Whole genome selection method and device of graph neural network, equipment and storage medium

Abstract

The invention relates to the technical field of biological information, and provides a whole genome selection method, device and equipment of a graph neural network and a storage medium, and the whole genome selection method of the graph neural network comprises the following steps: mapping single nucleotide polymorphism (SNP) to a gene level, and converting the SNP into a gene embedding vector; constructing a gene interaction network based on preset multi-source biological priori knowledge; inputting the target character into a graph neural network model to obtain a predicted value of the target character output by the graph neural network model; wherein the graph neural network model is determined based on a gene embedding vector and a gene interaction network. According to the method, the SNP is mapped to the gene level, and the gene interaction network is combined, so that the prediction result has clear biological significance; and the gene interaction network is constructed based on the preset multi-source biological priori knowledge, so that interaction information between genes can be fully utilized, and the accuracy of target character prediction is improved.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a method, apparatus, device, and storage medium for whole-genome selection using graph neural networks. Background Technology

[0002] In the field of plant breeding, genomic selection (GS) technology has become a key method to accelerate the breeding process. GS improves breeding efficiency by establishing predictive models between genotype and phenotype, enabling accurate assessment of the breeding value of non-phenotypical individuals.

[0003] Traditional genome-wide selection primarily relies on linear models for prediction. Classical models such as GBLUP (Genomic Best Linear Unbiased Prediction) and RRBLUP (Ridge Regression BLUP) assume that all marker effects follow the same normal distribution, estimating genetic relationships between individuals by constructing a genomic kinship matrix. Bayesian models (such as BayesA and BayesB) allow for different effect distributions for different markers, making them more suitable for handling major genes. Penalized regression methods such as LASSO regression and elastic networks address high-dimensional feature selection problems by introducing penalty terms.

[0004] In recent years, some new genome-wide selection methods have emerged using deep learning models. For example, DNNGP (Deep Neural Network Gaussian Process) integrates multi-omics data using multi-layer neural networks, demonstrating superior prediction accuracy and computational efficiency compared to traditional methods on large-scale breeding data.

[0005] However, DNNGP still has significant limitations: it lacks the ability to integrate gene functional networks, struggles to explicitly model complex interactions between genes, has limited biological interpretability, and its model architecture lacks flexibility. These limitations restrict DNNGP's potential for understanding the molecular mechanisms of complex traits and supporting precision breeding decisions. Summary of the Invention

[0006] This application provides a method, apparatus, device, and storage medium for whole-genome selection using graph neural networks. This method aims to solve the technical problems existing in the prior art, specifically the lack of effective integration of gene functional networks and the limited biological interpretability of the prior art, which leads to low prediction accuracy of target traits.

[0007] This application provides a genome-wide selection method using graph neural networks, comprising: mapping single nucleotide polymorphisms (SNPs) to the gene level and converting them into gene embedding vectors; constructing a gene interaction network based on pre-defined multi-source biological prior knowledge; inputting the target trait into the graph neural network model to obtain the predicted value of the target trait output by the graph neural network model; wherein the graph neural network model is determined based on the gene embedding vectors and the gene interaction network.

[0008] According to the genome-wide selection method of graph neural network provided in this application, single nucleotide polymorphisms (SNPs) are mapped to the gene level and converted into gene embedding vectors. The method includes: mapping by using an extended gene region definition to assign SNPs to corresponding genes; for each gene, sorting all SNPs mapped to that gene according to their physical location on the genome to form the SNP sequence of that gene; and inputting the SNP sequence of each gene into a preset sequence coding model to obtain the gene embedding vector corresponding to each gene output by the preset sequence coding model, wherein the gene embedding vector serves as the initial node feature of the graph neural network model.

[0009] According to the genome-wide selection method of graph neural networks provided in this application, multi-source biological prior knowledge includes gene co-expression networks, transcriptional regulation networks, and miRNA regulation networks; based on the preset multi-source biological prior knowledge, a gene interaction network is constructed, including: weighted fusion of gene co-expression networks, transcriptional regulation networks, and miRNA regulation networks to obtain a preliminary gene interaction network; and preprocessing the preliminary gene interaction network to obtain the final gene interaction network.

[0010] According to the genome-wide selection method of graph neural network provided in this application, when the hybrid individual is represented as a gene interaction network, the nodes are genes, the node features are gene SNP embedding vectors, and the edges are the interaction relationships between genes; wherein, each gene interaction network is associated with a true phenotypic value and a linear model prediction value; the graph neural network model is used to propagate and aggregate information on the gene interaction network to learn the nonlinear interaction between genes and output the prediction value of the target trait.

[0011] According to the genome-wide selection method using a graph neural network provided in this application, the predicted value of the target trait is obtained in the following manner: the target trait is predicted using a first model to obtain a linear predicted value; the target trait is predicted using a second model to obtain a residual predicted value; the linear predicted value and the residual predicted value are combined to obtain a predicted value; wherein, the first model is a kernel matrix linear model based on general combining ability (GCA) and specific combining ability (SCA); and the second model is a graph neural network model.

[0012] The whole-genome selection method for graph neural networks provided in this application further includes: in the gene interaction network, using the breeding value and cumulative heterozygous dominance effect at the gene level as node features.

[0013] This application also provides a genome-wide selection device for a graph neural network, comprising: a gene embedding vector module for mapping single nucleotide polymorphisms (SNPs) to the gene level and converting them into gene embedding vectors; a gene interaction network module for constructing a gene interaction network based on preset multi-source biological prior knowledge; and a prediction value module for inputting the target trait into the graph neural network model to obtain the predicted value of the target trait output by the graph neural network model; wherein the graph neural network model is determined based on the gene embedding vectors and the gene interaction network.

[0014] This application 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 whole-genome selection method of the graph neural network as described above.

[0015] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the whole-genome selection method of the graph neural network as described above.

[0016] This application also provides a computer program product, including a computer program that, when executed by a processor, implements a genome-wide selection method for graph neural networks as described above.

[0017] This application provides a genome-wide selection method, apparatus, device, and storage medium using graph neural networks. The method includes: mapping single nucleotide polymorphisms (SNPs) to the gene level and converting them into gene embedding vectors; constructing a gene interaction network based on pre-defined multi-source biological prior knowledge; inputting the target trait into the graph neural network model to obtain the predicted value of the target trait output by the graph neural network model; wherein the graph neural network model is determined based on the gene embedding vectors and the gene interaction network. Through the above methods, this application maps SNPs to the gene level and combines them with a gene interaction network, giving the prediction results clear biological significance; and constructing a gene interaction network based on pre-defined multi-source biological prior knowledge can fully utilize the interaction information between genes; the graph neural network model combines the information from gene embedding vectors and the gene interaction network, enabling a more comprehensive capture of the complex relationship between genes and the target trait, thereby improving the accuracy of target trait prediction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application 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 application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of the genome-wide selection method for graph neural networks provided in this application embodiment.

[0020] Figure 2 This is a schematic diagram of the actual process of the whole-genome selection method of graph neural network provided in the embodiments of this application.

[0021] Figure 3 This is a schematic diagram of the structure of the whole genome selection device of the graph neural network provided in the embodiments of this application.

[0022] Figure 4 This is a schematic diagram of the physical structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0025] Among related technologies, whole-genome selection technology has the following obvious limitations: 1. Insufficient utilization of structural information: Traditional linear models treat genomic data as independent features, ignoring the complex regulatory network structure between genes. Even deep learning models such as CNN and RNN mainly treat genomic data as one-dimensional sequences, failing to fully capture the multidimensional interactions between genes.

[0026] 2. Limited integration of prior biological knowledge: Relevant models rarely systematically integrate known prior knowledge of gene networks. This biological knowledge is crucial for understanding the genetic mechanisms of complex traits, but it is often overlooked in model construction.

[0027] 3. Insufficient Granularity of Feature Representation: Most current methods represent features at the SNP (Single Nucleotide Polymorphism) marker level, lacking embedding representation at the gene functional unit level. As functional units, genes have a more biologically significant impact on phenotypes due to their overall effects and interactions, but related technologies rarely model this level.

[0028] 4. Poor model interpretability: Deep learning models have superior predictive performance, but their "black box" nature makes it difficult to reveal the underlying biological mechanisms, which limits their application value in assisted molecular design breeding.

[0029] Based on this, this application provides a genome-wide selection method using graph neural networks. This method analyzes at the gene level and considers the interactions between genes, thus offering better biological interpretability. It allows for a deeper understanding of the intrinsic connections between genes and target traits from the perspective of gene networks, contributing to the revelation of biological genetic mechanisms and providing valuable information for research in related fields.

[0030] Please see Figure 1 , Figure 1 This is a schematic flowchart of the genome-wide selection method for graph neural networks provided in this embodiment. In this embodiment, the genome-wide selection method for graph neural networks may include steps S110 to S130, each step of which is as follows: S110: Maps single nucleotide polymorphisms (SNPs) to the gene level and converts them into gene embedding vectors.

[0031] S120: Construct a gene interaction network based on pre-defined multi-source biological prior knowledge.

