Brain region-gene network based evolutionary method and system for generative adversarial networks

Abstract

This invention relates to the field of artificial intelligence technology, and discloses a method and system for the evolution of brain region-gene networks based on generative adversarial networks, providing new technical guidance for the research, prediction, and prevention of intractable diseases. The method includes: obtaining the first and second brain region-gene networks corresponding to each sample before and after evolution; inputting the adjacency matrix and simplex structure information matrix of the first brain region-gene network before evolution into a generator for iterative updating of the adjacency matrix and simplex structure information matrix; a discriminator obtaining the adjacency matrix and simplex structure information matrix after the generator's iteration terminates, and the adjacency matrix and simplex structure information matrix of the second brain region-gene network after the corresponding sample evolution, then distinguishing between true and false input data, and performing backpropagation according to a loss function to update the learning matrix in the generator; and after training to convergence based on the dataset, the generator online detects the evolution results of the brain region-gene network in new sample data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an evolutionary method and system for brain region-gene networks based on generative adversarial networks. Background Technology

[0002] In existing brain region-gene networks, the processing of structural information is mostly limited to the connections between nodes.

[0003] However, during the development of a disease, there are often relatively stable combinations of higher-order functions that exist intermittently. Any such combination can influence other peripheral nodes, and the influence can be large or small. As time goes on, the superimposed influence will produce new changes. The scope of such changes includes, but is not limited to, causing the original higher-order functional combinations to grow stronger or shrink, or forming new, relatively stable higher-order functional combinations.

[0004] Currently, no existing large-scale medical models possess a mechanism capable of simulating the aforementioned evolutionary process. The development of such an evolutionary mechanism would be extremely valuable for researching, predicting, and preventing intractable diseases such as Alzheimer's. Summary of the Invention

[0005] The purpose of this invention is to disclose an evolutionary method and system for brain region-gene networks based on generative adversarial networks, providing new technical guidance for the research, prediction and prevention of intractable diseases such as Alzheimer's disease.

[0006] To achieve the above objectives, this invention discloses a brain region-gene network evolution method based on generative adversarial networks (GANs), wherein the GAN includes a generator and a discriminator; the method includes:

[0007] S1. Obtain the first and second brain regions-gene networks corresponding to each sample before and after evolution; the edges of the first and second brain regions-gene networks are used to connect two nodes whose feature correlation is greater than the first threshold; in all samples, the number of nodes and the feature source of each node are consistent.

[0008] S2. Detect all simplex sets of different dimensions in the first and second brain region-gene networks to obtain the simplex structure information matrix and the adjacency matrix between each simplex corresponding to the first and second brain region-gene networks respectively; among the different elements of the simplex structure information matrix, the structure information of zero-dimensional nodes is calculated based on degree centrality, and the structure information of one-dimensional edges and polygons of two or more dimensions is calculated based on multivariate mutual information.

[0009] S3. Input the adjacency matrix and simplex structure information matrix of the first brain region-gene network before evolution into the generator to perform iterative updates of the adjacency matrix and simplex structure information matrix, and then send the adjacency matrix and simplex structure information matrix after the iteration terminates to the discriminator; the specific process of a single iteration includes:

[0010] S31. Obtain the closed star-shaped neighborhood of each simplex based on the adjacency matrix before evolution or after the previous iteration. Then, update the structural information of each simplex based on the star-shaped closed adjacency matrix formed by each closed star-shaped neighborhood, the simplex structural information matrix before evolution or after the previous iteration, and the first type of learning matrix to obtain the intermediate state simplex structural information matrix. The first type of learning matrix is ​​used to allocate the interaction intensity of structural information between each simplex. The star-shaped closed adjacency matrix is ​​used to limit the scope of any simplex to the closed star-shaped neighborhood formed by that simplex.

[0011] S32. Select stable simplexes with structural information strength greater than the second threshold, and merge the stable simplexes with directly related structures into a supernode simplex complex, wherein the structural information strength of all nodes and edges in the supernode simplex complex is greater than the second threshold.

[0012] S33. Determine the external nodes that can be affected by the simple complex of each supernode based on the closed star neighborhood;

[0013] S34. Establish potential edges between all nodes inside the simple complex of each supernode and all external nodes within the influence range;

[0014] S35. Calculate the structural information of each potential edge based on the updated node structural information of the first type of learning matrix and the second type of learning matrix. The second type of learning matrix is ​​used to allocate the contribution relationship of each node's own structural information to the potential edge structural information.

