Disease marker subtype screening method based on molecular docking

Abstract

The invention provides a disease marker subtype screening method based on molecular docking, and belongs to the technical field of disease marker subtype screening, and the method comprises the following steps: firstly collecting multi-omics data of a patient and generating an expression profile and a regulation relation matrix, then constructing a molecular hypergraph network model, endowing nodes with features, and constructing hyperedges; and then calculating the marker association degree of hyperedges, extracting hierarchical features by using a hypergraph neural network, and identifying key molecular markers through node importance scores. And then verifying physical interaction by using a molecular docking technology, optimizing a network structure, and retraining the model to obtain a disease marker subtype classification model. And finally, verifying the performance of the model through an independent verification data set, and applying the verified model to disease subtype classification of a new patient. The method provided by the invention solves the problem of lack of verification of molecular physical interaction in the prior art, and more accurately screens out disease markers.

Description

Technical Field

[0001] This invention belongs to the field of disease biomarker subtype screening technology, and more specifically, relates to a disease biomarker subtype screening method based on molecular docking. Background Technology

[0002] With the deepening development of biomedical research, disease subtype classification is of great significance for achieving precision medicine. Traditional disease classification methods are mainly based on clinical phenotypes and pathological characteristics, which are difficult to fully reflect the molecular mechanisms of diseases. In recent years, the generation of multi-omics data has provided new opportunities for a deeper understanding of disease mechanisms. Existing disease biomarker screening methods mainly rely on statistical significance or feature importance scores from machine learning models, lacking verification of molecular physical interactions. This can lead to false positives in the screened disease biomarkers. Especially when integrating multi-omics data, how to effectively utilize information from different omics levels to improve the accuracy of biomarker screening remains an urgent problem to be solved. Summary of the Invention

[0003] In view of this, the present invention provides a disease biomarker subtype screening method based on molecular docking, which can solve the technical problem that the existing technology mainly relies on statistical significance or feature importance scores of machine learning models, lacks verification of molecular physical interactions, and may result in false positives in the screened disease biomarkers.

[0004] This invention is implemented as follows:

[0005] This invention provides a method for screening disease biomarker subtypes based on molecular docking, comprising the following steps:

[0006] S10. Collect transcriptomic data, proteomic data and metabolomic data of patient samples to form a multi-omics dataset, and generate an expression profile data matrix and a regulatory relationship matrix based on the multi-omics dataset;

[0007] S20. Construct a molecular hypergraph network model, using molecules as nodes, assigning initial feature matrices to the nodes based on the expression spectrum data matrix, constructing hyperedges based on the regulatory relationship matrix, and generating an initial hyperedge correlation matrix.

[0008] S30. Calculate the biomarker correlation degree of each hyperedge in the molecular hypergraph network using the biomarker correlation degree calculation function. The input of the biomarker correlation degree calculation function includes the expression spectrum data matrix, the regulatory relationship matrix, and the initial hyperedge correlation matrix. The output of the biomarker correlation degree calculation function is the hyperedge weight vector.

[0009] S40. Use a hypergraph neural network to perform feature learning and extract hierarchical marker features between molecules. The hierarchical marker features include a node feature matrix, a hyperedge feature matrix, and a global feature matrix.

[0010] S50. The importance of molecular nodes is ranked using a node importance scoring function. The input of the node importance scoring function includes the node feature matrix, the hyperedge feature matrix, the global feature matrix, and the initial hyperedge correlation matrix. The output of the node importance scoring function is a node importance score vector. Key molecular markers are identified based on the node importance score vector.

[0011] S60. Verify the physical interaction strength between the key molecular markers using molecular docking technology, obtain the molecular docking score matrix, and optimize the structure of the molecular hypergraph network using a hyperedge weight update function. The input of the hyperedge weight update function includes the hyperedge weight vector, the molecular docking score matrix, and the initial hyperedge correlation matrix. The output of the hyperedge weight update function is the optimized hyperedge weight matrix.

[0012] S70. Retrain the molecular hypergraph network model based on the optimized hyperedge weight matrix to obtain the disease biomarker subtype classification model.

[0013] S80. Validate the disease biomarker subtype classification model using an independent validation dataset to obtain the validated classification model;

[0014] S90. Apply the validated classification model to classify the new patient samples into disease subtypes.

[0015] Specifically, step S10 includes:

[0016] Step 101: Obtain transcriptomic, proteomic, and metabolomic data from patient samples to form a multi-omics dataset;

[0017] Step 102: Generate an expression profile data matrix based on the multi-omics dataset. The expression profile data matrix is ​​represented as an n-row m-column matrix, where n is the number of samples and m is the number of molecules. The matrix element represents the expression level of the j-th molecule in the i-th sample. The expression level is obtained through high-throughput sequencing experiments with a sequencing depth of not less than 30X.

