Abstract: Objective To develop a machine learning model based on the Adverse Outcome Pathway (AOP) framework, integrating Molecular Initiating Events (MIE) and Adverse Outcomes (AO) , for the efficient prediction of endocrine-disrupting toxicity of chemical ingredients in personal care product. Methods Employing binding affinity to estrogen receptor (ERα, ERβ) and androgen receptor (AR) as the MIEs, molecular features of chemical compounds were extracted. Five machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), and K-Nearest Neighbors (KNN) were utilized to establish predictive models.. The optimal model was applied to virtually screen 2 267 chemical ingredients found in personal care products. High-risk constituents were selected based on prediction scores, and their estrogenic activity was further authenticated via in vivo assays using transgenic medaka (Oryzias latipes). Results For the ER and AR binding endpoints, the LR and SVM models demonstrated the highest predictive performance, with balanced accuracies of 0.934 7 and 0.831 6, respectively. For the AO endpoint, the RF model exhibited superior performance, with a balanced accuracy of 0.828 5. In vivo experiments revealed that p-hydroxyacetophenone and o-cymen-5-ol exhibited significant estrogenic activity and the capacity to potentiate endogenous estrogen effects. Conclusion The establishment of an integrated framework of ' computational prediction → risk prioritization → experimental validation' enables the effective identification of endocrine-disrupting chemicals in personal care products. This methodology offers a robust and novel strategy for high-throughput safety screening of chemicals.

Key words: Endocrine disrupting chemicals(EDCs), Machine learning, Molecular initiating event (MIE), Personal care products safety, Virtual screening