Systems, methods, media, and products for crop genomics

Abstract

The present application relates to a computer system using a calculation model, and discloses a system, method, medium and product for crop genetic question answering. A system for crop genetic question answering comprises a computing resource, a storage unit, a large language model, a planning agent, an execution agent and a knowledge mining module. The storage unit is used for storing a knowledge graph. The large language model is used for: receiving an input; generating a task description; generating an answer to a crop genetic question based on a query output. The planning agent is used for: determining a target tool; generating a tool calling instruction. The execution agent is used for: generating a function calling instruction. The knowledge mining module is used for: providing a tool description document; calling the target tool to perform a structured query on the knowledge graph to obtain the query output. The system according to the present application overcomes the limitation that the large language model lacks professional knowledge in the field of crop genetics, and improves the accuracy of the large language model in answering crop genetic questions.

Description

Technical Field

[0001] This invention generally relates to computer systems utilizing computational models, and more specifically to systems, methods, media, and products for crop gene question answering. Background Technology

[0002] Crop gene function research is a core part of modern agricultural breeding. This research usually requires systematically summarizing the known functions, unknown areas, and scientific controversies of specific genes or agronomic traits. It relies heavily on researchers manually searching and cross-validating massive amounts of unstructured literature and structured databases, resulting in a large amount of manual operation and a very high knowledge threshold. It also suffers from technical bottlenecks such as low efficiency, long time consumption, and susceptibility to subjective experience.

[0003] Existing crop gene function research can directly utilize general-purpose large language models for open-domain question answering. These models, based on general knowledge acquired during pre-training, directly generate text answers to natural language questions posed by users. However, in the highly specialized field of crop genetics, the content generated by these general-purpose large language models lacks reliable data source support, making them prone to factual errors or omissions of crucial evidence, resulting in poor professional accuracy.

[0004] There is a need in this field for improvements to crop gene question-answering technology at at least one of the aforementioned levels. Summary of the Invention

[0005] This invention is provided to further improve crop gene question answering technology by leveraging large language models, agents, tools, and knowledge graphs for interaction.

[0006] One aspect of the present invention provides a system for crop gene question answering, comprising: computing resources; a storage unit configured to store a knowledge graph including a crop gene dataset; a large language model configured to invoke the computing resources to: receive input from a user, the input including a crop gene question; generate a task description based on the input; and generate an answer to the crop gene question based on a query output; a planning agent configured to invoke the large language model to: receive the input, the task description, and a tool description document; determine a target tool to be invoked from a plurality of crop gene tools based on the input, the task description, and the tool description document; and generate a tool invocation instruction based on the target tool, the tool invocation instruction including an indication of the target tool; and an execution agent configured to invoke the large language model to: receive the input, the task description, the tool description document, and the tool invocation instruction; and generate a function invocation instruction based on the input, the task description, the tool description document, and the tool invocation instruction, the function invocation instruction including the... The system includes a target tool instruction and structured information associated with the crop genetic question in a format conforming to the target tool; and a knowledge mining module configured to: provide the tool documentation; receive the function call instruction; and, based on the structured information in the function call instruction, invoke the target tool indicated in the function call instruction to perform a structured query on the knowledge graph to obtain the query output, wherein the answer to the crop genetic question is generated by the large language model based on the query output, wherein the knowledge mining module includes the plurality of crop genetic tools, the plurality of crop genetic tools including a multi-trait-gene mapping tool, the multi-trait-gene mapping tool being configured to: generate a structured query sequence based on the structured information in the function call instruction, the structured information including crops and multiple traits associated with the crop genetic question, the structured query sequence being used to perform a search for genes associated with the multiple traits of the crop and perform logical operations on the search results; and execute the structured query sequence on the knowledge graph to obtain the query output.

[0007] In the system described above, the planning agent is configured to invoke the large language model to: filter candidate tools associated with the crop gene problem from a list of crop gene tools, which includes the plurality of crop gene tools, using a directed acyclic graph; and determine the target tool from the candidate tools based on few-sample cue words and the crop gene problem.

[0008] In the system described above, in response to the tool invocation instruction indicating that the target tool is the multi-trait-gene mapping tool, the executing agent is configured to invoke the large language model to: identify the crop and the multiple traits from the input; and, based on the crop and the multiple traits, generate the function invocation instruction in accordance with the format specified in the tool description document of the multi-trait-gene mapping tool, the function invocation instruction including the indication of the multi-trait-gene mapping tool and structured information of the crop and the multiple traits conforming to the format of the multi-trait-gene mapping tool.

[0009] In the system described above, the multi-trait-gene mapping tool is configured to: map the multiple traits in the structured information of the function call instruction to standard trait annotation fields; generate the structured query sequence based on the crop in the structured information, the mapped standard trait annotation fields, and the template sequence; execute the structured query sequence on the database associated with the crop in the knowledge graph to obtain gene records associated with the multiple traits; and perform statistics on the gene records to generate a structured summary as the query output.

[0010] In the system described above, the plurality of crop gene tools include a gene function summarization tool, which is configured to execute the function call instruction on the knowledge graph to obtain the query output, wherein the structured information in the function call instruction includes a structured query sequence for performing a lookup of the function of genes associated with the crop gene question.

[0011] In the system described above, the gene function summarization tool is configured to: execute the structured query sequence in the function call instruction on the knowledge graph to obtain query results associated with the gene in different types of node data in the knowledge graph; and concatenate the query results into query text as the query output.

[0012] In the system described above, in response to the tool invocation instruction indicating that the target tool is the gene function summarization tool, the executing agent is configured to invoke the large language model to: identify crop and gene identifiers from the input; generate extended identifiers in other naming conventions based on the gene identifiers; and generate the function invocation instruction based on the crop, the gene identifiers, and the extended identifiers, in accordance with the format specified in the tool description document of the gene function summarization tool.

[0013] In the system described above, in response to the target tool to be invoked being identified as the multi-trait-gene mapping tool, the planning agent is configured to invoke the large language model to: generate a first tool invocation instruction based on the multi-trait-gene mapping tool, the first tool invocation instruction including an indication of the multi-trait-gene mapping tool; the execution agent is configured to invoke the large language model to: generate a first function invocation instruction based on the input, the task description, the first tool documentation, and the first tool invocation instruction, the first function invocation instruction including an indication of the multi-trait-gene mapping tool and first structured information related to the crop genetic problem in a format conforming to the multi-trait-gene mapping tool; the knowledge mining module is configured to: invoke the multi-trait-gene mapping tool to perform a first structured query on the knowledge graph based on the first structured information in the first function invocation instruction to obtain a first query output; the planning agent is further configured to invoke the large language model to: receive the first query output; in response to determining that the first query output includes information related to the multiple traits... The gene function summarization module is configured to: identify the target tool to be invoked as the gene function summarization tool; and generate a second tool invocation instruction based on the gene function summarization tool, the second tool invocation instruction including the indication of the gene function summarization tool; the executing agent is further configured to invoke the large language model to: receive the first query output, the second tool description document, and the second tool invocation instruction; and generate a second function invocation instruction based on the first query output, the second tool description document, and the second tool invocation instruction, the second function invocation instruction including the indication of the gene function summarization tool and second structured information associated with the first query output conforming to the format of the gene function summarization tool; the knowledge mining module is further configured to: receive the second function invocation instruction; and, based on the second structured information in the second function invocation instruction, invoke the gene function summarization tool to perform a second structured query on the knowledge graph to obtain a second query output, wherein the answer to the crop gene question is generated by the large language model based on the first query output and the second query output.

[0014] In the system described above, in response to the target tool to be invoked being identified as the gene function summarization tool, the planning agent is configured to invoke the large language model to: generate a first tool invocation instruction based on the gene function summarization tool, the first tool invocation instruction including an indication of the gene function summarization tool; the execution agent is configured to invoke the large language model to: generate a first function invocation instruction based on the input, the task description, a first tool documentation, and the first tool invocation instruction, the first function invocation instruction including an indication of the gene function summarization tool and first structured information related to the crop gene problem in a format conforming to the gene function summarization tool; the knowledge mining module is configured to: invoke the gene function summarization tool to perform a first structured query on the knowledge graph based on the first structured information in the first function invocation instruction to obtain a first query output; the planning agent is further configured to invoke the large language model to: receive the first query output; in response to determining that the first query output includes multiple traits associated with the gene, the target to be invoked... The tool is identified as the multi-trait-gene mapping tool; and a second tool invocation instruction is generated based on the multi-trait-gene mapping tool, the second tool invocation instruction including the indication of the multi-trait-gene mapping tool; the executing agent is further configured to invoke the large language model to: receive the first query output, the second tool documentation, and the second tool invocation instruction; and generate a second function invocation instruction based on the first query output, the second tool documentation, and the second tool invocation instruction, the second function invocation instruction including the indication of the multi-trait-gene mapping tool and second structured information associated with the first query output conforming to the format of the multi-trait-gene mapping tool; the knowledge mining module is further configured to: receive the second function invocation instruction; and, based on the second structured information in the second function invocation instruction, invoke the multi-trait-gene mapping tool to perform a second structured query on the knowledge graph to obtain a second query output, wherein the answer to the crop gene question is generated by the large language model based on the first query output and the second query output.

[0015] In the system described above, the large language model is trained based on the following: the mapping relationship between multiple gene identifiers of the same crop gene in different nomenclature schemes; and the mapping relationship between multiple expressions of the same trait and standard trait annotation fields.

[0016] Another aspect of the present invention provides a method for crop gene question answering, comprising the following steps: S1: receiving input from a user by a large language model, the input including a crop gene question; S2: generating a task description by the large language model based on the input; S3: invoking the large language model by a planning agent to determine a target tool to be invoked from among a plurality of crop gene tools included in the knowledge mining module based on the input, the task description, and a tool description document provided by the knowledge mining module; S4: invoking the large language model by the planning agent to generate a tool invocation instruction based on the target tool, the tool invocation instruction including an indication of the target tool; S5: invoking the large language model by an execution agent to generate a function invocation instruction based on the input, the task description, the tool description document, and the tool invocation instruction, the function invocation instruction including an indication of the target tool and a format of the target tool associated with the crop gene question. S6: The knowledge mining module, based on the structured information in the function call instruction, calls the target tool indicated in the function call instruction to perform a structured query on the knowledge graph to obtain a query output, wherein the knowledge graph is stored in a storage unit and includes a crop gene dataset; and S7: The large language model generates an answer to the crop gene question based on the query output, wherein the multiple crop gene tools include a multi-trait-gene mapping tool, the multi-trait-gene mapping tool being configured to: generate a structured query sequence based on the structured information in the function call instruction, the structured information including crops and multiple traits associated with the crop gene question, the structured query sequence being used to perform a search for genes associated with the multiple traits of the crop and to perform logical operations on the search results; and execute the structured query sequence on the knowledge graph to obtain the query output.

[0017] As described above, step S3 includes: S31: using a Directed Acyclic Graph (DAG) to screen candidate tools associated with the crop gene problem from a list of crop gene tools, the list of crop gene tools including the plurality of crop gene tools; and S32: determining the target tool from the candidate tools based on a few-sample cue words and the crop gene problem.

