Disease process based population feature extraction method and system using large models

Abstract

The application relates to the technical field of big data models, and discloses a lesion process-based group feature extraction method and system based on a big model, which is used for deducing a lesion process through changes of a directed graph and an edge weight and improving reliability of extracted group features. The method comprises the following steps: constructing a gene-brain region directed graph of each subject based on brain images and gene data; decomposing into strongly connected components, selecting root nodes to construct a directed tree; updating features through inter-layer and intra-layer transmission, pruning multi-in-degree nodes until the in-degree is 1 to obtain a key support tree; fusing the root nodes to obtain a key subgraph; generating a graph through diffusion deconvolution reconstruction, training a discriminator together with a real graph, sharing parameters to amplify disease feature differences; and after convergence, screening lesion brain regions and risk genes based on group statistical node importance and in combination with incremental search.

Description

Technical Field

[0001] This invention relates to the field of big data modeling technology, and in particular to a method and system for extracting population features of disease processes based on a big data model. Background Technology

[0002] Alzheimer's disease (AD) is a complex neurodegenerative disease whose pathogenesis involves the interaction between abnormal gene expression and brain region dysfunction. Research indicates that information transmission between genes and brain regions has a clear direction and specific regulatory mechanisms; when this information flow is impaired, it may lead to the onset and progression of the disease. Therefore, simultaneously analyzing genetic data and brain imaging data, and modeling the directed interaction between them, is of great significance for revealing the pathological mechanisms of AD.

[0003] Currently, computer-aided diagnosis and feature extraction methods for Alzheimer's disease (AD) mainly suffer from the following technical limitations:

[0004] Graph structure modeling ignores directionality: Traditional methods typically use undirected graphs (such as brain networks built based on Pearson correlation coefficients) to represent the connections between genes or brain regions, which cannot accurately describe the direction of information transmission. However, physiological processes such as gene regulation of brain regions and feedforward / feedback between brain regions are inherently directed, and undirected models will lose crucial directional information.

[0005] Static correlations are difficult to reflect dynamic regulation: Existing methods mostly use static correlations (such as Pearson coefficient and mutual information) to construct the edges between nodes, which cannot capture the time-lag causal regulatory relationships between genes and brain regions (such as Granger causality and transfer entropy).

[0006] Complex graph structures make it difficult to extract key features: Gene-brain region directed graphs are typically large in scale and densely connected, containing a large number of redundant or weakly correlated edges. Existing graph neural network (GNN) methods directly convolve the original graph, lacking an effective mechanism to extract key substructures with disease-discriminating capabilities from complex graphs, resulting in limited classification performance and difficulty in locating specific pathogenic genes and diseased brain regions.

[0007] In addition, most existing methods analyze gene data or image data separately, and the few fusion methods are mainly based on simple splicing or feature cascading, failing to make full use of the directed regulatory relationship between genes and brain regions for joint modeling. Summary of the Invention

[0008] The purpose of this invention is to disclose a method and system for extracting population features based on a large model of disease process, so as to interpret the disease process through the changes of directed graph and edge weights and improve the reliability of the extracted population features.

[0009] To achieve the above objectives, this invention discloses a method for extracting population features of disease processes based on a large model, comprising: Step S1: Based on the brain imaging data and gene data of each subject, construct directed brain region graphs and directed gene graphs, calculate the transfer entropy between genes and brain regions, and fuse them to obtain the gene sequence of each subject. Directed graphs of brain regions; Step S2: Based on genes The adjacency matrix of the directed graph of the brain region stores the genes of each subject. The directed graph of the brain region is decomposed into at least two strongly connected components that do not intersect; within each strongly connected component, the node with the highest importance is selected as the root node and a directed structure tree is constructed according to the directed edges; the strongly connected component is a single node or the largest set of nodes where "any two nodes can reach each other along the direction of the directed edges". Step S3: Perform inter-layer and intra-layer information transfer in each directed tree structure in sequence to update node feature information. After completing the intra-layer information transfer, first compare the weight difference of each in-degree edge for nodes with at least two in-degree edges, and prune the edge with the smallest weight before transferring information to the next layer. If, after pruning the bottom layer nodes, there are still nodes with an in-degree greater than 1, return to the root node to re-process the information transfer until the key support tree is pruned into a tree in which the in-degree of all nodes except the root node is 1. Step S4: Arrange the root nodes of each key support tree according to their genes. The key genes for each subject were obtained by fusing relationships in directed brain region graphs. Brain regions have directional maps, the key genes mentioned Directed subgraphs in brain regions carry edge weight information after information transmission and processing. Step S5: Incorporate key genes The directed subgraphs of the brain region are reconstructed into generated graphs through diffusing deconvolution; the real graphs and generated graphs are input into the discriminator, which outputs classification probabilities through directed graph convolution and fully connected layers; through adversarial training between the generator and the discriminator, the learnable parameter matrix shared by each directed tree structure amplifies the feature differences between patients and normal subjects during information transmission. Step S6: After the adversarial training between the generator and the discriminator converges, the importance scores of each brain region node and gene node are statistically analyzed based on the population samples. Then, the subset of brain regions and gene subsets that achieve the highest classification accuracy are selected by ranking the importance scores and using the incremental search method. These subsets are then identified as the diseased brain regions and risk genes that are clustered for the disease category.

