A method and system for dense subgraph query mining in genome dynamic interaction networks

Abstract

The present invention discloses a dense subgraph query and mining method and system for genomic dynamic interaction networks, which relates to the field of genomic data analysis. The method comprises: constructing a genomic dynamic interaction network based on a permutation test method; distributing and storing the genomic dynamic interaction network based on a network storage model and a network partitioning strategy; constructing an index structure for the stored genomic dynamic interaction network; the index structure comprises: a time dimension index and a hierarchical structure index; querying and mining a k-point connected dense subgraph based on the index structure; the k-point connected dense subgraph is used to characterize unknown disease marker sites. The present invention studies complex diseases from the perspective of genomic interaction networks, and realizes the mining of unknown disease marker sites by constructing and analyzing a k-point connected dense subgraph.

Description

Technical Field

[0001] The present invention relates to the field of genome data analysis, and in particular to a dense subgraph query and mining method and system for genome dynamic interaction networks. Background Art

[0002] In the field of complex disease research, genome sequencing data is one of the main bases for complex disease research, and single nucleotide polymorphism (SNP) is the main form of genome sequencing data, that is, the mutation of a single nucleotide (A, C, G and T) on the chromosome sequence, which is widely present in the population. With the development of high-throughput sequencing technology, the cost of genome sequencing has dropped rapidly, making it possible to study complex diseases from a genomic perspective. The main research method at present is genome-wide association study (GWAS), which compares the patient group and the normal group to find SNP sites associated with complex diseases.

[0003] However, GWAS has some problems. Association analysis based on single SNP sites will encounter serious genetic missing problems. As the research deepens, the current stage mainly studies the association between SNP site interactions and complex diseases.

[0004] Complex disease genome dynamic interaction networks are characterized by large network scale, diverse site roles, synergistic interaction relationships, and local structural variability. The existing two-site genome association analysis technology is not sufficient to support the accurate construction of large-scale genome dynamic interaction networks. Although dense subgraph query and mining technology in large graphs has been widely studied and a large number of excellent research results have been achieved, these results often do not take into account the characteristics of diverse site roles and volatile local structures in genome dynamic interaction networks. In addition, the research on k-point connected dense subgraphs has just started, and the understanding of subgraph properties is not deep enough, resulting in the inability of existing large graph management technology and dense subgraph query and mining algorithms to be directly applied to the query and mining of k-point connected dense subgraphs for complex disease genome dynamic interaction networks, affecting the mining of unknown disease marker sites, where k represents the number of vertices in the dense subgraph. Summary of the invention

[0005] Based on this, an embodiment of the present invention provides a dense subgraph query and mining method and system for genomic dynamic interaction networks, which studies complex diseases from the perspective of genomic interaction networks and mines unknown disease marker sites by constructing and analyzing k-point connected dense subgraphs.

[0006] To achieve the above purpose, the embodiment of the present invention provides the following solution:

[0007] A dense subgraph query and mining method for genomic dynamic interaction networks, comprising:

[0008] A genomic dynamic interaction network is constructed based on a permutation test method; the genomic dynamic interaction network uses SNP sites as vertices and the strength of association between site interactions and diseases as edges;

[0009] Based on a network storage model and a network partitioning strategy, the genome dynamic interaction network is distributedly stored; the network storage model includes: a distributed file system and a distributed memory system; the network partitioning strategy is determined based on the partitioning method of the minimum point cut set;

[0010] Constructing an index structure for the stored genomic dynamic interaction network; the index structure includes: a time dimension index and a hierarchical structure index;

[0011] Based on the index structure, a k-point connected dense subgraph is queried and mined; the k-point connected dense subgraph is used to characterize unknown disease marker sites; k represents the number of vertices of the dense subgraph.

[0012] Optionally, a genomic dynamic interaction network is constructed based on a permutation test method, specifically including:

[0013] Access to genomic data;

[0014] Extracting SNP sites from the genomic data;

[0015] The chi-square test method is used to compare the distribution differences of SNP site pairs between disease patients and healthy controls, and the permutation test method is used to determine the site pair p value; the site pair p value represents the significance of the association of site interaction; the site pair p value is used as the weight of the edge;

[0016] Constructing a preliminary genomic dynamic interaction network according to the SNP sites and the site-pair p-values;

[0017] Site interaction prediction is performed based on domain knowledge and graph embedding technology to add edges to the preliminary genome dynamic interaction network to obtain a refined genome dynamic interaction network;

[0018] The refined genome dynamic interaction network was taken as the final genome dynamic interaction network.

[0019] Optionally, a permutation test method is used to determine the site pair p value, specifically including:

[0020] Based on the site interaction reduction strategy with a flexible lower bound or the site interaction parallel calculation acceleration strategy, a permutation operation is performed on each SNP site pair to calculate the site pair p-value.

[0021] Optionally, based on a network storage model and a network partitioning strategy, the genome dynamic interaction network is distributedly stored, specifically including:

[0022] A hierarchical storage strategy of a distributed file system and a distributed memory system is used to hierarchically store the genome dynamic interaction network;

[0023] The network data in the genomic dynamic interaction network is divided by a method based on a minimum point cut set, and the vertices with connectivity higher than a set value in the hierarchically stored genomic dynamic interaction network are stored in the same storage node to realize distributed storage of the genomic dynamic interaction network.

