Plant polygene stress resistance collaborative prediction method and system based on federal map neural network

Abstract

The invention relates to the technical field of collaborative prediction, in particular to a plant polygene stress resistance collaborative prediction method and system based on a federal map neural network. The method comprises the following steps: modeling a multilayer heterogeneous graph according to acquired plant genome data; constructing a dynamic graph neural network model based on time sequence perception; training a dynamic graph neural network model by using the multi-layer heterogeneous graph, and performing distributed privacy calculation on the multi-layer heterogeneous graph by using a federated graph neural network; generalization training is carried out on the small sample scene data based on multi-task learning and a cross-species migration mechanism; and obtaining a gene function prediction result. Through a multi-layer heterogeneous graph fusion technology and a dynamic graph neural network design of time sequence perception, multi-dimensional biological information such as gene regulation, protein interaction, metabolic pathways and stress response can be captured at the same time, and time sequence change characteristics of gene expression in the plant stress response process can be captured at the same time.

Description

Technical Field

[0001] The present invention relates to the field of collaborative prediction technology, and in particular to a method and system for collaborative prediction of plant polygenic stress resistance based on a federated graph neural network. Background Art

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

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

[0004] In recent years, graph neural networks have demonstrated tremendous potential in processing complex network data, effectively learning high-order interactions between nodes. Federated learning, an emerging distributed machine learning paradigm, allows multiple participants to collaboratively train machine learning models without sharing raw data, providing a new technical approach to addressing data privacy and data silos. However, existing graph neural network methods are primarily designed for homogeneous graphs and struggle to handle heterogeneous relationships among multiple node types, such as those found in plant biological networks. Furthermore, existing federated learning algorithms primarily target Euclidean data (such as images and text) and lack specialized 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 the synergistic effects of multiple genes, and protect data privacy, so as to provide technical support for global plant molecular breeding cooperation. Summary of the Invention

[0006] In order to solve the above-mentioned technical defects and application limitations of existing plant gene function prediction technologies in processing multi-institutional data integration, multi-gene collaborative modeling and privacy protection, the present invention provides a plant multi-gene stress resistance collaborative prediction method and system based on a federated graph neural network.

[0007] In a first aspect, the present invention provides a method for collaboratively predicting plant polygenic stress resistance based on a federated graph neural network, which adopts the following technical solutions: A collaborative prediction method for plant polygenic stress resistance based on federated graph neural networks, including: Access to plant genome data; Modeling multi-layer heterogeneous graphs based on acquired plant genome data; Build a dynamic graph neural network model based on time series perception; Using multi-layer heterogeneous graphs to train dynamic graph neural network models, including using federated graph neural networks to perform distributed privacy computing on multi-layer heterogeneous graphs; generalization training for small sample scenario data based on multi-task learning and cross-species migration mechanisms; Obtain gene function prediction results.

[0008] Furthermore, the method of modeling a multi-layer heterogeneous graph based on the obtained plant genome data includes defining the plant biological network as a four-layer heterogeneous graph, including a gene regulation layer, a protein interaction layer, a metabolic pathway layer, and a stress response layer. For the gene nodes in the gene regulation layer, a feature vector is constructed by multi-dimensional information fusion. For the regulatory relationship edges in the gene regulation layer, a feature vector is constructed by multi-dimensional information fusion. ∈ E 1 , and its weight is calculated by the formula, which is expressed as: , Among them, the Pearson correlation coefficient Capturing genes i and j The linear expression correlation between Detecting nonlinear statistical dependencies in transcription factor binding strength Based on promoter sequence analysis and transcription factor binding site prediction, the weight coefficient , , A data-driven optimization strategy is adopted to determine the optimal parameter combination through cross-validation.

[0009] Furthermore, the method of modeling a multi-layer heterogeneous graph based on the obtained plant genome data also includes integrating the information of each layer of the four-layer heterogeneous graph using an inter-layer fusion strategy based on an attention mechanism. , attention weight The calculation is done using the self-attention mechanism, which is expressed as: , Among them, the linear transformation matrix Used to map node representations of different layers to a unified feature space, attention vector parameters The vector concatenation operation || is learned through end-to-end training and is used to combine the query and key vectors.

[0010] Furthermore, the construction of a dynamic graph neural network model based on time series perception includes modeling the gene regulatory network of the plant during the stress response process as a time series graph sequence: , Each time point t i Corresponding to a graph snapshot, a spatiotemporal graph convolutional network ST-GCN is constructed. The network architecture adopts an encoder-decoder structure and achieves deep feature learning by stacking multiple ST-GCN layers. The forward propagation process of a single ST-GCN layer is: , in, For the l The spatiotemporal feature tensor of the layer, Processing spatial neighborhood information for graph convolution operations, Handle timing dependencies for temporal convolution operations, , is a learnable parameter matrix; graph convolution operation Adopting improved GAT mechanism: , The attention weight is calculated by considering the time-related neighborhood importance : , in is the time-dependent edge eigenvector.

