Microorganism data semantic processing method and system based on graph model

Abstract

The invention discloses a microbial data semantic processing method and system based on a graph model, and relates to the technical field of knowledge graph and microbial information processing, and the method comprises the steps: obtaining microbial data from a heterogeneous data source, and carrying out the standardization processing; constructing a domain ontology model conforming to microbial taxonomy specifications; implementing entity disambiguation and unified identifier mapping by utilizing a pre-training language model; a semantic relation triple is extracted in a mode of combining remote supervision and deep learning; constructing a microorganism knowledge attribute graph and storing the graph in a graph database; knowledge reasoning and function prediction are carried out by using a graph attention network; the method supports natural language semantic retrieval, and solves the problems that heterogeneous microorganism data integration is difficult, the naming disambiguation precision is insufficient and the function association mining capacity is limited.

Description

Technical Field

[0001] This invention relates to the fields of knowledge graph and microbial information processing technology, specifically to a method and system for semantic processing of microbial data based on graph models. Background Technology

[0002] With the rapid development of high-throughput sequencing technology and microbiome research, a massive amount of microbial-related data resources have accumulated globally. These data resources are scattered across various types of databases and information systems, including strain information databases of microbial preservation institutions, sequence databases of genome sequencing centers, academic literature databases, and experimental record systems of various laboratories. However, due to historical reasons and differences in technical standards, these heterogeneous data sources suffer from severe data silos, hindering the in-depth integration and utilization of microbial data and the ability to discover knowledge.

[0003] Chinese patent CN118152578A discloses a method for constructing a knowledge graph. This method extracts structured data from a relational database and maps it to an ontology model, then combines named entity recognition and relation classification models to process unstructured text, thereby generating and updating knowledge graph triples. However, this method faces the following technical problems when applied to the field of microbiology: First, microbial nomenclature suffers from severe synonymy and synonymy issues. The same strain may use different naming forms in different data sources, including Latin scientific names, common names, and accession numbers. Existing entity recognition methods struggle to accurately establish equivalence relationships between different naming forms. Second, the microbial taxonomy is constantly being revised. The reclassification of genus and species-level taxonomic units leads to difficulties in mapping historical data to the current taxonomic system. Existing ontology model construction methods lack the ability to support the dynamic evolution of taxonomy. Third, the functional attributes and ecological characteristics of strains are often implicit in literature descriptions and experimental data. Existing relation extraction methods lack sufficient accuracy in recognizing complex semantic relationships unique to the microbial field, such as strain-phenotype associations and strain-habitat associations.

[0004] Therefore, there is an urgent need for a knowledge graph construction method that can effectively integrate heterogeneous microbial data, solve the problem of disambiguation in microbial naming, support the dynamic construction of microbial taxonomy ontology, and deeply mine the functional associations of strains, so as to improve the semantic level and knowledge service capabilities of microbial data resources. Summary of the Invention

[0005] The purpose of this invention is to provide a semantic processing method and system for microbial data based on graph models, so as to solve the technical problems of difficulty in integrating heterogeneous microbial data, insufficient accuracy of microbial naming disambiguation, and limited ability to mine functional associations of strains in the prior art.

[0006] To achieve the above objectives, this invention provides a semantic processing method for microbial data based on a graph model, comprising the following steps: a data acquisition and preprocessing step, acquiring heterogeneous microbial data from microbial strain databases, genome sequence databases, microbiology literature databases, and experimental record systems; performing format unification conversion, missing value imputation, and noise data filtering operations on the heterogeneous microbial data to generate a standardized microbial dataset; an ontology model construction step, constructing a microbial domain ontology model according to microbial taxonomy standards; the microbial domain ontology model includes a taxonomic hierarchical structure from the strain level through the species level, genus level to the family level, and an attribute relationship framework covering gene attributes, metabolite attributes, culture condition attributes, physiological characteristic attributes, and application function attributes; an entity recognition and linking step, extracting a candidate entity set from the standardized microbial dataset using a named entity recognition algorithm; addressing synonyms and polysemy in the candidate entity set; calculating entity disambiguation vector representations based on the microbial domain ontology model; and mapping each entity in the candidate entity set to the corresponding concept node in the microbial domain ontology model to generate a unified identifier mapping result; and a linking step. The system comprises the following steps: **Semantic Extraction Step:** Based on the unified identifier mapping results and standardized microbial datasets, a deep learning relation extraction model is used to identify semantic associations between strains and genes, strains and metabolites, strains and habitats, and strains and functions. This generates a set of semantic relation triples, and the confidence score of each triple in the set is calculated. **Knowledge Graph Construction Step:** The unified identifier mapping results are used as a node set, and the set of semantic relation triples is used as an edge set to construct a microbial knowledge attribute graph structure, which is then stored in a graph database storage format. **Knowledge Reasoning Step:** Based on the microbial knowledge attribute graph structure, graph neural network reasoning operations are performed to calculate the functional similarity between strains and predict the potential functional attributes of unlabeled strains, generating knowledge reasoning result data. This data includes strain function prediction data and metabolic pathway association prediction data. **Semantic Retrieval Step:** Natural language query requests input by users are received, converted into graph query statements, and graph traversal retrieval operations are performed on the microbial knowledge attribute graph structure. The semantic retrieval response results are then generated by combining the knowledge reasoning result data.

