Cervical cancer gene regulation network analysis method based on multi-scale graph convolution

Abstract

The application discloses a cervical cancer gene regulation network analysis method based on a multi-scale graph convolution, S1. Construct a multi-modal biological data set; S2. Construct an initial cervical cancer gene regulation network based on the multi-modal biological data set; S3. Feature extraction is performed on the initial cervical cancer gene regulation network by using an improved multi-scale graph convolution network; S4. The network structure, graph convolution kernel weight and hyperparameter of the improved multi-scale graph convolution network are globally and locally searched and optimized by using a zebra optimization algorithm, so that an optimized improved multi-scale graph convolution network model is obtained; S5. A high-confidence cervical cancer gene regulation network is generated; S6. The biological significance of key driver genes of cervical cancer in cervical cancer is verified in combination with a public database, and the analysis of the cervical cancer gene regulation network is realized. The application is helpful to improve the accuracy and biological credibility of the gene regulation network, and provides theoretical support for early diagnosis and individualized treatment of cervical cancer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cervical cancer, and in particular to a cervical cancer gene regulation network analysis method based on multi-scale graph convolution. BACKGROUND

[0002] With the development of bioinformatics and artificial intelligence technology, the analysis of gene regulation network plays an increasingly important role in the field of tumor research. Cervical cancer, as one of the common malignant tumors of women worldwide, is affected by various gene regulation mechanisms. Analyzing the cervical cancer gene regulation network can reveal key driver genes and provide theoretical support for precision medicine.

[0003] Currently, the construction of gene regulation network mainly relies on gene co-expression analysis, Bayesian network and random forest traditional computing methods. Traditional methods usually only focus on the linear or low-order nonlinear relationship between genes, making it difficult to fully capture complex multi-scale gene regulation patterns. In addition, existing methods are often based on a single data source and fail to fully integrate transcription factor binding data, epigenetic modification data and other multi-modal biological data, limiting the biological credibility of the regulation network.

[0004] In recent years, graph neural networks have been preliminarily applied in gene regulation network analysis. Compared with traditional methods, graph neural networks have stronger non-linear modeling capabilities and can more effectively learn the topology of gene regulation networks. However, most current graph neural network methods still have the following shortcomings: first, existing graph neural network models mostly use single-scale graph convolution, which cannot simultaneously learn local and global gene regulation information, resulting in insufficient comprehensiveness of feature extraction; second, the training process of graph neural networks relies on fixed hyperparameter configurations, making it difficult to adaptively optimize for different gene datasets and affecting the generalization ability of the model.

[0005] In summary, there is an urgent need for a new method to accurately analyze cervical cancer gene regulation networks under multi-scale information, integrate multi-modal biological data to improve the accuracy of the regulation network, and use optimization methods to improve the adaptability and robustness of graph neural networks in gene regulation network analysis, in order to address the limitations of current technology. SUMMARY

[0006] One object of the present application is to provide a cervical cancer gene regulation network analysis method based on multi-scale graph convolution. The present application helps to improve the accuracy and biological credibility of gene regulation networks, providing theoretical support for early diagnosis and personalized treatment of cervical cancer.

[0007] According to the cervical cancer gene regulation network analysis method based on multi-scale graph convolution of the present application, the following steps are included:

[0008] S1. Collect gene expression data, transcription factor binding data and epigenetic data related to cervical cancer, and preprocess to construct a multi-modal biological data set;

[0009] S2. Construct an initial cervical cancer gene regulation network based on the multi-modal biological data set;

[0010] S3. Use an improved multi-scale graph convolutional network to extract features from the initial cervical cancer gene regulation network, and fuse the features of each scale to generate multi-scale fusion features;

[0011] S4. Take the multi-scale fusion features and the training parameters of the improved multi-scale graph convolutional network as inputs, and use the Zebra optimization algorithm to globally and locally search and optimize the network structure, graph convolution kernel weight and hyperparameters of the improved multi-scale graph convolutional network, to obtain an optimized improved multi-scale graph convolutional network model;

[0012] S5. Use the optimized improved multi-scale graph convolutional network model to infer unknown gene regulation relationships in the cervical cancer gene regulation network, generate gene regulation relationship inference results, and evaluate the confidence of the inference results, and generate a high-confidence cervical cancer gene regulation network;

[0013] S6. Verify the biological significance of the key driver genes of cervical cancer in cervical cancer by combining public databases, and realize the analysis of the cervical cancer gene regulation network.

[0014] Optionally, the step S1 specifically comprises:

[0015] S11. Collect cervical cancer-related gene expression data X G , transcription factor binding data X T and epigenetic data X E ;

[0016] S12. Process the gene expression data X G , transcription factor binding data X T and epigenetic data X E uniformly, construct a multi-modal biological data set, and standardize the multi-modal biological data set to normalize its value range to the interval [0, 1];

[0017] S13. Fill in the missing values by completing the missing values of the standardized multi-modal biological data set;

[0018] S14. Reduce the dimensionality of the multi-modal biological data set after filling in the missing values, and extract the main information features;

[0019] S15. Filter the noise of the multi-modal biological data set after dimensionality reduction;

[0020] S16. Combining gene expression data, transcription factor binding data and epigenetic data, according to the gene number g i Data matching is performed to align the features of different data sources to construct the final multi-modal biological dataset:

[0021]

[0022] Optionally, the gene expression data is obtained by RNA sequencing, representing the expression level of genes in different samples, constituting the gene expression matrix X G The transcription factor binding data is obtained by transcription factor immunoprecipitation sequencing, representing the binding of a specific transcription factor to a target gene, constituting the transcription factor binding matrix X T The epigenetic data includes DNA methylation data and histone modification data, respectively from methylation chip sequencing and histone modification immunoprecipitation sequencing, for representing the epigenetic regulation state of genes, constituting the epigenetic matrix X E .

