蛋白质二级结构预测的人工神经网络方法研究

本文比较了五种神经网络方法预测蛋白质二级结构的准确率,并做出初步评价。五种神经网络分别是:误差反传前向网络(BP),径向基函数网络(RBF),广义回归神经网络(GRNN),串并联叠层网络(CF),Elman网络(ELM)。结果显示:GRNN的预测准确率达85.7%,优于其它网络。本文还讨论了训练集样本数及参数的优化对GRNN预测准确率的影响。     

径向基函数神经网络用于重叠色谱峰解析

 在高斯基径向基函数神经网络 (RBFNN)学习算法中引入了鲁棒性和随机全局寻优的两阶段遗传算法 :结构学习和参数优化。通过两阶段学习算法的交替使用 ,使网络具有结构自学习和参数优化的能力 ,而后将网络应用于组分数未知的重叠色谱峰解析。该方法具有不需人为干预 ,可自动确定网络结构即组分数的优点 ;并且解析精度较高 ,适用于多组分重叠色谱峰的解析 ;对完全重叠色谱峰也具有良好的解析能力。     

近红外光谱-径向基神经网络在异烟肼片无损定量分析中的应用

应用异烟肼片粉末的近红外漫反射光谱数据分别结合偏最小二乘法(PLS)和径向基神经网络(RBFNN)建立定量分析模型,并用所建模型对预测集样品进行了预测,结果表明:应用RBFNN所建立的定量分析模型优于PLS模型,相关系数(r)值由0.99593提高到0.99734,交互验证均方根误差(RMSECV)值由0.00523下降到0.00423,预测均方根误差(RMSEP)值由0.00614下降到0.00501。     

Prediction of the activity concentrations of 232Th, 238U and 40K in geological materials using radial basis function neural network

In this paper, three individual models and one generalized radial basis function neural network (RBFNN) model were developed for the prediction of the activity concentrations of primordial radionuclides, namely, ²³²Th, ²³⁸U and ⁴⁰K. To achieve this, gamma spectrometry measurements of 126 different geological materials were used in the development of the RBFNN models. The results indicated that individual and generalized RBFNN models are quite efficient in predicting the activity concentrations of ²³²Th, ²³⁸U and ⁴⁰K of geological materials.

     

大鼠胰腺及癌组织红外光谱连续小波特征提取及径向基人工神经网络识别

采用水平衰减全反射(HATR)傅里叶变换红外光谱法(FTIR)测定了SD大鼠胰腺正常组织与非正常组织的谱图,提出了一种新的基于FTIR的连续小波特征提取与径向基人工神经网络分类方法以提高FTIR对早期SD大鼠胰腺癌的诊断准确率。利用连续小波多分辨率分析法提取FTIR特征量,对于提取的特征量采用径向基函数神经网络进行模式分类。对SD大鼠的胰腺正常组织、早期癌组织及进展期癌组织的FTIR,利用连续小波多分辨率分析法提取9个特征量,进行RBF神经网络分类判断。当目标误差为0.01,径向基函数的分布常数为5时,网络达到最优化,总的正确识别率为96.67%。并对影响分类结果的网络参数、目标误差和分布常数对分类样品的影响做了讨论。实验结果表明:此方法对早期胰腺癌具有较高的诊断率。     

Extending the working calibration ranges of four hexachlorocyclohexane isomers in gas chromatography-electron capture detector by radial basis function neural network

A radial basis function neural network (RBFNN) method was developed for the first time to model the nonlinear calibration curves of four hexachlorocyclohexane (HCH) isomers, aiming to extend their working calibration ranges in gas chromatography-electron capture detector (GC-ECD). Other 14 methods, including seven parametric curve fitting methods, two nonparametric curve fitting methods, and five other artificial neural network (ANN) methods, were also developed and compared. Only the RBFNN method, with logarithm-transform and normalization operation on the calibration data, was able to model the nonlinear calibration curves of the four HCH isomers adequately. The RBFNN method accurately predicted the concentrations of HCH isomers within and out of the linear ranges in certified test samples. Furthermore, no significant difference (p > 0.05) was found between the results of HCH isomers concentrations in water samples calculated with RBFNN method and ordinary least squares (OLS) method (R² > 0.9990). Conclusively, the working calibration ranges of the four HCH isomers were extended from 0.08-60 ng/ml to 0.08-1000 ng/ml without sacrificing accuracy and precision by means of RBFNN. The outstanding nonlinear modeling capability of RBFNN, along with its universal applicability to various problems as a “soft” modeling method, should make the method an appealing alternative to traditional modeling methods in the calibration analyses of various systems besides the GC-ECD.     

