A gene encoding breeding prediction method and device based on graph clustering

Abstract

The application provides a gene coding breeding prediction method and device based on graph clustering, constructs a gene graph based on the strength of the correlation between genes, solves the gene graph by clustering to obtain the number of common regulatory gene groups and the gene group cluster number information of each gene, fuses the gene allele information and the gene group cluster number information to obtain the gene cluster coding of the sample, constructs a deep convolutional neural network based on the gene cluster coding information and the biological phenotype information to be predicted to optimize the gene breeding prediction performance. The application fully utilizes the gene interaction relationship network contained in the gene graph, can effectively extract the regulatory gene features for controlling the biological phenotype output, solves the problem of insufficient gene interaction relationship coding between the input coding layers of the classical model and the gene graph, guarantees the accuracy of the gene breeding prediction of the biological phenotype, and further improves the speed, efficiency and quality of the gene breeding, especially the yield.

Description

Technical Field

[0001] This invention mainly relates to the field of gene breeding prediction in precision molecular breeding of crops, and mainly relates to a gene coding breeding prediction method and device based on graph clustering. Background Technology

[0002] With the development of gene sequencing technology, laboratory technicians can obtain large-scale, multi-sample gene data with data mining application value through wet experimental processes such as sample collection, library preparation, and sequencing, based on sequencing, PCR (polymerase chain reaction), gene chips, and optical mapping. However, after whole-genome sequencing, the accuracy of genome prediction models is very low. Taking soybean as an example, soybeans contain approximately 60,000 genes, of which 80 million mutations have been found in 40,000 gene pairs. Genotype-predicted phenotypes can only be qualitatively described, not quantitatively analyzed. This greatly limits the quantity, speed, and quality of crop breeding, especially the improvement of yield.

[0003] To improve the accuracy of molecular breeding, current gene prediction methods for crop phenotypes mainly include traditional statistical analysis methods such as Bayesian methods, linear regression, and ridge regression. However, deep learning methods, which have achieved great success in speech, image, and natural language processing, have not performed well in crop breeding due to the limited sample size. Furthermore, gene data has a high dimensionality, making it difficult for traditional statistical analysis methods to quickly extract effective features from such high-dimensional gene feature data using feature selection methods. Therefore, existing popular methods are insufficient to address the high-dimensional, small-sample problem in crop molecular breeding.

[0004] To address the high-dimensional, small-sample problem in crop molecular breeding, innovative gene breeding prediction methods are needed to simultaneously solve the problems of feature selection and extraction of high-dimensional features and insufficient feature encoding of complex model sample gene maps. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a gene-encoding breeding prediction method and device based on graph clustering. This method utilizes the gene-interaction network contained in gene maps, extracts information on common regulatory genome clusters through graph clustering, and proposes a novel gene clustering encoding method that integrates gene allelic information and gene map clustering information. Furthermore, it leverages the weight sharing of deep convolutional neural networks to effectively extract regulatory gene features that control biological phenotypic output, thus solving the problem of insufficient encoding of gene-interaction relationships between gene maps in the input encoding layer of classical models and ensuring the accuracy of gene-encoding breeding prediction of biological phenotypes.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention discloses a gene-encoding breeding prediction method based on graph clustering, comprising the following steps:

[0008] Obtain genotype data and gene location information of the offspring to be predicted;

[0009] An undirected graph is constructed as a gene map based on the strength of correlations between genes in genotype data.

[0010] Clustering is performed on the gene map to obtain the number of genes that commonly regulate the genome and the genome cluster number for each gene;

[0011] The allele information and genome clustering number information corresponding to each gene in the genotype data are combined and concatenated to obtain the gene clustering code of the sample;

[0012] Gene clustering codes and gene location information are input into a gene-coded breeding prediction model to obtain the biophenotype information of the offspring to be predicted; based on the predicted biophenotype information of the offspring, a set of high-quality seeds is selected.

[0013] The gene-encoded breeding prediction model is trained based on a collected dataset, and each sample in the dataset includes the sample's gene clustering code, gene location information, and biophenotype information.

[0014] Preferably, the biological phenotypic information includes measurable information such as the quantity, quality, percentage, and classification of the target phenotypic, and the allele information to be encoded includes SNP alleles, such as homozygous 0 / 0, 1 / 1, and heterozygous 0 / 1.