[0015] S36. Compare the structural information of each potential edge with the third threshold, and determine the potential edge whose structural information is greater than the third threshold as a real edge and retain it; otherwise, delete it.

[0016] S37. Update the adjacency matrix and the simplex structure information matrix of the intermediate state according to the addition and deletion of each potential edge to obtain the adjacency matrix and simplex structure information matrix after the current iteration.

[0017] S4. The discriminator obtains the adjacency matrix and simplex structure information matrix after the generator iteration terminates, and the adjacency matrix and simplex structure information matrix of the second brain region-gene network after the corresponding sample evolution. Then, it distinguishes the authenticity of the two types of input data and performs backpropagation according to the loss function to update the learning matrix in the generator.

[0018] S5. After the generator and the discriminator are repeatedly trained on the dataset until the generative adversarial network converges, the trained generator is used to detect the evolution of the brain region-gene network in new sample data online.

[0019] Preferably, the method of the present invention further includes:

[0020] S6. The generator calculates the approximation of the adjacency matrix and simplex structure information matrix in the online detection results of the new sample with the standard adjacency matrix and standard simplex structure information matrix, respectively. Then, the two approximation calculation results are weighted to obtain the probability that the new sample evolves into the disease stage commonly corresponding to the standard adjacency matrix and the standard simplex structure information matrix.

[0021] Preferably, the simplex structure information matrix is ​​a diagonal matrix; the specific calculation formula for step S31 is:

[0022] ;

[0023] Among them, superscript Represents the current iteration number. Represents the state when the previous iteration was completed, when This represents the initial state corresponding to the first brain region-gene network. This represents the intermediate state of the current iteration process; Represents the simplex structure information matrix. For the first The first type of learning matrix in the next iteration has a dimension of ; This represents the total number of all simplexes in the first brain region-gene network. It is a matrix of all ones. A matrix whose diagonal elements are all 1s and all other elements are all 0s; "Represents the Hadama product operation," " represents multiplication; To determine the adjacency matrix based on the adjacency matrix at the end of the previous iteration The star-shaped closed adjacency matrix obtained from the simplex set;

[0024] The specific calculation formula for step S35 is as follows: ;in, express The only node set extracted The local matrix formed, To represent the supernode diffusion matrix of the external nodes that can be affected by the simple complex of each supernode, This indicates the transmission of structural information between nodes within a supernode; Represents the information matrix of the external nodes of the supernode; superscript It is the transpose operator; For the first The second type of learning matrix in the next iteration process; Indicates the first The matrix of structural information received by each potential edge during each iteration.

[0025] Preferably, the adjacency matrix between each simplex It is represented as:

[0026] ;

[0027] in, Indicates the first A simplex With the A simplex They are adjacent. Then it means and There is no adjacency relationship, and the adjacency relationship between simplexes of the same dimension is 0;

[0028] The dimension of the supernode diffusion matrix is , The number of nodes in the brain region-gene network; the first Elements of the node diffusion matrix during the next iteration Indicates the first The node and the first Each node has a potential edge; if no potential edge exists, then... .

[0029] Preferably, elements =1 indicates the first The simplex in the first... In a closed star-shaped neighborhood of a simplex, if the first... The simplex is not in the first... In a closed star-shaped neighborhood of a simplex, it is represented as =0.

[0030] To achieve the above objectives, the present invention also discloses a brain region-gene network evolution system based on generative adversarial networks, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the computer program.

[0031] The present invention has the following beneficial effects:

[0032] Generative adversarial networks (GANs) are relatively mature neural networks, and the detection of various simplexes and their relationships are based on edge-to-edge connections, which can be reliably implemented through algorithms. This invention, by redefining the initial structural information of each simplex, expands the process of updating the adjacency matrix and simplex structural information matrix during the generator's structural information update based on the learning matrix. This includes, but is not limited to, the innovative method of dynamically constructing, merging, and diffusing the structural information of supernode simplexes in a single iteration, as well as the inheritance and renewal relationships between adjacent iterations. This perfectly simulates the disease progression of Alzheimer's disease and other intractable diseases, ensuring the reliability of the evolution results. Furthermore, it provides new technical guidance for the research, prediction, and prevention of Alzheimer's disease and other intractable diseases.

