A multi-agent-based protein design knowledge graph construction method and system

Abstract

The present application relates to the technical field of protein design, and discloses a protein design knowledge graph construction method based on multi-agent, comprising the following steps: obtaining a natural language research topic input by a user; obtaining entity terms and intention objects by using a large language model, and generating a query strategy for a multi-source heterogeneous database based on the standardized entity terms and intention objects; obtaining original data from the multi-source heterogeneous database by using the query strategy; processing the parsed original data uniformly, and storing the processed data into a relational database; automatically extracting knowledge triples from the relational database and unstructured data, performing knowledge verification and deduplication, and storing the knowledge triples into a graph database; simultaneously, using a graph embedding algorithm and link prediction to perform knowledge completion and reasoning, and forming a domain-enhanced protein design knowledge graph; the present application solves the problems of low efficiency and poor generality of the existing protein design knowledge graph construction method.

Description

Technical Field

[0001] This invention relates to the field of protein design technology, and in particular to a method and system for constructing a protein design knowledge graph based on multiple agents. Background Technology

[0002] Protein design is a cutting-edge area in life sciences and drug development, and its research process heavily relies on massive, multi-source, and heterogeneous biomedical data. This data is scattered across structured databases (such as UniProt and PDB), unstructured literature documents (such as PubMed, and patent and clinical trial databases), and encompasses multi-dimensional knowledge including protein sequences, three-dimensional structures, functional annotations, mutation information, disease associations, and drug mechanisms of action. Efficiently extracting and integrating valuable information from these scattered and complex data sources has become a significant technical challenge driving protein design research. Currently, researchers still largely rely on manual methods for data retrieval, downloading, and integration when conducting specific protein research projects. This process suffers from technical pain points such as low data acquisition and preprocessing efficiency, severe semantic heterogeneity of diverse data, insufficient knowledge association and reasoning capabilities, and a lack of dynamic adaptability and customization. Therefore, there is an urgent need for an intelligent technical framework that integrates automated data acquisition, semantic parsing, entity recognition, cross-source fusion, and knowledge graph construction, capable of quickly and accurately constructing a domain knowledge base highly relevant to protein design based on user-input natural language research topics. Summary of the Invention

[0003] This invention provides a method and system for constructing a protein design knowledge graph based on multiple agents, in order to solve the problems of low construction efficiency and poor versatility in existing protein design knowledge graph construction methods.

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] In a first aspect, the present invention provides a method for constructing a protein design knowledge graph based on multiple agents, comprising the following steps:

[0006] S1: Obtain the natural language research topic input by the user;

[0007] S2: Utilize a large language model to perform entity recognition and intent analysis on the natural language research topic to obtain entity terms and intent objects. Standardize the entity terms based on entity standardization and semantic alignment to obtain standardized entity terms. Generate a query strategy for multi-source heterogeneous databases based on the standardized entity terms and intent objects.

[0008] S3: Execute the query strategy using an asynchronous concurrent scheduling mechanism, obtain raw data from multi-source heterogeneous databases in parallel using the query strategy, and perform adaptive parsing and temporary storage on the raw data. The raw data includes protein structured data and protein unstructured data. The protein structured data is data from a protein database, and the protein unstructured data is literature document data related to proteins.

[0009] S4: Perform mixed information extraction and cross-source entity alignment on the parsed raw data, and store the processed data in a relational database;

[0010] S5: Extract knowledge triples from the relational database using the GraphRAG framework, perform knowledge verification and deduplication, and store the knowledge triples in the graph database. At the same time, use graph embedding algorithms and link prediction to perform knowledge completion and reasoning, forming a domain-enhanced protein design knowledge graph.

[0011] Optionally, step S2 includes the following sub-steps:

[0012] S2.1: Design entity extraction prompts, use a large language model to perform entity recognition on the natural language research topic through prompts, extract biological entity categories and model named entity categories in the natural language research topic, use biological entity categories and model named entity categories as entity terms, and construct an entity list by one-to-one correspondence between biological entity categories and model named entity categories;

[0013] S2.2: Analyze the relationships between entity terms in the entity list and generate an intent object containing actions, goals, and context based on the relationships between entity terms;

[0014] S2.3: Use the BioBERT model to vectorize the identified entity terms, calculate the semantic similarity between the entity vectorization and the authoritative thesaurus terms, and when the semantic similarity is greater than a preset threshold, use the authoritative thesaurus terms as the standardized entity terms corresponding to the entity vectorization, and automatically generate executable query instructions for multi-source heterogeneous databases based on the standardized entity terms and intent objects.

[0015] Optionally, the semantic similarity calculation uses the cosine similarity formula:

[0016] ;

[0017] in, For cosine similarity, From the perspective of entity vectorization, A and B are the vector representations of entity terms and authoritative thesaurus terms, respectively.

[0018] Optionally, step S3 includes the following sub-steps:

[0019] S3.1: Use an asynchronous I / O model to create independent coroutines for each multi-source heterogeneous database API request, and use an event loop for concurrent scheduling to obtain raw data in parallel;

[0020] S3.2: Dynamically select and load the corresponding parser according to the original data format, perform adaptive parsing on the acquired original data, and adopt a breakpoint resume mechanism with retry and backoff strategies to ensure the stability of the acquired original data.

