JP5852902B2 - Gene interaction analysis system, method and program thereof - Google Patents

Description

  The present invention relates to a gene interaction analysis system, a method and a program for analyzing an effect caused by an interaction between genes.

In various species, it is well known that genes on the genome are involved in the expression of phenotypes that show the biological characteristics of individuals.
In addition, when a plurality of genes are considered, each individual gene does not affect the phenotype, but the combination of the plurality of genes may affect the phenotype. At this time, it is difficult to search for the effect of each gene and search for the mutual effect by the combination of a plurality of genes based on the search result.

In addition, a technology for exhaustively searching for mutations in the human genome has been put into practical use, and SNP (Single Nucleotide Polymorphism) of genes in the genome is used to identify human phenotypes, particularly the onset of diseases and individual differences in drug side effects. Based on single nucleotide polymorphisms or single nucleotide polymorphisms), it is possible to examine gene mutations genome-wide (see, for example, Patent Document 1).
In particular, elucidating the relationship between the onset of a disease (an example of phenotypic expression) and a SNP in a gene is important for enabling the prevention of the onset and maintaining human health.

As described above, the expression (quality) of phenotypes in the genotype (genotype) possessed as a human individual has the effect that multiple genes form a complex network or develop or progress disease. Expression or quantitative traits).
Therefore, elucidating the influence of interaction (epistasis) between SNPs on the expression of traits on a genome-wide basis can prevent the onset of diseases in advance and enable useful gene therapy even after the onset Is important.

JP 2005-527904 A

When searching for a combination of SNPs and SNPs that affect the expression of traits in the genome-wide manner described above, in the interaction between the different SNPs described above, the analysis takes into account the combinations of SNPs at a plurality of different loci, and a huge amount of data It becomes analysis.
For example, in the case of combinations of two SNPs, there are about 500 billion combinations that take two different SNPs from 900,000 types of SNPs. For this reason, when the conventional mathematical model is used, if the number of specimens used for confirming the effect is considered to be about 1000, the limit of the mathematical model itself or the computational complexity when the mathematical model is handled by a computer is exceeded. End up.

For this reason, the conventional method of searching for a combination of SNPs that affect the expression of a trait in a genome-wide manner performs filtering for selecting a part of the gene region of the genome-wide gene, and selects the selected gene region. By applying it to a mathematical model for analysis processing, analysis considering a plurality of loci is performed.
However, when this gene filtering is used, there may be a case where a gene in a region not selected by filtering is a gene having a SNP having a main influence on expression of a phenotype.
For this reason, as described above, when a gene that actually affects the onset of a disease is unfortunately filtered and removed from the search target, a combination of SNPs involved in the onset of the disease can be identified. Disappear.

  The present invention has been made in view of such circumstances, and easily searches for combinations of SNPs that affect the expression of traits in genome-wide genes without performing selection by gene filtering. It is an object of the present invention to provide a gene interaction analysis system capable of analyzing expression, a method and a program thereof.

  The present invention has been made to solve the above-described problems, and the gene interaction analysis system of the present invention is a gene that affects phenotypic expression from genotype data of genome-wide SNPs (single nucleotide polymorphisms). A gene interaction analysis system that comprehensively identifies SNP pairs and generates a regression model for analyzing the expression of traits from the SNP pairs, which is detected from n specimens. From the m types of SNPs selected, SNP pairs consisting of two different types of SNPs are selected sequentially, and each genotype (combination of dominant type and inferior type) of the SNPs in the SNP pair is a dummy consisting of two categories. Classify into a table of codes, classify each of the n samples for each dummy code, and belong to two categories and each category A table creation unit that creates a table indicating the specimen, and a single regression model using the specimen classified into each category by a single regression model using a dummy code in the table as a dummy variable. An evaluation value calculation unit that calculates an evaluation value indicating the strength of association with the expression of the trait for each dummy variable, and the dummy variables are arranged in descending order of the evaluation value indicating the strength of the association A ranking creating unit for generating a ranking, extracting the dummy variables up to a preset order, a dummy variable creating unit using the extracted dummy variables as dummy variables of the regression model, and the extracted dummy variables And a variable selection unit that calculates a regression coefficient to be multiplied by a penalized maximum likelihood method.

  In the gene interaction analysis system according to the present invention, the table creation unit uses a CDC (cell-wise dummy coding) as a coding method, and a CDC of a dummy code in which genotypes in one type of SNP are classified into two categories. A CDC table creation unit configured as a table and a dummy code CDC interaction table in which genotypes in two different types of SNPs are classified into two categories of a combination of one genotype and another genotype; An ADC table creation unit which uses ADC (adaptive dummy coding) as a coding method and uses a plurality of combinations of genotypes in two different types of SNPs as ADC interaction tables of dummy codes classified into two categories; Characterized in that it is al configured.

  In the gene interaction analysis system of the present invention, when the regression model is a logistic regression model, each of the CDC table, the CDC interaction table, and the ADC interaction table includes two categories and the quality of each category. In the single regression analysis of the dummy code of each of the CDC table and the CDC interaction table. A CDC likelihood calculating unit that calculates likelihood as the evaluation value; a CDC_p value calculating unit that calculates p value as the evaluation value in the single regression analysis of each dummy code of the CDC table and the CDC interaction table; , In the single regression analysis of each dummy code of the ADC interaction table Wherein there between ADC_p value calculation unit for calculating a p value as serial evaluation value.

  In the gene interaction analysis system of the present invention, the ranking creation unit generates the ranking by arranging the likelihood of each of the dummy codes in the CDC table and the CDC interaction table in the order of strong relevance, Further, the ranking is generated by arranging the p values of the dummy codes of the CDC table and the CDC interaction table in the order of the relevance, and the p of each dummy code of the ADC interaction table is generated. The ranking is generated by arranging values in the order of strong relevance.

  In the gene interaction analysis system of the present invention, when the regression model is a linear regression model, the CDC interaction table and the ADC interaction table include two categories and the quantitative traits of the specimen included in each of the categories. A CDC_t value calculation unit that calculates a t value as the evaluation value in the single regression analysis of each dummy code of the CDC interaction table, and the ADC mutual value An ADC_t value calculation unit that calculates a t value as the evaluation value in the single regression analysis of each dummy code in the action table.

  In the gene interaction analysis system according to the present invention, the ranking creation unit generates the ranking by arranging t values of the dummy codes of the CDC interaction table in the order of strong association, The ranking is generated by arranging t values of the dummy codes of the action table in the order of the relevance.

  The gene interaction analysis method of the present invention is a combination of SNPs having gene-gene interactions that affect phenotypic expression from genome-wide SNP (single nucleotide polymorphism) genotype data. A gene interaction analysis method for generating a regression model for identifying and analyzing the expression of a trait from the SNP pair, wherein the table creation unit sequentially selects m types of SNPs detected from n samples. A SNP pair consisting of two different types of SNPs is selected, and each genotype (combination of dominant type and inferior type) of the SNPs in the SNP pair is classified into a table of dummy codes consisting of two categories. A table for classifying each of the n samples into two categories and creating a table indicating two categories and the samples belonging to each category And the evaluation value calculation unit performs a single regression analysis using the specimen classified into each category by a single regression model using the dummy code in the table as a dummy variable, and the expression of the dummy variable and the character An evaluation value calculation process for calculating an evaluation value indicating the strength of relevance for each dummy variable, and a ranking creating unit, in order from the highest evaluation value indicating the strength of the relevance, A ranking creation process in which ranking is generated by arranging, and a dummy variable creation unit extracts the dummy variables up to a preset order, and uses the extracted dummy variables as dummy variables of the regression model And a variable selection step of calculating a regression coefficient to be multiplied by the extracted dummy variable by a penalized maximum likelihood method.

  The program of the present invention comprehensively identifies SNP pairs which are SNP combinations having gene-gene interactions that affect phenotypic expression from genotype data of genome-wide SNPs (single nucleotide polymorphisms). A gene interaction analysis program for generating a regression model for analyzing the expression of a trait from a pair, in which a computer starts from m types of SNPs detected from n specimens, and sequentially differs from two types of SNPs. SNP pairs are selected, and each genotype (combination of dominant type and inferior type) of the SNPs in the SNP pair is classified into a table of dummy codes consisting of two categories, and n specimens for each dummy code Creating means for classifying each of the two and creating a table showing two categories and the specimen belonging to each category The single regression model with the dummy code in the table as a dummy variable is used to perform a single regression analysis using the specimen classified into each category, and an evaluation showing the strength of the association between the dummy variable and the expression of the trait Evaluation value calculating means for calculating values for each dummy variable, ranking creating means for generating ranking by arranging the dummy variables in descending order of evaluation values indicating the strength of the relevance, preset order As dummy variable creation means for extracting the dummy variables up to and using the extracted dummy variable as a dummy variable of the regression model, a variable selection means for calculating a regression coefficient to be multiplied by the extracted dummy variable by a maximum likelihood method with penalties It is a program to make it function.

