Crop multi-omics phenotypic prediction methods and devices
By preprocessing and feature splitting of crop multi-omics data, performing parallel convolutional computation, and constructing a gene regulation network using graph neural networks, the problem of low accuracy in crop phenotypic prediction in existing technologies is solved, and efficient characterization and accurate prediction of complex genetic effects are achieved.
Patent Information
- Application Number
- CN202611109231.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies have low ability to characterize complex genetic effects and struggle to characterize inter-locus interactions when predicting crop phenotypes, resulting in low accuracy in crop phenotype predictions.
A crop multi-omics phenotypic prediction method is adopted. By preprocessing the genomic and transcriptomic data of the target crop, initial feature maps are extracted and feature splitting and parallel convolution calculation are performed. A gene regulatory network is constructed by combining graph neural networks to obtain joint feature representations and finally perform phenotypic prediction.
It improves the accuracy of crop phenotypic prediction results, enhances the ability to characterize complex genetic effects, provides comprehensive and complementary feature descriptions, and improves breeding efficiency.
Smart Images

Figure CN123072491A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart agriculture technology, and in particular to a method and apparatus for predicting crop multi-omics phenotypic traits. Background Technology
[0002] Prediction of complex agronomic traits is an important research topic in modern crop molecular breeding, and it is of great significance for improving the efficiency of early generation selection and shortening the breeding cycle. Complex agronomic traits such as plant height, flowering time, and yield are usually determined by genetic variation, gene expression regulation, and multi-level molecular interactions.
[0003] In related technologies, genome prediction methods are commonly used to predict crop phenotypic data. This method establishes association models between genotypes and phenotypes based on whole-genome molecular markers. It mainly reflects static genetic differences at the DNA (Deoxyribonucleic acid) sequence level, but has low characterization ability for complex genetic effects. Furthermore, when processing high-dimensional SNP (Single Nucleotide Polymorphism) data, the above association models are difficult to characterize the non-additive interaction information between loci. Traditional technologies also use machine learning and deep learning methods to predict crop phenotypic data. These methods rely on pre-screening of loci for specific traits, thus losing the association and interaction information between loci before modeling, resulting in low accuracy of crop phenotypic prediction results. Summary of the Invention
[0004] This invention provides a crop multi-omics phenotypic prediction method and apparatus to address the shortcomings of existing technologies in predicting crop phenotypes, such as low ability to characterize complex genetic effects and difficulty in characterizing inter-site interaction information, resulting in low accuracy of crop phenotype prediction results.
[0005] This invention provides a crop multi-omics phenotypic prediction method, comprising: The genomic data, transcriptome data, and agronomic trait data of the target crop are preprocessed to obtain genomic feature input information and transcriptome expression matrix; An initial feature map is extracted from the genomic feature input information, and the initial feature map is split along the channel dimension to obtain shallow bypass features and main branch input features; the main branch input features are subjected to parallel convolution calculation at different scales, and the parallel convolution calculation results and corresponding weights are weighted to obtain the main branch output features; The main branch output features and the shallow bypass features are spliced and mapped to obtain a genome feature representation, and the genome feature representation is fused with the transcriptome feature representation to obtain a joint feature representation; the transcriptome feature representation is determined based on the transcriptome expression matrix. The phenotypic information of the target crop is predicted based on the joint feature representation to obtain the phenotypic prediction result.
[0006] According to the crop multi-omics phenotypic prediction method provided by the present invention, the parallel convolution calculation results and corresponding weights are weighted to obtain the main branch output features, including: Basic local features are extracted from the main branch input features, and the basic local features are sequentially processed by pooling, one-dimensional convolution and non-linear activation to obtain attention weights; The parallel convolution calculation results corresponding to different scales are fused with the basic local features to obtain multi-scale fused features; The attention weights are used to perform gated residual weighting on the multi-scale fusion features to obtain the main branch output features.
[0007] According to the crop multi-omics phenotypic prediction method provided by the present invention, the transcriptome expression matrix includes multiple genes and the corresponding expression levels of each gene; The transcriptome features were obtained through the following steps: The regulatory relationships between different genes are determined based on the transcriptome expression matrix, and a gene regulation network is constructed using the multiple genes as network nodes and the regulatory relationships as network edges. The expression levels of each gene are used as the node features of the corresponding network nodes, and a graph neural network is used to perform structured feature learning along the network edges to obtain the transcriptome feature representation.
[0008] According to the crop multi-omics phenotypic prediction method provided by the present invention, the multiple genes include transcription factors and target genes; The construction of a gene regulatory network using the plurality of genes as network nodes and the regulatory relationships as network edges includes: The transcriptome expression matrix was processed using weighted gene co-expression network analysis to obtain the search range of candidate regulatory relationships; The directed regulatory relationship between the transcription factor and the target gene is determined based on the search range of the candidate regulatory relationship; The gene regulatory network is obtained by using the feature importance score as the initial weight of the candidate regulatory edges, using the initial weight to perform sparsification screening of the candidate regulatory edges, and then standardizing or monotonically transforming the weights of the screened regulatory edges.
[0009] According to the crop multi-omics phenotypic prediction method provided by the present invention, the method utilizes a graph neural network to perform structured feature learning along the network edges to obtain the transcriptome feature representation, including: Using each network node in the gene regulation network as the target node and the node in the gene regulation network that points to the target node along the corresponding network edge as the source node, the message features are calculated based on the node features of the source node and the weights of the corresponding network edges, and the message features are sent to the target node through the source node. The message features received by the target node from all source nodes are aggregated to obtain neighborhood aggregation features. The neighborhood aggregation features are mapped and transformed to obtain node update terms, and the node features of the target node are updated using the node update terms to obtain the node-level feature representation of the network node. The gating weights of each network node are calculated, and the node-level feature representations of all network nodes are weighted and summed using the gating weights to obtain the transcriptome feature representation.
