Multi-omics phenotype driven network construction method and system based on graph neural network

Abstract

The application provides a multi-omics phenotype driving network construction method and system based on a graph neural network, which comprises the following steps: constructing a multi-omics heterogeneous graph according to SNPs, genes, metabolites and phenotypes of a population; extracting population feature embeddings of each node in the heterogeneous graph by using a graph attention network, calculating an importance score of each edge based on population feature embeddings of nodes at two ends of the edge, and constructing a core sub-network by screening edges according to the importance score; replacing the population features of each node in the heterogeneous graph with individual features of each sample, constructing a specific heterogeneous graph of each sample, extracting individual feature embeddings of each node, screening key nodes, and screening phenotype response edges according to the correlation between individual feature embeddings of nodes at two ends of each edge in the core sub-network and the phenotype; and screening edges with both key nodes at two ends and being phenotype response edges from the core sub-network, generating a simplified and trait-regulated phenotype driving network, and providing precise target points for molecular breeding.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for constructing a multi-omics phenotypic driven network based on a graph neural network (GNN). Background Technology

[0002] In crop molecular breeding, efficiently and accurately identifying genes, metabolites, and their regulatory relationships closely related to target traits from massive multi-omics data is both a key focus and a challenge. Traditional methods, such as genome-wide association studies (GWAS), expression quantitative trait loci (eQTL) analysis, or co-expression network analysis, typically utilize only single-omics data, making it difficult to systematically capture cross-level molecular interactions. Furthermore, these methods are often based on population averages, neglecting heterogeneity among individuals and failing to reveal phenotype-driven dynamic regulatory networks.

[0003] In recent years, graph neural networks (GNNs) have been introduced into multi-omics ensemble analysis. However, existing techniques mostly focus on phenotypic prediction tasks, lacking optimization of the network structure itself, discovery of unknown associations, and fine characterization of individual-level phenotypic response patterns. Therefore, there is an urgent need for a method that can integrate multi-omics data, utilize graph neural networks to learn node representations, and ultimately construct a finely tuned network closely related to phenotypes. Summary of the Invention

[0004] This invention provides a method and system for constructing a multi-omics phenotypic-driven network based on graph neural networks, which addresses the shortcomings of existing technologies such as insufficient integration of multi-omics, neglect of individual heterogeneity, and difficulty in constructing phenotypic-driven networks. It integrates genomic, transcriptomic, metabolomic, and phenotypic data and uses graph neural networks to construct a phenotypic-driven multi-omics regulatory network.

[0005] This invention provides a method for constructing a multi-omics phenotypic driven network based on graph neural networks, comprising:

[0006] Construct a multi-omics heterogeneity map based on the population's SNPs, genes, metabolites, and phenotypes;

[0007] The graph attention network is used to extract the group feature embedding of each node in the multi-omics heterogeneous graph. Based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, the importance score of each edge is calculated. The edges are selected according to the importance score to construct the core sub-network.

[0008] The group features of each node in the multi-omics heterogeneous graph are replaced with the individual features of each sample to construct a specific heterogeneous graph for each sample. The graph attention network is then used to extract the individual feature embeddings of each node in the specific heterogeneous graph.

[0009] Key nodes are selected by embedding the individual characteristics of each node, and phenotypic response edges are selected by embedding the individual characteristics of the nodes at both ends of each edge in the core sub-network and their correlation with the phenotype.

[0010] From the core sub-network, edges whose two ends are both key nodes and are phenotypic response edges are selected to generate a phenotypic driving network.

[0011] According to the multi-omics phenotypic driven network construction method based on graph neural networks provided by the present invention, before calculating the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, the method further includes:

[0012] The link prediction model is trained based on the existing edges in the multi-omics heterogeneous graph.

[0013] Based on the trained link prediction model, new cross-omics associations in the multi-omics heterogeneous graph are predicted, and the corresponding edges are obtained.

[0014] According to the present invention, a method for constructing a multi-omics phenotypic driven network based on a graph neural network calculates the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, including:

[0015] Calculate the dot product between the group feature embeddings of the two endpoints of each edge in the multi-omics heterogeneous graph, and use it as the importance score of each edge.

