A digital twin experiment intelligent agent system for microbial detection

Abstract

The application provides a digital twin experiment intelligent agent system for microbial detection, comprising a research and development target access module, a field knowledge and data base module, a programmed integration framework module, a digital twin experiment environment module, a multi-agent collaboration module, a virtual experiment evaluation module, an experiment feedback optimization module and a result output module; structured intermediate results are continuously transmitted between the modules around the same detection research and development task. The application accesses the public database, professional analysis tool, experiment specification, internal experiment record and project knowledge required for microbial detection research and development through a unified programmed integration framework, and automatically collaborates through multiple professional agents in target discovery, detection method design, recognition molecule engineering, algorithm modeling, scheme evaluation and experiment feedback optimization, etc. Thus, multiple rounds of virtual design, simulation verification and scheme screening are completed before actual experiments, improving the efficiency, reliability and success rate of microbial detection method development.

Description

Technical Field

[0001] This invention relates to the fields of microbial detection, in vitro diagnostics research and development, digital twins, artificial intelligence agents, and bioinformatics computing, specifically to a digital twin experimental agent system for microbial detection. Background Technology

[0002] Microbial detection technologies are widely used in clinical diagnosis, food safety, environmental monitoring, disease control and prevention, drug development, and public health emergencies. When dealing with drug-resistant bacteria, emerging pathogens, low-abundance microorganisms, and complex mixed samples, researchers typically need to complete a significant amount of work, including candidate target screening, sequence conservation analysis, specificity verification, primer or probe design, antigen-antibody or aptamer screening, detection reaction system optimization, algorithm model establishment, and real-sample validation. This process involves multiple disciplines such as biology, computational chemistry, statistics, experimental automation, and regulatory standards, resulting in a long development chain, numerous variables, and high trial-and-error costs.

[0003] The existing microbial detection research and development process has the following main shortcomings:

[0004] 1. The R&D process is highly dependent on physical experiments and expert experience: Traditional methods often begin with experts identifying targets or protocols based on literature and experience, followed by gradual validation through numerous wet experiments. For emerging pathogens or drug-resistant bacteria where data is insufficient, early protocols have high uncertainty, easily leading to the repeated consumption of reagents, samples, manpower, and time.

[0005] 2. Fragmented and Disjointed Databases, Tools, and Knowledge Standards: Microbial detection research and development requires the simultaneous use of PubMed, NCBI, UniProt, PDB, KEGG, AlphaFold DB, BLAST, PyMOL, Rosetta, RDKit, statistical analysis tools, LIMS, ELN, and standard documents such as CLSI and ISO. Existing systems typically only solve a single problem and lack a procedural framework that can unify the orchestration, invocation, and tracking of heterogeneous tools and knowledge.

[0006] 3. Low efficiency of cross-disciplinary collaboration: Target discovery personnel, molecular engineers, detection method developers, algorithm modelers, and experimental verification personnel need to frequently exchange results and hypotheses. Due to inconsistencies in data formats, evaluation metrics, and technical languages, information loss easily occurs when solutions flow between different stages, making it difficult to establish reusable and traceable standardized protocols in the R&D process.

[0007] 4. The lack of a closed-loop update mechanism between virtual predictions and real experiments: Some existing tools can perform sequence alignment, structure prediction, or virtual screening, but their prediction results are usually used as a one-time reference and fail to continuously incorporate subsequent real experimental results. Virtual models cannot be continuously corrected based on experimental biases, causing subsequent projects to still rely on manual judgment.

[0008] 5. Incomplete evaluation dimensions: Current R&D management methods focus more on single technical indicators, such as sensitivity, specificity, or binding energy. However, actual testing products must also comprehensively consider sample type, reaction conditions, cost, throughput, time consumption, regulatory compliance, instrument compatibility, and mass production stability. The lack of a unified multi-objective evaluation mechanism may make a technical solution seem effective in the laboratory stage, but it may become unimplementable after entering the productization stage.

[0009] Therefore, there is an urgent need for a system that can organically integrate knowledge, tools, data and professional roles in microbial detection research and development, and drive a closed loop of virtual experiments through digital twins and multi-agent collaboration, so that researchers can obtain more thorough scheme simulations and risk assessments before putting physical experiments into place. Summary of the Invention

[0010] To address the aforementioned issues, this invention provides a digital twin experimental intelligent agent system for microbial detection. Taking microbial detection R&D goals as input, it forms a closed-loop process of "R&D goal input - virtual experiment design - simulation verification - experiment suggestion output - real result feedback - continuous model optimization" through steps such as project goal analysis, domain knowledge enhancement, tool programmatic scheduling, multi-agent collaborative reasoning, virtual experiment simulation, candidate solution evaluation, and real experiment feedback optimization.