  According to the present invention, in genome-wide gene analysis using a conventional statistical analysis method, genotype data is classified into two categories rather than a predetermined range in the entire gene, and the influence of the interaction between genotypes is strengthened. SNP genotype interaction in the entire human gene by selecting a dummy code that has a strong influence on the single regression analysis from the dummy data corresponding to each table. The influence on the qualitative trait or quantitative trait by can be identified, and a highly accurate regression analysis can be performed using a dummy code corresponding to the interaction between genes as an explanatory variable used for the regression analysis.

It is a block diagram which shows the structural example of the gene interaction analysis system by one Embodiment of this invention. An example of the SNP basic table in which the number of specimens of specimen phenotypes for each SNP genotype data is tabulated in SNP units stored in the SNP database 11 is shown. An example of the SNP interaction basic table showing the classification of the qualitative trait class of the specimen corresponding to the secondary interaction for each genotype data of two different types of SNPs in all SNPs is shown. An example of a CDC table showing a classification of a class of qualitative traits in a category for each category of a combination of any two of the three genotype data and the other one of the genotype data in the SNP basic table FIG. An example of a CDC table showing a classification of a class of qualitative traits in a category for each category of a combination of any two of the three genotype data and the other one of the genotype data in the SNP basic table FIG. It is a figure which shows an example of the CDC interaction table which divided | segmented the combination of the genotype data of two different types of SNP in the SNP interaction table shown in FIG. 3 into two categories. It is a figure which shows an example of the ADC interaction table which divided | segmented the combination of the genotype data of two different types of SNP in the SNP basic table shown in FIG. 2 into two categories. It is a conceptual diagram explaining the calculation of likelihood or p value when performing a single regression analysis for each CDC table and using the objective variable indicated by the CDC table. An example of a CDC table showing a classification of a class of qualitative traits in a category for each category of a combination of any two of the three genotype data and the other one of the genotype data in the SNP basic table FIG. It is a flowchart which shows the operation example of selection of a dummy variable in the regression model using a gene interaction analysis system, and calculation of a regression coefficient. It is a schematic block diagram which shows the structural example of the gene interaction analysis system by 2nd Embodiment of this invention. The structural example of the SNP interaction basic | foundation table in 2nd Embodiment is shown.

The present invention is intended to comprehensively analyze genome-wide effects on the expression of traits caused by gene-gene interactions in genome-wide genes, for example, the effects of gene-gene interactions on disease expression. The SIS (Sure Independence Screening) method is applied to the above algorithm. The SIS method is known as an optimal algorithm for case-control research because it has been theoretically proved by a generalized linear model.
The present invention uses this SIS method to code dummy variables between SNP (single nucleotide polymorphism) genotypes or between two different types of SNP genotypes (create a dummy code), and This algorithm combines ranking of dummy variables to be obtained, selection of dummy variables from the ranking up to a preset order, and calculation of a regression coefficient to be multiplied by the selected dummy variable. The present invention also relates to a technique for creating a regression model of phenotypic expression from genome-wide SNPs from data of small samples and high-dimensional variables using this algorithm.

<First Embodiment>
In this embodiment, the SNP genotype data used in the logistic regression model, which is a mathematical model, is extracted from the SNP genotype data of different types of loci extracted from a large number of specimens which are large-scale data. Regression analysis is performed using the extracted SNP genotype data.
That is, in the m-dimensional linear function in the genotype data of m types of SNPs shown in the following equation (1), the explanatory variable that is actually used is selected from the explanatory variables x _j (1 ≦ j ≦ m).

Then, using the selected q (<< m) types of explanatory variables x _i (1 ≦ i ≦ q), regression analysis is performed using the q-dimensional linear function of the following equation (2).

In the process of extracting the SNP genotype data used for regression analysis from the many SNP genotype data described above, the dummy code obtained by dividing the genotype data for each type of SNP extracted from the specimen into two categories is used. . Then, a 2 × 2 evaluation table is generated using this dummy code, a single regression model is created for each table, and an evaluation value such as likelihood or p-value is used for each single regression model using single regression analysis. Are ranked, and a dummy code having an evaluation value within a preset ranking (for example, up to 256th) is selected as a dummy variable (explanatory variable). Then, the regression coefficient β _j that is multiplied by this selected explanatory variable is extracted as 0. Then, 0 and does not extract the explanatory variable x _i which corresponds to the regression coefficient (non-zero), to produce a logistic regression model for a regression analysis of qualitative traits from genotype data of SNP.
Hereinafter, the configuration for performing the above-described processing will be described in detail.

Next, a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic block diagram showing a configuration example of a gene interaction analysis system corresponding to genome-wide genes according to the first embodiment of the present invention.
In FIG. 1, the gene interaction analysis system of this embodiment includes a basic table creation unit 1, a CDC (cell-wise dummy coding) table creation unit 2, an ADC (adaptive dummy coding) table creation unit 3, and a CDC likelihood calculation. Unit 4, CDC_p value calculation unit 5, ADC_p value calculation unit 6, ranking creation unit 7, dummy variable creation unit 8, variable selection unit 9, regression analysis processing unit 10, SNP database 11, CDC likelihood storage unit 12, CDC_p value A storage unit 13, an ADC_p value storage unit 14, and a ranking storage unit 15 are provided.

  In the SNP database 11, m (for example, SNPs at gene loci in genome-wide genes exceeding 500,000 locations) extracted from n specimens are assigned to each of n nenotype data. Stored in correspondence with the specimen identification information. Further, in the SNP database 11, each specimen has a binary qualitative trait, whether it is a predetermined disease, whether there is a side effect on a predetermined drug, or whether it is obese. Attribute data as an expression type is given. This SNP database 11 stores only specimens for which SNP genotype data is obtained in the process of extracting SNP genotype data, with no loss of data on the locus at which the SNP was detected in other specimens. . Therefore, the SNP genotype data in the present embodiment has data of samples that require general imputation removed.

  Next, FIG. 2 shows an example of an SNP basic table in which the number of specimens of the specimen phenotype is tabulated for each SNP genotype data in SNP units stored in the SNP database 11. In the SNP database 11, since this SNP basic table is created in SNP units existing at different loci, m pieces are stored for one qualitative trait. Here, in the SNP database 11, the SNP basic table is written and stored in correspondence with the SNP basic table identification information for identifying the SNP basic table. In FIG. 2, three attributes of SNP = 0, SNP = 1, and SNP = 2, for example, respectively correspond to genotypes AA, AB, and BB as SNP genotype data. Here, for example, A is the dominant type, B is the inferior type, and the genotype is a combination thereof. The SNP basic table in FIG. 2 shows whether each specimen for each genotype is affected (= 1) or not affected (non-affected = 0) by the person who collected the specimen as a phenotype. ) Is classified into one of the binary data classes.

When genotype SNP = 0, there are n00 unaffected specimens and n10 affected specimens. When genotype SNP = 1, there are n01 unaffected specimens and n11 affected specimens. When genotype SNP = 2, there are n02 non-affected specimens and n12 affected specimens. Here, in the present embodiment, “xx” in “nxx” is a subscript.
Here, n00 + n10 + n01 + n11 + n02 + n12 = n (total number of samples).
In other words, the SNP basic table is divided into 2 × 3 for n specimens by two classes of qualitative traits and three types of attributes using genotype data as category data for one type of SNP. It is a table.

  Returning to FIG. 1, the basic table creation unit 1 sequentially reads out the SNP genotype data of each sample from the SNP database 11 in units of SNPs, and classifies each sample into corresponding genotype data. Further, the basic table creation unit 1 classifies each genotype data specimen into two classes of qualitative traits, and sets the classified result as the SNP basic table shown in FIG. As described above, the basic table creation unit 1 generates the SNP basic table of FIG. 2 for all types (m) of SNPs.

  Next, FIG. 3 shows the SNP interaction basic table showing the classification of the class of qualitative traits of the specimen corresponding to the secondary interaction for each genotype data of two different types of SNPs in all types of SNPs. An example is shown. In FIG. 3, SNP1 and SNP2 are used as SNP types, SNP1 genotype data is SNP1 = 0, SNP1 = 1 and SNP1 = 2, SNP2 genotype data is SNP2 = 0, SNP2 = 1 and SNP2 = 2. That is, the SNP interaction basic table indicates that the specimen is affected (= 1) and not affected (= 1) for each combination of the SNP1 genotype (0, 1, 2) data and the SNP2 genotype (0, 1, 2). = 0). Here, genotype (0, 1, 2) indicates a combination of dominant type and inferior type of genotype (AA, AB, BB).

For example, in FIG. 3, for the combination of genotype data SNP1 = 0 and SNP2 = 0, the number of non-affected classes is n000 and the number of affected classes is n100. Further, for the combination of genotype data SNP1 = 2 and SNP2 = 1, the number of specimens in the non-affected class is n012, and the number of specimens in the affected class is n112.
Here, as in the SNP basic table, n000 + n001 + n002 + n010 + n011 + n012 + n020 + n021 + n022 + n100 + n101 + n102 + n110 + n111 + n112 + n120 + n121 + n122 = n (total number of samples). Here, in the present embodiment, “yyy” in “nyyy” is a subscript.