[0023] Optionally, the step S2 specifically includes:

[0024] S21. Constructing a node set V of the cervical cancer gene regulation network according to the final multi-modal biological dataset, wherein each gene g i is a node v i , and the gene node set V is defined;

[0025] S22. Calculating the correlation between genes according to the gene expression data The Pearson correlation coefficient R(i,j) between genes is calculated as a gene expression correlation measure:

[0026]

[0027] wherein, represents the expression value of gene g i in sample k, is the average expression value of gene g i in all samples, and m is the total number of samples;

[0028] S23. Calculating the regulation relationship weight W(i,j) between genes, integrating transcription factor binding data and epigenetic data The gene regulation relationship weight is defined by linear combination:

[0029]

[0030] wherein, α, β, γ are respectively the weighting parameters of gene expression data, transcription factor binding data and epigenetic regulation data on the influence of gene regulation relationship.

[0031] S24. Constructing a set of regulatory relationships E of the cervical cancer gene regulatory network according to the gene regulatory relationship weight W(i,j):

[0032] E={(v i ,v j ) | W(i,j)>θ};

[0033] wherein θ is a set regulatory relationship threshold, if W(i,j) exceeds the threshold, an edge (v i ,v j ) is added in the cervical cancer gene regulatory network to represent the regulatory relationship between the gene g i and the gene g j ;

[0034] S25. Representing the cervical cancer gene regulatory network in the form of an adjacency matrix A, wherein the matrix elements are defined as follows:

[0035]

[0036] wherein A(i,j) represents the regulatory relationship weight between the gene g i and the gene g j ;

[0037] S26. The initial cervical cancer gene regulatory network is represented as:

[0038] G=(V,E,A).

[0039] Optionally, the step S3 specifically comprises:

[0040] S31. Based on the initial cervical cancer gene regulatory network G, performing multi-scale expansion on the adjacency matrix A, introducing a dynamic weight normalization strategy, and constructing a weighted adjacency matrix A multi that is suitable for different scale feature extraction:

[0041]

[0042] wherein A s represents the s-order propagation matrix of the adjacency matrix, S is the maximum propagation order, and w s is the learned different scale propagation weight;

[0043] S32. Dynamically adjusting the receptive field of different scales by using adaptive convolution operation with variable scale kernel size, and defining a multi-scale convolution kernel parameter matrix to adapt to the feature propagation of different ranges at different scales:

[0044]

[0045] wherein H(l) is the gene feature matrix of the first layer, D multi is the multi-scale weighted degree matrix, A multi is the multi-scale weighted adjacency matrix, and σ(·) is a nonlinear activation function.

[0046] S33. Constructing local dynamic receptive field features based on different scale graph convolution calculations and global propagation features H global :

[0047]

[0048] wherein, and denote the local and global feature matrices, and denote the local and global weight parameters, is the global gene regulation propagation matrix, is the global degree matrix.

[0049] S34. Introducing a node-level adaptive attention mechanism to fuse the local dynamic receptive field features H local and the global propagation features H global , and calculating the attention weight λ i for each gene node v i for adaptively weighted fusion of features of two scales:

[0050]

[0051] wherein, [·∥·] represents a vector concatenation operation, W a is an adaptive attention weight matrix, L1 and L2 represent the number of layers of the local and global graph convolution networks respectively, and N(i) is the neighborhood node set of the node v i in the cervical cancer gene regulation network, λ i is calculated by Softmax normalization, so that the weight adapts to the importance of different gene nodes in the local and global features.

[0052] Optionally, the step S4 specifically comprises:

[0053] S41. Setting a Zebra optimization algorithm search population P, wherein each population individual Z i represents an improved multi-scale graph convolution network configuration, including key parameters for analyzing the cervical cancer gene regulation network:

[0054] Z i ={L, W multi , λ, α, β, γ, η};

[0055] Wherein, L is the number of improved multi-scale graph convolutional network layers, affects the depth information capture ability of cervical cancer gene regulation network, W multi is a multi-scale convolution kernel parameter matrix, controls the learning of different scale gene regulation patterns, and η is a learning rate.

[0056] S42. Calculate each population individual Z i The performance of the corresponding improved multi-scale graph convolutional network model on the cervical cancer gene regulation network analysis task, the fitness function is defined as follows:

[0057] F(Z i )=Loss(H fusion ,Y)+λ r ∥W multi ∥2;

[0058] Wherein, Loss(H fusion ,Y) reflects the error of the improved multi-scale graph convolutional network model in predicting the cervical cancer gene regulation relationship, λ r is a regularization coefficient, and ∥W multi ∥2 is an L2 regularization term of the convolution kernel weight.

[0059] S43. Simulate the cooperative behavior of zebra populations, use global search to explore the optimal improved multi-scale graph convolutional network parameters, and make the improved multi-scale graph convolutional network model analyze the local and global relationship in the cervical cancer gene regulation network:

[0060]

[0061] Wherein, is the individual with the optimal fitness in the current population, is a randomly selected individual in the population, and r1, r2 are random step factors of global search.

[0062] S44. On the basis of global search, locally optimize the optimal individual, and enhance the recognition ability of the improved multi-scale graph convolutional network model to the key driver genes of cervical cancer:

[0063]

[0064] Wherein, δ is the local search step, indicates the individual after optimization Corresponding fitness gradient.

[0065] S45. Termination condition judgment, if one of the following conditions is met in the optimization process, stop searching, otherwise continue optimization:

[0066]

[0067] Wherein, The current optimal individual obtained by global search, The optimal individual further optimized by local search, and ε is a set convergence threshold. When the optimized model has converged, further adjustment will not improve the accuracy of the analysis of the cervical cancer gene regulation network. T is the maximum number of iterations.

[0068] S46. Obtain the optimized improved multi-scale graph convolutional network model, and finally select the individual with the optimal fitness As the final optimized improved multi-scale graph convolutional network parameters.

