JP4421971B2 - Analysis engine exchange system and data analysis program - Google Patents

Description

  The present invention starts from univariate analysis such as t-test, Mann-Whitney U test, and Fisher's exact test, and multivariate analysis such as regression analysis, logistic regression analysis, variance analysis, discriminant analysis, principal component analysis, etc. Analytical engine replacement that narrows down significant variables efficiently from large-volume analysis data such as genetic data, especially for data mining methods such as statistical methods such as neural network, binary tree analysis method and support vector machine (SVM) The present invention relates to a type system and an analysis engine exchange type system program (data analysis program).

  First, the first related technique will be described.

In general, one of the purposes of statistical analysis of actual phenomena is to find and predict the relationship between various characteristics. In such a case, use a generalized linear model including regression analysis, logistic regression analysis, discriminant function, etc., or use data mining techniques such as SVM to find some relationship from the data, and for a certain variable It is common to make predictions. For example, a plurality of explanatory variables x ₁ , x ₂ ,. . . , X _p is analyzed. If you create a model formula that uses all the variables of the data, the generality of the model is lost, and it is highly likely that it will not be applicable when applied to other data. In particular, when there are many variables in the data, it is necessary to select as few optimal variables as possible so that the objective variable y can be explained well and to create a model formula. In general, the variables are selected so that the number of explanatory variables in the model formula is several to several tens. In such a case, a general statistical analysis system prepares a variable selection method or a brute force method, and can select a model that is considered to be optimal from a combination of various variable models.

  Next, a second related technique will be described.

  General statistical analysis systems and genetic analysis systems have various analysis methods such as regression analysis, logistic regression analysis, statistical methods such as generalized linear models including discriminant functions, and data mining methods such as SVM. Yes. However, since it is assumed that analysis is performed by specifying a combination of variables only once, in order to repeatedly analyze various combinations of variables for data having tens of thousands of variables, It is necessary to add special processing using programming functions and macro functions.

  Furthermore, since the process cannot be completed within a realistic time even if the combination of variables is indiscriminately examined, the number of variable combinations is reduced as much as possible so that more significant variable combinations are included. Although it is necessary to investigate the combination method efficiently, the above analysis system does not provide such an algorithm. For example, when trying to estimate a model of about five variables, the combination of five variables is as follows for 10,000 variables, and it is practically difficult to calculate all the combinations. is there.

₁₀₀₀₀ C ₅ = (10000 × 9999 × 9998 × 9997 × 9996) / 5! ≒ ¹⁰ 20/5!
In addition, there is no standard for how to select an optimal model combination or a significant variable combination from many estimated models for various combinations of variables. For this reason, in order to select an optimal model or variable combination in the above general analysis system, a special process must be incorporated by a programming function or a macro function.

  Further, a third related technique will be described.

  In order to select the optimal model combination or significant variable combination, rather than using only one type of method, a statistical method such as a generalized linear model or a data mining method such as SVM can be used. It is also necessary to use it. For example, in order to narrow down variables using different analysis methods A and B as shown in FIG. 8, special processing must be incorporated in the analysis system by a programming function or a macro function.

  Claim 1 of Patent Document 1 states that “the input value is set to a predetermined constant value except for one explanatory variable, and the influence relationship between the output value of the neural network output at that time and the objective variable is used in the statistical analysis. It is described that evaluation with the F value or t value to be performed is sequentially performed for all explanatory variables, and all unnecessary explanatory variables below a predetermined value are discarded.

In the 16th to 21st lines of the eighth column of Patent Document 2, “a variable (x ₄ ) whose reference value is maximized in step 54 (FIG. 3) is selected, and this variable is selected in step 51 (FIG. 3). the combination of the two variables that contain _{x 4 (x 4, x 1} ), the _{(x 4,} x 2), made each time come sequentially loop _(x 4, x _3), the reference value in step 52 "Calculate." However, the combination of two variables (x ₄ , x ₁ ), (x ₄ , x ₂ ), (x ₄ , x ₃ ) constitutes the selected (squeezed) variable is x ₄ is _{only, x 1, x 2, x} 3 is (is squeezed) is selected not a variable. That is, first, it is disclosed to select a plurality of significant variables from all variables, and then to create all combinations of at least two from a plurality of selected (narrowed) variables. There is no.