[0007] This invention also provides a graph-based semantic processing system for microbial data, comprising: a data acquisition module for acquiring heterogeneous microbial data from strain preservation databases, genome sequence databases, microbiology literature databases, and experimental record systems; performing format unification conversion, missing value imputation, and noise data filtering on the heterogeneous microbial data to generate a standardized microbial dataset; an ontology construction module for constructing a microbial domain ontology model according to microbial taxonomy standards; the microbial domain ontology model includes a taxonomic hierarchy from strain level through species level, genus level to family level, and an attribute relationship framework covering gene attributes, metabolite attributes, culture condition attributes, physiological characteristic attributes, and application function attributes; an entity recognition module for extracting a candidate entity set from the standardized microbial dataset using a named entity recognition algorithm; calculating entity disambiguation vector representations based on the microbial domain ontology model to address synonyms and polysemy in the candidate entity set; mapping each entity in the candidate entity set to the corresponding concept node in the microbial domain ontology model; and a relation extraction module for... The system utilizes a unified identifier mapping (UID) and a standardized microbial dataset. A deep learning relation extraction model is used to identify semantic associations between strains and genes, strains and metabolites, strains and habitats, and strains and functions. This generates a set of semantic relation triples, and the system calculates the relation confidence value for each triple. A graph construction module uses the UID mapping results as a node set and the semantic relation triples as an edge set to construct a microbial knowledge attribute graph structure, storing it in a graph database format. A knowledge reasoning module performs graph neural network reasoning operations based on the microbial knowledge attribute graph structure, calculating functional similarity between strains and predicting the potential functional attributes of unlabeled strains, generating knowledge reasoning result data. This data includes strain function prediction data and metabolic pathway association prediction data. A semantic retrieval module receives user-input natural language queries, converts them into graph query statements, performs graph traversal retrieval operations on the microbial knowledge attribute graph structure, and generates semantic retrieval response results based on the knowledge reasoning result data.

[0008] Compared with existing technologies, this invention has the following advantages: First, by constructing a multi-level ontology model that conforms to the norms of microbial taxonomy, a complete hierarchical classification structure from strain to family is established, which can effectively support the dynamic evolution of the microbial taxonomy system and the accurate mapping of historical data. Second, through an entity disambiguation mechanism based on a pre-trained language model, combined with the joint representation of contextual semantic features and ontology concept embedding, equivalent entities under different naming forms can be accurately identified, solving the serious problem of synonymy in the field of microbiology. Third, through a relation extraction method combining remote supervision and deep learning, various types of complex semantic relationships such as strain-gene, strain-metabolite, strain-habitat, and strain-function can be effectively identified, improving the depth of strain functional association mining. Fourth, through graph neural network reasoning and link prediction algorithms, potential relationships hidden in the knowledge graph can be discovered, enabling strain function prediction and metabolic pathway association analysis, expanding the application value of knowledge graphs. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall process of the semantic processing method for microbial data based on graph models provided in this embodiment of the invention.

[0010] Figure 2 This is a schematic diagram of the module architecture of the graph model-based semantic processing system for microbial data provided in this embodiment of the invention. Detailed Implementation

[0011] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0012] Reference Figure 1 As shown, this embodiment of the invention provides a semantic processing method for microbial data based on a graph model. This method constructs a microbial domain-specific ontology model, integrates heterogeneous microbial data sources, and utilizes deep learning technology to achieve entity recognition, relation extraction, and knowledge reasoning, ultimately constructing a high-quality microbial knowledge graph. The method includes the following steps:

[0013] Step S1: Data acquisition and preprocessing steps.

[0014] The core objective of the data acquisition and preprocessing steps is to obtain microbial-related data from multiple heterogeneous data sources and convert them into a unified, standardized format, laying the data foundation for subsequent ontology mapping and knowledge extraction.

[0015] In the specific implementation process, the data acquisition module first establishes connection interfaces with various microbial data sources. For strain preservation databases, the system obtains basic information records of strains through the database API interface, including fields such as strain preservation number, taxonomic name, isolation source, culture conditions, morphological characteristics, and physiological and biochemical properties. For genome sequence databases, the system downloads genome sequence files and their annotation information via FTP protocol or WebAPI, and parses sequence data in FASTA format and gene annotation data in GenBank format. For microbiology literature databases, the system uses a literature search API to obtain the titles, abstracts, keywords, and full text of documents. For the experimental recording system, the system reads experimental data files and parses information such as experimental parameters, measurement results, and experimental conditions.

[0016] The format standardization conversion process addresses the differences in data formats across various data sources. For structured strain records from microbial collections, the system converts them into a unified entity attribute table format. This format uses a key-value pair structure to store various attribute information, with each strain record containing standard fields such as a unique identifier, taxonomic name, isolation source, and culture conditions. For sequence files from genome sequence databases, the system uniformly converts FASTA and GenBank formats into a standardized sequence annotation format, which includes both sequence data and gene function annotation information. For unstructured text content from microbiology literature databases, the system uses natural language processing tools to segment and divide the text into sentences, parsing it into a semi-structured annotated document format for easier subsequent entity recognition processing. For experimental data from experimental record systems, the system converts various experimental parameters and measurement results into a key-value pair attribute list format.