[0032] S130: Input the target trait into the graph neural network model to obtain the predicted value of the target trait output by the graph neural network model; wherein, the graph neural network model is determined based on gene embedding vectors and gene interaction networks.

[0033] Single nucleotide polymorphisms (SNPs) are a common form of genetic variation in the genome. This embodiment maps SNPs to the gene level. This gene-level feature representation method not only reduces feature dimensionality but also preserves richer biological information, providing high-quality node features for subsequent graph neural network learning.

[0034] Genes do not exist in isolation within organisms; they interact in complex ways, such as through regulation and synergistic effects. In this embodiment, a gene interaction network can be constructed based on pre-defined multi-source biological prior knowledge. This network reflects the functional connections and mutual influences between genes.

[0035] Specifically, a gene interaction network is a graph structure consisting of genes as nodes and biological associations between genes as edges.

[0036] Furthermore, this embodiment also uses a graph neural network model, which is a neural network model specifically designed to process graph-structured data. Gene embedding vectors and gene interaction networks constitute graph-structured data.

[0037] In this context, the target trait is used as input. The graph neural network can learn the relationships between genes and the association between genes and the target trait. It can use the nodes (genes) and edges (relationships between genes) in the network to transmit information and extract features, thereby outputting the predicted value of the target trait.

[0038] In summary, this application provides a genome-wide selection method based on gene-level embedding using graph neural networks. This method effectively utilizes graph neural networks to capture complex interactions between genes and systematically integrates prior knowledge of gene networks to construct an algorithmic model capable of achieving fine-grained gene-level embedding representations. Compared to related methods, this application not only improves the accuracy of target trait prediction but also greatly enhances the biological interpretability of the model, providing a more efficient and accurate method for predicting breeding value in the field of plant breeding.

[0039] In some embodiments, the step of mapping single nucleotide polymorphisms (SNPs) to the gene level and converting them into gene embedding vectors may specifically include: An extended gene region definition is used for mapping, and SNPs are assigned to corresponding genes. For each gene, all SNPs mapped to that gene are sorted according to their physical location on the genome to form the SNP sequence of that gene. The SNP sequence of each gene is input into a preset sequence coding model to obtain the gene embedding vector corresponding to each gene output by the preset sequence coding model. The gene embedding vector serves as the initial node feature of the graph neural network model.

[0040] In this embodiment, expanding the gene region means including surrounding regions that may affect gene function. Since SNPs can exist in the coding region, non-coding region, and surrounding regulatory regions of a gene, expanding the gene region definition allows for a more comprehensive allocation of potentially function-related SNPs to their corresponding genes. This enables subsequent gene-based analyses to encompass richer genetic information.

[0041] The genome is the sum of all genetic material in an organism, and the physical location of SNPs on the genome contains important information. Sequencing SNPs mapped to the same gene according to their physical location on the genome, forming SNP sequences, helps preserve the relative positional relationships between SNPs. These positional relationships may reflect information such as gene structural characteristics and linkage disequilibrium between SNPs.

[0042] The pre-defined sequence encoding model is a trained model that converts SNP sequences into numerical gene embedding vectors. Gene embedding vectors are low-dimensional vector representations that compress and abstract information from SNP sequences, presenting gene features as vectors. Graph neural networks are neural networks specifically designed for graph-structured data. Using gene embedding vectors as initial node features, they leverage the network structure to learn complex interactions between genes, thereby better capturing the association between genes and target traits.

[0043] In this embodiment, SNP mapping by expanding gene region definitions enables more comprehensive collection of gene-related SNP information, allowing gene-level analysis to encompass more potential genetic information. SNPs are ordered by physical location to form SNP sequences, preserving their relative positional relationships and providing richer features for subsequent encoding and analysis. These features more accurately reflect gene structure and genetic characteristics, helping to improve the model's understanding and processing of gene information. Gene embedding vectors, as initial node features of the graph neural network model, provide high-quality input. Graph neural networks can utilize these features to learn complex interactions between genes, thereby more accurately predicting target traits and improving the model's prediction accuracy and generalization ability. Based on the refined embedding representation at the gene level, the model's output can be interpreted in conjunction with the gene's biological information. The SNP and positional information contained in the gene embedding vectors are related to gene function and structure, helping to understand the model's prediction results from a biological perspective and providing more valuable references for biological research and practical applications.

[0044] For example, SNP markers are mapped to the gene level and converted into gene embedding vectors. The specific steps are as follows: Step 1: Determining the gene-SNP mapping rules.

[0045] In this embodiment, an extended gene region definition can be used to map SNPs to genes. Specifically, if an SNP is located within 3000 bp upstream of the transcription start site, within the entire gene region, or within 3000 bp downstream of the transcription termination site, then the SNP is assigned to the corresponding gene. In this embodiment, this extended definition method can more comprehensively consider SNP information in the surrounding regions of the gene, which helps to uncover richer genetic information.

[0046] It is worth noting that in this embodiment, miRNAs can also be treated as genes. This is because miRNAs play an important role in biological processes such as gene regulation. Including them in the gene category for SNP mapping makes the entire mapping system more complete and comprehensively covers biologically relevant information, providing a richer data foundation for subsequent analysis.

[0047] Step 2: Construction of SNP sequences in gene regions.

[0048] After mapping SNPs to genes, for each gene, its corresponding SNP sequence needs to be constructed. Specifically, all SNPs mapped to that gene are ordered according to their physical location on the genome, thus forming the SNP sequence for that gene. This physical location-based ordering method preserves the spatial order information between SNPs, which is helpful for subsequent analysis of the positional relationships and interactions between SNPs.

[0049] For genes without SNP mappings, special padding markers can be used to ensure data consistency and integrity. This avoids missing or incomplete data in subsequent processing, ensuring the smooth progress of the entire data processing workflow.

[0050] The third step is to use Transformer encoding to generate gene embedding vectors.

[0051] The SNP sequence of each gene is input into the pre-trained Transformer encoder. The Transformer encoder can effectively capture the positional relationships and interdependencies between SNPs through the self-attention mechanism. The self-attention mechanism allows the model to weight information according to the correlation between SNPs, thereby better learning the complex association patterns between SNPs.

[0052] The Transformer model can effectively handle SNP sequences of varying lengths and learn complex association patterns between SNPs. After processing by the Transformer encoder, a 128-dimensional gene embedding vector can be generated. This high-dimensional embedding vector can more comprehensively represent the gene's feature information, providing richer feature representations for subsequent analysis and model training.

[0053] It should be noted that the SNP embedding method in this embodiment is flexible and can be adjusted according to specific needs and scenarios in practical applications. For example, when high computational efficiency is required, the Transformer encoder can be replaced with LSTM, which has higher computational efficiency; when pursuing higher performance, other dimensional Transformers can be selected, such as a 1028-dimensional Transformer; in addition, other embedding methods can be used as needed to meet different application requirements.

[0054] In some embodiments, multi-source biological prior knowledge includes gene co-expression networks, transcriptional regulatory networks, and miRNA regulatory networks; the step of constructing a gene interaction network based on pre-defined multi-source biological prior knowledge may specifically include: The gene co-expression network, transcriptional regulation network, and miRNA regulation network were weighted and fused to obtain a preliminary gene interaction network; the preliminary gene interaction network was then preprocessed to obtain the final gene interaction network.

[0055] In this embodiment, the gene co-expression network describes the correlation of expression levels among genes. When multiple genes exhibit similar expression patterns under different experimental conditions or in different samples, they may be involved in the same biological process or regulatory pathway.

[0056] Transcription is the first step in gene expression, and the transcriptional regulatory network illustrates the relationship between transcription factors (a class of proteins that bind to DNA and regulate gene transcription) and the target genes they regulate. Transcription factors can activate or inhibit the transcription of target genes, thereby affecting gene expression levels.

[0057] miRNAs (microRNAs) are a class of short non-coding RNAs that regulate gene expression by binding to target mRNAs (messenger RNAs), inhibiting mRNA translation, or promoting mRNA degradation. The miRNA regulatory network describes the interactions between miRNAs and target genes.

[0058] Different biological networks reflect the interactions between genes from different perspectives. Gene co-expression networks can be used to represent the correlation of gene expression levels, transcriptional regulation networks can be used to represent the regulatory relationship between transcription factors and target genes, and miRNA regulation networks can be used to represent the regulation of gene expression by miRNAs.

[0059] Therefore, this embodiment utilizes the weighted fusion of these three networks to comprehensively leverage their respective advantages and more fully reflect the interactions between genes. The weights can be determined based on factors such as the reliability and information richness of each network.

[0060] It should be noted that the preliminary gene interaction network may contain some noisy data, such as spurious connections or incomplete information. This embodiment uses a preprocessing step to clean and optimize this data.

[0061] Optionally, preprocessing operations may include removing low-confidence connections, filling in missing information, and normalizing the network. For example, in gene co-expression networks, there may be some weakly correlated connections due to experimental errors. By setting a confidence threshold, removing these low-confidence connections can improve the quality of the network.

