Method and system for inferring chicken gene-pathogen interaction network using graph neural network

Abstract

The application provides a method and system for inferring a chicken gene-pathogen interaction network by using a graph neural network, and relates to the technical field of data processing, which comprises the following steps: firstly, obtaining multiple omics raw data of a chicken individual at the same time scale; through cross-species unified graph structure construction processing, mapping gene and pathogen sequences to a unified topological space and constructing an initial interaction graph based on time series co-occurrence; combining continuous phenotype data and genetic pedigree information, performing dynamic graph representation learning under genetic constraints to obtain a graph representation with spatiotemporal characteristics and genetic generalization ability; through multi-scale topological structure optimization processing, screening out candidate interactions; finally, based on the graph auto-encoding discriminant structure of mutual information maximization, automatically identifying and retaining connections that are crucial to network reconstruction to generate an interaction network. The application realizes the inference of a gene-pathogen interaction relationship with high robustness and biological interpretability from multidimensional dynamic data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for inferring chicken gene-pathogen interaction networks using graph neural networks. Background Technology

[0002] Utilizing high-throughput sequencing data to analyze the interaction network between the host and pathogens has become an important direction for studying disease mechanisms and guiding disease-resistant breeding. Particularly in large-scale poultry farming, understanding the dynamic interactions between chicken genes and pathogens is crucial for disease control and genetic improvement. This scenario involves multi-omics data continuously collected from individual chickens, including gene transcripts, pathogen genomes, dynamic phenotypes, and genetic pedigree information. Its core characteristics lie in the inherent complexity of the data, including high dimensionality, heterogeneity, strong temporal correlations, and significant individual genetic differences.

[0003] Traditional research methods typically process data from different sources in steps. For example, they may first construct gene co-expression networks or pathogen phylogenetic trees independently, and then perform association matching based on statistical correlations or existing biological knowledge bases. These methods struggle to deeply integrate synergistic change patterns over time and lack effective mechanisms to distinguish signals generated by stable biological interactions from interference introduced by random noise or specific individual genetic backgrounds. As a result, the inferred interaction networks often exhibit poor stability and limited generalization ability, making it difficult to reveal core interaction patterns that are prevalent across different genetic backgrounds.

[0004] Therefore, how to automatically learn and extract gene-pathogen cross-species interaction networks with genetic robustness and temporal consistency from raw, multi-dimensional dynamic data has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for inferring chicken gene-pathogen interaction networks using graph neural networks, thereby improving the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0006] In a first aspect, this application provides a method for inferring chicken gene-pathogen interaction networks using graph neural networks, including:

[0007] Acquire raw transcript sequencing reads of all chicken individuals collected at multiple consecutive time points under the same preset time scale, raw whole genome sequencing reads of pathogens of the corresponding chicken individuals at the same time point, raw data of continuous phenotypic records of chicken individuals at all time points, and raw data of blood relations between chicken individuals;

[0008] Based on the acquired raw transcript sequencing reads and pathogen whole genome raw sequencing reads, cross-species unified graph structure construction was performed to obtain the initial gene-pathogen interaction graph of the fusion time-series signal.

[0009] Based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records, and the original data of blood relations between chicken individuals, dynamic graph representation learning processing under genetic constraints is performed to obtain a spatiotemporal graph representation with genetic generalization ability.

[0010] Based on the spatiotemporal graph representation, multi-scale topology optimization processing is performed to obtain a candidate interaction network of chicken genes and pathogens;

[0011] Based on the candidate interaction network, a saliency network generation process is performed. Based on a preset graph autoencoder discriminant structure that maximizes mutual information, edge connections that are more important than a preset threshold to the overall network response pattern are automatically identified and retained, resulting in the final inferred chicken gene-pathogen interaction network.

[0012] Secondly, this application also provides a system for inferring chicken gene-pathogen interaction networks using graph neural networks, comprising:

[0013] The acquisition unit is used to acquire raw transcript sequencing reads of all chicken individuals collected at multiple consecutive time points under the same preset time scale, raw genome sequencing reads of pathogens of the corresponding chicken individuals at the same time point, raw data of continuous phenotypic records of chicken individuals at all time points, and raw data of blood relations between chicken individuals.

[0014] The building unit is used to perform cross-species unified graph structure construction processing based on the acquired raw transcript sequencing read data and pathogen whole genome raw sequencing read data to obtain the initial gene-pathogen interaction map of fused temporal signals;

[0015] The learning unit is used to perform dynamic graph representation learning processing under genetic constraints based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records, and the original record data of blood relations between chicken individuals, to obtain a spatiotemporal graph representation with genetic generalization ability.

[0016] An optimization unit is used to perform multi-scale topology optimization processing based on the spatiotemporal graph representation to obtain a candidate interaction network of chicken genes and pathogens.

[0017] The generation unit is used to perform saliency network generation processing based on the candidate interaction network. Based on a preset graph autoencoder discriminant structure that maximizes mutual information, it automatically identifies and retains edge connections that are more important than a preset threshold to the overall network response pattern, thereby obtaining the final inferred chicken gene-pathogen interaction network.

[0018] The beneficial effects of this invention are as follows:

[0019] This invention acquires multimodal raw data of chicken individuals at continuous time points and directly maps gene and pathogen sequences to a unified topological space based on a cross-species unified graph structure to fuse temporal signals. Then, it utilizes dynamic graph representation learning under genetic constraints to enable the network to simultaneously absorb the temporal evolution of individual phenotypes and satisfy structural consistency constraints across different genetic backgrounds. Further multi-scale topological optimization distinguishes and strengthens core connection patterns that are stable in both spatiotemporal and genetic dimensions. Finally, a saliency network generation process based on maximizing mutual information automatically screens out biological pathway consistency connections crucial for reconstructing the overall pattern. This series of coherent processes allows this invention to directly infer, end-to-end, chicken gene-pathogen interaction networks with high genetic generalization ability and temporal robustness from raw data, effectively overcoming the problems of poor network stability and weak generalization ability caused by traditional methods due to step-by-step processing and the inability to deeply integrate spatiotemporal dynamics and genetic background information.

