KR102606267B1 - Target Prediction Method and System using correction based on prediction reliability - Google Patents

Abstract

The present invention relates to a target prediction method and system using correction techniques based on prediction reliability. Specifically, the present invention derives relevance information for each of a plurality of second entities related to the queried first entity, learns the training data to simultaneously derive the predictive reliability of the derived relevance information, and solves problems caused by biased distribution of the learning data. It is about the method and system that solved the problem.

Description

Target prediction method and system using correction technology based on prediction reliability {Target Prediction Method and System using correction based on prediction reliability}

The present invention relates to a target prediction method and system using correction techniques based on prediction reliability.

There are increasing attempts to increase the efficiency of new drug development by using artificial intelligence. A lot of effort has been made to develop artificial intelligence-based prediction models to search for new drugs that are effective in treating specific diseases, and the technology that can query the name of the disease to derive drugs related to the queried disease is a core technology in new drug development. It is becoming.

Meanwhile, drugs derived through a prediction model can be viewed as having a relationship with the queried disease through a specific relationship, and may have different relevance scores. However, the relevance score is different for each prediction model depending on what learning data it was trained with. Learning data may be biased in a certain area depending on the technology level at the time of training, and the prediction reliability distribution varies depending on the biased/unbiased area of the learning data.

Even though the prediction reliability is different for each drug found to be related to the queried disease, it is inappropriate to interpret the results by placing them on the same line.

Accordingly, the present inventors developed the present invention, which provides more accurate results by using a correction result value reflecting correction based on the prediction reliability of the result value derived from a prediction model generated through learning as the final output.

Korean Patent Document No. 10-2225278 (2021.03.10.)

In order to solve the above problem, the present invention derives relevance information for each of a plurality of second entities related to the queried first entity, learns the learning data to jointly derive the prediction reliability of the derived relevance information, and distributes the learning data bias. The purpose is to provide a method and system that solves the problems caused by.

In addition, the purpose is to provide a method and system that improves the accuracy of the generated prediction model by learning training data so that the step of deriving relevance information and the step of deriving prediction reliability are performed simultaneously.

In addition, the purpose is to provide a method and system with higher reliability compared to the prior art by correcting the relevance information (relevance score) by reflecting the derived prediction reliability and providing the user with the corrected relevance information that is the result of the correction.

In addition, the purpose is to provide a method and system that improves the accuracy of the generated prediction model by imposing constraints so that the weight matrix of the artificial neural network model is an orthogonal matrix in the process of learning training data to derive prediction reliability.

In addition, the purpose is to provide a method and system that can solve the issue of overconfidence by lowering the reliability of input without learning history by deriving prediction reliability.

In one embodiment of the present invention for achieving the above object, when a first entity is queried, a learning processor provides a plurality of second entities related to the queried first entity and each of the queried first entity and the second entity. A first learning step of training an artificial neural network to derive relevance information for each and the learning processor, when a first entity is queried, a plurality of second entities related to the queried first entity and each of the queried first entity A second learning step of training an artificial neural network to derive prediction reliability for each individual, wherein the artificial neural network has disease-related data defined as a first node and gene-related data defined as a second node, A target prediction method is provided in which drug-related data learns learning data including node data defined as a third node and a preset path for each node-pair.

In one embodiment, the first learning step and the second learning step may be performed simultaneously.

In one embodiment, the second learning step is to learn so that the prediction reliability is maximized for the training data, and the training data is different depending on the similarity between the queried first entity-derived second entity pair and the training data. A learning step may be further included to derive prediction reliability.

In one embodiment, the second learning step is to learn so that the prediction reliability is maximized for the training data, and the queried first entity - derived second entity pair is mutually dependent on whether or not the pair is included in the training data. A step of learning to derive different prediction reliability may be further included.

In one embodiment, the second learning step may include learning the training data so that the weight matrix of the artificial neural network becomes an orthogonal matrix.

In one embodiment, the weight matrix (Q) is: when, , where I is the identity matrix, inside may be a column vector of a parameter matrix that is a lower triangular matrix.

In one embodiment, the arbitrary first entity is a term belonging to any one type of disease, gene, or drug, and the prediction model outputs second entities belonging to a type different from the type of the input first entity. You can.

In one embodiment, an input step in which an arbitrary first object is input to an input device, a processor queries an input layer of the artificial neural network for an arbitrary first object, and the plurality of It may further include a derivation step in which two entities, the relevance information, and the prediction reliability are derived, and an output step in which the output device outputs the plurality of second entities, the relevance information, and the prediction reliability derived in the derivation step. .

In one embodiment, after the derivation step and before the output step, the processor further includes a correction step in which the processor corrects the relevance information for each of the plurality of second entities based on the derived prediction reliability to derive corrected relevance information. And, the output step may further include outputting the correction relevance information derived in the correction step.

In one embodiment, the relevance information includes a relevance score proportional to the relevance to the queried first entity, and the correction step includes correcting the relevance score so that the lower the derived prediction reliability is, the lower the relevance score is. More may be included.

In one embodiment, the learning data is calculated by a preset method among the embedding vectors of the first to third nodes and a plurality of paths included in the preset path type (metapath) for each node-pair. It may include some paths extracted in order of high path score.

In one embodiment, the training data includes an embedding vector output from the query encoding model as a pair of random entities is input to the query encoding model, and a path connecting the pair of random entities input to the query encoding model. As some of the extracted paths are input to the path encoding model, the embedding vector output from the path encoding model may include an embedding vector output from the query-path encoding model as it is input to the query-path encoding model. there is.

In one embodiment, the random entity pair includes entity embedding vectors for identifying each entity of the random entity pair, and a positional embedding vector for identifying the location of each of the entity embedding vectors. is input to the query encoding model in the form of a query embedding vector containing It can be input to the path encoding model in the form of a path embedding vector including embedding vectors for identifying each of the edge types of connecting edges, and positional embedding vectors for identifying the positions of each of the embedding vectors. .

In addition, the present invention is a system using the above-described pre-trained artificial neural network, which includes an input device configured to receive an arbitrary first object, and a query for an arbitrary first object input through the input device to the input layer of the artificial neural network. A target prediction system is provided, including a processor configured to derive a plurality of second entities, relevance information, and prediction reliability, and an output device configured to output a plurality of second entities, relevance information, and prediction reliability derived from the processor. .

In one embodiment, the processor may correct the relevance information based on the derived prediction reliability, and the corrected relevance information may be output through the output device.

Additionally, the present invention provides a computer program stored on a computer-readable recording medium for executing the above-described method.

According to the present invention, relevance information for each of a plurality of second entities related to the queried first entity is derived, and the learning data is learned to simultaneously derive the predictive reliability of the derived relevance information, thereby solving the problem caused by biased distribution of the learning data. do.

In addition, the accuracy of the generated prediction model is improved by learning the training data so that the step of deriving relevance information and the step of deriving prediction reliability are performed simultaneously.

In addition, the reliability is higher compared to the prior art by correcting the relevance information (relevance score) by reflecting the derived prediction reliability and providing the corrected relevance information that is the result of the correction to the user.

