Machine learning and data mining on graph data have been common since early 2000's. Organic compounds have been used as benchmark datasets, but graph mining methods are not so common among pharmaceutical researchers. In this talk, I will address graph mining methods and their applications.

View full abstract

In the fragment molecular orbital method (FMO) the total quantummechanical energy is decomposed into the energy of fragments (one-body property), pair interactions (two-body effect) and higher order effects (three-body etc), and useful information for drug design and other studies is obtained at the ab initio level. The results from several applications will be discussed.

View full abstract

We already defined the "discrimination" as an index representing the characteristics of the fingerprint which is one of the structure descriptor. Since the usage of each bit of fingerprints in compound database is reflect to the discrimination, it is necessary to clarify dataset dependency for the discrimination. In this study, we examined the dataset dependency of the discrimination by using various datasets. The discrimination values for some different compound databases were different among databases. Also we considered a molecular weight, which indicate numerically the size of the compound, as influential factors. Discriminations for datasets that are separated by molecular weight had different pattern among fingerprints. Those results show that the discrimination has dataset dependency.

View full abstract

We developed a website SDBS Strucrures Web (http://staff.aist.go.jp/hiroshi-watanabe/ssw/) that presents a library of chemical compounds in Spectral Database for Organic Compounds (SDBS). For designing the website, we tried to apply a simple guiding principle that Hasse diagram for chemical substructure relation. In this paper, we remind the principle and report our case study for a practical database SDBS.

View full abstract

The conventional stereochemistry is incapable of clarifying the difference between symmetry/asymmetry/pseudoasymmetry and chirality/achirality. Even though the terms of stereogenicity/astereogenicity are adopted in place of symmetry/asymmetry/pseudoasymmetry, the difference between the adopted terms and chirality/achirality cannot be clarified within the conventional stereochemistry. The concepts of RS-stereoisomeric groups and the corresponding stereoisograms are capable of integrating RS-stereogenicity/RS-astereogenicity and chirality/achirality.

View full abstract

On the first stage of development of new drugs, various lead compounds with high activity are required. To design such compounds, we focus on chemical spaces defined by structural descriptors. New compounds close to areas around which highly active compounds exist will show the same degree of activity. Therefore we have been developing a new system of structure generation for searching a target area in chemical spaces. First, highly active compounds are manually selected as initial seeds. Then, those seeds are entered to our generator and structures slightly different from the structures of the seeds are generated and pooled. Next seeds are selected from the new structure pool with the scores based on distance from target on the map. In this study, we used GVK data of ligand-binding affinity to verify the advantage of our generator. Visualization of the chemical space and structure generation were performed, and then the outputs were compared with test data. As a result, our generat or could produce many structures similar to the test data, which exist near the target area. This result shows that exploration of the target area on the chemical space was performed.

View full abstract

We extend the enumeration algorithm in our previous study for acyclic chemical structures, and propose an algorithm for enumerating chemical structures with benzene rings. These algorithms use a tree to represent a chemical structure, where nodes and edges represent atoms and bonds of the chemical structure, respectively. We include benzene rings in the enumeration by treating a benzene ring as a special kind of atom. To enumerate the structure, firstly, the extended algorithm gets chemical formula as the input, and calculates the number of benzene rings that the input chemical formula can have. Then, we generate non-redundant trees from the chemical formula using the algorithm in the previous study. Because we still do not know if nodes adjacent to a benzene ring have bonds with which carbon of that benzene ring, we generate a rule to assign each node to the carbon atom of a benzene ring. To test the performance of the proposed algorithm, we compared the computational time with that of MOLGEN, a well-known enumeration tool. The result shows that our algorithm is much faster than MOLGEN.