[0010] Preferably, in step S1, the causal relationship between gene sequences is calculated by Granger causality analysis, and a directed edge is established if a causal relationship exists; the information similarity between brain region time series is calculated by lag correlation coefficient, and a directed edge is established when the similarity is greater than a threshold; the causal relationship of gene regulation on brain regions is calculated by transfer entropy, and a directed edge is established if the causal relationship is valid.

[0011] Preferably, in step S2, the gene is... In the process of decomposing a directed graph of a brain region into at least two strongly connected components with no intersection, nodes with only incoming edges and nodes with only outgoing edges are first removed as strongly connected components. Then, a depth-first search algorithm is used to decompose the remaining strongly connected components. Isolated points after decomposition are also considered as strongly connected components.

[0012] Preferably, in step S3, the feature information of the root node remains at its initial value throughout the inter-layer and intra-layer information transmission process of the directed tree structure, and the weight of the edge is calculated using the Pearson correlation coefficient between the feature information of the two connected nodes; and the feature information of the single node constituting the strongly connected component remains at its initial value.

[0013] Preferably, both inter-layer and intra-layer information transmission first employ an aggregation method of "current node features + ∑(edge ​​weight × current directed neighbor node features conforming to inter-layer and intra-layer rules)" to obtain the aggregated feature vector. ∈ ; then, Right-multiply by the learnable parameter matrix shared by all directed tree structures. , ∈ The updated node features are obtained. ;matrix elements in The first feature represents the input feature. The i-th dimension of the output feature Linear contribution weights in each dimension For the input feature dimension, This is the output feature dimension.

[0014] Preferably, key genes The specific processing steps for carrying edge weight information in directed subgraphs of brain regions include: First, arrange the edge weight matrices of each key support tree diagonally and merge them into one matrix. Then, arrange the weight matrices of each root node according to their gene structure. The relationships in the directed graph of the brain region are established by creating connecting edges between the root nodes, and the edge weights between the root nodes are filled into the merged matrix to complete the fusion of the key support trees; the edge weights between the root nodes are all kept at their initial values.

[0015] Preferably, in the case of key genes The process of reconstructing a generative graph from a directed subgraph of a brain region using diffusing deconvolution is determined by the key gene-directed subgraph weight matrix of the brain region. go through Obtained by multiple diffusion deconvolutions. elements in This represents the weight of the directed edge from node i to node j; set up The input to the first-order diffraction deconvolution is , After diffusion deconvolution, the directed graph of the gene-brain region generated by the step absorbs each edge. The structural information of edge diffusion within the order neighborhood is used to obtain the updated weight matrix. The goal of diffusing deconvolution is to compute new weights for each possible directed edge in the graph. Let... For different nodes, for slave nodes To the node Target edge, traverse all nodes The calculation formula is: = ; in, yes Bias of order-diffusion deconvolution; It is the first The learnable kernel matrix of order-order diffraction deconvolution. It is the gene in step S1 The number of nodes in a directed graph of a brain region; Indicates in In the order-generating graph, from node Pointing to node The weight of the directed edge; Indicates in In the order-generating graph, from node Pointing to node The weight of the directed edge; Represents a node To the node The diffusion coefficient, corresponding to The Middle Line 1 The value of the column; node To the node The diffusion coefficient, corresponding to The Middle Line 1 The value of the column; It is an activation function ; , They are from the nodes To the node target side go through and The updated value after diffusing deconvolution.

[0016] Preferably, the method of the present invention further includes: extracting regulatory pathways along directed edges in the key support tree of each subject based on the lesion brain region and risk genes.

[0017] To achieve the above objectives, the present invention also discloses a population feature extraction system for pathological processes based on a large model, 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.

[0018] The present invention has the following beneficial effects: 1. Accurately model directional control relationships and preserve information flow.