[0011] Furthermore, the construction of a dynamic graph neural network model based on time series perception also includes using a multi-scale time modeling mechanism to capture dynamic features of different time scales, where the attention weight The calculation is expressed as: , the synergistic effect prediction mechanism based on graph attention identifies gene modules that work synergistically under specific stress conditions. For a given gene set , and its synergistic stress resistance score is calculated as: , in It's time t The gene interaction weights, It's genes i and j In time t The interaction intensity.

[0012] Furthermore, the distributed privacy computing of multi-layer heterogeneous graphs using federated graph neural networks includes adaptive aggregation based on the graph structure characteristics of each client, learning the similarity of the graph structures of each client through graph embedding, and then calculating the aggregation weight based on the structural similarity. The structural embedding vector is calculated for the graph structure of each client: , is a node i The representation vector of the last layer of GNN is used to calculate the structural similarity matrix between clients: , in P k Represents the client k The degree distribution of is the balance parameter, and finally based on the structural similarity, the adaptive aggregation weight is calculated: , in is the amount of data from client k, is the weight coefficient of structural similarity, and the final global model parameter update formula is: .

[0013] Furthermore, the distributed privacy computing on multi-layer heterogeneous graphs using federated graph neural networks also includes the introduction of a differential privacy protection mechanism in the federated learning framework, where node-level privacy protection adds Gaussian noise during the node feature update process: , Noise variance According to the privacy budget and sensitivity calculate: ; Add Laplacian noise to the attention weight calculation: , in is the sensitivity of the attention weight, is the edge-level privacy budget; gradient clipping and noise addition are applied when the model parameters are updated: , in is the gradient clipping threshold, is the gradient noise variance.

[0014] Furthermore, the generalization training of small sample scenario data based on multi-task learning and cross-species migration mechanism includes the cross-species generalization problem under small sample conditions, and the hierarchical design is used to decompose the plant gene function prediction into multiple interrelated subtasks, among which the shared representation layer learns the common features across species and tasks; the task-specific layer captures the unique information of a specific task or species; the fusion prediction layer integrates multi-level information for the final prediction, and sets a multi-task set. Each of these tasks For a specific prediction target, i The loss function of a task is defined as: , in is a shared parameter, is a task-specific parameter, is the prediction loss of the i-th task, is the regularization loss of task-specific parameters, is the cross-task consistency loss; the total multi-task loss function is balanced by the task weights: , Task weight Adopt a dynamic adjustment strategy and adaptively update based on the learning difficulty and importance of each task: , in is the learning rate for weight updates.

[0015] Furthermore, the generalization training of small sample scene data based on multi-task learning and cross-species migration mechanism also includes adopting an uncertainty quantification method based on deep integration to estimate the epistemic uncertainty and accidental uncertainty of the prediction by training multiple models with different initializations. for M An independently trained graph neural network model, for input samples x , the mean and variance of the predicted probability are: , The uncertainty-based active learning strategy selects the most informative samples for annotation: , in is a set of unlabeled samples, is a set of labeled samples, Measuring samples x and the diversity of labeled samples.

[0016] The second aspect is a plant multi-gene stress resistance collaborative prediction system based on a federated graph neural network, including: The data acquisition module is configured to acquire plant genome data; The heterogeneous graph module is configured to model a multi-layer heterogeneous graph based on the acquired plant genome data; The model building module is configured to build a dynamic graph neural network model based on time series perception; The model training module is configured to train a dynamic graph neural network model using a multi-layer heterogeneous graph, wherein a federated graph neural network is used to perform distributed privacy computation on the multi-layer heterogeneous graph; and generalization training is performed on small sample scenario data based on multi-task learning and cross-species transfer mechanisms; The prediction module is configured to obtain gene function prediction results.

[0017] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for collaborative prediction of plant multi-gene stress resistance based on a federated graph neural network.

[0018] In a fourth aspect, the present invention provides a terminal device comprising 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, and the instructions are suitable for being loaded and executed by the processor to implement the method for collaborative prediction of plant multi-gene stress resistance based on a federated graph neural network.