[0016] According to the present invention, a method for constructing a multi-omics phenotypic driven network based on a graph neural network, wherein the core sub-network is constructed by selecting edges according to the importance score, comprising:

[0017] Sort each type of edge in descending order of importance score;

[0018] The core subnetwork is constructed by selecting edges from a pre-defined proportion based on the edge sorting of each type.

[0019] According to the present invention, a method for constructing a multi-omics phenotypic driven network based on a graph neural network includes selecting key nodes by embedding the individual features of each node, comprising:

[0020] Calculate the cross-sample coefficient of variation of the individual feature embeddings of each node;

[0021] Calculate the correlation coefficients between each dimension of the individual feature embedding of each node and the phenotype, and take the maximum value of the correlation coefficients corresponding to each node as the phenotypic association strength of each node.

[0022] Key nodes are selected based on the cross-sample coefficient of variation and phenotypic association strength of each node.

[0023] According to the multi-omics phenotypic driven network construction method based on graph neural networks provided by the present invention, after calculating the correlation coefficients between each dimension of the individual feature embedding of each node and the phenotypic, the method further includes:

[0024] Based on the correlation coefficients between all dimensions embedded with the individual features of each node and the phenotype, construct the phenotypic response vector of each node;

[0025] Based on the phenotypic response vectors of each node, the nodes in the specific heterogeneous graph are clustered to obtain multiple node sets.

[0026] According to the present invention, a method for constructing a multi-omics phenotypic driven network based on a graph neural network is provided, which filters phenotypic response edges based on the correlation between the individual feature embeddings of the nodes at both ends of each edge in the core sub-network and the phenotypic characteristics, including:

[0027] Calculate the correlation between the dot product of individual feature embeddings of the nodes at both ends of each edge in the core sub-network and the phenotype;

[0028] Edges with a correlation greater than a preset threshold are selected as phenotypic response edges.

[0029] This invention also provides a multi-omics phenotypic driven network construction system based on graph neural networks, comprising:

[0030] The first building block is used to construct a multi-omics heterogeneity map based on the population's SNPs, genes, metabolites, and phenotypes.

[0031] The second construction module is used to extract the group feature embedding of each node in the multi-omics heterogeneous graph using a graph attention network, calculate the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, and select edges to construct a core sub-network based on the importance score.

[0032] The third construction module is used to replace the group features of each node in the multi-omics heterogeneous graph with the individual features of each sample, construct a specific heterogeneous graph for each sample, and use the graph attention network to extract the individual feature embeddings of each node in the specific heterogeneous graph.

[0033] The filtering module is used to embed and filter key nodes based on the individual characteristics of each node, and to filter phenotypic response edges based on the correlation between the individual characteristics of the nodes at both ends of each edge in the core sub-network and the phenotypic characteristics.

[0034] The generation module is used to select edges from the core sub-network that are both key nodes and are phenotypic response edges, and to generate a phenotypic driving network.

[0035] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-omics phenotypic driving network construction method based on graph neural networks as described above.

[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-omics phenotypic driving network construction method based on graph neural networks as described above.

[0037] The method and system for constructing multi-omics phenotypic driven networks based on graph neural networks provided by this invention have the following main advantages:

[0038] Integrating multi-omics information: through heterogeneous Figure 1 This one-time integration of four-dimensional molecular data and multiple types of known correlations allows for more comprehensive information utilization.

[0039] Network optimization and knowledge discovery: Graph attention networks are used for link prediction and core subnetwork extraction, which can discover new associations and simplify the network structure.

[0040] Achieving individualized analysis: An individual-specific embedding method is proposed to capture the dynamic changes of molecules among individuals without changing the graph topology, thus achieving a leap from the population to the individual.

[0041] Provides interpretable phenotypic associations: The invention’s concept of “phenotypic response vector” directly associates high-dimensional embeddings with phenotypes, providing a clear biological explanation for each node and functional module.

