Method for identifying traditional chinese medicine and western medicine perturbation atlas based on deep learning and protein interaction network topology

Abstract

The application provides a Chinese and Western medicine perturbation atlas recognition method based on deep learning and protein interaction network topology, and comprises the following steps: S1, collecting protein interaction data, constructing a protein interaction graph and a node index mapping table; S2, constructing a node set and a non-directional edge set; S3, constructing a node feature matrix; S4, constructing a deep graph structure regression model based on a graph attention neural network; S5, through training, the model obtains a continuous regression prediction value of the association between each protein node in the protein interaction network and a disease; S6, constructing a disease high-heat protein node set; S7, constructing a drug-protein-disease interaction network and calculating a coverage function of the drug; S8, based on a topological similarity and a redundancy penalty strategy, constructing a non-redundant drug perturbation atlas under different topological granularities; the drug perturbation atlas can provide a basis for the multi-pathway synergistic mechanism of the drug and the intervention ability of the drug on the disease state.

Description

Technical Field

[0001] This invention belongs to the fields of medicine and artificial intelligence technology, specifically relating to a method for identifying perturbation maps of traditional Chinese and Western medicines based on a deep learning graph attention neural network architecture and the topology of human protein-protein interaction networks. Background Technology

[0002] The occurrence and development of diseases are typically driven by multi-level, cross-pathway biological processes. While modern omics technologies provide a wealth of data on disease targets, existing methods still struggle to establish a scientific prioritization system among these targets. Furthermore, the molecular basis of diseases is often not caused by a single gene abnormality, but rather by the overall perturbation of multiple nodes and their interactions within a biological network, influenced by genetic variation, pathogenic mutations, epigenetic regulation, and microenvironmental signals. Therefore, traditional target screening strategies fail to reflect the hierarchical structure and dynamic regulatory characteristics of disease networks, limiting their ability to elucidate the core driving forces of diseases and construct system-level models.

[0003] In elucidating drug action mechanisms, the paradigm differences between traditional Chinese medicine (TCM) and Western medicine further increase the difficulty of modeling. Western medicine is typically characterized by single components, few targets, and well-defined mechanisms of action, with interventions concentrated on local pathways or key targets, making it suitable for fine-grained perturbation spectrum analysis. In contrast, TCM exhibits a systemic regulatory model involving multiple components, targets, and pathways, with its efficacy achieved through the synergistic effects of multiple nodes and pathways within the network. Due to significant functional similarities and redundancy among TCM targets, and their wide distribution, traditional methods struggle to accurately assess their network-level action characteristics, often yielding only fragmented target predictions that fail to reflect their comprehensive perturbation capabilities across multiple functional modules. Currently, there is a lack of a method capable of simultaneously characterizing the wide-area perturbation spectrum of TCM and the local regulatory spectrum of Western medicine within a unified protein-protein interaction network, which limits its application in scenarios such as comparing the mechanisms of TCM and Western medicine, studying combination therapies, and drug repositioning.

[0004] In recent years, advancements in deep learning, graph neural networks, and network topology analysis have enhanced the ability to model protein-protein interaction network structures and systemic perturbation effects. However, existing methods still struggle to meet the comprehensive needs of both disease and drug development. At the disease level, traditional network pharmacology struggles to simultaneously incorporate multiple factors such as genetic variation, pathogenic mutations, and epigenetic regulation, and also fails to analyze the cross-level interactions of targets within the network, thus limiting the network foundation for constructing perturbation spectra. At the drug level, existing computational methods struggle to simultaneously address the fine-grained local regulatory characteristics of Western medicine and the wide-area systemic perturbation patterns of traditional Chinese medicine. The lack of a framework capable of expressing the action patterns of both types of drugs in a unified topological space makes it difficult to compare systemic effects across drug types and to form structured, quantifiable perturbation spectrum representations. Therefore, a model integrating the feature learning capabilities of deep learning with the analytical capabilities of network topology is needed to achieve systematic and unified modeling of perturbation features of both Western and traditional Chinese medicine. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies in modeling disease networks and drug action mechanisms. It proposes a method for identifying perturbation maps of traditional Chinese medicine and Western medicine based on deep learning and protein interaction network topology. By constructing a unified network space, introducing a hierarchical topology modeling strategy, and combining the structural learning capabilities of graph neural networks, it achieves the joint identification of perturbation spectra of multi-component systems in traditional Chinese medicine and fine-tuning spectra of Western medicine. This results in quantifiable system-level perturbation maps applicable to multiple types of drugs, providing a new methodological foundation for comparative analysis and mechanistic research on the action characteristics of traditional Chinese and Western medicine.

[0006] To achieve the above objectives, this invention provides a method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology. The method includes the following steps:

[0007] S1: Collect human protein-protein interaction network data, which consists of several protein-protein interaction pairs. Each protein-protein interaction pair includes two interacting proteins, and each protein has a unique biological identifier. Summarize and remove duplicate protein identifiers from all interaction pairs to obtain a set of network protein nodes. The set of network protein nodes consists of proteins that have participated in at least one protein-protein interaction. Subsequently, assign a unique integer index to each protein in the set of network protein nodes to construct a protein node index mapping table for subsequent graph structure modeling and calculation.

[0008] S2: Map the protein interaction pairs obtained in S1 to edge indices in the graph structure. The edge indices are used to represent the connection relationship between two protein nodes in the protein interaction network. The obtained edge indices are then undirected so that each pair of protein interactions is represented in the form of bidirectional edges. This results in a standardized human protein interaction graph dataset consisting of a set of network protein nodes and a set of undirected protein edges, which serves as the structural input data for graph neural network modeling.

[0009] S3: Initialize node feature vectors for all protein nodes in the human protein-protein interaction network node set obtained in S2 to represent the relationship between the protein nodes and the disease to be analyzed. By default, the initial node feature vector of each protein node is set to zero vector, which is used to indicate that the node does not have disease association attributes in the initial state. Based on this, for protein nodes with known disease association, assign corresponding label values ​​with disease association strength according to the disease association evidence, and use the protein node features and label values ​​as supervised learning data to participate in the model training constructed in S4; for protein nodes that do not contain disease label values, their feature vectors remain unchanged as zero vectors. The protein node feature matrix is ​​constructed in the above manner.

[0010] S4: Based on the set of undirected edges of proteins obtained in S2 and the node feature matrix obtained in S3, a deep graph structure regression model based on graph attention neural network is constructed. The model is composed of multiple layers of graph attention convolutional layers (GATConv) stacked sequentially. The model uses a multi-head attention mechanism to weight and aggregate the node features associated with the disease according to the interaction relationship between proteins. Each graph attention convolutional layer iterates on the weighted and aggregated node features under the constraint of the protein interaction network structure, so that the protein and disease-associated node features in the protein interaction graph dataset are continuously learned and updated, forming a deep representation of the protein node features, and obtaining the optimal deep graph structure regression model.