[0069] Optionally, the step S5 specifically comprises:

[0070] S51. Perform gene regulation relationship reasoning based on the optimized improved multi-scale graph convolutional network model, and utilize the optimized multi-scale fusion feature H fusion And the optimized improved multi-scale graph convolutional network parameters Calculate the regulation relationship probability P(g i ,g j ) between the genes g i and g j .

[0071] S52. Generate a reasoning matrix R based on the gene regulation relationship probability P(g i ,g j ) obtained by reasoning:

[0072]

[0073] Where θ r is the probability threshold of gene regulation relationship determination. If P(g i ,g j ) > θ r , it is considered that the gene g i has a regulation relationship with the gene g j , otherwise, it is considered that there is no direct regulation relationship.

[0074] S53. Calculate the confidence score C(g i ,g j ) of the gene regulation relationship based on the reasoning matrix R and the known biological experimental data Y true :

[0075]

[0076] Where |R(i,j)∩Y true (i,j)| represents the number of matches between the inferred gene regulation relationship and the real experimental data, |R(i,j)| is the total number of inferred gene regulation relationships, and ε is a minimum value to prevent the denominator from being zero. C(g i ,gj The closer the value is to 1, the higher the reliability of the gene regulation relationship prediction.

[0077] S54. Confidence score C(g) based on gene regulatory relationships i ,g j Generate a high-confidence cervical cancer gene regulatory network:

[0078] G conf =(V,E) conf ,R conf );

[0079] Where V is the set of gene nodes, E conf For a set of high-confidence regulation relations, satisfying:

[0080] E conf ={(g i ,g j )∣C(g i ,g j )>θ c};

[0081] Where, θ c R is the confidence threshold. conf This is a high-confidence modulating relation weight matrix, containing all relationships that satisfy C(g) i ,g j )>θ c Gene regulatory relationships.

[0082] Optionally, step S7 specifically includes:

[0083] S71. From the high-confidence cervical cancer gene regulatory network G conf Calculate g for each gene i Keyness score S(g) in the network i The key score reflects the gene g. i Its central role in the cervical cancer gene regulatory network:

[0084]

[0085] Among them, N(g) i ) is gene g i The set of neighboring genes, R conf (i,j) represents the weight of the high-confidence gene regulatory relationship, C(g) i ,g j () represents the confidence score for gene regulatory relationships;

[0086] S72. Based on the critical score S(g) i All genes were sorted, and the top k genes with the highest scores were selected to form the set of key driver genes for cervical cancer, G.key ;

[0087] S73. Retrieve the cervical cancer key driver genes G key from the cancer-related gene database, and construct a gene annotation matrix A db :

[0088]

[0089] wherein A db (i,j) represents whether the gene g i in the database is marked as related to cervical cancer;

[0090] S74. Calculate the matching degree M(G key ,G db ) between the set of cervical cancer key driver genes G key and the set of cervical cancer-related genes G db in the public database:

[0091]

[0092] wherein |G key ∩G db | represents the number of intersection genes between the set of screened cervical cancer key driver genes G key and the set of cervical cancer-related genes G db in the public database, |G key | represents the total number of screened key genes, and the closer M(G key ,G db ) is to 1, the more biologically reliable the screened key genes are;

[0093] S75. Extract the data of cervical cancer patient samples and normal samples from the public database, and calculate the expression difference D(g i ) of the cervical cancer key driver gene g i in the cancer samples, wherein the greater the expression difference is, the more significant the expression change of the gene in the cancer samples is.

[0094] S76. Evaluate the biological reliability of the cervical cancer key driver genes based on the database verification results, and calculate the comprehensive reliability score C final (g i ) of the cervical cancer key driver genes:

[0095] C final (g i ) = w1S(g i ) + w2M(G key ,G db ) + w3D(g i ).

[0096] wherein w1, w2, w3 are weight coefficients of the criticality score, the database matching degree, and the expression change score, respectively.

[0097] S77. Based on the verification result, a cervical cancer key driver gene analysis report is generated to show the regulatory role, biological function and clinical significance of the key gene in cervical cancer, and the analysis of the cervical cancer gene regulatory network is completed.

[0098] The beneficial effects of the present application are:

[0099] (1) The present application constructs an improved multi-scale graph convolutional network. According to the topological structure characteristics of the cervical cancer gene regulatory network, different scale graph convolutional kernels are used to extract local gene interaction features and global gene regulatory patterns, and an adaptive attention mechanism is used for fusion, thereby improving the comprehensiveness and accuracy of gene regulatory network analysis. It can simultaneously learn local gene regulatory relationships and global gene regulatory patterns, making the structure of the regulatory network more reasonable.

[0100] (2) The present application introduces the zebra optimization algorithm into the training process of the improved multi-scale graph convolutional network. Through the cooperative strategy of global search and local optimization, the network structure, convolution kernel weight and hyperparameter are adaptively optimized, effectively solving the problem of unstable optimization effect of traditional multi-scale graph convolutional network model hyperparameters relying on manual adjustment. The zebra optimization algorithm uses the group cooperative mechanism to iteratively optimize the parameters, so that the network can maintain high generalization ability on different data sets.

