【作者】 苏怡；

【导师】 周鲁；

【作者基本信息】 四川大学 ， 物理化学， 2002， 硕士

【摘要】 合理药物设计需要科学地预见化合物的药效学和药动学性质，以往的新药设计过分注重药效学这一方面而忽视了药动学方面。使得许多药物因药动力学因素不理想的而被淘汰掉，造成巨大的浪费。药物的体内过程与药物的许多理化性质及结构因素相关。众所周知，药物血浆蛋白结合率是一个很重要的药代动力学因素。本文应用基于MATALAB的人工神经网络方法对61种抗生素进行了关于药物血浆蛋白结合率的定量结构-药物动力学关系研究(Quantitative Structure Pharmacokinetic Relationship,QSPR)。首先，建立了合理的含一个隐含层的BP神经网络。输出层为实验获得的药物血浆蛋白结合率；输入层为计算获得的药物分子的量化参数，理化参数和分子连接性指数等，共19个参数。输入和输出层数据均标准化到0～1之间。随机挑选51个样本为训练集。通过多次试错尝试隐含节点数确定为30；最大迭代次数为500,000次；其它参数确定为：学习速率1r=0.1：学习速率的正增长率为1r(i)=1.08；学习速率的负增长率为1r(d)=0.9；学习动量因子M=0.9。其次为了验证网络，本文进行leave-one-out分析。从训练集中移去一个化合物，以剩下的50个样本作为训练组，将网络训练好后再预测被移去的化合物的血浆蛋白结合率，重复计算51次。计算结果 四川大学硕土学位论文rt—0．9834，s—0刀3823。结果证明所建网络能够有效地进行药物血浆蛋白结合率与其结构的相关性分析。最后利用挑选剩下的10个样本进行预测，10个药物的血浆蛋白结合率的预测值能很好地与实验数据相吻合，研究结果表明本文所建立的神经网络模型对于研究药物血浆蛋白结合率的QSPR是合理有效的。也证明了人工神经网络功能强大，将成为药物 QSPR研究的一个有效的工具。更多还原

【Abstract】 Rational drug design requires reliable prediction of both pharmacokinetic and pharmadynamic properties. Many candidate drugs were eliminated through selection because of their bad phamacokinetic property, which caused great waste. Previously the pharmaceutical researchers took much attention on the facet of pharmacodynamic. However, the pharmacokinetic was overlooked. As well known, drug plasma protein binding is an important pharmacokinetic process. So this study demonstrates the application of neural network to research the QSPR of sixty-one antimicrobial agents. Firstly, the neural network used in our study is a three-layer back propagation system network including an input layer, a hidden layer and an output layer. The input layer of the neural network is compromised by the quantum chemistry parameter, physical-chemistry parameter and molecular connectivity index. The number of all the parameter is nineteen. The output of the neural network is the pharmakokinetic property -drug plasma protein binding(DPPB) of antimicrobial agents which is derived from the experimental data .All the input data and output data are normalized between 0.1 to 0.9. Fifty-one sample are randomly selected as training pair. By trial and error, the number of the hidden neurons isdetermined as thirty; The max-epoch for the training network is 500,000; In addition, other parameter of the neural network is that: learning rate 1r=0.1; increased frequency of the learning rate lr(i)=1.08; decreased frequency of the learning rate lr(d)=0.9;momentum=0.9. Secondly, the prediction ability of the network is tested by the leave-one-out method. In this procedure one compound is omitted from training set. After training, the input values of the left-out compound are fed into the network and an output (DPPB) is obtained by repeating the experiment for all the member of the training set one at a time. The computational result demonstrates that neural network is effective for the analysis of the relativity between the structure of the compound and their plasma protein binding. Finally, the test involved prediction of the pharmacokinetic properties(DPPB) of ten compound never seen by the network. The neural network predicted values show good agreement with the experimental values. This result indicates that the neural network described in this study is proper and effective for QAPR research on the drug plasma protein binding. Furthermore, neural networks are proved roust, flexibility and fitting to solve the complex nonlinear problems including QSPR study. The result also indicates that neural networks can be a powerful tool in exploration of quantitative structure-pharmacokinetic relationship.更多还原