[0011] S5: Based on the optimal depth map structure regression model obtained in S4, under the premise of keeping the parameters unchanged, the regression model is used to infer the degree of association between proteins with zero vector node features and the disease to be analyzed and assign them label values. At the same time, the node feature vectors and label values ​​of protein nodes that already have label values ​​are updated. Through the above calculation, the model finally obtains the continuous regression prediction value of the association between each protein node and the disease in the protein interaction network through the output layer. The continuous regression prediction label value is in the form of continuous real numbers and is used to represent the association strength between the protein node and the target disease. The higher the value, the higher the potential association degree and network regulation importance of the protein in the disease network.

[0012] S6: Based on the protein node-disease association score obtained in step S5, sort all protein node features in the protein interaction network from high to low score; simultaneously obtain or directly input the target protein node set of the drug to be analyzed from public databases (TCMSP, HIT 2.0, BATMAN 2.0), and count the number of target protein nodes of the drug to be analyzed; according to the sorting results, select a group of disease targets with scores from high to low and a number consistent with the number of protein targets of the drug (e.g., 100 drug targets, 100 disease targets selected simultaneously from high to low score), which constitutes the set of disease high-temperature protein nodes composed of the proteins with the highest association score in the protein network related to the disease to be analyzed and targeted by the drug to be analyzed.

[0013] S7: Based on the set of protein nodes in the high-intensity disease region obtained in S6, a drug-protein-disease interaction network is constructed. Taking the protein target of each drug in the protein interaction network as the source node, the shortest path distance between it and each protein node associated with the disease in the drug-protein-disease interaction network is calculated. The spatial scale of topology propagation is adjusted by setting the shortest path length and the distance decay coefficient. On this basis, the shortest path distance is mapped to a continuous weighted value to construct a weighted matrix of drug protein target – high-intensity disease protein node. This matrix is ​​used to quantitatively evaluate the potential regulatory strength of the drug on the high-intensity disease protein node at different topological granularities and to calculate the drug coverage function.

[0014] S8: Calculate the topological similarity between any two protein targets based on the protein interaction network structure. Introduce a stepwise greedy selection strategy with a redundancy penalty term. Adjust the repulsion strength of similar protein targets by adjusting the penalty coefficient. Protein targets that generate positive marginal coverage gain in the set of protein nodes in the high-intensity disease area are classified as effective contributing protein targets and included in the non-redundant target set. This set is used to construct drug perturbation maps at different topological granularities. The positive marginal coverage gain mentioned in this step refers to whether adding a new protein target to a previously selected set of protein targets allows the drug to act on more previously uncovered disease-related proteins, or strengthens the effect on key disease regions. If a new protein target, after being added, can influence new disease-related nodes through the protein interaction network, or significantly enhance the regulatory effect on the high-intensity disease area, then the target is considered to have brought positive marginal coverage gain. If, after adding the target, the main area of ​​action highly overlaps with existing targets, with almost no increase in new regulatory range or effect, then its marginal coverage gain is considered small or zero, and it is not considered an effective contributing target. The core of this algorithm lies in prioritizing the selection of targets with the broadest coverage. Taking inflammation-related targets as an example, if target A can influence 10 inflammation-related targets through interactions, while target B can only influence 5, and their effects overlap, then target A should be preferred in the drug intervention strategy. Step S8 aims to achieve this goal—screening out key targets like A while eliminating suboptimal options like B.

[0015] The human protein-protein interaction network data used in S1 of this invention has high flexibility in its source. It can come from public databases such as STRING, HIPPIE, IntAct, and László Barabási, or from laboratory-built interaction datasets, thus making it suitable for various data scales and application scenarios. This invention does not strictly limit the format of the interaction data; only a set of interaction pairs consisting of paired protein identifiers is needed to meet the model construction requirements.

[0016] In S3 of this invention, the disease target information used for supervised learning is also open-source, capable of integrating multiple sources of biomedical evidence, including genomic association studies, pathogenic mutation information, gene function perturbation experiments, transcriptomic differential analysis, and epigenetic regulatory evidence. As long as the relevant features can be mapped to the protein identifiers corresponding to network nodes, they can be used as continuous labels for training, enabling the model to adapt to different disease types, different phenotypes, and evidence systems from different sources.

[0017] In a preferred embodiment, the node label data described in S1 needs to be preprocessed before model training, specifically including: firstly, standardizing all protein identifiers (e.g., uniformly converting them to Entrez ID) to ensure consistency of the identification system between different data sources; then, checking the completeness and mappability of the label values, deleting protein nodes with missing labels or those that cannot correspond to the node index; and if necessary, normalizing the label values ​​to improve the numerical stability of the training process and the convergence efficiency of the model.

[0018] Dataset partitioning: The data was divided into training, validation and test sets in an 8:1:1 ratio using stratified sampling.

[0019] In a preferred embodiment, the graph attention neural network model described in S4 employs a deep graph structure model consisting of two layers of GATConv, and uses a multi-head attention mechanism to weighted aggregate the features of neighboring protein nodes. Its single-head attention weights... It can be calculated using the following formula:

[0020]

[0021] in, Represents a node The input feature vector, Represents the transformed nodes With nodes The concatenated vector of features is used to calculate the attention correlation between two nodes. Represents a node The set of neighboring nodes, that is, the set of nodes with respect to the node There exist all nodes connected in the graph structure.

[0022] The first layer of the model, GATConv, uses four attention heads to extract neighborhood topological features from different perspectives. The second layer, GATConv, uses a single-head attention and outputs a non-spliced ​​one-dimensional continuous value to achieve regression prediction of node risk scores. The model internally employs the ReLU activation function and adds Dropout between layers to enhance generalization ability and suppress overfitting.

[0023] In supervised learning, mean squared error is used as the loss function to measure the deviation between the predicted value and the true label. In this invention, the AdamW optimizer is used to update all trainable parameters, with a learning rate set to 0.01 and a weight decay coefficient set to 5 × 10⁻⁶. -4To balance training efficiency and model complexity, the maximum number of training epochs is set to 200. After each epoch, multiple metrics, including the coefficient of determination, mean squared error, root mean square error, and mean absolute error, are calculated on the validation set to dynamically evaluate model performance. When the validation set coefficient of determination is better than the previous best record, the current model parameters are saved. If the validation set coefficient of determination does not improve within 30 consecutive epochs, an early stopping mechanism is triggered, and training is terminated, effectively avoiding overfitting and optimizing the overall training process.

[0024] In a preferred embodiment, the set of protein nodes representing high-activity disease regions constructed in step S6 has a node count that is exactly the same as the number of drug targets. This design effectively avoids dimensional shifts or exaggerated coverage effects caused by the mismatch in scale between the two, ensuring that the disease space and the drug action space maintain a unified scale in subsequent analyses. Compared to the traditional method of selecting disease nodes with a fixed threshold or fixed proportion, this method uses the number of targets as a reference, resulting in higher comparability, interpretability, and cross-drug applicability. It is suitable for traditional Chinese medicines with multiple targets and Western medicines with fewer targets. This high-activity region set represents the core region with the highest activity in the disease network and can serve as a structural benchmark for evaluating the coverage ability of drug targets on key functional modules of the disease.