[0024] Optionally, a hierarchical storage strategy of a distributed file system and a distributed memory system is used to hierarchically store the genome dynamic interaction network, specifically including:

[0025] The dynamic interaction network of the genome after a set time is represented as a vertex set and an edge set;

[0026] Use hashing technology to distribute all vertices in the vertex set to storage nodes in the distributed memory system;

[0027] Allocate all edges in the edge set to corresponding storage nodes in the distributed memory system according to the mapping relationship between points and edges;

[0028] For the genomic dynamic interaction network before a set time, snapshots are generated regularly in chronological order, and the historical snapshot records are stored in a distributed file system.

[0029] Optionally, constructing an index structure for the stored genomic dynamic interaction network specifically includes:

[0030] Use an extensible B+ tree index structure to build a time dimension index for the master node connected to the storage node;

[0031] A hierarchical index is constructed for the storage nodes in the order of k values ​​from small to large; the k value represents the dispersion of the storage nodes.

[0032] Optionally, querying and mining a k-point connected dense subgraph based on the index structure specifically includes:

[0033] If the index structure is online, a single query point query, a multi-query point query or a top-r query is performed based on the index structure to obtain a preliminary k-point connected dense subgraph;

[0034] If the index structure is offline, a single query point local expansion method is used to perform a single query point query or a k-core filtering-based method is used to perform a multi-query point query to obtain a preliminary k-point connected dense subgraph;

[0035] The preliminary k-point connected dense subgraph is subjected to comparative subgraph mining and multi-layer subgraph mining to obtain the final k-point connected dense subgraph.

[0036] Optionally, after constructing the index structure for the stored genomic dynamic interaction network, the method further includes:

[0037] The index structure is maintained.

[0038] Optionally, maintaining the index structure specifically includes:

[0039] Maintaining the time dimension index using a life cycle strategy;

[0040] A batch processing strategy is adopted to maintain the hierarchical index.

[0041] The present invention also provides a dense subgraph query and mining system for genome dynamic interaction networks, comprising:

[0042] An interaction network construction module is used to construct a genomic dynamic interaction network based on a permutation test method; the genomic dynamic interaction network uses SNP sites as vertices and the strength of association between site interactions and diseases as edges;

[0043] A distributed storage module is used to perform distributed storage on the genome dynamic interaction network based on a network storage model and a network partitioning strategy; the network storage model includes: a distributed file system and a distributed memory system; the network partitioning strategy is determined based on a partitioning method of a minimum point cut set;

[0044] An index building module is used to build an index structure for the stored genomic dynamic interaction network; the index structure includes: a time dimension index and a hierarchical structure index;

[0045] A subgraph query and mining module is used to query and mine a k-point connected dense subgraph based on the index structure; the k-point connected dense subgraph is used to characterize unknown disease marker sites; k represents the number of vertices of the dense subgraph.

[0046] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0047] The embodiment of the present invention constructs a genomic dynamic interaction network based on a permutation test method, performs distributed storage of the genomic dynamic interaction network based on a network storage model and a network partitioning strategy, constructs an index structure for the stored genomic dynamic interaction network, and finally searches based on the index structure and mines a k-point connected dense subgraph, thereby mining unknown disease marker sites. The embodiment of the present invention studies complex diseases from the perspective of a genomic interaction network, realizes the mining of unknown disease marker sites by constructing and analyzing a k-point connected dense subgraph, improves the explanation of the pathogenesis of complex diseases, and provides a basis for prevention, diagnosis, and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0049] Figure 1 A flowchart of a dense subgraph query and mining method for a genomic dynamic interaction network provided by an embodiment of the present invention;

[0050] Figure 2 A flowchart of the construction process of a preliminary genomic dynamic interaction network provided in an embodiment of the present invention;

[0051] Figure 3 A flowchart of the construction process of a refined genomic dynamic interaction network provided by an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of a time-hierarchy dual-layer index structure provided by an embodiment of the present invention;

[0053] Figure 5 A schematic diagram of the relationship between the various parts of the overall solution provided by the embodiment of the present invention;

[0054] Figure 6 A structural diagram of a dense subgraph query and mining system for genomic dynamic interaction networks provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present invention will be described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0056] Existing methods have achieved certain research results in the construction, management, and dense subgraph query and mining of dynamic genomic interaction networks, but there are also many shortcomings. In terms of network construction, the existing two-site genome association analysis technology cannot effectively support the precise measurement of the significance of SNP site interactions and complex disease associations under large sample conditions. In terms of network management, the existing large-graph data management technology does not take into account the characteristics of diverse roles of genomic interaction network sites, collaborative interaction relationships, and volatile local structures, resulting in the inability of existing large-graph storage indexes and index maintenance technologies to directly support efficient query and mining of dense subgraphs. Specifically:

[0057] (1) In terms of the construction of dynamic interaction networks of complex disease genomes, existing research results have failed to achieve accurate measurement of the association between site interactions and complex diseases. Site interactions and the association between complex diseases are usually measured using statistical methods, and the huge number of interactions leads to serious false positive problems in multiple hypothesis testing results. In order to improve the quality of genomic interaction network construction, permutation tests are used to correct false positive results caused by multiple tests, but the large sample size causes a heavy computational burden. Therefore, how to design a fast multiple hypothesis testing correction method to achieve accurate measurement of site interactions and the association between complex diseases is an urgent problem to be solved.

