Methods, systems, equipment, and media for generating drug repositioning study reports

Abstract

This application relates to the field of pharmaceutical data analysis technology, specifically to a method, system, device, and medium for generating drug repositioning study reports. The method for generating drug repositioning study reports decomposes the drug repositioning request submitted by pharmaceutical companies into three sub-tasks: indication expansion analysis, target matching analysis, and pharmacodynamic simulation verification. These sub-tasks are respectively invoked by intelligent agents for indication expansion analysis, target matching analysis, and pharmacodynamic simulation verification. Through multi-agent collaborative calculation of the contribution values ​​of each stage, a repositioning study report containing potential ratings and confidence assessments is finally generated based on a drug-disease knowledge graph. Implementing the technical solution provided in this application improves the efficiency and reliability of drug repositioning analysis, providing pharmaceutical companies with scientific and credible decision support.

Description

Technical Field

[0001] This application relates to the technical field of pharmaceutical data analysis, specifically to methods, systems, equipment, and media for generating drug repositioning study reports. Background Technology

[0002] In the field of pharmaceutical R&D, drug repositioning has become an important strategy for pharmaceutical companies to reduce R&D risks and shorten development cycles. However, traditional repositioning methods rely on human experience or single algorithm models, which have the following pain points: (1) Insufficient data integration: multi-source data such as drug structure, target information, and disease pathways have not been effectively integrated, making it difficult to systematically explore potential drug-disease associations; (2) Lack of transparency in the decision-making process: existing analysis methods lack interpretable reasoning paths, leading pharmaceutical companies to doubt the confidence of repositioning recommendations; (3) High verification costs: the lack of accurate target prediction and efficacy pre-verification links makes subsequent experimental verification risk high and resources wasteful.

[0003] Therefore, there is an urgent need for a method that can integrate multi-source data, simulate multi-task collaborative reasoning, and output a research report with high credibility relocation suggestions. Summary of the Invention

[0004] This application provides a method, system, device, and medium for generating drug repositioning research reports. Through the collaborative work of multiple agents, it enables a deep understanding and accurate decision-making regarding drug repositioning, thereby improving the efficiency and reliability of drug repositioning analysis.

[0005] The first aspect of this application provides a method for generating a drug repositioning study report, applied to a study report generation platform, the method comprising:

[0006] In response to receiving a request from a pharmaceutical company for target drug relocation, the request is decomposed into an indication expansion analysis subtask, a target matching analysis subtask, and a pharmacodynamic simulation verification subtask based on the molecular structure characteristics and original indications of the target drug provided by the pharmaceutical company. An initial weight is assigned to each subtask according to the request, and the indication expansion analysis agent, target matching analysis agent, and pharmacodynamic simulation verification agent are called from a preset agent library according to the request.

[0007] The indication expansion analysis subtask triggers the indication expansion analysis agent, which calculates the correlation score between the target drug and the potential indication based on a preset disease-pathway knowledge graph. The agent generates an indication expansion list based on the correlation score and calculates the indication expansion analysis contribution value. The indication expansion list is then written to shared storage, and the target matching analysis agent is triggered.

[0008] The target matching analysis agent calculates the structural similarity of new targets based on the indication expansion list and drug-target data, generates a new target list based on the structural similarity of new targets, calculates the target matching analysis contribution value, writes the new target list into shared storage, and triggers the drug efficacy simulation verification agent.

[0009] The pharmacodynamic simulation verification agent calls the molecular docking simulation platform interface according to the new target list, performs molecular docking to generate a pharmacodynamic scoring vector, and calculates the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector.

[0010] The candidate disease association score is obtained by weighting and summing the contribution values ​​of the pharmacodynamic simulation validation, the indication expansion analysis, and the target matching analysis based on the initial weights.

[0011] Multi-hop paths are generated based on the target drug, the candidate disease association scores, and a preset drug-disease knowledge graph. Each path in the multi-hop paths is rated for potential and the confidence level of each path is calculated. A research report on the relocation of the target drug is generated based on the potential rating and the confidence level.

[0012] By adopting the above technical solution, drug retargeting is decomposed into three stages: "indication discovery → target prediction → efficacy verification," corresponding to the standard process of target-pathway-efficacy verification in clinical research. At each stage, a corresponding intelligent agent is invoked for processing. This decomposition is not a simple task split, but rather a trigger-based collaborative mechanism between intelligent agents designed based on the biological causal chain (molecular structure → target → disease pathway): the output of the preceding intelligent agent (such as an expanded indication list) serves as the input constraint for subsequent intelligent agents, avoiding invalid search spaces. The contribution value of each intelligent agent's output is quantified using a domain-specific method, with initial weights preset according to the importance of sub-tasks. During the collaborative process, the weights are dynamically adjusted through a reward function, forming a closed-loop optimization mechanism. Utilizing the rich semantic information and logical relationships of knowledge graphs, the generated decision results have higher scientific validity and reliability. It not only provides drug developers with clear decision-making directions but also provides strong support for subsequent clinical trials and drug promotion, reducing research and development risks and costs.

[0013] Optionally, generating a multi-hop path based on the target drug, the candidate disease association score, and a preset drug-disease knowledge graph includes:

[0014] A candidate disease set is formed based on all candidate diseases whose association scores are greater than a first preset threshold. A subgraph of the target candidate disease is located in the preset drug-disease knowledge graph. The target candidate disease is any one of the candidate diseases in the candidate disease set. The subgraph includes pathological pathways, targets and drug entities that are directly connected to the first target node corresponding to the target candidate disease.

[0015] Based on the subgraph, backtracking from the first target node to the target drug entity along the relational edges of the preset drug-disease knowledge graph to form an initial path set;

[0016] A second target node with a node centrality higher than a preset centrality is identified in the initial path set. The action probability distribution of the second target node is generated through a preset policy network. The optimal action is selected by combining the preset reward function. The initial path in the initial path set is adjusted according to the optimal action to obtain a multi-hop path.

[0017] By employing the aforementioned technical solution, a candidate disease set is formed by screening all candidate diseases whose association scores exceed a first preset threshold from a large pool of candidate diseases. This step focuses on diseases with a high potential association with the target drug, improving the targeting and efficiency of subsequent analyses. A subgraph of the target candidate disease is located within the preset drug-disease knowledge graph. This subgraph contains pathological pathways, targets, and drug entities directly connected to the first target node corresponding to the target candidate disease. This approach comprehensively acquires various information closely related to the candidate disease, including the pathological mechanisms of disease occurrence, possible targets, and related drugs, providing a rich data foundation for a deeper understanding of the relationship between the target drug and the candidate disease. The formation of the subgraph constructs a local information network, connecting the candidate disease with surrounding key elements, helping analysts intuitively grasp the core elements related to the disease and their interrelationships. Based on the subgraph, the path backtracks along the relationship edges of the preset drug-disease knowledge graph from the first target node to the target drug entity, forming an initial path set. This backtracking method systematically identifies all possible paths from the candidate disease to the target drug, covering various possible pathways and associations between the drug and the disease, ensuring the integrity and comprehensiveness of the initial path set. The system generates an action probability distribution for the second target node using a pre-defined policy network, selects the optimal action based on a pre-defined reward function, and adjusts the initial paths in the initial path set according to the optimal action to obtain a multi-hop path. This process incorporates reinforcement learning, making the path generation process intelligent and adaptive. The policy network can learn the optimal action strategy based on the structure of the knowledge graph and historical experience, while the reward function can evaluate and provide feedback on the effects of different actions, thereby guiding the path to be adjusted in a more reasonable and effective direction, improving the quality and reliability of the multi-hop path.

[0018] Optionally, the step of generating the action probability distribution of the second target node through a preset policy network, selecting the optimal action in combination with a preset reward function, and adjusting the initial paths in the initial path set according to the optimal action to obtain a multi-hop path includes:

[0019] An embedding vector for the second target node is generated using a graph neural network, and a basic probability distribution for the candidate action is generated based on the embedding vector. The candidate action is one of all allowed actions.

[0020] Calculate the state value of the second target node and the action value of the candidate action, and generate an advantage value based on the action value and the state value;

[0021] The probability distribution is obtained by adjusting the basic probability distribution based on the advantage value of all the candidate actions, the basic reward value is calculated based on the preset reward function, and the comprehensive reward value of each candidate action is obtained based on the probability distribution and the basic reward value.

[0022] Select the target action with the highest overall reward value. If the target action is to add a relationship, then add a relationship edge between the second target node and the original target node. If the target action is to replace a node, then replace the original target node with the nearest neighbor node of the original target node. The original target node is the node directly connected to the second target node, and the nearest neighbor node is the node with the highest correlation to the original target node, excluding the second target node.

[0023] By employing the aforementioned technical solution, a graph neural network is used to generate the embedding vector of the second target node. The graph neural network can automatically learn the local and global structural information of nodes in the knowledge graph, as well as the relationship features between nodes and other entities. In this way, complex node information is encoded into low-dimensional embedding vectors, accurately characterizing the features of the second target node and laying the foundation for generating a reasonable action probability distribution. The state value of the second target node and the action value of the candidate actions are calculated, and an advantage value is generated. The state value reflects the long-term benefit obtained by taking the optimal action in the current state, while the action value measures the direct benefit of a specific action in the current state. The advantage value comprehensively considers the additional benefit of the action relative to the current state, enabling a more scientific evaluation of the merits of each candidate action and providing an objective basis for selecting the optimal action. A new probability distribution is obtained by adjusting the basic probability distribution based on the advantage values ​​of all candidate actions, increasing the probability of selecting actions with high advantage values ​​and decreasing the probability of selecting actions with low advantage values. This adjustment method, based on the actual value of actions, guides the policy network to be more inclined to select actions beneficial to path optimization, improving the rationality and effectiveness of action selection.