[0010] According to the crop multi-omics phenotypic prediction method provided by the present invention, the step of updating the node features of the target node using the node update term to obtain the node-level feature representation of the network node includes: By adding the node features of the target node to the node update term through residual connection, the node-level feature representation of the network node is obtained; Wherein, when the target node does not have a corresponding source node in the gene regulation network, the neighborhood aggregation feature of the target node is configured as a zero vector.
[0011] According to a crop multi-omics phenotypic prediction method provided by the present invention, after obtaining the phenotypic prediction result value, the method further includes: The contribution of each input feature in the genomic feature input information and / or transcriptome expression matrix to the phenotype prediction result is calculated using the interpretive analysis method. Based on the contribution values, a set of high-contribution features is selected, and the set of high-contribution features is used as candidate functional genes related to the target agronomic trait.
[0012] The present invention also provides a crop multi-omics phenotypic prediction device, comprising: The preprocessing module is used to preprocess the genomic data, transcriptome data and agronomic trait data of the target crop to obtain genomic feature input information and transcriptome expression matrix; The feature calculation module is used to extract an initial feature map from the genomic feature input information, and perform feature splitting on the initial feature map along the channel dimension to obtain shallow bypass features and main branch input features; perform parallel convolution calculation on the main branch input features at different scales, and perform weighted calculation on the parallel convolution calculation results and corresponding weights to obtain the main branch output features; The feature fusion module is used to perform feature splicing and mapping on the main branch output features and the shallow bypass features to obtain a genomic feature representation, and to fuse the genomic feature representation with the transcriptome feature representation to obtain a joint feature representation; the transcriptome feature representation is determined based on the transcriptome expression matrix; The prediction module is used to predict the phenotypic information of the target crop based on the joint feature representation, and obtain the phenotypic prediction result.
[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the crop multi-omics phenotypic prediction method as described above.
[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the crop multi-omics phenotypic prediction method as described above.
[0015] The crop multi-omics phenotypic prediction method and apparatus provided by this invention eliminates the differences in scale and format between different omics data by preprocessing the original multi-source data of the target crop; it enhances the feature dimensionality of SNP sequences through nonlinear mapping and feature splitting mechanisms, and obtains non-additive interactions and synergistic effects between SNP sites in different length segments through multi-scale convolution and attention weighting mechanisms, thereby enhancing the ability to characterize complex genetic effects of crops; it provides comprehensive and complementary feature descriptions of complex traits through splicing mapping and cross-modal fusion strategies; and finally, it improves the accuracy of crop phenotypic prediction results by predicting phenotypic information of the target crop through joint feature representation. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0017] Figure 1 This is one of the flowcharts of the crop multi-omics phenotypic prediction method provided by the present invention.
[0018] Figure 2 This is the second flowchart of the crop multi-omics phenotypic prediction method provided by the present invention.
[0019] Figure 3 This is one of the structural schematic diagrams of the crop multi-omics phenotyping device provided by the present invention.
[0020] Figure 4 This is the second schematic diagram of the structure of the crop multi-omics phenotyping device provided by the present invention.
[0021] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0023] The following is combined Figures 1-4 The present invention describes a crop multi-omics phenotypic prediction method and apparatus.
[0024] Figure 1 This is one of the flowcharts illustrating the crop multi-omics phenotypic prediction method provided by the present invention, such as... Figure 1 As shown, the method includes the following steps: Step 100: Preprocess the genomic data, transcriptome data, and target agronomic trait data of the target crop to obtain genomic feature input information and transcriptome expression matrix.
[0025] In this step, the target crop can be a crop with complex agronomic traits, such as corn, wheat, or rice.
[0026] In this step, the genomic data are whole-genome SNP (single nucleotide polymorphism) molecular marker data of crop samples, used to characterize the underlying static genetic differences of the crop.
[0027] In this step, transcriptome data, which is gene expression data, is used to characterize the dynamic molecular interaction states regulated by environment and development.
[0028] In this step, the target agronomic trait data can be the specific phenotype of interest in breeding (such as plant height, flowering time, yield, etc.), and this trait data can be used as the prediction target (label) for model training.
[0029] In this embodiment, the SNP sequence is converted into a vector or matrix that can be input into a neural network (usually encoded as 0, 1, 2) to obtain genomic feature input information for subsequent extraction of genetic features.
[0030] In this embodiment, the transcriptome expression matrix can be a numerical matrix of gene expression levels after screening and normalization, which is used for subsequent construction of regulatory networks and extraction of expression features.
[0031] In this embodiment, the preprocessing includes, but is not limited to: quality checking of the raw data, removal of outliers, duplicates and low-quality data, encoding representation of genomic data, sorting expression levels of transcriptome data, standardizing target agronomic trait data to eliminate dimensions, imputing missing values, and standardizing or normalizing the processed data.
[0032] Specifically, after obtaining the genomic data, transcriptome data, and agronomic trait data of the target crop, the following operations are performed: S11: Data Acquisition: Acquire genotypic, transcriptomic, and phenotypic data of crop population samples; S12: Quality Inspection and Data Cleaning: Perform quality inspection on the acquired genomic data, transcriptome data, and phenotypic data to ensure the integrity and accuracy of the data, including confirming the consistency of sample identification, checking and correcting outliers, errors, and duplicates, and removing low-quality features from the transcriptome data; In a preferred embodiment, for genomic data, only biallelic SNP sites may be retained and screened according to the least allele frequency (MAF) and deletion rate threshold; for transcriptomic data, genes with a median expression level of zero may be removed.
[0033] S13: Missing Value Handling and Standardization: Missing sites in the genomic data are imputed, missing expression values in the transcriptome data are processed, and the genomic data, transcriptome data, and target agronomic trait data are standardized or normalized respectively. In a preferred embodiment, the expression values of each gene across samples can be subjected to a rank-based inverse normal transformation.
[0034] S14: Sample Alignment: Align genomic data, transcriptomic data and target agronomic trait data according to sample number to ensure that different omics data come from the same batch of crop samples, and construct a unified training dataset and test dataset accordingly.
[0035] S15: Constructing Model Input Data: Model input data includes two modalities: genomic and transcriptomic data.