[0019] In summary, the present invention has the following beneficial technical effects: 1. Improve the accuracy of gene function prediction, Through multi-layer heterogeneous graph fusion technology and time-aware dynamic graph neural network design, the present invention can simultaneously capture multi-dimensional biological information such as gene regulation, protein interaction, metabolic pathways and stress response, as well as the temporal change characteristics of gene expression during plant stress response.

[0020] 2. Achieve secure collaborative use of multi-institutional data, The federated learning framework designed by this paper solves the data silo problem in the field of plant genomics. Through a graph-aware federated aggregation algorithm and differential privacy protection mechanisms, research institutions can jointly train powerful predictive models without leaking original data.

[0021] 3. Reduce the cost and cycle of cultivating new varieties, Traditional plant breeding relies on extensive field trials and phenotyping, typically requiring 5-10 years to develop a new variety. This invention accurately predicts the synergistic stress-resistance effects of multi-gene combinations, enabling rapid computer-generated screening of optimal gene editing target combinations, reducing candidate gene screening time from years to days.

[0022] 4. Achieve effective transfer of gene function knowledge across species, The multi-task learning and cross-species transfer mechanism of this invention effectively transfers knowledge of gene function from known species to new or rare species. Through conservation modeling based on evolutionary distance and a meta-learning framework, gene function prediction accuracy for new species can be achieved with relatively high accuracy, even with small sample sizes (requiring only 20-50 labeled samples).

[0023] 5. Provide interpretable insights into biological mechanisms, Unlike traditional "black-box" machine learning methods, the prediction results of this invention, based on graph neural networks, are highly interpretable. Through attention mechanisms and graph structure analysis, the system can identify key regulatory pathways, core gene nodes, and important interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a schematic diagram of a method for collaboratively predicting plant multi-gene stress resistance based on a federated graph neural network in Example 1 of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below with reference to the accompanying drawings.

[0026] Example 1 Reference Figure 1 In this embodiment, a method for collaboratively predicting plant polygenic stress resistance based on a federated graph neural network includes: Access to plant genome data; Modeling multi-layer heterogeneous graphs based on acquired plant genome data; Build a dynamic graph neural network model based on time series perception; Using multi-layer heterogeneous graphs to train dynamic graph neural network models, including using federated graph neural networks to perform distributed privacy computing on multi-layer heterogeneous graphs; generalization training for small sample scenario data based on multi-task learning and cross-species migration mechanisms; Obtain gene function prediction results.

[0027] Specifically: S1 Multi-layer heterogeneous graph construction, To address the problem of unified modeling of heterogeneous biological data, this paper proposes a multi-layer heterogeneous graph fusion technology. The core concept of this technology is to decompose the complex biological network of plants into four functional layers and then construct a unified multi-layer heterogeneous graph representation through an inter-layer fusion mechanism. This hierarchical modeling strategy not only preserves the inherent structural characteristics of various types of biological information but also captures the correlations between different biological processes through cross-layer connections, thus laying a solid foundation for modeling the synergistic effects of multiple genes.

[0028] Traditional plant gene network analysis methods typically simply merge different types of biological data or use only a single data source, ignoring the hierarchical nature of biological systems and cross-layer interactions. For example, gene regulation is primarily reflected at the transcriptional level, while protein interactions occur at the post-translational level. Metabolic pathways reflect biochemical reaction networks, and stress responses involve coordination at multiple levels. By constructing a multi-layered heterogeneous graph, the present invention systematically integrates this multi-scale information while maintaining the biological significance of each layer.

[0029] S1.1 Design principles and structural definition of multi-layer heterogeneous graphs, The multi-layer heterogeneous graph designed in this paper 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 interrelated but functionally independent layers. This hierarchical design not only conforms to biological principles but also facilitates subsequent graph neural network processing. Define the plant biological network as a four-layer heterogeneous graph G ={ G 1, G 2, G 3, G 4}.

[0030] G 1=( V 1, E 1) represents the gene regulatory layer, V 1 is a gene node set, including transcription factor genes and target genes, E 1 is the regulatory relationship edge set, which represents the direct or indirect regulatory relationship between transcription factors and target genes. G 2=( V 2, E 2) Represents the protein interaction layer, reflecting molecular interactions at the post-translational level, V 2 is a collection of protein nodes, including various functional proteins such as enzymes, receptors, signaling molecules, etc. E 2 is the set of interaction edges, including protein-protein direct binding, enzyme-substrate interaction, signal transduction cascade and other relationships. G 3=( V 3, E3) Metabolic pathway layer, reflecting the metabolic reprogramming process of plants under stress conditions, V 3 is the metabolite node set, including primary metabolites, secondary metabolites and signal molecules, E 3 is the biochemical reaction edge set, which represents enzyme-catalyzed reactions, substrate-product conversion relationships, etc. G 4=( V 4, E 4) represents the stress response layer, which specifically captures the specific response patterns of plants under abiotic stress conditions. V 4 is a stress response gene set, including genes involved in ABA signal transduction, osmotic regulation, and reactive oxygen species scavenging processes. E 4 is a set of co-expression relationship edges, which is constructed based on transcriptome data under different stress conditions.

