Method and system for evaluating curative effect of traditional Chinese medicine prescription generated by large model based on network pharmacology

Abstract

The invention provides a network pharmacology-based large model generated traditional Chinese medicine prescription curative effect evaluation method and system, and the method comprises the steps: employing a machine learning model to generate candidate prescriptions, and obtaining all known active components of each traditional Chinese medicine in the prescriptions; predicting human body protein targets corresponding to the patient based on the known active components, and gathering all the human body protein targets to obtain a prescription target set P; obtaining human body protein targets related to the corresponding diseases, and gathering all the human body protein targets corresponding to the diseases to obtain a disease target set D; constructing a protein-protein interaction network; inputting the prescription target point set P and the disease target point set D into a protein-protein interaction network, and constructing a connected sub-network; calculating curative effect indexes of the connected sub-networks; and optimizing the prescription based on the curative effect index to complete the curative effect evaluation of the traditional Chinese medicine prescription. According to the method, calculation and experiments are closely combined, the blindness and the cost of experimental verification are remarkably reduced through a calculation priority strategy, and the research and development efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, computational medicine and bioinformatics, and specifically relates to a method and system for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology. Background Technology

[0002] In recent years, machine learning (ML) and large language models (LLMs) have been explored for the automatic generation of TCM prescriptions. Currently, large language models (LLMs) used for automatic TCM prescription generation can be categorized into general-purpose models and domain-specific models. General-purpose models, such as ChatGPT, possess powerful language generation capabilities but have limitations in the TCM field. Domain-specific models, on the other hand, are optimized for the TCM domain. For example, the Wuji TCM large model, through fine-tuning and training on specialized datasets, has improved the accuracy and practicality of prescription recommendations. However, existing LLMs still face shortcomings in TCM prescription generation, such as inaccurate understanding of TCM terminology, lack of real-world diagnostic experience, and poor model interpretability. Nevertheless, how to objectively and quickly evaluate the potential efficacy and scientific validity of these AI-generated prescriptions remains a key bottleneck.

[0003] Currently, the evaluation of large-scale TCM models largely relies on TCM test sets and expert scoring, lacking unified standards and comprehensive measurement of model performance. This limits their application and promotion in clinical practice, and there is a lack of a computationally driven pre-screening and evaluation system that does not depend on expensive wet laboratory experiments. TCM prescriptions (formulas) follow complex theories such as "principal, assistant, adjuvant, and guide," and their "multi-component, multi-target, and holistic regulatory" therapeutic mechanisms remain unclear. Traditional experimental methods (such as molecular biology) are costly and time-consuming to verify all mechanisms, making it difficult to systematically and quantitatively evaluate prescription efficacy. Network pharmacology offers a new perspective for understanding the mechanisms of action of TCM. It elucidates pharmacology at the systemic level by constructing a "drug component-target-disease" network. However, existing methods mostly focus on network construction and simple topology analysis, lacking quantitative, standardized, and comparable efficacy measurements of the overall network behavior of the prescription. Therefore, there is an urgent need in this field for a method that can map the components of TCM prescriptions to biological networks and quantify their therapeutic potential by calculating network parameters, thereby efficiently evaluating and optimizing TCM prescriptions (including those generated by large-scale models). Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a method for evaluating the efficacy of traditional Chinese medicine (TCM) prescriptions based on a large-scale model generated from network pharmacology. This method includes: generating candidate prescriptions using a machine learning model to obtain all known active ingredients of each TCM herb in the prescription; predicting the corresponding human protein targets based on the known active ingredients, and aggregating all human protein targets to obtain a prescription target set P; obtaining human protein targets related to the corresponding diseases, and aggregating all disease-related human protein targets to obtain a disease target set D; constructing a protein-protein interaction network; inputting the prescription target set P and the disease target set D into the protein-protein interaction network to construct a connected subnetwork; calculating the efficacy index of the connected subnetwork; and optimizing the prescription based on the efficacy index to complete the TCM prescription efficacy evaluation.

[0005] This invention also proposes a traditional Chinese medicine prescription efficacy evaluation system based on network pharmacology, which includes: a data acquisition module, a network construction and calculation module, a prescription evaluation and optimization module, and a user interface;

[0006] The data acquisition module retrieves component, target, and PPI data from the database;

[0007] The network construction and calculation module is used to construct the PPI network and perform network distance and Z-score calculations.

