Machine Learning-Based Models with High Accuracy and Broad Applicability Domains for Screening PMT/vPvM Substances

Area of Science:

• Environmental Chemistry
• Computational Toxicology
• Chemical Risk Assessment

Background:

• Persistent, mobile, and toxic (PMT) and very persistent and very mobile (vPvM) substances pose significant public health risks due to long-range transport.
• Effective risk prevention and mitigation require rapid, high-throughput screening methods for these hazardous substances.

Purpose of the Study:

• To develop and validate a machine learning-based screening system for the high-throughput classification of PMT/vPvM substances.
• To identify key molecular descriptors that predict the persistence and mobility of chemical substances.

Main Methods:

• Construction of five machine learning models (conventional, deep, and ensemble learning) using a dataset of 44,971 substances.
• Evaluation of model performance using metrics such as accuracy and AUROC, with LightGBM and XGBoost showing superior results.
• Identification of 'fr_halogen' (number of free halogen atoms) and 'MolLogP' (logarithm of partition coefficient) as critical descriptors for persistence and mobility, respectively.

Main Results:

• Machine learning models, particularly LightGBM and XGBoost, achieved classification metrics exceeding 0.900.
• The screening system demonstrated excellent generalization capability with an AUROC above 0.951.
• The system was successfully applied to classify persistent organic pollutants (POPs), prioritized PMT/vPvM substances, and pesticides.

Conclusions:

• The developed machine learning system provides an efficient and reliable tool for high-throughput screening of PMT/vPvM substances.
• This system can significantly aid in risk assessment and prioritization for managing emerging chemical contaminants.
• The study highlights the importance of molecular descriptors like fr_halogen and MolLogP in predicting environmental fate and transport.