Machine learning-based binary classification of insecticides for resistance risk modeling

The extensive use of insecticides has indeed led to an increasingly serious problem of pest resistance. However, indoor resistance experiments, despite providing the resistance multiplicity of insecticides, suffering from long experimental periods and difficulty in obtaining test materials. Nevertheless, the use of machine learning models to assess the potential resistance risk of insecticides quickly and reasonably is a promising approach that warrants further investigation. The study utilized the Arthropod Pesticide Resistance Database (APRD), British Crop Production Council (BCPC), and SPECS databases to carefully select samples with low structural similarity to form a training set. Six machine learning algorithms LDA (linear discriminant analysis), SVM (support vector machine), ANN (artificial neural network), DT (decision tree), RF (random forest), SOM (self-organizing map) were used to construct binary classification resistance risk models for insecticides. The prediction model's parameters were meticulously optimized based on the test set and rigorously verified with the optimal model against the external validation set. Among the single models, DT had a prediction accuracy of 84.62% in the external validation set. By utilizing a voting mechanism to combine the evaluation effects of the six models, we achieved a prediction accuracy of 78.95% for positive samples and 65% for negative samples. The model offers a confident evaluation of the potential resistance risk of new insecticides and guides the scientific use of insecticides to delay the development of resistance.