[0008] The prescription evaluation and optimization module is used to compare the efficacy indicators of different prescriptions and provide optimization suggestions;

[0009] The user interface is used to input prescriptions and disease information, and to visualize the output results.

[0010] The beneficial effects of this invention are:

[0011] This invention proposes for the first time a comprehensive and computable index system (statistical distance, shortest distance, Z-score) to quantify the overall therapeutic potential of traditional Chinese medicine prescriptions into a comparable numerical value. Through network analysis, this invention not only evaluates efficacy but also reveals the possible action modules and key targets of the prescription, providing computational evidence for understanding the scientific connotations of the "principal, assistant, adjuvant, and guide" principles in traditional Chinese medicine. This invention closely integrates computation and experimentation, significantly reducing the blind spots and costs of experimental verification through a computation-first strategy, thereby improving research and development efficiency. Attached Figure Description

[0012] Figure 1 This is an overall flowchart of the present invention;

[0013] Figure 2 This is a conceptual diagram of network distribution and shortest path in this invention;

[0014] Figure 3This is a schematic diagram of the Z-score calculation in this invention;

[0015] Figure 4 This invention relates to the overall interaction and data flow between the system and external entities.

[0016] Figure 5 This is an exploded view of the data processing procedure of the present invention;

[0017] Figure 6 This is a diagram illustrating the dynamic interaction process of task submission and data preparation during the prescription evaluation task of the present invention.

[0018] Figure 7 This is a diagram illustrating the dynamic interaction process between core calculations and result evaluation during the prescription evaluation task of this invention.

[0019] Figure 8 A PPT network diagram illustrating the therapeutic efficacy of the two prescription generation methods of this invention;

[0020] Figure 9 This is a structural diagram of the large-scale model capability assessment framework based on network efficacy quantification of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] A method for evaluating the efficacy of traditional Chinese medicine (TCM) prescriptions based on a large-scale model generated by network pharmacology is proposed. The method includes: generating candidate prescriptions using a machine learning model to obtain all known active ingredients of each TCM herb in the prescription; predicting the corresponding human protein targets for the patient based on the known active ingredients, and aggregating all human protein targets to obtain a prescription target set P; obtaining human protein targets related to the corresponding diseases, and aggregating all disease-related human protein targets to obtain a disease target set D; constructing a protein-protein interaction network; inputting the prescription target set P and the disease target set D into the protein-protein interaction network to construct a connected subnetwork; calculating the efficacy index of the connected subnetwork; and optimizing the prescription based on the efficacy index to complete the TCM prescription efficacy evaluation.

[0023] In this embodiment, predicting the corresponding human protein targets for patients based on known active ingredients includes:

[0024] Step 1: Obtain the active ingredients. For each Chinese herb in the candidate prescription, obtain all known active ingredient molecules from Chinese herbal medicine chemical composition databases (such as TCMSP, TCMID, BATMAN-TCM). The active ingredients are usually screened based on parameters such as oral bioavailability (OB) and drug-likeness (DL).

[0025] Step 2: Predict protein targets. Based on all active ingredients obtained in Step 1, predict their interactions with the corresponding human protein targets in the patient. This prediction process includes: ① Molecular structure standardization: converting the molecular structure of each active ingredient into a standardized molecular descriptor or fingerprint. ② Target prediction calculation: inputting the standardized molecular descriptor into a pre-constructed target prediction model to calculate its interaction probability or binding affinity with a series of human protein targets. The target prediction model can be a similarity-based prediction model (e.g., through the SwissTargetPrediction platform), a machine learning or deep learning-based prediction model, molecular docking simulation, or other prediction models. The principle of similarity-based prediction models is to calculate the similarity between the active ingredient and known ligands in chemical space, thereby inferring their potentially shared targets. Machine learning or deep learning-based prediction models predict new interactions by training on known ligand-target interaction data. Molecular docking simulation uses computational chemistry methods to simulate the binding energy between the active ingredient molecule and the three-dimensional structure of the protein target, screening out targets with binding energies below a specific threshold.

[0026] Step 3: Construct the prescription target set P. The human protein targets predicted for all active ingredients in Step 2 are deduplicated and merged to form a complete set, called the prescription target set P.

[0027] Step 4: Construct the disease target set D. Retrieve human protein targets closely related to the disease from disease-related databases (such as DisGeNET, OMIM, GeneCards, TTD), remove duplicates from all retrieved disease-related targets, and merge them to form the disease target set D.

