (0.852) in terms of ROC-AUC. For Ketone/Aldehyde compounds, SHAP and ICE analyses identified key molecular descriptors such as MIC4, ATSC2i, ATS4i and ETA_dEpsilon_C as critical determinants for toxicity, flammability, reactivity, and RW respectively. Notably, 100% of the hazardous chemicals list were predicted to be flammable, 99.5% toxic, 66.4% reactive, and only 0.4% exhibited RW. The results demonstrate the potential of machine learning models to provide efficient and scalable predictions, reducing the need for costly experimental testing while improving safety protocols for hazardous chemical management.