Contributions to the Development of Prediction Models for the Toxicity of Ionic Liquids

Abstract

Ionic Liquids (ILs) are a new class of innovative green solvents of significant importance from academic, industrial, and environmental perspectives due to their attractive properties. For designing any process involving ILs, understanding their toxicity is crucial. This work develops robust Quantitative Structure-Toxicity Relationship (QSTR) models to predict the toxicity of ILs, offering efficient in silico alternatives to traditional experimental assessments. Initially, this study developed Support Vector Machine (SVM) models optimized using three meta-heuristic algorithms: Dragonfly Algorithm (DA), Moth-Flame Optimization (MFO), and Grey Wolf Optimizer (GWO). These models utilized molecular features selected by Genetic Algorithm-Multiple Linear Regression (GA-MLR) and Boruta descriptor selection techniques. Among these, the DA-SVM model consistently demonstrated superior predictive accuracy and robustness. For instance, the DA-SVM model with GA-MLR descriptors achieved a global RMSE of 0.1787 and an R² of 0.9742 Subsequently, a distinct approach explored the employment of novel ARKA (Autoencoded Reduced Kernel-based Architecture) descriptors combined with advanced ensemble learning methods, XGBoost and Random Forest. Models built with these ARKA descriptors consistently outperformed their GA-MLR counterparts, showing significantly enhanced predictive power and generalization. For example, the RF-ARKA model achieved a global R2 of 0.9421 and R of 0.9707. Overall, the developed models offer reliable, interpretable, and efficient tools for predicting IL toxicity, supporting safer chemical design and contributing to sustainable chemistry.