  The 39th to 44th lines in the second column of Patent Document 3 indicate that “the number of explanatory variables used in one analysis is fixed, and an abnormal item (explanatory variable) is automatically generated using a known variable increase / decrease method. ) Is performed a plurality of times, and a final analysis is performed using only the items narrowed down in each analysis. The known variable increase / decrease method used in Patent Document 3 first selects one significant variable from all the variables, then fixes one selected variable and fixes the remaining variables. One variable is selected from among them, and two combinations are created. Therefore, also in this Patent Document 3, first, a plurality of significant variables are selected from all the variables, and then, at least two of all the selected (restricted) variables are selected. There is no disclosure of creating a combination.

  In the summary of Patent Document 4, there is a description “analyzing the relationship between genetic polymorphism site information and phenotype”.

JP 2000-31511 A JP-A-7-93284 JP 2002-110493 A JP 2003-67389 A Cox "Analysis of binary data" Asakura Shoten

The first problem with the related technology described above is that the number of explanatory variables is larger than the number of samples, and the algorithms used in statistical multivariate analysis (such as variable selection and brute force methods) are for gene expression analysis. It cannot be applied to data with a large amount of variables such as DNA chips and microarrays. In the variable selection method (forward selection) and variable increase / decrease method (stepwise selection) in the conventional variable selection method, when a variable is added to or removed from the model, the variable becomes statistically significant by adding or deleting it. Are simply added / deleted one by one, and variables cannot be selected efficiently in screening or the like that narrows down candidate variables from a large number of variables. In addition, in the variable selection method (backward selection), a model in which all variables are taken in first is necessary, but it is impossible to consider a model consisting of 10,000 explanatory variables. In the round robin method, in order to examine combinations of all variables, when the number of variables is p, 2 ^p −1 combinations of models are tried. When p is as large as 10,000, it cannot be calculated realistically.

  As a second problem, univariate statistical analysis is performed on the expression analysis results of individual genes for the gene expression analysis data examined at high throughput, and each gene expression result is evaluated. Importantly, no dedicated device has been developed to automatically repeat the specified analysis for tens of thousands of genes.

  As a third problem, statistical analysis methods applied to DNA chips and microarrays start from univariate analysis such as t-test, Mann-Whitney U test, and Fisher's exact test, regression analysis, logistic regression Not only multivariate analysis such as analysis, variance analysis, discriminant analysis, principal component analysis, but also a variety of data mining methods such as neural network, binary tree analysis method and support vector machine (SVM). An apparatus for efficiently processing data consisting of a large amount of variables such as a DNA chip and a microarray while combining or exchanging these methods has not been developed. First, there is no disclosure to select a plurality of significant variables from all variables, and then to create all combinations of at least two from a plurality of selected (narrowed) variables. .

  An object of the present invention is to provide an analysis engine exchange type system and a data analysis program for an analysis engine exchange type system that can eliminate the above-mentioned problems.

  The object of the present invention is to first select a plurality of significant explanatory variables from all the explanatory variables, and then select all of at least two of the selected (restricted) variables. It is an object of the present invention to provide an analysis engine exchange type system and a data analysis program for an analysis engine exchange type system that can be combined.

  The analysis engine exchange type system according to the present invention and the data analysis program according to the present invention are as follows.

[Claim 1] A data analysis device and an input device for inputting a data file to be analyzed to the data analysis device,
The data analysis apparatus includes an analysis engine control unit and an analysis engine unit,
When the analysis engine control unit receives the data file to be analyzed consisting of one objective variable y and p explanatory variables x ₁ , x ₂ ,..., X _p , the p explanatory variables All the combinations for extracting one explanatory variable from the data are sequentially put together with the objective variable in the analysis engine unit by _p C ₁ (= p) sets of data (y, x ₁ ), (y, x ₂ ), ..., (y, x _p )
The analysis engine unit performs a predetermined analysis on the p sets of data sent, and sends an analysis result to the analysis engine control unit,
The analysis engine control unit selects _p ′ explanatory variables x ′ ₁ ,..., X ′ _p that are higher in the result from the p explanatory variables based on the analysis result. Then, all the combinations for extracting two explanatory variables from the p ′ explanatory variables are sequentially put together with the objective variable into the analysis engine unit _{p ′} C ₂ (= p ′ × (p′−1). ) / 2) sets of data (y, x ′ ₁ , x ′ ₂ ), (y, x ′ ₁ , x ′ ₃ ),..., (Y, x ′ _p′−1 , x ′ _{p ′} ) And pass on as
The analysis engine unit performs another predetermined analysis on the (p ′ × (p′−1) / 2) sets of data sent, and outputs another analysis result to the analysis engine. To the control unit,
The analysis engine control unit selects, based on the other analysis result, from the p ′ explanatory variables, a lower number of explanatory variables than p ′, which is higher in the result. Engine replacement system.

