Prediction of skin sensitization using machine learning.

Area of Science:

• Toxicology
• Computational Chemistry
• In silico toxicology

Background:

• Growing global awareness of animal welfare necessitates the development of alternative testing methods for chemical safety.
• Skin sensitization assessment is crucial for product safety and regulatory compliance.

Purpose of the Study:

• To develop a practical machine learning model for predicting skin sensitization.
• To utilize physicochemical properties (surface tension, melting point, molecular weight) for model development.

Main Methods:

• A Random Forest (RF) model was developed using 297 chemical datasets with 6 physicochemical properties and local lymph node assay results.
• The model was validated against 45 fragrance allergens using performance metrics like f1-scores.
• Experimental validation was performed using the Direct Peptide Reactivity Assay (DPRA).

Main Results:

• The RF model demonstrated strong performance, outperforming Support Vector Machine (SVM), QSARs (ChemTunes, OECD Toolbox), and a linear model in various classification tasks (penal, ternary, binary).
• Recommended ternary classification based on the Global Harmonized System for more precise hazard information.
• Experimental validation with DPRA showed similar trends to the model's predictions.

Conclusions:

• The developed machine learning model provides an efficient and rapid method for screening chemical sensitization hazards.
• This in silico approach can significantly reduce reliance on animal testing for skin sensitization assessment.
• The study supports the use of physicochemical properties and machine learning for predicting toxicological endpoints.