HPGCN: A graph convolutional network-based prediction model for herbal heat/cold properties

Herbal properties are part of the fundamental theories of traditional Chinese medicine (TCM), which has been of great significance for herbal formulas and disease treatment in clinics for thousands of years. However, determining herbal properties, such as heat/cold, still relies on ancient books and the doctor's experience, which can present significant limitations. In this study, we propose an herbal property graph convolutional network (HPGCN) model by combining TCM theory, modern pharmacological mechanisms, prior knowledge of herbal properties, and intelligent algorithms, which can effectively predict herbal heat/cold properties. Based on protein-protein interactions (PPI) and herb-herb networks, 30 target genes were selected as features for herbal heat/cold property prediction. Compared to previous machine learning algorithms, the HPGCN obtained optimal classification prediction results for ACC, Recall, Precision, F1, and AUC indicators by 5-fold cross-validation on the training and test sets. The function of herbs predicted by HPGCN improved by 3 % in hit@k compared to predictions that did not account for herbal properties. Herbs with disputed heat/cold properties in ancient books (such as Pulsatilliae Radix and Menthae Herba) were predicted using recommended property probabilities. The proposed HPGCN model may have profound practical value and significance for elucidating the scientific mechanisms of herbal property theory and in herbal medicine development.