Additionally, in the process of learning training data to derive prediction reliability, the accuracy of the generated prediction model is improved by imposing constraints so that the weight matrix of the artificial neural network model is an orthogonal matrix.

In addition, by deriving the prediction reliability, the issue of overconfidence is resolved by lowering the reliability of input without learning history.

1 is a schematic block diagram for explaining a system according to an embodiment of the present invention.
FIG. 2 is a schematic block diagram illustrating a data processing process by a first processor in the system of FIG. 1.
Figure 3 is a diagram for explaining a knowledge graph constructed in a system according to an embodiment of the present invention.
Figure 4 is a diagram for explaining nodes and edges used in the process of building a knowledge graph according to the present invention.
FIG. 5 is a diagram illustrating a data processing process by a first processor in the system of FIG. 1.
FIG. 6 is a diagram illustrating a data processing process by a second processor in the system of FIG. 1.
FIG. 7 is a diagram illustrating an embedding vector input to the query encoding model of FIG. 6 and an embedding vector output accordingly.
FIG. 8 is a diagram for explaining an embedding vector input to the path encoding model of FIG. 6 and an embedding vector output accordingly.
Figure 9 is a diagram showing the embedding vectors input to the query-path encoding model of Figure 6, the embedding vectors output accordingly, and how they are learned.
Figure 10 is a diagram explaining the artificial neural network model learning process by a learning processor.
Figure 11 is a diagram to explain the query and derivation process for a trained artificial neural network model.
Figure 12 is a diagram for explaining aspects of output information according to an embodiment of the present invention.

Hereinafter, the present invention will be described in detail with reference to the attached drawings.

Hereinafter, the term “entity” means a word or symbol, and may include the name of the disease, the name of the gene (or protein), and the name of the drug. Likewise, “entity-pair” refers to data consisting of pairs of entities, and data consisting of different types of entities (disease-gene, gene-disease, disease-drug, drug-disease, gene-drug, drug-gene etc.).

Hereinafter, the term "gene" refers to an individual unit of genetic information composed of a specific base sequence in a genome composed of DNA or RNA, and genetic information composed of a specific amino acid sequence in a genome composed of protein as well as DNA and RNA. It is a concept that also includes individual units of . Additionally, protein, which is the result of translation of the genome, is also included in the term “gene” of the present invention.

Hereinafter, the term "target" refers to a second entity (predicted entity) related to the first entity queried, where both the first entity and the second entity belong to any one type of disease, gene, or drug, but The types of the first entity and the second entity are different (for example, if the first entity is a disease, the second entity is a gene or drug).

1. Description of the system

Referring to Figure 1, a system according to an embodiment of the present invention will be described in detail.

A system according to an embodiment of the present invention includes a model creation terminal 100 and a prediction terminal 200.

The model generation terminal 100 is a terminal configured to generate a prediction model used in the present invention.

Referring to FIG. 1, the model creation terminal 100 includes a communication unit 110, an input unit 120, a first processor 130, a second processor 140, a control unit 150, an output unit 160, and a memory ( 170) and a learning processor 180. That is, the model creation terminal 100 may be a computing device having communication/input/computation/output functions, and may be implemented as various electronic devices such as servers, desktop PCs, laptop PCs, tablet PCs, etc.

In one embodiment, the communication unit 110 is configured for communication with an external device, and data may be transmitted and received with the external device through the communication unit 110.

In one embodiment, the input unit 120 transmits commands or data to be used for components (e.g., processor, learning processor) of the model creation terminal 100 to an external (e.g., user) of the model creation terminal 100. It can be received from. The input unit 120 may include, for example, a microphone, mouse, keyboard, keys (eg, buttons), or a digital pen (eg, stylus pen).

In one embodiment, the control unit 150 may control one or more other components (eg, hardware or software components) of the model creation terminal 100 connected to the processor by executing software, for example.

In one embodiment, the first processor 130, the second processor 140, and the learning processor 180 may perform data processing or calculation functions, and as at least part of the data processing or calculation function, other components Commands or data received from may be stored in volatile memory, commands or data stored in volatile memory may be processed, and resultant data may be stored in non-volatile memory. Additionally, the learning processor 180 may include a hardware structure specialized for processing artificial intelligence models.

The output unit 160 can output information processed or calculated in the model creation terminal 100 to the outside, and a display, speaker, etc. may be included here.

The memory 170 can store various data used in the model creation terminal 100. Data may include, for example, input data or output data for software and instructions related thereto.

The prediction terminal 200, which will be described later, may also be implemented in the form of a computing device, and may include the components included in the model creation terminal 100 as is or at least some of them.

The learning processor 180 of the model creation terminal 100 learns training data, and when any first entity is queried, derives a plurality of second entities related to the queried first entity, and Relevance information for each second entity is derived, and a prediction model is created to derive prediction reliability for each second entity. Below, we will describe in detail how the learning data used to create the prediction model is structured.

Learning data is preprocessed and generated by the first processor 130 and the second processor 140 of the model creation terminal 100.

The first processor 130 defines nodes, edges, and paths, generates a knowledge graph consisting of the defined nodes and edges, embeds the generated knowledge graph, and uses the embedding result to create paths of arbitrary entity-pairs. It is configured to extract some paths of high importance.

Referring to FIG. 2, the first processor 130 includes a data collection unit 131, a natural language processing unit 132, a node definition unit 133, an ID granting unit 134, an edge definition unit 135, and a path definition unit. (136), an embedding unit 137, a path score calculation unit 138, and a path extraction unit 139.

The data collection unit 131 is configured to collect data from multiple databases (D1, D2, ... Dn). Data collected by the data collection unit 131 may be, for example, gene expression data, drug-protein binding data, data itemizing information described in papers, document data, etc., but is not limited to the above-mentioned forms and is disease-related data. , the format is not limited as long as it includes gene-related data and drug-related data.

For this purpose, the system according to the embodiment of the present invention can be connected to a plurality of databases (D1, D2, ... Dn) (can be connected to each other by the communication unit of the model creation terminal), and a plurality of databases (D1, D2 , … Dn) may be a public database, but is not limited to this and may also be a private database, and may include a thesis database, a medical information database, a pharmaceutical information database, and a search portal database.

The data collection unit 131 collects first data related to a disease, second data related to a gene, and third data related to a drug (compound) from each of a plurality of databases (D1, D2, ... Dn). can be collected. In addition, the data collection unit 131 may collect the path as fourth data that is different from the first to third data, and documents (e.g., papers, etc.) that serve as the basis for determining that there is a relationship between the data. ) can be collected as another type of fifth data (for example, the ID assigned to each relevant paper in a specific database) to refer to the paper.

The first data is data related to the disease, such as name data of the disease, anatomy data of the disease (for example, anatomical data of the body in which the disease occurs, in the case of liver cancer, this may include the liver), and May include disease symptom data. In other words, it is a concept that includes not only terms referring to the disease itself, but also all terms necessary to provide information related to the disease.

