A multi-dimensional traditional chinese medicine model system with large model and intelligent agent

Abstract

This invention relates to the field of traditional Chinese medicine (TCM) model technology, and discloses a multi-dimensional TCM large-scale model collaborative intelligent agent system, comprising: a data segmentation unit that performs semantic segmentation on TCM data and omics data, and uniformly encodes TCM nodes and their relationships; a first collaborative unit that, based on relationship identifiers, inputs prescription nodes and their associated component nodes into the large model, and generates paths for target nodes and pathway nodes based on the relationship identifiers; a second collaborative unit that inputs the component-target correspondence in the candidate action path dataset as a constraint into the small model, and solves in parallel the affinity results and molecular property results between components and targets; and a collaborative update unit that determines the decision output based on the screened candidate action path dataset, and updates the relationship identifiers, the large model, and the small model. This invention ensures the reliability of multi-dimensional TCM large-scale model collaboration.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine model technology, and more specifically, to a multi-dimensional collaborative intelligent system for large and small traditional Chinese medicine models. Background Technology

[0002] The discovery of new traditional Chinese medicine (TCM) drugs typically involves multiple stages, including formula composition analysis, component screening, mechanism of action deduction, and drug-likeness assessment. Its research objects include both unstructured textual data such as ancient TCM texts and clinical case records, and structured data such as modern pharmacological experimental data and multi-omics data. Currently, existing technologies model the data by introducing machine learning or deep learning models; for example, using language models for semantic parsing of TCM texts or employing small-scale computational models to calculate component properties. However, this approach can only be trained and deployed for a single task, lacks collaborative mechanisms between models, and the inference results generated by large-scale models are difficult to use as computable constraints. This leads to the data being utilized by smaller, branch-based models, making it difficult for the model's processing results to influence the upstream inference structure. This results in a fragmented overall processing flow. Furthermore, the data obtained during the experimental verification phase is not effectively fed back into the preceding data structure or model parameters, causing subsequent analysis to remain based on static knowledge or fixed models. This makes it difficult to form a continuous iterative processing process. In the parallel research scenario of discovering new traditional Chinese medicine drugs, it is difficult to establish a constrained data transfer relationship between inference models and computational models of multi-source traditional Chinese medicine data. There is a lack of collaborative processing mechanisms, and experimental verification data is difficult to integrate into the overall model structure, thus limiting the reliability of multi-task collaborative analysis in the discovery of new traditional Chinese medicine drugs.

[0003] Therefore, it is necessary to design a multi-dimensional collaborative intelligent agent system based on the size model of traditional Chinese medicine to solve the problems existing in the current technology. Summary of the Invention

[0004] In view of this, the present invention proposes a multi-dimensional TCM large and small model collaborative intelligent agent system, which aims to solve the problems of difficulty in establishing a constrained data transfer relationship between reasoning models and computational models of multi-source TCM data, lack of collaborative processing mechanism, and difficulty in integrating experimental verification data into the overall model structure, thus limiting the reliability of multi-task collaborative analysis in the discovery of new TCM drugs.

[0005] This invention proposes a multi-dimensional collaborative intelligent agent system based on large and small models of traditional Chinese medicine, comprising:

[0006] The data segmentation unit is configured to perform semantic segmentation on TCM data and omics data, extract TCM nodes based on drug name, ingredient name, target of action and action type, determine the association relationship based on the co-occurrence order and action description of the TCM nodes, uniformly encode the TCM nodes and association relationship, and determine the relationship identifier and knowledge structure. The knowledge structure includes prescription nodes, ingredient nodes, target nodes and pathway nodes.

[0007] The first collaborative unit is configured to input the prescription node and its associated component node into the large model based on the relationship identifier, and generate paths for the target node and pathway node based on the relationship identifier to determine the candidate action path dataset;

[0008] The second collaborative unit is configured to input the component-target correspondence in the candidate action path dataset as a constraint into the small model, and to solve the affinity results and molecular property results between the components and the target in parallel to determine the path evaluation results.

[0009] The collaborative update unit is configured to filter the candidate action path dataset based on the path evaluation results, determine the decision output based on the filtered candidate action path dataset, add the experimental verification data to the knowledge structure, and update the relation identifier, large model, and small model.

[0010] Furthermore, when performing semantic segmentation on TCM data and omics data, and extracting TCM nodes based on drug name, ingredient name, target of action, and type of action, the following steps are taken:

[0011] The TCM data includes TCM text data, clinical medical record text data, and pharmacological data;

[0012] The data segmentation unit segments the TCM text data and clinical medical record text data according to punctuation marks, drug names, and prescription structures to determine the text fragment set, and splits the pharmacological data and omics data according to field keys, sample identifiers, and indicator names to determine the record fragment set;

[0013] In the set of text fragments and the set of record fragments, based on entity localization using dictionary matching and synonym merging, candidate drug names, candidate ingredient names, candidate target of action, and candidate action type are determined, and ambiguity correction is performed on each candidate based on the context position and the combination of adjacent entities.

[0014] Each candidate for ambiguity correction is written into the Traditional Chinese Medicine (TCM) node table to determine the TCM node dataset. The TCM node table includes node category, source identifier, and location index.

[0015] Furthermore, when determining the association relationship based on the co-occurrence order and function description of the aforementioned traditional Chinese medicine nodes, the following is included:

[0016] The data segmentation unit determines node pairs based on the order of occurrence of the TCM nodes, encodes the cross-sentence distance and adjacent words of the node pairs to determine the sequence features, extracts the descriptive phrases of the effects between the node pairs and maps them to effect type labels to determine the descriptive features, and determines the relationship record based on the sequence features and descriptive features. The relationship record includes the starting node identifier, the ending node identifier, the effect type label and the source fragment identifier.

[0017] Relationship records with the same start node identifier, end node identifier, and function type label are merged, while the source fragment identifier is retained. The association relationship dataset is determined based on the merging result.

[0018] Furthermore, when uniformly encoding the aforementioned TCM nodes and relationships to determine relationship identifiers and knowledge structures, the following steps are included:

[0019] The data segmentation unit generates node codes for the TCM node dataset based on node category and standard name. The node codes include node category codes and name verification codes. The unit generates relationship identifiers for the association dataset based on the starting node identifier, ending node identifier, and function type label. The relationship identifiers include edge type codes and structure verification codes.

[0020] The adjacency index is determined based on the node codes and relation identifiers. The adjacency index includes a set of outgoing relation identifiers and a set of incoming relation identifiers corresponding to each node code. Based on the node category, the node codes are divided into prescription node codes, ingredient node codes, target node codes, and pathway node codes. The hierarchical connection direction of the code division is limited based on the adjacency index to determine the knowledge structure.

[0021] Furthermore, when inputting the prescription node and its associated component nodes into the large model based on the relationship identifier, the following steps are included:

[0022] The first collaborative unit takes the target prescription node as the starting point in the knowledge structure, extracts the first-order adjacent component nodes and corresponding relation identifiers, determines the prescription subgraph, and performs serialization encoding on the prescription subgraph to determine a triplet sequence containing prescription node encoding, component node encoding and relation identifier;

[0023] Using the edge type code of the relation identifier as a constraint, the triple sequence is input into the large model. The large model limits the extended node categories and edge types based on the constraint and outputs the relation identifier index.