Returning to FIG. 1, the basic table creation unit 1 sequentially selects combinations of two different types of SNPs from m SNPs stored in the SNP database 11.
Then, the basic table creation unit 1 performs, for each combination of genotype data in the SNP pair (for example, a combination of genotype data SNP1 = 0 and SNP2 = 0) in the selected SNP pair (hereinafter, SNP pair). Classify the specimen. Further, the basic table creation unit 1 classifies the samples of the combination of genotype data in each SNP pair into two classes of qualitative traits, and sets the classified result as the SNP interaction basic table having the configuration shown in FIG. As described above, the basic table creation unit 1 generates the SNP interaction basic table of FIG. 3 for two different types of SNPs from all types (m) of SNPs.
Therefore, since there are {m (m−1) / 2} combinations of two different types of SNPs of the m SNPs, the basic table creation unit 1 defines the SNP interaction basic table as {m (m−1). / 2} are generated.

  Next, FIG. 4 and FIG. 5 show a category for each category of a combination of any two of the three genotype data (data for each genotype) and another one of the genotype data in the SNP basic table. It is a figure which shows an example of the CDC table which shows the classification | category of the class of qualitative traits for the specimen in the inside. The CDC table shown in FIGS. 4 and 5 has a 2 × 2 contingency table in which two categories and the number of specimens in the categories are classified into two classes, and is stored in the SNP database 11.

  FIG. 4 shows that in SNP = 0, SNP = 1, and SNP = 2, which are genotype data of three types of attributes of SNP, genotype data SNP = 0 is a single category C1, and genotype data SNP = 1 and SNP. = 2 is an example of a table in which a combination of 2 is a category C2. In the case of FIG. 4, the CDC table focuses on genotype data SNP = 0, which is a single category C1. In FIG. 4, the number of specimens of the non-affected and affected classes in the category of genotype data SNP = 0 is the same as the SNP basic table of FIG. On the other hand, the number of specimens of the non-affected and affected classes in the category of the combination of genotype data SNP = 1 and SNP2 is the non-affected in each category of genotype data SNP = 1 and genotype data SNP2 in the SNP basic table of FIG. This is the result of adding the number of affected and affected samples.

  Further, FIG. 5 shows that in SNP = 0, SNP = 1, and SNP = 2, which are genotype data of three types of attributes of SNP, genotype data SNP = 1 is a single category C1, and genotype data SNP = 0. And a combination of SNP = 2 is an example of a table having category C2. In the case of FIG. 5, the CDC table focuses on genotype data SNP = 1, which is a single category C1. In FIG. 5, the number of specimens of the non-affected and affected classes in the category of genotype data SNP = 1 is the same as the SNP basic table of FIG. On the other hand, the number of specimens of the non-affected and affected classes in the category of the combination of genotype data SNP = 0 and SNP2 is the non-affected in each category of genotype data SNP = 0 and genotype data SNP2 in the SNP basic table of FIG. This is the result of adding the number of affected and affected samples.

  Returning to FIG. 1, the CDC table creation unit 2 reads the SNP basic table shown in FIG. 2 from the SNP database 11 in units of SNPs based on the SNP basic table identification information. Then, the CDC table creation unit 2 reclassifies the three types of three categories of genotype data into two types and one type in each read SNP basic table, and the two categories shown in FIG. 4 or FIG. 3 (= 3 (3-1) / 2) CDC tables are generated. Here, the combination of genotype data in the three CDC tables generated in one SNP includes genotype data SNP = 2 as a single category C1 and genotype SNP = 0, as shown in FIG. 4 and FIG. And SNP = 1 combination category C2.

  Here, the CDC table creation unit 2 calculates the number of non-affected and affected specimens in the category of the combination of two types of genotype data as the non-affected in each of the combined genotype data in the SNP basic table of FIG. The number of samples is added to obtain the number of unaffected samples in the combination category, and similarly, the number of affected samples in each of the combined genotype data is added to obtain the number of affected samples in the combination category. Further, the CDC table creation unit 2 assigns CDC table identification information to the created CDC table, associates this CDC table identification information with the CDC table indicated by this CDC table identification information, and sends it to the SNP database 11. Write and store.

  Next, FIG. 6 is a diagram illustrating an example of a CDC interaction table in which combinations of genotype data of two different types of SNPs in the SNP interaction basic table shown in FIG. 3 are divided into two categories. Similar to the CDC table, the CDC interaction table shown in FIG. 6 has a 2 × 2 configuration in which the number of specimens in the category is classified into two classes, and is stored in the SNP database 11. Yes. Since the SNP interaction basic table shown in FIG. 3 is a combination of three genotype data in each of two different types of SNPs, there are nine such combinations as can be seen from FIG.

As in the case of the CDC table from these nine combinations, one notable combination is set as a single category C1, and the remaining group of other combinations is set as a category C2. In the case of FIG. 6, the single combination is SNP = 1 and SNP2 = 1, and the conditions other than this combination in FIG. 3 are groups of other combinations.
Therefore, since there are nine combinations, there are nine unique combinations. As a result, the SNP interaction basic table of FIG. 3 is divided into nine CDC interaction tables.

  Returning to FIG. 1, the CDC table creation unit 2 reads the SNP interaction basic table shown in FIG. 3 from the SNP database 11 in combination units of two different types of SNPs. Then, the CDC table creation unit 2 reconfigures the nine types of genotype data in each of the read SNP interaction basic tables into two categories: one type of combination and the remaining eight types of combinations. Classify and generate eight CDC interaction tables consisting of the two categories shown in FIG.

  Here, the CDC table creation unit 2 calculates the number of non-affected and affected specimens in the category of combinations of eight types of genotype data, and sets each of the eight types of combinations grouped in the SNP interaction table of FIG. Is added to the number of unaffected specimens in the combination category, and similarly, the incidence in each of the eight grouped combinations is added to determine the number of affected specimens in the eight combination group categories. And Further, the CDC table creation unit 2 assigns CDC interaction table identification information to the created CDC interaction table, and this CDC interaction table identification information and the CDC interaction table indicated by this CDC interaction table identification information Are stored in the SNP database 11 in association with each other.

  Next, FIG. 7 is a diagram illustrating an example of an ADC interaction table in which a combination of genotype data of two different types of SNPs in the SNP basic table illustrated in FIG. 2 is divided into two categories. Similar to the CDC interaction table, the ADC interaction table shown in FIG. 7 has a 2 × 2 configuration in which the number of specimens in the category is classified into two classes, and is stored in the SNP database 11. Has been. The ADC interaction table shown in FIG. 7 uses a plurality of combinations of three genotype data in each of two different types of SNPs as dummy codes.

  For example, in the combination of genotype data divided into two categories (C1, C2) in FIG. 7, the category C1 is a combination of SNP1 = 0, SNP = 1, and SNP = 2, and SNP2 = 0 and SNP2 = 1. , A combination of SNP2 = 1 and SNP2 = 2, and a combination of SNP2 = 0 and SNP2 = 2, while category C2 is a geno other than category C1 already described in the SNP basic table. A group of type data. That is, the combination of any one of the genotypes of category C1 and any one of the genotypes of category C2 is set as one group, and the other eight combinations are set as the other group, and the number of samples of each class Is added.

  Returning to FIG. 1, the ADC table creation unit 3 selects different types of SNP combinations as SNP pairs from the SNP basic table stored in the SNP database 11, and sets all the genotype data in this SNP pair as two types. All combinations classified as categories are generated for each SNP pair. Then, the ADC table creation unit 3 reads out the number of unaffected and diseased samples of each of a plurality of genotype data in two different types of SNPs of category C1. Then, the ADC table creation unit 3 adds the number of specimens of the unaffected and affected classes in each genotype data in the combination of genotype data read out as the category C1 for each class.

Further, the ADC table creation unit 3 calculates the number of specimens of each class of non-affected and affected in the genotype data of the category C2 other than those included in the category C1 from the SNP basic table stored in the SNP database 11. read out. Then, the ADC table creation unit 3 adds the number of specimens of the non-affected and affected classes of the genotype data read out as the category C2 for each class, and generates the above-described ADC interaction table. Therefore, for each SNP pair, an ADC interaction table corresponding to all combinations classifying the genotype data of each of the two SNPs into two categories is generated.
Further, the ADC table creation unit 3 assigns CDC interaction table identification information to the created CDC interaction table, and the ADC interaction table identification information and the ADC interaction table indicated by the ADC interaction table identification information. Are stored in the SNP database 11 in association with each other.

The CDC likelihood calculating unit 4 sequentially reads out the CDC table indicated by the CDC table identification information from the SNP database 11, performs a single regression analysis using combinations of categories in the CDC table as explanatory variables, and calculates the likelihood. That is, two classes and two categories in the CDC table shown in FIG. 4 or FIG. 5 are extracted to form a 2 × 2 contingency table shown in FIG. FIG. 8 is a conceptual diagram illustrating the calculation of likelihood or p-value when performing a single regression analysis for each CDC table and using the explanatory variables indicated by the CDC table.
That is, the CDC likelihood calculation unit 4 sequentially reads out the CDC table from the SNP database 11 based on the CDC table identification information, and determines from the number of samples of the unaffected class and the affected class in each of the categories C1 and C2 in the read CDC table. The maximum log likelihood (hereinafter referred to as “likelihood”) of the logistic regression model is calculated using the following equation (3).

