Both classification problems and quantitative structure-activitity relationship (QSAR) analyses of drugs have been better studied by the multi-layer feedforward neural networks with back-propagation learning algorithm (BPNN), than the classical methods in pattern recognition such as linear multiple regression (LMR) or adaptive least square method (ALS), since those problems often involve nonlinear relationships between descriptors and the class (/activity).1,2) BPNN is surely powerful and interesting approach, however, it has the defects such as the problem of local minimum, overfitting etc.3,4) Recently Bayesian regularized neural networks (BRNN), 5,6) has been successfully applied for QSAR studies 7,8) even for massive sample data, 9) but this excellent framework is rather complicated and needs much computation time in general.Kohonen developed the unsupervised learning algorithm that simplified the mapping mechanism of the relatively homogeneous structures found in mammalian brains associated with the processing of sensory data, 10) and showed it possible to generate self-organizing map (SOM) of data. It is an essential characteristic of Kohonen's SOM that has the ability to project high-dimensional data onto two-dimensional visualized map which is suitable for easy analysis while preserving the most significant information. Therefore, SOM is applied to extensive problems that characteristics of complicated data structure in high-dimensional space can be grasped visually from two-dimensional maps, and plays an important role to analyze inherent data structure.The advantage of the SOM, compared with some other projection methods is that the algorithm is very simple, straightforward to implement, and fast to compute. In the field of pharmaceutical sciences, the SOM was applied for searching useful drugs. Anzali at al. generated SOMs as a two-dimensional representation of molecules to analyze the shape and surface properties of those three-dimensional molecules responsible for biological activity. After visual comparison of SOMs, they discovered the benzothiadiazole group as a surrogate for methylendioxyphenyl. 11,12) Tetko et al. used SOMs to compress the so many input CoMFA data in their three-dimensional QSAR studies. 13) In the former study, we obtained the SOM of norbornann derivatives and of carbonyl compounds, and showed that the classification of them into some groups was successfully achieved according to the clustering appeared in the gray map. 14)In this paper we use the SOM as a method to predict missing activity in the QSAR studies of carboquinone and benzodiazepine, somewhat different usage of SOM rather than the standard compression or visualization tool of data. Results of the calculation indicate that SOM is considered to be one of useful methods in QSAR study.The Generation of SOM and the Prediction of an Activity Kohonen's neural network consists of two layers, the first layer is the input layer of n-dimension and the second is the competitive layer where every neuron has ncomponents correspon...

Both classification problems and QSAR analyses of drugs have been better studied by the multi-layer feedforward neural networks with back-propagation learning than the classical methods in pattern recognition such as linear multiple regression or adaptive least square method, since those problems often involve nonlinear relationships between descriptors and the class (/activity). [1][2][3] Recently Bayesian regularized neural networks (BRNN), [4][5][6][7] which extends back-propagation learning algorithm in order to overcome its defects such as the problems of local trapping, overfitting etc. by introducing probabilistic treatment of the Bayesian inference technique for the synaptic weights, has been successfully applied for QSAR studies 8,9) even for massive sample data. 10)Analytical chemistry is often concerned with the classification of samples into groups on the basis of chemical results while such classification process is called diagnosis if these groups are pathological cases. Zhang et al. studied classification problems of thyroid diagnosis using back-propagation neural networks and showed its usefulness. 11)Kohonen developed the unsupervised learning algorithm that simplified the mapping mechanism of the relatively homogeneous structures found in mammalian brains associated with the processing of sensory data, 12) and showed it possible to generate self-organizing map (SOM) of data. In the field of pharmaceutical sciences, the SOM was applied for searching useful drugs. 13-15)When we apply the method of neural networks to some problems encountered in medical and pharmaceutical fields, it is necessary to analyze each problem in a variety of viewpoints under various conditions. It is expected that appropriate disposal and total unified cognition of an individual problem would be possible through those accumulation.Thyroid Disease Diagnosis The correct diagnosis of thyroid dysfunctions based on clinical and laboratory tests is important work and neural networks would be helpful. Coomans et al. analyzed the thyroid data which takes five hormons related to thyroid and includes 215 patients divided into 3 classes as hyperthyroid, hypothyroid and normal by using the method of linear discriminant analysis. 16)Now, with radioimmunoassay techniques it is possible to measure hormones T3, T4 and TSH in the blood very accurately, and the onset of thyroid disease can be surely confirmed.17) It is recently usual to measure two laboratory tests TSH and free T4 in diagnosing thyroid diseases. Then applying neural networks to such case seems to be not necessary now. Another various applications of the notable neural networks for a diagnosis support, however, would be possible and then to make reanalysis of these laboratory tests data should be important.In the present work we make reanalysis of the thyroid data, which was once analyzed by Coomans et al., using the recent two notable neural networks. The BRNN has a powerful soft pruning method called automatic relevance determination (ARD) 5,6) in addition to the importa...

No abstract

No abstract