JP3928050B2 - Base sequence classification system and oligonucleotide frequency analysis system - Google Patents

Description

  The present invention provides a base sequence classification system for creating a self-organizing map for classifying a base sequence into a biological classification based on the frequency of occurrence of each of a plurality of types of oligonucleotides in the base sequence, and This is an oligonucleotide frequency analysis system for analyzing the occurrence frequency bias of oligonucleotides (bias due to biological classification such as species and bias due to DNA sequence position) using the above self-organizing map. is there.

  Multivariate analysis, such as factor correspondence analysis and principal component analysis (PCA), has been used successfully to investigate genetic sequence differences. However, the clustering ability of the conventional multivariate analysis is insufficient when collectively analyzing a large amount of sequence data obtained from various genomes.

  Self-organizing map (SOM) developed by Kohonen using competitive neural networks is used for recognition of images, sounds, fingerprints, etc., and control of production processes for industrial products. (Non-Patent Document 1, Non-Patent Document 2). SOM is a non-linear mapping of multidimensional data onto a two-dimensional array of coupling weight vectors, and effectively preserves the topology of a high-dimensional data space. SOM is a powerful tool for clustering and visualizing high-dimensional complex data on a two-dimensional plane.

  In recent years, with the elucidation of genome information of various organisms, a vast amount of life information is being accumulated, and it is important from the aspect of drug development to solve the mystery of life from such life information using a computer, Applications of SOM are becoming popular. The inventors of the present application have proposed an improved SOM creation method improved from the conventional SOM creation method for genome information science (see Patent Document 1, Non-Patent Documents 3 and 4). This improvement is based on a batch learning SOM creation method that batch processes data input and learning so that the learning process and the map to be created (SOM) do not depend on the order of data input. Further, in the improved SOM creation method, the initial connection weight vector is defined using principal component analysis (PCA). Thus, the improved SOM does not depend on the initial conditions as well as the order of data entry.

For example, in Example 1 of Patent Document 1, based on the codon (trinucleotide) usage frequency of 16 types of microorganisms as high-dimensional input data, an SOM in which microbial genes are classified using an improved SOM is created. A method is disclosed.
International Publication No. WO 02/50767 A1 (released on June 27, 2002) Application of Self-Organizing Maps-Two-dimensional Visualization of Multidimensional Information ”(by Tokutaka Heizo, Satoru Kishida, Kikuro Fujimura, Kaibundo Publishing Co., Ltd., July 20, 1999, first edition, ISBN 4-303-73230-3) “Self Organizing Map” (T, Kohonen, Tokurataka Heizo, Kishida Satoru, Fujimura Kikuro, Shubringer Fairlark Tokyo, June 15, 1996, ISBN 4-431-70700) -XC3055) Kanaya, S., Kinouchi, M., Abe, T., Kudo, Y, Yamada, V., Nishi, T., Marl, H. and Ikemura, T. (2001) Analysis of codon usage diversity of bacterial genes with a self-organizing map (SOM): characterization of horizontally transferred genes with emphasis on the E. coli 0157 genome. Gene 276, 89-99. Abe, T., Kanaya, S., Kinouchi, M., Ichiba, V., Kozuki, T. and Ikemura, T. (2003) Informatics for unveiling hidden genome signatures.Genome Res. 13, 693-702.

  In the public DNA database, sequence data of only one of a pair of complementary DNA sequences is registered. In Patent Document 1 and Non-Patent Documents 3 and 4, an SOM is created using the appearance frequency of oligonucleotides in this sequence data.

  However, considering the overall characteristics of the oligonucleotide appearance frequency in the genome, the frequency of occurrence between complementary oligonucleotides (for example, between AAAC and GTTT) in one of the pair of complementary DNA sequences. The difference is not important for classification. Rather, since the appearance frequency of complementary oligonucleotides should be the same in the entire double-stranded DNA, the difference in the appearance frequency described above is that the created SOM accurately characterizes the entire double-stranded DNA. The result may not be reflected. In addition, since the creation of SOM requires a long calculation time, it is desired to reduce the calculation time as much as possible.

  A first object of the present invention is to provide a base sequence classification system capable of creating a self-organizing map in a short time without a significant decrease in classification ability.

  Moreover, in SOM created using the appearance frequency of the oligonucleotide in Patent Document 1 and Non-Patent Documents 3 and 4, microorganisms from which the base sequence is derived can be classified into a plurality of biological classifications. However, with this SOM, it is impossible to grasp the frequency of occurrence of individual oligonucleotides in a base sequence derived from an organism belonging to each biological classification. If the frequency of occurrence of individual oligonucleotides in a base sequence derived from an organism belonging to each biological class can be ascertained, for example, it is possible to find an important key oligonucleotide that separates biological classes, which is useful. is there.

  A second object of the present invention is to provide an oligonucleotide appearance frequency analysis system that makes it possible to grasp the appearance frequency of individual oligonucleotides in a base sequence derived from an organism belonging to each biological classification. is there.

  In the human genome, a full-length draft sequence has been determined, but there are still a large number of regions whose functions are unknown. It is extremely difficult to find a signal sequence (base sequence having a function such as recognizing a transcription factor) from an extremely long DNA base sequence such as human DNA and analyze the function. Therefore, if a position where a large number of signal sequences are present can be predicted from the DNA base sequence, it is useful for analyzing the function.

  A third object of the present invention is to provide an oligonucleotide frequency analysis system that makes it possible to predict a position where a large number of signal sequences are present from a DNA base sequence.

  In order to solve the above-described problem, the base sequence classification system of the present invention arranges the appearance frequency of each of a plurality of types of oligonucleotides in a base sequence as an input vector group on a multidimensional space, and these input vectors. A base sequence classification system that creates a self-organizing map by nonlinearly mapping a group onto a map in which a plurality of grid points are arranged and classifying the base sequence into each grid point. By adding the appearance frequencies of oligonucleotides forming a pair, the addition unit for calculating the appearance frequency of oligonucleotides for each pair, and the self-organizing map based on the appearance frequencies of oligonucleotides for each pair A self-organizing map creation unit is provided.

  In order to solve the above-described problem, the oligonucleotide appearance frequency analysis system of the present invention arranges the appearance frequencies at which multiple types of oligonucleotides each appear in a base sequence as an input vector group on a multidimensional space, and A self-organizing map creating unit that creates a self-organizing map by nonlinearly mapping an input vector group onto a map in which a plurality of lattice points are arranged and classifying the base sequence into each lattice point; And an appearance frequency map creating unit that creates an appearance frequency map representing information on the appearance frequency of the oligonucleotide for each lattice point for each oligonucleotide.

  The analysis system includes an expected value calculation unit that calculates an expected value of the appearance frequency of the oligonucleotide in the base sequence classified into each lattice point based on the mononucleotide composition in the base sequence classified into each lattice point. A normalization unit that normalizes the appearance frequency of the oligonucleotide in the base sequence classified into each lattice point by dividing by the expected value, and the appearance frequency map creation unit is normalized It is preferable that the appearance frequency map is created based on the appearance frequency of the oligonucleotide.

  The analysis system further includes an adding unit that calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequencies of the complementary pair of oligonucleotides, the self-organizing map creating unit and the appearance frequency It is preferable that the map creation unit creates a self-organizing map and an appearance frequency map based on the appearance frequency of oligonucleotides for each pair.

  In order to solve the above-described problem, the oligonucleotide frequency analysis system of the present invention is configured to determine the frequency of occurrence of each of a plurality of types of oligonucleotides in a plurality of fragment base sequences extracted from the same DNA sequence. Are placed on a multi-dimensional space, and these input vector groups are non-linearly mapped by self-organization from a multi-dimensional space onto a map on which a plurality of lattice points are placed. Frequency of occurrence of individual oligonucleotides on the DNA sequence based on the frequency of occurrence of individual oligonucleotides in the fragment base sequence classified into each lattice point And an appearance frequency distribution diagram creation unit for creating an appearance frequency distribution diagram showing the distribution of .

  Based on the mononucleotide composition in each fragment base sequence, the above analysis system includes an expected value calculation unit that calculates an expected value of the appearance frequency of the oligonucleotide in each fragment base sequence, and an oligonucleotide value in each fragment base sequence. A normalization unit that normalizes the appearance frequency by dividing by the expected value, and the appearance frequency distribution diagram creation unit creates an appearance frequency distribution diagram based on the appearance frequency of the normalized oligonucleotide It is preferable to do so.

  The analysis system further includes an adding unit that calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequencies of the oligonucleotides that make up each complementary pair, and the appearance frequency distribution map creation unit includes: It is preferable to create an appearance frequency distribution map based on the appearance frequency of oligonucleotides for each pair.

  According to the classification system of the present invention, a self-organizing map can be created based on the appearance frequency of oligonucleotides for each pair.

  Self-organizing maps are useful as bioinformatics tools for classifying many types of base sequences into biological classifications.

  In addition, if a self-organizing map is used, it is possible to efficiently predict how many types of microorganism-derived DNA exist in a mixed sample of bacterial DNA whose components are unknown. . Therefore, the self-organizing map is particularly useful for component analysis of a mixed sample containing a plurality of types of microorganisms such as a mixture of microorganisms in a natural environment that is difficult to culture.

  In addition, by using a self-organizing map, it is possible to classify base sequences without any information on biological classification. Therefore, SOM is useful for specifying to which strain group an organism (such as bacteria) whose species is completely unknown, only a part of the base sequence of its genome is known. Therefore, the self-organizing map is useful for searching for new and industrially useful bacteria.

In addition, if a self-organizing map is used, it is possible to find segments that are considered to have been introduced from other species through horizontal transmission from the genome.
According to the self-organizing map, details will be described later,
According to the classification system of the present invention, the adder calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequencies of the complementary pairs of oligonucleotides. Since the self-organizing map creating unit creates the self-organizing map based on the appearance frequency, the amount of calculation in the self-organizing map creating unit can be halved. Therefore, the computation time in the self-organizing map creating unit that generally requires a long computation time can be reduced to about half. In addition, the difference in frequency of appearance between two complementary pairs of oligonucleotides is not important for the classification of the base sequence, so it can be treated by treating the complementary pair of oligonucleotides as identical. Self-organizing maps can be created without noticeable reduction in classification ability.

  Therefore, according to the present invention, it is possible to provide a base sequence classification system capable of creating a self-organizing map in a short time without a decrease in classification ability.

  In the oligonucleotide appearance frequency analysis system of the present invention comprising a self-organizing map creating unit and an appearance frequency map creating unit, a self-organizing map is created and oligonucleotides corresponding to lattice points on the self-organizing map are created. It is possible to create an appearance frequency map that represents information related to the appearance frequency of. Since the self-organization map is separated into regions for each biological classification, if an appearance frequency map is viewed with reference to the self-organization map, individual base sequences derived from organisms belonging to each biological classification are included. Feature extraction of oligonucleotide appearance frequency can be performed. For example, the types of oligonucleotides that appear excessively in a base sequence derived from an organism belonging to a specific biological classification, the types of oligonucleotides that appear excessively in a base sequence derived from an organism belonging to a specific biological classification, etc. Can be grasped. Thus, for example, key oligonucleotides that are key to separating biological classifications can be found.

