이용빈 (서강대학교)

이용빈 (서강대학교)

오한빈 (서강대학교)

This study aims to develop a machine learning model that predicts the essential physicochemical properties of chemicals for regulatory models like K-simpleBox, in order to enable rapid and accurate risk assessment of hazardous substances in household chemical products. This contributes to establishing an evaluation framework for new substances lacking experimental data and enhancing the accuracy of existing regulatory models.

For the initial model development, this study selected Per- and Polyfluoroalkyl Substances (PFAS), which are significant due to their persistence in the environment, and built a model to predict the octanol-water partition coefficient (logKow/logD), a key indicator of environmental behavior. Training data was collected from the U.S. Environmental Protection Agency's (EPA) CompTox Chemicals Dashboard, and statistically reliable data were selected by comparing published experimental values (LOGD7.4) with predictions from an existing model (OPERA). Notably, a Weighted Regression method was applied to incorporate differences in data reliability into the model's learning process. The model used the XGBoost (eXtreme Gradient Boosting) algorithm, and optimal hyperparameters were determined through a Grid Search.

The developed model can be directly used to generate input data for regulatory models by rapidly and accurately predicting the properties of new substances that lack experimental data, enabling more efficient and scientific chemical management. Future plans include developing a prediction model for the octanol-air partition coefficient (logKoa) and expanding the research scope to other major substance groups, such as siloxanes and synthetic musks.