[0011] The specific plan is as follows:

[0012] A digital twin experimental intelligent agent system for microbial detection includes a research and development target access module, a domain knowledge and data foundation module, a programmatic integration framework module, a digital twin experimental environment module, a multi-agent collaboration module, a virtual experiment evaluation module, an experiment feedback optimization module, and a result output module; the modules continuously transmit structured intermediate results around the same detection research and development task;

[0013] The R&D target access module is used to transform the detection objects, sample types, expected performance and constraints proposed by users into executable tasks. Specifically, it receives the microbial detection R&D targets input by users and parses them to generate structured R&D task packages.

[0014] The domain knowledge and data foundation module is used to support credible reasoning in virtual experiments, transforming heterogeneous knowledge data into a unified knowledge representation, and providing a credible source of knowledge for reasoning;

[0015] The programmatic integration framework module is used to standardize and encapsulate various external databases and professional tools, and to uniformly handle call requests, parameter conversion and result normalization, thus solving the problems of scattered professional tools and inconsistent interfaces.

[0016] The multi-agent collaboration module is used to assign tasks and exchange results according to the research and development stage. It includes multiple agents with different professional capabilities and a task orchestrator. Among the agents, the target discovery agent is used to screen candidate detection targets and prioritize them. The task orchestrator assigns tasks according to the matching degree between sub-task requirements and agent capabilities. Each agent performs analysis around the assigned sub-task and submits structured results to the shared workspace to realize multi-agent collaborative reasoning.

[0017] The digital twin experimental environment module is used to map target microorganisms, candidate targets, detection reaction systems, identification molecules and algorithm models into computable objects to build a virtual experimental environment;

[0018] The virtual experiment evaluation module is used to conduct a comprehensive evaluation of candidate detection schemes based on multiple indicators, and output the scheme ranking, advantages, disadvantages and risk information;

[0019] The experiment feedback optimization module is used to receive real experimental results, compare the real results with the virtual prediction results, and update the parameters of the digital twin model.

[0020] The results output module is used to output virtual experiment suggestions and optimized detection schemes.

[0021] Furthermore, in the R&D target access module, the user or project-responsible intelligent agent inputs the R&D target into the system, such as "establish a rapid nucleic acid detection method for a certain drug-resistant bacterium" or "develop a highly specific antigen detection reagent for a certain emerging pathogen"; the R&D target access module parses the input content, extracts information such as the detection object, sample type, detection scenario, target performance, cost constraints, available instruments, expected delivery form and compliance requirements, and generates a structured R&D task package;

[0022] The R&D task package is sent to the task orchestrator; the task orchestrator automatically breaks it down into sub-tasks based on the nature of the task, such as target discovery, detection path selection, molecular identification design, reaction system simulation, algorithm modeling, risk assessment, and experimental verification suggestions; each sub-task is accompanied by input data, expected output, constraints, and a range of available tools to ensure that subsequent intelligent agents can collaborate in a unified context.

[0023] Furthermore, the domain knowledge and data foundation module integrates a public basic knowledge base and internal experimental data. The public basic knowledge base includes, but is not limited to, microbial genome databases, protein databases, structural databases, pathway databases, literature databases, drug resistance gene databases, and detection standard and specification databases. The internal experimental data includes experimental batch information in LIMS, experimental records in ELN, historical project plans, real sample verification results, failure cases, and process parameters.

[0024] The system transforms the aforementioned heterogeneous data into a unified knowledge representation. For structured data, it establishes associations between microorganisms, genes, proteins, targets, primers, probes, antibodies, aptamers, reaction conditions, sample types, and performance indicators using entity, relation, and attribute methods. For unstructured literature and experimental records, it extracts key conclusions, experimental conditions, limiting factors, and sources of evidence. All knowledge entries retain their source, time, scope of application, and confidence level so that the agent can distinguish between highly credible evidence and hypotheses to be verified during the subsequent reasoning process.

[0025] Furthermore, in the programmatic integration framework module, standardized tool encapsulation is established for each type of database, software, and internal system, and input parameters, output formats, error handling, log recording, and version information are uniformly defined; the intelligent agent does not need to directly care about the specific calling method of a certain tool, but only needs to make a calling request to the framework according to the task semantics, and the framework completes the permission verification, parameter conversion, execution scheduling, and result normalization.

[0026] For example, the target discovery agent calls BLAST for sequence-specific alignment, AlphaFold DB or PDB to obtain protein structure information, and KEGG to analyze related metabolic pathways; the molecular engineering agent calls PyMOL or Rosetta for binding region analysis and RDKit for small molecule or aptamer candidate structure evaluation; the algorithm modeling agent calls Python, scikit-learn, or statistical analysis tools to complete feature selection, model training, and performance validation. All tool calls are recorded as reproducible virtual experimental steps, thus forming a standardized digital experimental protocol.

