Gut microbe knowledge graph system

Abstract

A database structure obtained by means of information retrieval, and a reasoning system, which structure and system specifically relate to a gut microbe knowledge graph system, comprising: a gut microbe knowledge graph consisting of a gut microbe knowledge base, a gut microbe and small-molecule drug therapy association knowledge base, and a clinical medicine database; and a multimodal uncertainty reasoning system, using the gut microbe knowledge graph. The gut microbe knowledge graph system predicts potential diseases, drugs, genes, etc., which are associated with gut microbes.

Description

Intestinal microorganism knowledge graph system TECHNICAL FIELD

[0001] The present application relates to the fields of medicine, treatment, etc., in particular, to an intestinal microorganism field knowledge graph system and a reasoning system. BACKGROUND

[0002] The intestinal flora is the largest microbial community in the human body, and an adult human body carries about 200 grams of intestinal microorganisms (Gut Microbe), which contain about 30 times the genes of the human genome, and can even be considered as an important organ alone. The intestinal microorganism directly or indirectly acts on other important organs of the human body through the intestinal-brain axis, intestinal-liver axis and intestinal-lung axis, etc. Therefore, the stability and disorder of the intestinal microorganism are closely related to the health and diseases of human beings. They can help the human host to enhance the immune capacity, promote beneficial metabolism and control weight gain, etc., and also can lead to the occurrence and development of diseases such as autism, inflammatory bowel disease and diabetes. In recent years, with the emergence and popularization of high-throughput sequencing technology, the research threshold of human microbiome has been significantly reduced, and the multi-omics and clinical research of intestinal microorganism have become a very popular field at present, and many studies have carried out in-depth experimental verification and analysis on various correlations. However, due to the high variability of intestinal microorganism, the current existing findings may only be the tip of the iceberg, and more specific functional characteristics and action mechanisms of intestinal microorganism on human host still need further research and discovery.

[0003] In order to fully understand the important influence of intestinal microorganism on human health and diseases, in addition to reliable but time-consuming and laborious experimental methods and efficient but with great limitations of microbiome methods, researchers should also pay more attention to biomedical big data. With the high frequency of publication of frontier literature, the blooming of professional databases and the massive stacking of clinical data, biomedical has gradually shifted from the "evidence-based medicine" driven by subjective experience to the "precision medicine" driven by objective data. At present, a huge biomedical knowledge body has been formed, and how to effectively store them and carry out deep mining and fusion reasoning of knowledge will be a great challenge. In recent years, as a heterogeneous network, the knowledge graph (Knowledge Graph) gradually replaces the non-relational network which cannot distinguish complex biological effects, and is more and more applied to the knowledge system simulation in the field of biomedical. As a storage method of knowledge, the knowledge graph connects the mass of knowledge scattered in various carriers in the form of triplets (head entity, relationship, tail entity), and has received extensive attention from the academic and industrial circles.

[0004] However, the real world is changing rapidly, and the built knowledge graph is almost incomplete. It is very important to use existing knowledge to infer new reliable knowledge to improve the completeness and extension of knowledge. Among them, inductive reasoning based on knowledge graph embedding has become the mainstream in the field of biomedicine, which maps entities and relationships to a low-dimensional space, significantly improving computational efficiency, effectively alleviating data sparsity, and enabling the fusion of heterogeneous information. Knowledge graph embedding can be divided into translation models represented by RotatE, semantic models represented by HolE, and neural network models represented by ConvKB. Multi-modal knowledge graph embedding is another major development direction, such as RSN combining relationship paths, KEPLER combining entity descriptions, TKRL combining entity categories, and IKRL combining entity images. In addition, there are UKGE and UKGSE combining confidence scores for uncertain knowledge graph embedding. They play an important role in downstream tasks such as drug reutilization and disease-gene association prediction.

[0005] In recent years, many teams have tried to build some gut microbe knowledge graphs (GMKGs), such as Food4healthKG, KG4NH, MiKG4MD, and MGMLink, but there is no mature project in application. Looking at these works, although they have some enlightenment for the development of this field, they more or less have the following shortcomings:

[0006] (1) Most of them only analyze one or two diseases, especially mental health diseases, which have a very limited application range. The knowledge coverage should be expanded, not just for a single disease, but as much as possible for multiple diseases, because the knowledge systems of different diseases are sometimes interconnected;

[0007] (2) Most of them only extract structured knowledge from professional databases or manually read and extract knowledge from a small number of related literature, resulting in a small amount of data and insufficient completeness of knowledge, which cannot lead to reliable conclusions;

[0008] (3) After completing the construction work, most of them only visualize and query the knowledge graph based on graph databases such as Neo4j, but cannot infer reliable new knowledge based on existing knowledge. Therefore, it is also impossible to perform more complete and multi-modal reasoning based on annotation information other than triples;

[0009] (4) All constructed are deterministic knowledge graphs, a piece of knowledge only has two cases of existence or non-existence, and it is not reasonable to assign a confidence score or set a confidence score based on the real situation. Therefore, it is also impossible to perform more reasonable uncertain reasoning based on confidence scores;

[0010] (5) Traditional knowledge graph reasoning models can only predict existing relationships, that is, the knowledge graph is incomplete, and the prediction mode is very fixed;

[0011] (6) Traditional knowledge graph reasoning models can only predict existing entities, that is, hot start reasoning, and the prediction range is limited.

