A drug recommendation system and method based on multi-agent cooperation

Abstract

The application relates to the technical field of medical artificial intelligence, and discloses a drug recommendation system and method based on multi-agent cooperation. The system comprises the following steps: a retrieval agent performs structured processing and retrieval enhancement generation on extracted multi-source heterogeneous data to obtain a first drug candidate set related to a target disease; a pharmacological agent performs hierarchical quantitative scoring and reflection-driven rearrangement on the first drug candidate set to obtain a second drug candidate set; a clinical agent performs clinical quantitative scoring and diagnosis and treatment adaptation rearrangement on the first drug candidate set to obtain a third drug candidate set; a coordination agent performs dynamic weighted fusion on the first drug candidate set, the second drug candidate set and the third drug candidate set to obtain a drug comprehensive evaluation score matrix, and performs conflict resolution and precision rearrangement output on the drug comprehensive evaluation score matrix to obtain a drug recommendation result related to the target disease. The application significantly improves the accuracy, explainability and reliability of drug recommendation.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a drug recommendation system and method based on multi-agent collaboration. Background Technology

[0002] Drug prediction and recommendation, as a core research direction at the intersection of biomedical informatics and artificial intelligence, aims to adaptively mine potential treatment options from massive drug data, biomedical literature, and professional knowledge bases using deep neural network models, thereby providing efficient decision support for disease diagnosis and treatment.

[0003] Traditional drug recommendation methods primarily rely on rule-based systems, molecular docking simulations, and early machine learning models. While these provide fundamental support for drug discovery, their limitations in data processing and model expressive capabilities make them ill-suited to the complex, multi-source, heterogeneous nature of biomedical data. In recent years, drug discovery methods based on the fusion of deep learning and knowledge graphs have become the mainstream technological approach. These methods typically integrate multi-source biomedical databases such as DrugBank, ChEMBL, and PubMed to construct large-scale knowledge graphs covering core entities such as drugs, diseases, and targets. They then combine these with graph neural networks, representation learning, or semantic retrieval techniques to predict and uncover potential drug-disease associations. Of particular note are large language models such as the GPT series and LLaMA, which, with their powerful semantic understanding and generation capabilities, have been applied to zero-shot reasoning tasks such as drug relocation and combination therapy design, demonstrating advantages in simplifying model training processes.

[0004] However, existing technologies still have many key limitations, restricting the reliability and practicality of drug recommendation systems in precision medicine scenarios. First, drug recommendation methods based on static knowledge graphs heavily rely on the quality of graph construction and data integrity. Furthermore, the long update cycle of knowledge graphs makes it difficult to incorporate the latest biomedical research findings in a timely manner, resulting in insufficient novelty in the recommendation results. Second, while zero-shot drug reasoning using large language models avoids cumbersome feature engineering, the model reasoning process lacks interpretability and controllability. Its output is essentially a reproduction of existing knowledge in the training data, making it difficult to achieve innovative discoveries beyond the training data distribution. Moreover, due to the inherent characteristics of the model, the output results have a high risk of being illusory or factually incorrect, making direct application impossible. Furthermore, existing single-agent drug recommendation frameworks suffer from several drawbacks. Firstly, they lack functional generalization and specialized role division, failing to provide in-depth analysis of candidate drugs from multiple perspectives, including pharmacology and clinical medicine. Secondly, they lack effective iterative reflection mechanisms to dynamically optimize reasoning logic. Finally, most existing drug recommendation systems operate in an open-loop manner, lacking a closed-loop feedback mechanism between their predictions and real-world medical experimental validation. This prevents the models from self-correcting and continuously learning from experimental data, hindering their ability to adapt to the dynamic needs of personalized medicine and limiting their practical application in real-world clinical settings. Summary of the Invention

[0005] To address the lack of multi-perspective review and self-reflection capabilities in existing drug recommendation models, this invention provides a drug recommendation system and method based on multi-agent collaboration.

[0006] In a first aspect, embodiments of the present invention provide a drug recommendation system based on multi-agent collaboration, comprising:

[0007] The retrieval agent calls a biomedical database through a preset interface and performs structured processing and retrieval enhancement on the extracted multi-source heterogeneous data to obtain a first drug candidate set related to the target disease.

[0008] A pharmacological intelligent agent, based on a dynamically iterative pharmacological rule tree, performs hierarchical quantitative scoring and reflection-driven rearrangement on the first drug candidate set to obtain a second drug candidate set;

[0009] A clinical intelligent agent, which uses counterfactual reasoning and drug relationship modeling to perform clinical quantitative scoring and treatment adaptation rearrangement on the first drug candidate set to obtain a third drug candidate set;

[0010] A coordinating agent dynamically weights and fuses the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix. The coordinating agent then performs conflict resolution and fine-tuning on the comprehensive drug evaluation score matrix to output drug recommendation results related to the target disease.

[0011] Preferably, the method further includes introducing a global self-reflection mechanism, which is used to obtain the optimal drug recommendation result through multiple rounds of iterative execution based on the drug recommendation result;

[0012] The processing steps of the global self-reflection mechanism include:

[0013] The system receives the efficacy score and toxicity score output after the drug recommendation results are verified by medical experiments, and calculates the deviation value between the overall drug superiority and inferiority prediction results output by the coordinating agent and the standardized medical experiment results based on the efficacy score and the toxicity score.

[0014] The deviation value is compared with a threshold. If the deviation value exceeds a preset threshold, the parameters of the pharmacological agent, the clinical agent, and the coordination agent are optimized.

[0015] Preferably, the retrieval agent comprises a first configuration module, a first memory module, and a first execution module;

[0016] The first configuration module calls a biomedical database through a preset interface and extracts multi-source heterogeneous data related to the target disease.

[0017] The first memory module performs vector embedding on the unstructured text in the multi-source heterogeneous data to obtain a vector index library, and performs entity recognition and relation extraction on the multi-source heterogeneous data to obtain medical knowledge triples.

[0018] The first execution module performs a local search on the vector index library to obtain semantically similar data blocks, and summarizes the global search results for the medical knowledge triples and the semantically similar data blocks to generate a first drug candidate set related to the target disease.

[0019] Preferably, the pharmacological intelligent agent comprises a second configuration module, a second memory module, and a second execution module;

[0020] The second configuration module injects pharmacological assessment role prompts and initiates a rule tree structure generation task for the target disease;

[0021] The second memory module responds to the rule tree structure generation task by constructing a pharmacological rule tree based on the node set and leaf index candidates output by the stored large language model.

