Task assistance method and device, electronic equipment and storage medium

Abstract

The invention provides a task assistance method and device, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence, and the method comprises the steps: determining a task implementation process needed for completing an omics task through a coarse-grained planning group; traversing each process step in the task implementation process in sequence, and executing the following operations: determining specific implementation steps required for completing each process step through a fine-grained programming group; and executing the specific implementation steps through the action execution group to obtain a target omics result after the specific implementation steps are completed. According to the task assistance method and device, the electronic equipment and the storage medium provided by the invention, a method for carrying out omics research by combining a multi-agent system based on a large language model and bioinformatics research is provided, and an agent group grouping management mode is creatively designed; the interactive information redundancy and meaningless dialogue between the intelligent agents can be reduced, efficient and accurate information flow is ensured, and the efficiency and accuracy of omics research between the intelligent agents and the intelligent agents are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a task assistance method, device, electronic device, and storage medium. Background Art

[0002] In recent years, driven by the exponential growth of complex and multimodal biological data and the rapid development of artificial intelligence technology, the field of bioinformatics research is undergoing profound changes. Although the combination of bioinformatics and artificial intelligence technology has made remarkable progress, current life science research still faces many challenges. Especially when dealing with large-scale, high-dimensional, and multimodal biological data, advanced artificial intelligence technology is needed to support the scalable, adaptable, and interpretable processing of data. However, existing single-agent systems have limitations in coping with these complex tasks.

[0003] Although a single-agent system can support task planning and tool invocation to complete tasks specified by users, its capabilities are restricted by the performance of large language models when dealing with complex omics tasks. For example, when dealing with extremely long texts, multi-turn dialogue scenarios, or complex contexts, a single agent may misunderstand the user's intention due to limited memory retention length and understanding of context relevance, and provide responses that are not fully relevant. In addition, a single agent usually decomposes tasks into linear workflows, lacks the ability to analyze problems from multiple perspectives and macroscopically, and has low execution efficiency, making it difficult to respond promptly to dynamic environmental changes.

[0004] To overcome the limitations of single-agent systems, multi-agent systems based on large language models have emerged. A multi-agent system consists of multiple agents with different capabilities and roles that work together to simulate a complex real-world environment through interactions between agents, forming swarm intelligence to solve more dynamic and complex tasks. However, existing multi-agent systems also face some challenges, such as high system design and implementation difficulties, complex interaction processes prone to information redundancy, high computational resource consumption, and high customization costs. Summary of the Invention

[0005] The present invention provides a task assistance method, device, electronic device, and storage medium to solve the defects in the prior art that a single agent lacks the ability to analyze problems from multiple perspectives and macroscopically, has low execution efficiency, and a multi-agent system has high design and implementation difficulties, complex interaction processes, high computational resource consumption, and high customization costs.

[0006] The present invention provides a task assistance method, including the following steps: Obtain an omics task issued by a user; Determine the task implementation process required to complete the omics task through a coarse-grained planning group; Traverse each process step in the task implementation process in sequence and perform the following operations: Determine the specific implementation steps required to complete each of the process steps through the fine-grained planning group; execute the specific implementation steps through the action execution group to obtain the target omics results after completing the specific implementation steps; The coarse-grained planning group, the fine-grained planning group, and the action execution group are groups composed of multiple identical or different agents.

[0007] According to a task assistance method provided by the present invention, the coarse-grained planning group at least includes a coarse-grained planning agent, a first proxy agent, and a coarse-grained review agent; The determining of the task implementation process required to complete the omics task through the coarse-grained planning group includes: Generate an initial implementation process through the coarse-grained planning agent according to the omics task; Obtain the first review result of the initial implementation process by the coarse-grained review agent; If it is determined that the first review result includes process improvement suggestions, control the coarse-grained planning agent to generate a new initial implementation process according to the process improvement suggestions, and have the coarse-grained review agent re-review the new initial implementation process until it is determined that the first review result is passed; Use the initial implementation process when the first review result is passed as the task implementation process and send it to the first proxy agent to initiate a first human-computer interaction through the first proxy agent; The first human-computer interaction is used to obtain the adjustment suggestions of the user for the task implementation process, and the adjustment suggestions are used for the coarse-grained planning agent to re-adjust the initial implementation process.

[0008] According to a task assistance method provided by the present invention, the coarse-grained planning group is temporarily constructed based on the omics task, and different coarse-grained planning groups are constructed for different omics tasks.

[0009] According to a task assistance method provided by the present invention, the fine-grained planning group at least includes a fine-grained planning agent, a second proxy agent, and a fine-grained review agent; The determining of the specific implementation steps required to complete each of the process steps through the fine-grained planning group includes: Generate an initial implementation step through the fine-grained planning agent according to the process step; Obtain the second review result of the initial implementation step by the fine-grained review agent; If it is determined that the second review result includes step improvement suggestions, control the fine-grained planning agent to generate new initial implementation steps according to the step improvement suggestions, and re-review the new initial implementation steps through the fine-grained review agent until it is determined that the second review result is passed; Use the initial implementation steps when the second review result is passed as the specific implementation steps and send them to the second proxy agent, so that the second proxy agent initiates a second human-computer interaction; The second human-computer interaction is used to obtain the user's adjustment suggestions for the specific implementation steps, and the adjustment suggestions are used for the fine-grained planning agent to re-adjust the initial implementation steps.

[0010] According to a task assistance method provided by the present invention, the fine-grained planning group is temporarily constructed according to the process steps, and different process steps construct different fine-grained planning groups.

[0011] According to a task assistance method provided by the present invention, the action execution group at least includes a coding agent, a third proxy agent, and a debugging agent; Executing the specific implementation steps through the action execution group to obtain the target omics result after completing the specific implementation steps includes: Generating initial omics code by the coding agent according to the specific implementation steps; Debugging the initial omics code by the debugging agent, so that the coding agent regenerates new initial omics code according to the debugging result until it is determined that the debugging result of the debugging agent for the initial omics code is passed; Obtain the initial omics code when the debugging result is passed as the target omics code; Execute the target omics code through the third proxy agent to generate the target omics result.

[0012] According to a task assistance method provided by the present invention, after generating the initial omics code by the coding agent according to the specific implementation steps, it further includes: Initiating a third human-computer interaction through the third proxy agent; The third human-computer interaction is used to obtain the user's adjustment suggestions for the initial omics code.

[0013] According to a task assistance method provided by the present invention, after debugging the initial omics code by the debugging agent, it further includes: If the debugging agent determines that a preset trigger condition is met, initiate a fourth human-computer interaction through the third proxy agent; The fourth human-computer interaction is used to obtain the user's adjustment suggestions for the debugging results, and the adjustment suggestions are used for the coding agent to readjust the initial omics code.

[0014] According to a task assistance method provided by the present invention, the preset trigger conditions include: The debugging agent generates the same code error information when running the initial omics code twice in succession; Or, the number of code error messages generated by the debugging agent when running the initial omics code is greater than a preset threshold; Or, a preset sensitive word appears in the code error information generated by the debugging agent when running the initial omics code; Or, the debugging agent lacks necessary files or parameters when running the initial omics code.

[0015] According to a task assistance method provided by the present invention, it further includes: Receiving an information call request from any one of the coarse-grained planning group, the fine-grained planning group, or the action execution group; Responding to the information call request, establishing a connection for any one of the agents to call knowledge from the interdisciplinary knowledge base; The interdisciplinary knowledge base is constructed by extracting knowledge from various disciplines in the environment and encoding it in a form readable by agents.

[0016] According to a task assistance method provided by the present invention, it further includes: Receiving a self-learning information call request from any one of the coarse-grained planning group, the fine-grained planning group, or the action execution group; Responding to the self-learning information call request, sending historical files related to any one of the agents to any one of the agents for the any one of the agents to conduct reflective learning based on the historical files; The historical files include human-computer interaction information and / or task interaction logs.

[0017] According to a task assistance method provided by the present invention, after obtaining the target omics result after completing the specific implementation steps, it further includes: Storing the target omics result corresponding to each specific implementation step for an object.

[0018] The present invention also provides a task assistance device, including the following components: A task receiving unit, used to obtain an omics task issued by a user; A coarse-grained assistance unit, used to determine the task implementation process required to complete the omics task through a coarse-grained planning group; A fine-grained assistance unit is used to sequentially traverse each process step in the task implementation process and perform the following operations: Determine the specific implementation steps required to complete each of the process steps through a fine-grained planning group; execute the specific implementation steps through an action execution group to obtain the target omics results after completing the specific implementation steps; Among them, the coarse-grained planning group, the fine-grained planning group, and the action execution group are groups composed of multiple identical or different intelligent agents.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the program, it implements the task assistance method as described in any one of the above.