[Claim 2] In the analysis engine exchange type system according to claim 1,
An output device connected to the analysis engine control unit of the data analysis device;
The analysis engine control unit has a function of causing the output device to display the analysis result and the other display result.

[Claim 3] In the analysis engine exchange type system according to claim 1,
The analysis engine unit, for the p sets of data sent, as the predetermined analysis,
y = f (x _i ), i = 1, 2,..., p
P models are each estimated, and the analysis results are sent to the analysis engine control unit the degree of fit for the p models and the significance for the p explanatory variables,
The analysis engine control unit includes the p explanatory variables based on a comparison result with respect to a reference value of the degree of fit with respect to the p models and a comparison result with another reference value of significance with respect to the p explanatory variables. ., An analysis engine exchange type system characterized by selecting _p ′ explanatory variables x ′ ₁ ,..., X ′ _p in the top result.

[Claim 4] In the analysis engine exchange type system according to claim 3,
The analysis engine unit, for the (p ′ × (p′−1) / 2) sets of data sent, as the other predetermined analysis,
y = f (x _i , x _j ), i, j = 1, 2,..., p ′, i ≠ j
The estimation of (p ′ × (p′−1) / 2) models represented by the following is performed, and as another analysis result, (p ′ × (p′−1) / 2) The degree of fit for the model and the significance for (p ′ × (p′−1) / 2) explanatory variables are sent to the analysis engine controller,
The analysis engine control unit compares the (p ′ × (p′−1) / 2) models with a comparison result with respect to a reference value of the degree of fit for the (p ′ × (p′−1) / 2) models. Based on the comparison result of another significance value for the explanatory variable with respect to another reference value, a lower number of explanatory variables than the p ′ number are selected from among the p ′ explanatory variables. Analysis engine exchange type system.

[Claim 5] In the analysis engine exchange type system according to claim 4,
The analysis engine control unit selects a smaller number of explanatory variables than the p ′, and uses the selected small number of explanatory variables to determine the number of explanatory variables of the next model. An analysis engine characterized in that estimation of the next model is executed in a state where the number is increased by one, and a smaller number of explanatory variables are selected from the selected small number of explanatory variables based on the execution result. Interchangeable system.

[Claim 6] In the analysis engine exchange type system according to claim 1,
A recording medium connected to the data analysis device and recorded with a data analysis program is further included,
The data analysis program is read from the recording medium into the data analysis device, and controls the above-described operations of the analysis engine control unit and the analysis engine unit of the data analysis device. .

[Claim 7] A data analysis device having an analysis engine control unit and an analysis engine unit, an input device for inputting a data file to be analyzed to the data analysis device, and for causing the data analysis device to execute predetermined processing The data analysis program in an analysis engine exchange type system having a recording medium recording the data analysis program of
The predetermined process is:
When the analysis engine control unit receives the data file to be analyzed consisting of one objective variable y and p explanatory variables x ₁ , x ₂ ,..., X _p , the p explanatory variables All the combinations for extracting one explanatory variable from the data are sequentially put together with the objective variable in the analysis engine unit by _p C ₁ (= p) sets of data (y, x ₁ ), (y, x ₂ ), ..., the first step passing as (y, x _p ),
A second step in which the analysis engine unit executes a predetermined analysis on each of the p sets of data sent, and sends an analysis result to the analysis engine control unit;
Based on the analysis result, the analysis engine control unit selects _p ′ explanatory variables x ′ ₁ ,..., X ′ _p from the top of the p explanatory variables (p ′ <p). Then, all the combinations for extracting two explanatory variables from the p ′ explanatory variables are sequentially put together with the objective variable into the analysis engine unit _{p ′} C ₂ (= p ′ × (p′−1). ) / 2) sets of data (y, x ′ ₁ , x ′ ₂ ), (y, x ′ ₁ , x ′ ₃ ),..., (Y, x ′ _p′−1 , x ′ _{p ′} ) As a third step,
The analysis engine unit executes another predetermined analysis on the (p ′ × (p′−1) / 2) sets of data sent, and sends another analysis result to the analysis engine. A fourth step to send to the control unit;
A fifth step in which the analysis engine control unit selects, based on the other analysis result, from the p ′ explanatory variables, a lower number of explanatory variables than p ′, which are higher in the result. A data analysis program characterized by comprising.

[Claim 8] In the data analysis program according to claim 7,
When the analysis engine exchange type system further includes an output device connected to the analysis engine control unit of the data analysis device, the analysis engine control unit displays the analysis result and the other display result. A data analysis program further comprising the step of displaying on the output device.

