Definition of the molecular bases of cold and hot properties of traditional Chinese medicine through machine learning

Abstract

The purpose of this study is to confirm whether the 'cold-hot properties' of traditional Chinese medicine (TCM) have modern scientific connotations on the molecular level of compounds using machine learning (ML).In this study, 393 Chinese herbs were selected, modeling was based on three ML algorithms, and the voting fusion algorithm was used to construct the fusion model. According to the ROC area under the curve (AUC), the cold-hot herb property classification model was screened. The a priori and K-means algorithms were used to mine key molecules related to cold-hot properties.The AUC of the ML models and fusion models under the ROC curve after 10 rounds of folding cross-verification was above 0.75. Mining obtained 15 cold key compounds (CKCs), 26 hot key compounds (HKCs), a CKC structure similarity of 0.75 on average, and an HKC structure similarity of 0.69 on average. The average similarity between the molecular structures of CKCs compounds and HKCs compounds was 0.18, lower than 0.5. Twelve physical and chemical indices, such as molecular weight (MW) and molecular hydrophobicity (Alogp), had a statistically significant difference in molecular characteristics (P<0.05).In this study, the classification model and algorithm (RF&GNB Ensemble Classifier) of TCM cold-hot properties constructed from the molecular level of compounds are robust. At the same time, it was found that Chinese herbs with the same medicinal properties had a common molecular structure, while Chinese herbs with different medicinal properties had different molecular structures. In terms of physicochemical characteristics, compounds have statistically significant differences.