[0033] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0035] Figure 1 This invention discloses a brain region-gene network evolution method based on generative adversarial networks.

[0036] Figure 2 This is a schematic diagram of the generator's single iteration process disclosed in an embodiment of the present invention.

[0037] Figure 3 This is a flowchart illustrating the principle of a single iteration of the generator as disclosed in an embodiment of the present invention. Detailed Implementation

[0038] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0039] Example 1

[0040] This embodiment discloses a brain region-gene network evolution method based on generative adversarial networks (GANs), wherein the GAN includes a generator and a discriminator; such as Figure 1 As shown, the method in this embodiment includes:

[0041] S1. Obtain the first and second brain region-gene networks corresponding to each sample before and after evolution; the edges of the first and second brain region-gene networks are used to connect two nodes whose feature correlation is greater than the first threshold; in all samples, the number of nodes and the feature source of each node are consistent.

[0042] In this step, the Pearson correlation coefficient can be used to calculate the correlation. For ease of calculation, the feature dimensions of each node in the brain region-gene network are consistent. Since the initial feature dimensions of the brain region and genes are different (the brain region has a dimension of 87, while the gene dimension is 70 after processing), this embodiment truncates the brain region dimension to 70 during preprocessing to align with the gene dimension. The 87 features of the brain region are ordered by information importance; the first 70 dimensions cover the core information, and the correlation between the truncated data distribution and the gene features does not change significantly, thus avoiding information bias due to the loss of effective information.

[0043] S2. Detect all simplex sets of different dimensions in the first and second brain region-gene networks to obtain the simplex structure information matrix and the adjacency matrix between each simplex corresponding to the first and second brain region-gene networks respectively; among the different elements of the simplex structure information matrix, the structure information of zero-dimensional nodes is calculated based on degree centrality, and the structure information of one-dimensional edges and polygons of two or more dimensions is calculated based on multivariate mutual information.

[0044] The simplex is a basic geometric unit in algebraic topology, which is a minimal convex set enclosed by a number of vertices. Taking a k-dimensional simplex as an example, it is defined as consisting of k+1 vertices, and the line connecting any two points is part of the topological structure of the k-dimensional simplex.

[0045] In this step, the structural information is a scalar; taking the first brain region-gene network as an example, the first... A simplex Initial structural information The specific calculation formula can be:

[0046] ;

[0047] in, The calculation of the degree of a node. Indicates taking the current The number of elements in a simplex set of the same dimension. Used to output the dimensions of the simplex, when hour, Since it is a zero-dimensional simplex (i.e., nodes), its influence in the complex structure is measured by the degree centrality metric. In higher-order structures, the relationships between nodes are no longer simple first-order connections, and traditional metrics such as degree centrality are insufficient to characterize their role. Therefore, this step introduces Multivariate Mutual Information (MMI). MMI can capture redundant or unique information interactions in higher-order structures and better represent the influence of the simplex in the higher dimensions of complex structures. yes non-empty subset Represents the number of non-empty subsets. Representative subset Calculating the joint entropy on, for example: if subset ,but , The calculation of the joint probability distribution, where, This represents three distinct nodes.

[0048] Therefore, in this embodiment, the structural information of polygons above two dimensions can be extended and taken into account. These polygons include triangles, quadrilaterals and other types of polygons, and their specific distribution varies with each sample.

[0049] In this step, when the structure expands from nodes to higher dimensions (such as edges, triangles, or even higher-order simplexes), its role is no longer solely determined by the "number of connections." Simple dependency-related indicators inevitably lead to a degradation in information characterization. Mutual information does not characterize the connection strength itself, but rather the information relationship pattern between nodes within the simplex. Taking edges as an example, suppose there are two edges with the same degree. One connects two highly correlated nodes with highly repetitive information, while the other connects two functionally complementary nodes with relatively independent information. Although they are topologically equivalent, the latter introduces more discriminative joint constraints at the information level, resulting in higher mutual information.

[0050] Therefore, the mutual information in this step can be intuitively understood as: measuring whether the information between nodes within the simplex is redundant or whether it generates new information in a cooperative manner that cannot be interpreted by any single node alone. When cooperative information dominates, the influence of the corresponding simplex in the complex structure also increases, thus more accurately reflecting its actual role in higher-order structures.