[0020] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the task assistance method as described in any one of the above.

[0021] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the task assistance method as described in any one of the above.

[0022] The task assistance method, device, electronic device, and storage medium provided by the present invention provide a method for combining a multi-agent system based on a large language model with bioinformatics research for omics research and pioneeringly design a way of group management of intelligent agent groups, which can reduce information redundancy and meaningless conversations in the interaction between intelligent agents, ensure efficient and accurate information flow, and effectively improve the efficiency and accuracy of omics research between the two. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a flowchart of the task assistance method provided by the present invention.

[0025] Figure 2 It is a schematic diagram of the architecture of the task assistance system provided by the present invention.

[0026] Figure 3 It is a schematic diagram of the working process of the task assistance system provided by the present invention.

[0027] Figure 4 It is a schematic structural diagram of the task assistance device provided by the present invention.

[0028] Figure 5 It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0029] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] It should be noted that in the description of the present invention, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusions, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or device. Without further limitations, the element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element. The orientation or positional relationship indicated by the terms "upper", "lower", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation of the present invention. Unless otherwise clearly defined and limited, the terms "mounted", "connected" and "coupled" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the internal communication of two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0031] The terms "first", "second", etc. in the present invention are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object may be one or multiple. In addition, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the associated objects before and after.

[0032] Currently, how to combine the advantages of artificial intelligence technology and biological big data to drive bioinformatics research is a continuous and arduous pursuit. Researchers have professional knowledge in a single bioinformatics field, but lack interdisciplinary expertise and the scientific research ability of interdisciplinary integration. The effective application of artificial intelligence in bioinformatics usually requires a profound understanding of biological domain knowledge and advanced computing technologies. Therefore, it is necessary to strengthen interdisciplinary cooperation to bridge the domain gap and unlock the analytical capabilities of artificial intelligence-driven bioinformatics research.

[0033] In addition, the rapid accumulation of high-dimensional and multimodal biological data, characterized by large scale, complexity, diversity, noise, and context-dependence, has given rise to the need for advanced artificial intelligence technologies. Artificial intelligence should not only support the processing capabilities of biological big data in terms of scalability, adaptability, and interpretability, but also support the complementary integration with expert knowledge to promote the improvement of intelligence levels.

[0034] With the breakthrough development of deep learning technologies, especially the introduction of large language models, intelligent agents based on large language models have powerful natural language processing capabilities. They can autonomously perceive the external environment, think and make decisions, and call tools to execute decisions. This has become one of the most promising directions towards general artificial intelligence. Intelligent agents execute tasks in the form of perception-decision-execution. After execution, they reflect on and evaluate the effects, and then learn knowledge based on this and store it in the memory module to optimize the future action decisions of the intelligent agent. With their autonomous ability to perceive the external environment and think and make decisions, as well as their learning and adaptation abilities during the interaction with the environment, intelligent agents exhibit intelligent behaviors similar to humans. Human expertise can continuously improve and enhance intelligent agents, while intelligent agents generate new insights through the integration of multidisciplinary knowledge and the mining of biological big data, expanding human understanding of life sciences. By leveraging their complementary advantages, bioinformatics research has gained the ability to traverse unknown knowledge domains, generate testable hypotheses, identify new research directions, and optimize experimental designs to accelerate the pace of discovery, revealing previously unexplored associations and emerging features in complex biological systems.

[0035] Common single-agent frameworks include BabyAGI, Hugging, etc. These frameworks support task planning and tool invocation to complete tasks specified by users. For example, the Coscientist system can autonomously conduct chemical research, including automated planning and execution of chemical experiments, as well as predicting through models and optimizing reaction paths. Researchers complete instances by communicating with the Coscientist system based on large language models and perform six key tasks: planning the synthesis path of known compounds using public data, efficiently retrieving and navigating hardware documents, executing advanced instructions in a cloud laboratory, precisely controlling liquid handling equipment, simultaneously using multiple hardware modules to handle complex tasks and integrating multi-source data, and analyzing past experimental data for optimizing problem-solving.

[0036] Another example is that the ChemCrow system based on large language models can autonomously plan and implement tasks such as organic compound synthesis, drug discovery, and material design. The ChemCrow system enhances the system's ability to obtain and understand external information by integrating tools such as web search, literature search, and interactive programming environments, demonstrating the possibility of improving the ability to handle scientific tasks.

[0037] In the field of pathology, the multi-modal generative artificial intelligence-assisted system PathChat demonstrates the ability to perform human-computer interactive differential diagnosis by combining histological images, clinical data, and subsequent immunohistochemistry (IHC) results in cases of unknown primary cancer, aiming to assist pathologists in diagnosing and understanding complex cases and improving the accuracy and efficiency of pathological diagnosis. PathChat is designed to handle open-ended questions and multiple-choice questions related to pathology, demonstrating its potential in supporting pathological diagnosis.

[0038] Although large language models have promoted the development of agent technology, single agents still have some limitations. First, the capabilities of agents are restricted by the performance of large language models. For example, due to the poor quality of training data, agents are prone to making biased, incorrect, or outdated action decisions when dealing with complex tasks. When dealing with extremely long texts, multi-turn dialogue scenarios, or complex contexts, limited by the memory retention length of large language models and the understanding of prompt words and context relevance, agents may misunderstand the user's intention and provide responses that are not fully relevant. Second, although single agents can play multiple roles, they will quickly collapse to a specific perspective with the change of prompt words or the previous few rounds of conversations, thus lacking the ability to analyze problems from multiple perspectives and macroscopically. Finally, single agents usually decompose tasks into linear workflows, with low execution efficiency and difficulty in responding promptly to dynamic environmental changes.

[0039] Large language model-based single agents focus on internal thinking and decision-making processes and interactions with the external environment. Inspired by human group collaboration in solving complex tasks, large language model-based multi-agent systems focus on diverse agent groups, interactions between agents, and collective decision-making processes. Compared with single agents, multi-agent systems consist of multiple agents with different capabilities and roles working together, simulating complex real-world environments through interactions between agents, and making full use of the advantages of different agents to form swarm intelligence to solve more dynamic and complex tasks, such as software development, scientific experiments, and embodied intelligence.

[0040] In recent years, several multi-agent frameworks that have received much attention have emerged in the open-source community, such as CAMEL, AutoGen, etc. CAMEL uses prompt engineering techniques to define the roles of agents, promote autonomous collaboration between agents, and guide agent conversations towards task goals. AutoGen is a highly customizable multi-agent framework that supports researchers in customizing large language models, communication mechanisms, workflow management, etc. of agents to solve complex tasks.

[0041] However, multi-agent systems also face some challenges, mainly manifested in the complex interaction process when multi-agent systems face dynamic environments, where information redundancy and meaningless conversations are likely to occur, affecting task execution efficiency.

[0042] Advances in high-throughput sequencing technology and artificial intelligence have provided unprecedented opportunities for bioinformatics research. However, addressing the challenges brought about by the exponential growth of omics data and the rapid development of artificial intelligence technologies requires intelligent big biological data analysis capabilities and interdisciplinary knowledge-driven scientific insights.

[0043] Among them, omics is a comprehensive discipline that mainly studies the collection of various components of organisms, including genomes, transcriptomes, proteomes, metabolomes, etc., aiming to systematically and comprehensively analyze and study the various molecular compositions of organisms through high-throughput experimental techniques and data analysis methods.

[0044] The following combines Figures 1 - 5 Describe the task assistance method, device, electronic device, and storage medium provided by the present invention, aiming to introduce an innovative design concept of interdisciplinary cooperation, sharing knowledge resources, and continuous human-computer interaction feedback, especially providing innovative designs in aspects such as efficient human-computer interaction, agent group management, interdisciplinary knowledge empowerment, and agent continuous learning. These innovative designs strengthen the task assistance method and device provided by the present invention in aspects of language intelligence, social intelligence, knowledge empowerment, and learning intelligence, and can effectively make up for or at least alleviate some of the shortcomings existing in existing multi-agent systems.

[0045] Figure 1It is a schematic flowchart of the task assistance method provided by the present invention. As Figure 1 shown, its execution subject can be a processor, a computer, an industrial control computer, or other devices capable of data processing, storage, operation, and communication interaction with the external environment. For the convenience of description, it will be referred to as a coordination processor in the following embodiments. The provided task assistance method includes but is not limited to the following steps: First, obtain the omics task issued by the user.

[0046] The user first inputs a specific omics task through a user interface (UI) or application programming interface (API) that is communicatively connected to the coordination processor. For example, if the user hopes to assist in studying the gene mutations of a certain specific genetic disease, the task description input by the user may include the research purpose, expected output results, relevant sample information, etc.