[0024] Optionally, replacing the original target node with its nearest neighbor node includes:

[0025] Determine the set of neighboring nodes of the original target node, the set of neighboring nodes containing all entities that are directly connected to the original target node and are not the second target node;

[0026] Obtain the number of shared relationship types between the target neighbor node and the original target node, and calculate the ratio of the number of shared relationship types to the total number of relationship types of the original target node as the structural association degree. The target neighbor node is any node in the set of neighbor nodes.

[0027] Obtain the historical interaction frequency between the target neighbor node and the original target node, and calculate the functional correlation degree based on the historical interaction frequency;

[0028] The total correlation degree is calculated by weighting the structural correlation degree and the functional correlation degree, and the neighbor node with the highest total correlation degree is selected as the nearest target neighbor node.

[0029] The original target node in the initial path is replaced with the nearest target neighbor node to form a reconstructed path. The global reliability of the reconstructed path is calculated. If the global reliability is greater than a second preset threshold, the reconstructed path is retained.

[0030] By employing the aforementioned technical solution, a set of neighboring nodes for the original target node is determined. This set includes all entities directly connected to the original target node but not the second target node. This approach ensures comprehensive coverage of other entities directly related to the original target node when considering replacement nodes, without overlooking potentially important nodes, thus providing abundant candidate resources for subsequent selection of the nearest neighboring nodes. The structural relevance is calculated by obtaining the number of shared relationship types between the target neighboring node and the original target node and calculating the ratio of this number to the total number of relationship types of the original target node. This metric quantitatively reflects the similarity between the target neighboring node and the original target node in the knowledge graph structure. The evaluation of functional relevance incorporates historical interaction information, ensuring that node selection considers not only the structure of the knowledge graph but also its functional performance in practical applications. This comprehensive evaluation method, which considers both structure and function, better meets the needs of practical scenarios such as drug development, enabling the selection of neighboring nodes that are functionally and functionally closer to the original target node for replacement. By comprehensively weighing structural and functional relevance, neighboring nodes that are highly similar to the original target node in both structure and function can be selected for replacement. Such high-quality replacement nodes help maintain the rationality and effectiveness of the reconstructed path, reducing path logic errors or information loss caused by node replacement. Global reliability verification of the reconstructed path can filter out truly rational and effective paths, eliminating those that may contain logical errors or inaccurate information. This step helps optimize path selection results, providing more accurate and reliable path information for drug development decisions, and improving the scientific rigor and accuracy of those decisions.

[0031] Optionally, the step of calculating the correlation score between the target drug and the potential indication based on the disease-pathway knowledge graph through the indication expansion analysis agent, generating an indication expansion list based on the correlation score, and calculating the indication expansion analysis contribution value includes:

[0032] Based on the molecular structure of the target drug, a molecular graph embedding vector is generated;

[0033] Based on the disease-pathway knowledge graph, the biological pathway nodes associated with the original indication are located, and the pathway feature vectors of the biological pathway nodes are extracted.

[0034] Calculate the semantic similarity between the molecular graph embedding vector and the pathway feature vector, and obtain the association score based on the semantic similarity;

[0035] Filter out candidate diseases whose correlation scores exceed a third preset threshold, and generate an expanded list of indications based on the candidate diseases;

[0036] The pathways that the target drug may regulate in the candidate disease are identified as new pathways. The total pathways are determined based on the new pathways and the pathways that the target drug has been verified to regulate in the original indication. The proportion of the new pathways to the total pathways is calculated.

[0037] The contribution value of the indication expansion analysis is calculated based on the stated proportion and the stated correlation score.

[0038] By employing the aforementioned technical solution, molecular graph embedding vectors are generated based on the molecular structure of the target drug. This process utilizes graph embedding technology to transform complex molecular structural information into low-dimensional vector forms. This vector representation can capture key information such as the connection patterns and spatial structure between atoms in the molecule, providing an accurate and computable representation of drug features for subsequent matching with disease pathway features, avoiding the difficulties caused by directly processing complex molecular structures. The semantic similarity between the molecular graph embedding vector and the pathway feature vector is calculated, and an association score is obtained accordingly. The semantic similarity calculation method can measure the closeness of two vectors in the semantic space, thereby quantifying the similarity between the target drug and the biological pathway associated with the original indication. Candidate diseases with association scores exceeding a third preset threshold are screened, and an indication expansion list is generated. By setting reasonable thresholds, potential diseases with a high degree of matching with the target drug can be quickly and accurately screened from a large number of diseases, narrowing the scope of further research and improving the efficiency of drug indication discovery. The contribution value of indication expansion analysis provides an important decision-making basis for drug development projects. The higher the contribution value, the greater the research value and development potential of the target drug in the potential disease. Researchers can rationally allocate research resources and priorities based on the magnitude of the contribution value, prioritizing research on potential diseases with higher contribution values, thereby improving the efficiency and success rate of drug development.

[0039] Optionally, the step of calculating the structural similarity of new targets based on the expanded indication list and drug-target data using the target matching analysis agent, generating a new target list based on the structural similarity of the new targets, and calculating the target matching analysis contribution value includes:

[0040] The protein sequences of the target sites of the candidate diseases are analyzed, and three-dimensional structural embedding vectors are generated through a preset geometric deep learning model.

[0041] Functional domain similarity is determined based on the molecular graph embedding vector, structural similarity is determined based on the three-dimensional structure embedding vector, and the total similarity is calculated by weighting the functional domain similarity and the structural similarity.

[0042] Filter target points whose total similarity is greater than a fourth preset threshold, and generate a new target point list based on the target points;

[0043] Based on a preset protein-protein interaction network, the normalized betweenness centrality of the target is calculated, and the target matching analysis contribution value is calculated based on the normalized betweenness centrality and preset weights according to a preset contribution value quantification formula.

[0044] By employing the aforementioned technical approach, the protein sequences of potential disease targets are analyzed, which forms the basis for a deeper understanding of target structure and function. Protein sequences contain crucial information such as the amino acid composition and sequence of the target, providing raw data for the subsequent generation of three-dimensional structure embedding vectors, enabling target research to move from the sequence level to the structural level. Functional domain similarity is determined based on the molecular graph embedding vectors, and structural similarity is determined based on the three-dimensional structure embedding vectors, with a weighted average to obtain the total similarity. Functional domain similarity assesses the degree of matching between the drug molecule and the target's functional regions, reflecting the potential functional sites on which the drug may act on the target; structural similarity measures the similarity between the drug and the target from the perspective of the overall three-dimensional structure. By comprehensively considering the similarity of these two dimensions, the association between the drug and the target can be assessed more comprehensively and accurately, avoiding the limitations of single-dimensional assessment. Based on standardized betweenness centrality and preset weights, the target matching analysis contribution value is calculated using a preset contribution value quantification formula. This calculation method comprehensively considers the structural similarity between the target and the drug, as well as the importance of the target in the protein-protein interaction network, enabling a more comprehensive and reasonable assessment of the potential value and contribution of the target in drug development. The higher the contribution value, the greater the research significance and development potential of the target in the drug discovery and development process, providing an important reference for the decision-making of drug development projects.

[0045] Optionally, the step of calling the molecular docking simulation platform interface through the pharmacodynamic simulation verification agent according to the new target list, performing molecular docking to generate a pharmacodynamic scoring vector, and calculating the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector includes:

[0046] Molecular docking is performed to generate multimodal parameters, including binding energy, conformational stability score, number of hydrogen bonds and hydrophobic contact area, and a drug efficacy score vector is generated based on the multimodal parameters.

[0047] The sub-vectors of each dimension in the efficacy score vector are normalized, and the weighted sum of the normalized sub-vectors is performed to obtain the basic contribution value.

[0048] The delivery efficiency coefficient is calculated based on the hierarchical depth of the target in the disease pathway, and the basic contribution value is adjusted based on the delivery efficiency coefficient to obtain the efficacy simulation verification contribution value.

[0049] By employing the aforementioned technical solution, molecular docking is performed to generate multimodal parameters including binding energy, conformational stability score, number of hydrogen bonds, and hydrophobic contact area. Binding energy reflects the strength of the binding between the drug molecule and the target, and is one of the important indicators for evaluating drug efficacy. The conformational stability score reflects the structural stability of the drug-target complex after binding; a stable conformation helps the drug exert its effects continuously. The number of hydrogen bonds and the hydrophobic contact area describe the binding mode and tightness between the drug and the target from the perspective of intermolecular interactions. These multimodal parameters comprehensively capture key information about the drug-target interaction from different aspects, providing a rich data foundation for accurately evaluating drug efficacy. The delivery efficiency coefficient is calculated based on the hierarchical depth of the target in the disease pathway. Different hierarchical depths of the target in the disease pathway result in differences in the efficiency and complexity of the drug's effect on the disease process. The delivery efficiency coefficient can quantify these differences in the action pathway caused by different target hierarchical depths, providing a reasonable basis for adjusting the baseline contribution value. The contribution value for drug efficacy simulation validation is obtained by adjusting the baseline contribution value based on the delivery efficiency coefficient. This dynamic adjustment approach ensures that the contribution value of pharmacodynamic simulation validation considers not only the direct pharmacodynamic information of drug-target interaction, but also the impact of the target's location and pathway of action within the disease pathway. This comprehensive evaluation more accurately reflects the therapeutic effects and potential value of drugs targeting specific points, providing a more reliable reference for drug development decisions.