  The analysis system further includes an expected value calculation unit and a normalization unit, and the appearance frequency map creation unit creates an appearance frequency map based on the normalized appearance frequency of oligonucleotides. If so, the difference in the appearance frequency of the oligonucleotide between the base sequences classified into different lattice points can be accurately detected by separating from the bias of the mononucleotide composition in each base sequence. Therefore, it is possible to create an appearance frequency map that more accurately reflects the difference in the appearance frequency of oligonucleotides between organisms belonging to different biological classifications.

  The analysis system further includes an adding unit that calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequencies of the complementary pair of oligonucleotides, the self-organizing map creating unit and the appearance frequency If the map creation unit is configured to create a self-organizing map and an appearance frequency map based on the appearance frequency of oligonucleotides for each pair, the following effects can be further obtained. That is, the addition unit calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequency of the complementary pair of oligonucleotides, and the self-organization based on the appearance frequency of the oligonucleotide for each pair. Since the self-organizing map is created by the integrated map creating unit, the amount of calculation in the self-organizing map creating unit can be halved. Therefore, the computation time in the self-organizing map creating unit that generally requires a long computation time can be reduced to about half. In addition, the difference in frequency of appearance between two complementary pairs of oligonucleotides is not important for the classification of the base sequence, so it can be treated by treating the complementary pair of oligonucleotides as identical. Self-organizing maps can be created without noticeable reduction in classification ability. Furthermore, according to the above configuration, the addition unit calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequency of the complementary pair of oligonucleotides, and the appearance of the oligonucleotide for each pair. Since the appearance frequency map creation unit creates the appearance frequency map based on the frequency, the amount of calculation in the appearance frequency map creation unit can be halved. Therefore, it is possible to reduce the computation time in the appearance frequency map creating unit that generally requires a long computation time to about half. In addition, since the difference in appearance frequency between two complementary pairs of oligonucleotides is not important in the appearance frequency map, by treating the complementary pair of oligonucleotides as the same, An appearance frequency map can be created without degrading accuracy.

  In the oligonucleotide appearance frequency analysis system of the present invention comprising a self-organizing map creation unit and an appearance frequency distribution diagram creation unit, an appearance frequency distribution diagram showing the distribution of the appearance frequency of individual oligonucleotides on a DNA sequence is created. be able to. From this appearance frequency distribution diagram, it is possible to know a region where a specific oligonucleotide appears excessively or a region where a specific oligonucleotide appears too small on the DNA sequence. Some of these regions are considered to correspond to regions rich in signal sequences and gene-rich regions. Therefore, it is possible to predict the position of a region containing a lot of signal sequences or a gene-rich region based on the appearance frequency distribution diagram.

  The analysis system further includes an expected value calculation unit and a normalization unit, and the appearance frequency distribution diagram creation unit creates an appearance frequency distribution diagram based on the appearance frequency of the oligonucleotide normalized by the mononucleotide appearance frequency. With this configuration, the difference in the appearance frequency of oligonucleotides between fragment base sequences at different positions can be accurately detected by separating them from the bias of the mononucleotide composition in each base sequence. Therefore, it is possible to create an appearance frequency distribution map that more accurately reflects the difference in the appearance frequency of oligonucleotides between fragment base sequences at different positions.

  The analysis system further includes an adding unit that calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequencies of the oligonucleotides that make up each complementary pair, and the appearance frequency distribution map creation unit includes: If the configuration is such that an appearance frequency distribution map is created based on the appearance frequency of each pair of oligonucleotides, the following effects can be further obtained. That is, the addition unit calculates the appearance frequency of the oligonucleotide for each pair by adding the appearance frequency of the complementary pair of oligonucleotides, and the appearance frequency based on the appearance frequency of the oligonucleotide for each pair. Since the appearance frequency distribution diagram is created by the distribution diagram creation unit, the amount of calculation in the appearance frequency distribution diagram creation unit can be halved. Therefore, it is possible to reduce the calculation time in the appearance frequency distribution diagram creating unit that generally requires a long calculation time by about half. In addition, since the difference in appearance frequency between two complementary pairs of oligonucleotides is not important in the appearance frequency distribution diagram, it is possible to treat the complementary pair of oligonucleotides as the same. It is possible to create an appearance frequency distribution map without reducing accuracy.

  As shown in FIG. 1, the oligonucleotide appearance frequency analysis system (base sequence classification system) of the present embodiment includes a complementary data adding unit (adding unit) 1, an oligonucleotide appearance frequency data storage unit 2, and an expected value calculation. Unit 3, mononucleotide composition data storage unit 4, normalization unit 5, appearance frequency map creation unit 6, appearance frequency distribution diagram creation unit 7, storage unit 8 for storing various calculation data, annotation Data creation unit 9, SOM creation unit 10, input device 11 such as a keyboard and mouse for inputting data and instructing execution of calculation, and output device such as a display and a printer for performing display and printing 12.

  Among these, the complementary data addition unit 1, the expected value calculation unit 3, the normalization unit 5, the appearance frequency map creation unit 6, the appearance frequency distribution diagram creation unit 7, the annotation data creation unit 9, and the SOM creation unit 10 Then, the computer is caused to function as the complementary data adding unit 1, the expected value calculating unit 3, the normalizing unit 5, the appearance frequency map creating unit 6, the appearance frequency distribution diagram creating unit 7, the annotation data creating unit 9, and the SOM creating unit 10. Is a functional block realized by a computer program. For example, the SOM creation unit 10 can be realized by a computer and a collective learning SOM program for causing the computer to function as the SOM creation unit 10, for example, “XanaMine” manufactured by Xanagen Corporation. The storage unit 8 can be realized by a rewritable memory such as a RAM, and the oligonucleotide appearance frequency data storage unit 2 and the mononucleotide composition data storage unit 4 can be realized by a mass storage device such as a hard disk.

  The oligonucleotide appearance frequency data storage unit 2 stores appearance frequency data for each oligonucleotide in a plurality of base sequences supplied from a publicly available base sequence database or the like. The plurality of base sequences may be base sequences derived from different biological species, or may be a plurality of fragment base sequences extracted from the same DNA sequence. When using fragment base sequence data, it is preferable to align the length of the fragment base sequence to a certain length (for example, 10 kb or 100 kb). In the case of using a plurality of fragment base sequences extracted from the same DNA sequence, the fragment base sequences (regions) may partially overlap or not overlap at all. The length of each fragment base sequence is preferably 10 kb or more, and more preferably 100 kb or more.

  The oligonucleotide appearance frequency data stored in the oligonucleotide appearance frequency data storage unit 2 may be prepared by the following method, for example. Data of multiple DNA full-length base sequences or multiple fragment base sequences derived from the same DNA from existing genome information databases or full-length sequences or fragment base sequences of genes newly determined by various sequencing methods Get When using data of a plurality of fragment base sequences derived from the same DNA, data of a plurality of fragment base sequences obtained by cutting out a plurality of specific length regions from the full length base sequence may be used. Next, the appearance frequency of the oligonucleotide is obtained by counting the appearance frequency of the oligonucleotide in the full-length base sequence or fragment base sequence of the DNA based on the DNA base sequence data, and the oligonucleotide appearance frequency data storage unit 2 Remember. When the sequence to be analyzed is registered in an existing oligonucleotide appearance frequency database, the data is stored in the oligonucleotide appearance frequency data storage unit 2 from the existing database via the Internet or the like according to an instruction from the input device 11. You just need to capture it. In addition, storage of oligonucleotide appearance frequency data in the oligonucleotide appearance frequency data storage unit 2 can also be performed by a method of directly inputting a numerical value of the oligonucleotide appearance frequency using the input device 11. As a method of inputting a numerical value using the input device 11, manual input using a keyboard or the like as the input device 11, voice input, paper input, or the like can be used.

  The oligonucleotide appearance frequency is preferably the appearance frequency of oligonucleotides having 3 or more bases in order to improve the classification ability. On the other hand, if the appearance frequency of oligonucleotides with too many bases is used, the amount of data to be calculated becomes enormous. Therefore, the appearance frequency of the oligonucleotide is preferably the appearance frequency of 3 to 5 base oligonucleotide (trinucleotide, tetranucleotide, or pentanucleotide). In addition, the oligonucleotide appearance frequency data may be any data that includes a numerical value of the oligonucleotide appearance frequency for a plurality of oligonucleotides, but the oligonucleotides for all types of oligonucleotides having a specific number of bases (for example, 64 types of trinucleotides). It is preferable to include a numerical value of the nucleotide appearance frequency.

In addition, the appearance frequency of the above-mentioned oligonucleotides is the appearance of palindromic oligonucleotides having n bases N (n is an integer from 1 to n _max ; n _max is a natural number, for example, 3) not related to the palindrome. The frequency may be included. That is, for example, when analyzing the hexanucleotide appearance frequency as the oligonucleotide appearance frequency, the appearance frequency of a palindromic oligonucleotide having n bases N (n is an integer of 1 to 3) not related to the palindrome, For example, the appearance frequency of GGGNCCCC, GGGNNCCC, and GGGNNNNCCC may be included in the analysis target. In this case, although the amount of calculation increases, it becomes possible to create SOMs in which close species are separated into different regions.

  The complementary data adding unit 1 adds the appearance frequencies of the oligonucleotides forming a complementary pair, based on the appearance frequency data of each oligonucleotide stored in the oligonucleotide appearance frequency data storage unit 2. The frequency of occurrence of oligonucleotide for each pair is calculated, and data of the calculation result is output.

  The annotation data creation unit 9 performs annotation of the biological meaning of the fragment sequence obtained by referring to the base sequence database or the like. The SOM creation unit 10 arranges the appearance frequency of each of a plurality of types of oligonucleotides in the base sequence as an input vector group on a multidimensional space, and places these input vector groups on a map on which a plurality of grid points are arranged. A self-organizing map is created by non-linear mapping and self-organizing to classify the base sequence into each lattice point. The SOM creation unit 10 preferably performs self-organization by using this data as input vector data based on the data of the frequency of appearance of each complementary oligonucleotide supplied from the complementary data addition unit 1. Create a self-organizing map. By using the data on the frequency of appearance of each complementary oligonucleotide, the calculation time can be shortened and the influence of complementarity can be removed. Based on the appearance frequency data of each oligonucleotide stored in the oligonucleotide appearance frequency data storage unit 2, the SOM creation unit 10 performs self-organization using the data as input vector data. It is also possible to create a conversion map. The SOM creation unit 10 can output (display, print, etc.) the SOM as a two-dimensional or three-dimensional image by outputting the created SOM data via the output device 12. As an output form of the image, for example, on a two-dimensional image, the biological classification to which the base sequence classified into each grid point belongs is represented by color, and the number of base sequences classified into each grid point is expressed by concentration. Forms to be represented: forms in which the number of base sequences classified into each grid point is represented by the height of a bar on a three-dimensional image. In addition to the created SOM data, the SOM creation unit 10 includes identification information (ID) of the base sequence classified into each lattice point on the SOM, and the appearance frequency of the oligonucleotide in the base sequence classified into each lattice point Data and annotation information can be output. When the SOM was created based on the data of the appearance frequency of each pair of complementary oligonucleotides supplied from the complementary data adding unit 1, it was classified into each grid point output from the SOM creating unit 10. The data on the frequency of occurrence of oligonucleotides in the base sequence is the data on the frequency of appearance of oligonucleotides for each pair of complementary (the frequency of appearance of oligonucleotides for each pair of complementary supplied from the complementary data adding unit 1). Data).