[0031] Inter-layer connections are made through cross-layer edge sets ={ E 12 , E 13 , E 14 , E 23 , E 24 , E 34} indicates that they can capture causal relationships and regulatory cascades between different biological levels. For example, E 12 Connecting genes to the proteins they encode, E 23 Ligases and the metabolic reactions they catalyze, E 14 Connecting regulatory genes and stress-responsive genes, etc.

[0032] S1.2 Multi-dimensional construction strategy of node features, The present invention designs a comprehensive feature construction strategy based on the multi-dimensional characteristics of plant genes, which can fully utilize sequence information, expression information and functional annotation information. ∈ V 1 , whose feature vector is constructed by multi-dimensional information fusion: , Among them, the sequence features Extracting structural information of gene sequences through k-mer encoding technology; Expression characteristics Contains gene expression values ​​under different stress conditions, at different time points, and in different tissues; functional annotation features Indicates the biological processes, molecular functions, and cellular components in which genes are involved.

[0033] S1.3 Diversified calculation mechanism of edge weights, The present invention designs a diversified edge weight calculation mechanism that can comprehensively consider linear correlation, nonlinear relationship and biological prior knowledge. ∈ E 1 , whose weight is calculated by the following formula: , Among them, the Pearson correlation coefficient Capturing genes i and j The linear expression correlation between Detecting nonlinear statistical dependencies in transcription factor binding strength Provide biological evidence support based on promoter sequence analysis and transcription factor binding site prediction. , , Adopting data-driven optimization strategy, the optimal parameter combination is determined through cross-validation. S1.4 Attention mechanism for inter-layer information fusion, This paper designs an inter-layer fusion strategy based on the attention mechanism, which can adaptively learn the importance weights of information at different levels and achieve effective integration of information at each layer: , Attention weight Calculated using the self-attention mechanism: , Among them, the linear transformation matrix Used to map node representations of different layers to a unified feature space, attention vector parameters The vector concatenation operation || is learned through end-to-end training and is used to combine the query and key vectors.

[0034] This multi-layer heterogeneous graph construction technology integrates diverse biological information, including gene regulation, 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 inherent characteristics of various biological information but also effectively integrates multi-scale information through cross-layer connections and attention fusion mechanisms.

[0035] S2 Time-aware dynamic graph neural network design, To address the problem of dynamic modeling of multi-gene synergistic effects, the present invention designs a time-aware dynamic graph neural network architecture. This architecture can simultaneously capture the static topological structure and dynamic evolutionary characteristics of the gene regulatory network, in particular, the temporal changes in gene expression patterns and network reconstruction during plant stress response. Plant stress response is a highly dynamic biological process, and the regulatory relationships between genes will change significantly as the stress stimulus persists. For example, in the early stages of alkaline stress (1-3 hours), emergency response genes are mainly activated; in the mid-stage (6-12 hours), genes related to osmotic regulation and ion balance are activated; and in the late stages (24 hours and beyond), long-term adaptive genes are activated. This temporal gene regulatory network reconstruction phenomenon is an important feature of plant stress resistance and a key mechanism for the generation of multi-gene synergistic effects.

[0036] S2.1 Construction and representation of timing diagram sequences, The present invention models the gene regulatory network of plants in the stress response process as a time sequence diagram , where each time point t i Corresponds to a snapshot of the graph. This article sets T=5, corresponding to key time points such as 0h (control), 1h, 3h, 6h, 12h, and 24h. The graph at each time point G (t) =( V , E (t) , X (t) ) has the following characteristics: the node set V remains unchanged at all time points to ensure the consistency of the network structure; the edge set Dynamic adjustment based on the gene co-expression relationship at that time point; node feature matrix Contains the gene expression values ​​and cumulative change characteristics at this time point. t The edge weight calculation of takes into account the temporal changes of expression relevance: , in, represents the correlation coefficient of gene expression at time point t, is the distance correlation coefficient, capturing nonlinear relationships, is the time-delayed correlation, which takes into account the time lag effect of regulation.

