The integrated spectral database system (SDBS) including IR, Mass, 13CNMR, 1HNMR, Raman and ESR spectra compiled under a compound dictionary, was constructed in a main frame computer for 17 years from 1982. The computer was finished at the end of March, 1999. All the functions including the input tool, search and display systems were equipped in the main frame computer together with the databases. SDBS has been open in the Internet ( SDBSWeb: http://www.aist.go.jp/RIODB/SDBS/) from 1997 and the total accesses from the world are more than 10,000,000 at the end of 2000. To continue the database activities, new input tools by using a personal computer (PC) are necessary. Since the NMR spectral databases contain the chemical structures with spectral assignment, the database structure is complicated. In the present study, by using a PC new input tools have been created for the compound dictionary, and 1H and 13C NMR spectral databases. The DBMS of the Internet service is an Oracle in a workstation, and the functions of adding and updating compound information were made by the PC-Oracle. The input tools for the 1H and 13C NMR spectral databases have been made by using an ISIS/Base system as the manual input tools and a Java applet for the display the spectral patterns dynamically. The digital format to describe an NMR spectral pattern in our system is simple and consists of an initial value of x-axis (Hz and ppm units), an increment of x-axis (Hz and ppm units), number of y-axis, and values of y-axis (integer, normalized to 10,000).

View full abstract

PLS (partial least square regression) has been widely used in QSAR (quantitative structure-activity relationships) studies due to its robust character against the collinear descriptors and small rations of molecules/descriptors. Although PLS is useful, its major restriction is that only linear relation can be extracted from data. Since many structure-activity data are inherently nonlinear in nature, it is desirable to have a flexible method that can model any nonlinear relations. There has been a considerable interest in neural networks for nonlinear modeling. Among them, CNN (counter propagation neural networks) is considered to be a potential candidate because of small learning time and high reproduction compared to other methods. In this study, CNN was used to analyze the nonlinear QSAR data. In order to choose the optimal descriptors from the huge possible combinations, GA (genetic algorithms) was combined with CNN routine. The predictive explained variance for the prediction set was used as the fitness function of GA. From GA and CNN modeling, the totally 15 descriptors could be reduced to 9 descriptors and the prediction ability of the CNN model could be highly improved. Furthermore, external validation was carried out by use of the best CNN model and external data set. The results really indicated that the predicted values are closer to the observed ones compared to the original CNN model.

View full abstract

Artificial neural network (ANN) technique was applied to estimate solubility of several alternative compounds for CFCs and HCFCs. In this estimation method, we used both information on interaction parameters derived from molecular simulation code 'Cerius2/Blends (MSI)' and some physical properties such as liquid density, molecular volume and molecular weight. At first, this presented method succeeded in estimating the solubility of fluorinated compounds with n-Eicosane (C20H42). Furthermore, the solubility of fluorinated compounds with polyol PE-315 (aromatic polyol, OH value = 315mgKOH/g) was estimated accurately by adding maximum charge of hydrogen in fluorinated compounds as input parameter. Finally, constructed ANN was applied to screen new alternative compounds and HFE-236pc (CHF2CF2OCHF2) was selected as new candidates for blowing agents.

View full abstract

In the methodologies of 3D-QSAR such as CoMFA (Comparative Molecule Field Analysis), appropriate molecular alignment is definitely required for predictive data-modeling and correct analysis of the model. In this paper, the novel method for molecular alignment using Hopfield Neural Network (HNN) is proposed. Each molecule is represented by four chemical properties (Hydrophobic, Hydrogen-bonding donor, Hydrogen-bonding acceptor, Hydrogen-bonding donor/acceptor), and then properties among molecules are corresponded by HNN. In order to validate usefulness of this method, 12 pairs of 24 enzyme inhibitors are used as a test set, and our method could successfully reproduce the alignments obtained from X-ray crystallography.

View full abstract

It has been possible to practically use the computer-assisted synthesis design system such as AIPHOS and EROS in order to create synthetic routes of compounds. However, these systems produce many synthesis routes without giving information on which route should be applied first in experiments. It is common knowledge that transition states are keys for how to estimate easiness of chemical reactions and computational chemistry delivers that answer. However, it usually takes long time to obtain such information if organic chemists try to perform theoretical calculations by themselves. Thus, few organic chemists can use theoretical calculations in order to rank the synthetic routes for their experiments. In the present study, we developed a new system which makes it possible to rank the route easily by using a data base. The data base system, called "Data Base for Transition States, TSDB", combines computational and information chemistry. This paper describes the concept of TSDB and how to use it for ranking synthesis routes which a computer-assisted synthesis route design system such as AIPHOS and TOSP system creates.

