A collaborative prediction method and system for plant multi-gene stress resistance based on federated graph neural networks

Abstract

This invention relates to the field of collaborative prediction technology, and in particular to a method and system for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks. The method includes: modeling a multilayer heterogeneous graph based on acquired plant genome data; constructing a time-aware dynamic graph neural network model; training the dynamic graph neural network model using the multilayer heterogeneous graph, wherein distributed privacy computation is performed on the multilayer heterogeneous graph using a federated graph neural network; generalizing training on small sample scenario data based on multi-task learning and cross-species transfer mechanisms; and obtaining gene function prediction results. Through multilayer heterogeneous graph fusion technology and time-aware dynamic graph neural network design, this invention can simultaneously capture multi-dimensional biological information such as gene regulation, protein interactions, metabolic pathways, and stress responses, as well as the temporal changes in gene expression during plant stress responses.

Description

Technical Field

[0001] This invention relates to the field of collaborative prediction technology, and in particular to a collaborative prediction method and system for plant multi-gene stress resistance based on federated graph neural networks. Background Technology

[0002] Currently, plant gene function research faces the following major technical bottlenecks: First, traditional gene function prediction methods are mainly based on single-gene omics data (transcriptomics or proteomics), making it difficult to comprehensively capture the complex interactions and multi-level regulatory networks between genes. Second, existing methods mostly employ single-gene analysis strategies, lacking systematic modeling of multi-gene synergistic effects, and thus failing to effectively guide molecular design breeding for complex stress resistance traits. Third, plant gene function prediction heavily relies on large-scale, high-quality training data, but the data scale of individual research institutions is often limited, affecting the generalization ability of the models.

[0003] In addition, major research institutions, seed companies and breeding organizations around the world have accumulated a large amount of valuable plant genomics, transcriptomics and phenomics data. However, due to factors such as intellectual property protection, commercial competition and data privacy, these data cannot be directly shared and integrated, forming a serious "data silo" problem.

[0004] In recent years, graph neural networks have demonstrated great potential in processing complex network structures, effectively learning higher-order interactions between nodes. Federated learning, as an emerging distributed machine learning paradigm, allows multiple participants to collaboratively train machine learning models without sharing the original data, providing a new technical path to address data privacy and data silo issues. However, existing graph neural network methods are primarily designed for homogeneous graphs, making it difficult to handle heterogeneous relationships among various node types, such as gene-protein-metabolite relationships, in plant biological networks. Furthermore, existing federated learning algorithms are mainly geared towards Euclidean data (such as images and text), lacking dedicated federated aggregation mechanisms and privacy protection strategies for graph-structured data.

[0005] Therefore, there is an urgent need to develop a plant stress resistance prediction technology that can integrate heterogeneous biological data from multiple institutions, capture multi-gene synergistic effects, and protect data privacy, so as to provide technical support for global plant molecular breeding cooperation. Summary of the Invention

[0006] To address the technical shortcomings and application limitations of existing plant gene function prediction technologies in handling multi-institutional data integration, multi-gene collaborative modeling, and privacy protection, this invention provides a method and system for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks.

[0007] In a first aspect, the present invention provides a method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks, which adopts the following technical solution:

[0008] A collaborative prediction method for plant multi-gene stress resistance based on federated graph neural networks includes:

[0009] Obtain plant genome data;

[0010] Model a multilayer heterogeneity diagram based on the acquired plant genome data;

[0011] Construct a time-aware dynamic graph neural network model;

[0012] A dynamic graph neural network model is trained using multi-layer heterogeneous graphs. Specifically, a federated graph neural network is used to perform distributed privacy computation on the multi-layer heterogeneous graph. Generalization training is performed on small sample scene data based on multi-task learning and cross-species transfer mechanisms.

[0013] Gene function prediction results were obtained.

[0014] Furthermore, the process of modeling a multilayer heterogeneous graph based on the acquired plant genome data includes defining the plant biological network as a four-layer heterogeneous graph, comprising a gene regulation layer, a protein interaction layer, a metabolic pathway layer, and a stress response layer. For gene nodes in the gene regulation layer, feature vectors are constructed through multi-dimensional information fusion. For regulatory relationship edges in the gene regulation layer... ∈ E 1 The weight is calculated using the formula, expressed as:

[0015] ,

[0016] Among them, the Pearson correlation coefficient Capture genes i and j Linear expression correlation and mutual information between them Detecting nonlinear statistical dependencies and transcription factor binding strength Based on promoter sequence analysis and transcription factor binding site prediction, weighting coefficients , , A data-driven optimization strategy is adopted, and the optimal parameter combination is determined through cross-validation.

[0017] Furthermore, the step of modeling a multilayer heterogeneous graph based on the acquired plant genome data also includes integrating information from each layer of the four-layer heterogeneous graph using an attention-based interlayer fusion strategy. Attention weight Calculated using a self-attention mechanism, it is expressed as:

[0018] ,

[0019] Wherein, the linear transformation matrix Attention vector parameters are used to map node representations from different layers to a unified feature space. The vector concatenation operation (||) is learned through end-to-end training and is used to combine query and key vectors.

