四性、五味和归经对中药肝毒性预测价值的研究

Research of Four Properties,Five Flavors and Channel Tropism on Predicting Hepatotoxicity of Chinese Herbal Medicines

目的筛选四性、五味和归经中对中药肝毒性预测的重要因素,并对预测模型进行评估。方法文献检索肝毒性中药107味,非肝毒性中药431味。记录中药的四性(寒、热、温、凉和平)、五味(酸、苦、甘、辛、咸、淡和涩)和归经(肺经、胃经、脾经、大肠经、心经、小肠经、膀胱经、肾经、心包经、三焦经、胆经和肝经),进行肝毒性中药、非肝毒性中药与其四性、五味和归经的相关性检验,并将相关性变量纳入数据库,利用SPSS Modeler进行数据分析并建立模型,筛选重要预测变量。使用Med Calc绘制各个模型预测概率的受试者工作特征曲线(receiver operator characteristic curve,ROC曲线),筛选各个模型预测的最佳分界值。结果建立了Logistic回归模型、神经网络模型和贝叶斯网络模型,3种模型预测变量重要性排名的前2位均有辛味;3种模型最佳预测分界值均在0.2附近(默认0.5),3种模型预测中药肝毒性的能力均距最优模型有一定差距。结论辛味可能是预测中药肝毒性的重要因素;数据挖掘技术对于中药肝毒性预测研究以及中药毒理学研究具有一定的应用价值。

Objective To screen the predictive variables from four properties, five flavors and channel tropism for predicting the hepatotoxicity of Chinese herbal medicines, and to evaluate the prediction models. Methods 107 Chinese herbal medicines with hepatotoxicity and 431 Chinese herbal medicines without hepatotoxicity were included after systematic literature search. We recorded the four properties（cold, hot, warm, cool and neutral）, five flavors （sour, bitter, sweet, pungent, salty, bland and astringent） and channel tropism （lung meridian, large intestine meridian, stomach meridian, spleen meridian, heart meridian, small intestine meridian, bladder meridian, kidney meridian, pericardium meridials, triple energizer meridian, gall bladder meridian, and liver meridian） of the included herbs. Correlation between Chinese herbal medicinals with or without hepatotoxicity and their four properties, five flavors and channel tropism was investigated. And then the related variables were input to the database. SPSS Modeler was used for data mining, setting up model and screening predictive variables. MedCalc was applied to draw the ROC curve with prediction probability and to screen out the best cut-off value of each model. Results Three models had been set up, namely logistic regression model, neural network model and Bayesian network model. Pungent flavor was an important predictor, and the best cut-off value of each model was close to 0.2（the default is 0.5）. The prediction ability of three models had a certain gap away from the optimal model. Conclusion Pungent flavor might act as a core factor of predicting the hepatotoxicity of Chinese herbal medicines, and data mining might be a promising method for studying hepatotoxicity and toxicology of Chinese herbal medicine.