[0037] S2.2 Spatiotemporal Graph Convolutional Network Architecture, The spatiotemporal graph convolutional network (ST-GCN) designed in this paper can simultaneously process the spatial structure information and temporal evolution information of the graph. The network architecture adopts an encoder-decoder structure and achieves deep feature learning by stacking multiple ST-GCN layers. The forward propagation process of a single ST-GCN layer is defined as: , in, For the l The spatiotemporal feature tensor of the layer, Processing spatial neighborhood information for graph convolution operations, Handle timing dependencies for temporal convolution operations, , is the learnable parameter matrix.

[0038] Graph convolution operation Adopting the improved GAT mechanism: , The attention weight Taking into account the time-dependent neighborhood importance: , in It is a time-dependent edge feature vector, which contains information such as edge weight, control type and time decay factor.

[0039] S2.3 Temporal convolution and long short-term memory fusion mechanism, Temporal convolution operation Use causal convolution design to ensure that only historical information is used for prediction: , Where K is the size of the temporal convolution kernel, is the weight matrix corresponding to the time offset s. In order to better capture long-term temporal dependencies, the present invention introduces a gated recurrent unit mechanism: , in, , To reset the gate and update the gate, is a candidate hidden state, , , are learnable parameters.

[0040] S2.4 Multi-scale temporal modeling and attention pooling, Plant stress responses involve biological processes at multiple timescales, from minute-level signal transduction to hour-level gene expression changes to day-level physiological adaptation. This paper designs a multi-scale temporal modeling mechanism that can capture dynamic features at different timescales. Multi-scale temporal features are extracted using time windows of different sizes: , in Corresponding to different time scales. Multi-scale features are fused through an adaptive attention mechanism: , Attention weight Calculated as follows: , S2.5 Synergistic effect modeling and prediction mechanism, In order to achieve the synergistic effect of multiple genes, this paper designs a synergistic effect prediction mechanism based on graph attention, which can identify gene modules that work synergistically under specific stress conditions. , and its synergistic stress resistance score is calculated as: , in It's time t The gene interaction weights, It's genes i and j In time t The interaction strength is calculated by learning the node representation: , This time-aware dynamic graph neural network design captures the spatiotemporal evolution of gene regulatory networks, providing strong technical support for accurately predicting the synergistic stress resistance effects of multiple genes. The modular design of the network architecture enables it to adapt to different stress types and timescales, with good scalability and generalization capabilities.

[0041] S3 federated learning framework and privacy protection mechanism, To address privacy concerns in the joint utilization of multi-institutional data and the federated learning aggregation of graph-structured data, this paper 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 multi-institutional plant genetic data while ensuring data privacy and security.

[0042] 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 types, and edge connectivity patterns, exhibiting typical non-independent and identically distributed (Non-IID) characteristics.

[0043] Furthermore, plant genetic data holds 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 are ineffective at protecting node, edge, and subgraph 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 are needed.

[0044] S3.1 Overall Architecture Design of Federated Graph Neural Network, The federated graph neural network system designed in this paper adopts a parameter server architecture, consisting of a central aggregation server and multiple clients (participating institutions). The original data of each institution is always stored locally, and only the model parameters that have been processed for privacy protection are transmitted across the network. The system's training process includes the following key stages: Initialization phase: The central server initializes the global graph neural network model , the initial parameters Distributed to all participating clients. At the same time, the server also maintains a global graph structure template , used to unify the graph representation format of different clients. Local training phase: k Client-based local data Train the local model. The objective function of local training is designed as: , in It is the main mission loss, is the graph structure regularization loss, is the parameter regularization loss, and is the balance coefficient.

[0045] Aggregate update phase: Each client updates the privacy-preserving model Upload to the central server, the server executes the graph structure-aware federated aggregation algorithm and updates the global model parameters.

[0046] S3.2 Graph structure-aware federated aggregation algorithm, The heterogeneity of plant gene networks makes traditional federated averaging algorithms ineffective. This paper proposes a graph-structure-aware federated aggregation algorithm that can adaptively aggregate data based on the structural characteristics of each client's graph. The core idea of ​​the algorithm is to learn the similarities between the graph structures of each client using graph embedding technology and then calculate aggregation weights based on these structural similarities. The specific algorithm flow is as follows: First, a structural embedding vector is calculated for each client's graph structure: , in is a node iIn the last layer of GNN representation vector, the graph level embedding is obtained by average pooling. Then, the structural similarity matrix between clients is calculated: , The first one is the cosine similarity, which measures the similarity of the structural embeddings, and the second one is the KL divergence penalty based on the degree distribution. P k Represents the client k The degree distribution of Is a balancing parameter. Based on the structural similarity, the adaptive aggregation weight is calculated: , in is the amount of data from client k, is the weight coefficient of structural similarity. The final global model parameter update formula is: , S3.3 Design of graph neural networks with differential privacy protection, To protect the privacy of plant genetic data, this paper 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 (e.g., the expression values ​​of specific genes), but also edge-level privacy (e.g., regulatory relationships between genes) and subgraph-level privacy (e.g., specific regulatory modules). For graph neural networks, this paper implements privacy protection at multiple levels. Node-level privacy protection involves adding Gaussian noise to the node feature update process: , The noise variance According to the privacy budget and sensitivity calculate: , Edge-level privacy protection: Adding Laplace noise to attention weight calculation: , in is the sensitivity of the attention weight, is the edge-level privacy budget. Gradient-level privacy protection: Gradient clipping and noise addition are applied when updating model parameters: , in is the gradient clipping threshold, is the gradient noise variance.

