Mechanism Based Hierarchical Machine Learning for High-Throughput Quantitative Prediction of Estrogenic, Androgenic, and Thyroid Disruption Activities

Although qualitative predictions of endocrine-disrupting chemicals (EDCs) are well established, quantitative high-throughput models remain underdeveloped due to data heterogeneity and mechanistic complexity. To address this gap, we developed a mechanistically informed hierarchical framework that quantitatively predicts estrogenic, androgenic, and thyroid (EAT)-disruption activities. The framework consists of three key components: data refinement, fragment-based mechanistic clustering, and cluster-specific quantitative modeling. First, we eliminated 20.12% of false positives and 41.54% of false negatives through a five-step curation, resulting in a high-confidence data set for model development. Next, the predictive model couples a classifier that assigns EAT activity modes with an ensemble regressor for potency estimation. Compared to conventional models, Modeling based on high-confidence data sets and mechanism classification demonstrates improved performance (<i>R</i>² = 0.72-0.78, RMSE = 0.22-0.48 log₁₀(μM)). Mechanistic insights from molecular simulations across EAT receptors revealed that potent agonists feature aromatic cores with polar substituents, stabilizing hydrogen-bond networks and promoting helix 12 (H12) activation. In contrast, potent antagonists exhibit flexible chains or rigid polycyclic scaffolds that disrupt H12 orientation. Altogether, we offer a next-generation strategy for interpretable and accurate endocrine potency prediction.