[0028] In this embodiment, optimizing the prescription based on efficacy indicators includes: calculating the network efficacy Z-score for each of the multiple candidate prescriptions generated by the machine learning model or large language model; selecting the prescription with the highest Z-score as the optimal candidate; analyzing the prescription target that contributes the most to the network distance; and recalculating the Z-score of the new prescription by replacing or deleting some traditional Chinese medicines, thereby achieving computation-driven optimization of the prescription. Specifically, this includes:

[0029] S7.1. Multi-prescription scoring and initial screening. Calculate the network efficacy Z-score for each of the multiple candidate prescriptions generated by the machine learning model or large language model. Select the one or more prescriptions with the highest Z-scores to proceed to the subsequent optimization loop.

[0030] S7.2. Analysis of the Contribution of Key Network Targets and Traditional Chinese Medicine (TCM). For the selected high Z-score prescriptions, a contribution deconstruction analysis is performed, specifically including: ① Identifying key prescription targets: In the connected sub-network, identify several prescription targets that contribute the most to connecting the disease target set D, or to maintaining key network topological properties (such as betweenness centrality), and mark them as key prescription targets. ② Calculating the contribution of TCM: For each TCM in the prescription, calculate its contribution score. This score is calculated as the sum of the number or weights of all predicted targets included in the TCM that belong to the key prescription targets. The lower the contribution score, the less important the role of the TCM in the overall network efficacy.

[0031] S7.3. Computation-driven prescription optimization, based on the analysis results of S7.2, performs one or more of the following optimization operations: ① Removal operation: Remove traditional Chinese medicines (TCMs) with a contribution score below a preset threshold, or those with the lowest contribution score, from the prescription. ② Replacement operation: Replace TCMs with low contribution scores in the prescription with another TCM that has a similar function in TCM theory (such as properties and meridian tropism), but is predicted to contribute more key prescription targets or introduce new targets with higher proximity to the disease target set D network. ③ Dosage adjustment: Adjust the dosage of each TCM according to its contribution, and appropriately increase the weight of TCMs with high contributions (within the pharmacopoeia safe dosage range).

[0032] S7.4. Iterative Optimization: The new prescription generated after S7.3 is recalculated using steps S2 to S6 to obtain a new network efficacy Z-score. The changes in Z-score before and after optimization are compared. If the new prescription's Z-score significantly improves or meets the preset optimization target, the optimization is accepted; otherwise, it is rolled back or other optimization paths are tried. This process is repeated until the prescription's Z-score no longer significantly improves or reaches the preset number of iterations, thus outputting the final optimized prescription.

[0033] The core of steps S7.2 and S7.3 above lies in transforming the empirical question of "which herb to delete" into a quantitative decision-making problem based on network data. The judgment logic is as follows:

[0034] Step 1: Identify "redundant" or "minor" Chinese medicinal herbs. Rank all Chinese medicinal herbs using the aforementioned quantitative indicators. Those herbs that consistently rank low in contribution score and whose removal would have a negligible impact on network topology are marked as "candidates for deletion."

[0035] Step Two: Safety / Cost Verification. Before making a final decision on deletion, verification through an auxiliary knowledge base is required. ① Adjuvant / Guide Drug Verification: Is the drug the only "adjuvant / guide drug" in the prescription (e.g., the only licorice used to harmonize the other drugs)? If so, it may be retained even if its network contribution is low to comply with the principles of traditional Chinese medicine. ② Toxicity / Side Effect Check: If the drug is known to have toxicity or side effects and its contribution is low, it will be prioritized for deletion. ③ Cost Considerations: If the drug is expensive and its contribution is low, it will be prioritized for deletion to reduce costs.

[0036] Step 3: Execute deletion and verification. The system will execute the deletion operation based on a comprehensive rule (e.g., contribution score < threshold && network efficiency impact < threshold && non-core adjuvant drugs), and immediately verify whether the deletion truly improved the network efficacy of the prescription by recalculating the Z-score. The threshold is determined by domain knowledge of traditional Chinese medicine experts or traditional Chinese medicine pharmaceuticals.

[0037] In this embodiment, a method for evaluating the efficacy of traditional Chinese medicine prescriptions based on network pharmacology is provided. This method includes the following steps:

[0038] S1 constructs a set of targets for traditional Chinese medicine prescriptions and a set of targets for diseases.

