JP2021111141A - Feature network extraction device, computer program, feature network extraction method, and bayesian network analysis method - Google Patents

Abstract

To provide a feature network extraction device, a computer program, a feature network extraction method, and a Bayesian network analysis method capable of evaluating the relationship between samples or groups of samples in an estimated Bayesian network.SOLUTION: In the feature network extraction device, a processor assigns data to required nodes of a Bayesian network representing dependency relationships among a plurality of nodes to which each random variable is assigned using an acyclic directed graph. When calculating posterior probability of the nodes based on the assigned data, a calculation unit calculates a feature value of a link from a parent node to a child node based on a predetermined model that constitutes conditional probability when the probability variable of the parent node is given, so that a feature network is extracted from the Bayesian network based on the calculated feature values.SELECTED DRAWING: Figure 34

Description

The present invention relates to a feature network extraction device, a computer program, a feature network extraction method, and a Bayesian network analysis method.

The Bayesian network is one of the graphical models (statistical models using graph representation), and the causal relationship of multivariates is represented by the network (non-circular directed graph). By learning the structure of a Bayesian network from a large amount of data, the Bayesian network can be estimated and the causal relationship between multivariates can be estimated.

In Patent Document 1, a user creates a label object having an "expression" as a candidate for a node name or a definition area name on a GUI screen, and the label object arranged on the screen has a relationship (causality) between labels. A device that can easily create a Bayesian network by defining a relationship (relationship or propositional relationship) by operating a mouse is disclosed.

JP-A-2007-102717

However, even if a Bayesian network is estimated using a large amount of data, the estimated Bayesian network is a large amount of data that has commonality among variables hidden in the data (for example, a mass of data). Only some relationship is presumed, and for example, the relationship between individual samples or sample groups cannot be explained.

The present invention has been made in view of such circumstances, and is capable of evaluating the relationship between samples or sample groups in an estimated Bayesian network. Feature network extraction device, computer program, feature network extraction method, and Bayesian network. The purpose is to provide an analytical method.

The present application includes a plurality of means for solving the above problems. For example, the feature network extraction device displays a dependency relationship between a plurality of nodes to which each random variable is associated, and a non-circulating directed graph. The random variable of the parent node was given when calculating the posterior probability of the node based on the data given part that gives data to the required node of the Bayesian network represented by the above and the data given by the data given part. A calculation unit that calculates the feature amount of the link from the parent node to the child node based on a predetermined model that constitutes the conditional probability of the time, and a feature network from the Bayesian network based on the feature amount calculated by the calculation unit. It is provided with an extraction unit for extracting data.

According to the present invention, a feature network that characterizes the relationship between a sample or a sample group in an estimated Bayesian network can be extracted, and a sample or a sample group in the estimated Bayesian network can be evaluated.

It is a block diagram which shows an example of the structure of the feature network extraction apparatus of this embodiment. It is a schematic diagram which shows an example of a Bayesian network. It is a schematic diagram which shows an example of the nonparametric regression model using a B-spline. It is a schematic diagram which shows an example of a nonparametric Bayesian network. It is a schematic diagram which shows the 1st example of the feature amount of a branch. It is a schematic diagram which shows the 2nd example of the feature amount of a branch. It is a schematic diagram which shows the concept of ΔECv for the branch from the parent node of the variable X to the child node of the variable Y. It is a schematic diagram which shows the 3rd example of the feature amount of a branch. It is a schematic diagram which shows the 4th example of the feature amount of a branch. It is a schematic diagram which shows the 1st example of the extraction method of a feature network. It is a schematic diagram which shows the 2nd example of the extraction method of a feature network. It is a schematic diagram which shows the 3rd example of the extraction method of a feature network. It is a schematic diagram which shows the 1st example of the extracted feature network. It is a schematic diagram which shows the 1st example of the characterization of an individual by a feature network. It is a schematic diagram which shows the other structure of the ECv matrix. It is a schematic diagram which shows the 2nd example of the characterization of an individual by a feature network. It is a schematic diagram which shows the 2nd example of the extracted feature network. It is a schematic diagram which shows the 3rd example of the extracted feature network. It is a schematic diagram which shows the mechanism which can capture the gene of the immune system by the extracted feature network. It is a schematic diagram which shows the 3rd example of the characterization of an individual by a feature network. It is a schematic diagram which shows the 4th example of the characterization of an individual by a feature network. It is a schematic diagram which shows the 5th example of the characterization of an individual by a feature network. It is a schematic diagram which mapped the extracted feature network to the whole network. It is a schematic diagram which shows the 1st example of the relation between the extracted feature network and a DEG gene. It is a schematic diagram which shows the 2nd example of the relation between the extracted feature network and a DEG gene. It is a schematic diagram which shows the 4th example of the extracted feature network. It is a schematic diagram which shows the example which extracted the path related to the onset of chronic kidney disease (CKD). It is a schematic diagram which shows the example which extracted the path related to the onset of hypertension. It is a schematic diagram which shows the example of the two-disease-related network of CKD and hypertension in the case of having SNP. It is a schematic diagram which shows the example of the two-disease-related network of CKD and hypertension in the case of no SNP. It is a schematic diagram which shows the 1st example of the personal network of the onset of chronic kidney disease (CKD). It is a schematic diagram which shows the 2nd example of the personal network of the onset of chronic kidney disease (CKD). It is a schematic diagram which shows the 3rd example of the personal network of the onset of chronic kidney disease (CKD). Features It is a flowchart which shows an example of the processing procedure of a network extraction apparatus. It is a flowchart which shows an example of a feature network extraction process.

