Insight into the Mechanism of Machine Learning Models for Predicting Ionic Liquids Toxicity Based on Molecular Structure Descriptors

Abstract

The development and application of functionalized ionic liquids (ILs) are currently hot topics in chemical engineering. However, research on ILs toxicity is significantly lagging behind studies on their physical properties and applications. This study begins with the construction of ILs toxicity model, utilizing three types of descriptors to quantify ILs structures and developing four machine learning (ML) models for predicting toxicity to Daphnia magna. Guttmann coefficients are used to evaluate the diversity of ILs structures. Feature engineering is employed to optimize the inputs to the quantitative structure–activity relationship (QSAR) models, enhancing their ability to capture the relationship between ILs structures and toxicity. Grid search and cross-validation ensure model robustness and prevent overfitting. Results indicate that the random forest model based on RDKit descriptors performs best (R² = 0.975, RMSE = 0.222). SHAP analysis identifies key molecular features contributing to ILs toxicity, revealing that substructures around carbon atoms are crucial for ILs toxicity, while structures containing oxygen atoms can reduce toxicity. These findings offer insights for designing low-toxicity, environmentally friendly ILs and highlight the value of machine learning models in green chemistry and sustainability research.