[0062] In summary, this embodiment, by integrating prior biological knowledge from multiple sources, enables the gene interaction network to comprehensively consider the interactions between genes at multiple levels, including expression levels, transcriptional regulation, and miRNA regulation. This provides a more complete picture of the complex interactions between genes than a single network. This helps in discovering more potential gene regulatory pathways and biological processes. Preprocessing the initial gene interaction network removes noise and incomplete information, improving the network's quality and reliability. The optimized gene interaction network more accurately reflects the true interactions between genes, providing a more reliable foundation for subsequent biological research and analysis.

[0063] For example, integrating multi-source biological prior knowledge to construct a biologically meaningful gene interaction network specifically includes the following steps: Step 1: Integration of co-expression networks.

[0064] The construction of gene co-expression networks is based on multiple gene map datasets. Specifically, the internal JHAX gene map dataset (including 21,568 genes and 28,240,271 edges) and the publicly available B73 gene map dataset (including 22,303 genes and 28,968,592 edges) can be used to construct gene co-expression networks.

[0065] These two datasets provide a rich data foundation for gene co-expression networks from different sources. By integrating them, we can more comprehensively reflect the co-expression relationships between genes, laying the foundation for the subsequent construction of a complete gene interaction network.

[0066] Step 2: Integration of transcription factor regulatory networks.

[0067] Transcription factors play a crucial role in gene expression regulation. This step integrates transcription factor-target gene relationships from the MaizeGRN database, which contains 2,056 transcription factors and 35,790 target genes, totaling 34,483,530 regulatory relationships.

[0068] The regulatory relationship between transcription factors and target genes is an important component of the gene regulatory network. By integrating this information, we can gain a deeper understanding of the regulatory mechanisms of gene expression and further improve the structure of the gene interaction network.

[0069] Step 3: Integration of miRNA regulatory network.

[0070] miRNAs play a crucial role in the posttranscriptional regulation of gene expression. This step integrates miRNA-target gene regulatory relationships, which include 323 miRNAs and 3,954 target genes, totaling 8,944 regulatory edges.

[0071] The regulatory relationship between miRNAs and target genes adds information about post-transcriptional regulation to gene interaction networks, enabling the constructed networks to more comprehensively reflect the complex process of gene expression regulation.

[0072] Step 4: Network Convergence and Optimization.

[0073] After constructing the three networks mentioned above, they are weighted and fused. This weighted fusion method can reasonably allocate the contribution of each network in the final gene interaction network according to the importance and reliability of different networks.

[0074] To reduce network complexity, improve computational efficiency, and ensure that the network retains the most biologically significant information, this embodiment can also apply a network pruning algorithm to process the fused preliminary gene interaction network. By pruning, edges that contribute little to the overall network structure and biological significance are removed, ultimately constructing a gene interaction network suitable for graph neural network computation.

[0075] It should be noted that, in this embodiment, a relatively small-scale gene interaction network was chosen for proof-of-concept purposes. However, this application has good scalability and can handle very large-scale networks, supporting up to 20M+ nodes. Larger-scale networks can provide richer biological information, but they increase the complexity of the graph neural network (GNN) computational steps.

[0076] In practical applications, those skilled in the art can weigh the relationship between network size and computational complexity and select a gene interaction network of appropriate size for analysis based on specific research objectives, computing resources, and performance requirements.

[0077] In some embodiments, when a hybrid individual is represented as a gene interaction network, the nodes are genes, the node features are gene SNP embedding vectors, and the edges are the interaction relationships between genes; wherein, each gene interaction network is associated with a true phenotypic value and a linear model prediction value; the graph neural network model is used to propagate and aggregate information in the gene interaction network to learn the nonlinear interactions between genes and output the prediction value of the target trait.

[0078] In this embodiment, hybrid individuals are represented as gene interaction networks. Genes are the basic units of genetic information in organisms. In this gene interaction network, each gene is regarded as a node, so that genes can be treated as independent research units, which facilitates the analysis of the interrelationships between genes and their effects on target traits.

[0079] As mentioned earlier, single nucleotide polymorphisms (SNPs) are variations in a single nucleotide in the genome. Gene SNP embedding vectors are vector representations obtained by encoding and transforming gene-related SNP information. These vectors contain genetic variation information of genes and are input into graph neural networks as node features. They can provide the network with specific genetic characteristics of genes, helping the model learn the characteristics and functions of genes.

[0080] Genes do not exist in isolation; they interact within organisms through various molecular mechanisms, such as transcriptional regulation and signal transduction. The interactions between genes represented by edges reflect these complex interactions. Graph neural networks can utilize the information from these edges to propagate and aggregate information between nodes, thereby learning patterns of synergistic effects between genes.

[0081] It's important to note that phenotype refers to the observable characteristics exhibited by an organism, such as plant height and yield, or animal weight and coat color. True phenotypic values ​​are obtained through actual measurements and represent the organism's actual performance under specific environmental conditions. In model training, true phenotypic values ​​serve as the target value, used to measure the accuracy of the graph neural network model's predictions.

[0082] A linear model is a simple statistical model that assumes a linear relationship between variables. Linear model predictions are based on predictions of the target trait. Linking these predictions to gene interaction networks can serve as reference information to assist graph neural network models in their learning. For example, a linear model may capture a partial linear relationship between genes and the target trait, and graph neural networks can further learn non-linear interactions based on this.

[0083] The core function of graph neural networks is to perform information propagation and aggregation on the graph structure. In gene interaction networks, information from each node (gene) is transmitted to neighboring nodes along the edges (inter-gene interactions), while nodes aggregate information from neighboring nodes. Through multiple iterative processes of information propagation and aggregation, nodes can acquire global information about the entire network, thereby better understanding their role in the network and their relationships with other genes.

[0084] Gene interactions are often nonlinear, meaning that the effect of one gene on a target trait may be influenced by the combined effects of multiple other genes, and these interactions are not simply linear superpositions. Graph neural networks have powerful nonlinear modeling capabilities, and this embodiment can capture these complex nonlinear interactions by learning information from gene interaction networks.

[0085] Through information propagation and aggregation, graph neural networks can predict target traits based on learned inter-gene interaction patterns and output predicted values. These predicted values ​​can be used to evaluate the performance of hybrids, providing a basis for decision-making in practical applications such as breeding.

[0086] In this embodiment, graph neural networks can learn nonlinear interactions between genes. Compared to traditional linear models, they can more comprehensively and accurately capture the complex relationships between genes and target traits, thereby improving the accuracy of target trait prediction. This is of great significance for fields such as plant breeding and animal husbandry, as it can help screen hybrid individuals with superior traits. By analyzing the gene interaction patterns learned by graph neural networks, we can gain a deeper understanding of the regulatory mechanisms of genes within organisms. In the process of hybridization breeding, using graph neural networks to predict the target traits of hybrid individuals can help screen potential individuals in advance, reduce unnecessary breeding experiments and time costs, improve breeding efficiency, and accelerate the cultivation of superior varieties.

[0087] In some embodiments, in order to effectively capture nonlinear interactions between genes based on the constructed gene interaction network and gene embedding representation, a graph neural network model is designed. The specific steps and logic are as follows: Step 1: Determine the network architecture.

[0088] A multi-layer Graph Convolutional Network (GCN) structure is used as the basic architecture. After each GCN layer, batch normalization and layer normalization operations are connected. Batch normalization helps accelerate the model's convergence speed and reduce internal covariate bias, while layer normalization improves the model's stability across different samples. This approach significantly enhances the stability of model training.

[0089] In this embodiment, 3-4 hidden layers were tested, with the feature dimension of each layer set to 32, 64, or 128, respectively. This setting was intended to explore the impact of different number of layers and feature dimensions on model performance, thereby finding the network structure most suitable for capturing nonlinear interactions between genes.

[0090] Step 2: Introduce attention pooling mechanism.

[0091] Graph attention pooling is used to aggregate node-level representations into graph-level representations. In gene interaction networks, different genes may have varying importance for phenotypic prediction. Graph attention pooling enables the model to automatically learn the importance weights of these different genes, thereby more accurately aggregating information and enhancing the model's ability to capture complex relationships between genes.

[0092] Step 3: Apply residual learning strategies.

[0093] GNN models employ a residual learning strategy, focusing on learning the residuals between linear model predictions and the true phenotype. Linear models typically capture additive gene effects well, but their ability to capture nonlinear interactions between genes is limited. GNN models, however, can supplement the modeling of nonlinear interactions by learning residuals. The final prediction result is a weighted sum of the linear predictions and the GNN residual predictions; this design effectively combines the advantages of both linear and GNN models.

[0094] Step 4: Perform hyperparameter optimization.

[0095] To achieve optimal model performance, a Bayesian optimization method based on Hyperopt is used to systematically search for the following hyperparameters: hidden_channels: The value range is [32, 64, 128]. This parameter determines the feature dimension of each hidden layer. Different feature dimensions will affect the model's ability to express gene features.

[0096] num_layers: The value range is [3,4], which represents the number of hidden layers. The number of layers will affect the complexity of the model and its ability to capture non-linear relationships.

[0097] dropout_rate: The value ranges from [0.1, 0.3] and is used to prevent the model from overfitting by randomly dropping some neurons to enhance the model's generalization ability.