[0019] By constructing genes Directed graphs of brain regions (gene→gene, brain region→brain region, gene→brain region) were constructed, and directed edges were established using Granger causality and transfer entropy analysis. This overcame the technical deficiency of traditional undirected graphs in that they could not distinguish the direction of information transmission, and provided a foundation for the subsequent extraction of causal regulatory pathways.

[0020] 2. Decompose strongly connected components to reduce the difficulty of analyzing complex graphs.

[0021] The directed graph is decomposed into strongly connected, non-intersecting components, isolating tightly coupled groups of nodes and avoiding interference from complex global connections in key feature extraction. A strongly connected component is either a single node or the largest set of nodes where any two nodes are mutually reachable along directed edges, ensuring the integrity of matrix operations and the uniformity of feature updates.

[0022] 3. Simulate multi-hop information flow and adaptive pruning to highlight key structures.

[0023] Within strongly connected components, inter-layer and intra-layer propagation are performed hierarchically. After each intra-layer propagation, multi-in-degree nodes are pruned based on weight differences, removing only the edges with the smallest weights. If multi-in-degree nodes still exist, the process iterates again from the root node until all nodes outside the root node have an in-degree of 1, forming a key support tree. This process simulates the hierarchical propagation of biological information along the regulatory path and the gradual elimination of weak connections, effectively amplifying the feature differences between normal individuals and patients, and solving the problem that existing graph neural networks struggle to extract discriminative substructures from dense graphs.

[0024] 4. Integrate the key support tree and diffusing deconvolution to reconstruct the complete graph structure.

[0025] The key support trees are merged into a key gene through the root node relationship. The directed subgraphs in the brain region are then gradually restored by diffusing deconvolution, which preserves the sparsity of the key subgraphs and achieves consistency constraints with the original graph structure, thus facilitating the stable training of generative adversarial networks.

[0026] 5. Adversarial training drives the sharing of learnable parameters, enhancing feature differentiation.

[0027] Through adversarial game, the generator and discriminator automatically adjust the learnable parameter matrix shared by all directed tree structures, highlighting disease-related feature differences during information transmission. At the same time, the discriminator outputs classification probabilities, enabling accurate classification of normal, early MCI (Mild Cognitive Impairment), and AD.

[0028] 6. After model convergence, a key feature matrix for the population is obtained by weighted averaging of the key subgraph weight matrices of multiple similar samples. The importance score of each node is calculated, and then an incremental search method is used to automatically select the subset of brain regions and gene subsets that achieve the highest classification accuracy as the lesion brain regions and risk genes. This method does not rely on a subjectively set cutoff threshold, and the extraction results are statistically significant and reproducible.

[0029] In summary, this invention significantly improves the classification accuracy of Alzheimer's disease through a series of techniques, including directed graph decomposition, hierarchical information transfer and adaptive pruning, adversarial training, and group feature extraction. It can also unbiasedly identify disease-related risk genes and lesion brain regions, overcoming the shortcomings of existing technologies such as directional loss, key structure submersion, and subjective feature extraction, and has good prospects for clinical application.

[0030] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0031] 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: Figure 1 This is an overall framework diagram of the Directed Graph Generative Adversarial Network (DG-GAN) disclosed in the embodiments of the present invention.

[0032] Figure 2 This is a flowchart of the population feature extraction method based on a large model for the pathological process disclosed in the embodiments of the present invention.

[0033] Figure 3 The gene disclosed in the embodiments of this invention An example diagram illustrating the decomposition of a directed graph of a brain region into at least two strongly connected components with no intersection. Detailed Implementation

[0034] 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.

[0035] Example 1 This embodiment discloses a method for extracting population features of disease processes based on a large model, referring to... Figure 1 The diagram shows the overall framework of the Directed Graph Generative Adversarial Network (DG-GAN). It constructs a gene-brain region directed graph for each subject based on brain imaging and gene data; decomposes it into strongly connected components, selects root nodes to construct a directed tree; updates features through inter-layer and intra-layer propagation, prunes multi-in-degree nodes until the in-degree is 1, and obtains a key support tree; merges the root nodes to obtain a key subgraph; reconstructs the generated graph through diffusing deconvolution, and trains the discriminator together with the real graph, sharing parameters to amplify the differences in disease features; after convergence, it filters diseased brain regions and risk genes based on the importance of nodes in the population statistics, combined with incremental search.

[0036] Furthermore, referring to Figure 2 The method of this embodiment is further described in detail step by step as follows: Step S1: Based on the brain imaging data and gene data of each subject, construct directed brain region graphs and directed gene graphs, calculate the transfer entropy between genes and brain regions, and fuse them to obtain the gene sequence of each subject. Directed graph of brain regions.

