Development Of Artificial Neural Network Software Package And Its Application To Pharmaceutical Analysis

A MS-Windows based artificial neural network (ANN) software package was programmed with Visual Basic 6.0. Two programs compose this software package. One is Neural Computing Software (NCS), which providing some popular neural network computing capabilities, such as back-propagation artificial neural network (BP-ANN), self-organization feature mapping, Hopfield network, and so on, another program, namely BPManager, was coded combining with Access database which supplying straightforward management to original data, training results and other important output of BP-ANN. BPManager can also employ some useful work such as optimizing hidden node number, searching for the best network condition and simulating forecasting.Several projects relating to pharmaceutical analysis were conducted with this software package.BP-ANN was used to model and to predict the solutes migration time of vitamin B1 and other seven drugs according to both the varied experimental voltage and ionic strength of background electrolyte in capillary zone electrophoresis. "Leave-One-Out" method was applied and results show that neural networks fit the complex non-linear relation between experimental conditions and migration time of capillary zone electrophoresis well. NeuralABSTRACTnetwork modeling indicates better performance than regression modeling. BP-ANN was also employed to optimize the experimental conditions of capillary zone electrophoresis satisfactorily.BP-ANN and Principal Component Analysis (PCA) coupled with BP-ANN were applied in the recognition of Fufang Gancao Pian manufactured in six factories, more than 90% of the samples were clustered without error, which indicates strong capability of neural network in complex pattern recognition.Benzyl penicillin potassium and other twenty-four antibiotic drugs were recognized using neural networks. A simple method was employed in coding the infrared spectrum of these drugs, and BP-ANN was trained and then used to recognize the drugs. All drugs were successfully recognized, moreover, networks showed good compatibility capacity when perturbation was made on one input node intentionally.