Mrna vaccine sequence design system and method based on large language model intelligent agent

Abstract

The application discloses an mRNA vaccine sequence design system and method based on a large language model intelligent agent, and the mRNA vaccine sequence design system based on the large language model intelligent agent comprises a user interaction and task analysis module, a task planning and intelligent agent scheduling module, an RNA design skill library and skill description module, a multi-target evaluation and closed-loop optimization module and a data and model management module. The application integrates a large language model and various mRNA sequence analysis and optimization tools, automatically completes mRNA vaccine sequence generation, evaluation and optimization for a target antigen in a computer environment, is suitable for candidate sequence design and screening of a preventive or therapeutic mRNA vaccine, improves design efficiency, and reduces the degree of dependence on manual experience and manual script arrangement.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an mRNA vaccine sequence design system based on a large language model intelligent agent, a method for designing mRNA vaccine sequences based on a large language model intelligent agent, and an electronic device. Background Technology

[0002] The development of mRNA vaccines targeting a given pathogen or tumor typically involves multiple stages, including antigen screening and design, T-cell and B-cell epitope prediction, mRNA sequence engineering optimization (including coding and non-coding regions), delivery system design, and in vitro and in vivo validation. Among these, the computational design and optimization of the mRNA sequence itself is a crucial step in determining expression efficiency, stability, and safety, and has a significant impact on the final efficacy of the vaccine.

[0003] In the sequence design phase of mRNA vaccines, the relevant technical approach typically starts from the target protein or peptide sequence and follows a computational design strategy as follows: First, potential protective epitopes are screened from pathogen proteins or tumor antigens using immunoinformatics and structural biology methods. MHC I / II binding prediction tools (such as NetMHCpan, NetMHCIIpan, and various algorithms on the IEDB platform) are used to predict the binding affinity of candidate peptides to multiple HLA alleles to assess the potential of T-cell immune responses. Next, while maintaining the protein amino acid sequence, codon optimization is performed on the mRNA coding region to improve translation efficiency and protein expression levels. Commonly used indicators include the Codon Adaptation Index (CAI), GC content distribution, codon pair patterns, and avoiding certain unfavorable dinucleotide frequencies. Some studies employ deep generative models or specialized optimization algorithms to search for and optimize codon combinations. Then, local and global structural stability is estimated using RNA secondary structure prediction tools (such as RNAfold in the ViennaRNA suite). For example, algorithms such as LinearDesign, proposed in recent years, treat mRNA sequence design as an optimization problem that balances structural stability and codon usage in a huge combinatorial space, and use dynamic programming and other methods to significantly improve the stability and expression performance of full-length mRNA. Finally, the 5′UTR, 3′UTR, poly(A) tail, and end cap structures are engineered to improve ribosome loading, translation initiation, and mRNA half-life; immunogenicity, stability, and expression levels can be balanced by introducing chemical modifications such as N1-methylpseuuridine and optimizing local structures. However, this related technical approach often leads to a longer design cycle, and it is difficult to efficiently balance multiple objectives (such as TE>80% and off-target<5%).

[0004] Meanwhile, to support the aforementioned design approach, various computational tools and integrated platforms have been developed in related technologies. These include tools for structure and stability optimization, epitope and immunogenicity prediction, and mRNA vaccine design workflows and web platforms. These tools and platforms have improved the computational efficiency of mRNA vaccine design to some extent, enabling researchers to quickly obtain candidate mRNA sequences and conduct preliminary screening in terms of expression, stability, and immunogenicity.

[0005] Although the tools and platforms in related technologies have improved the computational efficiency of mRNA vaccine design to some extent, enabling researchers to quickly obtain candidate mRNA sequences and conduct preliminary screening, the following limitations still exist: 1. Fragmented tools and heterogeneous interfaces: mRNA vaccine design involves various computational tools and models, such as epitope prediction, codon optimization, secondary structure prediction, stability assessment, and immunogenicity and safety assessment. These tools mostly exist in the form of independent web services, command-line software, or script libraries, with different interface formats and input / output data structures. Researchers need to manually write scripts for interface and data conversion, which is difficult to use and prone to errors. Manual intervention usually accounts for more than 50%; 2. Fixed workflow and lack of flexible scheduling: The design process in related technologies mostly adopts a fixed "linear workflow," and the tool combination and calling order are pre-set during the platform development stage. Faced with the personalized needs of different pathogens, antigens, or indications (e.g., emphasizing immunogenicity, safety, or expression efficiency), the systems in related technologies are difficult to flexibly adjust. This often requires manual redesign of workflows and script modifications, resulting in a large workload and slow iteration. 3. Insufficient multi-objective optimization capabilities and imperfect closed-loop design: mRNA vaccine sequences require trade-offs among multiple indicators such as expression efficiency, structural stability, immunogenicity, safety, and manufacturability. Related technologies often optimize only one or a few indicators, lacking a unified multi-objective evaluation and optimization mechanism for mRNA vaccines. This makes it difficult to systematically iteratively design and screen candidate sequences within a unified framework, leading to conflicts between TE and off-target and hindering Pareto optimality. 4. Lack of automated design platforms centered on intelligent agents: Current mRNA vaccine design processes still heavily rely on bioinformatics experts manually selecting tools, writing scripts, and adjusting parameters, failing to fully utilize the capabilities of large language models in task understanding, process planning, and tool scheduling. Summary of the Invention

[0006] The present invention is proposed to address at least one of the aforementioned problems. According to a first aspect of the invention, an mRNA vaccine sequence design system based on a large language model intelligent agent is provided. The system includes: a user interaction and task parsing module, a task planning and intelligent agent scheduling module, an RNA design skill library and skill description module, a multi-objective evaluation and closed-loop optimization module, and a data and model management module.

[0007] The user interaction and task parsing module is used to receive vaccine design parameters and parse and transform the vaccine design parameters through a large language model to obtain a task description object.

[0008] The task planning and intelligent agent scheduling module is connected to the user interaction and task parsing module and the RNA design skill library and skill description module. It is used to construct a DAG workflow based on the task description object and the skills in the RNA design skill library and skill description module, call the skills in the RNA design skill library and skill description module through intelligent agents to generate candidate mRNA sequences, and dynamically adjust the skills selected from the RNA design skill library and skill description module according to the comprehensive scoring results during the execution of the DAG workflow.

[0009] The RNA design skill library and skill description module are connected to the multi-objective evaluation and closed-loop optimization module and the data and model management module. They are used to encapsulate the mRNA molecule design and multi-dimensional characteristic evaluation functions to obtain multiple skill modules, and to uniformly define, register and manage each of the skill modules.

[0010] The multi-objective evaluation and closed-loop optimization module is connected to the task planning and intelligent agent scheduling module. It is used to perform multi-dimensional index analysis and comprehensive evaluation on the candidate mRNA sequence, obtain the comprehensive score result, and feed the comprehensive score result back to the task planning and intelligent agent scheduling module.

[0011] The data and model management module is connected to the multi-objective evaluation and closed-loop optimization module, and is used to manage and store system data and model parameters and tool configurations related to each RNA design skill.

[0012] The system data includes the task description object and the candidate mRNA sequence.

[0013] In one embodiment of the present invention, the user interaction and task parsing module includes:

[0014] The input receiving submodule is used to receive vaccine design parameters input via natural language.

[0015] The task semantic parsing submodule, connected to the input receiving submodule, is used to perform semantic analysis on the vaccine design parameters through a large language model to obtain structured parsing results.

[0016] The task configuration construction submodule is connected to the task semantic parsing submodule and is used to convert the parsing result into the task description object.

[0017] The vaccine design parameters include target antigen information, indication scenarios, and design constraints.

[0018] In one embodiment of the present invention, the task planning and intelligent agent scheduling module includes:

[0019] The task planning submodule is connected to the user interaction and task parsing module and the RNA design skill library and skill description module. It is used to query relevant skill sets from the RNA design skill library and skill description module according to the task description object, and construct the DAG workflow by combining the dependencies between skills, input-output compatibility and resource constraints.

[0020] The intelligent agent decision-making submodule, connected to the task planning submodule, is used to analyze intermediate results through a large language model during the execution of the DAG workflow, obtain evaluation results, and dynamically adjust the skill call order, parameter configuration, and optimization strategy based on the evaluation results.

[0021] The scheduling and execution coordination submodule is connected to the RNA design skill library and skill description module and the multi-objective evaluation and closed-loop optimization module. It is used to issue invocation tasks, convert the data format of the input data of skill invocation, and send the skill invocation results and execution logs to the multi-objective evaluation and closed-loop optimization module.

[0022] The invocation task refers to invoking the tool interface in the RNA design skill library and skill description module.

[0023] In one embodiment of the present invention, the RNA design skill library and skill description module include:

[0024] The skill description submodule is used to define a structured skill description object for each skill module in the skill library.

[0025] The skill registration and discovery submodule, connected to the skill description submodule, is used to provide a skill registration interface, skill retrieval and skill filtering, and to verify the skill description object during the skill registration process.

[0026] The skill call interface submodule is connected to the multi-objective evaluation and closed-loop optimization module and the data and model management module. It is used to encapsulate a unified call interface, provide a standardized call method, encapsulate the skill call result into a unified result object, and send the result object to the multi-objective evaluation and closed-loop optimization module and the data and model management module.