[0018] Step 103: Generate a regulatory relationship matrix based on the multi-omics dataset. The regulatory relationship matrix is ​​represented as an m-row m-column matrix, and the matrix elements represent the regulatory strength between molecule i and molecule j, with values ​​ranging from -1 to 1. It is obtained through differential expression analysis and co-expression network analysis.

[0019] Specifically, step S20 includes:

[0020] Step 201: Construct a molecular hypergraph network model, using molecules as nodes;

[0021] Step 202: Assign an initial feature matrix to the node based on the expression spectrum data matrix;

[0022] Step 203: Construct hyperedges based on the regulation relationship matrix to generate an initial hyperedge correlation matrix with m rows and k columns, where k is the number of hyperedges and the matrix elements indicate whether numerator i belongs to hyperedge j. If it does, the value is 1, otherwise it is 0.

[0023] Specifically, step S30 includes:

[0024] Step 301: Calculate the marker correlation degree of each hyperedge in the molecular hypergraph network using the marker correlation degree calculation function;

[0025] Step 302: Use the expression spectrum data matrix, the regulation relationship matrix, and the initial hyperedge correlation matrix as inputs to the marker correlation degree calculation function;

[0026] Step 303: By calculating the correlation of molecular expression, regulatory strength and network topology characteristics within the hyperedge, and introducing a random error term, the hyperedge weight vector is obtained as the output.

[0027] Specifically, step S40 includes:

[0028] Step 401: Perform feature learning using a hypergraph neural network;

[0029] Step 402: Extract local molecular features through the first convolutional network to obtain the node feature matrix;

[0030] Step 403: Aggregate intermolecular interaction information through the second convolutional network to obtain the hyperedge feature matrix;

[0031] Step 404: Integrate global network information through the third convolutional network to obtain the global feature matrix and complete the hierarchical marker feature extraction.

[0032] Specifically, step S50 includes:

[0033] Step 501: Rank the molecular nodes by importance using the node importance scoring function;

[0034] Step 502: Use the node feature matrix, the hyperedge feature matrix, the global feature matrix, and the initial hyperedge correlation matrix as inputs to the node importance scoring function;

[0035] Step 503: Obtain the node importance score vector through multi-level feature weighted fusion and random error term;

[0036] Step 504: Identify key molecular markers based on the node importance score vector.

[0037] Specifically, step S60 includes:

[0038] Step 601: Verify the strength of the physical interaction between the key molecular markers using molecular docking technology;

[0039] Step 602: Obtain the molecular docking score matrix based on the strength of physical interactions;

[0040] Step 603: Optimize the structure of the molecular hypergraph network using the hyperedge weight update function. The input of the hyperedge weight update function includes the hyperedge weight vector, the molecular docking scoring matrix, and the initial hyperedge correlation matrix.

[0041] Step 604: By combining the original weights with the molecular docking results and considering the influence of the neighborhood, the optimized hyperedge weight matrix is ​​obtained.

[0042] Specifically, step S70 includes:

[0043] Step 701: Use the optimized hyperedge weight matrix as the model input;

[0044] Step 702: Retrain the molecular hypergraph network model and update the model parameters;

[0045] Step 703: Obtain the disease biomarker subtype classification model.

[0046] Specifically, step S80 includes:

[0047] Step 801: Obtain the independent validation dataset;

[0048] Step 802: Validate the disease biomarker subtype classification model using the independent validation dataset;

[0049] Step 803: Adjust the model based on the validation results to obtain the validated classification model.

[0050] Specifically, step S90 includes:

[0051] Step 901: Obtain expression profile data of new patient samples;

[0052] Step 902: Input the expression profile data into the validated classification model;

[0053] Step 903: Classify the new patient samples into disease subtypes using the validated classification model to obtain the classification results.

[0054] Furthermore, the molecular hypergraph network model is based on the graph attention network model, and a three-layer pyramid-shaped convolutional network structure is constructed on the basis of the hypergraph convolutional neural network. The three-layer pyramid-shaped convolutional network structure includes a first convolutional network, a second convolutional network, and a third convolutional network.

[0055] The first convolutional network is used to extract local molecular features. The input of the first convolutional network includes the node initial feature matrix and the initial hyperedge correlation matrix. The output of the first convolutional network is the node feature matrix.

[0056] The second convolutional network is used to aggregate intermolecular interaction information. The input of the second convolutional network includes the node feature matrix and the hyperedge weight vector. The output of the second convolutional network is the hyperedge feature matrix.

[0057] The third convolutional network is used to integrate global network information. The input of the third convolutional network includes the hyperedge feature matrix and the initial hyperedge correlation matrix. The output of the third convolutional network is the global feature matrix.