[0036] Specifically, for a single sample, its genomic modality (i.e., genomic feature input information) is represented as a length of The SNP encoding vector is shown in the following formula: ; in, Indicates the first The genome encoding of each SNP site.
[0037] Correspondingly, transcriptome modalities (i.e., the transcriptome expression matrix) is represented as a matrix of length [length missing]. Gene expression vector, as shown in the following formula: ; in, Indicates the first The expression values of each gene.
[0038] Step 200: Extract the initial feature map from the genomic feature input information, and perform feature splitting along the channel dimension of the initial feature map to obtain shallow bypass features and main branch input features; perform parallel convolution calculation on the main branch input features at different scales, and perform weighted calculation on the parallel convolution calculation results and corresponding weights to obtain the main branch output features.
[0039] In this step, the initial feature map is obtained by mapping the genomic feature input information to the channel dimension using a nonlinear transformation unit, and is used to initially extract sequence features.
[0040] In this embodiment, the nonlinear transformation unit can be composed of one-dimensional convolution, batch normalization, and activation functions.
[0041] In this step, the initial feature map is divided into two parts along the channel by feature splitting. The shallow bypass features are the features that are retained and do not participate in the subsequent deep convolution, which are used to maintain the integrity of the original genetic information in the low layer (to prevent information loss in the deep network). The main branch input features are the other part of the features that are fed into the subsequent complex convolution module, which are used to refine the local genetic patterns in the deep layer step by step.
[0042] In this embodiment, multiple one-dimensional deep convolutional kernels with different receptive fields (such as 3, 5, 7, 9, 11) can be used to extract local genetic information in SNP sequences in parallel, thereby adaptively capturing cross-scale local cooperative genetic features such as linkage disequilibrium (LD) caused by different physical distances.
[0043] In this embodiment, attention or gating mechanisms can be used to generate adaptive importance weights for features at different locations and weight them with the convolution result to obtain the main branch output features. These features are used to enhance the information representation of key variant sites and suppress redundant background noise.
[0044] Further, in step 200, the parallel convolution calculation results and corresponding weights are weighted to obtain the main branch output features, including: extracting basic local features from the main branch input features, and performing pooling, one-dimensional convolution and nonlinear activation on the basic local features in sequence to obtain attention weights; fusing the parallel convolution calculation results corresponding to different scales with the basic local features to obtain multi-scale fusion features; and using the attention weights to perform gated residual weighting on the multi-scale fusion features to obtain the main branch output features.
[0045] In this embodiment, the genomic feature input information G and the transcriptome expression matrix T are sent to the genomic branch and the transcriptome branch, respectively, and the multi-scale genomic context mixing module (MGCM) is used to extract the cross-scale local genetic features in the SNP sequence, i.e. the main branch output features.
[0046] Specifically, the process includes the following steps: S21: Initial feature mapping: Due to linkage disequilibrium (LD), genetic variations between adjacent sites usually exhibit local correlation that decreases with physical distance. In order to extract this distance-dependent genetic pattern, this embodiment introduces a multi-scale genome context mixing module in the genome branch to aggregate multi-scale local context information in parallel while keeping the sequence length unchanged.
[0047] For a given SNP feature vector (B is the batch size,) The sequence length (i.e., genomic feature input information) is first used to perform feature mapping using the nonlinear transformation unit CBS, generating a channel number of... initial feature map : ; CBS is a minimal nonlinear transform unit consisting of sequentially cascaded one-dimensional convolutions, batch normalization (BN), and SiLU activation functions. The kernel size and stride of these convolutions are both 1, and the dimensionality is maintained through automatic padding. CBS is represented as follows: .
[0048] S22: Feature splitting: In this embodiment, the initial feature map is divided into two parts along the channel dimension: ; in, This is a shallow bypass feature, which is directly retained in subsequent fusion stages to maintain the integrity of low-level information. Input features for the main branch; Sent by The main branch is composed of interconnected contextual hybrid subunits with identical structures and independent parameters, which are used to refine deeper local patterns step by step.
[0049] In this embodiment, the number is recorded as follows: The nonlinear transformation achieved by each sub-unit is And let the main branch start with input. Then the hierarchical propagation relationship of the main branch is: ; in, In the main branch The features output by the first context mixing subunit, i.e., after the first context mixing subunit... Local genetic characteristics obtained after primary refining.
[0050] S23: Multi-scale local pattern extraction and context aggregation: Within the sub-unit, a lightweight pre-transformation is first performed on the input to extract basic local features. : .
[0051] Subsequently, this embodiment employs multi-scale one-dimensional depthwise convolution to extract local patterns under different receptive fields, and adds them element-wise with the preceding features to obtain multi-scale fused features. : ; in, Indicates the kernel size as In this embodiment, a one-dimensional depthwise convolution is used, where K is the set of kernel sizes used in parallel. The number of elements in the set is the number of scales extracted in parallel.
[0052] In this embodiment, the network can adaptively capture the synergistic effects between sites at different scales through this parallel multi-branch structure.
[0053] S24: Cross-channel fusion and attention enhancement: This embodiment utilizes a nonlinear transformation unit to perform cross-channel information fusion on multi-scale fusion features, generating aggregated features. : ; To highlight key information and suppress redundant noise, this embodiment introduces a contextual attention mechanism. Using prior features as input, the input is sequentially processed through average pooling, depthwise convolution, and a sigmoid function to generate position-adaptive attention weights. : ; in, This is used to describe the importance of features at different locations.
[0054] In this embodiment, after obtaining the attention weights, gated residual connections are used for fusion, and then CBS is used to obtain the current unit output. : in, This indicates element-wise multiplication. For aggregation features The intermediate features obtained after gating and weighting the residuals are those that, while retaining the information of the aggregated features themselves, are superimposed with attention weights. The recalibrated characteristic components.
[0055] S25: Genome modality output: Completed in the main branch After the transformation, the shallow bypass features will be... Main branch input and output at each stage The genome features were spliced along the channel dimension and compressed using CBS to obtain the enhanced genome features. : .