Hereinafter, the present invention will be described with reference to the drawings showing the embodiments thereof. FIG. 1 is a block diagram showing an example of the configuration of the feature network extraction device 50 of the present embodiment. Features The network extraction device 50 includes a processor 51, an operation unit 52, an interface unit 53, a display panel 54, a recording medium reading unit 55, a ROM 56, a memory 57 (for example, RAM), and a storage unit 58. The storage unit 58 can store a pre-estimated Bayesian network model 581 and sample data 582. The feature network extraction device 50 may be composed of one device or a plurality of devices.

The processor 51 includes hardware such as a CPU (for example, a multi-processor in which a plurality of processor cores are mounted), a GPU (Graphics Processing Units), a DSP (Digital Signal Processors), and an FPGA (Field-Programmable Gate Arrays). It can be configured by combining.

The display panel 54 can be composed of a liquid crystal panel, an organic EL (Electro Luminescence) display, or the like.

The operation unit 52 is composed of, for example, a hardware keyboard, a mouse, or the like, and can operate an icon or the like displayed on the display panel 54 or input characters or the like. The operation unit 52 may be composed of a touch panel.

The interface unit 53 can acquire sample data, an estimated Bayesian network model, and the like from an external device or the like. The interface unit 53 has a wired communication function and a wireless communication function. The sample data and the Bayesian network model acquired via the interface unit 53 can be stored in the storage unit 58.

The recording medium reading unit 55 can, for example, read the recording medium M on which the computer program for which the procedure for extracting the feature network is defined is recorded, and store the read computer program in the storage unit 58. The computer program in which the procedure for extracting the feature network is defined may be acquired from an external device or the like via the interface unit 53.

The storage unit 58 may be composed of a hard disk, a flash memory, or the like. By reading the computer program stored in the storage unit 58 into the memory 57 and processing it by the processor 51, the feature network can be extracted.

The processor 51 can execute functions as a data addition unit, a calculation unit, an extraction unit, and a setting unit.

Before going into the description of the feature network extraction method by the feature network extraction device 50, first, the outline of the Bayesian network will be described as a premise.

FIG. 2 is a schematic diagram showing an example of a Bayesian network. The Bayesian network is one of the graphical models (statistical models using graph representation), and the causal relationship of multivariates is represented by the network (non-circular directed graph). In the figure, a circle indicates a node to which a random variable (also simply referred to as a “variable”) is associated, and an arrow indicates a branch (link or edge). The branches have a direction as shown by the arrow, and the node on the upstream side of the arrow is called the parent node, and the node on the downstream side of the arrow is called the child node. In the example of FIG. 2, six nodes are illustrated corresponding to _{the variables X 1} , X ₂ , X ₃ , X ₄ , X ₅ , and X _6.

Pr (X _1, X _2, X _3, X _4, X _5, X ₆ ) represents the joint probability (distribution) for the variables X ₁ , X ₂ , X ₃ , X ₄ , X ₅ , X _6. By exploring how this simultaneous probability can be decomposed, that is, conditional independence, Pr (X _1, X _2, X _3, X _4, X _5, X ₆ ) is Pr (X _j | It can be expressed as the product of conditional probabilities Pa (X _j)). Here, j is the index of the variable, and in the example of FIG. 2, p = 6. Pa (X _j ) is a set of variables corresponding to the parent node in the network of the _{variable X j.} Pr (X ₄ | X _1, X ₂ ) represents the conditional probability of the variable X ₄ given the values of the variables X ₁ and X _2.

In the example of FIG. 2 shows that the variable X _3, X ₄ are independently under variable X ₁ (i.e., independently is with the proviso that the value of variable X ₁ is known). The variables X ₅ and X ₆ are also independent under the _{variable X 3.} A variable that is conditional independence means that the variable that is the condition does not show a correlation when it is fixed to a specific value, which can be regarded as a causal relationship. Since Bayesian networks are estimated using a large amount of data, it is possible to estimate causal relationships between sets of common data.

FIG. 3 is a schematic diagram showing an example of a nonparametric regression model using a B-spline. When the relationships between variables are non-linear, what model to use is important. Nonparametric regression is a method of performing regression without assuming a specific functional form when the relationship between variables does not follow a specific functional form such as a linear expression or a polynomial and is unknown. The decomposition of the simultaneous probabilities for the variables X ₁ , X ₂ ..., X _{p is} expressed as the decomposition of the probability density function f (X _j | Pa (X _j )). The probability density function f (X _j | Pa (X _j )) can be constructed by a nonparametric regression model using a B-spline. As shown in FIG. 3, nodes of the variable X ₄ when the variable of the parent node and X _1, X _2, and data x ₄ variables X _4, variable X _1, the data x ₁ of X _2, and x ₂ In the _{meantime, the relationship x 4} = m ₁ (x ₁ ) + m ₂ (x ₂ ) + ε is established. m ₁ and m ₂ are smooth functions (non-linear functions), and ε is a numerical value that cannot be expressed by a model, and is also called a noise term. N (0, σ ² ) is a normal distribution with a mean of 0 and a variance of σ ^2.

FIG. 4 is a schematic diagram showing an example of a nonparametric Bayesian network. The nonparametric Bayesian network uses a B-spline nonparametric regression model as illustrated in FIG. 3 for the local probability distribution of the Bayesian network. Unlike a general Bayesian network as illustrated in FIG. 2, a nonparametric Bayesian network can handle non-linear continuous values.

