KR101927910B1 - System and method for predicting disease inforamtion using deep neural network - Google Patents

Abstract

The system for predicting disease information includes an expression-altering gene extractor for generating expression-altering gene information for each of a plurality of gene data sets based on gene-based expression distribution, an initial depth-based neural network based on a gene, cell function, A learning unit for generating the final neural network by learning the initial neural network based on the expression change gene information for each of the plurality of gene data sets; and a backward propagation unit for backward propagating the final neural network, And a search unit for searching for disease information related to the disease.

Description

TECHNICAL FIELD [0001] The present invention relates to a system and method for predicting disease information based on a neural network,

The present invention relates to a disease information prediction system and method.

In conventional disease studies, disease gene candidates were first selected, and disease genes were identified by comparing how the effects of changes in inhibition / activity or sequence variation of the candidate gene are similar to known disease patterns. However, these traditional disease studies have difficulties in predicting the candidate to be influenced, have to know the disease pattern in advance, and are difficult to interpret the disease mechanism.

In recent years, bioinformatics has been used to search for genes whose expression is significantly changed in patients, or genes that relate to previously discovered disease genes based on similarity and expression similarity of cellular functions, Large - scale disease gene discovery techniques are being studied. However, this method also does not provide information on the mechanism of disease induction, so it is not easy to interpret, and there are no clear criteria to detect candidates.

A problem to be solved by the present invention is to provide a system and method for learning in-depth neural networks using disease expression-altering genes and predicting disease information such as disease genes and disease-related functions based on in-depth neural networks.

The system for predicting disease information according to one embodiment of the present invention includes an expression change gene extractor for generating expression change gene information for each of a plurality of gene data sets based on an expression distribution by gene, A learning unit for generating an initial depth neural network based on an association relationship and learning the initial depth neural network based on expression change gene information for each of the plurality of gene data sets to generate a final depth neural network, And a search unit for searching for disease information related to the disease by backward propagation.

The expression-altering gene information may include a value indicative of whether the gene is an expression-altering gene for each gene, and the expression-altering gene may be a gene that exhibits a difference in expression value above a reference level relative to the expression distribution of the gene.

The extracting unit may generate expression change gene information for each of the plurality of gene data sets by comparing an expression value of each gene set of each gene data set with an expression distribution by the gene.

The plurality of gene data sets may include normal gene data sets extracted from a normal population and diseased gene data sets extracted from a patient population, and the extracting unit may calculate the gene expression distribution based on the normal gene data sets.

Wherein the learning unit generates the initial in-depth neural network composed of an input layer including a plurality of gene nodes, at least one hidden layer including a plurality of cell function nodes, and an output layer including a disease node, Determines an initial connection strength between a node of the input layer and a node of the hidden layer based on the data and determines an initial connection strength between the node of the hidden layer and the node of the output layer based on the relationship data between the cell function and the disease .

Wherein the learning unit inputs learning data to an input layer of the initial neural network to perform unsupervised learning on the initial neural network and generates an intermediate neural network in which the initial connection strength is updated by completing non- And the learning data may be expression change gene information for each of the plurality of gene data sets.

Wherein the learning unit learns the current layer by propagating and propagating in the current layer and then learning the next layer of the current layer when the learning of the current layer is completed using the Deep Belief Network method, You can update the initial connection strength.

Wherein the learning unit sets the learning data in the input layer of the intermediate neural network and sets an output value assigned to the type of learning data input to the input layer in the output layer of the intermediate neural network, And the final depth neural network can be generated by supervised learning of the intermediate neural network.

Wherein the type of the learning data includes any one of a first type indicating a normal gene data set extracted from a normal population and a second type indicating a disease gene data set extracted from the patient population, Learning data is input to the input layer of the intermediate neural network, a first output value is set to the output layer of the intermediate neural network to learn the intermediate neural network, and learning data corresponding to the second class is transmitted to the intermediate neural network The intermediate depth neural network can be learned by setting a second output value in the output layer of the intermediate neural network.

The learning unit compares the predicted value obtained by propagation calculation in the forward direction in the input layer of the intermediate neural network with the output value set in the output layer of the intermediate neural network to obtain an error and then back propagates in a direction of minimizing the error, Each connection strength of the neural network can be updated.