[0020] Other features and advantages of the present invention will be further described in the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart of the method for inferring chicken gene-pathogen interaction networks using graph neural networks as described in this embodiment of the invention;

[0023] Figure 2 This is a schematic diagram of the system structure for inferring the chicken gene-pathogen interaction network using a graph neural network, as described in an embodiment of the present invention.

[0024] In the diagram: 701, Acquisition Unit; 702, Construction Unit; 703, Learning Unit; 704, Optimization Unit; 705, Generation Unit. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Example 1:

[0028] This embodiment provides a method for inferring chicken gene-pathogen interaction networks using graph neural networks.

[0029] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4 and S5.

[0030] Step S1: Obtain raw transcript sequencing reads of all chicken individuals at multiple consecutive time points under the same preset time scale, raw whole genome sequencing reads of pathogens of the corresponding chicken individuals at the same time point, raw data of continuous phenotypic records of chicken individuals at all time points, and raw data of blood relations between chicken individuals;

[0031] Understandably, the raw transcript sequencing reads in this step are obtained using high-throughput RNA sequencing technology. Biological samples (such as peripheral blood or specific tissues) are collected from the same batch of chickens at multiple consecutive time points, and without any expression level standardization or gene annotation, a set of raw base sequence reads reflecting chicken gene transcription activity is directly obtained. For the raw whole-genome sequencing reads of pathogens from the same chicken individuals at the same time points, metagenomic sequencing is performed on similar samples (such as blood or tissue homogenates) collected at the same time. This aims to directly obtain the raw genome sequences of all pathogens in the samples without sequence assembly and species identification, thus preserving the original abundance and diversity information of the pathogen population. The raw data for continuous phenotypic records of chicken individuals at all time points are collected using standardized manual recording methods under natural exposure or rearing conditions, including measurements such as body temperature, activity level, feed intake, weight gain, and clinical symptoms, forming continuous time-series observations without discretization. The raw data for pedigree records among chicken individuals rely on parentage identification results obtained from existing pedigree archives at the farm to clearly record the genetic origin and kinship of all participating individuals.

[0032] Step S2: Based on the obtained raw transcript sequencing read data and pathogen whole genome raw sequencing read data, cross-species unified graph structure construction processing is performed to obtain the initial gene-pathogen interaction map of the fused temporal signal;

[0033] Understandably, this step uses a parameter-sharing heterogeneous projection mechanism to non-linearly map two types of original read data, which differ significantly in sequence composition and statistical characteristics, to a common latent space. This allows sequence fragments from different species to obtain comparable vector representations within this space, thus being defined as nodes in a unified topological space. The construction of edges between nodes does not rely on prior knowledge but is dynamically calculated based on the co-occurrence patterns of these node vectors across multiple chicken individuals and consecutive time points, paying particular attention to associations with temporally leading or lagging relationships. This initializes a graph structure that can represent entity coexistence and also implies potential causal or response sequences. The initial gene-pathogen interaction graph directly produced by this process is essentially a heterogeneous graph embedding temporal information, laying a data structure foundation for subsequent in-depth analysis that combines entity unity with dynamic associations. In this step, step S2 includes steps S21, S22, and S23.

[0034] Step S21: Perform cross-species sequence fragment numerical characterization processing based on the original transcript sequencing read data and the original pathogen whole genome sequencing read data. Specifically, by constructing statistical feature vectors for chicken gene transcript fragments and pathogen genome fragments respectively, a preliminary numerical representation of the sequence fragments is obtained. The statistical feature vectors are constructed based on their corresponding sequence composition and read coverage depth.

[0035] Understandably, this step first involves independently analyzing each chicken gene transcript read and pathogen genome read. For sequence composition, a fixed-length overlapping k-mer sliding window is used to traverse the entire read sequence, counting the frequency of all possible k-mer combinations to form a high-dimensional sequence profile vector. This vector characterizes the unique sequence pattern and base distribution preference of the read. Simultaneously, all sequencing read information from the same biological sample is integrated, and the original coverage count of the fragment in the sample is calculated by counting the number of reads mapped to the same genomic location or transcript region. Subsequently, the k-mer frequency vector and coverage depth count of each sequence fragment are standardized and concatenated to form a unified statistical feature vector. Standardization involves L2 normalization, and concatenation directly combines sequence composition features with coverage depth features into feature pairs.

[0036] Features are constructed by directly mining the intrinsic statistical properties of the reads themselves. Sequence composition features are extracted by focusing on the inherent patterns of base arrangement, which can reflect, to some extent, the functional region characteristics or species origin signature information of the fragment, without prior knowledge of its exact attribution. Read coverage depth, derived directly from raw observations in sequencing experiments, quantifies the frequency of the fragment's detection in a specific sample, serving as a direct proxy indicator of gene expression levels or pathogen load. Combining these two types of information to construct a statistical feature vector essentially creates a digital fingerprint for each sequence fragment, combining identity and state quantification, thereby transforming unstructured sequence data into a structured numerical representation. This method is particularly suitable for handling complex pathogen sequences and incomplete reference database coverage in metagenomic sequencing, as well as transcripts in the host transcriptome that have not yet been fully annotated, providing a compatible and comparable mathematical foundation for subsequently processing two types of biological entities from different sources within a unified space.

[0037] Step S22: Perform heterogeneous space unified projection processing based on the preliminary numerical representation of the sequence fragments. By normalizing the fragment feature vectors from chickens and pathogens and mapping them to a common latent space, a unified node embedding for chickens and pathogens is obtained.

[0038] Understandably, this step first performs dimensionless processing on the preliminary numerical representations of sequence fragments from chickens and pathogens—i.e., statistical feature vectors—to eliminate dimensional differences caused by variations in sequencing depth or fragment length, ensuring all features are on the same numerical scale. Subsequently, these normalized feature vectors are input into a parameter-shared projection function, typically implemented using a multi-layer neural network whose parameters are trained on both chicken and pathogen-derived data. This network uses a nonlinear transformation to map feature vectors, originally belonging to different statistical distributions, into a new, typically lower-dimensional, common latent space. In this common latent space, the mapping aims to ensure that biologically related chicken gene fragments and pathogen genome fragments have embedding vectors that are close in distance, while unrelated fragments are kept far apart, regardless of their original species labels.