[0022] The second execution module performs hierarchical quantitative scoring on the first drug candidate set based on the pharmacological rule tree to obtain the pharmacological score of each drug, and corrects the drug ranking priority determined by each pharmacological score through a local self-reflection mechanism to obtain the second drug candidate set. The local self-reflection mechanism is configured to correct the structure and node weights of the pharmacological rule tree based on medical experiment feedback.

[0023] Preferably, the clinical intelligent agent consists of a third configuration module, a third memory module, and a third execution module;

[0024] The third configuration module injects clinical assessment role prompts to determine the format of clinical input data and counterfactual reasoning rules.

[0025] The third memory module stores structured real case data conforming to the clinical input data format and counterfactual simulation samples generated by the stored large language model according to the counterfactual reasoning rules.

[0026] The third execution module performs clinical quantitative scoring on the first drug candidate set based on the structured real case data and the counterfactual simulation samples to obtain a comprehensive clinical score for each drug, and prioritizes the first drug candidate set according to each comprehensive clinical score to obtain a third drug candidate set.

[0027] Preferably, the coordinating agent consists of a fourth configuration module and a fourth execution module;

[0028] The fourth configuration module injects collaborative decision-making role prompts and sets conflict resolution rules related to the multi-agent multi-dimensional evaluation results.

[0029] The fourth execution module uses adaptive weights output by the stored large language model to perform weighted fusion on the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix. It then uses a pre-trained discriminant model to infer the comprehensive drug evaluation score matrix to obtain the comprehensive drug merit score for each drug. Finally, it combines the conflict resolution rules to perform fine-grained optimization on the comprehensive drug merit score for each drug to obtain drug recommendation results related to the target disease.

[0030] Preferably, the step of receiving the efficacy score and toxicity score output after the drug recommendation result has been verified by medical experiments, and calculating the deviation value between the overall drug superiority / inferiority prediction result output by the coordinating agent and the standardized medical experiment result based on the efficacy score and the toxicity score, includes:

[0031] The raw efficacy data and raw toxicity data output from the medical experiment are standardized to obtain efficacy scores and toxicity scores, and the toxicity scores are then converted into toxicity acceptable indicators.

[0032] The overall drug superiority / inferiority prediction results, efficacy scores, and toxicity acceptable indicators output by the coordinating agent are weighted and processed to obtain the deviation value between the overall drug superiority / inferiority prediction results and the standardized medical experiment results.

[0033] Preferably, the step of comparing the deviation value with a threshold, and if the deviation value exceeds a preset threshold, then optimizing the parameters of the pharmacological agent, the clinical agent, and the coordinating agent, includes:

[0034] Determine whether the deviation value exceeds a preset threshold. If so, update the pharmacological rule tree of the pharmacological agent and the counterfactual simulation sample of the clinical agent based on the deviation value.

[0035] Based on the adjusted pharmacological agent and the clinical agent, the adaptive weights of the coordinating agent are updated.

[0036] Secondly, embodiments of the present invention provide a drug recommendation method based on multi-agent cooperation, applied to the drug recommendation system based on multi-agent cooperation as described above, the method comprising:

[0037] The retrieval agent uses a preset interface to call a biomedical database and performs structured processing and retrieval enhancement on the extracted multi-source heterogeneous data to generate a first drug candidate set related to the target disease.

[0038] Using the pharmacological agent based on a dynamically iterative pharmacological rule tree, the first drug candidate set is subjected to hierarchical quantitative scoring and reflection-driven reordering to obtain the second drug candidate set;

[0039] Using the clinical agent to model drug relationships through counterfactual reasoning, the first drug candidate set is clinically quantitatively scored and re-ranked for diagnosis and treatment fit to obtain a third drug candidate set.

[0040] The coordinating agent dynamically weights and fuses the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix. Conflict resolution and fine-tuning of the comprehensive drug evaluation score matrix are then performed on the output to obtain drug recommendation results related to the target disease.

[0041] Preferably, after the step of using the coordinating agent to dynamically weight and fuse the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix, and performing conflict resolution and fine-tuning on the comprehensive drug evaluation score matrix to obtain drug recommendation results related to the target disease, the method further includes:

[0042] A global self-reflection mechanism is introduced, and the optimal drug recommendation result is obtained by iteratively executing the global self-reflection mechanism through multiple rounds based on the drug recommendation result.

[0043] The processing steps of the global self-reflection mechanism include:

[0044] The system receives the efficacy score and toxicity score output after the drug recommendation results are verified by medical experiments, and calculates the deviation value between the overall drug superiority and inferiority prediction results output by the coordinating agent and the standardized medical experiment results based on the efficacy score and the toxicity score.

[0045] The deviation value is compared with a threshold. If the deviation value exceeds a preset threshold, the parameters of the pharmacological agent, the clinical agent, and the coordination agent are optimized.

[0046] Compared with existing technologies, the drug recommendation system and method based on multi-agent collaboration of this invention has the following advantages: By performing structured processing and retrieval enhancement on multi-source heterogeneous data in biomedical databases through retrieval agents, candidate drugs related to target diseases can be efficiently screened from massive amounts of data, constructing a comprehensive and accurate first drug candidate set, laying a data foundation for subsequent evaluation; the dynamic iterative pharmacological rule tree of the pharmacological agent decomposes complex pharmacological mechanisms layer by layer into quantifiable indicators, continuously optimizing the tree structure and node weights under the drive of medical experiment feedback, achieving pharmacological evaluation that balances accuracy and interpretability; the clinical agent relies on counterfactual reasoning and drug relationship modeling driven by a large language model, making clinical evaluation more consistent with real diagnosis and treatment logic; the dynamic weighted fusion and conflict resolution mechanism of the coordinating agents effectively integrates the differentiated evaluation results of multiple agents, improving the robustness of drug recommendation ranking. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the structure of a drug recommendation system based on multi-agent collaboration according to an embodiment of the present invention;

[0048] Figure 2 This is another structural schematic diagram of a drug recommendation system based on multi-agent collaboration according to an embodiment of the present invention;

[0049] Figure 3 This is a flowchart illustrating a drug recommendation method based on multi-agent collaboration according to an embodiment of the present invention.

[0050] Figure 4 This is another schematic diagram of a drug recommendation method based on multi-agent collaboration according to an embodiment of the present invention. Detailed Implementation

[0051] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0052] In the description of this invention, it should be understood that the terms "first" and "second," etc., are used to distinguish different objects, rather than to describe a specific order.