[0021] S3.3: The original data is temporarily stored in a distributed file system or object storage, along with metadata files that record the source, timestamp, and version hash value, for use in tracing the source or verifying the consistency of the parsed data.

[0022] Optionally, step S4 includes the following sub-steps:

[0023] S4.1: First, based on rule-based methods, deterministic information is extracted from the unstructured text of the parsed raw data to obtain protein information. The deterministic information is protein mutation site information. The rule-based methods include regular expression method and domain keyword dictionary method. Then, based on the BERT question answering model, preset questions are set, and the preset questions and answers are used to extract complex semantic information from the protein information in the unstructured text of the raw data. The complex semantic information includes protein function description information and protein binding affinity numerical information.

[0024] S4.2: Construct an entity mapping table between protein information entity terms, use the entity mapping table to determine the target entity in the multi-source heterogeneous database, and normalize the target entity;

[0025] S4.3: Use a star schema to store normalized data in a relational database and enforce foreign key constraints and transaction control to ensure data consistency.

[0026] Optionally, step S5 includes the following sub-steps:

[0027] S5.1: Define a graph pattern in the field of protein design based on protein design requirements. The graph pattern includes protein node types and protein relationship types.

[0028] S5.2: Construct a knowledge corpus based on the aforementioned relational database;

[0029] S5.3: Use a large language model to extract knowledge triples that conform to the graph pattern from the corpus, and perform knowledge verification and deduplication. The structure of the knowledge triple is: entity term - relation - entity term.

[0030] S5.4: Store the knowledge triples in a graph database and use a graph embedding algorithm to predict links in the knowledge triples in the graph database to determine potential relationships between entity terms in the knowledge triples that do not appear in the graph database.

[0031] S5.5: Construct a protein design knowledge graph based on the potential relationships between the graph database and entity terms that do not appear in the graph database;

[0032] S5.6: Enhance the protein design knowledge graph by introducing domain-specific attributes, resulting in a domain-enhanced protein design knowledge graph. The domain-specific attributes include: protein physicochemical properties, conserved domains, and binding sites.

[0033] Secondly, embodiments of this application provide a protein design knowledge graph construction system based on multi-agent technology, including:

[0034] The topic parsing and semantic query generation module is used to receive natural language research topics input by users, perform entity recognition and intent analysis on the natural language research topics using a large language model to obtain entity terms and intent objects, automatically standardize the entity terms based on entity standardization and semantic alignment to obtain standardized entity terms, and generate query strategies for multi-source heterogeneous databases based on the standardized entity terms and intent objects.

[0035] The multi-source heterogeneous data parallel acquisition and preprocessing module is used to execute the query strategy using an asynchronous concurrent scheduling mechanism, acquire raw data from multi-source heterogeneous databases in parallel using the query strategy, and perform adaptive parsing and temporary storage on the raw data. The raw data includes protein structured data and protein unstructured data. The protein structured data is data from a protein database, and the protein unstructured data is literature document data related to proteins.

[0036] The structured knowledge extraction and relational database construction module is used to perform mixed information extraction and cross-source entity alignment on the parsed raw data, and store the processed data in a relational database.

[0037] A knowledge graph construction module for protein design is used to automatically extract knowledge triples from the relational database using the GraphRAG framework, perform knowledge verification and deduplication, and store the knowledge triples in the graph database. At the same time, graph embedding algorithms and link prediction are used to perform knowledge completion and reasoning to form a domain-enhanced protein design knowledge graph.

[0038] Optionally, the topic parsing and semantic query generation module is further configured as follows:

[0039] The design of entity extraction prompts uses a large language model to perform entity recognition on the natural language research topic through prompts, extracts biological entity categories and model named entity categories from the natural language research topic, uses biological entity categories and model named entity categories as entity terms, and constructs an entity list by one-to-one correspondence between biological entity categories and model named entity categories;

[0040] Analyze the relationships between entity terms in the entity list, and generate an intent object containing actions, goals, and context based on the relationships between entity terms;

[0041] The BioBERT model is used to vectorize the identified entity terms, and the semantic similarity between the entity vectorization and the authoritative thesaurus terms is calculated. When the semantic similarity is greater than a preset threshold, the authoritative thesaurus terms are used as the standardized entity terms corresponding to the entity vectorization. Based on the standardized entity terms and intent objects, executable query instructions for multi-source heterogeneous databases are automatically generated.

[0042] Optionally, the multi-source heterogeneous data parallel acquisition and preprocessing module uses an asynchronous I / O model for concurrent scheduling to acquire raw data in parallel;

[0043] Meanwhile, an adaptive parser is used to dynamically parse the data according to the original data format, adaptively parsing the acquired original data and implementing breakpoint resumption with retry and backoff mechanisms.

[0044] The structured knowledge extraction and relational database construction module uses a hybrid method based on rules and BERT question answering model to extract information and builds an entity mapping table to achieve cross-source entity alignment and associate IDs from different databases.

