A gene quantization representation method and system based on variational graph autoencoder

Abstract

The present invention discloses a gene quantitative characterization method and system based on variational graph autoencoder, the method comprising the following steps: constructing a directed graph of a protein interaction network according to protein interaction network data; constructing three tree graphs including a biological process tree graph, a cell component tree graph and a molecular function tree graph according to gene ontology data; respectively calculating the non-redundant shortest paths between the root nodes and the nodes of the target gene in the three tree graphs as the primary quantitative characterization of the target gene; fusing the primary quantitative characterizations of the target gene in the three tree graphs and then combining them with the directed graph of the protein interaction network to obtain a high-level directed graph of the protein interaction network; inputting the directed graph of the high-level protein interaction network into a graph neural network model based on variational graph autoencoder to realize high-level quantitative characterization of genes. The present invention realizes efficient quantitative characterization of single-cell genes based on PPI, GO and variational graph autoencoder, providing depth and precision for biomedical research.

Description

Technical Field

[0001] The present invention belongs to the technical field of bioinformatics, and in particular relates to a gene quantization representation method and system based on a variational graph autoencoder. Background Art

[0002] In recent years, single-cell sequencing technology has been widely used in cancer analysis. Single-cell sequencing technology can study gene expression at single-cell resolution, reveal cell heterogeneity and cell composition, and then explore the tumor immune microenvironment. Single-cell gene characterization and cell clustering are important tasks in this field, and different cell types and states are identified through cell clustering. Single-cell data contains expression information of thousands of genes, and the data dimension is extremely high. At the same time, single-cell data is very sparse, and the expression levels of most genes in most cells are close to zero. In addition, single-cell data contains a lot of technical noise, which needs to be effectively reduced.

[0003] Traditional clustering methods have limited effects when processing high-dimensional, sparse, and noisy data, while variational graph auto-encoders (VGAE) have significant advantages in processing complex networks. The Chinese patent application with publication number CN115798593A provides a single-cell identification method and device based on self-supervised clustering of graph neural networks. By targeting single-cell transcriptome data across batches, a gene-cell interaction relationship network is constructed through an algorithm model, and low-dimensional single-cell sequencing data of the cell interaction graph network is obtained. The low-dimensional single-cell sequencing data is input into the variational autoencoder for processing, and the optimal cell Gaussian latent variable representation is optimized, and self-supervised single-cell clustering is performed to complete single-cell type identification. However, this method has certain limitations for processing high-dimensional data, and it is difficult to efficiently and comprehensively capture the intrinsic connections and characteristics between genes. The Chinese patent application with publication number CN114783526A provides a deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder, which analyzes cell data by extracting latent features. The algorithm framework combines the variational graph autoencoder and the Gaussian mixture model, and extracts low-dimensional latent features of the data through deep learning methods to cluster the cell data and mine biological information, solving the problem of high-dimensional sparsity of cell data. However, it focuses more on capturing low-dimensional latent features of data through deep learning technology to achieve clustering, and relatively ignores the comprehensiveness of enhancing the representation directly from the data source level. It may not be able to fully retain or mine the potential global and local information in the cell data that is critical to biological processes, thereby affecting the depth and accuracy of subsequent bioinformatics analysis.

[0004] In summary, existing methods still have problems such as insufficient cell clustering, loss of key gene data, and insufficient gene additional information. Therefore, further research is urgently needed on gene quantitative characterization methods for multi-attribute single cells to promote the development of related application fields such as tumor biology. Summary of the invention

[0005] In view of the above, the purpose of the present invention is to provide a gene quantitative characterization method and system based on variational graph autoencoder, by combining protein interaction network and gene ontology information (including biological processes, cellular components, molecular functions) and other related data information in the form of graph data to obtain an improved protein interaction network, and on this basis, using variational graph autoencoder to extract global and local information of genes, thereby achieving efficient quantitative characterization of genes.

[0006] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention is as follows:

[0007] In a first aspect, an embodiment of the present invention provides a gene quantization characterization method based on a variational graph autoencoder, comprising the following steps:

[0008] According to the obtained protein interaction network data, the protein names are used as nodes, and the protein interaction scores are used as the weights of the edges between nodes to construct a directed graph of the protein interaction network;

[0009] Based on the acquired gene ontology data, the gene ontology terms are used as nodes, the hierarchical relationships between nodes are converted into the weights of the edges between nodes, the node numerical attributes are calculated according to the numerical sorting of the nodes, and the biological process tree diagram, cell component tree diagram and molecular function tree diagram are constructed;

[0010] The non-redundant shortest paths between the root nodes and the nodes of the target genes in the biological process tree diagram, the cell component tree diagram, and the molecular function tree diagram are calculated as the primary quantitative representation of the target genes;

[0011] The primary quantitative representations of the target gene in the biological process tree, cell component tree and molecular function tree are fused and further combined with the protein interaction network directed graph to obtain the advanced protein interaction network directed graph;

[0012] The directed graph of the high-level protein interaction network is input into the graph neural network model based on variational graph autoencoder to complete the high-level quantitative representation of genes.

[0013] Preferably, the method of constructing a directed graph of a protein interaction network based on the acquired protein interaction network data, taking the protein names as nodes and the protein interaction scores as the weights of the edges between the nodes, comprises:

[0014] The protein interaction network data are obtained, the protein name data in the protein interaction network data are used as the starting node and the ending node respectively, and the protein interaction score data in the protein interaction network data are used as the weight of the edge between the nodes, so as to construct a protein interaction network directed graph including the node names and the edge weights.

[0015] Preferably, the method comprises: using the gene ontology data obtained, taking the gene ontology terms as nodes, converting the hierarchical relationship between the nodes into the weights of the edges between the nodes, calculating the node numerical attributes according to the numerical sorting of the nodes, and constructing the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram, including:

[0016] Obtaining the basic ontology file of gene ontology, taking the gene ontology terms in the basic ontology file as nodes, converting the hierarchical relationship between the nodes in the basic ontology file into the weights of the edges between the nodes, and constructing a hybrid gene ontology node graph;

[0017] Using the subgraph() function of the networkx tool and the root node names of known biological processes, cellular components and molecular functions in the basic ontology file, the biological process tree, cellular component tree and molecular function tree were separated and extracted from the hybrid gene ontology node graph;

[0018] The actual numbers containing zero prefixes were extracted from the names of the nodes in the basic ontology file, and the actual numbers were sorted according to their numerical values ​​and converted into new numerical values. The new numerical values ​​were used as the numerical attributes of each node, and the numerical attributes of the nodes were added to the biological process tree diagram, cellular component tree diagram, and molecular function tree diagram to construct the biological process tree diagram, cellular component tree diagram, and molecular function tree diagram including nodes, node numerical attributes, and edge weights.