[0053] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by those skilled in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0054] It should be noted that the intelligent agent described in this invention refers to an independent intelligent unit with role-based positioning and modular architecture, which can be implemented through software programs, hardware chips, or a combination of software and hardware. Each intelligent agent independently performs data processing in its corresponding stage, and forms an intelligent collaborative system through a collaborative mechanism to support the accuracy and reliability of drug recommendations.

[0055] like Figure 1 The diagram shown is a structural schematic of a drug recommendation system based on multi-agent collaboration according to an embodiment of the present invention. (Refer to...) Figure 1 An embodiment of the present invention provides a drug recommendation system based on multi-agent collaboration, comprising:

[0056] The retrieval agent calls a biomedical database through a preset interface and performs structured processing and retrieval enhancement on the extracted multi-source heterogeneous data to generate a first drug candidate set related to the target disease.

[0057] Specifically, the retrieval agent consists of a first configuration module, a first memory module, and a first execution module.

[0058] The first configuration module calls a biomedical database through a preset interface and extracts multi-source heterogeneous data related to the target disease.

[0059] Employing RESTful APIs or JDBC adapters, the system connects to mainstream biomedical databases such as DrugBank and PubMed, enabling precise extraction of multi-source heterogeneous data related to target diseases through interface parameter configuration. Specifically, this multi-source heterogeneous data includes structured fields such as drug chemical structures, target information, indications, and toxicity data from the databases; unstructured text data such as abstracts, experimental methods, and research conclusions from medical literature; and semi-structured data such as database annotations and literature metadata.

[0060] Based on the characteristics of the acquired data source, an LLM hint template is constructed that includes prior information such as agent roles, data source credibility descriptions, and time windows, as well as retrieval constraints. Then, the extracted raw data is preprocessed.

[0061] Specifically, it calls open-source large models such as Llama 3.2 and Qwen3 to process the original fields of the drug database. ( For the first The original fields are expanded and reconstructed, redundant fields are merged, and auxiliary fields such as disease treatment scope and toxicological risk level are generated to obtain a complete drug entry. Simultaneously, a sliding window block algorithm is used to divide the literature collection. And the database's lengthy text is segmented into semantically complete paragraph blocks. To adapt to the requirements of large model context windows.

[0062] The first memory module performs vector embedding on unstructured text in multi-source heterogeneous data to obtain a vector index library, and performs entity recognition and relation extraction on multi-source heterogeneous data to obtain medical knowledge triples.

[0063] Specifically, the first memory module employs a two-layer memory storage mechanism:

[0064] One layer generates low-dimensional vectors from the segmented text and database records using a sentence vector representation model, and stores them in FAISS or Milvus vector databases to build a vector index library. The vector index library is periodically updated with new documents to ensure timeliness.

[0065] Another layer utilizes entity recognition and relation extraction components to extract from each data block. Formal medical knowledge triples ( The head is a physical entity, usually the name of a disease. For relationships, such as "can be treated" or "acts on". (The tail entity is usually a drug or target), and all triples are stored in a Neo4j or DGraph graph database to form a knowledge graph. (node ​​set) Includes diseases, drugs, targets, toxic agents, etc., and is used in edge collections. It represents the relationships between entities, and the knowledge graph retains the data source to ensure traceability.

[0066] The first execution module performs a local search of the vector index library to obtain semantically similar data blocks, and then summarizes the global search results for medical knowledge triples and semantically similar data blocks to generate a first drug candidate set related to the target disease.

[0067] The first execution module uses a search-enhanced generation (RAG) mechanism to screen candidate drugs. Specifically, it first queries the disease... Encoded as 3D feature vector The correlation between disease queries and drug vectors in the vector index is calculated using the cosine similarity formula, as shown below:

[0068]

[0069] in, Represents disease query vector Drug vectors in the vector index library Local semantic similarity score. Retain. ( (To set a threshold) semantically similar data blocks.

[0070] Then, a global search is performed based on the knowledge graph, and the shortest path hop count is calculated using the following formula:

[0071]

[0072] in, Indicates from disease node To drug node All reachable paths The minimum number of jumps included is used to obtain the global score, as shown below:

[0073]

[0074] in, Represents disease query vector Drug vectors in the vector index library The global structural similarity score.

[0075] The comprehensive search score is obtained through a weighted fusion formula:

[0076]

[0077] in, Represents disease query vector Drug vectors in the vector index library The overall search score, This indicates the adjustment parameter.

[0078] After sorting all drugs by score, select the top-scoring drugs. The drugs selected form the first drug candidate set:

[0079]

[0080]

[0081] in, This represents the first set of drug candidates related to the target disease. This indicates the ranking of the first drug candidate set based on comprehensive search scores. The candidate set is then summarized and output by calling LLM, which combines the prompt template from the first configuration module with the drug in position.

[0082] The pharmacological intelligent agent, based on a dynamically iterative pharmacological rule tree, performs hierarchical quantitative scoring and reflection-driven reordering on the first drug candidate set to obtain the second drug candidate set.

[0083] Specifically, the pharmacological intelligent agent consists of a second configuration module, a second memory module, and a second execution module.

[0084] The second configuration module injects pharmacological assessment role prompts and initiates a rule tree structure generation task for the target disease.

[0085] The second configuration module first injects pharmacological assessment role prompts into open-source large models such as Llama3.2 and Qwen3, clarifying that the main responsibility of the pharmacological agent is to construct a hierarchical and quantifiable drug assessment rule system based on the pathological mechanism of the disease. Then, it initiates a rule tree structure generation task, inputting task instructions into the large model through a preset instruction template: "Based on the pathological characteristics of the target disease, list the core pharmacological intervention mechanisms (such as inflammation regulation, metabolic pathway regulation, etc.), and construct a pharmacological rule tree according to the hierarchical structure of 'general criteria - primary mechanism - secondary sub-mechanism - quantifiable index', and output the node set and leaf index candidates." Based on the pharmacological domain knowledge stored inside the large model, it provides the core input for the initial rule tree construction.

[0086] It should be noted that, due to the lack of sufficient structured historical data in the initial stage, the knowledge within the large model was the main source of data.

[0087] The second memory module responds to the rule tree structure generation task by constructing a pharmacological rule tree based on the node set and leaf index candidates output by the stored large language model.