[Claim 9] In the data analysis program according to claim 7,
In the second step, the analysis engine unit performs the predetermined analysis on the p sets of data sent,
y = f (x _i ), i = 1, 2,..., p
Each of the estimations of the p models represented by: and sending the degree of fit to the p models and the significance of the p explanatory variables as analysis results to the analysis engine control unit,
The third step is based on the comparison result of the analysis engine control unit with respect to the reference value of the degree of fit with respect to the p models and the comparison result with respect to another reference value of the significance with respect to the p explanatory variables. A data analysis program characterized in that it is a step of selecting _p ′ explanatory variables x ′ ₁ ,..., X ′ _p that are higher in the result from the p explanatory variables.

[Claim 10] In the data analysis program according to claim 9,
In the fourth step, the analysis engine unit sends the (p ′ × (p′−1) / 2) sets of data sent as the other predetermined analysis,
y = f (x _i , x _j ), i, j = 1, 2,..., p ′, i ≠ j
The estimation of (p ′ × (p′−1) / 2) models represented by the following is performed, and as another analysis result, (p ′ × (p′−1) / 2) Sending the degree of fit to the model and significance for (p ′ × (p′−1) / 2) explanatory variables to the analysis engine controller,
In the fifth step, the analysis engine control unit compares the comparison result with the reference value of the degree of fit for the (p ′ × (p′−1) / 2) models and the (p ′ × (p ′ -1) / 2) Based on the comparison result with respect to another reference value of the significance for the explanatory variables, the explanatory variables of the top number of the p ′ explanatory variables are smaller than the p ′ explanatory variables. A data analysis program characterized by being a step of selecting.

[Claim 11] In the data analysis program according to claim 10,
The analysis engine control unit selects a smaller number of explanatory variables than the p ′ number and uses the selected small number of explanatory variables for the analysis engine unit to determine the number of explanatory variables of the next model. The method further includes the step of executing estimation of the next model in a state where the number is increased by one, and selecting a smaller number of explanatory variables from the selected small number of explanatory variables based on the execution result. A featured data analysis program.

  According to the present invention, first, a plurality of significant explanatory variables are selected from all the explanatory variables, and then, at least two of all the selected (restricted) variables are selected. An analysis engine exchange type system and a data analysis program for the analysis engine exchange type system that can be combined are obtained, and the analysis of the entire explanatory variable can be efficiently completed.

  Next, a first embodiment of the present invention will be described in detail with reference to the drawings.

  Referring to FIG. 1, an analysis engine exchange type system according to a first embodiment of the present invention includes a data analysis device 2 that operates under program control, and an input device that inputs a data file to be analyzed to the data analysis device 2. 1 and an output device 3 such as a display device or a printing device. The data file to be analyzed at this time consists of one objective variable and p explanatory variables. The data analysis device 2 includes an analysis engine control unit 21 and an analysis engine unit 22.

  The analysis engine control unit 21 extracts the objective variable and the selected explanatory variable of the given data file, and sends the data to the analysis engine unit 22. The analysis engine unit 22 performs a predetermined analysis on the sent data and sends the analysis result to the analysis engine control unit 21. In the output device 3, the analysis results sent from the analysis engine control unit 21 are sorted and displayed using statistics and parameters (for example, significance based on statistics for each explanatory variable). Based on the analysis result, the analysis engine control unit 21 selects the top p ′ explanatory variables (p ′ <p), increases the number of explanatory variables included in the next model, and analyzes from the input device 1. Run again. Based on the analysis result, the number of explanatory variables to be selected is reduced. By repeating the process, the number of explanatory variables to be included in the model is increased, but since the number of explanatory variables to be analyzed is decreased, the number of analysis executions is less than the brute force method. It becomes possible to finish the analysis of the whole variable.

  Next, the operation of the present embodiment will be described in detail with reference to FIG.

  The data in the data file to be analyzed is composed of one objective variable and p explanatory variables, as shown in Equation 1 below.

  The analysis engine control unit 21 receives the data file from the input device 1 and sequentially passes all combinations for extracting one explanatory variable from the p explanatory variables to the analysis engine unit 22 together with one objective variable. That is, p sets of data shown in the following Equation 2 are passed.

  The analysis engine unit 22 performs analysis such as regression analysis and logistic regression analysis on each set of data. In this case, p models of Equation 3 below are estimated. That is, the calculation is repeated p times.