[0024] Furthermore, when generating paths for the target nodes and pathway nodes based on the relationship identifiers to determine the candidate action path dataset, the process includes:

[0025] The first collaborative unit expands the adjacency index of the knowledge structure based on the large model, and the expansion order is as follows: prescription node, ingredient node, target node and pathway node.

[0026] During each expansion, adjacent items that satisfy the constraints are selected, several candidate paths are determined, a path string is determined based on the path code of each candidate path, and the corresponding relation identifier index is embedded in the path string. After embedding, deduplication is performed, and the candidate action path dataset is determined based on the deduplication result.

[0027] Furthermore, when inputting the component-target correspondence in the candidate action path dataset as a constraint into the small model, the following steps are included:

[0028] The second collaborative unit extracts path strings from the candidate action path dataset and extracts the component node code and target node code of each path string to determine the component-target pairing list.

[0029] The component nodes of each extracted path string are encoded and mapped to component vectors, and the target nodes of each extracted path string are encoded and mapped to site vectors. The component-target pairing list and the corresponding component vectors and site vectors are input into the small model. The small model uses the component-target pairing list as an index constraint to determine the source path string identifier.

[0030] Furthermore, when determining the path evaluation results by parallelly solving the affinity results and molecular property results between the components and the target, the following steps are included:

[0031] The small model performs parallel affinity calculations and molecular property calculations on the input data. The affinity calculation uses the component vector and the site vector as joint inputs and outputs pairwise affinity results. The molecular property calculation uses the component vector as inputs and outputs component-level property results.

[0032] The second collaborative unit backfills the pairing-level affinity results to the corresponding path strings according to the source path string identifier, and backfills the component-level property results to all path strings containing the component according to the component node code. It summarizes the pairing-level affinity results and component-level property results for each path string, determines the path evaluation vector, and determines the path evaluation vector and the summarized corresponding path string as the path evaluation result.

[0033] Furthermore, when filtering the candidate action path dataset based on the path evaluation results, the process includes:

[0034] The collaborative update unit groups the candidate action path dataset according to the same prescription node encoding based on the path evaluation vector of all path evaluation results, and removes path strings with inconsistent structure check codes or conflicting source fragment identifiers to determine the target action path dataset.

[0035] Furthermore, when adding experimentally verified data to the knowledge structure and updating the relation identifiers, large model, and small model, the process includes:

[0036] The collaborative update unit parses the experimental verification data to determine the experimental object identifier, measurement index identifier, experimental source identifier, and experimental result identifier. It maps the experimental object identifier to the source path string identifier in the target action path dataset. Based on the mapping result, it writes the experimental result identifier into the knowledge structure and the experimental source identifier into the association dataset. Based on the writing results of the experimental result identifier and the experimental source identifier, it determines the incremental sample set and inputs the incremental sample set into the large model and the small model for parameter updates.

[0037] Compared with existing technologies, the beneficial effects of this invention are as follows: Through the collaborative linkage of various units, a complete closed loop of data processing, path generation, quantitative evaluation, and iterative updates is established, solving the problems of lack of collaboration between models, inability to provide feedback on experimental data, and fragmented processes. This achieves the integration of various stages in the discovery of new traditional Chinese medicine (TCM) drugs, transforming multi-source heterogeneous data into a standardized knowledge structure. Precise collaboration between large and small models is achieved through relational identification, making the inference results of large-scale models the computational constraints of small-scale models. Simultaneously, experimental data provides feedback on the knowledge structure and model parameters, improving the utilization rate of data and models. Furthermore, candidate path generation is based on the logical constraints of the knowledge structure, path evaluation is supported by quantitative calculations, and iterative updates are verified by experimental data, ensuring the feasibility and drugability of new drug discovery paths and providing a reliable research direction for the discovery of new TCM drugs. Through dynamic knowledge structures and collaborative update mechanisms, multi-source TCM data can achieve constrained data transfer, and inference and computational models can collaborate efficiently, thus adapting to the parallel research needs of TCM drug discovery. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a functional block diagram of a multi-dimensional TCM large and small model collaborative intelligent agent system provided in an embodiment of the present invention. Detailed Implementation

[0040] 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.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] See Figure 1 As shown in some embodiments of this application, a multi-dimensional TCM large-scale model collaborative intelligent agent system includes:

[0043] The data segmentation unit is configured to perform semantic segmentation on TCM data and omics data, extract TCM nodes based on drug name, ingredient name, target of action, and type of action, determine the association based on the co-occurrence order and action description of TCM nodes, uniformly encode TCM nodes and associations, and determine relationship identifiers and knowledge structures. The knowledge structure includes prescription nodes, ingredient nodes, target nodes, and pathway nodes.

[0044] The first collaborative unit is configured to input the prescription node and its associated component node into the large model based on the relation identifier, and generate paths for the target node and pathway node based on the relation identifier to determine the candidate action path dataset.

[0045] The second collaborative unit is configured to input the component-target correspondence in the candidate action path dataset as a constraint into the small model, and to solve the affinity results and molecular property results between components and targets in parallel to determine the path evaluation results.

[0046] The collaborative update unit is configured to filter candidate action path datasets based on path evaluation results, determine decision outputs based on the filtered candidate action path datasets, add experimental verification data to the knowledge structure, and update relation identifiers, large models, and small models.

[0047] Specifically, TCM data encompasses various types of data involved in the discovery of new TCM drugs, including ancient TCM texts, clinical case records, and modern pharmacological experimental data. Omics data refers to multi-dimensional molecular data related to biological organisms, including genomics, transcriptomics, and proteomics data. For example, proteomics data on changes in intracellular protein expression after TCM components act on cells. TCM nodes are elements extracted from the data, categorized into formula nodes, component nodes, target nodes, and pathway nodes. These nodes form the foundation for constructing knowledge structures. Formula nodes represent nodes of specific traditional Chinese medicine formulas, such as the nodes corresponding to classic formulas or experimental formulas like "Ma Huang Tang" and "Liu Wei Di Huang Wan". Component nodes represent nodes of active ingredients or effective substances contained in traditional Chinese medicine, such as "ephedrine" and "pseudoephedrine" in Ma Huang Tang, and "astragaloside A" in Astragalus membranaceus. Target nodes represent nodes of the biological targets (mostly proteins, enzymes, receptors, etc.) of the effects of traditional Chinese medicine components, such as "cyclooxygenase-2 (COX-2)" for anti-inflammatory effects and "insulin receptor" for hypoglycemic effects. Pathway nodes represent nodes of signal transduction and metabolic regulation in the body, such as "NF-κB signaling" related to inflammation and "mitochondrial apoptosis" related to apoptosis. Relationship identifiers are unique identifiers obtained by uniformly encoding the relationships between traditional Chinese medicine nodes, used to clarify the type and logical direction of different node relationships. For example, "R001" represents "component-target binding relationship", and "R002" represents "target-pathway regulation relationship". The knowledge structure is a structured knowledge network composed of prescription nodes, ingredient nodes, target nodes, pathway nodes, and the relationships between nodes. It forms the foundation for intelligent agent data processing and model collaboration. The large model focuses on knowledge structure and relationship identification; it is a model that generates node path reasoning (a reasoning-oriented model). Its core is to uncover potential action paths between prescriptions, ingredients, targets, and pathways. The candidate action path dataset is generated by the large model and contains a collection of multiple potential prescription-ingredient-target-pathway action paths, such as datasets consisting of paths like "Ephedra Decoction → Ephedrine → Adrenaline Receptor → Sympathetic Nerve Excitation Pathway" and "Ephedra Decoction → Pseudoephedrine → Nasal Mucosal Vascular Smooth Muscle Receptor → Vasomotor Pathway." Smaller models focus on constraints and are computationally oriented, calculating component-target affinity and molecular properties. Affinity results, calculated by the small model, indicate the tightness of binding between the component and the target. Higher affinity means the component more easily binds to the target and exerts its effect; examples include the high affinity of ephedrine to the adrenergic receptor and the moderate affinity of astragaloside A to the insulin receptor. Molecular property results, calculated by the small model, reflect the intrinsic molecular characteristics of the herbal component, including lipid solubility, water solubility, stability, and toxicity. These serve as the basis for evaluating the drug-likeness of the component, such as the molecular property results showing moderate lipid solubility, strong stability, and no significant toxicity.Experimental validation data refers to the measured data obtained after verification through in vitro experiments (such as cell experiments and molecular docking experiments) and in vivo experiments (such as animal experiments). For example, in vitro experiments verify the binding rate of ephedrine to adrenaline receptors, and animal experiments verify the anti-inflammatory effect of this pathway.