[0098] learning_rate: The value ranges from [1e-5, 1e-3] (logarithmic scale), which controls the step size of parameter updates during model training. An appropriate learning rate helps the model converge quickly and avoid getting stuck in local optima.

[0099] weight_decay: The value ranges from [1e-6, 1e-4] (logarithmic scale) and is used to regularize the weights of the model to prevent overfitting.

[0100] batch_size: The value range is [32, 64], which represents the number of samples input to the model each time it is trained. An appropriate batch size can improve training efficiency and model stability.

[0101] It should be noted that GNN hyperparameter optimization can be performed on a larger hyperparameter space, which may contain better combinations of hyperparameters, thus obtaining more accurate model prediction results. Furthermore, the GNN model in this application is not the only option; in practical applications, it can be replaced with any other GNN design according to specific needs to better capture the nonlinear effects in hybrid prediction and enhance the model's adaptability and performance.

[0102] To improve the accuracy of hybrid combination trait prediction, this application integrates gene interaction networks and gene embedding representations into graph neural networks (GNNs). Through a multi-layer message passing mechanism, gene features are propagated and fused under biological association constraints, and nonlinear residual correction is performed based on baseline predictions. The logic and function of each part will be explained in detail below: I. Introduction of Gene Interaction Networks Gene interaction networks, serving as the structural foundation of the entire system, consist of nodes and edges. Nodes represent genes, while edges embody the biological relationships between genes. These biological relationships are an abstract expression of the complex interactions and regulation of genes within an organism.

[0103] In graph neural networks, the gene interaction network takes the form of an edge index matrix (edge_index) as input to the graph structure. The edge_index defines the adjacency relationships for message passing, determining the flow path of information between gene nodes. During message passing, each gene node only exchanges information with its neighboring nodes that are connected by edges in the gene interaction network. This mechanism preserves the topological constraints of biological networks, ensuring that information transmission conforms to the actual biological relationships between genes and avoiding irrational information exchange.

[0104] II. Introduction of Gene Embedding Representation Gene embedding representation is a numerical representation of gene features, derived from the SNP encoding corresponding to each gene. This SNP encoding contains the genotypic information of paternal 1 and paternal 2. These SNP encodings are processed by a SNP encoder (such as a linear layer, convolutional layer, or autoencoder) and converted into a fixed-dimensional vector representation, for example, 128 dimensions. This fixed-dimensional vector effectively captures the key features of the gene, providing a unified feature representation for subsequent analysis.

[0105] This embedding vector forms the initial node feature matrix of the graph neural network, with each row corresponding to the feature of a gene node. Therefore, the data input to the GNN is in the form of (X, edge_index), where X comes from the gene embedding and represents the feature information of each gene node; edge_index comes from the gene interaction network structure and specifies the connection relationship between nodes.

[0106] III. Integration Methods in GNN Architecture Feature Encoding: The gene embedding vector is input into the `genotype_encoder` for processing. The `genotype_encoder` employs linear transformation and normalization operations. Linear transformation converts the gene embedding vector into more representative features; normalization ensures the feature data has a uniform scale, improving the model's training stability and convergence speed. The features processed in this step serve as the initial node features for the GNN.

[0107] Message passing based on gene interaction networks: Each convolutional layer of a Gene Neural Network (GNN) (such as GCNConv) utilizes the gene interaction network topology defined by the edge_index to aggregate features from neighboring nodes to the current node. During the aggregation process, the flow of information is strictly restricted by the structure of the gene interaction network; information exchange can only occur between neighboring nodes connected by edges. This information aggregation method based on biological associations ensures that information flow conforms to the connection patterns of biological networks.

[0108] Graph-level Representation and Downstream Prediction: After multiple layers of message passing, the features of each gene node are updated. To obtain a holistic representation of the hybrid combination, an attentional-weighted global pooling method is used to integrate the features of all gene nodes. The attention mechanism automatically assigns weights based on the importance of gene nodes, highlighting the role of key genes, thereby obtaining a more representative global representation of the hybrid combination.

[0109] The graph is then concatenated with the GBLUP baseline predictions, and the final trait prediction results are output through the residual prediction module. The residual prediction module can perform nonlinear residual corrections on the baseline predictions, compensating for the shortcomings of the baseline prediction model and further improving prediction accuracy.

[0110] In summary, the gene interaction network provides structural support for the graph neural network, namely adjacency relationships, which determine the path of information transmission between nodes and ensure that information transmission conforms to biological realities. Gene embedding representations provide the initial node features for the graph neural network, serving as the "content" of information transmission and containing key gene feature information. Under the constraints of the gene interaction network structure, the graph neural network architecture utilizes gene embedding representations for multi-layer feature propagation and fusion, extracting higher-order representations of hybrid combinations, which are ultimately used for trait prediction. Through this organic combination, the advantages of both the gene interaction network and gene embedding representations are fully utilized, achieving high-precision prediction of hybrid combination traits.

[0111] In some embodiments, the predicted value of the target trait is obtained by: using a first model to predict the target trait to obtain a linear predicted value; using a second model to predict the target trait to obtain a residual predicted value; and combining the linear predicted value and the residual predicted value to obtain a predicted value; wherein the first model is a kernel matrix linear model based on general combining ability (GCA) and special combining ability (SCA); and the second model is a graph neural network model.

[0112] In hybridization breeding, general combining ability (GFA) refers to the average ability of a parent to produce offspring with a particular trait across a series of hybridization combinations; it reflects the additive effect of parental genes. Specific combining ability (SCA), on the other hand, refers to the deviation between the actual performance of a hybrid combination formed from two specific parents and the performance expected based on the GFA of the parents; it reflects the non-additive effect of genes, such as dominance and epistasis.

[0113] In this model, a kernel matrix is ​​constructed based on General Combining Ability (GCA) and Specific Combining Ability (SCA) to transform gene information into matrix form. Then, a linear model is used to predict the target trait. The linear model assumes a linear relationship between genes and the target trait; by fitting the kernel matrix to known phenotypic data, linear predictions are obtained. This model can capture some of the linear relationships between genes and the target trait, providing a foundation for subsequent predictions.

[0114] Graph neural networks (Graph Neural Networks) take gene interaction networks as input, with nodes representing genes, node features being gene SNP embedding vectors, and edges representing inter-gene interactions. They predict target traits by learning nonlinear interactions between genes through information propagation and aggregation. Since gene interactions are often nonlinear, Graph Neural Networks can capture these complex nonlinear relationships, compensating for the shortcomings of linear models.

[0115] Residuals refer to the difference between actual values ​​and linearly predicted values. Graph neural network models predict residuals by learning information from gene interaction networks. This means that graph neural networks primarily focus on the parts that linear models cannot explain, namely the impact of nonlinear interactions between genes on the target trait.

[0116] Therefore, in this embodiment, the linear prediction value is added to the residual prediction value to obtain the final target trait prediction value. This combination method fully utilizes the advantages of linear models and graph neural network models, considering both the linear relationship between genes and target traits and capturing nonlinear interactions, thereby improving the accuracy of prediction.

[0117] In summary, the embodiments of this application fully leverage the advantages of both linear models and graph neural network models by combining them. Linear models can quickly capture the linear relationship between genes and target traits, while graph neural network models can delve deeper into the nonlinear interactions between genes. The combination of the two makes the prediction results more accurate and better reflects the actual situation.

[0118] The relationship between different genes and target traits may have both linear and nonlinear components. This combined prediction approach can adapt to different types of gene-trait relationships, improving the model's generalization ability and adaptability, and making it suitable for predicting various complex biological systems and target traits.

[0119] In hybridization breeding, accurate prediction of target traits is crucial for selecting superior hybrid combinations. Combination prediction models can provide more reliable prediction results, helping breeders to more accurately screen hybrids with superior traits, improve breeding efficiency, and reduce breeding costs.

[0120] In some embodiments, the model training and prediction process may specifically include the following steps: I. Data preparation.

[0121] Before model training, the training set data needs to be properly organized and represented. Specifically, each hybrid individual in the training set is transformed into a gene interaction network. Within this gene interaction network: Nodes: Composed of genes, each gene exists as an independent node in the graph.

[0122] Node features: Represented using gene SNP embedding vectors. These vectors are obtained by encoding the SNP information of genes and can reflect the characteristic information of genes.

[0123] Edge: Represents the interaction between genes, which can be determined based on biological knowledge or related research.

[0124] Simultaneously, two important values ​​are associated with each gene interaction network: True phenotypic values: These are the actual phenotypic values ​​exhibited by hybrid individuals and serve as the target reference values ​​for model training.

[0125] Linear model prediction: The value obtained by making preliminary predictions of the traits of hybrid individuals using a pre-constructed linear model.

[0126] II. Training Process.

[0127] (a) Selection of loss function To measure the accuracy of model predictions, mean squared error (MSE) is used as the loss function. MSE is calculated by averaging the squared differences between the model's predicted values ​​and the true phenotypic values. By minimizing this loss function, the model's predicted values ​​can be made as close as possible to the true phenotypic values, thereby improving the model's prediction accuracy.

[0128] (ii) Use of the optimizer The Adam optimizer is used to update the model's parameters. Combining the advantages of Adagrad and RMSProp, the Adam optimizer adaptively adjusts the learning rate for each parameter, exhibiting good convergence speed and stability during training.