  In the case of regression analysis, the regression model equation is as shown in Equation 3 above, and the regression coefficient of the explanatory variable, the statistic indicating its significance, and the statistic indicating the fit of the model are calculated. The statistic that represents the fit of the model and the statistic that represents the significance of the regression coefficient can be defined arbitrarily. For example, the regression coefficient of each model and the statistic that represents the fit of the model are shown in Equation 4 below. Are defined as a multiple correlation coefficient and t-value and p-value as statistics representing the significance of the regression coefficient.

  The result of analyzing the p sets of data is the degree of fit for the p models and the significance for the p explanatory variables. The analysis engine control unit 21 selects a model and a variable based on these results by providing a reference as shown in Equation 5 below.

  This narrows down to p ′ explanatory variables (p ′ <p) out of the p explanatory variables.

  Here, the data file is composed of one objective variable and the above p ′ explanatory variables as shown in Equation 6 below.

The analysis engine control unit 21 receives the data file from the input device 1. Alternatively, the analysis engine control unit 21 creates the data file in the analysis engine control unit 21. Then, the analysis engine control unit 21 sequentially passes all combinations extracted from the p ′ explanatory variables to the analysis engine unit 22 together with one objective variable. That is, _{p ′} C ₂ = [p ′ × (p′−1) / 2! ] = [P ′ × (p′−1) / 2] sets of data are passed.

  The analysis engine unit 22 performs analysis such as regression analysis and logistic regression analysis on each set of data. In this case, p ′ × (p′−1) / 2 models shown in Equation 8 below are estimated.

  A statistical quantity indicating the degree of fit of each statistical quantity and model of regression analysis and logistic regression, and a statistical quantity indicating the significance of each explanatory variable can be obtained in the same manner using Equation 4 above. However, p = 2.

  Similarly, the explanatory variables are narrowed down to several tens according to the criterion shown in the following formula 9.

  Next, using the narrowed explanatory variables, the number of explanatory variables in the model is increased by 1, and the process is repeated. In this way, the explanatory variables are narrowed down to about 10 to 20, so that the relationship between each explanatory variable and the objective variable can be individually investigated.

  The above contents are shown in FIG.

  The analysis engine unit 22 can use any analysis method by replacing it with various statistical methods such as logistic regression analysis, discriminant function, t-test, and Mann-Whitney U test as well as regression analysis. The interface between the analysis engine unit 22 and the analysis engine control unit 21 makes this possible.

  An interface between the analysis engine unit 22 and the analysis engine control unit 21 is shown in FIG.

  In FIG. 3, the analysis engine control unit 21 functions as a repetitive control block, and the number of parameters for one analysis such as a repetitive number and a list 31 thereof are sent to the statistical analysis calculation unit via the parameter and data input unit 22a. To the analysis engine unit 22 acting as The analysis engine unit 22 sends a result 35 such as a statistic and a test statistic to the analysis engine control unit 21 via a result output unit 22b such as a statistic and a test statistic. In block 21a, the analysis engine control unit 21 performs result editing and output, and displays the result display 33 on the output device 3 (FIG. 1). In block 21a, the analysis engine control unit 21 further extracts a variable (explanatory variable) based on the reference value, and causes the output device 3 to display a list of the extracted variables as a variable list display 36 for each repetition.

  4 shows the data structure 31 in FIG. 3 and the data structure 35 in FIG.

  Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

  Referring to FIG. 5, the analysis engine exchange type system according to the second exemplary embodiment of the present invention typically includes a data analysis device 5 that is a CPU (Central Processing Unit) of a computer, and an input device similar to FIG. 1 and an output device 3. Further, the analysis engine exchange type system includes a recording medium 4 on which a data analysis program is recorded. The recording medium 4 may be either a portable type or a fixed type, and may be a magnetic disk, a semiconductor memory, a CD-ROM, or other recording media.

  In addition, a computer program (the above data analysis program) that can execute this method can be stored in a recording device of a computer connected to a network and transferred to another computer via the network. As a providing medium for providing a computer program (the above data analysis program) for executing the present algorithm, it can be distributed as a computer-readable medium in various formats, and is not limited to a specific type of medium. .

  The data analysis program is read from the recording medium 4 into the data analysis device 5, controls the operation of the data analysis device 5, and the data file shown in FIG. The same processing as that performed by the analysis device 2 is executed.

  Next, an embodiment of the present invention will be specifically described with reference to actual data. This example corresponds to the analysis engine exchange type system according to the first embodiment of FIG. As shown in FIG. 6, the analysis data are SNP (Single Nucleotide Polymorphism) data and clinical data, with or without side effects (1 or 0) as the objective variable (Y). SNP data is used as the explanatory variable (X). Here, in the SNP data, for example, two variables are assigned to one SNP as (A / A, A / T, T / T) = (10, 11, 01). As a result, the number of explanatory variables used for the analysis is about 5500 and the number of cases is 54.

