AOP-ontology-guided QSAR pipeline machine and deep learning: Prediction and mechanistic interpretation of female reproductive injury

: To explore the substances with estrogen interference effect, it is necessary to develop the optimal prediction model of inducing the interference effect of female reproduction, instead of the traditional research method of in vivo experiments. Focused on five adverse outcome pathways (AOPs) related to female reproductive damage, we established the prediction model of optimal machine learning or deep learning for each molecule-initiated events (MIE) and key event (KE) in each AOP. Morgan fingerprint combined with random forest had the best performance in predicting the interference effect of female reproduction. Besides, we found out which AOPs was stimulated by 1127 environmental hormones to achieve toxic effects, and the substructure of their compounds with the potential to trigger AOPs. Tributyltin compounds mostly interfered with female reproduction through AOP7, AOP549 and AOP289 pathways. Furthermore, the molecular docking study found the key proteins with good binding activity with those substructures. Among them, compounds containing nitrophenyl structure had strong binding activity with 5α-reductase and vitellogenin in AOP289. This study provided a new research perspective and method for predicting whether the compounds have female reproductive interference effect and exploring the toxicity mechanism.