QQ客服 官方微博 反馈留言

题名	改进的人工神经网络方法在药物构效关系研究中的应用
作者	张晓晨
学位类别	硕士
答辩日期	1994-06-30
授予单位	中国科学院研究生院
导师	周家驹
中文摘要	构效关系是药物分子设计的核心问题。人工神经网络方法（ANN）是计算机辅助药物分子设计软件中的一种人工智能方法，近几年在构效关系研究中提到应用，并取得了令人鼓舞的结果。本论文分析了人工网络适用于构效关系研究的原因，并从前人的应用中发现应用人工神经网络建模过程中容易发生的问题。这些问题主要是：（1）过拟合现象；及（2）人工神经网络建模中固有的不利之处，即人工神经网络模型通过连接权重表达自变量与因变量之间的关系，使得模型形式复杂，不便于分析自变量与因变量之间的因果关系。为了解决上述存在的问题，本论文从以下三个方面开展工作：1. 从网络输入住处的筛选，网络合结构的建立－隐节点数的合适选取，及利用交叉预报得到网络训练停止判据三个方面，解决人工神经网络在建模过程中出现的过拟合问题。2. 利用图示的方式将生物活性同每一个变量的关系表示出来。在考虑生物活性同某一变量关系的时候，其他变量分别固定在各自范围值的30,60或90％的数值上。这种作法尽管只是一种局部的表示方法，但是它能够较为直观地得到生物活性与每个物化参数的关系，而且也表示了变量共同作用对生物活性的影响。3. 作为计算机化学开放实验室药物分子设计软件系统的一个组成部分，研制了以防止过拟合交叉预报程序（CVANN）为核心的人工神经网络建模方法软件。本论文利用改进的人工神经网络模型对三个体系（包括定量和定性两类问题）进行了研究，均得到了令人满意的预报和分类结果，并将定量分析结果与Hansch分析结果进行比较，从而进一步分析和讨论了人工神经网络的优越性，证明了本论文所提出的三个改进方案的正确性和可行性，也说明了人工神经网络方法在药物分子设计中具有光明的前景。
英文摘要	Structure-Activity Relationship (SAR) is the key point of molecular design, understanding of SAR allows chemists to predict the bio-activity of a compound with the information obtained from chemical structure. As an artificial intelligence method in the molecular design software CASAC (Computer Aided Screening Agricultural Chemicals), Artificial Neural Network (ANN) shows attractive prospect in development of SAR models. Unlike the common statistic algorithms, regression and pattern recognition, Ann is powerful on modeling highly nonlinear relationships. In the last decade, the application of ANN in the field of SAR has had a rapid growing. It is demonstrated that this new technique is often superior to traditional statistic methods both on quantitative and qualitative problems. In this thesis, the reasons why ANN is suitable to SAR research are discussed, and some problems which occur in the ANN modeling are emphasized. These problems mainly are: (1) overfitting phenomena often prevents ANN to provide a reasonable prediction for the unknown compounds; (2) the ANN intrinsic shortcoming makes the model too complicated to obtain an explicit expression of the bio-activity and the physiochemical parameters. In the following, some improvements have been made to solve these problems. Firstly, improvements have been made on the scheme of ANN input variable screening, hidden node determination for network optimization and cross-validation convergence criterion. These improvements ensure the ANN model avoid the overfitting and provide a reliable prediction for the bio-activity of designed molecules. Secondly graphic representation is applied as a method in the SAR research, although no explicit expression can be obtained from the ANN model. Bio-activity can be achievable graphically by taking the network output as a function of each independent variable, while the rest of the variables are held at 30, 60 or 90% of their own responding ranges in the data set. The graphic representation depicts the variation of activity with one property while th others are held constant, and also manifests the strong intervariable coupling. Finally, an user-friendly ANN software, with core program CVANN, is developed. Based on the above-mentioned improvements, this software can be applied to develop the ANN models for highly nonlinear relationships in molecular design, with reliable prediction results. Three systems of compounds, including both quantitative and qualitative bio-activity data, are studied using the improved ANN method. The ANN models show good results compared with the regression methods. Discussions are give on the advantages and limitations of the ANN method, and also on the correctness and the feasibility of the improvement of the ANN method in this thesis. A bright future of ANN application in the field of molecular design is expected.
语种	中文
公开日期	2013-10-23
页码	62
内容类型	学位论文
源URL	[http://ir.ipe.ac.cn/handle/122111/3893]
专题	过程工程研究所_研究所（批量导入）
推荐引用方式 GB/T 7714	张晓晨. 改进的人工神经网络方法在药物构效关系研究中的应用[D]. 中国科学院研究生院. 1994.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。