中药保健食品原料肝毒性预测研究

目的 开发一种更优的评估中药肝毒性的方法，为中药保健食品原料安全性评价提供参考。方法 基于国际公开构建肝毒性定量构效关系（QSAR）模型训练集数据，形成训练集。使用Discovery Studio 4.5对训练集进行主成分分析和聚类分析，针对每一类使用朴素贝叶斯（NB）、逻辑回归（LR）、邻近算法（KNN）等9种机器学习的方法构建QSAR模型，并利用交叉验证方法对模型构建方法进行评估。最后使用最优类模型对113种中药保健食品原料包含的783个成分进行了预测，根据多模型加权平均概率筛选出肝毒性成分占比较大的中药。结果 根据对训练集化合物的分析，可以看出第3类模型的准确率为85%~91%，高于现有报道的中药成分肝毒性预测模型。对中药保健食品原料成分的分析发现肝毒性化合物48个、不具有肝毒性的化合物735个，肝毒性预测概率为0.15~0.30，说明中药保健食品原料肝毒性普遍较低。预测肝毒性成分所占比例较高的中药有茜草、番泻叶、当归、大黄、丹参、厚朴、川芎、桑枝、桑白皮、五味子等。结论 对训练集预先聚类，提高QSAR模型准确率，为中药安全性评价的方法学研究提供了新思路，为中药保健食品原料成分进一步研究提供重要参考。

Prediction of Hepatotoxicity of Raw Materials of Functional Food Containing Chinese Medicinal Herbs

Objective To provide a better way for assessing the hepatotoxicity of Chinese medicines and provide references for the safety evaluation of raw materials of functional food containing Chinese medicinal herbs.Methods The training set was formed based on the international public training set data of the quantitativestructure-activityrelationship (QSAR) model for hepatotoxicity. Discovery Studio 4.5 was used to conduct principal component analysis and cluster analysis for the training set, and nine machine learning methods, including Naive Bayes (NB), logistic regression (LR), and k-nearest neighbors algorithm (KNN), were used to build QSAR models for each class. The cross-validation method was used to evaluate the model-building methods. Finally, the optimal model was used to predict 783 components from 113 raw materials of functional food containing Chinese medicinal herbs, and the Chinese medicine with the most hepatotoxic components was selected according to the weighted average probability of multiple models.Results The analysis of the compounds in the training set showed that the accuracy of the third model was 85%-91%, which was higher than the existing reported models for predicting the hepatotoxicity of Chinese medicines. The analysis of raw materials of functional food containing Chinese medicinal herbs showed that 48 compounds were found to be hepatotoxic and 735 compounds were found to be non-hepatotoxic. The prediction probability of hepatotoxicity was 0.15-0.30, indicating that the hepatotoxicity of raw materials of functional food containing Chinese medicinal herbs was generally low.The Chinese medicines with a higher proportion of hepatotoxic components were Rubiae Radix et Rhizoma, Sennae Folium, Angelicae Sinensis Radix, Rhei Radix et Rhizoma, Salviae Miltiorrhizae Radix et Rhizoma, Magnoliae Officinalis Cortex, Chuanxiong Rhizoma, Mori Ramulus, Mori Cortex, Schisandrae Chinensis Fructus, and so on.Conclusion The pre-clustering of training sets can improve the accuracy of QSAR model. This paper provides a new idea for the methodological research of safety evaluation of Chinese medicine and important references for the profound research of raw materials of functional food containing Chinese medicinal herbs.