基于径向基神经网络和遗传算法的槽内式选择性超声电合成甲基苯甲醛

以混合二甲苯为原料, Mn(Ⅲ)为氧化剂, 硫酸溶液为电解质, 采用槽内式超声电合成甲基苯甲醛. 探讨了选择性电合成甲基苯甲醛的可能性, 通过径向基(RBF)神经网络和遗传算法(GA)对选择性电合成甲基苯甲醛3种异构体的比例、 电流效率与混合二甲苯的用量、 硫酸浓度和电流强度的关系建立预测模型, 并运用GA确定模型中RBF神经网络的目标均方误差(Goal)和径向基函数的分布(Spread). 然后根据预测模型, 使用GA对电合成条件进行优化, 分别获得了电合成产物中对位甲基苯甲醛占优、 邻位和对位甲基苯甲醛占优以及电流效率最高时的电合成条件. 当采用上述条件进行实验时, 模型给出的预测结果分别为: 对位甲基苯甲醛占优的质量分数可达90.01%, 邻位和对位甲基苯甲醛占优的质量分数为80.38%, 电流效率达到最高时的邻位、 间位和对位甲基苯甲醛的质量分数分别为16.80%, 8.43%和74.77%; 而与之相对应的实际实验结果分别为90.10%和79.91%, 以及17.20%, 8.49%和74.31%, 二者之间的最大相对误差小于±2.24%, 表明所建立模型的预测值与实测值基本吻合.     

Application of Near Infrared Diffuse Reflectance Spectroscopy with Radial Basis Function Neural Network to Determination of Rifampincin Isoniazid and Pyrazinamide Tablets

Partial least squares(PLS),back-propagation neural network(BPNN)and radial basis function neural network(RBFNN)were respectively used for estalishing quantative analysis models with near infrared(NIR)diffuse reflectance spectra for determining the contents of rifampincin(RMP),isoniazid(INH)and pyrazinamide(PZA)in rifampicin isoniazid and pyrazinamide tablets.Savitzky-Golay smoothing,first derivative,second derivative,fast Fourier transform(FFT)and standard normal variate(SNV)transformation methods were applied to pretreating raw NIR diffuse reflectance spectra.The raw and pretreated spectra were divided into several regions,depending on the average spectrum and RSD spectrum.Principal component analysis(PCA)method was used for analyzing the raw and pretreated spectra in different regions in order to reduce the dimensions of input data.The optimum spectral regions and the models' parameters were chosen by comparing the root mean square error of cross-validation(RMSECV)values which were obtained by leave-one-out cross-validation method.The RMSECV values of the RBFNN models for determining the contents of RMP,INH and PZA were 0.00288,0.00226 and 0.00341,respectively.Using these models for predicting the contents of INH,RMP and PZA in prediction set,the RMSEP values were 0.00266,0.00227 and 0.00411,respectively.These results are better than those obtained from PLS models and BPNN models.With additional advantages of fast calculation speed and less dependence on the initial conditions,RBFNN is a suitable tool to model complex systems.     

Prediction of quantitative calibration factors of some organic compounds in gas chromatography

A quantitative structure-property relationship (QSPR) methodology that involves multilinear (Hansch-type) and nonlinear (radial basis function neural network (RBFNN)) approaches was performed to correlate the quantitative molar calibration factors (f(M)) of 140 organic compounds against structural factors. The statistical characteristics provided by the multiple linear model (R(2) = 0.963; RMS = 0.089; AARD = 3.86% for test set) indicated satisfactory stability and predictive ability, while the predictive ability of the RBFNN model is somewhat superior (R(2) = 0.983; RMS = 0.075; AARD = 3.19% for test set). The multilinear model provided some insight into the main structure factors that modulate the quantitative calibration factor of the investigated compounds.     

近红外光谱无损定量分析吡嗪酰胺片

近红外光谱;径向基神经网络;吡嗪酰胺;定量分析