[0088] The second memory module receives the node set (containing overall criteria and pharmacological mechanism nodes at all levels) and leaf index candidates (quantifiable experimental indicators) from the output of the large model, and constructs a structured pharmacological rule tree according to the hierarchical relationship, formalized as follows:

[0089]

[0090] in, Representing a pharmacological rule tree, Represents a set of nodes, including the root node. (General principle: "Positive intervention effect of drugs on the pathological mechanism of disease"), internal nodes (Pharmacological mechanisms at various levels, such as "regulation of inflammatory response" and "inhibition of apoptosis pathway") leaf nodes (Quantifiable experimental indicators, such as "IL-6 inhibition rate" and "receptor binding affinity") It represents the set of parent-child relationships between nodes, and defines the subordinate relationships of nodes at each level.

[0091] During the construction of the pharmacological rule tree, the logic is refined through a recursive expansion mechanism:

[0092]

[0093] in, This indicates the upper-level rules. This represents the refined lower-level rules. Each lower-level rule corresponds to a specific experimental indicator of the drug. For example, if the upper-level rule is that the drug needs to affect the inflammatory response of the disease, then the lower-level rule might be that the drug's inhibition rate on a specific cytokine exceeds a threshold. The upper-level coarse-grained mechanism is decomposed layer by layer into lower-level fine-grained computable indicators, forming a total score hierarchical evaluation system. At the same time, the rule tree structure and the initial node weights (output by the large model in JSON format according to domain common sense) are stored in the module's built-in database to provide structural support for subsequent scoring.

[0094] The second execution module performs hierarchical quantitative scoring on the first drug candidate set based on the pharmacological rule tree to obtain the pharmacological score of each drug, and corrects the drug ranking priority determined by each pharmacological score through a local self-reflection mechanism to obtain the second drug candidate set.

[0095] The second execution module completes the quantitative scoring according to the hierarchical logic of "leaf node → internal node → root node".

[0096] The following is a detailed explanation of the hierarchical logic described above:

[0097] 1) Calculate the leaf node score:

[0098] For drugs Leaf node indicators The score As shown below:

[0099]

[0100] in, Indicates drug Leaf node indicators The observed values ​​are categorized into three sources: first, database fields, obtained through automatic extraction and standardization, such as activity values ​​reported in PubChem / ChEMBL; second, literature extraction, obtained through natural language processing (NLP) extraction and structuring, such as "IL-6 reduction of 65%" being extracted as an inhibition rate of 0.65; and third, experimental input, obtained through in vivo / in vitro medical experiments. When the same indicator has multiple sources, the priority order and application are based on medical experimental input > database fields > literature extraction. This indicates the maximum reference value for the leaf node index.

[0101] 2) Calculate the scores of internal nodes:

[0102] The drug is calculated using a weighted summation formula. Internal node metrics The score :

[0103]

[0104] in, Represents child nodes The weights satisfy , Indicates internal nodes The number of child nodes at the next level.

[0105] 3) Calculate the root node score:

[0106] The final score is based on the root node. As a medicine The pharmacological score.

[0107] Scoring and ranking based solely on the initial rule tree has limitations. To improve the accuracy and clinical relevance of pharmacological assessments, a local self-reflection mechanism is introduced into the pharmacological agent. This mechanism is configured to modify the structure and node weights of the pharmacological rule tree based on feedback from medical experiments.

[0108] Specifically, after initial ranking based on pharmacology scores, a local self-reflection mechanism is initiated: receiving unstructured text feedback from medical experiments. (Including efficacy and toxicity data from in vitro / in vivo experiments), With the current rule tree weight Task instructions ("Based on {experimental results} and the current weights, output a new weight vector (JSON format) and determine whether to add / prune nodes") Input multiple large open-source models, use a multi-expert voting mechanism to determine the update result, and obtain the corrected rule tree. With node weights Formalized as:

[0109]

[0110] The pharmacology scores are recalculated based on the updated rule tree, the drug ranking priorities are adjusted, and the top-ranked drugs are finally output. One drug, forming a second drug candidate set. The value needs to be determined by combining the evaluation accuracy of the pharmacological rule tree with the reliability of medical experiment feedback, and must satisfy... ( (This is the screening threshold for the first drug candidate set).

[0111] It should be noted that, The initial values ​​are generated by the open-source large model during the pharmacological rule tree construction phase. Pre-defined task instructions guide the large model to combine the pathological context of the target disease with common knowledge in the field of pharmacology, outputting an initial weight vector in JSON format. When the local self-reflection mechanism triggers rule tree structure updates based on feedback from medical experiments, Iterative corrections will be made simultaneously, and the weight update will immediately trigger the recalculation and sorting adjustment of the hierarchical quantitative scores of the first drug candidate set to ensure that the pharmacology scores are consistent with the updated rule tree evaluation system.

[0112] The clinical agent uses counterfactual reasoning and drug relationship modeling to perform clinical quantitative scoring and treatment adaptation rearrangement on the first drug candidate set to obtain the third drug candidate set.

[0113] Specifically, the clinical intelligent agent consists of a third configuration module, a third memory module, and a third execution module.

[0114] The third configuration module injects clinical assessment role prompts to determine the format of clinical input data and counterfactual reasoning rules.

[0115] The third configuration module uses open-source large models such as Llama3.2 and Qwen3 as semantic and inference engines. First, it injects clinical assessment role prompts into the large models, clarifying their main responsibility as evaluating the clinical efficacy, safety, and suitability of candidate drugs based on real-world clinical scenarios. Then, it defines a standardized clinical input data format, specifying that it must include structured fields such as patient baseline characteristics (age, gender), medical history information, laboratory indicators (such as biochemical indicators and imaging results), comorbidities, and previous medication records to ensure the standardization and completeness of the input data. At the same time, it formulates counterfactual inference rules through meta-configuration files, covering the definition of "monotherapy / combination / alternative therapy" scenarios, quantitative standards for clinical outcome variables (survival time, symptom relief rate, adverse reaction grade, etc.), and constraints for counterfactual sample generation (such as consistency requirements with the distribution of real case characteristics). The rules support dynamic iterative optimization based on feedback from subsequent medical experiments.

[0116] The third memory module stores structured real case data that conforms to the format of clinical input data and counterfactual simulation samples generated by the stored large language model according to the counterfactual reasoning rules.

[0117] The third memory module stores structured real-world case data, which comes from electronic medical records (EHR), real-world data (RWD), and clinical trial records. After data cleaning and de-identification, the data is structured and stored according to the input format defined by the third configuration module, forming a historical patient-drug response sample library, which is used to simulate drug effects in real-world medical scenarios.

