Pollution risk assessment by designing predictive binary classification models of substituted benzenes centered on...

Area of Science:

• Computational toxicology and cheminformatics.
• Development of predictive models for chemical safety assessment.

Background:

• Growing need for rapid chemical safety assessment by regulatory agencies.
• Limitations of traditional animal testing in terms of time and cost.
• Preference for quantitative structure-toxicity relationship (QSTR) models in chemical regulation.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting toxicity of benzene-derived compounds (BZCs).
• To identify a robust and interpretable model for QSTR applications.
• To develop a set of decision rules for BZC toxicity prediction.

Main Methods:

• Evaluation of four binary classification ML models: support vector machine, k-nearest neighbor, CART decision tree, and random forest.
• Utilized ClustOfvar for optimal feature selection and SMOTE for data balancing.
• External validation using fivefold cross-validation on 1416 benzene-derived compounds (708 non-toxic, 708 toxic).

Main Results:

• The CART decision tree (DT) model demonstrated superior performance.
• Achieved high accuracy metrics: Q=95.42%, Precision=96.60%, Recall=94.67%, F-score=95.62%, Specificity=96.27%, MCC=0.91, AUC=1.0.
• Identified 10 human-readable decision rules for predicting BZC toxicity.

Conclusions:

• The CART decision tree model is a robust and explanatory tool for QSTR.
• The developed methodologies can aid in filling data gaps and prioritizing chemical safety testing.
• Proposed approach facilitates efficient identification of hazardous organic chemicals.