[0129] (III) Residual Learning Mechanism During training, the GNN model focuses on the residual between the predicted linear values ​​and the true phenotypic values. The specific method for calculating the final predicted value is as follows: Final predicted value = linear predicted value + (learning weight × GNN residual).

[0130] The learned weights can be automatically learned during model training and are used to adjust the contribution of the GNN residuals to the final prediction results. Through this residual learning approach, the GNN model can focus on capturing the parts that linear models cannot accurately predict, thereby improving overall prediction performance.

[0131] III. Test Prediction.

[0132] First, a linear model is built based on the training set data. This linear model can be trained according to the linear relationship between gene features and phenotypic values. After training, the linear model is used to predict samples in the test set, obtaining a linear prediction value for each test sample.

[0133] Secondly, the test samples are input into the trained GNN model in the form of a gene interaction network. The GNN model will predict the residual between the linear prediction value and the true phenotypic value based on the previously learned gene interaction relationships and feature information, thereby obtaining the residual prediction value for each test sample.

[0134] Finally, the linear prediction value and the GNN residual prediction value are combined according to the above formula, i.e., final prediction value = linear prediction value + (learning weight × GNN residual), to obtain the final prediction result for each test sample. This final prediction result integrates the prediction information of the linear model and the nonlinear information captured by the GNN model, and can more accurately reflect the true phenotype of hybrid individuals.

[0135] It's worth noting that in constructing the gene interaction network, the resulting fixed network structure is presented in the form of `edge_index`, which defines the connections between nodes in the graph. The subsequently generated node feature matrix serves as X, with each row corresponding to the feature of a gene node. Together, they constitute the input data (X, `edge_index`) of the graph neural network. The entire process combines the fixed gene interaction network structure built based on prior knowledge with the node features generated from individual genotype data in the training set, inputting them into the graph neural network to uncover biological associations between genes and make relevant predictions.

[0136] This application aims to improve the accuracy of hybrid prediction models by systematically optimizing hybrid prediction parameters. In some embodiments, a grid search method can be used to optimize key parameters, and detailed steps for constructing the kernel matrix of the linear model GCA (General Combining Ability) / SCA (Specific Combining Ability) are provided. The method's flexibility is also considered to adapt to different application scenarios and data characteristics. The specific details are as follows: I. Grid Search Optimization of Hybrid Prediction Parameters This application employs a grid search method to systematically and comprehensively optimize parameters related to hybrid seed prediction. Accurate estimation of these parameters plays a crucial role in improving the accuracy of hybrid seed prediction models. Specific parameters and the search scope are as follows: (1) Heritability (h²): Represents the proportion of phenotypic variance explained by genetic factors. A higher h² value indicates a stronger influence of genetic factors on the trait. The search range is set to 0.1 to 0.9, with a step size of 0.1.

[0137] (2) General combining ability weight (w_gca): Used to quantify the additive effect of the parental lines, reflecting the average performance of the parents in different hybridization combinations. The search values ​​were set to 0.5, 0.7, and 0.9.

[0138] (3) Specific combining ability weight (w_sca): Used to capture non-additive effects, representing the deviation between the actual performance of the hybrid and the expected performance based on the parental GCA (general combining ability). The search values ​​are set to 0.1, 0.3, and 0.5.

[0139] It is worth noting that this grid search method can be replaced with other genetic force calculation methods to adapt to different application scenarios and data characteristics, so as to further optimize model performance.

[0140] II. Steps for constructing the GCA / SCA kernel matrix of a linear model The linear model part uses a kernel matrix-based method to predict hybrid phenotypes. The specific steps are as follows: (1) Data preparation: Data collection: Collect hybrid phenotypic data and parental SNP data.

[0141] Data preprocessing: The SNP data is pattern-filled, and its limit values ​​are set between [-1, 2]. This preprocessing helps improve the quality and consistency of the data.

[0142] Parental SNP extraction: Extract parental SNP data from the collected SNP data for subsequent construction of the kernel matrix.

[0143] (2) Kernel matrix construction: GCA nuclear matrix (K_gca): The GCA nuclear matrix is ​​obtained by calculating the average value of the parental nuclear matrix. This matrix mainly reflects the additive effect of the parental lines.

[0144] SCA kernel matrix (K_sca): used to capture the non-additive effects of specific hybridization combinations, reflecting the special performance of hybrids in combinations.

[0145] Combined kernel matrix: The GCA kernel matrix and SCA kernel matrix are combined according to their respective weights to obtain the combined kernel matrix K. The calculation formula is as follows: K=w_gca×K_gca+w_sca×K_sca.

[0146] The K matrix, a weighted combination of the GCA and SCA kernels, directly participates in the calculation of the variance-covariance matrix V, thus affecting model parameter estimation and prediction results. It is one of the core inputs of the GBLUP model. The K matrix model can utilize both GCA and SCA information simultaneously.

[0147] w_gca and w_sca are the weighting parameters for the GCA and SCA kernels, respectively, controlling the relative contributions of the two genetic effects in the combined kernel. By adjusting these two parameters, different GCA / SCA information fusion ratios can be explored, thereby optimizing prediction performance.

[0148] (3) Model training and evaluation: Model training: The SNP effect (β) is estimated using the training data by solving: K_train × β = y_train. Here, K_train is the combined kernel matrix of the training set, and y_train is the phenotypic data of the training set.

[0149] Prediction: The SNP effect (β) obtained from training is used to predict the test set. The calculation formula is: y_pred=K_test×β, where K_test is the combined kernel matrix of the test set and y_pred is the prediction result.

[0150] Evaluation: The predicted result y_pred is compared with the actual test set phenotypic data y_test, and the predictive performance of the model is evaluated using indicators such as PCC (Pearson correlation coefficient) and R² (coefficient of determination).

[0151] It should be noted that the linear prediction part of this application is highly flexible and can be flexibly replaced with other baseline models according to different prediction needs and data characteristics, so as to further improve the adaptability and prediction accuracy of the model.

[0152] It should be added that the GBLUP model involves several key concepts and their related meanings: K_train is the kernel matrix for training set, representing the genetic similarity between various hybridization combinations (or individuals) in the training set. This kernel matrix is ​​a similarity matrix calculated based on genotype data, constructed from paternal SNP information, and can be a GCA kernel, SCA kernel, or a weighted combination thereof.

[0153] β (beta) is the regression coefficient vector, which stands for "Regression coefficients (model parameters)," representing the genetic effect weights of each training sample estimated by the model. In GBLUP, β reflects the contribution of each training sample to the prediction, specifically obtained by solving the equation (V\beta=y).

[0154] y_train refers to the observed phenotypic values ​​for the training set, which are the actual measured phenotypic values ​​obtained in the training data, such as yield and plant height. It is the target variable for model fitting and is used to calculate prediction error and optimize parameters.

[0155] y_pred refers to the predicted phenotypic values, which are the phenotypic predictions calculated by the model based on genotype data and estimated genetic effects. These predictions can be compared with actual measurements to evaluate the model's predictive performance.

[0156] K_test is the kernel matrix between the test set and the training set. It represents the genetic similarity between each sample in the test set and each sample in the training set. It is used in the prediction phase to apply the genetic effect β estimated in the training set to the samples in the test set.

[0157] This application aims to improve the accuracy and effectiveness of hybrid phenotypic prediction by employing a residual learning strategy, cleverly combining the advantages of linear models and graph neural networks to construct a more accurate prediction model. The specific details are as follows: First, preliminary predictions are made using a linear model. By calculating the kernel matrix of the linear model GCA (general combination force) / SCA (specific combination force), preliminary linear prediction values ​​are obtained. Linear models, which construct kernel matrices based on the additive and non-additive effects of genes, can capture the linear relationships between genes quite well.

[0158] Next, Graph Neural Networks (GNNs) are introduced for residual learning. Since linear models may not fully capture the complex nonlinear interactions between genes, GNNs focus on learning the residuals between linear model predictions and the actual phenotype. Based on gene interaction networks, it mines higher-order associations and nonlinear features between genes through multi-layer message passing and aggregation mechanisms, thereby more accurately characterizing the influence of genes on phenotypes.

[0159] Finally, the linear predictions and the residual predictions from the graph neural network are weighted and combined to obtain the final prediction. The specific calculation formula is as follows: .

[0160] Where α is the learned weight coefficient. This represents the residual value predicted by the Graph Neural Network (GNN). This weighted combination method fully integrates the advantages of linear models and graph neural networks, retaining the stability and interpretability of linear models while utilizing the powerful ability of graph neural networks to capture nonlinear features, thereby improving the model's predictive performance.

[0161] In summary, this application, through a residual learning strategy, organically combines linear models and graph neural networks, providing a more accurate and effective method for predicting hybrid phenotypes.

[0162] In some embodiments, the following evaluation metrics can be used to measure model performance: (1) Coefficient of determination (R²): Used to measure how well a model fits the data and its ability to predict future results. The closer the R² value is to 1, the better the model fits the data. Its calculation formula is: R² = 1 - (SSE / SST).