In the example of FIG. 4, six nodes are illustrated corresponding to _{the variables X 1} , X ₂ , X ₃ , X ₄ , X ₅ , and X _{6, similar to the example of FIG.} In the equation shown in FIG. 4, i indicates the index of the sample and j indicates the index of the variable. In the example of FIG. 4, j = 1, 2, ..., 6. k indicates the index of the parent node. The function m _jk is a function from the parent node k to the child node node j. In the equation representing the _{function m jk , b lk} is a given M _jk B-spline basis functions, γ _lk is a coefficient parameter for the B-spline basis functions, and a nonparametric Bayesian network is estimated. Is fixed. The basis functions are not limited to B-spline basis functions, and other basis functions such as Fourier series, polynomial basis, regression spline basis, and wavelet basis may be used.

Next, the details of the feature network extraction device 50 will be described. In the present embodiment, the nonparametric regression model will be described as a predetermined model that constitutes the conditional probability when the random variable of the parent node is given, but the predetermined model is not limited to the nonparametric regression model. Further, a nonlinear function will be described as a predetermined function representing a predetermined model of the random variable of the parent node with respect to the random variable of the child node, but the predetermined function is not limited to the nonlinear function. In this embodiment, the Bayesian network is a nonparametric Bayesian network. In the following, the nonparametric Bayesian network will also be referred to as a Bayesian network.

Features The network extraction device 50 (processor 51) assigns data to the required nodes of a Bayesian network in which the dependency relationships between a plurality of nodes to which each random variable is associated are represented by using a non-circulating directed graph. When calculating the posterior probabilities of a node based on the data, the link from the parent node to the child node is based on a non-parametric regression model that constitutes the conditional probability given the random variables of the parent node. It is possible to perform a process of calculating a feature amount and a process of extracting a feature network from a Bayesian network based on the calculated feature amount. The feature network extraction device 50 of the present embodiment can extract a partial network of a Bayesian network estimated in advance as a feature network by using the feature amount. Hereinafter, each process will be described.

The processor 51 assigns the variable data of each node to the required nodes of the nonparametric Bayesian network. The required node can be appropriately determined depending on what kind of data is used and what kind of causal relationship between variables is to be obtained. The variable data includes, for example, various measured values of electronic chart data and medical examination data (including data related to medical practice, test data, data related to pharmaceuticals, etc.), data related to genes (gene expression data, epigenome data, proteome data, etc.). Genomic mutation data such as SNP (Single Nucleotide Polymorphism) and CNV (Copy Number Variations)) are included, but are not limited thereto. In addition, the data may be static data such that each sample is independent, such as an individual sample, and electronic medical records / health examination data and time-series expression data of drug administration that are regularly inspected and recorded. It may be dynamic / time series data such as.

When the processor 51 calculates the posterior probability (simultaneous probability) of another node based on the given data, the child from the parent node is based on the conditional probability when the random variable of the parent node is given. Calculate the feature quantity of the branch to the node. More specifically, the processor 51 calculates the feature amount of the branch from the parent node to the child node based on the function value of the nonlinear function representing the regression model of the random variable of the parent node with respect to the random variable of the child node.

Next, the feature amount of the branch (branch evaluation method) will be described. The feature amount can be defined based on the required formula as shown in FIGS. 5 to 9, and when extracting the feature network, a suitable one of the defined feature amounts can be used.

FIG. 5 is a schematic diagram showing a first example of the feature amount of a branch. As shown in FIG. 5, q parent nodes exist for the child node to which the variable y is associated, and the variables of each parent node are x ₁ , x ₂ , ..., X _q . In this case, there are q branches (links) between the child node and each corresponding parent node. Based on the nonparametric regression model, between the variables y and the variables x ₁ , x ₂ , ..., X _q , y = m ₁ (x ₁ ) + m ₂ (x ₂ ) + ... + m _q (x _q ) The relationship of + ε holds. Let _{the feature amount of x j} (j = 1 to q) to y be the branch contribution amount ECv (Edge Contribution value). The branch contribution amount ECv is defined as ECv (x _j → y) = m _j (x _j ). The branch contribution amount ECv is a function value of the _{function m j.} That is, the processor 51 can calculate the function value of the non-linear function as the feature amount of the branch. The branch contribution ECv of the sample group composed of a plurality of samples can be a statistical value (for example, average value, median value, etc.) of the branch contribution ECv of each sample.

FIG. 6 is a schematic view showing a second example of the feature amount of the branch. As shown in FIG. 6, q parent nodes exist for the child node to which the variable y is associated, and the variables of each parent node are x ₁ , x ₂ , ..., X _q . As in the case of FIG. 5, between the variables y and the variables x ₁ , x ₂ , ..., X _q , y = m ₁ (x ₁ ) + m ₂ (x ₂ ) + ... + m _q (x _q). ) + Ε. Let _{ΔECv be the feature amount of x j} (j = 1 to q) to y. Assuming that the ECvs for the data of the two samples A and B are ECv (x _j ^A → y ^A ) and ECv (x _j ^B → y ^B ), respectively, ΔECv is ΔECv (x _j → y, A, B) = It is defined as | ECv (x _j ^A → y ^A ) -ECv (x _j ^B → y ^B ) |. That is, when different sample data are assigned, the processor 51 sets a comparison value between the first function value of the nonlinear function based on the data of the first sample and the second function value of the nonlinear function based on the data of the second sample. It can be calculated as the feature amount of the branch.

FIG. 7 is a schematic diagram showing the concept of ΔECv for a branch from the parent node of the variable X to the child node of the variable Y. In the figure, the horizontal axis represents the value of the variable X, and the vertical axis represents the value of the variable Y. The value of the variable X can be a continuous value. The curve in the figure shows a nonparametric regression model between variables X and Y, and _{can be represented by Y = m 1} ^(Y) (X). In FIG. 7, the length of the arrow indicates ΔECv between two sample sets of a control sample group (for example, a sample group in which a specific symptom does not appear) and a target sample group (for example, a sample group in which a specific symptom appears). It is represented by. In the example of FIG. 7, ΔECv between two sample groups is shown, but ΔECv is not limited to comparison between two sample groups, and an individual (one sample) and another individual. It may be ΔECv between the individual and the average of all samples.