[0015] As a preferred method, clustering is performed on the gene map to obtain the number of commonly regulated genomes and the genome cluster number information for each gene, as follows:

[0016] The number of co-regulatory genomes, i.e. the number of gene clusters, is estimated based on the spatial distribution characteristics of gene maps;

[0017] Based on the estimated number of gene clusters, calculate the intra-cluster distance and inter-cluster distance for each gene to determine the cluster to which the gene belongs;

[0018] After clustering is completed, each gene cluster is assigned a unique clustering number, which serves as the genome clustering number for each gene in the corresponding gene cluster.

[0019] Preferably, the strength of the correlation between genes is generally obtained by calculating the similarity of SNP site strings of each pair of genes in multiple samples. Commonly used methods include Pearson correlation coefficient, Spearman correlation coefficient, Euclidean distance, cosine similarity, Manhattan distance, Hamming distance, edit distance, etc.; the weight of adjacent edges is generally represented by the strength of the correlation between genes or its normalized value.

[0020] Preferred gene clustering methods include spatial clustering (K-means, etc.), density clustering (DBSCAN, etc.), hierarchical clustering (bottom-up and top-down methods), and spectral clustering.

[0021] Preferred methods for estimating the number of gene clusters include statistical methods, random methods, exhaustive methods, and iterative methods. Among them, iterative methods mainly refer to the methods for determining the number of clusters by bottom-up or top-down iterative clustering in hierarchical clustering.

[0022] As a preferred method, the spectral clustering method mainly uses the Laplacian matrix and other computational graph connected components for clustering; the calculation methods for intra-class distance and inter-class distance include the gene similarity calculation method described above, and the intra-class and inter-class distances defined by graph connectivity and neighborhood features.

[0023] Preferably, the gene clustering number information can be provided by the clustering method itself, or by a random or sequential method.

[0024] The preferred method for integrating gene allelic information and genome clustering number information is string concatenation.

[0025] As a preferred approach, the structure of the gene-encoded breeding prediction model includes modules such as a gene clustering encoding input layer, an embedding layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, as well as strategies to improve the generalization ability of the neural network, including L1 / L2 regularization and Dropout, and optimization learning algorithms such as Adam.

[0026] Preferably, the input layer includes the gene clustering coding information obtained in step 4, or the gene clustering coding information plus gene location information, and the output layer includes a classification layer or regression layer related to the target task, or as a pre-trained multi-task classification and regression layer.

[0027] Preferably, the gene-encoded breeding prediction model is obtained through two-stage learning and training. In the first stage of learning, it serves as a pre-trained twin network, receiving encoded inputs from two gene strings and simultaneously learning differential and summation tasks at the output layer. In the second stage of learning, it serves as a pre-trained fixed-weight network layer for further training, participating in the fine-tuning of the target task.

[0028] As a preferred method, the selection of high-quality seed sets is to obtain a preferred seed set and its corresponding parent combinations by setting and optimizing reasonable thresholds.

[0029] The beneficial effects of this invention are as follows: Compared with the prior art, this invention provides a gene-encoding breeding prediction method based on graph clustering. First, it collects the phenotypic information and allele information of the organism to be predicted, required for precise molecular breeding. Then, it determines the gene map and adjacent edge weights based on the strength of inter-gene correlation. Next, it performs clustering on the gene map to obtain the number of commonly regulating genomes and the genome cluster number information for each gene. Then, it fuses the gene allele information and the genome cluster number information to obtain the gene cluster encoding of the sample. Finally, based on the gene cluster encoding information, or with the addition of gene location information and the phenotypic information of the organism to be predicted, it... This paper proposes a deep convolutional neural network to optimize gene breeding prediction performance. The method fully utilizes the gene interaction network contained in gene maps, extracts information on common regulatory genome clusters through graph clustering, proposes a novel gene clustering encoding method that integrates gene allelic information and gene map clustering information, and adds gene location information. By leveraging the weight sharing of deep convolutional neural networks, it can effectively extract regulatory gene features that control biological phenotypic output, solving the problem of insufficient encoding of gene interactions between gene maps in the input encoding layer of classical models, and ensuring the accuracy of gene breeding prediction for biological phenotypes. Attached Figure Description

[0030] Figure 1 A flowchart illustrating a gene-encoding breeding prediction method based on graph clustering, provided for an embodiment of the present invention;

[0031] Figure 2 A schematic diagram of cloud-edge-device collaborative deployment of a gene-encoding breeding prediction device based on graph clustering provided in an embodiment of the present invention;

[0032] Figure 3 A flowchart for collecting gene information provided in an embodiment of the present invention;

[0033] Figure 4 An architecture diagram of a single-stage deep convolutional neural network model provided in an embodiment of the present invention;

[0034] Figure 5 An architecture diagram of a two-stage deep convolutional neural network model provided in an embodiment of the present invention;

[0035] Figure 6 This is a structural block diagram of a gene-encoding breeding prediction device based on graph clustering provided in an embodiment of the present invention.

