Explainable machine learning models enhance prediction of PFAS bioactivity using quantitative molecular surface...

Area of Science:

• Computational toxicology
• Environmental chemistry
• Machine learning applications

Background:

• Per- and polyfluoroalkyl substances (PFAS) are widely used but pose significant health risks due to their toxicity.
• Existing quantitative structure-activity relationship (QSAR) models, including traditional machine learning approaches, have limitations in predicting PFAS bioactivity and interpretability.
• There is a critical need for advanced computational methods to assess PFAS risks effectively.

Purpose of the Study:

• To develop and validate a novel machine learning approach for predicting PFAS toxicity using quantitative molecular surface analysis (QMSA).
• To enhance the predictive performance and interpretability of models for PFAS bioactivity assessment.
• To provide a framework for high-throughput screening of emerging PFAS in aquatic environments.

Main Methods:

• Development of a machine learning approach leveraging quantitative molecular surface analysis (QMSA) descriptors derived from molecular electrostatic potential.
• Application of five machine learning models, including random forest, utilizing QMSA descriptors to predict PFAS bioactivity against five biological targets.
• Validation of model performance using metrics such as accuracy, AUC-ROC, F1-score, and MCC.
• In-depth analysis of important QMSA descriptors using SHAP (SHapley Additive exPlanations).
• Molecular docking and molecular dynamics simulations to investigate the interaction mechanisms between PFAS and target proteins.

Main Results:

• The QMSA-based machine learning models achieved outstanding predictive performance, outperforming previously reported models.
• Best model performance metrics included accuracy of 0.950 ± 0.017, AUC-ROC of 0.938 ± 0.012, F1-score of 0.734 ± 0.024, and MCC of 0.684 ± 0.111.
• SHAP analyses identified estimated density, molecular volume, positive surface area, and nonpolar surface area as key descriptors influencing PFAS binding.
• Molecular simulations confirmed the role of these descriptors in non-covalent interactions during PFAS binding to target proteins.

Conclusions:

• The QMSA-based machine learning framework offers superior performance and interpretability for predicting PFAS toxicity.
• This novel approach facilitates high-throughput and cost-effective screening of high-risk emerging PFAS in aquatic environments.
• The study aids in prioritizing PFAS for regulation and treatment, supporting water quality monitoring, risk assessment, and environmental management decisions.
• Enhanced understanding of the molecular mechanisms underlying PFAS bioactivity is achieved.