[0027] Furthermore, in the multi-agent collaborative module, the agents include a target discovery agent, a detection method design agent, a molecular engineering identification agent, an algorithm modeling agent, an experimental scheme evaluation agent, a compliance and risk review agent, and a project management agent;

[0028] The target discovery agent automatically invokes sequence alignment, conservation analysis, specificity analysis, structural accessibility analysis, and literature evidence retrieval tools based on the R&D task package to screen candidate detection targets from the genes, proteins, metabolic pathways, or surface structures of the target microorganisms. For nucleic acid detection scenarios, the system focuses on analyzing the conservation of the target sequence within the target strain, its differences in closely related species, and the feasibility of primer and probe design. For antigen or immune detection scenarios, the system focuses on analyzing protein expression stability, epitope exposure level, structural stability, and cross-reactivity risk.

[0029] To avoid selecting targets based on a single indicator, a comprehensive scoring mechanism for candidate targets is established:

[0030]

[0031] in, Indicates the first A comprehensive priority score for each candidate target Indicates the candidate target number; The normalized biological relevance score measures the degree of association between the target and the survival, drug resistance, virulence, or classification of the target microorganism. The normalized target specificity score is used to measure the target's ability to distinguish between target microorganisms and closely related species. The normalized evidence support score measures the degree of support for the target from literature, databases, and historical experiments. The normalized experimental feasibility score is used to measure the feasibility of designing primers, probes, antibodies, or aptamers around the target. This represents the normalized cross-reaction risk score; a higher value indicates a greater risk of false detection. , , , and These are the weighting coefficients for each scoring indicator, which are jointly determined by the task scenario, sample type, and project objectives.

[0032] This score is not the final conclusion, but rather the starting point for subsequent digital twin experiments. The system will send high-priority targets, along with their evidence chains, risk points, and recommended detection paths, to the detection method design agent and the identification molecular engineering agent.

[0033] The project's responsible agent maintains the overall R&D goals and constraints, while the other agents perform analyses around their respective sub-tasks and submit structured results to the shared workspace.

[0034] Furthermore, in the multi-agent collaboration module, the task orchestrator assigns tasks based on the matching degree between sub-task requirements and agent capabilities, and the assignment probability is calculated as follows:

[0035]

[0036] in, Indicates the first The sub-task was assigned to the first The probability of each agent; Indicates the first Capability requirement vector for each subtask; Indicates the first A capability profile vector for each intelligent agent; Indicates the first The historical credibility score of each agent in the current project; Indicates the first Current resource usage penalty for each agent; This represents the task assignment temperature coefficient, used to control the concentration of assignment results. This represents the total number of agents that can participate in the assignment. This represents the agent index during the normalization summation process.

[0037] Through the above mechanisms, the system can ensure professional matching while taking into account resource load and historical performance. Different agents communicate by sharing task states, evidence chains, and intermediate files. For example, after the target discovery agent outputs candidate proteins and their surface exposure areas, the molecular engineering agent generates aptamer or antibody candidate schemes accordingly; the detection method design agent designs the detection process by combining sample matrix, reaction conditions, and target detection limits; the algorithm modeling agent uses virtual data and historical data to build an interpretation model; and the evaluation agent performs multi-target comparisons of the complete scheme.

[0038] In the digital twin experimental environment module, the target microorganism is represented as a comprehensive object including genome, proteome, phenotype, drug resistance characteristics, epidemiological information, and sample background; candidate targets are represented as objects including sequence, structure, conservation, specificity, and accessibility; and detection methods are represented as objects including detection path, identification molecules, reaction conditions, instrument platform, interpretation algorithm, and quality control rules.

[0039] During virtual experiments, the system does not directly replace the final real-world verification, but rather performs computable simulations of multiple candidate schemes before the actual experiment. For example, for nucleic acid detection methods, the digital twin environment can simulate primer-probe matching, the impact of target sequence variations, amplification efficiency, the risk of non-specific amplification, and sample inhibition factors. For immunoassay methods, the digital twin environment can simulate antigen epitope exposure, the stability of recognition molecule binding, the risk of cross-reaction, and the signal window. For multi-index joint detection methods, the digital twin environment can simulate the impact of different target combinations on sensitivity and specificity.

