Endocrine disrupting chemicals (EDCs) are associated with various adverse health outcomes, thus necessitating high-throughput screening. However, current EDC screening models struggle to achieve high predictive performance and biological interpretability. Considering the predominant contribution of molecular interactions between EDCs and nuclear receptors (NRs) to molecular initiating events associated with adverse health outcomes, it is essential to integrate chemical-NRs interactions profiles into EDCs screening. Herein, we develop mechanism-driven models by integrating chemical-NRs interactions features of 372 434 chemicals with 16 NRs from large-scale molecular docking and chemical structural features, enabling comprehensive molecular representation of binding potential and physicochemical properties. Multiple models are constructed using diverse algorithms (XGBoost, NGBoost, Random Forest, TabPFN, and graph convolutional network), with SHAP analysis identifying key features of endocrine disruption to link predictions to chemical-NRs interactions. The optimal models achieved impressive performance (AUC: 0.815-0.995) and wide applicability with external validation against published EDCs inventory. The models are implemented in an open-access web server named EDC Profiler (http://www.edcprofiler.cn/service/edcprofiler), featuring interactive input interfaces, batch processing capabilities, and intuitive result visualization. The platform facilitates efficient EDCs screening and promotes new approach methods (NAMs) development for EDCs health risk assessment.
Qiu et al. (Tue,) studied this question.