Contributions to the development of prediction models for the toxicity of ionic liquids

Abstract

Ionic liquids (ILs) are a class of compounds with unique properties that make them highly valuable in various industrial and chemical processes, but their toxicity poses significant challenges for widespread use. This study investigates the prediction of the toxicity of ILs through quantitative structure–toxicity relationship (QSTR) modeling using a support vector machine (SVM) model enhanced with various optimization algorithms. A dataset comprising 304 ILs with toxicity measured in the leukemia rat cell line (IPC-81) and an additional 14 external validation points was employed. The model uses 13 molecular descriptors. Three optimization algorithms were constructed and evaluated: dragonfly algorithm (DA), moth–flame optimization (MFO), and gray wolf optimizer (GWO). Among them, the DA-optimized SVM model demonstrated superior predictive performance with a correlation coefficient (R) of 0.9871, a coefficient of determination (R2) of 0.9742, a root mean square error (RMSE) of 0.1787, and a mean squared error (MSE) of 0.0625. Additionally, the arithmetic residuals in K-groups analysis (ARKA) method was applied to reduce the dimensionality of the dataset and identify activity cliffs, areas where small changes in molecular structure result in significant shifts in toxicity. However, the DA-SVM model using the original 13 descriptors provided superior predictive accuracy compared to the ARKA-based model. The high predictive accuracy of the DA-optimized SVM model underscores its potential as a robust tool for QSTR modeling and for assessing the toxicity of ionic liquids.