[0037] In this step, when a node represents a brain region, its feature information is the time-series value of that brain region; when a node represents a gene, its feature information is the gene's coding sequence value. The causal relationship between gene sequences can be calculated using Granger causality analysis; if a causal relationship exists, a directed edge is established. The information similarity between brain region time series can be calculated using lag correlation coefficients; when the similarity exceeds a threshold, a directed edge is established. The causal relationship of gene regulation of brain regions can be calculated using transfer entropy; if a causal relationship is established, a directed edge is established.

[0038] For two gene sequences, the core idea of ​​the Granger causality test is: if using the past value of gene A to predict the current value of gene B is better than using only the past value of gene B itself (i.e. the prediction error is significantly reduced), then there is a causal relationship from gene A to gene B, that is, there is a directed edge from gene A to gene B.

[0039] The lag correlation coefficient measures the degree of linear correlation between the current value of one time series and the value of another time series at a past point in time (lag). In brain region time series analysis, this helps in infer the direction and delay of information transmission from one brain region to another. Assuming two brain region time series x(t) and y(t), the lag correlation coefficient calculates the correlation between x(t) and y(t) when y is lagged by k time steps. The Pearson correlation coefficient of y(k) is calculated. If the absolute value of this correlation coefficient is significantly greater than 0, it indicates that the past value of y can predict the current value of x, that is, there is a directed edge from y to x to represent the flow of information.

[0040] Transfer entropy is an asymmetric metric based on information theory, used to quantify the directed flow of information from one stochastic process (source) to another stochastic process (target). In gene-brain region regulation analysis, gene sequences are taken as the source G, and brain region time series are taken as the target X. By calculating the extent to which past values ​​of G can reduce the uncertainty of future values ​​of X (information beyond X's own history), we can determine whether a causal relationship exists from genes to brain regions.

[0041] Some existing papers involve the construction of the above three different types of directed graphs, which are existing technologies and will not be elaborated upon.

[0042] Step S2: Based on genes The adjacency matrix of the directed graph of the brain region stores the genes of each subject. The directed graph of the brain region is decomposed into at least two strongly connected components with no intersection; within each strongly connected component, the node with the highest importance is selected as the root node and a directed structure tree is constructed according to the directed edges; the strongly connected component is a single node or the largest set of nodes where "any two nodes can reach each other along the direction of the directed edge".

[0043] In this step, unlike the adjacency matrix of an undirected graph, the adjacency matrix in this embodiment can express three types of relationships: unidirectional, bidirectional, and unconnected. The elements of the adjacency matrix A are typically represented by 0 or 1 to indicate the presence or absence of a directed edge, but they are not necessarily symmetrical (i.e., the values ​​of the elements in the adjacency matrix representing the directed edge from node i to node j are not necessarily symmetrical). The element values ​​represent the directed edge from node j to node i. (Not necessarily equal); genes In the process of decomposing a directed graph of a brain region into at least two strongly connected components with no intersection, nodes with only incoming edges and nodes with only outgoing edges are first removed and regarded as independent strongly connected components. Then, a depth-search algorithm is used to decompose the remaining strongly connected components. The isolated points after decomposition are also regarded as strongly connected components.

[0044] Reference Figure 3 The image shows the first The gene-brain region directed graph connection structure of each subject and the directed graph of that gene-brain region were obtained. Based on the adjacency matrix and depth search algorithm, it was first found that the out-degree of node 9 in the adjacency matrix is ​​all 0, so node 9 forms a strong connected component 4 on its own. Next, the following strong connected components were extracted: strong connected component 1 consisting of nodes 0, 1, 2, 3 and 4; strong connected component 2 consisting of nodes 10, 11 and 12; and strong connected component 3 consisting of nodes 5, 6, 7 and 8.

[0045] Optionally, during the decomposition of strongly connected components, since the number of gene nodes in the gene-brain region directed network is much smaller than the number of brain region nodes, and a gene often interacts with multiple brain regions or genes, the connection complexity of gene nodes is often greater than that of brain region nodes. For the sake of algorithm efficiency, biological significance, and engineering simplification, gene nodes can be sorted in descending order of importance based on betweenness centrality (a term referring to the proportion of paths passing through a node in the shortest paths between all pairs of nodes in the network), and then the strongly connected components can be decomposed in order of gene node importance.