[0056] In this embodiment, it will also The layers are sequentially normalized, fully connected, ReLU activated, and Dropout applied, mapping the data to fixed-dimensional genomic modality embeddings, i.e., the main branch output features. : ; Among them, the fully connected Linear layer maps features to dimensions. , As output features of the genome branch, they are used for subsequent multi-omics feature fusion.
[0057] This embodiment utilizes multi-scale convolution and attention weighting mechanisms to adaptively capture non-additive interactions and synergistic effects between SNP sites in segments of different lengths, enhancing the ability to characterize complex genetic effects in crops. Furthermore, this process does not require pre-screening of sites, thus improving feature extraction efficiency.
[0058] Step 300: Perform feature splicing and mapping on the main branch output features and shallow bypass features to obtain the genome feature representation, and fuse the genome feature representation with the transcriptome feature representation to obtain the joint feature representation; the transcriptome feature representation is determined based on the transcriptome expression matrix.
[0059] In this step, the main branch output features and shallow bypass features can be merged in the channel dimension and mapped to a fixed-dimensional vector through a fully connected layer to obtain the corresponding genome feature representation, which is used to characterize the static genetic potential of crops.
[0060] In this embodiment, the transcriptome feature representation is a feature vector characterizing the dynamic molecular state of crops. This feature representation can be obtained by constructing a gene regulatory network based on the transcriptome expression matrix by building directed regulatory edges, and then extracting the expression levels of nodes by propagating along the edges using a graph neural network.
[0061] In this step, the two types of cross-modal features, namely genomic feature representation and transcriptome feature representation, can be comprehensively represented in a unified feature space to obtain the corresponding joint feature representation, which is used for the final downstream prediction task.
[0062] Further, in step 300, the transcriptome expression matrix includes multiple genes and their corresponding expression levels; transcriptome features are obtained through the following steps: Step 310: Determine the regulatory relationships between different genes based on the transcriptome expression matrix, and construct a gene regulation network with multiple genes as network nodes and regulatory relationships as network edges.
[0063] In this embodiment, the regulatory relationship includes the directed interaction between transcription factors (TFs) and their target genes, representing the control logic of gene expression in organisms.
[0064] In this embodiment, network nodes can be composed of individual genes in the matrix, and network edges are composed of inferred directed regulatory relationships between genes, with weights representing the intensity of regulation.
[0065] In this embodiment, a directed weighted gene regulation network is constructed based on the aforementioned network nodes and network edges, which serves as a structured prior constraint for feature extraction by the graph neural network.
[0066] Furthermore, in step 310, the multiple genes include transcription factors and target genes; a gene regulatory network is constructed using multiple genes as network nodes and regulatory relationships as network edges, including: (1) The transcriptome expression matrix was processed by weighted gene co-expression network analysis to obtain the search range of candidate regulatory relationships; It should be noted that the weighted gene co-expression network analysis method in this embodiment is used to identify stable gene co-expression modules, so as to constrain the search space of subsequent gene regulatory relationships and reduce the computational complexity of inference.
[0067] In this embodiment, the search scope for candidate regulatory relationships can be several gene co-expression modules divided by clustering. This search scope is used to infer the complex network of the whole genome and limit it to the local modules with highly related expression patterns, thereby avoiding invalid global calculations.
[0068] (2) Determine the directed regulatory relationship between transcription factors and target genes based on the search range of candidate regulatory relationships.
[0069] In this embodiment, transcription factors (TFs) are proteins that can bind to specific DNA sequences and thereby turn on or off the expression of other genes. Target genes are downstream genes whose transcriptional state is controlled by transcription factors.
[0070] In this embodiment, a directed regulatory relationship refers to a one-way control and controlled relationship from transcription factors to target genes, which is used to characterize biological causal information in a graph structure.
[0071] In this embodiment, within each co-expression module, the gene regulatory network inference method (GRNBoost2) is used to infer the regulatory relationship between transcription factors and their target genes through gradient boosting regression trees, and the feature importance score of each transcription factor to the target gene is used as the initial weight of the corresponding directed edge. .
[0072] (3) The feature importance score is used as the initial weight of the candidate regulatory edge. The candidate regulatory edge is sparsified by the initial weight and the weight of the selected regulatory edge is standardized or monotonically transformed to obtain the gene regulatory network.
[0073] In this embodiment, a regression tree model can be used to fit the data, quantify the contribution of a transcription factor to the prediction of target gene expression, and obtain the corresponding feature importance score, which is then used as the initial weight of the candidate edges.
[0074] In this embodiment, weakly correlated edges are removed and strongly correlated edges are retained by sparsification screening based on their scores. This screening process is used to control network density and suppress noise.
[0075] In this embodiment, the monotonic transformation includes, but is not limited to, logarithmic transformation, other monotonic transformations, or normalization methods. The initial edge weights can be standardized through this monotonic transformation.
[0076] In this embodiment, the retained edge weight values are smoothed and compressed through standardization or monotonic transformation to prevent numerical fluctuations caused by excessive weight variance during subsequent deep learning training.
[0077] Specifically, to control the sparsity of the gene regulatory network and suppress the interference of weakly correlated edges or noisy edges on subsequent message propagation, this embodiment performs the following steps: (1) The candidate directed edges inferred by GRNBoost2 in each co-expression module are adjusted according to their initial weights. Perform sparsity screening.
[0078] In this embodiment, only the top-ranked features based on their incoming edges can be retained for each target gene. Transcription factor regulation edge ( (For preset positive integers), or retain scores higher than the threshold. The directed edges ( You can take a specified quantile of the score distribution of all candidate edges and delete the remaining candidate edges.
[0079] (2) For each directed edge retained after sparsification, in order to reduce the numerical fluctuations and training instability that may be introduced by the larger edge weights during message propagation, the initial weights are standardized.
[0080] For example, using the logarithmic transformation shown in the following formula... The process is performed to obtain the transformed edge weights. : ; (3) According to Directional regulation relationship Network node set Obtain a directed weighted gene regulatory network for transcriptome branching. : ; in, Indicates transcription factor Target genes The directional regulatory relationship.