[0045] Optionally, the protein-oriented knowledge graph construction module is executed using the GraphRAG framework, which includes:

[0046] Define a mapping pattern in the field of protein design;

[0047] Construct a knowledge corpus;

[0048] Knowledge triples are extracted using a large language model through retrieval-enhancement-generation.

[0049] Perform knowledge verification and deduplication;

[0050] Store the triples in a graph database and use graph embedding algorithms for link prediction and knowledge completion.

[0051] The graph embedding algorithm uses the ComplEx or RotatE model.

[0052] Beneficial effects:

[0053] This invention provides a multi-agent-based protein design knowledge graph construction method. Through a multi-agent collaborative process (including topic parsing, data acquisition, knowledge extraction, and graph construction modules), it achieves fully automated processing from natural language research topics to customized knowledge graphs. Compared with traditional manual retrieval and script processing, this invention employs asynchronous concurrent scheduling, adaptive parsing, and entity normalization algorithms, improving data acquisition and processing efficiency by tens of times. By introducing the GraphRAG framework for knowledge extraction and graph embedding link prediction, the system can automatically extract semantically related triples from multi-source heterogeneous data and supports deep relationship discovery. The semantic parsing and intent recognition module based on a large language model enables the system to dynamically generate query strategies and graph construction processes based on user-input natural language topics. Combining a microservice architecture, the system supports rapid integration with new large-scale model APIs (such as DeepSeek and Qwen) and database sources (such as clinical trial data and patent databases), exhibiting excellent modular scalability and long-term adaptability. Through GraphRAG's "retrieval-enhancement-generation" mechanism, all extracted knowledge triples are generated based on the original corpus and validated and deduplicated using a domain-specific vocabulary, effectively avoiding the "illusion" problem of LLM. Simultaneously, graph embedding algorithms (such as ComplEx and RotatE) endow the graph with link prediction capabilities, automatically completing potential relationships and upgrading the knowledge base from passive storage to an intelligent engine for active reasoning. Attached Figure Description

[0054] Figure 1 This is a flowchart of a preferred embodiment of the protein design knowledge graph construction method based on multi-agent technology of the present invention;

[0055] Figure 2 This is a schematic diagram of the structure of a protein design knowledge graph construction system based on a preferred embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0058] Please see Figure 1 This application provides a method for constructing a protein design knowledge graph based on multi-agent technology, comprising the following steps:

[0059] S1: Obtain the natural language research topic input by the user;

[0060] S2: Utilize a large language model to perform entity recognition and intent analysis on the natural language research topic to obtain entity terms and intent objects. Standardize the entity terms based on entity standardization and semantic alignment to obtain standardized entity terms. Generate a query strategy for multi-source heterogeneous databases based on the standardized entity terms and intent objects.

[0061] S3: Execute the query strategy using an asynchronous concurrent scheduling mechanism, obtain raw data from multi-source heterogeneous databases in parallel using the query strategy, and perform adaptive parsing and temporary storage on the raw data. The raw data includes protein structured data and protein unstructured data. The protein structured data is data from a protein database, and the protein unstructured data is literature document data related to proteins.

[0062] S4: Perform mixed information extraction and cross-source entity alignment on the parsed raw data, and store the processed data in a relational database;

[0063] S5: Extract knowledge triples from the relational database using the GraphRAG framework, perform knowledge verification and deduplication, and store the knowledge triples in the graph database. At the same time, use graph embedding algorithms and link prediction to perform knowledge completion and reasoning, forming a domain-enhanced protein design knowledge graph.

[0064] Optionally, step S2 includes the following sub-steps:

[0065] S2.1: Design entity extraction prompts, use a large language model to perform entity recognition on the natural language research topic through prompts, extract biological entity categories and model named entity categories in the natural language research topic, use biological entity categories and model named entity categories as entity terms, and construct an entity list by one-to-one correspondence between biological entity categories and model named entity categories;

[0066] S2.2: Analyze the relationships between entity terms in the entity list and generate an intent object containing actions, goals, and context based on the relationships between entity terms;

[0067] S2.3: Use the BioBERT model to vectorize the identified entity terms, calculate the semantic similarity between the entity vectorization and the authoritative thesaurus terms, and when the semantic similarity is greater than a preset threshold, use the authoritative thesaurus terms as the standardized entity terms corresponding to the entity vectorization, and automatically generate executable query instructions for multi-source heterogeneous databases based on the standardized entity terms and intent objects.

[0068] Optionally, the semantic similarity calculation uses the cosine similarity formula:

[0069] ;

[0070] in, For cosine similarity, From the perspective of entity vectorization, A and B are the vector representations of entity terms and authoritative thesaurus terms, respectively.

[0071] Optionally, step S3 includes the following sub-steps:

[0072] S3.1: Use an asynchronous I / O model to create independent coroutines for each multi-source heterogeneous database API request, and use an event loop for concurrent scheduling to obtain raw data in parallel;

[0073] S3.2: Dynamically select and load the corresponding parser according to the original data format, perform adaptive parsing on the acquired original data, and adopt a breakpoint resume mechanism with retry and backoff strategies to ensure the stability of the acquired original data.

