基于数学模型的药物成分肾毒性预测

Predicting nephrotoxicity of drugs using mathematical models

目的:探索使用定量结构-活性关系(quantitative structure-activity relationship,QSAR)模型预测中药成分肾毒性。方法:从西药不良反应数据库(side effect resource,SIDER)收集得到了876个化合物(有肾毒不良反应记录的344个,无肾毒不良反应记录的532个),在对其描述符计算、挑选后,利用k近邻算法和支持向量机算法进行模型构建,然后用22个中药成分作为外部测试集进行模型验证评价。结果:利用k近邻算法当参数k为1时构建的模型对有肾毒数据集具有较好的预测能力,其对外部测试集中有肾毒成分的预测准确率也最高,达到64.29%。支持向量机算法构建的模型虽然有较高的自身预测准确率(80.25%),但是其对外部测试集的预测准确率仅有45.45%。结论:采用k近邻算法构建的模型对外部测试集22个中药成分具有较好的预测能力。

Objective： To establish computer models for the prediction of nephrotoxicity of small molecule drugs. Methods： Totally 876 small molecule drugs were collected from side effect resource （SIDER） database. Of these drugs 344 had adverse reactions related to nephrotoxicity, and the remaining drugs had no adverse reactions of nephrotoxicity. The predicting models by k-nearest neighbor （KNN） and support vector machine （SVM） were es- tablished based on calculation and selection of 2D descriptors. Twenty-two compounds from traditional Chinese medicines （TCM） were used as a test set to evaluate the prediction accuracy of the models. Results： The KNN model, if k = 1, had better accuracy for the training set itself, and also had the highest accuracy of 64.29% for the test set. The SVM model had an accuracy of 80.25% for the training set, but had a low accuracy, only 45.45% , for the test set. Conclusion： The KNN model of k = 1 can predict the nephrotoxicity of 22 TCM compounds with reasonable accuracy.