WO2025237327A1 - Gut microbe knowledge graph system - Google Patents

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

A Gut Microbiome Knowledge Graph System Technical Field

This invention relates to the fields of medicine and treatment, and specifically to a knowledge graph system and reasoning system in the field of gut microbiota.

Background Technology

The gut microbiota is the largest microbial community in the human body, with an adult carrying approximately 200 grams of gut microbes. These microbes contain about 30 times the genes of the human genome and could even be considered a vital organ in themselves. Gut microbes directly or indirectly influence other vital organs through pathways such as the gut-brain axis, gut-liver axis, and gut-lung axis. Therefore, the stability and dysregulation of the gut microbiota are closely related to human health and disease. They can help the human host enhance immunity, promote beneficial metabolism, and control weight gain, but they can also contribute to the development and progression of diseases such as autism, inflammatory bowel disease, and diabetes. In recent years, with the emergence and widespread use of high-throughput sequencing technology, the research threshold for the human microbiome has been significantly lowered. Multi-omics and clinical research on gut microbiota has become a very popular field, and many studies have conducted in-depth experimental verification and analysis of various associations. However, due to the high variability of gut microbiota, current findings may only be the tip of the iceberg; further research is needed to discover more about the specific functional characteristics and mechanisms of action of gut microbes in the human host.

To fully understand the crucial role of gut microbiota in human health and disease, researchers should pay close attention to biomedical big data, in addition to reliable but time-consuming experimental methods and efficient but limited microbiome approaches. With the frequent publication of cutting-edge literature, the proliferation of specialized databases, and the massive accumulation of clinical data, biomedicine has gradually shifted from subjective experience-driven "evidence-based medicine" to objective data-driven "precision medicine." A massive biomedical knowledge base has now formed, and effectively storing it and conducting in-depth knowledge mining and fusion reasoning will be a significant challenge. In recent years, knowledge graphs, as heterogeneous networks, have gradually replaced non-relational networks that cannot distinguish complex biological effects, and are increasingly used in the simulation of knowledge systems in the biomedical field. As a method of knowledge storage, knowledge graphs connect massive amounts of knowledge scattered across various carriers in the form of triples (head entity, relation, tail entity), attracting widespread attention from academia and industry.

However, the real world is constantly evolving, and existing knowledge graphs are almost always incomplete. Utilizing existing knowledge to infer new, reliable knowledge is crucial for improving the completeness and extensibility of knowledge. Inductive reasoning based on knowledge graph embedding has become mainstream in fields such as biomedicine. It maps entities and relationships to a low-dimensional space, significantly improving computational efficiency, effectively mitigating data sparsity, and enabling the fusion of heterogeneous information. Knowledge graph embedding can be further subdivided into translation models represented by RotatE, semantic models represented by HolE, and neural network models represented by ConvKB. Multimodal knowledge graph embedding is another major development direction, such as RSN combining relational paths, KEPLER combining entity descriptions, TKRL combining entity categories, and IKRL combining entity images. In addition, there are UKGE and UKGSE, which combine confidence scores for uncertain knowledge graph embedding. They have played an important role in downstream tasks such as drug reuse and disease gene association prediction.

In recent years, many teams have attempted to construct gut microbe knowledge graphs (GMKGs), such as Food4healthKG, KG4NH, MiKG4MD, and MGMLink, but no mature projects have yet been successfully implemented. While these works offer unique insights into the field, they also suffer from the following limitations to varying degrees:

(1) Most analyses only target one or two diseases, especially mental health diseases, which account for the majority, and their application scope is very limited. The scope of knowledge should be expanded, and it should no longer be limited to a single disease, but should be applied to multiple diseases as much as possible, because the knowledge systems of different diseases are sometimes interchangeable;

(2) Most of them only extract structured knowledge from professional databases or manually extract knowledge from a small number of relevant documents, resulting in a small amount of data, insufficient knowledge completeness, and inability to obtain convincing conclusions;

(3) After the construction work is completed, most of them only perform knowledge graph visualization and query based on graph databases such as Neo4j, and cannot obtain reliable new knowledge based on existing knowledge reasoning. Therefore, it is even more impossible to perform more complete and diverse multimodal reasoning based on annotation information other than triples;

(4) All constructed knowledge graphs are deterministic, with each piece of knowledge having only two possibilities: existence or non-existence. Confidence scores are not assigned based on actual circumstances, or the assignment of confidence scores is unreasonable. Therefore, it is even more impossible to conduct more reasonable uncertain reasoning based on confidence scores.

(5) Traditional knowledge graph reasoning models can only predict existing relationships, that is, the knowledge graph is incomplete and its prediction pattern is very fixed.

(6) Traditional knowledge graph reasoning models can only predict existing entities, which is hot-start reasoning, and its prediction range is relatively limited.