[0019] Preferably, the non-redundant shortest paths between the root nodes and the nodes of the target genes in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram are respectively calculated as the primary quantitative representation of the target genes, including:

[0020] Obtain an annotation file including the target gene name and its corresponding node name in the biological process tree diagram, cell component tree diagram and molecular function tree diagram, query the new value corresponding to each node name in the annotation file, search for the shortest path from each node to the root node in the biological process tree diagram, cell component tree diagram and molecular function tree diagram, record the new value of each node in the shortest path in sequence, and calculate the number of nodes included in the shortest path, finally delete the redundant paths included in the calculated shortest path, and obtain the non-redundant shortest path as the primary quantitative representation of the target gene.

[0021] Preferably, when the non-redundant shortest path is used as the primary quantitative representation of the target gene, the following optimization strategies are also included:

[0022] Based on the correlation coefficient analysis, a longest path among all non-redundant shortest paths is selected for each target gene as the primary quantitative representation of the target gene;

[0023] Or, based on multi-path analysis, multiple non-redundant shortest paths with a set length are selected for each target gene as the primary quantitative representation of the target gene;

[0024] Alternatively, based on cross-correlation analysis, multiple non-redundant shortest paths with a set length are selected for each target gene as the primary quantitative representation of the target gene.

[0025] Preferably, the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram are fused and further combined with the protein interaction network directed graph to obtain a high-level protein interaction network directed graph, including:

[0026] The primary quantitative representations of the target genes in the biological process tree diagram, the cellular component tree diagram, and the molecular function tree diagram were fused by splicing;

[0027] The primary quantitative representation of the fused target gene is combined with the directed graph of the protein interaction network, and the primary quantitative representation of the target gene is used as the node attribute to obtain a high-level directed graph of the protein interaction network including node name, edge weight and node attribute.

[0028] Preferably, the step of inputting the directed graph of the advanced protein interaction network into a graph neural network model based on a variational graph autoencoder to complete the advanced quantitative representation of genes includes:

[0029] The directed graph of the high-level protein interaction network is input into the graph neural network model based on variational graph autoencoder for unsupervised node embedding. The graph neural network model based on variational graph autoencoder includes an encoder and a decoder structure. The encoder extracts the intermediate features and then the decoder calculates the probability of the existence of edges between two nodes to obtain the reconstructed directed graph of the high-level protein interaction network, completing the high-level quantitative representation of genes, so that the high-level quantitative representation of genes includes protein interaction network information, gene ontology information and relationship information with other genes.

[0030] In a second aspect, to achieve the above-mentioned invention object, an embodiment of the present invention further provides a gene quantitative representation system based on a variational graph autoencoder, comprising: a protein interaction network directed graph construction module, a gene ontology tree graph construction module, a gene primary quantitative representation calculation module, an advanced protein interaction network directed graph construction module and a gene advanced quantitative representation extraction module;

[0031] The protein interaction network directed graph construction module is used to construct a protein interaction network directed graph according to the acquired protein interaction network data, using protein names as nodes and protein interaction scores as weights of edges between nodes;

[0032] The gene ontology tree diagram construction module is used to construct a biological process tree diagram, a cell component tree diagram and a molecular function tree diagram according to the acquired gene ontology data, taking gene ontology terms as nodes, converting the hierarchical structure relationship between nodes into the weight of the edge between nodes, calculating the node numerical attributes according to the numerical sorting of the nodes;

[0033] The gene primary quantitative representation calculation module is used to calculate the non-redundant shortest path between the root node and the node of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram as the primary quantitative representation of the target gene;

[0034] The advanced protein interaction network directed graph construction module is used to fuse the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram, and further combine them with the protein interaction network directed graph to obtain an advanced protein interaction network directed graph;

[0035] The gene high-level quantitative representation extraction module is used to input the high-level protein interaction network directed graph into the variational graph autoencoder for unsupervised node embedding and finally obtain the gene high-level quantitative representation.

[0036] In the third aspect, in order to achieve the above-mentioned purpose of the invention, an embodiment of the present invention also provides a gene quantization characterization device based on a variational graph autoencoder, comprising a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-mentioned gene quantization characterization method based on a variational graph autoencoder when executing the computer program.

[0037] In a fourth aspect, in order to achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a computer, the above-mentioned gene quantization characterization method based on variational graph autoencoder is implemented.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention combines the protein interaction network and gene ontology information (including biological processes, cell components, molecular functions) and other related data information in the form of graph data to obtain a directed graph of an advanced protein interaction network, and on this basis, a variational graph autoencoder is used to extract the global and local information of genes, thereby realizing efficient quantitative representation of genes. Among them, through the fusion of multi-level gene ontology information, the gene quantitative representation method is preliminarily defined to provide a standardized description for gene function; and the integration of the protein interaction network reveals the interaction between molecules in the cell, so that genes have mutual connections with other genes, further improving the gene feature representation ability; the variational graph autoencoder maps the fused advanced protein interaction network directed graph data to a continuous latent space through the encoder to capture the intrinsic characteristics between genes, and the decoder reconstructs the nodes and edges of the graph from the latent representation to realize unsupervised embedding of nodes, making the quantitative representation of genes more accurate and effective, and inferring cell states and trajectories by learning the complex relationships between cells, which is conducive to more finely distinguishing gene features of different types and functions.

[0040] In summary, single-cell gene quantitative representation based on protein interaction networks, gene ontology and variational graph autoencoders provides depth and precision for biomedical research. This integrated approach can deeply understand cellular heterogeneity, dynamically track biological processes, accurately map molecular functions and cellular components, help discover disease mechanisms and develop new drugs, promote the development of personalized medicine, and is of great value to the discovery of biomarkers and interdisciplinary research. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] 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 or the description of the prior art 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 creative work.