  This data is applied to a logistic regression model for analysis. A logistic regression model including p explanatory variables for n individuals assumes the following Equation 10.

Here, θ _i is the success probability for the individual, and λ _i is the logistic transformation of θ _i . α _ik is the value of the k-th explanatory variable for the individual i, and β _k is the regression coefficient on the logistic scale of the k-th explanatory variable.

Given binary response observation values y ₁ , y ₂ ,..., y _{n for n} individuals, the log likelihood is given by Equation 11 below.

とすると、この数式１２で定義された、L(β)を最大にするβの値を、求めるためには、L(β)を目的関数とする非線形最適化問題を解く必要がある。この解法としてここでは、Newton-Raphson法を用いる（非特許文献１を参照）。 here,

Then, in order to obtain the value of β defined by Equation 12 that maximizes L (β), it is necessary to solve a nonlinear optimization problem with L (β) as an objective function. Here, Newton-Raphson method is used as this solution (see Non-Patent Document 1).

  The model in the case of one explanatory variable has an analysis result as shown in FIG.

  In FIG. 7, one analysis result is shown for each row. In each column, “R_Variable No.” is a number representing an objective variable, and “E_Variable No.” is a number representing an explanatory variable. “Status” indicates whether or not the analysis processing was in error, and “X2L” is a test statistic of the logistic regression model. “B0” and “B” are a constant term and a regression coefficient, respectively. “T value” and “P value” are test statistic of each explanatory variable.

  In FIG. 7, the magnitudes of the p values with respect to the test statistic t value of the calculated logistic regression coefficient are sorted in ascending order, and can be seen in the order of the influence of the logistic regression coefficient.

Pearson's χ ² statistic, which is a test statistic of the logistic regression model, is calculated by the following Equation 13.

  The t value of each explanatory variable is calculated as in Equation 14 below.

  Here, s.e. () is the standard error of the elements in ().

  The P value can be calculated by obtaining the upper probability corresponding to the t value of the t distribution because the t value of the above equation 14 follows the t distribution with Np−1 degrees of freedom.

  From the above results, it can be seen which gene has a strong association with the presence / absence of side effects, and it is possible to narrow down the genes with strong association.

According to the first and second embodiments, various statistical analysis methods can be applied to data having a large amount of variables such as a DNA chip for gene expression analysis and a microarray. If the total number of variables is about 30,000,
One explanatory variable: 30,000 times Two explanatory variables: 500,000 times (about 1000 explanatory variables are extracted for one explanatory variable)
3 explanatory variables: 170,000 times (about 100 explanatory variables are extracted when 2 explanatory variables)
:
:
Thus, the processing can be completed within a realistic time.

  Furthermore, according to the first and second embodiments, candidate variables can be narrowed down from among a large number of variables more efficiently than existing variable selection methods. This is because candidate variables can be selected independently for each number of explanatory variables included in the model. Furthermore, the existing variable increasing method, variable increasing / decreasing method, and the like can be applied by expanding the above mathematical formula 3 to a model in which specific explanatory variables are fixed as in the following mathematical formula 15.

  In addition, according to the first and second embodiments, how to select the most appropriate model combination or significant variable combination from the models for many estimated combinations of variables. It is easy to set the standard of whether it is good. Select a statistical combination such as the estimated correlation coefficient or F value of the model itself, or a statistical value such as t value or p value for each variable, and select a combination of variables above (or below) the reference value. can do.

  Furthermore, according to the first and second embodiments, it is possible to extract a model from variable candidates that are wider than those of the existing variable selection method. In the existing variable increase / decrease method, decrease method, increase method, etc., only one model is extracted. However, since all the calculated model results are stored in the present invention, it is possible to set a standard for the model and display the top K analysis results to the user to examine the results. Furthermore, it is possible to proceed to the next model selection step using the explanatory variables included in the top K models, and it is possible to extract models from a wider range of variable candidates than the existing variable selection method.

  Further, according to the first and second embodiments, it is possible to use a plurality of analysis methods in a crosswise combination. Techniques applied to DNA chips and microarrays start with univariate analysis such as t-test, Mann-Whitney U test, and Fisher's exact test, and then regression analysis, logistic regression analysis, variance analysis, discriminant analysis, principal component There are a wide variety of statistical methods such as multivariate analysis such as analysis, neural network, binary tree analysis method and data mining method such as support vector machine (SVM), but by replacing the analysis engine part with these methods, explanatory variables It is possible to analyze by combining different methods when narrowing down.