[0039] The projection function is shown below: ;

[0040] in, For sequence fragments Uniform node embedding vectors in the common latent space The projection function is a parameter-shared function, typically implemented by a multilayer perceptron. Its parameters... During training, normalized chicken source data and pathogen source data are used together for optimization, without distinguishing between species.

[0041] The objective formula for optimizing the embedding space correlation is shown below:

[0042] ;

[0043] in, For the loss function based on contrastive learning, For a positive sample pair, Indicates from positive sample pairs set , Embedded vector and The Euclidean distance between them Let be the node embedding vector of the pathogen genome fragment in the common latent space. A negative sample fragment, derived from a randomly sampled sequence fragment. A collection of other fragments that have no obvious co-occurrence or known association , It is a natural exponential function. It is a logarithmic function.

[0044] Step S23: Perform edge connection initialization processing based on time-delay co-occurrence according to the unified node embedding and multiple continuous time point information. Calculate the time-delay mutual information of each node embedded in the cross-sample time series within a preset sliding time window, and dynamically construct the initial edge reflecting the response time sequence relationship to obtain the initial gene-pathogen interaction graph of the fused time sequence signal.

[0045] Understandably, this step first divides the time-series data of each chicken individual sample into a series of overlapping sliding time windows (1 second in length or other preset time periods) based on information from multiple consecutive time points. Within each window, the processing focuses on calculating the time-delay mutual information of the embedding vectors of any two nodes (representing a chicken gene fragment and a pathogen genome fragment, respectively) over time. This means it systematically evaluates the change in the amount of information contained in the state of the other node when the state of one node is delayed or advanced by a specific time unit. In this way, it is possible to quantify whether there is a directional information flow or predictive relationship between a pair of nodes, such as whether an early change in a pathogen characteristic predicts a subsequent change in the expression of a specific host gene. The calculated time-delay mutual information values ​​between all node pairs are then used to dynamically construct the edge connections of the initial graph, where an edge is established between node pairs with mutual information values ​​exceeding a certain threshold. The direction of the edge is determined by the dominant temporal direction of information transmission, and the weight is determined by the magnitude of the mutual information value. Ultimately, this process outputs an initial gene-pathogen interaction graph that incorporates temporal signals. The edges of this graph not only represent the relationships between entities, but also implicitly suggest the possible temporal order and causal orientation of their interactions, providing a crucial initial topological structure for subsequent dynamic modeling.

[0046] The formula for calculating mutual information is as follows:

[0047] ;

[0048] in, In the window Inside, node With nodes Between time delay The mutual information value below, Indicates a point in time. For nodes At the point of time The unified node embedding vector, For nodes At the point of time The unified node embedding vector, For nodes At the point of time With nodes At the point of time The joint probability distribution, This represents the marginal probability distribution of the embedded vector.

[0049] Step S3: Based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records, and the original data of blood relations between chicken individuals, perform dynamic graph representation learning processing under genetic constraints to obtain a spatiotemporal graph representation with genetic generalization ability.

[0050] Understandably, this step first fuses the initial gene-pathogen interaction map with the raw data of continuous phenotypic records. Through a specific network module, it models the correlation between the macroscopic phenotypic quantitative indicators observed at each time point and the state representation of nodes in the corresponding time-point graph. This drives the evolutionary path of node states to calibrate with the temporal trajectory of phenotypic changes, thereby encoding individual-level physiological or pathological state changes into the dynamic representation of the graph. Simultaneously, this processing makes in-depth use of the raw data on blood relations between chicken individuals. Its key operation is to construct a genetic distance matrix based on pedigree information. Based on this, a structured constraint is introduced during the training of the graph neural network. Specifically, during feature learning, it mandates that the graph representations of individuals with similar genetic backgrounds should have higher similarity in topological structure or node features, while the graph representations of individuals with distant genetic backgrounds should show differences. Through this training mechanism that combines fitting of phenotypic time series with genetic consistency constraints, the dynamic graph representation learning process not only captures the evolutionary patterns of the host-pathogen interaction network over time. In this step, step S3 includes steps S31, S32 and S33.

[0051] Step S31: Perform node state temporal alignment processing based on the initial gene-pathogen interaction diagram and the original data of the continuous phenotypic records. By associating the graph node state vector at each time point with the multi-dimensional phenotypic observation value of the corresponding chicken individual at that time point, the temporal evolution path of the node state is back-calibrated based on the preset recurrent neural network encoding the phenotypic time series, and the node state sequence is obtained synchronously calibrated with the phenotypic dynamics.

[0052] Understandably, this step first uses an LSTM recurrent neural network to encode the phenotypic time series of each chicken individual to capture its dynamic evolution patterns and state dependencies across time. Subsequently, the encoded phenotypic representation, containing temporal contextual information, is correlated with the corresponding time-point graph node state vectors through a correlation model. This modeling is not a simple feature concatenation, but rather uses a learnable attention or mapping mechanism to treat the dynamic change pattern of the phenotypic as a soft constraint, backpropagating and calibrating the temporal evolution path of the node state itself. The optimization objective is to maximize the explanation or prediction of the observed phenotypic trajectory by minimizing the phenotypic dynamics. After this calibration process, the final output node state sequence not only retains the original interaction information, but its change rhythm and trend are also synchronized with the macroscopic physiological or pathological phenotypic dynamics.

[0053] The formula for modeling the association between node state and phenotype is shown below:

[0054]

[0055] ;

[0056] in, The initial gene-pathogen interaction mapping at time points The node state vector, The scoring function (grey relational coefficient calculation function) calculates the correlation between node states and phenotypic codes. The total length of the time series. It is a sequence fragment. In time The generated calibration signal vector, Time delay The phenotypic encoding vector.

[0057] The node state timing calibration optimization formula is shown below:

[0058] ;

[0059] in, This is a quantified value representing the alignment between the graph node state sequence and the phenotypic time series. For time The node status, It is a multilayer perceptron with the following parameters: This is used to generate the calibrated node state. To make the calibration signal Compared to the node state at the previous time step splicing The weighting coefficients for the smoothing regularization term, It is the square of the L2 norm, that is, the square of the Euclidean distance.