[0039] S1.1 Input Traditional Chinese Medicine Prescription: Input a prescription consisting of multiple traditional Chinese medicines.

[0040] S1.2 Obtaining Chinese Medicine Components: Obtain all known active ingredients (small molecules) of each Chinese medicine in the prescription through Chinese medicine chemical component databases (such as TCMSP and TCMID).

[0041] S1.3 Mapping Protein Targets: Using a drug target prediction database (such as SwissTargetPrediction, BATMAN-TCM), predict and obtain the human protein targets corresponding to the above-mentioned active ingredients. The collection of all these target proteins constitutes the "prescription target set (P)".

[0042] S1.4 Define the disease target set: The set of human protein targets that are strongly correlated with specific diseases (such as psychosis, lung cancer) is obtained from disease-related databases (such as DisGeNET, OMIM) and constitutes the "disease target set (D)".

[0043] S2 constructs and analyzes protein-protein interaction (PPI) networks.

[0044] S2.1 Constructing the background PPI network: Download the complete human PPI network from authoritative databases (such as STRING, BioGRID) as the background network.

[0045] S2.2 Extracting Subnetwork: Using the union of the "Prescription Target Set (P)" and the "Disease Target Set (D)" as seed nodes, a connected subnetwork containing these nodes and their directly interacting neighbors is extracted from the background PPI network for subsequent analysis.

[0046] S3 Calculation of Network Efficacy Indicators is the core innovative step of this invention, which quantifies the efficacy of prescriptions by calculating network distance indicators at three levels.

[0047] S3.1 Calculate the statistical distance.

[0048] Calculate the distributional differences between the prescription target set (P) and the disease target set (D) in terms of network topological characteristics. Specifically, this may include: ① Degree centrality distance: Calculate the Jensen-Shannon divergence or Wasserstein distance between the degree centrality distributions of nodes in P and D. ② Betweenness centrality distance: Calculate the difference in betweenness centrality distributions of nodes in P and D. An effective prescription should have a similar "status" (e.g., hub status) with the disease target set D in the network, meaning their statistical distance should be small.

[0049] S3.2 Calculate the shortest path distance between network nodes.

[0050] For each target p in the prescription target set P, calculate the shortest path length from p to all targets in the disease target set D, and take the minimum value as the distance from p to D. Ultimately, the network distance from prescription P to disease D is defined as the average or median of all distances from p to D. This value directly reflects the physical proximity of the prescription's site of action to the core disease module in the network space. The shorter the distance, the more directly the prescription components may intervene in the disease process.

[0051] S3.3 Calculate the Z-score of network efficacy

[0052] (1) Constructing a random control: Randomly select a set of proteins from the background PPI network that are the same size as the prescription target set P, as the "random prescription target set (P_random)". Repeat this process N times (e.g., 1000 times).

[0053] (2) Calculate the zero distribution: For each P_random, calculate the "shortest distance of network nodes" to the disease target set D (as described in S3.2) to obtain N distance values, forming a zero distribution.

[0054] (3) Calculate the Z-score: Calculate the efficacy Z-score of prescription P using the following formula:

[0055]

[0056] Where Mean_random and SD_random are the mean and standard deviation of the zero distribution, respectively, and Distance_P is the true network distance of prescription P.

[0057] Z-scores eliminate the influence of network context structure. A Z-score significantly greater than 0 (e.g., > 1.65 or 2.0) indicates that the distance from prescription P to disease D is significantly smaller than the random expectation, meaning that the prescription has statistically significant potential efficacy.

[0058] S4 Evaluation and Experimental Design of Network-Based Therapeutic Effects.

[0059] S4.1 Evaluation of AI-Generated Prescriptions: Multiple candidate prescriptions generated by machine learning models or large language models are processed sequentially through S1 to S3, and the network efficacy Z-score of each prescription is calculated. The prescription with the highest Z-score is selected as the optimal candidate for subsequent experimental verification.

[0060] S4.2 Guided Prescription Optimization: Analyze the prescription components (traditional Chinese medicine) or targets that contribute most to network distance. By replacing or removing certain traditional Chinese medicines, the Z-score of the new prescription is recalculated, achieving computation-driven optimization of the prescription.

[0061] S4.3 Design a wet test verification scheme:

[0062] (1) High Z-score prescription: Prioritize in vitro cell experiments (such as detecting the viability, apoptosis, and inflammatory factor levels of disease-related cell models) and in vivo animal experiments (such as behavioral and histopathological improvements in disease model mice).