[0036] In the diagram: 11-Information collection module; 12-Gene clustering module; 13-Encoding pre-training module; 14-Continued training module; 15-Breeding prediction and screening module. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0038] To facilitate understanding of this embodiment, a gene coding breeding prediction method based on graph clustering disclosed in this embodiment of the invention will first be described in detail.

[0039] Example 1:

[0040] A gene-encoded breeding prediction method based on graph clustering is disclosed. This method combines graph clustering to encode genes, and then uses a gene-encoded breeding prediction model to predict the phenotypic information of the offspring to be predicted. Based on the predicted phenotypic information of the offspring, a set of high-quality seeds is selected. The gene-encoded breeding prediction model is trained on a collected dataset, and the training method is described in [reference needed]. Figure 1 Specifically, it includes:

[0041] S101. Collect the biological phenotypic information, genotype data and gene location information of each sample required for precision molecular breeding, and construct a dataset.

[0042] In embodiments of the present invention, the execution entity of the method is a computational breeding center. Specifically, if the computational breeding center is located on a computing cloud, then the computing cloud is the execution entity of the method; if the computational breeding center is located on a computing endpoint node, then the computing endpoint node is the execution entity of the method.

[0043] Specifically, in this embodiment of the invention, the computational breeding center predicts biological phenotypes based on high-dimensional genotype data, aiming to screen paternal and maternal parents based on the predicted values ​​to produce superior offspring. Ordinary gene coding, primarily based on numerical values, numerical mappings, or one-hot encoding, cannot effectively extract multi-neighborhood structural features from gene network maps. Graph clustering-based gene coding can effectively extract this structural information, thereby improving breeding prediction capabilities.

[0044] In one alternative implementation, refer to Figure 3 The collection of phenotypic and genotypic information required for breeding prediction includes: collecting biological tissue samples, extracting DNA / RNA, preparing samples for library construction, and sequencing.

[0045] S102. For each sample's genotype data, an undirected graph is constructed as a gene map based on the strength of inter-gene correlation. The strength of inter-gene correlation can be obtained by calculating the similarity of multiple SNP sites between pairs of genes in the genotype data. Methods include Pearson correlation coefficient, Jaccard correlation coefficient, Spearman correlation coefficient, Euclidean distance, cosine similarity, Manhattan distance, Hamming distance, edit distance, Chebyshev distance, Minkowski distance, and information entropy. The calculated similarity is the adjacent edge weight of the undirected graph. As one implementation scheme, the strength of inter-gene correlation is determined by the Pearson correlation coefficient, calculated as follows:

[0046]

[0047] Where X and K are vector representations of two genes, X i K i Let be the i-th component of the vector representation of genes X and K, respectively. is the mean of the corresponding gene. N is the dimension of the gene vector, where the weight of adjacent edges is taken from 0 to 1 in the correlation between genes, and 0 is taken for edges outside the interval, which are considered to be non-connected.

[0048] In one optional implementation, based on the gene string information and gene correlation calculation formula of the development set of samples, the correlation between each gene locus is calculated to obtain a heatmap of correlation between gene loci, namely the adjacent edge weight and adjacency matrix of gene loci.

[0049] S103. Perform clustering on the gene map constructed in step S102 to obtain the number of commonly regulated genomes and the genome clustering number information for each gene.

[0050] After constructing the gene map, unsupervised clustering is performed on the map. Gene clustering methods include spatial clustering, density clustering, hierarchical clustering, or spectral clustering, etc. Specifically, this step includes the following sub-steps:

[0051] The number of co-regulatory genomes, i.e. the number of gene clusters, is estimated based on the spatial distribution characteristics of gene maps. The methods for estimating the number of co-regulatory genomes are statistical methods, random methods, exhaustive methods, or iterative methods. Among them, iterative methods mainly refer to the method of determining the number of clusters by bottom-up or top-down iterative clustering in hierarchical clustering.