[0060] Step S32: Perform subgraph structure constraint processing based on genetic similarity according to the node state sequence and the original record data of the blood relationship between the chicken individuals. Construct a genetic distance matrix by parsing the blood relationship, and construct positive and negative sample pairs between the graph data of multiple chicken individuals based on the genetic distance matrix to obtain a set of node and edge representations that satisfy the genetic consistency constraint.

[0061] Understandably, this step first involves in-depth analysis of the raw records of blood relations among chicken individuals, transforming pedigree information into a quantified genetic distance matrix. This matrix precisely characterizes the closeness or distance between any two individuals within a specific genetic context. Subsequently, this matrix is ​​used to structure the graph data corresponding to multiple chicken individuals. The core operation is to systematically construct positive and negative sample pairs for training based on the magnitude of the genetic distance. Specifically, for individuals with a genetic distance less than a preset threshold, their respective subgraphs or node representations at the same time are constructed as positive sample pairs, encouraging the model to map them to similar positions in the feature space. Conversely, for individuals with a genetic distance greater than the preset threshold, their corresponding representations are constructed as negative sample pairs, forcing the model to learn to push their representations further apart. Through the introduction of this contrastive learning mechanism, the graph representation learning process is explicitly guided to focus on and strengthen recurring, conservative node associations and edge connections between individuals with different kinship relationships, while weakening accidental associations that only exist within a specific genetic context.

[0062] The formula for constructing the genetic distance matrix is ​​shown below:

[0063] ;

[0064] in, Genetic distance, For individuals A collection of genealogical ancestors. For individuals A collection of genealogical ancestors. The number of elements in the set.

[0065] Step S33: Perform spatiotemporal information fusion processing based on the node and edge representation set. Through a network structure with a preset fusion-gated temporal convolution and graph attention mechanism, aggregate the historical state of nodes in the time dimension and aggregate neighbor information in the spatial dimension based on the updated edge weights to obtain a spatiotemporal graph representation with genetic generalization ability.

[0066] Understandably, this step is implemented through a pre-defined neural network architecture that integrates a gated temporal convolution module and a graph attention mechanism. First, in the temporal dimension, the gated temporal convolution module operates on the state sequence of each node. It selectively aggregates the node's state information at different historical time points through controllable gating units, thereby extracting its cross-temporal dependency patterns and evolutionary trends. In the spatial dimension, the graph attention mechanism dynamically recalculates the attention weights of each edge in the graph based on the currently updated node representation. These weights reflect the importance of interactions between nodes in a specific spatiotemporal context, and then aggregate information from neighboring nodes based on these weights. These two processes are not independent but rather iteratively alternate within the network: the node states after temporal convolution provide updated node features for graph attention, while the spatial information aggregated by graph attention serves as one of the inputs for the next temporal convolution.

[0067] The gated temporal convolution update formula is shown below:

[0068] ;

[0069] in, To update the door, The Sigmoid activation function maps the output to the (0,1) interval. , and The weight matrix is ​​a learnable matrix. , and For learnable bias vectors, For nodes In the previous moment The hidden state is the memory unit of the temporal convolution. For nodes At the present moment Input features, It's a door reset. It is an element-wise multiplication operation. It is the hyperbolic tangent function. This represents the current candidate hidden state. For nodes At any moment The updated hidden state To update the door.

[0070] The processing formula for the spatiotemporal diagram representation is as follows:

[0071] ;

[0072] in, For nodes with his neighbors At any moment The original attention score, for Activation function For the transpose of the learnable parameter vector, The weight matrix is ​​a linear transformation matrix. For nodes At any moment The updated hidden state The attention weights are normalized. For the first 1 node For nodes The set of neighbors in the current graph structure. For nodes with his neighbors At any moment The original attention score, For nodes At any moment The representation after spatial attention aggregation, It is a non-linear activation function.

[0073] Step S4: Perform multi-scale topology optimization processing based on the spatiotemporal graph representation to obtain the candidate interaction network of chicken gene-pathogen;

[0074] Understandably, this step first refines the evaluation of first-order neighborhood relationships of nodes at the local graph scale using mechanisms such as graph attention, recalculating the reliability score of each edge as a direct interaction channel, thereby strengthening local connections that constitute tightly functional modules. Next, the process expands to the global scale, analyzing multi-step paths or higher-order structural patterns between nodes—for example, assessing whether two nodes frequently co-occur on longer paths or play similar topological roles—to identify entity pairs that, while not directly connected, may form stable biological associations through intermediate media. Finally, it comprehensively examines whether the various associations identified in previous steps (including local direct connections and global indirect connections) persist across different stages of the time series and in chicken subpopulations with different genetic backgrounds. Only connection patterns that exhibit high consistency under multiple conditions are preserved and integrated, ultimately forming a candidate interaction network that eliminates transient noise and individual-specific artifacts, has a clear hierarchical structure, and a more robust biological foundation. Step S4 includes steps S41, S42, and S43.

[0075] Step S41: Based on the spatiotemporal graph representation with genetic generalization ability, perform local direct interaction strength evaluation processing, quantify the importance weight of each edge in aggregating local neighborhood information through a preset attention network, and select the core edges that constitute a stable local cluster structure to obtain the enhanced local interaction subgraph.

[0076] Understandably, this step uses a pre-defined attention network for evaluation. This network doesn't assign static weights to all edges; instead, it dynamically calculates the relative importance of each edge in aggregating the local neighborhood information of the two nodes it connects, based on the specific representation of the nodes in the current spatiotemporal context. This calculation process essentially simulates the specificity and conditional dependence of biological interactions; for example, specific gene-pathogen interactions become more critical under certain physiological states. Based on the calculated dynamic importance weights, the process further filters and identifies and retains edges with consistently high weights (greater than or equal to a pre-defined importance threshold, e.g., greater than or equal to 0.8) that connect to form stable local cluster structures (such as tight groups or communities). These edges are considered core edges. This filtering criterion ensures that the retained connections are not only statistically significant but also constitute coherent functional modules in terms of topology. Ultimately, by removing a large number of low-weight or isolated edges, a local interaction subgraph with significantly reduced noise and enhanced key direct interaction relationships is obtained. This provides a clearer and more reliable foundation for subsequent differentiation of direct and indirect effects and for conducting higher-order topological analysis.