It is a block diagram of an analysis engine exchange type system by a 1st embodiment of the present invention. It is a flowchart which shows the process of narrowing down a variable in the analysis engine control part of the analysis engine exchange type system of FIG. FIG. 2 is a block diagram illustrating an interface between an analysis engine unit and an analysis engine control unit of the analysis engine exchange type system of FIG. 1. It is a figure which shows the data structure of the part 31 of FIG. 3, and the data structure of the part 35 of FIG. It is a block diagram of an analysis engine exchange type system by a 2nd embodiment of the present invention. It is the figure which showed the structure of the analysis data used for description of operation | movement of the Example corresponding to the analysis engine exchange type system of FIG. It is the figure which showed the analysis result in the said Example. It is a flowchart which shows the process for narrowing down a variable using a different analysis method.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Input device 2 Data analysis device 3 Output device 4 Recording medium 5 Data analysis device 21 Analysis engine control part 22 Analysis engine part

Claims (9)

A data analysis device and an input device for inputting a data file to be analyzed to the data analysis device;
The data analysis apparatus includes an analysis engine control unit and an analysis engine unit,
When the analysis engine control unit receives the data file to be analyzed consisting of one objective variable y and p explanatory variables x ₁ , x ₂ ,..., X _p , the p explanatory variables All the combinations for extracting one explanatory variable from the data are sequentially put together with the objective variable in the analysis engine unit by _p C ₁ (= p) sets of data (y, x ₁ ), (y, x ₂ ), ..., having a first control means passing as (y, x _p ),
The analysis engine unit includes a second control unit that executes predetermined analysis on each of the p sets of data that has been sent, and sends analysis results to the analysis engine control unit.
Based on the analysis result, the analysis engine control unit selects p ′ explanatory variables x ′ ₁ ,..., X ′ _{p ′} that are higher in the result from the p explanatory variables. Next, all combinations for extracting two explanatory variables from the p ′ explanatory variables are sequentially added to the analysis engine unit along with the objective variable in the analysis engine unit _{p ′} C ₂ (= p ′ × (p′− 1) / 2) sets of data (y, x ′ ₁ , x ′ ₂ ), (y, x ′ ₁ , x ′ ₃ ),..., (Y, x ′ _p′−1 , x ′ _{p ′} ) As a third control means
The analysis engine unit performs another predetermined analysis on the (p ′ × (p′−1) / 2) sets of data sent, and outputs another analysis result to the analysis engine. A fourth control means for sending to the control unit ;
The analysis engine control unit includes a fifth control unit that selects, based on the other analysis result, a lower number of explanatory variables than p ′, which is higher in the result, from the p ′ explanatory variables. Have
The predetermined analysis and the other predetermined analysis are analysis by regression analysis, the explanatory variable is genetic data, the target variable is presence / absence of side effect (1/0), Analysis engine exchange type system characterized by narrowing down genetic data strongly related to existence / non-existence of In the analysis engine exchange type system according to claim 1,
An output device connected to the analysis engine control unit of the data analysis device;
The analysis engine control system further comprises means for outputting the analysis result and the other analysis result to the output device. In the analysis engine exchange type system according to claim 1,
The second control means in the analysis engine unit, as the predetermined analysis for the p sets of data sent,
y = f (x _i ), i = 1, 2,..., p
P models are each estimated, and the analysis results are sent to the analysis engine control unit, as analysis results, the degree of fit for the p models and the significance for the p explanatory variables.
The third control means in the analysis engine control unit is based on a comparison result with respect to a reference value of the degree of fit for the p models and a comparison result with another reference value of significance for the p explanatory variables. the out of p number of explanatory variables, the results of the upper p 'number of explanatory variables x' _1, ... analysis engine switched system, characterized in that, it is to select the x _'p'. In the analysis engine exchange type system according to claim 3,
The fourth control means in the analysis engine unit, as the other predetermined analysis, for the (p ′ × (p′−1) / 2) sets of data sent,
y = f (x _i , x _j ), i, j = 1, 2,..., p ′, i ≠ j
The estimation of (p ′ × (p′−1) / 2) models represented by the following is performed, and as another analysis result, (p ′ × (p′−1) / 2) The degree of fit for the model and the significance for (p ′ × (p′−1) / 2) explanatory variables are sent to the analysis engine control unit,