[0074] S3.3: The original data is temporarily stored in a distributed file system or object storage, along with metadata files that record the source, timestamp, and version hash value, for use in tracing the source or verifying the consistency of the parsed data.

[0075] Optionally, step S4 includes the following sub-steps:

[0076] S4.1: First, based on rule-based methods, deterministic information is extracted from the unstructured text of the parsed raw data to obtain protein information. The deterministic information is protein mutation site information. The rule-based methods include regular expression method and domain keyword dictionary method. Then, based on the BERT question answering model, preset questions are set, and the preset questions and answers are used to extract complex semantic information from the protein information in the unstructured text of the raw data. The complex semantic information includes protein function description information and protein binding affinity numerical information.

[0077] S4.2: Construct an entity mapping table between protein information entity terms, use the entity mapping table to determine the target entity in the multi-source heterogeneous database, and normalize the target entity;

[0078] S4.3: Use a star schema to store normalized data in a relational database and enforce foreign key constraints and transaction control to ensure data consistency.

[0079] Optionally, step S5 includes the following sub-steps:

[0080] S5.1: Define a graph pattern in the field of protein design based on protein design requirements. The graph pattern includes protein node types and protein relationship types.

[0081] S5.2: Construct a knowledge corpus based on the aforementioned relational database;

[0082] S5.3: Through GraphRAG's retrieval-enhancement-generation and loop, a large language model is used to extract knowledge triples that conform to the graph pattern from the corpus, and knowledge verification and deduplication are performed. The structure of the knowledge triple is: entity term-relation-entity term.

[0083] S5.4: Store the knowledge triples in a graph database and use a graph embedding algorithm to predict links in the knowledge triples in the graph database to determine potential relationships between entity terms in the knowledge triples that do not appear in the graph database.

[0084] S5.5: Construct a protein design knowledge graph based on the potential relationships between the graph database and entity terms that do not appear in the graph database;

[0085] S5.6: Enhance the protein design knowledge graph by introducing domain-specific attributes, resulting in a domain-enhanced protein design knowledge graph. The domain-specific attributes include: protein physicochemical properties, conserved domains, and binding sites.

[0086] In this embodiment, regarding the issue of structured and unstructured protein data, the definitions and differences between them are as follows:

[0087] Structured data refers to data with a fixed format and clearly defined fields, primarily sourced from specialized biological databases, including:

[0088] Protein sequences, functional annotations, and accession numbers in the UniProt database.

[0089] 3D structure coordinate files, etc., in the PDB database.

[0090] This data is typically stored in standardized formats such as FASTA, XML, and JSON, with clearly defined fields that are easy for computers to parse and query.

[0091] Unstructured data refers to free text that lacks a predefined data model, and mainly includes:

[0092] Abstracts and full texts of scientific literature stored in the literature database.

[0093] This type of data typically contains rich semantic information such as the functional mechanisms and interaction backgrounds of proteins. However, due to the lack of a fixed format, computers cannot directly perform accurate logical retrieval or numerical calculations. They must rely on natural language processing techniques (such as the large model in this case) for semantic parsing.

[0094] Please see Figure 2 This application also provides a protein design knowledge graph construction system based on multi-agent technology, comprising:

[0095] The topic parsing and semantic query generation module is used to receive natural language research topics input by users, perform entity recognition and intent analysis on the natural language research topics using a large language model to obtain entity terms and intent objects, automatically standardize the entity terms based on entity standardization and semantic alignment to obtain standardized entity terms, and generate query strategies for multi-source heterogeneous databases based on the standardized entity terms and intent objects.

[0096] The multi-source heterogeneous data parallel acquisition and preprocessing module is used to execute the query strategy using an asynchronous concurrent scheduling mechanism, acquire raw data from multi-source heterogeneous databases in parallel using the query strategy, and perform adaptive parsing and temporary storage on the raw data. The raw data includes protein structured data and protein unstructured data. The protein structured data is data from a protein database, and the protein unstructured data is literature document data related to proteins.

[0097] The structured knowledge extraction and relational database construction module is used to perform mixed information extraction and cross-source entity alignment on the parsed raw data, and store the processed data in a relational database.

[0098] A knowledge graph construction module for protein design is used to automatically extract knowledge triples from the relational database using the GraphRAG framework, perform knowledge verification and deduplication, and store the knowledge triples in the graph database. At the same time, graph embedding algorithms and link prediction are used to perform knowledge completion and reasoning to form a domain-enhanced protein design knowledge graph.

[0099] Optionally, the topic parsing and semantic query generation module is further configured as follows:

[0100] The design of entity extraction prompts uses a large language model to perform entity recognition on the natural language research topic through prompts, extracts biological entity categories and model named entity categories from the natural language research topic, uses biological entity categories and model named entity categories as entity terms, and constructs an entity list by one-to-one correspondence between biological entity categories and model named entity categories;

[0101] Analyze the relationships between entity terms in the entity list, and generate an intent object containing actions, goals, and context based on the relationships between entity terms;

[0102] The BioBERT model is used to vectorize the identified entity terms, and the semantic similarity between the entity vectorization and the authoritative thesaurus terms is calculated. When the semantic similarity is greater than a preset threshold, the authoritative thesaurus terms are used as the standardized entity terms corresponding to the entity vectorization. Based on the standardized entity terms and intent objects, executable query instructions for multi-source heterogeneous databases are automatically generated.