[0027] The structured skill description object includes at least three of the following: skill identifier and version information, applicable object type, input parameter definition, output result definition, dependency relationship, performance and resource information, and invocation method.

[0028] In one embodiment of the present invention, the multi-objective evaluation and closed-loop optimization module includes:

[0029] The candidate sequence management submodule is used to manage the candidate mRNA sequences generated in each iteration round, assign a unique identifier to the candidate mRNA sequence, and maintain a set of candidate sequence states.

[0030] The multi-index evaluation submodule, together with the candidate sequence management submodule, the task planning and intelligent agent scheduling module, and the data and model management module, is used to call the evaluation skills in the RNA design skill library to perform multi-dimensional index analysis on the candidate mRNA sequences, obtain multiple typical indicators, and convert the typical indicators corresponding to each candidate mRNA sequence into EvaluationRecord objects in a standardized form and send them to the data and model management module.

[0031] The multi-objective comprehensive scoring submodule, connected to the multi-indicator evaluation submodule, is used to map multiple typical indicators into a comprehensive scoring result based on task weights and constraints and through a preset scoring strategy.

[0032] The closed-loop optimization control submodule, connected to the multi-objective comprehensive scoring submodule and the task planning and intelligent agent scheduling module, is used to screen the candidate mRNA sequences based on the comprehensive scoring results, construct feedback summary information for the screened mRNA sequences, and send the feedback summary information to the task planning and intelligent agent scheduling module. It is also used to control the termination conditions of the iteration process.

[0033] The candidate sequence state set includes a candidate set, a current primary selection set, and a historical elimination set. The termination conditions include reaching a preset scoring threshold, an upper limit on the number of iteration rounds, and insufficient improvement.

[0034] In one embodiment of the present invention, the data and model management module includes:

[0035] The Task and Configuration Management submodule is used to store the task description objects corresponding to each design task and record the time points of task creation, update and execution.

[0036] The sequence and evaluation result storage submodule is used to store the candidate mRNA sequences and their corresponding EvaluationRecord objects, and supports the retrieval of candidate mRNA sequences by screening conditions, wherein the screening conditions include: task ID, round number, and score range.

[0037] The Skills and Model Version Management submodule is used to manage the registration and version information of RNA design skills, record the historical changes of each RNA design skill, and manage the versions and parameters of each model within the system.

[0038] The logging and auditing submodule is used to record log information and support auditing and quality analysis of the design process.

[0039] In one embodiment of the present invention, the system further includes an experimental feedback integration module.

[0040] The experimental feedback integration module, connected to the multi-objective evaluation and closed-loop optimization module and the task planning and intelligent agent scheduling module, includes: an experimental data receiving submodule, a data standardization and mapping submodule, and a feedback interface submodule.

[0041] The experimental data receiving submodule is used to receive in vitro or in vivo experimental data of the mRNA vaccine candidate sequence.

[0042] The data standardization and mapping submodule is connected to the experimental data receiving submodule and is used to convert the experimental data into a unified data structure and map the data structure into experimental feedback parameters for updating the multi-objective evaluation strategy.

[0043] The feedback interface submodule is connected to the data standardization and mapping submodule, the multi-objective evaluation and closed-loop optimization module, and the task planning and intelligent agent scheduling module, and is used to send the experimental feedback parameters to the multi-objective evaluation and closed-loop optimization module and the task planning and intelligent agent scheduling module.

[0044] The experimental data include: expression levels, immune response strength, and safety indicators.

[0045] In one embodiment of the present invention, the system further includes a code generation and execution module.

[0046] The code generation and execution module includes: a code template and generation submodule, a secure execution environment submodule, and a result parsing and error handling submodule.

[0047] The code template and generation submodule are used to maintain multiple script templates related to RNA design skills, and generate or complete script code by the large language model according to the calling method and parameter definition in the skill description object.

[0048] The secure execution environment submodule is used to run the script generated by the code generation submodule in an isolated execution environment, and to record the standard output, error messages and return values ​​during the script execution process.

[0049] The result parsing and error handling submodule is used to parse the result file or standard output generated by the script execution, and also to organize the error information into prompts when errors or anomalies are detected, so that the large language model can provide correction suggestions or automatically generate a repaired script.

[0050] According to a second aspect of the present invention, a method for designing mRNA vaccine sequences based on a large language model intelligent agent is provided, applied to the aforementioned mRNA vaccine sequence design system based on a large language model intelligent agent, the method comprising:

[0051] The vaccine design parameters are obtained and sent to the user interaction and task parsing module. The user interaction and task parsing module parses and transforms the vaccine design parameters through a large language model to obtain a task description object.

[0052] The task planning and intelligent agent scheduling module constructs a DAG workflow based on the task description object and the skills in the RNA design skill library and skill description module, and generates candidate mRNA sequences by calling the skills in the RNA design skill library and skill description module through an intelligent agent.

[0053] The multi-objective evaluation and closed-loop optimization module performs multi-dimensional index analysis and comprehensive evaluation on candidate mRNA sequences to obtain a comprehensive score result, and feeds the comprehensive score result back to the task planning and intelligent agent scheduling module.

[0054] Determine whether the candidate mRNA sequence meets the preset design constraints or convergence conditions.

[0055] If the conditions are met, the current candidate mRNA sequence is output.

[0056] If the conditions are not met, the intelligent agent modifies, replaces, or regenerates the candidate mRNA sequence based on the comprehensive scoring result to form a new round of candidate mRNA sequences, and returns to the multi-objective evaluation and closed-loop optimization module to perform multi-dimensional index analysis and comprehensive evaluation of the candidate mRNA sequences.

[0057] During the execution of the DAG workflow, the task planning and intelligent agent scheduling module dynamically adjusts the skills selected from the RNA design skill library and skill description module based on the comprehensive scoring results.

[0058] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein when the computer program is executed by the processor, it implements any of the above-described mRNA vaccine sequence design methods based on a large language model intelligent agent.

[0059] The mRNA vaccine sequence design system and method based on a large language model intelligent agent provided by the present invention have the following beneficial effects:

[0060] 1. By integrating a large language model with various mRNA sequence analysis and optimization tools, the generation, evaluation, and optimization of mRNA vaccine sequences targeting antigens can be automatically completed in a computer environment. This is applicable to the design and screening of candidate sequences for preventive or therapeutic mRNA vaccines, thereby improving design efficiency and reducing reliance on human experience and manual script arrangement.

[0061] 2. By using a large language model intelligent agent as the central processing unit, the design task can be automatically understood and decomposed based on the vaccine design parameters provided by the user. In addition, this invention can replace a large amount of manual work in tool combination, parameter adjustment and intermediate result processing, reduce the reliance on bioinformatics engineers to write glue code, shorten the design cycle and reduce the probability of human error.

[0062] 3. By introducing an RNA design skill library and skill description module, this invention provides a unified structured description of various heterogeneous tools, shielding the differences in underlying implementation, reducing the difficulty of tool integration and maintenance, providing a foundation for subsequent automatic scheduling and expansion of new tools, and supporting on-demand dynamic discovery and combination, thereby realizing the unified abstraction and management of heterogeneous mRNA design tools.

[0063] 4. The system of the present invention has a multi-objective closed-loop optimization mechanism for mRNA vaccine design. Under the premise of comprehensively considering expression efficiency, stability, immunogenicity, safety and manufacturability, it helps to improve the overall quality of the final candidate sequence by iteratively generating and evaluating candidate mRNA sequences, and realizes an automated in silico optimization process.

[0064] 5. This invention utilizes a large language model intelligent agent combined with skill modules, and automatically selects skills, determines the calling order and parameter configuration according to the design goals and constraints of specific mRNA vaccine projects. This breaks through the limitations of traditional fixed linear workflows, improves the flexibility and customizability of the mRNA design process, and thus supports dynamic orchestration of workflows according to task requirements, rather than simply executing fixed linear workflows. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of the structure of an mRNA vaccine sequence design system based on a large language model intelligent agent according to an embodiment of the present invention;

[0067] Figure 2 This is a flowchart illustrating an embodiment of the mRNA vaccine sequence design method based on a large language model intelligent agent provided by the present invention.

[0068] Figure 3 A flowchart illustrating another embodiment of the present invention for a method of designing mRNA vaccine sequences based on a large language model intelligent agent;

[0069] Figure 4 This is a schematic diagram illustrating the technical effect verification of an embodiment of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the present invention more apparent, exemplary embodiments according to the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein. Based on the embodiments of the present invention described herein, all other embodiments obtained by those skilled in the art without inventive effort should fall within the protection scope of the present invention.

[0071] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0072] It should be understood that the invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0073] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0074] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0075] To improve the automation, flexibility, and multi-objective optimization capabilities of the mRNA vaccine candidate sequence design process, the first aspect of this invention provides an mRNA vaccine sequence design system based on a large language model intelligent agent, such as... Figure 1 As shown, the mRNA vaccine sequence design system based on a large language model intelligent agent includes: a user interaction and task parsing module 10, a task planning and intelligent agent scheduling module 20, an RNA design skill library and skill description module 30, a multi-objective evaluation and closed-loop optimization module 40, and a data and model management module 50.