[0058] The molecular hypergraph network model is structured as a cascaded connection of a multi-head attention mechanism. The first convolutional network, the second convolutional network, and the third convolutional network each contain an attention layer, a feature transformation layer, and a nonlinear activation layer. The node feature matrix, the hyperedge feature matrix, and the global feature matrix are preserved through skip connections. Finally, the output layer uses the softmax function for feature fusion.

[0059] Compared with existing technologies, the beneficial effects of the disease biomarker subtype screening method based on molecular docking provided by this invention are:

[0060] First, by introducing a hypergraph network structure, this method overcomes the limitation of traditional graph networks, which can only describe the relationship between two molecules. The design of hyperedges enables the model to simultaneously characterize the complex regulatory relationships between multiple molecules, providing a modeling framework that is more consistent with the actual situation of biological systems.

[0061] Secondly, an innovative three-layer pyramid-shaped convolutional network structure was constructed to achieve hierarchical feature extraction from local to global perspectives. The first layer of the convolutional network captures directly interacting molecular features, the second layer integrates neighborhood information, and the third layer extracts global network patterns. This progressive feature learning strategy significantly improves the model's ability to understand complex molecular regulatory patterns, enabling it to discover potential biomarkers that are difficult to identify using traditional methods.

[0062] Third, by introducing molecular docking technology to verify physical interactions, this method significantly improves the reliability of biomarker screening.

[0063] Fourth, validation results on multiple disease datasets demonstrate that, compared to existing techniques, this method not only improves the accuracy of disease subtype classification but also significantly enhances the interpretability of the classification results. By analyzing the key biomarkers identified by the model and their interaction networks, researchers can better understand the molecular mechanisms of disease development and progression.

[0064] In summary, this invention addresses the technical problem that existing technologies mainly rely on statistical significance or feature importance scores from machine learning models, lacking verification of molecular physical interactions, which may lead to false positives in the screened disease biomarkers. Attached Figure Description

[0065] Figure 1 A flowchart of the method provided by the present invention;

[0066] Figure 2 This is a schematic diagram of the molecular hypergraph network in Example 2;

[0067] Figure 3 This is a box plot of the subtype marker expression in Example 2. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0069] like Figure 1 The diagram shown is a flowchart of a disease biomarker subtype screening method based on molecular docking provided by this invention. The specific implementation details of each step are described below:

[0070] Step S10 is specifically implemented by acquiring multi-omics data from patient samples using high-throughput sequencing technology, including transcriptome data (sequencing depth ≥30X), proteome data, and metabolome data. Transcriptome sequencing is performed using a next-generation sequencing platform, and a sequencing depth of at least 30X ensures data reliability and accuracy. Based on this high-quality multi-omics data, an expression profile matrix is ​​generated to record the expression levels of each molecule in each sample. Simultaneously, a regulatory relationship matrix is ​​generated through differential expression analysis and co-expression network analysis to describe the strength of regulatory relationships between molecules. Differential expression analysis can identify molecules whose expression levels change significantly under different disease states, while co-expression network analysis reveals synergistic expression patterns between molecules. These two matrices provide comprehensive data support for the subsequent construction of a molecular hypergraph network model. The "X" in "30X" represents coverage or depth, and is a multiple sign.

[0071] The specific implementation of step S20 involves constructing a hypergraph network model with molecules as nodes. This model overcomes the limitation of traditional graph networks, which can only describe the relationship between two nodes, by introducing a hyperedge structure that enables the model to simultaneously characterize the complex relationships between multiple molecules. Specifically, an initial feature is assigned to each molecular node using an expression profile data matrix; these features reflect the expression patterns of the molecule in different samples. Hyperedges are constructed based on the regulatory relationship matrix to form an initial hyperedge association matrix, where each hyperedge connects a group of molecules with close regulatory relationships. This network structure design provides richer network topology information for subsequent biomarker feature extraction.

[0072] The specific implementation of step S30 involves calculating the biomarker correlation degree of each hyperedge in the molecular hypergraph network. This calculation process comprehensively considers several important factors: first, it analyzes the expression correlation of molecules within the hyperedges to examine the cooperative change patterns of molecular expression levels; second, it assesses the regulatory strength between molecules to quantify the mutual regulatory relationships between molecules; and third, it analyzes the network topological characteristics to consider the structural importance of molecules in the entire network. Simultaneously, a random error term is introduced to improve the robustness of the model, making the calculation results more stable and reliable. The hyperedge weight vector obtained through this step comprehensively reflects the importance and functional significance of different molecular combinations in the disease development process.