[0040] Furthermore, the virtual experiment evaluation module performs a comprehensive evaluation of each candidate solution, avoiding a focus on only a single technical indicator during the research and development process; for the first... The system calculates the multi-objective evaluation value for each candidate detection scheme:

[0041]

[0042] in, Indicates the first The comprehensive evaluation value of each candidate detection scheme; This represents the expected sensitivity index after normalization. This represents the expected specificity index after normalization. This represents a normalized process robustness index, used to measure the stability of the protocol's performance under different sample, batch, and reaction conditions. This represents the normalized implementation cost indicator, including reagent, instrument, consumable, and labor costs; This represents a normalized technical risk indicator, including risks such as cross-reactivity, sample suppression, structural instability, and regulatory restrictions. , , , and These represent the weight coefficients of each evaluation indicator;

[0043] The output includes not only the ranking results, but also the advantages, disadvantages, key risks, and suggested verification experiments for each solution. For solutions with similar scores but different risk types, the system suggests conducting small-scale verifications in parallel to reduce the omission of candidate solutions due to model bias.

[0044] Furthermore, in the experimental feedback optimization module, after the real experiment is completed, the experimenter returns the experimental results to the system via LIMS, ELN, or structured import methods; the feedback data includes sample source, experimental batch, reaction conditions, detection signal, positive / negative interpretation, failure reason, deviation explanation, and manual review conclusion; the real results are compared with the virtual prediction to identify the source of model error and update the parameters, weights, and evidence confidence in the digital twin environment;

[0045] The feedback update process is represented as follows:

[0046]

[0047] in, Indicates the first The parameter set of the digital twin model during round iteration; This represents the set of parameters after this round of feedback updates. Indicates the first The update step size of each iteration; This indicates calculating the gradient over the set of parameters; This indicates the number of real experimental samples used for feedback updates in this round; Indicates the actual experimental sample number; Indicates the first The credibility weight of a real experimental sample; Indicates the first A vector of the true experimental results for each sample; Indicates the system's response to the first... A vector of virtual prediction results for each sample; Indicates the parameter stability constraint coefficient; Represents the initial parameter set; This represents the L2 norm. This update method allows the system to absorb real experimental information while avoiding excessive model shift caused by a small number of abnormal experimental records.

[0048] Through the above technical solution, this invention integrates knowledge, tools, intelligent agents and experimental feedback in microbial detection research and development into a continuous closed loop, so that virtual experiments are no longer isolated computational analysis, but become a digital twin research and development environment that can continuously guide real experiments and constantly self-correct.

[0049] The beneficial effects of this invention are as follows: It integrates a unified, procedural framework with public databases, professional analysis tools, experimental specifications, internal experimental records, and project knowledge required for microbial detection research and development. Furthermore, it utilizes multiple specialized intelligent agents to automate and coordinate stages such as target discovery, detection method design, molecular engineering, algorithm modeling, scheme evaluation, and experimental feedback optimization. This allows for multiple rounds of virtual design, simulation verification, and scheme screening to be completed before actual experiments, thereby improving the efficiency, reliability, and success rate of microbial detection method development. Specific advantages are as follows:

[0050] (1) This invention completes multiple rounds of target screening, detection path comparison, molecular design identification, and scheme risk assessment in a digital twin experimental environment before real experiments, which can significantly reduce low-value trial-and-error experiments. Researchers can prioritize the verification of high-potential schemes recommended by the system, thereby shortening the cycle from the proposal of research and development goals to the determination of candidate schemes.

[0051] (2) The system identifies risks such as cross-reactivity, sample inhibition, structural inaccessibility, excessive cost, and unstable process in advance in a virtual environment, enabling researchers to eliminate obviously infeasible solutions before investing in reagents, samples, and instrument resources, thereby reducing the cost and failure risk of physical experiments. This effect is particularly pronounced in scenarios involving emerging pathogens or scarce sample resources.

[0052] (3) This invention encapsulates public databases, professional software, statistical tools, and internal experimental systems through a programmatic integration framework, enabling R&D resources that were originally scattered across different platforms, formats, and personnel to be uniformly scheduled by an intelligent agent. This approach improves tool reusability and makes the virtual experimental process traceable and reproducible.

[0053] (4) The multi-agent collaboration module organizes professional capabilities such as target discovery, detection method design, identification molecular engineering, algorithm modeling, and risk review into the same task context. Each agent interacts through standardized intermediate results, reducing information loss in manual communication, enabling complex detection R&D tasks to proceed with a clearer workflow and improving cross-professional collaboration efficiency.

[0054] (5) This invention continuously feeds real experimental results back to the digital twin model and corrects virtual prediction bias through a feedback update mechanism. As project data accumulates, the system can gradually learn empirical rules under specific testing platforms, sample types, and institutional experimental conditions, enabling subsequent R&D tasks to obtain higher quality initial suggestions and forming a sustainable R&D closed loop.

