The Utilization of Machine Learning Approaches in Predicting the Prominent Physicochemical Properties of Polychlorinated Biphenyl

Given the difficulties in experimental measurement, machine learning offers a viable alternative method to reduce and minimize costs and time. Machine learning enhances the effectiveness and efficiency of environmental pollutant monitoring, supporting better management and protection of ecosystems and public health. Random forest is indeed a significant machine learning method, widely used for various applications due to its robustness and versatility. Polychlorinated biphenyls (PCBs), a vital class of persistent organic pollutants, have garnered significant attention from the scientific community due to their detrimental impacts. In this study, a computational prediction model for octanol–water partition coefficient (logP) and bioconcentration factor (logBCF) of PCBs was developed by using random forest (RF), sparrow search algorithm random forest (SSA-RF), particle swarm optimization random forest (PSO-RF), and gray wolf optimization random forest (GWO-RF) methods. We performed a comprehensive validation, evaluation, and mechanistic explanation of the model for ensuring its reliability and applicability. Overall, the internal and external validation statistical parameters of the eight models have good robustness and predictive power. The Williams plots further show that all models are built in a wide range of application domains, and therefore they can be applied to unrecognized PCBs already in the environment to fill in the gaps in relevant experimental data. SSA-RF was superior to other methods, suggesting that it is more appropriate for computational studies of PCBs.