A development of a graph-based ensemble machine learning model for skin sensitization hazard and potency assessment

Area of Science:

• Toxicology
• Computational Chemistry
• Dermatology

Background:

• The European Union's ban on animal testing for cosmetics necessitates alternative methods for skin sensitization assessment.
• Defined approaches (DAs) based on the adverse outcome pathway (AOP) and machine learning show promise in replacing traditional animal tests.
• Existing DAs can be further improved by integrating diverse data sources and advanced modeling techniques.

Purpose of the Study:

• To develop and evaluate an ensemble prediction model for skin sensitization assessment.
• To integrate in silico parameters and in vitro assay data within a machine learning framework.
• To enhance the predictivity of defined approaches (DAs) for both skin sensitization hazard and potency.

Main Methods:

• Developed an ensemble prediction model using graph convolutional network (GCN) and multilayer perceptron (MLP).
• Integrated GCN-derived features with physicochemical properties and data from in vitro assays (direct peptide reactivity assay, KeratinoSens™, h-CLAT).
• Evaluated model performance on predicting human hazard and three potency classes (strong, weak, non-sensitizer) using a testing set.

Main Results:

• The ensemble model incorporating GCN, KeratinoSens™, and h-CLAT achieved 88% accuracy for human hazard classification.
• The best three-class potency model, using GCN and all three assays, reached 64% accuracy.
• Eleven out of sixteen candidate models showed equal or improved accuracy when GCN features were included.

Conclusions:

• The integration of GCN features significantly enhances the predictive capability of ensemble models for skin sensitization.
• The developed ensemble approach offers a more accurate assessment of skin sensitization hazard and potency.
• This study supports the advancement of non-animal testing strategies for cosmetic ingredients.