[0018] As described above, in response to the tool invocation instruction indicating that the target tool is the multi-trait-gene mapping tool, step S5 includes: S51: identifying the crop and the multiple traits from the input; and S52: generating the function invocation instruction based on the crop and the multiple traits, in accordance with the format specified in the tool documentation of the multi-trait-gene mapping tool, the function invocation instruction including the indication of the multi-trait-gene mapping tool and structured information of the crop and the multiple traits conforming to the format of the multi-trait-gene mapping tool.

[0019] As described above, the multi-trait-gene mapping tool is configured to: map the multiple traits in the structured information of the function call instruction to standard trait annotation fields; generate the structured query sequence based on the crop in the structured information, the mapped standard trait annotation fields, and the template sequence; execute the structured query sequence on the database associated with the crop in the knowledge graph to obtain gene records associated with the multiple traits; and perform statistics on the gene records to generate a structured summary as the query output.

[0020] As described above, the plurality of crop gene tools include a gene function summarization tool, which is configured to execute the function call instruction on the knowledge graph to obtain the query output, wherein the structured information in the function call instruction includes a structured query sequence for performing a lookup of the function of genes associated with the crop gene question.

[0021] As described above, the gene function summarization tool is configured to: execute the structured query sequence in the function call instruction on the knowledge graph to obtain query results associated with the gene in different types of node data in the knowledge graph; and concatenate the query results into query text as the query output.

[0022] As described above, in response to the tool call instruction indicating that the target tool is the gene function summarization tool, step S5 includes: S53: identifying crop and gene identifiers from the input; S54: generating extended identifiers in other naming conventions based on the gene identifiers; and S55: generating the function call instruction based on the crop, the gene identifiers, and the extended identifiers, in accordance with the format specified in the tool description document of the gene function summarization tool.

[0023] As described above, in response to the target tool to be invoked being determined to be the multi-trait-gene mapping tool, S4 includes: S41: generating a first tool invocation instruction based on the multi-trait-gene mapping tool, the first tool invocation instruction including an indication of the multi-trait-gene mapping tool; S5 includes: S56: generating a first function invocation instruction based on the input, the task description, the first tool documentation, and the first tool invocation instruction, the first function invocation instruction including an indication of the multi-trait-gene mapping tool and first structured information related to the crop gene problem in a format conforming to the multi-trait-gene mapping tool; S6 includes: S61: invoking the multi-trait-gene mapping tool to perform a first structured query on the knowledge graph to obtain a first query output based on the first structured information in the first function invocation instruction; S4 further includes: S42: in response to determining that the first query output includes information related to the multi-trait-gene mapping tool... The process involves: S43: generating a second tool invocation instruction based on the gene function summarization tool, the second tool invocation instruction including the instructions of the gene function summarization tool; S5 further includes: S57: generating a second function invocation instruction based on the first query output, the second tool documentation, and the second tool invocation instruction, the second function invocation instruction including the instructions of the gene function summarization tool and second structured information associated with the first query output conforming to the format of the gene function summarization tool; S6 further includes: S62: invoking the gene function summarization tool to perform a second structured query on the knowledge graph to obtain a second query output based on the second structured information in the second function invocation instruction; S7 includes: S71: generating an answer to the crop gene question by the large language model based on the first query output and the second query output.

[0024] As described above, in response to the target tool to be invoked being determined to be the gene function summarization tool, S4 includes: S44: generating a first tool invocation instruction based on the gene function summarization tool, the first tool invocation instruction including an indication of the gene function summarization tool; S5 includes: S58: generating a first function invocation instruction based on the input, the task description, the first tool documentation, and the first tool invocation instruction, the first function invocation instruction including an indication of the gene function summarization tool and first structured information related to the crop gene problem in a format conforming to the gene function summarization tool; S6 includes: S63: invoking the gene function summarization tool to perform a first structured query on the knowledge graph to obtain a first query output based on the first structured information in the first function invocation instruction; S4 further includes: S45: in response to determining that the first query output includes multiple traits associated with the gene, the tool to be invoked... The target tool is identified as the multi-trait-gene mapping tool; and S46: a second tool invocation instruction is generated based on the multi-trait-gene mapping tool, the second tool invocation instruction including the indication of the multi-trait-gene mapping tool; S5 further includes: S59: a second function invocation instruction is generated based on the first query output, the second tool documentation and the second tool invocation instruction, the second function invocation instruction including the indication of the multi-trait-gene mapping tool and second structured information associated with the first query output conforming to the format of the multi-trait-gene mapping tool; S6 further includes: S64: based on the second structured information in the second function invocation instruction, the multi-trait-gene mapping tool is invoked to perform a second structured query on the knowledge graph to obtain a second query output; S7 includes: S71: the large language model generates an answer to the crop gene question based on the first query output and the second query output.

[0025] As described above, the large language model is trained based on the following: the mapping relationship between multiple gene identifiers of the same crop gene in different nomenclature specifications; and the mapping relationship between multiple expressions of the same trait and standard trait annotation fields.

[0026] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0027] Another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any of the above embodiments.

[0028] The system and method of the present invention overcome the limitation of large language models lacking expertise in the field of crop genes. By interacting with an intelligent agent to invoke tools to retrieve knowledge graphs, the accuracy of large language models in answering crop gene questions is improved. Attached Figure Description

[0029] Various embodiments of the present invention are described in conjunction with the accompanying drawings.

[0030] Figure 1 This is a block diagram of a system for crop gene question answering according to some embodiments of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating the interaction between the system and the user according to some embodiments of the present invention.

[0032] Figure 3 This is a schematic diagram of a system for answering crop gene questions according to some embodiments of the present invention.

[0033] Figure 4 This is a flowchart of a method for crop gene question answering according to some embodiments of the present invention.

[0034] Figure 5 This is a flowchart of a first process associated with a method for crop gene question answering according to some embodiments of the present invention.

[0035] Figure 6 This is a flowchart of a second process associated with a method for crop gene question answering according to some embodiments of the present invention.

[0036] Figure 7 This is a flowchart of a third process associated with a method for crop gene question answering according to some embodiments of the present invention.

[0037] Figure 8 This is a flowchart of a fourth process associated with a method for crop gene question answering according to some embodiments of the present invention.

[0038] Figure 9 This is a flowchart of a fifth process associated with a method for crop gene question answering according to some embodiments of the present invention.

[0039] Figure 10 This is a flowchart of a sixth process associated with a method for crop gene question answering according to some embodiments of the present invention.

[0040] Figure 11 This is a block diagram of a computer-readable storage medium according to some embodiments of the present invention.

[0041] Figure 12This is a block diagram of a computer program product according to some embodiments of the present invention.

[0042] Figure 13 This is a performance comparison of the embodiments of the present invention with existing methods for answering crop gene questions. Detailed Implementation

[0043] In this application, the term "agent" refers to an agent capable of perceiving the environment and taking actions to perform specific goals. An agent primarily refers to software code. An agent can be executed by the system's computing resources. An agent can invoke corresponding models and tools through an Application Programming Interface (API) to interact with various forms of input or implement corresponding functions.

[0044] In this application, ordinal numbers such as "first," "second," and "third" are used to distinguish different instances of objects with the same name. The ordinal numbers "first," "second," and "third" do not indicate a relative order of the indicated objects in time, space, sequence, or other aspects.

[0045] According to one aspect of the present invention, a system for crop gene question answering is provided.

[0046] Figure 1 This is a block diagram of a system 100 for crop gene question answering according to some embodiments of the present invention.

[0047] System 100 can be a local or remote computer, server, etc. System 100 may include computing resources 110, storage unit 120, large language model 130, planning agent 140, execution agent 150, and knowledge mining module 160. Storage unit 120 can be configured to store knowledge graph 122. Knowledge graph 122 may include crop gene datasets.

[0048] In some embodiments, computing resources 110 may include a central processing unit (CPU), a graphics processing unit (GPU), and various other processing units or cores (e.g., arithmetic logic units, integer units, floating-point units, tensor units, ray tracing cores, etc.).

[0049] The large language model 130 can be configured to invoke computing resources 110 to perform corresponding operations. In some embodiments, the large language model 130 can be deployed locally on system 100. In some embodiments, the large language model 130 can be deployed remotely to system 100, for example, in the cloud.

[0050] The planning agent 140 and the execution agent 150 can be configured to call the large language model 130 via API to perform corresponding operations.

[0051] The knowledge mining module 160 can be configured to invoke various tools to interact with various forms of input or to perform corresponding functions. The knowledge mining module 160 may include multiple crop gene tools. In some embodiments, the crop gene tools can be implemented via software code. As an example, a crop gene tool may be an algorithm for performing structured queries on the knowledge graph 122.

[0052] Figure 2 This is a schematic diagram illustrating the interaction between the system and the user according to some embodiments of the present invention. Figure 3 This is a schematic diagram of a system for answering crop gene questions according to some embodiments of the present invention. Figure 2 Come to Figure 3 The processing flow will be explained.

[0053] In the first frame (310), the large language model addresses the crop gene reception problem. For example... Figure 2 As shown, the large language model 130 receives crop gene questions in natural language form from the user.

[0054] In the second frame (320), the planning agent invokes instructions based on the crop gene problem generation tool. For example... Figure 2 As shown, the planning agent 140 receives relevant inputs on crop gene problems from the large language model 130 and tool description documents from the knowledge mining module 160. Then, the planning agent 140 determines the target tool based on the inputs and tool description documents and generates tool invocation instructions.

[0055] In the third box (330), the agent generates function call instructions based on tool call instructions. For example... Figure 2 As shown, the execution agent 150 receives relevant inputs on crop gene problems from the large language model 130, receives tool documentation from the knowledge mining module 160, and receives tool invocation instructions from the planning agent 140. Then, the execution agent 150 generates function call instructions based on the inputs, tool documentation, and tool invocation instructions.

[0056] In the fourth box, at position 340, the knowledge mining module performs a structured query to obtain the query output. For example... Figure 2 As shown, the knowledge mining module 160 receives function call instructions from the executing agent 150. Based on the function call instructions, the knowledge mining module 160 invokes the target tool to perform a structured query on the knowledge graph 122, thereby obtaining the query output.

[0057] In box 5, at position 350, the large language model generates an answer to the crop gene question based on the query output. For example... Figure 2 As shown, the large language model 130 receives query output from the knowledge mining module 160. Based on the query output, the large language model 130 generates a natural language-compliant answer to the crop gene question and then returns the answer to the user.

[0058] Based on the above processing flow, the user submits a crop gene question in natural language to the large language model. The large language model then instructs the planning agent to determine the target tool corresponding to the question, generates function call instructions for the execution agent, and subsequently calls the target tool in the knowledge mining module to perform a structured query on the knowledge graph to obtain accurate query results. Finally, the large language model integrates the query results and returns the answer in natural language to the user.

[0059] return Figure 1 The large language model 130 can be configured to invoke computing resources 110 to perform corresponding operations. The large language model 130 can be configured to receive input from a user, including a crop gene question. The large language model 130 can be configured to generate a task description based on the input. The large language model 130 can be configured to generate an answer to the crop gene question based on the query output.

[0060] The planning agent 140 can be configured to invoke the large language model 130 to perform corresponding operations. The planning agent 140 can be configured to receive input, a task description, and a tool specification document. The planning agent 140 can be configured to determine the target tool to be invoked from among multiple crop genomic tools based on the input, task description, and tool specification document. The planning agent 140 can be configured to generate tool invocation instructions based on the target tool. The tool invocation instructions may include indications of the target tool.

