Objective To address the complex molecular structures and diverse ring systems characteristic of traditional Chinese medicine (TCM) compounds, this study aims to optimize the graph isomorphism network (GIN) to enhance its predictive capability for key molecular properties of TCM molecules. Methods Based on the GIN architecture, edge descriptors representing bond length strain and angle strain were introduced into the molecular graph, and multi-graph data were constructed by incorporating multiple favorable conformations, thereby developing a flexibility-ring enhanced graph neural network (FRGNN). The predictive performance of the proposed model on seven key molecular properties was evaluated using two TCM databases containing 37 822 molecules, and compared against three state-of-the-art (SOTA) graph neural network models and two basic graph neural network models. Results Compared with the second-best performing model, FRGNN achieved an average reduction of 8.63% in root mean square error across the seven property prediction tasks, for molecules containing polycyclic and macrocyclic structures, the predictive root mean square error was further reduced by 10.04%. Conclusion The proposed FRGNN model demonstrates superior performance over existing state-of-the-art models for small-molecule property prediction in the task of predicting key properties of TCM compounds, providing a novel and effective approach for the property prediction of complex natural products.

artificial intelligence  /  graph neural networks  /  traditional Chinese medicine molecules  /  Chinese medicine datasets  /  property prediction