The second data is data related to the gene, including name data of the gene, gene ontology data of the gene, anatomical data of the gene (e.g., body tissue information in which the gene is expressed, genes related to liver cancer). To find genes with high expression in the liver, liver may be included) and biological pathway data of the gene, and gene ontology data is the biological process data of the gene. , may include data on the cellular location of the gene (cellular component) and data on the molecular function of the gene. In other words, it is a concept that includes not only terms referring to the gene itself, but also all terms necessary to provide information related to the gene.

Anatomical data may be included in first data or second data. For example, if the data includes that gene A is expressed in tissue B, tissue B may be collected as second data that is gene-related data. , if the data includes that disease C occurs in tissue D, tissue D may be collected as first data that is disease-related data.

The third data is data related to the drug and may include name data of the drug, pharmacologic class data of the drug, and side effect data of the drug. In other words, it is a concept that includes not only terms referring to the drug itself, but also all terms necessary to provide information related to the drug.

However, it is not limited to the above types and can include data related to diseases, genes, drugs, and any data necessary to predict the relationship between diseases, genes, and proteins.

The natural language processing unit 132 extracts entities from the text included in the document data through a preset natural language processing algorithm from the document data collected by the data collection unit 131, and derives relationships between the entities. It is configured to do so.

The relationships between entities extracted by the natural language processing unit 132 and the derived entities may be defined as nodes and edges, respectively, and detailed descriptions will be provided later.

That is, the natural language processing unit 132 sets disease-related terms included in the document data as the first entity, gene-related terms as the second entity, and compound-related terms as the third entity. It is configured to recognize and extract terms that describe the relationship between the first entity to the third entity, respectively, as the fourth entity. In addition, the natural language processing unit 132 sets and extracts the entities necessary for prediction, such as recognizing and extracting the path included in the document data as the fifth entity and the data referring to the document that was the basis for the relevance judgment as the sixth entity. can do.

Additionally, the natural language processing unit 132 is configured to use the extracted entities to derive relationships between entities using a preset method.

Extraction of entities and derivation of relationships between entities by the natural language processing unit 132 according to the present invention may be performed using a pre-trained neural network model. That is, the neural network model is configured to, for example, learn learning data in which the first to sixth entities are each labeled, extract the first to sixth entities from the queried document data, and derive relationships between the entities. It can be.

According to the prior art, terms to be extracted are stored in advance in an index dictionary, and then only the pre-stored terms are extracted from the text. In this case, if the text contains a term that is not previously stored in the index dictionary, it cannot be extracted, and ultimately, the system can only be built within the existing known range.

However, in the case of the present invention, rather than extracting terms stored in the index dictionary, for example, the neural network model learns labeled training data to determine which part of the text corresponds to which of the set objects, so it is not pre-trained. For terms, it is possible to extract entities by considering the form of the term itself or its context. Therefore, it is possible to extract entities from categories known through existing papers as well as new categories and derive relationships between entities.

The node defining unit 133, the edge defining unit 135, and the path defining unit 136 are configured to define nodes and edges, which are components of the knowledge graph, and further define a path. do.

The node regulation unit 133 may group first data among the data collected by the data collection unit 131 into disease name data, disease anatomical data, disease symptom data, etc., and may group the collected second data into disease name data, disease anatomical data, disease symptom data, etc. Data can be grouped into name data of genes, biological process data of genes, anatomical data of genes, intracellular location data of genes, molecular function data of genes, biological pathway data of genes, etc., and collected third data. The types can be classified into a total of 11 groups by grouping them into drug name data, drug pharmacological classification data, and drug side effect data (see FIG. 4). However, the number is not limited to the above and various types of groups (eg, fifth data and sixth data) may be added.

In another embodiment, the node defining unit 113 may group each entity extracted through the natural language processing unit 112 according to a predetermined method and define each as a node.

That is, the node defining unit 133 divides the first to third entities extracted through the natural language processing unit 132 and the first to third data collected from multiple databases into the first to third nodes, respectively. It is defined as (see Figure 3). In another embodiment, the node defining unit 133 divides the first to third entities, the fifth entity, and the sixth entity, and the first to third data, fifth data, and sixth data into the first to third entities, respectively. It can also be defined as 5 nodes. As will be described later, the edge definition unit 135 defines the relationship between the first and third entities and the relationship between the first and third data derived through the natural language processing unit 132 as an edge. In addition, the edge defining unit 135 may define the relationship between the first to third objects, the fifth object and the sixth object, and the relationship between the first data to the third data, the fifth data and the sixth data as edges. there is. Examples of nodes defined in accordance with the present invention and edges connecting the nodes are shown in Figure 3.

Additionally, the node defining unit 133 defines the data included in the grouped data as each node according to its type.

That is, the node defining unit 133 defines the first data (object) as each node for each type, defines the second data (object) as each node for each type, and defines the third data (object) as each node. Each type is defined as a node, and the fifth data and sixth data are defined as nodes for each type.

The ID granting unit 134 is configured to grant a unique ID to each of the nodes defined by the node defining unit 133.

That is, the ID granting unit 134 according to the present invention assigns a unique ID to each arbitrary term representing each node, such as a synonym and abbreviation of the arbitrary term. It is configured to assign the same ID as the arbitrary term to terms that can be determined to be the same as .

Meanwhile, there may be cases where two or more IDs are assigned to any term. For example, alpha-fetoprotein is also referred to by the abbreviation AFP, and both alpha-fetoprotein and AFP can be assigned an ID of 174.

AFP is also a synonym for the gene TRIM26, that is, AFP may be given an ID of 7726, which is the same as the ID of TRIM26.

In other words, AFP is given two IDs, 174 and 7726. In this case, the ID granting unit 134 gives an ID (174) that matches the full name of the AFP, rather than an ID (7726) that matches the abbreviation. It is assigned as the AFP ID.

A unique ID is mapped and stored for each node in the memory 170, and the ID granting unit 134 assigns a unique ID to each node using the IDs stored in the memory 170. .

The edge defining unit 135 defines the relationship between nodes defined by the node defining unit 133 as an edge.

An edge refers to a connection relationship between nodes, and the edge defining unit 135 defines the relationship between nodes included in the collected data as an edge connecting the corresponding node-pair.

For example, if document data contains the text "In patients with breast cancer, lump symptoms may occur, and treatment may be performed using the hormonal drug tamoxifen," then there is a node called "breast cancer" and a node called "lumpy". One edge connecting the nodes may be defined, and one edge connecting the node “breast cancer” and the node “tamoxifen hormone drug” may be defined.

In this way, the edge defining unit 135 can define the relationship between nodes as an edge using the first data, second data, and third data collected by the data collection unit 131, and the node defining unit 133 and Likewise, defined edges can be grouped.

3 and 4 show edges defined by edge definition 135 and grouped and typed.

Referring to FIG. 3, the edges defined by the edge regulation unit 135 are a disease-gene relationship edge (Disease-Target), a gene-drug relationship edge (Target-Compound), and a disease-drug relationship edge (Disease-Compound). , can be divided into gene-related edges (Target-related), disease-related edges (Disease-related), and drug-related edges (Compound-related).

Figure 4 shows the edge type (metaedge) that categorizes each edge.