[0025] Specifically, this invention uses the continuous disease risk score output by the graph attention neural network model to quantify each node. The disease relevance is denoted as The drug target set is The number of its elements is Nodes are sorted from highest to lowest score, and the highest-scoring node is selected. The nodes form a set of disease high-fever zones. Its definition is as follows:

[0026]

[0027] in, Top ∣T∣ ( f () indicates ranking based on disease relevance. The nodes obtained. It maintains consistency with the number of drug targets while representing the most biologically active regions in the disease network, providing a structurally consistent control space for subsequent construction of drug-disease network intervention maps.

[0028] In a preferred embodiment, step S7 aims to establish a weighted topological mapping between drug targets and protein nodes in high-intensity disease regions to systematically describe the potential regulatory range and intensity of the drug on the core disease region. The shortest path reflects the propagation cost of biological signals in the protein-protein interaction network; therefore, this step focuses on each drug target... As the source node, calculate the distances from it to each node in the high-fever zone of the disease. shortest path distance This invention measures the accessibility and connectivity between drug targets and the pathological core from a topological perspective. To transform topological distance into a continuously quantifiable regulatory intensity, this invention introduces an exponential decay function to convert the distance... The mapping to the control intensity weight is as follows:

[0029]

[0030] in This is the distance attenuation coefficient. Target To the target point The shortest path length is adjusted by... In conjunction with the allowed shortest path length, different drug action modes are distinguished within a unified framework: when the shortest path length is set to a large value or... When the target is smaller, it can effectively affect disease nodes at greater distances, making it suitable for traditional Chinese medicine formulas that emphasize broad-area, multi-pathway regulation; conversely, when the shortest path length is smaller or... When the target is large, the model's regulatory effect will be more concentrated on the local structural modules around the target, which can better reflect the characteristics of precise targeting and local perturbation of Western medicine.

[0031] More preferably, after obtaining the regulatory weights of all targets and disease hotspot nodes, a target-disease weighted matrix is ​​constructed. This is used to systematically quantify the distribution of drug target sets in high-intensity disease areas. Furthermore, the coverage function is defined as follows:

[0032]

[0033] in, For drug target set, This function represents the disease relevance score for nodes in high-risk disease regions. It comprehensively reflects the maximum topological coverage capability of drugs in the core disease region, and can simultaneously integrate disease relevance and drug regulatory intensity.

[0034] In a preferred embodiment, step S8 aims to further reduce structural redundancy in the drug target set and construct a non-redundant effective target subset that can represent the regulatory characteristics of the drug system. Since some targets have highly similar neighborhood structures in the protein-protein interaction network, their functions and topological positions often highly overlap. Directly including them in the perturbation map would lead to repetitive regulatory information and reduce the model's ability to distinguish the true regulatory range of the drug. Therefore, this step, based on the protein-protein interaction network, determines a set of network nodes with a shortest path length of no more than 3 for each drug target, and calculates the Jaccard similarity between any two drug targets based on the degree of overlap between these node sets.

[0035]

[0036] in and Target points and The set of neighboring nodes. This similarity measure can simultaneously reflect the degree of functional overlap and the similarity of local network structure, and is an important indicator for measuring drug redundancy.

[0037] After obtaining the topological similarity between target points, to avoid duplicate selection of highly similar targets, the present invention further employs a stepwise selection strategy based on greedy search to construct a diverse and representative set of effective target points. This strategy first initializes the selected target point set as follows: Subsequently, in each iteration, targets that maximize the gain of the objective function are selected from the candidate targets that have not yet been selected. To simultaneously consider the coverage of the core disease region and the structural diversity of the target set, a redundancy penalty term is introduced into the objective function, in the following form:

[0038]

[0039] in, Indicate target point For the current target set The resulting marginal coverage gain; For this target point The maximum topological similarity with any single target in the selected target set; The redundancy penalty coefficient is used to adjust the suppression weight of structural repetition in the objective function. Regarding the selection of the similarity metric range, when a larger shortest path length is used, the similarity can reflect the degree of structural overlap of the target at a larger topological scale, which is beneficial for identifying core regulatory nodes with global action patterns. Conversely, when the shortest path length is smaller, the calculation results are more sensitive to local microstructural differences and are more suitable for measuring the fine-grained regulatory characteristics of monoclonal drugs or precisely targeted drugs in local modules. Meanwhile, the redundancy penalty parameter... The magnitude of the perturbation directly determines the bias of the final perturbation pattern: a larger value indicates a bias in the pattern. It places greater emphasis on structural diversity and tends to obtain a global target set with broad-area control capabilities; smaller This allows for the continuous inclusion of targets with high local relevance, thereby highlighting the perturbation pattern of the drug in the local action module.

[0040] In the greedy selection process described above, only when a certain target point satisfies... Only when a target is selected is it considered to have a positive coverage contribution to the existing target set and thus included in the final effective target set. Through this screening mechanism, the model can effectively reduce the redundancy between targets while improving the disease network coverage, thereby obtaining a representative target set with broad coverage and strong structural complementarity. This set constitutes the interference map of the drug on the disease network, which is the basis for subsequent efficacy evaluation and mechanism analysis.

[0041] In a preferred embodiment, the method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology according to the present invention further includes:

[0042] S9: Based on the drug perturbation map obtained in S8, it characterizes the coverage and intensity of drug effects on disease-related biological functional modules, assesses the overall regulatory characteristics of drugs at the disease network level, and provides a basis for the multi-pathway synergistic mechanism of drugs and their ability to intervene in disease states.

[0043] The disease-related biological functional modules described in S9 include, for example, the immune, metabolic, and inflammatory modules associated with tumors, and the inflammatory, immune, and bone function modules associated with rheumatoid arthritis.

[0044] To verify the advancement and practical value of the disease risk scoring method based on graph attention neural networks and its drug perturbation map construction strategy proposed in this invention, a control experiment was constructed on the same hardware platform (GPU RTX 5070 Ti 16GB ×1, CPU Intel® Ultra 7 processor 265K @ 5.50GHz, memory 96GB, storage 1TB). All control models were implemented using PyTorch and Scikit-learn to rigorously verify the performance advantages of the graph attention neural network model in this invention in disease correlation prediction and network structure representation capabilities.

[0045] This invention also proposes a system that applies the above-mentioned disease risk prediction and drug perturbation spectrum construction method, the system comprising the following functional modules:

[0046] (1) Disease target evidence integration module

[0047] This module is used to acquire, integrate, and continuously update multi-source disease-related evidence, including genome association studies, disease expression profile databases, experimentally validated pathogenic gene information, literature co-occurrence evidence, and network coupling scores. It is responsible for evidence standardization, missing data completion, and continuous labeling, providing the model with unified and reliable supervisory labels.