Existing technologies have reported some advancements in knowledge graph construction. For example, CN117292846A discloses a method and apparatus for constructing a gut microbiome knowledge graph. The method involves acquiring initial literature information related to gut microbiome research; extracting relevant entities, relationships between entities, and clinical annotation model information from the initial literature information; standardizing, organizing, and integrating the extracted entities to construct a data framework for the relevant entities; however, the gut microbiome knowledge graph constructed based on this data framework, relationships between entities, and clinical annotation model information does not include data cleaning or redundancy removal, nor does it construct a "multimodal uncertain reasoning system for knowledge graphs." It only extracts unidirectional relationships of information using a relatively simple extraction method, and the final result lacks further validation through a reasoning model.

CN117196028A discloses a method and system for producing medical knowledge graphs based on knowledge graphs. The main steps include constructing a graph production model, defining an entity-relationship-attribute framework for the knowledge graph, thereby connecting entities in the text with entities in the knowledge base to generate a triplet knowledge set for visualization. This patent establishes a triplet knowledge set, but it does not train or evaluate the generated data, nor does it calculate a function to obtain appropriate confidence scores. Therefore, it cannot train or evaluate the data, nor does it involve further constructing multimodal uncertain reasoning based on the triples.

CN114691896A discloses a method and apparatus for cleaning knowledge graph data. The method mainly includes training a knowledge graph embedding model and a triple classification model on the triples of the knowledge graph to be cleaned, and repairing erroneous triples using the global confidence score to obtain the cleaned knowledge graph. This patent only provides a commonly used data cleaning method for knowledge graphs and does not cover any specific application areas.

CN115080764A discloses a medical similar entity classification method and system based on knowledge graphs and clustering algorithms, solving the problem of tedious manual annotation for similar entity classification. This patent also primarily provides a training method for a medical database triplet dataset, classifying positive and negative samples to obtain an entity similarity classification model.

Summary of the Invention

Existing methods for constructing gut microbiome knowledge graphs often suffer from insufficient data cleaning, lack of multimodal feature fusion, and inability to perform reliable reasoning verification, thus limiting the accuracy and practicality of the knowledge graphs. This invention proposes a gut microbiome knowledge graph system that, through the construction of a high-quality knowledge base, dynamic confidence assessment, and multimodal uncertain reasoning, achieves accurate prediction and verification of the associations between gut microbiota and drugs, diseases, and other factors.

The first aspect of this invention relates to a method for constructing a knowledge graph.

1. Obtaining raw data for the "Gut Microbiome Knowledge Base" (bacterium-to-bacterium knowledge base)

The system integrates a gut microbiome knowledge base, obtains information from public databases, performs data cleaning and redundancy removal, and retains gut microbiome-related knowledge; it extracts entity tables and relation tables, and further includes the extraction of attribute tables; and it uses entity tables to uniformly match entities in relation tables and attribute tables to construct a complete "gut microbiome knowledge base".

Furthermore, entity tables and relational tables can be extracted from the entire database or a portion of the database, and attribute tables can be extracted from the entire database or a portion of the database after merging and deduplication.

The entity table refers to professional terms in the fields of gut microbiota, genetic engineering, protein engineering, enzyme engineering, and biochemistry, including but not limited to genes, proteins, RNA, small molecules, pathways, reactions, nucleotides, and variations. These terms can be expressed in various languages, such as Chinese, English, German, French, and Japanese. Specifically, each entity corresponds to at least one specialized term in the biomedical field, such as a proper noun; however, it is not necessarily a noun like "gene" or "protein." It can also be a verb, a verb-object phrase, such as "protein expression" or "gene modification," or a short sentence containing a few prepositions, such as "purify the protein," etc., but all of these can express a relatively complete meaning, though not sufficient to express the meaning of the entire sentence. Preferably, nouns are preferred entities; preferably, Chinese and/or English are preferred languages.

The relation table refers to the horizontal associations between the contents of each entity in the entity table, including but not limited to associations of therapeutic pathways, gene expression modes, chemical modifications, protein activation pathways, protein inhibition pathways, and protein auxiliary associations.

The attribute table refers to the inherent attributes, superordinate attributes, and subordinate attributes of each entity in the entity table, including but not limited to biological classification, text description and definition, genomic sequence, proteomics sequence, microbiological classification, 16S rRNA sequence, 168 rRNA sequence, etc.

The entities referred to here are the actual contents of entity tables, relation tables, and attribute tables, and not content that is a complete match of text. Based on the general understanding of text, entities are expressed in various ways, including translations of actual content in different languages, synonyms, near-synonyms, common aliases, general expressions, and changes in order that do not affect the understanding of meaning.