[0118] The third memory module stores counterfactual simulation samples, generated by an open-source large model based on counterfactual reasoning rules and internal clinical domain knowledge. It generates scenarios where medication use is scarce in real cases (such as medication use under specific complication combinations), indicating "if candidate drugs are used..." / Unused "If replaced with medicine" Counterfactual scenario samples, such as those for real cases, are used to address the issue of case sparsity. All generated samples must meet the consistency constraint with the distribution of characteristics of real cases to ensure the reliability of the simulation. Together, these two types of data provide comprehensive scenario support for clinical quantitative scoring.

[0119] The third execution module performs clinical quantitative scoring on the first drug candidate set based on structured real case data and counterfactual simulation samples to obtain a comprehensive clinical score for each drug, and prioritizes the first drug candidate set according to each comprehensive clinical score to obtain the third drug candidate set.

[0120] Specifically, the net clinical benefit of the drug is calculated through counterfactual reasoning, assuming a patient state vector. for:

[0121]

[0122] in, Indicates clinical characteristics. Clinical outcome variables. (e.g., symptom relief rate) for candidate drugs The outcomes after drug use were simulated respectively. Outcomes compared to no medication Net benefit is defined as:

[0123]

[0124] in, This represents the patient's state vector. When using candidate drugs Compared to the net benefit of not using the drug, It represents the mathematical expectation.

[0125] In clinical practice, complex interactions exist between drugs, including synergistic, antagonistic, and neutral effects. Furthermore, constructing a drug-drug relationship map is crucial. ( A collection of drugs, (For drug-drug relationships), through large-scale model analysis of medical literature and clinical trial data, synergistic / antagonistic relationships between drugs are extracted, and the synergistic gain is calculated as follows:

[0126]

[0127] in, This represents the patient's state vector. At that time, the drug With drugs The combined use of these technologies provides a synergistic gain compared to using either of them individually. When When the combined effect is better than the sum of the effects of single drugs, it indicates a synergistic effect; when When the combined effect is less than the sum of the effects of the single drugs, it indicates an antagonistic effect. When two or more drugs are used in combination, the effect is extended to the calculation of the combined gain of multiple drugs.

[0128] Furthermore, the benefits of monotherapy, combination therapy, and alternative therapy were evaluated separately:

[0129] 1) Monotherapy: Evaluate the net benefit of each candidate drug. ;

[0130] 2) Combination therapy, considering drug interactions The overall net benefit is:

[0131]

[0132] in, This represents the patient's state vector. At that time, the drug With drugs The overall net benefit of combined treatment.

[0133] 3) Alternative therapy: If a drug cannot be used due to contraindications in the first drug candidate set, the counterfactual benefit difference of similar drugs is calculated:

[0134]

[0135] in, This represents the patient's state vector. When using candidate drugs Alternative contraindications drugs The counterfactual effect is poor.

[0136] Finally, the clinical composite score is calculated using the comprehensive scoring formula:

[0137]

[0138] in, This represents the patient's state vector. At that time, candidate drugs Clinical composite score, and Represents the weighting coefficients, satisfying The first set of drug candidates is sorted from highest to lowest score, and the top-scoring candidates are selected. These drugs form a third drug candidate set. It should be noted that... The value needs to be determined by combining the assessment accuracy of the clinical agent and the size of the first drug candidate set, and must meet the following requirements: ( (This is the screening threshold for the first drug candidate set).

[0139] The coordinating agent dynamically weights and fuses the first, second, and third drug candidate sets to obtain a comprehensive drug evaluation score matrix. It then resolves conflicts and refines the output of the comprehensive drug evaluation score matrix to obtain drug recommendation results related to the target disease.

[0140] Specifically, the coordinating agent consists of a fourth configuration module and a fourth execution module.

[0141] The fourth configuration module injects collaborative decision-making role prompts and sets conflict resolution rules related to the multi-agent multi-dimensional evaluation results.

[0142] Specifically, the fourth configuration module injects collaborative decision-making role prompts into the stored open-source large models (such as Llama3.2, Qwen3, etc.), defining its main responsibility as integrating the multi-dimensional evaluation results of the retrieval agent, pharmacological agent, and clinical agent, resolving cross-perspective conflicts, and outputting drug recommendation schemes that take into account scientific validity, feasibility, and safety.

[0143] Furthermore, the fourth configuration module sets hierarchical conflict resolution rules based on the conflict characteristics of multi-agent evaluation (such as "significant advantages in pharmacological mechanism but unacceptable clinical toxicity" and "high data correlation but insufficient pharmacological evidence"). These rules include: priority constraint rules (clinical safety evaluation results have higher priority than pharmacological mechanism evaluation, and pharmacological mechanism evaluation has higher priority than data correlation evaluation), threshold calibration rules (hard thresholds are set for key indicators such as toxicity and efficacy; drugs exceeding the threshold have their ranking priority reduced), and experimental feedback adaptation rules (when the medical experimental verification results deviate significantly from the evaluation of a certain agent, the weight of that experimental feedback in conflict determination is increased). All rules are stored in a structured format, providing a clear basis for the fine-tuning optimization of the fourth execution module.

[0144] The fourth execution module uses adaptive weights output by the stored large language model to perform weighted fusion on the first, second, and third drug candidate sets to obtain a comprehensive drug evaluation score matrix. The pre-trained discriminant model infers the comprehensive drug evaluation score matrix to obtain the comprehensive drug merit score of each drug. The module then combines conflict resolution rules to perform fine-grained optimization of the comprehensive drug merit score of each drug to obtain drug recommendation results related to the target disease.

[0145] The processing procedure of the fourth execution module is explained in detail below:

[0146] First, obtain the adaptive weights and the initial weights. , , (These correspond to the fusion weights of the retrieval agent, pharmacological agent, and clinical agent, respectively.) Based on prior knowledge, in each subsequent iteration, the unstructured text from the medical experiment feedback will be used... With current weight Task instructions (Based on {experimental results} and the current agent score distribution) Output new fusion weights Input a large model and dynamically update the weights:

[0147]

[0148] Then, the scores of the three agents are weighted and fused based on the updated weights, assuming the drug... Search score Pharmacological score Clinical score The single-drug fusion score is calculated using the following formula:

[0149]

[0150] in, Indicates drug The individual drug fusion scores are used to calculate the fusion scores of all drugs. The fusion scores of all drugs constitute the comprehensive drug evaluation score matrix, which is a two-dimensional numerical matrix formed by each drug in the candidate drug set as a row and the individual drug fusion scores as columns.

