Developing Machine Learning Based Models for Prediction of Pesticide Properties using Molecular Descriptors

This study assessed the performance of six machine learning (ML) models which were developed to predict various properties of 888 pesticides based on molecular descriptors. The log-transformed soil organic carbon partition coefficient values obtained from the Pesticide Properties Database (PPDB) website and 690 molecular descriptors in 2D estimated from the molecular descriptor calculator Mordred were used as dependent and independent variables in the models, respectively. The constructed data set was split into training and test sets at an 8:2 ratio based on stratified sampling. Important variables which were responsible for the success of adopted ML models were selected based on a combination of filter and embedded methods. Also, hyperparameter optimization for each ML algorithm was done by random search. Results showed that the top-ranked variables such as SLogP, ZMIC1, and FilterltLogS determined from the variable importance significantly contributed to the improvement in model performance although their contribution highly varied from algorithm to algorithm. In addition, the random forest (RF) model achieved the highest prediction accuracy out of them in terms of all evaluation metrics such as R2. Identical results for model performance were also observed with hyperparameter tuning. The developed RF model exhibited relatively low error rates reaching around 1% against 5 target pesticides, as compared to those of the KOCWIN model in EPI Suite ranging roughly from 5% to 27%. We believe that a proposed methodology for the development of predictive models based on molecular descriptors helps end-users not only improve the accuracy of existing models but also refine their algorithms.