Prediction of skin sensitization potential using D-optimal design and GA-kNN classification methods.

Area of Science:

• Computational toxicology
• Cheminformatics
• Predictive modeling

Background:

• Skin sensitization is a critical endpoint in chemical safety assessment.
• Developing accurate predictive models for diverse chemical spaces remains challenging.
• Existing models often lack broad applicability and robustness.

Purpose of the Study:

• To develop global, predictive classification models for skin sensitization.
• To ensure models are applicable across the entire chemical space.
• To identify significant descriptors that predict skin sensitization potential.

Main Methods:

• Utilized a dataset of 255 diverse compounds and 450 calculated descriptors.
• Employed D-optimal design for optimal training and test set selection.
• Applied k-Nearest Neighbour classification (kNN) with Genetic Algorithms (GA-kNN) for descriptor selection and model building.

Main Results:

• Developed multiple stable and robust models (M1-M5).
• The best model (M1) achieved high accuracy: 84.3% (train), 87.2% (test), and 80.4% (external validation).
• Consensus prediction from models M1-M5 further improved accuracy across datasets.

Conclusions:

• The combination of D-optimal design and GA-kNN classification is a highly promising approach for predictive modeling.
• Selected descriptors accurately predict skin sensitization potential based on fundamental properties: lipophilicity, polarizability, shape, electrostatics, and reactivity.
• The developed models offer a stable, robust, and superior alternative to existing methods for assessing skin sensitization.