Specifically, the disease-gene relationship edge (Disease-Target) includes the gene-disease relationship edge type (associated) and the gene-disease regulation relationship edge type (downregulated_in, upregulated_in).

The gene-drug relationship edge (Target-Compound) includes the drug-gene binding relationship edge type (binds_to) and the drug-gene regulation relationship edge type (downregulated_by, upregulated_by).

The disease-drug relationship edge (Disease-Compound) includes drug-disease treatment relationship edge types (treats).

Gene-related edges (Target-related) are gene-anatomical data regulation/expression relationship edge types (expressed_low, expressed_in, expressed_high), gene covariation relationship edge types (covaries), and gene participation relationship edge types (biological_process, cellular_component, molecular_function). , involved_in), interrelationship edge types (PPI, PDI) between genes or proteins, and genetic interference-gene regulation relationship edge types (regulates).

Disease-related edges include disease-anatomical data relationship edge type (occurs_in), disease-symptom relationship edge type (presents), and disease co-occurrence similarity relationship edge type (mentioned_with).

Drug-related edges (Compound-related) include drug-side effect relationship edge type (causes), drug structural similarity relationship edge type (similar_to), and drug-pharmacological classification relationship edge type (categorized_in).

That is, the edge defining unit 135 can classify edges into 24 groups. However, it should be understood that the number is not limited to the above and various types of groups can be added.

The path defining unit 136 includes one or more, specifically two or more, edges defined by the edge defining unit 135, and defines the edges that are included as being connected to each other as a path.

More specifically, the path defining unit 136 is connected to the node defined by the node defining unit 133 and the edge defining unit 135 for each node-pair, with each node forming the node-pair as an end. What is connected to each other through defined edges is defined as a path.

In other words, the minimum unit path among the paths defined by the path regulation unit 136 may be composed of node - edge - node - edge - node, and in another example, node - edge - node - edge - node - edge - A path consisting of nodes, etc. may be defined.

More specifically, a path can be defined as a path in which node-pairs are connected to each other, but are connected by 2 or more and 5 or less edges (edges). More specifically, node-pairs are connected to each other, but are connected by 2 or more edges. Things (edges) connected to three or less edges can also be defined as a path. Node-pairs connected by four or more edges can be excluded from valid paths, because if nodes are connected to each other through multiple steps, the relationship can be considered weak.

For the paths defined by the path defining unit 136, a plurality of path types may be determined depending on the number of combinations of the number, order, and type (shown in FIG. 4) of edges constituting the path.

For example, the path "AKT1 - associates - Alzheimer's disease-resembles - Parkinson's disease" has the same path type as "Gene - associates - Disease - resembles - Disease". In other words, a path containing the A(type a)-B(type b) edge can be specified as (a,b) type, and the path containing the edge A(type a)-B(type b)-C(type c) Paths containing can be defined as (a,b,c) types and can be treated as different types.

Additionally, the path definition unit 136 may classify some of the multiple path types into preset path types (metapaths). As will be described later, path types that do not correspond to the preset path types are excluded from the learning process according to the present invention.

For example, the path regulation unit 136 may set a path type including an edge type in the order Disease -mentioned_with - Disease - associates_with - Gene as a preset path type among a plurality of path types, and in another example, Disease The path type including the edge type in the order of - treated_by - Compound - binds_to - Gene - interacts_with - Gene can be set to the preset path type. The present invention is not particularly limited to this, and a preset path type may be set by the system administrator. By learning only biologically meaningful paths among the paths connecting arbitrary node-pairs, learning efficiency and Accuracy can be improved.

In addition, the path regulation unit 136 selects an edge type in the order Disease - treated by - Compound - downregulates - Gene - regulated by - Gene among a plurality of path types determined according to the number, order, and number of combinations of types of edges. The path type including and the path type including the edge type in the order of Disease - downregulates - Gene - upregulated by - Compound - binds to - Gene may not be set to the preset path type. In this case, a path type that is excluded from the preset path type can be set by the system administrator, and biologically meaningless or less important paths among the paths connecting arbitrary node-pairs are excluded from the learning process, thereby learning Efficiency can be improved and calculation accuracy can be improved.

The embedding unit 137 includes a node defined by the node defining unit 133, an edge defined by the edge defining unit 135, an edge type (metaedge), a path defined by the path defining unit 116, and a preset Perform embedding for one or more of the path types (metapaths).

More specifically, the embedding unit 137 performs embedding for each of the node defined by the node defining unit 133 and the edge type defined by the edge defining unit 135.

Below, an example of an embedding method using the embedding unit 137 will be described.

First, the embedding unit 137 initializes all nodes defined by the node defining unit 133 into real vectors each composed of k random variables. Here k may be 128. However, it is not limited to this, and it is possible to initialize it with a real vector composed of various random variables such as 64, 256, 512, and 1024.

Next, all edge types defined by the edge defining unit 135 are initialized into real vectors each composed of k random variables. Here k may be 128. However, it is not limited to this, and it is possible to initialize it with a real vector composed of various random variables such as 64, 256, 512, and 1024.

Next, it is determined whether a random node-pair is connected to each other by an edge having an edge type defined by the edge definition unit 135, and injected as supervised learning label data. If any pair of nodes (source node, target node) are connected to each other by an edge having an edge type defined by the edge regulation unit 135, data of 1 will be injected, and if they are not connected to each other, data of 0 will be injected. It will be.

The prediction function, which takes three k-dimensional vectors (source node, target node, and edge type) as input, adjusts the k-dimensional vector to match the actual connection. Here, the prediction function may be a model such as TransE, HolE, or DistMult, but is not limited to this and various prediction function models may be applied to the present invention.

When adjustment is completed, k-dimensional real vectors corresponding to each node are calculated as the embedding results of the corresponding node and edge type.

In addition to the above-mentioned method, various embedding methods can be performed, and as a result of embedding by the embedding unit 137, each node can be mapped into one point in a k-dimensional space. Additionally, as a result of embedding by the embedding unit 137, not only can each of the first to third nodes be mapped in the k-dimensional space, but also edge types can be embedded together in the k-dimensional space.

The path score calculation unit 138 is configured to calculate the scores of edges included in the path according to a preset method and calculate the scores of the paths using the calculated edge scores.

The score of each edge included in the path is calculated using each node and edge type embedded in the embedding unit 137. That is, each edge included in the path has a k-dimensional real vector (mapping) of the corresponding edge type and k-dimensional real vectors of the start and end nodes of the edge, and the corresponding edge score can be calculated from these real vectors. As an example of a specific calculation method, the prediction function used in the embedding unit 137 can be applied, and the similarity of each node mapping can also be applied.

The calculation method based on the similarity of the node mappings is a method in which the higher the similarity between nodes mapped in a k-dimensional space, the higher the score is given to the edge connecting the corresponding nodes. As a similarity calculation method, a method of calculating the angle between vectors (more specifically, a method of calculating the cosine value of two vectors) can be applied. This is an example, so various methods of calculating similarity between vectors can be applied. I would say it can be done.