View full abstract

RNAi is a natural biological process involving gene silencing or regulation with siRNA and expected to be applied in the therapeutics of gene disorders. The delivery of siRNA into the cell is demonstrated using cationic lipid. The lipid is called ’Transfection reagents’and the development of new reagents has recently advanced. Some relation exists between the chemical structures of the reagents and their properties, but it is not clear yet quantitatively and the development of new reagents depends on the previous studies. In this study we collected the data of the transfection reagents from literature and constructed statistical models of QSAR models of transfection reagents in order to predict the chemical structures of high efficient reagents in future. The descriptors calculated for chemical structures and experimental parameters were used as explanatory variables, and their rate of gene expression was used as an objective variable. In this study we calculated descriptors for the structures of amines and carbon chains separately and constructed the regression models predicting the rate of gene expression. The value of q2 improved in PLS and SVR models compared with the value of q² calculated with the descriptors for the structures of transfection reagents. The descriptors derived from carbon chains had high correlations with the rate of gene expression compared with the correlations of the descriptors derived from amines.

View full abstract

Since there are many commercially available compounds which are toxic to humans, animals, and environment, Chemical Substances Control Law (CSCL) is enacted in order to regulate such compounds. However it is almost impossible to examine the toxicities of all the compounds experimentally. Thus, predicting the toxicities using Quantitative Structure-Toxicity Relationships (QSTR) Analysis as an alternative method of experimental toxicity test has been being tried. In this study, we tried to build a regression and a 2-class classification models in order to predict algae growth inhibition toxicities of the compounds, which have been determined experimentally, using the support vector machine (SVM) with common chemical descriptors as well as graph descriptors. We proposed to use the graph descriptors, which characterize partial graph structures, as complemental descriptors of chemical ones for building better QSAR models. As a result, although we could construct an enough 2-class classification model, graph descriptors were not as useful as chemical descriptors. However, it was suggested that graph descriptors have important information which were different from chemical descriptors’ one. Now, we are trying to build an enough regression model.

View full abstract

A membrane bioreactor (MBR) is a wastewater treatment process which uses activated sludge to remove organic substances from wastewater. The membrane is used to separate activated sludge from treated water in MBRs. The concentration of mixed liquor suspended solids (MLSS) is important for long-term fouling prediction and process control in MBRs, but it requires much time and cost to measure the concentration of MLSS. In this study, we developed a regression models between the concentration of MLSS and other variables such as operating conditions and water quality variables. To develop a widely-used and accurate model, we analyzed three data sets measured in different MBR plants. The model constructed with three data sets could achieve high predictive performance. From the results of the variable selection using a genetic algorithm-based partial least squares (GAPLS) method and the comparison of GAPLS models with different explanatory variables, it was suggested that the viscosity in the membrane tank and treated water quality variables are important for prediction of the concentration of MLSS.

View full abstract

Model predictive control is widely used as a process control method for a complicated multivariable process. However, optimization of control parameters is complicated and a data set for system identification cannot always be obtained in a real process. In order to solve these problems and perform more effective control, we propose a new process control method using soft sensor models. We refer to this method as inverse soft sensor-based feed forward (ISFF) control. Soft sensor models are constructed between a controlled variable (y) as an objective variable, and manipulated variables (U) and other process variables (X) as explanatory variables. The optimal control strategy of U which optimizes the objective function including y is determined with inverse analysis on the soft sensor models while considering X variables. The proposed method was applied to the change of a set point of a simulated CSTR system and the optimization of y of a simulated fed-batch fermentation process, and the validity of ISFF was confirmed.

View full abstract

Gas chromatography (GC) and two-dimensional GC (GC-GC) are widely used for separation, structure elucidation and quantitative analysis. In GC and GC-GC, the chemical structure is elucidated by comparing the measured retention time of each compound with the database. But, structure elucidation is infeasible if the retention time is not available from the database. Thus, quantitative structure retention relationship (QSRR) is proposed to predict the retention time from the structure. Some researchers constructed the QSRR models specialized for the limited types of compounds. In this study, we aim to construct the QSRR models that can predict the retention time of various compounds in GC-GC with high accuracy. In addition, we propose a structure elucidation method based on the inverse analysis with the models. First, the objective value of the retention time is set. Then, structure elucidation is accomplished by comparing the objective value with the predicted retention time of new structures. The prediction errors can be a problem in comparison between the predicted and objective values. We deal with this problem by setting an acceptable error for each compound based on the reliability of the predicted value. The analysis with the measurement data proved the effectiveness of the proposed method.