[0050] A second aspect of this application provides a system for generating drug repositioning study reports, including a decomposition module, a discovery module, a prediction module, a validation module, a calculation module, and an execution module, wherein:

[0051] The decomposition module is configured to respond to a request for target drug relocation submitted by a pharmaceutical company, decompose the request into an indication expansion analysis subtask, a target matching analysis subtask, and a pharmacodynamic simulation verification subtask based on the molecular structure characteristics and original indications of the target drug provided by the pharmaceutical company, assign initial weights to each subtask according to the request, and call the indication expansion analysis intelligent agent, target matching analysis intelligent agent, and pharmacodynamic simulation verification intelligent agent from a preset intelligent agent library according to the request.

[0052] The discovery module is configured to trigger the indication expansion analysis agent through the indication expansion analysis subtask, calculate the correlation score between the target drug and the potential indication based on a preset disease-pathway knowledge graph, generate an indication expansion list based on the correlation score and calculate the indication expansion analysis contribution value, write the indication expansion list to shared storage and trigger the target matching analysis agent;

[0053] The prediction module is configured to calculate the structural similarity of new targets based on the indication expansion list and drug-target data through the target matching analysis agent, generate a new target list based on the structural similarity of new targets and calculate the target matching analysis contribution value, write the new target list to shared storage and trigger the drug efficacy simulation verification agent;

[0054] The verification module is configured to use the pharmacodynamic simulation verification agent to perform molecular docking to generate a pharmacodynamic scoring vector based on the new target list, and to calculate the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector.

[0055] The calculation module is configured to perform a weighted summation of the contribution values ​​of the pharmacodynamic simulation verification, the indication expansion analysis, and the target matching analysis based on the initial weights to obtain a candidate disease association score.

[0056] The execution module is configured to generate multi-hop paths based on the target drug, the candidate disease association scores, and a preset drug-disease knowledge graph, to perform a potential rating on each path in the multi-hop paths, to calculate the confidence level of each path, and to generate a research report on the relocation of the target drug based on the potential rating and the confidence level.

[0057] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the foregoing.

[0058] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed, perform the method described in any of the preceding descriptions.

[0059] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0060] 1. Decompose drug repositioning requests into three sub-tasks: indication discovery, target prediction, and efficacy validation. Invoke appropriate agents to handle sub-tasks at different stages. Automatic data transfer between agents is achieved through shared storage (such as a database), avoiding redundant data parsing and shortening the decision-making cycle. An initial weight allocation (e.g., higher weight for efficacy validation) and a weighted summation mechanism based on contribution values ​​prioritize high-confidence results, reducing computational resource consumption on low-value paths.

[0061] 2. The target matching analysis agent integrates molecular graph embedding vectors (chemical properties) and three-dimensional structure embedding vectors (spatial conformation) to improve the biological rationality of target screening. The pharmacodynamic simulation validation agent generates multi-dimensional parameters such as binding energy and conformational stability through molecular docking, and generates weighted contribution values ​​after normalization, reducing the misjudgment rate caused by single-indicator bias. Based on the drug-disease knowledge graph, multi-hop paths are generated, and combined with the candidate disease association score screening mechanism to match indications, improving the accuracy of clinical relevance prediction.

[0062] 3. Each subtask outputs an independent contribution value (such as contribution value for indication expansion analysis and target matching analysis), clarifying the influence weight of each step on the final decision and supporting dual-dimensional result traceability based on "computational and biological mechanisms". In knowledge graph path generation, a node replacement strategy is adopted (such as replacing low-centrality nodes with high-relevance neighbor nodes), combined with a reward function to optimize path reliability and ensure that the decision path conforms to biological logic. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the method for generating a drug repositioning study report disclosed in an embodiment of this application;

[0064] Figure 2 This is a schematic diagram of the modules of the drug repositioning study report generation system disclosed in the embodiments of this application;

[0065] Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0066] Explanation of reference numerals in the attached diagram: 201, Decomposition module; 202, Discovery module; 203, Prediction module; 204, Verification module; 205, Calculation module; 206, Execution module; 301, Processor; 302, Communication bus; 303, User interface; 304, Network interface; 305, Memory. Detailed Implementation

[0067] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0068] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0069] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0070] This embodiment discloses a method for generating a drug repositioning study report. Figure 1 This is a flowchart illustrating the method for generating a drug repositioning study report disclosed in an embodiment of this application, which is applied to a study report generation platform, such as... Figure 1 As shown, the method includes the following steps:

[0071] S101. In response to receiving a request for target drug relocation submitted by a pharmaceutical company, the request is decomposed into an indication expansion analysis subtask, a target matching analysis subtask, and a pharmacodynamic simulation verification subtask based on the molecular structure characteristics and original indications of the target drug provided by the pharmaceutical company. An initial weight is assigned to each subtask according to the request, and the indication expansion analysis agent, target matching analysis agent, and pharmacodynamic simulation verification agent are called from a preset agent library according to the request.

[0072] S102. The indication expansion analysis subtask triggers the indication expansion analysis agent. The indication expansion analysis agent calculates the correlation score between the target drug and the potential indication based on the preset disease-pathway knowledge graph. The indication expansion list is generated according to the correlation score, and the indication expansion analysis contribution value is calculated. The indication expansion list is written to the shared storage and the target matching analysis agent is triggered.

[0073] S103. The target matching analysis agent calculates the structural similarity of new targets based on the indication expansion list and drug-target data, generates a new target list based on the structural similarity of new targets and calculates the target matching analysis contribution value, writes the new target list into shared storage and triggers the drug efficacy simulation verification agent.

[0074] S104. The drug efficacy simulation verification agent calls the molecular docking simulation platform interface according to the new target list, performs molecular docking to generate a drug efficacy score vector, and calculates the drug efficacy simulation verification contribution value according to the drug efficacy score vector.

[0075] S105. The candidate disease association score is obtained by weighting and summing the contribution values ​​of the pharmacodynamic simulation verification, the indication expansion analysis, and the target matching analysis according to the initial weights.

[0076] S106. Generate multi-hop paths based on the target drug, the candidate disease association scores, and the preset drug-disease knowledge graph. Assess the potential of each path in the multi-hop paths and calculate the confidence level of each path. Generate a research report on the relocation of the target drug based on the potential rating and the confidence level.