[0020] Furthermore, the construction of the time-aware dynamic graph neural network model includes modeling the gene regulatory network of plants during stress response as a time-series graph sequence:

[0021] ,

[0022] Each time point t i Corresponding to a graph snapshot, a spatiotemporal graph convolutional network ST-GCN is constructed, where the network architecture adopts an encoder-decoder structure, and deep feature learning is achieved through the stacking of multiple ST-GCN layers; the forward propagation process of a single ST-GCN layer is as follows:

[0023] ,

[0024] in, For the l The spatiotemporal feature tensor of the layer To process spatial neighborhood information in graph convolution operations, To handle temporal dependencies in temporal convolution operations, , A learnable parameter matrix; graph convolution operation Adopting an improved GAT mechanism:

[0025] ,

[0026] The attention weights are calculated by considering the time-dependent importance of neighborhoods. :

[0027] ,

[0028] in These are time-dependent edge feature vectors.

[0029] Furthermore, the construction of the time-aware dynamic graph neural network model also includes capturing dynamic features at different time scales using a multi-scale time modeling mechanism, wherein the attention weights... The calculation is expressed as:

[0030] A graph attention-based synergistic effect prediction mechanism identifies gene modules that work synergistically under specific stress conditions for a given set of genes. Its collaborative resilience score is calculated as follows:

[0031] ,

[0032] in It's time t Gene interaction weights, It's a gene. i and j In time t The intensity of interaction.

[0033] Furthermore, the distributed privacy computation using a federated graph neural network on multi-layer heterogeneous graphs includes adaptive aggregation based on the graph structure features of each client, learning the similarity of the graph structures of each client through graph embedding, and then calculating aggregation weights based on structural similarity. Specifically, a structural embedding vector is calculated for the graph structure of each client.

[0034] ,

[0035] It is a node i The representation vectors of the last layer GNN are used to calculate the structural similarity matrix between clients:

[0036] ,

[0037] in P k Indicates the client k degree distribution, These are balancing parameters; finally, based on structural similarity, adaptive aggregation weights are calculated.

[0038] ,

[0039] in This is the amount of data from client k. These are the weighting coefficients for structural similarity. The final global model parameter update formula is: .

[0040] Furthermore, the method of using federated graph neural networks to perform distributed privacy computation on multi-layer heterogeneous graphs also includes introducing a differential privacy protection mechanism in the federated learning framework, wherein node-level privacy protection adds Gaussian noise during the node feature update process:

[0041] ,

[0042] noise variance According to privacy budget and sensitivity calculate:

[0043] ;

[0044] Add Laplacian noise to the attention weight calculation:

[0045] ,

[0046] in It is the sensitivity of attention weight. It is a side-level privacy budget; gradient clipping and noise addition are applied when the model parameters are updated:

[0047] ,

[0048] in It is the gradient clipping threshold. It is the gradient noise variance.

[0049] Furthermore, the generalization training based on multi-task learning and cross-species transfer mechanisms for small-sample scenario data includes, for the cross-species generalization problem under small-sample conditions, a hierarchical design is adopted to decompose plant gene function prediction into multiple interrelated sub-tasks. Specifically, a shared representation layer learns general features across species and tasks; a task-specific layer captures unique information for a specific task or species; and a fusion prediction layer integrates multi-level information for final prediction, setting a multi-task set. Each of these tasks For a specific prediction target, for the first i For each task, the loss function is defined as:

[0050] ,

[0051] in These are shared parameters. These are mission-specific parameters. It is the prediction loss for the i-th task. It is the regularization loss of task-specific parameters. This is the cross-task consistency loss; the overall multi-task loss function is balanced by task weights:

[0052] ,

[0053] Task weight A dynamic adjustment strategy is adopted, which adaptively updates the algorithm based on the learning difficulty and importance of each task:

[0054] ,

[0055] in It is the learning rate for weight updates.

[0056] Furthermore, the generalization training of small sample scenario data based on multi-task learning and cross-species transfer mechanisms also includes employing a deep ensemble-based uncertainty quantification method. This involves training multiple models with different initializations to estimate the cognitive and accidental uncertainties in the predictions. for M Each independently trained graph neural network model, for input samples x The mean and variance of the predicted probabilities are as follows:

[0057] ,

[0058] An active learning strategy based on uncertainty selects the most informative samples for annotation:

[0059] ,

[0060] in It is an unlabeled sample set. It is a set of labeled samples. Measuring the sample x The diversity of labeled samples.

[0061] Secondly, a collaborative prediction system for plant multi-gene stress resistance based on federated graph neural networks includes:

[0062] The data acquisition module is configured to acquire plant genome data;

[0063] The heterogeneity graph module is configured to model multilayer heterogeneity graphs based on the acquired plant genome data;

[0064] The model building module is configured to build a time-aware dynamic graph neural network model.

[0065] The model training module is configured to train a dynamic graph neural network model using a multi-layer heterogeneous graph, wherein distributed privacy computation is performed on the multi-layer heterogeneous graph using a federated graph neural network; and generalization training is performed on small sample scene data based on multi-task learning and cross-species transfer mechanisms.

[0066] The prediction module is configured to obtain gene function prediction results.

[0067] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks.

[0068] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks.

[0069] In summary, the present invention has the following beneficial technical effects:

[0070] 1. Improve the accuracy of gene function prediction.

[0071] By employing multilayer heterogeneous graph fusion technology and time-aware dynamic graph neural network design, this invention can simultaneously capture multi-dimensional biological information such as gene regulation, protein interactions, metabolic pathways, and stress responses, as well as the temporal changes in gene expression during plant stress responses.