[0012] There are some reports on knowledge graph construction in the prior art. For example, CN117292846A relates to a method and device for constructing an intestinal microorganism knowledge graph, obtaining initial literature information related to intestinal microorganism research; performing information extraction processing on the initial literature information, extracting related entities, associated information between related entities, and clinical annotation model information, and performing standardization, arrangement and integration processing on the extracted related entities to construct a data framework of related entities; based on the data framework of related entities, the associated information between related entities and the clinical annotation model information, an intestinal microorganism knowledge graph is constructed. The obtained data is not cleaned and de-duplicated, and a "knowledge graph multi-modal uncertain reasoning system" is not constructed. Only one-way association of information is extracted, the extraction method is relatively simple, and the final result is not further verified by a reasoning model.

[0013] CN117196028A discloses a medical knowledge graph production method and system based on a knowledge graph, which defines the main steps as constructing a graph production model, defining an entity-relation-attribute framework of graph knowledge, connecting entities in text with entities in a knowledge base, generating a triple knowledge set for visual display. A triple knowledge set is established in this patent, but the generated data is not trained and evaluated, and a suitable confidence score is not obtained by a calculation function, so the data cannot be trained and evaluated, and further construction of a multi-modal uncertain reasoning of the triple is not involved.

[0014] CN114691896A discloses a knowledge graph data cleaning method and device, mainly including training a knowledge graph embedding model and a triple classification model for the triple of the knowledge graph to be cleaned, repairing the incorrect triple through the global confidence to obtain the cleaned knowledge graph. This patent only provides a general data cleaning method for a knowledge graph, and does not involve specific application fields.

[0015] CN115080764A discloses a medical similar entity classification method and system based on a knowledge graph and a clustering algorithm, which solves the problem of tedious manual annotation of similar entity classification. This patent also mainly provides a training method for a medical database triple data set, classifies positive samples and negative samples, and obtains an entity similarity classification model. SUMMARY

[0016] In the prior art, the construction method of the intestinal microorganism knowledge graph mostly has problems of insufficient data cleaning, lack of multi-modal feature fusion, and inability to perform reliable reasoning verification, resulting in limited accuracy and practicality of the knowledge graph. The present application proposes an intestinal microorganism knowledge graph system, which realizes accurate prediction and verification of the correlation between intestinal microorganisms and drugs, diseases, etc. by constructing a high-quality knowledge base, dynamic confidence evaluation, and multi-modal uncertain reasoning.

[0017] The first aspect of the present application relates to a knowledge graph construction method

[0018] 1 "Intestinal microorganism knowledge base" original data acquisition (bacteria-bacteria knowledge base)

[0019] Integrate the intestinal microorganism knowledge base, obtain information from public databases, perform data cleaning and redundancy removal, and retain intestinal microorganism-related knowledge; extract entity tables and relationship tables, further including extracting attribute tables; match entities in the entity table with the relationship table and the attribute table to construct a complete "intestinal microorganism knowledge base".

[0020] Further, the entity table and the relationship table can be extracted from all databases or selected from a part of the databases, and the attribute table can be extracted from all databases or selected from a part of the databases after merging and redundancy removal.

[0021] The entity table contains professional terms in the field of intestinal microorganisms, genetic engineering, protein engineering, enzyme engineering, and biochemistry, including but not limited to genes, proteins, RNA, small molecules, pathways, reactions, nucleotides, and variations. Various languages can be used for expression, such as Chinese, English, German, French, Japanese, etc. Specifically, each entity corresponds to at least one specialized term in the biomedical field, such as a specialized noun; but it is not necessarily a "gene" or "protein" term. It can also be a verb, a verb-object phrase, or a short sentence containing a few prepositions, such as "protein expression" or "gene modification", or a short sentence containing a few prepositions, such as "purifying the protein", etc. However, it can express a relatively complete meaning, but not enough to express a complete sentence. As a preferred embodiment, nouns are preferred entities; as a preferred embodiment, Chinese and / or English are preferred languages.

[0022] The relationship table represents the horizontal association between the contents of each entity in the entity table, including but not limited to the association of therapeutic pathways, the association of gene expression methods, the association of chemical modifications, the activation pathway of proteins, the inhibition pathway of proteins, and the auxiliary association of proteins.

[0023] The attribute table refers to the inherent attributes, superior attributes, subordinate attributes, etc. of each entity in the entity table, including but not limited to biological classification, text description and definition, genomics sequence, proteomics sequence, microbiology classification, 16S rRNA sequence, 168 rRNA sequence, etc.