  The created SOM is a new, powerful and sensitive tool for comparative genomic analysis. SOM, particularly for bacterial genomes, is useful. That is, SOM for bacterial genomes is particularly useful as a bioinformatics tool to classify DNA sequences obtained from a mixture of microorganisms in the natural environment that are difficult to culture into biological classifications. SOM is an efficient and powerful tool for extracting a wide variety of genomic information from vast amounts of bacterial sequences because of its very high classification ability.

  Most of the microorganisms in the environment are still unculturable. Thus, our understanding of the dominant microbial population in the environment is limited. In order to study the diversity of microorganisms in the environment and to search for completely new and industrially useful genes, sequencing DNA fragments derived from a mixture of microorganisms in the environment such as the extreme environment It is done by many groups. In such studies, DNA is extracted from a mixture of these microorganisms without culturing or cloning the species. Alternatively, the DNA sample is fragmented and cloned into a vector such as a sequencing vector, after which the vector sequence is determined. Therefore, it is important to know how many types of genomic DNA are present in the DNA mixture and how novel their sequences are. SOMs created using unsupervised neural network algorithms in the system of the present invention are powerful bioinformatics tools for obtaining information about genomic DNA in these DNA mixtures. That is, if an SOM is created using a known bacterial genome in advance and the region on the SOM is separated for each biological classification, the sequence of fragments obtained from a mixed sample of bacterial DNA whose components are unknown, Based on where the sequence of the fragment is located on the SOM, it is possible to efficiently predict what kind of microorganisms and how many microorganisms are present in the mixed sample. Therefore, SOM is particularly useful for component analysis of mixed samples containing a plurality of types of microorganisms, such as a mixture of microorganisms in the natural environment that are difficult to culture. Therefore, SOM is useful for searching for new and industrially useful bacteria.

  In addition, since the SOM can recognize the species-specific features (the key combination of oligonucleotide appearance frequency) that are signature features of each genome, an SOM is created using a known DNA sequence in advance. Separating the upper region by species allows species-specific classification of genomic sequences without any information about the species using SOM. Therefore, SOM is useful for identifying to which species a living organism (such as a bacterium) whose species is completely unknown belongs only to a part of the base sequence of its genome. Therefore, SOM is useful for searching for new and industrially useful bacteria.

  As described above, if genome sequences of a wide variety of taxonomic groups are accumulated, SOM can be applied to a broad range of bacterial sequence classifications, and is a powerful, non-culturable microbial mixed DNA sample ( For example, it is considered to be an extremely useful tool for classification of base sequences obtained from samples of seabed sediments.

  In addition, if an SOM is created using a known DNA sequence in advance and an exclusive region almost including a certain base sequence on the SOM is specified, each fragment base sequence cut out from a certain type of DNA sequence If the oligonucleotide appearance frequency is mapped on the SOM and the fragment base sequence corresponding to the position deviated from the specified exclusive region is searched, a segment considered to be introduced from another species can be found through horizontal transmission.

  The mononucleotide composition data storage unit 4 stores mononucleotide composition data in each base sequence to be analyzed. The expected value calculation unit 3 refers to the data of the mononucleotide composition in each base sequence to be analyzed stored in the mononucleotide composition data storage unit 4 and the mono value in the base sequence classified into each lattice point on the SOM. Bases classified into each lattice point based on the value of the mononucleotide composition (content of each mononucleotide in the base sequence classified into each lattice point) obtained for each lattice point The expected value of the appearance frequency of the oligonucleotide in the sequence is calculated. The normalization unit 5 normalizes the appearance frequency of the oligonucleotide in the base sequence classified into each lattice point on the SOM by dividing it by the expected value calculated by the expected value calculation unit 3.

  The appearance frequency map creating unit 6 creates an appearance frequency map representing information on the appearance frequency of the oligonucleotide for each lattice point for each oligonucleotide. The appearance frequency map creation unit 6 can output (display, print, etc.) as a two-dimensional or three-dimensional image by outputting the created appearance frequency map data via the output device 12. Yes. As an output form of the image, for example, a form in which the oligonucleotide appearance frequency of the base sequence classified into each grid point is represented by color and density on a two-dimensional image; In other words, the number of base sequences is expressed by the height of the bar. The appearance frequency map creation unit 6 creates an appearance frequency map based on the appearance frequencies of the oligonucleotides normalized by the normalization unit 5. The appearance frequency map creation unit 6 may create an appearance frequency map based on the appearance frequency data of each oligonucleotide stored in the oligonucleotide appearance frequency data storage unit 2. More preferably, the appearance frequency map is created based on the appearance frequency data for each pair of complementary oligonucleotides supplied from the creation unit 10. Thereby, the calculation time can be shortened and the influence of complementarity can be removed.

  When the base sequence group to be analyzed is a plurality of fragment base sequences extracted from the same genome sequence, the appearance frequency distribution map creation unit 7 generates individual oligonucleotides in the fragment base sequences classified into the lattice points. Based on the frequency, an appearance frequency distribution diagram showing the distribution of the appearance frequency of individual oligonucleotides on the genome sequence is created. The appearance frequency distribution chart creation unit 7 outputs the appearance frequency distribution chart as a one-dimensional or two-dimensional image by outputting data of the created appearance frequency distribution chart via the output device 12 (display, printing, etc.). Be able to. As an output form of an image, for example, on a bar graph corresponding to the genome sequence, a form in which the biological classification to which the fragment base sequence classified into each lattice point belongs is represented by color and concentration; on the two-dimensional image, the genome Examples include a form in which the position on the sequence is represented by x-axis coordinates, and the oligonucleotide appearance frequency of the fragment base sequence corresponding to the position is represented by y-axis coordinates. The appearance frequency distribution diagram creation unit 7 creates an appearance frequency distribution diagram based on the appearance frequencies of the oligonucleotides normalized by the normalization unit 5. It is preferable that the appearance frequency distribution diagram creating unit 7 creates an appearance frequency distribution diagram based on the data on the appearance frequency of each pair of complementary oligonucleotides supplied from the SOM creating unit 10. Thereby, the calculation time can be shortened and the influence of complementarity can be removed.

  Next, the SOM creation unit 10 and the map creation step executed thereby will be described in detail.

  As the SOM creation method in the map creation step, the SOM creation method by the self-organization method of Kohonen (hereinafter referred to as “Kohonen method”), the improved SOM creation method by the self-organization method described in Patent Document 1 (hereinafter, It is preferable to use an improved SOM creation method, although it may be simply referred to as “improved SOM creation method”.

  Here, the basic principle of the self-organization method for creating the SOM will be described. This self-organization method uses a neural network to non-linearly map (map) multi-dimensional input data from a high-dimensional space to a low-dimensional space. Data can be mapped to a low-dimensional space while maintaining the characteristics of the data. In this self-organization method, an input layer in a high-dimensional space where multi-dimensional input data is plotted, an output layer composed of a plurality of output neurons (lattice points) arranged in a grid in a low-dimensional space, A two-layer neural network is used. Then, using the input vector corresponding to the input value and the connection weight vector (neuron vector) representing the connection weight between the point corresponding to the input value and the output neuron, the connection weight vector is set to the initial value (initial connection weight vector). ), Learning is performed by correcting the coupling weight vector.

  In the self-organization method, learning is performed in consideration of the positional relationship between output neurons. In the output layer of a neural network, there is a relative positional relationship (distance relationship) between output neurons. Then, the connection weight vector is corrected for the connection weight vector of the output neuron corresponding to the connection weight vector closest to the input data vector and the output neuron in the vicinity thereof (to the connection weight vector other than the vicinity of the input data vector). But no corrections are made). Thereby, learning of the neural network is performed. The Euclidean distance is used to calculate the distance between the input vector and the coupling weight vector. The coupling weight vector is corrected so that the coupling weight vector approaches the input vector. For example, in order to bring the connection weight vector closer to the input vector, the difference between the connection weight vector of the input vector and the winner neuron is multiplied by a learning coefficient (the learning coefficient is 0 to 1) and then added to the original connection weight vector.

  In this way, the self-organization method considers the positional relationship of the output neurons, and maintains the distance relationship between the input data in the input data space (input layer) while maintaining the input data in the high-dimensional space (output layer). ) Can be mapped.

  The Kohonen method consists of the following three steps. Step 1: A vector (connection weight vector) on each neuron is initialized using a random value. Step 2: Select a neuron having a connection weight vector closest to the input vector. Step 3: Update the connection weight vector of the selected neuron and its neighborhood. Steps 2 and 3 are repeated for the number of input vectors. The learning is performed a predetermined number of times as one learning. After learning, the input vector is classified into the neuron having the closest connection weight vector. Kohonen's SOM can perform non-linear mapping while maintaining features from a group of input vectors in a high-dimensional space to a group of neurons arranged on a low-dimensional map.

  In this Kohonen method, the update to the connection weight vector is performed based on the classification of the connection weight vector for one input in Step 2 and Step 3, so that the later input vectors are separated more finely and the learning order of the input vectors is reduced. Different SOMs are created. Therefore, there is a possibility that a reproducible SOM cannot be obtained. In addition, since the initial connection weight vector setting in step 1 takes a random value, the structure of the random value affects the self-organizing map obtained after learning. It is reflected in. Therefore, there is a possibility that the structure of the input vector cannot be accurately reflected in the SOM. Furthermore, since the random value is taken in step 1, when the initial value is greatly different from the structure of the input vector, a very long learning time is required. In addition, since the connection weight vector is updated based on the classification of the connection weight vector for one input in Steps 2 and 3, the learning time becomes longer in proportion to the number of input vectors.

  On the other hand, the improved SOM creation method is a method of classifying input vector data into coupling weight vectors using a computer by a non-linear mapping method, and (a) using oligonucleotide appearance frequency data as multidimensional input vector data An input step; (b) a step of setting an initial coupling weight vector; (c) a step of classifying the input vector into each coupling weight vector; and (d) an input vector classified into each coupling weight vector and the coupling weight. A step of updating the connection weight vector so as to have a structure similar to the input vector classified in the vicinity of the vector, and (e) steps (c) and (d) until the learning number (repetition number) reaches the set learning number. And (e) classifying the input vector into a combination weight vector and outputting the result as an SOM That and a step.