For a path that includes n edges (n is an integer greater than 1), the score of the path can be calculated by adding up the edge scores of each of the n edges, and for a path that includes n+1 edges, n+1. The score of the corresponding path can be calculated by adding up the scores of each edge.

The path extraction unit 139 is configured to extract some of the paths defined by the path definition unit 136 based on the path score calculated by the path score calculation unit 138.

Specifically, the path extractor 139 extracts some paths for each preset path type (metapath) among the paths of any node-pair (entity-pair).

As described above, path types can be classified according to the number, order, and type of edges included in the path. For example, a path containing the edges A (type a) - B (type b) can be specified as type (a,b), and A (type a) - B (type b) - C (type c). Paths containing edges can be defined as (a,b,c) types and can be treated as different types.

More specifically, for paths with a preset path type among paths of a random node-pair, some paths are selected for each path type in descending order of score using the path score calculated by the path score calculation unit 138. Can be extracted, and as an example, five paths can be extracted for each path type. However, it should be understood that the number of paths is not limited to 5 and less than 5 or more than 5 paths can be extracted.

The second processor 140 is configured to generate training data to be learned by the learning processor 180 using the embedding vector calculated by the first processor 130 and the extracted path.

Referring to FIGS. 6 to 9, this will be described in detail.

The second processor 140 is characterized in that it generates learning data using a model that encodes time-series information.

A model that encodes time-series information refers to a model that compresses the input sequence into a single vector representation (context vector), but encodes the input sequence so that the time-series characteristics of the input sequence are revealed in the compressed vector representation. To reveal the time series characteristics of the input sequence, the present invention uses a positional vector.

In the present invention, two encoding models are largely used. The first is the first encoding model and the second is the second encoding model. A two-layer encoding model structure in which the value output from the first encoding model is input as the input value of the second encoding model (i.e., a stacked encoding model) ) was adopted. Here, the first encoding model is a model that encodes time-series information, and the second encoding model does not encode time-series information. A detailed explanation of this will be provided later.

The first encoding model can be divided into two encoding models.

The first is the query encoding model, which is used to specify which entity-pair connection the path used for learning is.

Referring to Figure 7, arbitrary pairs of entities are input to the query encoding model in the form of query embedding vectors.

The query embedding vector is: query vector (first query embedding vector) - an embedding vector that specifies one object among a pair of random objects (second query embedding vector) - specifies the other object among a pair of random objects It is composed in the order of the embedding vector (third query embedding vector).

The first query embedding vector is calculated using the query type embedding vector and the first positional embedding vector to identify the location of the query type embedding vector. Specifically, the query type embedding vector and the first positional embedding vector. It may be a summed embedding vector of . Each embedding vector can be expressed as a k-dimensional vector value. The query type embedding vector and the first positional embedding vector can be vector values of the same dimension, and the first query embedding vector can be generated through the process of simply adding them. You can.

The second query embedding vector identifies the first node type embedding vector indicating the type to which any one of the pairs of entities of types including diseases, genes, and drugs corresponds, and any one of the entities. It is calculated using an entity embedding vector to identify the location of the entity embedding vector, and a second positional embedding vector to identify the location of the entity embedding vector. Specifically, the sum of the first node type embedding vector, the entity embedding vector, and the second positional embedding vector is calculated. It can be. Taking FIG. 6 as an example, when an entity has Alzheimer's disease, a node type embedding vector indicating “disease” may correspond to the first node type embedding vector.

In the case of an entity embedding vector, it may be calculated using a first entity embedding vector and a second entity embedding vector, and more specifically, it may be a concatenate embedding vector of the first entity embedding vector and the second entity embedding vector. there is. Here, "combining" is a different concept from "summation" in which a k-dimensional embedding vector and a k-dimensional embedding vector are added to obtain the result of an embedding vector that maintains the k dimension. It is a different concept from a k/2-dimensional embedding vector It refers to the computational process of obtaining the result of a k-dimensional embedding vector by combining k/2-dimensional embedding vectors (i.e., obtaining a result in which the dimensions are summed). In the present invention, it is preferable that the embedding vectors generated through “combining” the embedding vectors generated through “summing” are embedding vectors of the same dimension.

The first entity embedding vector refers to an embedding vector that refers to (identifies) an entity among the embedding results generated by the embedding unit 137 of the first processor 130. That is, the first entity embedding vector corresponds to an embedding vector that refers to one of the above-mentioned entities among the node and edge type embedding results, and even if learning is performed by the learning processor 180, its value is fixed. It is an element (that does not change).

The second entity embedding vector, like the first entity embedding vector, refers to an embedding vector that refers to one of the entities. However, when learning is performed by the learning processor 180, it is an element whose value changes (the corresponding embedding vector value changes in the process of adjusting the weights to maximize prediction performance), and the randomly selected embedding vector value This can be applied to the second object embedding vector.

In the present invention, the first entity embedding vector (information based on already known knowledge) obtained using existing knowledge is set as a fixed value, and the second entity embedding vector, the value of which can be adjusted to maximize prediction performance during the learning process, is set as a fixed value. By utilizing both object embedding vectors, both advantages obtained by utilizing information based on existing knowledge and artificial intelligence are taken. Accordingly, the effect of significantly increasing prediction performance compared to the prior art is achieved.

The third query embedding vector is similar in shape to the second query embedding vector. However, the second query embedding vector is an embedding vector used to specify one entity among the pair of random entities, and the third query embedding vector is used to specify the other entity among the pair of random entities. It differs only in that it is an embedding vector.

That is, the third query embedding vector is a second node type embedding vector indicating the type to which another entity corresponds, an entity embedding vector for identifying another entity, and a third node type embedding vector for identifying the location of the entity embedding vector. It may be calculated using a sional embedding vector, and more specifically, it may be the sum of the second node type embedding vector, the entity embedding vector, and the third positional embedding vector.

In this way, the query embedding vector is composed of a first query embedding vector, a second query embedding vector, and a third query embedding vector, and when these are input to the query encoding model, a context vector, which is one compressed expression, is output. In other words, it is possible to specify the arbitrary object-pair through the context vector (embedding vector) output from the query encoding model, while distinguishing in which order the arbitrary object-pair is placed (even if it is a pair of objects A and B, A distinction can be made between A-B object pairs or B-A object pairs).

Next, the path embedding vector input to the path encoding model will be described in detail with reference to FIG. 8.

If the query embedding vector is an embedding vector for specifying an arbitrary entity-pair, the path embedding vector is different in that it is an embedding vector for specifying the paths of an arbitrary entity-pair, and the form of the embedding vector itself is similar.

The path embedding vector includes a path type embedding vector (first path embedding vector), an embedding vector specifying one entity among a pair of random entities (second path embedding vector), and an embedding vector specifying an edge type (third path embedding vector). path embedding vector), an embedding vector (fourth path embedding vector) specifying an object connected to one of the entities by an edge, an embedding vector (fifth path embedding vector) specifying an edge type, and an embedding vector specifying the other entity. may include an embedding vector (sixth path embedding vector).

