【作者】 李剑；

【导师】 陈德钊；

【作者基本信息】 浙江大学 ， 化学工程， 2005， 硕士

【摘要】 药物定量构效关系(Quantitative Structure-ActivityRelationships,QSAR)的研究对象是药物生理活性和分子结构参数间的量变规律。建立适当的数学模型对QSAR研究来说是非常重要的,通过合适的模型可以预测未知化合物的活性,探讨药物作用机理,对指导新药的设计与合成有重要意义。为此,本文提出两种新的非线性建模方法,并将它们应用于药物QSAR建模,结果显示这两种模型有良好泛化性能和稳定性。以下是本文的主要内容:(1) 分析了循环子空间回归(Cyclic Subspace Regression,CSR)的特点,并介绍了CSR的两种稳健算法。而后将径向基网络(Radial Basis FunctionsNetworks,RBFN)与循环子空间回归相结合构成了RBF-CSR方法,该方法既保留了神经网络的结构,又用数学方法直接求解,免去了人工神经网络冗长的训练过程,并可在更广的空间内寻找最优的网络参数,很好的解决了RBF网络训练量大,难以找到最优解的问题。(2) 针对简单遗传算法易早熟、全局寻优效率偏低等特点,设计了一种基于优进策略的遗传算法(Eugenic evolution Genetic Algorithm,EGA),在演进过程中自适应地改善个体地分布,适时引入确定性操作,以改善常规遗传算法的性能。并将EGA用于RBF-CSR网络的参数优化,构建了EGA-RBF-CSR方法。最后将EGA-RBF-CSR方法用于苯已酰胺类除草剂的定量构效关系建模,取得了良好的效果。(3) 回顾了统计学习理论研究的基本问题及主要内容。而后回顾了支持向量机(Support vector machine,SVM)原理和实现,简要介绍了国际上常用的各种SVM优化算法。介绍了最小二乘支持向量机(Least squares support vector machine,LSSVM)及其算法,该方法是标准支持向量机在二次损失函数下的一种扩展形式,它保留了SVM的一些优点,同时具有计算量小、求解速度快等特点。(4) 提出了一种基于最小二乘支持向量机的非线性偏最小二乘建模算法。该算法采用非线性迭代偏最小二乘(Nonlinear Partial Least Squares,NIPALS)方法从样本数据中成对地提取最优成分,使多变量系统降维,然后应用基于结构风险最小化准则的最小二乘支持向量机,对提取的成对成分进行非线性拟合,并加入基于误差的权值调整策略,构成EB-LSSVM-PLS模型。最后,将该方法应用于敏浙江大学硕士学位论文小样本的新型黄烷酮衍生物的QSAR建模,取得了满意的效果。 文章最后对所做的工作进行了总结,并在此基础上,提出了今后的研究发·展方向。关键词:径向基网络,循环子空间回归,定量构效关系,遗传算法,优进策略,参数优化,最小二乘支持向量机,偏最小二乘,基于误差修正,泛化性能谷更多还原

【Abstract】 Modeling is to abstract qualitative relationships of variables involved in objects, and is an important means of recognizing and describing the behavior of objects under studying. Many techniques are involved in modeling methods such as optimization, statistics, artificial intelligence, pattern recognition, machine learning, neural network, and so on.Building a proper model is important to Quantitative structure-activity relationships(QSAR) research. In this paper, two new approach were proposed and used to QSAR modeling. The two mothod are EGA-RBF-CSR and EB-LSSVMPLS.Compared to the other models, This two models not only holds on fine learning ability but also gives better prediction performance and steady capability.This article includes the following parts mainly:(1) By changing the dimension of solution space, cyclic subspace regression(CSR)gets a serial of regression models which contains least square regression,principal component regression,partial least squares regression and other medial regression outcomes.Through the cyclic subspace regression(CSR) being applied to determine the weight of Radial Basis Functions(RBF) networks,the RBF-CSR model is produced. The approach has the merit of RBF-PLS,and it can select the optimal model in wider range.(2)A eugenic evolution strategy was proposed to improve the efficiency of the conventional simple genetic algorithm (SGA) searching. The Eugenic evolution genetic algorithm (EGA) collects the population information along the evolution of children generations and constructs a deterministic optimization algorithm, which will be embedded in the evolution process at appropriate stage to speed up the local searching. The EGA was successfully applied to optimize the RBF-CSR parameter and EGA-RBF-CSR model was proposed. Finally, EGA-RBF-CSR was applied successfully to modeling quantitative structure-activity relationships.(3)Research contributions and major problems in statistical learning theory study(SLT) and support vector machine are reviewed.Then least squares support vector machine(LSSVM) was introduced. LSSVM is a modified version of SVM.The quadratic programming in original support vector machines is replaced by the linear set of equations.The approach has the merit of SVM, and it can compute fastly and stably.(4)A new nonlinear partial least squares algorithm embedded least squares support vector machine (LSSVM) into the regression framework of partial least squares (PLS) method was proposed in this paper. In this approach, LSSVM is used to fit the nonlinear inner relations between PLS components, thus a multi-input multi-output nonlinear modeling task was decomposed into linear outer relations and simple nonlinear inner relations that were performed by a number of single-input single-output LSSVM models. By using the universal approximation property of LSSVM, the PLS modeling method is generalized to a non-linear framework. Subsequently, to increase PLS components interpretative capability, the error-based weights updating procedure in the PLS input outer model was deduced and implemented in the LSSVMPLS regression framework. Finally, the EB-LSSVMPLS is applied to modeling QSAR.Finally,a brief review of this thesis is given.Some future research directions are highlighted.更多还原