[0077] The attention network includes an input layer, an attention coefficient calculation layer, and a normalization layer. The network outputs the dynamic attention weights of the edges at different times.

[0078] The formula for calculating the importance weight of local neighborhood information is as follows:

[0079] ;

[0080] in, For the edge At the point of time Dynamic attention weights, node At the point of time The spatiotemporal diagram represents, For the edge At the point of time The initial feature vector, For nodes At the point of time The spatiotemporal diagram represents, For the edge At the point of time The initial eigenvectors.

[0081] Step S42: Perform high-order path dependency mining processing based on the enhanced local interaction subgraph, capture and evaluate the potential regulatory or synergistic effects between nodes through multiple indirect paths through the attention network, and obtain a high-order relationship graph that reflects the global topological correlation.

[0082] Understandably, this step uses an enhanced local interaction subgraph as input, whose basic structure has filtered out a large number of noisy edges, thus enabling reliable higher-order analysis. First, an attention network captures and evaluates all possible multiple indirect paths between nodes. The network calculates an influence propagation score for each path with a length of two hops or longer, which comprehensively considers the strength of each connection segment in the current representation, as well as the overall topological efficiency of the path. Its key innovation lies in not simply enumerating and treating all paths equally, but dynamically learning the contribution weights of different paths to the potential functional associations between target nodes through an attention mechanism. This identifies node pairs that, while not directly connected, are functionally tightly coupled through a set of stable and coherent indirect paths. This process can reveal complex patterns such as co-regulatory modules, cascaded signaling pathways, or long-range collaboration in metabolic networks. Finally, all node pairs with significant potential regulatory or synergistic effects identified through this higher-order analysis, along with their quantified association strengths, are integrated to construct a novel graph structure—a higher-order relational graph reflecting global topological associations.

[0083] Step S43: Based on the cross-time and cross-sample information contained in the higher-order relation graph and the spatiotemporal graph representation, perform stable core connection pattern fusion extraction processing. By calculating the consistency of the occurrence of connection relationships in sample subsets with different time and genetic backgrounds, and using this as a criterion, fuse and screen local interaction edges and higher-order association edges to obtain the candidate interaction network of chicken gene-pathogen.

[0084] Understandably, this step first systematically evaluates all connections appearing in the local interaction subgraph and the higher-order relation graph. The core metric is calculating the stability of each connection across different consecutive time windows, as well as its reproducibility in subsets of chicken individuals from different genetic backgrounds (e.g., different families or strains). By combining consistency in the temporal dimension with consistency in the genetic dimension, a comprehensive stability score is formed. Subsequently, using this score as the core criterion, direct edges from the local subgraph and indirect connections from the higher-order relation graph are fused and uniformly screened: only those connections that exhibit high consistency in both the temporal and genetic dimensions are retained, such as those with both stability and consistency values ​​greater than a preset threshold (e.g., 0.8).

[0085] The formula for calculating the stability of each connection under different consecutive time windows is as follows:

[0086] ;

[0087] in, For each connection relationship within the time window The emergence of stability under these conditions The number of genetic background groups, This refers to the sequence number of the genetic background group. For the edge In the genetic background group In the middle, spanning all The average weight of each time window, For the edge In the genetic background group In the middle, spanning all The standard deviation of the weights of each time window. It is a very small constant, and greater than 0.

[0088] The formula for calculating the consistency of edge occurrence across different genetic backgrounds is shown below:

[0089] ;

[0090] in, For the edge Consistency score, For the edge In the time window In the middle, spanning all The standard deviation of the weights of each time window. For the edge In the time window In the middle, spanning all The average weight of each time window.

[0091] Step S5: Perform saliency network generation processing based on the candidate interaction network. Based on the preset graph autoencoder discriminant structure that maximizes mutual information, automatically identify and retain edge connections that are more important than a preset threshold to the overall network response pattern of the reconstruction, and obtain the final inferred chicken gene-pathogen interaction network.

[0092] Understandably, this step first uses an encoder to compress the node and edge structure of the candidate interaction network into a low-dimensional hidden representation. Then, a decoder attempts to reconstruct the key response patterns of the original network from this hidden representation, particularly the dependencies between node states. In this process, the mutual information maximization criterion is applied to measure the contribution of the presence or absence of each edge to maintaining a high level of information between the hidden representation and the overall response pattern of the original network, thus automatically calculating a quantified saliency score for each edge. This score directly reflects the importance of the edge connection for accurately reconstructing the overall network dynamics; edges with importance below a preset threshold are considered redundant or noisy connections and are filtered out. Through this automated discrimination mechanism aimed at information reconstruction fidelity, the network formed by the remaining edge connections not only inherits the topology of the candidate network but is also verified at the information theory level as the core framework for maintaining the key interaction patterns of the host-pathogen system. In this step, step S5 includes steps S51, S52, and S53.

[0093] Step S51: Perform edge connection importance quantification processing based on the candidate interaction network. By calculating the contribution of each edge in the network to the overall network representation under the objective of maximizing mutual information, a set of edges with initial saliency weights is obtained.

[0094] Understandably, this step takes the candidate interaction network as input and relies on a graph convolutional network based on maximizing mutual information to evaluate each edge in the network. First, the graph convolutional network maps the network's topology and node features to a low-dimensional latent space representation, and evaluates the impact of the presence or absence of each edge on the amount of information about the original network contained in the latent representation; that is, it calculates the change in mutual information caused by removing or retaining the edge. This change is directly quantified as the edge's contribution to the overall network representation, objectively reflecting the criticality of the connection in information transmission and maintaining the integrity of network functionality. Through this process, each edge in the candidate network is assigned an initial saliency weight, resulting in a set of edges with initial saliency weights.

[0095] The encoding formula of the encoder in this invention is as follows:

[0096] ;

[0097] in, This is the latent space representation matrix output by the encoder, where each row is the embedding vector of a node in the low-dimensional latent space. For the encoder function based on graph convolutional networks, This is a node feature matrix, where each row corresponds to the feature vector of a node (chicken gene or pathogen fragment). Let be the adjacency matrix of the candidate interaction network. These are the parameters of the encoder function for the graph convolutional network.