[0055] (6) This invention does not select a solution based on a single performance indicator, but comprehensively considers sensitivity, specificity, process robustness, implementation cost, and technical risks. This multi-objective evaluation method is more in line with the requirements of microbial detection products moving from laboratory research and development to practical application, and helps to avoid major defects in early technical solutions during the later transformation stage, thereby improving the comprehensiveness of solution evaluation.

[0056] (7) This invention can be applied to various scenarios such as pathogen nucleic acid detection, antigen detection, drug-resistant bacteria detection, environmental microbial monitoring, food microbial detection, and rapid detection of emerging infectious diseases. The knowledge base, tool packaging, and intelligent agent roles in the system are all scalable and can be continuously upgraded according to new detection technologies, new databases, and new experimental platforms. It has a wide range of applications and strong scalability.

[0057] (8) The system records every tool call, agent decision, virtual experiment input / output, and real experiment feedback, forming a complete R&D process chain. This mechanism is conducive to project review, quality management, regulatory review, and knowledge accumulation, enabling microbial detection R&D to gradually shift from experience-driven to data-driven and process-driven, and enhancing the standardization and auditability of the R&D process. Attached Figure Description

[0058] Figure 1 This is a system architecture diagram of the present invention.

[0059] Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0060] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0061] like Figure 1-2 As shown, this invention provides a digital twin experimental intelligent agent system for microbial detection, including a research and development target access module, a domain knowledge and data foundation module, a programmatic integration framework module, a digital twin experimental environment module, a multi-agent collaboration module, a virtual experiment evaluation module, an experiment feedback optimization module, and a result output module. Each module continuously transmits structured intermediate results around the same detection research and development task. The specific implementation process of this invention is described using the application scenario of "developing a rapid detection method for a certain drug-resistant bacteria." The specific implementation process includes the following:

[0062] (1) System deployment and basic configuration

[0063] A digital twin experimental agent system is deployed within the R&D institution. The system includes a front-end project workbench, a back-end task orchestration service, tool invocation service, knowledge base service, agent operation service, virtual experiment computation service, and experiment feedback service. The project workbench is used to input R&D goals, view virtual experiment processes, review candidate solutions, and import real experimental data. The task orchestration service is responsible for breaking down R&D goals into multiple sub-tasks. The tool invocation service encapsulates BLAST, structural analysis, molecular simulation, statistical modeling, and internal LIMS and ELN interfaces. The knowledge base service stores public database indexes, literature evidence, historical experiments, and testing specifications. The agent operation service is responsible for the reasoning and collaboration of various specialized agents.

[0064] During deployment, the first step is to configure knowledge sources for the target domain, including microbial genome databases, protein structure databases, drug resistance gene databases, literature indexes, testing standard and specification libraries, and historical project data from within the institution. Then, standard input / output templates and access permissions are configured for each type of tool. For example, the input template for sequence alignment tools includes the target sequence, alignment database, similarity threshold, and output fields; the input template for structure analysis tools includes protein IDs, structure files, target regions, and visualization parameters; and the input template for statistical modeling tools includes sample feature tables, label fields, validation methods, and performance metrics.

[0065] (2) R&D target input and task breakdown

[0066] The project leader inputs the research and development goal into the workbench: to develop a rapid detection method for clinical respiratory samples targeting a certain drug-resistant bacterium, requiring the ability to distinguish the target bacterium from common closely related bacteria within a short detection time, and to minimize reliance on expensive instruments. The research and development goal access module transforms this natural language description into a structured task package, which includes fields such as target microorganism, sample type, detection object, expected sensitivity, expected specificity, detection time, cost constraints, and instrument conditions.

[0067] The task orchestrator generates several sub-tasks based on the task package. The target discovery sub-task requires screening for specific and stable nucleic acid or protein targets for the target bacteria; the detection path design sub-task requires comparing the feasibility of routes such as nucleic acid amplification, antigen detection, and aptamer detection; the molecular engineering identification sub-task requires generating primer-probe, antibody, or aptamer schemes around candidate targets; the algorithm modeling sub-task requires establishing signal interpretation and quality control rules; the virtual experiment evaluation sub-task requires comprehensively ranking multiple complete schemes; and the feedback optimization sub-task requires updating model parameters after real experimental data is fed back.

[0068] (3) Implementation of candidate target screening

[0069] The target discovery agent first extracts candidate targets from the genome, proteome, and drug resistance-related databases of the target drug-resistant bacteria. For each candidate target, the system sequentially performs conservation analysis, exclusion analysis of closely related species, literature evidence retrieval, structural accessibility analysis, and experimental feasibility assessment. For sequence targets, the system focuses on examining their degree of conservation across different strains of the target bacteria, as well as their homology among common co-infecting bacteria and closely related bacteria; for protein targets, the system focuses on examining their expression stability, surface exposure areas, and potential cross-recognition risks.

