Machine learning-based prediction of toxicity of pesticide towards Americamysis bahia

Pesticides are toxic substances designed and widely applied throughout the world. However, their widespread use has received increasing attention from regulatory agencies due to the various acute and chronic effects they have on various organisms. In this study, QSTR (Quantitative Structure-Toxicity Relationship) models, based on nonlinear statistical techniques, have been established using five Machine Learning (ML) algorithms to predict the toxicity of pesticides on mysid shrimp (Americamysis bahia). The optimal nonlinear model (Random Forest, R² = 0.983) was verified by internal (leave one cross-validation) and external validations. The validation results (q_int² = 0.815 and q_ext² = 0.81) were satisfactory in predicting acute toxicity in the saltwater crustacean (A. bahia) compared to other models reported in the literature. In addition, this model also predicted the toxicity of some pesticides without experimental data. With a p(LC50) value of 12.102, bromadiolone was the most toxic compound. It is the active compound of the product RASTOP BLOCKS (rodenticide), which is classified in the category "Extremely Hazardous".