[0063] (2) Target validation: Based on the network analysis results, the core targets with the shortest distance from the disease target network are selected first, and molecular biological validation (such as Western Blot, ELISA, immunofluorescence) is performed to confirm the changes in their expression or activity.

[0064] This invention also provides a large-scale model capability assessment technique based on network efficacy quantification, comprising: treating traditional Chinese medicine prescriptions generated by different large models as "candidate drugs," and quantifying and scoring them using a unified and computable network pharmacology index (the core being the network efficacy Z-score). Finally, the capabilities of the large models themselves are ranked according to the scores. M: The set of large models participating in the assessment, M = {m1, m2, ..., mk}. D: The set of baseline diseases, D = {d1, d2, ..., dl}. P{ij}: The set of all prescriptions generated by model mi for disease dj. For simplicity, assuming each model generates n prescriptions for each disease, then P{ij} = {p{ij1}, p{ij2}, ..., p{ijn}}. Z(p): The network efficacy Z-score of prescription p. S(p, C): The average Jaccard similarity between prescription p and the classic prescription library C.

[0065] Step 1: Calculate the efficacy score. This score is based on the shortest distance between network nodes, but ultimately reflects the network efficacy Z-score. Because the Z-score already includes shortest distance information and has undergone statistical significance testing, it is a superior surrogate indicator for efficacy. The total efficacy score of model m_i is the average of the Z-scores of all the prescriptions it generates:

[0066]

[0067] Here, Ei is the efficacy score of model mi; the higher the value, the higher the average potential efficacy of the generated prescription.

[0068] Step 2: Calculate the stability score. This score measures the consistency of the efficacy of the prescriptions generated by the model.

[0069] a. Calculate the variance of the Z-scores of all prescriptions generated by model mi for all diseases.

[0070]

[0071] b. Take the reciprocal of the variance; the smaller the variance (the more stable), the higher the score.

[0072]

[0073] Where Si is the stability score of model mi. The larger the value, the more reliable the model output.

[0074] Step 3: Innovation Score. This score measures the model's ability to depart from classic prescriptions and create novel combinations.

[0075] A. For each prescription, calculate its Jaccard similarity to all prescriptions in the classic prescription library and take the average.

[0076] B. Calculate the average similarity of all generated prescriptions in model mi.

[0077]

[0078] C. Convert the average similarity score into an innovation score. The lower the similarity, the higher the innovation.

[0079]

[0080] Where Ii is the innovation score of model mi. The larger the value, the greater the difference between its prescription and the classic prescription, and the stronger its innovation.

[0081] Step 4: Indicator Standardization and Overall Ranking. Since the efficacy score, stability score, and innovation score have different dimensions and orders of magnitude, standardization is necessary. ① Standardization Process. Min-Max normalization is used to map each indicator to the [0, 1] interval. For any original score sequence {x1, x2, ..., xk} in the model set M, its normalized score is:

[0082]

[0083] Where max(x) and min(x) are the maximum and minimum values ​​in the score sequence, respectively. The normalized scores for efficacy, stability, and innovation are denoted as E'i, S'i, and I'i, respectively.

[0084] Overall score calculation. Weights are assigned to the three dimensions, and the weighted overall score for each model is calculated. Let the weight vector W = [w_E, w_S, w_I], and w_E + w_S + w_I = 1. Based on the setting W = [0.5, 0.3, 0.2], the overall score F_i of model m_i is:

[0085]

[0086] Final ranking. Models are sorted in descending order based on their overall score F_i.

[0087]

[0088] The model with the highest overall score F_i is considered to have the strongest prescription generation capability.

[0089] In this embodiment, a method for evaluating AI-generated traditional Chinese medicine prescriptions for treating sleep disorders (symptoms including palpitations) in psychotic disorders is evaluated.

[0090] Input: The large model generates the corresponding prescription "Astragalus membranaceus, Rehmannia glutinosa, Coptis chinensis, and Salvia miltiorrhiza" based on the actual patient information.

[0091] S1 execution: The active ingredients of four traditional Chinese medicines were queried from the TCMSP and HERB databases, yielding 50 ingredients such as "astragaloside A," "berberine," and "tanshinone." SwissTargetPrediction was used to predict the targets of these ingredients, resulting in a "prescription target set P," containing 120 unique proteins. Genes related to "Type 2 Diabetes" were retrieved from the DisGeNET database, resulting in a "disease target set D," containing 150 unique targets.