[0076] It should be noted that the above modules can be deployed on the same server or on different computing nodes as needed, communicating through an internal network or cloud service. The large language model can be implemented through local deployment or by calling external model inference services, and its output and input are managed by the task planning and intelligent agent scheduling module 20.

[0077] The user interaction and task parsing module 10 is used to receive vaccine design parameters and parse and transform the vaccine design parameters through a large language model (LLM) to obtain a task description object.

[0078] It should be noted that mRNA vaccines include, but are not limited to, mRNA sequences encoding proteins of viruses (such as SARS-CoV-2), bacteria, tumor-associated antigens, and other pathogens or pathologically related proteins.

[0079] As an example, large language models can analyze vaccine design parameters and extract standardized task requirements and constraints, such as extracting a translation efficiency (TE) threshold >80% and a list of human leukocyte antigen (HLA) alleles.

[0080] The task planning and intelligent agent scheduling module 20 is connected to the user interaction and task parsing module 10 and the RNA design skill library and skill description module 30. It is used to construct a DAG (Directed Acyclic Graph) workflow based on the task description object and the skills in the RNA design skill library and skill description module 30. It calls the skills in the RNA design skill library and skill description module 30 through intelligent agents to generate candidate mRNA sequences. During the execution of the DAG workflow, it dynamically adjusts the skills selected from the RNA design skill library and skill description module 30 based on the comprehensive scoring results.

[0081] As an example, during execution, the intelligent agent can also dynamically adjust the skill invocation order, parameter configuration, and subsequent optimization strategies based on intermediate results.

[0082] It is worth mentioning that by introducing intelligent agents, multiple dimensions of indicators such as codon usage, secondary structure, translation efficiency, immunogenicity, safety and manufacturability of mRNA sequences can be jointly optimized.

[0083] The RNA design skill library and skill description module 30 is connected to the multi-objective evaluation and closed-loop optimization module 40 and the data and model management module 50. This is used to encapsulate the functions of mRNA molecule design and multi-dimensional characteristic evaluation, resulting in multiple skill modules. Each skill module is then uniformly defined, registered, and managed.

[0084] It should be noted that the mRNA molecule design and multi-dimensional characteristic evaluation functions include: epitope prediction, codon optimization, mRNA structure prediction and stability evaluation, and immunogenicity and safety evaluation.

[0085] As an example, each skill module records its supported RNA types, input and output data types, applicable indicator dimensions, prerequisite dependencies, resource consumption information, and invocation methods through a structured skill description, achieving unified abstraction and management of heterogeneous tools. For instance, JSON (JavaScript Object Notation) metadata describes UTR (Untranslated Region) engineering constraints.

[0086] The multi-objective evaluation and closed-loop optimization module 40 is connected to the task planning and intelligent agent scheduling module 20. It is used to perform multi-dimensional index analysis and comprehensive evaluation of candidate mRNA sequences, obtain comprehensive scoring results, and feed back the comprehensive scoring results to the task planning and intelligent agent scheduling module 20.

[0087] As an example, multidimensional metrics include expression efficiency, structural stability, immunogenicity, safety, and manufacturability.

[0088] Specifically, the multi-objective evaluation and closed-loop optimization module 40 can call multiple skill modules to evaluate the candidate mRNA sequences generated by the intelligent agent, obtaining multi-dimensional indicators. Based on the multi-dimensional indicators and a preset or adaptive multi-objective evaluation strategy, the candidate sequences are comprehensively scored, ranked, and screened to obtain a comprehensive score result. The comprehensive score result is then fed back to the intelligent agent to guide the next round of sequence generation, local modification, or parameter adjustment, forming a closed-loop optimization process of "generation—evaluation—regeneration". For example, 3-5 rounds of Pareto iterations.

[0089] The data and model management module 50 is connected to the multi-objective evaluation and closed-loop optimization module 40 to manage and store system data and model parameters and tool configurations related to various RNA design skills.

[0090] As an example, the data and model management module 50 can also manage and store target antigen information, design task configurations, etc., thereby enabling the design process to be traceable and reproducible.

[0091] The system data includes: task description objects and candidate mRNA sequences.

[0092] It is worth mentioning that the mRNA vaccine sequence design system based on a large language model intelligent agent of the present invention takes a large language model intelligent agent as the core control unit, and abstracts and encapsulates various computational tools related to mRNA vaccine design into skill modules with unified structured descriptions. Through technical means such as task parsing, skill scheduling, multi-objective evaluation and closed-loop optimization, it automatically completes the generation, analysis and optimization of mRNA vaccine candidate sequences in a computer environment.

[0093] The mRNA vaccine sequence design system based on a large language model intelligent agent, as described in this invention, has the following advantages: 1. By integrating a large language model with various mRNA sequence analysis and optimization tools, it automatically generates, evaluates, and optimizes mRNA vaccine sequences targeting specific antigens within a computer environment. This system is suitable for candidate sequence design and screening of preventative or therapeutic mRNA vaccines, thereby improving design efficiency and reducing reliance on human experience and manual scripting. 2. By employing a large language model intelligent agent as the central processing unit, it can automatically understand and decompose design tasks based on user-provided vaccine design parameters. Furthermore, this invention can replace a significant amount of manual work in tool combination, parameter adjustment, and intermediate result processing, reducing reliance on bioinformatics engineers writing glue code, shortening the design cycle, and lowering the probability of human error. 3. By introducing an RNA design skill library and skill description module, this invention provides a unified structural description of various heterogeneous tools, shielding them from underlying implementation differences, reducing the difficulty of tool integration and maintenance, providing a foundation for subsequent automatic scheduling and expansion of new tools, and supporting on-demand dynamic discovery and combination, achieving unified abstraction and management of heterogeneous mRNA design tools. 4. The system of this invention features a multi-objective closed-loop optimization mechanism for mRNA vaccine design. By comprehensively considering expression efficiency, stability, immunogenicity, safety, and manufacturability, it iteratively generates and evaluates candidate mRNA sequences, helping to improve the overall quality of the final candidate sequences and achieving an automated in-silico optimization process. 5. This invention utilizes a large language model intelligent agent combined with skill modules, automatically selecting skills, determining the calling order, and configuring parameters according to the design goals and constraints of a specific mRNA vaccine project. This breaks through the limitations of traditional fixed linear workflows, improving the flexibility and customizability of the mRNA design process, thereby supporting dynamic workflow orchestration based on task requirements, rather than simply executing a fixed linear workflow.

[0094] It should be noted that the mRNA vaccine sequence design system based on a large language model intelligent agent proposed in this invention can be understood as a comprehensive software and hardware system deployed on a server, cloud platform, or local high-performance computing device. This system uses a large language model as its core control unit, encapsulates multiple computational tools related to mRNA vaccine design into skill modules, and completes the automated design and optimization of mRNA vaccine candidate sequences within a unified framework through skill description and scheduling mechanisms, multi-objective evaluation, and closed-loop optimization mechanisms.

[0095] In some embodiments, the user interaction and task parsing module 10 includes: an input receiving submodule, a task semantic parsing submodule, and a task configuration construction submodule.

[0096] The input receiving submodule is used to receive vaccine design parameters input via natural language.

[0097] The vaccine design parameters include target antigen information, indication scenarios, and design constraints.

[0098] As an example, the input receiving submodule is used to receive target antigen information submitted by the user. The target antigen information includes, but is not limited to: the amino acid sequence of the target protein, one or more pre-screened epitope fragments, and supplementary descriptions of existing vaccine construction protocols (e.g., whether adjuvant sequences, linker peptide sequences, etc. are included). It is also used to receive the user's requirements and constraints for mRNA vaccine design, such as target indications (preventive / therapeutic), priority optimization indicators (expression efficiency, safety, etc.), allowed mRNA length range, chemical modification preferences, etc.

[0099] The task semantic parsing submodule is connected to the input receiving submodule and is used to perform semantic analysis on vaccine design parameters through a large language model to obtain structured parsing results.

[0100] As an example, the task semantic parsing submodule performs semantic analysis on the user's natural language description based on a large language model or a dedicated natural language processing component, parsing unstructured text into structured parsing results.

[0101] The analysis results may include: standardized target antigen identifiers (e.g., protein name, database number), task type (single antigen / multi-antigen, multi-epitope fusion construction, etc.), priority optimization indicators and their weights or thresholds, specific design strategies that are required or prohibited (e.g., whether to enable a certain type of UTR element, whether to allow a certain codon), etc.

[0102] The task configuration building submodule is connected to the task semantic parsing submodule and is used to transform the parsing results into task description objects.

[0103] Preferably, the task description object can be a TaskProfile structure, which includes: basic information, such as task ID (Identifier), creation time, user ID; antigen information, such as protein sequence, epitope list, etc.; design constraints, such as coding region length, UTR selection range, GC (Guanine-Cytosine content) interval, etc.; optimization objectives, such as a list of multi-objective indicators and their priorities; scheduling preferences, such as whether there is a preference for using certain specific types of skill modules.