[0072] 2. Achieve secure and collaborative utilization of data from multiple organizations.

[0073] This invention presents a federated learning framework that addresses the data silo problem in plant genomics. Through a graph-aware federated aggregation algorithm and a differential privacy protection mechanism, research institutions can collaboratively train powerful prediction models without disclosing the original data.

[0074] 3. Reduce the cost and cycle of breeding new varieties.

[0075] Traditional plant breeding relies on extensive field trials and phenotypic identification, typically requiring 5-10 years to develop a new variety. This invention, by accurately predicting the synergistic stress resistance effects of multi-gene combinations, enables rapid computer-based screening of optimal gene-editing target combinations, reducing candidate gene screening time from years to days.

[0076] 4. To achieve effective transfer of gene function knowledge across species.

[0077] The multi-task learning and cross-species transfer mechanism of this invention can effectively transfer gene function knowledge from known species to new or rare species. Through conservation modeling based on evolutionary distance and a meta-learning framework, even under small sample conditions (requiring only 20-50 labeled samples), the gene function prediction accuracy of new species can still achieve relatively ideal results.

[0078] 5. Provides interpretable insights into biological mechanisms.

[0079] Unlike traditional "black box" machine learning methods, the prediction results based on graph neural networks in this invention have good interpretability. Through attention mechanisms and graph structure analysis, the system can identify key regulatory pathways, core gene nodes, and important interactions. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of a plant multi-gene stress resistance collaborative prediction method based on federated graph neural network according to Embodiment 1 of the present invention. Detailed Implementation

[0081] The present invention will be further described in detail below with reference to the accompanying drawings.

[0082] Example 1

[0083] Reference Figure 1 This embodiment of a method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks includes:

[0084] Obtain plant genome data;

[0085] Model a multilayer heterogeneity diagram based on the acquired plant genome data;

[0086] Construct a time-aware dynamic graph neural network model;

[0087] A dynamic graph neural network model is trained using multi-layer heterogeneous graphs. Specifically, a federated graph neural network is used to perform distributed privacy computation on the multi-layer heterogeneous graph. Generalization training is performed on small sample scene data based on multi-task learning and cross-species transfer mechanisms.

[0088] Gene function prediction results were obtained.

[0089] Specifically:

[0090] S1 multi-layer heterogeneous graph construction,

[0091] To address the challenge of unified modeling of heterogeneous biological data, this invention proposes a multilayer heterogeneous graph fusion technique. The core idea of ​​this technique is to decompose the complex biological network of plants into four functional levels, and then construct a unified multilayer heterogeneous graph representation through an interlayer fusion mechanism. This hierarchical modeling strategy not only preserves the inherent structural features of various biological information types but also captures the correlations between different biological processes through cross-layer connections, thus laying a solid foundation for modeling multi-gene synergistic effects.

[0092] Traditional plant gene network analysis methods often simply merge different types of biological data or use only a single data source, neglecting the hierarchical characteristics of biological systems and cross-layer interactions. For example, gene regulation is mainly reflected at the transcriptional level, while protein interactions occur at the post-translational level; metabolic pathways reflect biochemical response networks, and stress responses involve coordination at multiple levels. This invention, by constructing a multi-layer heterogeneous map, can systematically integrate this multi-scale information while preserving the biological significance of each layer.

[0093] S1.1 Design principles and structural definition of multilayer heterogeneous diagrams,

[0094] The multilayer heterogeneous graph designed in this invention fully considers the hierarchical characteristics of plant biological systems. Based on an in-depth analysis of plant stress resistance mechanisms, we decompose the complex biological network into four interconnected but functionally independent layers. This hierarchical design not only conforms to biological principles but also facilitates subsequent graph neural network processing. The plant biological network is defined as a four-layer heterogeneous graph. G ={ G 1, G 2, G 3, G 4}.

[0095] G 1=( V 1, E 1) Represents the gene regulatory layer. V 1 represents a set of gene nodes, containing transcription factor genes and target genes. E 1 represents the set of regulatory relationships, indicating the direct or indirect regulatory relationship between transcription factors and target genes. G 2=( V 2, E 2) Represents the protein-protein interaction layer, reflecting molecular interactions at the post-translational level. V 2 represents a collection of protein nodes, containing various functional proteins such as enzymes, receptors, and signaling molecules. E 2 represents the set of interaction edges, including protein-protein direct binding, enzyme-substrate interaction, signal transduction cascade, and other relationships. G 3=( V 3, E 3) Represents the metabolic pathway layer, reflecting the metabolic reprogramming process of plants under stress conditions. V 3 represents the set of metabolite nodes, including primary metabolites, secondary metabolites, and signaling molecules. E 3 represents the biochemical reaction edge set, indicating enzyme-catalyzed reactions, substrate-product transformation relationships, etc. G 4=( V 4, E 4) Represents the stress response layer, specifically capturing the specific response patterns of plants under abiotic stress conditions. V 4 represents a set of stress-responsive genes, including genes involved in processes such as ABA signal transduction, osmotic regulation, and reactive oxygen species scavenging. E 4 represents the set of co-expression relationship edges, constructed based on transcriptome data under different stress conditions.