Specifically, the original databases can be BioCyc, EcoCyc, MetaCyc, BacDive, BV-BRC, EMBL-EBI, NCBI-Taxonomy, and MicrobeWiki, etc. The number of gut microbiome-related knowledge retained is not limited; if a limit is needed, it can be 100-500, 200-400, 200-300, 200, or 300 gut microbiome-related knowledge. For example, the eight databases can first be cleaned and deredundantized, retaining knowledge related to 200 gut microbiome-related knowledge. After processing these separately, the BioCyc, EcoCyc, MetaCyc, BacDive, and BV-BRC databases can be merged and deredundantized, extracting entity and relational tables. Simultaneously, the EMBL-EBI, NCBI-Taxonomy, and MicrobeWiki databases can be merged and deredundantized, extracting attribute tables. Finally, the entities in the entity table are uniformly matched with those in the relation table and attribute table to form a complete "Gut Microbiome Knowledge Base" (see Figure 2, where "Zhuowei Knowledge Base" is the specific name of the "Gut Microbiome Knowledge Base" described above).

The "Gut Microbiome Knowledge Base" has been reviewed by domain experts, therefore its confidence score is set at 1.0.

2. Construction of the "Gut Microbiota Small Molecule Drug Therapy Association Knowledge Base" (Microbe-Human Knowledge Base)

This process includes retrieving relevant databases, indexing documents, obtaining a small database after indexing, dividing it into training and validation sets, using a non-generative pre-trained language model for training and evaluation, obtaining a three-class classification model, obtaining the probabilities of the three relationships through a function, and calculating confidence scores.

Specifically, the databases linking gut microbiota to small molecule drug treatment for human diseases are retrieved, relevant literature is downloaded, and the literature is indexed using annotation tools, AI technology, and manual indexing (preferably using annotation tools, AI technology for automatic annotation, and manual review of annotation results). After indexing, a small database is obtained, which is divided into training and validation sets. A non-generative pre-trained language model is used for training and evaluation to obtain a three-class classification model. The probabilities of the three relationships are obtained through functions, and confidence scores are calculated to finally obtain the "Gut Microbiota Small Molecule Drug Treatment Association Knowledge Base".

The annotation refers to extracting associations between gut microbiome entities, small molecule entities, and disease entities based on entities already annotated on the annotation tool platform. Specifically, it can be determined whether an entity is a potential entity for relationship extraction based on its co-occurrence location, interval, and frequency of occurrence in the article. The co-occurrence location can be within the same paragraph (with a line break as the pause), sentence (with a period as the pause), paragraph type, or field (such as abstract, title, keywords, background technology, methods, discussion, etc.). The interval can be a specific number of characters, such as 1-10, 2-6, 2-8 characters, or more, ensuring a specific order for multiple entities or allowing for unrestricted order. The frequency of occurrence refers to the number of times an entity appears in the article, such as once, twice, three times, etc. For example, if an entity appears only once, it can be considered to have low relevance. It can appear in the entire text, in a specific paragraph type, or in combination, such as once in the abstract and more than five times in the entire text. Users can choose one of the above criteria for annotation or use a combination of criteria for comprehensive judgment.

The comprehensive confidence score is ;

The final layer of the relation extraction model is a three-class classification model. A SoftMax function is used to obtain the probability that the paragraph belongs to one of the three relations. This probability can be directly used as the confidence score for non-unrelated triples. Assume the triples... There are a total of n sources, where y _i, j_i, and s_{_i} represent the nth, j_{_i} , and s_i sources, respectively. Article ( The source article's publication year, journal impact factor, and confidence score given by the relation extraction model are used. The overall confidence score of this triple is defined as follows: Where k _{_n} , k_{_yi} , and k_{_ji} represent linear segmentation weighting coefficients based on n, _yi , and _ji, respectively. These weighting coefficients are monotonically increasing, all less than 1, and the inflection points of the linear segments need to be defined manually based on experience. Finally, triples predicted as unrelated are deleted, and non-unrelated triples are predicted as related. Triples with confidence scores greater than or equal to a certain value (e.g., greater than or equal to 0.5) are retained, resulting in quadruplets containing "head entity, relation, tail entity, and confidence score".

For example, the association between gut microbiota and disease can be defined as activate or inhibit. Examples of non-unrelated triples include (a bacterium, activate, a disease), while quadruples obtained after further screening include (a bacterium, activate, a disease, confidence score 0.8).

The "retrieval of databases linking gut microbiota with small molecule drug treatment for human diseases" can be a PubTator3 raw XML file containing PubMed abstracts and PMC full texts;

The annotation tool platform can be PubTator3, officially developed by NCBI, used to annotate entities and some relationships in the article. The annotated entities include small chemical molecules, diseases, genes, species, and variants, as well as some relationships. However, the annotated relationships do not include knowledge from the "Gut Microbiome Knowledge Base". Of course, PubTator3 is just an example. In fact, there are many existing tool platforms that can achieve this, such as GNormPlus, MetaMap, BERN2, etc., and one or more can be selected.