[0048] Specifically, the data segmentation unit performs semantic segmentation on the collected TCM data (such as ancient TCM books and clinical case records) and omics data, breaking down long texts and discrete data into segments with independent semantic meaning to ensure accurate data parsing. Based on four dimensions—drug name, ingredient name, target of action, and type of action—TCM nodes are extracted from the semantic segments. The drug name dimension extracts formula nodes (e.g., "Guizhi Tang"), the ingredient name dimension extracts ingredient nodes (e.g., "Guizhi Oil"), the target of action dimension extracts target nodes (e.g., "prostaglandin E2 receptor"), and the type of action dimension helps clarify the logical association between nodes (e.g., "inhibition" and "activation"). The co-occurrence order and specific description of the effects of TCM nodes in the data are combined to determine the relationships between nodes. For example, from "Guizhi Oil inhibits prostaglandin E2 receptor and regulates inflammatory pathways," the relationship is determined as "Guizhi Oil (ingredient node) → Prostaglandin E2." The study establishes the association between "receptor (target node)" and "prostaglandin E2 receptor (target node) → inflammatory pathway (pathway node)," clearly defining the association type as "inhibition" and "regulation." All TCM nodes and associations are uniformly coded, with each association assigned a unique identifier. This ultimately constructs a knowledge structure encompassing formula nodes, component nodes, target nodes, pathway nodes, and inter-node relationships, forming a standardized knowledge network. This avoids the problems of inconsistent data formats and difficulty in integration and utilization from multiple sources in TCM new drug discovery. Through semantic segmentation and node extraction, unstructured text and structured data are transformed into a unified knowledge structure, providing a data carrier for subsequent collaborative processing of large and small models. Furthermore, the associations are determined based on the node co-occurrence order and function description, ensuring the logicality and accuracy of the knowledge structure and avoiding model inference biases caused by data clutter. The first collaborative unit, based on relation identifiers, accurately locates the formula nodes and their associated component nodes within the knowledge structure. This allows for the selection of all component nodes corresponding to a given formula. For example, "R004" locates the component nodes associated with "Guizhi Tang" such as cinnamon twig oil and paeoniflorin. The selected formula nodes and associated component nodes are then input into the large model, providing a clear starting point for inference. The large model generates paths for target nodes and pathway nodes within the knowledge structure based on relation identifiers. Combining the association types (such as binding and regulation) corresponding to the relation identifiers, it starts with the formula, deduces the possible target nodes based on the components, and then deduces the pathway nodes they belong to or regulate from the target nodes, forming a complete action path of formula-component-target-pathway. All derived potential paths are summarized to construct a candidate action path dataset, ensuring that the paths cover all possible pharmacological action directions of the formula. The precise linkage between the knowledge structure and the large model is achieved through relation identifiers, giving the path generation of the large model clear data constraints. Simultaneously, the path generation based on the node association logic of the knowledge structure ensures that the candidate paths conform to the pharmacological principles of traditional Chinese medicine, thereby improving the reliability of path generation.

[0049] Understandably, the second collaborative unit extracts the component-target correspondence for each path from the candidate action path dataset and inputs it as a constraint into the small model. For example, from the path "cinnamon twig oil → prostaglandin E2 receptor → inflammation pathway," it extracts the "inhibitory association between cinnamon twig oil (component node) and prostaglandin E2 receptor (target node)" as a constraint to limit the computational scope of the small model and ensure that the calculation fits the candidate path. Based on the constraint, the small model performs two parallel calculations simultaneously: first, it calculates the affinity results between the component and the target to assess the tightness of their binding and determine whether the component can effectively act on the target; second, it calculates the molecular properties of the component to assess its drug-likeness potential (e.g., whether it is easily absorbed, stable, or toxic). The small model combines the results of the two calculations to comprehensively evaluate each candidate path, thereby generating a path evaluation result. This solves the problem of the lack of collaborative mechanisms between traditional models and the difficulty in utilizing the inference results of big data models in small-scale models. By using the component-target correspondence of the candidate path as a constraint, it achieves... This system enables the collaborative operation of large-scale models (reasoning models) and small-scale models (computational models), allowing the inference results of big data models to be transformed into computational constraints for small-scale models. This improves the relevance and accuracy of the computational results. The collaborative update unit filters candidate action path datasets based on path evaluation results, eliminating unqualified candidate paths. Based on the filtered candidate action path datasets, it generates decision outputs to provide suggestions for the discovery of new traditional Chinese medicine drugs. For example, it recommends a certain path as a direction for experimental verification, adds the measured data obtained in the experimental verification stage to the knowledge structure to supplement new node relationships, updates the relationship identifiers in the knowledge structure based on the newly added experimental verification data, and adjusts the parameters of both the large and small models. For the large model, parameters include the upper limit of path length and the threshold of node jump logic, such as the maximum number of jump steps in "prescription-ingredient-target-pathway" to avoid generating excessively long and pharmacologically meaningless paths, or optimize the matching of target and pathway nodes to improve the fit between the path and the pharmacological laws of traditional Chinese medicine. For small models, parameters include the weights and criteria for various molecular properties. For example, if experiments verify that the toxicity of a component is lower than the model's prediction, the weight of "toxicity" can be reduced, or the toxicity threshold can be relaxed to improve the accuracy of molecular property results. After the parameters of the large model are updated, subsequent path generation will prioritize the reference to verified associations, thereby improving path generation efficiency. After the parameters of the small model are updated, the calculation accuracy of affinity and molecular properties will be further improved, thus forming a closed loop of data processing, path generation, quantitative evaluation, experimental verification, and model update. This solves the problems of experimental data not being able to be fed back to previous stages and the static nature of the model in traditional processes. Through the closed-loop update mechanism, dynamic iteration of TCM-related knowledge and models is achieved, improving the reliability of multi-task collaborative analysis. At the same time, the closed-loop process enables the various stages of TCM new drug discovery to be linked, breaking the fragmented process and ensuring the parallel research needs of TCM new drug discovery.