[0048] (2) Protein-protein interaction network construction module

[0049] This module loads public databases such as STRING, IntAct, HIPPIE, and László Barabási, or lab-built interaction data, and performs identifier normalization, redundancy removal, bidirectional edge construction, and network sparsification. It outputs node indices, feature matrices, and edge lists that meet the training requirements of graph attention neural networks.

[0050] (3) Disease Risk Map Modeling Module

[0051] Corresponding to steps S4–S6 of this invention, the process includes the construction, parameter setting, training, verification, and testing of a graph attention network, used to learn the continuous disease risk score of each node in the protein interaction network.

[0052] (4) Disease high fever zone identification module

[0053] According to step S7, with the number of drug targets as a reference, high-scoring nodes are selected from the risk score sequence output by the graph attention neural network to construct a set of high-risk disease zones with a uniform scale.

[0054] (5) Drug-Disease Topology Mapping Module

[0055] The shortest path calculation, exponential decay weighting, target-disease weighted matrix construction, and adjustable parameterization of the topology propagation space are used to perform step S8.

[0056] (6) Redundancy suppression and perturbation map generation module

[0057] Corresponding to step S9, the target neighborhood similarity calculation, the penalized greedy selection strategy, and the final construction of the interference map are performed, outputting the non-redundant key target set of the drug in the disease network.

[0058] The beneficial technical effects of this invention are reflected in the following aspects:

[0059] First, this invention introduces graph attention mechanisms into disease network risk prediction tasks. By directly learning the end-to-end graph structure on the protein-protein interaction network, it can effectively capture potential topological dependencies between genes and proteins, thus possessing stronger structural sensitivity and biological interpretability in the identification of disease driving factors. Compared to the limitations of traditional machine learning methods that rely on manual features and struggle to utilize network structure information, the model of this invention can automatically learn high-order dependencies between nodes, significantly improving prediction accuracy and stability.

[0060] Secondly, the disease high-intensity zone selection strategy based on drug target number constraints proposed in this invention effectively avoids the dimensional inconsistency problem caused by differences in target size among different drugs, and realizes a unified and comparable framework across drugs, prescriptions, and target sizes. Whether it is a Western medicine with few targets or a traditional Chinese medicine with many targets, this method can ensure the alignment of the disease space and the drug action space, laying a strict and fair comparison standard for subsequent perturbation analysis.

[0061] Secondly, by introducing a shortest path-based topological propagation mechanism and an exponential decay function, this invention enables drug-disease interaction modeling to move beyond direct target overlap and systematically measure potential regulatory pathways cascading through multiple shortest path lengths. This more closely reflects the properties of signal transduction, metabolic regulation, and network responses in actual biology. Furthermore, by adjusting the shortest path length and decay coefficient, the model can flexibly distinguish between the multi-target, broad-area action patterns of traditional Chinese medicine and the localized, precise regulatory patterns of Western medicine, significantly improving its adaptability and interpretability.

[0062] Furthermore, this invention introduces a Jaccard metric for target topological similarity and a redundancy penalty mechanism in the construction of the perturbation map. This automatically eliminates redundant targets with highly overlapping functions and structures, thereby retaining a set of effective targets with complementary structures and broad coverage. This improves the representativeness of the analysis results and enhances the analytical capability of the drug action space. Combined with a coverage function, this strategy generates a more compact and informative drug interference map, providing a solid basis for efficacy evaluation, mechanism inference, and combination drug design.

[0063] The drug perturbation map constructed in this invention is used to characterize the comprehensive intervention capabilities of drugs on multiple biological functional modules of a disease. The constructed drug perturbation map can reflect the coverage and intensity of drug effects on multiple biological functional modules of a disease. This allows for a systematic assessment of the overall regulatory characteristics of drugs at the disease network level, providing a network-level basis for understanding the multi-pathway synergistic mechanisms of drugs and their comprehensive intervention capabilities on complex disease states.

[0064] In summary, this invention organically combines graph neural networks, topological propagation, redundancy suppression, and a unified-scale strategy for constructing high-prevalence disease zones, forming a generalized disease network perturbation graph analysis framework applicable to both traditional Chinese and Western medicine. This framework possesses high accuracy, high interpretability, and high clinical application potential. Attached Figure Description

[0065] Figure 1 Flowchart of a method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology. Detailed Implementation

[0066] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result of the description. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of protection defined by the claims of the present invention.

[0067] The terms and meanings involved in this invention are as follows:

[0068] Protein-protein pairs: consist of two interacting proteins, each with a unique biological identifier.

[0069] Network protein node set: A protein dataset consisting of all proteins appearing in all protein interaction pairs collected by the network.

[0070] Protein Node Index Mapping Table: An index mapping table between protein nodes and said integers, constructed by assigning a unique integer index to each protein in the set of protein nodes in the network.

[0071] Edge index: Represents the connection between two protein nodes in a protein interaction network.

[0072] Undirected edge set of proteins: The resulting edge indices are undirected so that each pair of protein interactions is represented by a bidirectional edge, representing the relationship between proteins.

[0073] Standardized human protein interaction graph dataset: consists of a set of network protein nodes and a set of undirected edges between proteins.

[0074] Node feature vector: used to represent the relationship between the protein node and the disease.

[0075] Protein node feature matrix: Composed of node feature zero vector, node feature vector and label value, representing the relationship between protein and disease.

[0076] Neighborhood topology features: The topology features of other protein nodes that are connected to a protein node in a graph structure are also called neighborhood node features.

[0077] Example 1: Implementation of a disease node construction and risk scoring model using rheumatoid arthritis as an example.

[0078] Rheumatoid arthritis (RA) is a chronic, systemic autoimmune disease characterized by insidious onset, rapid progression, and high disability rates, requiring long-term treatment and continuous management. This disease not only causes significant joint pain and functional impairment but also imposes a severe economic and psychological burden on patients. Based on these disease characteristics, this embodiment constructs a network risk scoring model for RA-related nodes, providing a standardized disease characterization for subsequent drug perturbation spectrum identification.

[0079] 1. Collection and characterization of disease-related targets for rheumatoid arthritis

[0080] Figure 1 A flowchart of the method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology is provided. This embodiment uses rheumatoid arthritis as an example to illustrate a specific implementation of the invention. It should be noted that, based on the inventive concept of the invention and the above specific implementation, risk scoring models with other diseases as nodes can also be constructed.

[0081] Rheumatoid arthritis-related genes and the strength of evidence were collected from the GeneCards database. This database integrates multi-source disease association evidence, including gene expression changes, pathway enrichment, functional annotation, and literature co-occurrence frequency, providing a unified "gene-disease association score" for each gene. This score reflects the overall support of the gene in multi-level biological evidence; a higher score indicates a stronger association between the gene and rheumatoid arthritis. A total of 4,495 rheumatoid arthritis-related targets were obtained from GeneCards.