[0104] As an example, the task configuration building submodule is also used to store task description objects in the data and model management module 50 and pass them to the task planning and intelligent agent scheduling module 20 through an internal interface.

[0105] It should be noted that when the vaccine design parameters input by the user via natural language are incomplete or ambiguous, the user interaction and task parsing module 10 can also complete and standardize the incomplete or ambiguous user input according to a preset template. For example, when the user does not explicitly specify certain parameters, the system's default mRNA vaccine design strategy is adopted.

[0106] In this embodiment, vaccine design parameters are received through natural language interaction and transformed into structured task configurations through semantic parsing. This not only lowers the user's operational threshold but also ensures the standardization and completeness of the input information, effectively improving the efficiency and reliability of the mRNA vaccine design process.

[0107] In some embodiments, the task planning and intelligent agent scheduling module 20 includes: a task planning submodule, an intelligent agent decision-making submodule, and a scheduling and execution coordination submodule.

[0108] It should be noted that the task planning and intelligent agent scheduling module 20 is the core control module of the system of the present invention. With the support of the large language model, it can generate the mRNA vaccine sequence design process according to the task description object and make dynamic decisions and schedules during the execution process.

[0109] The task planning submodule is connected to the user interaction and task parsing module 10 and the RNA design skill library and skill description module 30. It is used to query relevant skill sets from the RNA design skill library and skill description module 30 based on the task description object, and construct a DAG workflow by combining the dependencies between skills, input-output compatibility and resource constraints.

[0110] Specifically, the task planning submodule can query the set of skills related to the current task from the RNA design skill library and skill description module 30 based on the requirements and constraints in the TaskProfile. Then, combining the dependencies between skills, input-output compatibility, and resource constraints, it constructs a candidate workflow structure, such as a directed acyclic graph with nodes and directed edges, where each node corresponds to one skill call or a set of skill calls. Workflow nodes may include: initial mRNA sequence generation, codon optimization, UTR combinatorial evaluation, secondary structure prediction, stability scoring, immunogenicity and safety assessment, etc.

[0111] As an example, the task planning submodule can call the large language model to perform comprehensive reasoning on information such as "task description + skill list + historical success cases" to output an initial workflow draft, which is then validated and corrected by the rule engine.

[0112] The intelligent agent decision-making submodule is connected to the task planning submodule. During the execution of the DAG workflow, it analyzes the intermediate results through a large language model to obtain evaluation results, and dynamically adjusts the skill call order, parameter configuration, and optimization strategy based on the evaluation results.

[0113] It should be noted that during the execution of the DAG workflow, the large language model, acting as an intelligent agent, can analyze intermediate results and determine the next action based on the evaluation results. For example, whether it is necessary to add a certain type of evaluation skill (e.g., add additional stability analysis when structural problems are found); whether it is necessary to adjust the parameters of a certain skill (e.g., increase the weight of a certain type of penalty factor); and whether it is necessary to terminate the current iteration and output the results.

[0114] As an example, the intelligent agent decision-making submodule can periodically obtain the comprehensive score results output by the multi-objective evaluation and closed-loop optimization module 40, encapsulate them into prompt information, and input them into the large language model to generate the next scheduling strategy.

[0115] The scheduling and execution coordination submodule is connected to the RNA design skill library and skill description module 30 and the multi-objective evaluation and closed-loop optimization module 40. It is used to issue call tasks, convert the data format of the input data of skill calls, and send the skill call results and execution logs to the multi-objective evaluation and closed-loop optimization module 40.

[0116] The task of invoking is to call the tool interface in the RNA design skill library and skill description module 30 or to call the script generated by the code generation and execution module.

[0117] It should be noted that preparing and converting the input data for each skill call ensures that it conforms to the format required in the skill description.

[0118] As an example, the scheduling and execution coordination submodule is also used to manage task queues and concurrent execution strategies, processing multiple candidate sequences or multiple skill calls in parallel when resources allow.

[0119] In this embodiment, through the synergistic effect between the task planning submodule, the intelligent agent decision-making submodule, and the scheduling and execution coordination submodule, the task planning and intelligent agent scheduling module 20 realizes the automatic conversion from high-level task objectives to specific tool call sequences, and maintains dynamically adjustable decision-making capabilities during execution.

[0120] In some embodiments, the RNA design skill library and skill description module 30 includes: a skill description submodule, a skill registration and discovery submodule, and a skill invocation interface submodule.

[0121] It should be noted that the RNA design skill library and skill description module 30 can be used to uniformly abstract, register and manage various computational tools related to mRNA vaccine design, which is the basis for realizing the "skill-based scheduling" mechanism of this invention.

[0122] The skill description submodule is used to define a structured skill description object for each skill module in the skill library.

[0123] The structured skill description object includes at least three of the following: skill identifier and version information, applicable object type, input parameter definition, output result definition, dependency relationship, performance and resource information, and invocation method.

[0124] Specifically, the skill description submodule can define a structured skill description object for each skill module. Example fields may include: skill identifier and version information: skill_id, version; applicable object types: such as "mRNA coding region optimization," "5′UTR sequence screening," "mRNA secondary structure prediction," "T cell epitope prediction," etc.; input parameter definitions: parameter name, type (sequence, number, enumeration, file path, etc.), constraints (length range, value range, etc.); output result definitions: result field names, types, and meanings (such as ΔG value, structural entropy, immunogenicity score, etc.); dependencies: a list of prerequisite skills or prerequisite data conditions (e.g., obtaining a certain type of structure prediction result must be obtained first); performance and resource information: estimated runtime, whether a GPU (Graphics Processing Unit) is required, whether it depends on an external Web API (Web Application Programming Interface), etc.; invocation method: invocation URL (Uniform Resource Locator), command line template, script entry function name, etc.

[0125] As an example, the skill description submodule can also be used to store skill descriptions in a uniform data format, such as JSON or a structured database table, so that the task planning and intelligent agent scheduling module 20 can parse and match them.

[0126] The Skill Registration and Discovery submodule is connected to the Skill Description submodule. It provides a skill registration interface, skill retrieval and skill filtering, and verifies the skill description object during the skill registration process.

[0127] As an example, the skill registration and discovery submodule provides a skill registration interface, allowing system administrators or developers to encapsulate new mRNA design tools as skills and add them to the skill library; during the registration process, the completeness and legality of the skill description are verified to avoid inconsistent parameters or unclear output; and skill retrieval and filtering functions are provided, supporting the retrieval of available skill lists based on conditions such as task objectives, applicable objects, input / output types, and runtime constraints.

[0128] Accordingly, the task planning submodule can obtain a set of skills that meet the specific task requirements through the query interface, and use the metadata in the skill descriptions to guide the construction of the workflow.

[0129] The skill call interface submodule is connected to the multi-objective evaluation and closed-loop optimization module 40 and the data and model management module 50. It is used to encapsulate a unified call interface, provide standardized call methods, encapsulate the skill call results into a unified result object, and send the result object to the multi-objective evaluation and closed-loop optimization module 40 and the data and model management module 50.

[0130] Specifically, the skill invocation interface submodule can encapsulate a unified skill invocation interface based on the invocation method in the skill description, and provide standardized invocation methods to the outside world, such as invoke_skill(skill_id, input_payload); it is responsible for parameter checking and format conversion before invocation, parsing, standardizing and error handling of tool output after invocation; and it encapsulates the skill invocation result into a unified result object (such as SkillResult) and passes it to the multi-objective evaluation and closed-loop optimization module 40 and the data and model management module 50.

[0131] In this embodiment, through the skill description submodule, the skill registration and discovery submodule, and the skill invocation interface submodule, the RNA design skill library and the skill description module 30 can abstract heterogeneous mRNA design tools into skills that can be uniformly scheduled by intelligent agents, thereby realizing the pluggability and scalability of the underlying tools.

[0132] In some embodiments, the multi-objective evaluation and closed-loop optimization module 40 includes: a candidate sequence management submodule, a multi-index evaluation submodule, a multi-objective comprehensive scoring submodule, and a closed-loop optimization control submodule.

[0133] It should be noted that the multi-objective evaluation and closed-loop optimization module 40 is mainly responsible for calculating and comprehensively evaluating candidate mRNA sequences in multiple dimensions, and realizing closed-loop optimization based on interaction with the intelligent agent.

[0134] The candidate sequence management submodule is used to manage the candidate mRNA sequences generated in each iteration round, assign a unique identifier to the candidate mRNA sequence, and maintain the candidate sequence state set.

[0135] The candidate sequence state set includes the candidate set, the current primary selection set, and the historical elimination set.

[0136] Specifically, the candidate sequence management submodule is used to manage the candidate mRNA sequences generated in each iteration round, associating each sequence with its source (e.g., initial generation, local optimization in a certain round) and the corresponding evaluation result.

[0137] The multi-index evaluation submodule, candidate sequence management submodule, task planning and intelligent agent scheduling module 20, and data and model management module 50 are used to call the evaluation skills in the RNA design skill library to perform multi-dimensional index analysis on candidate mRNA sequences, obtain multiple typical indicators, and convert the typical indicators corresponding to each candidate mRNA sequence into EvaluationRecord objects in a standardized form and send them to the data and model management module 50.