[0081] This embodiment provides prior constraints for inferring gene regulatory networks in large-scale crops through weighted gene co-expression network analysis, narrowing the search space for candidate regulatory relationships. By inferring the directed control logic of transcription factors on target genes within a limited module, and then sparsifying and reducing noise on the initial regulatory edges, combined with monotonic transformation to smooth the weight distribution, it achieves control over network size and suppression of weak correlation noise, thereby improving the quality of gene regulatory networks.
[0082] Step 320: Using the expression level of each gene as the node feature of the corresponding network node, and using a graph neural network to perform structured feature learning along the network edges to obtain the transcriptome feature representation.
[0083] In this embodiment, the node feature is the scalar or vector attribute corresponding to each node in the gene regulatory network, that is, the normalized expression level of a single gene.
[0084] In this embodiment, a graph neural network (GNN) is a deep learning model used to process graph-structured data, which can transmit information based on the connection relationships between nodes.
[0085] In this embodiment, the structured feature learning can be the propagation and aggregation of messages in the graph neural network. That is, a gene not only contains its own expression information, but also integrates the state information of the upstream transcription factors that regulate the gene. Finally, after the full graph feature is aggregated, the transcriptome feature representation is obtained, which is used to characterize the overall molecular regulatory state of the current sample.
[0086] This embodiment utilizes a graph neural network to perform message propagation, aggregation, and attention readout mechanisms along the directed edges of gene regulation. This enables the acquisition of structured expression features constrained by gene regulatory relationships, thereby improving the effectiveness and stability of transcriptome modality information in predicting complex agronomic traits.
[0087] Further, in step 320, a graph neural network is used to learn structured features along the network edges to obtain transcriptome feature representations, including: Step 321: Using each network node in the gene regulation network as the target node and the node in the gene regulation network that points to the target node along the corresponding network edge as the source node, calculate the message features based on the node features of the source node and the weights of the corresponding network edges, and send the message features to the target node through the source node.
[0088] In this embodiment, it is assumed that the transcriptome expression matrix contains If there are 1 gene, then the set of network nodes Represented as: ; In this embodiment, a directed weighted gene regulatory network is used as the graph structure. The topology and edge weights of the graph are shared among all samples. When performing forward propagation on a specific sample, it is only necessary to write the gene expression of that sample into the corresponding node.
[0089] Specifically, first, each node The initial input features are defined as the expression values of the corresponding genes. , is represented as: ; Subsequently, scalar features are projected onto a node encoder that shares parameters across all nodes. The hidden space is used to obtain the initial representation of the nodes. : .
[0090] Step 322: Aggregate the message features of all source nodes received by the target node to obtain the neighborhood aggregation features.
[0091] In this embodiment, graph convolution propagation and aggregation are performed, specifically in the first... During the layer information transmission process, the source node Along the directed edge To the target node The message characteristics are as follows: From the source node Layer representation With corresponding edge weight A joint decision can be expressed by the following formula: .
[0092] In this embodiment, the target node sums all incoming edge messages to obtain the neighborhood aggregation feature. , is represented as: .
[0093] in, It is the set of all directed regulatory edges in the gene regulatory network.
[0094] Step 323: Map and transform the neighborhood aggregation features to obtain the node update term, and use the node update term to update the node features of the target node to obtain the node-level feature representation of the network node.
[0095] In this embodiment, linear layer multiplication, batch normalization (BatchNorm), nonlinear activation (GELU), dropout regularization, and other operations can be sequentially performed on the neighborhood aggregation features to obtain the node update term.
[0096] In this embodiment, a residual connection can be used to add the original node features to the node update term to obtain a node-level feature representation. After multiple layers of "transmission-aggregation-update" loops, each gene node finally obtains a high-dimensional semantic vector.
[0097] Further, in step 323, the node features of the target node are updated using the node update term to obtain the node-level feature representation of the network node, including: adding the node features of the target node to the node update term through residual connection to obtain the node-level feature representation of the network node; wherein, when the target node does not have a corresponding source node in the gene regulation network, the neighborhood aggregation feature of the target node is configured as a zero vector.
[0098] In this embodiment, node-level feature representations are calculated through the following steps: (1) The neighborhood aggregation features are processed sequentially by unbiased linear transformation, batch normalization (BatchNorm), GELU activation and Dropout to generate node update terms. : ; in, For the first The layer is a learnable weight matrix used to perform a linear transformation on the neighborhood aggregated features. For the real number field OK A set of matrices of columns.
[0099] (2) The node features of the target node are obtained through residual connection. Add the updated node features to the node update terms to obtain the updated node features. , can be represented as: .
[0100] In this embodiment, for nodes without incoming edges, their neighborhood aggregation features are... When the vector is set to zero, the residual connection makes the update of the node mainly dominated by its own representation in the previous layer, thereby alleviating the propagation instability caused by the lack of neighborhood information.
[0101] (3) The above node characteristics through After stacking the layers, node-level feature representations are obtained. , is represented as: .
[0102] This embodiment, through residual connections and a zero-vector padding strategy for network nodes, avoids numerical propagation collapse or training instability caused by a lack of neighborhood information input, thereby improving the robustness of the model in representing complex crop molecular networks.
[0103] Step 324: Calculate the gating weights of each network node, and use the gating weights to perform a weighted summation of the node-level feature representations of all network nodes to obtain the transcriptome feature representations.
[0104] In this embodiment, after obtaining the node-level feature representation, a gated weighted attention readout mechanism can be used to generate graph-level transcriptome features, i.e., transcriptome feature representation. The specific steps of this feature representation are as follows: (1) Calculate the key vector for each node using the following formula. Gating weights AND value vector : ; ; ; in, , , All are learnable parameters. Used to characterize network nodes Contribution weights to the representation of the whole graph.
[0105] (2) Utilization The graph-level features are obtained by weighted summation of the value vectors of all network nodes. , is represented as: .
[0106] (3) Map the graph-level features to a unified feature fusion space to obtain the corresponding transcriptome feature representations. : ; In this embodiment, As an embedded representation of transcriptome modalities, it is input together with the genomic feature representation output from the genome branch into the subsequent feature fusion layer to obtain a joint feature representation.
