Development of Hybrid Machine Learning Model to Predict Intrinsic Clearance for Pesticides

Abstract

Predicting intrinsic clearance (Clint) is essential for understanding the pharmacokinetics of pesticides, as it directly influences dosing regimens and the overall behavior of chemicals within biological systems. This study aimed to develop and validate a hybrid machine learning model to accurately predict Clint for various pesticide categories, utilizing publicly available data from the U.S. Environmental Protection Agency's National Center for Computational Toxicology (EPA NCCT) High-Throughput Toxicokinetics (HTTK) dataset. Molecular descriptors were calculated using the PaDEL software, and relevant descriptors were selected using the k-nearest neighbors (kNN) algorithm based on their correlation with Clint values. Three machine learning models—Random Forest (RF), XGBoost, and Artificial Neural Networks (ANN)—were trained and evaluated across four pesticide categories: Total, Herbicides, Insecticides, and Fungicides. The Random Forest model achieved the highest R² value of 0.967 for Herbicides, while XGBoost outperformed the other models for Insecticides (R²=0.806) and Fungicides (R²=0.840), as well as the Total category (R²=0.744). Despite these promising results, the study faced limitations such as the treatment of outliers, the presence of excessive zeros in the dataset, and small sample sizes (e.g., n=20 for Herbicides), which could impact the accuracy of the models. The hybrid model, by selecting the optimal algorithm based on input chemical structures, demonstrates significant potential in predicting Clint for new chemicals, offering a rapid and reliable alternative to traditional in vivo methods. These findings contribute to the field of computational toxicology by enhancing the predictive capabilities of in silico models and supporting the development of safer chemical compounds.