[0058] (2) In terms of storage, indexing and maintenance of dynamic interaction networks of complex disease genomes, existing research results have failed to combine the characteristics of genome interaction networks for dynamic network management. Existing large-scale graph data management technologies fail to fully consider the serious data coupling caused by the diversity of network site roles and the synergy of interaction relationships in network storage and index construction. In addition, the variability of the local structure of the network leads to the inability of existing index structures to effectively manage the temporal attributes of data and the high cost of index maintenance. Therefore, how to design distributed data storage, index construction and index maintenance strategies suitable for the characteristics of genome dynamic interaction networks to achieve dynamic network management is an extremely challenging problem.

[0059] (3) In the query and mining of dynamic interaction networks of complex disease genomes, existing technologies cannot be directly used to solve the problem of efficient query and mining of k-point connected dense subgraphs. In the query and mining of genome interaction networks, there are currently only related studies on the mining of in-memory k-point connected dense subgraphs for static large graphs, but no consideration of the query and mining of k-point connected dense subgraphs in a dynamic large graph distributed environment due to the large scale of the network and the volatility of local structure. Moreover, the process of discovering k-point connected dense subgraphs involves a large number of vertex connectivity calculations and frequent data communications. Therefore, it is very necessary to design efficient query and mining of k-point connected dense subgraphs for dynamic genome interaction networks.

[0060] The purpose of the present invention is: first, the network is large in scale, and the problem of multiple tests easily leads to false positive results, so this embodiment solves the problem of accurate network construction; secondly, the dynamic maintenance of the network involves optimal graph partitioning and new index structure design to deal with the data coupling problem caused by the diversity of site roles and the synergy of interactive relationships; finally, the efficient discovery of k-point connected dense subgraphs is also a challenge, which can solve the performance bottleneck problem caused by frequent calculation of vertex connectivity. This embodiment aims to study complex diseases from the perspective of genomic interaction networks, and by constructing and analyzing k-point connected dense subgraphs, excavating unknown disease marker sites, and improving the explanation of the pathogenesis of complex diseases, providing a basis for prevention, diagnosis and treatment.

[0061] This embodiment proposes a method for constructing and analyzing a genomic dynamic interaction network, which uses SNP sites as network vertices, and site interactions and complex disease association strengths as network internal edges to describe the association between SNP sites from a global perspective. In order to better reflect the diversity of site roles and the synergy of interaction relationships, this embodiment introduces a k-point connected dense subgraph, which can reflect the synergy of interaction relationships in the network and also characterize the diversity of sites, because the same site may appear in different dense subgraphs.

[0062] Dense subgraph query algorithms mainly include dense subgraph queries on static large graphs and dynamic large graphs. The static large graph dense subgraph query algorithms mainly include dense subgraph queries based on query points and top-r dense subgraph queries based on result quantity constraints. For query based on query points, the user specifies one or more nodes in the graph as query content, and the query returns the dense subgraph containing these query nodes from the graph online. If the user does not know much about the structure and content of the graph, or wants to explore some patterns or relationships, it may be difficult to accurately select query nodes; incorrect or inappropriate query node selection may result in the returned subgraph having no practical significance. For top-r dense subgraph query, the aim is to find the top-r results that best meet a certain dense subgraph query constraint. If the query constraint is strict or the graph data scale is large, it may not be possible to obtain a sufficient number of results, thus affecting the comprehensive understanding of the graph; when it is necessary to select appropriate query constraints, different constraint selections may lead to different types of results, and selecting the correct constraint may require domain knowledge or a trial and error process.

[0063] Dense subgraph mining in dynamic large graphs refers to discovering dense subgraphs that meet given constraints from dynamic large graphs. Research on dynamic large graphs can be mainly divided into three categories: focusing only on changes at a certain moment; focusing on changes between the previous and next moments; and focusing on changes within a certain period of time. Only focusing on changes at a single moment may lead to an incomplete understanding of the entire dynamic process and a lack of global insight; focusing on changes between the previous and next moments cannot capture evolutionary trends on a longer time scale, and may not be sensitive enough to emergencies in certain situations; the choice of time period may affect the interpretability of the results. Too short a time period may not capture long-term trends, while too long a time period may lead to information loss.

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] See also Figure 1 The dense subgraph query and mining method for the genomic dynamic interaction network of this embodiment includes:

[0066] Step 101: construct a genomic dynamic interaction network based on a permutation test method; the genomic dynamic interaction network uses SNP sites as vertices and uses the strength of association between site interactions and diseases as edges.

[0067] The step 101 specifically includes:

[0068] (1) Obtain genomic data.

[0069] (2) Extract the SNP sites of the genomic data as the vertices of the graph.

[0070] (3) The chi-square test method is used to compare the distribution differences of SNP site pairs (two SNP sites) between disease patients and healthy controls, and the permutation test method is used to determine the site pair p-value; the site pair p-value represents the significance of the association between the site interactions; the site pair p-value is used as the weight of the edge.

[0071] Among them, performing a permutation operation on each SNP site pair to calculate the site pair p value has a very high computational and storage cost. Therefore, a permutation test method is used to determine the site pair p, specifically including: after performing a permutation test, a site interaction reduction strategy based on a flexible lower bound or a site interaction parallel calculation acceleration strategy is used to perform a permutation operation on each SNP site pair to calculate the site pair p value.

[0072] ① Site interaction reduction strategy based on flexible lower bound: The site pair p value of the statistic has a monotonic relationship with the frequency of pattern occurrence. The upper bound of the statistic is only related to the sample support of the pattern, but has nothing to do with the order of the class labels. Therefore, it is not necessary to calculate the statistics on all permutations to estimate the upper bound of the pattern P, and all patterns with the same support have the same upper bound. In the calculation of the pattern statistics, the lower bound of the current significant level is continuously improved to narrow the interval of pattern support, so that many patterns do not need to be permuted and the purpose of pattern reduction is achieved. Therefore, the site interaction reduction strategy based on the flexible lower bound can greatly speed up the calculation of the permutation test without affecting the accuracy of the results.