[0052] Based on the estimated number of gene clusters, calculate the intra-cluster distance and inter-cluster distance for each gene to determine the cluster to which the gene belongs;

[0053] After clustering is completed, each gene cluster is assigned a unique cluster ID, which serves as the genome cluster ID for each gene within that cluster. This gene cluster ID can be provided by the clustering method itself, or it can be provided randomly or sequentially.

[0054] In an optional implementation, based on the gene map corresponding to the gene locus adjacency matrix constructed above, the connected subgraph algorithm is applied to obtain the gene locus clustering results, connect isolated gene points together, number them sequentially, and obtain the number of commonly regulated genomes and the genome clustering number information for each gene.

[0055] S104. The allele information and genome clustering number information corresponding to each gene in the genotype data are fused together to obtain the gene clustering code of the sample.

[0056] After obtaining the gene clustering IDs, the gene clustering code of the fused sample can be obtained by concatenating each element. The gene clustering code method is shown below:

[0057]

[0058] Where S ij For the gene of the j-th component of sample i, SNP(S ij ) represents the corresponding allele characteristic, i.e., the original encoding of that gene in the genotype data. Group(S) ij () represents the corresponding cluster number. For fusion operators, it represents symbol concatenation operations.

[0059] S105. Based on gene clustering coding information, gene location information, and biological phenotypic information, construct and train a gene coding breeding prediction model to optimize gene breeding prediction performance.

[0060] After the gene clustering encoding of the samples is completed, the gene clustering encoding features of the samples are obtained. Based on these features, and with the addition of gene location information and the biological phenotype information to be predicted, a gene encoding breeding prediction model based on deep convolutional neural networks can be constructed. Dropout and L1 / L2 regularization strategies are added to optimize the gene breeding prediction performance.

[0061] In one alternative implementation, refer to Figure 4 By fusing gene allele information and gene clustering number information, the input and encoding layers (gene clustering encoding input layer) of the deep convolutional neural network are constructed in an element-wise splicing manner. Then, the embedding layer, SpatialDropout1D, 1D convolution, 1D max pooling, flattening, and fully connected are connected in sequence. Finally, the classification / regression output layer and phenotypic output related to the target task are connected.

[0062] In one alternative implementation, refer to Figure 5 The model construction and learning phases are divided into two stages: Stage 1 is the pre-training stage, where gene allele information, gene clustering number information, and gene location information are fused to form the input and encoding layers of a dual-channel, multi-task deep convolutional neural network. This is then fed into a shared twin network, with the output layer handling the difference and summation tasks. The entire network has two input channels, representing gene strings from two different samples. The output is a multi-task output: the difference task determines the polarity of the difference between the phenotypic values ​​of the gene strings in the left and right channels, while the summation task regresses the sum of the phenotypic values ​​of the gene strings in the left and right channels. The objective loss functions for the difference and summation tasks are Ld and Ls, respectively. - and L + Generally, L + The mean square error (MSE) and L can be used. - Cross-entropy can be taken. As shown below:

[0063]

[0064]

[0065] Where Y i1 Y i2 These are the actual sample label values. Let I(X) be the predicted value of the sample, and let I(X) be the indicator function, which has a value of 1 when X is true and a value of 0 when X is false; M represents the sample size; σ is the Sigmoid function, calculated as shown in the following formula:

[0066]

[0067] Where e is the natural constant, and its value is approximately 2.71828.

[0068] The multi-task objective loss function for Phase 1 is shown below:

[0069] L1=αL + +βL -

[0070] Where L + L - L and L are the loss functions for the summation task, the difference task, and the total task, respectively. α and β are the weight hyperparameters of the loss functions for the summation task and the difference task, respectively. The parameter values ​​can be determined by the grid search method.

[0071] Phase two is the continued training phase. The pre-trained shared twin network from Phase one is loaded, and its weights are fixed to prevent it from participating in the subsequent weight tuning. Then, the target task-related classification / regression output layer and phenotypic output are connected to construct the deep convolutional neural network for this phase. The model for this phase is trained and tuned based on gene string sample data and corresponding phenotypic data from the development set, resulting in the final target prediction model for the breeding task. For phenotypic regression prediction, the target loss function in Phase two is the mean squared error (MSE); for phenotypic classification prediction, the target loss function is the cross-entropy. Taking regression prediction as an example, the target loss function for Phase two is as follows:

[0072]

[0073] S106. Based on the constructed gene breeding prediction model, select a high-quality seed set, that is, optimize the optimal parent combination.