[0047] Then, determine the task implementation process required to complete the omics task through a coarse-grained planning group.

[0048] After receiving the omics task issued by the user, the coordination processor passes it to the coarse-grained planning group. The coarse-grained planning group can be composed of multiple agents with different domain knowledge and planning capabilities, and can initially disassemble and plan the omics task from a macroscopic perspective to obtain the task implementation process required to complete the omics task.

[0049] The multiple agents in the coarse-grained planning group can be selected from the agent cluster according to the requirements for completing the omics task. Through the cooperation of these agents, the main process steps required to complete the task according to the user task can be determined. For example, for the study of gene mutations in genetic diseases, the steps that may need to be executed include data collection, data preprocessing, gene mutation detection, mutation result analysis, etc.

[0050] The coarse-grained planning group feeds back the determined task implementation process to the coordination processor, and can also be synchronously displayed on the user interface through the coordination processor.

[0051] As an alternative embodiment, after receiving the omics task issued by the user, the coordination processor can first call a retrieval agent to perform a retrieval in the existing knowledge base and other prior arts. The retrieval agent is an agent that can comprehensively retrieve the prior arts.

[0052] If it is found through retrieval that there is an auxiliary method in the prior art (such as an existing knowledge base) that can completely solve the user's omics task, the retrieval result is directly used as the target omics result and fed back to the user through the coordination processor, and this task assistance ends.

[0053] If the retrieval agent does not retrieve an auxiliary method from the prior art that can completely solve the omics task input by the user, the coarse-grained planning group that has been constructed is used to analyze the implementation process of the omics task.

[0054] After obtaining the task implementation process determined by the coarse-grained planning group, the coordination processor will call the constructed fine-grained planning group and action execution group to sequentially traverse each process step in the task implementation process, and mainly perform the following operations: The specific implementation steps required to complete each of the said process steps are determined by the fine-grained planning group. That is, for each process step in the task implementation process, it is first passed to the fine-grained planning group. The fine-grained planning group is composed of multiple agents with detailed domain knowledge and execution capabilities, and can further disassemble and refine the process steps from a microscopic perspective.

[0055] The fine-grained planning group can determine the specific implementation steps required to complete the current process step according to the content and requirements of the current process step. For example, for the data preprocessing step, the specific implementation steps that may need to be performed include data cleaning, data formatting, data standardization, etc.

[0056] The fine-grained planning group returns the specific implementation steps determined for the current process step to the coordination processor, and then the coordination processor passes it to the action execution group. The action execution group is composed of multiple agents with actual operation and execution capabilities, and can perform relevant operations according to the requirements of the specific implementation steps.

[0057] Each agent in the action execution group will perform relevant operations and obtain results according to the content and requirements of the specific implementation steps. For example, for the data cleaning step, the operations that may need to be performed include removing invalid data, filling in missing values, handling abnormal data, etc.

[0058] After completing the execution of the specific implementation steps related to the current process step, the action execution group will return the obtained target omics result to the coordination processor.

[0059] After the coordination processor determines that the current process step has been executed, it continues to send the next process step to the fine-grained planning group for step division.

[0060] Iteratively execute the above steps until it is determined that all process steps in the entire task implementation process have been executed, and obtain the target omics results corresponding to each process step.

[0061] The task assistance method provided by the present invention provides a method for combining a multi-agent system based on a large language model with bioinformatics research for omics research, and innovatively designs a method for grouping and managing agent groups, which can reduce the information redundancy and meaningless conversations in the interaction between agents, ensure efficient and accurate information flow, and effectively improve the efficiency and accuracy of omics research between the two.

[0062] As an alternative embodiment, after the coordination processor obtains the target omics results after completing the specific implementation steps related to each process step, it further includes: Perform object storage on the target omics results corresponding to each of the specific implementation steps.

[0063] Taking a specific bioinformatics analysis process as an example, the coordination processor is configured to manage and coordinate the execution of multiple process steps related to omics tasks, and these process steps are designed to analyze specific genomic data.

[0064] First, the coordination processor sequentially triggers and executes each process step according to the process order of the task implementation process determined by the coarse-grained planning group. For example, sequentially execute 4 process steps including a quality control step, a sequence alignment step, a variant detection step, and a gene function annotation step. Among them, the quality control step cleans and filters out low-quality DNA sequence data through multiple specific implementation steps; the sequence alignment step aligns the filtered DNA sequence data with the reference genome through multiple specific implementation steps; the variant detection step identifies variant information in the sequence through multiple specific implementation steps, such as detecting single nucleotide polymorphisms (SNPs); the gene function annotation step provides a biological meaning explanation for the detected variants through multiple specific implementation steps.

[0065] Furthermore, the coordination processor will sequentially traverse the above 4 process steps according to the method provided in the above embodiment, and sequentially obtain the target omics results corresponding to each process step.

[0066] Specifically, whenever a process step is completed, the coordination processor retrieves the target omics results output by the action execution group. For example, after the quality control step, the retrieved data set is a cleaned high-quality DNA sequence data set; after the sequence alignment step, the retrieved result file includes the position information of each DNA sequence on the reference genome; after the variant detection step, the retrieved information list contains all detected variants; and after the gene function annotation step, the retrieved annotation information includes the gene name corresponding to each variant, functional impact, etc.

[0067] After the coordination processor retrieves the target omics results corresponding to each specific implementation step, it also performs an additional step of object storage for these target omics results. In the present invention, the result data is stored in a dedicated object storage system in the form of objects (such as files or data blocks) because the object storage system can provide high scalability, data persistence, and access speed, which is very suitable for storing and processing large-scale bioinformatics data.

[0068] The coordination processor can create a unique object identifier (such as an object key or file name) for the target omics results of each process step by calling the API interface of the object storage system, and upload the data related to the target omics results as the content of the object to the specified storage bucket (or container). At the same time, it can also add metadata to these objects, such as description information, creation time, step name, etc., for subsequent data retrieval and management.

[0069] Once the results of all process steps are securely stored in the object storage system, the coordination processor can further perform subsequent operations, such as generating a final report, triggering a downstream analysis process, or notifying the user that the results are ready.

[0070] The task assistance method provided by the present invention not only effectively manages and coordinates the analysis of omics tasks and the execution of process steps, but also ensures that the omics results generated by each process step can be securely and efficiently stored and managed, providing a solid foundation for subsequent data analysis and mining.

[0071] To more clearly describe the task assistance method provided by the present invention, the following will explain from the perspective of the system architecture on which the task assistance method runs.

[0072] Figure 2 is a schematic diagram of the architecture of the task assistance system provided by the present invention. As Figure 2 shown, the task assistance system on which the task assistance method provided by the present invention relies can be mainly regarded as consisting of three parts, including an agent cluster, a coordination layer, and an environment.

[0073] Among them, the agent cluster includes multiple agents. An agent is the core component unit of a multi-agent system. The agents mentioned in the subsequent embodiments of the present invention use a large language model as the control core, have autonomy, can perceive information in the environment and understand complex natural language instructions (such as the prompt words of experts in the environment), think and make decisions, and call tools to execute actions. Different agents can assume different roles and responsibilities, such as information collection, task planning, code generation, task execution, etc.

[0074] The coordination layer provides a bridge for agents to perceive the environment and feedback to affect the environment. The virtual coordination layer processor can be regarded as the execution subject of the task assistance method provided by the present invention, and can coordinate the behaviors of each agent simultaneously. On the basis of the task planning and allocation, resource scheduling, conflict resolution, cooperation strategy, etc. executed by existing agents, the present invention further develops functions such as agent group management strategy, more efficient human-agent interaction mechanism, interdisciplinary shared knowledge base, and agent continuous learning strategy in the coordination layer, effectively improving the task assistance ability.

[0075] The environment refers to the external conditions under which a multi-agent system operates. The state of the environment will affect the behavior of agents, and at the same time, the actions of agents will also change the environmental state. Specifically, the environment can include elements of the real world, such as users, instruments, devices, tools, etc., and can also include elements of the virtual world, such as data sets, documents, knowledge bases, software, etc.