[0073] ② Site interaction parallel calculation acceleration strategy: Since the calculation of site interaction and complex disease association measurement is loosely coupled, that is, the calculation of the current two-site statistical value does not depend on other sites or previous calculation results. Therefore, a parallel strategy is used to accelerate the calculation. First, the genomic data is divided into M pieces, so each segment has SNP sites. Then, the chi-square test method is used to calculate the statistical values ​​of the interaction between SNP sites within the shard and the statistical values ​​of the interaction between SNP sites between shards in parallel. It is worth noting that when calculating the statistical values ​​between shards, the communication order between shards needs to be set, and every two shards only communicate once, and ensure that as many shards as possible can communicate at the same time. Finally, by combining all the calculation results, the final statistically significant interaction pattern can be found.

[0074] (4) A preliminary genome dynamic interaction network is constructed based on the SNP sites and the site-pair p-values. Specifically, the SNP sites are used as graph vertices, and the site interactions and complex disease association strengths are used as graph edges to construct a complex disease genome interaction network. The construction process of the preliminary genome dynamic interaction network is as follows: Figure 2 shown.

[0075] (5) Site interaction prediction is performed based on domain knowledge and graph embedding technology to add edges to the preliminary genome dynamic interaction network to obtain a refined genome dynamic interaction network, and the refined genome dynamic interaction network is used as the final genome dynamic interaction network to achieve accurate construction of the genome dynamic interaction network. The construction process of the refined genome dynamic interaction network is as follows: Figure 3 As shown. Specifically:

[0076] ① Site interaction prediction based on domain knowledge.

[0077] As domain knowledge, biological pathways can guide organisms to complete a specific biological function. Since complex diseases are often caused by the combined action of pathways, a prediction method based on biological pathways is used, which is mainly divided into the following three steps: first, SNP sites are extracted for each biological pathway to be detected and the corresponding SNP site set is constructed; then, the correlation between the interaction of SNP sites and complex diseases in each two site sets is detected. This process is equivalent to constructing a relationship between the sites on both sides of the bipartite graph. Specifically, the discriminative pattern mining method is used to find SNP site pairs that appear many times in one type of sample and few times in another type of sample. Finally, the relationship between SNP sites with significant discriminative ability is added to the genome interaction network.

[0078] ② Site interaction prediction based on graph embedding technology.

[0079] First, using the genome interaction network data, nodes represent SNP sites, edges represent the interaction relationship between SNP sites, and construct an adjacency matrix to represent the connection between each SNP site and other sites; according to the mapping function f: Each site is mapped to a low-dimensional feature vector; the random walk sampling method is used to obtain the neighbor set of a given site, where random walk sampling includes two strategies: breadth-first and depth-first. Breadth-first sampling can obtain information about directly connected sites, and depth-first sampling can capture long-range paths near the site; then the maximum likelihood estimation method is used to model the probability of occurrence of the neighbor set of a given site, and the sampled neighbor set and the estimated parameter matrix are used to optimize the parameters by maximizing the likelihood function to obtain the optimized parameter matrix; then the mapping function and the parameter matrix are used to calculate the feature vector of each site; the cosine similarity between sites is calculated using the feature vector of the site, and the cosine similarity is used to measure the similarity between sites; according to the preset similarity threshold θ, if the similarity is greater than the threshold θ, there is an interactive relationship between the sites, and edges can be added to the genomic interaction network.

[0080] In the above step 101, multiple hypothesis testing techniques are first used to preliminarily discover SNP site interactions that are statistically significantly associated with complex diseases and to achieve effective measurement of site interaction relationships, and then domain knowledge-based and graph embedding techniques are used to achieve a detailed construction of SNP site interaction relationships. The construction of a complex disease genome dynamic interaction network in this step provides a theoretical guarantee for efficient dense subgraph query and mining.

[0081] Step 102: Based on the network storage model and the network partitioning strategy, the genome dynamic interaction network is distributedly stored.

[0082] The network storage model includes: a distributed file system and a distributed memory system; the network partitioning strategy is determined based on the partitioning method of the minimum point cut set.

[0083] Step 102 specifically includes:

[0084] (1) A hierarchical storage strategy of a distributed file system and a distributed memory system is used to hierarchically store the genome dynamic interaction network. Specifically:

[0085] The genomic dynamic interaction network after a set time is represented as a vertex set and an edge set; all vertices in the vertex set are assigned to storage nodes in the distributed memory system using hash technology; all edges in the edge set are assigned to corresponding storage nodes in the distributed memory system according to the mapping relationship between points and edges; for the genomic dynamic interaction network before a set time, snapshots are generated regularly in chronological order, and historical snapshot records are stored in a distributed file system.

[0086] (2) The network data in the genomic dynamic interaction network is divided by a method based on a minimum point cut set, and the vertices with a connectivity higher than a set value in the hierarchically stored genomic dynamic interaction network are stored in the same storage node to realize the distributed storage of the genomic dynamic interaction network.