[0107] In this embodiment, genomic feature representation Transcriptome feature representation is used to describe relatively stable patterns of genetic variation. Used to characterize the molecular regulatory state formed under the constraints of gene regulatory networks.
[0108] Specifically, given , The two types of features characterize the sample from the genetic and functional levels, respectively. In this embodiment, a parameter-free splicing strategy can be used to fuse them to obtain the corresponding joint feature representation. It can be expressed by the following formula: .
[0109] in, It also integrates genetic variation information with molecular functional status information.
[0110] In some embodiments, the above feature fusion process may also be implemented using weighted summation, gating fusion, or attention fusion.
[0111] This embodiment calculates and transmits weighted message features along the directed network edges, mathematically representing the directional regulatory intensity of transcription factors on target genes in a graph neural network. By summing and aggregating the incoming edge messages of the target node, the target gene state determined by multiple transcription factors in complex agronomic traits can be obtained. By introducing a feature update mechanism that includes nonlinear transformation and residual connections, the regulatory relationships between genes are obtained, thereby obtaining high-quality node features rich in structured regulatory context. By introducing a graph-level readout mechanism with attention-gated weights, gene noise unrelated to the target agronomic trait is adaptively filtered out and key functional nodes are focused, improving the quality of transcriptome feature representation.
[0112] Step 400: Predict the phenotypic information of the target crop based on the joint feature representation to obtain the phenotypic prediction results.
[0113] In this step, the joint feature representation can be input into a prediction network (such as a feedforward neural network). The network nonlinearly maps the joint feature representation into continuous numerical values of specific agronomic traits. Finally, the model outputs the predicted target value, i.e. the phenotypic prediction result, which is used to guide the selection of breeding generations or to calculate the loss by comparing with the true phenotype.
[0114] In this embodiment, predicting the target agronomic trait prediction result based on the joint feature representation includes the following steps: (1) The input to the prediction network is normalized, then linearly transformed and mapped to the hidden space. Finally, a nonlinear transformation is performed using GELU activation and Dropout regularization to obtain the phenotypic prediction result. The specific process is represented by the following formula: ; in, , , , These are the weight matrices and bias vectors for each layer. For the hidden layer dimension.
[0115] (2) When training the prediction network, optimize all parameters by minimizing the regression loss between the predicted value and the true phenotype.
[0116] For example, during network training, mean squared error (MSE) loss can be used, expressed as: ; in, These are true agronomic phenotypic values. These are the model's predicted values.
[0117] In some embodiments, end-to-end training can be performed using optimizers such as AdamW, and hyperparameters such as batch size, learning rate, weight decay coefficient, early stopping threshold, regression head hidden layer dimension, and Dropout can be determined by automatic hyperparameter grid search based on validation set correlation.
[0118] During the model application phase, the trained model can make predictions using only genomic data, only transcriptome data, or a fusion of both; the target agronomic traits include, but are not limited to, plant height, flowering time, and yield.
[0119] The crop multi-omics phenotypic prediction method provided in this invention preprocesses the original multi-source data of the target crop to eliminate the differences in scale and format between different omics data; it enhances the feature dimensionality of SNP sequences through nonlinear mapping and feature splitting mechanisms, and obtains non-additive interactions and synergistic effects between SNP sites in different length segments through multi-scale convolution and attention weighting mechanisms, thereby enhancing the ability to characterize complex genetic effects of crops; it provides comprehensive and complementary feature descriptions for complex traits through splicing mapping and cross-modal fusion strategies; and finally, it predicts the phenotypic information of the target crop through joint feature representation, thereby improving the accuracy of crop phenotypic prediction results.
[0120] Furthermore, after obtaining the phenotypic prediction results, the crop multi-omics phenotypic prediction method also includes: using an interpretation analysis method to calculate the contribution value of each input feature in the genomic feature input information and / or transcriptome expression matrix to the phenotypic prediction results; selecting a set of high-contribution features based on the contribution values, and using the set of high-contribution features as candidate functional genes related to the target agronomic traits.
[0121] In this embodiment, based on the trained multi-omics fusion prediction model, the SHAP interpretation analysis method is used to calculate the contribution value of each input feature to the prediction result. SHAP is based on the Shapley value in cooperative game theory, which regards the model's prediction of a single sample as the total benefit of all feature collaboration and distributes it fairly accordingly.
[0122] Specifically, for those containing Samples of each feature ,feature The Shapley value is defined as the weighted average of the predicted changes caused by adding it to all possible subsets of features. The absolute value represents the contribution intensity. Represented as: ; in, For the complete feature set, For those without features any subset of, For only subsets The expected model output when the features are known information.
[0123] In this embodiment, the obtained SHAP value To satisfy local accuracy, the following formula is used for calculation: ; in, This is the expected prediction based on the background data.
[0124] In this embodiment, since accurately calculating the Shapley value requires enumerating all feature subsets and the computational cost increases exponentially with the number of features, it is not feasible for high-dimensional deep models. Therefore, this embodiment can also use an approximation based on the expected gradient: from the background dataset Sampling baseline sample ,along With target sample The linear interpolation path between the inputs is used to integrate the gradient of the model output with respect to the input and take the expectation. It can also be calculated using the following formula: ; in, These are the interpolation coefficients.
[0125] In multi-omics fusion scenarios, SHAP values can be calculated for the input features of the genomic branch and the transcriptome branch respectively.
[0126] In this embodiment, a set of high-contribution gene features or a set of high-contribution loci are selected based on the contribution value. This set can be output as candidate functional genes or candidate loci related to the target trait.
[0127] In this embodiment, the above results can be corroborated by genetic association evidence. For example, genome-wide association analysis can be performed on the target trait for each population: the genotype matrix is standardized and the first five principal components are extracted by principal component analysis (PCA) as fixed-effect covariates to correct for population structure. Then, a single-label general linear model (GLM) is used for site-by-site regression, specifically expressed by the following formula: ; in, For the target phenotypic observations, The additive dose of the tested SNP, , The first Principal components of the genome and their partial regression coefficients The intercept is... The additive effect of the tested SNP, It represents the residual.
