Prediction models with multiple machine learning algorithms for POPs: The calculation of PDMS-air partition...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Toxicology

Background:

• The polydimethylsiloxane-air partition coefficient (K_PDMS-air) is crucial for passive sampling of persistent organic pollutants (POPs).
• Accurate K_PDMS-air values are essential for environmental monitoring and risk assessment of POPs.
• Existing quantitative structure-property relationship (QSPR) models often lack comprehensive applicability across diverse POP categories.

Purpose of the Study:

• To develop and validate robust QSPR models for predicting K_PDMS-air across multiple POP categories.
• To identify key molecular descriptors influencing the partitioning behavior of POPs.
• To provide a reliable tool for estimating K_PDMS-air values where experimental data is unavailable.

Main Methods:

• Development of 13 QSPR models using 244 POPs from 13 categories.
• Application of two descriptor selection methods (MLR, RF) and seven algorithms (MLR, LASSO, ANN, SVM, kNN, RF, GBDT).
• Rigorous internal and external validation using R_adj², Q_BOOT², and Q_ext² metrics.

Main Results:

• The Gradient Boosting Decision Tree (GBDT) model achieved R_adj² = 0.995, Q_BOOT² = 0.980, and Q_ext² = 0.951, outperforming other models.
• Mechanism explanation identified molecular size, branching, and bond types as key factors affecting partitioning.
• The developed models demonstrated superior fitting, robustness, and predictability.

Conclusions:

• The GBDT-based QSPR model offers a highly accurate and reliable method for predicting K_PDMS-air for a wide range of POPs.
• These models can effectively fill data gaps for experimental K_PDMS-air values, aiding environmental research.
• Understanding the relationship between molecular structure and partitioning behavior enhances the assessment of POPs distribution.