[0050] In some embodiments of this application, when performing semantic segmentation on TCM data and omics data, and extracting TCM nodes based on drug name, ingredient name, target of action, and type of action, the process includes: TCM data includes TCM text data, clinical medical record text data, and pharmacological data; the data segmentation unit segments the TCM text data and clinical medical record text data according to punctuation marks, drug names, and prescription structures to determine a set of text segments; and splits the pharmacological data and omics data according to field keys, sample identifiers, and indicator names to determine a set of record segments; in the set of text segments and the set of record segments, based on entity localization using dictionary matching and synonym merging, candidate drug names, candidate ingredient names, candidate targets of action, and candidate types of action are determined; ambiguity correction is performed on each candidate based on its contextual position and adjacent entity combination; and the ambiguity-corrected candidates are written into a TCM node table to determine a TCM node dataset; the TCM node table includes node category, source identifier, and location index.

[0051] Specifically, TCM data includes TCM textual data, clinical medical record textual data, and pharmacological data, covering both traditional literature and modern experimental data. Examples include prescription compatibility records in the *Treatise on Febrile and Miscellaneous Diseases* (TCM textual data), a physician's record of a patient's treatment with Huangqi Decoction (clinical medical record textual data), and experimental records of TCM components inhibiting tumor cell growth in vitro (pharmacological data). Punctuation marks are symbols used to break sentences in TCM texts and clinical medical record texts, including commas, periods, semicolons, and TCM-specific separators such as commas and semicolons in prescription compatibility, used to define pauses and boundaries in the text's semantics. Prescription structure refers to the compatibility of TCM prescriptions, including structural elements such as principal, assistant, adjuvant, and guide herbs, dosage ratios, and usage instructions. For example, a prescription might have the compatibility structure of "principal herb (Ephedra), assistant herb (Cinnamon twig), adjuvant herb (Apricot kernel), and guide herb (Licorice)." The text fragment set is a collection of independent text fragments formed after segmenting traditional Chinese medicine (TCM) text data and clinical medical record text data. Each fragment contains complete local semantics (such as a description of the efficacy of a single herb or a record of a set of prescription combinations). Field keys represent keywords in pharmacological and omics data, such as "experiment type" and "concentration of action" in pharmacological data, and "protein name" and "expression level" in omics data. Sample identifiers are used to distinguish different experimental samples and detection objects, such as "mouse number 001" in pharmacological experiments and "patient sample number S023" in omics detection. Indicator names are the names of specific detection indicators in pharmacological and omics data, such as "anti-inflammatory rate" and "cell survival rate" in pharmacological data, and "protein expression level" and "gene methylation degree" in omics data. Dictionary matching is the matching of entities in text fragments and record fragments based on a professional TCM dictionary. For example, dictionary matching locates "ephedra" as a herb name and "anti-inflammatory" as a type of action. Synonym merging unifies entities with the same meaning but different expressions within a fragment. For example, "Huangqi" and "Huangqi" (different names for the same herb) are merged into the same candidate herb name, and "inhibition" and "antagonism" are merged into the same candidate action type. The Traditional Chinese Medicine (TCM) node table stores a standardized table of candidates after ambiguity correction. It includes three main elements: node category, source identifier, and location index. The node category indicates the specific type of TCM node, corresponding to formula nodes, ingredient nodes, target nodes, and pathway nodes, clarifying the node's attributes. The source identifier indicates the data source corresponding to the TCM node, such as "TCM text - Volume 12 of *Compendium of Materia Medica*" or "Pharmacological data - in vitro anti-inflammatory experiment." The location index indicates the specific location information of the TCM node in the original data and corresponding fragment, such as the paragraph number of the text fragment or the field position of the record fragment.

[0052] Understandably, for TCM text data and clinical medical record text data, a segmentation strategy of punctuation marks + drug names + prescription structure is adopted. Punctuation marks are used as the foundation to define semantic pause boundaries, avoiding semantic breaks after segmentation. Drug names are used as anchors to segment the text around single herbs and compound prescription names, ensuring that drug-related information is presented in a concentrated manner. Prescription structure is then used as a logical guide, segmenting the text according to elements such as the principal, assistant, adjuvant, and guide herbs, usage, and dosage, so that each segment corresponds to a clear prescription logic. For pharmacological and omics data, a segmentation strategy of field keys + sample identifiers + indicator names is adopted. Field keys are used to lock the dimensions of the data to clarify the attributes to which the data belongs. By associating sample identifiers with specific detection objects, the system ensures data traceability back to the original data. It focuses on the core detection content through indicator names, breaking down complex data into corresponding independent records, ultimately forming a set of record fragments. This approach solves the problems of inconsistent splitting standards, semantic loss, and disorganized data from multiple sources. The differentiated splitting strategy aligns with the semantic logic of text data while adapting to the dimensions of structured data, improving the accuracy and relevance of data splitting. After generating the text fragment set and the record fragment set, the data segmentation unit, relying on a professional dictionary of traditional Chinese medicine, performs sentence-by-sentence matching on the content of both sets, accurately identifying drug names, ingredient names, target objects, and types of effects. The corresponding entity content and location are then identified. Synonym merging is then performed on the matched entities, unifying terms with the same meaning but different expressions to eliminate interference from "different names for the same medicine" and "different descriptions of the same thing." For example, "Huangqi" is merged into "Huangqi," and "antagonistic" is merged into "inhibitory," thus standardizing entity descriptions. Ultimately, candidate drug names, candidate ingredient names, candidate targets of action, and candidate types of action are selected. This solves the entity recognition bias caused by inconsistent expressions of traditional Chinese medicine terminology, ensuring the accuracy of entity localization. Ambiguity correction, on the one hand, combines the contextual position of each candidate, analyzing its surrounding language and sentence structure to determine the applicable scenario of the entity; for example, the candidate drug name "Danggui." In the TCM text "Angelica sinensis combined with Astragalus membranaceus for blood replenishment," the contextual semantics clearly point to the name of the medicine. On the other hand, by combining adjacent entity combinations, the semantics are verified by the located entities adjacent to each candidate. For example, the entity "licorice" has adjacent entities "ephedra" (the name of the medicine) and "harmonizing various medicines" (the type of action). Through the semantic association between entities, it can be clearly identified that "licorice" is the candidate name of the medicine, rather than other meanings. This effectively avoids the identification error of polysemous entities and avoids the chain deviation in subsequent knowledge structure construction and model reasoning caused by entity identification errors. After completing the ambiguity correction, the data segmentation unit writes all the corrected candidates into the TCM node table, laying the foundation for the data carrier for the subsequent process.