[0051] In this step, the adjacency matrix between each simplex It can be represented as:

[0052] ;

[0053] in, Indicates the first A simplex With the A simplex They are adjacent. Then it means and There is no adjacency relationship, and the adjacency relationship between simplexes of the same dimension is 0.

[0054] S3. Input the adjacency matrix and simplex structure information matrix of the first brain region-gene network before evolution into the generator to update and iterate the adjacency matrix and simplex structure information matrix, and then send the adjacency matrix and simplex structure information matrix after the iteration terminates to the discriminator.

[0055] like Figure 2 As shown, the specific process of a single iteration includes:

[0056] S31. Obtain the closed star-shaped neighborhood of each simplex based on the adjacency matrix before evolution or after the previous iteration. Then, update the structural information of each simplex based on the star-shaped closed adjacency matrix formed by each closed star-shaped neighborhood, the simplex structural information matrix before evolution or after the previous iteration, and the first type of learning matrix to obtain the intermediate state simplex structural information matrix. The first type of learning matrix is ​​used to allocate the interaction intensity of structural information between each simplex. The star-shaped closed adjacency matrix is ​​used to limit the scope of any simplex to the closed star-shaped neighborhood formed by that simplex.

[0057] In this step, the term "closed star neighborhood" is a common industry term and will not be elaborated upon.

[0058] Preferably, the simplex structure information matrix is ​​a diagonal matrix; the specific calculation formula for this step is:

[0059] ;

[0060] Among them, superscript Represents the current iteration number. Represents the state when the previous iteration was completed, when This represents the initial state corresponding to the first brain region-gene network. This represents the intermediate state of the current iteration process; Represents the simplex structure information matrix. For the first The first type of learning matrix in the next iteration has a dimension of ; This represents the total number of all simplexes in the first brain region-gene network. It is a matrix of all ones. A matrix whose diagonal elements are all 1s and all other elements are all 0s; "Represents the Hadama product operation," " represents multiplication; To determine the adjacency matrix based on the adjacency matrix at the end of the previous iteration The star-shaped closed adjacency matrix obtained from the simplex set.

[0061] Preferably, elements =1 indicates the first The simplex in the first... In a closed star-shaped neighborhood of a simplex, if the first... The simplex is not in the first... In a closed star-shaped neighborhood of a simplex, it is represented as =0.

[0062] The specific calculation formula described above in this step aims to achieve information propagation for each element within a star-shaped closed neighborhood. Then, the previously obtained structural information is added to the structural information received during the information propagation process to obtain the evolved and updated structural information for all elements. This structural information is stored in a structural information matrix I.

[0063] Since I is a diagonal matrix, the first value on the diagonal represents the structural information of the first element. D is a matrix with all elements equal to 1, that is, a matrix of all 1s. By multiplying D and I, we can obtain an intermediate matrix for calculation. The first row of this intermediate matrix stores the structural information of the first element in each block, and so on.

[0064] The weights used to learn the propagation of information among elements are essentially the strength of the interaction of structural information between simplexes. We can obtain the structural information that an element can pass to other elements, denoted as... The result is Y. Then the first row and second column of matrix Y represent the structural information that element 1 can pass to element 2.

[0065] Since the range of information transmitted by elements is unknown during the processing of Y, a star-shaped closed adjacency matrix U is introduced. U records the star-shaped closed neighborhood of each element. Multiplying U by Y yields the change in structural information for each element. Let the result of multiplying U by Y be matrix Q. The values ​​of each row in matrix Q are different, but each small block in the same row has the same value. For example, each block in the first row represents the structural information received by the first element in this evolution, and so on.

[0066] Furthermore, since matrix I is a diagonal matrix, in order to transform the received structural information Q into a diagonal matrix, a matrix E with all diagonal lines equal to 1 is introduced to transform Q into a diagonal matrix. Then, this matrix is ​​added to E, essentially adding the structural information from before the elements to the received structural information, thus obtaining the updated structural information matrix I. However, this structural information matrix I will change due to potential edge additions and deletions in subsequent steps, so it still only represents the simplex structural information matrix in an intermediate state during this iteration. Therefore, it is represented as... .