[0087] Appropriate data storage strategies can often improve the performance of data query and data mining. In practical applications, the network storage model and network partitioning strategy are as follows:

[0088] ① Network storage model. A hierarchical storage strategy of distributed file system and distributed memory system is used to realize the storage of genome interaction network. Since the topological structure of genome interaction network changes over time, the distributed memory system is used to store the newly generated genome interaction network for easy query and update. Specifically, the key-value structure is used to store network data, and the network is represented as a vertex set and an edge set. The hash technology is used to assign all vertices to storage nodes, and then the edges are assigned to the corresponding storage nodes according to the mapping relationship between points and edges; for the earlier genome interaction network, snapshots are generated regularly in the order of time, and the historical snapshot records are stored in the distributed file system. The advantage of this is that when comparative snapshot mining and multi-layer snapshot mining are performed later, the snapshots involved can be called into the memory system for processing at any time according to the time index.

[0089] ② Network partitioning strategy. The network data is partitioned based on the minimum point cut set to achieve distributed storage of the genomic dynamic interaction network. According to the properties of the k-point connected dense subgraph, vertices with high inter-vertex connectivity are more likely to be in the same k-point connected dense subgraph. In order to achieve efficient query and mining of k-point connected dense subgraphs, vertices with high connectivity need to be stored in the same storage node.

[0090] First, read the network data into the distributed memory, and then find the connected branches in the network. If the connected branches can be stored in the same storage node as a whole, they are directly stored; otherwise, find the minimum point cut set in the connected branch, and divide the connected branch into multiple subgraphs by removing the vertices in the minimum point cut set, that is, use the minimum point cut set to divide the current subgraph. And so on, until all subgraphs can be stored in the same storage node as a whole. Using the network minimum point cut set for partitioning can store the vertices in the dense subgraph with high point connectivity in the same storage node as much as possible, effectively reducing the communication cost of query and mining.

[0091] Step 103: constructing an index structure for the stored genomic dynamic interaction network; the index structure includes: a time dimension index and a hierarchical structure index.

[0092] The index structure is the cornerstone for supporting efficient query and mining of k-point connected dense subgraphs in the dynamic interaction network of complex disease genomes. This embodiment designs a time-hierarchical double-layer index structure, such as Figure 4 As shown, Figure 4 Part (a) shows the overall structure of the index. Figure 4 Part (b) shows the time dimension index, Figure 4 Part (c) shows the hierarchical index. The master node contains a time dimension index, based on which the snapshots of the genome interaction network corresponding to different moments can be located. Since the network structure divisions may be stored in multiple storage nodes, each moment may correspond to multiple storage nodes; the storage node contains a hierarchical index, which can effectively organize the network divisions in the storage node and realize efficient query and mining of k-point connected dense subgraphs.

[0093] Step 103 specifically includes:

[0094] (1) Time dimension index. Since the topology of the genomic dynamic interaction network changes dynamically with time, the index structure must reflect the time dimension of the network. The master node connected to the storage node uses an extensible B+ tree index structure to construct a time dimension index for the snapshots of the genomic interaction network in chronological order. The details are as follows: ① The time series data of the genomic interaction network is organized into a series of snapshots, each of which represents the network topology at a specific time point. ② A main B+ tree index is constructed, and each node contains a key value (timestamp) and a pointer to the corresponding time snapshot. ③ For each time snapshot, its timestamp is used as the key value, and the pointer to the time snapshot is inserted into the main B+ tree index. The insertion process will maintain the order of the B+ tree. When it is necessary to query the network snapshot within a specific time range, the main B+ tree index can be used to locate the leaf nodes that meet the time range. These leaf nodes store pointers to the time snapshots. Through the pointers stored in the leaf nodes, the corresponding time snapshot data can be retrieved from the storage to obtain the network topology. This method improves the efficiency of searching for network snapshots according to the time range, thereby supporting efficient k-point connected dense subgraph mining based on time snapshots.

[0095] (2) Hierarchical index. A hierarchical index is used in the storage nodes. The network partitions in the storage nodes are arranged in the order of k values ​​from small to large to construct a hierarchical tree index structure. The k value represents the dispersion of the storage nodes. The details are as follows: the network partitions in the storage nodes are arranged in the order of k values ​​from small to large; using the B+ tree structure, each node represents an index layer, which is used to organize dense subgraphs with different point connectivity. Each node of the B+ tree contains a keyword and a pointer to a child node; starting from the smallest k value, each network partition (dense subgraph) is inserted into the B+ tree; in the nodes of each index layer, a pointer to the next index layer is included to construct a hierarchical structure. Different layers in the index contain dense subgraphs with different point connectivity. It should be noted that the k-point connected dense subgraph has a hierarchical inclusion property. Therefore, the index compression algorithm can be used to further reduce the storage space of the index based on the inclusion relationship. That is, if the dense subgraph of the next layer has the same nodes as the dense subgraph of the previous layer, the same nodes in the next layer as in the previous layer will be deleted, thereby supporting efficient query and mining of k-point connected dense subgraphs.

[0096] In one example, after constructing the index structure, the method further includes: maintaining the index structure. The dynamic index maintenance strategy is a technical means to ensure efficient query of k-point connected dense subgraphs. The dynamic maintenance and update strategy of the time-hierarchical double-layer index is studied to achieve efficient query and mining of k-point connected dense subgraphs for complex disease dynamic genomic interaction networks.