[0061] The executive agent 150 can be configured to invoke the large language model 130 to perform corresponding operations. The executive agent 150 can be configured to receive input, a task description, tool documentation, and tool invocation instructions. The executive agent 150 can be configured to generate function invocation instructions based on the input, task description, tool documentation, and tool invocation instructions. The function invocation instructions may include indications of the target tool and structured information related to the crop genetic problem in a format consistent with the target tool.

[0062] The knowledge mining module 160 can be configured to provide tool documentation. The knowledge mining module 160 can be configured to receive function call instructions. The knowledge mining module 160 can be configured to, based on the structured information in the function call instructions, invoke the target tool indicated in the function call instructions to perform a structured query on the knowledge graph to obtain query output.

[0063] Answers to crop gene questions can be generated by the large language model 130 based on query output.

[0064] Multiple crop genetic tools may include multi-trait-gene mapping tools. Multi-trait-gene mapping tools can be configured to generate structured query sequences based on structured information in function call instructions. The structured information may include the crop and multiple traits associated with the crop genetic question. The structured query sequences can be used to perform searches for genes associated with multiple traits of the crop and to perform logical operations on the search results. Multi-trait-gene mapping tools can also be configured to execute structured query sequences on a knowledge graph to obtain query output.

[0065] Some embodiments of this invention propose a system for crop gene question answering that includes a large language model and an intelligent agent. This system interacts with the intelligent agent through the large language model to intelligently decompose user-posed natural language questions into structured query sequences targeting a knowledge graph, thereby invoking tools to perform structured queries on the knowledge graph. This system uses the knowledge graph as its factual basis, ensuring the accuracy and professionalism of the output content and significantly improving the efficiency of crop gene question answering. The knowledge mining module in some embodiments includes a multi-trait-gene mapping tool that performs logical operations on the search results and obtains query output based on the results of these operations. Compared to some existing technologies that retrieve all search results for local analysis and processing, this method saves local resources, has higher processing efficiency, and takes less time.

[0066] While existing large language models can answer questions by directly querying knowledge graphs, for highly specialized questions like crop genetics, they require complex chained steps. For example, a large language model might first query the set satisfying the first condition, then the set satisfying the second condition, and finally find the intersection in external code. This results in an excessively long reasoning chain for the large language model, making it highly susceptible to data loss or errors in intermediate steps.

[0067] Some embodiments of the present invention encapsulate complex execution logic within the crop gene tool of the knowledge mining module, allowing the large language model to be responsible only for parameter extraction without having to handle logical operations on the knowledge graph, thereby improving execution efficiency and answer accuracy.

[0068] In some embodiments, the planning agent may be given a problem classification function based on prompt words. For example, prompt words are set for the planning agent so that when the planning agent calls a large language model, the prompt words tell the large language model that its task is to identify the user's intent and classify it into a specific crop gene tool.

[0069] In some embodiments, the executing agent may be given instruction output functionality based on a prompt word. For example, a prompt word is set for the executing agent so that when the executing agent calls a large language model, the prompt word informs the large language model that its task is to output instructions in a structured form of a function call.

[0070] By using prompt words to assign roles to agents, large language models can be instructed to output structured instructions that conform to specific format requirements when called by agents, rather than free-form natural language questions and answers. This method enables collaboration and division of labor among multiple agents.

[0071] In some embodiments, the tool documentation may include tool use cases and parameter interfaces. Tool use cases inform the agent of the tool's functionality. Parameter interfaces inform the agent how to invoke the tool; for example, they may specify parameter types and formats. For instance, for a multi-trait-gene mapping tool, where the tool use case is to identify genes corresponding to multiple traits, the parameter interface specifies that the `crop` parameter receives the crop name, and the `trait_list` parameter receives the list of traits.

[0072] In some embodiments, the task description can be the content relayed to the planning and execution agents by a large language model based on the crop gene question proposed by the user. The crop gene question proposed by the user may have problems such as unclear references, lack of context, or brief descriptions. After the large language model rewrites the crop gene question proposed by the user into a task description, the user's intent can be conveyed to the planning and execution agents more clearly.

[0073] In some embodiments, the input provided by the large language model to the planning agent and the execution agent may consist only of crop genetic questions. When the crop genetic questions provided by the user fully reflect the user's intent, such as having complete information on crop species, genetic identifiers, or traits, the large language model can directly provide the crop genetic questions as user input to the planning agent and the execution agent.

[0074] In some embodiments, the input provided by the large language model to the planning agent and the execution agent may include a crop genetic question and the user's historical dialogue with the large language model. While the crop genetic question provided by the user may not fully reflect the user's intent—for example, it may lack information such as crop species, genetic identifiers, or traits—the historical dialogue between the user and the large language model may include information relevant to the crop genetic question. The large language model can provide both the crop genetic question and the historical dialogue with the large language model as user input to the planning agent and the execution agent, allowing them to find missing information related to the crop genetic question from the historical dialogue.

[0075] The planning agent can call upon a large language model and utilize the context learning technique of the large language model to use it as a text classifier for crop gene problems.

[0076] In some embodiments, the planning agent can be configured to invoke a large language model to filter candidate tools associated with the crop genetic problem from a list of crop genetic tools using a directed acyclic graph. The list of crop genetic tools may include multiple crop genetic tools. For example, a list of available crop genetic tools can be collected for the planning agent, including multi-trait-gene mapping tools, gene function summarization tools, gene co-expression tools, protein-protein interaction tools, etc., with each crop genetic tool serving as a classification label. The list of crop genetic tools can be dynamically adjusted, for example, by adding new tools or deleting existing tools.

[0077] Planning agents can utilize directed acyclic graphs (DAGs) to dynamically control the initial selection of tools. A DAG is a data structure consisting of nodes and directed edges connecting these nodes, and it contains no cycles. For example, a planning agent can maintain a logical dependency graph of multiple crop gene tools in a list of crop gene tools, and dynamically "enable" or "disable" specific tools based on preconditions for a particular problem. For instance, when the planning agent calls a large language model and determines that the user input does not contain an explicit "trait description," i.e., the precondition for using the "multi-trait-gene mapping" tool is not met, it can automatically disable the "multi-trait-gene mapping" tool in the crop gene tool list, allowing the large language model to classify only among other tools in the list, thereby improving classification accuracy. This strategy of dynamically reducing the search space significantly reduces the error rate of the large language model when selecting from complex sets of specialized tools.

[0078] In some embodiments, the planning agent can be configured to invoke a large language model to determine a target tool from candidate tools based on few-shot cue words and a crop gene problem. For example, few-shot cue words can be provided to the planning agent as recognition examples.

[0079] Here is an example of a few-sample suggestion for the "Multi-trait-gene mapping" tool:

[0080] {

[0081] Question: "Does a soybean gene have pleiotropic effects on both salt tolerance and drought tolerance?"

[0082] "tool": "Multi-trait-gene mapping",

[0083] "reasoning": "Users are asking about soybean genes with pleiotropic effects, making them suitable for multi-trait gene mapping tools."

[0084] }

[0085] For example, when the crop gene question is "Please summarize the genes that simultaneously affect corn starch content, sweetness, aroma and grain weight", since there are similar examples in the few sample prompts, the planning agent can call the large language model to automatically route the question to a multi-trait-gene mapping tool, rather than a general search tool.

[0086] Some embodiments of this invention utilize directed acyclic graphs (DAGs) to initially screen candidate tools and use few-sample cue words to determine the target tool from among the candidate tools, achieving task classification of crop gene problems based on a planning agent. Existing methods typically classify tasks based on keyword matching or simple semantic vector classification; for example, when the question contains the word "price," a price query tool is invoked. However, in the field of crop genes, users' query intentions are often implicit in complex scientific hypotheses, and existing methods cannot distinguish the deep differences between complex intentions such as "finding entities" and "understanding tasks." Some embodiments of this invention utilize the reasoning ability of large language models to act as a planner, enabling the identification of complex scientific research intentions, such as distinguishing between "formulaic queries (e.g., finding all drought-resistant genes)" and "descriptive queries (e.g., summarizing the function of a gene)," further improving the accuracy of the answer.

[0087] In some embodiments, in response to a tool invocation instruction indicating that the target tool is a multi-trait-gene mapping tool, the executing agent can be configured to invoke a large language model to identify crops and multiple traits from the input. The executing agent can be configured to invoke the large language model to generate function invocation instructions based on the crop and multiple traits, in accordance with the format specified in the tool documentation of the multi-trait-gene mapping tool. The function invocation instructions may include indications of the multi-trait-gene mapping tool and structured information about the crop and multiple traits conforming to the format of the multi-trait-gene mapping tool.

[0088] For example, when the crop gene problem is "Please summarize the genes that simultaneously affect corn starch content, sweetness, aroma, and grain weight", the executing agent can identify the species and trait list and generate the following function call instructions:

[0089] JSON

[0090] {

[0091] "tool": "Multi-trait-gene mapping",

[0092] "arguments": {

[0093] "crop": "corn",

[0094] "trait_list": [

[0095] "Starch content",

[0096] "Sweetness",

[0097] "aroma",

[0098] "Particle weight" ]

[0100] }

[0101] }

[0102] Among them, tool indicates the target tool determined by the planning agent through the tool call command, crop indicates the crop species, and trait_list indicates multiple traits.

[0103] In some embodiments of the present invention, all constraints are passed in at once by executing an intelligent agent in the form of a function call instruction. Subsequently, the knowledge mining module efficiently performs multi-condition matching on the knowledge graph, which greatly improves the response speed and accuracy of complex problems.

[0104] In some embodiments, the multi-trait-gene mapping tool can be configured to map multiple traits from structured information in a function call instruction to standard trait annotation fields. The multi-trait-gene mapping tool can be configured to generate structured query sequences based on the crop in the structured information, the mapped standard trait annotation fields, and the template sequence. The multi-trait-gene mapping tool can be configured to execute structured query sequences on a crop-associated database in a knowledge graph to obtain gene records associated with multiple traits. The multi-trait-gene mapping tool can be configured to perform statistical analysis on the gene records to generate a structured summary as query output.

[0105] As an example, a multi-trait-gene mapping tool can receive function call instructions from a planning agent, routing tasks to crop-associated databases in a knowledge graph based on crop species. For instance, the tool might receive a crop name (e.g., "Maize") and a list of traits (e.g., ["starch content", "sweetness", "grain weight"]). Then, since user descriptions of traits may differ from the knowledge graph, the tool can rewrite multiple traits in the trait list into standard trait annotation fields that conform to the representations in the knowledge graph. For example, it can invoke a large language model to rewrite and / or expand the traits. For instance, it can handle Chinese and English synonyms, such as expanding "grain weight" to match standard trait annotation fields like "thousand-grain weight" and "grain weight%". Next, the tool can populate a template sequence with the crop-associated database and the standard trait annotation fields to form a structured query sequence. For example, a structured query sequence can be a Structured Query Language (SQL) statement. For example, a template sequence can be a SELECT statement. The query `FROM gene_table WHERE (info LIKE '%trait A%') AND (info LIKE '%trait B%')` directive uses `gene_table` to specify the database associated with the crop, and `trait A` and `trait B` to specify the different traits to be queried. For example, in the above example, the resulting structured query sequence could be `SELECT`. The `maize_geneinfo` tool uses the following syntax: `FROM maize_geneinfo WHERE (info LIKE '%starch content%') AND (info LIKE '%sweetness%') AND (info LIKE '%grain weight%' OR info LIKE '%thousand-grainweight%' OR info LIKE '%grain weight%').`. This multi-trait gene mapping tool can perform structured query sequences on crop-related databases within a knowledge graph, filtering gene records from the full gene dataset that simultaneously satisfy multiple traits. Therefore, the tool can convert natural language traits into underlying database query statements, efficiently filtering gene records that simultaneously satisfy all trait descriptions within the knowledge graph's local database using logical operators. The logical operator `AND` can connect query conditions for different traits. The logical operator `OR` can expand the retrieval of different annotation fields for the same trait. The `LIKE` syntax enables fuzzy matching. Finally, the tool can aggregate and statistically analyze gene records, outputting structured summaries. For example, it can return the total number of genes involved, a list of key genes, and relevant references. Large language models can generate review reports that researchers can read based on structured summaries.

