Toxicity prediction of insecticides and pesticides via machine learning approach

Abstract

Pesticides are commonly used to protect crops, but their potential toxicity poses significant environmental and health risks. This study explores the effectiveness of seven machine learning (ML) models—Random Forest (RF), Extreme Gradient Boosting (XGB), Gradient Boosted Decision Tree (GBDT), Categorical Boosting (Catboost), Light Gradient-Boosting Machine (LGBM), stacked models (RF + XGB and RF + LGBM)—to predict key toxicity factors for pesticides. The models were designed to estimate the Bio-Concentration Factor (BCF), the n-octanol-water Partition Coefficient (Kow), and the Lethal Dose-50 (LD₅₀), using a dataset of 244 pesticides with over 160 features such as molecular weight, temperature, solubility, number of rings, and partition coefficient. A splitting of the dataset into 90 % training and 10 % testing sets. The RF + LGBM stacked model achieved the best performance for BCF prediction, with a coefficient of determination (R²) of 0.89 and a Mean Absolute Percentage Error (MAPE) of 12.72 %. Catboost excelled in predicting Kow with an R² of 0.88, a Mean square error (MSE) of 0.364, and an MAPE of 22.38 %. For LD50, the RF + XGB stacked model was the most accurate, with an R² of 0.75 and a MAPE of 8.5 %. Shapley Additive explanations (SHAP) analysis revealed that log P, water solubility, and SLogP were the most influential features across all models. This study demonstrates the power of machine learning for toxicity prediction while also setting the stage for future research in predictive toxicology, environmental monitoring, and sustainable pesticide regulation, ultimately contributing to more responsible and data-driven agricultural practices.