Calculating the similarity between prescriptions to find their new indications based on graph neural network

Abstract Background Drug repositioning has the potential to reduce costs and accelerate the rate of drug development, with highly promising applications. Currently, the development of artificial intelligence has provided the field with fast and efficient computing power. Nevertheless, the repositioning of traditional Chinese medicine (TCM) is still in its infancy, and the establishment of a reasonable and effective research method is a pressing issue that requires urgent attention. The use of graph neural network (GNN) to compute the similarity between TCM prescriptions to develop a method for finding their new indications is an innovative attempt. Methods This paper focused on traditional Chinese medicine prescriptions containing ephedra, with 20 prescriptions for treating external cough and asthma taken as target prescriptions. The remaining 67 prescriptions containing ephedra were taken as to-be-matched prescriptions. Furthermore, a multitude of data pertaining to the prescriptions, including diseases, disease targets, symptoms, and various types of information on herbs, was gathered from a diverse array of literature sources, such as Chinese medicine databases. Then, cosine similarity and Jaccard coefficient were calculated to characterize the similarity between prescriptions using graph convolutional network (GCN) with a self-supervised learning method, such as deep graph infomax (DGI). Results A total of 1340 values were obtained for each of the two calculation indicators. A total of 68 prescription pairs were identified after screening with 0.77 as the threshold for cosine similarity. Following the removal of false positive results, 12 prescription pairs were deemed to have further research value. A total of 5 prescription pairs were screened using a threshold of 0.50 for the Jaccard coefficient. However, the specific results did not exhibit significant value for further use, which may be attributed to the excessive variety of information in the dataset. Conclusions The proposed method can provide reference for finding new indications of target prescriptions by quantifying the similarity between prescriptions. It is expected to offer new insights for developing a scientific and systematic research methodology for traditional Chinese medicine repositioning.