[0067] S32. Select stable simplexes with structural information strength greater than the second threshold, and merge the stable simplexes with directly related structures into a supernode simplex complex, wherein the structural information strength of all nodes and edges in the supernode simplex complex is greater than the second threshold.

[0068] In this embodiment, a simplex complex is typically a combination of multiple simplexes. In special cases, if a single stable node simplex is isolated, it also belongs to the supernode simplex complex. The supernode simplex complex in this step is usually a combination of multiple adjacent or contained stable simplexes (in other words, a topological structure formed by combining multiple simplexes according to specific rules can be called a simplex complex), but in special cases, it can also be a single stable node.

[0069] S33. Determine the external nodes that can be affected by the simple complex of each supernode based on the closed star neighborhood.

[0070] S34. Establish potential edges between all nodes inside the simple complex of each supernode and all external nodes within the influence range.

[0071] S35. Calculate the structural information of each potential edge based on the updated node structural information of the first type of learning matrix and the second type of learning matrix. The second type of learning matrix is ​​used to allocate the contribution relationship between the structural information of each node and the structural information of the potential edge.

[0072] The specific calculation formula for this step can be: ;in, express The only node set extracted The local matrix formed, To represent the supernode diffusion matrix of the external nodes that can be affected by the simple complex of each supernode, This indicates the transmission of structural information between nodes within a supernode; Represents the information matrix of the external nodes of the supernode; superscript It is the transpose operator; For the first The second type of learning matrix in the next iteration process; Indicates the first The matrix of structural information received by each potential edge during each iteration.

[0073] In this step, the dimension of the supernode diffusion matrix is... , The number of nodes in the brain region-gene network; the first Elements of the node diffusion matrix during the next iteration Indicates the first The node and the first Each node has a potential edge; if no potential edge exists, then... .

[0074] S36. Compare the structural information of each potential edge with the third threshold, and determine the potential edge whose structural information is greater than the third threshold as a real edge and retain it; otherwise, delete it.

[0075] S37. Update the adjacency matrix and the simplex structure information matrix of the intermediate state according to the addition and deletion of each potential edge, and obtain the adjacency matrix and simplex structure information matrix after the current iteration.

[0076] In summary, the changes that occur in step S3 during the iteration process are twofold: first, the structural information of the elements in the complex; and second, the topological structure of the complex, namely, changes in topological properties such as adjacency relationships and star-shaped closed neighborhoods.

[0077] For ease of public understanding, please refer to Figure 3 The following is a further explanation of the above iterative process:

[0078] In brain region-gene simplex complexes (combinations of multiple simplexes, not elaborated further), simplexes propagate their structural information to all neighboring simplexes to update their structural information. Closures formed by simplexes with higher structural information create supernodes, achieving "lossless compression" of key structures and thus preserving relatively "stable" higher-order functional combinations during disease progression. Secondly, supernodes and their neighboring external nodes propagate structural information to the "potential edges" in the complex to filter out "real edges," thereby reconstructing the simplex and completing one simplex evolution. This simplex evolution process undergoes multiple iterations to fully capture the complex interaction patterns between different levels of biological characteristics during disease progression.

[0079] exist Figure 3 In the diagram illustrating the diffusion of structural information, the closed star-shaped neighborhood of the central node is circled. A supernode simplex can be considered a single stable point during the folding of its sub-complexes. There are two supernode simplexes (i.e., the local structures enclosed by closed arcs) in the diagram. The edges connecting them do not meet the stability condition and are therefore split into two. If the edges connecting the two supernode simplexes also meet the stability condition greater than the second threshold, then the two supernode simplexes need to be further merged into one. Similarly, during the unfolding of a supernode, any node originally connected to any node within the supernode simplex by an edge belongs to the external nodes affected by the corresponding supernode simplex, as determined in step S33 based on the closed star-shaped neighborhood. During the unfolding of a supernode, a small number of nodes fall within the influence range of multiple supernode simplexes, but the algorithm only considers edges and does not update the structural information of the affected external nodes, greatly reducing the processing complexity.

[0080] In other words, the essence of the generator iteration in this embodiment is to design a sub-complex folding mechanism to fold closely related simplexes into supernodes during the evolution process, in order to preserve highly synergistic functional combinations during disease development. A supernode unfolding mechanism is also designed to reconstruct the simplexes, thereby simulating changes in various functions and lesion areas during disease development. This evolutionary process of sub-complex folding and supernode unfolding continues multiple times. Each time the evolution progresses to the next step, the sub-complexes folded into supernodes will be inconsistent. This is also a manifestation of non-uniformity, meaning that relatively stable parts are not permanent but temporarily stable.