The searching unit may set at least one cell function associated with the disease or at least one gene associated with the disease by setting the output layer of the final neural network as an output value corresponding to the disease and back propagating.

Wherein the searcher sets the output layer of the final neural network to an output value corresponding to a non-disease and back propagates to search for at least one gene associated with non-disease associated with at least one cell function or non-disease, At least one cellular function associated with the disease and at least one cell function associated with the non-disease by extracting at least one cell function associated with the disease with a disease-specific cellular function, One gene may be compared to extract at least one gene associated only with the disease as a disease-specific gene.

A method for predicting disease information in a system according to another embodiment of the present invention is a method for predicting disease information in a system, comprising: comparing a gene expression value of each of a plurality of samples including disease samples with a normal expression distribution by gene; , A gene for a disease expression change in a corresponding sample, a step for extracting a gene having a significant change in a disease expression expression gene of the sample, Deep Neural Network), and searching for a cell function or gene related to the disease based on the learned neural network, wherein the sample is a gene data set including a gene expression value, and the sample The disease sample is classified into the disease sample and the normal sample, and the disease sample is classified into the patient of the target disease A gene derived data set, the normal sample is a set of data extracted from the normal gene.

The disease information prediction method may further include analyzing a normal expression distribution of each gene based on expression values of a plurality of normal samples.

The step of analyzing the normal expression distribution by gene may include extracting the plurality of normal samples from the microarray database, normalizing the expression values of the genes of the plurality of normal samples, And calculating a normal expression distribution by the gene.

Wherein the step of learning the depth neural network comprises the steps of: learning nonvisuality of the neural network on the basis of disease expression change gene information of each of the plurality of samples; and analyzing disease expression change gene information of each of the plurality of samples, Based neural network can be learned and learned.

A method for predicting disease information of a system according to another embodiment of the present invention is a method for predicting disease information of a system, comprising the steps of generating a first neural network in which an initial connection strength is set based on a relationship between genes, cell functions, and diseases, Generating second deep-network neural networks by unsupervised learning of the first neural network based on the learning data, generating second deep-layer neural networks based on the learning data, Supervised learning of the neural network to generate the third in-depth neural network. And searching for disease information related to the disease by backward propagating the third neural network, wherein the first neural network comprises an input layer including a plurality of gene nodes, a plurality of cell function nodes And at least one hidden layer including the node, and an output layer including the disease node.

Wherein the step of generating the first neural network comprises determining an initial connection strength between a node of the input layer and a node of the hidden layer based on the relationship data between the gene and the cell function, The initial connection strength between the node of the hidden layer and the node of the output layer can be determined.

The expression change gene information for the disease includes a value indicating whether the gene is an expression change gene for each gene, and the expression change gene may be a gene showing a difference in expression value above the reference gene expression distribution.

Wherein the step of searching for disease information related to the disease comprises the steps of setting an output value of the third neural network to a value corresponding to the disease, Extracting a node of the node, and outputting the gene or cell function corresponding to the extracted node as the disease information.

According to the embodiment of the present invention, the disease gene can be searched based on the disease expression-altering gene extracted through the microarray data analysis, so that the disease sorter capable of expressing various disease states in actual data can be constructed. According to the embodiment of the present invention, it is possible to make a disease classifier having a high predictive power by applying a depth neural network, which is a method not used in disease gene search, to generation of a disease classifier and applying various neural network learning techniques.

According to the embodiment of the present invention, it is possible to directly track the linkage with the disease mechanism which has not been solved in the conventional disease gene research, and thus provides the disease gene as well as the disease-related cell function, . According to the embodiment of the present invention, it is possible to discriminate a disease of a patient and to identify a disease gene and a disease-related cell function to select a drug target.

According to embodiments of the present invention, novel disease genes and novel disease-related functions can be discovered. According to the embodiment of the present invention, it is possible to discover a new drug target that can effectively control the disease mechanism, and thereby, the therapeutic drug development can be effectively performed. According to an embodiment of the present invention, a pathogenesis mechanism can be classified using a disease gene combination uncovered through back propagation, and a new accompanying diagnosis method can be developed.