FIG. 8 is a schematic view showing a third example of the feature amount of the branch. As shown in FIG. 8, q parent nodes exist for the child node to which the variable y is associated, and the variables of each parent node are x ₁ , x ₂ , ..., X _q . As in the case of FIG. 5, between the variables y and the variables x ₁ , x ₂ , ..., X _q , y = m ₁ (x ₁ ) + m ₂ (x ₂ ) + ... + m _q (x _q). ) + Ε. Let _{the feature amount of x j} (j = 1 to q) to y be the relative contribution RC. Relative contribution RC is defined as RC (x _j → y) = | m _j (x _j ) | / max | m _k (x _k ) |. The relative contribution RC is a value from 0 to 1. Here, k is 0 <k ≦ q. That is, the processor 51 is the ratio of the function value of the nonlinear function corresponding to the branch to the maximum value of the function values of each nonlinear function representing the regression model of the random variable of each of the plurality of parent nodes with respect to the random variable of the child node. Can be calculated as the feature amount of the branch. The relative contribution RC of the sample group composed of a plurality of samples can be a statistical value (for example, average value, median value, etc.) of the relative contribution RC of each sample.

FIG. 9 is a schematic view showing a fourth example of the feature amount of the branch. As shown in FIG. 9, q parent nodes exist for the child node to which the variable y is associated, and the variables of each parent node are x ₁ , x ₂ , ..., X _q . As in the case of FIG. 5, between the variables y and the variables x ₁ , x ₂ , ..., X _q , y = m ₁ (x ₁ ) + m ₂ (x ₂ ) + ... + m _q (x _q). ) + Ε. Let _{the feature amount of x j} (j = 1 to q) to y be the relative contribution rate RCr. The relative contribution rate RCr is defined as RCr (x _j → y) = | m _j (x _j ) | / Σ | m _k (x _k ) |. The relative contribution rate RCr is a value from 0 to 1. Here, Σ is the sum of k = 1 to q. That is, the processor 51 measures the ratio of the function value of the nonlinear function corresponding to the branch to the total value of the function values of each nonlinear function representing the regression model of the random variable of each of the plurality of parent nodes with respect to the random variable of the child node. It can be calculated as a feature amount of a branch. The relative contribution rate RCr of the sample group composed of a plurality of samples can be a statistical value (for example, average value, median value, etc.) of the relative contribution rate RCr of each sample.

The processor 51 can extract the feature network from the Bayesian network based on the calculated feature amount. Specifically, the processor 51 can extract a feature network including the branch when the feature amount of the branch is equal to or more than a predetermined threshold value. As described above, as the feature amount, a branch contribution amount ECv, ΔECv, a relative contribution degree RC, a relative contribution rate RCr, or the like can be used. Further, the threshold value does not have to be a fixed value, and may be changed according to the sample, or may be changed when the required node to which the data is given is changed. Further, the threshold value may be a required range determined by a combination of an upper limit value and a lower limit value. The feature quantity can quantify the important factors in the model that determine the variable y from the parent node (for example, variables x ₁ , x ₂ , ..., X _{q) to the child node (for example, variable y).} That is, by extracting a feature network using a feature quantity, a plurality of related paths showing relevance to a sample (for example, an individual or a specific disease) or a sample group in a pre-estimated Bayesian network (model). Can be extracted and samples or groups of samples in the estimated Bayesian network can be evaluated.

Next, a method of extracting the feature network will be described. FIG. 10 is a schematic diagram showing a first example of a feature network extraction method. As shown in the left figure, for convenience, it is assumed that the estimated Bayesian network is composed of 15 nodes. When the data of the sample A is given to the variable of the required node and the simultaneous probability of the variable of the node is calculated, the feature amount of the branch is calculated. In the example of FIG. 10, it is assumed that the branch contribution amount ECv is used as the feature amount. The branch contribution amount ECv of each branch is compared with the threshold value, and the branches whose branch contribution amount ECv is equal to or more than the threshold value are represented by thick lines. In this case, the branches can be specified in the order of indexes 3, 6, 8, 11, and 13, and the feature network connecting the specified branches can be extracted. Assuming that the variable of index 8 is the variable of the factor of interest among the variables, it is possible to identify another factor having a causal relationship with the factor of interest for sample A. In the example of FIG. 10, the feature network is extracted as one network, but it may be extracted as a plurality of independent networks, that is, a plurality of networks that are not connected to each other.

FIG. 11 is a schematic diagram showing a second example of the feature network extraction method. As in FIG. 10, for convenience, it is assumed that the estimated Bayesian network is composed of 15 nodes. When the data of the sample B is given to the variable of the required node and the simultaneous probability of the variable of the other node is calculated, the feature amount of the branch is calculated. In the example of FIG. 11, it is assumed that the branch contribution amount ECv is used as the feature amount. The branch contribution amount ECv of each branch is compared with the threshold value, and the branches whose branch contribution amount ECv is equal to or more than the threshold value are represented by thick lines. In this case, the branches can be specified in the order of indexes 2, 5, 8, 10, 12, and 15, and the feature network connecting the specified branches can be extracted. Assuming that the variable of index 8 is the variable of the factor of interest among the variables, other factors having a causal relationship with the factor of interest can be specified for sample B. Although one factor of interest is shown in FIGS. 10 and 11, there may be a plurality of factors of interest.