[0042] Figure 1 It is an overall flow chart of a gene quantization characterization method based on a variational graph autoencoder provided in an embodiment of the present invention;

[0043] Figure 2 is a protein interaction integration value distribution histogram drawn based on PPI data provided by an embodiment of the present invention;

[0044] Figure 3 is a schematic diagram of constructing a PPI directed graph provided by an embodiment of the present invention;

[0045] Figure 4 is a schematic diagram of a go-basic.obo file provided in an embodiment of the present invention;

[0046] Figure 5 It is a schematic diagram of directed relationship weights of gene ontology data nodes provided by an embodiment of the present invention;

[0047] Figure 6 is a schematic diagram of a hybrid GO node graph provided by an embodiment of the present invention;

[0048] Figure 7 It is a schematic diagram of a separated BP tree diagram, a CC tree diagram and a MF tree diagram without node attributes provided by an embodiment of the present invention;

[0049] Figure 8 It is a schematic diagram of a BP tree diagram, a CC tree diagram and an MF tree diagram with separated node attributes provided by an embodiment of the present invention;

[0050] Fig. 9 is a schematic diagram of calculating the Pearson correlation coefficient matrix provided by an embodiment of the present invention;

[0051] Fig.10 It is a box plot of the number of shortest paths (a) and the length of all shortest paths (b) of all genes provided by the embodiment of the present invention;

[0052] Fig.11 It is a flowchart of selecting h shortest paths of length j as a guarantee of primary gene quantization based on cross-correlation analysis provided by an embodiment of the present invention;

[0053] Fig.12 It is a schematic diagram of a high-level PPI directed graph after primary quantitative characterization of a fusion gene provided by an embodiment of the present invention;

[0054] Fig.13 It is a schematic diagram of a VGAE-based graph neural network model provided by an embodiment of the present invention;

[0055] Fig.14 This is a schematic diagram of tSNE dimensionality reduction comparative analysis based on single-cell gene quantitative characterization provided by an embodiment of the present invention;

[0056] Fig.15 The VGAE provided by the embodiment of the present invention embeds the ROC curve in the node of the advanced PPI directed graph;

[0057] Fig.16 It is a structural schematic diagram of a gene quantization characterization system based on a variational graph autoencoder provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0059] The inventive concept of the present invention is: in view of the problems of insufficient cell clustering, loss of key gene data and insufficient gene additional information in the prior art, the embodiment of the present invention provides a gene quantitative characterization method and system based on variational graph autoencoder, which combines the protein interaction network with the multi-dimensional information of biological processes, cell components and molecular functions of gene ontology, and constructs a comprehensive graph data structure. On this basis, the variational graph autoencoder is used to extract and encode the global and local information of genes. The primary gene quantitative characterization is defined by the fusion of multi-level gene ontology information, which provides a standardized description of gene function. The introduction of protein interaction network reveals the interaction between molecules in cells, further enriches the mutual connection between genes, and improves the ability of gene feature characterization. The encoder part of the variational graph autoencoder maps the fused high-level protein interaction network directed graph data into a continuous latent space, infers the cell state and trajectory by learning the complex relationship between cells, captures the intrinsic connection and characteristics between genes, and finally realizes efficient gene quantitative characterization.

[0060] Figure 1 is an overall flow chart of the gene quantization characterization method based on variational graph autoencoder provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides a gene quantization characterization method based on a variational graph autoencoder, comprising the following steps:

[0061] 1. Construct a PPI directed graph.

[0062] Protein-protein interaction networks (PPI) data are composed of proteins that interact with each other to participate in various aspects of life processes such as biological signal transmission, gene expression regulation, energy and material metabolism, and cell cycle regulation. PPI data can be obtained through a variety of public databases and online tools, such as STRING, BioGRID, IntAct, STRING-DB, and MENTHA. It should be noted that the quality and reliability of PPI data may vary depending on different databases and tools. Therefore, it is necessary to make targeted choices based on specific research questions and experimental conditions, and carefully check and verify the source and quality of the data. Table 1 gives the PPI data obtained from STRING (the protein1 and protein2 columns represent protein names, the neighborhood column represents genes that appear repeatedly in neighboring regions of the genome, the fusion column represents individual gene fusion events in each species, the cooccurence column represents whether cross-species connecting proteins exist, the coexpression column represents genes co-expressed in the same or other species, the experimental, database and textmining columns represent data from biological experiments, databases and text mining, respectively, and the combined_score column represents the protein interaction score obtained by integrating different sources of evidence).

[0063] Table 1 Protein interaction integration scores

[0064]

[0065] Draw a histogram for each column of data shown in Table 1, and the results are as follows Figure 2 As shown, it can be seen that only the combined_score column has no missing values ​​(0 indicates missing values), and the values ​​are relatively large compared to other columns (greater than 150, the larger the value, the better), and the textminning column data quality is second.

[0066] The protein names in the protein1 and protein2 columns in Table 1 are used as the start and end nodes respectively, and the data in the combined_score column (selection criteria: according to Figure 2 , less data missing, larger value) as the weight of the corresponding edge, construct the PPI network directed graph G PPI (like Figure 3 as shown).

[0067] Figure 3The PPI network directed graph only has node names (protein names, that is, gene names. According to the central dogma, there is a one-to-one correspondence between genes and proteins) and edge weights (pathway relationships), but no other biological features, that is, it lacks node attributes. Figure 3 When applied to node embedding, the resulting embedding vector can only reflect the local connection pattern of the node, but cannot capture the global biological properties of the node. This processing approach is equivalent to treating all nodes as equivalent and ignoring the possible biological differences between them, which is unreasonable in the field of bioinformatics. Therefore, this method may not provide sufficient information for downstream bioinformatics analysis tasks.

[0068] 2. Construct BP, CC and MF tree graphs based on GO node information.

[0069] To solve the above problems, Gene Ontology (GO) data related to biological process (BP), cellular component (CC), molecular function (MF), etc. are introduced to construct attribute information for nodes (protein / gene), so that these biological information can be incorporated into the node embedding when using the graph network, so that the final node embedding vector contains enough biological features.

[0070] 2.1. Obtain the basic ontology file.

[0071] go-basic.obo: Get the basic ontology file go-basic.obo from the official website of Gene Ontology, such as Figure 4 As shown, the latest gene ontology information can be obtained from it, including the functional annotations and associations of each gene.

[0072] An example of defining a root node based on the go-basic.obo file is as follows:

[0073] BP tree root node: BP_old_root_node ='GO:0008150', BP_old_root_node is the name of the defined BP tree root node, GO:0008150 is an example of a gene ontology term in the go-basic.obo file (a node without a parent node is a root node);

[0074] CC tree root node: CC_old_root_node ='GO:0005575', CC_old_root_node is the name of the defined CC tree root node, GO:0005575 is an example of a gene ontology term in the go-basic.obo file (a node without a parent node is a root node);

[0075] MF tree root node: MF_old_root_node ='GO:0003674', MF_old_root_node is the name of the defined MF tree root node, GO:0003674 is an example of a gene ontology term in the go-basic.obo file (a node without a parent node is a root node).