[0106] The multi-trait-gene mapping tool of some embodiments of the present invention supports one-time input of a trait list, significantly improving retrieval efficiency compared to existing methods that use a large language model to sequentially retrieve each trait and then find the intersection of all retrieval results. In the process of converting user input into a structured query sequence, existing methods typically use slot filling, that is, filling the user-input noun (e.g., "apple") into the statement WHERE name = 'name' (e.g., WHERE name = 'apple'). This requires that the user's input must strictly match the nouns stored in the database. However, in the field of crop breeding, due to the diversity of traits and the possibility that researchers may input different expressions for the same trait, it is extremely difficult to ensure that the user's input is completely consistent with the trait names in the database. The multi-trait-gene mapping tool of some embodiments of the present invention expands the description of traits, avoiding the situation where the user input is inconsistent with the knowledge graph and the corresponding gene cannot be retrieved, thus improving the query hit rate.

[0107] In some embodiments, multiple crop gene tools may include a gene function summarization tool. The gene function summarization tool may be configured to execute function call instructions on a knowledge graph to obtain query output. The structured information in the function call instructions may include a structured query sequence. The structured query sequence can be used to perform a lookup of the function of genes associated with a crop gene question.

[0108] Gene function summarization tools can retrieve gene function descriptions and related traits based on genes specified by the user. Since such descriptive queries are difficult to exhaustively list and cannot be constructed using template sequences, the executing agent generates structured query sequences during the function call instruction construction process.

[0109] In some embodiments, the gene function summarization tool can be configured to execute a structured query sequence in a function call instruction on a knowledge graph to obtain query results associated with genes in different types of node data within the knowledge graph. The gene function summarization tool can also be configured to concatenate the query results into query text, which is then used as the query output.

[0110] For example, gene function summarization tools can extract gene identifiers from structured query sequences within function call instructions, while simultaneously retrieving information from multiple data sources within a knowledge graph (e.g., textual information from gene databases, structured information from relational databases, etc.). For instance, gene function summarization tools can retrieve information from the knowledge graph at different levels: 1) Basic information layer: retrieving chromosome location, gene length, and strand direction; 2) Functional annotation layer: retrieving GO (Gene Ontology) entries, KEGG pathways, and protein domains; 3) Multi-omics data layer: retrieving differential expression data of transcriptomics and proteomics under different stress conditions; 4) Literature evidence layer: retrieving related academic paper titles and abstracts, etc. Gene function summarization tools can then concatenate the retrieved heterogeneous data (e.g., text, numerical values, lists, etc.) into query text according to a predefined template structure, providing it to a large language model. The large language model can organize the query output into a standard report, including specific sections such as biological function, molecular mechanism, phenotypic effect, and references, and can remove redundant data.

[0111] In some embodiments, in response to a tool invocation instruction indicating that the target tool is a gene function summarization tool, the executing agent can be configured to invoke a large language model to identify crop and gene identifiers from the input. The executing agent can be configured to invoke the large language model to generate extended identifiers in other naming conventions based on the gene identifiers. The executing agent can be configured to invoke the large language model to generate function invocation instructions based on the crop, gene identifiers, and extended identifiers, in accordance with the format specified in the tool description document of the gene function summarization tool.

[0112] For example, when the planning agent identifies the target tool as a gene function summarization tool, the execution agent can identify the crop rice and the gene identifier SGDP7 from the user's input natural language command "explain me what rice SGDP7 gene is?". The execution agent can then generate extended identifiers from other naming conventions based on the gene identifier SGDP7, such as potentially associated aliases (sgdp7), other gene identifier annotations (e.g., RapDB ID, MSU ID), chromosome number (Os07g), location identifier (LOC_Os07), etc. Then, based on the crop rice, the gene identifier SGDP7, and the extended identifiers (e.g., potentially associated aliases (sgdp7), other gene identifier annotations (e.g., RapDB ID, MSU ID, chromosome number (Os07g), location identifier (LOC_Os07) etc.), the execution agent can generate the following function call instructions:

[0113] SQL

[0114] SELECT FROM rice_geneinfo

[0115] Where gene LIKE '%SGDP7%'

[0116] OR info LIKE '%SGDP7%'

[0117] OR info LIKE '%sgdp7%'

[0118] OR gene LIKE '%Os07g%'

[0119] OR gene LIKE 'LOC_Os07%';

[0120] Here, `rice_geneinfo` indicates the database corresponding to the crop in the knowledge graph. `gene LIKE '%SGDP7%'`, `info LIKE '%SGDP7%'`, `info LIKE '%sgdp7%'`, `gene LIKE '%Os07g%'`, and `gene LIKE 'LOC_Os07%'` indicate the query statements generated by the executing agent based on the input, including the gene identifier and extended identifier, along with the corresponding fuzzy matching rules. The search conditions are expanded using OR logic to retrieve all entries related to the gene identifier `SGDP7` from the rice-related knowledge graph.

[0121] In some embodiments of this invention, before using the gene function summarization tool, an intelligent agent invokes a large language model. Leveraging the large language model's understanding of structured query language and biological entities and their relationships, the structured query sequence is deeply integrated with knowledge from the field of biological breeding, ensuring high recall and accuracy in biological queries. In the biological field, gene nomenclature and functional description exhibit high heterogeneity and inconsistency; a gene may have multiple aliases and different nomenclature systems. Relying solely on precise searches of gene identifiers, such as equivalence queries, can lead to the omission of a large amount of relevant information, resulting in low recall. Over-reliance on fuzzy matching, on the other hand, may introduce a large amount of irrelevant information, leading to low accuracy. Therefore, by utilizing the large language model's background knowledge and understanding of crop genes, users only need to describe their needs in natural language, and the system can intelligently switch or combine precise and fuzzy searches, significantly lowering the user's barrier to entry and improving query efficiency and satisfaction.

[0122] In some embodiments, a large language model can be trained based on the mapping relationship between multiple gene identifiers of the same crop gene in different nomenclature standards. Crop genomics suffers from complex historical annotation problems. For example, the same rice gene has completely different identifier systems in different international standard databases (such as MSU and RAP-DB), and non-standard common names are often used in the literature. Due to the polymorphism of gene identifiers, i.e., different names in different annotation systems, users cannot remember all the gene identifiers of a gene when inputting natural language questions, and the system cannot require users to input all the identifiers of a gene at the same time. By utilizing the prior knowledge of crop genes in a large language model, the chaotic natural language questions can be transformed into gene identifier annotations acceptable to a knowledge graph database.

[0123] In some embodiments, a large language model can be trained based on the mapping relationship between multiple expressions of the same trait and standard trait annotation fields. When users input natural language questions, the descriptions of traits are highly diverse and ambiguous. For example, a user might input "drought resistance," "water shortage tolerance," or "water stress tolerance," which may correspond to completely different text records at the underlying database level and cannot be fully covered by simple keyword matching. Utilizing the large language model's prior understanding of crop traits and its maintained list of standard trait annotation fields, these chaotic natural language questions can be transformed into trait expressions acceptable to the database. For example, a user's expression (such as "water shortage tolerance") can be automatically transformed into a database-acceptable standard trait (such as "drought resistance"). This allows the system to understand the semantic equivalence of biological concepts, thereby performing conceptual unification-based operations at the database level, rather than simply text matching.

[0124] In some embodiments, the large language model can generate answers to crop gene questions based on the query output according to a preset template. For example, the preset template is as follows:

[0125] • First-level heading: [Crop Name] Gene [Gene ID] Comprehensive Information Summary

[0126] • Basic Gene Information module: contains core identifiers such as gene ID, gene symbol, chromosome location, and UniProt number.

[0127] • Biological Role: A concise summary of the gene’s core function and molecular mechanism.

[0128] • Associated Traits and Research Studies module: Describes in detail the phenotypic changes (such as plant height and yield) caused by gene mutations or overexpression, as well as the metabolic pathways involved (such as the gibberellin pathway).

[0129] • Functional Annotations: Lists GO Gene Ontology entries and KEGG pathway information.

[0130] • References section (Research Literature): Lists the titles of key academic papers that support the above conclusions.

[0131] • Next Step Suggestion: Based on the currently known information, intelligently recommend directions for subsequent verification experiments.

[0132] In some embodiments, after the knowledge mining module completes the invocation of the target tool, the knowledge mining module can further invoke data manipulation tools to analyze the query output. For example, data manipulation tools include intersection operation tools, counting and statistics tools, and trend analysis tools.

[0133] For example, to address the question of finding genes that simultaneously influence both "drought resistance" and "plant height," after the knowledge mining module invokes a multi-trait-gene mapping tool and identifies specific genes matching both "drought resistance" and "plant height," the knowledge mining module can then call a counting and statistics tool to count genes matching a single trait, thereby assessing research interest. The knowledge mining module can provide the trait "drought resistance" to the counting and statistics tool, which can then execute the following structured query sequence:

[0134] SQL

[0135] SELECT COUNT(DISTINCT gene) FROM geneinfo

[0136] WHERE info LIKE '%drought%tolerance%' OR info LIKE '%drought resistance%';

[0137] Counting and statistical tools can use the COUNT(DISTINCT ...) aggregation function to return the number of duplicate genes, so as to quickly count the gene knowledge reserves of a knowledge graph about a specific trait.

[0138] In some embodiments, after the knowledge mining module completes the invocation of the target tool, the system can, based on the query output, invoke other crop gene tools through intelligent agent interaction to achieve secondary processing of the query output.

[0139] In some embodiments, in response to the target tool to be invoked being identified as a multi-trait-gene mapping tool, the planning agent 140 may be configured to invoke the large language model 130 to generate a first tool invocation instruction based on the multi-trait-gene mapping tool, the first tool invocation instruction including an indication of the multi-trait-gene mapping tool. The execution agent 150 may be configured to invoke the large language model 130 to generate a first function invocation instruction based on the input, task description, first tool documentation, and the first tool invocation instruction, the first function invocation instruction including an indication of the multi-trait-gene mapping tool and first structured information related to the crop genetic problem in a format conforming to the multi-trait-gene mapping tool. The knowledge mining module 160 may be configured to invoke the multi-trait-gene mapping tool to perform a first structured query on the knowledge graph based on the first structured information in the first function invocation instruction to obtain a first query output.