[0053] In some embodiments of this application, when determining the association relationship based on the co-occurrence order and function description of TCM nodes, the process includes: a data segmentation unit determining node pairs based on the occurrence order of TCM nodes, encoding the cross-sentence distance and adjacent words of the node pairs, determining the sequence features, extracting the function description phrases between node pairs and mapping them to function type labels, determining the description features, determining the relationship records based on the sequence features and description features, the relationship records containing the starting node identifier, the ending node identifier, the function type label and the source fragment identifier, merging relationship records with the same starting node identifier, ending node identifier and function type label, and retaining the source fragment identifier, and determining the association relationship dataset based on the merging result.

[0054] Specifically, a node pair is a combination of two TCM (Traditional Chinese Medicine) nodes arranged in the order of their appearance. For example, in the fragment "ephedrine acts on adrenaline receptors," the node pair formed in the order of appearance is "ephedrine (component node) - adrenaline receptor (target node)." The sentence-crossing distance is the distance between the sentences containing the two TCM nodes in a node pair. Sentence-crossing distance measures the strength of the positional association between nodes; the closer the interval, the stronger the semantic association. For example, if node A is in sentence 1 and node B is in sentence 2, the sentence-crossing distance is 1; if node A is in sentence 1 and node B is in sentence 4, the sentence-crossing distance is 3. Adjacent words are the words preceding and following the two nodes in a node pair, usually verbs, prepositions, adverbs, etc., which help determine the semantic direction and strength of the association between nodes. For example, the adjacent word for the node pair "ephedrine - adrenaline receptor" is "acts on," which helps determine that there is an actional association between the two. Sequence features are feature information that reflects the order of node appearance, positional association, and semantic tendency, formed by encoding the sentence-crossing distance and adjacent words of the node pair. Descriptive phrases, located between node pairs, are phrases or short sentences that describe the interaction and logical relationship between the two nodes, serving as the core basis for defining the action type. For example, "inhibitory binding" and "promoting activation" in the node pair "ephedrine-adrenergic receptor" are both descriptive phrases. Action type labels are fixed labels obtained by standardizing and mapping descriptive phrases, used to uniformly identify the logical type of action between nodes to avoid discrepancies in expression. For example, "inhibitory binding" and "antagonistic effect" are mapped to the label "inhibitory action," and "promoting activation" and "synergistic effect" are mapped to the label "activating action." Descriptive features are feature information transformed from action type labels, accurately representing the nature and logical relationship of the action between node pairs, and together with sequence features, supporting the determination of association relationships. For example, the descriptive feature corresponding to the label "inhibitory action" represents the existence of a mutually inhibitory association logic between nodes. The starting node identifier is a unique code assigned to the first node in the node pair, used to accurately locate the starting node and ensure record traceability. For example, the identifier "C001" is assigned to "ephedrine" as the starting node identifier for the relationship record. The endpoint node identifier is the code assigned to the node that appears later in the node pair. It is used to accurately locate the endpoint node and, together with the starting node identifier, locks the node pair.The source fragment identifier is a unique code assigned to the original fragment (text fragment or record fragment) corresponding to the relation record. The data segmentation unit traverses the generated TCM node dataset and synchronously associates the corresponding text fragment set and record fragment set, tracing the specific position and order of each TCM node in the original fragment. According to the natural order of the nodes in the fragment, nodes with semantic relationships are paired to form node pairs. Node pairs are not randomly combined, but rather nodes that appear adjacently or are close in interval and revolve around the same pharmacological logic (such as drug name and ingredient, ingredient and target, target and pathway) are preferentially combined to exclude unrelated node combinations (such as drug name nodes of different prescriptions). For example, in the fragment "Astragaloside A inhibits COX-2 and regulates the NF-κB signaling pathway", the node pairs "Astragaloside A-COX-2" and "COX-2-NF-κB signaling pathway" are combined in sequence, excluding other node combinations unrelated to this fragment. For each node pair, the descriptive phrases describing their interaction are precisely extracted. Focusing on the core statements within the node pair, phrases or short sentences that can explain the interaction method and logical relationship are selected, while irrelevant modifiers (such as modal particles and adjectives) are removed. Based on sequence and descriptive features, a complete relationship record is generated for each node pair, explicitly including the starting node identifier (the unique code of the first node to appear), the ending node identifier (the unique code of the second node to appear), the action type label, and the source fragment identifier. For all generated relationship records, those with identical starting node identifiers, ending node identifiers, and action type labels are merged, while the source fragment identifier is retained. This reduces the amount of data in the dataset while aggregating data from multiple sources with the same relationship, improving the reliability of the relationship. It transforms scattered nodes into a logically related dataset, thereby ensuring the synergy between data segmentation and the model and avoiding process fragmentation.

[0055] In some embodiments of this application, when uniformly encoding TCM nodes and relationships, and determining relationship identifiers and knowledge structures, the following steps are taken: a data segmentation unit generates node codes for the TCM node dataset based on node categories and standardized names. The node codes include node category codes and name verification codes. A relationship identifier is generated for the relationship dataset based on start node identifiers, end node identifiers, and action type labels. The relationship identifier includes edge type codes and structure verification codes. An adjacency index is determined based on the node codes and relationship identifiers. The adjacency index includes a set of outgoing edge relationship identifiers and a set of incoming edge relationship identifiers corresponding to each node code. The node codes are divided into prescription node codes, ingredient node codes, target node codes, and pathway node codes based on node categories. The hierarchical connection direction of the encoding division is limited based on the adjacency index to determine the knowledge structure.

[0056] Specifically, the data segmentation unit generates node codes for the TCM node dataset. Each node code consists of a node category code and a name check code. The node category code is a unique code assigned based on the node category (formula node, ingredient node, target node, pathway node) to quickly distinguish node types. For example, a "F" code is assigned to a formula node, a "C" code to an ingredient node, a "T" code to a target node, and a "P" code to a pathway node. The name check code is generated by encrypting and verifying the standardized name of the node. This ensures that the same standardized name corresponds to a unique node code, avoiding node confusion due to name duplication or errors. For example, "Ma Huang Tang" is a formula node, and its node code can be composed of an "F" code plus a corresponding name check code, which clarifies the category and ensures uniqueness, improving the accuracy of node management. Relationship identifiers are generated based on the start node identifier, end node identifier, and action type label in the association dataset. Each relationship identifier includes an edge type code and a structure check code. The edge type code is assigned a unique code to the action type label between nodes. For example, "inhibitory action" corresponds to the "Y1" code, and "activating action" corresponds to the "Y2" code. The structure check code is generated by verifying the combination of the start and end node identifiers and the edge type code to ensure the uniqueness and integrity of the relationship identifier. For example, the relationship identifier "Astragaloside A (component node) inhibits COX-2 (target node)" is composed of the corresponding edge type code and the structure check code, which solves the problem of confusing descriptions of different association relationships. The adjacency index is determined based on node codes and relationship identifiers. For each node code, the adjacency index generates a set of outgoing edge relationship identifiers and a set of incoming edge relationship identifiers. The outgoing edge relationship identifier set contains all relationship identifiers for the node as a starting point, and the incoming edge relationship identifier set contains all relationship identifiers for the node as an ending point. For example, the incoming edge set of the "COX-2" target node contains the relationship identifiers of all component nodes acting on it, and the outgoing edge set contains the relationship identifiers of its regulatory pathway nodes. This allows for rapid location of the upstream and downstream relationships of each node. Based on node categories, node codes are divided into formula node codes and component node codes. Point coding, target node coding, and pathway node coding rely on adjacency indexing to limit the hierarchical connection direction of the coding division. That is, the formula node coding can only be associated with the component node coding through outgoing edges, the component node coding can only be associated with the target node coding through outgoing edges, and the target node coding can only be associated with the pathway node coding through outgoing edges. At the same time, the source of incoming edges of each node coding is clearly defined, forming a hierarchical knowledge structure. This conforms to the pharmacological logic of "formula-component-target-pathway" in the discovery of new Chinese medicine drugs, and provides a standardized and structured knowledge carrier for the collaborative work of subsequent large and small models, avoiding logical confusion in model reasoning.