[0076] The go-basic.obo file contains information about three tree diagrams: BP tree diagram, CC tree diagram, and MF tree diagram, but all the information is mixed together, and the three trees cannot be separated from the text information alone. Fortunately, these three tree diagrams are independent, which means that there is no connection between the trees.

[0077] In view of this, we first extract the key information in the go-basic.obo file to construct all node pairs, and then build a hybrid GO node graph of three trees. G BP_CC_MF , and then use the subgraph() function of the networkx tool and the root node names (known) in the three tree graphs to get G BP_CC_MF Extract these three tree diagrams: BP tree diagram G BP 、CC Tree Map G CC and MF tree graph G MF The specific implementation process is shown in the following steps.

[0078] 2.2. Constructing the hybrid GO node graph G BP_CC_MF .

[0079] According to the go-basic.obo file features ( Figure 4 ), a target GO node starts with the gene ontology term '[Term]', and the second line is the id of the target GO node, which is represented by 'id: GO:*******' (* represents a positive integer between 0 and 9). Using these two features as markers, the id of the target GO node is accurately extracted: 'GO:*******'.

[0080] Key information under a '[Term]' also includes lines starting with the keyword 'is_a: GO:^^^^^^^' (^ represents a positive integer between 0 and 9) or 'relationship: part_of GO:#######' (# represents a positive integer between 0 and 9). The hierarchical relationship between nodes is represented by 'is_a' or 'part_of'. 'is_a' means that the target node 'GO:*******' is a type of node 'GO:^^^^^^^', for example, mitochondria is a type of organelle, and the relationship between the two is mitochondria is_a organelle. 'part_of' means that the target node 'GO:*******' is a component of the node 'GO:#######', for example, mitochondria is a component of the cytoplasm, and the relationship between the two is mitochondria part_of cytoplasm.

[0081] Therefore, according to the information under a '[Term]', we can extract: 1 target node; n (n=0,1,2...) 'is_a' nodes; m (m=0,1,2...) 'part_of' nodes. With 'is_a' and 'part_of' nodes as the starting nodes, the target node as the ending node, and the edge weight from 'is_a' to the target node as 0.4, and the edge weight from 'part_of' to the target node as 0.6, we finally form the node pairs shown in Table 2. Based on Table 2, we construct Figure 5 The directed graph data shown.

[0082] Table 2 Directed relationships and weights of gene ontology data nodes

[0083]

[0084] Repeat the above process for each '[Term]' flag in the go-basic.obo file, and finally get the node pair data between all target nodes and their respective 'is_a' and 'part_of' nodes. Combine all the node pair data to get a complete node pair data containing all nodes (at this time, it is no longer distinguished whether the node is a target node or an 'is_a' or 'part_of' node). Use this complete node pair data to construct the graph data to get the hybrid GO node graph. G BP_CC_MF ,like Figure 6 shown.

[0085] 2.3. Separately extract BP, CC and MF tree graphs.

[0086] Then use the subgraph() function of the networkx tool and the root node names of the three trees (known) from G BP_CC_MFSeparate and extract these three tree graphs:

[0087] BP Tree G BP =subgraph(BP_old_root_node);

[0088] CC Treemap G CC =subgraph(CC_old_root_node);

[0089] MF Treemap G MF =subgraph(MF_old_root_node);

[0090] The results are as follows Figure 7 shown.

[0091] 2.4. Numerize the node attributes of BP, CC and MF tree graphs.

[0092] Now we have obtained BP tree diagram, CC tree diagram and MF tree diagram Figure 3 There are three independent trees, but the nodes of each tree only have names, and they are still in the form of 'GO:xxxxxxx' (x represents a positive integer between 0 and 9). There are no node attributes and it cannot be directly applied to graph network analysis. Therefore, it is necessary to assign a numerical attribute feature to each node of each tree (the following takes the BP tree graph as an example to illustrate how to quantify the attributes of nodes in a tree graph. The processing methods of CC tree graph and MF tree graph are exactly the same).

[0093] Considering the naming method of GO nodes, 'xxxxxxx' is actually a number with a maximum of 7 digits. If it is less than 7 digits, the leading 0 is added, for example, 'xxxxxxx'='0000022'. Extract the 'xxxxxxxxx' field of all node names in the BP tree graph respectively, and extract the actual number num, discard the leading 0, for example, here num=22, and then sort all the nodes on the tree from small to large according to the num value; finally, use new_num=1, 2, 3, 4... as the numerical attribute of each node, as shown in Table 3, the new_num of node GO:000022 is 9, so that a BP tree graph with both edge weights and node numerical attributes is obtained, as shown in Table 3. Figure 8 As shown (note that the numerical attribute of the root node is not necessarily 1, nor does it mean that the smaller the value, the closer it is to the root node. This will not have a negative impact on subsequent analysis, because the focus is on the pathway relationship between nodes and the relative relationship between pathways. The absolute attributes of each node are not important).

[0094] Table 3 Comparison table of numerical attributes of gene ontology data nodes

[0095]

[0096] 3. Based on GO node tree diagram ( G BP , G CC and G MF ) Calculate the shortest path of genes.

[0097] The previous steps constructed the GO node tree diagrams of BP, CC and MF with node numerical attributes. Next, we need to construct the primary quantitative expression of genes based on the node positions of specific genes in the three trees (taking the BP tree diagram as an example, the calculation steps for the other two cases are exactly the same).

[0098] 3.1. Obtain BP, CC and MF data of genes.

[0099] To this end, we first obtained the following annotation files from the GO official website, as shown in Table 4. In Table 4, the 'SYMBOL' column represents the 'gene name', and the 'GO' column represents the corresponding gene in G BP The GO name of the last node on the tree, the 'ENTREZID' column indicates the unique assigned number of the gene, the 'EVIDENCE' column indicates the evidence code of the GO annotation, and the 'ONTOLOGY' column indicates the ontology category to which the GO annotation belongs. It should be pointed out that the so-called 'last node' here does not mean G BP The last leaf node on the tree is the gene that is known to be G BP The last node of a pathway in the tree. This node may be any node in the tree, including the root node. In addition, a 'gene' may correspond to multiple GO nodes, because a gene may have multiple expression pathways.

[0100] Table 4 GO node data of genes in BP tree

[0101]

[0102] 3.2. Calculate the shortest path of the GO node of the gene on the BP, CC and MF tree graphs.