Comparing FIG. 11 with the case of FIG. 10, there is a difference in the extracted feature networks between the samples A and B. In this way, important pathways (branch connections) for each sample (individual) can be extracted, and weighted personal networks in the estimated Bayesian network can be extracted. That is, the feature network that characterizes the sample or sample group in the estimated Bayesian network can be extracted, the sample or sample group in the estimated Bayesian network can be evaluated, and the characterization of the individual measurement data ( Explanation) is possible.

FIG. 12 is a schematic diagram showing a third example of the feature network extraction method. The processor 51 sets a plurality of required nodes of the Bayesian network. The required node settings can be made based on the user's specifications. In the example of FIG. 12, the upstream node S1 and the downstream node S2 are set as the setting nodes.

The processor 51 calculates the feature amount of one or a plurality of branches constituting each of a plurality of paths (pathways) from one set node (upstream node S1) to another node (downstream node S2). do. The processor 51 extracts a feature network including a path in which the feature amount of all the branches constituting the path is equal to or more than a predetermined threshold value among the plurality of paths.

In the example of FIG. 12, if the relative contribution RC of each of the six branches from the setting node S1 to the node S2 is RC1, RC2, RC3, RC4, RC5, RC6, the feature amount E of the entire six branches is , (RC1, RC2, RC3, RC4, RC5, RC6) can be calculated by the 6th root. Features can be calculated for other paths in the same way. If the 6th root of (RC1, RC2, RC3, RC4, RC5, RC6) is equal to or greater than the threshold value, the branch group shown by the thick line in FIG. 12 is extracted as the feature network connecting the setting node S1 and the node S2. NS.

Next, an example of the feature network extracted by using the above-mentioned extraction method will be described. FIG. 13 is a schematic diagram showing a first example of the extracted feature network. The estimated network shown on the left is an example of an EMT gene network, for example, the number of nodes (variables) is about 20,000 and the number of branches is about 300,000. EMT is an epithelial to mesenchymal transition, and when epithelial cells become EMT, they have the ability to move away from cancer cells and enter the blood to cause metastasis. EMT-related proteins are attracting attention as biomarkers for cancer. The EMT gene network is a network representing cells that have become EMT and cells that have not become EMT. The feature network shown on the right side is extracted from the estimated network using ΔECv as the feature amount of the branch. Specifically, the ΔECv between the EMT-ized cells and the non-EMT-ized cells is calculated, the branches whose calculated ΔECv is equal to or higher than a predetermined threshold value are specified, and the feature network composed of the specified branches is extracted. is doing. Features The number of nodes in the network is about 150 and the number of branches is about 120.

FIG. 14 is a schematic diagram showing a first example of characterization of an individual by a feature network. FIG. 14A is referred to as an ECv matrix, where each row shows a branch (branch index) of the feature network and each column shows a sample (individual) ECv. Each element of the matrix represents an ECv at each branch of each sample. Each sample here uses lung cancer patient sample data from a public database of cancer patients, and may not be used for estimating the Bayesian network and extracting the feature network. The sample group can be classified into two clusters, cluster (group1) and (group2), by a clustering method that collects samples whose values are close to the ECv matrix.

FIG. 14 (B) shows a survival time curve to which the survival time data included in the lung cancer data of the public database of cancer patients is applied to each of the two clusters. As shown in FIG. 14 (B), the results showed that the survival time of the patients belonging to one cluster was relatively long, and the survival time of the patients belonging to the other cluster was relatively short. That is, it was demonstrated that there is a large difference in prognosis (survival time) between the two clusters. In this way, the feature network enables the characterization and classification of individual data.

FIG. 15 is a schematic diagram showing another configuration of the ECv matrix. Each row shows the branch of the feature network (branch index), and each column shows the ECv of the sample (individual) for each subtype. Subtypes can be further subdivided for a particular cancer, such as the molecular subtype of gastric cancer. In the figure, subtypes T1, T2, and T3 are shown, but for example, CIN (Chromosomal Instability), MSI (Microsatellite Instability), EBV (Epstein Barr Virus), GS (Genomically Stable), etc. may be used. can. In the figure, the patterned portion represents a branch in which ΔECv is equal to or higher than the threshold value by combining two subtypes.

ΔECv for each subtype can be obtained, for example, as follows. That is, first, the gene network is estimated based on the publicly available gene expression data of gastric cancer patients. Next, the ECv of all branches is calculated for each sample. Then, for each of the four subtypes of gastric cancer defined in the literature (CIN, MSI, EBV, GS), the average value of ECv of each sample is calculated, and the difference between the subtypes is taken to obtain the two subtypes. The ΔECv between them can be calculated. Up to this point, the method is the same as the method for calculating ΔECv by comparing the sample that has been converted to EMT and the sample that has not been converted to EMT, that is, the comparison between the two groups. Gastric cancer data with four subtypes requires a feature network in other groups. This is possible, for example, by finding a branch whose ΔECv is greater than the threshold value with each of the other three subtypes for one subtype, and taking the union or intersection of the branches. This makes it possible to extract the feature network for each subtype. It is also possible to extract the feature network for each of the four subtypes by simply considering the other three subtypes as one large subtype for one subtype and comparing the two groups.

FIG. 16 is a schematic diagram showing a second example of characterization of an individual by a feature network. Similar to FIG. 14, FIG. 16 (A) is referred to as an ECv matrix, where each row shows a branch (branch index) of the feature network and each column shows the ECv of a sample (individual). Each element of the matrix represents an ECv at each branch of each sample. Each sample here uses gastric cancer patient sample data from a public database of cancer patients and includes four subtypes. That is, EBV is EB virus positive, MSI shows high frequency mutations in the microsatellite region, CIN shows abnormal somatic copy number, and GS shows other than that. FIG. 16 (A) can be classified into four categories by estimating the gene network using the data of the public database and clustering the ECv matrix of the upper branch, and roughly, four types of subs of the existing research. Indicates that it can be associated with a type. Further, as shown in FIG. 16 (B), it can be found that group 1 has a difference in survival time from other groups.