[0082] Genetic variation information was obtained from the GWAS Catalog database. This database systematically includes single nucleotide polymorphisms (SNPs) associated with rheumatoid arthritis (RA) identified in large-scale population studies and their corresponding genes. By extracting relevant genetic variations and their dominance ratios, the pathogenic risk of genes at the genetic level can be quantified; a higher dominance ratio indicates a more significant genetic contribution. A total of 740 genetic risk genes associated with RA were identified.

[0083] To simultaneously reflect the disease association strength and genetic susceptibility of genes, this embodiment constructs a comprehensive eigenvalue, combining the rheumatoid arthritis association score from GeneCards with the genetic risk advantage ratio provided by the GWAS Catalog. Specifically, let the gene... The GeneCards disease association score is Its genetic risk odds ratio in the GWAS Catalog is The comprehensive features are defined as follows:

[0084]

[0085] in, This represents the comprehensive disease risk characteristic value of a gene. For genes lacking a genetic variation advantage ratio, then... = 1, this comprehensive feature can highlight genes that have both multi-omics evidence and significant genetic risk, and improve the ability of subsequent graph attention neural networks to identify key disease drivers.

[0086] 2. Human protein-protein interaction network data collection

[0087] Protein-protein interaction data were obtained from four public databases: STRING, IntAct, HIPPIE, and a dataset released by Albert-László Barabási's team. To ensure data consistency, this embodiment deredundancy was performed on the interaction data in the above databases, and all protein identifiers were unified as Entrez IDs.

[0088] After preprocessing, five independent protein-protein interaction networks were constructed, each with proteins as nodes and interactions as edges. The STRING network contains 16,719 nodes and 715,269 edges; IntAct contains 16,135 nodes and 798,998 edges; HIPPIE contains 18,449 nodes and 810,973 edges; and the Albert-László Barabási dataset contains 17,934 nodes and 327,924 edges. All five interaction networks can serve as the structural basis for subsequent graph attention neural network modeling, enabling the representation of topological relationships between proteins from different data source dimensions.

[0089] 3. Construction of a rheumatoid arthritis risk gene scoring model based on graph attention neural network

[0090] In the model building phase, the aforementioned 4,495 disease-related genes and their comprehensive characteristics were first analyzed. The dataset is randomly divided into training, validation, and test sets in an 8:1:1 ratio to ensure consistent distribution across different subsets. A graph attention network model consisting of two layers of GATConv is then constructed, employing a multi-head attention mechanism to weighted aggregate neighborhood features, enhancing the model's ability to capture different topological relationships. The single-head attention weights are calculated as follows:

[0091]

[0092] in, Represents a node The input feature vector, Represents the transformed nodes With nodes The concatenated vector of features is used to calculate the attention correlation between two nodes. Represents a node The model uses a set of neighboring nodes. The first layer employs four attention heads to extract multi-view local topological features; the second layer uses a single-head attention and outputs a non-concatenated one-dimensional continuous value to complete the risk score regression. ReLU is used as the activation function, and a Dropout of 0.1 is set to improve generalization ability.

[0093] During training, mean squared error was used as the loss function, and the AdamW optimizer was employed (learning rate 0.01, weight decay coefficient 5×10⁻⁶). -4 The parameters are updated. The maximum number of training epochs is set to 200, and after each epoch, the coefficient of determination and other metrics are calculated on the validation set. The model is saved when the validation set performance reaches a new optimum; if there is no improvement after 30 consecutive epochs, an early stopping strategy is triggered to avoid overfitting.

[0094] To systematically evaluate the impact of different protein interaction databases on model performance, this embodiment employs five independent training pipelines to model the STRING, IntAct, HIPPIE, and Albert-László Barabási datasets respectively. Each pipeline maintains consistent node features, with only the interaction network structure changing depending on the data source. Each pipeline is run independently 10 times to obtain stable node scoring results and reduce random errors introduced during initialization. The training and scoring results are shown in Table 1.

[0095] Table 1. Performance comparison of graph attention neural network models trained based on different protein interaction networks

[0096]

[0097] The results show that the graph attention neural network constructed in this invention exhibits good disease risk prediction capabilities on different protein interaction databases, but its performance still varies under different network backgrounds. A comprehensive analysis of multiple metrics on both the test and validation sets reveals that the Albert-László Barabási network performs best overall: its validation set coefficient of determination (0.902 ± 0.097) is not only the highest among all networks, but its test set performance also remains stable (0.893 ± 0.077), with low mean squared error and mean absolute error. This indicates that the interaction topology provided by this network is most conducive to the graph attention neural network model capturing key node features related to disease drivers. In contrast, although the HIPPIE network achieves the highest mean coefficient of determination (0.904 ± 0.079) on the test set, its validation set performance is relatively volatile, indicating that while it has high coverage, there is still some heterogeneity in its local topology, making its generalization ability slightly lower than that of the Albert-László Barabási network. The STRING network exhibits stable performance, achieving high prediction accuracy on both the test and validation sets, indicating a good network structure quality and providing reliable neighborhood dependencies for the attention mechanism. Conversely, the IntAct network shows the most significant fluctuations on the validation set (coefficient of determination –2.604 ± 6.564), primarily due to the presence of numerous low-confidence edges in its interaction data, resulting in substantial local topological noise and impacting model convergence stability. However, its performance on the test set remains acceptable (coefficient of determination 0.773 ± 0.108), demonstrating the robustness of the proposed method even with noisy networks.

[0098] Overall, considering both the generalization performance on the validation set and the stability on the test set, the Albert-László Barabási network provided the best training support among the five types of interacting networks, enabling the graph attention neural network model of this invention to achieve the highest overall prediction performance. Therefore, the subsequent construction of high-fever zones, target coverage analysis, and perturbation map inference in this invention are all based on the training results of the Albert-László Barabási network to ensure the optimal performance of the analysis framework in terms of structural reliability and prediction accuracy.

[0099] The comparative results of the four types of networks show that the graph attention neural network model designed in this invention has outstanding cross-database generalization ability. The learned disease topology representation can maintain high prediction accuracy on different interaction networks, indicating that the extracted structural features have good universality. Meanwhile, the multi-pipeline training strategy effectively reduces the variance caused by random initialization, enabling the model to converge stably in ten independent runs and achieve good performance results in terms of both mean and standard deviation. Furthermore, the trend that higher network structure quality leads to better model performance is clearly demonstrated. High-confidence, multi-dimensional integrated protein interaction networks are more conducive to improving the accuracy of attention aggregation, enabling the model to more accurately identify the risk level of key disease nodes. Therefore, the disease risk prediction framework based on graph attention neural networks in this invention not only has high accuracy but also exhibits high robustness under different topological backgrounds, providing a reliable and consistent foundation for the subsequent construction of drug-disease perturbation maps.