[0103] First, query each node (i.e., each gene) in the GO column in Table 4 in Table 3 to obtain the new_num value of these nodes, and then use the Dijkstra algorithm to search each node to G BP Shortest path to root node in tree graph L shortest(The new_num value of each node in the short path is recorded in sequence) and the number of nodes Len( L shortest ), where Len() represents the function of calculating the length. Taking the 'NDUFAF7' gene in Table 4 as an example, this gene has two GO nodes (GO:0019918 and GO:0032981). Assuming that the corresponding new_num are 421 and 463 respectively (obtained from Table 3), the corresponding shortest paths obtained by the Dijkstra algorithm are [518, 107, 856, 421] and [518, 107, 463] respectively (see Figure 8 BP tree diagram), the shortest path lengths are 4 and 3 respectively, as shown in Table 5.

[0104] Table 5 Examples of the shortest path and length from the GO node to the root node of the 'NDUFAF7' gene in the BP tree

[0105]

[0106] 3.3. Identification and removal of redundant pathways for individual genes.

[0107] As mentioned above, a gene may have multiple GO nodes, that is, multiple shortest paths, and these paths may contain each other. Figure 8 One path L A =[518, 107, 856, 421], another path L B =[518,107, 856], obviously L B Completely L A Contains, called L B It is a redundant path and needs to be deleted. is_subsequence (seq1,seq2) The function determines whether two paths contain each other. seq1 and seq2 Representing two paths respectively, through two indexes seq1_idx and seq2_idx Come traverse seq1 and seq2 ,if seq1 The current element in seq2 If the current elements in are equal, then seq1_idx Increase, seq2_idx Increment after each comparison to ensure that seq2 If seq1_idx At the end of the traversal it is equal to seq1 The length of seq1 All elements in the seq2 , and no other values ​​are inserted in between. For all the shortest paths of a gene, is_ subsequence(seq1, seq2) By analyzing the function, all possible redundant paths of the gene can be deleted.

[0108] 4. Based on GO node tree diagram ( G BP , G CC and G MF )'s primary quantitative characterization of genes.

[0109] All genes were calculated above. G BP , G CC and G MF The non-redundant shortest paths on the tree diagram, but each gene may have multiple non-redundant shortest paths, the lengths and relationships between these paths are different, and the number of non-redundant paths between different genes is also different, which makes it difficult to quantitatively represent genes. To address this problem, the following three strategies are used (in G BP For example ，G CC and G MF The process is exactly the same).

[0110] 4.1. Based on the correlation coefficient analysis, one longest path among all the shortest paths was selected for each gene as the primary quantitative representation of the gene (Strategy 1).

[0111] The longer the shortest path of a gene's GO node on the three trees, the more thorough the understanding of the gene. Therefore, choosing the longest path among the shortest paths of all GO nodes of the gene as the quantitative representation of the gene can better contain the potential characteristics of the gene. For a gene, using this strategy, we need to discuss the following three situations.

[0112] A gene has only one GO node (i.e., only one shortest path):

[0113] At this point, the shortest path is the primary quantitative representation of the gene. G_P_F BP express( G_P_F:Gene Primary Features ).

[0114] A gene has multiple GO nodes (i.e., multiple shortest paths), but only one longest shortest path:

[0115] At this point, the unique longest shortest path is the primary quantitative representation of the gene, using G_P_F BP express.

[0116] A gene has multiple GO nodes (i.e., multiple shortest paths), but multiple longest shortest paths (they have the same length):

[0117] In this case, suppose a gene has S The longest shortest paths of the same length are calculated for each of these paths. r ij ( i,j=1,2,3...S ),in i and j Respectively S The path index in the path is used to obtain the Pearson correlation coefficient matrix R S×S (like Fig. 9 As shown, S =4).

[0118] R S×S Sum each column to get the sum vector of the Pearson correlation coefficient between a path and all other paths R 1×S , as shown in formula (1), where sum ( ) represents the sum function, axis =0 means sum each column.

[0119] R 1×S =sum (R S×S , axis= 0) (1)

[0120] correspond Figure 8 : R 1×4 = [2.79119568, 3.22595884, 3.17499814, 1.70141933].

[0121] Then use formula (2) to find R 1×S The position number of the maximum value k ,For example k = index _max( R 1×4 ) = 2,

[0122] k = index_max( R 1×S ) (2)

[0123] Finally, the aforementioned S The first k The pathway is used as the primary quantitative representation of the gene, that is, in this example, L 2 as the final primary quantitative representation:

[0124] G_P_F BP = [518, 107, 856, 208].

[0125] 4.2. Based on multi-path analysis, h shortest paths of length j are selected for each gene as the primary quantitative representation of the gene (Strategy 2).

[0126] The previous characterization of a gene using a single longest path has achieved relatively satisfactory results. However, this method cannot cover the path information between other GO nodes. If more path information can be integrated into the gene characterization process, it will undoubtedly significantly improve the accuracy and comprehensiveness of the gene characterization. As mentioned earlier, there are differences in the number of shortest paths between different genes. These differences reflect their unique biological pathway characteristics. In order to simplify the subsequent processing flow and ensure the consistency of the characterization, h shortest paths with a length of j are used to comprehensively characterize the genes. This method can more comprehensively capture the multidimensional characteristics of genes in biological pathways, thereby providing more accurate and rich information for subsequent bioinformatics analysis. Obviously, both h and j need to be determined according to actual conditions. The steps are as follows:

[0127] (1) Perform statistical analysis on the number of shortest paths of all genes, draw a box plot, and extract the values ​​at three quarters as the h value, such as Fig.10 As shown in (a), h = 5, that is, each gene is represented by 5 shortest paths. If the number of shortest paths of a gene itself is less than h, the missing paths are supplemented with a sequence of 0; if it is greater than h, h paths are extracted from the largest to the smallest in terms of path length.

[0128] (2) Perform statistical analysis on the shortest path lengths of all genes, draw a box plot, and extract the values ​​at three quarters as the j value, such as Fig.10As shown in (b), j=8, that is, the length of each path in the h paths selected for a gene is set to 8. If the length of a path in the h paths is less than j, the end of the path is padded with 0 to a length of j; if it is greater than j, a sub-path of length j is intercepted from the "end to the front" of the path as a replacement (the reason for intercepting from the back to the front is that all paths start with the root node, so the closer to the front, the more similar the different paths are, and the more different they are, the purpose is to show the differences).