Specifically, the training set and the validation set are divided according to a specific ratio, such as 6:2, 7:1, 8:1, 9:2, 9:1, or 8:2.

The term "training set" refers to the labeled subset of data used during the training of a machine learning model to optimize model parameters; the term "validation set" refers to the independent subset of data used to evaluate model performance, which is used to adjust hyperparameters or select the best model to prevent overfitting, and its division ratio corresponds to that of the training set.

The establishment of the "Gut Microbiota Small Molecule Drug Therapy Association Knowledge Base" mentioned in Figure 3 is shown in Figure 3.

3. Construct a "Clinical Medicine Database" (human-to-human knowledge base)

The aforementioned clinical medical database utilizes molecular biology data provided by authoritative clinical medical databases, such as the PMapp database, to offer knowledge related to disease diagnosis, medication guidelines, and disease prevention, thus providing knowledge support for the gut microbiota knowledge graph (see Figure 1, where GMKG-200 represents the "gut microbiota knowledge graph").

Furthermore, the database integrates multiple authoritative molecular biology and clinical medicine databases. Given the database's authority and the fact that it has been reviewed by field experts, a confidence score of 1.0 is defined. There is no limit to the number of databases; for example, it can be 50-100, or 55-90, such as 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, etc.

4. Construction of Gut Microbiome Knowledge Graph

The three knowledge bases, namely the "Gut Microbiome Knowledge Base," the "Gut Microbiome Small Molecule Drug Therapy Association Knowledge Base," and the "Clinical Medical Database," are combined, constructed, and connected to align small molecules, drugs, and disease entities, thereby obtaining the final gut microbiome knowledge graph.

The second aspect of this invention relates to a knowledge graph multimodal uncertainty reasoning system.

By utilizing the knowledge graph multimodal uncertain reasoning system of the "Gut Microbiome Knowledge Graph", potential associated diseases, drugs, genes, etc. can be predicted for gut microbiota.

First, each of the head entity, tail entity, and relation has a randomly initialized embedding vector, and each also has an embedding vector based on multimodal annotation, which is then normalized based on the representation matrix.

Secondly, 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;

Finally, the final vector representation of the triples is substituted into the basic knowledge graph embedding model and transformation function to generate the predicted confidence score, which is then compared with the true confidence score to define the loss function.

The term "hybrid expert system" refers to a machine learning architecture that integrates multi-source data through a dynamic gating network. This system deploys multiple expert networks (sub-models) in parallel, each processing specific types of data features (such as randomly initialized vectors and multimodal annotation vectors). It automatically learns the weight allocation of different expert networks through a trainable gating mechanism, and finally fuses the outputs of all expert networks 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 multimodal feature vectors to optimize the vectorized representation of entity relationships in knowledge graphs.

Specifically, the initialization vector is , and The embedding vector is and The extraction methods for multimodal feature vectors vary depending on the entity type. The vector normalization process can be performed on the intestinal microbial entity based on the gene expression matrix, or it can be performed on the entity based on the ontology category through multi-class vector average pooling, or it can be performed using encoding software.

For example: gene entities are processed by multi-class vector average pooling based on gene ontology categories; disease entities are encoded using PubMedBERT or BioLinkBERT based on text descriptions; protein entities are encoded using ProteinBERT based on amino acid sequences; and small molecule and drug entities are encoded using ChemBERT based on SMILES formulas.

The term "average pooling" represents a feature dimensionality reduction method, specifically referring to the arithmetic average of multi-class feature vectors in structured classification systems such as Gene Ontology (GO): feature vectors (e.g., GO term vectors) corresponding to multiple ontology categories belonging to the same gene entity are summed element-wise and then divided by the number of categories to generate a single feature vector with overall representativeness. In this invention, this method is used to fuse multi-dimensional ontology annotation information of gene entities, constructing a normalized gene representation vector by retaining the statistical features of each classification dimension, thus eliminating redundant information while maintaining the integrity of the ontology semantics.

Specifically, using the head entity vector For example: ,in , It is a relation-based weight vector, and it is also randomly initialized. (Tail entity vector) The calculation is similar. For an entity that does not exist in the "Gut Microbiome Knowledge Graph" (GMKG-200), simply... and Set as the zero vector. or That is, equal to zero. or Will be completely based on or Therefore, as long as multimodal annotations are extracted for any entity, this model can achieve cold-start inference.