[0076] The coarse-grained planning group is composed of multiple agents selected from the agent cluster that can cooperate to achieve the coarse-grained planning of omics tasks, mainly including but not limited to coarse-grained planning agents, first proxy agents, and coarse-grained review agents; The task implementation process required to complete the omics task determined by the coarse-grained planning group includes: The coarse-grained planning agent generates an initial implementation process according to the omics task; Obtain the first review result of the initial implementation process by the coarse-grained review agent; If it is determined that the first review result includes process improvement suggestions, control the coarse-grained planning agent to generate a new initial implementation process according to the process improvement suggestions, and let the coarse-grained review agent re-review the new initial implementation process until it is determined that the first review result is passed; When the first review result is passed, use the initial implementation process as the task implementation process and send it to the first proxy agent to initiate the first human-computer interaction through the first proxy agent; The first human-computer interaction is used to obtain the user's adjustment suggestions for the task implementation process, and the adjustment suggestions are used by the coarse-grained planning agent to re-adjust the initial implementation process.

[0077] Specifically, the coarse-grained planning agent is responsible for initially generating an implementation process (referred to as the initial implementation process) according to the input omics task.

[0078] After the coarse-grained planning agent completes the initial planning, an agent is responsible for initiating a human-computer interaction with the user. The purpose is to receive the user's adjustment suggestions for the initial implementation process and feedback these adjustment suggestions to the coarse-grained planning agent for subsequent process optimization.

[0079] The coarse-grained review agent is mainly used to review the initial implementation process generated by the coarse-grained planning agent. It mainly checks the rationality, feasibility of the initial implementation process, and whether there are potential errors or improvement spaces, etc. If the review result (referred to as the first review result) output by the coarse-grained review agent contains process improvement suggestions, then the coordination processor will forward them to the coarse-grained planning agent, and the coarse-grained planning agent can generate a new initial implementation process according to these process improvement suggestions and have it reviewed again by the coarse-grained review agent until the review result is passed.

[0080] Figure 3 It is a schematic diagram of the working process of the task assistance system provided by the present invention. The following will be combined with Figure 3 As shown, it will be described in detail how the task assistance method provided by the present invention uses the coarse-grained planning group to implement the determination process of the task implementation process.

[0081] Suppose the task description input by the user mainly includes "perform hierarchical clustering. The total number of clustering levels is 4. For the final human lung cell atlas, the marker genes of each cluster also need to be calculated", then it can be determined that the omics task input by the user is to perform hierarchical clustering on single-cell RNA sequencing data using a machine learning algorithm to identify different human lung cell types, hoping to assist in completing this process and providing the marker genes of each cluster.

[0082] The coarse-grained planning agent will perform a preliminary analysis of the omics task according to the omics task issued by the user and generate a preliminary implementation process. This preliminary implementation process describes the main steps required to complete the omics task and the logical relationship between these main steps, mainly including initial implementation processes such as data preprocessing, hierarchical clustering, and calculation of marker genes.

[0083] After receiving the initial implementation process, the coarse-grained review agent conducts a detailed review of it, mainly checking whether each process step in the process is clear, reasonable, and whether there are potential errors or omissions.

[0084] Suppose that during the review process, the coarse-grained review agent discovers two problems with hierarchical clustering in the initial implementation process: "1. Step combination and decomposition; 2. Method selection". That is, the coarse-grained review agent puts forward some process improvement suggestions and requires the coarse-grained planning agent to supplement or perfect the pointed-out problems.

[0085] According to the process improvement suggestions of the coarse-grained review agent, the coarse-grained planning agent adjusts the initial implementation process, supplements the specific algorithms in the hierarchical clustering steps (such as using the Leiden algorithm for hierarchical clustering) and the improvement of step combination and decomposition (such as initializing a loop in the third step to perform clustering at each level), and resubmits it to the coarse-grained review agent for review.

[0086] The above review process of the coarse-grained review agent and the improvement process of the coarse-grained planning agent may be iterated multiple times. Each iteration involves the coarse-grained planning agent adjusting the current preliminary implementation process according to the review suggestions, and the coarse-grained review agent re-reviewing the adjusted new preliminary implementation process.

[0087] When the coarse-grained review agent confirms that the finally generated preliminary implementation process is correct and gives a passing review result, the iteration process terminates. The initial implementation process at this time is determined as the final task implementation process, which includes clear implementation processes such as data preprocessing, hierarchical clustering, and calculation of marker genes.

[0088] After learning that the task implementation process has been determined, the coordination processor will call the first agent to receive this task implementation process and prepare to initiate a human-computer interaction with the user.

[0089] In the human-computer interaction stage, the first agent will display the task implementation process to the user and invite the user to put forward adjustment suggestions. The user can modify or perfect certain implementation processes or links in the task implementation process according to their professional knowledge and experience, such as the method for calculating marker genes.

[0090] After the user's adjustment suggestions, the optimized task implementation process by the coarse-grained planning agent will be submitted to the coarse-grained review agent for review and confirmation again. This is to ensure that the finally obtained task implementation process not only meets the requirements of the agent review but also satisfies the needs of the user.

[0091] When the coordination processor receives the final review result of passing given by the granularity review agent, the task implementation process is officially determined and can start to be executed. During the execution process, the coordination processor will trigger and execute each of them in turn according to the steps and logical relationships in the task implementation process, including data preprocessing, hierarchical clustering, calculation of marker genes, etc.

[0092] During the execution process, fine-grained planning and adjustment may also be required to ensure that each step can be executed precisely. For example, in the hierarchical clustering stage, the clustering algorithm and parameters are adjusted according to the actual situation of the data; in the stage of calculating marker genes, appropriate statistical methods and thresholds are selected to determine the marker genes, etc.

[0093] The task assistance method provided by the present invention is a complex process involving the collaboration, iteration, and human-computer interaction of multiple agents in the process of determining the task implementation process related to omics tasks. This process ensures that the finally determined task implementation process not only meets the requirements of objectivity in solving omics tasks but also can meet the subjective needs of users, thereby improving the efficiency and accuracy of omics research.

[0094] Based on the content of the above embodiments, as an alternative embodiment, the coarse-grained planning group is temporarily constructed based on the omics task, and different coarse-grained planning groups are constructed for different omics tasks.

[0095] Specifically, the task assistance method provided by the present invention optimizes the task execution process by constructing a coarse-grained planning group to improve the analysis efficiency and accuracy. The coarse-grained planning group is a group composed of multiple identical or different agents (such as artificial intelligence algorithms, machine learning models, expert systems, etc.), and they cooperate together to complete complex omics tasks.

[0096] In each execution process, by decomposing the omics task, each sub-task and its requirements can be identified, so as to determine the type and quantity of agents required according to the requirements of the sub-task, and one or more coarse-grained planning groups are temporarily constructed. Each coarse-grained planning group is composed of multiple identical or different agents, and these agents are selected and combined according to the requirements of the sub-task. The agents within the group complete the omics task through cooperation and communication.

[0097] Inside the coarse-grained planning group, specific tasks will be assigned according to the professional skills and experience of each agent. Each agent performs corresponding operations according to the assigned tasks, such as data preprocessing, algorithm application, result analysis, etc. During the execution process, the agents will also maintain necessary communication and cooperation to ensure the coherence and efficiency of the task. Each coarse-grained planning group integrates its respective results to form the final task implementation process.

[0098] Optionally, after completing the coarse-grained planning work of the omics task, the coarse-grained planning group can be disbanded, that is, the relevant resources (such as computing resources, storage devices, etc.) of each agent are recycled for serving other omics tasks.

[0099] For example, when the coordination processor receives that the omics task to be faced by the coarse-grained planning group is to deeply analyze a batch of genomics data, including gene variant detection, gene expression analysis, gene function annotation, etc., the coordination processor will temporarily construct a coarse-grained planning group, including multiple agents such as a variant detection algorithm agent, a data analysis agent, a computing resource agent, etc. All the above agents cooperate to complete steps such as data preprocessing, variant detection, expression analysis, and function annotation, and finally output a feasible task implementation process.

[0100] As another alternative embodiment, a coarse-grained planning group including multiple agents can be pre-constructed and repeatedly called each time an omics task is executed.

[0101] The task assistance method provided by the present invention can flexibly allocate various agent resources by constructing a coarse-grained planning group, optimize the task execution process, and improve the efficiency and accuracy of omics data analysis. This method is not only applicable to the fields of genomics and proteomics, but also can be extended to other omics fields and complex task scenarios.

[0102] Reference Figure 3 As shown, as an alternative embodiment, the fine-grained planning group at least includes a fine-grained planning agent, a second proxy agent, and a fine-grained review agent; Determining the specific implementation steps required to complete each of the process steps through the fine-grained planning group includes: Generating initial implementation steps by the fine-grained planning agent according to the process steps; Obtaining a second review result of the initial implementation steps by the fine-grained review agent; If it is determined that the second review result includes step improvement suggestions, controlling the fine-grained planning agent to generate new initial implementation steps according to the step improvement suggestions, and re-reviewing the new initial implementation steps by the fine-grained review agent until it is determined that the second review result is passed; Taking the initial implementation steps when the second review result is passed as the specific implementation steps and sending them to the second proxy agent for the second proxy agent to initiate a second human-computer interaction; The second human-computer interaction is used to obtain adjustment suggestions of the user for the specific implementation steps, and the adjustment suggestions are used for the fine-grained planning agent to re-adjust the initial implementation steps.