[0140] In some embodiments, the planning agent 140 may be further configured to invoke the large language model 130 to: receive a first query output; in response to determining that the first query output includes genes associated with multiple traits, identify the target tool to be invoked as a gene function summarization tool; and generate a second tool invocation instruction based on the gene function summarization tool, the second tool invocation instruction including an indication of the gene function summarization tool. The execution agent 150 may be further configured to invoke the large language model 130 to: receive the first query output, a second tool documentation, and a second tool invocation instruction; and generate a second function invocation instruction based on the first query output, the second tool documentation, and the second tool invocation instruction, the second function invocation instruction including an indication of the gene function summarization tool and second structured information associated with the first query output conforming to the format of the gene function summarization tool. The knowledge mining module 160 may be further configured to: receive the second function invocation instruction; and, based on the second structured information in the second function invocation instruction, invoke the gene function summarization tool to perform a second structured query on the knowledge graph to obtain a second query output, wherein the answer to the crop gene question is generated by the large language model 130 based on the first query output and the second query output.

[0141] Some embodiments of the present invention first call a multi-trait-gene mapping tool to find specific genes that satisfy multiple traits, and then call a gene function summary tool to query the function of specific genes. This allows the large language model to return answers to users that not only include specific genes associated with traits, but also include explanations of the gene functions of specific genes, providing users with comprehensive information related to crop gene issues and improving the user experience.

[0142] In some embodiments, the planning agent 140 may be further configured to invoke the large language model 130 to: receive a first query output; and, in response to determining that the first query output does not include genes associated with multiple traits, terminate the operation associated with the first query output. The answer to the crop gene question is generated by the large language model 130 based on the first query output. For example, when the first query result obtained by invoking a multi-trait-gene mapping tool does not find any genes, the large language model directly generates an answer based on the first query result without invoking other tools.

[0143] In some embodiments, in response to the target tool to be invoked being identified as a gene function summarization tool, the planning agent 140 may be configured to invoke the large language model 130 to generate a first tool invocation instruction based on the gene function summarization tool, the first tool invocation instruction including an indication of the gene function summarization tool. The execution agent 150 may be configured to invoke the large language model 130 to generate a first function invocation instruction based on input, task description, a first tool documentation, and the first tool invocation instruction, the first function invocation instruction including an indication of the gene function summarization tool and first structured information related to the crop gene problem in a format conforming to the gene function summarization tool. The knowledge mining module 160 may be configured to invoke the gene function summarization tool to perform a first structured query on the knowledge graph based on the first structured information in the first function invocation instruction to obtain a first query output.

[0144] In some embodiments, the planning agent 140 may be further configured to invoke the large language model 130 to: receive a first query output; in response to determining that the first query output includes multiple traits associated with genes, identify the target tool to be invoked as a multi-trait-gene mapping tool; and generate a second tool invocation instruction based on the multi-trait-gene mapping tool, the second tool invocation instruction including an indication of the multi-trait-gene mapping tool. The execution agent 150 may be further configured to invoke the large language model 130 to: receive the first query output, a second tool documentation, and a second tool invocation instruction; and generate a second function invocation instruction based on the first query output, the second tool documentation, and the second tool invocation instruction, the second function invocation instruction including an indication of the multi-trait-gene mapping tool and second structured information associated with the first query output conforming to the format of the multi-trait-gene mapping tool. The knowledge mining module 160 can be further configured to: receive a second function call instruction; and, based on the second structured information in the second function call instruction, call a multi-trait-gene mapping tool to perform a second structured query on the knowledge graph to obtain a second query output, wherein the answer to the crop gene question is generated by the large language model 130 based on the first query output and the second query output.

[0145] For example, the user input is "Please explain the function of the rice SGDP7 gene." The planning agent classifies this question as a "gene function summary" task, executes the agent's function generation call instruction, and the knowledge mining module calls the gene function summary tool to search for the function of SGDP7 in the knowledge graph and provides a first query output. The planning agent receives the first query output and finds multiple traits related to this gene, such as "drought resistance" and "plant height." The planning agent identifies the target tool as a multi-trait-gene mapping tool, executes the agent's function generation call instruction based on "drought resistance" and "plant height," and the knowledge mining module calls the multi-trait-gene mapping tool to search for genes that simultaneously affect the traits of "drought resistance" and "plant height" in the knowledge graph and provides a second query output. Finally, the large language model can generate an answer to the user's question based on the first and second query outputs.

[0146] Some embodiments of the present invention first call a gene function summary tool to query the traits affected by genes, and then call a multi-trait-gene mapping tool to find genes that satisfy multiple traits. This allows the large language model to return answers to users that not only include explanations of gene functions, but also other genes associated with the traits, providing users with comprehensive information related to crop gene issues and improving the user experience.

[0147] In some embodiments, the planning agent 140 may be further configured to invoke the large language model 130 to: receive a first query output; and, in response to determining that the first query output does not include multiple traits associated with genes, terminate the operation associated with the first query output. The answer to the crop gene question is generated by the large language model 130 based on the first query output. For example, when the first query result obtained by invoking a gene function summarization tool only retrieves a single trait, the large language model directly generates an answer based on the first query result without invoking other tools.

[0148] In some embodiments, the input may be a first input, and the large language model may be configured to invoke computational resources to: receive a second input from a user, the second input including a second question; determine the relevance of the second question to the biological domain; and, in response to the second question being irrelevant to the biological domain, generate an answer to the second question based on the second question.

[0149] Some embodiments of the present invention perform binary classification topic extraction on user input through a large language model to determine whether the user input belongs to "asking about biological questions" or "other irrelevant topics", thus avoiding unnecessary resource consumption caused by calling the agent when answering other irrelevant topics.

[0150] In some embodiments, crop gene datasets can be organized in a knowledge graph as triples. Triples can include nodes and edges, with the basic representation unit being node-edge-node. Nodes can include at least one of the following: gene, gene function, agronomic trait, molecular phenotype, environmental factor, experimental evidence, transcript, protein, tissue / organ, developmental stage, gene annotation (KEGG pathway, GO annotation (biological process, molecular function, cellular component)), protein domain, etc. The relationships between nodes are modeled as edges. Edges can also include attributes such as evidence, condition, and confidence, thus forming context-enhanced triple representations. The knowledge graph uses genes as core nodes, tightly linking nodes such as gene function, agronomic traits, molecular / cellular phenotype, environmental factors, experimental evidence, and literature sources through complex network relationships. For example, a gene node connected to a tissue / organ node can represent "gene A is expressed in the stem." A gene node connected to a drought stress experiment node can represent "gene B is differentially expressed in a drought stress experiment." The weight of the connecting edge reflects the frequency of this association; a higher edge weight indicates a stronger association.

[0151] In some embodiments, crops may include rice, corn, soybeans, Arabidopsis thaliana, etc.

[0152] In some embodiments, the large language model can be configured to: obtain literature related to crop genes; extract nodes and edges of crop gene datasets from the literatures; and construct a knowledge graph based on the nodes and edges of the crop gene datasets, wherein triples are mapped to text blocks in the literatures.

[0153] In some embodiments, the storage unit may include a relational database for storing structured data in the knowledge graph. For example, the relational database may be a DuckDB database, a PostgreSQL database, a MySQL database, etc. For example, the structured data may include gene function summaries, gene annotations, trait associations, and statistical features. Relational databases can efficiently store and query structured data; in scenarios requiring frequent queries, structured query sequences can efficiently access the structured data in the relational database. The knowledge mining module can be configured to query structured attribute information in the relational database using the invoked target tool.

[0154] In some embodiments, the storage unit may include a graph database for storing graph data in the knowledge graph. Graph data can represent complex, multidimensional relationships between nodes. Graph data may include co-expression and protein-gene interaction relationships, associations between genes and traits, environment, experiments, literature, and so on. Graph databases excel at handling complex relationships and reasoning between nodes. Graph databases can also interact with relational databases to provide multi-level association analysis. The knowledge mining module can be configured to query association information in the graph database using invoked target tools.

[0155] The system can allocate query requests to a relational database, a graph database, or coordinate the execution of queries using both, based on the classification of crop gene problems by the planning agent. For example, when a user queries "rice genes that simultaneously affect plant height and drought resistance," the system first filters the relational database using structured query sequences to identify a set of candidate genes annotated as simultaneously associated with "plant height" and "drought resistance." Subsequently, the system transfers this candidate set to the graph database for further analysis to determine if these genes exhibit co-expression relationships or reside in the same regulatory network module, thereby verifying their biological connections and identifying potential key genes. In this approach, the relational database handles efficient attribute filtering, while the graph database handles relational reasoning; the two form a seamless and complementary collaborative relationship during the query process.

[0156] The system according to some embodiments of the present invention interacts with an intelligent agent through a large language model, decomposing user-posed natural language questions into structured query sequences and invoking crop gene tools to perform structured queries on the knowledge graph. Since the answers to the questions originate from a validated knowledge graph, rather than the parameterized memory of the large language model itself, the accuracy of the answers is significantly improved. By automating the process to replace manual operations by human experts, the processing time for knowledge summarization tasks is greatly reduced. Compared with general large models, it can reliably complete professional tasks requiring complex structured queries and cross-analysis.

[0157] According to another aspect of the present invention, a method for crop gene question answering is provided.

[0158] Figure 4 This is a flowchart of a method for crop gene question answering according to some embodiments of the present invention. The method can be... Figure 1 The system 100 is executed.

[0159] The method may include step S1: receiving input from the user by a large language model, the input including crop gene questions.

[0160] The method may include step S2: generating a task description based on the input from a large language model.

[0161] The method may include step S3: the planning agent invokes the large language model to determine the target tool to be invoked from among the multiple crop gene tools included in the knowledge mining module based on the input, task description, and tool documentation provided by the knowledge mining module.

[0162] The method may include step S4: the planning agent invokes a large language model to generate a tool invocation instruction based on the target tool, the tool invocation instruction including an indication of the target tool.

[0163] The method may include step S5: the executing agent invokes a large language model to generate function call instructions based on the input, task description, tool documentation, and tool call instructions. The function call instructions include indications of the target tool and structured information related to the crop genetic problem in a format consistent with the target tool.

[0164] The method may include step S6: the knowledge mining module calls the target tool indicated in the function call instruction to perform a structured query on the knowledge graph based on the structured information in the function call instruction to obtain the query output, wherein the knowledge graph is stored in a storage unit and includes a crop gene dataset.

[0165] The method may include step S7: generating an answer to the crop gene question based on the query output using a large language model.

[0166] Multiple crop genetic tools may include multi-trait-gene mapping tools. These tools can be configured to generate structured query sequences based on structured information in function call instructions. This structured information includes the crop and multiple traits associated with the crop genetic question. The structured query sequences are used to perform searches for genes associated with the multiple traits of the crop and to perform logical operations on the search results. Multi-trait-gene mapping tools can also be configured to execute structured query sequences on knowledge graphs to obtain query output.

[0167] In some embodiments, the multi-trait-gene mapping tool can be configured to map multiple traits from structured information in a function call instruction to standard trait annotation fields. The multi-trait-gene mapping tool can be configured to generate structured query sequences based on the crop in the structured information, the mapped standard trait annotation fields, and the template sequence. The multi-trait-gene mapping tool can be configured to execute structured query sequences on a crop-associated database in a knowledge graph to obtain gene records associated with multiple traits. The multi-trait-gene mapping tool can be configured to perform statistical analysis on the gene records to generate a structured summary as query output.

[0168] In some embodiments, a large language model can be trained based on the mapping relationship between multiple gene identifiers of the same crop gene in different nomenclature schemes.

