利用树模型预测中药成分的肝毒性  

Predicting Hepatotoxicity of Compounds from Traditional Chinese Medicines Using Tree Models

作　　者：叶立 王新洲 朱永亮 金若敏[2] 叶祖光[3] 姚广涛[2] 刘敬阁 钱向平 

机构地区：[1]苏州润新生物科技有限公司,江苏苏州215123 [2]上海中医药大学药物安全评价研究中心,上海201203 [3]中国中医科学院中药研究所,北京100700

出　　处：《中国药学杂志》2014年第18期1583-1588,共6页Chinese Pharmaceutical Journal

基　　金：国家重点基础研究发展计划(973计划)资助项目(2009CB522807);国家自然科学基金资助项目(81173652)

摘　　要：目的建立能够预测中药成分肝毒性的定量构效关系（quantitative structure—activity relationship，QSAR）模型。方法从LTKB数据库和国内文献中收集得到了286个人工合成化合物和62个中药成分，作为建立预测模型的训练集，利用简单决策树、随机森林和推进式决策树3种树形算法进行模型构建。为验证模型的预测能力，选择了22个中药成分（外部测试集）进行肝毒性实验，然后将实验结果与模型预测的结果进行比较。结果3种树形算法模型均具有较好的自我预测能力，模型内部交叉验证（1eave—one—out和leave一10％一out）结果都在78％～85％之间；但简单决策树和随机森林算法构建的模型对无肝毒化合物预测的准确率显著低于对有肝毒化合物预测的准确率，显示了较大的对有肝毒化合物预测的偏爱性；而利用推进式决策树算法所构建模型的预测偏爱性较小，总体预测能力也较高（准确率82％）。因此选用推进式决策树模型作为最优模型，对外部测试集中22个中药成分的肝毒性进行预测，准确率达到73％，比单用人工合成化合物作为训练集所构建的模型有更高的预测准确率和更小的偏爱性。结论用人工合成化合物及中药成分作为训练集，建立了对中药成分的肝毒预测能力较高的推进式决策树模型。OBJECTIVE To build tree models for the prediction of hepatotoxicity of compounds from traditional Chinese medicines （TCM）. METHODS Three hundred and forty-eight compounds （256 with hepatotoxicity and 92 without） were collected from various databases and literatures and used as training set to build tree models. Twenty-two compounds identified from TCM were first tested for hepatotoxicity experimentally and then used as test set to evaluate the prediction accuracy of optimal tree models. RESULTS Models built with random forest algorithm had the highest overall predictive accuracy of 85% （leave-one-out） , but had much lower accuracy for hepatotoxicity negative compounds compared to hepatotoxicity positive compounds （more positive bias）. The model built with boosted decision tree had a similar overall predictive accuracy and a much less bias, and therefore was selected as the optimal model. The prediction accuracy of the 22 test samples was 73% by the optimal model. The optimal model based on the training set containing both synthetic and TCM compounds had less bias than an optimal model based on a training set containing only the synthetic compounds. CONCLUSION Tree models with high predictive accuracy are built based on a training set consisting of both synthetic and TCM compounds. The optimal models can predict the hepatotoxicity of TCM compounds with reasonable accuracy.

关 键 词：中药成分 肝毒性预测 推进式决策树 定量构效关系(QSAR) 

分 类 号：R913[医药卫生—药学]