[0151] To handle potential conflicts between agents, a hierarchical optimization framework is introduced to coordinate the agents. Let each drug... corresponding feature vector for:

[0152]

[0153] The coordinating agent estimates the overall superiority or inferiority of drugs across multiple dimensions by training a discriminant model. The discriminant model... A multilayer perceptron or a graph-based aggregation network can be used, and it can be trained with the following optimization objectives:

[0154]

[0155] in, This represents the set of candidate drugs participating in the comprehensive evaluation. Indicates drug The actual efficacy feedback value obtained after verification by in vitro / in vivo medical experiments is mapped to a discrete score in the range of [0, 0.5, 1] ​​by setting an efficacy threshold, where 0 corresponds to ineffective, 0.5 corresponds to partially effective, and 1 corresponds to completely effective. The output represents the discrete score of the drug under multidimensional indicators, that is, the overall drug quality score.

[0156] Finally, by combining the conflict resolution rules of the fourth configuration module, the overall merit score is finely ranked and optimized (e.g., drugs with "high pharmacological scores but clinical toxicity exceeding the threshold" are downgraded), and the optimal drug recommendation result is finally output, realizing the effective integration and conflict resolution of multi-agent evaluation results.

[0157] This invention employs a hierarchical collaborative agent architecture, constructing a multi-stage iterative drug recommendation pipeline of "retrieval and mining - professional analysis - collaborative integration" to address the issues of accuracy, applicability, and interpretability of drug recommendations within a closed loop of multi-source heterogeneous biomedical data and medical experimental feedback. The top layer is the retrieval agent, which, relying on semi-structured drug databases and unstructured scientific literature repositories, mines potential drug candidates from massive resources as needed. The middle layer is a parallel professional processing layer, encompassing pharmacological and clinical agents, which refine the initial candidate set from professional perspectives of pharmacological mechanism analysis and clinical efficacy simulation, respectively. The bottom layer is the coordination agent, responsible for integrating the outputs of multiple agents and completing the multi-dimensional evaluation and ranking of the drug candidate set. The system execution flow is as follows: the retrieval agent receives the target disease query and outputs a preliminary candidate set, which is simultaneously distributed to the pharmacological and clinical agents for parallel processing. The rearranged lists and scores from both are then input into the coordination agent for fusion and fine-tuning, ultimately generating the recommendation results.

[0158] It should be noted that although the agent framework belongs to a generalized known paradigm, this invention effectively improves the professional adaptability and iterative evolution capability of drug recommendation in three specific implementation details: the hierarchical collaborative mechanism of professional role division, the quantitative evaluation logic of pharmacological rule tree, and the experiment-driven adaptive weight update strategy.

[0159] To improve the alignment between drug recommendation results and real medical experiments and to achieve continuous iterative optimization of the system's recommendation capabilities, this invention introduces a global self-reflection mechanism in addition to coordinating the output of drug recommendation results by the intelligent agent. Based on the drug recommendation results, the optimal drug recommendation results are obtained by iteratively executing the global self-reflection mechanism in multiple rounds.

[0160] like Figure 2 The diagram shown is another structural schematic of a drug recommendation system based on multi-agent collaboration according to an embodiment of the present invention. The processing procedure of the global self-reflection mechanism in this embodiment includes the following steps:

[0161] 1) Receive the efficacy score and toxicity score output after the drug recommendation results are verified by medical experiments, and calculate the deviation value between the overall drug superiority and inferiority prediction results output by the coordinating agent and the standardized medical experiment results based on the efficacy score and toxicity score.

[0162] Specifically, the raw efficacy data and raw toxicity data output from medical experiments are standardized to obtain efficacy scores and toxicity scores, and the toxicity scores are then converted into acceptable toxicity indicators.

[0163] Based on the raw data types of medical experiments (such as in vitro cell inhibition rate and in vivo animal model survival improvement rate), the raw efficacy data are mapped to discrete values ​​in the range [0, 0.5, 1] ​​(0 corresponds to ineffective, 0.5 corresponds to partially effective, and 1 corresponds to completely effective), thus obtaining a standardized efficacy score. The raw toxicity data (such as adverse reaction grades and organ damage severity) are mapped to discrete values ​​in the range [0, 0.5, 1] ​​(0 corresponds to no toxicity, 0.5 corresponds to mild toxicity, and 1 corresponds to severe toxicity), thus obtaining a standardized toxicity score. .

[0164] Through formula The toxicity score is then converted inversely into an acceptable toxicity index. The closer a value is to 1, the lower the risk of clinical toxicity and the more acceptable the drug is.

[0165] Furthermore, the overall drug superiority / inferiority prediction results, efficacy scores, and toxicity acceptable indicators output by the coordinating agent are weighted and processed to obtain the deviation value between the overall drug superiority / inferiority prediction results and the standardized medical experiment results.

[0166] The difference between the predicted and experimental results is quantified using a bias formula:

[0167]

[0168] in, The drug that coordinates the output of the intelligent agent Based on the overall advantages and disadvantages prediction score, Indicates an acceptable level of toxicity. , Weighting coefficients representing efficacy and toxicity (satisfying) For example, in clinical settings , (To emphasize therapeutic efficacy).

[0169] Calculated using this formula That is, medicine The prediction-experimental bias value indicates that the larger the value, the greater the deviation between the prediction result and the experimental facts.

[0170] 2) Compare the deviation values ​​with threshold values. If the deviation value exceeds the preset threshold, optimize the parameters of the pharmacological agent, clinical agent, and coordination agent.

[0171] Specifically, it determines whether the deviation value exceeds a preset threshold. If so, it updates the pharmacological rule tree of the pharmacological agent and the counterfactual simulation sample of the clinical agent based on the deviation value.

[0172] Set deviation threshold If the drug deviation value If the sample is abnormal, it is marked as an anomaly, and the deviation is formalized into a text description, such as for drugs. The overall superiority / inferiority prediction score deviates from the experimental efficacy / toxicity results by more than a threshold, failing to meet clinical application expectations. In this embodiment, a preset threshold is set. It can be adapted and adjusted according to the experimental accuracy requirements.

[0173] Input the text into the main model of the pharmacology agent and prompt it to update the pharmacology rule tree (such as pruning leaf nodes that do not match experimental results, adding uncovered pharmacological mechanism evaluation nodes, and adjusting the weights of the corresponding nodes); at the same time, input the text into the main model of the clinical agent and prompt it to optimize the generation rules of counterfactual simulation (such as adjusting the simulation parameters of "drug-patient toxicity reaction"), update the counterfactual simulation sample library, and improve the accuracy of subsequent clinical evaluation.

[0174] Furthermore, based on the pharmacological agent and clinical agent with adjusted parameters, the adaptive weights of the coordinating agent are updated.