[0096] Inter-layer connections are achieved through cross-layer edge sets. ={ E 12 , E 13, E 14 , E 23 , E 24 , E 34 This indicates that they are able to capture causal relationships and regulatory cascades between different biological levels. For example, E 12 Connecting genes and the proteins they encode. E 23 Ligases and the metabolic reactions they catalyze. E 14 Connecting regulatory genes and stress response genes, etc.

[0097] S1.2 Multi-dimensional construction strategy for node features

[0098] This invention designs a comprehensive feature construction strategy targeting the multidimensional characteristics of plant genes, which can fully utilize sequence information, expression information, and functional annotation information. For gene nodes in the gene regulatory layer... ∈ V 1 Its feature vector is constructed through the fusion of multi-dimensional information:

[0099] ,

[0100] Among them, sequence features Structural information of gene sequences was extracted using k-mer encoding technology; expression characteristics were also analyzed. Includes gene expression values ​​under different stress conditions, at different time points, and in different tissues; functional annotation features. It represents the biological processes, molecular functions, and cellular components in which genes participate.

[0101] S1.3 Diversified calculation mechanism for edge weights,

[0102] This invention designs a diversified edge weighting mechanism that comprehensively considers linear correlation, nonlinear relationships, and prior biological knowledge. For regulatory relationships in gene regulatory layers... ∈ E 1 Its weight is calculated using the following formula:

[0103] ,

[0104] Among them, the Pearson correlation coefficient Capture genes i and j Linear expression correlation and mutual information between them Detecting nonlinear statistical dependencies and transcription factor binding strength Based on promoter sequence analysis and transcription factor binding site prediction, it provides supporting biological evidence. Weighting coefficients. , , A data-driven optimization strategy is adopted, and the optimal parameter combination is determined through cross-validation.

[0105] S1.4 Attention mechanism for inter-layer information fusion

[0106] This invention designs an inter-layer fusion strategy based on an attention mechanism, which can adaptively learn the importance weights of information at different levels, thereby achieving effective integration of information from each layer:

[0107] ,

[0108] Attention weight Calculation using a self-attention mechanism:

[0109] ,

[0110] Wherein, the linear transformation matrix Attention vector parameters are used to map node representations from different layers to a unified feature space. The vector concatenation operation (||) is learned through end-to-end training and is used to combine query and key vectors.

[0111] This multilayer heterogeneous graph construction technology integrates various biological information, including gene regulation, protein-protein interactions, metabolic pathways, and stress responses, providing a unified and rich data representation framework for subsequent dynamic graph neural network analysis. It not only preserves the intrinsic characteristics of various biological information types but also achieves effective integration of multi-scale information through cross-layer connections and attention fusion mechanisms.

[0112] S2 Time-Aware Dynamic Graph Neural Network Design

[0113] To address the dynamic modeling problem of multi-gene synergistic effects, this invention designs a time-aware dynamic graph neural network architecture. This architecture can simultaneously capture the static topology and dynamic evolutionary features of gene regulatory networks, particularly the temporal changes in gene expression patterns and network reconstruction phenomena during plant stress responses. Plant stress response is a highly dynamic biological process, and the regulatory relationships between genes change significantly with the duration of stress stimulation. For example, in the early stage of alkaline stress (1-3 hours), stress response genes are mainly activated; in the middle stage (6-12 hours), genes related to osmotic regulation and ion balance are activated; and in the late stage (24 hours and beyond), long-term adaptive genes are activated. This temporal gene regulatory network reconstruction phenomenon is an important characteristic of plant stress resistance and a key mechanism for the generation of multi-gene synergistic effects.

[0114] S2.1 Construction and Representation of Time Series Diagrams

[0115] This invention models the gene regulatory network of plants during stress response as a time-series sequence. each time point t i Each time point corresponds to a snapshot. This paper sets T=5, corresponding to key time points such as 0h (control), 1h, 3h, 6h, 12h, and 24h. Graphs for each time point are shown below. G (t) =( V , E (t) , X (t) It has the following characteristics: the set of nodes V remains unchanged at all time points, ensuring the consistency of the network structure; the set of edges... Dynamically adjusted based on gene co-expression relationships at that time point; node feature matrix. Includes gene expression values ​​and cumulative changes at that time point. Time point t The edge weight calculation takes into account the temporal changes in the expression of relevance:

[0116] ,

[0117] in, The correlation coefficient representing gene expression at time point t. It is the distance correlation coefficient, which captures non-linear relationships. It is a time-delay correlation, and the time lag effect of regulation should be considered.

[0118] S2.2 Spatiotemporal Graph Convolutional Network Architecture

[0119] The Spatiotemporal Graph Convolutional Network (ST-GCN) designed in this invention can simultaneously process the spatial structure information and temporal evolution information of graphs. The network architecture adopts an encoder-decoder structure, achieving deep feature learning through the stacking of multiple ST-GCN layers. The forward propagation process of a single ST-GCN layer is defined as follows:

[0120] ,

[0121] in, For the l The spatiotemporal feature tensor of the layer To process spatial neighborhood information in graph convolution operations, To handle temporal dependencies in temporal convolution operations, , is a learnable parameter matrix.

[0122] Graph convolution operations An improved GAT mechanism is adopted:

[0123] ,

[0124] Attention weight Time-dependent neighborhood importance was considered:

[0125] ,

[0126] in It is a time-dependent edge feature vector, containing information such as edge weight, regulation type, and time decay factor.