[0097] Maintaining the index structure specifically includes: maintaining the time dimension index using a life cycle strategy; maintaining the hierarchical structure index using a batch processing strategy. Specifically:

[0098] ① Time dimension index maintenance. The life cycle strategy is used to maintain the time dimension index, as follows: regularly or according to preset conditions, check whether the currently processed interactive network has failed. If the network fails, migrate the current interactive network from the distributed memory system to the distributed file system; after the interactive network is migrated to the distributed file system, update the time dimension index to reflect the new network storage location; first, find the corresponding index item from the time dimension index, identify the time period when the network fails, and then update the storage location information of the index item to the storage path in the distributed file system. The cost of index maintenance can be controlled by reasonably setting parameters such as detection frequency and failure threshold. The currently processed interactive network is stored in the distributed memory system. When the network fails, it is migrated to the distributed file system for storage, and the corresponding time dimension index is updated.

[0099] ②Hierarchical index maintenance. A batch processing strategy is used to maintain the hierarchical index, as follows: within a period of time, the data where update operations (addition, deletion, modification, etc.) occur is recorded, and these update operations are not performed immediately; the update operations of the collected data are grouped according to certain conditions to form an update batch; when the update batch meets the set conditions, the processing operation of the update batch is triggered; and the update frequency of the hierarchical index is reduced. For each data update operation in the update batch, the corresponding index maintenance operation is performed according to the type of operation and the affected local structure. When performing index maintenance operations, by using delayed update technology, it is ensured that only update operations that affect the final index structure are performed. In addition, when the local structure changes too much, after cost analysis, the local structure can be rebuilt. According to the above two technologies, the cost of index maintenance can be greatly reduced.

[0100] In the above steps 102 and 103, the distributed storage strategy of the genome dynamic interaction network is first introduced, and the network distributed storage is realized through the network storage model and partitioning strategy; the efficient query and mining of the k-point connected dense subgraph is realized through the index construction and dynamic maintenance and update of the genome dynamic interaction network. The complex disease genome dynamic interaction network management in this step provides technical support for the efficient query and mining of the k-point connected dense subgraph.

[0101] Step 104: query and mine a k-point connected dense subgraph based on the index structure; the k-point connected dense subgraph is used to characterize unknown disease marker sites.

[0102] k represents the number of vertices in the dense subgraph. The k-point connected dense subgraph can effectively model the dense regions closely related to complex diseases in the genomic interaction network. The k-point connected dense subgraph query can be used to retrieve the regions related to complex diseases. Subsequently, the k-point connected dense subgraph mining can be used to conduct in-depth analysis of the causes of complex diseases.

[0103] Step 104 specifically includes:

[0104] (1) If the index structure is online, an index-based subgraph query is implemented. The use of complex disease genome interaction network index structures can often accelerate the query of k-point connected dense subgraphs, thereby achieving efficient retrieval of complex disease-related regions. This step is studied from three perspectives: single query point query, multiple query point query, and top-r query, to support efficient query of k-point connected dense subgraphs, that is, single query point query, multiple query point query, or top-r query is performed based on the index structure to obtain a preliminary k-point connected dense subgraph. Top query is a common database query operation, which can filter out the first few data that meet the conditions according to a certain condition, and r represents the first r data that meet the conditions.

[0105] ① Single query point query. By traversing the current corresponding hierarchical index, we can find the point-connected dense subgraph that contains the query point and maximizes the k value. Using the k-point connected dense subgraph to prune the query path will greatly improve the query efficiency.

[0106] ②Multi-query point query. Traditional multi-query point query methods often assume that all query points belong to the same subgraph, but this is not the case. First, the query points are adaptively grouped using the truncated random walk technique. Then, for each query point in each group, a dense subgraph containing the points with the maximum k value is found. During the search process, the multi-query point query can be accelerated by finding the minimum common ancestor containing all query points in the group in the hierarchical index structure.

[0107] ③Top-r query. By traversing the index structure, we find all candidate sets of k-point connected dense subgraphs that meet the conditions, and then select the top-r results that best meet the conditions from the candidate set. In order to improve the query efficiency, we derive the lower bound of the top-r subgraph constraint by using the properties of the k-point connected dense subgraph, and prune the search space based on the lower bound, thereby narrowing the candidate result set and improving the efficiency of the top-r query.

[0108] (2) If the index structure is offline, a non-index-based subgraph query is implemented. Since the genome dynamic interaction network index may be unavailable offline, the non-index-based k-point connected dense subgraph query algorithm in this step supports efficient online query of k-point connected dense subgraphs. Specifically, a single query point local expansion method is used for single query point query or a k-core filtering-based method is used for multi-query point query to obtain a preliminary k-point connected dense subgraph.

[0109] ① Query point query. Using the single query point local expansion method, by visiting the neighbors of the query point, determine the range of the connected dense subgraph that contains the query point and maximizes the k value. Next, the query result is finally found by continuously tightening the subgraph.

[0110] ②Multi-query point query. First, the query points are grouped, and then subgraph queries are performed on the query points in each group. For the query points in the same group, the k-core filtering method is used to find the k-core connected subgraph that contains the query points in the group and makes k the largest. Then, using the properties of the k-point connected dense subgraph, the vertices that do not meet the conditions are continuously deleted from the subgraph, and finally the k-point connected dense subgraph that contains the query point set and makes the k value the largest is found.

[0111] (3) Perform comparative subgraph mining and multi-layer subgraph mining on the preliminary k-point connected dense subgraph to obtain the final k-point connected dense subgraph.

[0112] The dynamic genome interaction network has a time-varying topological structure. Mining k-point connected dense subgraphs from the dynamic genome interaction network can reflect the changing patterns of regions closely related to complex diseases, thereby helping to analyze the causes of complex diseases.