1 is a block diagram of a disease information prediction system according to an embodiment of the present invention.
2 is a diagram illustrating a depth-of-field network according to an embodiment of the present invention.
FIG. 3 is a flowchart of a disease expression change gene extraction method according to an embodiment of the present invention.
FIG. 4 is a diagram for explaining a disease expression-altering gene extraction method according to an embodiment of the present invention.
5 is a flowchart illustrating a method of learning a neural network according to an embodiment of the present invention.
6 is a view for explaining a depth learning method according to an embodiment of the present invention.
7 and 8 are flowcharts illustrating a disease information search method according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when an element is referred to as " comprising ", it means that it can include other elements as well, without excluding other elements unless specifically stated otherwise. Also, the terms " part, " " module, " and " module ", etc. in the specification mean a unit for processing at least one function or operation and may be implemented by hardware or software or a combination of hardware and software have.

FIG. 1 is a block diagram of a disease information prediction system according to an embodiment of the present invention, and FIG. 2 is a diagram illustrating a depth neural network according to an embodiment of the present invention.

Referring to FIG. 1, a disease information prediction system (hereinafter, referred to as a "prediction system") 100 generates a gene, a cell function, and a disease relationship as a Deep Neural Network. The prediction system 100 then backwards propagates the neural network to search for disease-related Disease genes or Disease-related functions. The prediction system 100 can then use the in-depth neural network to predict disease associated with a particular gene combination. Here, the in-depth neural network can be called a disease classifier.

The prediction system 100 includes a disease expression change gene extracting unit 110, a neural network learning unit 130, and a disease information searching unit 150. The prediction system 100 may further include a database unit 170 storing various information used by the disease expression change gene extracting unit 110 and the depth learning unit 130. [ The database unit 170 stores various information used in the disease expression change gene extracting unit 110 and the deep neural network learning unit 130 in the database unit 170. The database unit 170 is included in the prediction system 100 But may be implemented as a separate device or stored in a remote server. In this case, the prediction system 100 may connect to the database / server where necessary information is stored to obtain necessary information. Or the prediction system 100 may receive various kinds of information used in the disease expression change gene extracting unit 110 and the depth learning neural network learning unit 130 and may analyze based on the input information.

The disease expression change gene extracting unit 110 extracts an expression change gene (referred to as a " disease expression change gene ") for a disease. The disease expression change gene is input to the neural network learning unit 130, and the neural network learning unit 130 learns the neural network for the disease using the disease expression change gene. The disease expression change gene extracting unit 110 can extract expression change genes for various diseases.

The disease expression change gene extracting unit 110 extracts a disease-related expression change gene based on the gene expression distribution of the microarray data stored in the microarray database 171. The disease expression change gene selection unit 112 compares the normal expression distribution obtained from the normal samples in the microarray data with the gene value of the disease sample and indicates that the expression is significantly increased in the disease sample as compared with the normal expression distribution And extracts the decreasing gene as an expression-altering gene of the disease sample. At this time, the disease expression change gene selection unit 112 can extract not only the disease sample but also the normal expression gene of the normal sample by comparing the normal expression distribution by gene. Here, the microarray is a device for arranging gene fragments on a chip and analyzing gene expression extensively. The microarray database may be, for example, a database of the Gene Expression Omnibus. Wherein the sample is a gene data set comprising gene expression information, the normal sample is a gene data set extracted from a normal person, and the disease sample refers to a gene data set extracted from a patient of the disease.

The neural network learning unit 130 generates an initial neural network in which initial weights are input based on cell function information and disease gene information. Then, the depth neural network learning unit 130 learns the initial depth neural network based on the expression change gene information input from the disease expression change gene selection unit 112 to generate the final depth neural network. At this time, the neural network learning unit 130 generates the final neural network by performing unsupervised learning and supervised learning on the initial neural network based on the expression change gene information.