[0103] Optionally, the multi-source heterogeneous data parallel acquisition and preprocessing module uses an asynchronous I / O model for concurrent scheduling to acquire raw data in parallel;

[0104] Meanwhile, an adaptive parser is used to dynamically parse the data according to the original data format, adaptively parsing the acquired original data and implementing breakpoint resumption with retry and backoff mechanisms.

[0105] The structured knowledge extraction and relational database construction module uses a hybrid method based on rules and BERT question answering model to extract information and builds an entity mapping table to achieve cross-source entity alignment and associate IDs from different databases.

[0106] In this embodiment, the IDs associated with different databases are linked using UniProtAccessionNumber as the central anchor. UniProtAccessionNumber is the Uniprot database accession number, a unique identifier assigned to each entry in the Uniprot database. Uniprot is a comprehensive protein sequence and function database, providing a wealth of protein information, including sequence, structure, function, and interactions. The accession number is a unique identifier for each protein entry, facilitating retrieval and citation.

[0107] Optionally, the protein-oriented knowledge graph construction module is executed using the GraphRAG framework, which includes:

[0108] Define a mapping pattern in the field of protein design;

[0109] Construct a knowledge corpus;

[0110] Knowledge triples are extracted using a large language model through a retrieval-enhancement-generation loop.

[0111] Perform knowledge verification and deduplication;

[0112] Store the triples in a graph database and use graph embedding algorithms for link prediction and knowledge completion.

[0113] The graph embedding algorithm uses the ComplEx or RotatE model.

[0114] Example 1

[0115] Knowledge base construction for the topic of SARS-CoV-2 virus antibody design.

[0116] This embodiment takes the topic of "antibody design research against SARS-CoV-2 virus" as an example to describe in detail the specific workflow and technical implementation of the system of the present invention.

[0117] Step 1: Topic Analysis and Semantic Query Generation (Agent A)

[0118] Users input their research topic through the front-end interactive interface (developed based on Vue.js): "Antibody design research against SARS-CoV-2 virus".

[0119] 1) Entity and Intent Recognition: The system calls a large language model (such as DeepSeekAPI or QwenAPI) to identify core entities including SARS-CoV-2 (virus), antibody (biomolecule category), and design (research action). Semantic intent objects are then generated through prompting engineering.

[0120] Action: "Design";

[0121] Target: "Antibody";

[0122] Context: "SARS-CoV-2 virus";

[0123] 2) Entity standardization and alignment: The system maps the identified entities to authoritative thesaurus, for example, aligning "SARS-CoV-2" to the MeSH term D000086382 and generating a unique identifier.

[0124] 3) Query Strategy Generation: Based on semantic intent, the system automatically generates query objects for multiple databases (i.e., multi-source heterogeneous databases), for example:

[0125] PubMed query string: ("SARS-CoV-2"[MeSH]ANDantibody[Title / Abstract]ANDdesign);

[0126] UniProt query URL: https: / / rest.uniprot.org / uniprotkb / search?query=SARS-CoV-2.

[0127] 2. Step 2: Parallel Acquisition and Preprocessing of Multi-Source Heterogeneous Data (Agent B)

[0128] 1) The system initiates parallel requests to databases such as PubMed, UniProt, and PDB based on an asynchronous I / O concurrent scheduling mechanism.

[0129] 2) The system adaptively parses the returned results, for example, retrieving XML literature abstracts from PubMed, protein sequence information in JSON format from UniProt, and structure files (.pdb) from PDB. All raw data is temporarily stored and accompanied by versioned metadata (timestamp, source database, query strategy hash value) to ensure subsequent traceability or consistency verification.

[0130] 3. Step 3: Structured Knowledge Extraction and Relational Database Construction (Agent C)

[0131] 1) Information Extraction: Key information, such as "binding site of neutralizing antibody" and "impact of mutation on antibody efficacy," was extracted from PubMed literature abstracts using the BERT question-answering model. Simultaneously, rule-based methods were used to quickly identify sequence variations (e.g., p.N501Y).

[0132] 2) Entity alignment: Using UniProtAccession, a mapping relationship is established between the SARS-CoV-2 spike protein (SpikeProtein, P0DTC2) mentioned in the literature and the PDBID (such as 6M0J) in the structural database.

[0133] 3) Relational database import: The extracted entities (proteins, antibodies, mutations, diseases) and their corresponding relationships (such as "antibody-target binding") are stored in a PostgreSQL database to ensure data consistency and integrity.