[0128] In this embodiment, the Bonferroni threshold can be used. ( , To determine significant loci (based on the number of SNP markers tested), a suggestive threshold is used when no SNPs are detected. Based on this, the consistency between high |SHAP| sites and associated signals is compared, the enrichment of high |SHAP| genes in known functional validation genes (such as flowering regulation-related MADS-box transcription factors, florigen genes, etc.) is assessed, and the co-localization of high contribution features of genomic branches and transcriptomic branches in chromosomal physical locations is evaluated, thereby verifying the biological interpretability of the model predictions.
[0129] The crop multi-omics phenotypic prediction method provided in this invention uses SHAP interpretation analysis to attribute features of deep learning models, which improves the biological interpretability of complex models and can identify key regulatory nodes that drive the formation of specific agronomic traits. This provides accurate and high-value candidate target support for modern crop molecular breeding, the mining of superior gene resources, and targeted gene editing.
[0130] Figure 2 This is the second flowchart of the crop multi-omics phenotypic prediction method provided by the present invention. Figure 2 In the illustrated embodiment, the method further includes the following steps: S1. Obtain genomic, transcriptomic, and target agronomic trait data and perform preprocessing; S2, Genome Branch MGCM Multiscale Feature Extraction, i.e., obtaining the output features of the main branch; S3. Construct gene regulatory networks (WGCNA, GRNBoost2); S4. Transcriptome branch graph neural network feature learning; S5. Multi-omics feature mapping and fusion, that is, obtaining joint feature representation; S6. Output the prediction results, that is, output the phenotypic prediction results; S7 and SHAP interpretation analysis screened candidate functional genes.
[0131] The crop multi-omics phenotyping device provided by the present invention is described below. The crop multi-omics phenotyping device described below can be referred to in correspondence with the crop multi-omics phenotyping method described above.
[0132] Figure 3 This is one of the structural schematic diagrams of the crop multi-omics phenotypic prediction device provided by the present invention, such as... Figure 3 As shown, the device includes: The preprocessing module 310 is used to preprocess the genomic data, transcriptome data and target agronomic trait data of the target crop to obtain genomic feature input information and transcriptome expression matrix; The feature calculation module 320 is used to extract an initial feature map from the genomic feature input information, and perform feature splitting on the initial feature map along the channel dimension to obtain shallow bypass features and main branch input features; perform parallel convolution calculation on the main branch input features at different scales, and perform weighted calculation on the parallel convolution calculation results and corresponding weights to obtain the main branch output features; The feature fusion module 330 is used to splice and map the main branch output features and shallow bypass features to obtain a genome feature representation, and then fuse the genome feature representation with the transcriptome feature representation to obtain a joint feature representation; the transcriptome feature representation is determined based on the transcriptome expression matrix. The prediction module 340 is used to predict the phenotypic information of the target crop based on the joint feature representation and obtain the phenotypic prediction result.
[0133] The crop multi-omics phenotypic prediction device provided in this invention preprocesses the original multi-source data of the target crop to eliminate the differences in scale and format between different omics data; it enhances the feature dimensionality of SNP sequences through nonlinear mapping and feature splitting mechanisms, and obtains non-additive interactions and synergistic effects between SNP sites in different length segments through multi-scale convolution and attention weighting mechanisms, thereby enhancing the ability to characterize complex genetic effects of crops; it provides comprehensive and complementary feature descriptions of complex traits through splicing mapping and cross-modal fusion strategies; and finally, it predicts the phenotypic information of the target crop through joint feature representation, thereby improving the accuracy of crop phenotypic prediction results.
[0134] Figure 4 This is a second schematic diagram of the structure of the crop multi-omics phenotyping prediction device provided by the present invention, as shown below. Figure 4 As shown, the device also includes: a data acquisition and preprocessing module, a genome feature extraction module, a gene regulatory network construction module, a transcriptome feature extraction module, a multi-omics fusion prediction module, and an interpretation and analysis module.
[0135] The data acquisition and preprocessing module is used to acquire crop genome data, transcriptome data, and target agronomic trait data and perform preprocessing; the genome feature extraction module is used to extract local genetic features in SNP sequences using a multi-scale genome context mixing module; the gene regulation network construction module is used to construct a directed weighted gene regulation network by combining WGCNA and GRNBoost2; the transcriptome feature extraction module is used to construct a graph neural network to learn structured expression features constrained by regulatory relationships; the multi-omics fusion prediction module is used to fuse genome features and transcriptome features to model and output the prediction results of target agronomic traits; the interpretation and analysis module is used to identify key gene features and output candidate functional genes based on SHAP interpretation and analysis. Its implementation principle and the resulting technical effects are the same as those in the aforementioned method embodiments, and will not be repeated here.
[0136] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. The processor 510 can call logical instructions in the memory 530 to execute a crop multi-omics phenotypic prediction method. This method includes: preprocessing the genomic data, transcriptome data, and agronomic trait data of the target crop to obtain genomic feature input information and a transcriptome expression matrix; extracting an initial feature map from the genomic feature input information and performing feature splitting along the channel dimension of the initial feature map to obtain shallow bypass features and main branch input features; performing parallel convolution calculations on the main branch input features at different scales and weighting the parallel convolution calculation results and corresponding weights to obtain the main branch output features; performing feature concatenation and mapping on the main branch output features and shallow bypass features to obtain a genomic feature representation, and fusing the genomic feature representation with the transcriptome feature representation to obtain a joint feature representation; determining the transcriptome feature representation based on the transcriptome expression matrix; and predicting the phenotypic information of the target crop based on the joint feature representation to obtain phenotypic prediction results.
[0137] Furthermore, the logical instructions in the aforementioned memory 530 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 the present invention, 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 the present 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.