[0163] Where SSE is the sum of squared errors and SST is the total sum of squares. In simple linear regression, R² equals the square of the Pearson correlation coefficient (PCC).

[0164] (2) Pearson correlation coefficient (PCC): It is used to measure the linear correlation between two sets of data, and its value ranges from -1 to 1. PCC is the ratio of the product of the covariance and the standard deviation of two variables. This calculation method actually normalizes the covariance, so that the value of PCC has a uniform measurement standard.

[0165] In this application, PCC can be used to assess the correlation between predicted values ​​and actual phenotypic values. If the PCC value is close to 1 or -1, it indicates a strong linear relationship between the two; if it is close to 0, it indicates a weak linear relationship.

[0166] Through the above implementation scheme, this application realizes a genome-wide selection method based on gene-level embedding graph neural networks. The significant advantage of this method lies in its ability to effectively integrate genomic data with prior biological knowledge, combining gene-level information with known biological laws, thereby greatly improving the prediction accuracy of complex traits. Simultaneously, this method possesses an interpretable molecular mechanism, enabling in-depth analysis of the intrinsic relationship between genes and traits, providing a clear theoretical basis for the breeding process.

[0167] Furthermore, this application can also adopt a modular design, with each functional component / module possessing high flexibility and scalability. This means that when facing different breeding needs or data characteristics, various parts of the model can be easily adjusted and optimized, further enhancing the practicality and adaptability of the method.

[0168] In some embodiments, in a gene interaction network, the breeding value at the gene level and the cumulative heterozygous dominance effect can also be used as node features.

[0169] Breeding value is a measure of an individual gene's contribution to the genetic makeup of its offspring. Gene-level breeding value measures the potential impact of a single gene on the phenotypic expression of its offspring. In gene interaction networks, each gene has a unique breeding value, reflecting its importance in the transmission of genetic traits (such as crop yield and quality, animal growth rate, and meat quality).

[0170] In this embodiment, using gene-level breeding values ​​as node features allows graph neural networks to more accurately assess the potential contribution of each gene to the target trait during information propagation and aggregation.

[0171] For example, in crop breeding, if a gene has a high breeding value, it indicates that it plays an important role in increasing crop yield. Graph neural networks will pay more attention to the information of this gene during the learning process, thereby predicting the target trait more accurately.

[0172] Heterozygous dominance refers to the difference in gene expression between the heterozygous and homozygous states; that is, the heterozygous phenotype is superior to or different from that of two homozygous individuals. Cumulative heterozygous dominance, on the other hand, considers the combined effect of heterozygous states at multiple gene loci.

[0173] In hybrids, heterozygous dominance often has a significant impact on the expression of the target trait; for example, heterosis is a manifestation of heterozygous dominance.

[0174] In this embodiment, the cumulative heterozygous dominance effect can be used as a node feature, enabling the graph neural network to capture heterozygous interaction information between genes. This is because the heterozygous states between genes influence each other and work together to achieve the target trait.

[0175] Gene-level breeding values ​​and cumulative heterozygous dominance effects provide node features with more information about gene function and genetic effects. Compared to using only gene SNP embedding vectors, adding these two features allows nodes to carry more comprehensive gene-related information, helping graph neural networks to understand the biological significance of genes and their interactions more deeply.

[0176] Gene-level breeding values ​​and cumulative heterozygous dominance effects can reflect the complex nonlinear relationships between genes and target traits. Graph neural networks can utilize these features to further explore the nonlinear interactions between genes, thereby better capturing complex patterns in gene regulatory networks and improving the prediction accuracy of target traits.

[0177] For example, this application calculates the breeding value and cumulative heterozygous dominance effect at the gene level as the gene characteristics of the node, and the specific process is as follows: (1) Gene Breeding Value (GeneBV) Calculation: Represents the additive genetic effect inherited from the parents. For each gene, GeneBV is calculated as the average of the parental breeding values ​​for that gene: GeneBV=0.5×(gene_x1@gene_snp_effects+gene_x2@gene_snp_effects).

[0178] Here, @ is the matrix multiplication operator in Python (especially NumPy) syntax, gene_x1 and gene_x2 are genotype vectors (or matrices) representing the two parents, and gene_snp_effects is the effect value vector (or matrix) for each site (SNP).

[0179] gene_x1@gene_snp_effects means multiplying the genotype vector of parent 1 with the SNP effect vector and summing the results to obtain the additive breeding value of that parent. Similarly, gene_x2@gene_snp_effects obtains the additive breeding value of parent 2.

[0180] 0.5×(...) represents taking the average of the two values ​​to obtain the GeneBV (gene-level breeding value) of the hybrid combination.

[0181] (2) Calculation of cumulative heterozygous dominance effect (GCHDE): This represents the sum of the dominance effects of all heterozygous SNP sites within a gene. GCHDE=np.sum(gene_snp_effects[het_mask]).

[0182] Where het_mask is the mask for heterozygous SNPs, gene_snp_effects is the corresponding SNP effect value, and np.sum represents the summation operation.

[0183] In this embodiment, by calculating GeneBV and GCHDE, a molecular mechanism explanation can be provided for breeding decisions, and possible application scenarios can be explored. By using "breeding value at the gene level" and "cumulative heterozygous dominant effect" as node features, information on additive and dominant genetic effects can be directly introduced.

[0184] It's important to clarify that the "using SNP embeddings as node features" approach mentioned above involves first aggregating site-level SNP data to the gene level, for example, encoding and integrating multiple SNPs of the same gene. Then, a vector representation (embedding) is generated for each gene. Here, the node features are vectors obtained by mapping and embedding genotype data. When using graph neural networks (or similar models) for prediction, these gene embedding vectors are directly input into the model as node features, without introducing additional derived indicators such as breeding values ​​or dominance effects.

[0185] This differs from the "breeding value + heterozygous dominance effect" approach in the above embodiments, representing two different feature construction paths. Where there is no conflict, those skilled in the art can choose one or more of the above-mentioned node feature construction methods according to actual needs.

[0186] In some embodiments, an end-to-end modular design may be adopted.

[0187] End-to-end means that the entire process from inputting raw data to outputting the desired result can be coherent and consistent, without requiring extensive manual intervention or data transformation in the intermediate stages. Modular design, on the other hand, allows the entire system to be broken down into multiple relatively independent modules with specific functions. Each module is responsible for completing a specific task, and each module can be developed, tested, and maintained independently.

[0188] Each module focuses on solving a specific sub-problem, working together to achieve the overall system goal. Users can assemble different functional modules based on different application scenarios and data characteristics to gain a more comprehensive understanding of the biological processes of organisms and improve the accuracy of phenotypic predictions.

[0189] Optionally, a feedback mechanism can be incorporated into the modular design to enable collaboration and optimization between modules. For example, during model training, if the output of a certain module is found to be unsatisfactory, feedback information can be passed to the preceding modules for adjustment and optimization, thereby improving the performance of the entire system.

[0190] Optionally, parallel processing and distributed computing can be incorporated into the modular design. For large-scale genotypic data and complex model training tasks, parallel processing and distributed computing techniques can be used to improve overall processing speed and efficiency.

[0191] For example, different modules can be distributed across different computing nodes for parallel processing, or a distributed deep learning framework can be used to train the model in a distributed manner.

[0192] This embodiment also involves a message passing mechanism, which is performed in an orderly manner according to a specific process in each layer of the graph neural network. Specifically, each node first collects feature information from its neighboring nodes, and this feature information can be regarded as "messages". This collection process enables the node to obtain relevant information about its surrounding environment.

[0193] Subsequently, the node performs an aggregation operation on the feature "messages" collected from its neighbors. The purpose of aggregation is to integrate the different feature information of multiple neighboring nodes to form a comprehensive representation that reflects the overall characteristics of the neighboring nodes.

[0194] Finally, the nodes use the aggregation results to update their own representations. In this way, the nodes continuously adjust and improve their feature representations during message passing at each layer, enabling the entire graph neural network to learn the complex features and relationships in graph-structured data layer by layer.

[0195] The embodiments of this application propose a genome-wide selection method based on gene-level embedding graph neural networks, which solves the problems existing in related technologies through the following key technological innovations: (1) Gene-level SNP embedding representation: This application innovatively elevates SNP markers from the site level to the gene functional unit level for embedding representation. Specifically, the SNP sequences within a gene region are input into a Transformer encoder to obtain a high-dimensional embedding vector for each gene. This operation can effectively capture the complex association patterns between SNPs within a gene. This gene-level feature representation method not only reduces the feature dimensionality and the complexity of data processing, but also retains richer biological information, providing high-quality node features for subsequent graph neural network learning.

[0196] (2) Graph structure construction driven by prior biological knowledge: This application systematically integrates existing prior knowledge of gene networks, including gene co-expression networks, transcriptional regulation networks, and miRNA regulation networks. Based on this, a biologically meaningful gene interaction network is constructed. This graph structure design based on prior knowledge enables the model to fully utilize accumulated biological research results, thereby more accurately simulating complex regulatory relationships between genes.