Referring to FIG. 2, the depth-based neural network is a multi-layer network composed of an input layer, a hidden layer, and an output layer. Each of the input layer, the hidden layer, and the output layer is composed of a plurality of nodes. In the neural network, adjacent nodes of two layers are connected by a defined line of weight (weight). The value of each node of the hidden layer and the output layer can be determined by a value obtained by multiplying the value of the node of the previous layer connected to the node by the connection strength (weight value) of the line connected to the node to the nonlinear function. The neural network computes the error by comparing the predicted value obtained from the forward propagation calculation with the result to be learned, and updates the connection strength by back propagation in the direction of minimizing the error.

The neural network learning unit 130 generates a neural network by mapping a gene to a node of the input layer, mapping a cell function to a node of the hidden layer, and mapping the disease to a node of the output layer.

In a neural network, adjacent nodes of two layers are connected based on the relationship between nodes. The gene nodes of the input layer and the cell function nodes of the hidden layer are linked based on the relationship between gene and cell function. The neural network learning unit 130 extracts the association between the gene node of the input layer and the cell function node of the hidden layer based on the cell function database 173 storing the relation data between the gene and the cell function, . The cell function database may be, for example, a database of Gene Ontology.

Cell function nodes in the disease nodes and hidden layers of the output layer are linked based on the relationship between disease and cell function. When cell functions can be classified into a plurality of hierarchies, a plurality of hidden layers are generated hierarchically, and nodes of the hidden layer are connected based on the relationship between cell functions. The neural network learning unit 130 extracts the association between the cell function of the hidden layer and the disease of the output layer based on the cell function database 173 and the disease gene database 175 to determine the initial connection strength. The disease gene database can be, for example, a database of Online Mendelian Inheritance in Man.

The neural network learning unit 130 generates an initial neural network by connecting nodes of an input layer, a hidden layer, and an output layer with initial connection strengths.

The depth-of-neural network learning unit 130 pre-trains the initial depth-based neural network based on the disease samples and the expression change genes of the normal samples. The neural network learning unit 130 can perform the preliminary learning using the Deep Belief Network method. The depth neural network learning unit 130 may use the information on the expression change gene as the input value of the input layer and update the initial connection strength of the initial depth neural network through non-learning of the non-background. If the learning data is insufficient, the initial connection strength is adjusted by pre-learning by the non-bipartite learning method to reduce the influence of the initial weight (for example, 0 or 1) arbitrarily set on the final learning result.

The depth-of-neural network learning unit 130 generates a final depth-of-field neural network by learning the pre-learned depth-based neural network based on the disease samples and the expression change genes of the normal samples. The neural network learning unit 130 can adjust the connection strength of the neural network that has been previously learned through dropout and error back propagation in the pre-learned neural network.

The disease information search unit 150 searches back the disease-related cell function or disease gene by back propagating the learned final in-depth neural network. The disease information search unit 150 finds a cell function / gene that is activated through reverse propagation in the disease / normal output of the final depth neural network, and generates a disease-related function combination model / disease gene combination model through statistical analysis.

Specifically, the disease information search unit 150 extracts cell functions related to the disease by performing back propagation in the output layer node (value = 1) of the final depth neural network, and combines disease-related cell functions to form a disease- You can create a model. In addition, the disease information search unit 150 extracts cell functions related to non-disease (normal) by performing back propagation at the output layer node (value = 0) of the final depth neural network, and combines them to generate a normal function combination model have. The disease information search unit 150 can generate a disease-specific functional combination model by statistically analyzing a disease-associated function combination model and a normal function combination model. Here, the disease-specific function means a cell function that appears only in a disease-associated function combination model, and a disease-specific function combination model including only a disease-specific function can be provided as a disease-specific functional combination model.

The disease information search unit 150 performs back propagation at the output layer node (value = 1) of the final depth neural network to extract genes associated with the disease, and combines the genes associated with the disease to generate a disease-associated gene combination model. In addition, the disease information search unit 150 extracts genes associated with non-disease (normal) by performing back propagation at the output layer node (value = 0) of the learned final NN and combines them to generate a normal gene combination model . The disease information search unit 150 may generate a disease-specific gene combination model by statistically analyzing a disease-associated gene combination model and a normal gene combination model.