[0057] In some embodiments of this application, when inputting a prescription node and its associated component nodes into a large model based on relation identifiers, the process includes: a first collaborative unit extracting first-order adjacent component nodes and corresponding relation identifiers from the knowledge structure, starting from the target prescription node, determining a prescription subgraph, and performing serialization encoding on the prescription subgraph to determine a triple sequence containing prescription node encoding, component node encoding, and relation identifiers. Using the edge type code of the relation identifier as a constraint, the triple sequence is input into the large model. The large model limits the extended node categories and edge types based on the constraints and outputs the relation identifier index.

[0058] In some embodiments of this application, when generating paths for target nodes and pathway nodes based on relation identifiers to determine candidate action path datasets, the process includes: a first collaborative unit expanding the adjacency index of the knowledge structure layer by layer based on the large model, with the expansion order being prescription nodes, ingredient nodes, target nodes, and pathway nodes. During each expansion, adjacency items that satisfy the constraints are selected, several candidate paths are determined, a path string is determined based on the path code of each candidate path, and the corresponding relation identifier index is embedded in the path string. After embedding, deduplication is performed, and the candidate action path dataset is determined based on the deduplication result.

[0059] Specifically, the first collaborative unit focuses on the target prescription node in the knowledge structure (i.e., the TCM node corresponding to the prescription for which new drug research is to be carried out, such as the "Ma Huang Tang node"), extracts the first-order adjacent component nodes of this node (i.e., component nodes directly associated with the target prescription node, such as the ephedrine and pseudoephedrine nodes directly associated with the Ma Huang Tang node) and their corresponding relationship identifiers, and integrates these nodes and relationships to form a prescription subgraph (that is, a knowledge substructure with the target prescription node as the core, containing only its directly associated component nodes and corresponding relationships). Then, the prescription subgraph is serialized and encoded, thereby transforming it into a triple sequence containing prescription node code, component node code, and relationship identifier (e.g., "F001 (Ma Huang Tang node code), C001 (ephedrine node code), R00...). The system uses an ordered combination of "1 (relationship identifiers for inclusion-type effects)" as a constraint. The edge type code in the relationship identifier (the core code of the relationship identifier, corresponding to the type of effect between nodes, such as edge type code B1 for "inclusion-type effect" and edge type code J1 for "combination-type effect") is used as a constraint. The triple sequence is input into the large model. The large model strictly limits the expandable node categories (only target nodes and pathway nodes are allowed to be expanded, and irrelevant node types are excluded) and edge types (only pharmacologically related edge types such as combination and regulation are allowed to be expanded, and meaningless edge types are excluded) based on this constraint. The output is a relationship identifier index used to accurately locate the association relationship identifier. By using the edge type code constraint, the large model is prevented from expanding without direction, ensuring that the input data focuses on the core component association of the target prescription, thus improving the targeting and efficiency of the model inference. The first collaborative unit, based on the relation identifier index output by the large model, expands layer by layer in the adjacency index of the knowledge structure (which records the set of outgoing and incoming edge relation identifiers corresponding to each node's code) in a fixed order of "formula node → ingredient node → target node → pathway node". Each expansion only filters adjacency items that satisfy the edge type code constraint (i.e., outgoing / incoming edge association items corresponding to the node in the adjacency index; for example, for ingredient nodes, only target node adjacency items corresponding to "combination type" edge type codes are selected). This generates several candidate paths (e.g., "Ephedra Decoction → Ephedrine → Adrenaline Receptor → Sympathetic Nerve Excitation Pathway" and "Ephedra Decoction → Pseudoephedrine → Nasal Mucosal Vascular Smooth Muscle Receptor → Vasodilator Pathway"). (etc.); then, a unique path code is assigned to each candidate path and converted into a path string (the string form of the path code). The relation identifier index is embedded into the path string. The embedded path string is deduplicated (to remove duplicate path strings and avoid dataset redundancy). The candidate action path dataset is determined based on the deduplication result. The layer-by-layer expansion strategy strictly follows the pharmacological logic hierarchy to ensure that the candidate paths conform to the actual mechanism of the action of traditional Chinese medicine. The edge type code constraint further filters invalid paths. The final generated candidate action path dataset not only covers the pharmacological path of the target prescription, but also ensures the standardization and traceability of data processing, thereby improving the overall collaborative efficiency and reasoning accuracy of the intelligent agent.

[0060] In some embodiments of this application, when the component-target correspondence in the candidate action path dataset is input as a constraint into the small model, the method includes: a second collaborative unit extracting path strings from the candidate action path dataset, extracting the component node code and target node code of each path string, determining a component-target pairing list, mapping the component node code of each extracted path string to a component vector, mapping the target node code of each extracted path string to a site vector, and inputting the component-target pairing list and the corresponding component vector and site vector into the small model, wherein the small model uses the component-target pairing list as an index constraint to determine the source path string identifier.

[0061] In some embodiments of this application, when parallel solving of affinity results and molecular property results between components and targets to determine path evaluation results, the following steps are included: a small model performs parallel affinity calculations and molecular property calculations on the input data, wherein the affinity calculation uses component vectors and site vectors as joint inputs and outputs pair-level affinity results, and the molecular property calculation uses component vectors as inputs and outputs component-level property results; a second collaborative unit backfills the pair-level affinity results to the corresponding path strings according to the source path string identifier, and backfills the component-level property results to all path strings containing the component according to the component node encoding; the pair-level affinity results and component-level property results are summarized for each path string to determine the path evaluation vector; and the path evaluation vector and the summarized corresponding path strings are determined as the path evaluation results.