FIG. 17 is a schematic diagram showing a second example of the extracted feature network. In FIG. 17, of the four subtypes, the branch extracted by ΔECv with each of the other subtypes (CIN, GS, MSI) for EBV (the threshold value for extracting ΔECv is, for example, 0.5). It is a feature network extracted by taking the intersection (the common branch of the difference between the two groups) out of (which can be done).

FIG. 18 is a schematic diagram showing a third example of the extracted feature network. In FIG. 18, of the four subtypes, the ECv of the other three subtypes (CIN, GS, MSI) is extracted by ΔECv with the average with respect to EBV (the threshold value for extracting ΔECv is, for example, for example. It is a feature network extracted by (which can be 0.5).

In this way, it is not necessary to estimate the network for each subtype, and it is not necessary to compare the network structures for each subtype. ECv comparison allows the extraction of characteristic branches of subtypes from a single gene network without network structural comparison. Moreover, since it is not necessary to compare the structure of the network, it is possible to extract the characteristic branches of the subtype even when the number of samples of the specific subtype is small and the structure cannot be compared. As described above, it is possible to extract the difference in mechanism for each cancer subtype.

FIG. 19 is a schematic diagram showing a mechanism by which genes of the immune system can be captured by the extracted feature network. Due to EB virus infection such as Helicobacter pylori, cytokines (Cytokine) work via receptors to move the immune system. In this case, it was found that genes related to the immune system, which are thought to be driven by EB virus infection, are contained in the feature network extracted based on ECv. That is, it is suggested that the immune system is moved by viral infection, and the structural change of the signal transduction system obtained by structural comparison with a known molecular signature can be estimated by the feature network.

FIG. 20 is a schematic diagram showing a third example of characterization of an individual by a feature network. Similar to FIG. 14, FIG. 20 refers to the ECv matrix, where each row represents a branch (branch index) of the feature network and each column represents a sample (individual) ECv. Each element of the matrix represents an ECv at each branch of each sample. Each sample here uses data from patients with pancreatic cancer from TCGA (The Cancer Genome Atlas). From the samples of 153 patients with pancreatic cancer, 14 samples with a reliable prognosis and 14 samples with a bad prognosis are determined in advance. Branches with a large difference in the average value of ECv between the two groups with good and bad prognosis are extracted, and the ECv matrix of all 28 samples is clustered based on the value of ECv. In the figure, the darkly marked samples in each row are 14 samples with a good prognosis, and the brightly marked samples are 14 samples with a poor prognosis. The threshold value of ΔECv when extracting branches is 1.0.

FIG. 21 is a schematic diagram showing a fourth example of characterization of an individual by a feature network. In FIG. 21, the threshold value of ΔECv when extracting the branches is 0.75. Except for the threshold value, it is the same as in FIG. 20.

FIG. 22 is a schematic diagram showing a fifth example of characterization of an individual by a feature network. FIG. 22 shows the result of clustering by expanding from 28 samples to 153 samples. The threshold value of ΔECv when extracting branches is 0.75. As shown in FIGS. 20 to 22, it is shown that the group is roughly divided into a good group and a bad group.

FIG. 23 is a schematic diagram in which the extracted feature network is mapped to the entire network. In the figure, the darkly marked part indicates the feature network.

For gene network analysis, DEG (Differentially expressed genes), that is, a method for extracting genes having different expression differences is used. Hereinafter, the relationship between the method and the feature network according to the present embodiment will be described.

FIG. 24 is a schematic diagram showing a first example of the relationship between the extracted feature network and the DEG gene. In the figure, the extracted feature network is shown with the threshold value of ΔECv for extracting branches set to 1.0. There are 20 Top20DEG genes, which have a large expression difference between the two groups, good and bad (for example, the difference is 1 or more as a fold change). Of the 20 DEG genes, 5 are present (circled) if the distance from the feature network is within a predetermined value (for example, 1), and the 5 DEG genes are in the downstream direction of the feature network. You can see that.

FIG. 25 is a schematic diagram showing a second example of the relationship between the extracted feature network and the DEG gene. In the figure, the extracted feature network is shown with the threshold value of ΔECv for extracting branches set to 0.75. It can be seen that 13 of the 20 DEG genes are within a predetermined value (for example, 1) from the feature network, and 10 of these DEG genes are in the downstream direction of the feature network. From FIGS. 24 and 25, the genes obtained by the gene network analysis method for extracting genes with different expression are located downstream of the characteristic network, and the difference created from the difference in the characteristic network is the difference in the expression of each gene. It is considered that it can be estimated as.

FIG. 26 is a schematic diagram showing a fourth example of the extracted feature network. Although not shown, health survey data of residents in a region are used to define multiple critical diseases and estimate a single Bayesian network. FIG. 26 shows an ECv calculated as a branch feature amount using the data of the subject A and the subject B from the estimated network, and the feature network is extracted by extracting the branches whose ECv is equal to or higher than a predetermined threshold value. Is. Categories include, for example, age, gender, social background, lifestyle, health survey test values, genetic information, and the like. From FIG. 26, it can be seen what each of the two subjects has a disease and what is a common disease.