[0138] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the crop multi-omics phenotypic prediction method provided by the above methods. The method includes: preprocessing the genomic data, transcriptome data, and target agronomic trait data of the target crop to obtain genomic feature input information and transcriptome expression matrix; extracting an initial feature map from the genomic feature input information and performing feature splitting along the channel dimension of the initial feature map to obtain shallow bypass features and main branch input features; performing parallel convolution calculations on the main branch input features at different scales and performing weighted calculations on the parallel convolution calculation results and corresponding weights to obtain main branch output features; performing feature splicing and mapping on the main branch output features and shallow bypass features to obtain a genomic feature representation, and fusing the genomic feature representation with the transcriptome feature representation to obtain a joint feature representation; determining the transcriptome feature representation based on the transcriptome expression matrix; and predicting the phenotypic information of the target crop based on the joint feature representation to obtain phenotypic prediction results.
[0139] 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.
[0140] 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.
[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting crop multi-omics phenotypic traits, characterized in that, include: The genomic data, transcriptome data, and agronomic trait data of the target crop are preprocessed to obtain genomic feature input information and transcriptome expression matrix; An initial feature map is extracted from the genomic feature input information, and the initial feature map is split along the channel dimension to obtain shallow bypass features and main branch input features; The main branch input features are subjected to parallel convolution calculations at different scales, and the results of the parallel convolution calculations and the corresponding weights are weighted to obtain the main branch output features. The main branch output features and the shallow bypass features are spliced and mapped to obtain a genome feature representation, and the genome feature representation is fused with the transcriptome feature representation to obtain a joint feature representation; the transcriptome feature representation is determined based on the transcriptome expression matrix. The phenotypic information of the target crop is predicted based on the joint feature representation to obtain the phenotypic prediction result.
2. The crop multi-omics phenotypic prediction method according to claim 1, characterized in that, The parallel convolution calculation results and corresponding weights are weighted to obtain the main branch output features, including: Basic local features are extracted from the main branch input features, and the basic local features are sequentially processed by pooling, one-dimensional convolution and non-linear activation to obtain attention weights; The parallel convolution calculation results corresponding to different scales are fused with the basic local features to obtain multi-scale fused features; The attention weights are used to perform gated residual weighting on the multi-scale fusion features to obtain the main branch output features.
3. The crop multi-omics phenotypic prediction method according to claim 1, characterized in that, The transcriptome expression matrix includes multiple genes and the corresponding expression levels of each gene; The transcriptome features were obtained through the following steps: The regulatory relationships between different genes are determined based on the transcriptome expression matrix, and a gene regulation network is constructed using the multiple genes as network nodes and the regulatory relationships as network edges. The expression levels of each gene are used as the node features of the corresponding network nodes, and a graph neural network is used to perform structured feature learning along the network edges to obtain the transcriptome feature representation.
4. The crop multi-omics phenotypic prediction method according to claim 3, characterized in that, The multiple genes include transcription factors and target genes; The construction of a gene regulatory network using the plurality of genes as network nodes and the regulatory relationships as network edges includes: The transcriptome expression matrix was processed using weighted gene co-expression network analysis to obtain the search range of candidate regulatory relationships; The directed regulatory relationship between the transcription factor and the target gene is determined based on the search range of the candidate regulatory relationship; The gene regulatory network is obtained by using the feature importance score as the initial weight of the candidate regulatory edges, using the initial weight to perform sparsification screening of the candidate regulatory edges, and then standardizing or monotonically transforming the weights of the screened regulatory edges.
5. The crop multi-omics phenotypic prediction method according to claim 3, characterized in that, The method of learning structured features along the edges of a graph neural network to obtain the transcriptome feature representation includes: Using each network node in the gene regulation network as the target node and the node in the gene regulation network that points to the target node along the corresponding network edge as the source node, the message features are calculated based on the node features of the source node and the weights of the corresponding network edges, and the message features are sent to the target node through the source node. The message features received by the target node from all source nodes are aggregated to obtain neighborhood aggregation features. The neighborhood aggregation features are mapped and transformed to obtain node update terms, and the node features of the target node are updated using the node update terms to obtain the node-level feature representation of the network node. The gating weights of each network node are calculated, and the node-level feature representations of all network nodes are weighted and summed using the gating weights to obtain the transcriptome feature representation.
6. The crop multi-omics phenotypic prediction method according to claim 5, characterized in that, The step of updating the node features of the target node using the node update item to obtain the node-level feature representation of the network node includes: By adding the node features of the target node to the node update term through residual connection, the node-level feature representation of the network node is obtained; Wherein, when the target node does not have a corresponding source node in the gene regulation network, the neighborhood aggregation feature of the target node is configured as a zero vector.
7. The crop multi-omics phenotypic prediction method according to any one of claims 1-6, characterized in that, After obtaining the phenotypic prediction result value, the method further includes: The contribution of each input feature in the genomic feature input information and / or transcriptome expression matrix to the phenotype prediction result is calculated using the interpretive analysis method. Based on the contribution values, a set of high-contribution features is selected, and the set of high-contribution features is used as candidate functional genes related to the target agronomic trait.
8. A crop multi-omics phenotypic prediction device, characterized in that, include: The preprocessing module is used to preprocess the genomic data, transcriptome data and agronomic trait data of the target crop to obtain genomic feature input information and transcriptome expression matrix; The feature calculation module is used to extract an initial feature map from the genomic feature input information and perform feature splitting on the initial feature map along the channel dimension to obtain shallow bypass features and main branch input features; The main branch input features are subjected to parallel convolution calculations at different scales, and the results of the parallel convolution calculations and the corresponding weights are weighted to obtain the main branch output features. The feature fusion module is used to perform feature splicing and mapping on the main branch output features and the shallow bypass features to obtain a genomic feature representation, and to fuse the genomic feature representation with the transcriptome feature representation to obtain a joint feature representation; the transcriptome feature representation is determined based on the transcriptome expression matrix; The prediction module is used to predict the phenotypic information of the target crop based on the joint feature representation, and obtain the phenotypic prediction result.
9. 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 crop multi-omics phenotypic prediction method as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the crop multi-omics phenotypic prediction method as described in any one of claims 1 to 7.
