【作者】 茅晓延；

【导师】 朱永平；

【作者基本信息】 浙江大学 ， 卫生毒理学， 2005， 硕士

【摘要】 目的 化学致癌物/致突变物的定量结构-活性相关研究(Quantitative StructureActivity Relationship,QSAR)自八十年代以来一直是化学和研究领域的热点,并取得了一系列卓有成效的研究成果。N-亚硝基化合物是一类重要的哺乳动物致癌物,其构效关系并不是简单的线性关系,而是错综复杂的非线性关系。若使用传统的线性回归方法往往不能满足要求。所建模型较难契合化合物的实际构效关系。 人工神经网络(Artificial Neural Network,ANN)是近年来发展起来的模拟人脑生物过程的人工智能技术,具有较强的模拟多元非线性体系的能力,拟合效果非常好。其中应用最广泛的是反向传输神经网络(Back-PropagationNetwork,BP)。它具有强大的学习能力和泛化能力,可较好避免QSAR线性化带来的不足问题。 但传统BP神经网络存在收敛速度太慢、变量间存在相关、过拟合等缺陷。本研究旨在寻找成熟可靠的描述符来表征N-亚硝基化合物结构特性,探索改进BP神经网络的方法,并将其应用于N-亚硝基化合物致癌活性的QSAR研究。在同样样本集基础上,将本研究的预测和回判结果与朱永平的Fisher判别分析方法的研究结果进行比较。更多还原

【Abstract】 ObjectiveSince 1980’s, the chemical carcinogenicity/mutagenicity quantitative structure activity relationship (QSAR) studies have been a focus of attention in the chemistry and interrelated field, and have yielded exciting achievements. N-nitroso Compounds is a kind of important mammalian carcinogens. Its quantitative structure-activity relationship is not a simple linear relationship, but a pretty complex non-linear relationship. If we use traditional linear regression, it usually can’t satisfy the requirements, and the established model have difficulty in according with the actual structure-activity relationship.Artificial Neural Network (ANN), simulating biologic process of human brain, has being one of artificial intelligence technologies developed in recent years. ANN is accomplished in simulating multi non-linear system and have pretty good fitting effect. Back-Propagation Neural Network (BP) is widely used because of its strong learning and generalization ability.However, there are also disadvantages in traditional BP Neural Network, such as slow constringency velocity, correlation between variables, over-fitting. Theaim of our study was to find mature and reliable descriptor in token of structural characteristic of N-nitroso Compounds and apply modified BP Neural Network in QSAR study on carcinogenicity of N-nitroso Compounds. Additionally, we compared the result of our study with Zhu Yong ping’s result performed by Fisher’s discrimination analysis on the same sample.MethodsCarcinogenicity of N-nitroso Compounds data were extracted from Gold’s research report published in 5 plots. There are 94 carcinogens and 19 non-carcinogens in this study. Training sets includes 95 chemicals published in 1,2,3 plot. Test sets includes 18 chemicals published in 4,5 plot.We took 14 physiochemical parameters, i.e. critical pressure(Pc), critical temperature(Tc), Tb/Tc(0), critical compressibility factor(Zc), critical temperature(Tc), normal boiling point(Tb), the molal liquid volume at the normal boiling point(Vb), acentric factor(ω), heart of vaporization( A Hvb), liquid density( ρ L), vapor density( ρ v), liquid surface tension( σ ), parachor([P]) and thermal conductivity of organic liquid(λ) and 1 dummy variable as descriptor in token of structure characteristic of N-nitroso compounds.We modifed BP Neural Network by filtration parameters through partial correlation analysis, normalization data and determining hidden layer number, technical parameters infulencing constringency velocity,et al. Finnaly, we established and trained BP model to retrospective-predict and predict carcinogenicity of N-nitroso compounds. All program was develolped by MATLAB simulation software.ResultsOn basis of parameters filtration and normalization, we used training setsincluding 95 samples to successfully establish a BP Neural Network model which structure was 7:8:1. We repeated training the model. When "net.trainParam.epochs" reached 500, and "net.trainParam.goal" reached 0.0009, we stopped training and take for success.We applied successful trained BP Neural Network into retrospective-predicting training sets. The accuracy of retrospective-prediction of training sets was 98.95%(94/95), and percentages of true positive, false positive, true negative was respectively 98.72%(77/78), 0、 100.00%(17/17),l .28%(1/78).Similarly, we appllied the model into predicting testing sets including 18 samples. The accuracy of prediction was 100.00%(16/16), and percentages of true positive, false positive, true negative was respectively 50.00%(l/2), 50.00%(l/2), 0(0/16).On basis of same training and test sets, our result was compared with Zhu Yong ping’s result predicted by Fisher’s discrimination analysis. It shows that the accuracy, false positive and true negative of retrospective-prediction in BP Neural Network is better than that of Fisher’s discrimination analysis, and have statistical significance in Chi-square test ( x2=15.71, P<0.005) .But true positive and false negative have no statistical significance differences in two studies. There was also更多还原