[0100] Example 2: Perturbation map mining of Guizhi Shaoyao Zhimu Decoction in the treatment of rheumatoid arthritis

[0101] Guizhi Shaoyao Zhimu Decoction, originating from the *Synopsis of Prescriptions of the Golden Chamber*, is a classic formula for treating arthralgia and other joint diseases such as weakness and paralysis. Composed of cinnamon twig, peony root, anemarrhena rhizome, licorice root, ephedra, ginger, and atractylodes macrocephala, it has been shown to possess significant anti-inflammatory, anti-immune abnormality, and joint-protective effects. It effectively improves symptoms such as joint swelling, pain, and limited mobility, and in a rheumatoid arthritis model, it demonstrates multi-level regulatory capabilities on immune inflammatory pathways, cytokine networks, and abnormal synovial cell activation.

[0102] 1. Collection and standardization of target points for Guizhi Shaoyao Zhimu Decoction

[0103] The target targets of Guizhi Shaoyao Zhimu Decoction were derived from the Hit2.0 database. This database integrates literature and experimental evidence results with cross-annotation from multiple databases, systematically reflecting the potential targets of traditional Chinese medicine components. In this embodiment, target targets were searched for all medicinal materials in the formula, and the obtained targets were deredundant and their identifiers were standardized. All targets were uniformly converted into Entrez IDs to ensure consistency with the disease network and protein interaction network structure in Example 1. Finally, 842 candidate targets related to Guizhi Shaoyao Zhimu Decoction were obtained.

[0104] 2. Guizhi Shaoyao Zhimu Decoction as an intervention in rheumatoid arthritis: drug-disease network construction

[0105] To systematically evaluate the network perturbation effect of Guizhi Shaoyao Zhimu Decoction on rheumatoid arthritis, this embodiment first uses the node continuous risk score generated by the graph attention model constructed in Example 1 as a disease relevance indicator to quantify the pathological contribution of each node in the protein interaction network. Subsequently, based on the 842 standardized target points obtained from Guizhi Shaoyao Zhimu Decoction, the number of nodes in the high-fever zone of the disease is set to be consistent with the number of drug target points to avoid weighting bias caused by differences in the number of nodes, ensuring comparability and uniform scale between the disease space and the drug action space in subsequent coverage analysis.

[0106] The specific construction process is as follows: All nodes in the protein-protein interaction network are sorted from high to low according to their disease relevance scores, and the nodes with the highest scores are selected. The nodes form a set of disease high-fever zones. This is used to characterize the core regions in a disease network that have the highest overall activity and are most critical in driving pathological processes. Its formula is as follows:

[0107]

[0108] in, Top ∣T∣ ( f () indicates ranking based on disease relevance. The node.

[0109] The resulting set of high-intensity disease regions maintains consistency between the number of nodes (186) and the number of drug targets (186), while also reflecting the core active areas of the disease network. This lays a structural foundation for subsequent target coverage analysis and perturbation map construction.

[0110] 3. Analysis of the perturbation map of rheumatoid arthritis treated with Guizhi Shaoyao Zhimu Decoction

[0111] To systematically quantify the pharmacological perturbation effect of Guizhi Shaoyao Zhimu Decoction on rheumatoid arthritis, a set of genes in high-temperature regions was first selected from the whole genome network based on disease prediction scores. And calculate each target in the drug. With each gene in the high-temperature zone shortest path distance To improve the biological interpretability of topological propagation, the shortest path length is set to < 3 (i.e., ...). Based on this, an exponential decay function is introduced to map the topological distance to continuous control intensity:

[0112]

[0113] The distance attenuation coefficient is set to... = 1.0, = 0; if the shortest path between two nodes exceeds 3, then define Based on this, a target-disease weighted matrix can be constructed. This is used to quantify the potential regulatory effects of each target on genes in high-temperature regions of the disease.

[0114] Meanwhile, in order to evaluate the comprehensive regulatory ability of the target set of Guizhi Shaoyao Zhimu Decoction on various biological functional modules of rheumatoid arthritis, this embodiment further defines a coverage function:

[0115]

[0116] in, For drug target set, This is the disease relevance score for genes in high-temperature regions. This function measures the maximum topological coverage of the target set in the core disease region. Based on this, to avoid highly overlapping regulatory ranges between drug targets, this embodiment uses a protein-protein interaction network to determine the set of network nodes for each drug target whose shortest path length is no greater than 3, and calculates the Jaccard similarity between any two targets:

[0117]

[0118] in and Target points and The set of neighboring nodes with a shortest path length of 3 is used to measure the similarity of their local network structures.

[0119] To simultaneously optimize disease coverage and target diversity, this embodiment employs a greedy selection strategy with redundancy penalties. In each iteration, targets that maximize the following objective function are selected from the unselected candidate targets:

[0120]

[0121] in, Target For the current target set The resulting marginal coverage gain This is the similarity between the target and the target with the highest similarity in the selected target set. In this embodiment, the redundancy penalty coefficient is set to... .

[0122] Through continuous iterative selection And update the set This can generate a sequence of targets sorted by marginal contribution. Only if a target satisfies... Only when a target meets the screening criteria is it considered to have a positive contribution to the existing coverage and be included in the drug perturbation spectrum. Ultimately, a total of 186 targets met the screening criteria and were included in the network perturbation spectrum of Guizhi Shaoyao Zhimu Decoction for the intervention of rheumatoid arthritis (Table 2), constituting the key regulatory structure of this formula in the major biological modules of the disease.

[0123] Table 2. Network perturbation spectrum of Guizhi Shaoyao Zhimu Decoction in intervention of rheumatoid arthritis

[0124]

[0125]

[0126]

[0127] KEGG enrichment analysis (adj.p < 0.05) was performed on the 186 effective perturbation targets of Guizhi Shaoyao Zhimu Decoction obtained in the final screening, which systematically revealed the main intervention pathways of this formula in rheumatoid arthritis. The results showed that the significantly enriched pathways could be classified into six core biological functional modules: immune regulation, signal transduction, lipid metabolism, cell function regulation, nervous system function, and endocrine regulation, among which immune and inflammation-related pathways were dominant (Table 3).