Specifically, the final vector representation of the triples can be... Substitute the basic knowledge graph embedding model and transformation function Generate confidence scores for predictions The loss function is defined by comparing the loss function with the true confidence score.

For example, TransE represents the distance after translation in planar space: RotatE represents the distance after rotation in complex space: ,in Represents the element-wise product of vectors. It indicates the connection between the real and imaginary parts of a complex number.

In the Gut Microbiome Knowledge Graph (GMKG-200), gut microbiota are directly linked to diseases and drugs, and predicting new potential associations is equivalent to performing a simple knowledge graph completion task. However, gut microbiota are not directly linked to human genes; they can only be indirectly linked through multiple steps via diseases, drugs, and small molecules, which traditional knowledge graph link prediction methods cannot achieve. By using confidence scores to homogenize existing and new inferred relationships, and only needing to comprehensively calculate the expected probability, confidence score estimation and association ranking of long-chain relationships can be effectively achieved. (See Figure 5)

The third aspect of this invention relates to a gut microbiome knowledge graph system.

The gut microbiome knowledge graph system includes a "gut microbiome knowledge base", a "gut microbiome small molecule drug therapy association knowledge base", and a "clinical medical database".

Furthermore, the "Gut Microbiome Knowledge Base" contains information on the knowledge relationships between bacteria obtained after extraction and processing from the original database, and the "Gut Microbiome Small Molecule Drug Therapy Association Knowledge Base" contains information on gut microbiome and small molecule drug therapy for human diseases obtained after extraction and processing from the original database.

Furthermore, the "intestinal microbiome knowledge base" is constructed using the method for obtaining raw data of the "intestinal microbiome knowledge base" as described in the first aspect of the present invention, the "intestinal microbiome small molecule drug therapy association knowledge base" is constructed using the construction method of the "intestinal microbiome small molecule drug therapy association knowledge base" as described in the first aspect of the present invention, and the "clinical medical database" is constructed using the "construction of clinical medical database" method described in the first aspect of the present invention.

Furthermore, the gut microbiome knowledge graph system also includes a knowledge graph multimodal uncertain reasoning system, used to predict potential associated diseases, drugs, genes, etc. for gut microbiota.

Furthermore, the gut microbiome knowledge graph system also includes an image database that can be used for querying, providing visualized query results, and enabling open-source data sharing.

Image databases can be open-source graph databases, commercial/cloud graph databases, or RDF/semantic web graph databases. All of these databases enable visual queries and open-source data sharing. The specific choice depends on factors such as data scale, query complexity, and deployment environment. Open-source graph databases include ArangoDB, Neo4j, and JanusGraph; commercial/cloud graph databases include Amazon Neptune and Microsoft Azure Cosmos DB; and RDF/semantic web graph databases include Virtuoso and Stardog.

The fourth aspect of this invention relates to a knowledge graph system construction apparatus.

The knowledge graph system construction apparatus includes:

The data acquisition module is used to acquire raw data related to gut microbiota, specifically, to acquire raw database information for the "Gut Microbiota Knowledge Base", "Gut Microbiota Small Molecule Drug Therapy Related Knowledge Base", and "Clinical Medical Database".

The data processing module includes cleaning and deduplicating data, retaining target-related knowledge, and extracting entity tables, relation tables, and attribute tables; or using annotation tool platforms, AI technology, and manual indexing to index the acquired data.

The data training module is used to divide the labeled data into training and validation sets, train and evaluate it using a non-generative pre-trained language model, obtain a three-class classification model, obtain the probabilities of the three relationships through a function, and calculate the confidence score.

The data matching module is used to uniformly match entities in the entity table with entities in the relationship table and attribute table; and to obtain matched data after data processing based on confidence scores.

Furthermore, it also includes a data reasoning module, which is used to import the acquired data into the knowledge graph multimodal uncertain reasoning system to predict potential associated diseases, drugs, genes, etc. for gut microbiota.

Specifically, it includes a vector normalization processing module. Each of the head entity, tail entity, and relation has a randomly initialized embedding vector, as well as an embedding vector based on multimodal annotation. This module performs vector normalization processing based on the expression matrix.

Specifically, it also includes a hybrid expert system module, which uses a dynamic gating network of a hybrid expert system to fuse the two embedding vectors of each entity to obtain its final vector representation;

Specifically, it also includes a comparison module, which is used to input the final vector representation of the triples into the basic knowledge graph embedding model and transformation function to generate the predicted confidence score, and compare it with the true confidence score to define the loss function.

Furthermore, the knowledge graph system construction device may also include a display module, which can be used for querying, providing visualized query results, and enabling open-source data sharing.

The fifth aspect of this invention relates to computer programs and storage media.

