[in Japanese]

View full abstract

SIGMA is a general purpose text database management system and is now working at the Computer Center, Kyushu University for large scale full text databases, natural language analysis and bibliographic retrievals. SIGMA searches a 2MB file in 1 CPU second or in 3 seconds even if disk accessing time is included. Since Genetic information databases are of complicated formats and contain very long sequence information, converting them to full text databases and retrieving them by SIGMA system is useful for analyzing the Genetic Information.
We have converted the text files of Genetic Sequence Data Bank (Genbank) and Protein Sequence database (PIR) and efficiently maintained and retrieved the converted files as full text databases.

View full abstract

Information or embedded patterns can be extracted by via minimum energy techniques or maximum entropy methods. We show how to combine these two approaches via Bayes' theorem so that the problem can then be cast into that of minimizing the Helmholtz free energy. This problem can then be solved an optimization algorithm such as simulated annealing which as we show yields the minimum Helmholtz free energy state rather than the minimum energy state. We will also briefly discuss the Bayes'theorem as a model for learning.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

The prediction of globular protein secondary structure is studied by the neural network. The protein secondary structure is allocated to each residue by using Kabsch and Sander's DSSP and the neural network is trained to learn the protein secondary structures. In the input layer of the neural network, the sequence of residues can be allowed to be 24 characters which are 20 amino acides, chain break (!), B, X and Z. The 3, 4 and 8 state classification scheme of secondary structures are considered. In each case, the percent of correct prediction is calculated. Furthermore, the overlearning effect in the protein secondary structure prediction is discussed. In addition, the application of the neural network with the modular architecture to the prediction of the protein secondary structure is presented. The results of the neural network with the modular architecture and thesimple three layer perceptron are compared.

View full abstract

It is maintained that active use of a stochastic model is essential for analyzing the amino acid sequence of protein or the base sequence of DNA. Here we shall discuss the amino acid sequence of protein. Using HMM, which is a type of stochastic model, an attempt was made to predict the secondary structure of protein. In this process, secondary structures such as helix, sheet and turn are each learned by descrete HMM, and the output probabilities from the stochastic model is used to determine which part of the sequence (whose structure is unknown) corresponds to which structure. With a model in which one amino acid is taken as the output symbol, favorable results could not be obtained, but when two consecutive amino acids were used as the output symbol, good results could be obtained in estimating helix and sheet.

View full abstract

A procedure to detect similar local structures of proteins from Ca coodinates ispresented. First, the conformations of peptide segments of seven residues long are approximated by a limited number of representatives, each of which is assigned a symbol. Thus, the entire conformation of a protein is represented by a symbol string. Then, the comparison of these symbol strings using sequence alignment technique gives pairs of similar local structures. These pairs are considered candidates of structural motifs. The application of the procedure to the analysis of 93 proteins gave 858 pairs of similar local structures, which inculded several well-known structural motifs such as nucleotide-binding βαβ motif and calcium-binding EF hand. The characterization of amino acid patterns of similar local structures given by the procedure should be useful for the development of protein structure prediction based on the acquisition of empirical rules from a large-scale database.

View full abstract