[0017] Missing value imputation employs different strategies based on the data field type and business meaning. For missing values ​​in numeric fields, the system uses the median of the same data type; for missing values ​​in categorical fields, the system uses the mode; and for missing values ​​in critical fields that cannot be automatically imputed, the system marks them as awaiting manual review. Noisy data filtering operations include deleting records with obvious formatting errors, filtering outliers exceeding reasonable ranges, and removing duplicate records.

[0018] After the above processing, the system generates a standardized microbial dataset. This dataset uses a unified data model to store microbial data from all sources. Each record contains metadata such as data source identifier, original identifier, standardized attribute set, and data quality score.

[0019] Step S2: Ontology model construction steps.

[0020] The core objective of the ontology model construction step is to establish a microbial domain ontology model based on microbial taxonomy standards, providing semantic framework support for subsequent entity linking and relation extraction.

[0021] In constructing the hierarchical classification structure, the system referenced the International Code of Bacterial Nomenclature and Bergey's Manual of Systematic Bacteriology to establish a complete classification hierarchy from the strain level to the family level. The strain level is at the bottom of the hierarchical structure, representing specific microbial isolates. Each strain entity includes instance-level attributes such as accession number, isolation date, and isolation location. The species level is the basic unit of taxonomy; each species entity includes attributes such as species name, type strain reference, and species characteristic description. The genus level is the superordinate taxonomic unit of the species; each genus entity includes attributes such as genus name, genus characteristic description, and type species reference. The family level is the superordinate taxonomic unit of the genus; each family entity includes attributes such as family name and family characteristic description. The levels are connected by genus-family relationships, forming a strict hierarchical inclusion relationship.

[0022] To support the dynamic evolution of microbial taxonomy, the ontology model also incorporates a taxonomic revision history mechanism. When a genus or species undergoes taxonomic recombination, the system establishes a recombination relationship between the original and new taxa within the ontology model, and records the revision time and the literature on which the revision was based, thereby achieving an accurate mapping of historical data to the current taxonomic system.

[0023] In constructing the attribute relationship framework, the system defined the core entity types and their attributes in the microbial domain. Attributes for strain entities include preservation number, taxonomic name, isolation source, culture conditions, morphological characteristics, and physiological and biochemical properties. Attributes for gene entities include gene name, gene sequence, gene function annotation, and gene ontology terminology. Attributes for metabolite entities include compound name, molecular formula, chemical structure, and bioactivity. Attributes for habitat entities include habitat name, geographical location, and environmental parameters. Attributes for functional entities include functional name, functional description, and functional category.

[0024] The system further defines the relationship types and constraints between entity types. Between strain entity types and gene entity types, coding relationship types are defined, representing the relationship of encoding specific genes in the strain's genome, and regulatory relationship types, representing the regulatory relationship of gene expression. Between strain entity types and metabolite entity types, production relationship types are defined, representing the relationship that the strain can synthesize specific metabolites, and degradation relationship types are defined, representing the relationship that the strain can decompose specific compounds. Between strain entity types and habitat entity types, habitat relationship types are defined, representing the natural habitat of the strain, and isolation relationship types, representing the isolation source of the strain. Between strain entity types and functional entity types, functional relationship types are defined, representing the biological functions possessed by the strain, and process participation relationship types, representing the biological processes in which the strain participates.

[0025] Each relation type is configured with relation attribute constraints, including relation direction constraints (specifying the source entity type and target entity type of the relation), relation cardinality constraints (specifying whether the relation is one-to-one, one-to-many, or many-to-many), and relation transitivity constraints (specifying whether the relation has transitivity).

[0026] Step S3: Entity recognition and linking steps.

[0027] The core objective of the entity recognition and linking step is to identify microbial-related entities from a standardized microbial dataset and link the identified entities to the corresponding concept nodes in the microbial domain ontology model, thereby solving the common problems of synonymy and polysemy in microbial nomenclature.

[0028] In the entity recognition stage, the system utilizes a deep learning-based named entity recognition model to extract candidate entities from the text content of a standardized microbial dataset. This named entity recognition model employs a bidirectional long short-term memory network combined with a conditional random field architecture, capable of recognizing various types of entities, including strain names, gene names, metabolite names, habitat names, and functional terms. For field values ​​from structured data sources, the system directly adds them to the candidate entity set as candidate entities.

[0029] During the entity disambiguation phase, the system addresses synonyms and polysemy in the candidate entity set. Synonyms refer to the same biological entity being referred to by different names in different data sources; for example, *Escherichia coli* may be represented as *Escherichiacoli*, *E. coli*, or *Escherichia coli*, among other forms. Polysemy refers to the same name referring to different biological entities; for example, certain strain numbers may refer to different strains in different collections of cultural relics.