[0175] The update differences between the pharmacological agent (rule tree structure / weight changes) and the clinical agent (counterfactual simulation parameter adjustments) are formalized into text descriptions and input into the main model of the coordinating agent; combined with the task instruction "Adjust the retrieval agent according to the update differences between the pharmacological and clinical agents (…)" ), pharmacological intelligent agents ( ), clinical intelligent agent ( The "fusion weights" are dynamically updated by the large model to enhance the weight ratio of agents that are more closely matched with the experimental results, ensuring that the subsequent multi-agent fusion results are more in line with the facts of medical experiments.

[0176] It should be noted that after the above parameters are optimized, the system enters the next round of iterative recommendation process, forming a closed-loop iterative mechanism of recommendation results - medical experiment - bias calculation - agent parameter optimization - new round of recommendation. Its iterative relationship can be formalized as follows:

[0177]

[0178] in, Indicates the first The results of the drug recommendations for the round, Indicates the first Medical experimental feedback corresponding to the round of recommendations This describes an iterative recommendation process that coordinates the intelligent agent using a global self-reflection mechanism. Through multiple iterations, the deviation between the predicted results and actual medical experiments is continuously reduced until the deviation value stably converges to a preset threshold. Within this range, the final output is the optimal drug recommendation result, which approximates the optimal choice in a real medical experiment.

[0179] This invention discloses a drug recommendation system based on multi-agent collaboration. Through a retrieval agent, it performs structured processing and retrieval enhancement on multi-source heterogeneous data from a biomedical database, efficiently screening candidate drugs related to target diseases from massive amounts of data to construct a comprehensive and accurate first drug candidate set, laying a data foundation for subsequent evaluation. The pharmacology agent's dynamic iterative pharmacology rule tree decomposes complex pharmacological mechanisms layer by layer into quantifiable indicators, continuously optimizing the tree structure and node weights under the drive of medical experiment feedback, achieving pharmacological evaluation that balances accuracy and interpretability. The clinical agent relies on counterfactual reasoning and drug relationship modeling driven by a large language model, making clinical evaluation more closely aligned with real-world diagnostic and treatment logic. The dynamic weighted fusion and conflict resolution mechanism of the coordinating agent effectively integrates the differentiated evaluation results of multiple agents, improving the robustness of drug recommendation ranking.

[0180] like Figure 3 The diagram shown is a flowchart illustrating a drug recommendation method based on multi-agent cooperation according to an embodiment of the present invention. This drug recommendation method based on multi-agent cooperation is applied to the drug recommendation system based on multi-agent cooperation described above. (Refer to...) Figure 3 This invention provides a drug recommendation method based on multi-agent collaboration, comprising the following steps:

[0181] S1. Using a retrieval agent to call a biomedical database through a preset interface, and performing structured processing and retrieval enhancement on the extracted multi-source heterogeneous data to obtain a first drug candidate set related to the target disease.

[0182] S2. Using a pharmacological agent based on a dynamically iterative pharmacological rule tree, hierarchical quantitative scoring and reflection-driven reordering are performed on the first drug candidate set to obtain the second drug candidate set.

[0183] S3. Using a clinical agent to model the relationship between counterfactual reasoning and drugs, the first drug candidate set is clinically quantitatively scored and re-ranked for diagnosis and treatment fit to obtain the third drug candidate set.

[0184] S4. The coordinated intelligent agent dynamically weights and fuses the first, second, and third drug candidate sets to obtain a comprehensive drug evaluation score matrix. The comprehensive drug evaluation score matrix is ​​then conflict-resolved and refined to produce a drug recommendation result related to the target disease.

[0185] like Figure 4 As shown, this is another flowchart illustrating a drug recommendation method based on multi-agent collaboration according to an embodiment of the present invention. (Refer to...) Figure 4According to an embodiment of the present invention, a drug recommendation method based on multi-agent cooperation, after step S4, further includes the following step:

[0186] S5. Introduce a global self-reflection mechanism. Based on the drug recommendation results, the optimal drug recommendation results are obtained by iteratively executing the global self-reflection mechanism in multiple rounds.

[0187] The process of the global self-reflection mechanism includes:

[0188] The system receives efficacy and toxicity scores from drug recommendations after verification through medical experiments. Based on these scores, it calculates the deviation between the overall drug superiority / inferiority prediction results output by the coordinating agent and the standardized medical experiment results. The deviation is compared to a threshold. If the deviation exceeds a preset threshold, the parameters of the pharmacological agent, clinical agent, and coordinating agent are optimized.

[0189] It should be noted that the specific limitations of a drug recommendation method based on multi-agent collaboration are the same as those for a drug recommendation system based on multi-agent collaboration mentioned above. The two have the same function and role, and will not be repeated here.

[0190] In summary, this invention provides a drug recommendation system and method based on multi-agent collaboration. By using a retrieval agent to perform structured processing and retrieval enhancement on multi-source heterogeneous data from biomedical databases, it can efficiently screen candidate drugs related to target diseases from massive amounts of data, constructing a comprehensive and accurate first drug candidate set, laying a data foundation for subsequent evaluation. The pharmacological agent's dynamic iterative pharmacological rule tree decomposes complex pharmacological mechanisms layer by layer into quantifiable indicators, continuously optimizing the tree structure and node weights under the drive of medical experiment feedback, achieving pharmacological evaluation that balances accuracy and interpretability. The clinical agent relies on counterfactual reasoning and drug relationship modeling driven by a large language model, making clinical evaluation more closely aligned with real-world diagnostic and treatment logic. The dynamic weighted fusion and conflict resolution mechanism of the coordinating agent effectively integrates the differentiated evaluation results of multiple agents, improving the robustness of drug recommendation ranking.

[0191] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on describing the differences from other embodiments. In particular, the method embodiments are basically similar to the system embodiments, so the description is relatively simple; refer to the description of the method embodiments for relevant details. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0192] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A drug recommendation system based on multi-agent collaboration, characterized in that, include: The retrieval agent calls a biomedical database through a preset interface and performs structured processing and retrieval enhancement on the extracted multi-source heterogeneous data to obtain a first drug candidate set related to the target disease. A pharmacological intelligent agent, based on a dynamically iterative pharmacological rule tree, performs hierarchical quantitative scoring and reflection-driven rearrangement on the first drug candidate set to obtain a second drug candidate set; A clinical intelligent agent, which uses counterfactual reasoning and drug relationship modeling to perform clinical quantitative scoring and treatment adaptation rearrangement on the first drug candidate set to obtain a third drug candidate set; A coordinating agent dynamically weights and fuses the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix. The coordinating agent then performs conflict resolution and fine-tuning on the comprehensive drug evaluation score matrix to output drug recommendation results related to the target disease. The system also includes a global self-reflection mechanism, which is used to obtain the optimal drug recommendation result through multiple rounds of iterative execution based on the drug recommendation result. The processing steps of the global self-reflection mechanism include: The system receives the efficacy score and toxicity score output after the drug recommendation results are verified by medical experiments, and calculates the deviation value between the overall drug superiority and inferiority prediction results output by the coordinating agent and the standardized medical experiment results based on the efficacy score and the toxicity score. The deviation value is compared with a threshold. If the deviation value exceeds a preset threshold, the parameters of the pharmacological agent, the clinical agent, and the coordination agent are optimized.