[0098] The formula for calculating the contribution in this invention is as follows:

[0099]

[0100] ;

[0101] in, For hidden representation With input network structure Mutual information estimates between them To calculate the expectation of a jointly distributed sample, To distinguish joint distributions A discriminator for positive samples. The discriminator first identifies the hidden representations... With input network structure The vectorized representations are concatenated (element-wise multiplication), then non-linearly transformed through one or more fully connected layers, and finally a sigmoid output layer is used to produce a score. To calculate the expectation of the product samples of the marginal distribution, To distinguish marginal distribution products A discriminator for negative samples, the discriminator first classifies the hidden representations... With input network structure The vectorized representations are concatenated (element-wise multiplication), then non-linearly transformed through one or more fully connected layers, and finally a sigmoid output layer is used to produce a score. For the edge The contribution of [the entity / organization] to the overall network representation. To remove the edge The mutual information is obtained from subsequent network calculations.

[0102] Step S52: Perform edge recalibration based on cross-condition stability according to the edge set with initial significance weights. By statistically analyzing the variation of the weights of each edge in different time windows and different genetic background sample subsets, screen out the edges with variation less than a preset threshold in different time windows and different genetic backgrounds to obtain the core edge subset.

[0103] Understandably, this step first divides the data according to different time windows, and then re-evaluates the weight performance of each edge in the subsets of each time window. Simultaneously, it divides the data into subsets based on the bloodline data of individual chickens, representing different genetic backgrounds, and evaluates the edge weights in these subsets as well. Subsequently, the standard deviation of the weights is calculated through statistical analysis. A stable core interaction should maintain relatively constant importance under different observation conditions, exhibiting low variability. The system will filter out edges with variability less than a preset theoretical threshold; these edges are considered connections that maintain stable importance over dynamic time processes and in diverse genetic backgrounds. Through this recalibration process, the final output subset of core edges further filters out connections that, while contributing high information, are conditionally specific.

[0104] The formulas for calculating the edge weights for different time windows are shown below:

[0105] ;

[0106] in, For the edge In the time window The average weight under, For time window The absolute value of the sample set below, For time window The following sample set, Each sample is defined by time points individual chickens The logo, For evaluating latent space representation A discriminator for the degree of matching between the graph structure and the graph structure, wherein the discriminator For individuals At the point of time The latent space representation. For individuals At the point of time Includes edges The vectorized representation of .

[0107] Among them, the method used to evaluate latent space representation The discriminant for the degree of matching between the graph structure and the graph structure is shown below:

[0108] ;

[0109] in, It is the Sigmoid activation function. This is an operation that expands the latent space representation into column vectors and then transposes them. The weight matrix is ​​a learnable matrix. To make individuals At the point of time Includes edges The vectorization of the operation is expanded into a column vector. It is a learnable bias scalar.

[0110] Step S53: Perform network generation processing oriented towards biological pathway consistency based on the core edge subset. With the goal of reconstructing core edge connections through a preset graph autoencoder structure, and perform iterative optimization based on preset chicken immune or metabolic pathways as sparsity constraints. By retaining only connections that conform to the prior pathway structure, the final inferred chicken gene-pathogen interaction network is obtained.

[0111] Understandably, this step begins with a core edge subset that has undergone multiple rounds of stability screening, placing it within an optimization framework that incorporates prior knowledge for final structural refinement. This framework is based on a pre-defined graph autoencoder structure, whose training objective is to accurately reconstruct the connection patterns defined by the core edge subset. Simultaneously, existing knowledge of chicken immune, metabolic, or stress response pathways is encoded as sparsity constraints and introduced into the autoencoder's optimization process, for example, by penalizing new connections incompatible with any known pathway structure. During iterative optimization, the encoder-decoder structure learns the reconstruction of core edges while fine-tuning network connections under the guidance of constraints. This may involve merging redundant paths, pruning isolated connections, and strengthening edges that provide a coherent explanation of prior pathway knowledge. Ultimately, through this joint optimization driven by core pattern reconstruction and constrained by biological consistency, the system outputs a final inferred chicken gene-pathogen interaction network that retains the core data-driven findings while maintaining overall topological consistency with existing biological cognitive frameworks. This completes the transformation from complex data into a knowledge network with biological insights.

[0112] The processing formula for the graph autoencoder is shown below:

[0113] ;

[0114] in, To minimize the loss function, To reconstruct the loss function, The loss function is the path consistency constraint. To balance the parameters, Let be the latent space representation matrix of the encoder output. for function, It is a symmetric normalized adjacency matrix. The node feature matrix, This is the learnable weight matrix of the first layer of the encoder. This is the learnable weight matrix for the second layer of the encoder. Let the transpose of the latent space representation matrix of the encoder output be denoted as . The number of edges in the core edge subset. For node pairs, Represents node pairs Belongs to the core edge subset , Represents the reconstruction matrix The Middle Line number Column elements, The number of edges in the negative sample edge set. Represents node pairs Belonging to the negative sample edge set , This represents the total number of nodes in the graph, including all chicken gene fragments and pathogen genome fragments. For the path mask matrix, This is the final inferred edge set of the chicken gene-pathogen interaction network. The probability threshold is set to 0.5.

[0115] Example 2:

[0116] like Figure 2 As shown, this embodiment provides a system for inferring chicken gene-pathogen interaction networks using graph neural networks. See [link to documentation]. Figure 2 The system includes an acquisition unit 701, a construction unit 702, a learning unit 703, an optimization unit 704, and a generation unit 705.

[0117] The acquisition unit 701 is used to acquire the original transcript sequencing read data of all chicken individuals collected at multiple consecutive time points under the same preset time scale, the original sequencing read data of the pathogen whole genome of the corresponding chicken individuals at the same time point, the original data of the continuous phenotypic records of chicken individuals at all time points, and the original record data of the blood relationship between chicken individuals.

[0118] Construction unit 702 is used to perform cross-species unified graph structure construction processing based on the acquired raw transcript sequencing read data and pathogen whole genome raw sequencing read data to obtain the initial gene-pathogen interaction map of fused temporal signals;

[0119] Learning unit 703 is used to perform dynamic graph representation learning processing under genetic constraints based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records and the original record data of blood relations between chicken individuals, to obtain a spatiotemporal graph representation with genetic generalization ability.

