The development of a computational model for predicting sensitization potential in cosmetics

The analysis and testing allergenic ingredients in cosmetics are crucial for quality and safety control. This study presents an improved method based on machine learning ensemble models, a dynamic weighted combination prediction model for sensitization prediction of cosmetic ingredients, aiming to increase prediction accuracy. We constructed an expanded dataset by integrating data from human, animal, and non-animal sources. The final curated dataset consisted of 2,862 unique cosmetic ingredients. And all sensitization outcomes were uniformly standardized into binary labels following OECD TG 406 criteria. Building on the work of Fuadah et al. in 2024, we developed a more flexible prediction framework. This framework first employed LazyPredict to screen multiple benchmark models, which were then integrated via an ensemble method. Our approach, leveraging multiple data sources and models, achieved improved performance compared to previous studies. The new model achieved AUC scores between 0.86 and 0.91 for predicting cosmetic ingredient sensitization. The result is based on internal validation only, and further external validation is needed. A Mann–Whitney U test based on Bootstrap resampling confirmed that the distribution of AUC performance for the new model was statistically superior to that of the previous model, a conclusion supported by the calculated Cliff’s Delta effect size. This resulted in an approximately 10% reduction in prediction error compared to the model trained solely on the original LLfNA dataset. This study verifies the complementary nature of different models and data sources, thereby providing a robust methodological foundation for cosmetic safety assessment.