[0127] S2.3 The mechanism for fusing temporal convolution and long short-term memory.

[0128] Temporal convolution operation By employing a causal convolutional design, we ensure that only historical information is used during prediction:

[0129] ,

[0130] Where K is the size of the temporal convolution kernel. This is the weight matrix corresponding to the time offset s. To better capture long-term time dependencies, this invention introduces a gated recurrent unit mechanism:

[0131] ,

[0132] in, , To reset and update the door, In the candidate hidden state, , , These are learnable parameters.

[0133] S2.4 Multi-scale temporal modeling and attention pooling

[0134] Plant stress responses involve biological processes across multiple timescales, ranging from minute-level signal transduction to hourly gene expression changes and day-level physiological adaptations. This invention designs a multi-scale temporal modeling mechanism capable of capturing dynamic features at different timescales. These multi-scale temporal features are extracted using time windows of varying sizes.

[0135] ,

[0136] in Corresponding to different time scales. Multi-scale features are fused through an adaptive attention mechanism:

[0137] ,

[0138] Attention weight Calculated in the following way:

[0139] ,

[0140] S2.5 Synergistic Effect Modeling and Prediction Mechanism

[0141] To achieve multi-gene synergistic effects, this invention designs a graph attention-based synergistic effect prediction mechanism, capable of identifying gene modules that function synergistically under specific stress conditions. Given a set of genes... Its collaborative resilience score is calculated as follows:

[0142] ,

[0143] in It's time t Gene interaction weights, It's a gene. i and j In time t The interaction strength is calculated using the learned node representations.

[0144] ,

[0145] This time-aware dynamic graph neural network design captures the spatiotemporal evolution characteristics of gene regulatory networks, providing strong technical support for the accurate prediction of multi-gene synergistic stress resistance effects. The modular design of the network architecture enables it to adapt to different stress types and time scales, exhibiting good scalability and generalization capabilities.

[0146] S3 Federated Learning Framework and Privacy Protection Mechanism

[0147] To address the privacy concerns of collaborative data utilization across multiple institutions and the challenges of federated learning aggregation of graph-structured data, this invention designs a federated learning framework specifically for graph neural networks. This framework proposes a graph-structure-aware federated aggregation algorithm and a differential privacy protection mechanism, enabling effective collaborative modeling of plant genetic data from multiple institutions while ensuring data privacy and security.

[0148] Traditional federated learning algorithms are primarily designed for independent and identically distributed (IID) Euclidean data, such as image classification and text processing tasks. However, plant gene network data has a complex graph topology, and data from different institutions vary significantly in graph size, node type, and edge connection patterns, exhibiting typical non-independent and identically distributed (Non-IID) characteristics.

[0149] Furthermore, plant genetic data has significant scientific research value, involving sensitive information such as germplasm resources, breeding techniques, and intellectual property. Existing federated learning privacy protection mechanisms are primarily designed for vector data and cannot effectively protect node privacy, edge privacy, and subgraph structure privacy in graph-structured data. Therefore, a new federated learning framework and privacy protection mechanism specifically tailored to the characteristics of plant genetic network data is needed.

[0150] S3.1 Overall architecture design of federated graph neural networks,

[0151] The federated graph neural network system designed in this invention adopts a parameter server architecture, comprising a central aggregation server and multiple clients (participating institutions). Each institution's raw data is always stored locally, and only privacy-preserving model parameters are transmitted over the network. The system's training process includes the following key stages:

[0152] Initialization Phase: The central server initializes the global graph neural network model. , set initial parameters Distribute to all participating clients. Simultaneously, the server maintains a global graph structure template. This is used to standardize the graph representation format across different clients. Local training phase: [The following is a separate, unrelated section:] k Each client is based on local data Train a local model. The objective function for local training is designed as follows:

[0153] ,

[0154] in The main task loss, It is the graph structure regularization loss. It is the parameter regularization loss. and It is the balance coefficient.

[0155] Aggregation update phase: Each client updates the model after privacy protection processing. The data is uploaded to the central server, where the server executes a graph-aware federated aggregation algorithm to update the global model parameters.

[0156] S3.2 Graph Structure Aware Federated Aggregation Algorithm

[0157] The heterogeneity of plant gene networks makes traditional federated averaging algorithms ineffective. This invention proposes a graph structure-aware federated aggregation algorithm that can adaptively aggregate based on the graph structure features of each client. The core idea of ​​the algorithm is to learn the similarity of the graph structures of each client through graph embedding technology, and then calculate the aggregation weight based on the structural similarity. The specific algorithm flow is as follows. First, calculate the structure embedding vector for the graph structure of each client:

[0158] ,

[0159] in It is a node i The representation vectors of the last GNN layer are used to obtain graph-level embeddings through average pooling. Then, the structural similarity matrix between clients is calculated:

[0160] ,

[0161] The first term is cosine similarity, which measures the similarity of structural embeddings. The second term is a KL divergence penalty term based on the degree distribution. P k Indicates client k degree distribution, These are balancing parameters. Based on structural similarity, adaptive aggregation weights are calculated:

[0162] ,

[0163] in This is the amount of data from client k. These are the weighting coefficients for structural similarity. The final global model parameter update formula is:

[0164] ,

[0165] S3.3 Differential Privacy-Preserving Graph Neural Network Design

[0166] To protect the privacy of plant genetic data, this invention introduces a differential privacy protection mechanism specifically designed for graph neural networks within a federated learning framework. This mechanism protects not only node-level privacy (such as the expression values ​​of specific genes), but also edge-level privacy (such as regulatory relationships between genes) and subgraph-level privacy (such as specific regulatory modules). For graph neural networks, this invention adds privacy protection at multiple levels. Node-level privacy protection involves adding Gaussian noise during node feature updates:

[0167] ,

[0168] Wherein noise variance According to privacy budget and sensitivity calculate:

[0169] ,

[0170] Side-level privacy protection: Adding Laplacian noise to attention weight calculation:

[0171] ,

[0172] in It is the sensitivity of attention weight. This is a borderline privacy budget. Gradient-level privacy protection: applies gradient clipping and noise addition during model parameter updates.

[0173] ,

[0174] in It is the gradient clipping threshold. It is the gradient noise variance.

[0175] S3.4 Privacy Budget Allocation and Optimization Strategies

[0176] This invention proposes a privacy budget allocation strategy based on importance sampling, prioritizing the protection of parameters and structures that have a significant impact on model performance. Total privacy budget. Distribute in the following manner:

[0177] ,

[0178] Budget allocation at each level is based on sensitivity analysis:

[0179] ,

[0180] The sensitivity is estimated using the influence function:

[0181] ,

[0182] S3.5 Secure Aggregation and Communication Optimization

[0183] To further enhance system security and efficiency, this invention utilizes a secure multi-party computation protocol for parameter aggregation and communication compression technology to reduce network transmission overhead. The secure aggregation protocol employs a secret-sharing-based method, decomposing each client's model update into multiple secret fragments:

[0184] ,

[0185] in Is sent to the j The secret fragment of an aggregate node satisfies:

[0186] ,

[0187] Communication compression employs an importance-based sparsity strategy:

[0188] ,

[0189] in It's the compression ratio. TopK Operations retain the one with the largest absolute value Proportional parameters.

[0190] This federated learning framework provides a complete technical solution for privacy-preserving collaborative modeling of plant genetic data through graph structure-aware aggregation algorithms, multi-level differential privacy protection, and secure communication mechanisms. The modular design of the framework enables it to adapt to multi-institutional collaboration scenarios of different scales and types.

[0191] S4 multi-task learning and cross-species migration mechanism

[0192] To address the problem of cross-species generalization under small sample conditions, this invention designs a multi-task learning and cross-species transfer mechanism. The core of this mechanism lies in achieving effective transfer of gene function knowledge between different species through a meta-learning framework, and improving the model's prediction accuracy and generalization ability in small sample scenarios through multi-task joint training.

[0193] Plant gene functions exhibit significant conservation and species specificity during evolution. On the one hand, homologous genes often have similar functions across different species; for example, stress-resistance-related transcription factor families are prevalent in legumes. On the other hand, different species have developed unique adaptive mechanisms over long periods of evolution, such as the salt gland structures and related genes unique to halophytes. Traditional machine learning methods often process data for each species independently, failing to effectively utilize cross-species functional conservation information, resulting in a significant decrease in prediction accuracy when facing new species or rare stress conditions.

[0194] S4.1 hierarchical multi-task learning architecture,

[0195] The multi-task learning architecture proposed in this invention adopts a hierarchical design, decomposing plant gene function prediction into multiple interrelated sub-tasks. This design can capture common features among different tasks while preserving the specific information of each task. The architecture consists of three layers: a shared representation layer, a task-specific layer, and a fusion prediction layer. The shared representation layer learns general features across species and tasks; the task-specific layer captures unique information for a specific task or species; and the fusion prediction layer integrates multi-layered information for the final prediction.

[0196] Set up a multi-task set Each of these tasks This corresponds to specific prediction objectives, such as salt tolerance prediction, drought tolerance prediction, and alkali tolerance prediction. For the first... i For each task, the loss function is defined as:

[0197] ,

[0198] in These are shared parameters. These are mission-specific parameters. It is the prediction loss for the i-th task. It is the regularization loss of task-specific parameters. This is the cross-task consistency loss. The overall multi-task loss function is balanced by task weights:

[0199] ,

[0200] Task weight A dynamic adjustment strategy is adopted, which adaptively updates the algorithm based on the learning difficulty and importance of each task:

[0201] ,

[0202] in It is the learning rate for weight updates.

[0203] S4.2 A cross-species transfer framework based on meta-learning

[0204] To achieve effective cross-species knowledge transfer, this invention introduces a model-independent meta-learning framework. The goal of this framework is to learn a good model initialization, enabling the model to quickly adapt to new species or tasks with a limited number of samples. The meta-learning training process consists of two layers: an inner loop and an outer loop. The inner loop simulates the rapid adaptation process on new tasks, while the outer loop optimizes the adaptation capability.

[0205] Suppose support set and query set Representing tasks A small number of labeled and test samples. Inner loop update process:

[0206] ,

[0207] in It is the inner loop learning rate. The parameter is The graph neural network model. The outer loop updates the performance of the adapted parameters on the query set:

[0208] ,

[0209] in It is the outer loop learning rate. N It represents the number of training tasks.

