ExSERA: The explainable machine learning model for skin sensitization risk assessment

The development of new approach methodologies using machine learning offers promising alternatives to animal testing for toxicity assessment. However, to be suitable for regulatory use, machine learning models must ensure transparency and interpretability. In this study, we developed the explainable machine learning model for skin sensitization risk assessment (ExSERA), a regression model designed to predict murine Local Lymph Node Assay (LLNA) EC3 values, quantitative indicators of skin sensitization potency. Using XGBoost, we trained the model on 154 substances from OECD TG 497 and externally validated it with 38 substances from the Expansion of the Cosmetics Europe Skin Sensitization Database. The model incorporated 21 variables, including results from three in vitro assays (DPRA, KeratinoSens™, and h-CLAT), molecular descriptors, and structural alerts. Within the applicability domain, 80% of internal validation and 65% of external validation predictions fell within a five-fold range of the measured LLNA EC3 values. Variable contribution analysis using Shapley Additive exPlanations showed that all three in vitro assays were the most influential contributors, with higher toxic activity correlating with stronger predicted sensitization. This interpretable model, grounded in the mechanistic understanding of skin sensitization, demonstrates potential for regulatory application as a defined approach.