[0073] The specific implementation of step S40 involves using an innovative three-layer pyramid-shaped convolutional network structure for feature learning. The first convolutional layer focuses on extracting local molecular features, capturing the correlation patterns between directly interacting molecules. The second convolutional layer is responsible for aggregating intermolecular interaction information, integrating molecular interaction relationships over a larger scope. The third convolutional layer integrates global network information, understanding the complex regulatory networks between molecules from a holistic perspective. This hierarchical feature extraction architecture can capture molecular marker feature patterns from the microscopic to the macroscopic level from the bottom up, providing rich feature representations for subsequent classification tasks.

[0074] Step S50 is implemented by ranking the importance of molecular nodes using a node importance scoring function. This scoring process employs a multi-level feature fusion strategy, comprehensively considering local features of nodes, neighborhood features transmitted through hyperedges, and global features reflecting the overall network characteristics. By designing a reasonable feature weighting scheme, the contributions of features at different levels are balanced in the final score. Simultaneously, a random error term is introduced to improve the robustness of the score and avoid overfitting to features of a specific dataset. Based on this comprehensive scoring result, molecular biomarkers that play a key regulatory role in disease development can be accurately identified.

[0075] The specific implementation of step S60 involves using molecular docking technology to verify and quantify the physical interactions between key molecular markers. Molecular docking analysis considers multiple aspects, including the three-dimensional structural information of molecules, binding site characteristics, and interaction energies, generating a molecular docking score matrix reflecting the strength of intermolecular interactions. Based on this score matrix, a hyperedge weight update function is used to optimize the molecular hypergraph network structure. The optimization process not only considers the original hyperedge weights but also incorporates the physical interaction information provided by the molecular docking results, while maintaining the continuity of the network structure by considering neighborhood influences. By introducing a random error term, the model's accuracy in describing molecular interaction patterns and its generalization ability are further improved.

[0076] The specific implementation of step S70 is to retrain the molecular hypergraph network model based on the optimized hyperedge weight matrix. This step updates the model parameters through an iterative optimization process, enabling the model to more accurately capture the physical interaction features and functional association patterns between molecules. The optimized model, while maintaining its original classification ability, enhances its sensitivity to intermolecular physical interactions, thus obtaining a classification model capable of accurately segmenting disease biomarker subtypes. This model optimization strategy based on physical interaction validation significantly improves the interpretability and reliability of the classification results.

[0077] Step S80 involves comprehensively validating the disease biomarker subtype classification model using an independent validation dataset. The validation process measures model performance using multiple evaluation metrics, including classification accuracy, sensitivity, and specificity. Based on the validation results, key parameters and network structure of the model are specifically adjusted and optimized to ensure good generalization ability. This validation and optimization process is a crucial step in ensuring the reliability of the model in real-world clinical applications; rigorous independent validation improves the credibility of the model's prediction results.

[0078] The specific implementation of step S90 involves inputting the expression profile data of new patient samples into a validated classification model to achieve accurate classification of disease subtypes. This step first requires standardizing the expression profile data of the new patients to make it comparable to the model training data. Then, the processed data is input into the optimized classification model, and the disease subtype to which the patient belongs is determined based on the model's prediction results. This application process realizes the transformation of the model from research and development to clinical practice, providing important scientific evidence for the accurate diagnosis, subtyping, and personalized treatment planning of diseases.

[0079] Specifically, the principle of this invention is:

[0080] From a data perspective, this method integrates multi-dimensional molecular information from the transcriptome, proteome, and metabolome. This integration of multi-omics data is fundamental to accurate subtyping, as disease development often involves coordinated regulation at multiple molecular levels. Transcriptome data (sequencing depth ≥30X) obtained through high-throughput sequencing ensures the accuracy of gene expression information, while proteome and metabolome data provide important supplementary information at the functional level.

[0081] From a modeling perspective, the core reason for employing a hypergraph network structure in this method is that interactions between biomolecules are typically many-to-many relationships. Traditional graph networks can only describe the connection between two nodes, and this simplified modeling approach loses a significant amount of biological information. Hyperedges in a hypergraph, however, allow for the simultaneous connection of multiple nodes, and this mathematical representation better reflects the actual state of biomolecular regulation. For example, transcriptional regulation often involves the synergistic effects of multiple molecules such as transcription factors, cofactors, and target genes; this complex regulatory relationship can be accurately represented by a single hyperedge.

[0082] The principle behind using a three-layer pyramidal convolutional network structure for feature extraction lies in the distinct hierarchical nature of biomolecular networks. Direct interactions at the local level, functional modules at intermediate levels, and global regulatory networks—features at different scales—all contain important biological information. The design of the three-layer convolutional network is based on this biological understanding, achieving a comprehensive understanding of molecular regulatory networks by extracting and integrating features at different scales layer by layer.

[0083] The principle behind introducing molecular docking technology for physical verification is based on the fact that computationally predicted molecular interactions require physicochemical verification to confirm their reliability. Molecular docking can simulate three-dimensional structural matching and energy calculation between molecules, providing physical evidence for the predicted interactions. This verification mechanism not only eliminates false-positive interaction predictions but also provides a quantitative assessment of interaction strength, offering a reliable basis for subsequent network optimization.

