Human activities have introduced tens of thousands of chemicals into water systems around the world which has significantly impacted water quality and aquatic ecosystems. The aim of this study was to develop an in silico QSAR model, capable of predicting the aquatic toxicity of pesticides in terms of a lethal dose (LD50) for fish without requiring the use of in vivo testing. A large data set of 230 diverse pesticides, including fungicides, herbicides and insecticides, with experimentally measured LD50 values was used to develop a predictive QSAR model. Each pesticide molecule was described using 62 calculated molecular descriptors. These descriptors were then related to the LD50 values via an Artificial Neural Network. Sensitivity analysis was used to select descriptors that best describe the model. The developed model included 13 molecular descriptors related to lipophilicity, hydrogen binding and polarity. Note the value of the predictive squared correlation coefficient (q2) for the final model was 0.748, demonstrating the model’s predictability. In the domain of QSAR studies, a q2 value above 0.5 renders a model to be predictive. The model could therefore be used to accurately screen a wide range of compounds without the need for actual compound synthesis and to prioritize potentially toxic compounds for further testing.
Keywords: ANNs, aquatic toxicity, LD50, pesticides, QSARs.