[0046] Step S3: Perform inter-layer and intra-layer information transfer in each directed tree structure in sequence to update node feature information. After completing the intra-layer information transfer, first compare the weight difference of each in-degree edge for nodes with at least two in-degree edges, and prune the edge with the smallest weight before transferring information to the next layer. If, after pruning the bottom-level nodes, there are still nodes with an in-degree greater than 1, return to the root node to re-process the information transfer until the key support tree is pruned so that the in-degree of all nodes except the root node is 1.

[0047] Similarly, based on the principle that gene importance is higher than brain region importance, in the directed tree structure corresponding to strongly connected components, the gene with the highest importance is selected as the root node, referring to... Figure 1 The root node is node 0. According to the hierarchical relationship of directed edges, nodes 1, 3, and 2 are determined as the initial first-level nodes, and node 4 is the initial second-level node. The initial intra-level information transmission direction of the first level is divided into node 2 to node 3 and node 3 to node 1.

[0048] In this step, both inter-layer and intra-layer information transfer first use an aggregation method of "current node features + ∑(edge ​​weight × current directed neighbor node features that conform to inter-layer and intra-layer rules)" to obtain the aggregated feature vector. ∈ ; then, Right-multiply by the learnable parameter matrix shared by all directed tree structures. , ∈ The updated node features are obtained. ;matrix elements in The first feature represents the input feature. The i-th dimension of the output feature Linear contribution weights in each dimension For the input feature dimension, The output feature dimension is defined as follows: Edge weights are calculated using the Pearson correlation coefficient between the feature information of the two connected nodes. If the feature information of a related node has already incorporated features from the upper-level nodes based on inter-layer propagation, the updated feature after incorporating the upper-level node features is used during inter-layer feature propagation. Edge weights during intra-layer propagation are also calculated based on this updated feature using the Pearson correlation coefficient. Feature propagation within the same layer does not distinguish the order of propagation; only the direction of flow is determined by the directed edges. The updated feature is then propagated to the next layer of nodes, and the initial feature of the next layer node is the updated feature after completing the previous round of information propagation. A round is counted only when the root node propagates from top to bottom to the lowest layer. As a complement, the original node features of genes and brain regions are aligned to the same dimension during preprocessing using methods such as projection to facilitate subsequent data processing.

[0049] Still with Figure 1 For example, after the updated feature information of node 3 is passed to node 1, the edge weights corresponding to the two in-degree edges of node 3 are calculated. It is determined that the edge weight from node 1 to node 3 is less than the edge weight from node 2 to node 3. Therefore, the edge between node 3 and the root node 0 is pruned. Next, node 1 compares the edge weights of the in-degree edges originating from the root node 0 and the in-degree edges originating from the root node 3. Based on the comparison result, the in-degree edges from node 3 to node 1 are pruned, thus reducing node 3 to a child node of node 2. In this embodiment, the feature information of the root node remains at its initial value throughout the inter-layer and intra-layer information transmission process of the directed tree structure. Therefore, the directed edges from node 1 to the root node 0 are pruned to obtain the final key support tree.

[0050] Step S4: Arrange the root nodes of each key support tree according to their genes. The key genes for each subject were obtained by fusing relationships in directed brain region graphs. Brain regions have directional maps, the key genes mentioned The directed subgraphs in the brain region carry edge weight information after information transmission processing.

[0051] In this step, key genes The specific processing procedure for the directed subgraph of the brain region carrying edge weight information can be as follows: First, the edge weight matrices of each key supporting tree are arranged diagonally and merged into one matrix. Then, each root node is processed in the gene... In the directed graph of brain regions, connections between root nodes are established, and the edge weights are filled into the merged matrix to complete the fusion of key support trees. Furthermore, in key genes... In directed subgraphs of brain regions, the feature information of single nodes constituting strongly connected components is preserved at their initial values.

[0052] Step S5: Incorporate key genes The directed subgraphs of the brain region are reconstructed into generated graphs through diffusing deconvolution; the real graphs and generated graphs are input into the discriminator, which outputs classification probabilities through directed graph convolution and fully connected layers; through adversarial training between the generator and the discriminator, the learnable parameter matrix shared by each directed tree structure amplifies the feature differences between patients and normal subjects during information transmission.

[0053] This step involves key genes. The process of reconstructing a generative graph from a directed subgraph of a brain region using diffusing deconvolution is determined by the key gene-directed subgraph weight matrix of the brain region. go through Obtained by multiple diffusion deconvolutions. elements in The weight represents the directed edge weight from node i to node j, used to characterize the correlation strength. Positive values ​​indicate positive correlation / activation regulation, while negative values ​​indicate negative correlation / inhibition regulation.

