机器学习方法用于激素敏感脂肪酶抑制剂活性预测

摘要

对激素敏感脂肪酶,我们构建了表征分子组成、电荷、拓扑、几何结构及物理化学性质等特征的1559个描述符,通过Fischer Score排序过滤和Monte Carlo模拟退火法相结合进行变量筛选得到35个描述符,然后分别用支持向量学习机(SVM)、人工神经网络(ANN),k-近邻(k-NN),连续核密度估计(CKD)和逻辑回归(LR)等机器学习方法建立了激素敏感脂肪酶抑制剂的分类预测模型.对于训练集的200个样本,通过五重交叉验证,各机器学习方法对正样本,负样本和总样本的平均预测精度分别在78.0％-94.0％,69.0％-91.0％和73.5％-92.5％；通过y-scrambling方法验证SVM模型是否偶然相关,结果正样本,负样本和总样本的平均预测精度分别在60.0％-74.0％,58.0％-71.0％和61.0％-69.5％,明显低于实际所建模型的预测精度,表明所建模型不存在偶然相关；对52个没有参与建模的外部独立测试样本,各机器学习方法对正样本,负样本和总样本的预测精度分别在84.6％-92.3％,88.5％-92.3％和86.5％-92.3％.所建模型中,SVM,CKD和LR较好,且明显高于其他文献报道结果.%A total of 1559 molecular descriptors including constitutional, charge distribution, topological, geometrical, and physico-chemical descriptors were calculated to encode the hormone-sensitive lipase inhibitors. The number of 35 molecular descriptors Was selected using a hybrid filler/wrapper approach combing Fischer Score and Monte Carlo simulated annealing, then classificationmodels for hormone-sensitive lipase inhibitors were built based on support vector machine ( SVM) , artificial neural networks (ANN) ,k-nearest neighbor( k-NN ) ,continuous kernel discrimination(CKD)and logistic regression ( LR) methods. For 200 samples in training set, average prediction accuracies of 78. 0% -94. 0% ,69. 0% -91. 0% and 73. 5% -92. 5%for positive, negative, and total samples, respectively, were obtained through 5-fold cross validation. Average prediction accuracies of 60. 0% -74. 0% , 58. 0% -71. 0%and 61. 0%-69. 5%for positive,negative,and total samples,respectively,were obtained by using y-scrambling method,indi-cating that there was no chance correlation on our models. For an external test of 52 samples which were not used in models building, prediction accuracies of 84. 6%-92. 3% ,88. 5% -92. 3% and 86. 5%-92. 3% for positive, negative, and total samples, respectively, were obtained. The prediction accuracies by all machine learning methods,especially by SVM method,in this study were far better than literature results.