[0030] The calculation process for entity disambiguation vector representation is as follows: First, the system obtains the context text fragments of each candidate entity in the candidate entity set within the standardized microbial dataset. These context text fragments contain a sequence of words within a certain range before and after the candidate entity. The system then encodes the context text fragments using a pre-trained language model to generate context semantic feature vectors. This pre-trained language model is domain-adaptive trained based on a large-scale microbiology literature corpus and is capable of capturing the semantic features of microbiological terms.

[0031] Secondly, the system extracts the definition description text and hierarchical path information for each concept node from the microbial domain ontology model. The definition description text includes the standard name of the concept, a list of synonyms, and a definition description. The hierarchical path information contains the path sequence of the concept from the root node to the current node in the ontology hierarchy. The system uses a pre-trained language model to jointly encode the definition description text and hierarchical path information to generate ontology concept embedding vectors.

[0032] Then, the system calculates the cosine similarity between the context semantic feature vector and the embedding vectors of each ontology concept in the microbial domain ontology model. The formula for calculating the cosine similarity is as follows:

[0033] ,

[0034] in: It is a contextual semantic feature vector; It is an ontology concept embedding vector; It is the result of the inner product of two vectors; It is the Euclidean norm of the context semantic feature vector; It is the Euclidean norm of the ontological concept embedding vector; It is the calculated cosine similarity value, which ranges from negative one to positive one.

[0035] When the cosine similarity score exceeds a preset similarity threshold (e.g., set to 0.85), the system maps the corresponding candidate entity to the concept node with the highest similarity. For candidate entities whose cosine similarity score does not exceed the preset similarity threshold, the system marks them as pending manual confirmation, waiting for domain experts to review and confirm them.

[0036] After the above processing, the system generates a unified identifier mapping result. Each candidate entity is assigned a unique ontology concept identifier. Different names referring to the same biological entity in different data sources are mapped to the same ontology concept identifier, thereby realizing a unified representation of entities.

[0037] Step S4: Relation extraction step.

[0038] The core objective of the relation extraction step is to identify semantic relationships between entities from a standardized microbial dataset, construct a set of semantic relation triples, and provide data support for the edge set of the knowledge graph.

[0039] In the remote supervised training sample generation phase, the system automatically generates training samples based on the defined relation types in the microbial ontology model and known relation instances in publicly available microbial databases. The specific process is as follows: The system obtains known strain-gene, strain-metabolite, strain-habitat, and strain-function relation instances from publicly available microbial databases (such as UniProt, KEGG, NCBITaxonomy, etc.). For each known relation instance, the system retrieves text sentences containing both the head and tail entities of that relation in a standardized microbial dataset and labels these sentences as positive examples of that relation type. Simultaneously, the system generates negative examples, which consist of randomly paired entity pairs and their co-occurring sentences. Through this process, the system automatically generates a remote supervised training sample set, eliminating the need for manual annotation.

[0040] In the construction phase of the deep neural network for relation classification, the system designed a network architecture comprising an entity pair encoding layer, a relation representation layer, and a classification output layer. The entity pair encoding layer receives entity pairs and their context sentences as input, encodes the sentences using a pre-trained language model, and highlights the semantic information of entity positions through an attention mechanism, outputting a joint semantic representation vector of the entity pairs. The relation representation layer performs a nonlinear transformation on the joint semantic representation vector of the entity pairs to extract high-level semantic features related to the relations. The classification output layer employs a multi-label classifier, outputting the probability distribution of entity pairs belonging to each relation type.

[0041] The system trains a deep neural network for relation classification using a remotely supervised training sample set. Since noisy labels may exist in the remotely supervised training samples, the system employs a multi-instance learning strategy, treating all co-occurring sentences of the same entity pair as a single packet for training and prediction at the packet level, thereby reducing the impact of noisy labels.

[0042] During the relation extraction and prediction phase, the system inputs all entity pairs and their corresponding contexts from the unified identifier mapping results into the trained relation classification deep neural network to obtain the relation type prediction results for each entity pair. For relation types with prediction probabilities exceeding a preset relation threshold (e.g., set to 0.7), the system generates corresponding semantic relation triples in the format (head entity identifier, relation type, tail entity identifier). The system also records the relation confidence score for each triple, which is determined by the output probability of the relation classification deep neural network.

[0043] After the above processing, the system generates a set of semantic relation triples, which contains all high-confidence semantic relations extracted from the standardized microbial dataset, covering various relation types such as strain-gene, strain-metabolite, strain-habitat, and strain-function.

[0044] Step S5: Knowledge Graph Construction Steps.

[0045] The core objective of the knowledge graph construction process is to organize the unified identifier mapping results and semantic relation triples into an attribute graph structure and store them in a graph database to form a queryable microbial knowledge graph.

[0046] During the attribute graph structure construction phase, the system converts each entity in the unified identifier mapping result into nodes in the graph. Each node contains a unique identifier, a node type label, and node attribute fields. The node type label corresponds to the ontology concept type to which the entity belongs, such as a strain node, gene node, metabolite node, habitat node, or functional node. The node attribute fields store the entity's various attribute values, and the schema of the attribute fields is determined by the attribute definition of that concept type in the ontology model.