View full abstract

Many kinds of nonlinear regression methods have been developed to construct predictive models even in the existence of the nonlinear relationship between objective variables and explanatory variables. However, even when very accurate regression models are constructed, the constructed models exhibit poor predictive performance for new data. Therefore, these regression models must be validated to quantify their predictive ability and allow the appropriate model and hyperparameters to be selected. We propose predictive performance criteria for nonlinear regression models without cross-validation. The proposed criteria are the determination coefficient and the root-mean-square error for the midpoints between k-nearest-neighbor data points. These criteria can be used to evaluate predictive ability after the regression models are updated, whereas cross-validation cannot be performed in such a situation. The proposed method is effective and helpful in handling big data when cross-validation cannot be applied. We analyze numerical simulation data, aqueous solubility data and toxicity data, and confirm that the proposed criteria enable the predictive ability of the nonlinear regression models to be appropriately quantified.

View full abstract

One of the most fundamental questions in chemistry is how structural modification of a molecule by substituents affects macroscopic properties of the molecular condensed system. For example, Itoh et al. reported reduced viscosity in fluorinated derivatives of the 1-butyl-3-methylimidazolium (bmim) salts. At first, we suspected that the reduced viscosity was due to a modification of molecular motion induced by the substituted atoms, because the rotational potential calculated for the fluorinated derivative of the bmim cation was different from that for the bmim cation. But, unfortunately, we couldn’t assign the difference of the rotational potential to the reduced viscosity. This fact implies that the substituent effect on the macroscopic properties absolutely demands the approach for many-body problems. In order to answer the question, we have applied the 3-state Potts model to a dimer system of 1-methyl-3-propylimidazolium chloride (pmimCl) derivatives and calculated interaction matrixes by the DFT method. Now we report the mean field approximation of the N-body Potts model that suits for conformation-modulated interaction in the pmimCl.

View full abstract

Although there are many methods which theoretically calculate pKa values of carboxylic acids, it has been very difficult to calculate the values consistent with experimental ones for some acids. In the present study, we consider model systems in which 8 water molecules surround around the carboxylic group and regarded the process as the ionization reaction. Free energies of reactions were calculated at the B3LYP/6-311++G(d,p) level of theory and contributions from solvents were evaluated using the CPCM method. The present method can calculate pKa values close to experimental ones for almost the acids we adopted in the present study.

View full abstract

We have been developing the hydrolysis prediction system using quantum chemical calculation and experiment in order to add a new view in the biodegradation prediction using activation free energy of hydrolysis reaction derived from the reaction analysis of acid and alkaline hydrolysis of esters through quantum chemical calculation. I investigated the correlation between experimental values (ΔG^‡_obs) and calculated values (ΔG^‡_inf). As a result, It was a relatively good correlation (R²=0.87) in acidic condition. On the other hand, It was low correlation (R²=0.67) in basic condition. In particular, the substances susceptible to solvent effects are large standardized residual. These results indicate that we have to calculate the ΔG^‡_inf　including the solvent effects.

View full abstract

It is widely known that the structural/physico-chemical complementarities in each protein surface is very important in protein-protein interaction (PPI) interface, since the stability of the complex is affected by the complementarity. And, a decrease in the existence ratio of charged amino acid residues and an increase in that of non-polar residues on the PPI interface have been reported compared with on the surface of protein. However, it has been also reported that there is no difference in the existence ratio of Arg between protein surface and PPI interface. So, we analyzed the complementarities focusing on Arg at PPI interface using 143 PDB data and we picked up the charged amino acid residue’s pairs and the surrounding residues existing within 5Å in a PPI interface to generate the model structures for our calculations. Then, to analyze the inter-residue interactions between charged amino acid residue’s pair and the surrounding residues, we have performed ab initio molecular orbital calculations and density functional calculations.

View full abstract

The interactions between aromatic residues in protein structures have been extensively studied. It was shown that aromatic clusters formed by three or more aromatic rings were found in many protein structures and they were important for structure stability, folding, protein-protein recognition, and ligand binding. In drug design, we have a strong interest in protein-ligand interactions. Previously, we found that there were aromatic clusters formed by three or more aromatic rings in protein-ligand complexes as well as in protein structures. To reveal attractive and repulsion interactions in aromatic clusters and between aromatic clusters and the surrounding residues, we performed ab initio density functional calculations. To generate the model structures for our calculations, we analyzed aromatic clusters in 201 protein-ligand complexes data of PDBbind-CN(http://www.pdbbind.org.cn/) and picked up the aromatic clusters and the surrounding residues existing within 7Å in the protein-ligand complexes.