[0070] The system ranks candidate targets according to a comprehensive scoring mechanism in the technical solution. For rapid clinical testing scenarios, specificity and experimental feasibility are given high weights, while cross-reactivity risk is deducted from the score. Targets with higher scores will proceed to the next round of testing pathway design. The system also retains the evidence chain for each target, including database source, literature abstract, comparison results, structural analysis results, and risk statement.

[0071] (4) Co-design of detection methods and recognition molecules

[0072] After receiving high-priority targets, the detection method design agent proposes candidate detection paths based on sample type and instrument constraints. If the target is a conserved and specific nucleic acid sequence, the system prioritizes generating isothermal amplification or fluorescent probe detection schemes; if the target is a stably expressed protein with good epitope exposure, the system generates antibody or aptamer recognition schemes; if a single target has a mutation risk, the system generates a multi-target joint detection scheme.

[0073] The molecular engineering agent identifies candidate molecules based on the detection pathway. For nucleic acid detection, the system generates multiple sets of primers and probes, checking amplified fragment length, dimer risk, melting temperature consistency, and non-specific matching risk. For protein detection, the system generates antibody or aptamer screening suggestions based on epitope regions, spatial accessibility, and conservation, and evaluates binding stability through structural simulation. The output of the molecular engineering agent is fed back to the detection method design agent, which then supplements the reaction conditions, quality control design, and preliminary experimental procedures.

[0074] (5) Multi-agent collaborative execution

[0075] In this embodiment, the task orchestrator assigns tasks based on sub-tasks and agent capability profiles. The target discovery sub-task is led by the target discovery agent, the structural analysis task is assisted by the molecular engineering identification agent, the model interpretation task is led by the algorithm modeling agent, and the risk review task is led by the compliance and risk review agent. The system calculates the participation probability of each agent according to the task assignment mechanism in the technical solution and determines the actual executor based on resource usage.

[0076] Each agent submits standardized results in the shared workspace. The target discovery agent submits a list of candidate targets and a chain of evidence; the molecular engineering identification agent submits candidate primers, probes, or identifying molecules and a risk statement; the detection method design agent submits the detection process, reaction conditions, and quality control settings; the algorithm modeling agent submits signal interpretation rules and a model validation plan; and the risk review agent submits cross-reactivity, regulatory restrictions, and sample suitability risks. The project-leading agent summarizes these results to form a complete candidate solution that can be used in the virtual experimental environment.

[0077] (6) Virtual experiment and scheme evaluation implementation

[0078] The digital twin experimental environment generates virtual experimental objects based on complete candidate schemes. For nucleic acid detection schemes, the system simulates the impact of different strain variations, interference from closely related species, sample inhibition factors, and fluctuations in reaction conditions on amplification efficiency and interpretation results. For protein or aptamer detection schemes, the system simulates target epitope variations, binding stability, nonspecific adsorption, and signal window changes. Each round of virtual experiments records input parameters, tool versions, model versions, output results, and anomaly information.

[0079] The virtual experiment evaluation module calculates the comprehensive evaluation value of candidate schemes according to the multi-objective evaluation mechanism in the technical solution. If a scheme has high expected sensitivity but a high risk of cross-reaction, the system will lower its comprehensive ranking and provide a list of interfering bacteria that need further validation. If a scheme has low cost and high stability, but its detection limit is slightly higher than the target requirement, the system will suggest adjusting the reaction time, the concentration of the identified molecule, or the signal amplification method. Finally, the system outputs the top-ranked schemes to the user, each including the target source, detection path, identified molecule, reaction conditions, expected performance, main risks, and recommended validation experiments.

[0080] (7) Implementation of real experimental feedback

[0081] Researchers select several candidate schemes based on system recommendations to conduct small-scale real-world experiments. After the experiments are completed, the experimental batches, sample information, detection signals, positive / negative interpretations, reasons for failure, and manual review conclusions are imported into the system via LIMS or ELN. The experimental feedback optimization module compares the real results with the virtual prediction results to identify the sources of deviation. For example, if a target is stable in the virtual simulation but shows signal suppression in the real sample, the system will increase the weight of matrix-related risks for that sample; if a primer combination produces non-specific amplification in the real experiment, the system will decrease the experimental feasibility score for the corresponding target or sequence region.

[0082] The system corrects the parameters and evidence confidence levels of the digital twin model according to the feedback update mechanism in the technical solution. The updated model is used for the next round of candidate solution generation and virtual experiment evaluation, enabling the system to gradually accumulate R&D experience for specific institutions, specific sample types, and specific testing platforms. After multiple iterations, R&D personnel can reduce the number of low-value physical experiments and concentrate resources on candidate solutions with a higher probability of success.