[0054] set up The input to the first-order diffraction deconvolution is , After diffusion deconvolution, the directed graph of the gene-brain region generated by the step absorbs each edge. The structural information of edge diffusion within the order neighborhood is used to obtain the updated weight matrix. The goal of diffusing deconvolution is to compute new weights for each possible directed edge in the graph. Let... For different nodes, for slave nodes To the node Target edge, traverse all nodes The calculation formula is: = ; in, yes Bias of order-diffusion deconvolution; It is the first The learnable kernel matrix of order-order diffraction deconvolution. It is the gene in step S1 The number of nodes in a directed graph of a brain region; Indicates in In the order-generating graph, from node Pointing to node The weight of the directed edge; Indicates in In the order-generating graph, from node Pointing to node The weight of the directed edge; Represents a node To the node The diffusion coefficient, corresponding to The Middle Line 1 The value of the column; node To the node The diffusion coefficient, corresponding to The Middle Line 1 The value of the column; It is an activation function ; , They are from the nodes To the node target side go through and The updated value after diffusing deconvolution.

[0055] In summary, During the process of diffraction deconvolution, Each edge in the brain region-gene hypermap has been absorbed. Structural information of edge diffusion within the order neighborhood, at which point key genes... In a directed subgraph of a brain region, each edge diffuses its own structural information to edges within its first-order neighborhood; that is, each edge in the network graph can absorb its own structural information. The structural information of edges within the order neighborhood. During the deconvolution process, the in-degree and out-degree edges that were pruned in each node of the component convolution are gradually restored. Each key edge in the key support tree diffuses its own information to the other potential edges, thereby obtaining... Gene generation-brain region directed graph.

[0056] Step S6: After the adversarial training between the generator and discriminator converges, the importance scores of each brain region node and gene node are statistically analyzed based on the population samples. Then, the subsets of brain regions and genes that achieve the highest classification accuracy are selected by ranking the importance scores and using an incremental search method. These subsets are identified as the diseased brain regions and risk genes associated with the disease categories. Further, based on the diseased brain regions and risk genes, regulatory paths are extracted along directed edges in the key support tree of each subject.

[0057] In this step, the loss function can be constructed from adversarial loss, classification loss, and generator-assisted loss, which are conventional techniques in this field. Backpropagation and parameter iteration mechanisms, as well as importance score calculation and incremental search methods, are existing technologies and will not be elaborated upon.

[0058] In this embodiment, the purpose of pruning to obtain the key support tree is to extract discriminative features: by simulating information transmission, comparing edge weights, and deleting weak edges, the key substructures (tree-like, with each node having an in-degree ≤1) that contribute most to disease classification are extracted from the complex original graph. This key support tree is sparse and has removed redundant connections. It amplifies the differences between normal individuals and patients, serves as the direct basis for subsequent classification and feature extraction, and reduces noise to avoid interfering with the classifier.

[0059] The purpose of diffusing deconvolution (reconstructing the generated graph) is to meet the adversarial training requirements of generative adversarial networks: the discriminator needs to compare the real graph with the generated graph. The real graph is the original, dense gene-brain region directed graph (containing all edges). Therefore, the generator must output a graph of the same size and structure as the real graph (i.e., the same number of nodes, and the same edge weight matrix size (referring to row and column dimensions, not element values), rather than only outputting a sparse key subgraph. Otherwise, the discriminator cannot make a fair comparison. Furthermore, the generated graph after diffusing deconvolution is only used for the discriminator's true / false judgment and not for subsequent extraction of diseased brain regions / risk genes. Feature extraction is still based on the key subgraph obtained through pruning.

[0060] The generated graph after diffusing deconvolution is not a simple reconstruction of the original graph, but rather a recognizable graph learned by the model. Its weight distribution differs from the original graph, but its overall statistical properties are similar. Through adversarial training, the generator is forced to make the generated graph increasingly closer to the real graph, thereby compelling the key subgraphs extracted by the generator to contain sufficient discriminative information.

[0061] In conclusion, Figure 1 The end-to-end learning process shown involves pruning and reconstruction, which are two in-series modules of the same generator and are indispensable. The key graph obtained by pruning purifies the information of the original graph, while diffusing deconvolution is the output format conversion that the generator must perform. The combination of the two enables the generator to extract discriminative features and output generated graphs compatible with the real graph, thereby completing adversarial training and extracting reliable group features from the key graph obtained by pruning through importance calculation and statistics.

[0062] Example 2 This embodiment discloses a population feature extraction system for pathological processes based on a large model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method.

[0063] In summary, the disease process extraction method and system based on a large model for population features disclosed in the embodiments of the present invention have the following beneficial effects: 1. Accurately model directional control relationships and preserve information flow.