View full abstract

A rooted digraph is regarded as a protonated water (PW) cluster, and a hydrogen bond (H-B) matrix can represent a rooted digraph. All possible topology-distinct patterns corresponding to PW clusters containing up to 8 water molecules were enumerated by using H-B matrices. We apply the Hamiltonian algorithm combined with ab initio molecular orbital calculations to the optimization of molecular structures for PW clusters up to octamers.

View full abstract

Straightfoward molecular dynamics predictions were attempted on the Cis-Trans isomerization rate of benzylideneanilines in solution, where solute-solvent interactions plays crucial role in regulating the chemical equilibrium under extremely high pressure.

View full abstract

This work aims at a molecular design system based on computer simulations, auto data processing and data mining. From the result of performance evaluation using standard methods on Gaussian09, we confirmed that we can deal with hydrocarbon molecules including several carbon atoms.

View full abstract

The aromaticity of transition metal complexes with the π conjugated chelate ligands is interesting in physical and chemical nature of the metal atom. It is considered that the interaction between the d orbital of metal atom and π bond of ligand causes the aromaticity. Especially two d orbitals probably play important role for the aromaticity. But the role of the d orbitals for the aromaticity was unknown well. In this conference, we will present the aromaticity of transition metal complexes with a π conjugated-ligand by ab initio molecular orbital theory. The estimation of the aromaticity was performed by the value of IDA. Because other aromatic indexes as NICS and HOMA can’t be use for complexes and /or multi-cyclic compounds. IDA is an index of the quantitative aromatic property based on each bond weight by CiLC method. We studied systematically transition metal complexes for many transition metal atoms. To study the variation of the aromaticity for ligands (NH ₃, PH ₃, CO) were also selected for all ligands of the complex expect for the chelate ligand. As the results, the aromaticity can be characterized the interaction between d orbital of metal atom and π orbital of ligands.

View full abstract

The intramolecular [2+2+2] reaction is one of the famous pericyclic reactions. Although reaction mechanism is important and has attracted a lot of attention from both experimental and theoretical researcher, the reaction mechanism is still not clear. In this study, we have studied the ring-opening reaction mechanism of tris-cyclopropacyclohexane (TPH) and tris-cyclobutacyclohexane (TBH) to interpret the intramolecular [2+2+2] reaction mechanism. On each system, we calculated the ring-opening reaction pathways from each structural isomer because there are many structural isomers for reactant. All structures were determined with analytically calculated energy gradients at the CASSCF method with the 6-31G* basis set, the energy was estimated by a MRMP2 method with the 6-311+G** basis set. The reaction from TBH has lager activation energy barriers than the reaction from TPH. The C-C Bond of TPH is easier to cleave than that of TBH because the ring distortion of a three-membered ring is larger than that of a four-membered ring. In addition, it was found that the conformation of the central six-membered ring affects the activation energy and the reaction mechanism.

View full abstract

Cycloalkanes “C_nH_2n” composed of hydrocarbon and having one or more rings of carbon atoms in their structure are typical organic cyclic compounds. The physical properties of cycloalkane are similar to those of alkane, but the melting point (m.p.) of cycloalkane having one ring, representative physical property, is found to strongly on the number of carbon atoms; the m.p. increases as the cycloalkane-size increases and has local minima at n = 5, 7, 9, 11 and 13. According to our latest study, semi-empirical MO calculation was used to calculate the stable structure of cycloalkane, it was revealed to relate strongly the m.p. of cycloalkanes with their molecular symmetry. In the present study, the size dependence of the m.p. of cycloalkanes was discussed in terms of comparison with the specific heat obtained from ab initio calculation. We found that the size dependence of the m.p. of cycloalkanes was caused by their structures mainly and by their specific heat slightly.