[0138] As an example, the multi-index evaluation submodule can call evaluation skills in the RNA design skill library to perform multi-index analysis on candidate mRNA sequences according to the instructions issued by the task planning and intelligent agent scheduling module 20.

[0139] Typical indicators may include, but are not limited to: expression efficiency-related indicators: codon fitness, translation efficiency prediction, ribosome occupancy, etc.; structure and stability indicators: global / local secondary structure stability (ΔG), structural entropy, UTR structural features, etc.; immunogenicity and safety indicators: potential epitope distribution, undesirable immune elements, potential off-target risk, etc.; manufacturability indicators: sequence length, GC content, repetitive sequences and difficult-to-synthesize fragments, etc.; each indicator is stored in a standardized form as an EvaluationRecord object and associated with the corresponding candidate sequence.

[0140] The multi-objective comprehensive scoring submodule is connected to the multi-indicator evaluation submodule. It is used to map multiple typical indicators into a comprehensive scoring result based on task weights and constraints and through a preset scoring strategy.

[0141] Preferably, the comprehensive scoring result is a comprehensive score or a ranking result.

[0142] Specifically, the preset scoring strategies can include different modes such as weighted summation, hierarchical screening, constraint satisfaction judgment, and Pareto front selection. Among them, the weights and screening thresholds can be adaptively adjusted based on experimental feedback or historical experience.

[0143] The closed-loop optimization control submodule is connected to the multi-objective comprehensive scoring submodule and the task planning and intelligent agent scheduling module 20. It is used to screen candidate mRNA sequences based on the comprehensive scoring results, construct feedback summary information for the screened mRNA sequences, and send the feedback summary information to the task planning and intelligent agent scheduling module 20. It is also used to control the termination conditions of the iteration process.

[0144] The termination conditions include reaching a preset scoring threshold, the maximum number of iterations, and insufficient improvement.

[0145] As an example, the closed-loop optimization control submodule can select a set of candidate sequences with better or more representative performance based on the comprehensive scoring results, and construct feedback summary information, which includes indicator statistics, advantages and disadvantages, etc. The feedback summary information is then passed to the task planning and intelligent agent scheduling module 20, and the large language model determines the sequence generation strategy for the next round, such as local mutation, codon substitution, UTR substitution, structure-oriented optimization, etc.

[0146] In this embodiment, the candidate sequence management submodule, multi-index evaluation submodule, multi-objective comprehensive scoring submodule, and closed-loop optimization control submodule are used to achieve systematic optimization of mRNA vaccine candidate sequences on multi-dimensional indicators, integrating the originally scattered evaluation steps into a unified closed-loop process.

[0147] In some embodiments, the data and model management module 50 includes: a task and configuration management submodule, a sequence and evaluation result storage submodule, a skill and model version management submodule, and a log and audit submodule.

[0148] It should be noted that the data and model management module 50 is mainly used to support the operation of each module and provide reliable data storage and version control for the mRNA vaccine design process.

[0149] The Task and Configuration Management submodule stores the task description objects corresponding to each design task and records the time points of task creation, update and execution to support task query and reproduction.

[0150] As an example, the Task and Configuration Management submodule stores the TaskProfile for each design task, including antigen information, design constraints, optimization goals, etc.

[0151] The sequence and evaluation results storage submodule is used to store candidate mRNA sequences and their corresponding EvaluationRecord objects, and supports the retrieval of candidate mRNA sequences by filtering criteria, including: task ID, round number, and score range.

[0152] It should be noted that all candidate mRNA sequences and their EvaluationRecords are stored, including sequences and corresponding indicators generated in each round of iteration.

[0153] The Skills and Model Version Management submodule is used to manage the registration and version information of RNA design skills, record the historical changes of each RNA design skill, and manage the versions and parameters of each model within the system.

[0154] Specifically, the Skills and Model Version Management submodule manages the versions and parameters of the large language model and other internal prediction models, ensuring that design tasks executed at different points in time can be traced back to the corresponding model version.

[0155] The logging and auditing submodule is used to record log information and supports auditing and quality analysis of the design process.

[0156] Specifically, the logging and auditing submodule is used to record log information for key steps such as skill invocation, intelligent agent decision-making, and multi-objective evaluation.

[0157] In this embodiment, the task and configuration management submodule, sequence and evaluation result storage submodule, skill and model version management submodule, and log and audit submodule provide reliable data storage and version control support for mRNA vaccine design through full-process management of task configuration, sequence results, skill model versions, and logs, ensuring that the design process is queryable, reproducible, and auditable; at the same time, it supports accurate retrieval of candidate sequences and traceability of model versions, effectively improving the standardization, controllability, and overall operational efficiency of the vaccine design process.

[0158] In some embodiments, such as Figure 1 As shown, the mRNA vaccine sequence design system based on a large language model intelligent agent also includes: an experimental feedback integration module 60.

[0159] The experimental feedback integration module 60 is connected to the multi-objective evaluation and closed-loop optimization module 40 and the task planning and intelligent agent scheduling module 20, and includes: an experimental data receiving submodule, a data standardization and mapping submodule and a feedback interface submodule.

[0160] As an example, the experimental feedback integration module 60 can be used to receive in vitro or in vivo experimental data of mRNA vaccine candidate sequences, store the experimental results in a structured manner, and map them to experimental feedback parameters that update multi-objective evaluation strategies or design constraints. This allows the intelligent agent to iteratively optimize in subsequent design tasks by incorporating real experimental feedback. For example, backfilling IC50 values ​​to adjust TE weights.

[0161] Specifically, the experimental data receiving submodule is used to receive in vitro or in vivo experimental data of mRNA vaccine candidate sequences.

[0162] The experimental data include: expression level, immune response intensity, safety indicators, etc.

[0163] It should be noted that the experimental data receiving submodule supports importing experimental data from the experimental data management system, electronic experimental record book, or user-uploaded files.

[0164] The data standardization and mapping submodule is connected to the experimental data receiving submodule. It is used to convert experimental data into a unified data structure and map the data structure into experimental feedback parameters for updating multi-objective evaluation strategies.

[0165] As an example, the data standardization and mapping submodule can transform experimental data from different sources and in different formats into a unified data structure and associate it with corresponding candidate sequences; it can also map experimental indicators into parameters that can be used to update multi-objective evaluation strategies, such as adjusting the weights of certain evaluation indicators or modifying screening thresholds.

[0166] The feedback interface submodule is connected to the data standardization and mapping submodule, the multi-objective evaluation and closed-loop optimization module 40, and the task planning and intelligent agent scheduling module 20, and is used to send experimental feedback parameters to the multi-objective evaluation and closed-loop optimization module 40 and the task planning and intelligent agent scheduling module 20.

[0167] As an example, the feedback interface submodule provides the processed experimental feedback information to the multi-objective evaluation and closed-loop optimization module 40 and the task planning and intelligent agent scheduling module 20; it also allows the intelligent agent to update subsequent task planning strategies based on experimental data, such as giving more importance to safety or expression efficiency.

[0168] In this embodiment, through the experimental feedback integration module 60, the present invention can incorporate real experimental results into the design closed loop during the design process, realize the continuous improvement process of "computational design - experimental verification - strategy update", promote the systematic accumulation and reuse of mRNA vaccine design experience, improve the optimization accuracy, and facilitate continuous optimization and knowledge accumulation by combining experimental data.

[0169] In some embodiments, such as Figure 1 As shown, the mRNA vaccine sequence design system based on a large language model intelligent agent also includes: a code generation and execution module 70.

[0170] The code generation and execution module 70 includes: a code template and generation submodule, a secure execution environment submodule, and a result parsing and error handling submodule.

[0171] As an example, the code generation and execution module 70 can be used in a secure and controlled operating environment to automatically generate script code that calls the underlying RNA computing tools based on skill descriptions, execute the script, and parse the output results. Furthermore, when errors or exceptions occur, it combines a large language model to perform error analysis and script modification, thereby further reducing the reliance on manually written glue code.

[0172] Specifically, the code template and generation submodule are used to maintain multiple script templates related to RNA design skills, and generate or complete script code by the large language model according to the calling method and parameter definition in the skill description object, so as to realize the call to the underlying tools.

[0173] The script templates include Python, R, or Shell script templates.

[0174] The Secure Execution Environment submodule is used to run scripts generated by the code generation submodule in an isolated execution environment, avoiding security risks to the main system, and recording standard output, error messages and return values ​​during script execution.

[0175] The result parsing and error handling submodule is used to parse the result file or standard output generated by the script execution, convert it into a standard SkillResult object, and also to organize the error information into a prompt when an error or exception is detected, so that the large language model can provide correction suggestions or automatically generate a repaired script and re-execute it.

[0176] In this embodiment, by introducing a code generation and execution module 70, the system of the present invention can still flexibly call various underlying tools even when the skill description is relatively abstract, thereby enhancing the scalability and automation of the entire system. At the same time, it reduces the burden of manually writing glue code.

