KR100575495B1 - Method for Extracting and Inferring the Interaction of Biological Components, Inferring Program for Performing the Method and Recording Medium thereof - Google Patents

Abstract

The present invention relates to a method for extracting and inferring interactions between biological objects in a text related to biology, a program for executing the method, and a recording medium storing the program. We perform text analysis using, deduce high-level interaction information from extracted primary interaction information, and visualize and show interaction relations between objects.

Bio Text, Bioinformatics, Interaction, Genes, Proteins

Description

Methods for Extracting and Inferring the Interaction of Biological Components, Inferring Program for Performing the Method and Recording in Biological Texts Medium kind}

1 is a structural diagram showing an overview of an interaction relationship inference system according to the present invention.

2 is a structural diagram showing an outline of a text analysis step;

3 is a structural diagram showing an overview of entity name recognition.

4 is a structural diagram showing an overview of surface parsing.

5 is a structural diagram showing an outline of interaction relationship extraction;

6 is a structural diagram showing an overview of the relationship inference and visualization steps.

7 is a structural diagram illustrating an embodiment of an interface for visualization;

The present invention relates to a method for extracting and inferring interactions between biological objects in a text related to biology, a program for executing the method, and a recording medium storing the program. Extracting and logically inferring the interaction relations of various terms appearing in the literature, and identifying the interactions and similarities between them and visually displaying them. A program for execution and a recording medium storing the program.

Bioinformatics is a general field of research that intends to efficiently produce, manage, and utilize information processing techniques used in fields such as computer science, mathematics, and statistics as the amount of information dealing with biology increases rapidly.

In the present invention, we will look at biotext mining in the field of bioinformatics.

Biotext mining in the present invention is to apply data mining technology to bio-related texts, and is to obtain necessary knowledge through effective processing of an explosion of bio-related literature.

That is, after the success of the genome project, various research results related to biotechnology (BT) such as biology have been published, and the quantitative increase in the number of related documents is gradually accelerating. In the field of biology, new types of protein or gene names and new research literature on the relationship between them are constantly flowing, making it difficult for scholars and researchers in the field to obtain the information they want. Therefore, the importance of biotext mining techniques for extracting meaningful information from BT-related literature databases is increasingly emphasized. Therefore, if the text mining technology is applied to the biology-related field that has a lot of problems with access to information, it can be contributed to the improvement of research efficiency in the biological research field. In particular, since most of the recent research results exist in the form of electronic documents or databases that can be accessed online, the foundation environment for the efficient use of such mining technology is already established.

Data mining technology applied to the biotechnology field requires natural language processing technology for analyzing texts and data mining-related learning algorithms that infer high-dimensional relationships because biological documents are in natural language. The useful information extracted from bio-related texts by applying these techniques is the interaction between proteins and proteins, or between genes and genes. By automatically extracting interactions between entities from text and storing them in a database, you can search for specific proteins or genes and view all the proteins and genes that interact with them in graphs, and each of them You will know if you have it. This can help biology researchers in many ways.

A conventionally disclosed system is MedStract. MedStract is a system that automatically extracts interaction information between objects and uses only solid natural language processing technology without inference process. The system recognizes individual names and parts-of-speech using dictionaries that accompany UMLS thesaurus, taking into account words that are characteristic of biotext. MedStract applies several independent automata step by step to recognize noun phrases and verb phrases, and extracts the interaction information according to the defined pattern. It provides a web-based user interface where keyword search results are given in the form of tables and graphs.

Next, GENIES is a natural language processing system that extracts molecular pathways from biological literature. This is a variation of MedLEE as one of the modules that make up GeneWay. GENIES uses Term taggers to identify protein or gene names, preprocessors to determine sentences, words and phrases, parsers to check for proper interactions using grammar of constraint rules and semantic patterns, and errors in parsing. It consists of an error recovery module that handles using heuristics. The system constructs a pathway using the interaction information between the extracted entities.

BIOBIBLIOMETRICS is a system for retrieving and visualizing information from biological literature databases using gene names. The system was developed from the assumption that if two genes have related biological functions, they often air in the biological literature. In the literature DB, the similarity was obtained from the degree of air of two genes, and it was determined that it was related if it was above a certain threshold. In fact, when a biological researcher searches for a particular gene, other genes related to that gene are searched for and visualized, which helps to study the relationship between genes.

