Confidently Uncertain: Probabilistic Machine Learning to Predict Soil Biotransformation Half-Lives.

Area of Science:

• Environmental chemistry
• Computational toxicology
• Predictive modeling

Background:

• Predicting chemical environmental persistence from molecular structure is vital for regulatory screening and developing sustainable chemicals.
• Limited biotransformation half-life data hinders accurate persistence prediction, and existing models often lack generalizability.
• Reliable estimates of prediction confidence are essential for robust environmental risk assessments.

Purpose of the Study:

• To develop a probabilistic model for predicting soil biotransformation half-lives of chemicals.
• To provide well-calibrated probability distributions for chemical persistence, enabling accurate assessment of environmental risk.
• To enhance the identification of potentially persistent organic pollutants.

Main Methods:

• A Gaussian Process Regressor was employed, trained on 867 mean pesticide half-lives with associated data uncertainty.
• The model predicts probability distributions rather than single half-life values, incorporating uncertainty estimates.
• Probabilistic predictions were validated on pesticide transformation products and a database of global chemicals.

Main Results:

• The probabilistic model generates well-calibrated probability distributions for soil biotransformation half-lives.
• Despite moderate overall performance, predictions demonstrate high reliability when confidence is high.
• The model successfully identified known and suspected persistent chemicals in applied datasets.

Conclusions:

• The developed probabilistic model offers a reliable approach to estimate chemical persistence in soil environments.
• The model's ability to provide confidence estimates enhances its utility in regulatory decision-making and chemical design.
• Accessible as an online app and Python library, the model supports diverse applications in environmental science and chemical safety.