[0113] ① Contrast subgraph mining. Construct a difference network that can reflect the significant changes between the previous and next network time snapshots. Specifically, obtain the previous and next network time snapshots, each snapshot represents a genome interaction network at a time point; for each network time snapshot, calculate the difference with the network at the previous time point by comparing the addition, deletion and attribute changes of nodes; construct a difference network, in which the newly added nodes and edges are regarded as differences, and the deleted nodes and edges are also regarded as differences; filter the nodes and edges in the difference network, and only retain those nodes and edges that have significant changes in the difference network. The hierarchical property of the k-point connected dense subgraph requires that the connectivity between the vertices in the subgraph meets certain requirements. This property is used to continuously peel off the subgraph areas with low connectivity in the network according to the hierarchy of the network, and finally accurately find the maximum k-point connected dense subgraph, also called the contrast subgraph. In order to further improve the efficiency of subgraph mining, an approximate contrast subgraph mining algorithm based on the relaxation and tightening strategy is designed. By first relaxing the constraints on the connectivity of the subgraph, a coarse-grained candidate contrast subgraph is found, and then the connectivity requirements are tightened to extract fine-grained candidate results from the previously discovered candidate subgraphs.

[0114] ② Multi-layer subgraph mining. First, the life cycle of vertices and edges in the multi-layer network snapshot is determined according to the time attribute, and then the vertices and edges in the multi-layer network snapshot are reduced using the given continuous appearance time or periodic change threshold to obtain a coarse-grained candidate subgraph. Next, according to the properties of the k-point connected dense subgraph, the degree of the vertex and the connectivity between the vertices are used as conditions to perform a branch and bound search on the candidate subgraph until a multi-layer maximal k-point connected dense subgraph that meets the appearance time and subgraph structure constraints is found.

[0115] Starting from the key issues of complex disease genome dynamic interaction network construction, storage indexing and maintenance, and dense subgraph query and mining, this embodiment studies the relevant theories, technologies and methods for k-point connected dense subgraph query and mining for genome dynamic interaction network, and provides new solutions for the analysis and diagnosis of complex diseases. The construction of genome dynamic interaction network is the basis of research. By measuring the correlation between SNP site interaction and complex diseases, the interaction network is constructed to provide a data basis for the subsequent genome dynamic interaction network management; the storage indexing and maintenance of the genome interaction network is the core of dynamic network management, providing technical support for dense subgraph query and mining; the query and mining of k-point connected dense subgraph is the ultimate research goal, allowing users to efficiently query and mine areas closely related to complex diseases, providing assistance for the analysis and diagnosis of complex diseases. The properties of k-point connected dense subgraphs always run through the network management and dense subgraph query and mining process. As Figure 5 As shown, the overall technical concept of this embodiment includes: genome dynamic interaction network construction, genome dynamic interaction network management, and k-point connected dense subgraph query and mining.

[0116] In order to execute the methods corresponding to the above embodiments and achieve corresponding functions and technical effects, a dense subgraph query and mining system for genomic dynamic interaction networks is provided below.

[0117] See also Figure 6 , the system comprising:

[0118] The interaction network construction module 601 is used to construct a genomic dynamic interaction network based on the permutation test method; the genomic dynamic interaction network uses SNP sites as vertices and the association strength between site interactions and diseases as edges.

[0119] The distributed storage module 602 is used to perform distributed storage on the genome dynamic interaction network based on a network storage model and a network partitioning strategy; the network storage model includes: a distributed file system and a distributed memory system; the network partitioning strategy is determined based on the partitioning method of the minimum point cut set.

[0120] The index building module 603 is used to build an index structure for the stored genome dynamic interaction network; the index structure includes: a time dimension index and a hierarchical structure index.

[0121] The subgraph query and mining module 604 is used to query and mine a k-point connected dense subgraph based on the index structure; the k-point connected dense subgraph is used to characterize unknown disease marker sites; k represents the number of vertices of the dense subgraph.

[0122] All the above embodiments have the following advantages: achieving accurate measurement of the association between SNP site interaction and complex disease; studying the storage and indexing of genome interaction networks, designing network storage models, indexing structures and dynamic maintenance strategies; studying k-point connected dense subgraph query and mining algorithms, achieving efficient subgraph query and mining; and designing and implementing a k-point connected dense subgraph query and mining prototype system for complex disease genome dynamic interaction networks. In this system, the distributed management part is responsible for the maintenance of data storage nodes and the transmission of messages between nodes; the data management part is responsible for the construction of the network, the storage and indexing of the network; the query and mining part implements the query and mining algorithm of the k-point connected dense subgraph, performs analysis, diagnosis and prediction of complex diseases and outputs visualization results.