[0084] The principle behind the hyperedge weight update strategy for network structure optimization is that the validation results of molecular docking should be fed back into the network model to improve its ability to describe real biological processes. By fusing the molecular docking score with the original network weights and considering the influence of neighborhood, the model can be gradually optimized to create a network structure that better reflects physical reality.

[0085] The model's generalization ability stems from its multi-layered feature learning architecture and physically validated optimization strategy. Multi-layered feature extraction enables the model to capture biological patterns at different scales, while physical validation ensures the reliability of these features. This design allows the model to not only accurately classify known samples but also effectively handle novel patient data.

[0086] The following provides a specific embodiment 1 of the present invention. The specific implementation of each step in this embodiment 1 is described in detail below: The specific implementation of step S10 is as follows:

[0087] First, transcriptomic, proteomic, and metabolomic data from patient samples were acquired to form a multi-omics dataset. Transcriptomic data were obtained through high-throughput sequencing with a sequencing depth of at least 30X.

[0088] Then, based on the multi-omics dataset, an expression profile data matrix E = {e} is generated. ij} n×m Where j represents the expression level of the j-th molecule in the i-th sample; n is the number of samples; and m is the number of molecules.

[0089] Furthermore, it is also necessary to generate a regulatory relationship matrix R = {r} based on the aforementioned multi-omics dataset. ij} m×m Among them, r ij The value represents the regulatory strength between molecule i and molecule j, and its range is [-1, 1]. It is obtained through differential expression analysis and co-expression network analysis.

[0090] Through the above steps, a multi-omics dataset containing expression profiles and regulatory relationships was obtained, laying the foundation for the subsequent construction of molecular hypergraph network models.

[0091] The specific implementation method of step S20 is as follows:

[0092] First, a molecular hypergraph network model is constructed, with molecules as nodes.

[0093] Then, based on the expression spectrum data matrix E, the node is assigned an initial feature matrix E. 0 .

[0094] Next, based on the aforementioned regulation relation matrix R, hyperedges are constructed to generate an initial m×k hyperedge correlation matrix H = {h ij}. Among them, h ij Indicates whether numerator i belongs to hyperedge j. If it does, the value is 1, otherwise it is 0; k is the number of hyperedges.

[0095] Through the above steps, a hypergraph network model with molecules as nodes was constructed, and initial features were assigned to the nodes and hyperedges, laying the foundation for subsequent marker association calculation and feature learning.

[0096] The specific implementation method of step S30 is as follows:

[0097] First, the correlation function of markers is used. Calculate the marker correlation degree for each hyperedge in the molecular hypergraph network. Where w i e represents the weight of the i-th hyperedge; i Let corr(E) be the set of molecules contained in the i-th hyperedge. j E k ) represents the correlation coefficient between molecule j and molecule k; deg(j) represents the degree of molecule j; α, β, γ are learnable parameters with values ​​ranging from [0,1]; δ represents the normality of the expression. The random error term.

[0098] The marker correlation calculation function considers the correlation of molecular expression within the hyperedge, the intensity of regulation, and the network topology characteristics, and introduces a random error term to improve the robustness of the model.

[0099] The specific implementation method of step S40 is as follows:

[0100] First, a hypergraph neural network is used for feature learning. Specifically, a first convolutional network extracts local molecular features to obtain a node feature matrix. The input to the first convolutional network includes the node initial feature matrix E. 0 And the initial hyperedge incidence matrix H. Where, D v D is a diagonal matrix with node degree. e Let Θ be a hyperedge degree diagonal matrix, W be a hyperedge weight diagonal matrix, and Θ be a hyperedge weight diagonal matrix. 1 Let σ be the learnable parameter matrix of the first layer, and σ be the activation function.

[0101] Then, the intermolecular interaction information is aggregated through a second convolutional network to obtain the hyperedge feature matrix F. 2 =σ(A1F 1 Θ 2 )+σ(A2F 1 The input to this second convolutional network includes the node feature matrix F. 1 And the hyperedge weight vector. Where A1 and A2 are attention weight matrices, Θ 2 This is the learnable parameter matrix for the second layer.

[0102] Finally, the global network information is integrated through a third convolutional network to obtain the global feature matrix F. 3 =σ(MLP(concat[F 1 ,F 2 ])Θ 3 The input to this third convolutional network includes the hyperedge feature matrix F. 2 And the initial hyperedge correlation matrix H. Where MLP is a multilayer perceptron, concat is the feature concatenation operation, and Θ... 3 This is the learnable parameter matrix for the third layer.

[0103] The three-layer pyramid-shaped convolutional network structure described above completes the hierarchical extraction of marker features from local to global perspectives, laying the foundation for subsequent node importance assessment.