[0210] S4.3 Modeling of gene function conservation among species,

[0211] To better utilize cross-species functional conservation information, this invention designs a conservation modeling mechanism based on evolutionary distance. This mechanism guides the knowledge transfer process through phylogenetic tree information, enabling species with closer evolutionary relationships to share more functional information. Let the set of species be... The evolutionary distance matrix between species is ,in Indicates species and The evolutionary distance between species. Based on the evolutionary distance, the similarity weight between species is calculated:

[0212] ,

[0213] in It is a distance-sensitive parameter. Cross-species functional prediction is achieved through weighted ensembles:

[0214] ,

[0215] in Based on species The probability that a model trained on data predicts for sample x.

[0216] S4.4 Domain Adaptive and Adversarial Training Mechanism

[0217] To mitigate the impact of species-specific domain differences, this invention introduces a domain adaptation mechanism based on adversarial training. This mechanism enables the feature extractor to learn species-independent universal representations through adversarial learning. The domain adaptation module contains the feature extractor. and Domain Discriminator The feature extractor aims to learn species-independent gene representations, while the domain discriminator attempts to determine the species from which a sample originates based on these feature representations. The objective function for adversarial training is:

[0218] ,

[0219] At the same time, in order to maintain the performance of task prediction, it is also necessary to minimize the task loss:

[0220] ,

[0221] in It is a task classifier. This is the classification loss function. The overall training objective is:

[0222] ,

[0223] S4.5 Uncertainty Quantification and Active Learning Strategies

[0224] In few-shot learning scenarios, quantifying the predictive uncertainty of a model is crucial for guiding active learning and improving model reliability. This invention employs a deep ensemble-based uncertainty quantification method, estimating the cognitive and accidental uncertainties of predictions by training multiple models with different initializations. Let... for M Each independently trained graph neural network model, for input samples x The mean and variance of the predicted probabilities are as follows:

[0225] ,

[0226] An active learning strategy based on uncertainty selects the most informative samples for annotation:

[0227] ,

[0228] in It is an unlabeled sample set. It is a set of labeled samples. Measuring the sample x The diversity of labeled samples.

[0229] Example 2

[0230] This embodiment provides a plant multi-gene stress resistance collaborative prediction system based on federated graph neural networks, including:

[0231] The data acquisition module is configured as follows:

[0232] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks.

[0233] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned plant multi-gene stress resistance collaborative prediction method based on federated graph neural networks.

[0234] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks, characterized in that, include: Obtain plant genome data; Model a multilayer heterogeneity diagram based on the acquired plant genome data; This includes defining the plant biological network as a four-layer heteromorphic graph and using an attention-based interlayer fusion strategy to integrate information from each layer of the four-layer heteromorphic graph; the four-layer heteromorphic graph includes a gene regulation layer, a protein interaction layer, a metabolic pathway layer, and a stress response layer; Constructing a time-aware dynamic graph neural network model; including modeling the gene regulatory network of plants during stress response as a time-series graph sequence. each time point t i For each graph snapshot, a spatiotemporal graph convolutional network ST-GCN is constructed, where the network architecture adopts an encoder-decoder structure, and deep feature learning is achieved by stacking multiple ST-GCN layers; It also includes using multi-scale time modeling mechanisms to capture dynamic features at different time scales; A graph attention-based synergistic effect prediction mechanism identifies gene modules that function synergistically under specific stress conditions; A dynamic graph neural network model is trained using multi-layer heterogeneous graphs. Specifically, federated graph neural networks are used to perform distributed privacy computation on multi-layer heterogeneous graphs. Generalization training is performed on small sample scenario data based on multi-task learning and cross-species transfer mechanisms. Gene function prediction results were obtained.

2. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 1, characterized in that, For gene nodes in the gene regulation layer, feature vectors are constructed through multi-dimensional information fusion; for regulatory relationship edges in the gene regulation layer... ∈ E 1 The weight is calculated using the formula, expressed as: ， in, i and j These represent two different gene nodes in the gene regulatory layer; Pearson correlation coefficient Capture genes i and j Linear expression correlation and mutual information between them Detecting nonlinear statistical dependencies and transcription factor binding strength Based on promoter sequence analysis and transcription factor binding site prediction, weighting coefficients , , A data-driven optimization strategy is adopted, and the optimal parameter combination is determined through cross-validation.

3. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 2, characterized in that, The method of integrating information from each layer of a four-layer heterogeneous graph using an attention-based inter-layer fusion strategy is represented as follows: Attention weight Calculated using a self-attention mechanism, it is expressed as: ， Where k represents the k-th layer, u is the node index, v is the layer index taking values ​​from 1 to 4, and the linear transformation matrix is... It is used to map node representations from different layers to a unified feature space. Indicates the dimensions of the output and input; attention vector parameters The vector concatenation operation (||) is learned through end-to-end training and is used to combine query and key vectors. Indicates the first k The original feature vector of the u-th node in the layer; Let be the updated feature vector of node u at layer l.

4. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 3, characterized in that, The forward propagation process of a single ST-GCN layer is as follows: ， in, For the first l The spatiotemporal feature tensor of the layer To process spatial neighborhood information in graph convolution operations, To handle temporal dependencies in temporal convolution operations, , A learnable parameter matrix; graph convolution operation Adopting an improved GAT mechanism , This represents the set of neighboring nodes of node p at time t, where the attention weights are calculated by considering the temporal importance of the neighborhood. : ， This represents the attention weight vector in the graph attention mechanism. Let L represent the learnable linear matrix used for node feature transformation in the l-th network layer. express t At the moment of l Nodes in a layered network z eigenvectors; express t At this moment l Nodes in a layered network p eigenvectors; These are time-dependent edge feature vectors.

5. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 4, characterized in that, The method of capturing dynamic features at different time scales using a multi-scale time modeling mechanism includes extracting multi-scale time features through time windows of different sizes: ， in Corresponding to different time scales This represents a one-dimensional convolution operation performed along the time axis. Multi-scale features are fused through an adaptive attention mechanism: ， Attention weight Calculated in the following way: ； in This represents the learnable weight matrix used for transforming scale features. Transpose of the learnable attention context vector used to compute attention scores; The bias vector is represented; the graph attention-based synergistic effect prediction mechanism identifies gene modules that synergistically function under specific stress conditions, and does so in the following manner: for a given set of genes Its collaborative resilience score is calculated as follows: ， in It is time t Gene interaction weights, It's a gene. and In time t The intensity of interaction.

6. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 5, characterized in that, The method of using federated graph neural networks to perform distributed privacy computation on multi-layer heterogeneous graphs includes adaptive aggregation based on the graph structure features of each client, learning the similarity of the graph structures of each client through graph embedding, and then calculating aggregation weights based on structural similarity. Specifically, a structural embedding vector is calculated for the graph structure of each client. ， Indicates client c 1. Graph data; Representation diagram The set of nodes; It is a node n The representation vectors of the last layer GNN are used to calculate the structural similarity matrix between clients: ， in Indicates the cth * Graph structure embedding vectors for each client; and Let these represent the degree distributions of the graphs for clients c1 and c2, respectively. These are balancing parameters; finally, based on structural similarity, adaptive aggregation weights are calculated. ， in, It is the dataset of client c. Given its data volume, C is the weight coefficient for structural similarity, M is the total number of clients, and the final global model parameter update formula is: 。 7. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 6, characterized in that, The method of using federated graph neural networks to perform distributed privacy computation on multi-layer heterogeneous graphs also includes introducing a differential privacy protection mechanism into the federated learning framework, wherein Gaussian noise is added during the node feature update process to protect node privacy. ， Represents all nodes directly connected to node i; noise variance According to privacy budget and sensitivity calculate: ； The relaxation parameter in differential privacy represents the maximum probability of allowing leakage; Laplace noise is added in the attention weight calculation. ， in It is the sensitivity of attention weight. It is a side-level privacy budget; gradient clipping and noise addition are applied when the model parameters are updated: ， in, This represents the original gradient of client c. yes L2 norm, This represents the gradient of client c after privacy protection processing, where T is the gradient clipping threshold. It is the gradient noise variance; Represents the identity matrix.

8. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 7, characterized in that, The method for generalization training on small-sample scenario data based on multi-task learning and cross-species transfer mechanisms includes a hierarchical design for cross-species generalization under small-sample conditions. This design decomposes plant gene function prediction into multiple interrelated sub-tasks. Specifically, a shared representation layer learns general features across species and tasks; a task-specific layer captures unique information for a specific task or species; and a fusion prediction layer integrates multi-level information for final prediction, setting a multi-task set. Each of these tasks For a specific prediction target, for the first task_idx For each task, the loss function is defined as: ， in, Indicates the first task_idx The regularization weight coefficient for each task; Indicates the first task_ idx Consistency constraint weights for each task; These are shared parameters. These are mission-specific parameters. It is the first task_idx Predicted loss for each task, It is the regularization loss of task-specific parameters. This is the cross-task consistency loss; the overall multi-task loss function is balanced by task weights: ， Among them, task weight A dynamic adjustment strategy is adopted, which adaptively updates the algorithm based on the learning difficulty and importance of each task: ， in It is the learning rate for weight updates. Indicates the first iter Task during the next iteration task_idx The loss value.

9. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural networks according to claim 8, characterized in that, The generalization training based on multi-task learning and cross-species transfer mechanisms for small sample scenario data also includes employing a deep ensemble-based uncertainty quantification method. This method estimates the cognitive and accidental uncertainties of predictions by training multiple models with different initializations. for M Each independently trained graph neural network model, for input samples x The mean and variance of the predicted probabilities are as follows: ， Indicates by parameters Defined model f For input samples x The predicted output; an active learning strategy based on uncertainty selects the most informative samples for annotation: ， in This represents the optimal sample selected during active learning; Indicates diversity weight; It is an unlabeled sample set. It is a set of labeled samples. Measuring the sample x The diversity of labeled samples.

10. A plant multi-gene stress resistance collaborative prediction system based on federated graph neural networks, executing the plant multi-gene stress resistance collaborative prediction method based on federated graph neural networks as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire plant genome data; The heterogeneity graph module is configured to model multilayer heterogeneity graphs based on the acquired plant genome data; The model building module is configured to build a time-aware dynamic graph neural network model. The model training module is configured to train a dynamic graph neural network model using a multi-layer heterogeneous graph, wherein distributed privacy computation is performed on the multi-layer heterogeneous graph using a federated graph neural network; and generalization training is performed on small sample scene data based on multi-task learning and cross-species transfer mechanisms. The prediction module is configured to obtain gene function prediction results.

Images