人工智能算法在生物药剂学中的应用

【作者】 陈敏燕；

【导师】 梁文权；

【作者基本信息】 浙江大学 ， 药剂学， 2008， 硕士

【摘要】 人工神经网络（Artificial Neural Network,ANN）是对人脑若干基本特性通过数学方法进行的抽象和模拟,是一种模仿人脑结构及其功能的非线性信息处理系统。BP（Back-Propagation）网络是一种采用反向传播学习算法的神经网络,对于在任意闭区间内的一个连续函数都可以用单隐层的BP网络逼近,一个三层的BP网络就可以完成任意的n维到m维的映射。遗传算法（Genetic Algorithm,GA）以生物进化过程为背景,模拟生物进化的步骤,将繁殖、杂交、变异、竞争和选择等概念引入到算法中,通过维持一组可行解,并对可行解进行重新组合,改进可行解在多维空间内的移动轨迹或趋向,最终走向最优解。它克服了传统优化方法容易陷入局部极值的缺点,是一种全局优化算法。本文第一部分将神经网络应用到药物透过人皮肤的渗透性能预测研究中。采集两组数据样本,用Matlab自编程序分别建立了预测药物经皮渗透速率BP网络和预测经皮渗透系数的BP网络模型。两个模型的输入变量相同,为药物的分子量（Mr）,油水分配系数(logK_ow),供氢数（Hd）和受氢数（Ha）。输出变量分别为药物经皮渗透速率(logJ_max)和经皮渗透系数（logKp）。前者模型测试组预测logJ_max与实际logJ_max的相关系数平方为0.997,回归直线的斜率为0.970。后者模型测试组预测得到的logKp与实际logKp的相关系数为0.95,均方误差平方根（RMSE）为0.37。本研究表明BP神经能够很好地预测药物透过皮肤的渗透性能。第二部分旨在建立一个生物药剂学分类系统（BCS）预测模型。药物在BCS中的分类取决于其溶解性和渗透性。故首先建立药物特征溶解度模型和绝对生物利用度模型。用多元线性回归和神经网络建立溶解度的线性模型和非线性模型,通过模型预测值与实际值的相关系数（R）、均方误差平方根（RMSE）和AIC比较,神经网络模型比逐步线性回归模型优越。用神经网络建立两种预测药物口服绝对生物利用度的模型。前者的输入变量为7个已能明确解释的理论参数,后者的输入变量有15个,包括上述变量和用遗传算法筛选出来的8个变量。通过比较,后者的预测效能优越。采集16个BCS分类明确的药物样本,用自行建立的溶解度模型和生物利用度模型预测分类,溶解度预测准确度为93.8%,渗透性预测准确度为81.2%,生物药剂学分类系统预测准确度为75.0%。第三部分将神经网络应用到药动学参数的预测,用BP网络预测新生儿丁胺卡那霉素消除速率常数。建立梯度下降BP网络（GD-BP-NN）,贝叶斯标准化BP网络（BR-BP-NN）和遗传BP网络（G-BP-NN）,以23例新生儿静脉滴注丁胺卡那霉素相关临床资料为研究对象,研究新生儿胎龄、日龄和体重对丁胺卡那霉素消除速率的影响,预测患者的消除速率常数,并比较三种网络的预测精度和运行效率。结果表明GD-BP-NN,BR-BP-NN和G-BP-NN测试组预测值和实验计算值的相关系数分别为0.92,0.91和0.98;测试组均方误差平方根（RMSE）分别是0.020,0.024和0.010;相同的预测精度条件下,GD-BP-NN,BR-BP-NN和G-BP-NN分别运行了2000,219和82步。从中可以看出遗传算法对BP神经网络权值和阈值进行优化,从而克服了BP神经网络训练速度慢,容易陷入局域极小和全局搜索能力弱等缺点,故G-BP-NN预测效果更好。更多还原

【Abstract】 Artificial neural network（ANN） is a type of mathematical model that simulates the biological nervous system and draws on analogues of adaptive biological learning.It is known to be a powerful tool to simulate various non-linear systems.The most popular ANN architecture is the multiplayer perception that generally trains the input-output relationship using a back propagation of error algorithm.Genetic algorithm（GA） is a kind of search and optimized algorithm that has been produced from stimulated biologic heredities and long evolutionary processes of creatures.Three elemental operators of the genetic algorithm are selection,crossover,and mutation.GA is capable of global optimum searching and it is a parallel process to population change and provides intrinsic parallelism.In the first section,Back-Propagation（BP） neural network was used to predict transdermal flux(1ogK_ow) and permeation coefficient（logK_p） of the compounds through the human skin.Two ANNs had different output neuron.However,the input neurons were the same,namely molecule weight（Mr）,octanol-water partition coefficient (logK_ow),hydrogen-bond donor（Hd） and hydrogen-bond acceptor（Ha）.For the first model,the correlation coefficient with predicted logJ_max and actual logJ_max of the testing drugs was 0.997.For the second model,the correlation coefficient between actual logK_p and calculated logK_p and root mean squared error（RMSE） of the testing set were 0.95 and 0.37.It demonstrated that BP neural network was of great help in predicting human skin permeability.The second section was aimed to establish a model for predicting biopharmaceutics classification system（BCS）.Solubility and permeability of the drug decide which BCS class it belongs to.First intrinsic solubility model and absolute bioavailability model were set up.Linear algorithm and ANN were applied and compared for the solubility model and the later was chosen for higher coefficient（R）,lower RMSE and lower AIC. For human oral bioavailability model,two ANNs were established.The main difference was that the second had additional eight input parameters chosen by GA,besides the seven input parameters of the first ANN which can be explained theoretically.And the second one was selected for preferable predicting efficiency.In the BCS prediction stage, sixteen durgs were chosen and self-established solubility and bioavailability models were used for prediction.The prediction accuracy of solubility,bioavailability,and BCS class were 93.8%,81.2%and 75.0%,respectively.In the third section,BP neural network was used for predicting the elimination rate constant of amikacin in neonates.Gradient descent backpropagation neural network （GD-BP-NN）,bayesian regularized backpropagation neural network（BR-BP-NN） and genetic backpropagation neural network（G-BP-NN） were established.The data of amikacin serum concentrations and clinical information of 23 neonates were used to train, validate,and test the models,the effects of gestational age（w）,postnatal age（d） and body weight（kg） were analyzed.The prediction precision and running efficiency were compared between the three models.Correlation coefficient of experiment calculative data and predicted data with GD-BP-NN,BR-BP-NN and G-BP-NN for the testing set were 0.92,0.91 and 0.98,respectively.Root mean squared error（RMSE） were 0.020, 0.024 and 0.010.Under the same prediction precision,running epochs of GD-BP-NN, BR-BP-NN and G-BP-NN were 2000,219 and 82.G-BP-NN is much better in the prediction of elimination rate constant of amikacin in neonates.It can overcome the weakness of BP-NN,such as the slow training speed and the vulnerability to local area pole smallness.更多还原