However, existing analysis systems require manual data entry for new gene names and protein names that are constantly being published, which requires a great deal of time and money. In addition, there is a limit in extracting a formal interaction relationship between biotexts.

This, in turn, reveals the limitations of information processing capabilities of researchers in the field of biotechnology and causes problems that reduce the efficiency of research-related information collection.

The present invention has been made to solve such a problem, it performs a text analysis using natural language processing technology for the text data to be analyzed, and extracts the high-level interaction information from the extracted first interaction information. The purpose of the present invention is to provide a method of extracting and inferring interactions between biological entities in a biological text that infers and visualizes the interaction relations between the objects, a program for executing the method, and a recording medium storing the programs. .

The present invention for solving the above problems is a method of extracting and inferring interactions between biological objects in a biological text, the text analysis step of performing the analysis using the natural language processing technology on the biotext; Inferring secondary interaction information from the first interaction information extracted as a result of the text analysis and expressing the second interaction information in a network, and generating a hypothesis for the network; And visualizing the interaction relationship between the biotexts using the network structure between the biotexts.

The text analyzing step includes assigning and storing accurate part-of-speech information to each biotext in the document; Attaching a boundary recognition tag of an entity name using a TO notation (Term or Other) to the document, and generating a single support vector machine (SVM) model by considering the TO as a boundary class; Combining syntactically related texts into a basic phrase based on the boundary of the part-of-speech information and the entity name; Finding another basic phrase having a dependency relationship with a verb based on the combined state of the basic phrases and expressing a sentence relationship such as a main subject and a target case; Analyzing a pattern by analyzing a sentence containing a verb specified in the document or a sentence including a verb used at a predetermined frequency or more in the document, and extracting primary interaction information using the analyzed pattern information. It includes; step.

In the relation inference step, in order to extract the secondary interaction information through the analysis of the primary interaction information, the inference of the association rule and the higher relationship inference are performed and learned by using data mining and machine learning algorithms, and the inferred association Organizing rule information and input data into an interactive network; Development of a clustering model for inference of correlation rules between hierarchical and non-hierarchical groups and hierarchical grouping, and statistical hypothesis of inference associations through analysis of relationships extracted using DBN (Dynamic Bayesian Network) model It comprises; generating.

The visualization step may be characterized by using the visualization tool including a graph, diagram, network structure to represent the extracted and inferred interaction relationship with the reliability confirmation element.

The biotext includes a gene name, a protein name.

The verb is a verb that frequently appears in biological documents, including activate, inhibit, associate, bind.

Another embodiment of the present invention for solving the above problems is a program for executing a method of extracting and inducing interactions between biological objects in the biological text.

Another embodiment of the present invention for solving the above problems is a recording medium storing a program for executing a method of extracting and inferencing interaction between biological objects in the biological text.

It will be described below with reference to the drawings.

Constant research in the field of biology has led to many new protein and gene names. Therefore, it is difficult to automatically extract information related to a new object from text unless the existing dictionary is complemented. However, it is too much human and time consuming to update the dictionary manually. Therefore, even if the object does not exist in the dictionary, the machine learning method is actively used to automatically recognize and extract information about the object. However, in order to perform machine learning effectively, a corpus with attached information on interaction between objects is required.

Currently, the GENIA project in Japan, a representative study on building resources for biotext analysis, is building a learning corpus to automatically extract the names of proteins and genes from related research literature and recognize the relationships between them. So far, we have built a small learning corpus to extract the names of proteins and genes. The learning corpus in a GENIA project is built on a completely passive method, which requires a long time and effort to build a practically useful corpus. Therefore, there is a need for a methodology for effectively constructing a large amount of learning corpus while minimizing manual work.

In the present invention, resources are constructed by experts in the field of biology only for literature related to a specific topic in the field of biology. Experts first identify the objects of interest in the literature and determine the semantic class to which they belong. It defines what an eventual verb represents an event that is an interaction between objects, and adds a tag if the relationship between objects is expressed as an eventual verb in the literature. At this time, two or more people review one document so as to maintain objectivity as much as possible. The tagged documents are used as learning corpus in the object name recognition and interaction relationship extraction.