FIG. 3 is a flowchart illustrating a method for extracting a disease expression-altering gene according to an embodiment of the present invention, and FIG. 4 is a diagram for explaining a disease expression-altering gene extraction method according to an embodiment of the present invention.

Referring to FIG. 3, the disease expression change gene selection unit 112 fetches microarray data including a normal sample and a disease sample in the microarray database 171 (S110).

The disease expression change gene selection unit 112 analyzes the normal expression distribution by gene using normal samples (S120). The disease expression change gene selection unit 112 can normalize microarray data using Quantile Normalization. As shown in Table 1, the disease expression alteration gene selection unit 112 determines a gene expression level of a disease sample (Cancer-1, Cancer-2, etc.) related to a specific normal (Normal-1 to Normal- Normalizes the expression value. Expression changes according to normal expression distribution by gene An example of gene judgment is shown in Fig.

Sample gene Normal-1 ... Normal-N Cancer-1 Cancer-2 ... Gene 1 0.5 ... 2 One ... ... Gene 2 One ... 4 3 ... ... Gene 3 3 ... 8 5 ... ... Gene 4 2 ... 3 4 ... ... Gene 5 One ... 9 10 ... ... ... ... ... ... ... ... ...

The disease expression altering gene selection section 112 compares the normal expression distribution of each gene with the gene expression level of each sample and determines a gene whose expression is increased or decreased compared with the normal sample based on the normal expression distribution by gene, And extracted as an expression-altering gene (S130). Expression-altering genes are extracted on a sample-by-sample basis, and the sample types can be divided into normal and diseased samples. As shown in Table 2, the disease expression alteration gene selection unit 112 displays a value (hereinafter referred to as " expression change indicator ") indicating whether the expression gene is a gene for each sample. Referring to Table 2, when the expression change indicator is " 1 ", it means an expression change gene associated with a specific disease.

Sample gene Normal-1 ... Normal-N Cancer-1 Cancer-2 ... Gene 1 0 ... 0 0 ... ... Gene 2 0 ... One One ... ... Gene 3 0 ... 0 0 ... ... Gene 4 0 ... 0 One ... ... Gene 5 0 ... 0 0 ... ... ... ... ... ... ... ... ...

4, the disease expression-altering gene selection unit 112 compares an expression distribution of each gene with a gene expression value of a disease sample (Cancer-1), and determines a gene having a significant change (for example, Gene 2, and Gene 4) as expression-altering genes. For example, the disease expression altering gene selection unit 112 can obtain the mean and standard deviation of each gene in a normal sample, and search for a distribution assuming a normal distribution. For example, a gene having an expression value of less than -6.45 or greater than 26.45 in an average of 10 and a standard deviation of 10 can be determined as an expression-altering gene. FIG. 5 shows a method of learning a neural network according to an embodiment of the present invention. 6 is a diagram for explaining a depth learning method according to an embodiment of the present invention.

Referring to FIG. 5, the SNN learning unit 130 generates an initial SNN in which genes, cell functions, and diseases are arranged at nodes of a corresponding layer, and nodes between layers are connected at an initial connection strength (S210). The initial depth neural network consists of an input layer containing a plurality of genetic nodes, at least one hidden layer containing a plurality of cell functional nodes, and an output layer corresponding to a disease node. The input layer has as many nodes as the number of genes, the hidden layer has as many nodes as the number of cell functions, and the output layer can be composed of a single node (disease). In the initial depth neural network, all the nodes between adjacent layers are connected, and a nonlinear function (for example, a rectified linear unit) can be used as an activation function in the node. The neural network learning unit 130 inputs an initial connection strength for a connection between nodes based on the association of a specific gene, a specific cell function, and a disease. The initial connection strength can be set to any value, for example, 0 or 1. For example, if the relationship between the gene of a specific node and the cell function of a specific node is stored in the cell function database 173, the depth neural network learning unit 130 inputs the connection strength between the two nodes as " 1 " , And if the relationship between the gene of a specific node and the cell function of a specific node is not stored in the cell function database 173, the connection strength between these two nodes can be input as " 0 ". The neural network learning unit 130 extracts the association between the cell function and the disease based on the cell function database 173 and the disease gene database 175. If the cell function of the specific node is related to the disease of the output layer node (for example, the significance value is 0.05 or less ) , the depth neural network learning unit 130 sets the connection strength of the cell function node and the disease node to " 1 " If there is no association, the connection strength between the cell function node and the disease node can be input as " 0 ". In this case, the neural network learning unit 130 extracts a gene associated with the cell function 1 and a gene associated with the disease 1, and can calculate the significance value through Fischer's accuracy test on the number of common genes relative to the total number of genes.