[0047] S3.4 Privacy Budget Allocation and Optimization Strategy, This paper proposes a privacy budget allocation strategy based on importance sampling, which prioritizes the protection of parameters and structures that have a greater impact on model performance. Distribute as follows: , Budget allocations at each level are based on sensitivity analysis: , The sensitivity is estimated by the influence function: , S3.5 Security aggregation and communication optimization, To further enhance the security and efficiency of the system, this paper uses a secure multi-party computation protocol for parameter aggregation and communication compression technology to reduce network transmission overhead. The secure aggregation protocol uses a secret sharing-based approach to decompose each client's model update into multiple secret fragments: , in Is sent to j The secret fragments of the aggregation nodes satisfy: , Communication compression uses an importance-based sparsification strategy: , in is the compression ratio, TopK The operation retains the largest absolute value Scale parameter.

[0048] 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.

[0049] S4 Multi-task learning and cross-species transfer mechanism, To address the problem of cross-species generalization under small sample sizes, this paper designs a multi-task learning and cross-species transfer mechanism. The core of this mechanism is to effectively transfer gene function knowledge between different species through a meta-learning framework, and to improve the model's prediction accuracy and generalization ability in small sample sizes through multi-task joint training.

[0050] Plant gene functions exhibit remarkable conservation and species specificity during evolution. On the one hand, homologous genes often share similar functions across species, such as the widespread presence of stress-resistance transcription factor families in legumes. On the other hand, different species have developed unique adaptive mechanisms over the course of long-term evolution, such as the salt gland structure and associated genes unique to halophytes. Traditional machine learning methods often process data for each species independently, failing to effectively leverage cross-species functional conservation. Consequently, prediction accuracy significantly decreases when faced with novel species or rare stress conditions.

[0051] S4.1 Hierarchical Multi-Task Learning Architecture, The multi-task learning architecture proposed in this paper employs a hierarchical design, breaking down plant gene function prediction into multiple interrelated subtasks. This design captures common features across tasks while preserving task-specific information. The architecture comprises three layers: a shared representation layer, a task-specific layer, and a fused prediction layer. The shared representation layer learns common features across species and tasks; the task-specific layer captures information unique to a specific task or species; and the fused prediction layer integrates this multi-layered information for the final prediction.

[0052] Setting up multitasking collections Each of these tasks Corresponding to specific prediction targets, such as salt tolerance prediction, drought tolerance prediction, alkali tolerance prediction, etc. i The loss function of a task is defined as: , in is a shared parameter, is a task-specific parameter, is the prediction loss of the i-th task, is the regularization loss of task-specific parameters, is the cross-task consistency loss. The total multi-task loss function is balanced by the task weights: , Task weight Adopt a dynamic adjustment strategy and adaptively update based on the learning difficulty and importance of each task: , in is the learning rate for weight updates.

[0053] S4.2 Meta-learning-based cross-species transfer framework, To achieve effective cross-species knowledge transfer, this paper introduces a model-agnostic meta-learning framework. This framework aims to learn a good model initialization that enables the model to quickly adapt to new species or tasks with a small number of samples. The meta-learning training process is divided into two layers: an inner loop and an outer loop. The inner loop simulates rapid adaptation to new tasks, while the outer loop optimizes adaptability.

[0054] Assume support set and queryset Represents tasks A small number of labeled samples and test samples. Inner loop update process: , in is the inner loop learning rate, The parameter is The outer loop updates the performance of the adapted parameters on the query set: , in is the outer loop learning rate, N is the number of meta-training tasks.