[0077] When a system receives a target drug relocation request from a pharmaceutical company, it breaks down the entire relocation request into three sub-tasks based on the target drug's molecular structural characteristics (such as atomic composition, chemical bond types, and spatial conformation) and its original indication (the disease for which the drug was originally approved). These sub-tasks are indication expansion analysis, target matching analysis, and pharmacodynamic simulation verification. Simultaneously, based on the specific analytical requirements of the request, an initial weight is assigned to each sub-task, and the corresponding agent is invoked from a pre-defined agent library to execute the analysis task. This task decomposition method breaks down the complex target drug relocation problem into multiple relatively independent sub-tasks with clear objectives, facilitating separate processing and research. Assigning initial weights ensures that the relative importance of different sub-tasks in the entire relocation process is reflected during subsequent comprehensive evaluation, while the invocation of agents allows for flexible configuration of analytical capabilities. The indication expansion analysis agent utilizes the target drug's molecular structural characteristics and original indication to obtain a correlation score through a series of complex calculations and analyses (such as calculating molecular graph embedding vectors, extracting pathway feature vectors, and calculating semantic similarity). Based on this score, potential diseases are screened, an indication expansion list is generated, and the contribution value of indication expansion analysis is calculated. Then, the indication expansion list is written to shared storage, and the target matching analysis agent is triggered. From the perspective of molecular structure and original indications, potential diseases that may have a therapeutic relationship with the target drug are sought, providing a disease range for subsequent target prediction. The contribution value of indication expansion analysis is calculated to evaluate the effectiveness and importance of this subtask in discovering potential diseases. After the target matching analysis agent is triggered, it calculates the structural similarity (including functional domain similarity and structural similarity) between new targets and the target drug based on the indication expansion list and a pre-set drug-target database. Target targets are screened based on structural similarity, a new target list is generated, and the contribution value of target matching analysis is calculated. Then, the new target list is written to shared storage, and the pharmacodynamic simulation verification agent is triggered. Within the identified potential disease range, new targets that may interact with the target drug are sought, providing specific target objects for subsequent pharmacodynamic verification. The contribution value of target matching analysis is calculated to evaluate the accuracy and importance of this subtask in predicting new targets. The pharmacodynamic simulation validation agent performs molecular docking operations based on a list of new targets, generating multimodal parameters (such as binding energy, conformational stability score, number of hydrogen bonds, and hydrophobic contact area), and then generates a pharmacodynamic score vector based on these parameters. The pharmacodynamic score vector is then normalized, weighted, and summed to obtain a baseline contribution value. The delivery efficiency coefficient is then calculated based on the hierarchical depth of the target in the disease pathway, and the baseline contribution value is adjusted to finally obtain the pharmacodynamic simulation validation contribution value. This validates the interaction between the target drug and the new targets, and assesses the drug's therapeutic potential for underlying diseases through these targets.Calculating the contribution value of pharmacodynamic simulation validation is to comprehensively measure the importance of the pharmacodynamic validation results in the entire relocation process. Based on the initial weights assigned to each subtask in the previous steps, the contribution values ​​of pharmacodynamic simulation validation, indication expansion analysis, and target matching analysis are weighted and summed to obtain the candidate disease association score. Considering the results of the three subtasks—indication discovery, target prediction, and pharmacodynamic validation—a comprehensive score is given to measure the degree of association and potential therapeutic likelihood between the target drug and the candidate disease. Multi-hop paths are generated based on the target drug, candidate disease association scores, and a pre-defined drug-disease knowledge graph. A multi-hop path refers to the connection path from the target drug to the candidate disease in the knowledge graph, which may involve multiple intermediate nodes (such as targets, biological pathways, etc.). Decision results are generated based on these multi-hop paths to determine whether to relocate the target drug to the candidate disease for treatment. Utilizing the rich information in the knowledge graph, the relationship between the target drug and the candidate disease is further analyzed, and combined with the candidate disease association score, a final relocation decision is made. The relocation decision can be output based on the total path score and logical closure of the multi-hop paths. Repositioning decisions can be exemplified by the following: Aspirin, by inhibiting PPARδ and blocking the Wnt pathway, may treat colorectal cancer (92% confidence level), and in vitro cell experiments are recommended. The total pathway score consists of a pathway score and an association score, reflecting the statistical significance of biological evidence: the pathway score (60% weighting) is derived from the product of the sub-task contributions (indication discovery × target prediction × efficacy validation), for example, pathway score = 0.7 (indication contribution) × 0.8 (target contribution) × 0.9 (efficacy contribution) = 0.504; the association score (40% weighting) is based on the entity association strength of the knowledge graph: efficacy score (molecular docking binding energy ≤ -7.0 kcal / mol → conversion score 0.3), pathway matching degree (KEGG enrichment analysis P value < 0.01 → conversion score 0.4), association score = 0.3 + 0.4 = 0.7. The total path score is calculated as (0.504 × 0.6) + (0.7 × 0.4) = 0.302 + 0.28 = 0.582. Logical closure verification ensures the path conforms to biomedical common sense, avoiding "high-scoring but unreliable" paths. It can verify whether the aspirin → PPARδ → Wnt pathway → colorectal cancer form a closed chain: checking whether PPARδ regulates the Wnt pathway (verified using the KEGG PATHWAY database), and confirming whether Wnt pathway activation drives colorectal cancer (verified using TCGA tumor genomics data). If aspirin simultaneously activates the oncogenic pathway COX-2, the path needs to be downweighted; if no conflict is detected, the logical integrity score is 0.95 (out of 1.0), and high weights are assigned to key nodes (PPARδ) using graph attention (GAT). Confidence score = total path score × logical integrity × attenuation factor. The attenuation factor is adjusted based on the reliability of the data source (e.g., experimental data weight > simulated data weight).The confidence level is calculated as 0.582 × 0.95 × 0.98 × 100% ≈ 92%. Based on threshold grading: if the confidence level is greater than or equal to 90%, a strong recommendation is given (directly proceed with the experiment); if the confidence level is between 70% and 90%, a moderate recommendation is given (further data is required); if the confidence level is less than or equal to 70%, a weak recommendation is given (only as a research direction). The final research report will clearly indicate the potential level and confidence level of each recommended path and provide differentiated follow-up research suggestions for different levels of recommendations. For example, high-potential, high-confidence paths are recommended to be prioritized for experimental verification, while medium-level paths are recommended to be further evaluated after supplementing data. This tiered recommendation mechanism provides clear priority guidance for pharmaceutical companies' R&D decisions, effectively improving the efficiency of R&D resource allocation.

[0078] Epidemiological data for the target disease, including annual incidence (I), prevalence (P), and patient population size, can be obtained from the Global Burden of Disease (GBD) database and pharmaceutical market intelligence platforms. This data allows for the acquisition of annual treatment costs (C) and the compound annual growth rate (CAGR) of the market size for that disease area, enabling the calculation of the disease market value factor M: M = log 10(P×C)×(1+G). Establish a weight mapping system for clinical development stages: preclinical stage (0.2), Phase I clinical trial (0.4), Phase II clinical trial (0.7), Phase III clinical trial (0.9), and marketed targets (1.0). Count the number of investigational drugs targeting the target (D) from the clinical trial database. Calculate the competition intensity index CI: CI=1-min(D / 50, 0.8), where 50 is the set competition saturation threshold. Count the number of patents related to this path mechanism (Pat) from the patent database and calculate the patent density factor PDF: PDF=min(Pat / 100, 0.9), reflecting the degree of patent barriers in this technology field. Calculate the mechanism complexity MC based on the path length (L): MC=max(0, 1-(L-2) / 10), the more complex the path, the higher the development difficulty. Comprehensive potential score = M×stage weight×CI×(1-PDF)×MC. A grading threshold is set: a score ≥ 0.7 indicates a high-potential path, 0.4 ≤ score < 0.7 indicates a medium-potential path, and a score < 0.4 indicates a low-potential path. Path confidence is calculated through multi-source evidence fusion to ensure scientific reliability: Literature Evidence Score (LE): LE = min(number of high-quality journal articles / 5, 1.0) × 0.6 + min(total number of articles / 20, 1.0) × 0.4. The system maintains a list of high-quality journals, which is mainly defined based on the internationally recognized journal citation report's quartiles and impact factors. The number of high-quality journal articles refers to the number of articles published in this high-quality journal list. An experimental evidence weighting system is established: human clinical trials (1.0), animal models (0.7), and cell experiments (0.4). Experimental Evidence Score (EE): EE = Σ(experiment type weight × number of experiments) / max(total number of experiments, 1). Multi-omics data fit analysis: multi-omics datasets supporting the path are extracted from a preset database. The cumulative contribution rate of eigenvalues ​​in the correlation coefficient matrix is ​​calculated to obtain the consistency score (CS): CS = the proportion of eigenvalues ​​with a cumulative contribution rate ≥ 0.8. Clinical phenotype association validation: The statistical significance of disease-phenotype associations is calculated using electronic medical record data. Clinical association score (CAS): CAS = -log 10(P-value) / 10, where the P-value comes from the correlation test. An expert-annotated gold standard dataset is constructed, containing experimentally validated biological pathways. The topological similarity (TS) between the path to be evaluated and the gold standard dataset is calculated: TS = Graph structure similarity × Node function similarity. The path is ensured to conform to known biological laws. Confidence level = (LE × 0.3 + EE × 0.3) × 0.5 + CS × 0.2 + CAS × 0.2 + TS × 0.1. Confidence levels are set: ≥0.8 is high confidence, 0.6-0.8 is medium confidence, and <0.6 is low confidence. In the final research report, the potential rating and confidence assessment form a two-dimensional decision matrix: high-potential-high-confidence paths are the core recommendation, and priority should be given to R&D resources; high-potential-medium confidence paths are the key validation directions, and supplementary experiments are recommended; medium-potential-high confidence paths are alternative solutions, and follow-up progress should be monitored; other combined paths are exploratory discoveries, and continued monitoring is recommended.

[0079] Optionally, generating a multi-hop path based on the target drug, the candidate disease association score, and a preset drug-disease knowledge graph includes:

[0080] A candidate disease set is formed based on all candidate diseases whose association scores are greater than a first preset threshold. A subgraph of the target candidate disease is located in the preset drug-disease knowledge graph. The target candidate disease is any one of the candidate diseases in the candidate disease set. The subgraph includes pathological pathways, targets and drug entities that are directly connected to the first target node corresponding to the target candidate disease.

[0081] Based on the subgraph, backtracking from the first target node to the target drug entity along the relational edges of the preset drug-disease knowledge graph to form an initial path set;

[0082] A second target node with a node centrality higher than a preset centrality is identified in the initial path set. The action probability distribution of the second target node is generated through a preset policy network. The optimal action is selected by combining the preset reward function. The initial path in the initial path set is adjusted according to the optimal action to obtain a multi-hop path.

[0083] All candidate diseases with a correlation score greater than a first preset threshold are selected to form a candidate disease set. The first preset threshold is set according to actual needs and is used to screen candidate diseases with high potential for treatment. In the preset drug-disease knowledge graph, for any target candidate disease in the candidate disease set, its corresponding subgraph is located. The subgraph contains pathological pathways, targets, and drug entities directly connected to the first target node corresponding to the target candidate disease. A knowledge graph is a data model that uses a graph structure to represent entities and the relationships between them. In the drug-disease knowledge graph, entities can be drugs, diseases, targets, pathological pathways, etc., and relationships represent the interactions or associations between them. For example, if a target drug called "Drug A" (which has been authorized) is being studied, and after the previous steps, a series of candidate diseases and their correlation scores are obtained, such as disease X (correlation score 0.8), disease Y (correlation score 0.6), and disease Z (correlation score 0.3), and the first preset threshold is set to 0.5, then the candidate disease set includes disease X and disease Y. With disease X as the target candidate disease, in the drug-disease knowledge graph, the first target node corresponding to disease X is directly connected to the pathological pathway "inflammatory response pathway", the target "protein B" and "drug C". These connected elements and their connection relationship with disease X constitute the subgraph of disease X.