[0103] Next, based on the fact that the task implementation process has been determined using the coarse-grained planning group, the specific implementation steps of how to further refine the task implementation process of each process step using the fine-grained planning group will be elaborated in detail.

[0104] The fine-grained planning group includes at least three types of agents, such as fine-grained planning agents, proxy agents (referred to as the second proxy agents for ease of description), and fine-grained auditing agents.

[0105] Among them, the fine-grained planning agents are mainly responsible for generating initial specific implementation steps (referred to as initial implementation steps) according to each process step determined by the coarse-grained planning group. These initial implementation steps need to be detailed enough to directly guide subsequent task execution.

[0106] The fine-grained auditing agents are mainly used to audit the initial implementation steps generated by the fine-grained planning agents to ensure that they meet the relevant requirements, safety standards, and other relevant specifications of the omics tasks. If any problems or areas for improvement are found, step improvement suggestions will be provided.

[0107] The second proxy agents mainly act as a bridge for interaction with users. They are responsible for receiving the specific implementation steps output by the fine-grained planning group and initiating human-machine interaction with users through the human-machine interaction interface to obtain user feedback and adjustment suggestions.

[0108] Next, it will be described in detail how the above three types of agents specifically cooperate to achieve the formulation of specific implementation steps related to each process step.

[0109] Based on the task implementation process already determined by the coarse-grained planning group, the fine-grained planning agents generate preliminary refined specific implementation steps (referred to as initial implementation steps) according to the detailed description and context information of each process step. These initial implementation steps are required to include information such as specific operation steps, required resources, and expected results.

[0110] The fine-grained auditing agents receive the initial implementation steps generated by the fine-grained planning agents and conduct a detailed audit. The audit content includes but is not limited to the feasibility, safety, efficiency of the steps, and whether they meet the overall goals of the task.

[0111] If any problems or areas for improvement are found through the audit by the fine-grained auditing agents, the fine-grained auditing agents will generate step improvement suggestions and return them to the fine-grained planning agents through the coordination processor. For example, these step improvement suggestions involve reordering of steps, reallocation of resources, adjustment of operation details, etc.

[0112] In the above situation, the fine-grained planning agents will revise the initial implementation steps they generated according to the step improvement suggestions provided by the fine-grained auditing agents to generate new initial implementation steps. The new initial implementation steps will be submitted to the fine-grained auditing agents for review again.

[0113] The process from generating the initial implementation steps to reviewing the initial implementation steps is repeated until the fine-grained review agent confirms that the initial implementation steps generated currently by the fine-grained planning agent fully meet the task requirements, that is, the second review result is passed.

[0114] Once the coordination processor learns that the second review result is passed, it will send the initial implementation steps generated by the fine-grained planning agent to the second proxy agent as the final specific implementation steps.

[0115] The coordination processor will notify the second proxy agent to initiate a human-machine interaction with the user (referred to as the second human-machine interaction at this time), display these specific implementation steps to the user, and invite the user to put forward adjustment suggestions.

[0116] At this time, the user can fine-tune the implementation steps or put forward other suggestions according to their own experience and needs. The second proxy agent will feedback the user's adjustment suggestions to the fine-grained planning agent through the coordination processor, so that the fine-grained planning agent can further adjust and optimize the specific implementation steps according to the adjustment suggestions feedback by the user, ensuring that the steps not only meet the objective requirements of the omics task but also satisfy the actual needs of the user.

[0117] The task assistance method provided by the present invention realizes the further refinement and optimization of the task implementation process through the collaborative work of each agent in the fine-grained planning group. The fine-grained planning agent is responsible for generating the initial implementation steps, the fine-grained review agent is responsible for reviewing and providing improvement suggestions, and the second proxy agent is responsible for interacting with the user to obtain the user's feedback and adjustment suggestions. This collaborative work method effectively improves the efficiency and accuracy of the execution of the omics task, and especially can meet the actual needs of the user.

[0118] In omics research, such as genomics, transcriptomics or proteomics, etc., its data analysis tasks usually involve multiple complex and interdependent implementation steps. Each implementation step may require different professional knowledge, computing resources and review criteria. In order to improve the flexibility and efficiency of omics task understanding and splitting, the present invention provides a method for dynamically constructing a fine-grained planning group based on process steps.

[0119] First, the entire omics task is decomposed into a series of ordered process steps through a coarse-grained planning group. For example, in genomic data analysis, these steps may include data reception and preprocessing, sequence alignment, variant detection, functional annotation and result reporting, etc. Each step clearly defines its input, output and required professional skills.

[0120] Then, according to the execution requirements of each process step, the required agents are dynamically screened to construct a coarse-grained planning group. For example, for the process step of data reception and preprocessing, a fine-grained planning group including a data cleaning agent, a format conversion agent, and a preprocessing review agent is constructed. Among them, the data cleaning agent is responsible for removing low-quality data, the format conversion agent ensures that the data format is compatible with subsequent analysis software, and the preprocessing review agent performs quality checks on the preprocessing results.

[0121] For another example, for the process step of sequence alignment, a fine-grained planning group including an alignment algorithm selection agent, an alignment execution agent, and an alignment result review agent is reconstructed. Among them, the alignment algorithm selection agent selects a suitable alignment algorithm according to data characteristics and alignment requirements, the alignment execution agent performs the alignment operation, and the alignment result review agent verifies the accuracy and integrity of the alignment result. Through the collaboration and information transfer mechanism between groups, the continuity of omics task execution and the consistency of data are achieved.

[0122] Optionally, during the execution of an omics task, if different fine-grained planning groups are constructed for each process step, all the fine-grained planning groups can collaborate and transfer information through a coordination processor, including providing a unified communication protocol and data interface.

[0123] The task assistance method provided by the present invention realizes the flexible and efficient execution of omics data analysis tasks by dynamically constructing fine-grained planning groups based on process steps. Each process step constructs a corresponding fine-grained planning group according to specific requirements, ensuring the pertinence and professionalism of task execution.

[0124] As an alternative embodiment, the action execution group at least includes an encoding agent, a third proxy agent, and a debugging agent; Executing the specific implementation step through the action execution group to obtain the target omics result after completing the specific implementation step includes: Generating initial omics code according to the specific implementation step by the encoding agent; Debugging the initial omics code by the debugging agent so that the encoding agent regenerates new initial omics code according to the debugging results until it is determined that the debugging result of the debugging agent for the initial omics code is passed; Obtaining the initial omics code when the debugging result is passed as the target omics code; Executing the target omics code by the third proxy agent to generate the target omics result.

[0125] The task assistance method provided by the present invention provides a way to execute specific implementation steps through an action execution group and obtain the target omics results after completing these steps. Among them, the action execution group includes at least three types of agents, such as an encoding agent, a proxy agent (denoted as the third proxy agent for ease of description), and a debugging agent. These agents work together to execute the specific implementation steps related to each process step and finally generate the target omics results.

[0126] Taking the content shown below Figure 3 as an example, it will be described in detail how the action execution group executes the specific implementation steps and obtains the target omics results after completing the specific implementation steps.

[0127] The encoding agent generates an initial omics code (referred to as the initial omics code) according to the specific implementation steps. Specifically: The encoding agent first imports the scampy library and sets the dataset variable. Then, it reads a dataset named integrated_dataset_copy.h5ad using the sc.read_h5ad('integrated_dataset_copy.h5ad') command. If the key 'SCANVI' exists in this dataset, the encoding agent will generate corresponding code to print the value corresponding to this key.

[0128] The debugging agent debugs the above initial omics code to ensure its correctness and reliability. Specifically: if the detection fails, a debugging result KeyError will be generated, prompting "No 'neighbors' in uns". This KeyError indicates that the neighborhood graph has not been calculated yet, which is a prerequisite for the Leiden clustering algorithm. That is, after the debugging agent captures this error, it will analyze and determine the cause of the error. Then, the debugging result will be fed back to the encoding agent through the coordination processor for the encoding agent to regenerate a new initial omics code according to the debugging result.

[0129] The process of generating the initial omics code and adjusting the initial omics code may be repeated multiple times until the debugging agent determines that the debugging result of the initial omics code passes, that is, the code has no errors and can be executed correctly.