[0104] The specific implementation method of step S50 is as follows:

[0105] First, using the node importance scoring function The molecular nodes are ranked by importance. Where s i Score the importance of the i-th node; Let N(i) be the feature vector of node i; N(i) is the set of hyperedges associated with node i. f is the eigenvector of the j-th hyperedge; global Here, θ1, θ2, and θ3 are the global feature vectors; θ1, θ2, and θ3 are the feature fusion coefficients, satisfying θ1 + θ2 + θ3 = 1; and η is a variable that follows a normal distribution. The random error term.

[0106] The node importance scoring function adopts a multi-level feature weighted fusion approach, which considers both local and global information and introduces an error term to deal with data noise.

[0107] Then, key molecular markers are identified based on the node importance score vector.

[0108] The specific implementation method of step S60 is as follows:

[0109] First, molecular docking technology was used to verify the strength of the physical interactions between the key molecular markers, and the molecular docking scoring matrix D = {d ij}

[0110] Then, the hyperedge weight update function is used. The molecular hypergraph network was structurally optimized. This represents the updated weight of the i-th hyperedge. The original weights; d i Let λ be the molecular docking score corresponding to hyperedge i; λ be the weight update coefficient, with a value range of [0,1]; μ be the neighborhood influence factor, with a value range of [0,0.5]; and ξ be a normally distributed ξ. The random error term.

[0111] The hyperedge weight update function combines the original weights with the molecular docking results, takes into account the influence of the neighborhood, and handles experimental errors through an error term.

[0112] The specific implementation method of step S70 is as follows:

[0113] First, based on the optimized hyperedge weight matrix W new As input to the model.

[0114] Then, the molecular hypergraph network model is retrained, and the model parameters are updated.

[0115] Ultimately, a disease biomarker subtype classification model was obtained.

[0116] This step involves retraining the classification model using the optimized hypergraph network structure, enabling it to more accurately classify disease biomarker subtypes.

[0117] The specific implementation methods of steps S80 and S90 are the same as those of the present invention, and will not be described in detail again.

[0118] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: Researchers conducted a research project on ovarian cancer patients, aiming to discover key molecular markers and their subtype characteristics related to the disease by comprehensively analyzing the patients' transcriptome, proteome and metabolome data, so as to provide a basis for clinical diagnosis and personalized treatment.

[0119] First, the research team collected biological samples from 50 ovarian cancer patients and 20 healthy controls, and obtained transcriptomic, proteomic, and metabolomic data from these samples using high-throughput sequencing technology. After data preprocessing and standardization, a multi-omics dataset containing 70 samples and 18,426 molecular features was finally formed. Specific data details are shown in Table 1.

[0120] Table 1 Overview of Multi-omics Datasets

[0121] Data types Sample size Number of molecules Transcriptome 70 15428 proteomics 70 2564 metabolomics 70 434 total 70 18426

[0122] Next, the research team constructed a molecular hypergraph network model based on this multi-omics dataset. First, a 70×15428 expression profile matrix E was generated from the transcriptome data, where e ij Let represent the expression level of the j-th gene in the i-th sample. Simultaneously, using differential expression analysis and co-expression network analysis, a 15428×15428 regulatory relationship matrix R was generated, where r... ij This represents the regulatory strength between gene i and gene j, with values ​​ranging from -1 to 1. For example... Figure 2 As shown, this is a network graph illustrating the structure of a molecular hypergraph network. The graph contains multiple nodes and edges, with different colors representing different types of molecules, and edge thickness representing interaction strength. The graph also includes mathematical expressions for node features, edge weights, and hyperedge correlation.

[0123] Then, these 18426 molecules were used as nodes, and initial features were assigned to the nodes based on the expression profile data matrix. Hyperedges were constructed based on the regulatory relationship matrix, thus forming a molecular hypergraph network model. Specifically, the initial hyperedge correlation matrix H is a 15428×k matrix, where k is the number of hyperedges, and h is the number of hyperedges. ij This indicates whether gene i belongs to hyperedge j; if it does, the value is 1, otherwise it is 0.

[0124] Next, the research team used the marker association function to calculate the correlation degree. The marker correlation degree for each hyperedge was calculated. Where, w i e represents the weight of the i-th hyperedge; i Let corr(E) be the set of genes contained in the i-th superedge. j E k ) represents the correlation coefficient between gene j and gene k; deg(j) represents the degree of gene j; α, β, γ are learnable parameters with values ​​ranging from [0,1]; δ represents the normal distribution. The random error term. Through this step, the marker correlation degree of each hyperedge is obtained.