[0062] Specifically, the second collaborative unit extracts path strings (path-encoded strings) from the candidate action path dataset and extracts component node codes and target node codes from each path string, integrating them to form a component-target pairing list (used to record the correspondence between components and targets in each path, such as the pairing of ephedrine node codes with adrenaline receptor node codes, and the pairing of pseudoephedrine node codes with nasal mucosal vascular smooth muscle receptor node codes). Then, it maps the component node codes of each path string to component vectors (converting the physicochemical properties and structural features of component nodes into vector form, such as converting the molecular structure and activity features of ephedrine into component vectors), and maps the target node codes to site vectors (converting the binding sites and structural features of target nodes into vector form, such as converting the binding site features of adrenaline receptors into site vectors). The component-target pairing list, the corresponding component vectors, and the site vectors are then input into the small model. The small model uses the component-target pairing list as an index constraint to accurately associate each component-target pairing with the path string, determining the source path string identifier (corresponding to the path string). A unique identifier for the path string is used to trace the path to which the pairing result belongs, avoiding confusion between calculation results from different paths. For example, the unique identifier of the original path string associated with the pairing of ephedrine and adrenaline receptor is used. Index constraints ensure the association between component-target pairing and path, avoiding data misalignment. At the same time, vector mapping transforms unstructured component and target features into model calculation forms, laying the foundation for parallel solution. The small model simultaneously performs parallel affinity calculations and molecular property calculations on the input data. Affinity calculation uses component vectors and site vectors as joint inputs, focusing on the binding characteristics of components and targets, and outputs pairing-level affinity results (results on the binding tightness of single component-target pairings, such as the high affinity result of ephedrine and adrenaline receptor, and the low affinity result of a component and target). Molecular property calculation uses only component vectors as inputs, analyzes the physicochemical properties of the components themselves, and outputs component-level property results (results on the molecular properties of single components, covering drug-related characteristics such as lipid solubility, stability, and toxicity, such as the component-level property result of ephedrine being "moderately lipid-soluble and without significant toxicity").The second collaborative unit backfills the pairing-level affinity results into the corresponding path strings according to the source path string identifier, ensuring that each path string corresponds only to the affinity result of its own component-target pairing. At the same time, it backfills the component-level property results into all path strings containing that component node according to the component node code. Then, it summarizes the corresponding pairing-level affinity results and component-level property results for each path string, and integrates the summarized quantitative results to form a path evaluation vector. The path evaluation vector can comprehensively reflect the feasibility of the path, covering multiple dimensions such as affinity and molecular properties. The path evaluation vector and the corresponding path string are jointly determined as the path evaluation result. The index constraint realizes the precise correlation between the calculation results and the path. Parallel computing improves the efficiency of the calculation and avoids the time-consuming problem of serial computing. It ensures that the path evaluation results can comprehensively and quantitatively reflect the feasibility and druggability of each candidate path, and makes up for the problem of the disconnect between quantitative results and path in traditional technologies.

[0063] In some embodiments of this application, when screening candidate action path datasets based on path evaluation results, the following steps are included: a collaborative update unit groups the candidate action path datasets according to the same prescription node encoding based on the path evaluation vector of all path evaluation results, and removes path strings with inconsistent structure check codes or conflicting source fragment identifiers to determine the target action path dataset.

[0064] In some embodiments of this application, when adding experimental verification data to the knowledge structure and updating the relation identifier, large model, and small model, the process includes: a collaborative update unit parsing the experimental verification data, determining the experimental object identifier, measurement index identifier, experimental source identifier, and experimental result identifier, mapping the experimental object identifier to the source path string identifier in the target action path dataset, writing the experimental result identifier into the knowledge structure based on the mapping result, writing the experimental source identifier into the relation dataset, determining the incremental sample set based on the writing results of the experimental result identifier and the experimental source identifier, and inputting the incremental sample set into the large model and small model for parameter updates.

[0065] Specifically, the collaborative update unit, based on the path evaluation vectors in all path evaluation results, groups all path strings in the candidate action path dataset according to the same prescription node code (e.g., prescription node code F001 for Ephedra Decoction), ensuring that all paths in the same group belong to the same target prescription. Then, it removes path strings with inconsistent structural check codes (codes used in relation identifiers to verify the integrity of the association structure; inconsistent structural check codes for the same association indicate a logical error in the association) or conflicting source fragment identifiers (unique identifiers for data fragments; conflicting source fragment identifiers indicate a contradiction in the data basis of the path). The system filters out logically consistent path strings with conflict-free data to determine the target action path dataset. By grouping and focusing on paths of the same prescription and verifying the consistency of structure and source, invalid paths with logical errors and conflicting data are eliminated, improving the reliability of the target action path dataset. The collaborative update unit parses the experimental verification data, extracting and determining the experimental object identifier (a unique identifier corresponding to the component, target, or path targeted in the experiment, such as the exclusive identifier for "ephedrine-adrenergic receptor pairing"), measurement index identifier (the identifier corresponding to the core indicators detected in the experiment, such as "component-target affinity" and "component toxicity"), and the actual... This method verifies source identifiers (the origin of the experiment, such as a TCM laboratory or a core journal article) and result identifiers (standardized identifiers of experimental results, such as "high affinity," "low toxicity," or "no activity"). It maps the experimental object identifiers to the corresponding source path string identifiers in the target action path dataset, ensuring that experimental results are accurately attributed to their respective paths. Based on the mapping results, the experimental result identifiers are written into the knowledge structure, supplementing the empirical verification results of the paths and improving the empirical evidence of the knowledge structure. Simultaneously, the experimental source identifiers are written into the association dataset to enhance the credibility of the associations. This process is based on the writing of experimental result identifiers and experimental source identifiers. The input results are integrated to form an incremental sample set (a standardized sample set composed of newly added experimental data, containing complete information such as experimental subjects, indicators, sources, and results). This incremental sample set is input into the large and small models to update parameters, enabling the large and small models to learn from the newly added experimental validation data to optimize reasoning and computational logic. This transforms the knowledge structure from theoretical correlation to empirical correlation, thereby continuously optimizing the path generation logic of the large model and the quantitative calculation accuracy of the small model, forming a closed loop of data-model-experiment-update. This avoids the staticization of models and knowledge structures and improves the reliability of multi-dimensional medical large and small models working together.

[0066] In summary, the beneficial effects of this invention are as follows: Through the collaborative linkage of various units, a complete closed loop of data processing, path generation, quantitative evaluation, and iterative updates is established, solving the problems of lack of collaboration between models, inability to provide feedback on experimental data, and fragmented processes. This achieves the integration of various stages in the discovery of new traditional Chinese medicine (TCM) drugs, transforming multi-source heterogeneous data into a standardized knowledge structure. Precise collaboration between large and small models is achieved through relational identification, making the inference results of large-scale models the computational constraints of small-scale models. Simultaneously, experimental data provides feedback on the knowledge structure and model parameters, improving the utilization rate of data and models. Furthermore, candidate path generation is based on the logical constraints of the knowledge structure, path evaluation is supported by quantitative calculations, and iterative updates are verified by experimental data, ensuring the feasibility and drugability of new drug discovery paths and providing a reliable research direction for the discovery of new TCM drugs. Through dynamic knowledge structures and collaborative update mechanisms, multi-source TCM data can achieve constrained data transfer, and inference and computational models can collaborate efficiently, thus adapting to the parallel research needs of TCM drug discovery.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-dimensional traditional Chinese medicine size model collaborative agent system, characterized in that, include: The data segmentation unit is configured to perform semantic segmentation on TCM data and omics data, extract TCM nodes based on drug name, ingredient name, target of action and action type, determine the association relationship based on the co-occurrence order and action description of the TCM nodes, uniformly encode the TCM nodes and association relationship, and determine the relationship identifier and knowledge structure. The knowledge structure includes prescription nodes, ingredient nodes, target nodes and pathway nodes. The first collaborative unit is configured to input the prescription node and its associated component node into the large model based on the relationship identifier, and generate paths for the target node and pathway node based on the relationship identifier to determine the candidate action path dataset; The second collaborative unit is configured to input the component-target correspondence in the candidate action path dataset as a constraint into the small model, and to solve the affinity results and molecular property results between the components and the target in parallel to determine the path evaluation results. The collaborative update unit is configured to filter the candidate action path dataset based on the path evaluation results, determine the decision output based on the filtered candidate action path dataset, add the experimental verification data to the knowledge structure, and update the relation identifier, large model, and small model.