[0130] When the debugging result of the initial omics code by the debugging agent passes, the initial omics code is determined as the target omics code. This means that the initial omics code has been fully verified and tested and can be safely used to generate the target omics results.

[0131] By executing the target omics code, the third proxy agent generates the target omics results. The target omics results can be data analysis reports, gene expression maps, protein interaction networks, etc.

[0132] It should be noted that during the process of the action execution group executing the generation of the target omics code, human-computer interaction will be dynamically executed, that is, during the whole process, not only do the agents within the action execution group communicate and cooperate with each other, but they can also interact with external users or other systems to obtain more information and guidance.

[0133] In addition, the task assistance method provided by the present invention is to orderly analyze all process steps in the task implementation process through the fine-grained planning group and the action execution group. That is, after completing the analysis of the current process step and obtaining its corresponding target omics results, repeat the analysis of the next process step until the analysis of all process steps is completed and the target omics results corresponding to each process step are obtained.

[0134] The task assistance method provided by the present invention realizes the complete process from specific implementation steps to target omics results through the collaborative work of the coding agent, the debugging agent and the third proxy agent, improving the efficiency and accuracy of omics analysis or data processing tasks.

[0135] In the data- and intelligence-intensive scientific research paradigm, both users as human experts and agents are core participants in scientific research work. To support a close human-machine collaborative work mode, the task assistance method provided by the present invention designs a user-friendly and efficient human-computer interaction mechanism.

[0136] Compared with the previous human-computer interaction interface with a fixed and static human-computer interaction mechanism for the opening time, the present invention provides a dynamic human-computer interaction mechanism, that is, when certain trigger conditions are met, a time window for human-computer interaction is automatically opened to provide high-real-time interaction.

[0137] This dynamic human-computer interaction mechanism is widely used in the agents responsible for action execution, which helps to cope with complex and dynamic execution processes. Combining the use of static human-computer interaction in all agent groups and dynamic human-computer interaction in the action execution group, under the condition of ensuring that the multi-agent system can autonomously execute tasks, unnecessary human participation and intervention are minimized.

[0138] The following will specifically describe how to implement static human-computer interaction and how to implement dynamic human-computer interaction in the action execution group. Specifically, the third proxy agent is equipped with a Python kernel and a code execution environment, responsible for executing the target omics code and generating the target omics results. At the same time, the third proxy agent also provides two types of human-computer interaction methods for the action execution group, including static human-computer interaction methods and dynamic human-computer interaction methods.

[0139] As an alternative embodiment, the static human - machine interaction performed by the action execution group is after the coding agent generates the initial omics code according to the specific implementation steps, and specifically includes: Initiating a third human - machine interaction through the third proxy agent; The third human - machine interaction is used to obtain the user's adjustment suggestions for the initial omics code.

[0140] Specifically, after the initial omics code is generated by the coding agent, the third proxy agent will initiate a third human - machine interaction for the user to view and evaluate the initial omics code and at the same time put forward adjustment suggestions. For example, the user can interact with the third proxy agent through a Graphical User Interface (GUI), a Command Line Interface (CLI), or other interaction methods. During the interaction process, the user can view the content, structure, and execution results of the initial omics code and put forward modification opinions according to their own needs or expectations.

[0141] The debugging agent receives the adjustment suggestions put forward by the user through the third human - machine interaction and will perform corresponding debugging and modification on the initial omics code, such as correcting errors in the code, optimizing the code structure, adjusting parameter settings, etc.

[0142] Of course, in the above process, the debugging agent may perform multiple iterations according to the user's adjustment suggestions until an initial omics code that meets the user's needs and has a passed debugging result is determined.

[0143] When the debugging agent determines that the debugging result of the initial omics code is passed, the initial omics code is determined as the target omics code. Then, it continues to be the responsibility of the third proxy agent to execute the target omics code and generate the target omics result.

[0144] As another alternative embodiment, the dynamic human - machine interaction performed by the action execution group may occur after the debugging agent debugs the initial omics code, and specifically includes: If the debugging agent determines that a preset trigger condition is met, then initiate a fourth human - machine interaction through the third proxy agent; The fourth human - machine interaction is used to obtain the user's adjustment suggestions for the debugging result, and the adjustment suggestions are used for the coding agent to readjust the initial omics code.

[0145] Specifically, the debugging agent first receives the initial omics code generated by the encoding agent and debugs it. The debugging process may include checking aspects such as the syntax, logic, and execution efficiency of the code to ensure that the code can run correctly and achieve the expected results.

[0146] During the debugging process, the debugging agent will record the debugging results, including key information such as whether the code runs successfully, the execution time, and resource consumption. These key information will be used as the basis for subsequent judgment on whether to initiate the fourth human-computer interaction.

[0147] After completing the debugging, the debugging agent will judge whether to initiate the fourth human-computer interaction according to the preset trigger conditions. If the debugging agent judges that the preset trigger conditions are met, it will enter the next operation; otherwise, it will continue to execute other tasks or wait for new instructions.

[0148] If the preset trigger conditions are met, the debugging agent will initiate the fourth human-computer interaction through the third proxy agent, aiming to obtain the user's adjustment suggestions for the debugging results in order to further optimize the initial omics code. The fourth human-computer interaction can adopt an interaction method similar to the third human-computer interaction, such as GUI, CLI, or voice interaction, etc.

[0149] During the fourth human-computer interaction, the user will analyze and evaluate the debugging results based on their professional knowledge and experience, and at the same time give adjustment suggestions for the debugging results. These adjustment suggestions may involve aspects such as code modification, parameter adjustment, and algorithm optimization.

[0150] After receiving the user's adjustment suggestions, the encoding agent will make corresponding modifications and optimizations to the initial omics code to ensure the correctness, efficiency, and scalability of the initial omics code to meet the user's needs and expectations. The code after re-adjustment will be verified and tested by the debugging agent again to ensure that its quality and performance meet the preset standards and requirements.

[0151] As an alternative embodiment, the above preset trigger conditions may include, but are not limited to: The debugging agent generates the same code error information when running the initial omics code twice in succession; Or, the number of code error information generated by the debugging agent when running the initial omics code is greater than the preset threshold; Or, a preset sensitive word appears in the code error information generated by the debugging agent when running the initial omics code; Or, the debugging agent lacks necessary files or parameters when running the initial omics code.

[0152] When the debugging agent runs the initial omics code twice in succession, if it detects that the generated code error messages are exactly the same, this may indicate that there are problems in the code that the debugging agent cannot continuously solve. At this time, it is considered that the trigger condition is met, and the fourth human-computer interaction is initiated through the third proxy agent to request the user to provide further adjustment suggestions.

[0153] If the number of code error messages generated by the debugging agent when running the initial omics code exceeds a preset threshold (for example, the number of errors exceeds 5), this indicates that the code may have multiple problems or be unstable. At this time, it can also be considered that the trigger condition is met, and the fourth human-computer interaction process is started to request the user to provide adjustment suggestions for solving this phenomenon based on professional knowledge and experience.

[0154] In some cases, specific error messages may indicate serious code problems or potential security risks. Therefore, if the error messages generated by the debugging agent when running the initial omics code contain preset sensitive words (such as the actual gene names included in the dataset), the fourth human-computer interaction can be immediately triggered so that the user can intervene quickly and solve the problem (if necessary, request the user to manually specify the name in the error).

[0155] If the debugging agent finds that necessary files or parameters are missing when running the initial omics code, this usually means that the code cannot continue to execute or the execution result may be inaccurate. In this case, the debugging agent can also consider that the trigger condition is met and initiate the fourth human-computer interaction through the third proxy agent to prompt the user to check and provide the missing files or parameters.

[0156] Of course, the description of the trigger condition needs to be specifically set according to the actual situation. For example, the trigger condition can also include: the debugging result does not meet the preset quality standard, the code execution time exceeds the threshold, the resource consumption is too large, etc. By setting reasonable preset trigger conditions and combining with a dynamic human-computer interaction mechanism, the efficient debugging and optimization of the initial omics code are realized. This method not only improves the automation degree of code generation but also provides a more convenient and efficient omics analysis experience for users.

[0157] The task assistance method provided by the present invention enhances the cognitive abilities of agents with different roles based on knowledge empowerment by constructing a shared interdisciplinary knowledge base.

[0158] As an alternative embodiment, the task assistance method provided by the present invention further includes: Receiving an information call request from any one of the coarse-grained planning group, the fine-grained planning group, or the action execution group; Responding to the information call request and establishing a connection for any one of the agents to call knowledge from the interdisciplinary knowledge base; The interdisciplinary knowledge base is constructed by extracting knowledge from various disciplines in the environment and encoding it in a form readable by agents.