[0101] (3) Based on the optimized improved multi-scale graph convolutional network model, the present application uses a confidence evaluation mechanism to screen the inferred gene regulatory relationships, and constructs a high-confidence gene regulatory network. Then, combined with the public cancer gene database, the biological verification of the key driver gene is carried out, which improves the reliability of the gene regulatory relationship. By calculating the criticality score of the gene in the regulatory network and combining the database matching degree to calculate its biological reliability, the cervical cancer key driver gene can be more accurately identified. BRIEF DESCRIPTION OF DRAWINGS

[0102] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0103] Figure 1 A flowchart of a cervical cancer gene regulatory network analysis method based on multi-scale graph convolution is provided. DETAILED DESCRIPTION

[0104] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0105] Reference Figure 1 A multi-scale graph convolution-based cervical cancer gene regulation network analysis method includes the following steps:

[0106] S1. Collect gene expression data, transcription factor binding data, and epigenetic data related to cervical cancer, and preprocess them to construct a multi-modal biological data set;

[0107] S2. Construct an initial cervical cancer gene regulation network based on the multi-modal biological data set;

[0108] S3. Use an improved multi-scale graph convolution network to extract features from the initial cervical cancer gene regulation network, and fuse the features of each scale to generate multi-scale fusion features;

[0109] S4. Take the multi-scale fusion features and the training parameters of the improved multi-scale graph convolution network as inputs, and use the Zebra optimization algorithm to globally and locally search and optimize the network structure, graph convolution kernel weights, and hyperparameters of the improved multi-scale graph convolution network, to obtain an optimized improved multi-scale graph convolution network model;

[0110] S5. Use the optimized improved multi-scale graph convolution network model to infer unknown gene regulation relationships in the cervical cancer gene regulation network, generate gene regulation relationship inference results, evaluate the confidence of the inference results, and generate a high-confidence cervical cancer gene regulation network;

[0111] S6. Verify the biological significance of the key driver genes in cervical cancer in combination with public databases to analyze the cervical cancer gene regulation network.

[0112] In this embodiment, step S1 specifically includes:

[0113] S11. Collect cervical cancer-related gene expression data X G , transcription factor binding data X T , and epigenetic data X E ;

[0114] S12. Process the gene expression data X G , transcription factor binding data X T , and epigenetic data X E uniformly to construct a multi-modal biological data set, and standardize the multi-modal biological data set to normalize its value range to the [0, 1] interval;

[0115] S13. Perform missing value imputation on the standardized multi-modal biological dataset to fill in missing values;

[0116] S14. Perform dimensionality reduction on the missing value imputed multi-modal biological dataset to extract main information features;

[0117] S15. Perform noise filtering on the dimensionally reduced multi-modal biological dataset;

[0118] S16. Combine gene expression data, transcription factor binding data, and epigenetic data, according to gene number g i Perform data matching to align features from different data sources and construct the final multi-modal biological dataset:

[0119]

[0120] In this embodiment, gene expression data is obtained by RNA sequencing, representing the expression level of genes in different samples, constituting a gene expression matrix X G Transcription factor binding data is obtained by transcription factor immunoprecipitation sequencing, representing the binding of a specific transcription factor to a target gene, constituting a transcription factor binding matrix X T Epigenetic data includes DNA methylation data and histone modification data, respectively from methylation chip sequencing and histone modification immunoprecipitation sequencing, used to represent the epigenetic regulation state of genes, constituting an epigenetic matrix X E .

[0121] In this embodiment, step S2 specifically includes:

[0122] S21. Construct a node set V of the cervical cancer gene regulation network according to the final multi-modal biological dataset, where each gene g i is a node v i , and the gene node set V is defined;

[0123] S22. Calculate the correlation between genes, and calculate the Pearson correlation coefficient R(i,j) between genes as a measure of gene expression correlation according to gene expression data

[0124]

[0125] wherein, represents the expression value of gene g i in sample k, is the average expression value of gene g i in all samples, and m is the total number of samples;

[0126] S23. Calculate the regulation relationship weight W(i,j) between genes, integrating transcription factor binding data​ and epigenetic data The gene regulation relationship weight is defined in a linear combination manner:

[0127]

[0128] Wherein, a, b, g are the weighted parameters of the influence of gene expression data, transcription factor binding data, and epigenetic regulation data on gene regulation relationship, respectively.

[0129] S24. Constructing the regulation relationship set E of the cervical cancer gene regulation network according to the gene regulation relationship weight W(i,j):

[0130] E={(v i ,v j ) | W(i,j)>0};

[0131] Wherein, 0 is the set regulation relationship threshold, if W(i,j) exceeds the threshold, the edge (v i ,v j ) is added in the cervical cancer gene regulation network to represent the regulation relationship between gene g i and gene g j ;

[0132] S25. Representing the cervical cancer gene regulation network in the form of adjacency matrix A, wherein the matrix element is defined as follows:

[0133]

[0134] Wherein, A(i,j) represents the regulation relationship weight between gene g i and gene g j ;

[0135] S26. The initial cervical cancer gene regulation network is represented as:

[0136] G=(V,E,A).

[0137] In this embodiment, step S3 specifically comprises:

[0138] S31. Based on the initial cervical cancer gene regulation network G, performing multi-scale expansion on the adjacency matrix A, introducing dynamic weight normalization strategy, and constructing the weighted adjacency matrix A multi suitable for different scale feature extraction:

[0139]

[0140] Wherein, A s represents the s-order propagation matrix of the adjacency matrix, S is the maximum propagation order, and w s is the learned different scale propagation weight.

[0141] S32. Adopting adaptive convolution operation with variable scale kernel size to dynamically adjust receptive field of different scales, defining multi-scale convolution kernel parameter matrix Make it adapt to different ranges of feature propagation at different scales:

[0142]

[0143] Where H (l) is the gene feature matrix of the lth layer, D multi is the multi-scale weighted degree matrix, A multi is the multi-scale weighted adjacency matrix, and σ(·) is a nonlinear activation function.

[0144] S33. Construct local dynamic receptive field features based on graph convolution calculation at different scales And global propagation features H global :

[0145]

[0146] Where, And represent the local and global feature matrices, And represent the local and global weight parameters, is the global gene regulation propagation matrix, is the global degree matrix.