The fifth control means in the analysis engine control unit includes a comparison result with respect to a reference value of the degree of fit for the (p ′ × (p′−1) / 2) models and the (p ′ × (p ′ -1) / 2) Based on the comparison result with respect to another reference value of the significance for the explanatory variables, the explanatory variables of the top number of the p ′ explanatory variables are smaller than the p ′ explanatory variables. analysis engine switched system, characterized in that to select. In the analysis engine exchange type system according to claim 1,
A recording medium that records a data analysis program to be read by the data analysis device;
The data analysis program is read from the recording medium into the data analysis device, and the data analysis program controls the analysis engine control unit of the data analysis device and first to fifth control means of the analysis engine unit . An analysis engine exchange type system characterized by As an analysis engine control unit and an analysis engine unit in an analysis engine exchange type system having a data analysis device having an analysis engine control unit and an analysis engine unit, and an input device for inputting a data file to be analyzed to the data analysis device A data analysis program for causing a computer to function ,
Before SL analysis engine control unit, one objective variable y and p number of explanatory variables x _{1 of,} x 2, _..., consisting of x _p, when receiving the data file serving as the analysis target, the p number of Description All combinations for extracting one explanatory variable from the variables are sequentially added to the analysis engine unit along with the objective variable in the _p C ₁ (= p) sets of data (y, x ₁ ), (y, x ₂ ). ,..., (Y, x _p )
A second step in which the analysis engine unit executes a predetermined analysis on each of the p sets of data sent, and sends an analysis result to the analysis engine control unit;
Based on the analysis result, the analysis engine control unit selects p ′ explanatory variables x ′ ₁ ,..., X ′ _{p ′} that are higher in the result from the p explanatory variables. Next, all combinations for extracting two explanatory variables from the p ′ explanatory variables are sequentially added to the analysis engine unit along with the objective variable in the analysis engine unit _{p ′} C ₂ (= p ′ × (p′− 1) / 2) sets of data (y, x ′ ₁ , x ′ ₂ ), (y, x ′ ₁ , x ′ ₃ ),..., (Y, x ′ _p′−1 , x ′ _{p ′} ) As a third step,
The analysis engine unit executes another predetermined analysis on the (p ′ × (p′−1) / 2) sets of data sent, and sends another analysis result to the analysis engine. A fourth step to send to the control unit;
A fifth step in which the analysis engine control unit selects, based on the other analysis result, from the p ′ explanatory variables, a lower number of explanatory variables than p ′, which are higher in the result; Including processing ,
The predetermined analysis and the other predetermined analysis are analysis by regression analysis, the explanatory variable is genetic data, the target variable is presence / absence of side effect (1/0), A data analysis program characterized by narrowing down genetic data strongly related to the presence or absence of. In the data analysis program according to claim 6,
When the analysis engine exchange type system further includes an output device connected to the analysis engine control unit of the data analysis device, the analysis engine control unit displays the analysis result and the other analysis result. A data analysis program further comprising a step of outputting to the output device. In the data analysis program according to claim 6,
In the second step, the analysis engine unit performs the predetermined analysis on the p sets of data sent,
y = f (x _i ), i = 1, 2,..., p
Each of the estimations of the p models represented by: and sending the degree of fit to the p models and the significance of the p explanatory variables as analysis results to the analysis engine control unit,
The third step is based on the comparison result of the analysis engine control unit with respect to the reference value of the degree of fit with respect to the p models and the comparison result with respect to another reference value of the significance with respect to the p explanatory variables. A data analysis program characterized by the step of selecting _{p ′} explanatory variables x ′ ₁ ,..., X ′ _{p ′} that are higher in the result from the p explanatory variables. In the data analysis program according to claim 8,
In the fourth step, the analysis engine unit sends the (p ′ × (p′−1) / 2) sets of data sent as the other predetermined analysis,
y = f (x _i , x _j ), i, j = 1, 2,..., p ′, i ≠ j
The estimation of (p ′ × (p′−1) / 2) models represented by the following is performed, and as another analysis result, (p ′ × (p′−1) / 2) Sending the degree of fit to the model and significance for (p ′ × (p′−1) / 2) explanatory variables to the analysis engine controller,
In the fifth step, the analysis engine control unit compares the comparison result with the reference value of the degree of fit for the (p ′ × (p′−1) / 2) models and the (p ′ × (p ′ -1) / 2) Based on the comparison result with respect to another reference value of the significance for the explanatory variables, the explanatory variables of the top number of the p ′ explanatory variables are smaller than the p ′ explanatory variables. A data analysis program characterized by being a step of selecting.

Images