[0024] The entity refers to the actual content in the entity table, relationship table and attribute table, rather than the content that matches the text completely. Based on the general understanding of the text, the entity contains various language translations, synonyms, near synonyms, common aliases, general expressions, and order changes that do not affect the understanding of the meaning.

[0025] Specifically, the original database can be BioCyc, EcoCyc, MetaCyc, BacDive, BV-BRC, EMBL-EBI, NCBI-Taxonomy and MicrobeWiki, etc. Specifically, the retained knowledge of intestinal microorganisms is not limited to the number of intestinal microorganisms. If necessary, it can be 100-500, 200-400, 200-300, 200, 300 kinds of intestinal microorganism-related knowledge. For example, first, clean and de-duplicate 8 databases, retain 200 kinds of intestinal microorganism-related knowledge, and process them separately. After that, combine and de-duplicate BioCyc, EcoCyc, MetaCyc, BacDive and BV-BRC, and extract the entity table and relationship table from them. At the same time, combine and de-duplicate EMBL-EBI, NCBI-Taxonomy and MicrobeWiki, and extract the attribute table from them. Finally, match the entities in the entity table, relationship table and attribute table to form a complete "intestinal microorganism knowledge base" (see Figure 2, "Zhuo micro knowledge base" in Figure 2 is the specific name of the "intestinal microorganism knowledge base" described above)

[0026] The "intestinal microorganism knowledge base" has been reviewed by experts in the field, so the confidence score is set to 1.0.

[0027] 2 Construction of "Intestinal Microorganism Small Molecule Drug Treatment Association Knowledge Base" (Bacteria-Human Knowledge Base)

[0028] It includes searching related databases, indexing literature, obtaining a small database after indexing, dividing it into training set and validation set, training and evaluating using non-generated pre-training language model, obtaining three classification models and obtaining the probability of three relationships through function, and obtaining confidence score through calculation.

[0029] Specifically, a database of intestinal microorganisms associated with small molecule drugs for treating human diseases is retrieved, associated literature is downloaded, and the literature is indexed. The indexing is performed using an annotation tool platform, AI technology, and manual indexing, and the results are further annotated using an annotation tool platform, automatically annotated using AI technology, and manually reviewed. After indexing, a small database is obtained, which is divided into a training set and a validation set. A non-generated pre-trained language model is used for training and evaluation to obtain a three-classification model and the probabilities of the three relationships through a function. The confidence score is calculated, and the "intestinal microorganism-small molecule drug treatment association knowledge base" is finally obtained.

[0030] The annotation refers to extracting the association between intestinal microorganism entities and small molecule entities, disease entities based on the entities annotated by the annotation tool platform. Specifically, the potential entities can be determined based on the position, interval, and frequency of occurrence of the entities in the article. The same position can be in the same paragraph (with a carriage return as a pause), sentence (with a period as a pause), paragraph type, or field (such as abstract, title, keywords, background technology, method, discussion, etc.). The interval can be a specific number of characters, such as 1-10, 2-6, or 2-8 characters, or more characters, which can ensure a specific order or no limit order for multiple entities. The frequency of occurrence refers to the frequency of occurrence of entities in the article, such as 1, 2, 3, etc. For example, if it appears only once, it can be determined to have a low correlation, which can be in the full text, in a specific paragraph type, or in a combination, such as appearing once in the abstract and 5 times or more in the full text. One of the above determination criteria can be selected for annotation, or multiple determination criteria can be used for comprehensive judgment.

[0031] The comprehensive confidence score is ;

[0032] The last layer of the relationship extraction model is a three-classification model, which obtains the probability of the paragraph belonging to the three relationships through a SoftMax function. This probability can be directly used as the confidence score of the non-Unrelated triple. Suppose the triple has n literature sources, y i , j i , and s i represent the publication year, journal impact factor, and confidence score given by the relationship extraction model of the i th literature source, respectively. The comprehensive confidence score of the triple is defined as , where k n , k yi , and k ji represent the weights based on n, yi and j i linear piecewise weighting coefficients, which are monotonically increasing, all less than 1 and need to be artificially defined by experience; finally, the triplets predicted as Unrelated are deleted, the non-Unrelated triplets are kept, and the triplets with confidence scores greater than or equal to a certain value (for example, greater than or equal to 0.5) are kept, to obtain four-tuples containing "head entity, relationship, tail entity, confidence score".

[0033] For example, for the association between intestinal microorganisms and diseases, the non-Unrelated triplets can be defined as activate or inhibit, such as (certain bacteria, activate, certain disease), and the four-tuples obtained after further screening, such as (certain bacteria, activate, certain disease, confidence score 0.8).

[0034] The "database for searching the association between intestinal microorganisms and small molecule drugs for treating human diseases" can be a PubTator3 original xml file containing PubMed abstracts and PMC full texts;

[0035] The annotation tool platform can be PubTator3 developed by NCBI as a tool platform for annotating entities and part of the relationships in the article; the annotated entities include chemical small molecules, Disease diseases, Gene genes, Species species, and Variant variants, and other related entities and part of the relationships; but the annotated relationships do not contain the relevant knowledge of the "intestinal microorganism knowledge base"; of course, PubTator3 is only an example, and in fact there are multiple tool platforms that can be implemented in the prior art, such as GNormPlus, MetaMap, BERN2, etc., which can be selected one or more.