[0177] In an exemplary architecture, users can access the system's front-end interface via a browser or client program. The front-end sends the target antigen information and mRNA vaccine design requirements input by the user to the back-end server, where they are preprocessed by the user interaction and task parsing module 10. Subsequently, the task planning and intelligent agent scheduling module 20 drives the entire design process. The RNA design skill library and skill description module 30 provides the scheduling module with a list of available skills and skill metadata. The multi-objective evaluation and closed-loop optimization module 40 performs index calculations and iterative optimization on candidate sequences. The data and model management module 50 is responsible for the storage and management of data and configuration throughout the entire process. The experimental feedback integration module 60 and the code generation and execution module 70 are used to enhance system functionality.

[0178] In addition, this invention also provides an mRNA vaccine sequence design method based on a large language model intelligent agent, which is applied to the aforementioned mRNA vaccine sequence design system based on a large language model intelligent agent, such as... Figure 2 As shown, the mRNA vaccine sequence design method based on a large language model intelligent agent includes:

[0179] S1: Obtain vaccine design parameters and send them to the user interaction and task parsing module. The user interaction and task parsing module parses and transforms the vaccine design parameters through a large language model to obtain a task description object.

[0180] S2, the task planning and intelligent agent scheduling module constructs a DAG workflow based on the task description object and the skills in the RNA design skill library and skill description module, and generates candidate mRNA sequences by calling the skills in the RNA design skill library and skill description module through the intelligent agent.

[0181] S3, the multi-objective evaluation and closed-loop optimization module, performs multi-dimensional index analysis and comprehensive evaluation on candidate mRNA sequences, obtains comprehensive scoring results, and feeds back the comprehensive scoring results to the task planning and intelligent agent scheduling module.

[0182] S4, determine whether the candidate mRNA sequence meets the preset design constraints or convergence conditions.

[0183] S5, if satisfied, output the current candidate mRNA sequence.

[0184] S6. If the conditions are not met, the intelligent agent modifies, replaces, or regenerates the candidate mRNA sequences based on the comprehensive scoring results to form a new round of candidate mRNA sequences and returns to step S3.

[0185] During the execution of the DAG workflow, the task planning and intelligent agent scheduling module dynamically adjusts the skills selected from the RNA design skill library and skill description module based on the comprehensive scoring results.

[0186] For other specific implementations of the mRNA vaccine sequence design method based on a large language model intelligent agent according to the embodiments of the present invention, please refer to the specific implementation of the mRNA vaccine sequence design system based on a large language model intelligent agent according to the above embodiments of the present invention.

[0187] Next, through a typical and detailed embodiment, the implementation process, technical details, and interaction logic of the mRNA vaccine sequence design method based on a large language model intelligent agent of the present invention are illustrated in practical applications. This embodiment takes the design of a prophylactic mRNA vaccine against the SARS-CoV-2 virus spike protein (UniProt ID: P0DTC2, full-length 1273 amino acid residues) as an example, aiming to generate an optimized mRNA sequence with a length not exceeding 4000 nucleotides (nt), meeting the following key constraints: translation efficiency (TE) greater than 80%, GC content between 40-60%, immunogenicity off-target risk less than 5%, and secondary structure stability (minimum free energy ΔG) less than -20 kcal / mol. This design is particularly suitable for rapidly responding viral variants and supports multiple HLA allele coverage (such as A02:01 and B07:02) to enhance the broad spectrum of T cell immune responses.

[0188] The system is deployed on a cloud platform (such as an AWS EC2 instance equipped with an NVIDIA A100 GPU accelerator), and the LLM core uses a finely tuned version of the mRNABERT model (based on pre-trained data of more than 18 million mRNA sequences, and integrates a Flash Attention mechanism to efficiently process long sequence inputs). The skill library integrates the latest open-source and commercial tools available in 2025, including GEMORNA (a generative Transformer model for de novo mRNA sequence design, which can improve SARS-CoV-2 spike protein expression levels by 41 times), RiboNN (a convolutional neural network model for cell-specific TE prediction, covering more than 140 human and mouse cell types, with a prediction accuracy 2 times higher than the benchmark model), LinearDesign (a dynamic programming algorithm for mRNA structure reconstruction, achieving a 3-fold increase in expression in spike protein optimization), NetMHCpan-4.2 (a deep learning framework for MHC I / II binding affinity prediction, integrating the IEDB database to assess off-target risk), and mRNAdesigner (a web server tool for joint optimization of CDS (Coding DNA Sequence) and UTR, which can precisely adjust GC content to 54% and improve the codon adaptation index (CAI) to 0.96). The entire end-to-end process takes less than 24 hours and supports 3-5 rounds of iterative optimization of the initial 500 candidate sequences.

[0189] For ease of description, Table 1 summarizes the key parameters and expected results of the embodiments (based on simulation verification data from RiboNN and LinearDesign):

[0190] Table 1: Key Parameters and Expected Results

[0191]

[0192] This embodiment demonstrates that the present invention achieves highly efficient multi-objective automation in mRNA vaccine design, outperforming existing benchmarks such as BNT162b2 or mRNA-1273 (TE improvement of only 20-25%). Figure 3 As shown, the process is mainly divided into six main steps, each of which is further subdivided into sub-steps, emphasizing direction A (skill description and DAG scheduling) and direction B (Pareto multi-objective optimization), with optional integration of direction C (experimental feedback integration) and direction D (code generation and execution).

[0193] Step 1: User Interaction and Task Parsing. This step is responsible for extracting and standardizing the mRNA vaccine design requirements from the user's natural language or structured input, ensuring that the system understands personalized constraints and avoids optimization biases caused by ambiguity. The entire step takes less than 1 minute and relies on the semantic parsing capabilities of LLM.

[0194] Sub-step 1.1: Input Reception Users submit core data through the system's web front-end interface (an interactive form built on React.js). Specifically, users upload or directly input the amino acid sequence of the SARS-CoV-2 spike protein (FASTA format, e.g., starting with "MVNILLF...", 1273 residues in length), and provide a natural language description. Additionally, users can upload optional attachments, such as a CSV file of an existing UTR sequence library (containing beta-globin 5'UTR and hGH 3'UTR variants). The front-end performs initial validation: checking sequence format compliance (e.g., only A / C / G / U / T bases, no illegal characters), length limit (<4000 nt), and HLA allele validity (cross-referenced to the IEDB database). After successful validation, the data is packaged into a JSON payload and sent to the backend API endpoint (e.g., / api / v1 / parse_task) via an HTTPS POST request.

[0195] Sub-step 1.2: The semantic parsing backend calls the LLM core (mRNABERT fine-tuning model) for semantic analysis. The system constructs a dedicated prompt template: "Based on the following user input [complete text, including sequence and description], parse the mRNA vaccine design task: extract antigen details (ID, sequence type), constraints (TE threshold > 80%, GC range 40-60%, off-target < 5%, ΔG < -20 kcal / mol), optimization objectives and their weights (immunogenicity 0.4, TE 0.3, stability 0.2, manufacturability 0.1), and handle fuzzy terms (such as the default UTR library). Output a standardized JSON TaskProfile object." After LLM processing, a structured output is generated, for example: the antigen identifier is UniProt P0DTC2, the task type is "full-length spike prophylaxis vaccine", and the priority index weight is immunogenicity dominant (0.4 weight). If the HLA allele is not specified in the input, LLM automatically completes the common coverage set (based on the top 10% alleles in IEDB prevalence). The parsing accuracy reaches over 95% through internal cross-validation, ensuring no constraint conflicts (e.g., the GC range is compatible with the TE target).

[0196] Sub-step 1.3: Configuration, Construction, and Verification. The LLM output is converted into an internal TaskProfile object, and logical verification is performed: Constraint consistency is checked (e.g., if the GC limit is <40%, the system prompts the user to correct); default values ​​are completed (e.g., poly(A) tail length 120nt, head-end cap structure ARCA). After successful verification, the TaskProfile is persistently stored in the data management module (MongoDB collection tasks, index task_id). Finally, the object is asynchronously passed to the task planning module via an internal message queue (Redis pub / sub channel). See Table 2 for the fields, example values, and descriptions of the example TaskProfile.

[0197] Table 2: TaskProfile Field Definitions and Example Values

[0198]

[0199] Step 2: Task Planning and Intelligent Agent Scheduling. This step dynamically constructs a directed acyclic graph (DAG) workflow (direction A) based on TaskProfile, enabling skill dependency resolution and parallel scheduling to ensure computational efficiency (e.g., GPU resource allocation). The entire step takes 5-10 minutes and processes the planning of 500 sequences.

[0200] Sub-step 2.1: The task planning system first queries the RNA design skill library (PostgreSQL table skills), filters the subset of skills that match the task (SQL condition: rna_type='mRNA_vaccine' AND metrics CONTAINS 'TE,off_target,delta_g'), and returns 10-15 candidates (such as GEMORNA for generation and RiboNN for TE evaluation). The LLM core then generates an initial DAG draft, guided by a prompt: "Construct the DAG workflow for [TaskProfile] and [skill list]: nodes correspond to skill calls (e.g., CDS generation → UTR optimization → structure prediction), edges represent dependencies (e.g., UTR is required after CDS), prioritize parallel evaluation to minimize time, and support branches such as immune screening independent of TE." The generated DAG includes the core node sequence: Node 1 (GEMORNA: de novo CDS generation, input protein sequence + GC constraints, output 500 CDS variants); Node 2 (OptimumGene: codon optimization, dependent on node 1, output CAI>0.95 sequence); Parallel branch node 3 (mRNAdesigner: UTR engineering, dependent on node 2, output 5' / 3'... UTR combination, improving ribosome loading); Node 4 (RNAfold: preliminary ΔG prediction, dependent on node 2); Node 5 (LinearDesign: global stability reconstruction, dependent on nodes 3 and 4, outputting full-length mRNA variants); Node 6 (NetMHCpan: immunogenicity screening, dependent on node 5); Node 7 (RiboNN: TE prediction, dependent on node 5). DAG optimization considers resources (e.g., RiboNN prioritizes GPU), with a total of 8-12 nodes.