[0047] The system converts each triple in the semantic relation triple set into edges in a graph. Each edge connects a source node and a target node and includes a relation type label, a relation confidence score, and a relation source attribute. The relation type label corresponds to the relation type defined in the ontology model. The relation confidence score records the confidence score calculated during the relation extraction phase. The relation source attribute records the data source information for this relation, facilitating subsequent tracing and verification.

[0048] During the graph database storage phase, the system imports the microbial knowledge attribute graph structure into the graph database system. The system establishes node and edge index structures to support efficient queries. The node index structure includes a classification index based on node type (facilitating the retrieval of all nodes of a specific type) and a full-text search index based on node attributes (facilitating fuzzy queries based on attribute values). The edge index structure includes a fast query index based on relation type (facilitating the retrieval of all edges of a specific relation type).

[0049] After the above processing, the system generates a microbial knowledge attribute graph structure and completes graph database storage, forming a complete microbial knowledge graph that supports graph operations such as graph traversal query, path retrieval, and subgraph matching.

[0050] Step S6: Knowledge Reasoning Step.

[0051] The core objective of the knowledge reasoning step is to perform reasoning operations based on the microbial knowledge attribute graph structure, discover the knowledge hidden in the graph, and generate strain function prediction data and metabolic pathway association prediction data.

[0052] In the node embedding learning phase, the system utilizes a graph attention network to learn the representations of each node in the microbial knowledge attribute graph structure. The graph attention network adaptively learns the contribution weights of neighboring nodes to the central node through an attention mechanism, aggregating neighborhood information to generate node-level embedding vector representations. The computational process of node embedding learning is as follows:

[0053] For any node in the graph Its node-level embedding vector representation Calculated by aggregating neighbor node information:

[0054] ,

[0055] in: It is a node In the Layer embedding vector representation; It is a node The set of neighboring nodes; It is the first Neighbor nodes in the layer For nodes Attention weight coefficient; It is the first The learnable weight matrix of the layer; Neighboring nodes In the Layer embedding vector representation; It is a non-linear activation function.

[0056] Attention weight coefficient The calculation formula is as follows:

[0057] ,

[0058] in: These are learnable attention vector parameters; This represents a vector concatenation operation; It is the activation function of the rectified linear unit with leakage.

[0059] In the strain similarity calculation phase, the system calculates the strain similarity metric between any two strain nodes based on node-level embedded vector representation. The strain similarity metric comprehensively reflects the genomic similarity, metabolic function similarity, and niche similarity between strains. For any two strain nodes... and The strain similarity metric values The calculation formula is as follows:

[0060] ,

[0061] in: and They are strains and The node-level embedding vector representation.

[0062] In the link prediction phase, the system uses a link prediction algorithm based on node-level embedding vector representation to predict potential relationship edges that do not yet exist but may exist in the microbial knowledge attribute graph structure. For any two nodes... and and relation types The system calculates triples. Probability score of existence:

[0063] ,

[0064] in: It is a relation type The corresponding relation matrix parameters; It is the Sigmoid activation function, which maps the score to the interval between zero and one.

[0065] When the probability score exceeds the preset link prediction threshold (e.g., set to 0.8), the system adds the potential relationship to the strain function prediction data, indicating that the system predicts that the relationship may exist.

[0066] In the metabolic pathway association analysis phase, the system constructs a metabolic pathway association graph structure based on the connection relationships between metabolite nodes, strain nodes, and enzyme nodes in the microbial knowledge attribute graph structure. The system then uses a subgraph matching algorithm to identify complete metabolic pathway patterns within the metabolic pathway association graph structure, discovering the associations between strains and metabolic pathways.

[0067] After the above processing, the system generates knowledge reasoning result data, including strain function prediction data and metabolic pathway association prediction data, which expands the knowledge coverage of the knowledge graph.

[0068] Step S7: Semantic retrieval step.

[0069] The core objective of the semantic retrieval step is to support users in querying microbial knowledge graphs in natural language, converting natural language queries into graph query statements and returning retrieval results.

[0070] During the query intent recognition phase, the system analyzes the user's natural language query request to determine the query type. The system supports three main query types: entity attribute query (querying the attribute values ​​of a specific entity), relationship path query (querying the relationship paths between entities), and aggregation statistics query (statistically counting the number of entities or attribute distributions that meet specific conditions). The system uses a rule-based and machine learning-based intent recognition model to classify query types.

[0071] During the semantic parsing phase, the system extracts query entities, query relationships, and query constraints from the natural language query request. Query entities are extracted from the query text using named entity recognition and entity linking technologies and mapped to corresponding nodes in the microbial knowledge attribute graph structure. Query relationships are extracted from the query text using relation term recognition and relation mapping technologies and mapped to relation types defined in the ontology model. Query constraints include attribute value constraints, relation depth constraints, and result quantity constraints.

[0072] During the graph query generation phase, the system generates corresponding graph query statements based on the query type, query entities, query relationships, and query constraints. The graph query statements use query language specifications supported by the graph database (such as Cypher query language), enabling query operations such as node matching, relationship traversal, attribute filtering, and result aggregation.