View full abstract

In order to estimate the kinetics of reaction between OH radical (kOH) and fluorinated formates, which are primary products from the oxidation of hydrofluoroethers (HFEs), the C-H bond dissociation enthalpies (BDE) for fluorinated formates were calculated by using density functional theory (DFT). B3LYP exchange correlation function and 6-311G(d,p) basis set were used to estimate BDE for CnF2n+1OC(O)H: (n=2-4) and CnHF2nOC(O)H: (n=1-3). Since the experimentally measured activation energy (Eact) of the hydrogen abstraction by OH radical for fluorinated formates has never been reported, we estimated the relation between BDE and Eact from the computed values, which had been calculated by using G2(MP2), for syn and anti isomers of CF3OC(O)H. Furthermore, computed Arrhenius A-factor of CF3OC(O)H was adopted to other all formates because lower level calculations with UHF/6-311G(d,p) was revealed that Arrhenius A-factors did not depend on the chain length of formates. Finally, kOH and the life time concerning with decomposition by OH radical (tOH) were estimated with using Arrhenius A-factor and Eact, which were evaluated using BDE, for ten fluorinated formates.

View full abstract

Biochemical IC family containing micro fluid devices has been proposed and developed to contribute to biotechnology. We can produce any functions by using biochemical IC for analyzing chemical reaction, and arrange connection of reactors, concentrators and valves with conserving volume of specimen and processing time. Then to evaluate the connection, we have developed simulation software for flow movement and chemical reaction under mathematical theory based on Petri Net. We present an example of simulation and evaluate validity of this software.

View full abstract

Biochemical IC family containing micro fluid devices must be designed in the connection arrangement of reactors, concentrators and valves for effective reaction in volume of specimen. We have developed simulation CAD software with GUI to optimize the connection by considering volume of reactor, power of micropump, amount of required products and time limitation under simulation theory. We evaluate validity of this software by introducing a concrete example of optimization and simulation.

View full abstract

In order to apply the ensemble learning method to the regression analysis, three new procedures "Arcing_RA1, 2, 3 (Adaptive Resampling and Combining for Regression Analysis) " are proposed. We tested their ability by using Friedman Test ; f(x) = 10sin(pix1x2) + 20(x2-0.5)2 + 10x4 + 5x5 + N(1,0) (x1-5: random number from 0 to 1, N(1,0): normal deviation ). For the training data and test data, 200 and 1000 data were generated, respectively. In the case of single three-layered Neural Network, over learning was observed when the number of neurons in hidden layer (Ny) was over 10. Therefore, the Neural Network, whose Ny was set to 5, was chosen as a weak learner. As a result of the ensemble learning using Arcing_RA1-3, the average of absolute deviation of both training and test data were decreased about 20%. In addition, over learning was not observed. From these results, it was confirmed that Arcing_RA is useful as ensemble learning method for the regression analysis.

View full abstract

The cascade model is a rule induction methodology that uses the level-wise expansion of a lattice. An attribute-value pair is expressed as an item, and every node in the lattice is specified by an itemset and its supporting instances. If the distribution of the class attribute values suddenly changes along a link in the lattice, the link is represented as a rule "IF item-along-link added on itemset-on-upper-node, THEN class-i". The strength of the rule is measured by the BSS value of the link. In the SAR (structure activity relationship) study, we generate many linear substructure patterns like "NH-C-C-C-OH" from the data set of molecular structures. Matching between a pattern and a molecule leads to an item [pattern: y or n] depending on whether the pattern exists in the molecule. Application of the cascade model to this item data set gives us SAR rules. The resulting rules are examined by referring to the supporting molecular structures. Several rules have led to valuable working hypotheses, including the importance of steric hindrance to the coplanarity of NO2 to the activity level.

View full abstract

The Apriori-based graph mining method is an extension of the Apriori algorithm for association rule mining. It constructs a lattice of graph nodes, in which a node at the k-th level of the lattice has k vertices and the number of supporting instances exceeds a user-specified minimum support. The method can devise a rule "IF subgraph Ga is in transaction G, the union of subgraphs GaUGb is also contained in G with a ceratin confidence level". When we give a transaction consisting of a chemical graph and virtual vertices expressing molecular properties, we can obtain rules representing structure activity relationships. The method was used to analyze mutagenicity data for 230 aromatic nitro compounds. Several interesting substructures were found to affect the mutagenicity.

View full abstract