[0042] Focusing on the core regulatory network: The final phenotypic driven network is both concise and highly correlated with the target traits, providing verifiable molecular targets for breeding. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the multi-omics phenotypic driven network construction method based on graph neural networks provided by the present invention;

[0045] Figure 2 This is a schematic diagram of the overall process of the multi-omics phenotypic driven network construction method based on graph neural networks provided by the present invention;

[0046] Figure 3 This is a schematic diagram of the structure of the multi-omics phenotypic driven network construction system based on graph neural networks provided by the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0048] The following is combined Figure 1 This invention describes a method for constructing a multi-omics phenotypic driven network based on a graph neural network, comprising:

[0049] Step 101: Construct a multi-omics heterogeneity map based on the population's SNPs, genes, metabolites, and phenotypes;

[0050] We collected SNP genotypes, gene expression, metabolite content, and quantitative phenotype data from the population and constructed a four-dimensional weighted heterogeneous graph. Nodes include SNPs, genes, metabolites, and phenotypes. Edges include SNP-SNP linkage, SNP-gene association, gene-gene co-expression, gene-metabolite association, SNP-metabolite association, metabolite-metabolite association, and three types of known molecular-to-phenotype associations, totaling nine types of edges, and assigned weights based on biological priors or statistical tests.

[0051] Step 102: Extract the group feature embedding of each node in the multi-omics heterogeneous graph using a graph attention network. Calculate the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph. Select edges based on the importance score to construct a core sub-network.

[0052] Message passing can be performed using a two-layer heterogeneous graph attention network (GAT) to learn a low-dimensional dense embedding vector for each node.

[0053] The dot product of the embedded nodes at both ends of each existing edge can be calculated as the importance score of the edge. A certain percentage (e.g., 20%) of the edges with the highest scores can be retained according to edge type to construct a simplified core sub-network.

[0054] Step 103: Replace the group features of each node in the multi-omics heterogeneous graph with the individual features of each sample, construct a specific heterogeneous graph for each sample, and use the graph attention network to extract the individual feature embeddings of each node in the specific heterogeneous graph.

[0055] In this network, the individual characteristics of SNP nodes are genotype values ​​(scalars); the individual characteristics of gene nodes and metabolite nodes are two-dimensional vectors composed of the mean of the sample under control conditions and the treatment-control difference. Different node types are uniformly projected onto the hidden dimensions of the graph attention network through type-specific linear transformations.

[0056] Construct a specific heterogeneous graph for each individual: retain the same edge structure and weights as the population graph, but replace the population features of the nodes with the individual's own molecular measurements and perform population standardization.

[0057] The encoder in the trained graph attention network generates personalized embedding vectors for each node of each individual.

[0058] Step 104: Filter key nodes based on the individual characteristics of each node, and filter phenotypic response edges based on the correlation between the individual characteristics of the nodes at both ends of each edge in the core sub-network and the phenotypic characteristics.

[0059] Key nodes and phenotypic response edges are selected based on the individual feature embeddings of each node. For example, key nodes are selected based on the cross-sample coefficient of variation (CV) of the individual feature embeddings of each node, and edges whose average correlation between the individual feature embeddings of the two endpoints and the phenotypic trait is greater than a set value are selected as phenotypic response edges.

[0060] Step 105: Select edges from the core sub-network that are both key nodes and are phenotypic response edges to generate a phenotypic driving network.

[0061] Edges that are both critical nodes and are themselves responsive edges are selected from the core subnetwork to form the final phenotype-driven network. The complete flowchart is as follows: Figure 2 As shown.

[0062] This embodiment constructs a four-dimensional heterogeneous graph, trains a graph attention network, extracts individual-specific embeddings, defines phenotypic response vectors, and filters key nodes and edges to ultimately build a concise phenotypic-driven network that is regulated by traits. This provides precise targets for molecular breeding and realizes a systematic analysis from a global multi-omics network to an individualized phenotypic-driven network, significantly improving the accuracy and interpretability of breeding target screening.

[0063] Based on the above embodiments, this embodiment further includes the following step before calculating the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph:

[0064] The link prediction model is trained based on the existing edges in the multi-omics heterogeneous graph.

[0065] Based on the trained link prediction model, new cross-omics associations in the multi-omics heterogeneous graph are predicted, and the corresponding edges are obtained.

[0066] Using existing SNP-gene edges as positive samples and randomly sampling non-existent edges as negative samples, we train the link predictor and optimize the encoder parameters.