[0081] In this step, the number of iterations can be flexibly set according to the different disease stages inferred from different datasets, typically ranging from 100 to 150. Furthermore, the second and third thresholds remain consistent across different iteration processes. Generally, the third threshold differs from the first threshold because the dimensions of comparison are drastically different: one is based on feature approximation, and the other on the strength of the accepted two-node structural information.

[0082] S4. The discriminator obtains the adjacency matrix and simplex structure information matrix after the generator iteration terminates, as well as the adjacency matrix and simplex structure information matrix of the second brain region-gene network after the corresponding sample evolution. Then, it distinguishes between the true and false of the two types of input data and performs backpropagation according to the loss function to update the learning matrix in the generator.

[0083] S5. After the generator and the discriminator are repeatedly trained on the dataset until the generative adversarial network converges, the trained generator is used to detect the evolution of the brain region-gene network in new sample data online.

[0084] Furthermore, the method in this embodiment also includes:

[0085] S6. The generator calculates the approximation of the adjacency matrix and simplex structure information matrix in the online detection results of the new sample with the standard adjacency matrix and standard simplex structure information matrix, respectively. Then, the two approximation calculation results are weighted to obtain the probability that the new sample evolves into the disease stage commonly corresponding to the standard adjacency matrix and the standard simplex structure information matrix.

[0086] Example 2

[0087] The present invention also discloses a brain region-gene network evolution system based on generative adversarial networks, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the computer program.

[0088] In summary, the methods and systems disclosed in the above embodiments of the present invention have the following beneficial effects:

[0089] Generative adversarial networks (GANs) are relatively mature neural networks, and the detection of various simplexes and their relationships are based on the edge-to-edge connections, which can be reliably implemented through algorithms. This invention, by redefining the initial structural information of each simplex, innovatively discloses how supernode simplexes dynamically construct, merge, and diffuse their structural information during a single iteration in the process of updating structural information based on the learning matrix. It also discloses the inheritance and renewal relationships between adjacent iterations, perfectly simulating the disease progression of Alzheimer's disease and other intractable diseases, ensuring the reliability of the evolution results. This provides new technical guidance for the research, prediction, and prevention of Alzheimer's disease and other intractable diseases.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A brain region-gene network evolution method based on generative adversarial networks, wherein the generative adversarial network includes a generator and a discriminator; characterized in that, Includes the following steps: S1. Obtain the first and second brain regions-gene networks corresponding to each sample before and after evolution; the edges of the first and second brain regions-gene networks are used to connect two nodes whose feature correlation is greater than the first threshold; in all samples, the number of nodes and the feature source of each node are consistent. S2. Detect all simplex sets of different dimensions in the first and second brain region-gene networks to obtain the simplex structure information matrix and the adjacency matrix between each simplex corresponding to the first and second brain region-gene networks respectively; among the different elements of the simplex structure information matrix, the structure information of zero-dimensional nodes is calculated based on degree centrality, and the structure information of one-dimensional edges and polygons of two or more dimensions is calculated based on multivariate mutual information. S3. Input the adjacency matrix and simplex structure information matrix of the first brain region-gene network before evolution into the generator to update and iterate the adjacency matrix and simplex structure information matrix, and then send the adjacency matrix and simplex structure information matrix after the iteration terminates to the discriminator. S4. The discriminator obtains the adjacency matrix and simplex structure information matrix after the generator iteration terminates, and the adjacency matrix and simplex structure information matrix of the second brain region-gene network after the corresponding sample evolution. Then, it distinguishes the authenticity of the two types of input data and performs backpropagation according to the loss function to update the learning matrix in the generator. S5. After the generator and the discriminator are repeatedly trained on the dataset until the generative adversarial network converges, the trained generator is used to detect the evolution of the brain region-gene network in new sample data online.