[0092] S2 execution: Download the human PPI network from the STRING database, use the P∪D protein as the seed, and extract its first-level neighbors to form a subnetwork containing 5000 nodes and 65000 edges.

[0093] S3 execution:

[0094] S3.1: Calculate the degree centrality distributions of P and D. The Jensen-Shannon divergence is 0.15, indicating that the two are similar in pivotality.

[0095] S3.2: Calculate the shortest distance from P to D, with an average value of 2.8.

[0096] S3.3: Randomly select 120 proteins from the human PPI network as P_random, repeat 1000 times, calculate their average distance to D, and obtain a null distribution (mean = 4.1, standard deviation = 0.3). Then the Z score = (4.1 - 2.8) / 0.3 = 4.33.

[0097] S4 Evaluation and Decision: Z-score = 4.33 > 2.0, indicating that this AI prescription has a highly significant potential anti-diabetic therapeutic effect at the computational level. Compared with two other AI prescriptions with Z-scores of 1.1 and 0.5, this prescription was selected for subsequent experiments. Network analysis indicated that AKT1, INS, and IL6 are core targets connecting the prescription and the disease. Experimental Design: In a mouse sleep model, the study verified whether the extract of this prescription could improve the sleep quality of mice, focusing on detecting the expression and phosphorylation levels of proteins such as AKT1, INS, and IL6.

[0098] S5. Large-scale model capability assessment based on network efficacy quantification: Traditional Chinese medicine prescriptions generated by different large-scale models are regarded as "candidate drugs," and they are quantitatively scored using a unified and computable network pharmacology index (the core being the network efficacy Z-score). Finally, the capabilities of the large-scale models are ranked according to their scores.

[0099] A system for evaluating the efficacy of traditional Chinese medicine prescriptions based on network pharmacology. The system includes: a data acquisition module, a network construction and computation module, a prescription evaluation and optimization module, and a user interface.

[0100] The data acquisition module retrieves component, target, and PPI data from the database;

[0101] The network construction and calculation module is used to construct the PPI network and perform network distance and Z-score calculations.

[0102] The prescription evaluation and optimization module is used to compare the efficacy indicators of different prescriptions and provide optimization suggestions;

[0103] The user interface is used to input prescriptions and disease information, and to visualize the output results.

[0104] The implementation methods of the system and the method of the present invention are the same.

[0105] Figure 1 The overall flowchart of the method of this invention is as follows: First, the traditional Chinese medicine prescription and the target disease are input, entering the data acquisition and preprocessing stage. This stage acquires data from multiple databases, including traditional Chinese medicine component databases (such as TCMSP, TCMID), target prediction databases (such as SwissTargetPrediction), disease target databases (such as DisGeNET, OMIM), and PPI network databases (such as STRING, BioGRID). Next, core network analysis is performed to construct a "prescription-disease" PPI subnetwork, calculating network efficacy indicators, including statistical distance, shortest distance between network nodes, and network efficacy Z-score. Finally, the evaluation and decision-making stage is entered, evaluating the AI-generated prescription, guiding prescription optimization, and designing a wet experimental validation plan.

[0106] Figure 2 Network distribution and shortest path conceptual diagram. The diagram contains multiple protein nodes, such as proteins A, B, C, D, E, F, and G, which are connected by lines to represent interactions. Proteins A and B belong to the prescription target set P, while proteins F and G belong to the disease target set D. The diagram illustrates the shortest path distance from the prescription target set to the disease target set, for example, from A2 to B1 and then to C1, demonstrating the connection and action pathway between drug targets and disease targets, used to analyze the mechanism of drug action on diseases.

[0107] Figure 3The diagram illustrating Z-score calculation demonstrates the randomization process and compares the true network distance of prescriptions with the zero distribution, providing a visual representation of the Z-score calculation. First, we have the true prescription distance (Distance_P) and the random distance zero distribution (including the mean (Mean_random) and standard deviation (SD_random)). Both are input into the Z-score calculation, which uses the formula: Z-score = Mean_random - Distance_P / SD_random. Simultaneously, the diagram showing the zero distribution and true values ​​displays the random zero distribution curve, which includes the random distance mean (Mean_random) and the true prescription distance (Distance_P). By comparing these, the position of the true prescription distance within the random zero distribution can be visually presented, thus allowing for the assessment of its significance.