A computer program for implementing the gut microbiota knowledge graph construction program designed in the first to fourth aspects of the present invention;

A computer storage medium can store the gut microbiome knowledge graph construction program designed according to the first to fourth aspects of the present invention, as well as the gut microbiome knowledge graph established thereon.

In summary, this invention constructs a gut microbiome knowledge graph (GMKG) with broad knowledge coverage (multi-topic), high knowledge completeness (multi-source), rich knowledge modalities (multi-modal), and confidence scores. Based on a knowledge graph embedding model, a multi-modal uncertainty reasoning system is established, utilizing multi-modal annotations and confidence scores to achieve cold-start and long-chain reasoning, improving the prediction range and interpretability of traditional reasoning models. This not only helps gut microbiome researchers summarize existing authoritative and cutting-edge knowledge (e.g., constructing a Neo4j graph database for convenient querying and searching), but also serves as a large-scale screening tool to discover new research directions from a forward-looking perspective for subsequent in-depth experimental verification (publishing confidence scores and ranking lists of new reasoning knowledge). Furthermore, it will enable more accurate and rapid prevention and treatment of various diseases in clinical decision-making (publishing trained entity and relation embedding vectors as pre-training vectors for downstream clinical tasks).

Beneficial effects

This invention addresses two major challenges in current biomedical reasoning: cold-start and long-chain reasoning. It achieves cold-start reasoning by extracting multimodal annotations of biomedical entities and utilizing a hybrid expert system to supplement the knowledge gaps in triples, thus enhancing predictive ability for small and zero-sample scenarios. Long-chain reasoning is achieved by calculating the comprehensive confidence score of relational paths using confidence scores. The resulting relational paths improve the interpretability of biomedical knowledge, acting similarly to pathways. This invention overcomes four main shortcomings: narrow knowledge coverage, low knowledge completeness, fixed knowledge structure, and simplistic knowledge reasoning.

Attached Figure Description

Figure 1 is an overview of the gut microbiota knowledge graph system.

Figure 2 shows the data processing of the "Gut Microbiome Knowledge Base".

Figure 3 shows the establishment of the "Gut Microbiota Small Molecule Drug Therapy Association Knowledge Base".

Figure 4 shows the construction of a unified knowledge base for small molecules, drugs, and disease entities.

Figure 5 shows a knowledge graph multimodal uncertainty reasoning system.

Detailed Implementation

For 200 common gut microbiota species in the human body and commonly used in detection, a gut microbiota knowledge graph was constructed and a multimodal uncertainty reasoning system was built and operated.

Building a Gut Microbiome Knowledge Base

A large-scale comprehensive gut microbiome knowledge base, named "Zhuowei Knowledge Base," was constructed by integrating eight biological categories, molecular pathways, and multi-omics public databases. The Zhuowei Knowledge Base fully downloads and integrates eight original databases: BioCyc, EcoCyc, MetaCyc, BacDive, BV-BRC, EMBL-EBI, NCBI-Taxonomy, and MicrobeWiki. First, each of the eight databases underwent data cleaning and redundancy removal, retaining only knowledge related to 200 gut microorganisms. After individual processing, the BioCyc, EcoCyc, MetaCyc, BacDive, and BV-BRC databases were merged and redundancy removed, extracting entity and relational tables. Simultaneously, the EMBL-EBI, NCBI-Taxonomy, and MicrobeWiki databases were merged and redundancy removed, extracting attribute tables. Finally, the entity tables were uniformly matched with entities in the relational and attribute tables to form the complete gut microbiome knowledge base (Figure 1 shows the Zhuowei Knowledge Base as an example).

Construction of a "Knowledge Base for Small Molecule Drug Therapy Related to Gut Microbiota"

Existing databases contain very little information about gut microbiota interactions and their associations with small molecules and diseases. Downloading the full PubTator3 raw XML file, which includes PubMed abstracts and the full text of the PMC, can supplement this information. PubTator3 is an annotation tool platform officially developed by NCBI, which annotates Chemical, Disease, Gene, Species, and Variant entities and some relationships in articles. However, the annotated relationships do not include Species entities, meaning there is a lack of knowledge related to gut microbiota.

First, based on the entities already labeled in PubTator3, if a gut microbiome entity appears in the same paragraph as another gut microbiome entity, small molecule entity, or disease entity, it is considered a potential entity pair for relation extraction. Then, 2000 paragraphs from each of the three categories that meet the criteria are randomly selected for relation labeling. Each labeling uses three types of relations: besides "Unrelated," the relations between gut microbes are "Symbiotic" or "Exclusion," between gut microbes and small molecules are "Secrete" or "Consume," and between gut microbes and diseases are "Activate" or "Inhibit."