[0169] In some embodiments, a large language model can be trained based on the mapping relationship between multiple expressions of the same trait and standard trait annotation fields.

[0170] Figure 5 This is a flowchart of a first process associated with a method for crop gene question answering according to some embodiments of the present invention. This first process may be... Figure 1 The planning agent 140 in the middle calls the large language model 130 to execute, and can be Figure 4The specific implementation of step S3 in the method is described, but the scope of the present invention is not limited thereto.

[0171] The first process may include step S31: using a directed acyclic graph to screen candidate tools associated with crop genetic problems from a list of crop genetic tools, which includes multiple crop genetic tools.

[0172] The first process may include step S32: identifying the target tool from candidate tools based on few-sample cue words and crop gene questions.

[0173] Figure 6 This is a flowchart of a second process associated with a method for crop gene question answering according to some embodiments of the present invention. This second process may be... Figure 1 The executing agent 150 calls the large language model 130 to execute, and can be Figure 4 The specific implementation of step S5 in the method is described, but the scope of the present invention is not limited thereto.

[0174] In response to a tool invocation instruction indicating that the target tool is a multi-trait-gene mapping tool, the second process may include step S51: identifying crops and multiple traits from the input.

[0175] The second process may include step S52: Based on the crop and multiple traits, generate function call instructions in accordance with the format specified in the tool description document of the multi-trait-gene mapping tool. The function call instructions include the instructions of the multi-trait-gene mapping tool and the structured information of the crop and multiple traits in accordance with the format of the multi-trait-gene mapping tool.

[0176] In some embodiments, multiple crop gene tools may include a gene function summarization tool. The gene function summarization tool may be configured to execute function call instructions on a knowledge graph to obtain query output. The structured information in the function call instructions includes a structured query sequence used to perform a lookup of the function of genes associated with a crop gene question.

[0177] In some embodiments, the gene function summarization tool can be configured to execute a structured query sequence in a function call instruction on a knowledge graph to obtain query results associated with genes in different types of node data within the knowledge graph. The gene function summarization tool can also be configured to concatenate the query results into query text, which is then used as the query output.

[0178] Figure 7 This is a flowchart of a third process associated with a method for crop gene question answering according to some embodiments of the present invention. This third process may be... Figure 1 The executing agent 150 calls the large language model 130 to execute, and can be Figure 4The specific implementation of step S5 in the method is described, but the scope of the present invention is not limited thereto.

[0179] In response to a tool invocation instruction indicating that the target tool is a gene function summarization tool, the third process may include step S53: identifying crop and gene identifiers from the input.

[0180] The third process may include step S54: generating extended identifiers in other naming conventions based on the gene identifier.

[0181] The third process may include step S55: generating function call instructions based on the crop, gene identifier, and extended identifier, in accordance with the format specified in the tool description document of the gene function summarization tool.

[0182] Figure 8 This is a flowchart of a fourth process associated with a method for crop gene question answering according to some embodiments of the present invention. This fourth process may be... Figure 1 The system 100 executes this process. This fourth process involves an embodiment where the query output requires further processing.

[0183] In some embodiments, the fourth process may include: executing a first execution sequence in response to the target tool to be invoked being identified as a multi-trait-gene mapping tool.

[0184] In some embodiments, the fourth process may include: executing a second execution sequence in response to the target tool to be invoked being identified as a gene function summarization tool.

[0185] Figure 9 This is a flowchart of a fifth process associated with a method for crop gene question answering according to some embodiments of the present invention. This fifth process may be... Figure 1 System 100 executes, and can be Figure 8 The specific implementation of the first execution sequence in the fourth process is described, but the scope of the present invention is not limited thereto.

[0186] In response to the target tool to be invoked being identified as a multi-trait-gene mapping tool, the fifth process may include step S41: generating a first tool invocation instruction based on the multi-trait-gene mapping tool, the first tool invocation instruction including an indication of the multi-trait-gene mapping tool.

[0187] Step S41 can be Figure 4 The specific implementation of step S4 in the method is described, but the scope of the present invention is not limited thereto.

[0188] The fifth process may include step S56: generating a first function call instruction based on the input, task description, first tool documentation, and first tool call instruction. The first function call instruction includes instructions for the multi-trait-gene mapping tool and first structured information related to crop genetic issues in a format consistent with the multi-trait-gene mapping tool.

[0189] Step S56 can be Figure 4 The specific implementation of step S5 in the method is described, but the scope of the present invention is not limited thereto.

[0190] The fifth process may include step S61: based on the first structured information in the first function call instruction, calling a multi-trait-gene mapping tool to perform a first structured query on the knowledge graph to obtain a first query output.

[0191] Step S61 can be Figure 4 The specific implementation of step S6 in the method is described herein, but the scope of the present invention is not limited thereto.

[0192] The fifth process may include step S42: in response to determining that the first query output includes genes associated with multiple traits, identifying the target tool to be invoked as a gene function summary tool.

[0193] The fifth process may include step S43: generating a second tool invocation instruction based on the gene function summarization tool, the second tool invocation instruction including instructions for the gene function summarization tool.

[0194] Steps S42 and S43 can be Figure 4 The specific implementation of step S4 in the method is described, but the scope of the present invention is not limited thereto.

[0195] The fifth process may include step S57: generating a second function call instruction based on the first query output, the second tool documentation, and the second tool call instruction. The second function call instruction includes instructions from the gene function summarization tool and second structured information associated with the first query output that conforms to the format of the gene function summarization tool.

[0196] Step S57 can be Figure 4 The specific implementation of step S5 in the method is described, but the scope of the present invention is not limited thereto.

[0197] The fifth process may include step S62: based on the second structured information in the second function call instruction, calling the gene function summary tool to perform a second structured query on the knowledge graph to obtain a second query output.

[0198] Step S62 can be Figure 4 The specific implementation of step S6 in the method is described herein, but the scope of the present invention is not limited thereto.

[0199] The fifth process may include step S71: generating an answer to the crop gene question by the large language model based on the output of the first query and the output of the second query.

[0200] Step S71 can be Figure 4 The specific implementation of step S7 in the method is described, but the scope of the present invention is not limited thereto.

[0201] Figure 10 This is a flowchart of a sixth process associated with a method for crop gene question answering according to some embodiments of the present invention. This sixth process may be... Figure 1 System 100 executes, and can be Figure 8 The specific implementation of the second execution sequence in the fourth process is described, but the scope of the present invention is not limited thereto.

[0202] In response to the target tool to be invoked being identified as a gene function summarization tool, the sixth process may include step S44: generating a first tool invocation instruction based on the gene function summarization tool, the first tool invocation instruction including an indication of the gene function summarization tool.

[0203] Step S44 can be Figure 4 The specific implementation of step S4 in the method is described, but the scope of the present invention is not limited thereto.

[0204] The sixth process may include step S58: generating a first function call instruction based on the input, task description, first tool description document and first tool call instruction, the first function call instruction including instructions for the gene function summarization tool and first structured information related to crop gene issues in a format consistent with the gene function summarization tool.

[0205] Step S58 can be Figure 4 The specific implementation of step S5 in the method is described, but the scope of the present invention is not limited thereto.

[0206] The sixth process may include step S63: based on the first structured information in the first function call instruction, calling the gene function summary tool to perform a first structured query on the knowledge graph to obtain a first query output.

[0207] Step S63 can be Figure 4 The specific implementation of step S6 in the method is described herein, but the scope of the present invention is not limited thereto.

[0208] The sixth process may include step S45: in response to determining that the first query output includes multiple traits associated with genes, identifying the target tool to be invoked as a multi-trait-gene mapping tool.

[0209] The sixth process may include step S46: generating a second tool invocation instruction based on the multi-trait-gene mapping tool, the second tool invocation instruction including instructions for the multi-trait-gene mapping tool.

[0210] Steps S45 and S46 can be Figure 4 The specific implementation of step S4 in the method is described, but the scope of the present invention is not limited thereto.

[0211] The sixth process may include step S59: generating a second function call instruction based on the first query output, the second tool documentation, and the second tool call instruction. The second function call instruction includes an indication of the multi-trait-gene mapping tool and second structured information associated with the first query output that conforms to the format of the multi-trait-gene mapping tool.

[0212] Step S59 can be Figure 4 The specific implementation of step S5 in the method is described, but the scope of the present invention is not limited thereto.

[0213] The sixth process may include step S64: based on the second structured information in the second function call instruction, calling the multi-trait-gene mapping tool to perform a second structured query on the knowledge graph to obtain the second query output.

[0214] Step S64 can be Figure 4 The specific implementation of step S6 in the method is described herein, but the scope of the present invention is not limited thereto.

[0215] The sixth process may include step S71: generating an answer to the crop gene question by the large language model based on the output of the first query and the output of the second query.

[0216] Step S71 can be Figure 4 The specific implementation of step S7 in the method is described, but the scope of the present invention is not limited thereto.

[0217] According to another aspect of the present invention, a computer-readable storage medium is provided.

[0218] Figure 11 This is a block diagram of a computer-readable storage medium 1100 according to some embodiments of the present invention.

[0219] A computer-readable storage medium 1100 stores a computer program 1150. When executed by a processor, the computer program 1150 implements the above-mentioned... Figures 4-10 The steps of each method or process described.

[0220] According to another aspect of the present invention, a computer program product is provided.

[0221] Figure 12This is a block diagram of a computer program product 1200 according to some embodiments of the present invention.

[0222] Computer program product 1200 may include computer program 1150. Computer program 1150, when executed by a processor, implements the above-mentioned... Figures 4-10 The steps of each method or process described.

[0223] The following are examples of interactions between embodiments of the present invention and users based on real system operation logs:

[0224] User-input question:

[0225] What research has been conducted on the soybean gene GmW82.16G017200?

[0226] The corresponding output of SeedLLM (a Wiki-style overview translated into Chinese):

[0227] Comprehensive information on soybean gene GmW82.16G017200

[0228] 1. Basic genetic information

[0229] Wm82-NJAU Gene ID: GmW82.16G017200

[0230] Gene symbol: GmLHY1a

[0231] Chromosomal location: Chromosome 16

[0232] UniProt entry: I1MKB1

[0233] 2. Biological Function: GmLHY1a is a member of the LHY / CCA1 transcription factor family and a key component of the plant circadian rhythm. This gene encodes a MYB transcription factor, which primarily regulates plant height and internode length in soybean.

[0234] 3. Related Traits and Research Progress: Major studies on GmW82.16G017200 (GmLHY1a) have focused on its regulatory effect on plant height:

[0235] Plant height and internode length: CRISPR / Cas9-mediated targeted mutation studies showed that the GmLHY gene alters plant height. Compared to the wild type, quadruple mutants exhibited a significantly dwarfing phenotype.

[0236] Gibberellin (GA) pathway involvement: Studies have confirmed that the GmLHY gene is involved in the gibberellin biosynthesis pathway. Mutant plants showed lower levels of endogenous GA3.

[0237] 4. Key References

[0238] CRISPR / Cas9-mediated targeted mutation of GmLHY genes alters plant height and internode length in soybean.

[0239] 5. Recommendations for the next step of the experiment: In order to further explore the role of this gene in plant height regulation, it is recommended to conduct follow-up studies to detect its expression pattern under different environmental conditions, or to study its interaction with the biological clock and other components in the GA signaling pathway.