[0159] During the execution of a task, any agent in the coarse-grained planning group, fine-grained planning group, or action execution group may need to obtain relevant knowledge or information from the interdisciplinary knowledge base according to the requirements of the current task. At this time, the agent will send an information call request to the system.

[0160] After receiving the information call request from any agent, the coordination processor will immediately respond. This response process includes identifying the request source, parsing the request content, and determining the specific knowledge or information in the interdisciplinary knowledge base that needs to be called. Then, according to the parsed request content, a connection for the agent to call knowledge from the interdisciplinary knowledge base will be established. This connection ensures that the agent can efficiently access and obtain the required knowledge or information.

[0161] For example, during coarse-grained planning, the granularity planning agent starts its process by retrieving relevant information from the interdisciplinary knowledge base, and the retrieved data is integrated into the operation framework of the granularity planning agent as supplementary hints, so as to increase the insight of the granularity planning agent based on knowledge and enrich the decision-making process with existing knowledge support, thereby improving its planning quality.

[0162] Among them, the interdisciplinary knowledge base is constructed by extracting knowledge from various disciplines in the environment, such as research papers in multiple disciplinary fields, and encoding it in a form readable by agents. This means that the knowledge or information in the interdisciplinary knowledge base is stored in a format that agents can understand and apply, thereby supporting knowledge calls and decision-making by agents during task execution.

[0163] The task assistance method provided by the present invention takes into account that bioinformatics involves multidisciplinary intersections. Therefore, a shared interdisciplinary knowledge base is constructed, and professional knowledge in different fields is encoded and stored in a form understandable by agents. The interdisciplinary knowledge empowers the agents to extract relevant information from the knowledge base according to task requirements and give reasonable decision-making plans, significantly improving the efficiency and decision-making ability of agents during task execution. At the same time, it promotes the sharing and integration of interdisciplinary knowledge and enhances the flexibility and scalability of the system.

[0164] In order to quickly adapt to new challenging life science research tasks, the task assistance method provided by the present invention designs an agent continuous learning strategy, that is, reflective learning (also called self-learning).

[0165] As an optional embodiment, the task assistance method may further include: Receive a self - learning information invocation request from any agent in the coarse - grained planning group, the fine - grained planning group, or the action execution group; In response to the self - learning information invocation request, send the historical files related to the any agent to the any agent for the any agent to conduct reflective learning based on the historical files; The historical files include human - machine interaction information and / or task interaction logs.

[0166] During task execution or planning, any agent in the coarse - grained planning group, the fine - grained planning group, or the action execution group may generate a need to review and learn from past experiences or historical information according to the current task requirements or its own development needs. At this time, the agent will send a self - learning information invocation request.

[0167] After receiving the self - learning information invocation request from any agent, the coordination processor will immediately respond. This response process includes identifying the request source, parsing the request content, and determining the type and content of the historical files to be sent.

[0168] According to the parsed request content, the historical files related to the requesting agent will be sent to the agent. These historical files include, but are not limited to, human - machine interaction information and / or task interaction logs. Human - machine interaction information records the interaction process between the agent and the user, including the user's instructions, feedback, and the agent's responses, etc.; task interaction logs record various state changes, decision - making processes, and result outputs of the agent during task execution.

[0169] After receiving the historical files, the agent will conduct reflective learning based on these files. By reviewing and analyzing past interaction information and task logs, the agent can discover its own deficiencies and errors in task execution or planning, and then adjust and optimize its own strategies and behaviors.

[0170] For example, the coarse - grained review agent in the coarse - grained planning group, the fine - grained review agent in the fine - grained planning group, the debugging agent in the action execution group, etc. They are responsible for checking the plans, steps, and codes generated by the coarse - grained planning agent, the fine - grained planning agent, and the coding agent. By adopting the reflective learning provided by the present invention and through multiple rounds of interaction, their own capabilities can be effectively optimized, and they can automatically provide modification suggestions to improve the quality of the answers.

[0171] The task - assisting method provided by the present invention enables the agent to conduct reflective learning from its own behaviors, so as to continuously optimize its own decision - making and execution capabilities over time, which helps to improve intensive intelligence.

[0172] Finally, it should be emphasized that the large language model of each agent in the task-assisted method provided by the present invention can adopt large language models of different types and versions, such as the LLaMa series, etc., and the agent can be implemented using various frameworks, such as LangChain, etc.

[0173] In addition, the task-assisted method provided by the present invention can be applied to the construction of a human lung tissue cell atlas, and of course can also be extended to other bioinformatics analysis tasks, such as multi-species, multi-tissue type, and multi-omics data analysis.

[0174] Figure 4 is a schematic diagram of the structure of the task assisting device provided by the present invention, such as Figure 4 As shown, it mainly includes: A task receiving unit 41 is used to obtain the omics task issued by the user; A coarse-grained auxiliary unit 42, used to determine the task implementation process required to complete the omics task through a coarse-grained planning group; The fine-grained auxiliary unit 43 is used to sequentially traverse each process step in the task implementation process and perform the following operations: The specific implementation steps required to complete each of the process steps are determined through a fine-grained planning group; the specific implementation steps are executed through an action execution group to obtain the target omics results after completing the specific implementation steps.

[0175] Among them, the coarse-grained planning group, the fine-grained planning group and the action execution group are groups composed of multiple identical or different intelligent agents.

[0176] It should be noted that the task assistance device provided by the present invention can execute the task assistance method described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0177] The task assisting device provided by the present invention provides a method for conducting omics research by combining a multi-agent system based on a large language model with bioinformatics research, and innovatively designs a way of grouping and managing agent groups, which can reduce information redundancy and meaningless dialogues in interactions between agents, ensure efficient and accurate information flow, and effectively improve the efficiency and accuracy of omics research between the two.

[0178] In order to effectively prove the feasibility and advancement of the auxiliary method and device provided by the present invention, they were evaluated from multiple perspectives, mainly including task planning evaluation, code quality evaluation, and task execution efficiency evaluation. The relevant evaluation methods and corresponding evaluation results are briefly described below.

[0179] To fully evaluate the performance of the task assistance method and its device provided by the present invention, a wide range of evaluation metrics were used. For example, the system constructed using the present invention (hereinafter referred to as this system) was applied to all bioinformatics tasks in the workflow of constructing a large-scale human lung tissue cell atlas, and the execution results of this system were compared with those of the control group and the ablation group.

[0180] Among them, the control group consists of AutoBA and AutoGen. Among them, AutoBA is an influential single AI agent in the field of bioinformatics, and AutoGen is an influential multi-agent system. The ablation group includes variants of this system that exclude knowledge empowerment related to interdisciplinary knowledge bases and exclude reflective learning. By comparing the execution results, it can be determined that this system has achieved state-of-the-art overall performance in performing the above-mentioned omics tasks.

[0181] In addition, the task planning evaluation adopted was used to analyze the ability of this system in decomposing bioinformatics tasks and formulating task implementation processes. The evaluation metrics include clarity, technical feasibility, structure, comprehensiveness, rationality of method selection, risk management, data management, scientific rigor, resource allocation, and scalability. Obviously, this system has achieved state-of-the-art task planning performance in all tasks. Compared with AutoBA and AutoGen of the control group, the task implementation processes generated by this system have higher executability in bioinformatics research tasks. This further confirms that through the close cooperation between multiple agents and users, intensive intelligence helps to design excellent scientific research plans. In addition, the comparison with the ablation group also shows that knowledge empowerment, self-reflective learning, and literature retrieval can improve task planning ability.

[0182] The code quality evaluation assesses the ability of the system to understand the specific implementation steps required for each of the said process steps. The evaluation metrics include accuracy, efficiency, integrity, readability, logic, portability, maintainability, usability, reusability, robustness, security, and scalability. Similarly, this system has achieved state-of-the-art code quality performance in all omics tasks, demonstrating a more comprehensive understanding of the process steps. The comparison with the control group shows that this system can generate higher-quality code for the process steps. The comparison with the ablation group also shows that knowledge empowerment and self-reflective learning are beneficial to the system in code generation.

[0183] The execution efficiency evaluation assesses the step execution ability of this system from aspects such as task integrity, user intervention, running time, and computational resource cost. The results show that this system can complete all omics tasks according to the generated specific implementation steps, while the methods of the control group cannot complete the execution of most specific implementation steps. User intervention represents the frequency of user participation in guiding. This system completes all omics tasks with very little user intervention, highlighting a high level of automation. It is worth noting that AutoBA has no user intervention during the execution process, but its task integrity is low, indicating that during the execution of complex tasks, appropriate user intervention can prevent task interruption or failure due to misunderstanding or insufficient information. AutoGen has the same number of user interventions as the system in all tasks, but their tasks are not completed. This also confirms the effectiveness of the human-computer interaction mechanism designed for this system, enabling this system to handle complex bioinformatics tasks. The comparison with the ablation group shows that knowledge empowerment, self-reflection learning, and literature retrieval can effectively reduce user intervention. In all tasks, the overall running time of the system is shorter than that of AutoGen, but the task integrity is higher. Although the overall running time of AutoBA is less, its task integrity is the lowest. In terms of resource consumption, the computational resource consumption of each system is within an acceptable range. 128 CPU cores, 32 GB of memory, and a GPU with 24 GB of memory can meet the usage requirements of the system.