In this equation (3), as shown in FIG. 8, the number of specimens belonging to the non-affected class in category C1 is a, and the number of specimens belonging to the affected class is c. In addition, the number of samples belonging to the non-affected class in category C2 is b, and the number of samples belonging to the affected class is d.
Further, the CDC likelihood calculating unit 4 writes and stores the likelihood of the CDC table indicated by the CDC table identification information in the CDC likelihood storage unit 12 in association with each of the CDC table identification information.
Here, since there are 3m CDC tables in the genotype data of a single SNP, 3m likelihood is calculated.

Similarly, the CDC likelihood calculating unit 4 sequentially reads the CDC interaction table indicated by the CDC interaction table identification information from the SNP database 11, and the unaffected class in each of the categories C1 and C2 in the read CDC interaction table. And the likelihood of the logistic regression model is calculated from the number of specimens in the affected class using the above equation (3).
Then, the CDC likelihood calculating unit 4 associates the CDC likelihood table with the likelihood of the CDC interaction table indicated by the CDC interaction table identification information in association with the CDC interaction table identification information indicating each CDC interaction table. It is written and stored in the degree storage unit 12.
Here, there are m (m-1) / 2 combinations of two different SNPs, and these are classified into nine, so the number of CDC interaction tables is 9m (m-1) / 2. Therefore, the likelihood is calculated as 9m (m-1) / 2.

  The CDC_p value calculation unit 5 sequentially reads the CDC table indicated by the CDC table identification information from the SNP database 11, performs a single regression analysis using the combination of the categories of the CDC table as a dummy variable, that is, an explanatory variable, and calculates the p value. To do. That is, the CDC_p value calculation unit 5 calculates the p value (corresponding to the CDC table from the number of specimens of the non-affected class and the affected class in each of the categories C1 and C2 in the CDC table using the following equation (4) Hereinafter, it is referred to as CDC_p value).

In this equation (4), as shown in FIG. 8, the number of specimens belonging to the non-affected class in category C1 is a, and the number of specimens belonging to the affected class is c. In addition, the number of samples belonging to the non-affected class in category C2 is b, and the number of samples belonging to the affected class is d.
Further, the CDC_p value calculation unit 5 writes the p value of the CDC table indicated by the CDC table identification information in the CDC_p value storage unit 13 so as to correspond to each of the CDC table identification information indicating the CDC table.
Here, since there are 3m CDC tables in the genotype data of a single SNP, 3m is calculated as the p value.

Similarly, the CDC_p value calculation unit 5 sequentially reads out the CDC interaction table indicated by the CDC interaction table identification information from the SNP database 11, and sets the unaffected class in each of the categories C1 and C2 in the read CDC interaction table. The p-value of the logistic regression model is calculated from the number of samples in the affected class using the above equation (4).
Then, the CDC_p value calculation unit 5 associates each of the CDC interaction table identification information indicating the CDC interaction table with the CDC interaction table identification information and indicates the p value of the CDC interaction table indicated by the CDC interaction table identification information to the CDC_p value storage unit 13. To write and memorize.
Here, there are m (m-1) / 2 combinations of two different SNPs, and these are classified into nine, so the number of CDC interaction tables is 9m (m-1) / 2. Therefore, the p value is calculated as 9m (m-1) / 2.

Similarly to the CDC_p value calculation unit 5, the ADC_p value calculation unit 6 converts the two classes and the two categories in the ADC interaction table shown in FIG. 7 into a 2 × 2 table shown in FIG. To calculate the p-value for each ADC interaction table.
That is, the ADC_p value calculation unit 6 sequentially reads out the ADC interaction tables of the same two different types from the SNP database 11 according to the ADC interaction table identification information, and each category C1 and C2 in each of the read ADC interaction tables. BA (balanced accuracy) is calculated from the number of specimens of the non-affected class and the affected class using the following equation (5).

Further, the ADC_p value calculation unit 6 selects an ADC interaction table corresponding to a combination of genotype data of two different types of SNPs in which the calculated BA has a maximum value. For example, the ADC_p value calculation unit 6 calculates BA for each of the ADC interaction tables of all combinations of genotype data of SNP1 and SNP2 by the above equation (5), and the combination of genotype data having the maximum BA The ADC interaction table corresponding to is selected.
Here, the ADC_p value calculation unit 6 selects m combinations having the maximum BA value from among the combinations of the genotype data of each of these two SNPs in the above-described two different SNP sets. All of the two sets in the type of SNA are sequentially read out from the SNP database 11.
Further, since the ADC_p value calculation unit 6 selects one ADC interaction table for each set of two different SNPs, m (m−1) is obtained from all the ADC interaction tables stored in the SNP database 11. ) / 2 ADC interaction tables are selected.

Then, the ADC_p value calculation unit 6 calculates the slope of the logistic regression model using the above equation (4) from the number of samples of the unaffected class and the affected class in each of the categories C1 and C2 in the selected ADC interaction table. The p-value for (regression coefficient) is calculated.
The ADC_p value calculation unit 6 is associated with each of the ADC interaction table identification information indicating the ADC interaction table, and the ADC_p value storage unit is associated with the p value of the ADC interaction table indicated by the ADC interaction table identification information. 14 is written and stored.
Here, since there are m (m−1) / 2 combinations of genotype data in a plurality of SNPs, that is, the number of ADC interaction tables is m (m−1) / 2, the p value is m ( m-1) / 2 are calculated.

As described above, the likelihood and the p value obtained from each of the CDC table and the CDC interaction table are {3m + 9m (m−1) / 2}.
Further, the p value obtained from m (m−1) / 2 ADC interaction tables is m (m−1) / 2 as described above.
In the ranking described below, which one to select is determined by user settings. That is, whether the dummy variable of CDC or ADC is used as the explanatory variable, and if CDC is used as the dummy code, the user sets the evaluation value for ranking when using the likelihood or the p-value. Change arbitrarily according to the subject you want to evaluate.

  The ranking creation unit 7 corresponds to the CDC table identification information and the CDC interaction table identification information when the ranking setting condition that the user inputs from the input means (keyboard or the like) is the likelihood of the evaluation value using CDC. All the stored likelihoods are read out from the CDC likelihood storage unit 12 and sorted in descending order of likelihood to obtain a ranking. Here, the higher the likelihood, the higher the probability that it will be applied as an explanatory variable. Therefore, the ranking creating unit 7 sets the read likelihoods in order from the first to the {3m + 9m (m−1) / 2} th. Ranking up to.

  Here, the ranking creating unit 7 calculates the likelihood obtained from the obtained first to {3m + 9m (m−1) / 2} -th CDC tables and the CDC interaction table, that is, the CDC likelihood ranking, from the CDC likelihood. Each of the CDC table identification information and the CDC interaction table identification information corresponding to the degree ranking is written and stored in the ranking storage unit 15 in descending order of numerical values.

  Further, when the ranking setting condition that the user inputs from the input means is a p-value using CDC, the ranking creation unit 7 stores the evaluation value corresponding to the CDC table identification information and the CDC interaction table identification information. All the p values that have been set are read from the CDC_p value storage unit 13 and sorted in ascending order of the p values to obtain a ranking. Here, when the null hypothesis that the smaller the p-value is, the less the relationship with the character is assumed, the probability that the observed value is generated is small, that is, the hypothesis that there is no relationship is unnatural. 7 ranks the read p-value from the first to the {3m + 9m (m−1) / 2} -th in ascending order.

  Here, the ranking creating unit 7 calculates the ranking of the p value obtained from the obtained first to {3m + 9m (m−1) / 2} -th CDC table and the CDC interaction table, that is, the CDCp value (CDC_p value). Each of the CDC table identification information and the CDC interaction table identification information corresponding to the ranking of the CDCp value is written and stored in the ranking storage unit 15 in order from the top.

  In addition, when the ranking setting condition that the user inputs from the input unit uses ADC, the ranking creation unit 7 uses the ADC_p value storage unit 14 to store the p value stored corresponding to the ADC interaction table identification information. Read all, sort in ascending order of p-value, and perform a process of obtaining the ranking. Here, the smaller the p-value, the smaller the probability that an observed value will occur, that is, the hypothesis that there is no relevance, is assumed, assuming the null hypothesis that there is no association with a trait, as in the calculation of the CDCp value. Since it is natural, the ranking creating unit 7 ranks the read p values from the first to the {3m + 9m (m−1) / 2} -th in ascending order.

  Here, the ranking creation unit 7 calculates the ADCp value from the obtained first to {3m + 9m (m−1) / 2} -th ADC interaction tables, that is, the ranking of ADCp values (ADC_p values). Each of the ADC interaction table identification information corresponding to the ranking is written and stored in the ranking storage unit 15 in order from the top.