[0084] Based on the subgraph obtained from the localization, the system backtracks from the first target node (i.e., the node corresponding to the target candidate disease) to the target drug entity along the edges of the pre-defined drug-disease knowledge graph. Backtracking means starting from the disease and searching for connecting paths towards the drug along the edges in the knowledge graph representing relationships such as drug-disease, target-disease, and pathological pathway-disease. Through this backtracking process, multiple paths from the target candidate disease to the target drug can be found, forming an initial path set. Continuing the example above, the system backtracks from the first target node, disease X. In the knowledge graph, it is found that "drug A" regulates the "inflammatory response pathway" by affecting "protein B" (target), thereby affecting disease X; simultaneously, "drug A" may also be associated with disease X through other unknown intermediate links. Thus, two initial paths are obtained: path 1 is "drug A - protein B - inflammatory response pathway - disease X"; path 2 may be "drug A - [unknown intermediate link] - disease X". These two paths constitute the initial path set. The system then identifies the second target node in the initial path set whose node centrality is higher than the pre-defined centrality. Node centrality is a metric that measures the importance of a node in a graph. Preset centrality is set based on the characteristics of the knowledge graph and the analysis requirements, used to filter nodes that have a significant impact on the path. A preset policy network generates the probability distribution of actions for the second target node. The policy network is a machine learning model that predicts the probability distribution of different actions that can be taken against a given node based on the input node information and path context. Actions can include adding new intermediate nodes, deleting unnecessary nodes, and adjusting the connections between nodes. A preset reward function is used to select the optimal action. The reward function evaluates the impact of each action on the path quality; for example, an action that makes the path more consistent with biological common sense, more concise, or better explains the relationship between drugs and diseases will receive a higher reward value. The action with the highest probability and the largest reward value is selected as the optimal action based on the reward function. The initial paths in the initial path set are adjusted based on the optimal action to obtain multi-hop paths. Multi-hop paths are paths containing multiple intermediate nodes, which can more comprehensively demonstrate the complex relationship between drugs and diseases. For example, in the initial path set, assuming that the node centrality of "protein B" is higher than the preset centrality, it is determined as the second target node. The pre-defined policy network generates an action probability distribution based on information such as the position of "protein B" in the path and its connections with other nodes. For example, the probability of adding a new intermediate node "signaling molecule D" is 0.3, the probability of deleting "protein B" is 0.1, and the probability of keeping it unchanged is 0.6. After evaluation by the pre-defined reward function, it is found that adding "signaling molecule D" can better explain how drug A regulates the inflammatory response pathway by affecting protein B, thereby affecting disease X. Therefore, the action of adding "signaling molecule D" receives a higher reward value.Based on the optimal action (adding "signaling molecule D"), the initial path 1 is adjusted to obtain the multi-hop path "drug A-protein B-signaling molecule D-inflammatory response pathway-disease X". This path shows the mechanism of action between drug A and disease X in more detail.

[0085] Optionally, the step of generating the action probability distribution of the second target node through a preset policy network, selecting the optimal action in combination with a preset reward function, and adjusting the initial paths in the initial path set according to the optimal action to obtain a multi-hop path includes:

[0086] An embedding vector for the second target node is generated using a graph neural network, and a basic probability distribution for the candidate action is generated based on the embedding vector. The candidate action is one of all allowed actions.

[0087] Calculate the state value of the second target node and the action value of the candidate action, and generate an advantage value based on the action value and the state value;

[0088] The probability distribution is obtained by adjusting the basic probability distribution based on the advantage value of all the candidate actions, the basic reward value is calculated based on the preset reward function, and the comprehensive reward value of each candidate action is obtained based on the probability distribution and the basic reward value.

[0089] Select the target action with the highest overall reward value. If the target action is to add a relationship, then add a relationship edge between the second target node and the original target node. If the target action is to replace a node, then replace the original target node with the nearest neighbor node of the original target node. The original target node is the node directly connected to the second target node, and the nearest neighbor node is the node with the highest correlation to the original target node, excluding the second target node.

[0090] A graph neural network (Graph Neural Network) is used to process the second target node, generating its embedding vector. Graph Neural Networks capture the local and global structural information of nodes within a graph, encoding the information of the node and its surrounding environment into a low-dimensional vector, the embedding vector. Based on this embedding vector, a fully connected layer or other suitable neural network structure is used to generate the basic probability distribution of candidate actions. A candidate action is one of all possible actions that can adjust the path, such as adding a relationship, deleting a node, or replacing a node. For example, in the initial path "Drug A - Target B - Disease C" of a drug-disease knowledge graph, target B is the second target node. A graph neural network is used to process target B, obtaining an embedding vector, assumed to be [0.2, 0.5, 0.3] (this is a simplified illustration; the actual vector dimension is higher). A fully connected layer maps this embedding vector to the probability space of the candidate actions. Suppose there are three candidate actions: adding a relationship, deleting a node, and replacing a node. The generated basic probability distribution is [0.4, 0.3, 0.3], meaning the probability of adding a relationship is 0.4, the probability of deleting a node is 0.3, and the probability of replacing a node is 0.3. Calculate the state value of the second target node. The state value reflects the node's contribution to the final goal in its current state and can be estimated using a value network (such as a deep Q-network). Calculate the action value of each candidate action. The action value represents the expected reward gained from transitioning from the current state to the next state after performing the action. Generate an advantage value based on the action value and state value. The advantage value measures how much additional benefit a particular action brings relative to the average level of the current state. The formula is usually Advantage Value = Action Value - State Value. Continuing the example above, suppose the state value of target point B is estimated to be 0.6 using a value network. For the action of adding a relationship, suppose its action value is 0.8 after simulation or evaluation; the action value of deleting a node is 0.5; and the action value of replacing a node is 0.7. The advantage value for adding a relationship is 0.8 - 0.6 = 0.2; the advantage value for deleting a node is 0.5 - 0.6 = -0.1; and the advantage value for replacing a node is 0.7 - 0.6 = 0.1. The base probability distribution is adjusted based on the advantage values ​​of all candidate actions to obtain a new probability distribution. Generally, a softmax function can be used to give actions with larger advantage values ​​a higher probability. The base reward value is calculated based on a preset reward function. This reward function can be designed according to the specific task; for example, adding an action that more reasonably explains the drug-disease relationship would receive a higher reward. The comprehensive reward value for each candidate action is calculated based on the new probability distribution and the base reward value. The comprehensive reward value takes into account both the initial probability of the action and the potential reward for performing that action.For example, using a soft maximization function to adjust the base probability distribution [0.4, 0.3, 0.3] for the advantage value [0.2, -0.1, 0.1], we obtain a new probability distribution [0.45, 0.25, 0.3]. Assuming the preset reward function stipulates that adding a reasonable relationship rewards 1 point, deleting a node rewards 0 points, and replacing a node rewards 0.5 points, then the base reward value is [1, 0, 0.5]. The comprehensive reward value = new probability distribution × base reward value (this can be understood as multiplying corresponding elements and then summing them; simply put, the comprehensive reward value for each action is the product of the new probability and the base reward). Therefore, the comprehensive reward value for adding a relationship is 0.45 × 1 = 0.45; the comprehensive reward value for deleting a node is 0.25 × 0 = 0; and the comprehensive reward value for replacing a node is 0.3 × 0.5 = 0.15. The target action with the highest comprehensive reward value is selected. If the target action is adding a relationship, a relationship edge is added between the second target node and the original target node. The original target node is the node directly connected to the second target node. If the target action is node replacement, then the original target node is replaced with its nearest neighbor. The nearest neighbor is the directly connected node with the highest degree of association with the original target node, excluding the second target node. The degree of association can be measured by calculating the weight of the edges between nodes, the number of common neighbors, etc.

[0091] Optionally, replacing the original target node with its nearest neighbor node includes:

[0092] Determine the set of neighboring nodes of the original target node, the set of neighboring nodes containing all entities that are directly connected to the original target node and are not the second target node;

[0093] Obtain the number of shared relationship types between the target neighbor node and the original target node, and calculate the ratio of the number of shared relationship types to the total number of relationship types of the original target node as the structural association degree. The target neighbor node is any node in the set of neighbor nodes.

[0094] Obtain the historical interaction frequency between the target neighbor node and the original target node, and calculate the functional correlation degree based on the historical interaction frequency;

[0095] The total correlation degree is calculated by weighting the structural correlation degree and the functional correlation degree, and the neighbor node with the highest total correlation degree is selected as the nearest target neighbor node.

[0096] The original target node in the initial path is replaced with the nearest target neighbor node to form a reconstructed path. The global reliability of the reconstructed path is calculated. If the global reliability is greater than a second preset threshold, the reconstructed path is retained.