  In the improved SOM creation method described above, the sequential processing algorithm “classify one input vector into (initial) coupling weight vectors” in the Kohonen method is used, and after classifying all input vectors into coupling weight vectors, By changing to the batch processing learning algorithm of “updating vectors”, a reproducible SOM can be obtained, and even if the number of input vectors increases, the calculation time can be reduced.

  In the step (a), the input vector data may be data of K input vectors (K is a positive integer of 3 or more) having M dimensions (M is a positive integer).

  In the step (b), the initial coupling is performed by reflecting the characteristics of the multidimensional space distribution of the multidimensional input vector obtained by the unsupervised multivariate analysis method in the arrangement and elements of the initial coupling weight vector. It is preferable to set a weight vector. As a result, the structure of the input vector can be accurately reflected in the SOM, and the learning time can be shortened. As the unsupervised multivariate analysis method, principal component analysis, multidimensional scale construction method, or the like can be used.

  In the step (c), as a method for classifying the input vector into each coupling weight vector, a classification method based on a similarity measure selected from a measure including a distance, an inner product, and a direction cosine can be used. Examples of the distance include Euclidean distance.

  In the step (d), the batch processing is also used for updating the connection weight vector so that the input vector classified into each connection weight vector and the input vector classified in the vicinity of the connection weight vector have a similar structure. A learning algorithm can be used.

  It is preferable that the processes of the above steps, particularly the processes using the collective processing learning algorithm, are arithmetically processed using a parallel computer. Thereby, calculation time can be shortened.

  Hereinafter, each step of the improved SOM creation method using the SOM creation unit 10 will be described in detail.

[Step (a)]
First, oligonucleotide appearance frequency data of a plurality of base sequences stored in the oligonucleotide appearance frequency data storage unit 2 is input to the SOM creation unit 10 as multidimensional input vector data.

The input vector data is usually K (K is the number of base sequences) input vectors {x ₁ , x ₂ ,..., X _k , ..., x _K } (k = 1, 2,. ..., K). Each input vector x _k is an M-dimensional vector (M is the number of oligonucleotide segments), and can be expressed by the following equation (1).

xk = { _xk1 , _xk2 ,..., _xkM } (1)
(Where x _k1 , x _k2 ,..., X _kM are oligonucleotide appearance frequencies)
For example, if the oligonucleotide appearance frequency is 64 types of trinucleotides and a plurality of microorganisms are classified based on the trinucleotide usage frequency of the full-length base sequences of K types of DNA, The trinucleotide usage frequency of the full length base sequence of DNA is digitized into 64 dimensions, and the digitized 64 dimensions data is set as an input vector.

  When classifying a base sequence into multiple biological classes, oligos related to a sufficient number of DNA base sequences for each biological class (eg, species) to analyze the characteristics of each biological class more accurately It is preferable to input nucleotide appearance frequency data as input vector data. About hundreds to tens of thousands of DNA base sequences (full-length sequences or fragment base sequences) to be analyzed are usually prepared for each biological classification. It should be noted that the input vector can usually be set according to the ordinary method described in Non-Patent Document 1, Non-Patent Document 2, and the like.

[Step (b)]
Next, a neural network is constructed using a computer, and an initial connection weight vector is set. On the output layer in the neural network, output neurons (lattice points) are arranged in a D-dimensional lattice according to the dimension D of the SOM to be created (D is a positive integer; D <M). The initial connection weight vector of each output neuron (grid point) is also arranged in a D-dimensional grid.

The initial connection weight vector can be set based on a random value as in the SOM creation method described in Non-Patent Document 1, Non-Patent Document 2, and the like. If you want to accurately reflect the structure of the input vector in the SOM or reduce the learning time, rather than setting the initial coupling weight vector based on the random number value, you can use many methods such as principal component analysis and multidimensional scaling. An initial coupling weight vector is set based on _K input vectors {x ₁ , x ₂ ,..., X _K } data having M dimensions set in the step (a) by using a variable analysis method. It is preferable. The initial connection weight vector set in this way is a set of P connection weight vectors {W ⁰ ₁ , W ⁰ ₂ ,..., W ⁰ _P } arranged in a D-dimensional lattice pattern. In this case, each connection weight vector can be expressed by the following equation (2).

W ⁰ _i = F {x ₁ , x ₂ ,..., X _K } (2)
In the expression (2), i is i = 1, 2,. Further, _F in the formula _{(2) {x 1, x} 2, ..., x K} is the input vector _{{x 1, x 2, ...} , x K} representing the transfer function of the initial combining weights Bekutoruhe from . As a specific example, a method of setting initial coupling weight vectors in a two-dimensional (D = 2) and three-dimensional (D = 3) lattice will be described. In accordance with this method, the initial coupling weight vector can be set in a D-dimensional lattice pattern.
(1) Method of setting initial coupling weight vectors in a two-dimensional (D = 2) lattice (when creating a two-dimensional SOM)
These are obtained by performing principal component analysis on _K input vectors {x ₁ , x ₂ ,..., X _K } having M dimensions to obtain a first principal component vector and a second principal component vector. the principal component vectors and _{b 1} and _{b 2.} Based on these two principal component vectors, principal components Z _1k = b ₁ x _k and Z _2k = b ₂ x _k for K input vectors are obtained (k = 1, 2,..., K). _{{Z 11, Z 12, ...} , Z 1k, ..., Z 1K} and _{{Z 21, Z 22, ...} , Z 2k, ..., Z 2K} respectively sigma ₁ standard deviation And σ ₂ .

An average value of the input vectors is obtained, and the obtained average value is defined as x _ave .

A two-dimensional lattice point on the output layer is represented by coordinates on a two-dimensional plane (output layer), that is, ij (i = 1, 2,... I; j = 1, 2,..., J). In this case, a coupling weight vector W ⁰ _ij is placed on a two-dimensional lattice point (ij). The values of I and J may be integers of 3 or more. J is preferably the largest among integers smaller than I × σ ₂ / σ ₁ . The value of I may be appropriately set according to the number of input vector data. The value of I is usually 50 to 1000, for example 100.

W ⁰ _ij can be defined by equation (3).

（２）３次元（Ｄ＝３）の格子状に初期結合重みベクトルを設定する方法
（３次元のＳＯＭを作成する場合）
上述（１）の主成分分析において、第１主成分ベクトルと第２主成分ベクトルに加えて、第３主成分ベクトルを求め、得られた第１主成分ベクトル、第２主成分ベクトル、および第３主成分ベクトルをそれぞれ、ｂ_１、ｂ_２、およびｂ_３とする。これら３つの主成分ベクトルをもとに、主成分Ｚ_１ｋ＝ｂ_１ｘ_ｋ、Ｚ_２ｋ＝ｂ_２ｘ_ｋ、およびＺ_３ｋ＝ｂ_３ｘ_ｋを求める。｛Ｚ_１１，Ｚ_１２，...，Ｚ_１ｋ，...，Ｚ_１Ｋ｝、｛Ｚ_２１，Ｚ_２２，...，Ｚ_２ｋ，...，Ｚ_２Ｋ｝、および｛Ｚ_３１，Ｚ_３２，...，Ｚ_３ｋ，...，Ｚ_３Ｋ｝の標準偏差をそれぞれσ_１、σ_２、およびσ_３とする。３次元の格子点をｉｊｌ（ｉ＝１，２，...，Ｉ；ｊ＝１，２，...，Ｊ；ｌ＝１，２，...，Ｌ）で表現し、３次元の格子点（ｉｊｌ）上に結合重みベクトルＷ^０ _ｉｊｌを置く。Ｉ、Ｊ、Ｌの値は、３以上の整数であれば良い。Ｊは、Ｉ×σ_２／σ_１よりも小さい整数のなかで最大のもの、Ｌは、Ｉ×σ_３／σ_１よりも小さい整数のなかで最大のものが好ましい。Ｉの値は、入力ベクトルのデータ数に応じて適宜設定すればよい。Ｉの値は、通常５０〜１０００であり、例えば１００である。Ｗ^０ _ｉｊｌは、下式（４）により定義することができる

(2) Method of setting initial coupling weight vectors in a three-dimensional (D = 3) grid (when creating a three-dimensional SOM)
In the above principal component analysis (1), in addition to the first principal component vector and the second principal component vector, a third principal component vector is obtained, and the obtained first principal component vector, second principal component vector, Let the three principal component vectors be b ₁ , b ₂ , and b ₃ , respectively. Based on these three principal component vectors, principal components Z _1k = b ₁ x _k , Z _2k = b ₂ x _k , and Z _3k = b ₃ x _k are obtained. {Z ₁₁ , Z ₁₂ , ..., Z _1k , ..., Z _1K }, {Z ₂₁ , Z ₂₂ , ..., Z _2k , ..., Z _2K }, and {Z ₃₁ , Z ₃₂ ,..., Z _3k ,..., Z _3K } are standard deviations σ ₁ , σ ₂ , and σ ₃ , respectively. A three-dimensional lattice point is expressed by ijl (i = 1, 2,..., I; j = 1, 2,..., J; l = 1, 2,..., L). The connection weight vector W ⁰ _ijl is placed on the grid point (ijl) of The values of I, J, and L may be integers of 3 or more. J is preferably the largest integer smaller than I × σ ₂ / σ ₁ , and L is preferably the largest integer smaller than I × σ ₃ / σ ₁ . The value of I may be appropriately set according to the number of input vector data. The value of I is usually 50 to 1000, for example 100. W ⁰ _ijl can be defined by the following equation (4)

〔ステップ（ｃ）〕
すべての入力ベクトル｛ｘ_１，ｘ_２，...，ｘ_Ｋ｝を、各結合重みベクトルヘ分類する。

[Step (c)]
All input vectors {x ₁ , x ₂ ,..., X _K } are classified into the respective connection weight vectors.

Specifically, all input vectors {x ₁ , x ₂ ,..., X _K } are learned (corrected) t times using a similarity measure (distance, inner product, direction cosine, etc.). binding weight vector, of P arranged in D-dimensional lattice shape after ^{_{W t 1, W t 2,}} ..., to one of W ^t _P, it is classified using the computer. Here, t represents the number of times of learning (epoch), that is, how many times step (d) is executed before step (c). When learning is performed a total of T times (T represents the set learning number), t = 0, 1, 2,. . . T. The i-th coupling weight vector at the t-th epoch (during the t-th learning, that is, after execution of the t-th step (d)) can be represented by Wti. Here, i = 1, 2,. . . , P. When t = 0 (when step (c) is executed for the first time), the connection weight vectors W ^t ₁ , W ^t ₂ ,..., W ^t _P are the initial connection weight vectors set in step (b). It corresponds to. The classification of each input vector can be performed by calculating the Euclidean distance with each coupling weight vector W ^t _i and assigning the input vector to the coupling weight vector having the minimum Euclidean distance. In the case of a coupling weight vector arranged on a two-dimensional lattice point (ij), W ^t _i can be expressed as W ^t _ij .