[0129] h and j were determined earlier (taking h=5, j=8 as an example), and then the path extraction for each gene needs to be performed based on these two parameters. As shown in Table 6, all the shortest paths of a gene (arranged from large to small by length) have a total number of paths of 7>h=5, and the longest path length is 11>j=8. According to the above regulations, the first five paths are selected, but the length of the first four of them exceeds 8, so 8 nodes are intercepted from the back to the front as replacement paths for the original path; the length of path 5 is 4, which is less than 8, so 0 is added after it to a length of 8 (the shaded part in the table). Finally, the final multi-path primary characterization results of the gene shown in Table 7 are obtained.

[0130] Table 6 All shortest paths of a gene

[0131]

[0132] Table 7 Final multi-path primary representation of a gene extracted based on h=5 and j=8

[0133]

[0134] Finally, the two-dimensional matrix composed of h paths shown in Table 7 is used as the primary quantitative representation matrix of the gene:

[0135] G_P_F BP = [[421, 65, 843, 98, 561, 4216, 23, 1],

[0136] [208, 615, 34, 4605, 3425, 450, 15, 32],

[0137] [2016, 609, 418, 40, 545, 498, 361, 8],

[0138] [463, 26, 84, 961, 2458, 3054, 31, 91],

[0139] [518, 35, 151, 78, 0, 0, 0, 0]].

[0140] 4.3. Based on cross-correlation analysis, h shortest paths of length j are selected as the primary quantitative representation of the gene (Strategy 3).

[0141] When the total number of paths M of a gene is much larger than h (for example, L>2h), Strategy 2 will not be able to select the most representative h paths well, affecting the final gene characterization results. In this case, when selecting the target path, follow the principle of "the greater the difference in the path, the more important it is". There is reason to believe that "similar paths represent similar functional characteristics". To this end, first fill all the paths of the gene with a length of L (the longest path length of the gene) by adding 0 to the tail, and obtain an M×L two-dimensional path matrix H M×L , where the vertical direction is the path data M and the horizontal direction is the path length L. Then, we perform pairwise cross-correlation calculations on the paths to obtain the path cross-correlation matrix K M×M , further sum the matrix column by column to obtain a vector F of length M 1×M Note that the value in this vector represents the correlation between the gene path corresponding to the value and other paths. The larger the value, the stronger the correlation between the path and its path, and the smaller the value, the smaller the correlation. Based on this information, from F 1×M Select the path corresponding to the minimum value and determine the number of 0s added to the end of the path. The final selection criteria are: 1×M The smaller the value in , the better the number of zeros added to the path. According to this principle, h paths are determined, and the final length j of these h paths is still obtained according to Strategy 2. According to the above process, the primary quantitative representation matrix of the gene is finally obtained. G_P_F BP , the process is as follows Fig.11 shown.

[0142] 5. Based on GO node tree diagram ( G BP , G CC and G MF ) of the primary quantitative characterization of genes fused.

[0143] The BP tree diagram was used as an example to illustrate how to obtain the primary quantitative representation of a gene. G_P_F BP According to the same method, the primary quantitative representation of the gene on the CC tree diagram and MF tree diagram was obtained respectively. G_P_F CC and G_P_F MF . It is necessary to unify these three quantities into one quantity G_P_F , so as to facilitate subsequent calculations, the horizontal or vertical splicing method is adopted, as shown in equations (3) and (4), respectively, that is:

[0144] G_P_F = [ G_P_F BP , G_P_F CC , G_P_F MF ] (3)

[0145] G_P_F = [ G_P_F BP ; G_P_F CC ; G_P_F MF ] (4)

[0146] 6. Combine primary quantitative characterization of genes with PPI data.

[0147] Step 1 constructs the initial PPI directed graph, and steps 2-5 construct the primary quantitative representation of genes based on GO node paths. This step uses the primary quantitative representation of genes as the node features of the PPI directed graph to achieve information fusion of the two, thus obtaining high-level PPI graph data with GO node path information. Figure 3 Upgraded to Fig.12 Results shown.

[0148] 7. Advanced quantitative representation of genes based on variational graph autoencoders.

[0149] In the previous step, a multi-path representation of genes was first constructed from GO information, and it was fused with the PPI directed graph as a node (gene) feature to form a high-level PPI directed graph. However, the node features at this time do not have the PPI graph structure information, nor do they embed information about other surrounding nodes. For this purpose, a graph neural network model was constructed using the variational graph auto-encoders (VGAE) framework to perform unsupervised node embedding on high-level PPI graph data, so that the node representation has PPI, GO, and relationship information with other genes at the same time.

[0150] VGAE is a deep learning framework that combines a graph neural network (GNN) and a variational autoencoder (VAE). Its core idea is to model the generation process of graph data as a probabilistic process and optimize the model parameters through variational inference. VGAE consists of two core components, the encoder and the decoder. It can simultaneously learn the structure of the graph and the representation of node features, and can be used for unsupervised learning tasks. Fig.13As shown, in the VGAE-based graph neural network model provided in the embodiment, three layers of GCNConv( ) (i.e., GCN1, GCN2, and GCN3) and one layer of Dropout module are used to construct an encoder. GCNConv( ) is a function or module for performing convolution operations on graph data. It accepts the adjacency matrix and node features of the graph as input, constructs a new feature representation by convolving the features of each node with the features of its neighboring nodes, and outputs an updated node feature representation. Dropout( ) regularization is used in the model to reduce overfitting, and the decoder adopts a graph decoder, that is, the probability of an edge between two nodes is calculated to reconstruct the graph, where φ ( ) represents the sigmoid function, A represents a high-level PPI directed graph, X Indicates converting the advanced PPI directed graph into matrix input. Z represents the extracted intermediate feature representation, Z T express Z The transpose of X ’ represents the reconstructed feature matrix, A ’ Represents the reconstructed high-level PPI directed graph.

[0151] 8. Example test analysis.

[0152] In order to verify the actual effect of the gene quantification characterization method based on variational graph autoencoder provided in the embodiment of the present invention, the "pbmc3k" classic single-cell sequencing data set derived from peripheral blood mononuclear cells (PBMC) was used for testing. The pbmc3k data contains 2700 cells and 13714 genes. The original data contains three files: barcodes.tsv, features.tsv, and matrix.mtx. After preprocessing, it is converted into a matrix with rows and columns representing genes and cells respectively.