[0067] Based on the trained link predictor, the connection probability of unconnected cross-omics node pairs is calculated to predict new potential associations.

[0068] Based on the above embodiments, this embodiment calculates the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, including:

[0069] Calculate the dot product between the group feature embeddings of the two endpoints of each edge in the multi-omics heterogeneous graph, and use it as the importance score of each edge.

[0070] Based on the above embodiments, this embodiment constructs a core sub-network by filtering edges according to the importance score, including:

[0071] Sort each type of edge in descending order of importance score;

[0072] The core subnetwork is constructed by selecting edges from the top preset percentage (e.g., 40%) of each type of edge sorting.

[0073] Based on the above embodiments, this embodiment embeds and filters key nodes according to the individual characteristics of each node, including:

[0074] Calculate the cross-sample coefficient of variation (CV) of the individual feature embeddings of each node to measure the differences between individuals;

[0075] Calculate the correlation coefficients between each dimension of the individual feature embedding of each node and the phenotype, and take the maximum value of the correlation coefficients corresponding to each node as the phenotypic association strength of each node.

[0076] Key nodes are selected based on the cross-sample coefficient of variation and phenotypic association strength of each node.

[0077] When the phenotype is a continuous quantitative trait, the correlation coefficients between the dimensions of the individual feature embeddings at each node and the phenotype can be Pearson correlation coefficients. When the phenotype is a categorical trait, dotted bicollinear correlation coefficients or logistic regression are used instead of Pearson correlation coefficients.

[0078] Nodes with cross-sample coefficient of variation less than a preset value (e.g., 0.05) and phenotypic association strength greater than the median phenotypic association strength can be designated as key nodes.

[0079] Based on the above embodiments, this embodiment, after calculating the correlation coefficients between each dimension of the individual feature embedding of each node and the phenotype, further includes:

[0080] Based on the correlation coefficients between all dimensions embedded with the individual features of each node and the phenotype, construct the phenotypic response vector of each node;

[0081] Based on the phenotypic response vectors of each node, the nodes in the specific heterogeneous graph are clustered to obtain multiple node sets.

[0082] The vector formed by the correlation coefficients between all dimensions of the individual features embedded in each node and the phenotype is used as the phenotypic response vector.

[0083] Nodes can be ranked based on their effects to reduce computational burden. Their phenotypic response vectors are used, and nodes with similar response patterns are grouped into the same functional module through standardization and clustering algorithms (such as K-Means). Each module reflects a set of molecules that respond collaboratively to phenotypic changes.

[0084] The final output includes the predicted new associations, core subnetworks, a list of key nodes, phenotypic response modules, and the final phenotypic driving network.

[0085] Based on the above embodiments, this embodiment filters phenotypic response edges according to the correlation between the individual feature embeddings of the nodes at both ends of each edge in the core sub-network and the phenotypic characteristics, including:

[0086] Calculate the correlation between the dot product of individual feature embeddings of the nodes at both ends of each edge in the core sub-network and the phenotype;

[0087] Edges with a correlation greater than a preset threshold are selected as phenotypic response edges.

[0088] The following is an example of screening breeding targets for the area trait of rapeseed diseases, which includes the following steps:

[0089] Data preparation: Multi-omics data were collected from 138 rapeseed germplasm accessions, including 47,090 SNP genotypes (0 / 1 / 2 encoding), z-score-normalized gene expression and metabolite content (with control and treatment for each material), and the quantitative phenotype "disease_area". Nine weighted edges were constructed based on public databases such as linkage disequilibrium (LD) and eQTL.

[0090] Heterogeneous graph construction: Each SNP, gene, and metabolite is treated as an independent node, and the phenotype is treated as a virtual node. SNP features are genotype scalar values, while gene and metabolite features are 2D vectors composed of the control mean and treatment-control differences. All features are standardized using the population mean and standard deviation. Edge weights are logarithmically compressed and Min-Max normalized.