[0036] Specifically, the training set and the validation set are divided according to a certain ratio, for example, 6:2, 7:1, 8:1, 9:2, 9:1, and 8:2.

[0037] The term "training set" represents a subset of annotated data used in the machine learning model training process to optimize model parameters; the term "validation set" represents an independent subset of data used to evaluate model performance, which is used to prevent overfitting by adjusting hyperparameters or selecting the best model, and the division ratio corresponds to the training set.

[0038] The establishment of the "intestinal microorganism small molecule drug treatment association knowledge base" of FIG. 3 is shown in FIG. 3.

[0039] 3. Constructing a "clinical medical database" (human-human knowledge base)

[0040] The clinical medicine database adopts molecular biology to provide data from a clinical medicine authority database, such as a PMapp database, to provide disease diagnosis, medication guidelines, disease prevention, and other related knowledge, and to provide knowledge support for the intestinal microorganism knowledge graph (see FIG. 1, where GMKG-200 shown in FIG. 1 is the "intestinal microorganism knowledge graph")

[0041] Further, the database integrates multiple molecular biology and clinical medicine authority databases, and since the database is authoritative and has been reviewed by experts in the field, the confidence score is defined as 1.0. The number of databases does not need to be limited, for example, it can be 50-100, or 55-90, for example, it can be 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, and the like.

[0042] 4 Construction of intestinal microorganism knowledge graph

[0043] The "intestinal microorganism knowledge base", "intestinal microorganism small molecule drug treatment correlation knowledge base", and "clinical medicine database" are combined to construct a docking and alignment of small molecules, drugs, and disease entities, and an intestinal microorganism knowledge graph is obtained.

[0044] The second aspect of the present application relates to a knowledge graph multi-modal uncertainty reasoning system

[0045] The knowledge graph multi-modal uncertainty reasoning system of the "intestinal microorganism knowledge graph" is used to predict potential associated diseases, drugs, genes, and the like for intestinal microorganisms.

[0046] First, the head entity, tail entity, and relationship each have a randomly initialized embedding vector, and also each have an embedding vector based on multi-modal annotation, and the vector is normalized based on the expression matrix;

[0047] Second, a dynamic gating network of a hybrid expert system is used to fuse the two embedding vectors of each entity to obtain its final vector representation;

[0048] Finally, the final vector representation of the triple is brought into the basic knowledge graph embedding model and the conversion function to generate the predicted confidence score, which is compared with the true confidence score to define the loss function.

[0049] The term "mixed expert system" represents a machine learning architecture that integrates multi-source data through a dynamic gating network. The system deploys multiple expert networks (sub-models) in parallel, each handling a specific type of data feature (such as randomly initialized vectors and multi-modal annotation vectors). Through a trainable gating mechanism, the system automatically learns the weight distribution of different expert networks and finally fuses the outputs of each expert network to generate a unified vector representation of entities. In this invention, this module is specifically used to coordinate the fusion of initial embedding vectors and multi-modal feature vectors to optimize the vectorization of entity relationships in the knowledge graph.

[0050] Specifically, the initialization vector is , and The embedding vector is and The extraction method of multi-modal feature vectors for different entity types is different. The vector normalization process can be a vector normalization process based on the gene expression matrix for intestinal microorganism entities, a multi-class vector average pooling process based on ontology categories for entities, or encoding processing using encoding software.

[0051] For example: gene entities perform multi-class vector average pooling based on gene ontology categories, disease entities perform encoding processing using PubMedBERT or BioLinkBERT based on text descriptions, protein entities perform encoding processing using ProteinBERT based on amino acid sequences, and small molecule and drug entities perform encoding processing using ChemBERTa based on SMILES.

[0052] The term "average pooling" represents a feature dimensionality reduction method, specifically an arithmetic average operation on multi-class feature vectors in structured classification systems such as Gene Ontology (GO). The feature vectors corresponding to multiple ontology categories of the same gene entity (such as GO term vectors) are added element by element and divided by the number of categories to generate a single feature vector that represents the whole. In this invention, this method is used to fuse the multi-dimensional ontology annotation information of gene entities, and a normalized gene representation vector is constructed by retaining the statistical characteristics of each classification dimension, which eliminates redundant information and maintains the integrity of ontology semantics.

[0053] Specifically, taking the head entity vector as an example: , where , is a weight vector based on the relationship, which is also randomly initialized. The calculation of the tail entity vector is the same. For an entity that does not exist in the "Gut Microorganism Knowledge Graph" (GMKG-200), only And is set to zero vector, Or is equal to zero, Or will be completely based on Or generated. Therefore, as long as any entity extracts multimodal annotations, this model can realize cold start reasoning.