Since the path length between any object-pair can vary (some paths may consist of only two edges, while others may consist of three or more edges), the length of the path embedding vector can vary. However, it may have a minimum path embedding vector of the first path embedding vector to the sixth path embedding vector (the first path embedding vector to the sixth path embedding vector is 2 edges between any object-pair and 1 (corresponds to an embedding vector when it has a path leading from node (referring to one entity) - edge - node - edge - node (referring to another entity), containing nodes).

The first path embedding vector is calculated using a path type embedding vector and a fourth positional embedding vector to identify the position of the path type embedding vector, and specifically, the sum of the path type embedding vector and the fourth positional embedding vector. It can be a vector.

The second path embedding vector includes a first node type embedding vector indicating the type of the one entity, an entity embedding vector for specifying the one entity, and a fifth positional embedding vector for identifying the location of the entity embedding vector. It is calculated using a vector, and specifically may be the sum embedding of the first node type embedding vector, the entity embedding vector, and the fifth positional embedding vector. The entity embedding vector of the second path embedding vector is calculated using the first entity embedding vector and the second entity embedding vector as described above. Specifically, it may be a combined embedding vector of the first entity embedding vector and the second entity embedding vector. there is.

The third path embedding vector is an edge included in the path and a first edge type embedding vector for specifying the edge type to which the edge connected to one of the entities corresponds, and a first edge type embedding vector for identifying the location of the first edge type embedding vector. It is calculated using a 6-positional embedding vector, and specifically, it may be the sum of the first edge-type embedding vector and the sixth positional embedding vector.

The fourth path embedding vector is a second node type embedding vector indicating the type of the object connected to the one entity by an edge, an entity embedding vector for specifying the entity, and a seventh path for identifying the location of the entity embedding vector. It is calculated using a sional embedding vector, and specifically, it may be the sum of the second node type embedding vector, the entity embedding vector, and the seventh positional embedding vector.

The fifth path embedding vector is an embedding vector for specifying an edge connected to the entity specified in the fourth path embedding vector, and has a similar form to the third path embedding vector.

The sixth path embedding vector is an embedding vector for specifying another entity and has a similar form to the second path embedding vector and the fourth path embedding vector.

In this way, if the path embedding vector includes the first to sixth path embedding vectors (may include more path embedding vectors depending on the number of edges making up the path), and these are input to the path encoding model, one A context vector, which is a compressed representation, is output. In other words, it is possible to specify the path between arbitrary object-pairs through the context vector (embedding vector) output from the path encoding model, while also specifying which nodes and which edges and in what order the path is composed.

The embedding vector output from the query encoding model for each random object-pair and the embedding vectors output from the path encoding model (since there are multiple paths, an embedding vector can be output for each path) are all included in the query-path encoding model. is input, and a context vector, a single compressed representation, is output from the encoder.

Referring to FIG. 6, the embedding vector output from the query encoding model and the embedding vectors output from the path encoding model are respectively input as embeddings of the query-path encoding model (as the first query embedding to the third query embedding into the query encoding model). It is similar to inputting the constructed query embedding and inputting the path embedding composed of the first to sixth path embeddings into the path encoding model).

Meanwhile, while the embedding vectors input to the query encoding model and the path encoding model each include positional embeddings, the embedding vectors input to the query-path encoding model do not include positional embeddings. The embedding vectors input to the query-path encoding model are intended to specify high-importance paths between arbitrary entity-pairs, and are independent without having a time-series relationship with each other.

In other words, the context vector (embedding vector) output from the query-path encoding model corresponds to an embedding vector that reflects only the highly important paths between arbitrary entity-pairs.

Meanwhile, for any individual-pair, there are multiple biologically meaningful pathways. In the present invention, among the preset path types for each entity-pair, only the path extracted by the path extractor 139 is input to the query-path encoding model.

The embedding vector output from the query-path encoding model contains all information on the highly important path among i) a pair of random entities and ii) a path between the pair of random entities. That is, in the present invention, the learning data trained by the learning processor 180 in a random artificial neural network model includes i) a random entity-pair ii) a high-importance path among the paths between the random entity-pair, When any first entity is queried, the learning data is used to derive second entities related to the queried entity, relationship information between the queried first entity and each second entity, and prediction reliability for each second entity. You learn.

Here, the artificial neural network models are DNN (Deep Neural Network), CNN (Convolutional Neural Network), DCNN (Deep Convolutional Neural Network), RNN (Recurrent Neural Network), RBM (Restricted Boltzmann Machine), DBN (Deep Belief Network), and SSD. It may be a model based on (Single Shot Detector), MLP (Multi-layer Perceptron), or Attention Mechanism, but is not limited to this and various artificial neural network models can be applied to the present invention.

The learning processor 180 is configured to generate a prediction model by training the artificial neural network on the training data pre-processed and extracted from the first processor 130 and the second processor 140.

The learning processor 180 trains the artificial neural network model on the embedding vector output from the query-path encoding model. The embedding vector output from the query-path-encoding model is important in terms of the preset path type (metapath) of the entity-pair. Contains route information. More specifically, when a first entity is queried, the learning processor 180 creates an artificial neural network model to derive a plurality of second entities related to the queried first entity, and relevance information and prediction reliability for each second entity. Let them learn.

Here, the relevance information may be in the form of a score, and more specifically, may be a score in the form of a decimal point with a size of 0 to 1. The closer it is to 1, the higher the probability that the second entity is related to the queried first entity, and the closer it is to 0, the higher the probability that it is not related.

In other words, the artificial neural network model calculates the score of each second entity in the context of the relevance and importance of the “queried entity (first entity)” - “predicted entity (second entity)”. In other words, among the possible paths between the first entity and the second entity (e.g., disease - target), find a path that belongs to a preset path type (metapath), and determine the relationship between the first entity and the second entity for each path. Then calculate the weight. At this time, a high weight may be assigned if the path is related to the first entity and the second entity, and a low weight may be assigned if the path is unrelated to the first entity and the second entity. Here, the first entity and the second entity are different types. In other words, if the first entity belongs to the “disease” type, it means that the second entity belongs to the “gene” or “drug” type.

Next, multiple paths are merged into one real vector based on the calculated weights.

Next, relevance information can be calculated using a multilayer perceptron (MLP) that uses the merged real vector and the queried first and second entity embeddings as input.

Meanwhile, the learning data preprocessed and extracted by the first processor 130 and the second processor 140 is based on data collected from the database (D), and more specifically, the level of technology/type of database at the time of data collection. Depending on this, the distribution of learning data in a certain area may be different. In other words, because it relies on knowledge known at the time of data collection, areas that are not widely known at the time (e.g., disease areas with significantly fewer clinical cases) may require less training data than areas that are widely known (e.g., diabetes). There is no choice but to do so. As a result, even if the artificial neural network was trained, a problem of low prediction reliability occurred in some areas.

In order to overcome the technical limitations caused by the above-described deviation in the distribution of learning data, a process of forcibly equalizing the number of learning data was introduced. For example, we attempted to solve the problem by introducing a process to match the number of learning data in areas with a lot of data and areas with little data. However, when the number of training data is forcibly matched, there is a problem that the prediction reliability of the results derived from the trained model deteriorates.