[0074] For trait classification tasks, the optimal parental combination is determined by directly selecting seed pools predicted to fall under the specified trait category based on the constructed gene breeding prediction model. For trait regression, i.e., numerical prediction tasks, the optimal parental combination is determined by selecting seed pools whose predicted values ​​reach or exceed a specified threshold based on the constructed gene breeding prediction model. For trait regression tasks, the specified screening threshold can be obtained through comprehensive optimization of the screening ratio and the experimental field size.

[0075] Existing gene input feature encoding methods typically consider numerical or one-hot encoding formats. For example, numerical formats usually map 0 / 0, 0 / 1, and 1 / 1 to -1, 0, and 1, or to 0, 1, and 2, failing to express complex gene network features. One-hot encoding merely discretizes the original features without increasing their information content. Graph clustering-based gene encoding, however, integrates original allele features and jointly regulating genomic features, thereby capturing higher-level graph neighborhood structure features and significantly improving the predictive ability of deep neural network-based models. This method fully utilizes the gene interaction network contained in gene maps, extracts jointly regulating genomic clustering information through graph clustering, and proposes a novel gene clustering encoding method that integrates gene allele information and gene map clustering information. Furthermore, by leveraging the weight sharing of a dual-channel, multi-task deep convolutional neural network, it can effectively extract regulatory gene features that control biological phenotypic output, addressing the problem of insufficient encoding of gene interactions between gene maps in the input encoding layer of classic models, and ensuring the accuracy of gene breeding predictions of biological phenotypes.

[0076] Example 2:

[0077] A gene-coding breeding prediction device based on graph clustering, referencing Figure 2 and Figure 6 The device includes:

[0078] The information collection module 11 is used to collect phenotypic information and genotypic data of the organisms to be predicted, which are required for precision molecular breeding. This module is mainly completed by a smart terminal.

[0079] The gene clustering module 12 is used to determine the gene map and adjacent edge weights based on the strength of correlation between genes, and to perform clustering on the gene map to obtain the number of commonly regulated genomes and the genome cluster number information for each gene. This module is completed by cloud-edge-device collaborative computing.

[0080] The encoding pre-training module 13 is used to fuse gene allele information, genome clustering number information, and gene location information, and to construct a dual-channel multi-task deep convolutional neural network based on the fused encoding information. It simultaneously performs differential and summation tasks and optimizes network weights for dual-gene string augmentation data, resulting in a shared twin network and its weights for subsequent training. This module is completed through cloud-edge-device collaborative computing.

[0081] The continued training module 14 is used to load the shared twin network and its weights obtained from the encoding pre-training module, solidify the weights, and continue training and fine-tuning for classification / regression tasks related to breeding target traits to obtain a prediction model for breeding target traits. This module is completed by cloud-edge-device collaborative computing.

[0082] The breeding prediction and screening module 15 is used to screen high-quality seed sets, i.e., optimize the optimal parent combination, based on the gene breeding prediction model built by the continued training module. This module is completed by cloud-edge-device collaborative computing.

[0083] In the gene-encoding breeding prediction device based on graph clustering of this invention, training set and prediction set data required for gene breeding prediction are first collected. The training set data includes the phenotypic information of the organism to be predicted and the allele information to be encoded, which are required for precise molecular breeding. The prediction set data includes the allele information of the seeds to be predicted. Then, the correlation between different gene loci is calculated, and a gene adjacency matrix and graph are constructed based on the strength of the correlation. The gene graph is then clustered to obtain the number of commonly regulating genomes and the genome cluster number information of each gene. Next, a dual-channel multi-task deep convolutional neural network is constructed by fusing gene allele information, genome cluster number information, and gene location information for pre-training to obtain a shared twin network and its weights. Then, the shared twin network and its weights are loaded and the weights are fixed. The network is then trained and optimized for classification / regression tasks related to breeding target traits to obtain a prediction model for breeding target traits. Finally, a high-quality seed set is selected based on the gene breeding prediction model, i.e., the optimal parent combination is optimized. This device can fully utilize the gene interaction network contained in the gene map, extract the clustering information of the genome that jointly regulates the genome through graph clustering, and propose a gene clustering encoding method that integrates gene allelic information and gene map clustering information. By utilizing the weight sharing of a dual-channel multi-task deep convolutional neural network, it can effectively extract the regulatory gene features used to control the output of biological phenotypes, solve the problem of insufficient encoding of gene interaction relationships between gene maps in the input encoding layer of classical models, and ensure the accuracy of gene breeding prediction of biological phenotypes.