Input vector _{{x 1, x 2, ...} , x K} may be classified into ^W _{t i} by parallel processing for each input vector _{x k.}

[Step (d)]
For each neuron spectrum W ^t _i , the connection weight vector W ^t _i has a similar structure to the input vector (x _k ) classified into the connection weight vector and the input vector classified in the vicinity of the connection weight vector. Update.

That is, let S _{i ′} be a set of input vectors belonging to a lattice point where a specific coupling weight vector W ^t _{i ′} is positioned. _'N number of vector ^x _t 1 belonging to the ^{_{(S i' S i),}} x t 2 (S i ') _{^{and, ..., x t N (S}} i'), from the ^{W t} _{i ', S i} A new connection weight vector (W ^{t + 1} _{i ′} ) reflecting the structure of the input vector belonging to _“′ is obtained by the function G of the following equation (5), so that the neuron vector W ^t _{i ′} (i ′ = 1, 2) , ..., P) is updated.

W ^{t + 1} _{i ′} = G (W ^t _{i ′} , x ^t ₁ (S _{i ′} ), x ^t ₂ (S _{i ′} ),..., X ^t _N (S _{i ′} )) (5)
As a specific example, the update of the connection weight vector Wtij set in a two-dimensional lattice shape will be described. The same can be done for the coupling weight vectors set in other D-dimensional grids.

Belong to W ^t _ij coupling weight vectors input vector x _k are arranged in a two-dimensional grid-shaped, W ^t _ij S the set of input vectors belonging to the lattice points in the vicinity of the grid points are positioned are _ij when a _{was, N ij} inputs the vector ^x _t 1 belonging to _{^{S ij (S ij), x}} t 2 (S ij), ..., from _x t Nij _{(S ij)} and ^W _{t ij,} the _{S ij} By obtaining a new joint weight vector x ^{t + 1} _ij that reflects the input vector structure to which it belongs, the joint weight vector W ^t _ij can be updated.

ここで、Ｎ_ｉｊは、Ｓ_ｉｊに分類された入力ベクトルの総数である。

Here, N _ij is the total number of input vectors classified as S _ij .

  α (t) is a learning coefficient (0 <α (t) <1) for the t-th epoch when the set number of learning times is set to T, and uses a monotonically decreasing function. More preferably, it can obtain | require by the following Formula (7).

設定学習回数Ｔは、入力ベクトルのデータ数に応じて適宜設定すればよい。設定学習回数Ｔは、通常１０〜１０００であり、例えば１００である。

The set learning frequency T may be set as appropriate according to the number of input vector data. The set learning frequency T is usually 10 to 1000, for example 100.

More preferably, the neighborhood set S _ij is a lattice point i′j ′ that satisfies the conditions of i−β (t) ≦ i ′ ≦ i + β (t) and j−β (t) ≦ j ′ ≦ j + β (t). A set of classified input vectors _xij . β (t) is a number that determines the neighborhood, and is obtained by, for example, Expression (8).

β (t) = max {0,25−t} (8)
The connection weight vectors {W ^t ₁ , W ^t ₂ ,..., W ^t _P } can be updated by parallel processing for each connection weight vector W ^t _i .

[Step (e)]
It is determined whether or not the number of learnings (the number of repetitions of step (d)) t has reached the set learning number T. If the learning number t has not reached the set learning number T, the process returns to step (c), (C) and step (d) are performed again. That is, the learning is performed by repeating Step (c) and Step (d) until the learning number t reaches the set learning number T. When the learning number t reaches the set learning number T, the process proceeds to the next step (f).

[Step (f)]
After the learning is completed, the input vector x _k is classified into the connection weight vector W ^t _i by the computer according to the method of step (c), and the result is output. The input vector x _k is classified on the basis of the classification criterion represented by “Z” that reflects the structure of the input vector. That is, when a plurality of input vectors are classified into the same coupling weight vector, it can be seen that the vector structures of these input vectors are very similar. The input vectors {x ₁ , x ₂ ,..., X _K } can be classified by parallel processing for each input vector x _k .

  An SOM is created according to the classification result output in the above step. The created SOM can be visualized by outputting (displaying, printing, etc.) from the output device 12. Creation and display of the SOM can be performed according to the methods described in Non-Patent Document 1, Non-Patent Document 2, and the like. For example, an input vector classification result obtained by setting a coupling weight vector to a two-dimensional lattice point can be displayed as a two-dimensional SOM. Specifically, an appropriate label is assigned to each grid point based on the attribute of the input vector attributed to each grid point of the coupling weight vector having a two-dimensional grid point, and then this label is assigned to the two-dimensional grid. It can be displayed as SOM by screen display or printing. It is also possible to display the value of the total number of input vectors assigned to each grid point as a SOM by screen display or printing on a two-dimensional grid.

As a computer used in each of the above steps, a computer having a high calculation speed is preferable. The steps (a) to (f) need not be performed using the same computer. That is, the result obtained in a certain step may be output to another computer, and the next step may be performed by the computer. In addition, the arithmetic processing of steps (steps (c) to (f)) that can be performed in parallel can be performed in parallel using a computer having a multi-CPU or a plurality of computers. The conventional SOM creation method employs a sequential processing learning algorithm and cannot perform parallel processing. However, the improved SOM creation method employs a batch processing learning algorithm and can perform parallel processing. is there. By enabling parallel processing, it is possible to greatly reduce the calculation time for classifying input vectors. That is, the time required to process the above 6 steps by one processor is T1, T2, T3, T4, T5, and T6. If C processors are processed in parallel, ideally, the time required for each step is , Tl, T2, T3 / C, T4 / C, T5 / C, T6.
T1 + T2 + T3 + T4 + T5 + T6- {T1 + T2 + (T3 + T4 + T5) / C + T6} = (1-1 / C) (T3 + T4 + T5)
Calculation time can be shortened by time.

  Examples of the present invention are shown below.

[Example 1]
(SOM creation method)
In Examples 1 to 3, two-dimensional and three-dimensional SOMs were created by the following method according to the improved SOM creation method described in Patent Document 1 using the system of the above embodiment. The initial binding weight vector was defined by principal component analysis (PCA) instead of random values. This is based on the fact that principal component analysis can classify gene sequences into known biological classifications when analyzing relatively small amounts of sequences. The connection weight vectors (W _ij ) were arranged in a two-dimensional lattice represented by i (= 0, 1,..., I−1) and j (= 0, 1,. I was set to 250. J was defined as the closest integer greater than (σ ₂ / σ ₁ ) × 250 (where σ ₁ and σ ₂ are the standard deviations of the first and second principal components, respectively). The connection weight vector was set and updated by the method described in Non-Patent Document 4. The collective learning SOM program “XanaMine” used in this example was obtained from Xanagen Corporation.

  The sequence data to be analyzed was obtained from “GenBank” (http://www.ncbi.nlm.nih.gov/Genbank/). When the number of undetermined nucleotides (N) in the fragment base sequence cut out from the full length sequence exceeds 10% of the total length (window size; 10 kb or 100 kb) of the fragment base sequence, the fragment base sequence is analyzed. Excluded from the subject. When the number of undetermined nucleotides (N) in the fragment base sequence is 10% or less of the window size, the frequency of occurrence of oligonucleotides is normalized to the length excluding undetermined nucleotides (N) and analyzed. Included in subject.

(SOM for 13 eukaryotic genomes)
In order to investigate the clustering ability of SOM for eukaryotic sequences, the inventors first started by extracting from the 13 eukaryotic genome sequences (total 3 Gb) 300 non-overlapping 300, The frequency of occurrence of trinucleotide, tetranucleotide, and pentanucleotide in 000 10 kb fragment base sequences and about 300,000 100 kb fragment base sequences shifted by 10 kb was analyzed. These genomic sequences include human (Homo sapiens), puffer (Fugu rubripes), zebrafish (Danio rerio), rice (Oryza sativa), Arabidopsis thaliana, Medicago truncatula, Drosophila melanogaster (asterisk) , Caenorhabditis elegans, Dictyostelium discoideum, Plasmodium falciparum, Entamoeba histolytica, Schizosaccharomyces pombe, and cesacsiae bacillus cerevisiae including. For humans, sequences from chromosomes 2, 6, 7, 13, 14, 20, 21, 22, X, and Y, for which nearly complete sequence data are available, were analyzed.

  Then, an improved SOM adapted to genomic information science was created for these fragment base sequences by the method described in Non-Patent Document 4. First, the frequency of occurrence of oligonucleotides (frequency of occurrence of trinucleotides, tetranucleotides, and pentanucleotides) in 300,000 10-kb fragment base sequences was analyzed by principal component analysis, and the first and second principal components were used. An initial coupling weight vector arranged as a two-dimensional grid was set. The set learning count was set to 80. After 80 learning cycles, the oligonucleotide appearance frequency of the 10 kb fragment base sequence can be represented by a final bond weight vector arranged in a two-dimensional grid, and as a result, an SOM was created. The resulting SOM showed a clear species-specific separation. The sequence was clustered primarily in species-specific regions.

  The generated SOM is shown in color by chromatic color, and the grid points including arrays derived from a plurality of species are shown in black by color output (color printing, color display, etc.) ( Not shown). Classification by the initial vector set by the first and second principal components (principal component analysis) is achieved within the SOM (hereinafter abbreviated as “10 kb Tri-SOM”) regarding the trinucleotide appearance frequency of the 10 kb fragment base sequence. By comparison with the classified classification, it was clearly found that sequences from a single species were clustered much more closely within the 10 kb Tri-SOM. Species clustering was further strengthened within SOM (hereinafter abbreviated as “10 kb Tetra-SOM”) and pentanucleotide SOM (hereinafter abbreviated as “10 kb Penta-SOM”) regarding the frequency of appearance of tetranucleotides in the 10 kb fragment base sequence. For example, within 10 kb Tri-SOM, within 10 kb Tetra-SOM, and within 10 kb Penta-SOM, 94%, 97%, and 98% of the human sequence were classified as human regions, respectively.

  In the DNA database, only one of a pair of complementary sequences is registered. Considering the overall characteristics of oligonucleotide appearance frequency within the genome, the difference in appearance frequency between complementary oligonucleotides (eg between AAAC vs. GTTT) is not significant. It should also be noted that the creation of SOMs related to tetranucleotide appearance frequency and pentanucleotide appearance frequency requires a long calculation time.

  Therefore, as an attempt to reduce the calculation time, using the appearance frequency of the degenerate set obtained by adding the appearance frequencies of the complementary oligonucleotide pairs, the SOM regarding the tetranucleotide appearance frequency of the 10 kb fragment base sequence, and the 100 kb fragment An SOM (hereinafter abbreviated as “100 kb DegePenta-SOM”) relating to the appearance frequency of pentanucleotides in the base sequence was prepared. This allowed the computation time to be halved without a noticeable decrease in clustering ability.