The depth neural network learning unit 130 pre-learns the initial depth neural network based on the expression change gene (S220). The depth neural network learning unit 130 inputs the expression change genes for each sample extracted from the disease expression change gene selection unit 112 and pre-learns the initial depth neural network by the non-background learning method through radio wave / back propagation. Here, since the initial connection strength of initial neural network is input as "0" or "1", the initial connection strength is adjusted by using the expression change gene as learning data. The neural network learning unit 130 performs the pre-learning using the deep trust neural network method. 6, when the neural network learning unit 130 uses the disease sample of Table 2 as the learning data, the neural network learning unit 130 inputs the node value corresponding to the expression change gene (Gene 2, Gene 4, etc.) as 1, The node value corresponding to the gene (Gene 1, Gene 3, Gene 5, etc.) is input as zero. The neural network learning unit 130 updates the initial connection strength while propagating / back propagating in each layer using the deep trust neural network method. Here, the deep trust neural network method is a learning method for each layer by a non-background learning method. When a layer of learning ends, the connection strength of the layer is updated and the learning of the next layer of the layer after the learning is progressed step by step. The in-depth trust neural network method can be performed using known methods. For example, in the deep trust neural network method, an input having a node value corresponding to the expression change gene as "1" and a node value corresponding to the remaining gene as "0" is generated for each layer using a dropout and error back propagation method Learn.

The neural network learning unit 130 generates a final neural network by learning the neural network that has been previously learned based on the expression change gene (S230). Here, the depth neural network learning unit 130 inputs a node value corresponding to an expression change gene of each sample as "1" and a node value corresponding to the remaining genes as "0". The neural network learning unit 130 sets the node value of the output layer to "1" or "0" based on the sample type, and performs dropout and error back propagation using the node values of the input layer and the output layer, Learning the deepened neural network. For example, in case of learning with a disease sample, the neural network learning unit 130 sets the node value of the output layer to "1", and uses the node values of the input layer and the output layer to perform the dropout and error back propagation, It learns the neural network. In the case of learning with a normal sample, the neural network learning unit 130 sets the node value of the output layer to " 0 ", and learns the in-depth neural network through dropout and error back propagation using the node values of the input layer and the output layer.

In this way, the deep-network neural network learning unit 130 adjusts (learns) the connection strength between nodes of the input layer, the hidden layer, and the output layer through repetitive propagation and back propagation to generate a final depth-based neural network. The neural network learning unit 130 receives a normal sample and a disease sample as training samples, calculates an error between the resultant value output through the propagation of the training sample and the discrimination value of the actual training sample, and then reduces the error The classifier is learned by adjusting the connection strength between nodes.

In the learning process, each node of the hidden layer is known to represent the characteristics of input information, ranging from simple features to complex features. The prediction system 100 can generate and track the cell functions of the hidden layer based on the expression change genes generated based on the microarray data of the normal group and the patient group. In particular, the prediction system 100 can track the nodes of the hidden layer to identify complex disease characteristics such as cell functions associated with the disease. Specifically, since the prediction system 100 learns the in-depth neural network so as to classify diseases from genes, diseases corresponding to arbitrary combinations of genes can be identified. In addition, each node of the hidden neural network hidden layer corresponds to the cell function. Cell function can be defined as a combination of specific expression change genes based on the linkage between the node of the hidden layer and the gene node of the input layer. Thus, the prediction system 100 can track activated cell functions corresponding to any combination of genes, and can also track cell functions associated with the disease.

7 and 8 are flowcharts illustrating a disease information search method according to an embodiment of the present invention.

Referring to FIG. 7, the disease information search unit 150 is requested to search for cell functions related to diseases (S310).

