Toxicity prediction of pesticide to bumblebee and honey bee based on machine learning methods

Bumblebee (Bombus spp.) and honey bee (Apis mellifera L.) are important pollinating insects in nature. In recent years, large-scale unreasonable use of pesticides has caused the continuous decline of bumblebee and honey bee populations in many regions of the world. In order to evaluate the toxicity of pesticides to bumblebee and honey bee in a better way, the common toxicity data of 61 pesticides were collected. The classification models for the acute contact toxicity LD₅₀ of bumblebee and honey bee were established using 12 molecular fingerprints combined with 8 machine learning algorithms, respectively. The results showed that the predictive accuracy of the classification model for the acute contact toxicity of bumblebee and honey bee was 86.7% and 80.0%, respectively. Three algorithms (Random Forest, Neural Network, and Support Vector Machine) and three fingerprints (Fingerprinter, Klekota-Roth Count, and Extend) had a better predictive ability in our study. In addition, the prediction models of bumblebee and honey bee were used to carry out cross-toxicity predictions with the predictive accuracy of 72.9% and 66.7%, respectively. The toxicity prediction model of bumblebee is more accurate in predicting honey bee toxicity than the toxicity prediction model of honey bee for the prediction of bumblebee toxicity. Our study has provided theoretical guidance to design low-toxicity compounds and predict the cross-toxicity for the different insect targets.