【作者】 余杰；

【导师】 程翼宇； 吴永江；

【作者基本信息】 浙江大学 ， 生物化工， 2002， 硕士

【摘要】 中药材质量分析信息处理方法研究近年来正成为化学、生物及药学界的一个研究热点，吸引了越来越多的关注。由于仪器分析等辅助实验手段的发展和推动，中药组成及药效数据急剧扩增，且具有高维、小样本、多变量、非线性、交互作用强等特点。本文针对复杂的中药分析体系，将化学信息学的思想和方法应用于中药材质量分析信息处理研究中，各类计算技术包括神经元网络、小波变换、遗传算法、进化学习、支持向量机及可视化技术等被有机地综合运用，以有效地解决中药质量鉴别评价难题。本文涉及的工作主要有以下几个方面： (1)针对中药材定性分类鉴别问题，引入了机器学习领域中最新的支持向量机技术，并将其与决策树算法相结合，将二类别支持向量分类器拓展到多类别情况下，从而应用于中药材多个产地、等级的分类鉴别，结果明显优于传统的统计模式识别方法及神经元网络方法，特别是有效克服了神经元分类严重“过拟合”的缺陷，具有较强的推广和泛化能力，适合于高维小样本的中药分类鉴别。 (2)将中药组效关系研究从定性进一步拓展到定量的层次上，将西药定量构效关系的概念引入到中药研究中，提出了中药定量组效关系建模的新方法，把神经元网络同遗传算法、进化学习策略集成并进行改进和优化，并成功地应用于中药组效关系建模中，与单一神经元网络及遗传算法相比，该方法具有更优的训练、预测精度及模型可信度，可实现中药药效的定量智能预测。 (3)针对中药成分分析数据剧增，解析难度增大，将科学计算可视化技术引入到中药化学分析数据处理中，发展了一类仪器分析数据可视化新方法，采用小波(包)变换、核主成分分析等计算方法对药材进行隐含化学指纹特征抽提，再可视化表征为基于二维灰度图的虚拟指纹图谱，从而将抽象分析量测数据变换为直观的计算机图像，实现中药材质量的整体直观鉴别与评价。该方法被成功应用于中药色谱、光谱分析数据处理，具有常规分类鉴别方法不可替代的优点。更多还原

【Abstract】 In recent years, research on the methods of information processing for quality analysis of Traditional Chinese Medicine (TCM) has become an active area in the fields of Chemistry, Biology and Pharmacy, and gained increasing attention. With the development of experimental techniques like modern instrumental analysis, the data sets of chemical composition and drug effect of TCM are multiplying rapidly. Furthermore, they have indicated the characteristics of high dimension, small samples, multiple variables and strong interaction. Aimed at such complex analytical systems, the ideas and approaches in Chemoinformatics were introduced into the information processing of quality analysis for TCM. A variety of computing techniques, including artificial neural networks (ANN), wavelet transform (WT), genetic algorithms (GA), evolutionary learning, support vector machines (SVM) and visualization, were ingeniously integrated to deal with the challenging issue in the identification of TCM. The main contents of my thesis are listed as follows:(1) Aimed at the qualitative identification of TCM, the up-to-date technique in the area of machine learning, namely support vector machines, was incorporated with decision tree (DT) algorithm and extended to multi-class cases. Compared with traditional statistical pattern recognition paradigms and neural networks, the combined SVM-DT strategy can effectively overcome the drawback of ANN to over-fit and has powerful capability of generation. Its successful applications in the classification of geographic origins and quality grades for TCM have proved its great potential for the identification of high-dimensional small samples.(2) The QCAR investigation of TCM was further extended from the qualitative to quantitative level. The concept of quantitative structure-activity relationship (QSAR) in western medicines was introduced into the research of TCM, and a novel QCAR modeling method was further proposed. Through the integration of ANN, GA and evolutionary learning, the new strategy was successfully employed to the QCAR modeling of TCM. The results showed that itis superior to single ANN or GA approach in training accuracy, predicting accuracy and model reliability. Therefore it provides a new effective tool for the intelligent prediction of pharmacological activities of TCM.(3) Due to the increased quantity of analytical data for TCM and the difficulty of data interpretation, visualization technique was adopted to the measured analytical data processing. A novel analytical data visualization method was developed, in which the hidden chemical features were extracted using WT and kernel principal component analysis (kPCA), respectively, and then transformed into the virtual fingerprint by the two-dimensional grayscale images. In this way, the abstract data sets can be represented as computer-based images and the medicinal samples may be identified and appraised visually. The successful applications of this method in the data processing of chromatogram and spectrogram demonstrated its significant advantages over canonical identification approaches.更多还原