[0147] S34. Introducing node-level adaptive attention mechanism, fusing local dynamic receptive field features H local And global propagation features H global Calculate attention weight λ i For each gene node v i , used to adaptively weight and fuse features of two scales:

[0148]

[0149] Where [·∥·] represents vector concatenation operation, W a is the adaptive attention weight matrix, L1 and L2 represent the number of layers of local and global graph convolution networks respectively, and N(i) is the neighborhood node set of node v i In the cervical cancer gene regulation network, λ i is calculated by Softmax normalization, so that its weight adapts to the importance of different gene nodes in local and global features.

[0150] In this embodiment, step S4 specifically comprises:

[0151] S41. Set the zebra optimization algorithm search population P, where each population individual Z i Represent an improved multi-scale graph convolutional network configuration, including the key parameters for analyzing the cervical cancer gene regulatory network:

[0152] Z i ={L, W multi , λ, α, β, γ, η};

[0153] Where L is the number of layers of the improved multi-scale graph convolutional network, affecting the depth information capture ability of the cervical cancer gene regulatory network, W multi is the multi-scale convolution kernel parameter matrix, controlling the learning of different scale gene regulatory patterns, and η is the learning rate;

[0154] S42. Calculate the performance of each population individual Z i The corresponding improved multi-scale graph convolutional network model in the cervical cancer gene regulatory network analysis task, the fitness function is defined as follows:

[0155] F(Z i ) = Loss(H fusion , Y) + λ r ||W multi ||2;

[0156] Where Loss(H fusion , Y) reflects the error of the improved multi-scale graph convolutional network model in predicting the cervical cancer gene regulatory relationship, λ r is the regularization coefficient, and ||W multi ||2 is the L2 regularization term of the convolution kernel weight;

[0157] S43. Simulate the cooperative behavior of zebra population, use global search to explore the optimal improved multi-scale graph convolutional network parameters, make the improved multi-scale graph convolutional network model analyze the local and global relationship in the cervical cancer gene regulatory network:

[0158]

[0159] Where, is the individual with the best fitness in the current population, is a randomly selected individual in the population, and r1, r2 are random step factors for global search;

[0160] S44. On the basis of global search, locally optimize the optimal individual, enhance the recognition ability of the improved multi-scale graph convolutional network model to the key driver genes of cervical cancer:

[0161]

[0162] Where δ is the local search step, the optimized individual the corresponding fitness gradient

[0163] S45. Termination condition judgment, if one of the following conditions is met during optimization, stop searching, otherwise continue optimization:

[0164]

[0165] wherein, the current optimal individual obtained by global search, the optimal individual further optimized by local search, ∈ is the set convergence threshold, indicates that the optimized model has converged, further adjustment will not improve the accuracy of cervical cancer gene regulation network analysis, T is the maximum number of iterations;

[0166] S46. Obtain the optimized improved multi-scale graph convolutional network model, and finally select the individual with the optimal fitness as the final optimized improved multi-scale graph convolutional network parameters.

[0167] In this embodiment, step S5 specifically comprises:

[0168] S51. Based on the optimized improved multi-scale graph convolutional network model, infer the gene regulation relationship, and use the optimized multi-scale fusion feature H fusion and the optimized improved multi-scale graph convolutional network parameters to calculate the regulation relationship probability P(g i , g j ) between gene g i and gene g j ;

[0169] S52. Based on the inferred gene regulation relationship probability P(g i , g j ), generate an inference matrix R:

[0170]

[0171] wherein, θ r is the probability threshold of gene regulation relationship determination, if P(g i , g j )> θ r , it is considered that gene g i has a regulation relationship with gene g j , otherwise it is considered that there is no direct regulation relationship;

[0172] S53. Based on the inference matrix R and the known biological experimental data Y true , calculate the confidence score C(g i , g j ) of the gene regulation relationship:

[0173]

[0174] wherein |R(i,j)∩Y true | represents the matching number between the inferred gene regulation relationship and the real experimental data, |R(i,j)| is the total number of the inferred gene regulation relationship, ∈ is a minimum value for preventing the denominator from being zero, C(g i ,g j ) is closer to 1, the higher the credibility of the gene regulation relationship prediction is.

[0175] S54. generating a high-confidence cervical cancer gene regulation network based on the confidence score C(g i ,g j ):

[0176] G conf =(V,E conf ,R conf );

[0177] wherein V is a gene node set, E conf is a high-confidence regulation relationship set, and satisfies:

[0178] E conf ={(g i ,g j )∣C(g i ,g j )>θ c};

[0179] wherein θ c is a confidence threshold, R conf is a high-confidence regulation relationship weight matrix, and contains all gene regulation relationships satisfying C(g i ,g j )>θ c .

[0180] In the embodiment, step S7 specifically comprises:

[0181] S71. calculating a key score S(g conf ) of each gene g i in the network from the high-confidence cervical cancer gene regulation network G i , the key score reflecting the core role of the gene g i in the cervical cancer gene regulation network:

[0182]

[0183] wherein N(g i ) is a gene g ineighborhood gene set R conf (i,j) is the weight of high-confidence gene regulatory relationship, C(g i ,g j ) is the confidence score of gene regulatory relationship;

[0184] S72. Rank all genes according to the criticality score S(g i ), and select the top k genes with the highest scores to form the cervical cancer critical driver gene set G key .

[0185] S73. Retrieve the biological characteristics of the cervical cancer critical driver genes G key from the cancer-related gene database, and construct the gene annotation matrix A db :

[0186]

[0187] where A db (i,j) represents whether the gene g i in the database is labeled as related to cervical cancer;

[0188] S74. Calculate the matching degree M(G key ,G db ) between the cervical cancer critical driver gene set G key and the cervical cancer-related gene set G db in the public database:

[0189]

[0190] where |G key ∩G db | represents the number of intersection genes between the screened cervical cancer critical driver gene set G key and the cervical cancer-related gene set G db in the public database, |G key | represents the total number of screened critical genes, and the closer the matching degree M(G key ,G db ) is to 1, the more reliable the screened critical genes are biologically;

[0191] S75. Extract the data of cervical cancer patient samples and normal samples from the public database, and calculate the expression difference D(g i ) of the cervical cancer critical driver gene g i in the cancer samples. The greater the expression difference, the more significant the expression change of the gene in the cancer samples.