[0134] 4. Step 4: Knowledge Graph Construction for Protein Design (AgentD)

[0135] 1) GraphRAG-based knowledge extraction: Utilizing the GraphRAG framework, relevant fragments are retrieved from structured databases and unstructured documents, and knowledge triples are automatically extracted using LLM-enhanced generation. For example:

[0136] (AntibodyX,INTERACTS_WITH,SARS-CoV-2SpikeProtein);

[0137] (SpikeProtein,HAS_MUTATION,N501Y);

[0138] (N501Y,ASSOCIATED_WITH, antibody neutralization escape);

[0139] 2) Knowledge verification and deduplication: Automatically filter duplicate triples to ensure uniqueness.

[0140] 3) Graph storage and reasoning: The triples are stored in the Neo4j graph database, and link prediction is performed using graph embedding algorithms (such as RotatE) to discover potential unverified relationships, such as predicting that "antibody Y may have reduced binding affinity with mutant strain N501Y".

[0141] After discovering potential unverified relationships, new triples are constructed using these relationships. For example, if "protein A interacts with protein B" and "protein B interacts with protein C", then the potential unverified relationship "protein A interacts with protein C" can be discovered. At this point, a new triple "protein A-interaction-protein C" can be generated based on the potential relationship. Incorporating the new triple into the protein design knowledge graph can provide new design ideas for protein design. At the same time, to avoid potential relationships interfering with known protein design ideas, the triples constructed from potential relationships need to be different from the triples constructed from known relationships. This difference can be reflected in the triple information. For example, when extracting triples from known relationship structures, the font or ID is normal, while when extracting triples constructed from potential relationships, the font or ID is special.

[0142] 4) Domain Enhancement and Application: Add antibody physicochemical properties (molecular weight, affinity value) as node attributes to the graph, and introduce structural domain nodes (such as RBD regions) to form a multi-level knowledge graph for protein design, providing intelligent knowledge support for subsequent antibody optimization and new drug development.

[0143] Through the above implementation process, the system of this invention can transform natural language research topics end-to-end into a structured, reasonable protein design knowledge graph, achieving a significant improvement in the efficiency of multi-source heterogeneous data integration and the value of knowledge application.

[0144] Example 2

[0145] Construction of a knowledge base based on the theme of drug design targeting the HER2 protein.

[0146] This embodiment takes "targeted drug design research against HER2 protein" as an example to illustrate in detail the workflow and technical implementation of the system of the present invention in practical applications.

[0147] Step 1: Topic Analysis and Semantic Query Generation (Agent A)

[0148] The user inputs their research topic through the front-end interface: "Design and research of targeted drugs for HER2 protein". The system then performs the following operations:

[0149] 1) Entity and intent recognition.

[0150] The system calls a large language model (such as DeepSeekAPI or QwenAPI) to identify the core entities and research intent:

[0151] Action: "Design"

[0152] Target: "HER2 protein"

[0153] Context: "Targeted drugs".

[0154] 2) Entity standardization and alignment.

[0155] The system maps the identified entities to authoritative thesaurus:

[0156] "HER2 protein" is aligned to UniProtAccessionP04626;

[0157] Align "drugs" to MeSH term D000068206;

[0158] Meanwhile, BioBERT is used to vectorize entities and perform semantic matching with the domain vocabulary to solve the problems of synonyms, abbreviations and non-standard naming.

[0159] 3) Generation of multi-database query strategies.

[0160] The system automatically generates multi-source queries based on semantic intent and standardized entities:

[0161] PubMed query string: ("HER2"[MeSH]ANDdrug[Title / Abstract]ANDdesign);

[0162] UniProt query URL: https: / / rest.uniprot.org / uniprotkb / search?query=HER2;

[0163] PDB query URL: https: / / files.rcsb.org / download / 1N8Z.pdb.

[0164] 2. Step Two: Parallel Acquisition and Preprocessing of Multi-Source Heterogeneous Data (Agent B)

[0165] 1) The system uses asynchronous I / O concurrent scheduling to initiate parallel requests to databases such as PubMed, UniProt, and PDB.

[0166] 2) Adaptive parsing of response data:

[0167] PubMed documents are parsed using XML format to extract abstracts and keywords;

[0168] UniProt returns protein information in JSON format;

[0169] PDB provides structure files (.pdb).

[0170] 3) The original data is temporarily stored in the local distributed file system with versioned metadata (acquisition time, source, query strategy hash value) to ensure traceability.

[0171] 3. Step Three: Structured Knowledge Extraction and Relational Database Construction (Agent C)

[0172] 1) Information extraction.

[0173] The BERT question-answering model was used to extract protein function, drug binding sites, and the effects of mutations from PubMed literature abstracts;

[0174] Use regular expressions to quickly identify mutation information (such as p.L755S).

[0175] 2) Cross-source entity alignment.

[0176] A mapping was established between the HER2 protein (UniProtP04626) mentioned in the literature and PDBID (such as 1N8Z);

[0177] Construct an entity mapping table to achieve "one entity, one unique ID".

[0178] 3) Importing data into relational databases.

[0179] Entities (proteins, drugs, mutations, diseases) and their relationships (such as "drug-target binding") are stored in a PostgreSQL database to ensure data consistency and integrity.

