用量子化学方法预测PCDFs的Ah受体结合能力(1)——神经网络模型

Predicting Ah Receptor Binding Data of PCDFs Using Quantum Chemical Method (1)——Artificial Neural Network Approach

利用AM1半经验量子化学方法计算135种多氯联苯并呋喃(PCDFs)的12种结构-性质参数,以30种已有文献值的PCDFs为训练样本,采用误差反向传播人工神经网络(BP-ANN)方法建立PCDFs芳烃(Ah)受体结合能力的定量结构活性关系(QSAR)方程,对训练样本集以外的3个测试样本的检验结果表明,模型具有相当高的精度(训练样本最大相对误差小于5％,检验样本最大相对误差小于7％),且误差频数符合正态分布,表明该模型可用于未知样本的定量预测,据此给出目前尚没有文献值的其余102种PCDFs和Ah受体结合能力的预测值。

Twelve kinds of quatum chemical parameters for all 135 PCDFs are calculated as structural descriptors using AMI semiempirical method. Based on these descriptors and observed Ah receptor binding data of 30 PCDFs in training sample set, a back-propagation artificial neural network is established. The model shows that a high accuracy with the maximum relative error is less than 5 % for training samples and 7 % for test samples,and the distribution of error frequency follows the normal distribution. Results show the applicability of the model in predicting the samples outside the training sample set. Ah receptor binding data of 102 PCDFs are given.