[0054] Specifically, the final vector representation of the triple is substituted into the basic knowledge graph embedding model and the conversion function to generate a predicted confidence score , and compare it with the real confidence score to define the loss function.

[0055] For example, TransE represents the distance after translation in the plane space: , RotatE represents the distance after rotation in the complex space: , wherein represents the vector element product, represents the connection of the real part and the imaginary part of the complex number.

[0056] In the "gut microbial knowledge graph (GMKG-200)", gut microbes are directly connected with diseases and drugs, and predicting new potential associations is equivalent to performing a simple knowledge graph completion task. However, gut microbes have no direct association with human genes, and can only be indirectly associated through diseases, drugs, and small molecules in multiple steps, and traditional knowledge graph link prediction cannot be achieved. Homogenization of existing and new reasoning relationships based on confidence scores only needs to integrate the probability expectation to effectively estimate the confidence score and correlation degree of long-chain relationships. (See Figure 5)

[0057] The third aspect of the present application relates to a gut microbial knowledge graph system

[0058] The gut microbial knowledge graph system includes a "gut microbial knowledge base", a "gut microbial small molecule drug treatment association knowledge base", and a "clinical medical database";

[0059] Further, the "gut microbial knowledge base" contains knowledge association information between bacteria and bacteria obtained after extraction and processing from the original database, and the "gut microbial small molecule drug treatment association knowledge base" contains information about the relationship between gut microbes and small molecule drug treatment of human diseases obtained after extraction and processing from the original database.

[0060] Further, the "gut microbiota knowledge base" is obtained by the method of the "gut microbiota knowledge base" raw data in the first aspect of the application, the "gut microbiota small molecule drug treatment correlation knowledge base" is obtained by the construction method of the "gut microbiota small molecule drug treatment correlation knowledge base" in the first aspect of the application, and the "clinical medical database" is constructed by the construction method of the "clinical medical database" in the first aspect of the application.

[0061] Further, the gut microbiota knowledge graph system further comprises a knowledge graph multi-modal uncertain reasoning system for predicting potential associated diseases, drugs, genes and the like of gut microbiota.

[0062] Further, the gut microbiota knowledge graph system further comprises an image database, which can be used for query, providing visual query results, and data open source sharing.

[0063] The image database can be an open source graph database, or a commercial / cloud graph database, or an RDF / semantic web graph database, and the database can realize visual query and data open source sharing. The specific selection depends on factors such as data size, query complexity, deployment environment, etc. Open source graph databases such as ArangoDB, Neo4j, JanusGraph, commercial / cloud graph databases such as Amazon Neptune, Microsoft Azure Cosmos DB, and RDF / semantic web graph databases such as Virtuoso, Stardog, etc.

[0064] The fourth aspect of the application relates to a knowledge graph system construction device

[0065] The knowledge graph system construction device comprises:

[0066] The data acquisition module is used for acquiring raw data related to gut microbiota, and specifically, the raw database information acquisition is performed to obtain the "gut microbiota knowledge base", the "gut microbiota small molecule drug treatment correlation knowledge base", and the "clinical medical database";

[0067] The data processing module comprises cleaning, de-duplicating, retaining target related knowledge, extracting entity tables, relationship tables, and attribute tables, or using a labeling tool platform, AI technology, and manual indexing to index the acquired data;

[0068] The data training module is used for dividing the labeled data into a training set and a validation set, training and evaluating using a non-generative pre-training language model, obtaining a three-classification model and the probability of three relationships through a function, and calculating a confidence score;

[0069] The data matching module is used for matching entities in the entity table unified matching relationship table and the attribute table; and matched data is obtained through data processing by a confidence score.

[0070] Further, the data reasoning module is used for importing the obtained data into a knowledge graph multi-modal uncertain reasoning system to predict potential associated diseases, drugs, genes and the like of intestinal microorganisms.

[0071] Specifically, the vector normalization processing module is used for normalizing vectors based on expression matrices, and each of the head entity, the tail entity and the relationship has a randomly initialized embedding vector and a multi-modal annotation based embedding vector.

[0072] Specifically, the hybrid expert system module is used for adopting a dynamic gating network of a hybrid expert system to fuse the two embedding vectors of each entity to obtain a final vector representation of the entity.

[0073] Specifically, the comparison module is used for bringing the final vector representation of the triple into a basic knowledge graph embedding model and a conversion function to generate a predicted confidence score, comparing the predicted confidence score with a real confidence score, and defining a loss function.

[0074] Further, the knowledge graph system construction device can further include a display module, which can be used for querying, providing visual query results, and realizing open source sharing of data.

[0075] The fifth aspect of the present application relates to a computer program and a storage medium

[0076] A computer program is used for implementing the intestinal microorganism knowledge graph construction program designed in the first to fourth aspects of the present application.

[0077] A computer storage medium can store the intestinal microorganism knowledge graph construction program designed in the first to fourth aspects of the present application and the established intestinal microorganism knowledge graph.