Next, a specific example of the usage pattern of the present invention will be described. Municipalities such as cities and companies belonging to health insurance associations carry out health examinations with the aim of maintaining the health of residents and employees and early detection of illnesses. As a result of such a health examination, a large number of health survey data can be collected. In addition, hospitals and clinics can also collect patient data when examining or treating patients. By using the feature network extraction method of the present invention, it is possible to evaluate the relationship between a large number of samples or sample groups such as residents, employees, and patients.

Below, the existing nodes make it easy to analyze the Bayesian network estimated from the medical examination data (data for 727 people for 4 years from 2014 to 2017) measured by Hirosaki COI (Center of Innovation). An example is shown in which contraction is performed, a feature network is extracted, and a causal relationship (related path) for each desired disease and each individual is extracted. When the estimated Bayesian network is a general discrete model, it can be applied to a continuous Bayesian network by performing preprocessing used in machine learning such as 1-hot conversion.

FIG. 27 is a schematic diagram showing an example in which a path related to the onset of chronic kidney disease (CKD) is extracted, and FIG. 28 is a schematic diagram showing an example in which a path related to the onset of hypertension is extracted. In FIGS. 27 and 28, in order to extract the related paths, the geometric mean upper path is used by utilizing the relative contribution rate RCr described above. When extracting related paths, only the paths leading from lifestyle-related diseases to specific diseases (chronic kidney disease and hypertension in the example of the figure) are extracted.

FIG. 29 is a schematic diagram showing an example of a network related to two diseases of CKD and hypertension with SNP, and FIG. 30 is a schematic diagram showing an example of a network related to two diseases of CKD and hypertension without SNP. FIG. 29 shows the intersection of both chronic kidney disease (CKD) and hypertension in the presence of SNPs, i.e. personal genome (gene) mutation data. FIG. 30 shows the intersection of both chronic kidney disease (CKD) and hypertension in the absence of SNPs. As shown in FIGS. 29 and 30, a common associated path for both chronic kidney disease (CKD) and hypertension becomes observable.

An example of an individual path extracted based on the relative contribution rate RCr for each individual on the chronic kidney disease (CKD) onset-related path illustrated in FIG. 27 will be described below.

FIG. 31 is a schematic diagram showing a first example of a personal network of chronic kidney disease (CKD) onset, and FIG. 32 is a schematic diagram showing a second example of a personal network of chronic kidney disease (CKD) onset, FIG. 33. Is a schematic diagram showing a third example of a personal network of chronic kidney disease (CKD) onset. The first example shown in FIG. 31 is a path for a woman in her 70s, and it can be seen that a drinking-related path and a stress / sleep-related path are effective from the viewpoint of developing chronic kidney disease. The second example shown in FIG. 32 is a path for a man in his fifties, and it can be seen that a path related to heart disease is effective from the viewpoint of developing chronic kidney disease. The third example shown in FIG. 33 is a path for a man in his 60s, and it can be seen that a diabetes-related path is effective from the viewpoint of developing chronic kidney disease. As shown in FIGS. 31 to 33, it is possible to clearly observe that the effective path differs for each individual.

FIG. 34 is a flowchart showing an example of the processing procedure of the feature network extraction device 50. For convenience, the subject of processing will be described below as the processor 51. The processor 51 acquires sample (individual) data (S11) and assigns the acquired data to a required node of the Bayesian network (S12).

The processor 51 starts calculating posterior probabilities of nodes other than the required node (S13), and calculates the feature amount of the link (branch or edge) (S14). The processor 51 determines the presence / absence of another sample (S15), and if there is another sample (YES in S15), the processor 51 continues the processing after step S11.

When there is no other sample (NO in S15), the processor 51 extracts the feature network based on the calculated feature amount (S16), and ends the process.

Next, the extraction of the feature network in step S16 described above will be described. FIG. 35 is a flowchart showing an example of the feature network extraction process. The processor 51 calculates the average of the feature amounts (for example, ECv) of each branch for each group (S161), and calculates ΔECv, which is the difference in ECv between the groups, for each branch (S162).

The processor 51 extracts a branch whose ΔECv is larger than the threshold value (S163). The processor 51 determines the presence or absence of another group (S164), and if there is another group (YES in S164), the union or intersection of the branches extracted from all the other groups for each group. (S165), and the process of step S166 described later is performed.

When there is no other group (NO in S164), the processor 51 constructs a feature network from the extracted branches (S166) and ends the process.

Features The network extraction device 50 can also be realized by using a computer equipped with a CPU (processor), RAM, and the like. A computer program (which can be recorded on the recording medium M) that defines the processing procedure as shown in FIGS. 34 and 35 is read by the recording medium reading unit 55 provided in the computer, the read computer program is loaded into the RAM, and the computer is loaded. By executing the program on the CPU (processor), the feature network extraction device 50 can be realized on the computer.

As described above, according to the present embodiment, the causal relationship of an individual or an individual sample that cannot be explained by the Bayesian network beyond the limit point of the Bayesian network that the characteristics (causal relationship) of the entire data can be explained. Can be explained by using the branch evaluation method of estimated Bayesian network and branch features.

In the present embodiment, the predetermined model used for the Bayesian network is not limited to the nonparametric regression model. For example, the predetermined model may be an addition model or a multiplication model. In the case of the additive model, there are some functions m1, m2, ... For the parent variables x1, x2, ..., And the child variables y = m1 (x1) + m2 (x2) + ... are represented by "sum". Can be done. The functions m1 (x1), m2 (x2), ... May be required functions, and the values of the functions m1 (x1), m2 (x2), ... Can be ECv. Further, if m1 (x) = x, the predetermined function becomes a linear function and can be a linear model. Further, in the case of a multiplication model, it can be expressed by "multiplication" such as child variables y = m1 (x1), m2 (x2), .... The predetermined function is not limited to the non-linear function, but may be a linear function.