[0083] (8) Results output and traceability management

[0084] The system ultimately outputs a complete R&D recommendation report, which includes the results of R&D goal analysis, candidate target ranking, detection path comparison, molecular identification design suggestions, virtual experiment records, scheme evaluation results, risk warnings, recommended experimental verification plans, and subsequent optimization directions. Each conclusion is linked to its data source, tool call records, agent reasoning process, and experimental feedback records.

[0085] Through the above implementation process, those skilled in the art can configure the system according to actual databases, experimental platforms, and project requirements, and apply it to different microbial detection R&D tasks. This implementation method demonstrates that the present invention can not only provide auxiliary analysis for individual R&D stages, but also connect the entire process of target discovery, scheme design, virtual verification, real feedback, and model optimization.

[0086] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A digital twin experimental intelligent agent system for microbial detection, characterized in that, It includes a research and development target access module, a domain knowledge and data foundation module, a programmatic integration framework module, a digital twin experimental environment module, a multi-agent collaboration module, a virtual experiment evaluation module, an experiment feedback optimization module, and a result output module; the modules continuously transmit structured intermediate results around the same detection and research and development task; The R&D target access module is used to receive the microbial detection R&D targets input by the user and parse them to generate a structured R&D task package; The domain knowledge and data foundation module is used to transform heterogeneous knowledge data into a unified knowledge representation, providing a reliable source of knowledge for reasoning. The programmatic integration framework module is used to standardize and encapsulate various external databases and professional tools, and to uniformly handle call requests, parameter conversion, and result normalization. The multi-agent collaboration module is used to assign tasks and exchange results according to the research and development stage. It includes multiple agents with different professional capabilities and a task orchestrator. The target discovery agent among the agents is used to screen candidate detection targets and prioritize them. The task orchestrator assigns tasks based on the matching degree between sub-task requirements and agent capabilities. Each agent performs analysis around the assigned sub-task and submits structured results to the shared work area, realizing multi-agent collaborative reasoning. The digital twin experimental environment module is used to map target microorganisms, candidate targets, detection reaction systems, identification molecules and algorithm models into computable objects to build a virtual experimental environment; The virtual experiment evaluation module is used to conduct a comprehensive evaluation of candidate detection schemes based on multiple indicators, and output the scheme ranking, advantages, disadvantages and risk information; The experiment feedback optimization module is used to receive real experimental results, compare the real results with the virtual prediction results, and update the parameters of the digital twin model. The results output module is used to output virtual experiment suggestions and optimized detection schemes.

2. The digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the R&D target access module, the user or project-responsible intelligent agent inputs the R&D target into the system; the R&D target access module parses the input content, extracts information such as the detection object, sample type, detection scenario, target performance, cost constraints, available instruments, expected delivery form and compliance requirements, and generates a structured R&D task package; The R&D task package is sent to the task orchestrator; the task orchestrator automatically breaks it down into sub-tasks based on the nature of the task, such as target discovery, detection path selection, molecular identification design, reaction system simulation, algorithm modeling, risk assessment, and experimental verification suggestions; each sub-task is accompanied by input data, expected output, constraints, and a range of available tools.

3. The digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, The domain knowledge and data foundation module integrates a public basic knowledge base and internal experimental data. The public basic knowledge base includes, but is not limited to, microbial genome databases, protein databases, structural databases, pathway databases, literature databases, drug resistance gene databases, and detection standard and specification databases. The internal experimental data includes experimental batch information in LIMS, experimental records in ELN, historical project plans, real sample verification results, failure cases, and process parameters. The system transforms the aforementioned heterogeneous data into a unified knowledge representation; For structured data, we establish associations between microorganisms, genes, proteins, targets, primers, probes, antibodies, aptamers, reaction conditions, sample types, and performance indicators using entity, relation, and attribute methods. For unstructured literature and experimental records, we extract key conclusions, experimental conditions, limiting factors, and sources of evidence. All knowledge entries retain their source, time, scope of application, and confidence level.

4. The digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the programmatic integration framework module, standardized tool encapsulation is established for each type of database, software, and internal system, uniformly defining input parameters, output formats, error handling, log recording, and version information. The intelligent agent does not need to directly care about the specific calling method of a certain tool; it only needs to make a calling request to the framework according to the task semantics, and the framework completes the permission verification, parameter conversion, execution scheduling, and result normalization. All tool calling processes are recorded as reproducible virtual experimental steps, thereby forming a standardized digital experimental protocol.

