Skin Sensitization Potency Prediction of Ingredients in Hair Colorants Using in silico Models of Machine Learning

Abstract

Skin sensitization caused by cosmetics, including quasi-drugs, is an extremely serious problem and can become a health hazard. Although several in vitro methods for skin sensitization have been established by the Organization for Economic Cooperation and Development (OECD), no practical method for predicting skin sensitization potency has been found which can serve as an alternative to animal testing. In this study, an in silico model is developed which predicts skin sensitization potency by combining in vitro OECD test guidelines and information of the physical properties of chemicals to evaluate hair dye ingredients without animal testing. A dataset published by Cosmetics Europe was used to develop the in silico model. The EC3 value of the mouse local lymph node assay (LLNA), which is used as an indicator of skin sensitization potency, was the objective variable; in vitro test values, physical properties, and chemical information obtained from the molecular descriptor calculation software MOE constituted the explanatory variables. CatBoost was adopted as the machine-learning approach. Bandrowski’s base (BB), a trimer of the typical oxidative dye para-phenylenediamine (pPD), was used as the test material. In the predictive model, the mean-predicted EC3 value for BB was 0.33%, which meant a strong sensitizer. Furthermore, it was lower than the mean model prediction EC3 value (0.66%) of pPD, which is considered weaker sensitization potency than BB in LLNA. This indicates that BB was correctly predicted as a stronger skin sensitizer than pPD in the proposed in silico model.