However, in the present invention, rather than using a preprocessing process according to the prior art that forcibly resolves the learning data distribution deviation, the data is learned to derive prediction reliability, and the relevance information according to the derived prediction reliability is corrected, and the corrected relevance information is provided. Technical limitations due to deviations in learning data distribution were overcome by providing to users. This is explained in detail.

When a first entity is queried, the learning processor 180 according to an embodiment of the present invention uses an artificial neural network to derive a plurality of second entities related to the queried first entity, and relevance information and prediction reliability for each second entity. Learning the model has been described above.

Here, prediction reliability is an indicator indicating how reliable the relevance information of the second entity derived from the learned artificial neural network model is. The higher the reliability of the relevance information, the higher the prediction reliability (for example, a value close to 1 through normalization), and the lower the reliability of the relevance information, the lower the prediction reliability (e.g., a value close to 1 through normalization). For example, with normalization, it has a value close to 0). Meanwhile, unreliability, which has the opposite concept of prediction reliability, can be introduced, and can be calculated through the formula: non-reliability = (1 - prediction reliability value). Here, a higher unreliability may mean that the reliability of the relevant relevance information is low, and a lower unreliability may mean that the reliability of the relevant relevance information is high.

The learning processor 180 learns the training data so that the prediction reliability has the maximum value, that is, the prediction reliability is close to 1. Specifically, it learns using the embedding vector obtained by inputting the learning data into the query-path encoding model. You may.

Here, if the training data is trained so that the prediction reliability has the maximum value for all arbitrary inputs, there is a problem that it is difficult to classify the data according to the prediction reliability. Additionally, in the case of conventionally learned models, overconfidence problems may occur. In other words, although the derived entity has low relevance to the queried entity, a problem occurs in which relevance information with a high score is derived.

In an embodiment of the present invention, in order to solve the above problems, learning may be performed to have different prediction reliability depending on whether or not a given input (first entity - second entity pair) includes training data. More specifically, learning may be performed so that when a given input is included in the training data, it has a value of 1, and when it is not included, it has a value between 0 and 1 but less than 1. More specifically, learning can be done to have different prediction reliability depending on the similarity between the given input and the training data, and learning can be done to have a value proportional to the similarity.

Through the above-described learning process, the problem of overconfidence in the learned model can be solved, and as the prediction reliability is output together with relevance information, it is possible to check the reliability of the information output from the prediction model.

Meanwhile, in the present invention, training data is learned by constraining the weight matrix of the artificial neural network model to be an orthogonal matrix.

Artificial neural networks generally include one or more hidden layers between an input layer and an output layer, and during the learning process, nodes included in the input layer, hidden layer, and output layer (included in the artificial neural network structure) The connection strength (weight) between the nodes (meaning nodes and a different concept from the nodes defined in the node definition unit 133 described above) is adjusted. By imposing constraints so that these weights are orthogonal to each other, the stability and performance of the model can be improved.

The output layer of the artificial neural network according to an embodiment of the present invention includes a first output node related to relevance information, and a second output node related to prediction reliability.

In the present invention, training data is learned so that the matrix of weights (weight matrix) leading from the input node to the second output node is an orthogonal matrix, which can further improve the accuracy of deriving prediction reliability.

Here, the weight matrix (Q) of the artificial neural network can be expressed as follows.

[Formula 1]

when, ego,

Here, I is the identity matrix, inside is the column vector (real vector) of the parameter matrix, is the weight matrix of the artificial neural network. Here, the parameter matrix may be a lower triangular matrix.

Meanwhile, in the present invention, the step of learning the training data to derive relevance information (first learning step) and the step of learning the training data to derive prediction reliability (second learning step) are performed simultaneously, thereby generating the relevance information and prediction. The accuracy of deriving reliability is improved.

The artificial neural network in the present invention includes Deep Neural Network (DNN), Convolutional Neural Network (CNN), Deep Convolutional Neural Network (DCNN), Recurrent Neural Network (RNN), Restricted Boltzmann Machine (RBM), Deep Belief Network (DBN), It may be a model based on SSD (Single Shot Detector), MLP (Multi-layer Perceptron), or Attention Mechanism, but is not limited to this and various artificial neural network models can be applied to the present invention.

The prediction terminal 200 receives the artificial neural network model generated by the model generation terminal 100, and when a random first entity is input from the outside through the input unit 220, the processor 250 A first entity is input into an artificial neural network model, and through the output layer of the artificial neural network model, a plurality of second entities related to a random first entity, relationship information and prediction reliability for each second entity are derived. Information derived through the output layer of the artificial neural network model may be output in the form of information visible from the outside through the output unit 260 of the prediction terminal 200.

In one example, the processor 250 of the prediction terminal 200 may correct relevance information based on prediction reliability derived from an artificial neural network model and derive corrected relevance information that is a result of the correction.

Relevance information may be expressed as a relevance score in the form of a number, and the processor 250 may correct the final relevance score so that the lower the prediction reliability, the lower the final relevance score. For example, the corrected relevance information may be calculated through the formula (relevance information x prediction reliability), and in another example, it may be calculated through the formula (relevance information - unreliability). However, it is not limited to the above example, and any formula that ensures that the higher the value of prediction reliability is, the closer the value of the corrected relevance information is to the value of the (initial) relevance information can be applied to the present invention.

Figure 12 shows an example in which the correction technique proposed by the present invention is applied.

Alzheimer's disease was queried in the artificial neural network model, and secondary entities of GRIN2A, GRIN2B, PPARG, ADRB3, and PTGS2 were derived. Relevance information and prediction reliability are derived for each second entity, and the ranking is sorted and output according to the size of the relevance information.

Referring to FIG. 12, the corrected relevance information, which is a result of correcting the relevance information based on the derived prediction reliability, is derived again in the form of a score. The ranking is rearranged and output according to the correction relevance information, and you can see that it is different from the ranking before correction (before correction: GRIN2A - GRIN2B - PPARG - ADRB3 - PTSG2, after correction: GRIN2A - PTGS2 - PPARG - GRIN2B - ADRB3) ). Through this, system users can identify second entities with higher reliability.

2. Description of method

With reference to FIGS. 5, 10, and 11, the prediction method according to an embodiment of the present invention will be described in detail. Since the learning method of the artificial neural network model used in the prediction method has been described above, only the key parts of each step will be briefly explained below.

First, a method of generating a prediction model by the model creation terminal 100 will be described in detail with reference to FIGS. 5 and 10.

The data collection unit 131 collects at least one of first data related to a disease, second data related to a gene or protein, and third data related to a drug from a plurality of databases (D1 to Dn) (S51).

Next, the node defining unit 133, the edge defining unit 135, and the path defining unit 136 define nodes, edges, and paths using the collected first to third data (S52).

Next, the embedding unit 137 performs embedding so that the specified node and edge types are real vectorized in a k-dimensional space (S53).

Next, the path score calculation unit 138 calculates the score of each edge included in the path using the embedding result, and calculates the path score of each path using the calculated edge score (S55).

