From deep learning-based toxicity prediction to theoretical simulation-based mechanism analysis: A case study of 33 typical pesticides and their transformation products.

Pesticides and their transformation products (TPs) have long been recognized as classical environmental pollutants in agricultural systems, posing receptor-responsive toxicity to fish. Pesticide TPs are numerous and have not been uniformly collected, and their receptor-responsive toxicity remains largely unidentified. Herein, 33 typical pesticides and their 393 TPs were used as examples. A total of 26 typical pesticides and 175 TPs were preidentified as having potential receptor-responsive toxicity using similarity and residual toxicity methods. A DToxNet prediction model (toxicity prediction model based on deep focusing network, DToxNet) was constructed to predict reproductive toxicity, developmental toxicity, and endocrine-disrupting toxicity, with accuracies exceeding 85%. The predicted results provided a comprehensive picture of the three types of receptor-responsive toxicity of typical pesticides, including profenofos, imidacloprid, and fenitrothion, and their TPs. For the predicted results, molecular dynamics simulation was used to reveal interaction modes and interaction levels. The results elucidated the toxicity mechanisms driven by protein conformational changes, hydrophobic and hydrogen-bonding forces, and the classification of binding energy strengths. Following an extensive review of the receptor-responsive toxicity preidentified results, DToxNet model prediction results, and molecular dynamics simulation findings, a prioritized control list of 30 typical pesticides with receptor-responsive toxicity in their parent forms or TPs was developed. The list included 9, 9, and 12 pesticides with high, medium, and low toxicity, respectively. The DToxNet developed in this study can efficiently identify the receptor-responsive toxicity of typical pesticides and their TPs, and the identification and integration of receptor-responsive toxicity data can support the priority management of pesticides in agriculture.