The present invention utilizes a published bio-related database for effective resource construction. Bio-related literature for information extraction is obtained from MEDLINE, an open database provided by the National Library of Medicine. REBASE, GenBank, etc. are used for gene names, and PDB, PIR, SWISS-PROT, etc. are used for protein names. In addition, the semantic classification system of bio-related terms such as UMLS is used for the semantic classification of each individual.

1. System Configuration

1 is a structural diagram showing an outline of an interaction relationship inference system according to the present invention.

The ultimate goal of a biotext mining system is to implement a system that analyzes biotext documents to extract formal interactions between biological elements such as genes and proteins. For this purpose, the system can be composed of text analysis, relationship inference module, and network visualization module.

1-1. Text analysis module

2 is a structural diagram showing an outline of a text analysis step.

The text analysis module is the beginning of the biotext mining system, and performs text analysis using natural language processing techniques for text data to be mined. This module attaches part-of-speech and syntactic information to documents used as input for biotext mining, recognizes individual names such as gene names, and extracts information on primary interactions among them.

The interaction information between biological entities such as genes or proteins extracted in this way is transferred to the relation inference module, which is the next step, to obtain more extended interaction information.

1-1-1. Entity Name Recognition

Conceptually, entity name recognition can be thought of as two problems: (1) the recognition of the boundary of an entity name that separates the boundary of the entity name, and (2) the determination of the semantic class that determines the semantic class of the recognized entity name. Generally, these problems are considered classification problems and object recognition is recognized using methods such as Hidden Markov Model (HMM), Support Vector Machine (SVM), Maximum Entropy (ME) model, etc. The module is developed. However, due to the lack of learning corpus, the systems developed to date have not shown satisfactory performance. In particular, the problem of lack of learning materials tends to be more serious when one wants to solve the boundary recognition of class names and class decisions. This leads directly to performance, causing degradation in accuracy as well as processing efficiency. Therefore, the two tasks need to be separated and accessed. In addition, a method of improving recognition accuracy by using a terminology dictionary or a manually generated terminology recognition rule is also attempted.

For these two problems, this study uses a method that separates the boundary recognition of the entity name and the semantic class determination of the entity name, and combines the existing entity name dictionary and machine learning method. The machine learning method utilizes the SVM that has performed well in many existing classification problems, and selects the appropriate features for each task from parts of speech, spelling, and internal and external lexical context information to develop recognizers and semantic classifiers. .

In the case of detailed boundary recognition using SVM, one or more SVMs are attached using the one-vs-rest method by attaching the boundary recognition tag of the object name using TO (Term, Others) notation to the learning document, and considering T / O as the boundary class. Create a model and recognize boundaries. After the boundary of the entity name is recognized, the semantic classification of the entity name is performed targeting only the recognized entity names. At this time, 22 SVM classifiers are studied in one-vs-rest manner and 22 individual names of GENIA corpus are performed.

3 is a structural diagram showing an overview of entity name recognition.

However, developing object name recognition and classifiers through machine learning from small learning sets is prone to errors. In particular, in the case of recognizing individual names in two stages as in this study, errors in the boundary recognition stage have a fatal effect on the semantic classification stage. Therefore, in this study, we improve the performance by correcting the error of the boundary recognizer by using the already established entity name dictionary information.

1-1-2. Superficial parsing

If full parsing is performed to analyze the structure of the preprocessed sentences, the accuracy and the speed of analysis may be slow. Therefore, it is necessary to perform minimal parsing to extract only the information necessary for the next step, event analysis. The minimal parsing here refers to recognizing basic phrases in sentences and determining dependencies between them. Basic phrases are combined with other basic phrases to generate sentence structure through automatically learned syntax constraint rules and manually written constraint rules. The grammatical relationship between the basic phrases is also determined in this process.

The problem of structural analysis of sentences that cannot be solved by syntax constraint rules alone can be solved by using statistical information learned from large corpus. The constraint rules described above allow a significant degree of automatic acquisition from the Penn Treebank. In addition, for vocabularies that are frequently used in the field of biology, the rule can be manually added or modified in the automatically learned rule set for accurate analysis. Applying a constraint rule may result in a plurality of candidates for parsing. In this case, structural significance can be solved by using statistical information learned from corpus.

