基于深度神经网络的DYRK1A抑制剂预测模型研究

Deep Neural Network-based Models to Predict DYRK1A Inhibitors

目的:构建基于深度神经网络(deep neural network,DNN)的双底物特异性酪氨酸磷酸化调节激酶1A(DYRK1A)抑制剂预测模型,为筛选DYRK1A抑制剂提供计算工具。方法:从ChEMBL数据库中收集了DYRK1A抑制剂和非抑制剂共927个,通过随机采样10次建立了10组训练集和测试集,并计算出每个化合物的2种不同的分子特征,即MACCS指纹和Morgan2指纹,采用深度神经网络算法建立了20个模型,通过比较研究确定其中性能最佳的模型,并且对模型进行了Y-随机化检验、应用域和相似度图分析。结果:基于第2对训练-测试集Morgan2指纹的DNN模型(DNN1_Morgan2)性能最佳,对测试集内化合物的分类准确度为0.821,马修斯相关系数和ROC曲线下面积分别为0.647和0.917。结论:此最优模型可用于DYRK1A抑制剂的活性预测、虚拟筛选,并用于先导化合物的设计及优化。

Objective:This study is aimed to build a deep neural network(DNN)-based model to predict dual specificity tyrosine phosphorylation-regulated kinase 1A(DYRK1A)inhibitors,so as to provide a computational tool for virtual screening of DYRK1A inhibitors.Methods:A total of 927 DYRK1A inhibitors and non-inhibitors were collected from the ChEMBL database and 10 pairs of training sets and test sets were generated by random sampling for 10 times.Two different molecular features,namely MACCS fingerprints and Morgan2 fingerprints,were calculated for each compound,and then 20 models were built with the DNN algorithm.The best-performing model was determined by the comparative study and was further studied by the Y-randomization tests,applicability domain and similarity maps analysis.Results:The DNN model based on the 2nd pair of training and test sets and Morgan2 fingerprints(namely DNN1_Morgan2),showed the best performance.The accuracy of the optimal model on the test set was 0.821.The Matthews correlation coefficient and the ROC AUC value were 0.647 and 0.917,respectively.Conclusion:The optimal model can be used for activity prediction and virtual screening of DYRK1A inhibitors,as well as lead compounds design and optimization.