[0153] In order to highlight the effectiveness of the gene quantization representation method based on variational graph autoencoder provided by the embodiment of the present invention, Node2Vec (Node2Vec is a graph embedding technology based on random walk, which is used to generate low-dimensional vector representations of nodes in the graph, capture local and global structural relationships between nodes, and is widely used in graph analysis tasks such as node classification and link prediction) is first used to perform unsupervised node embedding on the original PPI directed graph, that is, only the edge information in the PPI is used without the node attribute features (primary representation of gene paths based on GO information). The result after node embedding is reduced in dimension using the tSNE method to obtain Fig.14As shown in (a), the genes are clustered into many small groups, indicating that the PPI graph structure alone can be used to characterize gene characteristics to a certain extent. However, overall, the genes are distributed in a uniform circular area, indicating that there is not enough information to support further classification. Fig.14 (b) is the result of tSNE dimensionality reduction directly using the primary quantitative representation of genes obtained from GO data, excluding the original PPI information. Fig.14 The effect shown in (a) has been significantly improved, with fewer total clusters, clearer boundaries, and less randomness in the overall distribution. It can be seen that the primary gene representation constructed based on GO data is able to better characterize gene characteristics. Fig.14 (c) is the result of unsupervised node embedding and tSNE on the high-level PPI directed graph with primary gene representation as node feature using the VGAE-based graph neural network model, which is similar to the previous Fig.14 (a) and Fig.14 Comparing the two results in (b), Fig.14 In (c), the boundaries between gene clusters are reduced. This is because the graph data that integrates PPI and GO information contains more information at the same time, and the connection between genes is strengthened. In addition, the overall distribution randomness is further reduced, indicating that the combination of PPI, GO, and VGAE enhances information extraction as a whole. Fig.15 It is the ROC curve of 100 epochs of unsupervised node embedding on the advanced PPI directed graph using VGAE for pbmc3k data.

[0154] The experiments on pbmc3k show that the method proposed in the embodiment of the invention is effective to construct a primary gene representation based on gene links from the BP, CC and MF three tree diagrams of GO data, combine it with the PPI directed graph as a node feature to form an advanced PPI directed graph, and then use VGAE for unsupervised node embedding to construct an advanced gene representation.

[0155] In summary, the embodiment of the present invention provides a gene quantitative characterization method based on variational graph autoencoder, which utilizes the advantages of variational graph autoencoder (VGAE) in processing complex networks, and combines protein interaction network (PPI) and gene ontology (GO) information (including biological process BP, cell component CC and molecular function MF), providing a new idea for single-cell gene characterization. In the PPI network, nodes represent proteins (corresponding to related genes), and edges represent the connections between proteins. PPI reveals the interactions between proteins through experimental data and computational predictions. The PPI network helps to reveal the complex molecular interactions inside cells, thereby gaining a deeper understanding of cell functions and states. GO provides a framework that enables researchers to describe the functions of genes in a standardized way, thereby facilitating comparisons and data integration between different studies. Through GO classification, researchers can better understand the role of genes in organisms and how they participate in complex biological processes. VGAE is a deep learning framework that combines graph neural networks and variational autoencoders. It learns low-dimensional potential representations of graph data through encoders, reconstructs graph structures and node features with decoders, and optimizes the representation of latent space using variational inference. A primary gene representation is constructed based on the pathway from the GO node tree and used as the node feature of the PPI graph data. The VGAE-based graph neural network model is then used to learn the improved PPI graph data to capture the local and global information between genes, thereby achieving efficient gene quantitative representation. This result will effectively improve the accuracy and efficiency of single-cell gene representation and cell clustering, which is of great significance in the study of tumor microenvironment and further promote the development of tumor biology and related fields.

[0156] Based on the same inventive concept, Fig.16 As shown, an embodiment of the present invention also provides a gene quantitative representation system 1600 based on a variational graph autoencoder, including: a protein interaction network directed graph construction module 1601, a gene ontology tree graph construction module 1602, a gene primary quantitative representation calculation module 1603, an advanced protein interaction network directed graph construction module 1604 and a gene advanced quantitative representation extraction module 1605.

[0157] The protein interaction network directed graph construction module 1601 is used to construct a protein interaction network directed graph according to the obtained protein interaction network data, using protein names as nodes and protein interaction scores as weights of edges between nodes;

[0158] The gene ontology tree diagram construction module 1602 is used to construct a biological process tree diagram, a cell component tree diagram and a molecular function tree diagram according to the acquired gene ontology data, using gene ontology terms as nodes, converting the hierarchical structure relationship between the nodes into the weight of the edges between the nodes, calculating the node numerical attributes according to the numerical sorting of the nodes, and constructing the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram;

[0159] The gene primary quantitative representation calculation module 1603 is used to calculate the non-redundant shortest path between the root node and the node of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram as the primary quantitative representation of the target gene;

[0160] The high-level protein interaction network directed graph construction module 1604 is used to fuse the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram, and further combine them with the protein interaction network directed graph to obtain a high-level protein interaction network directed graph;

[0161] The gene high-level quantitative representation extraction module 1605 is used to input the high-level protein interaction network directed graph into the graph neural network model based on the variational graph autoencoder to complete the gene high-level quantitative representation.

[0162] Based on the same inventive concept, an embodiment of the present invention also provides a gene quantization characterization device based on a variational graph autoencoder, comprising a memory and one or more processors, the memory being used to store a computer program, and the processor being used to implement the above-mentioned gene quantization characterization method based on a variational graph autoencoder when executing the computer program.

[0163] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a computer, the above-mentioned gene quantization characterization method based on variational graph autoencoder is implemented.

[0164] It should be noted that the gene quantization characterization system based on variational graph autoencoder, the gene quantization characterization device based on variational graph autoencoder and the computer-readable storage medium provided in the above embodiments all belong to the same inventive concept as the gene quantization characterization method based on variational graph autoencoder. The specific implementation process is detailed in the embodiment of the gene quantization characterization method based on variational graph autoencoder, which will not be repeated here.

[0165] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A gene quantization representation method based on variational graph autoencoder, characterized in that: The following steps are involved: According to the obtained protein interaction network data, the protein names are used as nodes, and the protein interaction scores are used as the weights of the edges between nodes to construct a directed graph of the protein interaction network; Based on the acquired gene ontology data, the gene ontology terms are used as nodes, the hierarchical relationships between nodes are converted into the weights of the edges between nodes, the node numerical attributes are calculated according to the numerical sorting of the nodes, and the biological process tree diagram, cell component tree diagram and molecular function tree diagram are constructed; The non-redundant shortest paths between the root nodes and the nodes of the target genes in the biological process tree diagram, the cell component tree diagram, and the molecular function tree diagram are calculated as the primary quantitative representation of the target genes; After fusing the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram, the primary quantitative representations are further combined with the directed graph of the protein interaction network to obtain a high-level directed graph of the protein interaction network, specifically including: fusing the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram by splicing; combining the primary quantitative representations of the fused target gene with the directed graph of the protein interaction network, using the primary quantitative representations of the target gene as node attributes, and obtaining a high-level directed graph of the protein interaction network including node names, edge weights and node attributes; The directed graph of the high-level protein interaction network is input into the graph neural network model based on the variational graph autoencoder to complete the high-level quantitative representation of the gene, specifically including: inputting the directed graph of the high-level protein interaction network into the graph neural network model based on the variational graph autoencoder for unsupervised node embedding, the graph neural network model based on the variational graph autoencoder includes an encoder and a decoder structure, the encoder extracts the intermediate features and then the decoder calculates the probability of the existence of edges between two nodes to obtain a reconstructed directed graph of the high-level protein interaction network, completing the high-level quantitative representation of the gene, so that the high-level quantitative representation of the gene includes protein interaction network information, gene ontology information and relationship information with other genes.