Two rounds of automatic annotation were performed using GPT-4. Paragraphs with inconsistent results were then manually reviewed to obtain the final annotation results. When using GPT-4 for automatic annotation, prompt words were added before paragraphs to assist in text generation, and entity annotation tokens were added before and after the entity context for marking (gut microbial 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$).

Then, the labeled small dataset is divided into training and validation sets in an 8:2 ratio, and fine-tuned and evaluated using non-generative pre-trained language models such as PubMedBERT or BioLinkBERT. The last layer of this relation extraction model is a three-class classification model, which uses a SoftMax function to obtain the probability that the paragraph belongs to one of the three relations. This probability can be directly used as the confidence score for non-unrelated triples. Assuming triples... There are a total of n sources, where y _i, j_i, and s_{_i} represent the nth, j_{_i} , and s_i sources, respectively. Article ( The source article's publication year, journal impact factor, and confidence score given by the relation extraction model are used. The overall confidence score of this triple is defined as follows: , where k _{_n} , k_{_yi} , and k_{_ji} represent linear piecewise weighting coefficients based on n, _yi , and j _, respectively. These weighting coefficients are monotonically increasing, all less than 1, and the inflection points of the linear pieces need to be defined manually based on experience.

(3) Constructing a gut microbiome knowledge graph (GMKG-200)

A multimodal, uncertain gut microbiome knowledge graph was constructed based on three parts: microbe-microbe, microbe-human, and human-human, and named "GMKG-200".

The knowledge about bacteria is provided by the Zhuowei Knowledge Base, which is structured and extracted from eight authoritative databases. It has already passed the manual review of researchers in the early stage, so the confidence score is directly defined as 1.0.

The human-to-human knowledge base, namely the "clinical medical database," is provided by the PMapp database, which integrates 61 authoritative molecular biology and clinical medical databases previously built by the inventors. The confidence score is also defined as 1.0, which can provide complete knowledge support for the gut microbiome knowledge graph (GMKG-200).

Knowledge about gut microbiota and humans is automatically extracted from PubTator3, which introduces some extraction error. A comprehensive confidence score is calculated by combining information from all sources during relation extraction, effectively measuring this uncertainty. This section includes three relation types defined during annotation: (Gut Microbiota, Symbiotic or Exclusion, Gut Microbiota), (Gut Microbiota, Secrete or Consume, Small Molecules), and (Gut Microbiota, Activate or Inhibit, Diseases). PubTator3 connects the Zhuowei Knowledge Base and the PMapp Knowledge Base; simply aligning small molecule, drug, and disease entities allows for the final construction of the gut microbiota knowledge graph (GMKG-200).

(4) Knowledge graph multimodal uncertain reasoning system

A knowledge graph multimodal uncertainty reasoning system is constructed based on triples of the gut microbiome knowledge graph (GMKG-200) to predict potential associated diseases, drugs, and genes for gut microbiota.

First, the head entity, tail entity, and relation each have a randomly initialized embedding vector. , and In addition, the head entity and the tail entity each have an embedding vector based on multimodal annotations. and The methods for extracting multimodal feature vectors differ for different entity types. Gut microbial entities are processed using vector normalization based on gene expression matrices; gene entities are processed using multi-class vector average pooling based on gene ontology categories; disease entities are encoded using PubMedBERT or BioLinkBERT based on text descriptions; protein entities are encoded using ProteinBERT based on amino acid sequences; and small molecule and drug entities are encoded using ChemBERT based on SMILES.

Secondly, 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, with the head entity vector as the first example. For example: ,in , It is a relation-based weight vector, and it is also randomly initialized. (Tail entity vector) The calculation is similar. For an entity that does not exist in GMKG-200, simply... and Set as the zero vector. or It equals zero. or Will be completely based on or Therefore, as long as multimodal annotations are extracted for any entity, this model can achieve cold-start inference.

Finally, the final vector representation of the triples. Substitute the basic knowledge graph embedding model and transformation function Generate confidence scores for predictions The loss function is defined by comparing the result with the true confidence score. For example, TransE represents the distance translated in planar space: RotatE represents the distance after rotation in complex space: ,in Represents the element-wise product of vectors. It indicates the connection between the real and imaginary parts of a complex number.

(5) Image database, which can be used for querying, providing visual query results, and enabling open-source sharing of data.

Furthermore, the Neo4j graph database can be used to achieve the above functions.

In the Gut Microbiome Knowledge Graph (GMKG-200), gut microbes are directly linked to diseases and drugs, making the prediction of new potential associations akin to a simple knowledge graph completion task. However, gut microbes are not directly linked to human genes; associations are only established indirectly through multiple steps via diseases, drugs, and small molecules, which traditional knowledge graph link prediction methods cannot achieve. By using confidence scores to homogenize existing and new inferred relationships, and by comprehensively calculating probability expectations, confidence score estimation and association ranking of long-chain relationships can be effectively realized.