[0078] In summary, the present application constructs an intestinal microbial knowledge graph (GMKG) with a wide coverage of knowledge (multi-theme), high completeness of knowledge (multi-source), rich modalities of knowledge (multi-modal), and a confidence score. Based on the knowledge graph embedding model, a knowledge graph multi-modal uncertainty reasoning system is established, which uses multi-modal annotation and confidence score to realize cold start and long chain reasoning, and improve the prediction range and interpretability of traditional reasoning model. Not only can it help intestinal microbial researchers to summarize existing authoritative and cutting-edge knowledge (such as constructing Neo4j graph database for convenient query and search), but also can be used as a large-scale screening tool to discover new research directions from a prospective perspective and then conduct further experimental verification (publish the confidence score and ranking list of new reasoning knowledge), and can realize more accurate and fast prevention and treatment of various diseases in clinical decision-making (publish the entity and relationship embedding vectors trained as pre-training vectors for downstream clinical tasks).

[0079] Advantages

[0080] The present application solves two major difficulties of current biomedical reasoning: cold start and long chain reasoning. The application uses multi-modal annotation of biomedical entities to realize cold start reasoning by supplementing the knowledge loss of triples with a hybrid expert system, and enhances the prediction ability for small samples and zero samples. The comprehensive confidence score of the relationship path is calculated by the confidence score to realize long chain reasoning, which will improve the explainability of biomedical knowledge and play a similar role to the pathway. It overcomes the four main defects of narrow knowledge coverage, low knowledge completeness, fixed knowledge structure and simple knowledge reasoning. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 is a system overview of the intestinal microbial knowledge graph.

[0082] Figure 2 is the processing data processing of "intestinal microbial knowledge base".

[0083] Figure 3 is the establishment of "intestinal microbial small molecule drug treatment related knowledge base".

[0084] Figure 4 is the construction of small molecule, drug and disease entity unified knowledge base.

[0085] Figure 5 is a knowledge graph multi-modal uncertainty reasoning system. DETAILED DESCRIPTION

[0086] For 200 kinds of common and commonly used intestinal microorganisms in human body, the construction and operation of the intestinal microbial knowledge graph and the multi-modal uncertainty reasoning system are carried out.

[0087] Construction of intestinal microbial knowledge base

[0088] Integrating 8 biological categories, molecular pathways and multi-omics public databases, a large comprehensive intestinal microorganism knowledge base was constructed and named "Zhuo Micro Knowledge Base". Zhuo Micro Knowledge Base will download and integrate BioCyc, EcoCyc, MetaCyc, BacDive, BV-BRC, EMBL-EBI, NCBI-Taxonomy and MicrobeWiki these 8 original databases. First, the data cleaning and redundancy removal of the 8 databases were carried out separately, and only the knowledge related to 200 intestinal microorganisms was retained. After separate processing, the BioCyc, EcoCyc, MetaCyc, BacDive and BV-BRC databases were merged and de-duplicated, and the entity table and relationship table were extracted from them. At the same time, the EMBL-EBI, NCBI-Taxonomy and MicrobeWiki databases were merged and de-duplicated, and the attribute table was extracted from them. Finally, the entity table was used to match the entities in the relationship table and the attribute table to form a complete intestinal microorganism knowledge base (shown as Zhuo Micro Knowledge Base in Figure 1 as an example).

[0089] Construction of "Intestinal Microorganism Small Molecule Drug Treatment Association Knowledge Base"

[0090] There is little knowledge about the interaction of intestinal microorganisms, and their association with small molecules and diseases in existing databases. The PubTator3 original xml file containing PubMed abstracts and PMC full text was downloaded completely to supplement these relationships. PubTator3 is an annotation tool platform developed by NCBI, which annotates Chemical, Disease, Gene, Species and Variant entities and some relationships in the article, but the annotated relationships do not contain Species entities, that is, there is no knowledge about intestinal microorganisms.

[0091] First, according to the annotated entities of PubTator3, if an intestinal microorganism entity appears in the same paragraph as another intestinal microorganism entity or a small molecule entity or a disease entity, it will be considered as a set of potential entity pairs for relationship extraction. Then, 2000 paragraphs of each of the three types of paragraphs that meet the conditions were randomly selected for relationship annotation, each of which was annotated as three types of relationships, in addition to Unrelated, between intestinal microorganisms and intestinal microorganisms Symbiotic or Exclusion, between small molecules Secrete or Consume, and between diseases Activate or Inhibit.

[0092] Two rounds of automatic annotation were performed using GPT-4, and the final annotation results were obtained by manual review for paragraphs with inconsistent results. When using GPT-4 for automatic annotation, a prompt word is added before the paragraph to assist text generation, and entity annotation tokens are added before and after the entity context (intestinal microorganism entities are @gm-1$, @ / gm-1$, @gm-2$, and @ / gm-2$, small molecules are @sm-1$ and @ / sm-1$, and diseases are @di-1$ and @ / di-1$).

