Predicting binding affinities of liquid crystal monomers: An activity cliffs-driven multidimensional feature fusion model

Liquid crystal monomers (LCMs) have emerged as novel endocrine disrupting chemicals that affect the growth, development, and metabolism of organisms by binding to nuclear hormone receptors (NHRs). However, the studies on the impact of LCMs' molecular features on their binding affinities remain limited. In this study, considering the challenge of activity cliffs in linear quantitative structure-activity relationship modeling, a multidimensional feature fusion model was developed to predict the binding affinities of 1173 LCMs to 15 NHRs. This model integrated the automatic feature extraction capability of Message Passing Neural Network (MPNN) and the interpretability of Categorical Boosting (CatBoost). The results indicated that stratified splitting of the activity cliffs into both training and test sets enhanced the model's learning and generalization compared to assigning them to only one set. Activity cliffs analysis revealed that the binding affinities of LCMs enhanced with the length of the carbon chain and the number of cyclohexyl and phenyl rings. The interpretability of the MPNN_CatBoost model revealed that the total number of cyclohexyl and phenyl rings, polarity, and volume were common features influencing the binding affinities of LCMs with different NHRs. These findings offer a comprehensive understanding of the molecular features influencing the potential endocrine-disrupting effects of LCMs, thereby providing a scientific foundation for the design and development of more environmentally friendly alternatives.