[0128] In terms of immune regulation, the enriched pathways of rheumatoid arthritis, cytokine-receptor interactions, JAK-STAT signaling pathway, and multiple inflammation-related pathways are highly consistent with previously reported basic and clinical studies of Guizhi Shaoyao Zhimu Decoction. Clinical and animal experiments have shown that Guizhi Shaoyao Zhimu Decoction can correct the immune imbalance in rheumatoid arthritis patients and alleviate joint inflammation by regulating the Th17 / Treg cell balance and the JAK2 / STAT3 signaling pathway, downregulating the expression of pro-inflammatory factors such as IL-17, IL-22, IL-6, and TNF-α, and upregulating anti-inflammatory factors such as TGF-β and IL-10. Furthermore, in a collagen-induced arthritis (CIA) model, studies have demonstrated that Guizhi Shaoyao Zhimu Decoction can regulate the cGAS-STING-NF-κB axis and related cytokines, significantly reducing serum IL-6 levels and alleviating synovial inflammation. This corroborates the significant enrichment of NF-κB-related pathways observed in this invention. In other inflammatory joint disease models (such as gouty arthritis), this formula has also been shown to downregulate TLR2 / 4-MyD88 and NF-κB pathway-related molecules, and inhibit the release of various inflammatory mediators. This suggests that the broad-spectrum regulatory effect of Guizhi Shaoyao Zhimu Decoction on the innate immune receptor-NF-κB axis has certain commonalities, providing evidence to support the Toll-like receptor, NOD-like receptor, and multiple inflammatory pathways enriched in this invention. Regarding functional modules related to bone destruction and bone protection, the enrichment of pathways such as "osteoclast differentiation," "necrotizing apoptosis," and "apoptosis" obtained in this invention is highly consistent with previous studies on the prevention and treatment of bone destruction in rheumatoid arthritis using Guizhi Shaoyao Zhimu Decoction. Existing animal experiments have shown that this formula can significantly inhibit rheumatoid arthritis-related bone destruction. Its mechanism is closely related to the regulation of the NK / NKL / OPG axis of rheumatoid arthritis and the reduction of osteoclast formation and activity. In CIA rats, Guizhi Shaoyao Zhimu Decoction upregulates OPG and downregulates the expression of NK cells in rheumatoid arthritis, thus exerting a bone-protective effect. This is in good agreement with the "osteoclast differentiation" and inflammation-bone destruction-related signaling pathways enriched in this invention.

[0129] The enrichment of lipid metabolism (such as arachidonic acid metabolism, fat absorption, and adipokines signaling pathways), endocrine regulation (prolactin, relaxin, estrogen, insulin, thyroid hormones, GnRH, etc.), and nervous system functional modules (neurotrophic factors, dopamine synapses, cholinergic synapses, etc.) further demonstrates that Guizhi Shaoyao Zhimu Decoction possesses broad-spectrum systemic intervention capabilities. Rheumatoid arthritis inherently exhibits pathological characteristics involving multiple biological couplings of immunity, metabolism, neuroendocrine systems. The aforementioned enrichment results suggest that the effective perturbation of this formula is not limited to inflammatory targets but also deeply involves the body's energy metabolism, neuroimmune regulation, and endocrine balance, providing molecular mechanism support for its overall regulatory effect. Furthermore, studies from the perspective of articular cartilage and synovial angiogenesis suggest that Guizhi Shaoyao Zhimu Decoction can improve cartilage damage in rheumatoid arthritis and inhibit pathological pannus formation through the JAK2 / STAT3 / VEGF pathway. This result is consistent with the enrichment of the JAK-STAT pathway, cytokine-receptor interaction, and some angiogenesis-related pathways in the KEGG of this invention, indicating that the key targets obtained by the present invention based on network perturbation spectrum screening do indeed cover the classic immune-inflammatory-bone destruction pathway of Guizhi Shaoyao Zhimu Decoction in rheumatoid arthritis, which has been reported in the literature. This confirms the biological rationality and clinical relevance of the perturbation map mining method at the pathway level.

[0130] Table 3. Biological functional modules of Guizhi Shaoyao Zhimu Decoction in intervention of rheumatoid arthritis

[0131]

[0132]

[0133] .

Claims

1. A method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein-protein interaction network topology, characterized in that, The method includes the following steps: S1: Collect human protein-protein interaction network data, which consists of several protein-protein interaction pairs. Each protein-protein interaction pair includes two interacting proteins, and each protein has a unique biological identifier. Summarize and remove duplicate protein identifiers from all interaction pairs to obtain a set of network protein nodes. The set of network protein nodes consists of proteins that have participated in at least one protein-protein interaction. Subsequently, assign a unique integer index to each protein in the set of network protein nodes to construct a protein node index mapping table for subsequent graph structure modeling and calculation. S2: Map the protein interaction pairs obtained in S1 to edge indices in the graph structure. The edge indices are used to represent the connection relationship between two protein nodes in the protein interaction network. The obtained edge indices are then undirected so that each pair of protein interactions is represented in the form of bidirectional edges. This results in a standardized human protein interaction graph dataset consisting of a set of network protein nodes and a set of undirected protein edges, which serves as the structural input data for graph neural network modeling. S3: Initialize node feature vectors for all protein nodes in the human protein-protein interaction network node set obtained in S2 to represent the relationship between the protein nodes and the disease to be analyzed. By default, the initial node feature vector of each protein node is set to zero vector, which is used to indicate that the node does not have disease association attributes in the initial state. Based on this, for protein nodes with known disease association, assign corresponding label values ​​with disease association strength according to the disease association evidence, and use the node feature vector and label values ​​as supervised learning data to participate in the model training constructed in S4; for protein nodes that do not contain disease label values, their feature vectors remain unchanged as zero vectors. The protein node feature matrix is ​​constructed in the above manner. S4: Based on the set of undirected edges of proteins obtained in S2 and the node feature matrix obtained in S3, a deep graph structure regression model based on graph attention neural network is constructed. The model is composed of multiple layers of graph attention convolutional layers stacked sequentially. The model uses a multi-head attention mechanism to weight and aggregate the node features associated with the disease according to the interaction relationship between proteins. Each graph attention convolutional layer iterates on the weighted and aggregated node features under the constraint of the protein interaction network structure, so that the node features associated with the disease of proteins in the protein interaction graph dataset are continuously learned and updated, forming a deep representation of the protein node features, and obtaining the optimal deep graph structure regression model. S5: Based on the optimal depth map structure regression model obtained in S4, under the premise of keeping the parameters unchanged, the regression model is used to infer the degree of association between proteins with zero vector node features and the disease to be analyzed and assign them label values. At the same time, the node feature vectors and label values ​​of protein nodes that already have label values ​​are updated. Through the above calculation, the model finally obtains the continuous regression prediction value of each protein node in the protein interaction network and the disease through the output layer. The label of the continuous regression prediction value is in the form of continuous real numbers, which is used to represent the association strength between the protein node and the target disease. The higher the value, the higher the potential association degree and network regulation importance of the protein in the disease network. S6: Based on the protein node-disease association score obtained in step S5, sort all protein node features in the protein interaction network from high to low score; simultaneously obtain or directly input the set of protein nodes targeted by the drug to be analyzed from a public database, and count the number of protein nodes targeted by the drug to be analyzed; according to the sorting results, select a set of disease targets with scores from high to low and a number consistent with the number of protein targets of the drug, to form a set of high-temperature protein nodes in the disease high-temperature zone used to characterize the drug to be analyzed, the set of high-temperature protein nodes in the disease high-temperature zone is composed of the protein with the highest association score in the protein network related to the disease to be analyzed; S7: Based on the set of protein nodes in the high-intensity disease region obtained in S6, a drug-protein-disease interaction network is constructed. Taking the protein target of each drug in the protein interaction network as the source node, the shortest path distance between the drug and each protein node associated with the disease in the drug-protein-disease interaction network is calculated. The spatial scale of topology propagation is adjusted by setting the shortest path length and the distance decay coefficient. On this basis, the shortest path distance is mapped to a continuous weighted value using an exponential decay function. A drug protein target – disease high-intensity disease protein node target – disease weighted matrix is ​​constructed to quantitatively evaluate the potential regulatory strength of the drug on the disease high-intensity disease protein nodes at different topological granularities and to calculate the drug coverage function. S8: Calculate the topological similarity between any two protein targets based on the protein interaction network structure. Introduce a stepwise greedy selection strategy with a redundancy penalty term. Adjust the repulsion strength of similar protein targets by adjusting the penalty coefficient. Classify protein targets that generate positive marginal coverage gain in the set of protein nodes in the high-temperature zone of the disease as effective contributing protein targets and include them in the non-redundant target set. The set is used to construct drug perturbation maps at different topological granularities.

2. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, S1 includes a preprocessing step for the protein identifiers before model training: first, all protein identifiers are standardized, then the identifiers are checked for completeness and mappability, missing identifiers or protein nodes that cannot be matched with node indices are deleted, and stratified sampling is used to divide the node data with identifiers into training set, validation set and test set in a ratio of 8:1:

1.

3. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, The depth graph structure regression model described in S4 employs a two-layer GATConv depth graph structure model and uses a multi-head attention mechanism to weighted aggregate the protein node features in the neighborhood. Its single-head attention weights... Calculate using the following formula: ， in, Represents a node The input feature vector, Represents the transformed nodes With nodes The concatenated vector of features is used to calculate the attention correlation between two nodes. Represents a node The set of neighboring nodes, that is, the set of nodes with respect to the node There exist all nodes connected in the graph structure; The first layer of the model, GATConv, sets up four single attention heads to extract the topological features of the neighborhood based on protein and disease association features and network protein structure features. The second layer, GATConv, uses single-head attention to output non-spliced ​​one-dimensional continuous values ​​to achieve regression prediction of node risk scores. The model uses the ReLU activation function and adds Dropout between layers to enhance the model's generalization ability and suppress overfitting.

4. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 3, is characterized in that... In the training described in S4, mean squared error is used as the loss function to measure the deviation between the predicted value and the true label. The AdamW optimizer is used to update all trainable parameters, with the learning rate set to 0.01 and the weight decay coefficient set to 5 × 10⁻⁶. -4 The maximum number of training epochs for the model is set to 200. After each epoch, the coefficient of determination, mean squared error, root mean square error, and mean absolute error are calculated on the validation set to dynamically evaluate the model performance. When the coefficient of determination on the validation set is better than the previous best record, the current model parameters are saved. When the coefficient of determination on the validation set does not improve within 30 consecutive epochs, the early stopping mechanism is triggered and training is terminated, thereby effectively avoiding overfitting and optimizing the overall training process.

5. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, The construction of the disease high-risk zone set described in S6 involves using the continuous disease risk score output by the depth map structure regression model to quantify each node. The disease relevance is denoted as The set of protein targets for the drug is The number of its elements is Sort the nodes by score from highest to lowest, and select the node with the highest score. The nodes form a set of disease high-fever zones. Its definition is as follows: M = Top ∣T∣ ( fi )， in, Top ∣T∣ ( fi () indicates ranking based on disease relevance. The nodes obtained It maintains consistency with the number of drug targets while representing the most biologically active regions in the disease network, providing a structurally consistent control space for subsequent construction of drug-disease network intervention maps.

6. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, The formula for the exponential decay function described in S7 is as follows: ， in This is the distance attenuation coefficient. Protein target To protein target The shortest path length is adjusted by... Using the allowed shortest path length, different drug action modes are distinguished within a unified framework; The target-disease weighted matrix Where T is the set of drug targets, M is the set of disease high-intensity areas, and the drug coverage function is: ， in, A collection of protein targets for drugs. The disease relevance score is given to protein nodes in the high-intensity region of the disease.

7. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, The topological similarity calculation method described in S8 is as follows: Based on the protein-protein interaction network, for each drug target, a set of network nodes with a shortest path length of no more than 3 is determined, and based on the degree of overlap between these node sets, the Jaccard similarity between any two drug targets is calculated. ， in and Target points and The set of neighboring nodes.

8. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, The stepwise greedy selection strategy in S8 is as follows: initialize the set of selected target points to... Subsequently, in each iteration, targets that maximize the gain of the objective function are selected from the candidate targets that have not yet been selected. To balance the coverage of the core disease region with the structural diversity of the target set, a redundancy penalty term is introduced into the objective function, which takes the following form: ， in, Indicate target point For the current set of protein targets The resulting marginal coverage gain; For this target point The maximum topological similarity with any single target in the selected target set; Here, is the redundancy penalty coefficient, used to adjust the inhibition weight of structural repetition in the objective function. T is the set of drug targets. In the greedy selection process, only when a target satisfies ... At that time, it is considered to have a positive coverage contribution to the existing target set and is therefore included in the final effective target set.

9. The method for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in claim 1, characterized in that, The method further includes: S9: Based on the drug perturbation map obtained in S8, it characterizes the coverage and intensity of drugs on the biological functional modules of diseases, assesses the overall regulatory characteristics of drugs at the disease network level, and provides a basis for the multi-pathway synergistic mechanism of drugs and their ability to intervene in disease states.

10. A system for identifying perturbation maps of traditional Chinese and Western medicines based on deep learning and protein interaction network topology as described in any one of claims 1-9, the system comprising the following functional modules: (1) Disease target evidence integration module, which is used to acquire, integrate and continuously update multi-source disease-related evidence, including genome association studies, disease expression profile databases, experimentally verified pathogenic gene information, literature co-occurrence evidence, network coupling degree scores, and to standardize, complete missing information and continuously process labels for evidence, providing unified supervision labels for the model; (2) Protein interaction network construction module, which is used to load public databases or laboratory self-built interaction data, and perform identifier standardization, redundancy removal, bidirectional edge construction and network sparsification processing, and output node index, feature matrix and edge list that meet the training requirements of graph attention neural network; (3) Disease risk graph modeling module, which performs the S4–S5 tasks, including the construction, parameter setting, training, verification and testing process of graph attention network, and learns the continuous disease risk score of each node in the protein interaction network; (4) Disease high-temperature zone identification module, executes the S6 task, selects high-scoring nodes from the risk score sequence output by the graph attention neural network with reference to the number of drug targets, and constructs a set of protein nodes in disease high-temperature zones with a uniform scale. (5) Drug-disease topology mapping module, which performs shortest path calculation, exponential decay weighting, target-disease weighting matrix construction and adjustable parameterization of topology propagation space in S7; (6) Redundancy suppression and perturbation map generation module: Execute S8 task, introduce a stepwise greedy selection strategy with redundancy penalty term, adjust the repulsion strength of similar protein targets by adjusting the penalty coefficient, classify protein targets that generate positive marginal coverage gain in the high-temperature area of ​​the disease as effective contributing protein targets, and include them in the non-redundant target set to construct drug perturbation maps under different topological granularities.

Images