Machine learning for predicting halogen radical reactivity toward aqueous organic chemicals

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Machine Learning Applications

Background:

• Organic pollutants pose environmental risks, and understanding their reactivity is crucial for risk assessment.
• Predicting the reactivity of organic pollutants, particularly their reaction rate constants (logk) with free radicals, is essential for environmental remediation and management.
• Traditional methods for determining pollutant reactivity are often time-consuming and resource-intensive.

Purpose of the Study:

• To develop accurate and efficient machine learning (ML) models for predicting the rate constants (logk) of free radical-mediated organic pollutant reactivity.
• To investigate the effectiveness of different molecular descriptors, specifically Morgan fingerprint (MF) and Mordred descriptors (MD), in combination with various ML algorithms.
• To address the challenge of limited sample size by implementing a data combination strategy to enhance prediction accuracy and mitigate overfitting.

Main Methods:

• Employed Morgan fingerprint (MF) and Mordred descriptors (MD) to represent molecular structures.
• Utilized a series of machine learning models, including Light Gradient Boosting Machine (LightGBM) and Random Forest (RF).
• Implemented a data combination strategy to create a unified dataset, improving model performance and generalizability.
• Applied SHapley Additive exPlanations (SHAP) for model interpretability and identified key molecular features influencing reactivity.
• Conducted applicability domain analysis to ensure the reliability of predictions for new compounds.

Main Results:

• The LightGBM model with MF and the RF model with MD, trained on the unified dataset, were identified as the optimal predictive models.
• SHAP analysis revealed that the MF-LightGBM model effectively captured the impact of electron-withdrawing/donating groups, while the MD-RF model highlighted the importance of autocorrelation, walk count, and information content descriptors.
• The study emphasized the significant contribution of pH to pollutant reactivity predictions.
• Applicability domain analysis confirmed the broad and reliable predictive capability of the developed models for diverse chemical structures.

Conclusions:

• Machine learning, particularly when integrating effective descriptors and data strategies, offers a powerful approach for accurately predicting organic pollutant reactivity.
• The developed models provide a robust and efficient tool for environmental risk assessment and the design of remediation strategies.
• A practical web application was successfully developed for calculating logk, facilitating wider accessibility and application of the research findings.