[0201] Sub-step 2.2: The Intelligent Agent Decision LLM acts as an agent to monitor intermediate results during DAG execution (e.g., after node 2, if 20% of the sequences have a CAI < 0.9, prompt: "Analyze [Intermediate Evaluation Summary], decide on DAG adjustment: if TE is initially < 70%, insert LinearDesign to reconstruct the node and reroute; if off-target > 4%, prioritize the NetMHCpan branch."). Decisions are recorded as log entries, such as "Decision: Insert additional UTR mutation assessment, reason: initial batch ΔG > -18 kcal / mol in 15% of sequences." This sub-step ensures adaptability, with a dynamic branching rate < 10%.

[0202] Sub-step 2.3: Scheduling and Execution. The scheduler and coordinator distributes tasks to the execution queue (Celery distributed task framework), preparing the input load for each node (e.g., node 1: {"protein_seq": "spike_aa","num_variants":500,"gc_target":50}). Parallel processing is supported (e.g., nodes 3 and 4 run simultaneously, multi-process pool <10 minutes / batch), and resources are monitored (GPU utilization >80%). Execution logs include timestamps and status (e.g., "Node 1 completed: 500 CDS generated, average length 3200nt"). In case of an exception (e.g., API timeout), the system automatically retryes 3 times.

[0203] Step 3: Initial Candidate Sequence Generation. This step utilizes generative skills to generate diverse initial batches, covering the mRNA design search space (>10^600 variants), focusing on CDS and UTR engineering. Time required: <5 minutes.

[0204] Sub-step 3.1: CDS and UTR Generation of GEMORNA. The skill first executes node 1, inputting the antigen sequence and constraints from the TaskProfile, generating 500 de novo CDS variants (incorporating S-2P mutations to stabilize the prefusion conformation, resulting in a 41-fold increase in baseline expression). Subsequently, the mRNAdesigner skill (node ​​3) combines the UTRs (selecting beta-globin 5' UTR and hGH3' UTR from the library, <200 nt in length), inserting a poly(A) tail (120 nt) and an ARCA cap structure. Example generated sequence: seq_id="gen_001", full-length mRNA="AUG[optimized_CDS 3120 nt]...[3'UTR 150 nt]AAAAA...A", initial GC=52%.

[0205] Sub-step 3.2: Preliminary Preprocessing. The OptimumGene skill (node ​​2) applies human codon preference optimization (CAI>0.95, avoiding unfavorable dinucleotides such as UU / UA), filtering out variants with high synthesis difficulty (repetitive sequences>2%). The LinearDesign skill (node ​​4) preliminarily reconstructs hotspots (e.g., stem-loop regions), ensuring ΔG>-18 kcal / mol as the starting point for iteration. The batch is reduced to 400 valid sequences, with a rejection rate of 20% (based on length / CAI threshold).

[0206] Step 4: Multi-objective evaluation. This step involves parallel invocation of assessment skills, generating indicator vectors and performing preliminary screening (direction B), taking 10-15 minutes per batch.

[0207] Sub-step 4.1: Indicator calculations are performed according to DAG nodes 6-7. The NetMHCpan skill is used to assess immunogenicity (input translated peptide, output off-target risk: weak MHC-binding peptide ratio <5%, HLA A02:01 / B07:02 coverage); the RiboNN skill is used to predict TE (lung epithelial cell model >80%, 140 cell type baseline); RNAfold / mRNAdesigner is used to calculate ΔG (global + local stability, <-20 kcal / mol); manufacturing indicators are evaluated using custom rules (GC balance, repeatability <2%, synthesis fraction <moderate). Each sequence generates a complete EvaluationRecord, for example, seq_id="gen_001": TE=0.82, ΔG=-22.5 kcal / mol, off-target=3.2%, GC=52%.

[0208] Sub-step 4.2: Preliminary comprehensive screening using the NSGA-II variant algorithm (adaptive weighted Pareto front). Input all records and generate a sorted list (Top-100 sequences, balanced TE vs. off-target). Threshold filtering removes unqualified items (e.g., TE < 75% or off-target > 5%), with a retention rate of 70%. See Table 3 for the screening results.

[0209] Table 3: Example of Filtering Results

[0210]

[0211] Step 5: Closed-loop iterative optimization. Based on evaluation feedback, LLM drives sequence mutation, achieving convergence in 3-5 rounds (direction B), with a total time of <15 hours.

[0212] Sub-step 5.1: Feedback generates batch statistics from the optimization module (e.g., "Batch 1: TE mean 76%, off-target mean 4.5%; 30% of sequences ΔG>-20, it is recommended to modify UTR hotspots"), which are then packaged into a summary report and passed to LLM.

[0213] Sub-step 5.2: LLM Mutation Generation LLM prompt: "Based on [Evaluation Abstract] and [Task Profile], generate mutations for low-performance sequences (e.g., TE < 80%): Apply N1-methylpseuuridine modification, codon substitution, or UTR variants to produce 50 new sequences + 50 local modifications, ensuring constraint compliance." Output a new batch (e.g., Round 2: 10% base substitution rate, 5 UTR combinations).

[0214] Sub-step 5.3: Iteration Control and Convergence. Repeat steps 4-5.2: Round 2: TE increases to 81%, off-target decreases to 3.5%; Round 3: Insert LinearDesign reconstruction, TE reaches 84%; Round 4: Convergence (improvement <5%, Pareto front stable). Optional direction C: If importing in vitro data (CSV format: seq_id, measured TE=78%, IC50=low), map and update weights (TE+0.1), triggering an additional round. Example of direction D: LLM generates a script description ("Execute RNAfold on seq_001 with --T=37C"), and runs it in a sandbox. Total iterations: 4 rounds, retaining the Top-50 sequences.

[0215] Step 6: Output and Report Generation. The top 10 Pareto sequences are finally selected, and a comprehensive report is generated, taking less than 2 minutes.

[0216] Sub-step 6.1: Sequence filtering and export. Select the winners from the Pareto front (e.g., seq_id="opt_007": TE=88%, off-target=2.1%, ΔG=-25.0) and export the FASTA file ("opt_007: AUG...polyA").

[0217] Sub-step 6.2: Report Generation. The report includes: indicator heatmap (TE vs. ΔG scatter plot), DAG execution path log, and optimization suggestions (e.g., "opt_007 is compatible with LNP delivery, in vitro validation is recommended"). It is stored in the data module and can be downloaded via API.

[0218] like Figure 4 As shown, through RiboNN and LinearDesign simulations, this embodiment achieves a 32% increase in TE, a 64% reduction in off-target, and >95% immune coverage (multi-HLA) in spike protein optimization, significantly outperforming the existing mRNA-1273 benchmark. Real-world deployment can further integrate in vitro feedback to improve generalization.

[0219] It is worth mentioning that, through the mRNA vaccine sequence design system and method based on intelligent agents of large language models, this invention realizes task understanding, tool scheduling, multi-objective evaluation and closed-loop optimization of mRNA vaccine sequence design tasks under a unified framework, providing an intelligent, scalable and general design platform for mRNA vaccine research and development.

[0220] In addition, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described mRNA vaccine sequence design method based on a large language model intelligent agent.

[0221] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of this application. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of this application. All such changes and modifications are intended to be included within the scope of this application as claimed in the appended claims.

[0222] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0223] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0224] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0225] Similarly, it should be understood that, in order to streamline this application and aid in understanding one or more of the various inventive aspects, features of this application may sometimes be grouped together in a single embodiment, figure, or description thereof in the description of exemplary embodiments of this application. However, this approach should not be construed as reflecting an intention that the claimed application requires more features than are expressly recited in each claim. Rather, as reflected in the corresponding claims, its inventive point lies in solving the corresponding technical problem with features fewer than all features of a single disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0226] Those skilled in the art will understand that, apart from the mutual exclusion of features, all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or elements of any method or apparatus so disclosed may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0227] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0228] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules according to the embodiments of this application. This application can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0229] It should be noted that the above embodiments are illustrative of this application and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0230] The above description is merely a specific embodiment or illustration of the embodiments of this application. The scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. The scope of protection of this application shall be determined by the scope of the claims.

