Improved defined approaches for predicting skin sensitization hazard and potency in humans

Area of Science:

• Toxicology
• Computational Chemistry
• Dermatology

Background:

• The EU ban on animal testing for cosmetics necessitates alternative methods for skin sensitization assessment.
• Machine learning (ML) in defined approaches (DA) shows promise but faces challenges like data imbalance and limited database information.
• Improving the predictivity of DAs is crucial for reliable, animal-free safety evaluations.

Purpose of the Study:

• To enhance the predictivity of DAs for skin sensitization assessment using data-rebalancing ensemble learning and a comprehensive database.
• To develop and evaluate ML models for predicting both human hazard and three-class potency of cosmetic ingredients.
• To compare the performance of novel DAs against existing methods like the Local-Limb Node Assay (LLNA).

Main Methods:

• Application of data-rebalancing ensemble learning, specifically bagging with Support Vector Machine (SVM).
• Utilized a novel and comprehensive Cosmetics Europe database for training and testing ML models.
• Developed 12 SVM-bagging models for hazard prediction and 12 SVM models for potency prediction using 96 training and 32 test substances.

Main Results:

• The best hazard prediction model (hazard-DA) achieved 90.63% accuracy on the test set using SVM-bagging with all variables.
• The best potency prediction model (potency-DA) achieved 68.75% accuracy on the test set using SVM alone.
• Both developed DAs outperformed LLNA and other ML-based DAs, with potency-DA offering more detailed assessment.

Conclusions:

• SVM-bagging-based DAs significantly enhance hazard prediction accuracy through data rebalancing.
• Detailed categorization of sensitizers may offset data imbalance for potency assessment, making standalone SVM effective.
• The improved DAs represent promising, animal-free tools for robust skin sensitization risk assessment in cosmetics.