Novel graph neural network reveals binding mechanisms and environmental risks of PAHs interaction with estrogen...

Polycyclic aromatic hydrocarbons (PAHs) are widespread environmental contaminants that threaten ecosystems and human health by binding to estrogen receptor β (ERβ) and disrupting endocrine function. Accurately identifying and predicting the interactions between PAHs and ERβ remains a key challenge in environmental science. To address this, we propose a Multi-Scale Dual-Stream Graph Attention Network (MS-DSGAT) for predicting PAHs-ERβ binding affinity. MS-DSGAT outperforms traditional machine learning models, achieving the highest prediction accuracy (R² = 0.95) while offering strong interpretability. MS-DSGAT assigns Positional Attention Weights (P_AW) to atoms in each PAH molecule, highlighting the critical influence of functional groups such as hydroxyl (-OH), amino (-NH₂), and sulfonic acid (-SO₃H) on binding affinity. These insights provide valuable guidance for targeted molecular modifications. Virtual screening of 6357 external chemicals using MS-DSGAT identified approximately 6.6 % of the chemicals as high-affinity binders and 66.4 % as moderate binders. Molecular docking results further validate the model's interpretations, confirming functional groups as key determinants of binding affinity. By leveraging molecular graph representation, MS-DSGAT effectively predicts PAHs-ERβ interactions and can be extended to study other ligand-receptor interactions to identify potential endocrine disruptors, toxicants, and related compounds.