[0197] (3) Graph neural network captures nonlinear gene interaction: This application adopts a graph neural network architecture and performs information aggregation and updating on the gene interaction network through message passing mechanism. This method can effectively model the nonlinear interaction between genes and breaks through the limitations of traditional linear models and sequence models. When dealing with the prediction of complex traits controlled by multiple genes, traditional models often have difficulty accurately capturing the complex regulatory patterns between genes, while the method of this application can adapt well to this situation and show stronger adaptability and accuracy.

[0198] (4) End-to-end modular design: This application adopts an end-to-end modular architecture, in which each functional module (such as gene embedding module, graph neural network module, prediction module, etc.) can be flexibly replaced and optimized. This design gives the system a high degree of scalability, which can be customized according to the characteristics of different crops and traits, adapt to diverse breeding needs, and improve the versatility and practicality of the method.

[0199] (5) Interpretability Analysis Framework: This application designs an interpretability analysis framework based on graph attention mechanism, which can identify gene nodes and inter-gene interactions that make important contributions to the prediction results. This interpretability not only improves the transparency of the model and enables a clear understanding of the model's decision-making process, but also provides valuable biological insights for molecular design breeding, helping breeders understand the genetic mechanisms of complex traits.

[0200] Through the above-mentioned technological innovations, this application effectively solves the problems of insufficient utilization of structural information, limited integration of biological prior knowledge, insufficient granularity of feature representation, and poor model interpretability in related technologies, and provides a new method for whole-genome selection that has both high predictive accuracy and good biological interpretability.

[0201] Please see Figure 2 , Figure 2 This is a schematic diagram of the actual process of the whole-genome selection method of graph neural network provided in the embodiments of this application.

[0202] In this embodiment, an end-to-end modular design can be adopted, including multiple core functional modules. These core modules form a complete data processing and prediction pipeline, realizing the entire process from raw genotype data to phenotype prediction. The specific scheme is as follows: 1. Input file: Input genotype and phenotype datasets.

[0203] 2. Grid search: Perform a grid search for general and special cohesion force weights to optimize model parameters.

[0204] 3. Gene-level SNP embedding representation: Gene-level SNPs are embedded and represented.

[0205] 4. Construct a linear model: Use the GCA / SCA kernel matrix to construct a linear model.

[0206] 5. Calculation of Genetic Breeding Value / GCHDE and Linear Prediction Value Linear predictions are obtained through linear model calculations, which can incorporate the gene breeding value / GCHDE calculation to provide a molecular mechanism explanation for breeding decisions.

[0207] 6. Graph-based GNN residual learning: Residual learning is performed using graph neural networks (GNN).

[0208] 7. Final predicted value The final predicted value is obtained by combining the outputs of the linear model and the GNN model. .

[0209] 8. Evaluation metrics: R² & PCC (Pearson correlation coefficient) are used as evaluation metrics to measure the predictive performance of the model.

[0210] In summary, the genome-wide selection method based on gene-level embedding graph neural networks proposed in this application has the following technical advantages: I. Enhanced biological significance (1) Gene-level analysis: Traditional genome-wide selection methods are mainly based on SNP markers and lack biological function interpretation. This application maps SNPs to the gene level and combines them with gene interaction networks to give the prediction results clear biological significance, providing a functional explanation for breeding decisions.

[0211] (2) Functional knowledge integration: It successfully integrates multi-source biological prior knowledge such as gene expression, transcription factor regulation and miRNA regulation, bridging the gap between genotype-phenotype association and molecular function research, and providing a new perspective for understanding the molecular mechanisms of complex traits.

[0212] (3) From marker association to functional analysis: The traditional marker effect estimation is transformed into gene functional network analysis, which enables breeding selection to shift from "knowing what" to "knowing why", providing functional guidance for precision breeding.

[0213] II. Improvements to the Black Box Problem in Deep Learning (1) Enhanced interpretability: Through gene-level embedding and the attention mechanism of graph neural networks, this application alleviates the "black box" problem of deep learning to a certain extent, making the model prediction results traceable to specific genes and their interaction networks.

[0214] (2) Biological hypothesis verification: The gene importance and interaction patterns learned by the model can be compared and verified with known biological knowledge, providing clues for further functional research and forming a virtuous cycle of computational prediction and biological verification.

[0215] (3) Transparent decision-making basis: Compared with traditional deep learning methods, the technical solution of this application can provide more transparent decision-making basis, enabling breeders to understand which genes and their interactions have an important impact on the target trait, thereby enhancing the credibility and practicality of the model.

[0216] III. The bridging role of research and application (1) Connecting basic research and application: This application has successfully built a bridge between whole-genome selection and gene function research, enabling the results of the two fields to promote each other and accelerating the transformation process from gene function discovery to breeding application.

[0217] (2) Two-way knowledge flow: On the one hand, functional genomics knowledge is integrated into the prediction model through gene interaction networks; on the other hand, the gene importance and interaction patterns learned by the model provide new hypotheses for functional genomics research, forming a two-way flow of knowledge.

[0218] (3) Interdisciplinary value: This application reflects the deep integration of bioinformatics, artificial intelligence and molecular breeding, providing a new paradigm for interdisciplinary cooperation and promoting the innovative application of technologies and methods in various fields.

[0219] IV. Flexibility and Scalability of the Technical Framework (1) Advantages of modular design: This application adopts a highly modular design, and each functional component (such as SNP embedding, gene interaction network construction, GNN architecture, etc.) can be upgraded or replaced independently to meet the needs of different application scenarios.

[0220] (2) Prior knowledge expansion capability: The gene-level design enables the system to easily integrate new prior knowledge such as gene function annotation, expression profile, and protein interaction. As biological research deepens, the biological basis of the model can be continuously enriched and improved.

[0221] (3) Flexibility of technical plug-ins: Each module can be regarded as an independent technical plug-in. For example, gene-level SNP embedding can be replaced by LSTM or other dimensions of Transformer, GNN can be replaced by other network architectures, and parameter optimization can use different search strategies, providing flexible choices for applications of different scales and needs.

[0222] V. Application Prospects (1) Applicability to multiple species: The framework of this application is not limited to specific species and can be widely applied to various breeding fields such as crops, animal husbandry and aquaculture, with broad application prospects.

[0223] (2) Foundation for multi-omics data integration: Gene-level design provides a natural interface for future integration of transcriptomics, proteomics, metabolomics and other multi-omics data, which can support more comprehensive molecular breeding research.

[0224] (3) Assisted gene editing design: By identifying key genes and their interaction networks, this application can provide support for target selection and effect prediction for gene editing breeding, and promote the development of precision breeding technology.

[0225] VI. Practical Application Value (1) Breeding decision support: Although the computational complexity is high, by providing gene-level interpretation and prediction, this application can help breeders design hybrid combinations and selection strategies more effectively and improve breeding efficiency.

[0226] (2) Improved prediction performance: On the test dataset, this application improves the prediction accuracy by at least 5% compared to the traditional linear model; this improvement is of practical significance for the prediction of complex states.

[0227] (3) Knowledge accumulation effect: With the continuous enrichment of biological prior knowledge and the continuous optimization of the algorithm, the prediction performance of this application is expected to be further improved, reflecting the long-term value of knowledge-driven methods.

[0228] In summary, the genome-wide selection method based on gene-level embedding graph neural networks proposed in this application has significant advantages in enhancing the biological significance of prediction results, improving the black-box problem of deep learning, and building a bridge between research and application by innovatively mapping SNP markers to the gene level and integrating prior biological knowledge. Its value in biological interpretation, knowledge integration, and technological scalability provides a new technical path and approach for the development of precision breeding.

[0229] It should also be noted that the genome-wide selection method based on gene-level embedding graph neural networks proposed in this application has a unique technical approach. The following discusses several possible alternative solutions and their comparison with this application: First, traditional machine learning methods Traditional machine learning methods such as random forests and gradient boosting trees can be used for genome-wide selection prediction. These methods are simple to implement, computationally efficient, and may achieve good predictive performance on certain datasets. However, these methods struggle to effectively model complex interactions between SNPs, especially long-distance dependencies and higher-order interactions. More importantly, they cannot integrate prior biological knowledge such as gene functional networks, resulting in predictions that lack biological interpretability and fail to provide functional guidance for breeding decisions.

[0230] Second, end-to-end deep learning methods End-to-end deep neural networks (such as CNNs, RNNs, or Transformers) can directly model SNP data. These methods have powerful expressive capabilities and may achieve high prediction accuracy in some cases. However, these methods are usually "black box" models, making it difficult to provide biologically meaningful explanations; at the same time, they lack the modular design of the scheme in this application, making it difficult to flexibly integrate constantly updated biological knowledge and unable to perform component-level optimization and adjustment for different application scenarios.

[0231] Third, pure genomic statistical methods GBLUP or Bayes-based methods, which utilize mixed linear models, are classic approaches to genome-wide selection. These methods have a solid statistical foundation and relatively high computational efficiency. However, they are typically based on simplified additive genetic models, making it difficult to capture complex nonlinear effects and gene interactions. More importantly, these methods struggle to integrate prior knowledge such as gene functional networks, cannot explain predictions at the molecular mechanism level, and lack the modular design advantages of the approach proposed in this application, making it difficult to flexibly evolve with updates in biological knowledge.