[0097] In a knowledge graph, identify all entities directly connected to the original target node that are not the second target node. The set of these entities is the neighbor node set. This is to exclude the second target node (because subsequent association calculations are based on other neighbor nodes) and focus on nodes that might be used for replacement. Suppose an initial path in the drug-disease knowledge graph is "Drug A - Target B (Second Target Node) - Disease C (Original Target Node)," and Disease C is also directly connected to "Symptom D" and "Treatment E." Then the neighbor node set for Disease C is {Symptom D, Treatment E}. For any target neighbor node in the neighbor node set, obtain the number of shared relation types between it and the original target node. Relation types in a knowledge graph represent different connections between entities, such as "treatment," "association," and "impact." Calculate the total number of relation types for the original target node, i.e., the total number of relation types the original target node is connected to all other nodes in the knowledge graph. Divide the number of shared relation types by the total number of relation types to obtain the structural association degree. The structural association degree reflects the similarity between the target neighbor node and the original target node in the graph structure. Continuing the example above, disease C has a total of 3 relation types (relation types connected to target B, symptom D, and treatment E). Assume symptom D shares 2 relation types with disease C (e.g., "cause" and "accompany"), and treatment E shares 1 relation type with disease C ("used for treatment"). For the target neighbor node symptom D, its structural correlation degree = 2 ÷ 3 ≈ 0.67. For the target neighbor node treatment E, its structural correlation degree = 1 ÷ 3 ≈ 0.33. Obtain the interaction frequency between the target neighbor node and the original target node in historical data. Historical interaction frequency can be calculated by counting the number of times these two nodes appear simultaneously in past studies, case reports, etc., recorded in the database. Based on the historical interaction frequency, obtain the functional correlation degree through a preset calculation method (e.g., directly normalizing the interaction frequency). Functional correlation degree reflects the closeness of the connection between the target neighbor node and the original target node at the actual functional or application level. Suppose that in past medical research, disease C and symptom D were mentioned together 50 times, and disease C and treatment E were mentioned together 20 times. For ease of comparison, the interaction frequency is normalized to the range of 0-1, assuming a maximum interaction frequency of 100 times. After normalization: the functional relevance of symptom D = 50 ÷ 100 = 0.5. The functional relevance of treatment E = 20 ÷ 100 = 0.2. Based on pre-defined weights for structural and functional relevance (the weights can be determined based on specific tasks and experience, for example, a structural relevance weight of 0.6 and a functional relevance weight of 0.4), a weighted average is calculated to obtain the total relevance.The neighbor node with the highest total relevance is selected as the nearest target neighbor node. This node is most similar to or most closely related to the original target node in both structure and function. Using a structural relevance weight of 0.6 and a functional relevance weight of 0.4, the total relevance of symptom D is calculated as follows: 0.67 × 0.6 + 0.5 × 0.4 = 0.402 + 0.2 = 0.602. The total relevance of treatment method E is calculated as follows: 0.33 × 0.6 + 0.2 × 0.4 = 0.198 + 0.08 = 0.278. Since 0.602 > 0.278, symptom D is selected as the nearest target neighbor node. The original target node in the initial path is replaced with the nearest target neighbor node to form the reconstructed path. The global reliability of the reconstructed path is calculated. Global reliability can be calculated in various ways, such as considering the credibility of all nodes and edges in the path, and the logical rationality of the path. If the global reliability is greater than the second preset threshold (this threshold is set based on actual needs and path quality requirements), the reconstructed path is retained; otherwise, it may be necessary to reselect replacement nodes or abandon the replacement operation. Replace "Disease C" with "Symptom D" in the initial path "Drug A-Target B-Disease C" to obtain the reconstructed path "Drug A-Target B-Symptom D". Assuming that the global reliability of this reconstructed path is calculated to be 0.8 through a comprehensive evaluation of the credibility and logicality of each element in the path, and the second preset threshold is set to 0.7, because 0.8 > 0.7, this reconstructed path is retained.

[0098] Optionally, the step of obtaining a correlation score based on the molecular structural characteristics and the original indication through the indication expansion analysis agent, generating an indication expansion list based on the correlation score, and calculating the indication expansion analysis contribution value includes:

[0099] Based on the molecular structure of the target drug, a molecular graph embedding vector is generated;

[0100] Based on the disease-pathway knowledge graph, the biological pathway nodes associated with the original indication are located, and the pathway feature vectors of the biological pathway nodes are extracted.

[0101] Calculate the semantic similarity between the molecular graph embedding vector and the pathway feature vector, and obtain an association score based on the semantic similarity;

[0102] Filter out candidate diseases whose correlation scores exceed a third preset threshold, and generate an expanded list of indications based on the candidate diseases;

[0103] The pathways that the target drug may regulate in the candidate disease are identified as new pathways. The total pathways are determined based on the new pathways and the pathways that the target drug has been verified to regulate in the original indication. The proportion of the new pathways to the total pathways is calculated.

[0104] The contribution value of the indication expansion analysis is calculated based on the stated proportion and the stated correlation score.

[0105] The molecular structure of the target drug is transformed into a vector form, namely a molecular graph embedding vector. The molecular structure contains key information such as the types and connections of atoms. This method converts the complex molecular structure into a numerical vector that a computer can process and compute, facilitating subsequent operations such as similarity comparisons with other vectors. Assuming the target drug is a small molecule compound with a specific cyclic structure and multiple functional groups, a graph neural network (GNN) model is used. The molecular structure is input into the model, and after multi-layer neural network computation and feature extraction, a 128-dimensional molecular graph embedding vector is output, for example, [0.23, -0.15, 0.47, ..., 0.12], where each dimension represents a feature of the molecular structure in different aspects. Based on a pre-defined disease-pathway knowledge graph, which integrates the association information between diseases and related biological pathways, biological pathway nodes associated with the original indication are found in the graph, and the corresponding pathway feature vectors are extracted from the graph. If the original indication is a certain type of cancer, the biological pathways related to this cancer are searched in the disease-pathway knowledge graph, revealing associations such as "cell proliferation regulation pathway" and "DNA damage repair pathway". For the "cell proliferation regulation pathway" node, its corresponding pathway feature vector is extracted, assumed to be a 64-dimensional vector, such as [0.35, 0.12, -0.28, ..., 0.09]. A suitable similarity calculation method is used to calculate the semantic similarity between the molecular graph embedding vector and the pathway feature vector. Semantic similarity measures the proximity of two vectors in semantic space, reflecting the closeness of the association between the target drug's molecular structure and a specific biological pathway. Based on the calculated semantic similarity, an association score is obtained according to certain rules or formulas; a higher score indicates a greater likelihood of the drug being associated with that pathway and related diseases. The cosine similarity algorithm can be used to calculate the similarity between the molecular graph embedding vector and the "cell proliferation regulation pathway" feature vector. A third preset threshold is set to filter out candidate diseases whose association scores exceed this threshold. These candidate diseases are considered to have potential indications due to their high association with the target drug at the molecular structure and pathway levels. Assuming the third preset threshold is set to 60 points, the indication expansion list includes two diseases, such as ["Disease A (related to the cell proliferation regulation pathway)", "Disease B (related to another pathway)"]. We analyze the pathways that the target drug may regulate in candidate diseases and identify them as new pathways. Simultaneously, we identify pathways that the target drug has already validated its regulation in the original indication. We combine the new pathways and the validated regulated pathways to form the total pathway, and calculate the proportion of the new pathway in the total pathway. This proportion reflects the extent to which the target drug expands its pathway regulation in the potential indication (candidate disease) relative to the original indication.The target drug has been validated to regulate three pathways in its original indication (a certain type of cancer): the cell proliferation regulation pathway, the DNA damage repair pathway, and the apoptosis pathway. In disease A on the indication expansion list, two new pathways were found to be potentially regulated by the target drug: the cell migration regulation pathway and the angiogenesis regulation pathway. The total number of pathways is 3 + 2 = 5, and the proportion of new pathways is 2 ÷ 5 = 0.4. Based on the previously calculated proportion of new pathways and the correlation score, the contribution value of the indication expansion analysis is calculated using a weighted formula. The weight of the proportion of new pathways is set to 0.6, and the weight of the correlation score is set to 0.4. For disease A, the proportion of new pathways is 0.4, and the correlation score is 72 (converted to decimal form, 0.72). Therefore, the contribution value of the indication expansion analysis = 0.6 × 0.4 + 0.4 × 0.72 = 0.24 + 0.288 = 0.528. This contribution value can be compared with the contribution values ​​of other potential diseases to identify diseases more likely to become new indications for the target drug.

[0106] Optionally, the step of calculating the structural similarity of new targets based on the expanded indication list and drug-target data using the target matching analysis agent, generating a new target list based on the structural similarity of the new targets, and calculating the target matching analysis contribution value includes:

[0107] The protein sequences of the target sites of the candidate diseases are analyzed, and three-dimensional structural embedding vectors are generated through a preset geometric deep learning model.

[0108] Functional domain similarity is determined based on the molecular graph embedding vector, structural similarity is determined based on the three-dimensional structure embedding vector, and the total similarity is calculated by weighting the functional domain similarity and the structural similarity.

[0109] Filter target points whose total similarity is greater than a fourth preset threshold, and generate a new target point list based on the target points;

[0110] Based on a preset protein-protein interaction network, the normalized betweenness centrality of the target is calculated, and the target matching analysis contribution value is calculated based on the normalized betweenness centrality and preset weights according to a preset contribution value quantification formula.

[0111] Protein sequences, containing the order of amino acids, are fundamental to protein structure and function. A pre-defined geometric deep learning model captures the three-dimensional structural features of proteins, transforming the protein sequence into a three-dimensional structural embedding vector. Using the previously generated molecular graph embedding vector of the target drug, a pre-defined algorithm or model (such as a functional domain annotation-based similarity comparison algorithm) is employed to determine the functional domain similarity between the drug and the target. Functional domain similarity reflects the degree of functional association between the target drug and the target. Assuming the molecular graph embedding vector of the target drug reflects functional features related to a certain signal transduction, comparison reveals that the target protein also possesses a functional domain related to this signal transduction. The calculated functional domain similarity is 0.6 (typically ranging from 0 to 1, with values ​​closer to 1 indicating higher similarity). Based on the previously generated three-dimensional structural embedding vector, an appropriate similarity metric (such as cosine similarity or Euclidean distance) is used to calculate the structural similarity between the target drug and the target disease. Cosine similarity is used to calculate the similarity between the molecular graph embedding vector of the target drug and the 3D structural embedding vector of the candidate disease target. Based on pre-defined weights for functional domain similarity and structural similarity (e.g., functional domain similarity weight 0.4, structural similarity weight 0.6), a weighted average is calculated to obtain the total similarity. The total similarity comprehensively reflects the degree of functional and structural association between the target drug and the candidate disease target. A fourth pre-defined threshold is set to screen target targets with a total similarity greater than the fourth pre-defined threshold. These target targets, due to their high functional and structural similarity to the target drug, are considered potential new targets. The screened target targets are then used to generate a new target list according to certain rules (e.g., similarity from high to low). A pre-defined protein-protein interaction network is used, which describes the interactions between proteins within the cell. Betweenness centrality is an indicator of a protein's importance in the network, reflecting its criticality in information transmission and functional regulation. The betweenness centrality of the target site in the protein-protein interaction network is calculated and standardized (e.g., by dividing the betweenness centrality by the maximum betweenness centrality of all nodes in the network, ensuring the value is between 0 and 1), resulting in a standardized betweenness centrality. Based on the standardized betweenness centrality and preset weights (e.g., a standardized betweenness centrality weight of 0.3 and a total similarity weight of 0.7), the target matching analysis contribution value is calculated according to a preset contribution value quantification formula (e.g., contribution value = standardized betweenness centrality × its weight + total similarity × its weight). The target matching analysis contribution value comprehensively measures the importance and potential of the target site in target prediction.