5. A digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the multi-agent collaborative module, the agents include a target discovery agent, a detection method design agent, a molecular engineering identification agent, an algorithm modeling agent, an experimental scheme evaluation agent, a compliance and risk review agent, and a project management agent. The target discovery agent automatically invokes sequence alignment, conservation analysis, specificity analysis, structural accessibility analysis, and literature evidence retrieval tools based on the R&D task package to screen candidate detection targets from the genes, proteins, metabolic pathways, or surface structures of the target microorganisms. For nucleic acid detection scenarios, the system focuses on analyzing the conservation of the target sequence within the target strain, its differences in closely related species, and the feasibility of primer and probe design. For antigen or immune detection scenarios, the system focuses on analyzing protein expression stability, epitope exposure level, structural stability, and cross-reactivity risk. To avoid selecting targets based on a single indicator, a comprehensive scoring mechanism for candidate targets is established: in, Indicates the first A comprehensive priority score for each candidate target Indicates the candidate target number; The normalized biological relevance score measures the degree of association between the target and the survival, drug resistance, virulence, or classification of the target microorganism. The normalized target specificity score measures the target's ability to distinguish between target microorganisms and closely related species. The normalized evidence support score measures the degree of support for the target from literature, databases, and historical experiments. The normalized experimental feasibility score measures the feasibility of designing primers, probes, antibodies, or aptamers around the target. This represents the normalized cross-reaction risk score; a higher value indicates a greater risk of false detection. , , , and These are the weighting coefficients for each scoring indicator, which are jointly determined by the task scenario, sample type, and project objectives. The project's responsible agent maintains the overall R&D goals and constraints, while the other agents perform analyses around their respective sub-tasks and submit structured results to the shared workspace.

6. The digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the multi-agent collaboration module, the task orchestrator assigns tasks based on the matching degree between sub-task requirements and agent capabilities. The assignment probability is calculated as follows: in, Indicates the first The sub-task was assigned to the first The probability of each agent; Indicates the first Capability requirement vector for each subtask; Indicates the first A capability profile vector for each intelligent agent; Indicates the first The historical credibility score of each agent in the current project; Indicates the first Current resource usage penalty for each agent; This represents the task assignment temperature coefficient, used to control the concentration of assignment results. This represents the total number of agents that can participate in the assignment. This represents the agent index during the normalization summation process.

7. The digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the digital twin experimental environment module, the target microorganism is represented as a comprehensive object including genome, proteome, phenotype, drug resistance characteristics, epidemiological information, and sample background; candidate targets are represented as objects including sequence, structure, conservation, specificity, and accessibility; and detection methods are represented as objects including detection path, identification molecules, reaction conditions, instrument platform, interpretation algorithm, and quality control rules. During the virtual experiment, the system does not directly replace the final real verification, but performs computable deductions on multiple candidate solutions before the actual experiment.

8. A digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the virtual experiment evaluation module, each candidate solution is comprehensively evaluated to avoid focusing on only a single technical indicator during the R&D process; for the first The system calculates the multi-objective evaluation value for each candidate detection scheme: in, Indicates the first The comprehensive evaluation value of each candidate detection scheme; This represents the expected sensitivity index after normalization. This represents the expected specificity index after normalization. This represents a normalized process robustness index, used to measure the stability of the protocol's performance under different sample, batch, and reaction conditions. This represents the normalized implementation cost indicator, including reagent, instrument, consumable, and labor costs; This represents a normalized technical risk indicator, including risks such as cross-reactivity, sample suppression, structural instability, and regulatory restrictions. , , , and These represent the weight coefficients of each evaluation indicator; The output includes not only the ranking results, but also the advantages, disadvantages, key risks, and suggested verification experiments for each solution. For solutions with similar scores but different risk types, the system suggests conducting small-scale verifications in parallel to reduce the omission of candidate solutions due to model bias.

9. A digital twin experimental intelligent agent system for microbial detection according to claim 1, characterized in that, In the experimental feedback optimization module, after the real experiment is completed, the experimenter sends the experimental results back to the system via LIMS, ELN, or structured import. The feedback data includes sample source, experimental batch, reaction conditions, detection signal, positive and negative interpretation, failure reason, deviation explanation, and manual review conclusion. The real results are compared with the virtual prediction to identify the source of model error and update the parameters, weights, and evidence confidence in the digital twin environment. The feedback update process is represented as follows: in, Indicates the first The parameter set of the digital twin model during round iteration; This represents the set of parameters after this round of feedback updates. Indicates the first The update step size of each iteration; This indicates calculating the gradient over the set of parameters; This indicates the number of real experimental samples used for feedback updates in this round; Indicates the actual experimental sample number; Indicates the first The credibility weight of a real experimental sample; Indicates the first A vector of the true experimental results for each sample; Indicates the system's response to the first... A vector of virtual prediction results for each sample; Indicates the parameter stability constraint coefficient; Represents the initial parameter set; This represents the L2 norm.

Images