Next, based on the path score calculated by the path score calculation unit 138, the path extraction unit 139 extracts a plurality of paths for each preset path type of the node-pair (S55). Paths extracted by the path extractor 139 are used for learning, and paths that are not extracted (paths of low importance) are excluded from learning.

Next, the second processor 140 generates learning data including the embedding vectors of the first to third nodes generated in each step of FIG. 5 and the paths extracted by the path extractor 139 (S101) . Specifically, the embedding vector obtained by inputting data including the embedding vectors of the first to third nodes and the path extracted by the path extractor 139 into the query-path encoding model may be used as learning data.

Next, when the first entity is queried, the learning processor 180 derives a plurality of second entities related to the queried first entity and relevance information (relevance score) for each of the queried first entity and each second entity. Training data is trained on the artificial neural network (S102). A relevance derivation model is created through the corresponding learning step.

Next, when the first entity is queried, the learning processor 180 trains the artificial neural network with training data to derive the prediction reliability for each second entity related to the queried first entity (S103). Through this learning step, a predictive reliability derivation model is created.

In one embodiment of the present invention, steps S102 and S103 may be performed simultaneously. In other words, the learning processor 180 learns the training data to obtain a plurality of second entities related to the queried first entity, relationship information for the queried first entity and each second entity, and the queried first entity and Training data is trained in an artificial neural network to simultaneously derive the prediction reliability for each second entity. In other words, the relevance derivation model and the prediction reliability derivation model can be created as one prediction model rather than as separate models (S104).

Referring to FIG. 11, a random first entity (disease, drug, or gene) is queried in a predicted model (artificial neural network model) on which learning has been completed (S111), and a plurality of second entities, each related to the queried first entity, are queried (S111). Relevance information and prediction reliability for each second entity are derived (S112).

Next, the relevance information is corrected based on the derived prediction reliability (S113), and the corrected relevance information can be output in a form visible to the user (S114).

The prediction method according to an embodiment of the present invention described above may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the medium may be specially designed and constructed for the present invention or may be known and usable by those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to carry out the operations of the present invention, and vice versa.

Above, the present invention has been described with reference to the embodiments shown in the drawings so that those skilled in the art can easily understand and reproduce the present invention, but these are merely illustrative examples, and various modifications and equivalent alternatives can be made by those skilled in the art from the embodiments of the present invention. It will be appreciated that embodiments are possible. Therefore, the scope of protection of the present invention should be determined by the claims.

100: Model creation terminal
110: Department of Communications
120: input unit
130: first processor
131: Data collection department
132: Natural language processing unit
133: Node regulation part
134: ID grant unit
135: Edge regulation part
136: Path regulation unit
137: Embedding part
138: Path score calculation unit
139: Path extraction unit
140: second processor
150: control unit
160: output unit
170: memory
180: Running processor
200: prediction terminal
210: Department of Communications
220: input unit
230: control unit
240: memory
250: processor
260: output unit

Claims (16)

As an arbitrary first entity is input into the target prediction model, a derivation step in which the processor derives a plurality of second entities related to the input first entity, relationship information and prediction reliability for each of the first entity and each second entity; and
A correction step in which the processor derives corrected relevance information by correcting relevance information for each of a plurality of second entities related to the input first entity, based on each derived prediction reliability, and
The target prediction model is,
When a first entity is queried, the learning processor trains an artificial neural network to derive a plurality of second entities related to the queried first entity and relationship information between the queried first entity and each second entity. step; and
The learning processor trains an artificial neural network to derive, when a first entity is queried, a plurality of second entities related to the queried first entity and prediction reliability for the queried first entity and each second entity. Created by the learning phase;
The artificial neural network includes node data in which disease-related data is defined as a first node, gene-related data is defined as a second node, and drug-related data is defined as a third node, and a preset path for each node-pair. learning training data,
Target prediction method.
According to paragraph 1,
The first learning step and the second learning step are performed simultaneously,
Target prediction method.
According to paragraph 1,
The second learning step is,
Learning to maximize the prediction reliability for the training data, further comprising learning to derive different prediction reliability depending on the similarity between the queried first entity-derived second entity pair and the learning data,
Target prediction method.
According to paragraph 1,
The second learning step is,
Learning so that the prediction reliability is maximized for the training data, but further comprising learning to derive different prediction reliability depending on whether the queried first entity - derived second entity pair is included in the learning data. ,
Target prediction method.
According to paragraph 4,
The second learning step is,
Comprising the step of learning the training data so that the weight matrix of the artificial neural network becomes an orthogonal matrix,
Target prediction method.
According to clause 5,
The weight matrix (Q) is,
when, ego,
Here, I is the identity matrix, inside is the column vector of the parameter matrix, which is a lower triangular matrix, ,
Target prediction method.
According to paragraph 1,
The first entity is a term belonging to any one type of disease, gene, or drug,
The prediction model outputs second entities belonging to a type different from the type of the input first entity,
Target prediction method.
According to paragraph 1,
Further comprising: an output step of outputting the plurality of second entities, the relevance information, and the prediction reliability derived in the derivation step from an output device,
Target prediction method.
According to clause 8,
The correction step is performed after the derivation step and before the output step,
The output step further includes outputting correction relevance information derived from the correction step,
Target prediction method.
According to paragraph 1,
The relevance information includes a relevance score proportional to the relevance to the queried first entity,
The correction step further includes the step of correcting the relevance score so that the lower the derived prediction reliability, the lower the relevance score.
Target prediction method.
According to paragraph 1,
The learning data is,
Embedding vectors of first to third nodes; and
Among the plurality of paths included in the preset path type (metapath) for each node-pair, some paths are extracted in order of high path score calculated by a preset method; including,
Target prediction method.
According to clause 11,
The learning data is,
As a pair of random entities is input to the query encoding model, the embedding vector output from the query encoding model and some of the paths extracted among the paths connecting the pair of random entities input to the query encoding model are path encoded. An embedding vector output from the path encoding model as input to the model includes an embedding vector output from the query-path encoding model as input to the query-path encoding model,
Target prediction method.
According to clause 12,
The pair of random entities is,
To the query encoding model in the form of a query embedding vector including entity embedding vectors for identifying each of the entities of the arbitrary entity pair and positional embedding vectors for identifying the location of each of the entity embedding vectors. is entered,
The path connecting the pair of arbitrary entities is,
Embedding vectors for identifying objects included in the path between the arbitrary pair of objects, each of the edge types of edges connecting the objects, and a positional signal for identifying the location of each of the embedding vectors. Input to the path encoding model in the form of a path embedding vector including embedding vectors,
Target prediction method.
A system using an artificial neural network pre-trained through any one of claims 1 to 13,
an input device configured to receive an arbitrary first object as input;
To query the input layer of the artificial neural network for any first entity input through the input device to derive a plurality of second entities, relevance information, and prediction reliability, and to correct the relevance information based on the derived prediction reliability. configured processor; and
Including; an output device configured to output a plurality of second entities, relevance information, and prediction reliability derived from the processor;
Target prediction system.
According to clause 14,
Correction relevance information is output through the output device,
Target prediction system.
Stored on a computer-readable recording medium to execute the method according to any one of claims 1 to 13,
computer program.

Images