In this study, to use the feature-based statistical approach, we extract feature and statistical information from Penn Treebank, a large-scale learning material with syntax analysis. The base chunking results and statistical information determine the dependencies between the base phrases, and at the same time attach a grammatical function tag.

4 is a structural diagram showing an overview of surface syntax analysis.

At this time, WordNet, an English word ontology, is used to solve the lack of data generated by eliminating existing features or adding new features to create a set of suitable features and building a statistics-based module using a learning corpus. To utilize them.

1-1-3. Interaction Relationship Extraction

5 is a structural diagram showing an outline of interaction relationship extraction.

This step extracts the interactions from the surface parsing results. The basic method of extracting interaction relationships from parse results is to use manually written rules. In this case, it is necessary to construct a pattern for extracting the interaction relationship, which is built by a biological expert. However, there are limitations in building the necessary patterns manually, which requires the development of a method that can be automatically extended.

In this study, we create an interaction relationship extraction pattern using syntax analysis information to extract the interaction relationship that the user ultimately wants from biological documents. In other words, from the biological documents, we extract the eventual verbs that show interactions such as 'activate' or 'inhibit' among the verbs that appear at high frequency together with the gene or protein individual name, and then analyze the pattern and use the analyzed pattern information. You will create the interaction relationship extraction pattern manually. [Table 1] below shows the created pattern.

verb pattern activate  NP activate NP NP be activated by NP inhibit  NP inhibit NP NP be inhibit by NP associate  NP associate with NP NP associate with by NP bind  NP bind (to) NP NP binding NP

[Table 1] Examples of Patterns According to Verbs

However, there is a limit in manually extracting pattern information. A method for semi-automatically acquiring a pattern for extracting interaction relationships can be used. FIG. 5 shows a method for semi-automatically expanding the interaction relationship extraction pattern with the help of an expert. This method automatically extracts the pattern candidate for event extraction from the result of the event recognizer and extends the pattern by presenting and verifying the result of the event extraction using the extracted patterns to experts in the relevant field.

In addition, anaphora resolution improves the performance of the interaction relationship extraction and finally determines the format in which the interaction is expressed formally. It is about how to express the interactions between individuals in a biological document, which defines how to express static and dynamic relationships between individuals and should be verified to include the information that biologists actually need.

1-2. Relational Reasoning Module

6 is a structural diagram showing an outline of relationship inference and visualization steps.

In this module, inference and clustering of association rules using data mining algorithm and machine learning algorithm is inferring high-level interaction information from primary interaction information extracted from text analysis. This module supports the function of expressing such interaction information as network and generating hypothesis about expressed network.

The hypothesis generated in this module refers to a more generalized association rule that can comprehensively express interaction relationship extraction rules and inferred rules. This association rule is learned using Apriori-ppi, a data mining algorithm modified by combining Bayesian network and Naive Bayes Classifier. Then use SVM to determine the class of each interaction relationship. At this time, a hierarchical or non-hierarchical clustering algorithm is developed to develop a large classification model of the association rules, and define the classification of the association rules according to the large classification model. The clustering information and association rule information are used to determine the inference range of the interaction relationship and to perform high-level inference between biological entities such as protein-protein or gene-gene.

In this way, the entire interaction relationship information, including the relationships inferred using various algorithms, is organized and stored separately in the local DB, and then delivered to the network visualization module, which is a module for providing more efficient information transfer.

1-2-1. Association rule learning

It is a step that uses machine learning method to infer higher order relationship through analysis of interaction relationship. For this, we study the association rule inference machine using Bayesian network and simple Bayesian classifier. And we use SVM to develop a classification model for each interaction relationship group. However, machine learning algorithms such as SVM and Bayesian network are limited in performance improvement due to lack of learning materials. To complement this, data mining techniques such as IB clustering are used to discover associations between genes or protein groups, to discover interactions between genes and proteins, and to combine them with machine learning. Finally, in this step, high-level interaction relationships are extracted by combining association rule discovery using data mining techniques and machine learning clustering and inference techniques used in text mining.

