OpenTox Virtual Conference 2026

From Biodegradation to Hydrolysis: An Integrated Machine Learning Strategy for Environmental Fate Prediction of Organic Chemicals

The environmental fate of organic chemicals is governed by a complex interplay of transformation processes that determine their persistence, exposure, bioaccumulation potential, mobility, and subsequent ecological risks. Among various processes, biodegradation and hydrolysis represent two key mechanisms that control the transformation and environmental persistence of chemicals in both aquatic and terrestrial ecosystems. However, experimentally assessing these endpoints is resource-intensive and is not always feasible for the growing number of chemicals entering the environment. Here, we have proposed a framework that combines quantitative structure–property relationship (QSPR), read-across, read-across structure–property relationship (RASPR), and machine learning approaches to determine environmental fate endpoints. We collected experimental values for chemical biodegradability and hydrolysis from https://www.echemportal.org/echemportal/. For biodegradation assessment, a curated set of 286 and 79 compounds was collected according to OECD Test Guidelines 302 and 307, and the compounds were used to classify chemicals according to their inherent aerobic biodegradability in aquatic systems and their biotransformation potential in soil, respectively.[1] For hydrolytic stability assessment, 351 compounds according to OECD Test Guideline 111 were utilized to characterize the hydrolytic behaviour of chemicals at pH 4, 7, and 9, respectively.[2] Different data balancing approaches were utilized to balance the number of active and inactive chemicals in the dataset. For the biodegradation datasets, the support vector classifier (SVC) algorithm showed the highest performance, with a test set accuracy and Matthews correlation coefficient (MCC) of 0.77 and 0.49 for inherent aerobic biodegradability in aquatic systems and 0.84 and 0.64 for the biotransformation potential in soil, while for the hydrolysis dataset, the categorical boosting (CatBoost) algorithm showed a test set accuracy and MCC of 0.82 and 0.65, respectively. SHapley Additive exPlanations (SHAP) analysis was used to explore the individual feature contribution and to provide mechanistic insights into the modeled descriptors. Finally, the developed models were disseminated through user-friendly Python based predictive platforms (BiodegPred and HydroFate) to facilitate rapid screening of chemicals from their molecular structures without requiring extensive experimental testing. These computational tools can support chemical prioritization, environmental risk assessment, regulatory decision-making, and the development of safer and more environmentally sustainable chemicals. Overall, the integrated modeling framework demonstrates the potential of machine learning, QSPR, and read-across methodologies to provide a practical computational framework for assessing complementary environmental transformation processes and advancing predictive environmental fate assessment.

Keywords: OECD TG; environmental fate; predictive tool; classification modeling

1. Pandey, S.K., Pore, S. and Roy, K. (2025) Science of The Total Environment, 1000, 180376.

2. Pandey, S.K., Pore, S. and Roy, K. (2026) Science of The Total Environment, 1013, 181306.