The dummy variable creation unit 8 selects the ranking corresponding to the ranking setting condition that the user inputs from the input means from the ranking created by the ranking creation unit 7, selects the top r, for example, 256 from this ranking, The dummy code corresponding to the selected table is used as an explanatory variable, and a multiple logistic regression model shown in the following equation (6) is configured. In this embodiment, selecting the top r identifies the strength of the influence on the qualitative trait (quantitative trait in the second embodiment) due to the SNP genotype interaction in the entire human gene. (Ie, identifying genotype combinations in the SNP as dummy variables used in the regression model).
For example, when the user uses ADC as a ranking setting condition, the dummy variable creation unit 8 searches the ranking storage unit 15 for the ADC interaction table identification information, and stores the ADC interaction table identification information in correspondence with the ADC interaction table identification information. The ranking of the ADCp value being read is read out.

Then, the dummy variable creation unit 8 selects the top r ADC interaction table identification information from the ranking of the read ADCp values, and the dummy code that becomes the category of the ADC interaction table indicated by the ADC interaction table identification information Is an explanatory variable.
Here, the numerical value r to be set is 256 as an example because of the calculation limit, but it is desirable that the numerical value r be determined by the following equation (7).

  For example, if 256 dummy variables selected by the dummy variable creation unit 8 include a combination of SNP genotype data in the CDC table shown in FIG. 9, 6 unaffected individuals out of a total of 15 subjects. There are 9 people affected. This FIG. 9 shows the classification of the qualitative trait class of the specimen in the category for each category of the combination of any two of the three genotype data and the other one of the genotype data in the SNP basic table. It is a figure which shows an example of the CDC table shown.

  In the case of the CDC table of FIG. 8, since there are 6 non-affected persons and 9 affected persons, the response variable vector is y = (0,0,0,0,0,0,1,1). , 1, 1, 1, 1, 1, 1, 1), and the explanatory variable vector uses a dummy code such as a CDC table, a CDC interaction table, or an ADC interaction table as an explanatory variable (that is, a dummy variable). A dummy variable is determined by giving 0 or 1 according to the column of each table which is a 2 × 2 contingency table. As an example, the vector of the dummy variable is x = (0, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 1, 0, 1).

Returning to FIG. 1, the variable selection unit 9 sets the dummy variable selected by the dummy variable creation unit 8 as an explanatory variable in a logistic regression model described later.
That is, the variable selection unit 9 applies the explanatory variable vector in the logistic regression model to the selected dummy variables in the 256 2 × 2 contingency tables, and is a matrix of 256-dimensional explanatory variables made up of 256 dummy variables. An explanatory variable matrix is obtained. Then, the variable selection unit 9 substitutes this explanatory variable matrix into a general logistic regression equation to generate a logistic regression model equation represented by equation (6).

Next, the variable selection unit 9 obtains an objective variable for the regression analysis, for example, by using an algorithm based on the SEE (Smooth threshold Estimating Equation) method in selecting a regression coefficient. Therefore, a determination process is performed as to whether or not there is an effect as an explanatory variable.
Here, the variable selection unit 9 performs n (total number of samples) in which 256 logistic regression models are arranged vertically as the explanatory variable vector X in the processing of SEE which is a penalized regression method (penalized maximum likelihood method). A design matrix of rows × 256 columns is generated.
In addition, the variable selection unit 9 sets ^ β _j ⁽⁰⁾ in the equation (6) as an initial value of a regression coefficient by which the j-th dummy variable is multiplied, and a tuning parameter λ (a smoothing parameter that is a degree of local variation of the estimated curve) Is controlled to generate a set A (λ) represented by the following equation (8).

The variable selection unit 9 uses the value of the regression coefficient β _j that maximizes (maximizes) the expression (9) for the regression coefficient β _j satisfying j∈A (λ) in the following expression (9). Ask. The variable selection unit 9 sets the regression coefficient β _j not belonging to A (λ) to 0 as a sparse solution (regression method with penalties).

  In the above equation (9), l (β) is a log likelihood function of the logistic regression model shown in the following equation (10).

Here, the variable selection unit 9 uses the tuning parameter λ ₂ in the L2 penalized logistic regression model as a tuning parameter used when obtaining the initial estimated value ^ β _j ⁽⁰⁾ used in the equation (8). Here, when the variable selection unit 9 obtains the initial estimated value ^ β _j ⁽⁰⁾ , δ _j (λ) in the equation (9) is set to 0 using the L2 penalty logistic regression model (9) Extract the initial estimate ^ β _j ⁽⁰⁾ that maximizes the equation. Newton-Raphson solves the maximization problem as a tuning parameter λ _{2 in} equation (9) using a logistic regression model with an L2 penalty to which a normalization term of the sum of squares of the initial estimate ^ β _j ⁽⁰⁾ is added In the update step of the (Newton-Raphson) method, the minimum value at which the inverse matrix is not singular is used. Therefore, in this embodiment, the tuning parameter λ _{2 is set} to 1 / n (total number of samples).

In addition, the variable selection unit 9 substitutes the tuning parameter λ into the following equation (11) by setting β _j where the above-described equation (9) is maximized as ^ β _j , and sets EBIC (Extended Bayes Information Criterion) to calculate.

Here, the variable selection unit 9 obtains the above equation (11), that is, the tuning parameter λ that minimizes the EBIC, and obtains β _j that maximizes the SEE equation (9) using the tuning parameter λ. In Equation (11), γ is a constant defined in the interval [0, 1], and the numerical value p uses {9m (m−1) / 2 + 3m} when it is CDC, and {ADC m (m-1) / 2} is used.
Similarly, as a reference for selecting the tuning parameter λ, an AIC represented by the following expression (12) and a BIC represented by the expression (13) may be used. Also in the following formula | equation, regarding the numerical value of (beta) and p, it is the same as that of (11) Formula.

The variable selection unit 9 obtains the regression coefficient β _j of the top r dummy variables x in the ranking supplied from the dummy variable creation unit 8 by the above-described processing, and sends the obtained regression coefficient β _j to the regression analysis processing unit 10. Output.

Regression analysis processing unit 10, the regression coefficient beta _j corresponding to each of r pieces of dummy variables supplied from the variable-selecting part 9, and extracts the regression coefficient beta _j is 0.
Further, the regression analysis processing unit 10 is configured by removing a dummy variable corresponding to the regression coefficient βj that is 0 in the equation (6) including r dummy variables and including a dummy variable in which the regression variable βj is not 0. Generate a logistic regression model.
Thereby, the regression analysis processing unit 10 substitutes the data of the dummy variable generated from the SNP genotype of the new specimen using the logistic regression model obtained as described above, and the disease of the patient of the new specimen is affected. Find the logit of the probability of being. Here, if the logit value is positive, the probability of morbidity is 50% or more, and the higher the value, the higher the probability. On the other hand, if the value of logit is negative, the probability of morbidity is 50% or more, and the probability decreases as the absolute value increases.

Further, there are merits and demerits in the case of using CDC or ADC as the above-described dummy variable coding method and the coding method conventionally known in the field of genetics.
As a conventional coding method, the merit is that when a model with an increased risk of 0, 1, 2 depending on the number of alleles in SNP is appropriate, the distribution according to the statistics can be accurately approximated. Can control the first type of error.
However, as a demerit, since various patterns such as inferiority, dominance, and additivity are mixed as genotypes in reality, it is not easy to overlook significant interactions by analyzing with a single coding method. It becomes a cause. This demerit is a problem of this embodiment.

Therefore, in the present embodiment, it is necessary to select either the CDC or the ADC according to the object to be analyzed by the user, or both to select a regression model that can obtain a more accurate result. is there.
As a merit of CDC, by cutting out 2 × 2 contingency table centered on each cell of contingency table, the influence of the observation cell of decimal example is eliminated, and it corresponds to the pattern of genotypes or multiple genotype interaction it can.
On the other hand, as a disadvantage of CDC, there is a possibility that the interaction is evaluated redundantly in that one interaction is treated as an independent factor of a plurality of 2 × 2 contingency tables.

Moreover, the merit of ADC eliminates the influence of the decimal example observation cell by the same principle as CDC. That is, by over-searching a set of 2 × 2 contingency tables created by regrouping each cell (group as a category classified by genotype) of the interaction between different types of SNPs, the interaction is overlooked. The possibility of is greatly reduced.
On the other hand, as a disadvantage of the ADC, if the increase in variance due to regrouping is significant in the optimal 2 × 2 contingency table finally obtained, the odds ratio calculated from it can be reliable. Absent.

Next, in the ranking performed by the ranking generation unit 7, dummy variables generated using CDC are ranked using likelihood, dummy variables generated using CDC are generated using p-value, and are generated using ADC. The merit and demerit of each ranking using the dummy variable p is shown below.
(A) Combination of CDC and Likelihood Merits are interpreted as a framework for selecting dummy variables, and can be said to be a consistent and natural method that leads to subsequent dummy variable selection parts. Statistics used in the original SIS.
Disadvantages include less experience in epidemiological studies than the p-value of the 2 × 2 contingency table odds ratio. In addition, there is a possibility that the interaction is evaluated redundantly in that the genotype interaction of one SNP is decomposed into a plurality of factors called 2 × 2 contingency tables and these are treated as independent factors.

