【作者】 左之利；

【导师】 周鲁；

【作者基本信息】 四川大学 ， 物理化学， 2002， 硕士

【摘要】 计算机辅助药物设计的主要目标就是为了从理论上预建可与机体重要功能分子(蛋白质、核酸、酶、离子通道等)发生作用的化学物质，达到治疗疾病的目的。本文利用神经网络方法，进行了药物动力学参数-生物利用度与药物理化和量化参数之间的相关性分析。我们选用了已知生物利用度的不同系列的39种药物，分别计算了它们理化参数：分子的体积(V)，分子折射率(R)，疏水性参数(LgP)，水合能(H)，分子极化度(P)，量化参数：前线轨道的能量EHOMO和ELUMO，共七种参数。神经网络为三层动量BP网络，最大迭代次数为500,000次，隐含层节点数为36，初始学习速率为0.11，学习增长率为1.07，学习减少率为0.8，学习动量为0.9。为了测试网络的预测能力，也进行了Leave-one-out分析。首先从样本组中移去一个化合物作为预测组，以剩下的38个化合物作为训练组，将网络训练好后，再预测被移去的化合物的生物利用度。这样重复39次，得到整个数据组的预测生物利用度。我们还计算了预测结果与生物利用度平均值的残差和预测结果与生物利用度的残差。预测结果与生物利用度的残差小于0.05的样本共29个，占总样本数的74.36％；残差小于0.10的样本数为33个，占总样本数的84.62％；残差小于0.20的样本数共36个，占总样本数92.31％。这些结果证明，神经网络方法能对药物的生物利用度与药物 四川大学硕士学位论文理化参数和量化参数的关联关系提供比较可靠的预测，同时，计算结果也说明我们所选择的七个参数与药物的生物利用度有良好的相关性。更多还原

【Abstract】 Computer Aided Drug Design majors in establishing chemical compounds in advance through theory in order to cure ailments. The compounds can interact with the important functional molecules in the body, such as protein, nucleic acid, enzyme and ionic channels, et al. This study demonstrates the application of neural networks to correlative analysis between bioavialability and parameters of physicochemistry and quantum chemistry, such as volume(v), refraction(R), lipophilicity(LgP), polarization(P), the frontier orbital energy EHOMO and ELUMO About 39 different sets of compounds with bioavialabilities which were derived from experiment by others are chose, and the according parameters are calculated. BP neural network with three layers is designed. The max-epoch for the network is 500,000. The number of hidden neurons is 36. The learning rate is 0.11. The increasing learning rate is 1.07. The decreasing learning rate is 0.8. The learning momentum is 0.9. To test the predicting ability of network, the analysis of Leave-one-out is carried out. The bioavialability of predicting group is predicted when the training group is trained well. All the predicted bioavialability are acquired after repeat 39 times. The deviation between predicted bioavialability and theaverage value of experiment, and deviation between predicted bioavialability and the value of experiment are calculated. The number of whose later deviation below 0.05 is 29, about 74.36% of all. The number below 0.10 is 33, about 84.62% of all. The number below 0.20 is 36, about 92.31% of all. All the results prove that the neural network can provide reliable prediction of the correlation between bioavialability and various parameters of drugs, and that the seven parameters we choose are well correlate with bioavialability.更多还原