[0120] The optimization unit 704 is used to perform multi-scale topology optimization processing based on the spatiotemporal graph representation to obtain a candidate interaction network of chicken genes and pathogens.

[0121] The generation unit 705 is used to perform saliency network generation processing based on the candidate interaction network. Based on a preset graph autoencoder discriminant structure that maximizes mutual information, it automatically identifies and retains edge connections that are more important than a preset threshold to the overall network response pattern, thereby obtaining the final inferred chicken gene-pathogen interaction network.

[0122] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for inferring chicken gene-pathogen interaction networks using graph neural networks, characterized in that, include: Acquire raw transcript sequencing reads of all chicken individuals collected at multiple consecutive time points under the same preset time scale, raw whole genome sequencing reads of pathogens of the corresponding chicken individuals at the same time point, raw data of continuous phenotypic records of chicken individuals at all time points, and raw data of blood relations between chicken individuals; Based on the acquired raw transcript sequencing reads and pathogen whole genome raw sequencing reads, cross-species unified graph structure construction was performed to obtain the initial gene-pathogen interaction graph of the fusion time-series signal. Based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records, and the original data of blood relations between chicken individuals, dynamic graph representation learning processing under genetic constraints is performed to obtain a spatiotemporal graph representation with genetic generalization ability. Based on the spatiotemporal graph representation, multi-scale topology optimization processing is performed to obtain a candidate interaction network of chicken genes and pathogens; Based on the candidate interaction network, a saliency network generation process is performed. Based on the preset graph autoencoder discriminant structure that maximizes mutual information, edge connections that are more important than a preset threshold to the overall network response pattern are automatically identified and retained, resulting in the final inferred chicken gene-pathogen interaction network. Based on the acquired raw transcript sequencing reads and pathogen whole-genome sequencing reads, a cross-species unified graph structure was constructed to obtain the initial gene-pathogen interaction map of the fusion time-series signal, including: Cross-species sequence fragment numerical characterization processing was performed based on the original transcript sequencing read data and the original whole genome sequencing read data of the pathogen. Specifically, statistical feature vectors were constructed for chicken gene transcript fragments and pathogen genome fragments to obtain preliminary numerical representations of the sequence fragments. The construction of the statistical feature vectors was based on their corresponding sequence composition and read coverage depth. Based on the preliminary numerical representation of the sequence fragments, heterogeneous space unified projection processing is performed. By normalizing the fragment feature vectors from chickens and pathogens and mapping them to a common latent space, unified node embeddings for chickens and pathogens are obtained. Based on the unified node embedding and multiple consecutive time point information, edge connection initialization processing based on time delay co-occurrence is performed. By calculating the time delay mutual information of each node embedded in the cross-sample time series within a preset sliding time window, and using this to dynamically construct the initial edge reflecting the response time sequence relationship, the initial gene-pathogen interaction graph of the fused time sequence signal is obtained.

2. The method for inferring chicken gene-pathogen interaction networks using graph neural networks according to claim 1, characterized in that... Based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records, and the original data of blood relations among chicken individuals, dynamic graph representation learning processing under genetic constraints is performed to obtain a spatiotemporal graph representation with genetic generalization ability, including: Based on the initial gene-pathogen interaction diagram and the original data of the continuous phenotypic records, the node state time sequence alignment process is performed. The graph node state vector at each time point is associated with the multi-dimensional phenotypic observation value of the corresponding chicken individual at that time point to form a model. The phenotypic time sequence is encoded based on a preset recurrent neural network and the temporal evolution path of the node state is back-calibrated to obtain the node state sequence that is synchronously calibrated with the phenotypic dynamics. Based on the node state sequence and the original record data of the blood relationship between the chicken individuals, a subgraph structure constraint processing based on genetic similarity is performed. A genetic distance matrix is ​​constructed by parsing the blood relationship, and positive and negative sample pairs are constructed between the graph data of multiple chicken individuals based on the genetic distance matrix to obtain a set of node and edge representations that satisfy the genetic consistency constraint. Spatiotemporal information fusion processing is performed based on the node and edge representation set. Through a network structure with a preset fusion-gated temporal convolution and graph attention mechanism, the historical state of nodes is aggregated in the time dimension, and the neighbor information is aggregated in the spatial dimension according to the updated edge weights, to obtain a spatiotemporal graph representation with genetic generalization ability.

3. The method for inferring chicken gene-pathogen interaction networks using graph neural networks according to claim 1, characterized in that... Based on the spatiotemporal graph representation, multi-scale topology optimization is performed to obtain a candidate interaction network for chicken genes and pathogens, including: The local direct interaction strength evaluation process is performed based on the spatiotemporal graph representation with genetic generalization ability. The importance weight of each edge in aggregating local neighborhood information is quantified through a preset attention network, and the core edges that constitute a stable local cluster structure are selected accordingly to obtain the enhanced local interaction subgraph. Based on the enhanced local interaction subgraph, high-order path dependency mining is performed. The attention network is used to capture and evaluate the potential regulatory or synergistic effects between nodes through multiple indirect paths, resulting in a high-order relationship graph that reflects the global topological correlation. Based on the cross-time and cross-sample information contained in the higher-order relation graph and the spatiotemporal graph representation, a stable core connection pattern fusion extraction process is performed. By calculating the consistency of the occurrence of connection relationships in sample subsets with different time periods and genetic backgrounds, and using this as a criterion, local interaction edges and higher-order association edges are fused and screened to obtain a candidate interaction network of chicken gene-pathogen.

4. The method for inferring chicken gene-pathogen interaction networks using graph neural networks according to claim 1, characterized in that... The saliency network generation process based on the candidate interaction network includes: The importance of edge connections is quantified based on the candidate interaction network. By calculating the contribution of each edge in the network to the overall network representation under the objective of maximizing mutual information, a set of edges with initial saliency weights is obtained. Based on the edge set with initial significance weights, edge recalibration processing based on cross-condition stability is performed. By statistically analyzing the variation of the weights of each edge in different time windows and different genetic background sample subsets, edges with variation less than a preset threshold in different time windows and different genetic backgrounds are selected to obtain the core edge subset. Based on the core edge subset, a network generation process oriented towards biological pathway consistency is performed. The core edge connections are reconstructed using a preset graph autoencoder structure, and iterative optimization is performed based on preset chicken immune or metabolic pathways as sparsity constraints. By retaining only connections that conform to the prior pathway structure, the final inferred chicken gene-pathogen interaction network is obtained.