In this embodiment, the Bayesian network can also be applied to a discrete model. When the Bayesian network is a discrete model, it can be applied to a continuous model by performing a general preprocessing performed by machine learning called 1-hot conversion. For example, when the variable X takes A, B, and C, the 1-hot conversion is divided into three variables, "X is A", "X is B", and "X is C". , If applicable, 1 can be used as the value of each variable, otherwise 0 can be used as the value of each variable. Also, "X is C" can be expressed if both "X is A" and "X is B" are 0, so variables in N categories are converted to 1-hot. May be performed with the variable of N-1.

The feature network of this embodiment is not limited to application to a medical-related Bayesian network. For example, the feature network of this embodiment can be applied to advertisement provision using a Bayesian network, marketing research, questionnaire analysis, application to system failure diagnosis, and the like. For example, analysis using a conventional Bayesian network can explain the causal relationship of rough attribute data such as the age and gender of a user, but cannot explain the causal relationship of an individual or an individual sample. By applying this embodiment, it is possible to use a branch evaluation method of estimated Bayesian network and branch features, and it is possible to explain the causal relationship between individual or individual samples, and to user modeling and human modeling. When applied, it will be possible to analyze in detail down to the individual level.

In the Bayesian network analysis method of the present embodiment, a feature network is extracted from a required Bayesian network using the above-mentioned Bayesian network extraction device, and a sample or a sample group in the Bayesian network is obtained based on the extracted feature network. Can be evaluated. In this case, the required Bayesian network can represent a multivariate causal relationship for at least one of the medical data, advertising data, marketing data and survey data, while the multivariate causal relationship for the other data. It may represent.

50 Features Network extraction device 51 Processor 52 Operation unit 53 Interface unit 54 Display panel 55 Recording medium reader 56 ROM
57 Memory 58 Storage 581 Bayesian network model 582 Sample data

Claims (13)

A data addition unit that assigns data to the required nodes of a Bayesian network that expresses the dependency relationships between multiple nodes to which each random variable is associated using a non-circulated directed graph, and
When calculating the posterior probability of a node based on the data given by the data addition unit, the child from the parent node is based on a predetermined model that constitutes the conditional probability when the random variable of the parent node is given. A calculation unit that calculates the feature amount of the link to the node,
A feature network extraction device including an extraction unit that extracts a feature network from the Bayesian network based on a feature amount calculated by the calculation unit. The calculation unit
The feature network according to claim 1, wherein the feature amount of the link from the parent node to the child node is calculated based on the function value of a predetermined function representing a predetermined model of the random variable of the parent node with respect to the random variable of the child node. Extractor. The calculation unit
The feature network extraction device according to claim 2, wherein the function value of the predetermined function is calculated as the feature amount of the link. The calculation unit
When the data addition unit assigns data of different samples, a comparison value between the first function value of the predetermined function based on the data of the first sample and the second function value of the predetermined function based on the data of the second sample. The feature network extraction device according to claim 2, wherein the feature amount of the link is calculated. The calculation unit
The feature of the link is the ratio of the function value of the predetermined function corresponding to the link to the maximum value of the function values of each predetermined function representing the predetermined model of the random variable of each of the plurality of parent nodes with respect to the random variable of the child node. The feature network extraction device according to claim 2, which is calculated as a quantity. The calculation unit
The ratio of the function value of the predetermined function corresponding to the link to the total value of the function values of each predetermined function representing the predetermined model of the random variable of each of the plurality of parent nodes with respect to the random variable of the child node is used as the feature quantity of the link. The feature network extraction device according to claim 2 for calculation. The extraction unit
The feature network extraction device according to any one of claims 2 to 6, wherein when the feature amount of the link calculated by the calculation unit is equal to or more than a predetermined threshold value, the feature network including the link is extracted. A setting unit for setting a plurality of required nodes of the Bayesian network is provided.
The calculation unit
The feature amount of the entire one or a plurality of links constituting each of the plurality of paths from one node to the other node set in the setting unit is calculated.
The extraction unit
The feature network extraction device according to any one of claims 2 to 6, wherein a feature network including a path in which the feature amount of the entire link constituting the path is equal to or more than a predetermined threshold value is extracted from the plurality of paths. .. The predetermined model includes a nonparametric regression model.
The feature network extraction device according to any one of claims 2 to 8, wherein the predetermined function includes a non-linear function. The feature network is extracted from the required Bayesian network by using the feature network extraction device according to any one of claims 1 to 9.
Evaluate a sample or sample group in the Bayesian network based on the extracted feature network.
Bayesian network analysis method. The Bayesian network analysis method according to claim 10, wherein the Bayesian network represents a multivariate causal relationship with respect to at least one of medical data, advertising data, marketing data, and questionnaire data. On the computer
The process of assigning data to the required nodes of the Bayesian network, which expresses the dependency relationships between multiple nodes to which each random variable is associated using a non-circulated directed graph, and
When calculating the posterior probability of a node based on the given data, the link from the parent node to the child node is based on a predetermined model that constitutes the conditional probability given the random variable of the parent node. Processing to calculate the feature amount and
A computer program that executes a process of extracting a feature network from the Bayesian network based on the calculated feature amount. It is a feature network extraction method by computer.
Computer
Data is added to the required nodes of the Bayesian network in which the dependency relationships between multiple nodes to which each random variable is associated are represented using an uncirculated directed graph.
When calculating the posterior probability of a node based on the given data, the link from the parent node to the child node is based on a predetermined model that constitutes the conditional probability given the random variable of the parent node. Calculate the feature amount,
A feature network extraction method for extracting a feature network from the Bayesian network based on the calculated feature quantity.

Images