[0055] S4.3 Modeling of gene function conservation among species, In order to better utilize the functional conservation information across species, the present invention designs a conservation modeling mechanism based on evolutionary distance. This mechanism guides the knowledge transfer process through phylogenetic tree information, so that species with closer evolutionary relationships can share more functional information. Suppose the species set is , the evolutionary distance matrix between species is ,in Indicates species and The evolutionary distance between them. Based on the evolutionary distance, the similarity weight between species is calculated: , in is the distance sensitivity parameter. Function prediction across species is achieved through weighted integration: , in Based on species The predicted probability of sample x by the model trained with the data.

[0056] S4.4 Domain Adaptation and Adversarial Training Mechanism, In order to reduce the impact of domain differences between species, the present invention introduces a domain adaptation mechanism based on adversarial training. This mechanism enables the feature extractor to learn a species-independent universal representation through adversarial learning. The domain adaptation module includes a feature extractor and domain discriminator The goal of the feature extractor is to learn species-independent gene representations, while the domain discriminator attempts to determine the species of the sample based on the feature representations. The objective function of adversarial training is: , At the same time, in order to maintain task prediction performance, it is also necessary to minimize task loss: , in is a task classifier, is the classification loss function. The overall training goal is: , S4.5 Uncertainty Quantification and Active Learning Strategies, In small sample learning scenarios, the quantification of model prediction uncertainty is of great significance for guiding active learning and improving model reliability. This paper adopts an uncertainty quantification method based on deep integration to estimate the epistemic uncertainty and aleatoric uncertainty of prediction by training multiple models with different initializations. for M An independently trained graph neural network model, for input samples x , the mean and variance of the predicted probability are: , The uncertainty-based active learning strategy selects the most informative samples for annotation: , in is a set of unlabeled samples, is a set of labeled samples, Measuring samples x and the diversity of labeled samples.

[0057] Example 2 This embodiment provides a plant multi-gene stress resistance collaborative prediction system based on a federated graph neural network, including: The data acquisition module is configured as follows: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a method for collaboratively predicting plant multi-gene stress resistance based on a federated graph neural network.

[0058] A terminal device includes 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, and the instructions are suitable for being loaded and executed by the processor to implement a method for collaborative prediction of plant multi-gene stress resistance based on a federated graph neural network.

[0059] 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, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for collaborative prediction of plant polygenic stress resistance based on federated graph neural network, characterized in that: include: Access to plant genome data; Modeling multi-layer heterogeneous graphs based on acquired plant genome data; Build a dynamic graph neural network model based on time series perception; Using multi-layer heterogeneous graphs to train dynamic graph neural network models, including using federated graph neural networks to perform distributed privacy computing on multi-layer heterogeneous graphs; generalization training for small sample scenario data based on multi-task learning and cross-species migration mechanisms; Obtain gene function prediction results.

2. A method for collaborative prediction of plant polygenic stress resistance based on federated graph neural network according to claim 1, characterized in that: The method of modeling a multi-layer heterogeneous graph based on the obtained plant genome data includes defining the plant biological network as a four-layer heterogeneous graph, including a gene regulation layer, a protein interaction layer, a metabolic pathway layer, and a stress response layer. For the gene nodes in the gene regulation layer, a feature vector is constructed by multi-dimensional information fusion. For the regulatory relationship edges in the gene regulation layer, a feature vector is constructed by multi-dimensional information fusion. ∈ E 1. Calculate its weight through the formula, expressed as: ， Among them, the Pearson correlation coefficient Capturing genes i and j The linear expression correlation between Detecting nonlinear statistical dependencies in transcription factor binding strength Based on promoter sequence analysis and transcription factor binding site prediction, the weight coefficient , , A data-driven optimization strategy is adopted to determine the optimal parameter combination through cross-validation.

3. The method for collaborative prediction of plant polygenic stress resistance based on federated graph neural network according to claim 2, characterized in that: The method of modeling a multi-layer heterogeneous graph based on the obtained plant genome data also includes integrating information of each layer of the four-layer heterogeneous graph using an inter-layer fusion strategy based on an attention mechanism. , attention weight The calculation is done using the self-attention mechanism, which is expressed as: ， Among them, the linear transformation matrix Used to map node representations of different layers to a unified feature space, attention vector parameters The vector concatenation operation || is learned through end-to-end training and is used to combine the query and key vectors.