(B) Combination of CDC and p-value The merit is a safe statistic because the p-value of the odds ratio of the 2 × 2 contingency table is widely used in epidemiological studies and many good properties are known.
The disadvantage is that there is a possibility that the interaction is evaluated redundantly in that the genotype interaction of one SNP is decomposed into a plurality of factors called 2 × 2 contingency tables and treated as independent factors.
(C) Combination of ADC and p-value The advantage is that each cell of SNP interaction is adaptively regrouped, and a single statistic representative of the interaction is constructed, so that redundancy existing in the CDC in processing Sex can be removed.
Disadvantages do not take into account the increase in dispersion due to regrouping of each cell of SNP interaction, which can lead to a fake ranking result.

In the selection of regression coefficients performed by the variable selection unit 9, SCAD (smoothly clipped absolute device) and SEE used in the present embodiment have the following advantages and disadvantages.
The advantage of SCAD is that the initial value of the estimated amount of the regression coefficient is not required.
On the other hand, the disadvantage of SCAD is that there is no highly accurate and simple algorithm for finding a regression coefficient.
SEE is used in the present embodiment in order to solve the above-mentioned disadvantages of SCAD, and an initial value estimation amount of a regression coefficient is necessary. However, there is a simple algorithm with high accuracy for appropriate solution.

Next, in the selection of the regression coefficient performed by the variable selection unit 9, a method for obtaining the tuning parameter by AIC for setting the regression coefficient for SEE, a method for obtaining the tuning parameter by BIC for setting the regression coefficient for SEE, and the regression coefficient for SEE The merits and demerits of each of the methods for obtaining the tuning parameters by EBIC with respect to the setting are shown below.
(A) Method obtained by AIC Since it is known that the theoretically preferable property of consistency of variable selection is not guaranteed, use of AIC is not recommended for the purpose of variable selection.
(B) Technique obtained by BIC It is known that the merit of variable selection is guaranteed when the dimension of the variable is relatively small. Also, tuning parameters such as EBIC are not included. In addition, there are many empirical achievements regarding use.
The disadvantage is that the consistency of variable selection may be broken when the dimension of the variable is large.

(C) Method obtained by EBIC A merit is that consistency of variable selection is established even when the dimension of the variable is large.
The disadvantage is that false positives and false negatives need to be controlled by tuning parameters, but there is no general reliable method for appropriately adjusting them.
The coincidence of the variable selection described above means that when the true regression coefficient is divided into non-zero and zero, this division can be correctly determined, that is, as the number of samples increases more accurately, This means that the probability of correct discrimination converges to 1.

  As described above, in the present embodiment, genotypes are classified into two categories for each SNP, and classified into two genotype categories for each combination of two different SNPs, and each is used as a dummy variable. A dummy variable having a high degree of contribution to the determination that the patient is affected is selected from the dummy variables, and a regression coefficient of each dummy variable is obtained. For this reason, when analyzing the interaction affecting the morbidity between different SNPs, a huge amount of data was required to determine the dummy variable and the regression coefficient for this dummy variable for the generation of the regression model. On the other hand, in this embodiment, since the dummy variable is selected and used according to the contribution, the number of samples for generating the regression model can be reduced as compared with the conventional case.

Next, the flow of processing of dummy variable selection and regression coefficient calculation in the regression model using the gene interaction analysis system in the first embodiment will be described with reference to FIG. FIG. 10 is a flowchart showing an operation example of selecting a dummy variable and calculating a regression coefficient in a regression model using the gene interaction analysis system.
Step S1:
The basic table creation unit 1 reads genotype data for each SNP of the specimen from the SNP database 11 and generates a SNP basic table for each SNP.
And the basic table preparation part 1 attaches | subjects the SNP basic table identification information which identifies this SNP basic table to every created SNP basic table, and makes it correspond to a SNP basic table, and SNP basic table identification with this SNP basic table Information is made to correspond and written into the SNP database 11 for storage.

Step S2:
The basic table creation unit 1 reads out genotype data for each SNP of the specimen from the SNP database 11, and determines an SNP interaction basic table for determining dummy variables for each genotype interaction between two different SNPs. Generate.
And the basic table preparation part 1 attaches | subjects the SNP interaction basic table identification information which identifies this SNP interaction basic table to every created SNP interaction basic table, and makes this SNP interaction basic table correspond, and this The SNP interaction basic table identification information is associated with the SNP interaction basic table, and is written and stored in the SNP database 11.

Step S3:
A control unit (not shown) in the gene interaction analysis system performs processing in the case of CDC when a signal for selecting CDC or ADC as a dummy variable coding method is supplied from an external device to the gene interaction analysis system. In the case of ADC, the process proceeds to step S10.

Step S4:
The CDC table creation unit 2 sequentially reads out the SNP basic table to which the SNP basic table identification information is attached from the SNP database 11 for each SNP, performs genotype data grouping, and a CDC table that is a 2 × 2 partitioned table. Is generated.
Then, the CDC table creation unit 2 attaches CDC table identification information for identifying the CDC table to each created CDC table, associates the CDC table with the CDC table, associates the CDC table identification information with the CDC table, Write to the database 11 and store it.

Step S5:
The CDC table creation unit 2 sequentially reads out the SNP interaction basic table to which the SNP interaction basic table identification information is attached from the SNP database 11 for each of two different sets of SNPs, and determines combinations of nine types of genotype data. Dividing into two categories, a CDC interaction table which is a 2 × 2 contingency table is generated.
Then, the CDC table creation unit 2 attaches CDC table identification information for identifying the CDC table to each created CDC table, associates the CDC table with the CDC table, associates the CDC table identification information with the CDC table, Write to the database 11 and store it.

Step S6:
A control unit (not shown) in the gene interaction analysis system receives a signal for selecting likelihood or p value (p value of odds ratio) as an evaluation value for ranking dummy variables from an external device to the gene interaction analysis system. If supplied, the process proceeds to step S7 if likelihood, while the process proceeds to step S8 if p-value.

Step S7:
The CDC likelihood calculating unit 4 sequentially reads out the CDC table to which the CDC table identification information is attached, and calculates the likelihood indicating the strength of the relationship between the SNP and the disease by the equation (3).
Then, the CDC likelihood calculating unit 4 associates the obtained likelihood with the CDC table identification information, and writes and stores it in the CDC likelihood storage unit 12.

Step S8:
The CDC_p value calculation unit 5 sequentially reads the CDC table to which the CDC table identification information is attached, and calculates the p value indicating the strength of the relationship between the SNP and the disease by the equation (4).
Then, the CDC_p value calculation unit 5 associates the obtained p value with the CDC table identification information, writes it in the CDC_p value storage unit 13, and stores it.

Step S9:
The ADC table creation unit 3 sequentially reads out the SNP basic table to which the SNP basic table identification information is attached from the SNP database 11 for every two different sets of SNPs, and all of the genotype data of the two different sets of SNPs. The number of groupings corresponding to the combination is performed, and an ADC interaction table that is a 2 × 2 contingency table for each grouping is generated.
Then, the CDC table creation unit 2 attaches ADC interaction table identification information for identifying the ADC interaction table to each created ADC interaction table, and associates the ADC interaction table with the ADC interaction table. At the same time, the ADC interaction table identification information is made to correspond and written into the SNP database 11 and stored.

Step S10:
The ADC_p value calculation unit 6 sequentially reads out the ADC interaction table to which the ADC interaction table identification information is attached, and calculates a p value indicating the strength of the relationship between the SNP and the disease according to the equation (4).
Then, the ADC_p value calculation unit 6 associates the obtained p value with the ADC interaction table identification information, and writes and stores them in the ADC_p value storage unit 14.

Step S11:
The ranking creation unit 7 corresponds to a coding method (CDC or ADC) when setting a dummy variable set by the user and an evaluation value (likelihood or p value) for ranking when the coding method is CDC. Then, ranking processing is performed.
That is, when the coding method is CDC and the evaluation value is likelihood, the ranking creating unit 7 sequentially reads out the likelihood for each CDC table and CDC interaction table from the CDC likelihood storage unit 12, and the likelihood Perform ranking processing to arrange in descending order.
Then, the ranking creating unit 7 arranges the CDC table identification information and the CDC interaction table identification information in the order of the ranking arrangement, and writes and stores them in the ranking storage unit 15.

On the other hand, when the coding method is CDC and the evaluation value is a p value, the ranking creating unit 7 sequentially reads the p value for each CDC table and each CDC interaction table from the CDC_p value storage unit 13, and the p value is small. Perform ranking processing to arrange in order.
Then, the ranking creation unit 7 arranges the CDC table identification information and the CDC interaction table identification information in the order of the ranking arrangement, and writes and stores them in the ranking storage unit 15 as in the case of likelihood.
Further, when the coding method is ADC, the ranking creating unit 7 sequentially reads p values for each ADC interaction table from the ADC_p value storage unit 14 and performs ranking processing for arranging the p values in ascending order.
Then, as in the case of CDC, the ranking creating unit 7 arranges the ADC interaction table identification information in the order of ranking arrangement, and writes and stores them in the ranking storage unit 15.

