Multiple machine learning algorithms assisted QSPR models for aqueous solubility: Comprehensive assessment with...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Machine Learning in Environmental Science

Background:

• Aqueous solubility is a critical environmental property for assessing hydrophobicity, ecological risk, and toxicity of organic pollutants.
• Quantitative Structure-Property Relationship (QSPR) models are used to predict aqueous solubility, but lack standardized evaluation procedures.
• Existing QSPR models face challenges in comprehensive assessment, hindering the selection of optimal predictive tools.

Purpose of the Study:

• To propose and validate the CRITIC-TOPSIS comprehensive assessment method for environmental QSPR models.
• To evaluate the predictive performance of various machine learning (ML) algorithms for aqueous solubility prediction.
• To identify the most accurate and externally competitive QSPR models for predicting aqueous solubility (log Kws).

Main Methods:

• Development of 39 QSPR models using 13 ML algorithms across 4 algorithm families and 3 descriptor screening methods.
• Application of the CRITIC-TOPSIS comprehensive assessment method, integrating multiple statistical parameters.
• Evaluation of model predictive accuracy and external competitiveness.

Main Results:

• The CRITIC-TOPSIS method was successfully applied to comprehensively evaluate QSPR models for aqueous solubility.
• MLR-1, XGB-1, DNN-1, and kNN-1 models showed superior predictive accuracy within their respective algorithm tribes.
• The XGBoost model combined with the SRM descriptor screening method (XGB-1, C = 0.599) emerged as the optimal model for predicting aqueous solubility.

Conclusions:

• The proposed CRITIC-TOPSIS approach provides a robust framework for evaluating environmental property prediction models.
• The optimized XGBoost model offers a reliable and accurate method for predicting the aqueous solubility of organic chemicals.
• This comprehensive evaluation strategy can guide the selection of superior QSPR models in environmental research.