Laboratory of Chemometrics, Faculty of Chemistry, University of Mazandaran, Babolsar, Iran.
J Sep Sci. 2011 Nov;34(22):3216-20. doi: 10.1002/jssc.201100544. Epub 2011 Oct 20.
In this work, quantitative structure-retention relationship (QSRR) approaches were applied for modeling and prediction of the retention index of 282 amino acids (AAs) and carboxylic acids (CAs). Descriptors that were used to encode structural features of molecules in a data set were calculated by using the Dragon software. The genetic algorithm (GA) and stepwise multiple linear regression (MLR) methods were used to select the most relevant descriptors. Then support vector machine (SVM), artificial neural network (ANN) and multiple linear regression were utilized to construct nonlinear and linear quantitative structure-retention relationship models. The obtained results using these techniques revealed that nonlinear models were much better than other linear ones. The GA-ANN model has the average absolute relative errors (AARE) of 0.054, 0.059 and 0.100 for training, internal and external test set. Applying the tenfold cross-validation procedure on GA-AAN model obtained the statistics of Q(2)=0.943, which revealed the reliability of this model.
在这项工作中，定量构效关系（QSRR）方法被应用于建模和预测 282 种氨基酸（AAs）和羧酸（CAs）的保留指数。用于在数据集编码分子结构特征的描述符是通过使用 Dragon 软件计算得到的。遗传算法（GA）和逐步多元线性回归（MLR）方法被用于选择最相关的描述符。然后，支持向量机（SVM）、人工神经网络（ANN）和多元线性回归被用于构建非线性和线性定量构效关系模型。使用这些技术得到的结果表明，非线性模型比其他线性模型要好得多。GA-ANN 模型对于训练集、内部测试集和外部测试集的平均绝对相对误差（AARE）分别为 0.054、0.059 和 0.100。在 GA-AAN 模型上应用十重交叉验证程序得到了 Q²=0.943 的统计数据，这表明了该模型的可靠性。