[0073] During the query execution and result fusion phase, the system executes graph query statements on the microbial knowledge attribute graph structure to obtain graph path traversal results. The system further fuses these graph path traversal results with knowledge reasoning results, adding the potential relationships and predictions obtained through reasoning as supplementary information to the search results. The system then ranks the fused results by relevance and generates semantic search response results to return to the user.

[0074] Step S8: Knowledge update step.

[0075] To maintain the timeliness and accuracy of the microbial knowledge graph, the system optionally performs a knowledge update step. This step periodically retrieves incremental data from microbial strain databases, genome sequence databases, microbiology literature databases, and experimental record systems. The system then performs entity recognition and linking steps, as well as relation extraction steps, on the incremental data, generating an incremental entity set and an incremental relation set.

[0076] The system detects conflicts between the incremental entity set and existing nodes in the microbial knowledge attribute graph structure. When attribute value conflicts exist, the system determines the version to retain based on the reliability of the data source and the freshness of the timestamp. The reliability of the data source is predefined based on the authority and accuracy of the data source, with data from authoritative databases having higher reliability. When reliability is the same, data with a more recent timestamp is retained first.

[0077] The system integrates the incremental entity set and the incremental relation set into the microbial knowledge attribute graph structure, and triggers the incremental reasoning update operation of the knowledge reasoning step to update the embedding vector representation and link prediction results of the affected nodes.

[0078] Reference Figure 2As shown, this embodiment of the invention also provides a semantic processing system for microbial data based on a graph model. This system corresponds to the method described in Embodiment 1 and includes a data acquisition module, an ontology construction module, an entity recognition module, a relation extraction module, a graph construction module, a knowledge reasoning module, and a semantic retrieval module.

[0079] The data acquisition module establishes data connections with microbial culture databases, genome sequence databases, microbiology literature databases, and experimental record systems to obtain microbial-related data from these heterogeneous data sources. Internally, the data acquisition module includes a data source adapter component, a format conversion component, a missing value handling component, and a noise filtering component. The data source adapter component is responsible for establishing connections with various data sources and acquiring raw data. The format conversion component converts raw data in different formats into a unified standardized format. The missing value handling component fills in missing fields in the data. The noise filtering component removes malformed and outlier data, ultimately generating a standardized microbial dataset.

[0080] The ontology construction module builds a microbial ontology model according to microbial taxonomy standards. Internally, the module includes a taxonomic hierarchy construction component and an attribute relationship definition component. The taxonomic hierarchy construction component establishes a taxonomic hierarchy structure from strain level through species level, genus level, and family level, in accordance with the International Code for Bacterial Nomenclature, and supports the recording and querying of taxonomic revision history. The attribute relationship definition component defines the attribute patterns of entity types such as strains, genes, metabolites, habitats, and functions, as well as the relationship types and constraints between entity types.

[0081] The entity recognition module processes standardized microbial datasets using named entity recognition (NAME) and entity disambiguation algorithms. Internally, the module comprises a named entity recognition component, a context encoding component, and an entity linking component. The named entity recognition component identifies candidate entities from the text data. The context encoding component generates contextual semantic feature vectors for candidate entities using a pre-trained language model. The entity linking component calculates the similarity between the contextual semantic feature vectors and the ontology concept embedding vectors, mapping candidate entities to ontology concept nodes and generating unified identifier mapping results.

[0082] The relation extraction module extracts semantic relations from a standardized microbial dataset using remote supervision and deep learning techniques. Internally, the module includes a training sample generation component, a relation classification network component, and a relation prediction component. The training sample generation component automatically generates training data through remote supervision. The relation classification network component constructs and trains a deep neural network for relation classification. The relation prediction component uses the trained model to predict the relationship type and confidence level between entity pairs, generating a set of semantic relation triples.

[0083] The graph construction module organizes the unified identifier mapping results and semantic relation triples into a microbial knowledge attribute graph structure. Internally, the graph construction module includes a graph structure generation component and a graph database storage component. The graph structure generation component converts entities into nodes and relation triples into edges, constructing the attribute graph structure. The graph database storage component imports the attribute graph structure into a graph database and establishes node and edge indexes.

[0084] The knowledge reasoning module uses graph neural network technology to perform reasoning operations on the microbial knowledge attribute graph structure. Internally, the module includes a node embedding learning component, a similarity calculation component, and a link prediction component. The node embedding learning component uses a graph attention network to learn the embedding vector representation of nodes. The similarity calculation component calculates the similarity measure between strains based on the embedding vectors. The link prediction component predicts potential relationships that may exist in the graph, generating knowledge reasoning result data.

[0085] The semantic retrieval module receives natural language queries from users and returns search results. Internally, it includes an intent recognition component, a semantic parsing component, a graph query generation component, and a result fusion component. The intent recognition component determines the query type. The semantic parsing component extracts query entities, relationships, and constraints. The graph query generation component generates graph database query statements. The result fusion component merges and sorts the graph query results with the inference results to generate the semantic retrieval response.