[0064] By constructing genes Directed graphs of brain regions (gene→gene, brain region→brain region, gene→brain region) were constructed, and directed edges were established using Granger causality and transfer entropy analysis. This overcame the technical deficiency of traditional undirected graphs in that they could not distinguish the direction of information transmission, and provided a foundation for the subsequent extraction of causal regulatory pathways.

[0065] 2. Decompose strongly connected components to reduce the difficulty of analyzing complex graphs.

[0066] The directed graph is decomposed into strongly connected, non-intersecting components, isolating tightly coupled groups of nodes and avoiding interference from complex global connections in key feature extraction. A strongly connected component is either a single node or the largest set of nodes where any two nodes are mutually reachable along directed edges, ensuring the integrity of matrix operations and the uniformity of feature updates.

[0067] 3. Simulate multi-hop information flow and adaptive pruning to highlight key structures.

[0068] Within strongly connected components, inter-layer and intra-layer propagation are performed hierarchically. After each intra-layer propagation, multi-in-degree nodes are pruned based on weight differences, removing only the edges with the smallest weights. If multi-in-degree nodes still exist, the process iterates again from the root node until all nodes outside the root node have an in-degree of 1, forming a key support tree. This process simulates the hierarchical propagation of biological information along the regulatory path and the gradual elimination of weak connections, effectively amplifying the feature differences between normal individuals and patients, and solving the problem that existing graph neural networks struggle to extract discriminative substructures from dense graphs.

[0069] 4. Integrate the key support tree and diffusing deconvolution to reconstruct the complete graph structure.

[0070] The key support trees are merged into a key gene through the root node relationship. The directed subgraphs in the brain region are then gradually restored by diffusing deconvolution, which preserves the sparsity of the key subgraphs and achieves consistency constraints with the original graph structure, thus facilitating the stable training of generative adversarial networks.

[0071] 5. Adversarial training drives the sharing of learnable parameters, enhancing feature differentiation.

[0072] The generator and discriminator engage in adversarial game, which automatically adjusts the learnable parameter matrix shared by all directed tree structures. This highlights disease-related feature differences during information transmission, while the discriminator outputs classification probabilities, enabling accurate classification of normal, early-stage MCI, and AD.

[0073] 6. After model convergence, a key feature matrix for the population is obtained by weighted averaging of the key subgraph weight matrices of multiple similar samples. The importance score of each node is calculated, and then an incremental search method is used to automatically select the subset of brain regions and gene subsets that achieve the highest classification accuracy as the lesion brain regions and risk genes. This method does not rely on a subjectively set cutoff threshold, and the extraction results are statistically significant and reproducible.

[0074] In summary, this invention significantly improves the classification accuracy of Alzheimer's disease through a series of techniques, including directed graph decomposition, hierarchical information transfer and adaptive pruning, adversarial training, and group feature extraction. It can also unbiasedly identify disease-related risk genes and lesion brain regions, overcoming the shortcomings of existing technologies such as directional loss, key structure submersion, and subjective feature extraction, and has good prospects for clinical application.