Step S12:
The dummy variable creating unit 8 performs dummy variable selection processing corresponding to the case where the coding method for setting the dummy variable set by the user is CDC or ADC.
That is, when the coding method is CDC, the dummy variable creation unit 8 sequentially reads r in the order of arrangement in the CDC table identification information and CDC interaction table identification information stored in the ranking storage unit 15. Then, the classified categories of the read r CDC table genotype data are set as dummy variables.
Similarly, when the coding method is ADC, the dummy variable creation unit 8 sequentially reads r pieces in the order of arrangement in the arrangement of the ADC interaction table identification information stored in the ranking storage unit 15 and reads the r The classified categories of the genotype data of the individual ADC interaction tables are set as dummy variables.

Step S13:
Regression analysis processing unit 10 generates the dummy the set dummy variables of variable creation unit 8 as regression model as explanatory variables x _j (6) formula.
Then, the variable selection unit 9 calculates the regression coefficient β _j of each explanatory variable x _{j in} the expression (6) generated by the regression analysis processing unit 10. At this time, the user sets which of AIC, BIC, and EBIC to use as an evaluation value for setting the tuning parameter λ when calculating the regression coefficient β to the gene interaction analysis system.

Step S14:
The regression analysis processing unit 10 substitutes the regression coefficient β _j calculated by the variable selection unit 9 for each of the explanatory variables x _{j in} the equation (6), and generates a regression model for obtaining the corresponding objective variable.

<Second Embodiment>
The first embodiment described above is a process for the logistic regression method. The second embodiment is directed to dummy variable selection and regression coefficient determination processing when creating a linear regression model.
Next, embodiments of the present invention will be described with reference to the drawings. FIG. 11 is a schematic block diagram showing a configuration example of a gene interaction analysis system according to the second embodiment of the present invention. In this embodiment, the gene interaction analysis system includes a basic table creation unit 1, a CDC table creation unit 2, an ADC table creation unit 3, a CDC_t value calculation unit 16, an ADC_t value calculation unit 17, a ranking creation unit 7, a dummy A variable creation unit 8, a variable selection unit 9, a regression analysis processing unit 10, an SNP database 11, a CDC_t value storage unit 18, an ADC_t value storage unit 19, and a ranking storage unit 15 are provided.

Hereinafter, the configuration and operation of the second embodiment different from those of the first embodiment will be described. First, the difference from the first embodiment is that there is no combination of single SNP genotypes in the CDC coding method. In the second embodiment, the CDC interaction table and the ADC interaction table in the first embodiment are generated.
The SNP database 11 stores quantitative trait data (for example, body weight, blood glucose level, etc.) and genotype data of each SNP corresponding to each specimen.

  FIG. 12 shows a configuration example of the SNP interaction basic table in the second embodiment. FIG. 12 shows an example of a SNP interaction basic table showing classification of qualitative trait classes of specimens corresponding to secondary interactions in genotype data of two different types of SNPs in all SNPs. Yes. In FIG. 12, SNP1 and SNP2 are used as SNP types, SNP1 genotype data is SNP1 = 0, SNP2 and SNP = 2, and SNP2 genotype data is SNP2 = 0, SNP2 = 1 and SNP2 = 2. And That is, the SNP interaction basic table indicates the integrated value of the quantitative trait of the specimen for each combination of the SNP1 genotype data and the SNP genotype data.

For example, in FIG. 12, the number of specimens is n01 for the combination of genotype data SNP1 = 0 and SNP2 = 0, and the total value of the numerical values of quantitative traits in these n01 specimens is s01. Further, for the combination of genotype data SNP1 = 2 and SNP2 = 1, the number of specimens is n11, and the total value of the numerical values of quantitative traits in these n11 specimens is s11.
Here, n00 + n01 + n02 + n10 + n11 + n12 + n20 + n21 + n22 = n (total number of samples).

Returning to FIG. 1, the basic table creation unit 1 sequentially selects combinations of two different types of SNPs from m SNPs stored in the SNP database 11.
Then, the basic table creation unit 1 performs, for each combination of genotype data in the SNP pair (for example, a combination of genotype data SNP1 = 0 and SNP2 = 0) in the selected SNP pair (hereinafter, SNP pair). Classify the specimen. Further, the basic table creation unit 1 classifies the samples of the combination of genotype data in each SNP pair into two classes of qualitative traits, and uses the classified result as the SNP interaction basic table shown in FIG. As described above, the basic table creation unit 1 generates the SNP interaction basic table of FIG. 12 for two different types of SNPs from all types (m) of SNPs.

  Returning to FIG. 11, the CDC table creation unit 2 sequentially reads the SNP interaction basic table shown in FIG. 12 from the SNP database 11. Then, the CDC table creation unit 2 reconfigures the nine types of genotype data in each of the read SNP interaction basic tables into two categories: one type of combination and the remaining eight types of combinations. Classification is performed, and eight CDC interaction tables (category classification similar to the CDC interaction table in the first embodiment) including two categories are generated. Here, the CDC table creation unit 2 adds the total value of the numerical values of the quantitative traits in each of the grouped combinations to the combination of the eight types of genotype data, and the total value of the quantitative traits in the category of the combination And Further, the CDC table creation unit 2 assigns CDC interaction table identification information to the created CDC interaction table, and this CDC interaction table identification information and the CDC interaction table indicated by this CDC interaction table identification information Are stored in the SNP database 11 in association with each other.

The ADC table creation unit 3 sequentially selects combinations of two different types of SNPs from the m SNPs stored in the SNP database 11 as SNP pairs.
Then, the ADC table creation unit 3 sums the values of the quantitative traits of each specimen of the genotype data of two different types of SNPs of category C1 in the SNP pair.
Further, the ADC table creation unit 3 sums the values of the quantitative traits of each specimen of the genotype data of the category C2 other than those included in the category C1 from the SNP basic table stored in the SNP database 11. To do.
Next, the ADC table creation unit 3 generates an ADC interaction table (first implementation) consisting of the total value of the quantitative traits of the category C1 and this category C1, and the total value of the quantitative traits of the category C2 and this category C1. Category classification similar to the ADC interaction table in the form) is generated.
Then, the ADC table creation unit 3 assigns ADC interaction table identification information to the generated ADC interaction table, and the ADC interaction table identification information and the ADC interaction table indicated by the ADC interaction table identification information. Are stored in the SNP database 11 in association with each other.

  The CDC_t value calculation unit 16 sequentially reads the CDC interaction table indicated by the CDC interaction table identification information from the SNP database 11, performs a single regression analysis using the combination of the categories of the CDC interaction table as an explanatory variable, and the t value Is calculated. That is, the CDC_t value calculation unit 16 creates a CDC interaction table from the total value of quantitative traits in each of the categories C1 and C2 in the CDC interaction table using the t value calculation method of the following equation (14). A corresponding t value (hereinafter referred to as a CDCt value) is calculated.

  In the case of quantitative trait regression analysis, a linear regression analysis model is used, but the optimal classification of two categories is detected in the same manner as the logistic regression model already described. In the classification of these two categories, the difference between the average values of the quantitative traits in each of the two categories, that is, the t value as a statistical fee used in the test of the difference between the average values of the two groups is used. Yes. In the above equation (14), the average value y1 of the total value of quantitative traits in each of the categories C1, the average value y2 of the total value of quantitative traits in the category C2, the number of samples n1 included in the category C1, The number of samples n2 included in the category C2 is used.

  That is, the CDC_t value calculation unit 16 substitutes the number of specimens and the total value of the quantitative traits in each of the categories C1 and C2 for each CDC interaction table in the following equation (14) to obtain the t value Is calculated. As a result, the equation (14) becomes the following equation (15). In the equation (15), the category C1 of the combination of genotype data of SNP1 = 0 and SNP2 = 1 and the other genotype data 8 The t value with the category C2 composed of the combinations is calculated.

  Further, the CDC_t value calculation unit 16 associates each of the CDC interaction table identification information indicating the CDC interaction table with the CDC interaction table identification information, and stores the t value of the CDC interaction table indicated by the CDC interaction table identification information in the CDC_t value storage unit 18. Write and memorize it.

  The ADC_t value calculation unit 17 sequentially reads out the ADC interaction table indicated by the ADC interaction table identification information from the SNP database 11, performs a single regression analysis using the combination of the categories of the ADC interaction table as an explanatory variable, and outputs the t value. Is calculated. That is, the ADC_t value calculation unit 17 corresponds to the ADC interaction table using the t value calculation method of the above formula (14) from the total value of the quantitative traits in each of the categories C1 and C2 in the ADC interaction table. The t value (hereinafter referred to as ADCt value) is calculated.

  That is, the ADC_t value calculation unit 17 calculates the t value by substituting the number of specimens and the total value of the quantitative traits in each of the categories C1 and C2 for each CDC interaction table with respect to the above equation (14). To do. As a result, the equation (14) becomes the following equation (16). In the equation (16), the category C1 of the combination of genotype data of SNP1 = 0 and SNP2 = 1 and SNP1 = 2 and SNP2 = 1 and T value with category C2 which consists of a combination of genotype data other than is calculated.

Here, the ADC_t value calculation unit 17 selects an ADC interaction table having the maximum value for each SNP pair calculated by the equation (16), and sets the selected ADC interaction table as a dummy variable of the SNP pair. Classification.
Further, the ADC_t value calculating unit 17 associates each of the ADC interaction table identification information indicating the ADC interaction table selected in each SNP pair with the t value of the ADC interaction table indicated by the ADC interaction table identification information. , Write to the ADC_t value storage unit 19 for storage.