  The GC content (G + C%) obtained from the coupling weight vector for each lattice point in the 10 kb Tri-SOM is reflected on the horizontal axis of the 10 kb Tri-SOM and increases from the left to the right of the 10 kb Tri-SOM. The high GC content sequence is located to the right of the 10 kb Tri-SOM. Similar results were obtained for 10 kb Tetra-SOM and 10 kb Penta-SOM. Sequences with the same GC content were separated by a complex combination of oligonucleotide frequencies, resulting in species-specific separation. Within the 10 kb SOM, intraspecific separation is evident. For example, humans were segregated into two major regions within 10 kb Tri-SOM and 10 kb Tetra-SOM. However, within the 10 kb Penta-SOM, human sequences were classified into a single contiguous region. This indicates that despite the wide variation between human 10 kb sequences, SOM recognizes a common feature of pentanucleotide frequency in human sequences.

  Next, SOMs relating to the appearance frequency of trinucleotide, tetranucleotide, and pentanucleotide in about 300,000 100 kb fragment base sequences (abbreviated as “100 kb Tri-SOM”, “100 kb Tetra-SOM”, and “100 kb Penta-SOM”, respectively). These three were collectively referred to as “100 kb SOM”).

  The generated SOM is displayed in color by color output (color printing, color display, etc.), and lattice points including arrays derived from a single species are displayed in different chromatic colors for each species, and lattice points including arrays derived from a plurality of species. Is shown in black. FIG. 2 shows the result of converting this color output result to a black and white image. 2 (a), 2 (b), 2 (c), and 2 (d) show 100 kb Tri-SOM, 100 kb Tetra-SOM, 100 kb Penta-SOM, and 100 kb DegePenta-SOM, respectively. In FIG. 2, C is a nematode region, A is an Arabidopsis region, R is a rice region, D is a Drosophila region, F is a puffer region, Z is a zebrafish region, and H is a human region. Indicates.

  Within 100 kb SOM, inter-species separation was more prominent (not intra-species separation) than within SOM for 10 kb fragment base sequences. Within the 100 kb Tetra-SOM and 100 kb Penta-SOM, all species had one major region (FIG. 2 (b) and FIG. 2 (c)). In addition, the seed region was surrounded by a white continuous grid that did not contain genomic sequences. The species-specific lattice vectors differ between regions even near the region boundaries, and the seed boundaries can be mainly automatically drawn based on the white continuous lattice. Examining the 100 kb SOM in detail, there were several small regions consisting of a small number of sequences with specific features. For example, a small area of Arabidopsis (area indicated by A in FIG. 2) located between the rice area (area indicated by R in FIG. 2) and the puffer area (area indicated by F in FIG. 2) is mainly composed of kinematics. It consists of sequences derived from the somatic region and the subcentromeric region. Intraspecific segregation analysis can provide in-depth information on the detailed structure of individual genomes.

[Example 2]
SOM was able to recognize species-specific combinations of oligonucleotide appearance frequencies, which are typical characteristics of the genome of each species, and identify characteristic appearance frequency patterns. The appearance frequency (observed value) of each oligonucleotide at each lattice point in 100 kb SOM was calculated, and normalized with the expected value of the appearance frequency of each oligonucleotide expected from the mononucleotide composition at each lattice point. The oligonucleotide appearance frequency (ratio of observed value / expected value) normalized at each lattice point is expressed in a two-dimensional lattice-like map similar to SOM (appearance frequency map) for each oligonucleotide. Created for each. Information on the appearance frequency (observed value / expected value ratio R) of the normalized oligonucleotide at each lattice point on the two-dimensional map can be expressed by the color of the lattice point on the two-dimensional map, for example, when outputting in color.

  As an example, lattice points where the frequency of occurrence of oligonucleotides is excessive with respect to the expected value (oligonucleotides appear excessively), that is, lattice points with a ratio of observed value / expected value sufficiently larger than 1, are shown in red. In addition, a lattice point where the appearance frequency of the oligonucleotide is less than the expected value (an oligonucleotide appears too small), that is, a lattice point whose observation value / expected value ratio is sufficiently smaller than 1, is shown in blue. In addition, a lattice point where the appearance frequency of the oligonucleotide is about the same as the expected value (the oligonucleotide appears at the expected value level), that is, a lattice point where the ratio of the observed value / expected value is near 1, is shown in white. Then, the color density of the grid point is set to increase as the ratio of the observed value / expected value goes away from 1.

  3 and 4 show typical examples of the appearance frequency map output in color as described above and converted into a black and white image. 3B is a CAGT appearance frequency map, and FIG. 3C is an AATT appearance frequency map. As a control, 100 kb Tetra-SOM is shown in FIG. FIG. 3 also shows a scale indicating the relationship between the observed value / expected value ratio R and the black density. 4 (b) is an appearance frequency map showing the total appearance frequency of ACAGG and CCTGT, FIG. 4 (c) is an appearance frequency map showing the total appearance frequency of CGACG and CGTCG, and FIG. 4 (d) is a diagram of CGAAA and TTTCG. It is an appearance frequency map which shows the total appearance frequency. As a control, 100 kb DegePenta-SOM is shown in FIG. FIG. 4 also shows a scale indicating the relationship between the observed value / expected value ratio R and the black density.

  In addition, in the appearance frequency map, information on the appearance frequency (observed value / expected value ratio) of the normalized oligonucleotide at each lattice point is obtained in another manner, for example, a three-dimensional three-dimensional representation of a two-dimensional lattice map. You may express by the height etc. in a bar graph.

  The normalization of the oligonucleotide appearance frequency described above makes it possible to examine the oligonucleotide appearance frequency at each lattice point separately from the difference in mononucleotide composition. For example, the difference in the appearance frequency of oligonucleotides containing CG and GC between the base sequences can be detected sensitively by separating from the difference in the GC content between the base sequences. In the appearance frequency map for various tetranucleotides and pentanucleotides, the boundary between the overappearing region and the underappearing region almost exactly coincided with the species boundary. FIG. 3 shows a characteristic example regarding species separation. AATT appeared in excess in rice, Drosophila melanogaster, and nematodes, underexpressed in puffer fish and zebrafish, and moderately appeared in humans and Arabidopsis. CAGT appeared in excess in all three vertebrates (human, puffer, and zebrafish), but under-represented in both two plants (rice and Arabidopsis).

  In addition, an appearance frequency map was created based on a value obtained by adding the appearance frequencies of complementary tetranucleotide pairs. The created appearance frequency map was almost identical to the appearance frequency map based on the appearance frequency of one of the complementary pair of tetranucleotides. Therefore, only the appearance frequency map based on one tetranucleotide of the complementary pair is shown.

  As for the appearance frequency map of pentanucleotide, an example of the appearance frequency map related to DegePenta-SOM is illustrated (FIG. 4). In all three vertebrates (human, puffer, and zebrafish), (ACAGG + CCTGT) and (CGACG + CGGTCG) were over- and under-represented, respectively. (CGAAA + TTTCG) appeared in excess in Drosophila melanogaster and nematodes, and appeared moderately in fission yeast. SOM can achieve species classification by utilizing a complex combination of many oligonucleotides for sequence separation.

Example 3
(Intraspecies differences in human genome sequences)
To elucidate the biological meaning of the characteristic tetranucleotide in Tetra-SOM with respect to the bacterial genome, the inventors have used a palindrome using a restriction enzyme system in bacteria producing 4-base restriction enzymes. The correlation of the appearance frequency of tetranucleotides in the palindrome was examined. The tetranucleotide of the restriction enzyme recognition site (cleavage site) is characteristically under-represented in the genome of bacteria producing 4-base restriction enzymes. SOM correctly recognized the biological properties of this bacterial genome. SOM was able to classify genomic sequences into known biological classifications (in the case of FIG. 2 as species) without using any information other than the frequency of oligonucleotide appearance. Since SOM has a very high classification ability, it should be a powerful information science tool for extracting a wide variety of genome information. To examine SOM's ability to classify within-genome differences within one genome, the inventors analyzed 2.8 Gb high quality draft sequences from the human genome. The inventors of the present application have classified 10-kb non-overlapping sequences obtained from a 2.8 Gb human sequence and Tetra-SOM and Penta-SOM relating to a 100-kb sequence shifted by 10 kb into each lattice point. Was output as a three-dimensional image representing the height of the bar.

  The inventors of the present application normalize the frequency of appearance of individual oligonucleotides (tetranucleotides, etc.) at each lattice point of 10 kb SOM by the mononucleotide composition of each lattice point, and determine the appearance frequency of the normalized oligonucleotide. Calculated. The appearance frequency of the normalized oligonucleotide is a ratio of the value (observation value) of the appearance frequency of the oligonucleotide on the SOM to the expected value of the appearance frequency of the oligonucleotide expected from the mononucleotide composition. Based on the appearance frequency of the normalized oligonucleotide, an appearance frequency map was created in the same manner as described above. The appearance frequency map of each tetranucleotide included an appearance frequency map that appeared in an insufficient amount over the entire area (AACG appearance frequency map), and an appearance frequency map that appeared in an excessive amount over the entire area (TTCC appearance frequency map). Next, the inventors of the present application appear prominently in a limited part within 10 kb SOM (small over-appearance area surrounded by a large under-appearance area in the appearance frequency map), but under-appear over the entire area within 100 kb SOM. We focused on tetranucleotides. One type of these examples corresponded to multiple tetranucleotides containing one CG and two C or G. For this type, a similar local over-occurrence pattern was observed (type A). These tetranucleotides corresponded to components of the GC box, a well-characterized transcription signal. The components of the TATA box, TTAA, ATAA, and ATTA, had a pattern similar to that of type A tetranucleotides, but their sequences were completely different (type B). Other characteristic patterns were also observed, some of which were similar.

  In order to examine the biological meaning of the local over-occurrence pattern of tetranucleotides, the frequency of occurrence of each normalized tetranucleotide at the lattice point (each normalized tetra in the 10 kb sequence classified at this lattice point). (Which represents the frequency of occurrence of nucleotides) was plotted along the chromosome 21q sequence. The results are shown in FIG. In FIG. 5, the appearance frequency of each tetranucleotide is displayed in the same display color as in FIGS. 3 (b) and 3 (c). FIG. 5 also shows the location of the gene.

  The distribution pattern of type A and type B tetranucleotides and several other tetranucleotides were plotted along the chromosome 21q sequence. In the gene-rich region, the observation that type A and type B have a similar pattern and a pronounced appearance (red and white) was observed, indicating that these nucleotides are TATA boxes that are typical examples of transcriptional regulatory signals and It is consistent with the view that it is the core sequence of the GC box. In various parts of the chromosome, the distribution pattern of GATC and AGTA was similar to that of type A and type B. All of these tetranucleotides also appeared at high levels in gene-rich regions on chromosomes 20 and 22 (not shown). This result suggests that these tetranucleotides are likely to be signal sequences or components thereof related to gene expression regulation or function. It has been found that the frequency of occurrence of dinucleotides, trinucleotides, and tetranucleotides within a genome is highly correlated. Due to this basic genomic feature, the inventors of the present application could easily identify a characteristic tetranucleotide. The tetranucleotide appeared under-represented in many regions in Tetra-SOM, whereas its constituent trinucleotide appeared rather in excess in Tri-SOM. The underappearance of the tetranucleotide was not due to the underappearance of its constituent trinucleotides, so the underappearance was observed for the sequence of tetranucleotides in the bacterial genome as observed for the restriction enzyme recognition site sequences. It is thought to reflect the biological meaning. ATTG belongs to this type, and the distribution was concentrated in a gene-poor region (FIG. 5).