[0112] Optionally, the step of calling the molecular docking simulation platform interface through the pharmacodynamic simulation verification agent according to the new target list, performing molecular docking to generate a pharmacodynamic scoring vector, and calculating the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector includes:

[0113] Molecular docking is performed to generate multimodal parameters, including binding energy, conformational stability score, number of hydrogen bonds and hydrophobic contact area, and a drug efficacy score vector is generated based on the multimodal parameters.

[0114] The sub-vectors of each dimension in the efficacy score vector are normalized, and the weighted sum of the normalized sub-vectors is performed to obtain the basic contribution value.

[0115] The delivery efficiency coefficient is calculated based on the hierarchical depth of the target in the disease pathway, and the basic contribution value is adjusted based on the delivery efficiency coefficient to obtain the efficacy simulation verification contribution value.

[0116] Computer simulations are used to perform molecular docking between potential drug molecules and each target in a new target list. Molecular docking simulates the binding process of drug molecules and target proteins in three-dimensional space. By calculating the interaction forces between the two, the stability and affinity of their binding are predicted. Binding energy is an important indicator of the stability of drug molecule binding to the target, reflecting the energy released or absorbed during binding. The lower the binding energy, the more stable the drug molecule binding to the target and the stronger the interaction. Conformational stability scoring is used to assess the conformational stability of drug molecules at the target binding site. By analyzing the conformational changes of drug molecules after docking, and comprehensively considering factors such as atomic positions, bond lengths, and bond angles, a comprehensive stability score is given. The higher the score, the more stable the conformation of the drug molecule. Hydrogen bonds are one of the important interaction forces between drug molecules and targets, and they have a significant impact on the specificity and stability of binding. The number of hydrogen bonds formed between drug molecules and targets is counted; more hydrogen bonds generally indicate a stronger interaction. Hydrophobic interactions are also an important driving force for drug molecule binding to targets. The hydrophobic contact area between the drug molecule and the target is calculated. A larger hydrophobic contact area indicates a stronger hydrophobic interaction, which is beneficial for the binding of the drug molecule to the target. The four multimodal parameters obtained from the above calculations—binding energy, conformational stability score, number of hydrogen bonds, and hydrophobic contact area—are combined into a vector in a certain order, namely the efficacy score vector. This vector comprehensively reflects multiple aspects of the interaction between the drug molecule and the target. For example, the constructed efficacy score vector can be [-8.5, 0.8, 3, 120]. Since the parameters of each dimension in the efficacy score vector have different dimensions and numerical ranges, the sub-vectors of each dimension need to be normalized to make them comparable. According to the importance of each parameter in evaluating efficacy, corresponding weights are assigned to the sub-vectors of each dimension. Then, the weighted summation of the normalized sub-vectors is performed to obtain the basic contribution value. The hierarchical depth of the target in the disease pathway reflects its importance in the disease occurrence and development process and its key position in signal transduction. Generally speaking, the deeper the target is, the more critical its role is, and the higher the signal transduction efficiency generated after the drug molecule interacts with it. Based on the target's hierarchical depth in the disease pathway, a delivery efficiency coefficient is calculated using a specific algorithm or model. For example, a simple rule can be established where the delivery efficiency coefficient is multiplied by a fixed factor (e.g., 1.2) for each additional level of hierarchical depth. Multiplying the calculated delivery efficiency coefficient by the baseline contribution value yields the pharmacodynamic simulation validation contribution value. This contribution value comprehensively considers the interaction between the drug molecule and the target, as well as the target's importance in the disease pathway, thus more accurately reflecting the target's potential value as a drug target.

[0117] This embodiment also discloses a system for generating drug repositioning study reports. Figure 2This is a schematic diagram of the modules of the drug repositioning study report generation system disclosed in the embodiments of this application, such as... Figure 2 As shown, the system includes a decomposition module 201, a discovery module 202, a prediction module 203, a verification module 204, a calculation module 205, and an execution module 206, wherein:

[0118] The decomposition module 201 is configured to respond to a request for target drug relocation submitted by a pharmaceutical company, decompose the request into an indication expansion analysis subtask, a target matching analysis subtask, and a pharmacodynamic simulation verification subtask based on the molecular structure characteristics and original indications of the target drug provided by the pharmaceutical company, assign initial weights to each subtask according to the request, and call the indication expansion analysis intelligent agent, the target matching analysis intelligent agent, and the pharmacodynamic simulation verification intelligent agent from a preset intelligent agent library according to the request.

[0119] Discovery module 202 is configured to trigger indication expansion analysis agent through the indication expansion analysis subtask, calculate the correlation score between the target drug and potential indications based on a preset disease-pathway knowledge graph, generate an indication expansion list based on the correlation score and calculate the indication expansion analysis contribution value, write the indication expansion list to shared storage and trigger target matching analysis agent;

[0120] Prediction module 203 is configured to calculate the structural similarity of new targets based on the indication expansion list and drug-target data through the target matching analysis agent, generate a new target list based on the structural similarity of new targets and calculate the target matching analysis contribution value, write the new target list to shared storage and trigger the drug efficacy simulation verification agent;

[0121] Verification module 204 is configured to generate a pharmacodynamic scoring vector by performing molecular docking based on the new target list through the pharmacodynamic simulation verification agent, and calculate the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector;

[0122] The calculation module 205 is configured to perform a weighted summation of the contribution values ​​of the pharmacodynamic simulation verification, the indication expansion analysis, and the target matching analysis based on the initial weights to obtain a candidate disease association score.

[0123] The execution module 206 is configured to generate multi-hop paths based on the target drug, the candidate disease association scores, and a preset drug-disease knowledge graph, perform potential rating on each path in the multi-hop paths, calculate the confidence level of each path, and generate a research report on the relocation of the target drug based on the potential rating and the confidence level.

[0124] It should be noted that the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0125] This embodiment also discloses an electronic device, as shown in the reference. Figure 3 The electronic device may include: at least one processor 301, at least one communication bus 302, user interface 303, network interface 304, and at least one memory 305.

[0126] The communication bus 302 is used to enable communication between these components.

[0127] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0128] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0129] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and by calling data stored in the memory 305.

[0130] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include non-transitory computer-readable storage medium. The memory 305 may be used to store instructions, programs, code, code sets, or instruction sets. Figure 3 As shown, the memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface 303 module, and an application program for generating drug repositioning research reports.

[0131] exist Figure 3 In the electronic device shown, the user interface 303 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 301 can be used to call the application program storing the method for generating drug repositioning research reports in the memory 305. When executed by one or more processors 301, the electronic device performs one or more methods as described in the above embodiments.

[0132] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the disclosure in this specification. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure.

Claims

1. A method for generating a drug repositioning study report, characterized in that, Applied to a research report generation platform, the method includes: In response to receiving a request from a pharmaceutical company for target drug relocation, the request is decomposed into an indication expansion analysis subtask, a target matching analysis subtask, and a pharmacodynamic simulation verification subtask based on the molecular structure characteristics and original indications of the target drug provided by the pharmaceutical company. An initial weight is assigned to each subtask according to the request, and the indication expansion analysis agent, target matching analysis agent, and pharmacodynamic simulation verification agent are called from a preset agent library according to the request. The indication expansion analysis subtask triggers the indication expansion analysis agent, which calculates the correlation score between the target drug and the potential indication based on a preset disease-pathway knowledge graph. The agent generates an indication expansion list based on the correlation score and calculates the indication expansion analysis contribution value. The indication expansion list is then written to shared storage, and the target matching analysis agent is triggered. The target matching analysis agent calculates the structural similarity of new targets based on the indication expansion list and drug-target data, generates a new target list based on the structural similarity of new targets, calculates the target matching analysis contribution value, writes the new target list into shared storage, and triggers the drug efficacy simulation verification agent. The pharmacodynamic simulation verification agent calls the molecular docking simulation platform interface according to the new target list, performs molecular docking to generate a pharmacodynamic scoring vector, and calculates the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector. The candidate disease association score is obtained by weighting and summing the contribution values ​​of the pharmacodynamic simulation validation, the indication expansion analysis, and the target matching analysis based on the initial weights. Multi-hop paths are generated based on the target drug, the candidate disease association scores, and a pre-defined drug-disease knowledge graph. Each path in the multi-hop paths is rated for potential, and the confidence level of each path is calculated. A research report on the relocation of the target drug is generated based on the potential rating and the confidence level. The step of generating a multi-hop path based on the target drug, the candidate disease association score, and the preset drug-disease knowledge graph includes: A candidate disease set is formed based on all candidate diseases whose association scores are greater than a first preset threshold. A subgraph of the target candidate disease is located in the preset drug-disease knowledge graph. The target candidate disease is any one of the candidate diseases in the candidate disease set. The subgraph includes pathological pathways, targets and drug entities that are directly connected to the first target node corresponding to the target candidate disease. Based on the subgraph, backtracking from the first target node to the target drug entity along the relational edges of the preset drug-disease knowledge graph to form an initial path set; A second target node with a node centrality higher than a preset centrality is identified in the initial path set. An action probability distribution for the second target node is generated using a preset policy network. The optimal action is selected based on a preset reward function. The initial paths in the initial path set are then adjusted according to the optimal action to obtain multi-hop paths. The process of calculating the correlation score between the target drug and potential indications based on the disease-pathway knowledge graph using the indication expansion analysis agent, generating an indication expansion list based on the correlation score, and calculating the indication expansion analysis contribution value includes: Based on the molecular structure of the target drug, a molecular graph embedding vector is generated; Based on the disease-pathway knowledge graph, the biological pathway nodes associated with the original indication are located, and the pathway feature vectors of the biological pathway nodes are extracted. Calculate the semantic similarity between the molecular graph embedding vector and the pathway feature vector, and obtain the association score based on the semantic similarity; Filter out candidate diseases whose correlation scores exceed a third preset threshold, and generate an expanded list of indications based on the candidate diseases; The pathways that the target drug may regulate in the candidate disease are identified as new pathways. The total pathways are determined based on the new pathways and the pathways that the target drug has been verified to regulate in the original indication. The proportion of the new pathways to the total pathways is calculated. The contribution value of the indication expansion analysis is calculated based on the stated proportion and the stated correlation score.