[0091] Graph Neural Network Training and Optimization: A two-layer heterogeneous graph attention network was constructed (layer 1 with 2 heads, layer 2 with 1 head, outputting 32-dimensional L2-normalized embeddings). Using SNP-gene edges as positive samples and an equal number of random edges as negative samples, a binary cross-entropy loss was used to train the encoder and link predictor (dot product + sigmoid). After training, candidate pairs were randomly sampled for each type of cross-omics edge, and the 50 pairs with the highest dot product scores were selected as the predicted new associations. The dot products of all existing edge embeddings were calculated, and the top 20% of high-scoring edges were retained stratified by edge type, resulting in a core subnetwork of approximately 196,810 edges.

[0092] Individual-specific embeddings and phenotypic association: For each sample, an individualized heterogeneous graph is constructed (reusing global edges and replacing them with the features of that sample). The trained encoder is used to extract the 32-dimensional individual embeddings of the nodes. The cross-sample coefficient of variation (CV) of each node embedding is calculated, and the correlation coefficients between each dimension of the embedding and the phenotype are calculated to obtain the phenotypic response vector (32-dimensional correlation coefficients). The maximum absolute correlation coefficient and the p-value are recorded.

[0093] Phenotypic-driven network construction: 11697 key nodes with p < 0.05 and CV > median were selected. For each edge in the core sub-network, the correlation between its embedded dot product sequence and phenotype was calculated, and 15303 response edges with p < 0.05 were selected. Edges with both ends being key nodes and themselves being response edges were further selected, ultimately constructing a phenotype-driven network containing approximately 4497 ​​edges.

[0094] Phenotypic response module analysis revealed that, using an independent ridge regression model effect ranking, the top 5000 SNPs and top 5000 genes were selected, with all metabolites retained. After standardizing the phenotypic response vectors of these nodes, K-Means (k=10) clustering was used to obtain 10 functional modules (profile coefficient 0.534), each containing multiple node types.

[0095] Results output: The output includes the predicted new associations, core subnetworks, key nodes, response edges, module partitioning, and the final phenotypic driven network file, providing highly reliable targets for subsequent functional validation and molecular breeding.

[0096] The following describes the multi-omics phenotypic driving network construction system based on graph neural networks provided by this invention. The multi-omics phenotypic driving network construction system based on graph neural networks described below can be referred to in correspondence with the multi-omics phenotypic driving network construction method based on graph neural networks described above.

[0097] like Figure 3 As shown, the system includes a first construction module 301, a second construction module 302, a third construction module 303, a filtering module 304, and a generation module 305, wherein:

[0098] The first building module 301 is used to construct a multi-omics heterogeneity map based on the SNPs, genes, metabolites and phenotypes of the population;

[0099] The second construction module 302 is used to extract the group feature embedding of each node in the multi-omics heterogeneous graph using a graph attention network, calculate the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, and filter the edges to construct a core sub-network based on the importance score.

[0100] The third construction module 303 is used to replace the group features of each node in the multi-omics heterogeneous graph with the individual features of each sample, construct a specific heterogeneous graph for each sample, and use the graph attention network to extract the individual feature embeddings of each node in the specific heterogeneous graph.

[0101] The filtering module 304 is used to embed and filter key nodes based on the individual characteristics of each node, and to embed and filter phenotypic response edges based on the correlation between the individual characteristics of the nodes at both ends of each edge in the core sub-network and the phenotypic characteristics.

[0102] The generation module 305 is used to select edges from the core sub-network that are both key nodes and are phenotypic response edges to generate a phenotypic driving network.

[0103] This embodiment constructs a four-dimensional heterogeneous graph, trains a graph attention network, extracts individual-specific embeddings, defines phenotypic response vectors, and filters key nodes and edges to ultimately build a concise phenotypic-driven network that is regulated by traits. This provides precise targets for molecular breeding and realizes a systematic analysis from a global multi-omics network to an individualized phenotypic-driven network, significantly improving the accuracy and interpretability of breeding target screening.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a multi-omics phenotypic driven network based on graph neural networks, characterized in that, include: Construct a multi-omics heterogeneity map based on the population's SNPs, genes, metabolites, and phenotypes; The graph attention network is used to extract the group feature embedding of each node in the multi-omics heterogeneous graph. Based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, the importance score of each edge is calculated. The edges are selected according to the importance score to construct the core sub-network. The group features of each node in the multi-omics heterogeneous graph are replaced with the individual features of each sample to construct a specific heterogeneous graph for each sample. The graph attention network is then used to extract the individual feature embeddings of each node in the specific heterogeneous graph. Key nodes are selected by embedding the individual characteristics of each node, and phenotypic response edges are selected by embedding the individual characteristics of the nodes at both ends of each edge in the core sub-network and their correlation with the phenotype. From the core sub-network, edges whose two ends are both key nodes and are phenotypic response edges are selected to generate a phenotypic driving network.