2. The drug recommendation system based on multi-agent cooperation according to claim 1, characterized in that, The retrieval agent consists of a first configuration module, a first memory module, and a first execution module; The first configuration module calls a biomedical database through a preset interface and extracts multi-source heterogeneous data related to the target disease. The first memory module performs vector embedding on the unstructured text in the multi-source heterogeneous data to obtain a vector index library, and performs entity recognition and relation extraction on the multi-source heterogeneous data to obtain medical knowledge triples. The first execution module performs a local search on the vector index library to obtain semantically similar data blocks, and summarizes the global search results for the medical knowledge triples and the semantically similar data blocks to generate a first drug candidate set related to the target disease.

3. The drug recommendation system based on multi-agent cooperation according to claim 1, characterized in that, The pharmacological intelligent agent consists of a second configuration module, a second memory module, and a second execution module; The second configuration module injects pharmacological assessment role prompts and initiates a rule tree structure generation task for the target disease; The second memory module responds to the rule tree structure generation task by constructing a pharmacological rule tree based on the node set and leaf index candidates output by the stored large language model. The second execution module performs hierarchical quantitative scoring on the first drug candidate set based on the pharmacological rule tree to obtain the pharmacological score of each drug, and corrects the drug ranking priority determined by each pharmacological score through a local self-reflection mechanism to obtain the second drug candidate set. The local self-reflection mechanism is configured to correct the structure and node weights of the pharmacological rule tree based on medical experiment feedback.

4. The drug recommendation system based on multi-agent cooperation according to claim 1, characterized in that, The clinical intelligent agent consists of a third configuration module, a third memory module, and a third execution module; The third configuration module injects clinical assessment role prompts to determine the format of clinical input data and counterfactual reasoning rules. The third memory module stores structured real case data conforming to the clinical input data format and counterfactual simulation samples generated by the stored large language model according to the counterfactual reasoning rules. The third execution module performs clinical quantitative scoring on the first drug candidate set based on the structured real case data and the counterfactual simulation samples to obtain a comprehensive clinical score for each drug, and prioritizes the first drug candidate set according to each comprehensive clinical score to obtain a third drug candidate set.

5. The drug recommendation system based on multi-agent cooperation according to claim 1, characterized in that, The coordinating agent consists of a fourth configuration module and a fourth execution module; The fourth configuration module injects collaborative decision-making role prompts and sets conflict resolution rules related to the multi-agent multi-dimensional evaluation results. The fourth execution module uses adaptive weights output by the stored large language model to perform weighted fusion on the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix. It then uses a pre-trained discriminant model to infer the comprehensive drug evaluation score matrix to obtain the comprehensive drug merit score for each drug. Finally, it combines the conflict resolution rules to perform fine-grained optimization on the comprehensive drug merit score for each drug to obtain drug recommendation results related to the target disease.

6. The drug recommendation system based on multi-agent cooperation according to claim 1, characterized in that, The process involves receiving the efficacy score and toxicity score output after the drug recommendation result has been verified by medical experiments, and calculating the deviation between the overall drug superiority / inferiority prediction result output by the coordinating agent and the standardized medical experiment result based on the efficacy score and the toxicity score, including: The raw efficacy data and raw toxicity data output from the medical experiment are standardized to obtain efficacy scores and toxicity scores, and the toxicity scores are then converted into toxicity acceptable indicators. The overall drug superiority / inferiority prediction results, efficacy scores, and toxicity acceptable indicators output by the coordinating agent are weighted and processed to obtain the deviation value between the overall drug superiority / inferiority prediction results and the standardized medical experiment results.

7. The drug recommendation system based on multi-agent cooperation according to claim 1, characterized in that, The step of comparing the deviation value with a threshold, and if the deviation value exceeds a preset threshold, then optimizing the parameters of the pharmacological agent, the clinical agent, and the coordinating agent, includes: Determine whether the deviation value exceeds a preset threshold. If so, update the pharmacological rule tree of the pharmacological agent and the counterfactual simulation sample of the clinical agent based on the deviation value. Based on the adjusted pharmacological agent and the clinical agent, the adaptive weights of the coordinating agent are updated.

8. A drug recommendation method based on multi-agent cooperation, characterized in that, The method, applied to the drug recommendation system based on multi-agent collaboration as described in any one of claims 1 to 7, comprises: The retrieval agent uses a preset interface to call a biomedical database and performs structured processing and retrieval enhancement on the extracted multi-source heterogeneous data to generate a first drug candidate set related to the target disease. Using the pharmacological agent based on a dynamically iterative pharmacological rule tree, the first drug candidate set is subjected to hierarchical quantitative scoring and reflection-driven reordering to obtain the second drug candidate set; Using the clinical agent to model drug relationships through counterfactual reasoning, the first drug candidate set is clinically quantitatively scored and re-ranked for diagnosis and treatment fit to obtain a third drug candidate set. The coordinating agent dynamically weights and fuses the first drug candidate set, the second drug candidate set, and the third drug candidate set to obtain a comprehensive drug evaluation score matrix. Conflict resolution and fine-tuning of the comprehensive drug evaluation score matrix are then performed to output drug recommendation results related to the target disease. A global self-reflection mechanism is introduced, and the optimal drug recommendation result is obtained by iteratively executing the global self-reflection mechanism through multiple rounds based on the drug recommendation result. The processing steps of the global self-reflection mechanism include: The system receives the efficacy score and toxicity score output after the drug recommendation results are verified by medical experiments, and calculates the deviation value between the overall drug superiority and inferiority prediction results output by the coordinating agent and the standardized medical experiment results based on the efficacy score and the toxicity score. The deviation value is compared with a threshold. If the deviation value exceeds a preset threshold, the parameters of the pharmacological agent, the clinical agent, and the coordination agent are optimized.

Images