[0084] The computer program product of the gene-encoding breeding prediction method and apparatus based on graph clustering provided in this invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0086] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion 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 invention. 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.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A gene-encoding breeding prediction method based on graph clustering, characterized in that, Includes the following steps: Obtain genotype data and gene location information of the offspring to be predicted; An undirected graph is constructed as a gene map based on the strength of correlations between genes in genotype data. Clustering is performed on the gene map to obtain the number of genes that commonly regulate the genome and the genome cluster number for each gene; The allele information and genome clustering number information corresponding to each gene in the genotype data are combined and concatenated to obtain the gene clustering code of the sample; Gene clustering codes and gene location information are input into a gene-coded breeding prediction model to obtain the biophenotype information of the offspring to be predicted; based on the biophenotype information of the predicted offspring, a set of high-quality seeds is selected. The gene-encoded breeding prediction model is trained on a collected dataset. Each sample in the dataset includes gene clustering codes, gene location information, and biophenotype information. The model is trained in two stages. The first stage uses a shared bridging network with a dual-channel gene clustering input layer, accepting gene clustering codes from two samples respectively. The output layer simultaneously learns a difference task and a summation task. The network input consists of two channels, representing gene strings from two different samples. The output is a multi-task output: the difference task determines the polarity of the difference between the phenotypic values ​​of the gene strings in the left and right channels, while the summation task regresses the sum of the phenotypic values ​​of the gene strings in the left and right channels. The objective loss functions for the difference and summation tasks are L0 and L1, respectively. - and L + As shown below: Where Y i1 Y i2 These are the actual sample label values. , Let X be the predicted value of the sample, and I(X) be the indicator function, which has a value of 1 when X is true and a value of 0 when X is false. Indicates the number of samples; The Sigmoid function is calculated using the following formula: Where e is the natural constant; the multi-task objective loss function L1 for stage one is shown below: in and The weight hyperparameters are the loss functions for the addition and difference tasks, respectively. The second stage is based on the fixed network parameters trained in the first stage, with only one gene clustering encoding input layer receiving the gene clustering encoding and gene location information from one sample to participate in the fine-tuning learning of the target task until training is completed. For the regression prediction task of phenotype, the target loss function of the second stage is the mean squared error (MSE). For the classification prediction task of phenotype, the target loss function of the second stage is the cross-entropy.

2. The method according to claim 1, characterized in that, The strength of the correlation between genes is obtained by calculating the similarity of multiple SNP sites between pairs of genes in the genotype data. The methods include Pearson correlation coefficient, Jaccard correlation coefficient, Spearman correlation coefficient, Euclidean distance, cosine similarity, Manhattan distance, Hamming distance, edit distance, Chebyshev distance, Minkowski distance, and information entropy. The calculated similarity is used as the weight of the adjacent edge to construct an undirected graph.

3. The method according to claim 1, characterized in that, Clustering was performed on the gene map to obtain the number of commonly regulated genomes and the genome cluster number information for each gene, as detailed below: The number of co-regulatory genomes, i.e. the number of gene clusters, is estimated based on the spatial distribution characteristics of gene maps; Based on the estimated number of gene clusters, calculate the intra-cluster distance and inter-cluster distance for each gene to determine the cluster to which the gene belongs; After clustering is completed, each gene cluster is assigned a unique clustering number, which serves as the genome clustering number for each gene in the corresponding gene cluster.

4. The method according to claim 3, characterized in that, Gene clustering methods include spatial clustering, density clustering, hierarchical clustering, or spectral clustering.

5. The method according to claim 3, characterized in that, Methods for estimating the number of co-regulatory genomes include statistical methods, random methods, exhaustive methods, or iterative methods. Among them, iterative methods mainly refer to the method of determining the number of clusters by bottom-up or top-down iterative clustering in hierarchical clustering.

6. The method according to claim 1, characterized in that, Gene clustering number information is provided by the clustering method itself, or by random or sequential methods.

7. The method according to claim 1, characterized in that, The biological phenotypic information includes the quantity, quality, percentage, or classification of the target phenotypic.

8. The method according to claim 1, characterized in that, The gene-encoded breeding prediction model includes a gene clustering encoding input layer, an embedding layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

9. A gene-encoding breeding prediction device based on graph clustering, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the gene-encoding breeding prediction method based on graph clustering as described in any one of claims 1-8.

Images