The ranking creation unit 7 sequentially reads t values of the CDC interaction table from the CDC_t value storage unit 18 based on the CDC interaction table identification information.
Then, the ranking creating unit 7 sorts the read t values in descending order, arranges the CDC interaction table identification numbers in this order, and writes and stores them in the ranking storage unit 15.

In addition, the ranking creating unit 7 sequentially reads t values of the ADC interaction table from the ADC_t value storage unit 19 based on the ADC interaction table identification information.
Then, the ranking creating unit 7 sorts the read t values in descending order, arranges the ADC interaction table identification numbers in this order, and writes and stores them in the ranking storage unit 15.

The dummy variable creation unit 8 selects the ranking corresponding to the ranking setting condition that the user inputs from the input means from the ranking created by the ranking creation unit 7, selects the top r, for example, 256 from this ranking, A dummy regression code corresponding to the selected table is used as an explanatory variable to construct a linear regression model.
For example, when the user uses CDC as the ranking setting condition, the dummy variable creation unit 8 searches the ranking storage unit 15 for the CDC interaction table identification information, and stores it in correspondence with the CDC interaction table identification information. The ranking of the CDCp value being read is read out. Then, the dummy variable creation unit 8 selects the top r CDC interaction table identification information from the read ranking of the CDCp values, and the dummy variable that becomes the category of the CDC interaction table indicated by the CDC interaction table identification information Is an explanatory variable.

On the other hand, if the user uses ADC as the ranking setting condition, the dummy variable creation unit 8 searches the ranking storage unit 15 for the ADC interaction table identification information, and stores the ADC interaction table identification information in correspondence with the ADC interaction table identification information. The ranking of the ADCp value being read is read out. Then, the dummy variable creation unit 8 selects the top r ADC interaction table identification information from the ranking of the read ADCp values, and the dummy variable that becomes the category of the ADC interaction table indicated by the ADC interaction table identification information Is an explanatory variable.
In the present embodiment, the processes of the variable selection unit 9 and the regression analysis processing unit 10 are the same as those in the first embodiment, and thus the description thereof is omitted.

  Further, a program for realizing the function of the gene interaction analysis system in FIG. 1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into the computer system and executed. An analysis process of gene interaction may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design and the like within a scope not departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 1 ... Basic table creation part 2 ... CDC table creation part 3 ... ADC table creation part 4 ... CDC likelihood calculation part 5 ... CDC_p value calculation part 6 ... ADC_p value calculation part 7 ... Ranking creation part 8 ... Dummy variable creation part 9 ... Variable selection unit 10 ... regression analysis processing unit 11 ... SNP database 12 ... CDC likelihood storage unit 13 ... CDC_p value storage unit 14 ... ADC_p value storage unit 15 ... ranking storage unit 16 ... CDC_t value calculation unit 17 ... ADC_t value calculation unit 18 ... CDC_t value storage unit 19 ... ADC_t value storage unit

Claims (8)

Genome-wide SNP (single nucleotide polymorphism) genotype data is a combination of SNPs that have gene-gene interactions that affect phenotypic expression, comprehensively identifying SNP pairs, and character expression from the SNP pair A gene interaction analysis system for generating a regression model to analyze,
From m types of SNPs detected from n specimens, SNP pairs consisting of two different types of SNPs are selected in sequence, and each genotype (combination of dominant type and inferior type) of each SNP in the SNP pair is selected. A table creation unit that classifies a dummy code table including two categories, classifies each of the n samples for each dummy code, and creates a table indicating the two categories and the samples belonging to each category. When,
The single regression model with the dummy code in the table as a dummy variable is used to perform a single regression analysis using the specimen classified into each category, and an evaluation showing the strength of the association between the dummy variable and the expression of the trait An evaluation value calculation unit for calculating a value for each dummy variable;
A ranking creating unit that generates a ranking by arranging the dummy variables in order of the evaluation value indicating the strength of the relevance;
Extracting the dummy variables up to a preset order, and making the extracted dummy variables dummy variables of the regression model,
A variable selection unit that calculates a regression coefficient to be multiplied by the extracted dummy variable by a maximum likelihood method with penalties. The table creation unit
Using CDC (cell-wise dummy coding) as a coding method, a CDC table of dummy codes obtained by classifying genotypes in one type of SNP into two categories, and genotypes in two different types of SNPs A CDC table creation unit as a CDC interaction table of dummy codes classified into two categories of combinations of genotypes and other genotypes;
Using ADC (adaptive dummy coding) as a coding method, an ADC table creation unit that makes a plurality of combinations of genotypes in two different types of SNPs into an ADC interaction table of dummy codes classified into two categories, and The gene interaction analysis system according to claim 1, wherein If the regression model is a logistic regression model,
Each of the CDC table, the CDC interaction table, and the ADC interaction table is a 2 × 2 table including two categories and the number of the specimens belonging to two classes of qualitative traits in each category. Yes,
The evaluation value calculation unit
A CDC likelihood calculating unit that calculates likelihood as the evaluation value in the single regression analysis of each dummy code of the CDC table and the CDC interaction table;
A CDC_p value calculation unit that calculates a p value as the evaluation value in the single regression analysis of each dummy code of the CDC table and the CDC interaction table;
The gene interaction analysis system according to claim 2, wherein the system is an ADC_p value calculation unit that calculates a p value as the evaluation value in the single regression analysis of each dummy code of the ADC interaction table. The ranking generation unit generates the ranking by arranging the likelihoods of the dummy codes of the CDC table and the CDC interaction table in the order of strong association, and the CDC table and the CDC interaction table. The ranking is generated by arranging the p values of each dummy code in the order of strong association, and the p values of the dummy codes of the ADC interaction table are arranged in the order of strong association. The gene interaction analysis system according to claim 3, wherein the ranking is generated. When the regression model is a linear regression model,
The CDC interaction table and the ADC interaction table are two categories and a table that is an addition value of the quantitative traits of the specimen included in each of the categories;
The evaluation value calculation unit
A CDC_t value calculation unit that calculates a t value as the evaluation value in the single regression analysis of each dummy code of the CDC interaction table;
The gene interaction analysis system according to claim 2, wherein the system is an ADC_t value calculation unit that calculates a t value as the evaluation value in the single regression analysis of each dummy code of the ADC interaction table. The ranking generation unit generates the ranking by arranging the t values of the dummy codes of the CDC interaction table in the order of the relevance, and the t values of the dummy codes of the ADC interaction table. The gene interaction analysis system according to claim 5, wherein the ranking is generated by arranging the sequences in the order of strong association. Genome-wide SNP (single nucleotide polymorphism) genotype data is a combination of SNPs that have gene-gene interactions that affect phenotypic expression, comprehensively identifying SNP pairs, and character expression from the SNP pair A gene interaction analysis method for generating a regression model to be analyzed,
The table creation unit selects SNP pairs consisting of two different types of SNPs sequentially from m types of SNPs detected from n samples, and each genotype (dominant type and inferiority) of the SNPs in the SNP pair. A table indicating the two categories and the samples belonging to the respective categories by classifying each of the n samples for each dummy code. The table creation process to create,
The evaluation value calculation unit performs a single regression analysis using the specimen classified into each category using a single regression model with the dummy code in the table as a dummy variable, and the relationship between the dummy variable and the expression of the trait An evaluation value calculation process for calculating an evaluation value indicating the strength of each dummy variable;
A ranking creation process in which the ranking creation unit generates a ranking by arranging the dummy variables in descending order of evaluation values indicating the strength of the relevance;
A dummy variable creating unit extracts the dummy variables up to a preset order and uses the extracted dummy variables as dummy variables of the regression model; and
A variable selection section, wherein the variable selection section includes a variable selection step of calculating a regression coefficient to be multiplied by the extracted dummy variable by a penalized maximum likelihood method. Genome-wide SNP (single nucleotide polymorphism) genotype data is a combination of SNPs that have gene-gene interactions that affect phenotypic expression, comprehensively identifying SNP pairs, and character expression from the SNP pair A gene interaction analysis program for generating a regression model to be analyzed,
Computer
From m types of SNPs detected from n specimens, SNP pairs consisting of two different types of SNPs are selected in sequence, and each genotype (combination of dominant and inferior types) of each SNP in the SNP pair is selected. Table creation means for classifying into a table of dummy codes composed of two categories, classifying each of the n samples for each dummy code, and creating a table indicating the two categories and the samples belonging to each category ,
The single regression model with the dummy code in the table as a dummy variable is used to perform a single regression analysis using the specimen classified into each category, and an evaluation showing the strength of the association between the dummy variable and the expression of the trait Evaluation value calculating means for calculating a value for each dummy variable;
Ranking generating means for generating a ranking by arranging the dummy variables in descending order of evaluation values indicating the strength of the relevance;
Dummy variable creating means for extracting the dummy variables up to a preset order and using the extracted dummy variables as dummy variables of the regression model;
A program for functioning as variable selection means for calculating a regression coefficient to be multiplied by the extracted dummy variable by a maximum likelihood method with penalties.

Images