2. The brain region-gene network evolution method based on generative adversarial networks according to claim 1, characterized in that, In step S3, the specific process of a single iteration includes: S31. Obtain the closed star-shaped neighborhood of each simplex based on the adjacency matrix before evolution or after the previous iteration. Then, update the structural information of each simplex based on the star-shaped closed adjacency matrix formed by each closed star-shaped neighborhood, the simplex structural information matrix before evolution or after the previous iteration, and the first type of learning matrix to obtain the intermediate state simplex structural information matrix. The first type of learning matrix is ​​used to allocate the interaction intensity of structural information between each simplex. The star-shaped closed adjacency matrix is ​​used to limit the scope of any simplex to the closed star-shaped neighborhood formed by that simplex. S32. Select stable simplexes with structural information strength greater than the second threshold, and merge the stable simplexes with directly related structures into a supernode simplex complex, wherein the structural information strength of all nodes and edges in the supernode simplex complex is greater than the second threshold. S33. Determine the external nodes that can be affected by the simple complex of each supernode based on the closed star neighborhood; S34. Establish potential edges between all nodes inside the simple complex of each supernode and all external nodes within the influence range; S35. Calculate the structural information of each potential edge based on the updated node structural information of the first type of learning matrix and the second type of learning matrix. The second type of learning matrix is ​​used to allocate the contribution relationship between the structural information of each node and the structural information of the potential edge. S36. Compare the structural information of each potential edge with the third threshold, and determine the potential edge whose structural information is greater than the third threshold as a real edge and retain it; otherwise, delete it. S37. Update the adjacency matrix and the simplex structure information matrix of the intermediate state according to the addition and deletion of each potential edge, and obtain the adjacency matrix and simplex structure information matrix after the current iteration.

3. The brain region-gene network evolution method based on generative adversarial networks according to claim 1 or 2, characterized in that, Also includes: S6. The generator calculates the approximation of the adjacency matrix and simplex structure information matrix in the online detection results of the new sample with the standard adjacency matrix and standard simplex structure information matrix, respectively. Then, the two approximation calculation results are weighted to obtain the probability that the new sample evolves into the disease stage commonly corresponding to the standard adjacency matrix and the standard simplex structure information matrix.

4. The brain region-gene network evolution method based on generative adversarial networks according to claim 2, characterized in that, The simplex structure information matrix is ​​a diagonal matrix; the specific calculation formula for step S31 is: ； Among them, superscript Represents the current iteration number. Represents the state when the previous iteration was completed, when This represents the initial state corresponding to the first brain region-gene network. This represents the intermediate state of the current iteration process; Represents the simplex structure information matrix. For the first The first type of learning matrix in the next iteration has a dimension of ; This represents the total number of all simplexes in the first brain region-gene network. It is a matrix of all ones. A matrix whose diagonal elements are all 1s and all other elements are all 0s; "Represents the Hadamard product operation," " represents multiplication; To determine the adjacency matrix based on the adjacency matrix at the end of the previous iteration The star-shaped closed adjacency matrix obtained from the simplex set; The specific calculation formula for step S35 is as follows: ;in, express The only node set extracted The local matrix formed, To represent the supernode diffusion matrix of the external nodes that can be affected by the simple complex of each supernode, This indicates the transmission of structural information between nodes within a supernode; Represents the information matrix of the external nodes of the supernode; superscript It is the transpose operator; For the first The second type of learning matrix in the next iteration process; Indicates the first The matrix of structural information received by each potential edge during each iteration.

5. The brain region-gene network evolution method based on generative adversarial networks according to claim 4, characterized in that, Adjacency matrix between simplexes It is represented as: ； in, Indicates the first A simplex With the A simplex They are adjacent. Then it means and There is no adjacency relationship, and the adjacency relationship between simplexes of the same dimension is 0; The dimension of the supernode diffusion matrix is , The number of nodes in the brain region-gene network; the first Elements of the node diffusion matrix during the next iteration Indicates the first The node and the first Each node has a potential edge; if no potential edge exists, then... .

6. The brain region-gene network evolution method based on generative adversarial networks according to claim 4, characterized in that, elements =1 indicates the first The simplex in the first... In a closed star-shaped neighborhood of a simplex, if the first... The simplex is not in the first... In a closed star-shaped neighborhood of a simplex, it is represented as =0.

7. The brain region-gene network evolution method based on generative adversarial networks according to claim 3, characterized in that, The disease stage referred to is the disease stage of Alzheimer's disease.

8. A brain region-gene network-based evolutionary system based on generative adversarial networks, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1 to 7.

Images