基于机器学习的利用药物标签信息定量预测药物-药物相互作用

Machine learning-based quantitative prediction of drug drug interaction using drug label information

目的用现有药代动力学(PK)药物相互作用(DDI)信息的数据库,构建出可用于预测AUC倍数变化(FC)的机器学习模型,用于探究对现有DDI预测的可能,为临床用药提供一定的合理建议。方法从美国食品药品监督管理局(FDA)认证的药品标签中提取DDI的PK数据和AUC倍数变化的数据。通过DrugBank检索出DDI有关的多肽和药效学(PD)信息,用蛋白质资源(UniProt)对相关多肽ID进行药物类型(PPDT)标识,用矩阵归一化的代码生成便于分析的多维向量数据。PPDT对AUC的影响和所产生的倍数变化作为因变量,进行机器学习模型构建。用均方根误差(RMES)值最小的模型进行模型构建,训练出袋装决策树(Bagged)预测模型。利用训练好的模型对部分药物检验,检测模型的预测性别。通过查阅现有的有关检测DDI对的文献研究结果,对预测值进行分析比较,对模型进行评价。结果检验模型药物对共16对,分别为16种药物对他克莫司的影响,发现对DDI的有无预测准确率为81.25%;预测结果根据FDA标准分类强弱,结果表明,DDI强弱预测,偏离较大的预测较少。结论模型预测DDI的有无评价一般;但对DDI的强弱分类后,对DDI的预测结果较好,预测结果说明模型预测性能对于在临床试验之前进行潜在的DDI评估具有一定的参考价值。

Objective To construct machine learning models that can be used to predict AUC fold change(FC)using a database of existing pharmacokinetic(PK)and drug-drug interaction(DDI)information,which can be used to explore the possibility of predicting existing drug interactions and to provide certain rational recommendations for clinical drug use.Methods PK data of DDIs and AUC fold change data were extracted from FDA-approved drug labels.Peptide and pharmacodynamic(PD)information related to drug interactions were retrieved through DrugBank,and PPDT identification of relevant peptide IDs was performed using Protein Resource(UniProt),and a matrix normalization code was used to generate multidimensional vector data that were easy to analysis.The effect of PPDT on the AUC,and the resulting multiplicity change was used as the dependent variable for machine learning model construction.The model with the smallest root mean square error(RMES)value was used for model construction to train a bagged decision tree(Bagged)prediction model.The models were tested using the trained models for some of the drug tests.The models were evaluated by reviewing the available literature findings on detection of drug interaction pairs and analyzing and comparing the predicted values.Results A total of 16 pairs of model drug pairs were tested for the effects of 16 drugs on tacrolimus,and it was found that the accuracy of the prediction of the presence or absence of drug interactions was 81.25%;the prediction results were classified according to the FDA standard classification of the strong and weak for the strength of drug interactions,and the results showed that the prediction of the strength of drug interactions,with a large deviation from the larger prediction was less.Conclusion The evaluation of the model to predict the presence or absence of drug interactions was general;however,after classifying the strengths and weaknesses of drug interactions,the prediction of drug interactions was better,and the prediction results indicated that the model prediction performance has a certain reference value for potential DDI assessment before clinical trials.