目的 通过构建早期预测药源性急性间质性肾炎(drug-induced acute interstitial nephritis，DI-AIN)的计算模型，筛选中药致DI-AIN成分。方法 从文献和SIDER等数据库中收集得到了125个致DI-AIN的药物和122个未导致DI-AIN的药物，作为预测模型的训练集，基于人工神经网络(artificial neural network，ANN)和支持向量机(support vector machine，SVM)2种算法构建模型。通过文献各选择45种药物进行验证，以评估最优模型的预测性能。并将其应用于筛选10种中药的DI-AIN成分。结果 共筛选得到207种分子描述符参与建模，其中，ANN和SVM算法搭建的最优模型分别包含112和80种分子描述符，2个模型的特异度、灵敏度、准确度均在84%以上。使用2种算法搭建的最优模型进行外部验证，准确度均在90%以上。ANN联合SVM模型预测为DI-AIN的中药成分有雷公藤甲素、水苏碱、京尼平苷等。结论 首次建立中药成分致DI-AIN早期预测的计算模型，具有良好的预测能力，对于中药肾毒性预测研究及中药毒理学研究具有一定的应用价值。

Objective To construct computational models for early prediction of drug-induced acute interstitial nephritis (DI-AIN) to screen the ingredients of traditional Chinese medicine (TCM)-induced DI-AIN. Methods A total of 125 DI-AIN-causing drugs and 122 non-DI-AIN-causing drugs were collected from literatures and SIDER database as training sets for prediction models. The models were constructed through the algorithms of artificial neural network (ANN) and support vector machine (SVM). In order to evaluate the prediction performance of the optimal model, a total of 45 DI-AIN-causing drugs and 15 non-DI-AIN-causing drugs were collected as external validation set through literatures. Two models were applied to screen the DI-AIN components of 10 TCMs. Results After filtering, 207 molecular descriptors were obtained to construct the model. The optimal models built by ANN and SVM algorithms contained 112 and 80 molecular descriptors respectively, and the specificity, sensitivity and accuracy of two models were all above 84%. The ANN and SVM models were used for external validation with accuracy above 90%. ANN and SVM models predicted that the TCM ingredients of DI-AIN included triptolide, stachydrine, geniposide. Conclusion This study established the computational model with reasonable accuracy for the early prediction of DI-AIN caused by the ingredients of TCM for the first time. It might be a promising method for studying nephrotoxicity and toxicology of TCM.

R285

国家自然科学基金资助项目(81573825)；国家中医药管理局青年岐黄学者支持项目