The disease information search unit 150 extracts a cell function that is activated by reverse propagation of the neural network whose output is set as a disease, as a disease-related cell function (S320). The in - depth neural network is a multi - layer network generated by learning the linkage of genes, cell functions, and diseases based on disease expression change genes. The disease information search unit 150 sets the output layer node of the neural network to " 1 ", and searches for the activated node of the layer (hidden layer) corresponding to the cell function by back propagating from the output layer node. For example, the disease information search unit 150 may track the node values of each node in the hidden layer, and may reverse-track the genes constituting the node whose node value is equal to or greater than the reference value. Associated cell function can be predicted by cell function associated with the disease.

The disease information search unit 150 combines the extracted disease-related cell functions to generate a disease-related function combination model (S330).

The disease information search unit 150 extracts a cell function that is activated by back propagation of a neural network with normal output (non-disease) as a normal cell function (S340). The disease information searching unit 150 sets the output layer node of the neural network to " 0 ", and searches for the activated node of the layer (hidden layer) corresponding to the cell function by back propagating from the output layer node.

The disease information search unit 150 combines the extracted normal cell functions to generate a normal function combination model (S350).

The disease information search unit 150 extracts cell functions included only in the disease-associated function combination model based on the disease-associated function combination model and the normal function combination model to generate a disease-specific functional combination model (S360).

Although it is described that each step progresses sequentially, it can be processed in parallel, and the order of the steps can be changed.

Referring to FIG. 8, the disease information search unit 150 is requested to search for a gene associated with the disease (S410).

The disease information search unit 150 extracts a gene that is activated by back propagation of the neural network whose output is set as disease (S420). The disease information searching unit 150 sets the output layer node of the depth neural network to " 1 ", and searches for an activated node of the layer (input layer) corresponding to the gene by back propagating from the output layer node. For example, the disease information search unit 150 may track the node value of each node of the input layer, and may predict a gene corresponding to a node whose node value is equal to or greater than a reference value as a disease gene.

The disease information search unit 150 combines the extracted disease genes to generate a disease gene combination model (S430).

The disease information search unit 150 extracts a gene that is activated by back propagation of a neuron network whose output is set to normal (non-diseased) as a normal gene (S440). The disease information search unit 150 sets the output layer node of the neural network to " 0 ", and searches for the activated node of the layer (input layer) corresponding to the gene by back propagating from the output layer node.

The disease information search unit 150 combines the extracted normal genes to generate a normal gene combination model (S450).

The disease information search unit 150 extracts genes included only in the disease gene combination model based on the disease gene combination model and the normal gene combination model to generate a disease-specific gene combination model (S460).

Although it is described that each step progresses sequentially, it can be processed in parallel, and the order of the steps can be changed.

In this manner, the disease information search unit 150 can search disease-related cell functions that can propagate the target disease to the output, or search disease genes that can propagate the target disease to the output, to the learned neural network . Therefore, the disease information search unit 150 can infer and interpret disease pathway based on disease-related cell functions and disease genes.

According to the embodiment of the present invention, the disease gene can be searched based on the disease expression-altering gene extracted through the microarray data analysis, so that the disease sorter capable of expressing various disease states in actual data can be constructed. According to the embodiment of the present invention, it is possible to make a disease classifier having a high prediction ability by applying a neural network, which has not been used for disease gene search, to the generation of a disease classifier and applying various neural network learning techniques.

According to the embodiment of the present invention, it is possible to directly track the linkage with the disease mechanism which has not been solved in the conventional disease gene research, and thus provides the disease gene as well as the disease-related cell function, . According to the embodiment of the present invention, it is possible to discriminate a patient's disease, and to identify a disease gene and a disease-related cell function to select a drug target.

According to embodiments of the present invention, novel disease genes and novel disease-related functions can be discovered. According to the embodiment of the present invention, it is possible to discover a new drug target that can effectively control the disease mechanism, and thereby, the therapeutic drug development can be effectively performed. According to an embodiment of the present invention, a pathogenesis mechanism can be classified using a disease gene combination uncovered through back propagation, and a new accompanying diagnosis method can be developed.