[0180] 4. Step Four: Knowledge Graph Construction for Protein Design (AgentD)

[0181] 1) Knowledge extraction based on GraphRAG: Utilizing GraphRAG's retrieval-enhancement-generation mechanism, knowledge triples are extracted from structured databases and literature, for example:

[0182] (Drug A, INTERACTS WITH, HER2 protein);

[0183] (HER2 protein, HAS_MUTATION, L755S) and;

[0184] (L755S,ASSOCIATED_WITH, drug resistance).

[0185] 2) Knowledge verification and deduplication.

[0186] Automatically filter duplicate triples to ensure the uniqueness of the spectrum.

[0187] 3) Graph storage and reasoning.

[0188] The triples are stored in the Neo4j graph database;

[0189] Link prediction can be performed using graph embedding algorithms (such as RotatE) to infer potential relationships, such as predicting that "drug B may reduce the binding affinity of the L755S mutant".

[0190] Similarly, as a potential relationship discovered during drug design, if it is inferred that "drug B may reduce the binding affinity to the L755S mutant strain," a new triplet will be constructed from this inferred potential relationship. The new triplet will be incorporated into the protein design knowledge graph, which can provide new design ideas for drug design. At the same time, in order to avoid potential relationships interfering with known drug design ideas, the triplet constructed from the potential relationship needs to be different from the triplet constructed from the known relationship framework. This difference can be reflected in the triplet information. For example, when extracting the triplet constructed from the known relationship framework, it is in a normal font or with a normal ID, while when extracting the triplet constructed from the potential relationship, it is in a special font or with a special ID.

[0191] 4) Domain enhancement and application.

[0192] Add physicochemical properties (molecular weight, affinity, etc.) to drug nodes;

[0193] Added HER2 domain and binding site nodes, which are associated with protein nodes;

[0194] Construct a multi-level knowledge graph for drug design to provide intelligent knowledge support for drug optimization and new drug development.

[0195] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for constructing a multi-agent based protein design knowledge graph, characterized in that, The method comprises the following steps: S1: obtaining a natural language research topic input by a user; S2: performing entity recognition and intent analysis on the natural language research topic by using a large language model to obtain entity terms and intent objects, performing standardization processing on the entity terms based on entity standardization and semantic alignment to obtain standardized entity terms, and generating a query strategy for a multi-source heterogeneous database based on the standardized entity terms and the intent objects; S3: executing the query strategy using an asynchronous concurrent scheduling mechanism, obtaining raw data from the multi-source heterogeneous database in parallel using the query strategy, and performing adaptive parsing and temporary storage on the raw data, wherein the raw data includes protein structured data and protein unstructured data, the protein structured data is data in a protein database, and the protein unstructured data is literature document data related to proteins; S4: performing mixed information extraction and cross-source entity alignment processing on the parsed raw data, and storing the processed data to a relational database; S5: extracting knowledge triples from the relational database by using a GraphRAG framework, performing knowledge verification and deduplication, storing the knowledge triples to a graph database, and using graph embedding algorithms and link prediction to perform knowledge completion and reasoning, thereby forming a domain-enhanced protein design knowledge graph. 2.The multi-agent based protein design knowledge graph construction method of claim 1, wherein, The S2 comprises the following sub-steps: S2.1: designing entity extraction prompt words, performing entity recognition on the natural language research topic by using a large language model through the prompt words, extracting biological entity categories and model named entity categories in the natural language research topic, taking the biological entity categories and the model named entity categories as entity terms, and constructing the biological entity categories and the model named entity categories into an entity list in a one-to-one correspondence; S2.2: analyzing the relationships between the entity terms in the entity list, and generating intent objects containing actions, targets and contexts based on the relationships between the entity terms; S2.3: performing entity vectorization on the recognized entity terms by using a BioBERT model, calculating semantic similarity of entity vectorization and authoritative glossary terms, taking an authoritative glossary term as a standardized entity term of an entity vectorization corresponding entity term when the semantic similarity is greater than a preset threshold, and automatically generating executable query instructions for a multi-source heterogeneous database based on the standardized entity terms and the intent objects. 3.The multi-agent based protein design knowledge graph construction method of claim 2, wherein, The semantic similarity calculation adopts a cosine similarity formula: ； wherein, is the cosine similarity, is the vector angle of entity vectorization, A and B are the vector representations of entity terms and authority vocabulary terms, respectively. 4.The multi-agent based protein design knowledge graph construction method of claim 1, wherein, The S3 comprises the following sub-steps: S3.1: creating independent coroutines for each multi-source heterogeneous database API request using an asynchronous I / O model, and performing concurrent scheduling through an event loop to obtain raw data in parallel; S3.2: dynamically selecting and loading corresponding parsers according to raw data formats, adaptively parsing the obtained raw data, and using a breakpoint resume mechanism with a retry and backoff strategy to ensure stability of raw data acquisition; S3.3: temporarily storing the raw data in a distributed file system or an object storage, and attaching metadata files recording source, timestamp and version hash value, for tracing or consistency checking of the parsed data. 5.The multi-agent based protein design knowledge graph construction method of claim 1, wherein, The S4 comprises the following sub-steps: S4.1: First, deterministic information in protein information is extracted from unstructured text in parsed raw data based on a rule method, wherein the deterministic information is protein mutation point information, and the rule method comprises a regular expression method and a domain keyword dictionary method; then, a preset question is set based on a BERT question and answer model, and complex semantic information is extracted from protein information in unstructured text in raw data using the preset question and the question answer, wherein the complex semantic information comprises protein function description information and protein binding affinity numerical information; S4.2: An entity mapping table between protein information entity terms is constructed, and a target entity is determined in a multi-source heterogeneous database using the entity mapping table, and the target entity is normalized; S4.3: The normalized data is stored in a relational database using a star model, and foreign key constraints and transaction controls are forced to ensure data consistency. 6.The multi-agent based protein design knowledge graph construction method of claim 1, wherein, The S5 comprises the following sub-steps: S5.1: Defining a graph pattern of the protein design field based on protein design requirements, wherein the graph pattern comprises protein node types and protein relationship types; S5.2: Constructing a knowledge corpus based on the relational database; S5.3: Extracting knowledge triples conforming to the graph pattern from the corpus using a large language model, and performing knowledge verification and deduplication, wherein the structure of the knowledge triples is: entity term-relation-entity term; S5.4: Storing the knowledge triples into a graph database, and using a graph embedding algorithm to perform link prediction on the knowledge triples in the graph database to determine potential relationships between entity terms in the knowledge triples that do not appear in the graph database; S5.5: Constructing a protein design knowledge graph based on the graph database and the potential relationships between entity terms that do not appear in the graph database; S5.6: Introducing domain-specific attributes into the protein design knowledge graph to enhance the protein design knowledge graph, and obtaining a domain-enhanced protein design knowledge graph, wherein the domain-specific attributes include protein physicochemical properties, conserved domains, and binding sites. 7.A multi-agent based protein design knowledge graph construction system, characterized by, It comprises: A topic analysis and semantic query generation module for receiving a user input natural language research topic, performing entity recognition and intent analysis on the natural language research topic using a large language model to obtain entity terms and intent objects, automatically standardizing the entity terms based on entity standardization and semantic alignment to obtain standardized entity terms, and generating a query strategy for a multi-source heterogeneous database based on the standardized entity terms and the intent objects; A multi-source heterogeneous data parallel acquisition and preprocessing module for executing the query strategy using an asynchronous concurrent scheduling mechanism, acquiring raw data from a multi-source heterogeneous database in parallel using the query strategy, and performing adaptive parsing and temporary storage on the raw data, wherein the raw data comprises protein structured data and protein unstructured data, the protein structured data is data in a protein database, and the protein unstructured data is literature document data related to proteins; A structured knowledge extraction and relational database construction module is configured to perform hybrid information extraction and cross-source entity alignment on the parsed raw data, and store the processed data into a relational database. A knowledge graph construction module for protein design is configured to automatically extract knowledge triples from the relational database using a GraphRAG framework, perform knowledge verification and deduplication, and store the knowledge triples into a graph database, while using a graph embedding algorithm and link prediction to perform knowledge completion and reasoning, thereby forming a domain-enhanced knowledge graph for protein design.