[0093] Then, the annotated small dataset was divided into training and validation sets in an 8:2 ratio, and non-generative pre-trained language models such as PubMedBERT or BioLinkBERT were used for fine-tuning training and evaluation. The last layer of the relationship extraction model is a three-classification model, which obtains the probability of the paragraph belonging to three relationships through a SoftMax function. This probability can be directly used as the confidence score of the non-Unrelated triplets. Assuming that the triplet n i , y i , and s i represent the publication year, journal impact factor, and confidence score given by the relationship extraction model, respectively. of the th n , k yi , and k ji represent the linear segmentation-based weighting coefficients based on n, y i , and j i , respectively. The weighting coefficients are monotonically increasing, and they are all less than 1. The turning points of linear segmentation need to be defined manually and empirically.

[0094] (3) Construction of the Intestinal Microorganism Knowledge Graph (GMKG-200)

[0095] A multi-modal uncertain intestinal microorganism knowledge graph was constructed based on the three parts of bacteria-bacteria, bacteria-human, and human-human, and was named "GMKG-200".

[0096] The bacteria-bacteria knowledge was provided by the Zhiwu knowledge base, which was structured from 8 authoritative databases and had been manually reviewed by researchers in the previous study, so the confidence score was directly defined as 1.0.

[0097] ​The human-human knowledge base, i.e., the "clinical medicine database", is provided by the PMapp database previously constructed by the inventors, which is a fusion of 61 authoritative molecular biology and clinical medicine databases, and the confidence score is also defined as 1.0, which can provide complete knowledge support for the intestinal microorganism knowledge graph (GMKG-200).

[0098] The microbe-human knowledge is automatically extracted from PubTator3, and there is a certain extraction error. When extracting the relationship, the comprehensive confidence score is calculated by combining all source information, which can well measure this uncertainty. This part contains 3 types of relationship types defined when annotating the relationship: (intestinal microorganism, Symbiotic or Exclusion, intestinal microorganism), (intestinal microorganism, Secrete or Consume, small molecule), and (intestinal microorganism, Activate or Inhibit, disease). PubTator3 connects the Zhuo Wei knowledge base and the PMapp knowledge base, and only by aligning the small molecule, drug and disease entities can the intestinal microorganism knowledge graph (GMKG-200) be finally constructed.

[0099] (4) Knowledge graph multi-modal uncertain reasoning system

[0100] Based on the intestinal microorganism knowledge graph (GMKG-200), a knowledge graph multi-modal uncertain reasoning system is constructed to predict potential associated diseases, drugs and genes for intestinal microorganisms.

[0101] First, the head entity, tail entity and relationship each have a randomly initialized embedding vector , and . In addition, the head entity and the tail entity each have an embedding vector based on multi-modal annotation and , and the extraction methods of multi-modal feature vectors for different entity types are different. The intestinal microorganism entity is based on the gene expression matrix for vector normalization processing, the gene entity is based on the gene ontology category for multi-class vector average pooling processing, the disease entity is based on the text description and uses PubMedBERT or BioLinkBERT for encoding processing, the protein entity is based on the amino acid sequence and uses ProteinBERT for encoding processing, and the small molecule and drug entity is based on SMILES and uses ChemBERTa for encoding processing.

[0102] Second, a dynamic gating network of a hybrid expert system is used to fuse the two embedding vectors of each entity to obtain its final vector representation. Take the head entity vector as an example: , where , is a relation-based weight vector, which is also randomly initialized. The tail entity vector is calculated in the same way. For an entity that does not exist in GMKG-200, set and to zero vector, or is equal to zero, or will be completely based on or generation. Therefore, as long as the multi-modal annotation of any entity is extracted, this model can realize cold start reasoning.

[0103] Finally, the final vector representation of the triple is substituted into the base knowledge graph embedding model and the conversion function to generate the predicted confidence score and compare it with the real confidence score to define the loss function. For example, TransE represents the distance after translation in the plane space: , RotatE represents the distance after rotation in the complex space: , where represents the vector element product, represents the connection of the real part and the imaginary part of the complex number.

[0104] (5) Image database, which can be used for query, provide visual query results, and data open source sharing.

[0105] Further, Neo4j graph database can be used to realize the above functions.

[0106] In the gut microbiota knowledge graph (GMKG-200), gut microbiota is directly connected with diseases and drugs, and predicting new potential associations is equivalent to performing a simple knowledge graph completion task. However, gut microbiota has no direct association with human genes and can only be indirectly associated through diseases, drugs, and small molecules in multiple steps, which cannot be achieved by traditional knowledge graph link prediction. Homogenization of existing and new reasoning relationships based on confidence scores only requires the comprehensive calculation of probability expectations to effectively estimate the confidence scores and correlation degrees of long-chain relationships.

