To explore the building and application of the prediction model of cold/hot properties of Chinese herbal medicines.Methods:A total of 646 Chinese herbal medicines，involving 10 053 compounds，were obtained from the traditional Chinese medicine information database.The molecular fingerprints and molecular graph representation coding of Chinese medicinal compounds were established，and the graph convolutional neural network(GCN)，K nearest neighbor(KNN)，decision tree(DT)，random forest(RF)，and support vector machine(SVM) algorithms were constructed，respectively.The hyperparameters of each model were adjusted.The data set was divided into a training set and a test set，and the accuracy，precision，recall，and F value of each model were determined to evaluate the model prediction performance.The predicted compounds with high probability were used to reveal the potential differences in biological mechanisms between cold and hot properties.The cell proliferation-toxicity test was carried out to examine the protective effect of the compounds with a cold property predicted based on the weighted score on PC-12 cells exposed to oxygen-glucose deprivation and reperfusion(OGD/R).Results:The GCN model performed well in the prediction of cold/hot properties of Chinese herbal medicines.The compounds with a high probability of cold and hot property classification screened by the GCN model from the representative cold/hot medicines showed a total of 413 targets，which were associated with 17 pathways.The cell experiment results showed that the components with a cold property demonstrated a protective effect on the PC-12 cells exposed to OGD/R as the weighted score decreased.Conclusion:In the cold/hot property prediction task，the GCN model based on molecular graph representation outperforms the conventional machine learning model based on molecular fingerprint representation，which can provide algorithm support for exploring the “property-structure” relationship of Chinese herbal medicines.