[0125] Then, the research team used a hypergraph neural network for feature learning and extracted the node feature matrix F. 1 Hyperedge feature matrix F 2 and the global characteristic matrix F 3Specifically, local gene features are extracted through the first convolutional network, inter-gene interaction information is aggregated through the second convolutional network, and global network features are integrated through the third convolutional network, ultimately resulting in a three-level molecular representation.

[0126] Next, the research team used a node importance scoring function. The importance of molecular nodes is ranked. Where, s i Score the importance of the i-th node; Let N(i) be the feature vector of node i; N(i) is the set of hyperedges associated with node i. f is the eigenvector of the j-th hyperedge; global Here, θ1, θ2, and θ3 are the global feature vectors; θ1, θ2, and θ3 are the feature fusion coefficients, satisfying θ1 + θ2 + θ3 = 1; and η is a variable that follows a normal distribution. The random error term was determined. Based on the node importance score vector, the research team identified 200 key molecular biomarkers.

[0127] To further verify the role of these key molecular markers, the research team used molecular docking technology to calculate the strength of their physical interactions, obtaining a 200×200 molecular docking score matrix D. Then, the team used a hyperedge weighting update function... The structure of the molecular hypergraph network was optimized. Among other things, This represents the updated weight of the i-th hyperedge. The original weights; d i Let λ be the molecular docking score corresponding to hyperedge i; λ be the weight update coefficient, with a value range of [0,1]; μ be the neighborhood influence factor, with a value range of [0,0.5]; and ξ be a normally distributed ξ. The random error term.

[0128] Based on the optimized hyperedge weight matrix, the research team retrained the molecular hypergraph network model, obtaining a classification model capable of accurately identifying ovarian cancer biomarker subtypes. To evaluate the generalization performance of this classification model, the team randomly selected 20 ovarian cancer patients and 10 healthy controls as independent validation sets, inputting their data into the trained classification model for prediction. Validation results showed that the classification model achieved an accuracy of over 90% in identifying ovarian cancer biomarker subtypes, outperforming existing single-atom data analysis methods.

[0129] Next, the research team applied this classification model to a new sample of ovarian cancer patients, successfully dividing these patients into three different biomarker subtypes. As shown in Table 2, these three subtypes showed significant differences in the expression levels of certain key molecular markers such as BRCA1, TP53, and KRAS, suggesting that they may represent different pathological mechanisms and treatment response characteristics.

[0130] Table 2 Key molecular characteristics of three ovarian cancer biomarker subtypes

[0131] Molecular markers Subtype 1 Subtype 2 Subtype 3 BRCA1 High expression low expression medium TP53 High expression low expression medium KRAS low expression High expression medium PTEN High expression low expression medium CCNE1 medium High expression low expression

[0132] like Figure 3 As shown, this is a box plot illustrating the expression levels of five key molecular markers (BRCA1, TP53, KRAS, PTEN, and CCNE1) across three different subtypes. Different colors are used to distinguish between subtypes, and the Y-axis uses a log2FC scale to represent relative expression levels. Each box plot displays the median, quartiles, and outliers. Through further clinical follow-up and experimental validation, the research team found that: subtype 1 patients have a relatively better prognosis and may be suitable for PARP inhibitor treatment; subtype 2 patients have a poorer prognosis and may be suitable for MEK inhibitors or immune checkpoint inhibitors; while subtype 3 patients fall between the two and may require combination therapy. These findings provide a basis for clinicians to develop personalized treatment plans.

[0133] It should be noted that the variables involved in this invention are explained in detail in Table 3 below.

[0134] Table 3. Variable Explanation Table

[0135]

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A molecular docking-based disease marker subtype screening method, characterized by, The method comprises the following steps: S10, collecting transcriptome data, proteome data and metabolome data of patient samples to form a multi-omics dataset, and generating an expression profile data matrix and a regulatory relationship matrix based on the multi-omics dataset; S20, constructing a molecular hypergraph network model, taking molecules as nodes, assigning the nodes an initial feature matrix based on the expression profile data matrix, constructing hyperedges based on the regulatory relationship matrix, and generating an initial hyperedge correlation matrix; S30, calculating the marker correlation degree of each hyperedge in the molecular hypergraph network by using a marker correlation degree calculation function, the input of the marker correlation degree calculation function comprising the expression profile data matrix, the regulatory relationship matrix and the initial hyperedge correlation matrix, and the output of the marker correlation degree calculation function being a hyperedge weight vector; S40, performing feature learning by using a hypergraph neural network to extract hierarchical marker features among molecules, the hierarchical marker features comprising a node feature matrix, a hyperedge feature matrix and a global feature matrix; S50, performing importance ranking on the molecular nodes by using a node importance scoring function, the input of the node importance scoring function comprising the node feature matrix, the hyperedge feature matrix, the global feature matrix and the initial hyperedge correlation matrix, the output of the node importance scoring function being a node importance score vector, and identifying key molecular markers based on the node importance score vector; S60, verifying the physical interaction strength between the key molecular markers by using molecular docking technology to obtain a molecular docking score matrix, and performing structure optimization on the molecular hypergraph network by using a hyperedge weight updating function, the input of the hyperedge weight updating function comprising the hyperedge weight vector, the molecular docking score matrix and the initial hyperedge correlation matrix, and the output of the hyperedge weight updating function being an optimized hyperedge weight matrix; S70, retraining the molecular hypergraph network model based on the optimized hyperedge weight matrix to obtain a disease marker subtype classification model; S80, verifying the disease marker subtype classification model by using an independent verification dataset to obtain a verified classification model; S90, applying the verified classification model to disease subtype classification of new patient samples.

