Development of Pesticide-Likeness Scores and Models for Predicting Pesticide Activity of Molecular Scaffolds with Machine Learning.

To develop pesticide molecules efficiently, machine learning models to predict the activity of pesticides on specific species have been proposed; however, the construction of activity target models for individual species is required because of the wide range of species that are active targets of pesticides, which is insufficient. In this study, we propose pesticide-likeness scores for chemical structures in a more comprehensive manner without relying on specific active targets. Combined data sets that are independent of the species of the active target were prepared; classification models that predict positive or negative activity from chemical structures are prepared, and the pesticide-likeness scores are calculated as the probability of positiveness with the constructed models. The pesticide-likeness scores were used as molecular descriptors in activity prediction models to achieve the same level of prediction accuracy as general molecular descriptors. In addition, the activity of pesticides is greatly affected by molecular scaffolds, and it is important in pesticide molecular design to remove scaffolds that provide inactive molecules no matter what substructures are added and to propose promising scaffolds that provide active molecules. We constructed machine learning models between the activity of scaffolds and chemical structures of scaffolds and proposed promising scaffolds by inputting chemical structures of scaffolds with unknown activity into the constructed models and selecting the scaffolds with high predicted activity. Furthermore, we obtained new knowledge by visualizing the contribution of fingerprints in activity prediction on the proposed scaffolds.