[0123] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0124] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A dense subgraph query and mining method for genomic dynamic interaction networks, characterized in that: include: Constructing a genomic dynamic interaction network based on the permutation test method; The genome dynamic interaction network uses SNP sites as vertices and the strength of association between site interactions and diseases as edges; Based on a network storage model and a network partitioning strategy, the genome dynamic interaction network is distributedly stored; The network storage model includes: a distributed file system and a distributed memory system; the network partitioning strategy is determined based on the partitioning method of the minimum point cut set; Constructing an index structure for the stored genomic dynamic interaction network; the index structure includes: a time dimension index and a hierarchical structure index; Based on the index structure, query and mine a k-point connected dense subgraph; the k-point connected dense subgraph is used to characterize unknown disease marker sites; k represents the number of vertices of the dense subgraph; Based on the network storage model and network partitioning strategy, the genome dynamic interaction network is distributedly stored, specifically including: A hierarchical storage strategy of a distributed file system and a distributed memory system is used to hierarchically store the genome dynamic interaction network; The network data in the genome dynamic interaction network is divided by a method based on a minimum point cut set, and the vertices with a connectivity higher than a set value in the hierarchically stored genome dynamic interaction network are stored in the same storage node to realize the distributed storage of the genome dynamic interaction network; The hierarchical storage strategy of the distributed file system and the distributed memory system is used to hierarchically store the genome dynamic interaction network, specifically including: The dynamic interaction network of the genome after a set time is represented as a vertex set and an edge set; Use hashing technology to distribute all vertices in the vertex set to storage nodes in the distributed memory system; Allocate all edges in the edge set to corresponding storage nodes in the distributed memory system according to the mapping relationship between points and edges; For the genomic dynamic interaction network before a set time, snapshots are generated regularly in chronological order, and the historical snapshot records are stored in a distributed file system.

2. The dense subgraph query and mining method for genomic dynamic interaction networks according to claim 1 is characterized in that: The genomic dynamic interaction network is constructed based on the permutation test method, including: Access to genomic data; Extracting SNP sites from the genomic data; The chi-square test method is used to compare the distribution differences of SNP site pairs between disease patients and healthy controls, and the permutation test method is used to determine the site pair p value; the site pair p value represents the significance of the association of site interaction; the site pair p value is used as the weight of the edge; Constructing a preliminary genomic dynamic interaction network according to the SNP sites and the site-pair p-values; Site interaction prediction is performed based on domain knowledge and graph embedding technology to add edges to the preliminary genome dynamic interaction network to obtain a refined genome dynamic interaction network; The refined genome dynamic interaction network was taken as the final genome dynamic interaction network.

3. The dense subgraph query and mining method for genomic dynamic interaction networks according to claim 2 is characterized in that: The permutation test method is used to determine the site pair p value, including: Based on the site interaction reduction strategy with a flexible lower bound or the site interaction parallel calculation acceleration strategy, a permutation operation is performed on each SNP site pair to calculate the site pair p-value.

4. The dense subgraph query and mining method for genomic dynamic interaction networks according to claim 1 is characterized in that: Construct an index structure for the stored genomic dynamic interaction network, including: Use an extensible B+ tree index structure to build a time dimension index for the master node connected to the storage node; A hierarchical index is constructed for the storage nodes in the order of k values ​​from small to large; the k value represents the dispersion of the storage nodes.

5. The dense subgraph query and mining method for genomic dynamic interaction networks according to claim 1 is characterized in that: Querying and mining k-point connected dense subgraphs based on the index structure specifically includes: If the index structure is online, a single query point query, a multi-query point query or a top-r query is performed based on the index structure to obtain a preliminary k-point connected dense subgraph; If the index structure is offline, a single query point local expansion method is used to perform a single query point query or a k-core filtering-based method is used to perform a multi-query point query to obtain a preliminary k-point connected dense subgraph; The preliminary k-point connected dense subgraph is subjected to comparative subgraph mining and multi-layer subgraph mining to obtain the final k-point connected dense subgraph.

6. The dense subgraph query and mining method for genomic dynamic interaction networks according to claim 1 is characterized in that: After constructing the index structure for the stored genomic dynamic interaction network, it also includes: The index structure is maintained.

7. The dense subgraph query and mining method for genomic dynamic interaction networks according to claim 6 is characterized in that: Maintaining the index structure specifically includes: Maintaining the time dimension index using a life cycle strategy; A batch processing strategy is adopted to maintain the hierarchical index.

8. A dense subgraph query and mining system for genomic dynamic interaction networks, characterized in that: include: Interaction network construction module, used to construct genomic dynamic interaction network based on permutation test method; The genome dynamic interaction network uses SNP sites as vertices and the strength of association between site interactions and diseases as edges; A distributed storage module, used for performing distributed storage of the genome dynamic interaction network based on a network storage model and a network partitioning strategy; The network storage model includes: a distributed file system and a distributed memory system; the network partitioning strategy is determined based on the partitioning method of the minimum point cut set; An index building module is used to build an index structure for the stored genomic dynamic interaction network; the index structure includes: a time dimension index and a hierarchical structure index; A subgraph query and mining module, used for querying and mining a k-point connected dense subgraph based on the index structure; the k-point connected dense subgraph is used to characterize unknown disease marker sites; k represents the number of vertices of the dense subgraph; Based on the network storage model and network partitioning strategy, the genome dynamic interaction network is distributedly stored, specifically including: A hierarchical storage strategy of a distributed file system and a distributed memory system is used to hierarchically store the genome dynamic interaction network; The network data in the genomic dynamic interaction network is divided by a method based on a minimum point cut set, and the vertices with a connectivity higher than a set value in the hierarchically stored genomic dynamic interaction network are stored in the same storage node to realize the distributed storage of the genomic dynamic interaction network; The hierarchical storage strategy of the distributed file system and the distributed memory system is used to hierarchically store the genome dynamic interaction network, specifically including: The dynamic interaction network of the genome after a set time is represented as a vertex set and an edge set; Use hashing technology to distribute all vertices in the vertex set to storage nodes in the distributed memory system; Allocate all edges in the edge set to corresponding storage nodes in the distributed memory system according to the mapping relationship between points and edges; For the genomic dynamic interaction network before a set time, snapshots are generated regularly in chronological order, and the historical snapshot records are stored in a distributed file system.

Images