Skin sensitisation testing in practice: Applying a stacking meta model to cosmetic ingredients

Area of Science:

• Toxicology
• Computational Chemistry

Background:

• Regulatory bodies increasingly seek non-animal testing methods for hazard identification, particularly for cosmetics.
• Existing in vitro and in silico methods show promise but require robust validation for broad acceptance.

Purpose of the Study:

• To evaluate the performance of a novel defined approach integrating multiple data sources for skin sensitization hazard identification.
• To assess the accuracy of this approach specifically for cosmetic raw materials compared to traditional in vivo data.

Main Methods:

• A defined approach integrating three in vitro assays (DPRA, KeratinoSens™, U-SENS™), two in silico tools (TIMES SS, TOXTREE), and physicochemical parameters was utilized.
• A stacking meta-model (version 5) was developed, incorporating these diverse data streams.
• Model predictions were compared against existing Local Lymph Node Assay (LLNA) data for 219 substances, predominantly cosmetic ingredients.

Main Results:

• The defined approach demonstrated high accuracy, ranging from 85% to 91% for cosmetic categories.
• Incorporating the TIMES SS in silico model significantly improved the model's ability to discriminate between sensitizers and non-sensitizers.
• The model successfully classified 68 non-sensitizers, 86 weak/moderate sensitisers, and 65 strong sensitisers.

Conclusions:

• The stacking meta-model represents a valuable tool for the next generation of risk assessment frameworks in toxicology.
• The approach offers reliable predictions for skin sensitization hazards, supporting the move away from animal testing in the cosmetics industry.
• Confidence in predictions is enhanced by the probabilistic output of the stacking meta-model.