[0232] Fourth, the single network method Using graph neural networks alone without combining them with linear models may be another technical approach. This method can model gene interactions, but it may not be able to effectively capture the additive effects of genes, especially when training data is limited. In addition, single-network methods lack the modular design of the approach in this application, making it difficult to flexibly adjust the components according to different application scenarios, and also unable to balance the contributions of additive and interaction effects through model fusion strategies as in the approach in this application.

[0233] Compared to the aforementioned alternatives, the technical approach proposed in this application has the following irreplaceable advantages: (1) Modular design advantage: The solution of this application adopts a highly modular design. Each functional component (such as SNP embedding, gene interaction network construction, GNN architecture, etc.) can be upgraded or replaced independently to adapt to the needs of different application scenarios. This is a key feature that other alternative solutions do not have.

[0234] (2) Biological knowledge integration capability: Through gene-level design, this scheme can naturally integrate biological prior knowledge such as gene functional networks, so that the prediction results have clear biological significance, while most alternative schemes are unable to effectively utilize this valuable prior information.

[0235] (3) Balance and practicality: The proposed solution achieves a good balance between prediction accuracy, computational efficiency, biological interpretability and technical scalability, making it more suitable for the needs of actual breeding applications. Other alternative solutions often only have advantages in one aspect and are difficult to meet multiple needs.

[0236] In summary, although there are multiple technical approaches that can be applied to whole-genome selection, the graph neural network algorithm based on gene-level embedding proposed in this application has significant advantages in terms of comprehensive performance and practical value through its unique modular design and ability to integrate biological knowledge, providing a technical solution with great potential for precision breeding.

[0237] This application also provides a whole-genome selection device for graph neural networks. The whole-genome selection device for graph neural networks provided in this application is described below. The whole-genome selection device for graph neural networks described below can be referred to in correspondence with the whole-genome selection method for graph neural networks described above.

[0238] Please see Figure 3 , Figure 3 This is a schematic diagram of the genome-wide selection device for a graph neural network provided in an embodiment of this application. In this embodiment, the genome-wide selection device for a graph neural network may include a gene embedding vector module 310, a gene interaction network module 320, and a prediction value module 330.

[0239] Gene embedding vector module 310 is used to map single nucleotide polymorphisms (SNPs) to the gene level and convert them into gene embedding vectors; Gene interaction network module 320 is used to construct a gene interaction network based on preset multi-source biological prior knowledge; The prediction module 330 is used to input the target trait into the graph neural network model and obtain the predicted value of the target trait output by the graph neural network model; wherein, the graph neural network model is determined based on gene embedding vectors and gene interaction networks.

[0240] In some embodiments, the gene embedding vector module 310 is specifically used for: An extended gene region definition is used for mapping, and SNPs are assigned to corresponding genes. For each gene, all SNPs mapped to that gene are sorted according to their physical location on the genome to form the SNP sequence of that gene. The SNP sequence of each gene is input into a preset sequence coding model to obtain the gene embedding vector corresponding to each gene output by the preset sequence coding model. The gene embedding vector serves as the initial node feature of the graph neural network model.

[0241] In some embodiments, multi-source biological prior knowledge includes gene co-expression networks, transcriptional regulatory networks, and miRNA regulatory networks; the gene interaction network module 320 can specifically be used for: The gene co-expression network, transcriptional regulation network, and miRNA regulation network were weighted and fused to obtain a preliminary gene interaction network; the preliminary gene interaction network was then preprocessed to obtain the final gene interaction network.

[0242] In some embodiments, when a hybrid individual is represented as a gene interaction network, the nodes are genes, the node features are gene SNP embedding vectors, and the edges are the interaction relationships between genes; wherein, each gene interaction network is associated with a true phenotypic value and a linear model prediction value; the graph neural network model is used to propagate and aggregate information in the gene interaction network to learn the nonlinear interactions between genes and output the prediction value of the target trait.

[0243] In some embodiments, the predicted value of the target trait is obtained by: using a first model to predict the target trait to obtain a linear predicted value; using a second model to predict the target trait to obtain a residual predicted value; and combining the linear predicted value and the residual predicted value to obtain a predicted value; wherein the first model is a kernel matrix linear model based on general combining ability (GCA) and special combining ability (SCA); and the second model is a graph neural network model.

[0244] In some embodiments, the gene interaction network module 320 can also be used for: In the gene interaction network, the breeding value at the gene level and the cumulative heterozygous dominance effect are used as node features.

[0245] On the other hand, this application also provides an electronic device, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the physical structure of the electronic device provided in the embodiments of this application, such as... Figure 4 As shown, the electronic device may include memory 420, processor 410, and a computer program stored in memory 420 and executable on processor 410. When processor 410 executes the program, it can implement a genome-wide selection method for graph neural networks, which may include: Single nucleotide polymorphisms (SNPs) are mapped to the gene level and converted into gene embedding vectors. Based on pre-defined multi-source biological prior knowledge, a gene interaction network is constructed. The target trait is input into a graph neural network model to obtain the predicted value of the target trait output by the graph neural network model. The graph neural network model is determined based on the gene embedding vectors and the gene interaction network.

[0246] Optionally, the electronic device may further include a communication bus 430 and a communication interface 440, wherein the processor 410, the communication interface 440, and the memory 420 communicate with each other via the communication bus 430. The processor 410 can call the computer program in the memory 420 to execute the genome-wide selection method for graph neural networks provided by the above methods.

[0247] Furthermore, the logical instructions in the aforementioned memory 420 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0248] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the whole genome selection method of the graph neural network provided by the above methods. The steps and principles of the method have been described in detail in the above methods and will not be repeated here.

[0249] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the genome-wide selection method for graph neural networks provided by the above methods. The steps and principles of this method have been described in detail in the above methods and will not be repeated here.

[0250] Non-transitory computer-readable storage media can be any available medium or data storage device that can be accessed by a processor, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0251] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0252] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0253] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application.

Claims

1. A genome-wide selection method using a graph neural network, characterized in that, include: Single nucleotide polymorphisms (SNPs) are mapped to the gene level and converted into gene embedding vectors; Based on pre-defined multi-source biological prior knowledge, a gene interaction network is constructed. The target trait is input into a graph neural network model to obtain the predicted value of the target trait output by the graph neural network model; wherein the graph neural network model is determined based on the gene embedding vector and the gene interaction network.

2. The genome-wide selection method for graph neural networks according to claim 1, characterized in that, The process of mapping single nucleotide polymorphisms (SNPs) to the gene level and converting them into gene embedding vectors includes: The mapping is performed using an extended gene region definition, assigning SNPs to corresponding genes; For each gene, all SNPs mapped to that gene are sorted according to their physical location on the genome to form the SNP sequence of that gene; The SNP sequence of each gene is input into a preset sequence coding model to obtain the gene embedding vector corresponding to each gene output by the preset sequence coding model, wherein the gene embedding vector is used as the initial node feature of the graph neural network model.

3. The genome-wide selection method for graph neural networks according to claim 1, characterized in that, The multi-source biological prior knowledge includes gene co-expression networks, transcriptional regulatory networks, and miRNA regulatory networks; The construction of a gene interaction network based on pre-defined multi-source biological prior knowledge includes: The gene co-expression network, the transcriptional regulation network, and the miRNA regulation network are weighted and fused to obtain a preliminary gene interaction network. The preliminary gene interaction network is preprocessed to obtain the gene interaction network.

4. The genome-wide selection method for graph neural networks according to claim 1, characterized in that, In the case where a hybrid individual is represented as the gene interaction network, the nodes are genes, the node features are gene SNP embedding vectors, and the edges are the interaction relationships between genes; wherein, each gene interaction network is associated with a true phenotypic value and a linear model prediction value. The graph neural network model is used to propagate and aggregate information in the gene interaction network to learn the nonlinear interactions between genes and output predicted values ​​of the target trait.

5. The genome-wide selection method for graph neural networks according to claim 1, characterized in that, The predicted value of the target trait is obtained in the following way: The target trait is predicted using the first model to obtain a linear prediction value; The target trait is predicted using the second model to obtain the residual prediction value; The predicted value is obtained by combining the linear predicted value and the residual predicted value. The first model is a kernel matrix linear model based on general cohesion force (GCA) and special cohesion force (SCA); the second model is a graph neural network model.

6. The genome-wide selection method for graph neural networks according to claim 1, characterized in that, Also includes: In the gene interaction network, the breeding value at the gene level and the cumulative heterozygous dominance effect are used as node features.

7. A genome-wide selection device for a graph neural network, characterized in that, include: The gene embedding vector module is used to map single nucleotide polymorphisms (SNPs) to the gene level and convert them into gene embedding vectors. The gene interaction network module is used to construct gene interaction networks based on pre-defined multi-source biological prior knowledge. The prediction module is used to input the target trait into the graph neural network model to obtain the predicted value of the target trait output by the graph neural network model; wherein the graph neural network model is determined based on the gene embedding vector and the gene interaction network.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the whole-genome selection method for the graph neural network as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the genome-wide selection method for graph neural networks as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the genome-wide selection method for graph neural networks as described in any one of claims 1 to 6.

Images