2. The method for generating a drug repositioning study report according to claim 1, characterized in that, The step of generating the action probability distribution of the second target node through a preset policy network, selecting the optimal action in combination with a preset reward function, and adjusting the initial paths in the initial path set according to the optimal action to obtain a multi-hop path includes: An embedding vector for the second target node is generated using a graph neural network, and a basic probability distribution for the candidate action is generated based on the embedding vector. The candidate action is one of all allowed actions. Calculate the state value of the second target node and the action value of the candidate action, and generate an advantage value based on the action value and the state value; The probability distribution is obtained by adjusting the basic probability distribution based on the advantage value of all the candidate actions, the basic reward value is calculated based on the preset reward function, and the comprehensive reward value of each candidate action is obtained based on the probability distribution and the basic reward value. Select the target action with the highest overall reward value. If the target action is to add a relationship, then add a relationship edge between the second target node and the original target node. If the target action is to replace a node, then replace the original target node with the nearest neighbor node of the original target node. The original target node is the node directly connected to the second target node, and the nearest neighbor node is the node with the highest correlation to the original target node, excluding the second target node.

3. The method for generating a drug repositioning study report according to claim 2, characterized in that, The step of replacing the original target node with its nearest neighbor node includes: Determine the set of neighboring nodes of the original target node, the set of neighboring nodes containing all entities that are directly connected to the original target node and are not the second target node; Obtain the number of shared relationship types between the target neighbor node and the original target node, and calculate the ratio of the number of shared relationship types to the total number of relationship types of the original target node as the structural association degree. The target neighbor node is any node in the set of neighbor nodes. Obtain the historical interaction frequency between the target neighbor node and the original target node, and calculate the functional correlation degree based on the historical interaction frequency; The total correlation degree is calculated by weighting the structural correlation degree and the functional correlation degree, and the neighbor node with the highest total correlation degree is selected as the nearest target neighbor node. The original target node in the initial path is replaced with the nearest target neighbor node to form a reconstructed path. The global reliability of the reconstructed path is calculated. If the global reliability is greater than a second preset threshold, the reconstructed path is retained.

4. The method for generating a drug repositioning study report according to claim 1, characterized in that, The process of calculating the structural similarity of new targets based on the expanded indication list and drug-target data using the target matching analysis agent, generating a new target list based on the structural similarity of the new targets, and calculating the target matching analysis contribution value includes: The protein sequences of the target sites of the candidate diseases are analyzed, and three-dimensional structural embedding vectors are generated through a preset geometric deep learning model. Functional domain similarity is determined based on the molecular graph embedding vector, structural similarity is determined based on the three-dimensional structure embedding vector, and the total similarity is calculated by weighting the functional domain similarity and the structural similarity. Filter target points whose total similarity is greater than a fourth preset threshold, and generate a new target point list based on the target points; Based on a preset protein-protein interaction network, the normalized betweenness centrality of the target is calculated, and the target matching analysis contribution value is calculated based on the normalized betweenness centrality and preset weights according to a preset contribution value quantification formula.

5. The method for generating a drug repositioning study report according to claim 4, characterized in that, The process of the pharmacodynamic simulation verification agent calling the molecular docking simulation platform interface based on the new target list to perform molecular docking and generate a pharmacodynamic scoring vector, and calculating the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector, includes: Molecular docking is performed to generate multimodal parameters, including binding energy, conformational stability score, number of hydrogen bonds and hydrophobic contact area, and a drug efficacy score vector is generated based on the multimodal parameters. The sub-vectors of each dimension in the efficacy score vector are normalized, and the weighted sum of the normalized sub-vectors is performed to obtain the basic contribution value. The delivery efficiency coefficient is calculated based on the hierarchical depth of the target in the disease pathway, and the basic contribution value is adjusted based on the delivery efficiency coefficient to obtain the efficacy simulation verification contribution value.

6. A system for generating drug repositioning study reports, characterized in that, The method for generating a drug repositioning study report as described in any one of claims 1-5 includes a decomposition module, a discovery module, a prediction module, a validation module, a calculation module, and an execution module, wherein: The decomposition module is configured to respond to a request for target drug relocation submitted by a pharmaceutical company, decompose the request into an indication expansion analysis subtask, a target matching analysis subtask, and a pharmacodynamic simulation verification subtask based on the molecular structure characteristics and original indications of the target drug provided by the pharmaceutical company, assign initial weights to each subtask according to the request, and call the indication expansion analysis intelligent agent, target matching analysis intelligent agent, and pharmacodynamic simulation verification intelligent agent from a preset intelligent agent library according to the request. The discovery module is configured to trigger the indication expansion analysis agent through the indication expansion analysis subtask, calculate the correlation score between the target drug and the potential indication based on a preset disease-pathway knowledge graph, generate an indication expansion list based on the correlation score and calculate the indication expansion analysis contribution value, write the indication expansion list to shared storage and trigger the target matching analysis agent; The prediction module is configured to calculate the structural similarity of new targets based on the indication expansion list and drug-target data through the target matching analysis agent, generate a new target list based on the structural similarity of new targets and calculate the target matching analysis contribution value, write the new target list to shared storage and trigger the drug efficacy simulation verification agent; The verification module is configured to use the pharmacodynamic simulation verification agent to perform molecular docking to generate a pharmacodynamic scoring vector based on the new target list, and to calculate the pharmacodynamic simulation verification contribution value based on the pharmacodynamic scoring vector. The calculation module is configured to perform a weighted summation of the contribution values ​​of the pharmacodynamic simulation verification, the indication expansion analysis, and the target matching analysis based on the initial weights to obtain a candidate disease association score. The execution module is configured to generate multi-hop paths based on the target drug, the candidate disease association scores, and a preset drug-disease knowledge graph; to perform a potential rating on each path in the multi-hop paths; to calculate the confidence level of each path; and to generate a research report on the relocation of the target drug based on the potential rating and the confidence level. The step of generating a multi-hop path based on the target drug, the candidate disease association score, and the preset drug-disease knowledge graph includes: A candidate disease set is formed based on all candidate diseases whose association scores are greater than a first preset threshold. A subgraph of the target candidate disease is located in the preset drug-disease knowledge graph. The target candidate disease is any one of the candidate diseases in the candidate disease set. The subgraph includes pathological pathways, targets and drug entities that are directly connected to the first target node corresponding to the target candidate disease. Based on the subgraph, backtracking from the first target node to the target drug entity along the relational edges of the preset drug-disease knowledge graph to form an initial path set; A second target node with a node centrality higher than a preset centrality is identified in the initial path set. An action probability distribution for the second target node is generated using a preset policy network. The optimal action is selected based on a preset reward function. The initial paths in the initial path set are then adjusted according to the optimal action to obtain multi-hop paths. The process of calculating the correlation score between the target drug and potential indications based on the disease-pathway knowledge graph using the indication expansion analysis agent, generating an indication expansion list based on the correlation score, and calculating the indication expansion analysis contribution value includes: Based on the molecular structure of the target drug, a molecular graph embedding vector is generated; Based on the disease-pathway knowledge graph, the biological pathway nodes associated with the original indication are located, and the pathway feature vectors of the biological pathway nodes are extracted. Calculate the semantic similarity between the molecular graph embedding vector and the pathway feature vector, and obtain the association score based on the semantic similarity; Filter out candidate diseases whose correlation scores exceed a third preset threshold, and generate an expanded list of indications based on the candidate diseases; The pathways that the target drug may regulate in the candidate disease are identified as new pathways. The total pathways are determined based on the new pathways and the pathways that the target drug has been verified to regulate in the original indication. The proportion of the new pathways to the total pathways is calculated. The contribution value of the indication expansion analysis is calculated based on the stated proportion and the stated correlation score.

7. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-5.

Images