Graph convolutional network learning model based on new integrated data of the protein-protein interaction network and network pharmacology for the prediction of herbal medicines effective for the treatment of metabolic diseases

Abstract

Chronic metabolic diseases constitute a group of conditions requiring long-term management and hold significant importance for national public health and medical care. Currently, in Korean medicine, there are no insurance-covered herbal prescriptions designated primarily for the treatment of metabolic diseases. Therefore, the objective of this study was to identify herbal prescriptions from the existing pool of insurance-covered options that could be effective in treating metabolic diseases. This research study employed a graph convolutional network learning model to analyze PPI network constructed from network pharmacology, aiming to identify suitable herbal prescriptions for various metabolic diseases, thus diverging from literature-based approaches based on classical indications. Additionally, the derived herbal medicine candidates were subjected to transfer learning on a model that binarily classified the marketed drugs into those currently used for metabolic diseases and those that are not for data-based verification. GCN, adept at capturing patterns within protein-protein interaction (PPI) networks, was utilized for classifying and learning the data. Moreover, gene scores related to the diseases were extracted from GeneCards and used as weights. Due to the absence of any prior research using similar data and learning structures, an alternative evaluation method of our pre-trained model was deemed necessary. The performance of the pre-trained model was validated through 5-fold cross-validation and bootstrapping with 100 iterations. Furthermore, to ascertain the superior performance of our proposed model, the number of layers was varied, and the performance of each was evaluated. Our proposed model structure achieved outstanding performance in classifying drugs, with an average precision of 96.68%, recall of 97.18%, and an F1 score of 96.74%. The trained model predicted that the most effective decoction would be Jowiseunggi-tang for hyperlipidemia, Saengmaegsan for hypertension, and Kalkunhaeki-tang for type 2 diabetes. This study is the first of its kind to integrate GCN with weighted PPI network data to classify herbal prescriptions by their potential for usage on certain diseases.