2. The multi-dimensional traditional Chinese medicine size model collaborative agent system according to claim 1, characterized in that, When performing semantic segmentation on TCM data and omics data, and extracting TCM nodes based on drug name, ingredient name, target of action, and type of action, the following is included: The TCM data includes TCM text data, clinical medical record text data, and pharmacological data; The data segmentation unit segments the TCM text data and clinical medical record text data according to punctuation marks, drug names, and prescription structures to determine the text fragment set, and splits the pharmacological data and omics data according to field keys, sample identifiers, and indicator names to determine the record fragment set; In the set of text fragments and the set of record fragments, based on entity localization using dictionary matching and synonym merging, candidate drug names, candidate ingredient names, candidate target of action, and candidate action type are determined, and ambiguity correction is performed on each candidate based on the context position and the combination of adjacent entities. Each candidate for ambiguity correction is written into the Traditional Chinese Medicine (TCM) node table to determine the TCM node dataset. The TCM node table includes node category, source identifier, and location index. 3.The multi-dimensional TCM size model collaborative agent system according to claim 2, characterized in that, When determining the association relationship based on the co-occurrence order and function description of the TCM nodes, the following are included: The data segmentation unit determines node pairs based on the order of occurrence of the TCM nodes, encodes the cross-sentence distance and adjacent words of the node pairs to determine the sequence features, extracts the descriptive phrases of the effects between the node pairs and maps them to effect type labels to determine the descriptive features, and determines the relationship record based on the sequence features and descriptive features. The relationship record includes the starting node identifier, the ending node identifier, the effect type label and the source fragment identifier. Relationship records with the same start node identifier, end node identifier, and function type label are merged, while the source fragment identifier is retained. The association relationship dataset is determined based on the merging result.

4. The multi-dimensional traditional Chinese medicine size model collaborative agent system according to claim 3, characterized in that, When uniformly encoding the TCM nodes and their relationships, and determining the relationship identifiers and knowledge structures, the following are included: The data segmentation unit generates node codes for the TCM node dataset based on node category and standard name. The node codes include node category codes and name verification codes. The unit generates relationship identifiers for the association dataset based on the starting node identifier, ending node identifier, and function type label. The relationship identifiers include edge type codes and structure verification codes. The adjacency index is determined based on the node codes and relation identifiers. The adjacency index includes a set of outgoing relation identifiers and a set of incoming relation identifiers corresponding to each node code. Based on the node category, the node codes are divided into prescription node codes, ingredient node codes, target node codes, and pathway node codes. The hierarchical connection direction of the code division is limited based on the adjacency index to determine the knowledge structure.

5. The multi-dimensional TCM large and small model collaborative intelligent agent system according to claim 4, characterized in that, When inputting the prescription node and its associated component nodes into the large model based on the relationship identifier, the following steps are included: The first collaborative unit takes the target prescription node as the starting point in the knowledge structure, extracts the first-order adjacent component nodes and corresponding relation identifiers, determines the prescription subgraph, and performs serialization encoding on the prescription subgraph to determine a triplet sequence containing prescription node encoding, component node encoding and relation identifier; Using the edge type code of the relation identifier as a constraint, the triple sequence is input into the large model. The large model limits the extended node categories and edge types based on the constraint and outputs the relation identifier index.

6. The multi-dimensional TCM large and small model collaborative intelligent agent system according to claim 5, characterized in that, When generating paths for the target nodes and pathway nodes based on the relationship identifiers to determine the candidate action path dataset, the process includes: The first collaborative unit expands the adjacency index of the knowledge structure based on the large model, and the expansion order is as follows: prescription node, ingredient node, target node and pathway node. During each expansion, adjacent items that satisfy the constraints are selected, several candidate paths are determined, a path string is determined based on the path code of each candidate path, and the corresponding relation identifier index is embedded in the path string. After embedding, deduplication is performed, and the candidate action path dataset is determined based on the deduplication result.

7. The multi-dimensional TCM large and small model collaborative intelligent agent system according to claim 6, characterized in that, When inputting the component-target correspondence in the candidate action path dataset as a constraint into the small model, the following is included: The second collaborative unit extracts path strings from the candidate action path dataset and extracts the component node code and target node code of each path string to determine the component-target pairing list. The component nodes of each extracted path string are encoded and mapped to component vectors, and the target nodes of each extracted path string are encoded and mapped to site vectors. The component-target pairing list and the corresponding component vectors and site vectors are input into the small model. The small model uses the component-target pairing list as an index constraint to determine the source path string identifier.

8. The multi-dimensional TCM large and small model collaborative intelligent agent system according to claim 7, characterized in that, When determining the path evaluation results by parallelly solving the affinity and molecular property results between the components and the target, the following are included: The small model performs parallel affinity calculations and molecular property calculations on the input data. The affinity calculation uses the component vector and the site vector as joint inputs and outputs pairwise affinity results. The molecular property calculation uses the component vector as inputs and outputs component-level property results. The second collaborative unit backfills the pairing-level affinity results to the corresponding path strings according to the source path string identifier, and backfills the component-level property results to all path strings containing the component according to the component node code. It summarizes the pairing-level affinity results and component-level property results for each path string, determines the path evaluation vector, and determines the path evaluation vector and the summarized corresponding path string as the path evaluation result.

9. The multi-dimensional TCM large and small model collaborative intelligent agent system according to claim 8, characterized in that, When filtering the candidate action path dataset based on the path evaluation results, the following steps are included: The collaborative update unit groups the candidate action path dataset according to the same prescription node encoding based on the path evaluation vector of all path evaluation results, and removes path strings with inconsistent structure check codes or conflicting source fragment identifiers to determine the target action path dataset.

10. The multi-dimensional TCM large and small model collaborative intelligent agent system according to claim 9, characterized in that, When adding experimentally validated data to the knowledge structure and updating the relation identifiers, large model, and small model, the process includes: The collaborative update unit parses the experimental verification data to determine the experimental object identifier, measurement index identifier, experimental source identifier, and experimental result identifier. It maps the experimental object identifier to the source path string identifier in the target action path dataset. Based on the mapping result, it writes the experimental result identifier into the knowledge structure and the experimental source identifier into the association dataset. Based on the writing results of the experimental result identifier and the experimental source identifier, it determines the incremental sample set and inputs the incremental sample set into the large model and the small model for parameter updates.

Images