8. The multi-agent based protein design knowledge graph construction system of claim 7, wherein, The subject analysis and semantic query generation module is further configured to: design entity extraction prompt words, use a large language model to perform entity recognition on the natural language research subject through the prompt words, extract biological entity categories and model named entity categories in the natural language research subject, and construct the biological entity categories and the model named entity categories into an entity list in a one-to-one correspondence as entity terms; analyze the relationships between the entity terms in the entity list, and generate an intent object containing actions, targets, and contexts based on the relationships between the entity terms; use a BioBERT model to perform entity vectorization on the recognized entity terms, calculate the semantic similarity between the entity vectorization and the authoritative vocabulary terms, and when the semantic similarity is greater than a preset threshold, use the authoritative vocabulary terms as the standardized entity terms of the corresponding entity terms of the entity vectorization, and automatically generate executable query instructions for the multi-source heterogeneous database based on the standardized entity terms and the intent object. 9.The multi-agent based protein design knowledge graph construction system of claim 7, wherein, The multi-source heterogeneous data parallel acquisition and preprocessing module uses an asynchronous I / O model for concurrent scheduling and parallel acquisition of raw data. At the same time, an adaptive parser is used to dynamically parse data according to the format of the raw data, and the acquired raw data is adaptively parsed, and a breakpoint resume mechanism with retry and backoff mechanisms is implemented. The structured knowledge extraction and relational database construction module uses a hybrid method based on a rule-based method and a BERT question and answer model for information extraction, and constructs an entity mapping table to realize cross-source entity alignment and associate the IDs of different databases.

10. The multi-agent based protein design knowledge graph construction system of claim 7, wherein, The knowledge graph construction module for protein design uses a GraphRAG framework, which includes: defining a graph schema for the field of protein design; constructing a knowledge corpus; extracting knowledge triples using a large language model through retrieval-enhancement-generation; performing knowledge verification and deduplication; storing the triples into a graph database and using a graph embedding algorithm for link prediction and knowledge completion; The graph embedding algorithm uses a ComplEx or RotatE model.

Images