View full abstract

Magnesium has interesting properties suitable for acting negative electrodes in rechargeable battery. The electrodeposition from THF solution with Grignard reagent is one of the potential methods to make good electrodes. In this study, dissociation of Grignard reagents in solution was analysed using values from quantum mechanical (QM) calculations. Geometry optimizations were performed for Grignard reagents with several substituents, such as methyl, ethyl n-propyl, n-pentyl, n-octyl, and phenyl groups. As we observed ionic conductivities of these reagents dissolved in THF, a PLS regression was applied in order to explain the observed property on the basis of calculated values such as dissociation energies of Mg-Br and Mg-C bonds and their distances, molecular volumes, maximum radius of molecules, r(max), and orbital energies of HOMO and LUMO as explanatory variables. The GA-PLS analysis selected r(max) and LUMO energies as explanatory variables for the PLS regression. The correlation equation from the PLS analysis was used to analyse the mechanism of magnesium deposition in THF with Grignard solution.

View full abstract

We have proposed a conformational code for the description of conformations of all kinds of chemical compounds based on structural analysis using vibrational circular dichroism (VCD) of chiral bioactive compounds and have developed the auto-conversion techniques of conformational information needed for structural homology between proteins, which can be represented by 20 symbols of amino acid residues. The fuzzy search and data mining techniques using “h”, “s”, and “o” fragment patterns to convert 3D data to 1D data reflect well the shape of main chain and show the characteristic fragment patterns “hhshshh” in the armadillo repeat region of β-catenin. Characteristic fragment patterns “shhshss”, which are sandwiched between two amino acid residues His and Pro, as HisXaaXaaXaaXaaXaaPro (Xaa: an amino acid residue), have been found in light chains of various types of immunogloblins and MHC class I and II molecules including HFE. However, human IgM rheumatoid factor, GM-CSF autoantibody, TSHR autoantibody, CD1D, and MIC-A,B do not clearly indicate these patterns. In this presentation, the comparison result of the new method with the DSSP method to standardize secondary structure assignment is also discussed.

View full abstract

We developed an automatic modeling system that analyzes experimental data on the cross-sectional shapes of films deposited on substrates with nanometer- or micrometer-sized trenches. The system identifies a suitable reaction model to describe the film deposition. The inference engine used by the system to model the reaction mechanism was designed using bio-inspired algorithms. We studied the effect of the bio-inspired algorithms on the system.

View full abstract

We proposed a novel calculation method to reproduce the deposition profiles in CVD reactors. We formulated the governing equations of the batch-type reactor with round-shaped substrates and showed that these reactors had a few basis functions, which consist of the concentration profiles of the deposition species under various conditions.

View full abstract

We applied support vector machine (SVM) classifiers for the musk-order molecules and their analogues. The EVA descriptor based on the vibrational eigenvalues was used as a molecular descriptor. The effects of parameters of EVA and the kernel functions used by the SVM classifiers were discussed.

View full abstract

In the previous study, the experimental design to satisfy the desired property in developing products based on probability that satisfy target was developed, but previous studies is only suitable when do one experiment at the same time. In this study, we developed the experimental design which can use when do some experiments at the same time based on previous studies. For avoiding to select candidates which are similar to experimental parameters which do experiment at same time, we select candidates at regular intervals. Changing the interval of candidates, the proposed method is applied to simulation dates and we compared the interval of candidates and the number of experiments which need satisfy target. The interval of candidates and the number of experiment have no clear relation, but the relationships the interval of candidates ,the speed of improving predictability and the predictability which need satisfy target have same tendencies in the all data.

View full abstract

Today, there are many kinds of grades in synthetic resins of domestic resin industry and this fact makes it difficult for plastic manufacturers to choice proper one. So, we developed visualization technique of the material characteristics that help us with Self-organizing map in order to grasp their similarities of resin grades. We made Self-organizing maps on 32 grades transparent PP (polypropylene) and 142 grades styrene elastomer each other. Self-organizing map is one of multivariate analysis technique. It can contract multidimensional information to two dimensions and help us to figure out entire data. On the map, the similarities among all grades were shown as distance, and they were grouped into some classes. The map of each physical property suggested grade characteristics by classes. To show an application of the maps, we try to consider the alternatives of transparent PP grade for certain product with them and search out appropriate one. We will report the details on poster.

View full abstract