[0086] The modules described above interact through data interfaces, forming a complete semantic processing pipeline for microbial data. The output of the data acquisition module serves as input to the ontology construction and entity recognition modules. The output of the ontology construction module serves as input to the entity recognition and relation extraction modules. The output of the entity recognition module serves as input to the relation extraction and graph construction modules. The output of the relation extraction module serves as input to the graph construction module. The output of the graph construction module serves as input to the knowledge reasoning and semantic retrieval modules. The output of the knowledge reasoning module serves as input to the semantic retrieval module.

[0087] The semantic processing method and system for microbial data based on graph models provided in this invention effectively solves the technical problems of difficult integration of heterogeneous microbial data, insufficient accuracy of microbial naming disambiguation, and limited ability to mine functional associations of strains by constructing a microbial domain-specific ontology model, achieving high-precision entity disambiguation and relation extraction, and using graph neural networks for knowledge reasoning. This improves the semantic level and knowledge service capabilities of microbial data resources.

[0088] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A semantic processing method for microbial data based on graph models, characterized in that, Includes the following steps: The data acquisition and preprocessing steps involve obtaining heterogeneous microbial data from microbial strain preservation databases, genome sequence databases, microbiology literature databases, and experimental record systems, and performing format unification conversion, missing value imputation, and noise data filtering operations on the heterogeneous microbial data to generate a standardized microbial dataset. The ontology model construction steps are as follows: a microbial domain ontology model is constructed according to the microbial taxonomy standard. The microbial domain ontology model includes a classification hierarchy structure from the strain level through the species level, genus level to the family level, and an attribute relationship framework covering gene attributes, metabolite attributes, culture condition attributes, physiological characteristic attributes and application function attributes. The entity recognition and linking step involves extracting a candidate entity set from the standardized microbial dataset using a named entity recognition algorithm. For the synonyms and polysemy in the candidate entity set, the entity disambiguation vector representation is calculated based on the microbial domain ontology model. Each entity in the candidate entity set is mapped to the corresponding concept node in the microbial domain ontology model to generate a unified identifier mapping result. The relation extraction step, based on the unified identifier mapping result and the standardized microbial dataset, uses a deep learning relation extraction model to identify semantic associations between strains and genes, strains and metabolites, strains and habitats, and strains and functions, generates a set of semantic relation triples, and calculates the relation confidence value for each triple in the set of semantic relation triples. The knowledge graph construction steps are as follows: the unified identifier mapping result is used as a node set, the semantic relation triple set is used as an edge set, a microbial knowledge attribute graph structure is constructed, and the microbial knowledge attribute graph structure is stored in a graph database storage format. The knowledge reasoning step involves performing graph neural network reasoning operations based on the microbial knowledge attribute graph structure to calculate the functional similarity between strains and predict the potential functional attributes of unlabeled strains, generating knowledge reasoning result data. The knowledge reasoning result data includes strain function prediction data and metabolic pathway association prediction data. The semantic retrieval step involves receiving a natural language query request input by the user, converting the natural language query request into a graph query statement, performing a graph traversal retrieval operation on the microbial knowledge attribute graph structure, and generating a semantic retrieval response result by combining the knowledge reasoning result data.

2. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the data acquisition and preprocessing steps, the format unification conversion operation includes: The structured strain records from the microbial culture bank database are converted into a unified entity attribute table format; Convert FASTA and GenBank format sequence files from genome sequence databases into a standardized sequence annotation format; Parse unstructured text content from microbiology literature databases into semi-structured annotated document format; Convert experimental parameter data from the experimental recording system into a key-value pair attribute list format.

3. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the entity recognition and linking step, the calculation of entity disambiguation vector representation based on the microbial domain ontology model includes: Obtain the contextual text fragments of each candidate entity in the candidate entity set in the standardized microbial dataset, and encode the contextual text fragments using a pre-trained language model to generate contextual semantic feature vectors; The definition description text and hierarchical path information of each concept node are extracted from the microbial domain ontology model. The definition description text and hierarchical path information are jointly encoded using the pre-trained language model to generate ontology concept embedding vectors. Calculate the cosine similarity between the context semantic feature vector and the embedding vectors of each ontology concept in the microbial domain ontology model. When the cosine similarity exceeds a preset similarity threshold, map the corresponding candidate entity to the concept node with the highest similarity and generate the entity disambiguation vector representation.

4. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the ontology model construction step, the construction of the attribute relationship framework includes: Define the coding and regulatory relationships between strain entity types and gene entity types; Define the types of generation and degradation relationships between strain entity types and metabolite entity types; Define the habitat relationship type and the isolation relationship type between strain entity type and habitat entity type; Define functional relationship types and participation process relationship types between strain entity types and functional entity types; Set relation attribute constraints for each relation type, including relation directionality constraints, relation cardinality constraints, and relation transitivity constraints.

5. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the relation extraction step, the step of identifying semantic associations using a deep learning relation extraction model includes: Based on the defined relation types in the microbial domain ontology model and known relation instances in the public microbial database, a remote supervision training sample set is automatically generated through remote supervision. A relation classification deep neural network is constructed, which includes an entity pair encoding layer, a relation representation layer, and a classification output layer. The relation classification deep neural network is trained using the remote supervised training sample set. The entity pairs and their corresponding contexts in the unified identifier mapping result are input into the trained relation classification deep neural network to obtain the relation type prediction result and the relation confidence value for each entity pair.

6. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the knowledge graph construction step, storing the microbial knowledge attribute graph structure in a graph database storage format includes: Each entity in the unified identifier mapping result is stored as a node record in a graph database, and a node type label and node attribute field are configured for each node record; Each triple in the set of semantic relation triples is stored as an edge record in a graph database, and each edge record is configured with a relation type label, a relation confidence value, and a relation source attribute; Establish a node index structure and an edge index structure. The node index structure includes a classification index based on node type and a full-text search index based on node attributes. The edge index structure includes a fast query index based on relation type.

7. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the knowledge reasoning step, the execution of graph neural network reasoning operations based on the microbial knowledge attribute graph structure includes: A graph attention network is used to aggregate neighborhood information of each node in the microbial knowledge attribute graph structure to generate a node-level embedding vector representation. The strain similarity metric between any two strain nodes is calculated based on the node-level embedded vector representation. The strain similarity metric comprehensively reflects the genomic similarity, metabolic function similarity, and niche similarity between strains. The link prediction algorithm is used to predict potential relationship edges that do not yet exist but may exist in the microbial knowledge attribute graph structure based on the node-level embedding vector representation, thereby generating the strain function prediction data. Based on the connection relationships between metabolite nodes, strain nodes, and enzyme nodes in the microbial knowledge attribute graph structure, a metabolic pathway association graph structure is constructed, and a subgraph matching algorithm is used to identify complete metabolic pathway patterns.

8. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, In the semantic retrieval step, converting the natural language query request into a graph query statement includes: The natural language query request is subjected to intent recognition to determine whether the query type is an entity attribute query, a relationship path query, or an aggregation statistics query. The semantic parsing model is used to extract query entities, query relationships, and query constraints from the natural language query request. Based on the query type, query entity, query relationship, and query constraints, a corresponding graph query statement is generated, which conforms to the graph database query language specification. The graph query statement is executed on the microbial knowledge attribute graph structure to obtain the graph path traversal results. The graph path traversal results are then fused and sorted with the knowledge reasoning results data to generate the semantic retrieval response results.

9. The semantic processing method for microbial data based on graph models according to claim 1, characterized in that, It also includes a knowledge update step, which includes: Incremental data is periodically obtained from the aforementioned strain preservation database, genome sequence database, microbiology literature database, and experimental record system; The entity recognition and linking step and the relationship extraction step are performed on the incremental data to generate an incremental entity set and an incremental relationship set; The system detects conflicts between the incremental entity set and existing nodes in the microbial knowledge attribute graph structure. When there are attribute value conflicts, the system determines the version to retain based on the reliability of the data source and the age of the timestamp. The incremental entity set and the incremental relation set are merged into the microbial knowledge attribute graph structure, triggering the incremental reasoning update operation of the knowledge reasoning step.

10. A graph-based semantic processing system for microbial data, used to implement the graph-based semantic processing method for microbial data as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire heterogeneous microbial data from microbial strain preservation databases, genome sequence databases, microbiology literature databases, and experimental record systems, and to perform format unification conversion, missing value imputation, and noise data filtering operations on the heterogeneous microbial data to generate a standardized microbial dataset. The ontology construction module is used to construct a microbial domain ontology model according to the microbial taxonomy standard. The microbial domain ontology model includes a classification hierarchy structure from the strain level through the species level, genus level to the family level, and an attribute relationship framework covering gene attributes, metabolite attributes, culture condition attributes, physiological characteristic attributes and application function attributes. The entity recognition module is used to extract a candidate entity set from the standardized microbial dataset using a named entity recognition algorithm. For the synonyms and polysemy in the candidate entity set, it calculates the entity disambiguation vector representation based on the microbial domain ontology model, maps each entity in the candidate entity set to the corresponding concept node in the microbial domain ontology model, and generates a unified identifier mapping result. The relation extraction module is used to identify semantic associations between strains and genes, strains and metabolites, strains and habitats, and strains and functions based on the unified identifier mapping results and the standardized microbial dataset using a deep learning relation extraction model, generate a set of semantic relation triples, and calculate the relation confidence value for each triple in the set of semantic relation triples. The graph construction module is used to construct a microbial knowledge attribute graph structure by taking the unified identifier mapping result as a node set and the semantic relation triple set as an edge set, and storing the microbial knowledge attribute graph structure in a graph database storage format. The knowledge reasoning module is used to perform graph neural network reasoning operations based on the microbial knowledge attribute graph structure, calculate the functional similarity between strains and predict the potential functional attributes of unlabeled strains, and generate knowledge reasoning result data, which includes strain function prediction data and metabolic pathway association prediction data. The semantic retrieval module is used to receive natural language query requests input by users, convert the natural language query requests into graph query statements, perform graph traversal retrieval operations on the microbial knowledge attribute graph structure, and generate semantic retrieval response results by combining the knowledge reasoning result data.

Images