TCMRGAT: Relational graph attention networks for predicting stroke treatment efficacy of traditional Chinese medicine prescriptions.

Background: Stroke is a serious neurological disorder that poses a global health challenge. Traditional Chinese Medicine (TCM) prescriptions have shown potential in its treatment. However, TCM prescriptions typically involve a wide variety of botanical drugs, and the efficacy of different combinations varies, with underlying patterns remaining unclear. This study aims to develop a model to predict the efficacy of TCM prescriptions for stroke, so as to deepen understanding of the underlying mechanisms of botanical drug therapies. Methods: We collected stroke-related TCM data, including prescriptions, botanical drugs, metabolites, and targets, from TCM classics and the HERB database. A generative adversarial network (GAN) was used to augment imbalanced data, and constructed a heterogeneous network. Then, we initialized node features and performed neighborhood feature learning using a relational graph attention network (RGAT) to predict TCM prescription efficacy. We compared our method, named RGAT for TCM prescription efficacy prediction (TCMRGAT), with other models. Results: TCMRGAT achieved an accuracy of 0.843 and an area under curve (AUC) of 0.853 on balanced data, outperforming competing methods. Ablation experiments confirmed the effectiveness of GAN-based data augmentation. Case studies using RGAT and GPT-4 highlighted the model's potential in real-world applications. Analysis of post-training attention weight changes revealed potential key botanical drug-metabolite relationships, suggesting they may be directly associated with stroke treatment. Conclusion: TCMRGAT aids in predicting prescription efficacy and identifying key metabolite s for stroke treatment. This study provides valuable insights into the use of Traditional Chinese Medicine for stroke and offers a promising direction for future research.