(362u) High-Throughput Screening for Identifying Potential Chemical Exposure Scenarios at End-of-Life Stage

Environmental regulatory entities like the U.S. Environmental Protection Agency and inter-government bodies like the Organization for Economic Cooperation and Development have tried to deal with the presence of chemicals of concern used in industrial processes or incorporated as part of consumer and commercial products. The correct identification of the pathway taken by hazardous chemicals is crucial to understanding the events associated with exposure scenarios. Nonetheless, the main challenge nowadays is the time-consuming task of collecting comprehensive data to describe a chemical exposure scenario, which is especially challenging considering the ever-increasing inventory of chemicals circulating in the world marketplace. The above task is even more demanding at the end-of-life (EoL) stage, where chemical flow traceability is more difficult and there is a higher level of epistemic uncertainty about where a chemical can go. The increase in computer power and the upward trend in cloud computer utilization make machine learning an alternative for resolving complex problems. Based on machine learning and multi-criteria decision-making techniques, this work proposes a workflow for the development of data-driven models called Quantitative Structure-Transfer Relation (QSTR) models to predict the probabilities of EoL activities for chemicals. The QSTR models are inspired by the operation of the well-known QSAR models that are widely used in chemoinformatics and toxicology. The QSTR models do not only use molecular descriptors but also features associated with chemical unit price, generator industry sector, amount of chemical transferred, environmental policy stringency, and gross value added by sector, enabling them to incorporate aspects that can affect the context for implementing and EoL activity for a chemical. A total of 10 QSTR models are selected, built using random forests, and tuned by Bayesian optimization with the Gaussian process. The model's external validation on the test dataset provides accuracy ranging from 67 to 97% and test accuracy (f1 score) between 69 and 97%. These QSTR models can be deployed and incorporated into a model for understanding the EoL management chain behavior, obtaining chemical flow analysis, and identifying potential EoL activities and exposure scenarios for chemicals of concern. Thus, QSTRs allow environmental regulatory decision-making and identify whether a chemical product satisfies safety compliance before going into the world marketplace, considering life cycle thinking.