[0108] Figure 4 The overall interaction and data flow between the system and external entities. This includes a Level 0 context graph and a Level 1 decomposition graph. The Level 0 context graph illustrates the interaction between the system and external entities. These external entities include traditional Chinese medicine / compound databases (such as TCMSP, TCMID), target prediction platforms (such as SwissTargetPrediction), disease and PPI databases (such as DisGeNET, STRING), and researchers. These external entities provide the system with data such as traditional Chinese medicine component data, component-target mapping data, disease target data, and PPI network data. Researchers query the system for prescription and disease information, and the system returns network efficacy reports, including Z-scores, core targets, and optimization suggestions.

[0109] Figure 5 Data processing breakdown diagram. This diagram illustrates the main data processing steps and data flow of the system. External entities are researchers / users who submit prescription and disease query requests to the system. The system consists of three main processing steps: data acquisition and integration (P1), network construction and indicator calculation (P2), and result evaluation and report generation (P3). Data storage areas include a local TCM knowledge base (DS1), a local PPI network database (DS2), a temporary network data area (DS3), and a historical evaluation result database (DS4). P1 reads data from DS1 and DS2, integrates it, and then passes the standardized target set to P2. P2 constructs the network, calculates indicators (such as Distance_P and Z-scores), and stores the temporary results in DS3. P3 generates a comprehensive efficacy evaluation report, returns it to the user, stores the report in DS4, and simultaneously reads historical data for comparative analysis.

[0110] Figure 6The prescription evaluation task is a dynamic interactive process of task submission and data preparation. The system interacts with users (researchers) and internal modules to complete the task. Users submit traditional Chinese medicine prescriptions and target disease information through the user interface (UI). The UI forwards the request to the data management module (DataMgr). DataMgr sequentially queries the traditional Chinese medicine component database, predicts component targets, obtains disease-related targets, and constructs a PPI sub-network. Upon completion, DataMgr notifies the UI that the data is ready, and the UI displays "Data preparation complete" to the user, marking the end of the task submission and data preparation phase and preparing for subsequent network computation and evaluation optimization.

[0111] Figure 7 The prescription evaluation task involves a dynamic interaction between core computation and outcome evaluation. In Phase Two, the user initiates network efficacy calculation, and the user interface (UI) invokes the network computation engine (NetCalc). NetCalc sequentially calculates statistical distances and shortest distances between network nodes, performs randomized Z-score calculations, and finally returns all network indicator results to the UI. In Phase Three, the UI requests the evaluation and optimization module (EvalMgr) to interpret the results and provide suggestions. EvalMgr generates an efficacy evaluation report, identifies core targets, generates experimental design suggestions, and returns a comprehensive report to the UI. The UI visualizes the results for the user, including the network graph, Z-score bar chart, list of core targets, and experimental validation suggestions.

[0112] Figure 8 This study evaluates the efficacy of two prescription generation methods (distance-based and Z-score-based) using PPI networks. The distance-based method assesses the drug's efficacy by calculating the shortest path distance between the drug target and the disease target. A shorter distance indicates a stronger association between the drug and disease targets, potentially leading to better efficacy. The Z-score-based method calculates the Z-score by comparing the actual distance between the drug and disease targets with the distance from a random background distribution. A higher Z-score indicates a more significant therapeutic effect and a more specific association between the drug and disease targets.

[0113] Figure 9A framework for evaluating the capabilities of large-scale models based on network efficacy quantification is presented. The flowchart outlines the process of generating TCM prescriptions using large-scale models and conducting unified efficacy evaluation. First, a unified evaluation task is defined. Then, prescriptions are generated in parallel, including general large-scale models (such as GPT-4, Wenxin Yiyan, and Tongyi Qianwen) and domain-specific large-scale models (such as professional fine-tuning models and Wuji TCM large-scale models). The generated prescriptions enter the unified network pharmacology evaluation stage, where the prescription and disease are input, targets are acquired and standardized, and the network efficacy Z-score is calculated. Next, data is collected and statistically analyzed, and scores and rankings are performed from multiple dimensions, including efficacy score (mean Z-score), stability score (Z-score variance), and innovativeness score (similarity to classic prescriptions). Finally, the models are ranked according to their comprehensive capabilities, and the final ranking of large-scale models is output, providing a basis for selecting and optimizing large-scale models.