2. The gene quantization characterization method based on variational graph autoencoder according to claim 1, characterized in that: The method comprises: constructing a directed graph of a protein interaction network based on the obtained protein interaction network data, taking the protein names as nodes, taking the protein interaction scores as the weights of the edges between the nodes, and comprising: The protein interaction network data are obtained, the protein name data in the protein interaction network data are used as the starting node and the ending node respectively, and the protein interaction score data in the protein interaction network data are used as the weight of the edge between the nodes, so as to construct a protein interaction network directed graph including the node names and the edge weights.

3. The gene quantization characterization method based on variational graph autoencoder according to claim 1, characterized in that: According to the obtained gene ontology data, the gene ontology terms are used as nodes, the hierarchical relationship between the nodes is converted into the weight of the edge between the nodes, the node numerical attributes are calculated according to the numerical sorting of the nodes, and the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram are constructed, including: Obtaining the basic ontology file of gene ontology, taking the gene ontology terms in the basic ontology file as nodes, converting the hierarchical relationship between the nodes in the basic ontology file into the weights of the edges between the nodes, and constructing a hybrid gene ontology node graph; Using the subgraph() function of the networkx tool and the root node names of known biological processes, cellular components and molecular functions in the basic ontology file, the biological process tree, cellular component tree and molecular function tree were separated and extracted from the hybrid gene ontology node graph; The actual numbers containing zero prefixes were extracted from the names of the nodes in the basic ontology file, and the actual numbers were sorted according to their numerical values ​​and converted into new numerical values. The new numerical values ​​were used as the numerical attributes of each node, and the numerical attributes of the nodes were added to the biological process tree diagram, cellular component tree diagram, and molecular function tree diagram to construct the biological process tree diagram, cellular component tree diagram, and molecular function tree diagram including nodes, node numerical attributes, and edge weights.

4. The gene quantization characterization method based on variational graph autoencoder according to claim 3 is characterized in that: The non-redundant shortest paths between the root nodes and the nodes of the target genes in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram are calculated as the primary quantitative representation of the target genes, including: Obtain an annotation file including the target gene name and its corresponding node name in the biological process tree diagram, cell component tree diagram and molecular function tree diagram, query the new value corresponding to each node name in the annotation file, search for the shortest path from each node to the root node in the biological process tree diagram, cell component tree diagram and molecular function tree diagram, record the new value of each node in the shortest path in sequence, and calculate the number of nodes included in the shortest path, finally delete the redundant paths included in the calculated shortest path, and obtain the non-redundant shortest path as the primary quantitative representation of the target gene.

5. The gene quantization characterization method based on variational graph autoencoder according to claim 1 or 4, characterized in that: When using the non-redundant shortest path as the primary quantitative representation of the target gene, the following optimization strategies are also included: Based on the correlation coefficient analysis, a longest path among all non-redundant shortest paths is selected for each target gene as the primary quantitative representation of the target gene; Or, based on multi-path analysis, multiple non-redundant shortest paths with a set length are selected for each target gene as the primary quantitative representation of the target gene; Alternatively, based on cross-correlation analysis, multiple non-redundant shortest paths with a set length are selected for each target gene as the primary quantitative representation of the target gene.

6. A gene quantization representation system based on variational graph autoencoder, characterized in that: include: Protein interaction network directed graph construction module, gene ontology tree graph construction module, gene primary quantitative representation calculation module, advanced protein interaction network directed graph construction module and gene advanced quantitative representation extraction module; The protein interaction network directed graph construction module is used to construct a protein interaction network directed graph according to the acquired protein interaction network data, using protein names as nodes and protein interaction scores as weights of edges between nodes; The gene ontology tree diagram construction module is used to construct a biological process tree diagram, a cell component tree diagram and a molecular function tree diagram according to the acquired gene ontology data, taking gene ontology terms as nodes, converting the hierarchical structure relationship between nodes into the weight of the edge between nodes, calculating the node numerical attributes according to the numerical sorting of the nodes; The gene primary quantitative representation calculation module is used to calculate the non-redundant shortest path between the root node and the node of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram as the primary quantitative representation of the target gene; The advanced protein interaction network directed graph construction module is used to fuse the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram, and then further combine them with the protein interaction network directed graph to obtain the advanced protein interaction network directed graph, specifically including: fusing the primary quantitative representations of the target gene in the biological process tree diagram, the cell component tree diagram and the molecular function tree diagram respectively in a splicing manner; combining the primary quantitative representations of the fused target gene with the protein interaction network directed graph, using the primary quantitative representations of the target gene as node attributes, and obtaining the advanced protein interaction network directed graph including node names, edge weights and node attributes; The gene high-level quantitative representation extraction module is used to input the high-level protein interaction network directed graph into the graph neural network model based on the variational graph autoencoder to complete the gene high-level quantitative representation, specifically including: inputting the high-level protein interaction network directed graph into the graph neural network model based on the variational graph autoencoder for unsupervised node embedding, the graph neural network model based on the variational graph autoencoder includes an encoder and a decoder structure, the encoder extracts intermediate features and then the decoder calculates the probability of edges between two nodes to obtain a reconstructed high-level protein interaction network directed graph, completing the gene high-level quantitative representation, so that the gene high-level quantitative representation includes protein interaction network information, gene ontology information and relationship information with other genes.

7. A gene quantization characterization device based on a variational graph autoencoder, comprising a memory and one or more processors, wherein the memory is used to store a computer program, characterized in that: The processor is used to implement the gene quantization characterization method based on variational graph autoencoder according to any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a computer, the gene quantization characterization method based on the variational graph autoencoder according to any one of claims 1 to 5 is implemented.

Images