Claims (12)

A gut microbiome knowledge graph system, characterized in that it comprises a gut microbiome knowledge base, a gut microbiome small molecule drug therapy association knowledge base, and a gut microbiome knowledge graph constructed from a clinical medical database, as well as a knowledge graph multimodal uncertain reasoning system utilizing the gut microbiome knowledge graph; The construction of the gut microbiota small molecule drug therapy association knowledge base includes searching relevant databases, indexing literature, obtaining a small database after indexing, dividing it into training and validation sets, training and evaluating it using a non-generative pre-trained language model, obtaining a three-class classification model, obtaining the probabilities of the three relationships through a function, and calculating the confidence score; the comprehensive confidence score is... Furthermore, the last layer of the relation extraction model is a three-class classification model. After passing through a SoftMax function, the probability of the paragraph belonging to the three relations is obtained, and this probability is directly used as the confidence score of the triple. The knowledge graph multimodal uncertain reasoning system is constructed through the following steps: First, each of the head entity, tail entity, and relation has a randomly initialized embedding vector and an embedding vector based on multimodal annotation. Vector normalization 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. Finally, the final vector representation of the triple is fed into the basic knowledge graph embedding model and transformation function to generate the predicted confidence score, which is compared with the true confidence score to define the loss function. According to claim 1, the construction of the gut microbiome knowledge graph system includes obtaining information from a public database, performing data cleaning and redundancy removal, and retaining gut microbiome-related knowledge; extracting entity tables and relation tables, further including extracting attribute tables; and constructing a complete "gut microbiome knowledge base" by uniformly matching entities in the relation tables and attribute tables with the entity tables. According to the gut microbiome knowledge graph system of claim 2, entity tables and relation tables are extracted from the entire database/or a portion of the database is selected for extraction. Attribute tables are extracted from the entire database/or a portion of the database is merged and redundancy removed. The annotation refers to extracting the associations between gut microbiome entities and small molecule entities and disease entities based on entities already annotated by the annotation tool platform; and determining whether an entity is a potential entity based on its location, interval, and frequency of occurrence in the article for relation extraction. According to the gut microbiome knowledge graph system of claim 1, the gut microbiome small molecule drug therapy association knowledge base assumes a triplet. There are a total of n sources, where y _i, j_i, and s_{_i} represent the nth, j_{_i} , and s_i sources, respectively. The publication year of the source article, the journal's impact factor, and the confidence score given by the relation extraction model; the overall confidence score of this triple is defined as follows: , where k _{_n} , k _{_yi} , and k_{_ji} represent linear piecewise weighted coefficients based on n, _yi , and _ji , respectively, and the weighted coefficients are monotonically increasing. According to claim 1 or 4, the training set and validation set are divided in a ratio of 6:2, 7:1, 8:1, 9:2, 9:1, or 8:2. The confidence score for the gut microbiome knowledge base and the clinical medical database is set to 1.0. According to claim 1, in the gut microbiome knowledge graph system, 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. According to the gut microbiome knowledge graph system of claim 6, the 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. According to the gut microbiome knowledge graph system of claim 7, the final vector representation of the triples is... Substitute the basic knowledge graph embedding model and transformation function Generate confidence scores for predictions The loss function is defined by comparing the loss function with the true confidence score. According to the gut microbiome knowledge graph system of claim 8, TransE represents the distance after translation in planar space: RotatE represents the distance after rotation in complex space: ,in Represents the element-wise product of vectors. It indicates the connection between the real and imaginary parts of a complex number. A device for constructing a gut microbiome knowledge graph includes: The data acquisition module is used to acquire raw data related to gut microbiota, specifically, to acquire raw database information for the "Gut Microbiota Knowledge Base", "Gut Microbiota Small Molecule Drug Therapy Related Knowledge Base", and "Clinical Medical Database". The data processing module includes cleaning and deduplicating data, retaining target-related knowledge, and extracting entity tables, relation tables, and attribute tables; or using annotation tool platforms, AI technology, and manual indexing to index the acquired data. The data training module is used to divide the labeled data into training and validation sets, train and evaluate it using a non-generative pre-trained language model, obtain a three-class classification model, obtain the probabilities of the three relationships through a function, and calculate the confidence score. The data matching module is used to uniformly match entities in the entity table with entities in the relationship table and attribute table; and to obtain matched data after data processing based on confidence scores. The data reasoning module is used to import the acquired data into the knowledge graph multimodal uncertain reasoning system to predict potential related diseases, drugs, genes, etc. for gut microbiota. A computer program for loading and constructing the gut microbiome knowledge graph as described in claims 1-9. A computer storage medium for carrying the computer program of claim 11.