Fundamental studies on developing the prediction model for photolytic half-lives of chemicals.

Abstract

It is significant to understand the physical properties and dynamics of chemicals in the environment. However, experimenting these properties for such huge amounts of compounds is almost impossible since it takes too much cost and time. Hence, we often use the QSPR (Quantitative Structure- Property Relationship) as an alternative approach. The purpose of current study is to construct a prediction model for the photolytic half-lives and meanwhile to estimate some physicochemical properties which are important to predict the photolytic half-lives. We used the Random Forest, a widely used machine learning algorithm. As a result, HOMO_LUMO gap, charge-related descriptors, and some other descriptors were selected to be significant. The accuracies of the model constructed with the ten most important descriptors were approximately 66% and 73% for training and test sets respectively.