  Considering the characteristic oligonucleotides in the genome, those that occur in excess as well as those that are under-represented, a variety of factors, including DNA conformation, context-dependent mutations, and DNA modifications, are responsible It is thought that there is. With respect to sequences that appear in excess, the preference for sequences recognized by abundant DNA binding proteins must be considered. Because interspecific and intraspecific separation using SOM is very clear, SOM is a basic guideline for understanding the detailed molecular mechanisms that have been used to characterize individual genome sequences during evolution. Should provide.

(Determination of gene signal sequence)
A wide variety of oligonucleotide sequences function as gene signal sequences (eg, regulatory signal sequences for gene expression). The finding in FIG. 2 etc. that various tetranucleotides (eg type A and type B) have characteristics adapted to the transcription signal sequence, SOM performs characterization of the signal sequence of the gene and computer prediction (in-silico prediction) It shows the possibility of becoming a new tool for. The signal sequence of a gene, such as a transcription signal sequence, is typically longer than a tetranucleotide. Therefore, longer oligonucleotide analysis is required to test this possibility. Therefore, in the same manner as in the case of tetranucleotides, pentanucleotides that appear in a small amount in the human genome were examined. As a result, there was an example in which the distribution pattern was different between the gene rich region and the gene poor region. The above-mentioned local and specific appearance pattern was often clearer than the tetranucleotide pattern. This suggests that many of the above tetranucleotides are constituents of signal sequences (eg, GC boxes) with longer sequence lengths.

  The signal sequence recognition mechanism and the appearance level of each oligonucleotide sequence throughout the genome are considered to be related to each other. If an oligonucleotide sequence has significant activity, such as high affinity for binding to a specific target protein, the appearance of that oligonucleotide sequence is biased from random appearance and may vary significantly throughout the genome. I will. For example, oligonucleotide sequences with strong binding activity for transcription factors appear under-compared compared to random appearance across many regions of the genome, but will be more frequent within gene regulatory regions. Such signal sequences are under-expressed across the entire SOM with a wide window (eg, 100 kb SOM), but appear more frequently at restricted sites in a SOM with a narrower window (eg, 10 kb SOM). In the opposite case where an oligonucleotide sequence that has binding activity for a known transcription factor appears at random or higher frequency throughout the genome, the combination with other adjacent signal components The sequence should be an absolute prerequisite for functioning as a regulatory signal sequence. The frequency of appearance of oligonucleotide sequences with factor binding activity across the entire genome is the basic information for understanding the mutual role of each oligonucleotide within the combination unit for transcriptional regulation (to identify differences in specificity determination). Will contribute). SOM data on the level of oligonucleotides with factor binding activity will allow to classify different behaviors of different classes of signal sequences (which can be visualized on SOMs with different window sizes). With reference to the behavior of categorized signal sequences collected from well-studied organisms, a computational method useful for predicting signal sequences in the genome that have been sequenced but with little additional experimental data (In-silico method) can be developed.

  In preparation for this approach, pentanucleotides known to have the potential to bind transcription factors were characterized as follows. First, pentanucleotides known as binding sequences for human transcription factors were searched in the TRANSFAC database (http://transfac.gbf.de/TRANSFAC). In the database, there were a total of 22 pentanucleotides reported in the literature as factor binding sequences. However, when these sequences were checked in detail with reference to the MATRIX table in the above database, many of these were components of longer signal sequences that could be selected as the main determining sequence for transcription factor binding. Were four pentanucleotides, NF-Y binding site CCAAT, GATA-1 factor binding site GATAA, KLF binding site CACCC, and NF-1 binding site TGGCA.

  Then, an appearance frequency distribution pattern (appearance frequency map) of these four pentanucleotides in each of 10 kb Penta-SOM and 100 kb Penta-SOM of the human 2.8 Gb sequence is created, and compared with the distribution pattern of the sequence serving as the core of the GC box did. GATAA was under-represented in many areas, both 10 kb SOM and 100 kb SOM. CCAAT appeared under-represented in many regions of 100 kb SOM, but appeared significantly in the restricted region of 10 kb SOM. A detailed comparison of the CCAAT distribution with the distribution of the core component of the GC box showed that the CCAAT-rich region in 10 kb SOM is clearly distinguished from the region with many GC box sequences. This is consistent with the finding that CCAAT (rather than the GC box sequence) was more prevalent in the gene poor region than in the gene rich region of chromosome 21q (FIG. 5). CACCC and TGGCA appeared in excess across many regions of 10 kb SOM and 100 kb SOM. Such frequently occurring sequences are sequences that require other additional specific sequences for normal regulation of the target gene or can correspond to binding sequences for abundant DNA binding factors. . Systematic classification of known signal sequences with respect to their level of occurrence across the genome has been possible since genome sequencing, and a novel method for analyzing the molecular mechanisms underlying accurate signal sequence recognition. provide.

  SOM can visualize the frequency of oligonucleotide appearance in a wide variety of genomes on one map. Appearance frequency patterns (appearance frequency maps) in 10 kb Penta-SOM and 100 kb Penta-SOM (FIG. 2C) of each of the four pentanucleotides described above for 13 eukaryotes were prepared. GATAA was under-represented in all of the above eukaryotes except for Plasmodium falciparum and D. gondii. CCAAT appeared underrepresented in many areas of all three vertebrates (Fugu, zebrafish, human), but two plants (rice and Arabidopsis) and two invertebrates (Nematoda, Drosophila) ) Appeared in excess. These results may provide basic information to understand the mechanism of signal sequence recognition in individual species and the evolutionary process that establishes the signal sequence recognition system. SOM manifests oligonucleotides that are characteristically biased from a random distribution on a two-dimensional map. In addition, since genomic sequences with specific features were self-organized on the map, the genomic positions of all such sequences could be plotted along the chromosome (FIG. 5). Characterizing various species of known signal sequences with sufficient experimental data using SOM and systematically classifying them against a genome that has been determined but otherwise has little experimental data It will be possible to develop a computerized method for predicting the most useful signal sequences. Since the number of such genomes is growing rapidly, the importance of developing such computerized prediction methods is also increasing.

Example 4
In the following examples, the system of the above embodiment is used for 17,000 fragment base sequences (segments) and 5 kb fragment base sequences which are obtained from 81 bacterial genomes with known full-length sequences and do not overlap each other. Using the trinucleotide appearance frequency and the tetranucleotide appearance frequency, two-dimensional and three-dimensional SOMs were constructed according to the improved SOM preparation method described in Patent Document 1. As a first step for obtaining an initial coupling weight vector, as described in Non-Patent Document 4, the appearance frequency of over 17,000 non-overlapping segments was analyzed by principal component analysis (PCA). The set learning count was set to 100. After 100 learning cycles, the oligonucleotide appearance frequency of the fragment base sequence was substantially projected onto the bond weight vector in the SOM. The created SOM output a lattice image including fragment base sequences derived from a single species in a chromatic color, and outputs a lattice image including a plurality of species fragment base sequences in black (not shown). In the SOM prepared using the frequency of appearance of tetranucleotide in the 5 kb fragment base sequence, the fragment base sequences of most species were separated into a plurality of species-specific non-overlapping regions.

  The analysis of the bond weight vector of each lattice point is performed using SOM created using the dinucleotide appearance frequency, SOM created using the trinucleotide appearance frequency, and SOM created using the tetranucleotide appearance frequency (each “Di−”). The GC content (G + C%) for each joint weight vector in “SOM”, “Tri-SOM”, and “Tetra-SOM”) is projected mainly on the horizontal axis of the SOM and goes from left to right of the SOM. Therefore increased. In SOM, AT-rich bacterial sequences were distributed on the left and GC-rich bacterial sequences on the right. Importantly, sequences with the same GC content (G + C%) were separated by a complex combination of oligonucleotide frequencies, resulting in species-specific separation. In other words, many 10 kb fragment base sequences in each genome have a combination of oligonucleotides reflecting each genome like a signature. The SOM can specify a signature as a representative binding weight vector.

  170,000 non-overlapping 1 kb fragment base sequences were similarly analyzed to prepare SOM. In this case, the degree of separation was somewhat smaller, but species-specific separation was again observed. This shows that species-specific features (signatures) can be detected even in the main group of 1 kb fragment sequences.

  Classification into individual species is not important at the beginning of the classification of diverse microorganisms that cannot be cultivated, but phylogenetic classification is important. Therefore, when the classification into 12 main phylogenetic groups by 2D and 3D SOM was tested, a good classification was actually detected. Specifically, when SOM regarding the frequency of oligonucleotide appearance was applied to the classification of about 100 bacteria, about 90% of the 5 kb bacterial sequence was converted into 12 phylogenetic groups, namely, alpha proteobacteria, beta proteo Bacteria (Betaproteobacteria), γproteobacteria, δproteobacteria, Archaea (Archaea), Chlamydia, Firmicutes, Actinobacteria, Fusoacteria ), Super thermophilic gram-negative anaerobic bacilli (Thermotogae), etc. Using the SOM created above, systematic estimation of difficult-to-culture microorganisms obtained from the environment was performed. In GenBank, the base sequences of environmental microorganisms whose biological species are not specified are registered. Classification was made into phylogenetic groups using 660 non-rDNA sequences other than the ribosomal RNA gene (rDNA) sequence having a base length of 1 kb or more among the base sequences for which no biological species were specified. Compared to rDNA sequences, non-rDNA sequences are difficult to classify into phylogenetic groups by conventional homology analysis. Of the 660 non-rDNA sequences, 343 were derived from microorganisms collected from the rumen. In the classification of these 343 sequences on the SOM, most of them are classified into the area of anaerobic microorganisms such as methanogens and sulfur-degrading bacteria. Met. Therefore, it was shown that it is possible to estimate the phylogeny and diversity of microbial groups mixed in the environment.

  SOM often resulted in the separation of each species into two major regions of approximately equal numbers of grid points. Region separation within the same species by SOM was related to transcription direction. Separation within the same species was dependent on the length of the sequence used in the SOM. For genomes where a fairly long range of neighboring genes have the same transcriptional direction, segregation within the same species was prominent even in SOMs analyzed for frequency of occurrence within the 10 kb sequence. However, in genomes where the transcription direction changes frequently even in a short range of genomes, such as a few kb, separation within the same species is less pronounced in a 10 kb SOM.

Example 5
In the case of genomic sequences of short fragments such as 1 kb, the direction of transcription is often unknown, and the distinction of transcription direction is not important and may cause complexity for identification between species. Absent. Taking this into account, the frequency of appearance of a pair of complementary tetranucleotides was added.