The embodiments of the present invention described above are not implemented only by the apparatus and method, but may be implemented through a program for realizing the function corresponding to the configuration of the embodiment of the present invention or a recording medium on which the program is recorded.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, It belongs to the scope of right.

Claims (10)

As a disease information prediction system,
The expression level of each gene contained in each of the plurality of gene data sets is compared based on the expression distribution of each gene and a gene showing a difference in expression value above the reference gene expression level is determined as an expression change gene, An expression-altering gene extracting unit for generating expression-altering gene information indicating the expression-altering gene for each gene contained in the set,
Based on the association of genes, cell functions, and diseases, an initial neural network consisting of an input layer containing a plurality of genetic nodes, at least one hidden layer containing a plurality of cell functional nodes, and an output layer comprising disease nodes is generated A learning unit for generating a final depth neural network by performing non-geographic learning and guidance learning on the initial depth-of-field neural network based on the expression change gene information for each of the plurality of gene data sets; and
Wherein the output layer of the final depth neural network is set to an output value corresponding to a specific disease or steady state and is reverse-biased, and cell functions or at least one cell function activated in the final neural network, Or a search unit for extracting a gene,
The learning unit
Determining an initial connection strength between a node of the input layer and a node of the hidden layer on the basis of relational data between the gene and the cell function and determining an initial connection strength between the node of the hidden layer and the node of the output layer Lt; RTI ID = 0.0 > a < / RTI >
Wherein the mobile station is configured to learn the current layer by propagating and propagating back and forth between adjacent layers sequentially from the input layer and updating the initial connection strength by sequentially learning the next layer of the current layer when learning of the current layer is completed, Prediction system. The method of claim 1,
The search unit
Related cell function combination model by combining the cell functions searched in the specific disease state and combining the searched cell functions in the normal state to generate a non-disease related cell function combination model,
Wherein said disease-related cell function combination model and said non-disease-associated cell function combination model are compared with cell functions to extract at least one cell function associated with said specific disease as disease-specific cell function. The method of claim 1,
The search unit
A disease-associated gene combination model is generated by combining the genes searched in the specific disease state, a non-disease-associated gene combination model is generated by combining the genes found in the normal state,
Wherein the disease-associated gene combination model is compared with genes included in the non-disease-associated gene combination model to extract at least one gene associated with the specific disease as a disease-specific gene. The method of claim 1,
Wherein the plurality of gene data sets comprises normal gene data sets extracted from a normal population and diseased gene data sets extracted from a patient population,
The expression change gene extracting unit
And calculating an expression distribution of each gene based on the normal gene data sets. delete The method of claim 1,
The learning unit
Learning data is input to the input layer of the initial depth-of-field network to perform unsupervised learning of the initial depth-of-neural network, and an intermediate depth-of-neural network in which the initial connection strength is updated is completed by completing non-
Wherein the learning data is expression gene information for each of the plurality of gene data sets. The method of claim 6,
The learning unit
The method comprising the steps of: propagating and propagating in the current layer to learn the current layer and learning the next layer of the current layer when learning of the current layer is completed; and using the Deep Belief Network method, The disease information prediction system. The method of claim 6,
The learning unit
Setting the learning data to an input layer of the intermediate neural network and setting an output value assigned to the type of learning data input to the input layer to an output layer of the intermediate neural network, And generating the final in-depth neural network by supervised learning of the middle-depth neural network based on the value of the threshold value. 9. The method of claim 8,
Wherein the type of the learning data includes any one of a first type indicating a normal gene data set extracted from a normal population and a second type indicating a disease gene data set extracted from the patient population,
The learning unit
Wherein when learning data corresponding to the first kind is input to the input layer of the intermediate neural network, a first output value is set in the output layer of the intermediate neural network to learn the intermediate neural network, Wherein learning data is input to the input layer of the intermediate neural network, and a second output value is set to the output layer of the intermediate neural network to learn the intermediate neural network. 9. The method of claim 8,
The learning unit
The predicted value obtained by calculating propagation in the forward direction in the input layer of the intermediate neural network is compared with the output value set in the output layer of the intermediate neural network to obtain the error and then the back propagation is performed in the direction of minimizing the error, A disease information prediction system that updates link strength.

Images