论文部分内容阅读
运用三层误差反向传播网络对５１种胺类有机物进行了结构－毒性关系的研究．选入的结构参数为分子连接性指数（０Ｘｖ，１Ｘｖ），信息理论指数（ＲＩＣ，ＳＲＩＣ，ＲＣＩＣ）及分子量等６种均可通过分子拓扑图直接计算获得的指标．毒性参数选用大鼠经口ＬＤ５０，根据其大小将样本分为３类：高毒、中毒、低毒，在神经网络模型上作出差别归类，并分别对每类进行定量预测．分类预测正确率为９０％，定量预测标准差分别为０．０８３、０．１０８、０．１１６．结果表明：神经网络对急性毒性ＬＤ５０具有良好预测效果，大大优于多元回归分析和判别分析．同时本文还讨论了影响网络性能的一些因素，提出了相应改善措施．
The structure-toxicity relationship of 51 amines was studied by using three-layer error propagation network. The selected structure parameters are the molecular connectivity index (0Xv, 1Xv), information theory index (RIC, SRIC, RCIC) and molecular weight of six kinds of molecular topological map can be directly calculated indicators. Toxicity parameters rat oral LD50, according to the size of the sample will be divided into three categories: high toxicity, poisoning, low toxicity, in the neural network model to make a differential classification, and each class were quantitatively predicted. The accuracy of classification prediction was 90%, and the standard deviation of quantitative prediction were 0.083,0.108,0.116 respectively. The results show that neural network has a good predictive effect on acute LD50, which is much better than multivariate regression analysis and discriminant analysis. At the same time, this paper also discusses some factors that affect network performance, and puts forward corresponding improvement measures.