[0184] In summary, the task assistance method and its device provided by the present invention have been deeply evaluated from three main aspects: task planning, code generation quality, and task execution efficiency. Of course, more evaluation indicators can be adopted to more comprehensively evaluate the performance of the system, such as intelligent grading, etc.

[0185] Figure 5 It is a schematic structural diagram of the electronic device provided by the present invention, as Figure 5 shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540. Among them, the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can execute the task assistance method through the logical instructions in the memory 530. The method includes: obtaining the omics task issued by the user; determining the task implementation process required to complete the omics task through the coarse-grained planning group; traversing each process step in the task implementation process in sequence, and performing the following operations: determining the specific implementation steps required to complete each process step through the fine-grained planning group; executing the specific implementation steps through the action execution group to obtain the target omics result after completing the specific implementation steps.

[0186] In addition, when the logical instructions in the above-mentioned memory 530 can be implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0187] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the task assistance method provided in the above-mentioned various embodiments. The method includes: obtaining an omics task issued by a user; determining a task implementation process required to complete the omics task through a coarse-grained planning group; sequentially traversing each process step in the task implementation process and performing the following operations: determining specific implementation steps required to complete each process step through a fine-grained planning group; and executing the specific implementation steps through an action execution group to obtain a target omics result after completing the specific implementation steps.

[0188] On another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is configured to execute the task assistance method provided in the above-mentioned various embodiments. The method includes: obtaining an omics task issued by a user; determining a task implementation process required to complete the omics task through a coarse-grained planning group; sequentially traversing each process step in the task implementation process and performing the following operations: determining specific implementation steps required to complete each process step through a fine-grained planning group; and executing the specific implementation steps through an action execution group to obtain a target omics result after completing the specific implementation steps.

[0189] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0190] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A task assistance method, characterized in that: include: Obtain the omics tasks assigned by the user; Determine the task implementation process required to complete the omics task through a coarse-grained planning group; Traverse each process step in the task implementation process in turn and perform the following operations: Determine the specific implementation steps required to complete each of the process steps through a fine-grained planning group; execute the specific implementation steps through an action execution group to obtain the target omics results after completing the specific implementation steps; The coarse-grained planning group, the fine-grained planning group and the action execution group are groups composed of multiple identical or different intelligent agents.

2. The task assistance method according to claim 1, characterized in that: The coarse-grained planning group at least includes a coarse-grained planning agent, a first proxy agent and a coarse-grained review agent; The task implementation process required to complete the omics task is determined by the coarse-grained planning group, including: Generate an initial implementation process according to the omics task through the coarse-grained planning agent; Obtaining a first audit result of the coarse-grained audit agent on the initial implementation process; If it is determined that the first audit result includes a process improvement suggestion, the coarse-grained planning agent is controlled to generate a new initial implementation process according to the process improvement suggestion, and the coarse-grained audit agent re-audits the new initial implementation process until the first audit result is determined to be passed; Sending the initial implementation process when the first audit result is passed as the task implementation process to the first agent intelligent body, so as to initiate a first human-computer interaction through the first agent intelligent body; The first human-computer interaction is used to obtain the user's adjustment suggestions for the task implementation process, and the adjustment suggestions are used by the coarse-grained planning agent to readjust the initial implementation process.

3. The task assistance method according to claim 2, characterized in that: The coarse-grained planning group is temporarily constructed based on the omics task, and different omics tasks construct different coarse-grained planning groups.

4. The task assistance method according to claim 1, characterized in that: The fine-grained planning group includes at least a fine-grained planning agent, a second proxy agent and a fine-grained audit agent; The specific implementation steps required to complete each of the process steps are determined by the fine-grained planning group, including: Generate initial implementation steps according to the process steps through the fine-grained planning agent; Obtaining a second audit result of the fine-grained audit agent on the initial implementation step; If it is determined that the second review result includes a step improvement suggestion, the fine-grained planning agent is controlled to generate new initial implementation steps according to the step improvement suggestion, and the new initial implementation steps are reviewed again by the fine-grained review agent until the second review result is determined to be passed; Sending the initial implementation steps when the second audit result is passed as the specific implementation steps to the second agent intelligent body, so that the second agent intelligent body initiates the second human-computer interaction; The second human-computer interaction is used to obtain the user's adjustment suggestions for the specific implementation steps, and the adjustment suggestions are used by the fine-grained planning agent to readjust the initial implementation steps.

5. The task assistance method according to claim 4, characterized in that: The fine-grained planning group is temporarily constructed according to the process steps, and different process steps construct different fine-grained planning groups.

6. The task assistance method according to claim 1, characterized in that: The action execution group includes at least a coding agent, a third agent agent and a debugging agent; The specific implementation steps are executed by the action execution group to obtain the target omics results after completing the specific implementation steps, including: Generate an initial omics code through the coding agent according to the specific implementation steps; The initial omics code is debugged by the debugging agent, so that the encoding agent regenerates a new initial omics code according to the debugging result, until it is determined that the debugging result of the initial omics code by the debugging agent is passed; Acquire an initial omics code when the debugging result is passed as the target omics code; The target omics code is executed by the third agent to generate the target omics result.

7. The task assistance method according to claim 6, characterized in that: After the encoding agent generates the initial omics code according to the specific implementation steps, it also includes: Initiate a third human-computer interaction through the third agent; The third human-computer interaction is used to obtain the user's adjustment suggestion for the initial omics code.

8. The task assistance method according to claim 7, characterized in that: After the initial omics code is debugged by the debugging agent, the method further includes: If the debugging agent determines that the preset triggering condition is met, a fourth human-computer interaction is initiated through the third proxy agent; The fourth human-computer interaction is used to obtain the user's adjustment suggestions for the debugging results, and the adjustment suggestions are used by the coding agent to readjust the initial omics code.

9. The task assistance method according to claim 8, characterized in that: The preset trigger conditions include: The debugging agent generates the same code error information when running the initial omics code twice consecutively; Alternatively, the number of code error messages generated by the debugging agent when running the initial omics code is greater than a preset threshold; Alternatively, a preset sensitive word appears in the code error information generated by the debugging agent when running the initial omics code; Alternatively, the debugging agent lacks necessary files or parameters when running the initial omics code.

10. The task assisting method according to claim 1, characterized in that: Also includes: Receiving an information call request from any agent in the coarse-grained planning group, the fine-grained planning group, or the action execution group; In response to the information call request, establishing a connection for any of the intelligent agents to call knowledge from the interdisciplinary knowledge base; The interdisciplinary knowledge base is constructed by extracting knowledge of various disciplines from the environment and encoding it in a form readable by the intelligent agent.

11. The task assistance method according to claim 1, characterized in that: Also includes: Receiving a self-learning information call request from any agent in the coarse-grained planning group, the fine-grained planning group, or the action execution group; In response to the self-learning information call request, a history file related to any of the intelligent agents is sent to any of the intelligent agents, so that any of the intelligent agents can perform reflective learning based on the history file; The history file includes human-computer interaction information and / or task interaction log.

12. The task assistance method according to claim 1, characterized in that: After obtaining the target omics results after completing the specific implementation steps, it also includes: Object storage is performed on the target omics results corresponding to each of the specific implementation steps.

13. A task assisting device, characterized in that: include: A task receiving unit, used to obtain the omics tasks issued by the user; A coarse-grained auxiliary unit, used to determine the task implementation process required to complete the omics task through a coarse-grained planning group; The fine-grained auxiliary unit is used to sequentially traverse each process step in the task implementation process and perform the following operations: Determine the specific implementation steps required to complete each of the process steps through a fine-grained planning group; execute the specific implementation steps through an action execution group to obtain the target omics results after completing the specific implementation steps; Among them, the coarse-grained planning group, the fine-grained planning group and the action execution group are groups composed of multiple identical or different intelligent agents.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the task assistance method according to any one of claims 1 to 12 is implemented.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the task assistance method according to any one of claims 1 to 12 is implemented.

16. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the task assistance method according to any one of claims 1 to 12 is implemented.

Images