[0192] S76. Based on the database verification result, the biological reliability of the key driver genes of cervical cancer is evaluated, and a comprehensive reliability score C of the key driver genes of cervical cancer is calculated final (g i ):

[0193] C final (g i )=w1S(g i )+w2M(G key ,G db )+w3D(g i );

[0194] wherein w1, w2, and w3 are weight coefficients of the key score, the database matching degree, and the expression change score, respectively.

[0195] S77. Based on the verification result, a key driver gene analysis report of cervical cancer is generated, the regulatory role, biological function, and clinical significance of the key genes in cervical cancer are displayed, and the analysis of the gene regulatory network of cervical cancer is completed.

[0196] Example 1

[0197] On April 15, 2024, the research team conducted gene regulatory network analysis on a batch of cervical cancer patient samples, aiming to screen out key driver genes related to the occurrence, progression, and treatment resistance of cervical cancer. The research data included 960 cervical cancer patient samples and 320 normal cervical tissue samples.

[0198] In recent years, although HPV infection has been identified as the main cause of cervical cancer, there are still significant differences in the progression speed, drug resistance, and prognosis of cancer in different patients, which indicates that in addition to HPV infection, other abnormal gene regulatory networks may play a key role in the occurrence of cervical cancer. In this embodiment, the cervical cancer gene regulatory network analysis method based on multi-scale graph convolution combined with Zebra optimization algorithm of the present application is used to conduct in-depth analysis on this batch of data, and the biological role of the key genes is verified.

[0199] In the experiment, the research team first extracted cervical cancer gene expression data, transcription factor binding data, and epigenetic modification data from the database, supplemented with additional cervical cancer patient data, and standardized all the data to construct the final data set. The specific data are as follows:

[0200] Cervical cancer samples: 960 cases;

[0201] Normal cervical tissue samples: 320 cases;

[0202] Gene expression data: covering 16800 genes, after sample normalization, the gene expression level is represented by FPKM value;

[0203] Transcription factor binding data: covering 270 known cervical cancer related transcription factors, used to identify regulatory relationships between genes;

[0204] DNA methylation data: containing 486500 CpG sites, used to infer epigenetic regulation of gene expression.

[0205] The research team uses the method of the application to construct a cervical cancer gene regulatory network, and uses an improved multi-scale graph convolutional network for modeling. In traditional gene co-expression analysis methods, the regulatory relationship between genes is usually based on Pearson correlation or Spearman rank correlation coefficient. However, in the analysis of the regulatory relationship of the key genes E6-E7 (HPV virus genes)-CCND1 of cervical cancer, the correlation coefficient of the traditional method fluctuates greatly, resulting in some important regulatory pathways being ignored. The method of the application successfully and stably learns the E6 / E7->CCND1->CDK4 / 6 regulatory pathway through multi-scale graph convolution modeling, and improves the stability of the regulatory network.

[0206] The method of the application finally constructs a cervical cancer gene regulatory network containing 8720 gene nodes and 47560 gene regulatory relationships, and optimizes the hyperparameters of the multi-scale GCN model through the Zebra optimization algorithm to improve the generalization ability of the model.

[0207] After the improved multi-scale graph convolutional network is constructed, the research team uses the method of the application to screen the key driver genes of cervical cancer. In the traditional gene expression difference analysis method (log2FC analysis), the Top10 key genes obtained are shown in Table 1:

[0208] Table 1 Top 10 key genes screened by the traditional method

[0209] Ranking Gene Traditional method (log2FC value) 1 TP53 3.8 2 CDKN2A 3.5 3 CCND1 3.1 4 MYC 2.7 5 EGFR 2.3 6 RB1 2.0 7 NOTCH1 1.8 8 KRAS 1.6 9 TERT 1.5 10 AKT1 1.3

[0210] However, during database cross-validation, the research team found that RB1 and KRAS did not show stable expression differences (P value> 0.05) in multiple cervical cancer datasets, indicating that the traditional method may have screened false positive genes.

[0211] Using the method of the application, through multi-scale feature learning + high confidence gene regulatory relationship reasoning, the Top10 key driver genes of cervical cancer are re-screened, as shown in Table 2:

[0212] Table 2 Top 10 key genes screened by the method of the application

[0213]

[0214]

[0215] In order to verify the role of the genes screened by the method in cervical cancer, the research team carried out expression analysis on the 10 genes in the cervical cancer samples of TCGA and GEO databases, and found that the expression level change was more significant than the result of the traditional method, P value < 0.001, and the stable expression was maintained in multiple independent data sets.

[0216] In order to verify the advantages of the application, the research team carried out a comparative experiment with the traditional method, and the experimental results are shown in Table 3:

[0217] Table 3 Comparative experimental results of different methods in gene regulatory network analysis

[0218]

[0219] The experimental results show that the method of the application is obviously superior to the traditional method in the accuracy of key gene screening and the reliability of gene regulatory relationship prediction, and provides a more efficient analysis tool for cervical cancer gene regulation research.

[0220] The improved multi-scale graph convolutional network is constructed, different scale graph convolutional kernels are used to extract local gene interaction features and global gene regulation modes according to the topological structure characteristics of the cervical cancer gene regulation network, and the self-adaptive attention mechanism is used for fusion, so that the comprehensiveness and accuracy of the gene regulation network analysis are improved, the local gene regulation relationship (transcription factor directly acting on target gene) and the global gene regulation mode (key driver gene affecting the whole gene network) can be learned at the same time, and the structure of the regulation network is more reasonable.