1-2-2. Association Rule Clustering

In this stage, we develop the classification rules for association rules through hierarchical and non-hierarchical clustering, and develop a clustering model for inferring the correlation between the major classification groups. In addition, we develop a statistical hypothesis generation model of inference associations through the analysis of relationships extracted using Dynamic Bayesian Network (DBN) model.

1-3. Network visualization module

7 is a structural diagram illustrating an embodiment of an interface for visualization.

This module is a module that visualizes the interaction relationship between individuals such as proteins and genes extracted from the system. In this module, the weights of interactions between biological entities calculated using the association information extracted from the previous text analysis module and the relation inference module, and the link between the original document that performed the extraction or the frequency of association occurrences in the document, etc. Various information can be expressed in various ways so that the user can check the reliability of the association information provided by the system.

This module visualizes and extracts extracted relationships using graphs, diagrams, and network structures. It provides a unified interface for visualizing the interrelationships between extracted and inferred entities.

In addition, it provides the function to verify the reliability by searching and referencing the source and the relationship to the object and provides the function to present the supporting document for the request of the association information.

Biological research and genetic information are expected to explode in the future, and related research results will explode, which will increase the need for biotext data mining technology that can automatically summarize and summarize the research results of biological research literature. Therefore, biotext mining-related technologies are expected to increase in importance as an essential technology for the biotechnology industry to be active. Biotext mining technology is known to significantly accelerate the timing of acquiring specific results using bioinformatics technology.

First of all, the BT-related technologies are expected to grow rapidly by drastically reducing the time and financial costs of BT-related R & D such as new drug development based on biotext mining technology. The development of automatic extension of related resources such as GeneBank through recognition can be accelerated. In addition, it is possible to intelligentize the literature search service through clustering and real-time automatic classification of biotechnology-related literature sets retrieved from the user-supplied keywords, and the development of language processing technology using surface syntax analysis and codification resolution It is expected to contribute to revitalization of multilingual biotext document retrieval system and translation system research.

Although the present invention has been described above, the scope of the present invention is not limited to these embodiments, and the scope of the present invention may be easily modified by those skilled in the art.

According to the present invention, researchers in a BT-related field can systematically collect information on various biological terms, automatically extract relationships between individuals, and infer high-dimensional relationships using primarily extracted relationships. As a result, the extracted interaction relationship can be visualized for easy identification.

Claims (8)

In the method of deriving the interaction relationship between the biotexts included in the biological document, Assigning and storing accurate part-of-speech information to each biotext in the document, attaching a boundary recognition tag to the document using a TO notation, and considering the TO as a boundary class; Generating, recognizing a boundary, combining syntactically related texts into one basic phrase based on the part-of-speech information and the boundary of the entity name, and relying on a verb based on the combined state of the basic phrases Searching for other basic phrases that have a relationship, expressing sentence relations such as subjective and objective, and sentences containing pre-specified verbs in the document or sentences containing verbs used at a certain frequency in the document. Analyze the pattern by analyzing, and extracting the first interaction information by using the analyzed pattern information W, and the text analysis step of performing an analysis using the natural language processing technology to the bio-text; In order to extract the secondary interaction information through the analysis of the primary interaction information, the inference of the association rules and the inference of the superordinate relations are conducted and learned using data mining and machine learning algorithms, and the inferred association information and input data are inferred. Develop a clustering model for the inference of correlation between the classification rules and the classification of the association rules through hierarchical or non-hierarchical clustering, and inferring the analysis of the extracted relationships using the DBN model. Generating a statistical hypothesis of the association, extracting the secondary interaction information from the first interaction information extracted as a result of the text analysis, and expressing the second interaction information in a network, and generating a hypothesis for the network. Wow; Represent the interaction relationship using a visualization tool including a graph, a diagram, and a network structure, and represent the original document of the relationship extracted as a reliability confirmation element, and use the network structure between the biotexts to Visualization step of visually expressing the interaction relationship; method for extracting and inferencing interaction between biological objects in the biological text. delete delete delete The method of claim 1, Wherein said biotext is a biological term comprising a gene name and a protein name. The method of claim 1, The verb is A method for extracting and inferring interactions between biological objects in biological texts, including verbs that frequently appear in biological documents, including activate, inhibit, associate, and bind. A recording medium storing a program for executing a method of extracting and inferring interactions between biological objects in the biological text of claim 1. delete

Images