5. A system for inferring chicken gene-pathogen interaction networks using graph neural networks, characterized in that, include: The acquisition unit is used to acquire raw transcript sequencing reads of all chicken individuals collected at multiple consecutive time points under the same preset time scale, raw genome sequencing reads of pathogens of the corresponding chicken individuals at the same time point, raw data of continuous phenotypic records of chicken individuals at all time points, and raw data of blood relations between chicken individuals. The building unit is used to perform cross-species unified graph structure construction processing based on the acquired raw transcript sequencing read data and pathogen whole genome raw sequencing read data to obtain the initial gene-pathogen interaction map of fused temporal signals; The learning unit is used to perform dynamic graph representation learning processing under genetic constraints based on the initial gene-pathogen interaction map, the original data of continuous phenotypic records, and the original record data of blood relations between chicken individuals, to obtain a spatiotemporal graph representation with genetic generalization ability. An optimization unit is used to perform multi-scale topology optimization processing based on the spatiotemporal graph representation to obtain a candidate interaction network of chicken genes and pathogens. The generation unit is used to perform saliency network generation processing based on the candidate interaction network. Based on the preset graph autoencoder discriminant structure that maximizes mutual information, it automatically identifies and retains edge connections that are more important than a preset threshold to the overall network response pattern of the reconstruction, and obtains the final inferred chicken gene-pathogen interaction network. The building unit includes: The first construction subunit is used to perform cross-species sequence fragment numerical characterization processing based on the original transcript sequencing read data and the original pathogen whole genome sequencing read data. Specifically, by constructing statistical feature vectors for chicken gene transcript fragments and pathogen genome fragments respectively, a preliminary numerical representation of the sequence fragments is obtained. The statistical feature vectors are constructed based on their corresponding sequence composition and read coverage depth. The second construction subunit is used to perform heterogeneous space unified projection processing based on the preliminary numerical representation of the sequence fragments. By normalizing the fragment feature vectors from chickens and pathogens and mapping them to a common latent space, a unified node embedding of chickens and pathogens is obtained. The third construction subunit is used to perform edge connection initialization processing based on time-delay co-occurrence according to the unified node embedding and multiple consecutive time point information. By calculating the time-delay mutual information of each node embedded in the cross-sample time series within a preset sliding time window, the initial edge reflecting the response time sequence relationship is dynamically constructed to obtain the initial gene-pathogen interaction graph of the fused time sequence signal.

6. The system for inferring chicken gene-pathogen interaction networks using graph neural networks according to claim 5, characterized in that, The learning unit includes: The first learning subunit is used to perform node state temporal alignment processing based on the initial gene-pathogen interaction diagram and the original data of the continuous phenotypic records. It models the association between the graph node state vector at each time point and the multi-dimensional phenotypic observation value of the corresponding chicken individual at that time point. The phenotypic time series is encoded based on a preset recurrent neural network and the temporal evolution path of the node state is back-calibrated to obtain the node state sequence that is synchronously calibrated with the phenotypic dynamics. The second learning subunit is used to perform subgraph structure constraint processing based on genetic similarity according to the original record data of the node state sequence and the blood relationship between the chicken individuals. It constructs a genetic distance matrix by parsing the blood relationship, and constructs positive and negative sample pairs between the graph data of multiple chicken individuals based on the genetic distance matrix to obtain a set of node and edge representations that satisfy the genetic consistency constraint. The third learning subunit is used to perform spatiotemporal information fusion processing based on the node and edge representation set. Through a network structure with a preset fusion-gated temporal convolution and graph attention mechanism, it aggregates the historical state of nodes in the time dimension and aggregates neighbor information in the spatial dimension based on the updated edge weights to obtain a spatiotemporal graph representation with genetic generalization ability.

7. The system for inferring chicken gene-pathogen interaction networks using graph neural networks according to claim 5, characterized in that, The optimization unit includes: The first optimization subunit is used to perform local direct interaction strength evaluation processing based on the spatiotemporal graph representation with genetic generalization ability. It quantifies the importance weight of each edge when aggregating local neighborhood information through a preset attention network, and selects the core edges that constitute a stable local cluster structure to obtain the enhanced local interaction subgraph. The second optimization subunit is used to perform high-order path dependency mining based on the enhanced local interaction subgraph. The attention network is used to capture and evaluate the potential regulatory or synergistic effects between nodes through multiple indirect paths, and to obtain a high-order relationship graph that reflects the global topological correlation. The third optimization subunit is used to perform stable core connection pattern fusion extraction processing based on the cross-time and cross-sample information contained in the higher-order relationship graph and the spatiotemporal graph representation. By calculating the consistency of the occurrence of connection relationships in sample subsets with different time and genetic backgrounds, and using this as a criterion, local interaction edges and higher-order association edges are fused and screened to obtain candidate interaction networks of chicken genes-pathogens.

8. The system for inferring chicken gene-pathogen interaction networks using graph neural networks according to claim 5, characterized in that, The generation unit includes: The first generation subunit is used to perform edge connection importance quantification processing based on the candidate interaction network. By calculating the contribution of each edge in the network to the overall network representation under the objective of maximizing mutual information, an edge set with initial saliency weights is obtained. The second generation subunit is used to perform edge recalibration processing based on cross-condition stability according to the edge set with initial significance weights. By statistically analyzing the degree of variation of the weights of each edge in different time windows and different genetic background sample subsets, edges with a degree of variation less than a preset threshold in different time windows and different genetic backgrounds are selected to obtain the core edge subset. The third generation subunit is used to perform network generation processing oriented towards biological pathway consistency based on the core edge subset. It aims to reconstruct the core edge connections through a preset graph autoencoder structure and performs iterative optimization based on preset chicken immune or metabolic pathways as sparsity constraints. By retaining only the connections that conform to the prior pathway structure, the final inferred chicken gene-pathogen interaction network is obtained.

Images