  Appearance of 1 kb sequences of 81 species of bacteria, with the frequency of occurrence of oligonucleotides forming complementary pairs of dinucleotides (2 consecutive bases), trinucleotides (3 consecutive bases), and tetranucleotides (4 consecutive bases) added. Using the frequency (degenerate appearance frequency), the same SOM creation processing (frequency analysis) as in Example 4 was performed. That is, in order to eliminate the influence of complementarity of genome sequences, two complementary pairs of oligonucleotides (for example, AA and TT) are regarded as the same, and the appearance frequency obtained by adding these appearance frequencies is used. The SOM creation process was performed. For example, an SOM was prepared in the same manner as in Example 4 except that 32 pairs of complementary trinucleotide appearance frequency data was used instead of 64 trinucleotide appearance frequency data. As a result, the same result as in Example 4 was obtained. Separation within the same species was also reduced.

Example 6
Genomic segments introduced through horizontal transmission from distantly related organisms are known to retain the sequence characteristics of the donor genome and can be distinguished from those of the received genome. The inventors have previously shown that SOM is useful for identifying horizontally transmitted genes and, importantly, for predicting the donor genome of the transmitted genes. (Non-Patent Document 3). In SOM, there may be a unique grid point located far from the main area of each species. Sequences with oligonucleotide frequencies that are distinctly different from the main region should correspond at least in part to portions of the genome that are transmitted horizontally from other genomes.

  In order to visualize these sequences with different oligonucleotide compositions from the regions of their species themselves, the inventors have identified the region of E. Coli and the region of the closely related bacterium S. typhimurium. A 10 kb sequence from E. Coli located outside of both was examined. Excluding the sequence of the S. typhimurium region, the second highest number of spots in the E. Coli sequence were found in the Y pestis region. Subsequently, attention was focused on five sequences present within the Y pestis region commonly found in dinucleotides, trinucleotides, and tetranucleotides. Within these sequences were 37 known genes (23 of which had significant homology to the Y pestis gene). For example, at the amino acid level, 6 out of 23 proteins encoded by these genes have an identity level exceeding 60%, and the one with the highest identity level has an identity of 80%. Had a level. This identity was significantly higher compared to the average identity value of 40% calculated for the vertical pair E. Coli protein and Y pestis protein. In addition, three genes were homologous to the phage encoded gene and one gene was homologous to the transposon gene. These findings support the prediction that these genes may have been transmitted horizontally from other organisms to the E. Coli genome.

Example 7
It was found that species-specific features of the frequency of occurrence of palindromic tetranucleotides were observed. For example, restriction enzyme target tetranucleotides encoded by their own genomes were particularly under-represented. Furthermore, palindromic oligonucleotides are known to be target sites for various proteins such as transcriptional regulatory proteins. The frequency and distribution of palindromic oligonucleotides may clearly differ from the results predicted for these oligonucleotides to appear randomly. It is therefore interesting to focus on palindromic oligonucleotides. Note that in the case of palindromic hexanucleotides the oligonucleotide type is 64, and in the case of palindromic octanucleotides the oligonucleotide type is 256. In other words, even long sequences with important biological meaning could be noted.

  The results of the analysis clearly revealed species separation and the frequency of palindromic oligonucleotides more clearly represented species-specific characteristics. In the case of a restriction enzyme recognition sequence, and in the case of a recognition sequence further comprising a DNA binding protein, there are examples in which the internal base is not involved in the recognition. For example, the recognition sequence of Tth1111 is GACN3GTC (wherein any base can be selected at the “N” position). Taking this into account, the inventors have included the following types of partially palindromic oligonucleotides in the analysis. That is, a palindromic oligonucleotide having n bases not related to the palindrome (n is an integer of 1 to 3), for example, GGGNNCCC, GGGNNCCC, and GGGNNCNCC are analyzed as palindromic oligonucleotides used for recognition Included. When true hexanucleotide (n = 0) was added to the analysis target at the recognition site of the oligonucleotide of this palindrome, 256 (64 × 4) variables could be analyzed.

  This analysis allows attention to oligonucleotides that may have a high biological specificity. Thus, SOM can efficiently identify species with the same type of restriction enzyme or DNA binding protein with similar recognition sequences. Therefore, it is considered that SOM can strongly separate related species (bacteria of different species belonging to γ-proteobacteria) as well as phylogenetic classification. It is possible to subdivide a wide variety of base sequences by first classifying the base sequences into phylogenetic groups using ordinary oligonucleotide SOM and subdividing them using palindromic oligonucleotides. become.

  As described above, the classification system of the present invention classifies the base sequence into the biological classification, analyzes the components of the mixed sample containing a plurality of kinds of organisms, searches for new and industrially useful bacteria, and the like in the genome sequence. It can be used to predict segments introduced from other species through horizontal transmission.

  In addition, as described above, the analysis system of the present invention can be used for searching for important oligonucleotides that are the key to separating biological classifications, searching for regions containing many signal sequences in genomic sequences, and the like.

It is a block diagram which shows the structure of the oligonucleotide appearance frequency analysis system which concerns on one Embodiment of this invention. It is a figure which shows the example of the SOM produced in the example of implementation of this invention, (a) is the SOM created using the trinucleotide appearance frequency, (b) is the SOM created using the tetranucleotide appearance frequency (C) shows the SOM created using the pentanucleotide appearance frequency, and (d) shows the SOM created using the pentanucleotide appearance frequency obtained by adding the appearance frequencies of two pentanucleotides forming a complementary pair. It is a figure which shows the example of the SOM created in the example of implementation of this invention, and the appearance frequency map, (a) is SOM created using the tetranucleotide appearance frequency, (b) is the appearance frequency map of CAGT, ( c) shows an AATT appearance frequency map. It is a figure which shows the example of the SOM created in the example of implementation of this invention, and an appearance frequency map, (a) is using the pentanucleotide appearance frequency which added the appearance frequency of two pentanucleotides which make a complementary pair. The created SOM, (b) shows the total frequency map of ACAGG and CCTGT, (c) shows the total frequency map of CGACG and CGTCG, and (d) shows the total frequency map of CGAAA and TTTCG. It is a tetranucleotide appearance frequency distribution map which shows distribution of the normalized appearance frequency of some tetranucleotide on the chromosome 21q.

Explanation of symbols

1 Complementary data adder (adder)
2 Oligonucleotide appearance frequency data storage unit 3 Expected value calculation unit 4 Mononucleotide composition data storage unit 5 Normalization unit 6 Appearance frequency map creation unit 7 Appearance frequency distribution diagram creation unit 10 SOM creation unit

Claims (4)

The frequency of appearance of each of multiple types of oligonucleotides in a base sequence is arranged as an input vector group on a multidimensional space, and these input vector groups are non-linearly mapped onto a map on which a plurality of grid points are arranged. A base sequence classification system that creates a self-organizing map by self-organizing a base sequence into each lattice point,
An oligonucleotide frequency data storage unit that stores data on the frequency of appearance of each of a plurality of types of oligonucleotides in a plurality of base sequences;
By extracting the data on the appearance frequency of each oligonucleotide stored in the oligonucleotide appearance frequency data storage unit and adding the appearance frequencies of the complementary oligonucleotides, An adder that calculates the frequency of appearance and outputs data of the frequency of occurrence of the calculated oligonucleotide for each pair ;
-Out based on the data of the frequency of occurrence of oligonucleotides for each pair of output from the adding section, the self-organizing by performing the self-organizing the frequency of occurrence of oligonucleotides for each pair as the input vector group A base sequence classification system comprising a self-organizing map creating unit that creates a map and outputs data of the created self-organizing map. An oligonucleotide frequency data storage unit that stores data on the frequency of appearance of each of a plurality of types of oligonucleotides in a plurality of base sequences;
By extracting the data on the appearance frequency of each oligonucleotide stored in the oligonucleotide appearance frequency data storage unit and adding the appearance frequencies of the complementary oligonucleotides, An adder that calculates the frequency of appearance and outputs data of the frequency of occurrence of the calculated oligonucleotide for each pair;
Based on the appearance frequency data of each pair of oligonucleotides output from the adder, the appearance frequency of each pair of oligonucleotides is arranged on a multidimensional space as an input vector group, and these input vector groups are A self-organizing map that non-linearly maps onto the map where the grid points are placed and classifies the above base sequence into each grid point , creates a self-organizing map, and outputs the data of the created self-organizing map An organizational map creation department;
Based on the appearance frequency data of each pair of oligonucleotides output from the above adder, an appearance frequency map is created for each oligonucleotide that shows information about the appearance frequency of each pair of oligonucleotides for each lattice point. And an appearance frequency map creating unit that outputs data of the created appearance frequency map, and an oligonucleotide appearance frequency analysis system. A mononucleotide composition data storage section storing mononucleotide composition data in each base sequence to be analyzed;
Retrieve data mononucleotide composition in each base sequence of the analyte that is stored in the mononucleotide composition data storage unit, based on the mononucleotide composition in nucleotide sequences classified into each grid point extracted, each grid point An expected value calculation unit that calculates the expected value of the appearance frequency of the oligonucleotide in the base sequence classified into, and outputs the calculated expected value;
Receiving the frequency data of each pair of oligonucleotides output from the adder and the expected value calculated by the expected value calculator, the frequency of occurrence of oligonucleotides in the base sequence classified into each lattice point Is further divided by the expected value, and a normalization unit that outputs normalized appearance frequency data ,
3. The oligonucleotide according to claim 2, wherein the appearance frequency map creation unit creates an appearance frequency map based on the appearance frequency data of the oligonucleotide normalized by the normalization unit. Appearance frequency analysis system. An oligonucleotide appearance frequency data storage unit storing data of appearance frequencies at which multiple types of oligonucleotides each appear in a plurality of fragment base sequences extracted from the same DNA sequence;
Take out the data of the appearance frequency of the oligonucleotide stored in the above-mentioned oligonucleotide appearance frequency data storage unit, and calculate the appearance frequency of the oligonucleotide for each pair by adding the appearance frequency of the complementary pair of oligonucleotides And an adding unit that outputs data of the appearance frequency of the oligonucleotide for each pair calculated,
Based on the data on the frequency of appearance of each pair of oligonucleotides output from the adder, the frequency of appearance of each pair of oligonucleotides is placed in a multidimensional space as an input vector group, and these input vector groups are multidimensional. By creating a self-organizing map in which the above-mentioned fragment base sequence is classified into each lattice point by creating a non-linear mapping by self-organization onto a map in which a plurality of lattice points are arranged from space, the self-organization created A self-organizing map creation unit that outputs map data;
Based on the appearance frequency of each oligonucleotide in the fragment base sequence classified into each lattice point, an appearance frequency distribution diagram showing the distribution of the appearance frequency of each oligonucleotide on the DNA sequence is created, and the created appearance frequency An appearance frequency distribution map creation unit for outputting distribution map data,
The appearance frequency distribution map creation unit creates an appearance frequency distribution map based on the appearance frequency data of each pair of oligonucleotides output from the addition unit. Frequency analysis system.

Images