[0240] Figure 13 This is a performance comparison of the embodiments of the present invention with existing methods for answering crop gene questions. The embodiments of the present invention were rigorously compared with mainstream international technologies on an evaluation set of professional questions related to crop gene traits. In terms of answer accuracy, the embodiments of the present invention scored significantly higher than general-purpose large language models (DeepSeek-V3, DeepSeek-R1, OpenAI-GPT5), mainstream intelligent agents (WebLangChain, OpenManus, DeerFlow), research platforms (ScienceOne, SciMaster), and human experts. Regarding answer time, under the premise of obtaining answers of the same quality level, the embodiments of the present invention consumed significantly less time than other methods. In terms of efficiency (the ratio of score to time), the embodiments of the present invention significantly outperformed other methods.

[0241] The following are examples of the embodiments of the present invention and the general large model DeepSeek-V3's responses to the same crop gene question, "Please summarize the genes that simultaneously affect corn starch content, sweetness, aroma, and grain weight."

[0242] Examples of responses to embodiments of the present invention are as follows:

[0243] A summary of maize genes related to starch content, sweetness, aroma content, and grain weight.

[0244] Nine genes were found to simultaneously affect the starch content, sweetness, aroma content, and grain weight of corn.

[0245] These genes include:

[0246] Zm00001eb159060 (Sh2, sh2, Zm00001d044129, Zmsh2, SH2, ZmSH2)

[0247] Zm00001eb174590 (su1, SU1, Su1)

[0248] Zm00001eb374090 (sh1, Zm00001d045042, Sh1, ZmSH1)

[0249] Zm00001eb235570 (bt1, ZmBT1, BT1, ZmBt1)

[0250] Zm00001eb392880 (ZmSUS1, sus1, Zm00001d047253, SUS1, Sus1)

[0251] Zm00001eb279740 (su2, Su2)

[0252] Zm00001eb242610 (ae1, Ae1, ZmAE1)

[0253] Zm00001eb083790 (mn1, Mn1)

[0254] Zm00001eb341550 (fl3, Fl3)

[0255] Gene annotation source:

[0256] literature: fl3, ae1, Sh2, ZmSUS1, Fl3

[0257] wos: sh2, bt1, su2, su1, sh1, ae1, sus1, mn1, Zm00001d045042,Zm00001d044129, Zm00001d047253, Mn1, Ae1, SH2, Sh2, Su1, ZmBT1, Sh1, Su2,Sus1, BT1, ZmSH2, ZmSH1, ZmSUS1, ZmAE1

[0258] pubmed: sh2, sh1, su2, su1, bt1, mn1, ae1, SU1, Sh2, SUS1, Ae1,Zmsh2, Su2, Sh1, Su1, ZmBT1, BT1, Mn1, Sus1, ZmBt1

[0259] References:

[0260] The Maize Imprinted Gene Floury3 Encodes a PLATZ Protein Required for tRNA and 5S rRNA Transcription through Interaction with RNA Polymerase III

[0261] EXPRESSION OF ALCOHOL-SOLUBLE ENDOSPERM PROTEINS IN MAIZE SINGLE AND DOUBLE MUTANTS

[0262] Soluble Sugar Content and Molecular Marker Detecting of Sweet-Related Alleles in Sweet Corn Inbred Lines

[0263] Source-sink relations in maize mutants with starch-deficient endosperms

[0264] Genetic diversity of starch synthesis genes of Chinese maize (Zea mays L.) with SNAPs

[0265] Investigation for properties of endosperm starches of maize containing ae gene in genetic background

[0266] Genetic diversity of maize kernel starch-synthesis genes with SNAPs

[0267] QTLs for enzyme activities and soluble carbohydrates involved instarch accumulation during grain filling in maize

[0268] Regulation of coronatine on the grain filling characteristics and starch synthesis in maize kernels

[0269] Genetic diversity analysis of Chinese fresh corn hybrids using SNP

[0270] Here is an example of a DeepSeek-V3 response:

[0271] Genes affecting corn starch content, sweetness, aroma, and kernel weight typically involve multiple metabolic pathways and regulatory networks. The following is a summary of known key genes and their functions:

[0272] ---

[0273] ### 1. Genes related to starch content

[0274] - Waxy (Wx1)

[0275] - Encodes granule-bound starch synthase (GBSS), responsible for amylose synthesis. The mutant (wx1) results in waxy maize (amylose deficiency, amylopectin increase).

[0276] - Shrunken2 (Sh2)

[0277] - Encodes the large subunit of ADP-glucose pyrophosphorylase (AGPase), which regulates the rate-limiting step in starch synthesis. Mutations result in grain shrinkage and reduced starch content.

[0278] - Brittle2 (Bt2)

[0279] - Encodes the small subunit of AGPase, which works synergistically with Sh2 to affect starch accumulation.

[0280] - Amylose Extender (Ae1)

[0281] - Encodes starch branching enzyme IIb (SBEIIb), mutations lead to changes in the structure of amylopectin and an increase in the content of amylose.

[0282] ---

[0283] ### 2. Sweetness-related genes

[0284] - Sugary1 (Su1)

[0285] - Encodes starch debranching enzyme (ISA1). The mutant accumulates phytoglycogen and water-soluble polysaccharides, increasing sweetness (common sweet corn).

[0286] - Shrunken1 (Sh1)

[0287] - Encodes sucrose synthase, which affects the conversion of sucrose into starch; mutations lead to sucrose accumulation (enhancing sweetness).

[0288] - Brittle1 (Bt1)

[0289] - Encodes the ADP-glucose transporter; mutations lead to sugar accumulation.

[0290] - Recessive mutant genes (se1, su2, du1, etc.)

[0291] - For example, super sweet corn (sh2 mutant) has a significantly increased sucrose content due to impaired sugar metabolism.

[0292] ---

[0293] ### 3. Aroma-related genes

[0294] - Aromatic Gene (fgr / bad2)

[0295] - Encodes betaine aldehyde dehydrogenase (BADH2), which degrades 2-acetyl-1-pyrrolline (2-AP, a key compound in popcorn aroma). Mutants (such as Thai jasmine corn) accumulate 2-AP due to BADH2 inactivation.

[0296] - Lipoxygenase (LOX) family

[0297] - Catalyzes the oxidation of fatty acids to produce volatile substances such as aldehydes and ketones, which affect the flavor of corn.

[0298] ---

[0299] ### 4. Particle weight-related genes

[0300] - ZmGW2

[0301] - It negatively regulates cell division; its homologous gene is known to affect grain weight in rice, and a similar function may be conserved in maize.

[0302] - ZmSWEET4c

[0303] - It regulates sugar transport, affecting grain filling and final grain weight.

[0304] - Kernel Row Number (KRN)

[0305] - Regulating the number of rows in the ear indirectly affects the weight of a single grain.

[0306] - ZmACS7

[0307] - Ethylene synthase gene, which affects grain size during development.

[0308] ---

[0309] ### Pleiotropic genes (genes that affect multiple traits simultaneously)

[0310] - Opaque2 (O2)

[0311] - Transcription factors regulate the synthesis of prolactin; mutants reduce protein content and may also alter starch structure and grain weight.

[0312] - ZmMYB14

[0313] - It participates in phenylpropane metabolism and may simultaneously affect aroma (volatile phenols) and particle weight (by regulating carbon distribution).

[0314] ---

[0315] ### Breeding applications

[0316] - Sweet corn breeding : Combines sh2 (high sugar) and fgr (aroma) mutations.

[0317] - High starch varieties : Optimize the expression of AGPase genes such as Sh2 and Bt2.

[0318] - Molecular marker-assisted selection The SNP markers of the above genes are used to accelerate trait aggregation.

[0319] For more specific research progress or species differences (such as waxy corn vs. common corn), please supplement the literature.

[0320] As can be seen from the above example answers, compared with the general large language model, the embodiments of the present invention can accurately identify genes related to multiple traits and provide information sources, while the general large language model can only list related genes for a single trait sequentially, failing to meet the user's requirement regarding genes related to multiple traits. Therefore, the embodiments of the present invention significantly improve the accuracy of answers to crop gene questions.

[0321] Embodiments of the invention have been described with reference to the accompanying drawings. These embodiments are illustrative and not restrictive.

Claims

1. A system for crop gene question answering, characterized in that, include: Computing resources; A storage unit is configured to store a knowledge graph, which includes a crop gene dataset. The large language model is configured to invoke the computing resources to: Receive input from the user, including questions about crop genes; Generate a task description based on the input; as well as The query output is used to generate an answer to the crop gene question. The planning agent is configured to invoke the large language model to: Receive the input, the task description, and the tool documentation; Based on the input, the task description, and the tool documentation, determine the target tool to be invoked from among multiple crop gene tools; as well as Based on the target tool, a tool invocation instruction is generated, wherein the tool invocation instruction includes an indication of the target tool; The executing agent is configured to invoke the large language model to: Receive the input, the task description, the tool documentation, and the tool invocation command; as well as Based on the input, the task description, the tool documentation, and the tool invocation instructions, a function invocation instruction is generated. The function invocation instruction includes an indication of the target tool and structured information related to the crop genetic problem in a format consistent with the target tool. as well as The knowledge mining module is configured as follows: Provide documentation for the aforementioned tools; Receive the function call instruction; as well as Based on the structured information in the function call instruction, the target tool indicated in the function call instruction is invoked to perform a structured query on the knowledge graph to obtain the query output, wherein the answer to the crop gene question is generated by the large language model based on the query output. The knowledge mining module includes multiple crop gene tools, including a multi-trait-gene mapping tool, which is configured to: A structured query sequence is generated based on the structured information in the function call instruction. The structured information includes crops and multiple traits associated with the crop genetic problem. The structured query sequence is used to perform a search for genes associated with the multiple traits of the crop and to perform logical operations on the search results. The structured query sequence is executed on the knowledge graph to obtain the query output.

2. The system as described in claim 1, characterized in that, The planning agent is configured to invoke the large language model to: Candidate tools associated with the crop gene problem are screened from a list of crop gene tools using a directed acyclic graph, the list of crop gene tools including the plurality of crop gene tools; as well as The target tool is determined from the candidate tools based on the few-sample cue words and the crop gene problem.

3. The system as described in claim 1, characterized in that, In response to the tool invocation instruction indicating that the target tool is the multi-trait-gene mapping tool, the executing agent is configured to invoke the large language model to: Identify the crop and the multiple traits from the input; as well as Based on the crop and the multiple traits, the function call instruction is generated in accordance with the format specified in the tool description document of the multi-trait-gene mapping tool. The function call instruction includes the instructions of the multi-trait-gene mapping tool and the structured information of the crop and the multiple traits in accordance with the format of the multi-trait-gene mapping tool.

4. The system as described in claim 1, characterized in that, The multi-trait-gene mapping tool is configured to: Map the multiple traits in the structured information of the function call instruction to the standard trait annotation field; The structured query sequence is generated based on the crop, the mapped standard trait annotation fields, and the template sequence in the structured information; The structured query sequence is executed on the database associated with the crop in the knowledge graph to obtain gene records associated with the multiple traits; as well as The gene records are statistically analyzed to generate a structured summary, which is then used as the query output.

5. The system as described in claim 1, characterized in that, The plurality of crop gene tools includes a gene function summarization tool, which is configured to: The function call instruction is executed on the knowledge graph to obtain the query output. The structured information in the function call instruction includes a structured query sequence, which is used to perform a search for the function of genes associated with the crop gene problem.