[0075] 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 lesion process large model-based population feature extraction method, characterized in that, include: Step S1: based on the brain image data and gene data of each subject, construct a brain region directed graph and a gene directed graph, and calculate the transfer entropy between genes and brain regions, and fuse to obtain the gene brain region directed graph; Step S2: Based on genes The adjacency matrix of the directed graph of the brain region stores the genes of each subject. The directed graph of the brain region is decomposed into at least two strongly connected components with no intersection; within each strongly connected component, the node with the highest importance is selected as the root node and a directed structure tree is constructed according to the directed edges; the strongly connected component is a single node or the largest set of nodes where "any two nodes can reach each other along the direction of the directed edge". Step S3: Perform inter-layer and intra-layer information transfer in each directed tree structure in sequence to update node feature information. After completing the intra-layer information transfer, first compare the weight difference of each in-degree edge for nodes with at least two in-degree edges, and prune the edge with the smallest weight before transferring information to the next layer. If, after pruning the bottom layer nodes, there are still nodes with an in-degree greater than 1, return to the root node to re-process the information transfer until the key support tree is pruned into a tree in which the in-degree of all nodes except the root node is 1. Step S4: Arrange the root nodes of each key support tree according to their genes. The key genes for each subject were obtained by fusing relationships in directed brain region graphs. Brain regions have directional maps, the key genes mentioned Directed subgraphs in brain regions carry edge weight information after information transmission and processing. Step S5: Incorporate key genes The directed subgraph of the brain region is reconstructed into a generated graph through diffusing deconvolution; the real graph and the generated graph are input into the discriminator, and the output is the classification probability after being processed by directed graph convolution and fully connected layer; Through adversarial training between the generator and the discriminator, the learnable parameter matrix shared by each directed tree structure amplifies the feature differences between patients and normal subjects during the information transmission process. Step S6: After the adversarial training between the generator and the discriminator converges, the importance scores of each brain region node and gene node are statistically analyzed based on the population samples. Then, the subset of brain regions and gene subsets that achieve the highest classification accuracy are selected by ranking the importance scores and using the incremental search method. These subsets are then identified as the diseased brain regions and risk genes that are clustered for the disease category. Among them, in the key genes The process of reconstructing a generative graph from a directed subgraph of a brain region using diffusing deconvolution is determined by the key gene-directed subgraph weight matrix of the brain region. go through Obtained by multiple diffusion deconvolutions. elements in This represents the weight of the directed edge from node i to node j; set up The input to the first-order diffraction deconvolution is , After diffusion deconvolution, the directed graph of the gene-brain region generated by the step absorbs each edge. The structural information of edge diffusion within the order neighborhood is used to obtain the updated weight matrix. The goal of diffusing deconvolution is to compute new weights for each possible directed edge in the graph. Let... For different nodes, for slave nodes To the node Target edge, traverse all nodes The calculation formula is: = ； in, yes Bias of order-diffusion deconvolution; It is the first The learnable kernel matrix of order-order diffraction deconvolution. It is the gene in step S1 The number of nodes in a directed graph of a brain region; Indicates in In the order-generating graph, from node Pointing to node The weight of the directed edge; Indicates in In the order-generating graph, from node Pointing to node The weight of the directed edge; Represents a node To the node The diffusion coefficient, corresponding to The Middle Line 1 The value of the column; node To the node The diffusion coefficient, corresponding to The Middle Line 1 The value of the column; It is an activation function ; , They are from the nodes To the node target side go through and The updated value after diffusing deconvolution.

2. The method for extracting population features based on a large model of the pathological process according to claim 1, characterized in that, In step S1, causal relationships between gene sequences are calculated using Granger causality analysis, and directed edges are established if causal relationships exist; information similarity between brain region time series is calculated using lag correlation coefficient, and directed edges are established when the similarity is greater than a threshold; causal relationships of gene regulation on brain regions are calculated using transfer entropy, and directed edges are established if causal relationships are valid.

3. The method for extracting population features based on a large model of the pathological process according to claim 1, characterized in that, In step S2, the gene In the process of decomposing a directed graph of a brain region into at least two strongly connected components with no intersection, nodes with only incoming edges and nodes with only outgoing edges are first removed as strongly connected components. Then, a depth-first search algorithm is used to decompose the remaining strongly connected components. Isolated points after decomposition are also considered as strongly connected components.

4. The method for extracting population features based on a large model of the pathological process according to claim 1, characterized in that, In step S3, the feature information of the root node remains at its initial value throughout the inter-layer and intra-layer information transmission process of the directed tree structure, and the weight of the edge is calculated using the Pearson correlation coefficient between the feature information of the two connected nodes; and the feature information of the single node constituting the strongly connected component remains at its initial value.

5. The method for extracting population features based on a large model of the pathological process according to claim 4, characterized in that, Both inter-layer and intra-layer information transmission first use an aggregation method of "current node features + ∑(edge ​​weight × current directed neighbor node features that conform to inter-layer and intra-layer rules)" to obtain the aggregated feature vector. ∈ ; then, Right-multiply by the learnable parameter matrix shared by all directed tree structures. , ∈ The updated node features are obtained. ;matrix elements in The first feature represents the input feature. The i-th dimension of the output feature Linear contribution weights for each dimension For the input feature dimension, This is the output feature dimension.

6. The method for extracting population features of the pathological process based on a large model according to claim 1, characterized in that, Key genes The specific processing steps for carrying edge weight information in directed subgraphs of brain regions include: First, arrange the edge weight matrices of each key support tree diagonally and merge them into one matrix. Then, arrange the weight matrices of each root node according to their gene structure. The relationships in the directed graph of the brain region are established by creating connecting edges between the root nodes, and the edge weights between the root nodes are filled into the merged matrix to complete the fusion of the key support trees; the edge weights between the root nodes are all kept at their initial values.

7. The method for extracting population features of a large model based on the pathological process according to any one of claims 1 to 6, characterized in that, Also includes: Based on the diseased brain regions and risk genes, regulatory pathways were extracted along directed edges in the key support tree of each subject.

8. A system for extracting population features of a pathological process based on a large model, 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