4. The method for collaborative prediction of plant polygenic stress resistance based on federated graph neural network according to claim 3, characterized in that: The construction of a dynamic graph neural network model based on time series perception includes modeling the gene regulatory network of plants in the stress response process as a time series graph sequence , where each time point t i Corresponding to a graph snapshot, a spatiotemporal graph convolutional network (ST-GCN) is constructed. The network architecture adopts an encoder-decoder structure and achieves deep feature learning by stacking multiple ST-GCN layers. The forward propagation process of a single ST-GCN layer is: ， in, For the l The spatiotemporal feature tensor of the layer, Processing spatial neighborhood information for graph convolution operations, Handle timing dependencies for temporal convolution operations, , is a learnable parameter matrix; graph convolution operation Adopting improved GAT mechanism: ， The attention weight is calculated by considering the time-related neighborhood importance : ， in is the time-dependent edge eigenvector.

5. The method for collaborative prediction of plant polygenic stress resistance based on federated graph neural network according to claim 4, characterized in that: The construction of a dynamic graph neural network model based on time sequence perception also includes using a multi-scale time modeling mechanism to capture dynamic features of different time scales, where the attention weight The calculation is expressed as: ， The synergistic effect prediction mechanism based on graph attention identifies gene modules that work synergistically under specific stress conditions. For a given gene set, , and its synergistic stress resistance score is calculated as: ， in It's time t The gene interaction weights, It's genes i and j In time t The interaction intensity.

6. A method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural network according to claim 5, characterized in that: The distributed privacy computing of multi-layer heterogeneous graphs using federated graph neural networks involves adaptive aggregation based on the graph structure characteristics of each client, learning the similarity of the graph structures of each client through graph embedding, and then calculating the aggregation weight based on the structural similarity. The structural embedding vector is calculated for the graph structure of each client: ， is a node i The representation vector of the last layer of GNN is used to calculate the structural similarity matrix between clients: ， in P k Represents the client k The degree distribution of is the balance parameter, and finally based on the structural similarity, the adaptive aggregation weight is calculated: ， in is the amount of data from client k, is the weight coefficient of structural similarity, and the final global model parameter update formula is: .

7. The method for collaborative prediction of plant polygenic stress resistance based on federated graph neural network according to claim 6, characterized in that: The distributed privacy computing of multi-layer heterogeneous graphs using federated graph neural networks also includes the introduction of a differential privacy protection mechanism in the federated learning framework, where node-level privacy protection adds Gaussian noise during the node feature update process: ， Noise variance According to the privacy budget and sensitivity calculate: ； Add Laplacian noise to the attention weight calculation: ， in is the sensitivity of the attention weight, is the edge-level privacy budget; gradient clipping and noise addition are applied when the model parameters are updated: ， in is the gradient clipping threshold, is the gradient noise variance.

8. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural network according to claim 7, characterized in that: The method is based on multi-task learning and cross-species migration mechanism to perform generalization training on small sample scenario data, including the cross-species generalization problem under small sample conditions, using a hierarchical design to decompose plant gene function prediction into multiple interrelated subtasks, wherein the shared representation layer learns common features across species and tasks; the task-specific layer captures the unique information of a specific task or species; the fusion prediction layer integrates multi-level information for the final prediction, and sets a multi-task set. Each of these tasks For a specific prediction target, i The loss function of a task is defined as: ， in is a shared parameter, is a task-specific parameter, is the prediction loss of the i-th task, is the regularization loss of task-specific parameters, is the cross-task consistency loss; the total multi-task loss function is balanced by the task weights: ， Task weight Adopt a dynamic adjustment strategy and adaptively update based on the learning difficulty and importance of each task: ， in is the learning rate for weight updates.

9. The method for collaborative prediction of plant multi-gene stress resistance based on federated graph neural network according to claim 8, characterized in that: The generalization training of small sample scene data based on multi-task learning and cross-species migration mechanism also includes adopting an uncertainty quantification method based on deep integration to estimate the epistemic uncertainty and accidental uncertainty of the prediction by training multiple models with different initializations. for M An independently trained graph neural network model, for input samples x , the mean and variance of the predicted probability are: ， The uncertainty-based active learning strategy selects the most informative samples for annotation: ， in is a set of unlabeled samples, is a set of labeled samples, Measuring samples x and the diversity of labeled samples.

10. A plant multi-gene stress resistance collaborative prediction system based on federated graph neural network, characterized in that: include: The data acquisition module is configured to acquire plant genome data; The heterogeneous graph module is configured to model a multi-layer heterogeneous graph based on the acquired plant genome data; The model building module is configured to build a dynamic graph neural network model based on time series perception; The model training module is configured to train a dynamic graph neural network model using a multi-layer heterogeneous graph, wherein a federated graph neural network is used to perform distributed privacy computation on the multi-layer heterogeneous graph; and generalization training is performed on small sample scenario data based on multi-task learning and cross-species transfer mechanisms; The prediction module is configured to obtain gene function prediction results.

Images