[0221] The zebra optimization algorithm is introduced into the training process of the improved multi-scale graph convolutional network, the network structure, convolution kernel weight and hyperparameter are adaptively optimized through the cooperative strategy of global search and local optimization, the problem of unstable optimization effect of the traditional multi-scale graph convolutional network model is effectively solved, the zebra optimization algorithm uses the group cooperative mechanism to iteratively optimize the parameters, so that the network can maintain high generalization ability on different data sets.

[0222] Based on the optimized improved multi-scale graph convolutional network model, the confidence evaluation mechanism is used to screen the gene regulation relationship obtained by reasoning, and the high-confidence gene regulation network is constructed, then the biological verification of the key driver gene is combined with the public cancer gene database, the reliability of the gene regulation relationship is improved, the key score of the gene in the regulation network is calculated, and the biological credibility is calculated by combining the database matching degree, so that the cervical cancer key driver gene can be more accurately identified.

[0223] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for analyzing cervical cancer gene regulatory network based on multi-scale graph convolution, characterized in that, The method comprises the following steps: S1. Collecting gene expression data, transcription factor binding data and epigenetic data related to cervical cancer, and preprocessing to construct a multi-modal biological data set; S2. Constructing an initial cervical cancer gene regulation network based on the multi-modal biological data set; The step S2 specifically comprises: S21. Construct a node set V of the cervical cancer gene regulatory network according to the final multi-modal biological dataset, wherein each gene as a node , define a gene node set V; S22. Calculate the correlation between genes, based on gene expression data The Pearson correlation coefficient R(i, j) between genes is calculated as a measure of gene expression correlation: ； in, Indicates gene The expression value in sample k, For genes The average expression value across all samples, where m is the total number of samples; S23. Calculate the regulatory relationship weight W(i, j) between genes, fuse transcription factor binding data and epigenetic data , define the gene regulatory relationship weight in a linear combination way: ； wherein, , , are the weighted parameters of the influence of gene expression data, transcription factor binding data, epigenetic regulation data on the gene regulatory relationship, respectively; S24. Constructing a regulation relationship set E of the cervical cancer gene regulation network according to the gene regulation relationship weight W(i, j): ； wherein, is a set threshold of the regulatory relationship, and if W(i, j) exceeds the threshold, an edge is added in the cervical cancer gene regulatory network to represent the regulatory relationship between the gene and the gene . S25. Representing the cervical cancer gene regulation network in the form of an adjacency matrix A, wherein the matrix elements are defined as follows: ； wherein A(i, j) represents the weight of the regulatory relationship between genes and genes ​ S26. The initial cervical cancer gene regulation network is represented as: G = (V, E, A); S3. Extracting features from the initial cervical cancer gene regulation network using an improved multi-scale graph convolutional network, and fusing the features of each scale to generate multi-scale fusion features; The step S3 specifically comprises: S31. Based on the initial cervical cancer gene regulation network G, the multi-scale expansion is performed on the adjacency matrix A, a dynamic weight normalization strategy is introduced, and a weighted adjacency matrix suitable for feature extraction of different scales is constructed : ； wherein, denotes an s-th order propagation matrix of the adjacency matrix, S is a maximum propagation order, is a learned different scale propagation weight; S32. Adopting adaptive convolution operation with variable scale kernel size to dynamically adjust receptive field of different scales, defining multi-scale convolution kernel parameter matrix to adapt to different ranges of feature propagation at different scales: ； wherein, is a gene feature matrix for the l-th layer, is a multi-scale weighted degree matrix, is a multi-scale weighted adjacency matrix, is a non-linear activation function; S33. Constructing local dynamic receptive field features based on different scale graph convolutional computation and global propagation features : ； ； in, and These represent the local and global feature matrices, respectively. and These represent the local and global weight parameters, respectively. This is a global gene regulation and propagation matrix. S34. Introduce a node-level adaptive attention mechanism to incorporate local dynamic receptive field features. and global propagation characteristics Perform fusion on each gene node. Calculate attention weights Used for adaptive weighted fusion of features at two scales: ； ； wherein, denotes a vector concatenation operation, is an adaptive attention weight matrix, and denote the number of local and global graph convolutional network layers, respectively, is a node a set of neighborhood nodes in the cervical cancer gene regulatory network, is computed by Softmax normalization, so that its weight adapts to the importance of different gene nodes in local and global features; S4. Taking the multi-scale fusion features and the training parameters of the improved multi-scale graph convolutional network as inputs, using the Zebra Optimization Algorithm to perform global and local search optimization on the network structure, graph convolution kernel weight and hyperparameters of the improved multi-scale graph convolutional network, and obtaining an optimized improved multi-scale graph convolutional network model; S5. Using the optimized improved multi-scale graph convolutional network model to infer unknown gene regulation relationships in the cervical cancer gene regulation network, generating gene regulation relationship inference results, and performing confidence evaluation on the inference results, and generating a high-confidence cervical cancer gene regulation network; S6. Verifying the biological significance of the key driver genes of cervical cancer in cervical cancer in combination with public databases, and realizing the analysis of the cervical cancer gene regulation network.

2. The method of claim 1, wherein the method is based on multi-scale graph convolutional network for cervical cancer gene regulatory network analysis. The step S1 specifically comprises: S11. Collecting gene expression data related to cervical cancer , transcription factor binding data and epigenetic data ; S12. Unify the gene expression data , transcription factor binding data and epigenetic data , construct a multi-modal biological dataset, normalize the multi-modal biological dataset so that its value range is normalized to the interval [0, 1]; S13. Completing the missing values of the standardized multi-modal biological data set to fill in the missing values; S14. Dimensionality reduction on the multi-modal biological data set after missing value completion to extract main information features; S15. Noise filtering on the multi-modal biological data set after dimensionality reduction; S16. Combining gene expression data, transcription factor binding data and epigenetic data, according to gene number Data matching was performed to align features from different data sources to construct the final multi-modal biological dataset: 。 3. The method of claim 1, wherein the method is based on multi-scale graph convolutional network for cervical cancer gene regulatory network analysis. The gene expression data is obtained by RNA sequencing, representing the expression level of genes in different samples, constituting a gene expression matrix The transcription factor binding data is obtained by transcription factor immunoprecipitation sequencing, representing the binding of a specific transcription factor to a target gene, constituting a transcription factor binding matrix The epigenetic data includes DNA methylation data and histone modification data, respectively from methylation chip sequencing and histone modification immunoprecipitation sequencing, for representing the epigenetic regulation state of genes, constituting an epigenetic matrix .