6. The system as described in claim 5, characterized in that, The gene function summarization tool is configured to: The structured query sequence in the function call instruction is executed on the knowledge graph to obtain query results associated with the gene in different types of node data in the knowledge graph; as well as The query results are concatenated into query text, which is then used as the query output.

7. The system as described in claim 5, characterized in that, In response to the tool invocation instruction indicating that the target tool is the gene function summarization tool, the executing agent is configured to invoke the large language model to: Identify crop and genetic identifiers from the input; Generate extended identifiers in other naming conventions based on the gene identifier; as well as Based on the crop, the gene identifier, and the extended identifier, the function call instruction is generated in accordance with the format specified in the tool description document of the gene function summarization tool.

8. The system as described in claim 5, characterized in that, In response to the target tool being identified as the multi-trait-gene mapping tool, The planning agent is configured to invoke the large language model to: A first tool invocation instruction is generated based on the multi-trait-gene mapping tool, the first tool invocation instruction including the instruction of the multi-trait-gene mapping tool; The executing agent is configured to invoke the large language model to: Based on the input, the task description, the first tool documentation, and the first tool call instruction, a first function call instruction is generated. The first function call instruction includes the instructions of the multi-trait-gene mapping tool and first structured information related to the crop gene problem in a format conforming to the multi-trait-gene mapping tool. The knowledge mining module is configured as follows: Based on the first structured information in the first function call instruction, the multi-trait-gene mapping tool is invoked to perform a first structured query on the knowledge graph to obtain a first query output; The planning agent is further configured to invoke the large language model to: Receive the output of the first query; In response to determining that the first query output includes genes associated with the plurality of traits, the target tool to be invoked is determined to be the gene function summary tool; as well as A second tool invocation instruction is generated based on the gene function summarization tool, the second tool invocation instruction including the instruction of the gene function summarization tool; The executing agent is further configured to invoke the large language model to: Receive the first query output, the second tool documentation, and the second tool invocation command; as well as Based on the first query output, the second tool documentation, and the second tool call instruction, a second function call instruction is generated. The second function call instruction includes the instructions of the gene function summarization tool and second structured information associated with the first query output that conforms to the format of the gene function summarization tool. The knowledge mining module is further configured to: Receive the second function call instruction; and Based on the second structured information in the second function call instruction, the gene function summary tool is invoked to perform a second structured query on the knowledge graph to obtain a second query output, wherein the answer to the crop gene question is generated by the large language model based on the first query output and the second query output.

9. The system as described in claim 5, characterized in that, In response to the target tool being identified as the gene function summarization tool, The planning agent is configured to invoke the large language model to: A first tool invocation instruction is generated based on the gene function summarization tool, the first tool invocation instruction including the instruction of the gene function summarization tool; The executing agent is configured to invoke the large language model to: Based on the input, the task description, the first tool documentation, and the first tool invocation instruction, a first function invocation instruction is generated. The first function invocation instruction includes the instructions of the gene function summarization tool and first structured information related to the crop gene problem in a format conforming to the gene function summarization tool. The knowledge mining module is configured as follows: Based on the first structured information in the first function call instruction, the gene function summary tool is invoked to perform a first structured query on the knowledge graph to obtain a first query output; The planning agent is further configured to invoke the large language model to: Receive the output of the first query; In response to determining that the first query output includes multiple traits associated with the gene, the target tool to be invoked is determined to be the multi-trait-gene mapping tool; as well as A second tool invocation instruction is generated based on the multi-trait-gene mapping tool, the second tool invocation instruction including the instruction of the multi-trait-gene mapping tool; The executing agent is further configured to invoke the large language model to: Receive the first query output, the second tool documentation, and the second tool invocation command; as well as Based on the first query output, the second tool documentation, and the second tool call instruction, a second function call instruction is generated. The second function call instruction includes the instructions of the multi-trait-gene mapping tool and second structured information associated with the first query output that conforms to the format of the multi-trait-gene mapping tool. The knowledge mining module is further configured to: Receive the second function call instruction; and Based on the second structured information in the second function call instruction, the multi-trait-gene mapping tool is invoked to perform a second structured query on the knowledge graph to obtain a second query output, wherein the answer to the crop gene question is generated by the large language model based on the first query output and the second query output.

10. The system as claimed in claim 1, characterized in that, The large language model was trained based on the following: The mapping relationship between multiple gene identifiers of the same crop gene in different nomenclature standards; and Mapping relationship between multiple expressions of the same trait and standard trait annotation fields.

11. A method for crop gene question answering, characterized in that, Includes the following steps: S1: The large language model receives input from the user, which includes crop gene questions; S2: The large language model generates a task description based on the input; S3: The planning agent calls the large language model to determine the target tool to be called from among the multiple crop gene tools included in the knowledge mining module based on the input, the task description, and the tool description document provided by the knowledge mining module; S4: The planning agent invokes the large language model to generate a tool invocation instruction based on the target tool, the tool invocation instruction including an indication of the target tool; S5: The executing agent invokes the large language model to generate a function call instruction based on the input, the task description, the tool documentation, and the tool call instruction. The function call instruction includes an indication of the target tool and structured information related to the crop gene problem in a format consistent with the target tool. S6: The knowledge mining module, based on the structured information in the function call instruction, invokes the target tool indicated in the function call instruction to perform a structured query on the knowledge graph to obtain query output, wherein the knowledge graph is stored in a storage unit and includes a crop gene dataset; and S7: The large language model generates an answer to the crop gene question based on the query output. Among them, the plurality of crop gene tools include a multi-trait-gene mapping tool, which is configured to: A structured query sequence is generated based on the structured information in the function call instruction. The structured information includes crops and multiple traits associated with the crop genetic problem. The structured query sequence is used to perform a search for genes associated with the multiple traits of the crop and to perform logical operations on the search results. The structured query sequence is executed on the knowledge graph to obtain the query output.

12. The method as described in claim 11, characterized in that, S3 includes: S31: Using a directed acyclic graph, candidate tools associated with the crop gene problem are screened from a list of crop gene tools, the list of crop gene tools including the plurality of crop gene tools; and S32: Based on the few-sample cue words and the crop gene question, determine the target tool from the candidate tools.

13. The method as described in claim 11, characterized in that, In response to the tool invocation instruction indicating that the target tool is the multi-trait-gene mapping tool, S5 includes: S51: Identify the crop and the plurality of traits from the input; and S52: Based on the crop and the multiple traits, generate the function call instruction according to the format specified in the tool description document of the multi-trait-gene mapping tool. The function call instruction includes the instructions of the multi-trait-gene mapping tool and the structured information of the crop and the multiple traits that conforms to the format of the multi-trait-gene mapping tool.

14. The method as described in claim 11, characterized in that, The multi-trait-gene mapping tool is configured to: Map the multiple traits in the structured information of the function call instruction to the standard trait annotation field; The structured query sequence is generated based on the crop, the mapped standard trait annotation fields, and the template sequence in the structured information; The structured query sequence is executed on the database associated with the crop in the knowledge graph to obtain gene records associated with the multiple traits; as well as The gene records are statistically analyzed to generate a structured summary, which is then used as the query output.

15. The method as described in claim 11, characterized in that, The plurality of crop gene tools includes a gene function summarization tool, which is configured to: The function call instruction is executed on the knowledge graph to obtain the query output. The structured information in the function call instruction includes a structured query sequence, which is used to perform a search for the function of genes associated with the crop gene problem.

16. The method as described in claim 15, characterized in that, The gene function summarization tool is configured to: The structured query sequence in the function call instruction is executed on the knowledge graph to obtain query results associated with the gene in different types of node data in the knowledge graph; as well as The query results are concatenated into query text, which is then used as the query output.

17. The method as described in claim 15, characterized in that, In response to the tool invocation instruction indicating that the target tool is the gene function summarization tool, S5 includes: S53: Identify crop and gene identifiers from the input; S54: Generate extended identifiers in other naming conventions based on the gene identifier; and S55: Based on the crop, the gene identifier, and the extended identifier, generate the function call instruction in accordance with the format specified in the tool description document of the gene function summary tool.

18. The method as described in claim 15, characterized in that, In response to the target tool being identified as the multi-trait-gene mapping tool, S4 includes: S41: Generate a first tool invocation instruction based on the multi-trait-gene mapping tool, wherein the first tool invocation instruction includes an instruction from the multi-trait-gene mapping tool; S5 includes: S56: Based on the input, the task description, the first tool description document, and the first tool call instruction, generate a first function call instruction, wherein the first function call instruction includes the indication of the multi-trait-gene mapping tool and first structured information related to the crop gene problem in a format conforming to the multi-trait-gene mapping tool; S6 includes: S61: Based on the first structured information in the first function call instruction, the multi-trait-gene mapping tool is invoked to perform a first structured query on the knowledge graph to obtain a first query output; S4 further includes: S42: In response to determining that the first query output includes genes associated with the plurality of traits, the target tool to be invoked is identified as the gene function summarization tool; and S43: Generate a second tool invocation instruction based on the gene function summarization tool, wherein the second tool invocation instruction includes an instruction from the gene function summarization tool; S5 further includes: S57: Based on the first query output, the second tool description document, and the second tool call instruction, generate a second function call instruction, wherein the second function call instruction includes the instructions of the gene function summarization tool and second structured information associated with the first query output that conforms to the format of the gene function summarization tool; S6 further includes: S62: Based on the second structured information in the second function call instruction, the gene function summary tool is invoked to perform a second structured query on the knowledge graph to obtain a second query output; S7 includes: S71: The large language model generates an answer to the crop gene question based on the first query output and the second query output.

19. The method as described in claim 15, characterized in that, In response to the target tool being identified as the gene function summarization tool, S4 includes: S44: Generate a first tool invocation instruction based on the gene function summarization tool, wherein the first tool invocation instruction includes an instruction from the gene function summarization tool; S5 includes: S58: Based on the input, the task description, the first tool description document, and the first tool call instruction, generate a first function call instruction, wherein the first function call instruction includes the instructions of the gene function summarization tool and first structured information related to the crop gene problem in a format conforming to the gene function summarization tool; S6 includes: S63: Based on the first structured information in the first function call instruction, the gene function summary tool is invoked to perform a first structured query on the knowledge graph to obtain a first query output; S4 further includes: S45: In response to determining that the first query output includes multiple traits associated with the gene, the target tool to be invoked is determined to be the multi-trait-gene mapping tool; and S46: Generate a second tool invocation instruction based on the multi-trait-gene mapping tool, the second tool invocation instruction including the indication of the multi-trait-gene mapping tool; S5 further includes: S59: Based on the first query output, the second tool description document, and the second tool call instruction, generate a second function call instruction. The second function call instruction includes the indication of the multi-trait-gene mapping tool and second structured information associated with the first query output that conforms to the format of the multi-trait-gene mapping tool. S6 further includes: S64: Based on the second structured information in the second function call instruction, the multi-trait-gene mapping tool is invoked to perform a second structured query on the knowledge graph to obtain a second query output; S7 includes: S71: The large language model generates an answer to the crop gene question based on the first query output and the second query output.

20. The method as described in claim 11, characterized in that, The large language model was trained based on the following: The mapping relationship between multiple gene identifiers of the same crop gene in different nomenclature standards; and Mapping relationship between multiple expressions of the same trait and standard trait annotation fields.

21. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 11-20.

22. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 11-20.

Images