2. The method for constructing a multi-omics phenotypic driven network based on graph neural networks according to claim 1, characterized in that, Before calculating the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, the following steps are also included: The link prediction model is trained based on the existing edges in the multi-omics heterogeneous graph. Based on the trained link prediction model, new cross-omics associations in the multi-omics heterogeneous graph are predicted, and the corresponding edges are obtained.

3. The method for constructing a multi-omics phenotypic driven network based on graph neural networks according to claim 1, characterized in that, The importance score of each edge is calculated based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, including: Calculate the dot product between the group feature embeddings of the two endpoints of each edge in the multi-omics heterogeneous graph, and use it as the importance score of each edge.

4. The method for constructing a multi-omics phenotypic driven network based on graph neural networks according to claim 1, characterized in that, The core sub-network is constructed by selecting edges based on the importance scores, including: Sort each type of edge in descending order of importance score; The core subnetwork is constructed by selecting edges from a pre-defined proportion based on the edge sorting of each type.

5. The method for constructing a multi-omics phenotypic driven network based on graph neural networks according to claim 1, characterized in that, Key nodes are selected based on their individual characteristics, including: Calculate the cross-sample coefficient of variation of the individual feature embeddings of each node; Calculate the correlation coefficients between each dimension of the individual feature embedding of each node and the phenotype, and take the maximum value of the correlation coefficients corresponding to each node as the phenotypic association strength of each node. Key nodes are selected based on the cross-sample coefficient of variation and phenotypic association strength of each node.

6. The method for constructing a multi-omics phenotypic driven network based on graph neural networks according to claim 1, characterized in that, After calculating the correlation coefficients between each dimension of the individual feature embedding of each node and the phenotype, the following is also included: Based on the correlation coefficients between all dimensions embedded with the individual features of each node and the phenotype, construct the phenotypic response vector of each node; Based on the phenotypic response vectors of each node, the nodes in the specific heterogeneous graph are clustered to obtain multiple node sets.

7. The method for constructing a multi-omics phenotypic driven network based on graph neural networks according to claim 1, characterized in that, Based on the correlation between the individual feature embeddings of the two endpoints of each edge in the core sub-network and the phenotype, phenotypic response edges are selected, including: Calculate the correlation between the dot product of individual feature embeddings of the nodes at both ends of each edge in the core sub-network and the phenotype; Edges with a correlation greater than a preset threshold are selected as phenotypic response edges.

8. A multi-omics phenotypic driven network construction system based on graph neural networks, characterized in that, include: The first building block is used to construct a multi-omics heterogeneity map based on the population's SNPs, genes, metabolites, and phenotypes. The second construction module is used to extract the group feature embedding of each node in the multi-omics heterogeneous graph using a graph attention network, calculate the importance score of each edge based on the group feature embedding of the nodes at both ends of each edge in the multi-omics heterogeneous graph, and select edges to construct a core sub-network based on the importance score. The third construction module is used to replace the group features of each node in the multi-omics heterogeneous graph with the individual features of each sample, construct a specific heterogeneous graph for each sample, and use the graph attention network to extract the individual feature embeddings of each node in the specific heterogeneous graph. The filtering module is used to embed and filter key nodes based on the individual characteristics of each node, and to filter phenotypic response edges based on the correlation between the individual characteristics of the nodes at both ends of each edge in the core sub-network and the phenotypic characteristics. The generation module is used to select edges from the core sub-network that are both key nodes and are phenotypic response edges, and to generate a phenotypic driving network.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-omics phenotypic driven network construction method based on graph neural networks as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-omics phenotypic driven network construction method based on graph neural networks as described in any one of claims 1 to 7.

Images