A QSTR framework for terrestrial oral acute toxicity prediction: A machine learning approach motivated by the search for sustainable chemical alternatives
Ziad Merghad, Anaële Lefeuvre, Johan Merzouki, Serge Maison · Next Sustainability · 2025
This work presents a Quantitative Structure-Toxicity Relationship (QSTR) framework to predict acute oral toxicity for a diverse range of organic compounds. Motivated by the urgent need to screen safe industrial chemicals, such as replacements for ozone-depleting Halon 1301, this study leverages a large dataset filtered from the Registry of Toxic Effects of Chemical Substances (RTECS) database. Binary, multi-class, and regression models were developed and compared using XGBoost and Random Forest algorithms. To address data imbalance in the binary classification ('toxic' vs. 'non-toxic'), techniques including Tomek Links, SMOTE, and Cost-Sensitive Learning were evaluated. The Cost-Sensitive XGBoost model coupled to a cost-sensitive data rebalancing method was identified as the most balanced and suitable for safety screening, achieving the highest precision for the non-toxic class (0.93) and recall for the toxic class (0.67). A rigorously defined Applicability Domain using PCA and UMAP ensures reliable predictions for new chemicals. While developed for oral toxicity, this model serves as a valuable first-pass screening tool to flag molecules with low inherent toxicity, providing crucial preliminary insights in contexts where route-specific data, such as for inhalation, is scarce. The framework presented offers a step towards accelerating the design and selection of safer, more sustainable chemicals.