Claims

1. An intestinal microorganism knowledge graph system, characterized in that, The intestinal microorganism knowledge graph comprises an intestinal microorganism knowledge base, an intestinal microorganism small molecule drug treatment correlation knowledge base, and a clinical medicine database, and a knowledge graph multi-modal uncertain reasoning system using the intestinal microorganism knowledge graph; The construction of the intestinal microorganism small molecule drug treatment correlation knowledge base comprises searching a related database, indexing literature, obtaining a small database after indexing, dividing into a training set and a verification set, training and evaluating by using a non-generated pre-training language model, obtaining a three-classification model and obtaining the probabilities of three relationships by a function, and obtaining a confidence score through calculation; the comprehensive confidence score is Further, the last layer of the relationship extraction model is a three-classification model, and a SoftMax function is used to obtain the probability that the paragraph belongs to three kinds of relationships, and the probability is directly used as the confidence score of the triple; The knowledge graph multi-modal uncertain reasoning system is constructed through the following steps: First, the head entity, tail entity and relationship each have a randomly initialized embedding vector, and each also has an embedding vector based on multi-modal annotation, and vector normalization processing is performed based on the expression matrix; second, a dynamic gating network of a hybrid expert system is used to fuse the two embedding vectors of each entity to obtain its final vector representation; and finally, the final vector representation of the triple is brought into a basic knowledge graph embedding model and a conversion function to generate a predicted confidence score, which is compared with the real confidence score to define a loss function.

2. The intestinal microorganism knowledge graph system of claim 1, wherein the construction of the intestinal microorganism knowledge base comprises obtaining information from public databases, performing data cleaning and redundancy removal, and retaining intestinal microorganism related knowledge; extracting entity tables and relationship tables, and further comprising extracting attribute tables; and constructing a complete "intestinal microorganism knowledge base" by uniformly matching entities in the entity tables with entities in the relationship tables and the attribute tables.

3. The intestinal microorganism knowledge graph system of claim 2, wherein the entity tables and the relationship tables are extracted from all databases or selected from a part of the databases, and the attribute tables are extracted by merging and removing redundancies from all databases or selected databases.

4. The gut microbiota knowledge graph system of claim 1, the gut microbiota small molecule drug treatment association knowledge base, hypothetical triplets n total number of literature sources, y i , j i and s i respectively represent the publication year, journal impact factor and confidence score given by the relationship extraction model of the nth source article; define the comprehensive confidence score of the triplet as , wherein k n , k yi and k ji respectively represent the linear piecewise weighting coefficients based on n, y i and j i , the weighting coefficients are monotonically increasing.

5. The intestinal microorganism knowledge graph system of claim 1 or 4, wherein the training set and the validation set are divided in a ratio of 6:2, 7:1, 8:1, 9:2, 9:1, or 8:2, and the confidence score of the intestinal microorganism knowledge base and the clinical medicine database is set to 1.

0.

6. The gut microbiome knowledge graph system according to claim 1, wherein in the knowledge graph multimodal uncertain reasoning system: the initialization vector is... , and The embedding vector is and The extraction methods for multimodal feature vectors vary depending on the entity type. Vector normalization is the normalization of gut microbial entities based on the gene expression matrix, while vector normalization refers to the average pooling of multi-class vectors based on the ontology category, or encoding processing using encoding software.

7. The gut microbiome knowledge graph system according to claim 6, head entity vector : ,in , It is a relation-based weight vector, randomly initialized; tail entity vector. The calculation is similar; for an entity that does not exist in the "Gut Microbiome Knowledge Graph" (GMKG-200), it is necessary to... and Set as the zero vector. or That is, equal to zero. or Will be completely based on or generate.

8. The gut microbiome knowledge graph system of claim 7, generating a final vector representation of the triplets into the base knowledge graph embedding model and a transformation function a confidence score for the prediction and comparing to the true confidence score defines a loss function.

9. The gut microbiome knowledge graph system of claim 8, TransE represents a distance after translation on a plane space: , RotatE represents a distance after rotation on a complex space: , where represents a vector element product, represents a connection of a real part and an imaginary part of a complex number.

10. An intestinal microorganism knowledge graph construction device, comprising: a data acquisition module configured to acquire original data related to intestinal microorganisms, specifically, to acquire original database information for constructing an "intestinal microorganism knowledge base", an "intestinal microorganism small molecule drug treatment correlation knowledge base", and a "clinical medicine database"; a data processing module configured to clean and remove redundancies from the data, retain target related knowledge, extract entity tables, relationship tables, and attribute tables, or use an annotation tool platform, AI technology, and manual indexing to index the acquired data; and a data storage module configured to store the cleaned and indexed data. The data training module is used for dividing the labeled data into a training set and a verification set, training and evaluating by using a non-generative pre-training language model, obtaining a three-classification model, and obtaining probabilities of three relationships through a function, and obtaining a confidence score through calculation; The data matching module is used for matching entities in the entity table, the relationship table and the attribute table, and obtaining matched data through data processing according to the confidence score. The data reasoning module is used for importing the obtained data into a knowledge graph multi-modal uncertain reasoning system, and predicting potential associated diseases, drugs, genes and the like of intestinal microorganisms.

11. A computer program for loading the intestinal microorganism knowledge graph according to claims 1-9.

12. A computer storage medium for carrying the computer program according to claim 11.

Images