2. The method of claim 1, wherein, The step S10 specifically comprises: Step 101, obtaining transcriptome data, proteome data and metabolome data of patient samples to form a multi-omics dataset; Step 102, generating an expression profile data matrix based on the multi-omics dataset, the expression profile data matrix being represented as an n-row m-column matrix, wherein n is the number of samples, m is the number of molecules, and the matrix element represents the expression of the jth molecule in the ith sample, the expression being obtained by high-throughput sequencing experiment, and the sequencing depth being not less than 30X; Step 103, generating a regulatory relationship matrix based on the multi-omics dataset, the regulatory relationship matrix being represented as an m-row m-column matrix, the matrix element representing the regulatory strength between molecule i and molecule j, the value range being -1 to 1, and the regulatory strength being obtained by differential expression analysis and co-expression network analysis.

3. The method of claim 1, wherein, The step S20 specifically comprises: Step 201, constructing a molecular hypergraph network model, taking molecules as nodes; Step 202, assigning a node initial feature matrix to the node based on the expression profile data matrix; Step 203, constructing a hyperedge based on the regulatory relationship matrix, generating an initial hyperedge association matrix of m rows and k columns, where k is the number of hyperedges, and the matrix elements represent whether molecule i belongs to hyperedge j, taking the value of 1 if it belongs to, otherwise 0.

4. The method of claim 1, wherein, The step S30 specifically comprises: Step 301, calculating the marker correlation degree of each hyperedge in the molecular hypergraph network using a marker correlation degree calculation function; Step 302, taking the expression profile data matrix, the regulatory relationship matrix, and the initial hyperedge association matrix as inputs of the marker correlation degree calculation function; Step 303, obtaining a hyperedge weight vector as output by calculating the molecular expression correlation, regulatory strength, and network topology characteristics in the hyperedge, and introducing a random error term.

5. The method of claim 1, wherein, The step S40 specifically comprises: Step 401, feature learning using a hypergraph neural network; Step 402, extracting local molecular features through a first convolutional network to obtain a node feature matrix; Step 403, aggregating intermolecular interaction information through a second convolutional network to obtain a hyperedge feature matrix; Step 404, integrating global network information through a third convolutional network to obtain a global feature matrix, and completing hierarchical marker feature extraction.

6. The method of claim 1, wherein, The step S50 specifically comprises: Step 501, sorting the molecular nodes by importance using a node importance scoring function; Step 502, taking the node feature matrix, the hyperedge feature matrix, the global feature matrix, and the initial hyperedge association matrix as inputs of the node importance scoring function; Step 503, obtaining a node importance score vector by multi-level feature weighted fusion and a random error term; Step 504, identifying key molecular markers based on the node importance score vector.

7. The method of claim 1, wherein, The step S60 specifically comprises: Step 601, verifying the physical interaction strength between the key molecular markers using molecular docking technology; Step 602, obtaining a molecular docking score matrix according to the physical interaction strength; Step 603, structurally optimizing the molecular hypergraph network using a hyperedge weight update function, the inputs of the hyperedge weight update function including the hyperedge weight vector, the molecular docking score matrix, and the initial hyperedge association matrix; Step 604, obtaining an optimized hyperedge weight matrix by combining the original weight and the molecular docking result and considering the neighborhood influence.

8. The method of claim 1, wherein, The step S70 specifically comprises: Step 701, taking the optimized hyperedge weight matrix as model input; Step 702, retraining the molecular hypergraph network model and updating model parameters; Step 703, obtaining a disease marker subtype classification model.

9. The method of claim 1, wherein, The step S80 specifically comprises: Step 801, obtaining an independent validation data set; Step 802, verifying the disease marker subtype classification model using the independent validation data set; Step 803, adjusting the model according to the verification result to obtain a verified classification model.

10. The method of claim 1, wherein, The step S90 specifically comprises: Step 901, obtaining expression profile data of a new patient sample; Step 902, inputting the expression profile data into the verified classification model; Step 903, classifying the disease subtype of the new patient sample by the verified classification model to obtain a classification result.

Images