[0114] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large-scale model generated from network pharmacology, characterized in that, include: A machine learning model was used to generate candidate prescriptions and obtain all known active ingredients of each Chinese herb in the prescription. Based on the known active ingredients, predict the corresponding human protein targets for the patient, and collect all human protein targets to obtain the prescription target set P; obtain the human protein targets related to the corresponding diseases, collect all human protein targets corresponding to the diseases to obtain the disease target set D. Constructing protein-protein interaction networks; The prescription target set P and the disease target set D are input into a protein-protein interaction network to construct a connected subnetwork; Calculate the efficacy index of the connected subnetwork; optimize the prescription based on the efficacy index, and complete the efficacy evaluation of the traditional Chinese medicine prescription.

2. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 1, characterized in that, Based on the prediction of the corresponding human protein targets of patients using known active ingredients, including: Step 1: For each Chinese herb in the candidate prescription, obtain all known active ingredient molecules from the Chinese herbal medicine chemical composition database; Step 2: Predict the interactions between all active ingredients and the corresponding human protein targets of the patient; Step 3: Remove duplicates from all the predicted human protein targets of the active ingredients and merge them to form a complete set, which is used as the prescription target set P. Step 4: Retrieve human protein targets closely related to the disease from the disease-related database, remove duplicates from all retrieved disease-related targets and merge them to form a disease target set D.

3. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 1, characterized in that, Constructing a protein-protein interaction network involves: obtaining a complete human PPI network from a database and using this network as a background network; using data from the union of the prescription target set P and the disease target set D as seed nodes, and inputting the seed nodes into the background network to obtain a connected subnet containing the seed nodes and their corresponding neighbor nodes.

4. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 1, characterized in that, The efficacy indicators for connected subnetworks include statistical distance, shortest distance between network nodes, and network efficacy Z-score.

5. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 4, characterized in that, The calculation of statistical distance includes: calculating the Jensen-Shannon divergence of the degree centrality distributions of nodes in the prescription target set P and the disease target set D; and calculating the difference in the betweenness centrality distributions of nodes in P and D based on the Jensen-Shannon divergence.

6. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 4, characterized in that, Calculating the shortest distance between network nodes includes: for each target p in the prescription target set P, calculating the shortest path length from it to all targets in the disease target set D, and taking the minimum value as the distance from p to D; the network distance from prescription P to disease D is defined as the average of all distances from p to D; and the minimum value is selected from all distances as the shortest distance between network nodes.

7. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 4, characterized in that, The calculation of the network efficacy Z-score includes: Step 1: Randomly select a set of proteins from the background PPI network that are the same size as the prescription target set P, and use this as the random prescription target set P_random; repeat this process N times. Step 2: For each P_random, calculate its shortest distance to the network nodes of the disease target set D, obtain N distance values, and form a null distribution; Step 3: Calculate the efficacy Z-score of prescription P based on the zero distribution.

8. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 7, characterized in that, The formula for calculating the Z-score of the efficacy of prescription P is: Among them, Mean random and SD random These are the mean and standard deviation of the null distribution, respectively. P It is the true network distance of prescription P.

9. The method for evaluating the efficacy of traditional Chinese medicine prescriptions based on a large model generated from network pharmacology according to claim 1, characterized in that, Optimizing prescriptions based on efficacy indicators includes: calculating the network efficacy Z-score of each candidate prescription generated by a machine learning model or a large language model; selecting the prescription with the highest Z-score as the optimal candidate; analyzing the prescription target that contributes the most to the network distance; replacing or deleting Chinese medicines in the prescription based on the prescription target that contributes the most, and recalculating the Z-score of the new prescription to achieve computation-driven optimization of the prescription.

10. A system for evaluating the efficacy of traditional Chinese medicine prescriptions based on network pharmacology, the system being used to execute the method for evaluating the efficacy of traditional Chinese medicine prescriptions generated by a large model based on network pharmacology as described in any one of claims 1 to 9, characterized in that, The system includes: a data acquisition module, a network construction and calculation module, a prescription evaluation and optimization module, and a user interface; The data acquisition module retrieves component, target, and PPI data from the database; The network construction and calculation module is used to construct the PPI network and perform network distance and Z-score calculations. The prescription evaluation and optimization module is used to compare the efficacy indicators of different prescriptions and provide optimization suggestions; The user interface is used to input prescriptions and disease information, and to visualize the output results.

Images