4. The method of claim 1, wherein the method is based on multi-scale graph convolutional network for cervical cancer gene regulatory network analysis. The step S4 specifically comprises: S41. Set the zebra optimization algorithm search population P, wherein each population individual Represent an improved multi-scale graph convolutional network configuration, including key parameters for analyzing cervical cancer gene regulatory network: ； Wherein, L is the number of improved multi-scale graph convolutional network layers, which affects the depth information capturing ability of the cervical cancer gene regulation network, is a multi-scale convolution kernel parameter matrix, which controls the learning of different scale gene regulation patterns, is a learning rate; S42. Calculate each population individual The performance of the corresponding improved multi-scale graph convolutional network model on the cervical cancer gene regulation network analysis task, the fitness function is defined as follows: ； wherein, reflecting the error of the improved multi-scale graph convolutional network model in predicting the gene regulation relationship of cervical cancer, is a regularization coefficient, is an L2 regularization term of the convolution kernel weight; S43. Simulating the cooperative behavior of zebra populations, using global search to explore the optimal improved multi-scale graph convolutional network parameters, so that the improved multi-scale graph convolutional network model analyzes the local and global relationships in the cervical cancer gene regulation network: ； wherein, is the individual with the best fitness in the current population, is a randomly selected individual from the population, , is a random step factor for the global search. S44. On the basis of global search, locally optimizing the optimal individual to enhance the recognition ability of the improved multi-scale graph convolutional network model to the key driver genes of cervical cancer: ； wherein, is a local search step size, denotes the optimized individual corresponding fitness gradient; S45. Termination condition judgment, if one of the following conditions is met in the optimization process, stop searching, otherwise continue optimization: ； wherein, is the current optimal individual obtained by global search, is the optimal individual further optimized by local search, is the set convergence threshold, indicating that the optimized model has converged, and further adjustment will not improve the accuracy of the analysis of the cervical cancer gene regulation network, and T is the maximum number of iterations. S46. Obtain the optimized improved multi-scale graph convolutional network model, and finally select the individual with the optimal fitness as the final optimized improved multi-scale graph convolutional network parameters.

5. The method of claim 1, wherein the method is based on multi-scale graph convolutional network for cervical cancer gene regulatory network analysis. The step S5 specifically comprises: S51. performing gene regulation relationship reasoning based on the improved multi-scale graph convolution network model after optimization, and utilizing the multi-scale fusion features after optimization and the improved multi-scale graph convolution network parameters after optimization to calculate the regulation relationship probability between and the genes . ; S52. Obtain the gene regulation relationship probability based on the inference Generate the inference matrix R: ； wherein, is a probability threshold for determining the gene regulation relationship, if , it is considered that the gene has a regulation relationship with the gene , otherwise, it is considered that there is no direct regulation relationship; S53. Based on the inference matrix R and known biological experimental data Computing a confidence score for a gene regulatory relationship : ； wherein, represents the number of matches between the inferred genetic regulatory relationships and the real experimental data, is the total number of inferred genetic regulatory relationships, is to prevent the denominator from being a minimum value of zero, the closer to 1, the higher the credibility of the prediction of the genetic regulatory relationship; S54. Confidence score based on gene regulatory relationship Generating a high-confidence cervical cancer gene regulatory network: ； wherein V is a set of gene nodes, is a set of high-confidence regulatory relationships, satisfying: ； wherein, is a confidence threshold, is a high confidence regulatory relationship weight matrix containing all gene regulatory relationships that satisfy .

6. The method of claim 5, wherein the method is based on multi-scale graph convolutional analysis of cervical cancer gene regulatory network. The step S7 specifically comprises: S71. From the high-confidence cervical cancer gene regulatory network calculate the keyness score of each gene in the network The keyness score reflects the key role of the gene in the cervical cancer gene regulatory network: ； wherein, is a gene of a neighborhood gene set, is a weight for a high confidence gene regulatory relationship, is a confidence score for a gene regulatory relationship; S72. Based on key scoring All genes were sorted, and the top k genes with the highest scores were selected to form a set of key driver genes for cervical cancer. ; S73. Retrieving cervical cancer key driver genes in cancer related gene databases biological properties and construct a gene annotation matrix : ； wherein, indicates whether a gene is flagged as being associated with cervical cancer in the database; S74. Calculating the cervical cancer key driver gene set Matching degree between the cervical cancer related gene set in public database :​ ； wherein, represents the number of intersection genes between the set of key driver genes of cervical cancer screened out and the set of cervical cancer related genes in public database , represents the total number of key genes screened out, the matching degree The closer to 1, the more reliable the key genes screened out biologically; S75. Extract the data of cervical cancer patient samples and normal samples from public databases, calculate the cervical cancer key driver genes Expression difference in cancer samples The greater the expression difference, the more significant the expression change of the gene in the cancer sample; S76. Evaluate the biological credibility of the cervical cancer key driver genes based on the database verification results, and calculate the comprehensive credibility score of the cervical cancer key driver genes : ； wherein, , , respectively are the weight coefficients of the criticality score, the database matching degree, and the expression change score; S77. Generating a cervical cancer key driver gene analysis report based on the verification result, showing the regulation role, biological function and clinical significance of the key genes in cervical cancer, and completing the analysis of the cervical cancer gene regulation network.

Images