Claims

1. An mRNA vaccine sequence design system based on a large language model intelligent agent, characterized in that, The system includes: a user interaction and task parsing module, a task planning and intelligent agent scheduling module, an RNA design skill library and skill description module, a multi-objective evaluation and closed-loop optimization module, and a data and model management module. The user interaction and task parsing module is used to receive vaccine design parameters and parse and transform the vaccine design parameters through a large language model to obtain a task description object. The task planning and intelligent agent scheduling module is connected to the user interaction and task parsing module and the RNA design skill library and skill description module. It is used to construct a DAG workflow based on the task description object and the skills in the RNA design skill library and skill description module, call the skills in the RNA design skill library and skill description module through intelligent agents, generate candidate mRNA sequences, and dynamically adjust the skills selected from the RNA design skill library and skill description module according to the comprehensive scoring results during the execution of the DAG workflow. The RNA design skill library and skill description module are connected to the multi-objective evaluation and closed-loop optimization module and the data and model management module. They are used to encapsulate the mRNA molecule design and multi-dimensional characteristic evaluation functions to obtain multiple skill modules. Each skill module is uniformly defined, registered and managed. The multi-objective evaluation and closed-loop optimization module is connected to the task planning and intelligent agent scheduling module. It is used to perform multi-dimensional index analysis and comprehensive evaluation on the candidate mRNA sequence, obtain the comprehensive score result, and feed the comprehensive score result back to the task planning and intelligent agent scheduling module. The data and model management module is connected to the multi-objective evaluation and closed-loop optimization module, and is used to manage and store system data and model parameters and tool configurations related to each RNA design skill. The task planning and intelligent agent scheduling module includes: The task planning submodule is connected to the user interaction and task parsing module and the RNA design skill library and skill description module. It is used to query relevant skill sets from the RNA design skill library and skill description module according to the task description object, and construct the DAG workflow by combining the dependencies between skills, input-output compatibility and resource constraints. The intelligent agent decision-making submodule, connected to the task planning submodule, is used to analyze intermediate results through a large language model during the execution of the DAG workflow, obtain evaluation results, and dynamically adjust the skill call order, parameter configuration, and optimization strategy based on the evaluation results. The scheduling and execution coordination submodule is connected to the RNA design skill library and skill description module and the multi-objective evaluation and closed-loop optimization module. It is used to issue invocation tasks, convert the data format of the input data of skill invocation, and send the skill invocation results and execution logs to the multi-objective evaluation and closed-loop optimization module. The system data includes the task description object and the candidate mRNA sequence, and the task invocation is to invoke the tool interface in the RNA design skill library and skill description module.

2. The mRNA vaccine sequence design system based on a large language model intelligent agent according to claim 1, characterized in that, The user interaction and task parsing module includes: The input receiving submodule is used to receive vaccine design parameters input via natural language. The task semantic parsing submodule, connected to the input receiving submodule, is used to perform semantic analysis on the vaccine design parameters through a large language model to obtain structured parsing results; The task configuration construction submodule is connected to the task semantic parsing submodule and is used to convert the parsing result into the task description object; The vaccine design parameters include target antigen information, indication scenarios, and design constraints.

3. The mRNA vaccine sequence design system based on a large language model intelligent agent according to claim 2, characterized in that, The RNA design skill set and skill description module include: The skill description submodule is used to define a structured skill description object for each skill module in the skill library; The skill registration and discovery submodule, connected to the skill description submodule, is used to provide a skill registration interface, skill retrieval and skill filtering, and to verify the skill description object during the skill registration process; The skill call interface submodule is connected to the multi-objective evaluation and closed-loop optimization module and the data and model management module. It is used to encapsulate a unified call interface, provide a standardized call method, and encapsulate the skill call result into a unified result object, and send the result object to the multi-objective evaluation and closed-loop optimization module and the data and model management module. The structured skill description object includes at least three of the following: skill identifier and version information, applicable object type, input parameter definition, output result definition, dependency relationship, performance and resource information, and invocation method.

4. The mRNA vaccine sequence design system based on a large language model intelligent agent according to claim 1, characterized in that, The multi-objective evaluation and closed-loop optimization module includes: The candidate sequence management submodule is used to manage the candidate mRNA sequences generated in each iteration round, assign a unique identifier to the candidate mRNA sequence, and maintain a set of candidate sequence states. The multi-index evaluation submodule, together with the candidate sequence management submodule, the task planning and intelligent agent scheduling module, and the data and model management module, is used to call the evaluation skills in the RNA design skill library to perform multi-dimensional index analysis on the candidate mRNA sequences, obtain multiple typical indicators, and convert the typical indicators corresponding to each candidate mRNA sequence into EvaluationRecord objects in a standardized form and send them to the data and model management module. The multi-objective comprehensive scoring submodule, connected to the multi-indicator evaluation submodule, is used to map multiple typical indicators into a comprehensive scoring result based on task weights and constraints and through a preset scoring strategy. The closed-loop optimization control submodule is connected to the multi-objective comprehensive scoring submodule and the task planning and intelligent agent scheduling module. It is used to screen the candidate mRNA sequences according to the comprehensive scoring results, construct feedback summary information for the screened mRNA sequences, and send the feedback summary information to the task planning and intelligent agent scheduling module. It is also used to control the termination conditions of the iteration process. The candidate sequence state set includes a candidate set, a current primary selection set, and a historical elimination set. The termination conditions include reaching a preset scoring threshold, an upper limit on the number of iteration rounds, and insufficient improvement.

5. The mRNA vaccine sequence design system based on a large language model intelligent agent according to claim 4, wherein the data and model management module comprises: The Task and Configuration Management submodule is used to store the task description objects corresponding to each design task and record the time points of task creation, update and execution; The sequence and evaluation result storage submodule is used to store the candidate mRNA sequences and their corresponding EvaluationRecord objects, and supports the retrieval of candidate mRNA sequences by screening conditions, wherein the screening conditions include: task ID, round number and score range; The Skills and Model Version Management submodule is used to manage the registration and version information of RNA design skills, record the historical changes of each RNA design skill, and manage the version and parameters of each model within the system. The logging and auditing submodule is used to record log information and support auditing and quality analysis of the design process.

6. The mRNA vaccine sequence design system based on a large language model intelligent agent according to claim 3, the system further includes: Experimental feedback integration module; The experimental feedback integration module is connected to the multi-objective evaluation and closed-loop optimization module and the task planning and intelligent agent scheduling module, and includes: an experimental data receiving submodule, a data standardization and mapping submodule and a feedback interface submodule. The experimental data receiving submodule is used to receive in vitro or in vivo experimental data of the mRNA vaccine candidate sequence. The data standardization and mapping submodule is connected to the experimental data receiving submodule and is used to convert the experimental data into a unified data structure and map the data structure into experimental feedback parameters for updating the multi-objective evaluation strategy. The feedback interface submodule is connected to the data standardization and mapping submodule, the multi-objective evaluation and closed-loop optimization module, and the task planning and intelligent agent scheduling module, and is used to send the experimental feedback parameters to the multi-objective evaluation and closed-loop optimization module and the task planning and intelligent agent scheduling module. The experimental data include: expression levels, immune response strength, and safety indicators.

7. The mRNA vaccine sequence design system based on a large language model intelligent agent according to claim 1, the system further comprising: Code generation and execution module; The code generation and execution module includes: a code template and generation submodule, a secure execution environment submodule, and a result parsing and error handling submodule; The code template and generation submodule are used to maintain multiple script templates related to RNA design skills, and generate or complete script code by the large language model according to the calling method and parameter definition in the skill description object. The secure execution environment submodule is used to run the script generated by the code generation submodule in an isolated execution environment, and to record the standard output, error messages and return values ​​during the script execution process; The result parsing and error handling submodule is used to parse the result file or standard output generated by the script execution, and also to organize the error information into prompts when errors or anomalies are detected, so that the large language model can provide correction suggestions or automatically generate a repaired script.

8. A method for designing mRNA vaccine sequences based on a large language model intelligent agent, characterized in that, The method, applied to the mRNA vaccine sequence design system based on a large language model intelligent agent as described in any one of claims 1-7, comprises: The vaccine design parameters are obtained and sent to the user interaction and task parsing module. The user interaction and task parsing module parses and transforms the vaccine design parameters through a large language model to obtain a task description object. The task planning and intelligent agent scheduling module constructs a DAG workflow based on the task description object and the skills in the RNA design skill library and skill description module, and generates candidate mRNA sequences by calling the skills in the RNA design skill library and skill description module through an intelligent agent; The multi-objective evaluation and closed-loop optimization module performs multi-dimensional index analysis and comprehensive evaluation on candidate mRNA sequences to obtain a comprehensive score result, and feeds the comprehensive score result back to the task planning and intelligent agent scheduling module. Determine whether the candidate mRNA sequence meets the preset design constraints or convergence conditions; If the conditions are met, output the current candidate mRNA sequence; If the requirements are not met, the intelligent agent modifies, replaces, or regenerates the candidate mRNA sequence based on the comprehensive scoring result to form a new round of candidate mRNA sequences, and returns to the steps of the multi-objective evaluation and closed-loop optimization module to perform multi-dimensional index analysis and comprehensive evaluation of the candidate mRNA sequences. During the execution of the DAG workflow, the task planning and intelligent agent scheduling module dynamically adjusts the skills selected from the RNA design skill library and skill description module based on the comprehensive scoring results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the computer program is executed by the processor, it implements the mRNA vaccine sequence design method based on a large language model intelligent agent as described in claim 8.

Images