A soybean root system drought gene regulatory relationship prediction method

Abstract

The application discloses a soybean root system drought gene regulation relationship prediction method, belongs to the technical field of bioinformatics, and aims at solving the problems of insufficient utilization of drought response characteristics and local directed regulation relationship in existing gene regulation relationship prediction and difficulty in effectively sorting candidate directed regulation edges. The application constructs a transcription factor pointing to a target gene directed regulation graph, extracts expression statistical characteristics and drought response characteristics, carries out directed neighborhood aggregation, dimension-by-dimension gate fusion, basic edge scoring and directed regulation context correction, obtains an active regulation score of a candidate directed regulation edge under a drought condition and sorts the candidate directed regulation edge, can make the candidate edge score utilize node expression information, prior regulation direction and local directed context information at the same time, and is suitable for soybean root system drought response gene regulation relationship prediction, plant transcriptome data analysis, candidate gene regulation network construction and other fields.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically involving the prediction technology of gene regulation relationships in soybean root drought. Background Technology

[0002] Drought stress affects gene expression in soybean roots and further alters the regulatory relationships between transcription factors and target genes. With the accumulation of transcriptome sequencing data and plant regulatory databases, inferring gene regulatory networks based on expression data and prior regulatory relationships has become an important technique for analyzing crop stress response mechanisms. Existing methods typically use correlation, mutual information, regression models, tree models, or graph learning models to infer potential regulatory relationships from gene expression profiles, or construct regulatory networks by combining known transcription factor-target gene relationships.

[0003] In recent years, graph neural networks, graph attention networks, multi-view graph models, hypergraph models, and link prediction methods have also been used to construct gene regulatory networks. These methods can utilize the topological relationships between nodes to propagate and update gene node representations, and calculate the probability of regulatory relationships between gene pairs based on node representations. Some methods also introduce second-order neighbors, higher-order graph structures, subgraph representations, or multi-source biological information to improve the completeness of gene regulatory network inference.

[0004] However, most existing methods focus on restoring the general gene regulatory network structure, typically treating gene regulatory relationships as general link prediction objects, and failing to adequately utilize expression response features under specific stress conditions. For the soybean root drought response scenario, merely determining whether regulatory relationships might exist between genes is insufficient to further reflect whether transcription factor-target gene candidate regulatory relationships might be active or enhanced under drought conditions. Furthermore, conventional link prediction methods mainly rely on the representation of nodes at both ends of candidate edges for scoring, failing to adequately consider local directed regulatory structures such as two-hop regulatory paths around candidate directed regulatory edges, shared upstream regulatory relationships, and shared downstream target relationships.

[0005] Therefore, in predicting gene regulatory relationships in soybean roots under drought conditions, a method is needed that can simultaneously utilize soybean root transcriptome expression data, drought response characteristics, prior transcriptional regulatory relationships, and the local directed context of candidate directed regulatory edges, in order to more effectively score and rank candidate transcription factor-target gene regulatory relationships under drought conditions. Summary of the Invention

[0006] To address the shortcomings of existing technologies in predicting soybean root drought gene regulatory relationships, which fail to adequately utilize drought response characteristics and local directed regulatory relationships, and struggle to effectively rank candidate directed regulatory edges, this invention proposes the following solution: A method for predicting gene regulatory relationships in soybean root drought includes: S1. Obtain soybean root transcriptome expression data, transcription factor set, prior transcriptional regulatory relationship data, and trained candidate edge scoring model; S2. Construct a directed regulatory graph of transcription factors pointing to target genes based on prior transcriptional regulation relationship data, with genes as nodes in the directed regulatory graph and candidate regulatory relationships of transcription factors to target genes as directed edges in the directed regulatory graph. S3. Extract the expression statistical features and drought response features of each gene node based on the soybean root transcriptome expression data, and construct a gene node feature matrix based on the expression statistical features and drought response features of each gene node. S4. Input the directed regulation graph and gene node feature matrix into the trained candidate edge scoring model, and perform directed neighborhood aggregation on the gene node feature matrix based on the directed regulation graph to obtain the topological context representation of each gene node. S5. Perform attribute mapping on the gene node feature matrix to obtain the attribute representation of each gene node, and calculate the dimension-wise gate vector based on the topological context representation and attribute representation. S6. The topological context representation and the attribute representation of each gene node are weighted and fused according to the dimension-wise gating vector to obtain the fused node representation of each gene node. S7. Construct candidate directed regulatory edges. The source nodes of the candidate directed regulatory edges belong to the set of transcription factors, and the target nodes of the candidate directed regulatory edges are gene nodes. S8. Generate the basic edge score of the candidate directed control edge based on the source node fusion node representation and the target node fusion node representation of the candidate directed control edge. S9. For the source node and target node of the candidate directed control edge, determine the set of two-hop control path nodes that point from the source node to the target node through the intermediate node, the set of shared upstream control nodes that point to both the source node and the target node, and the set of shared downstream target nodes that both the source node and the target node point to. Construct the directed control context features of the candidate directed control edge based on the set of two-hop control path nodes, the set of shared upstream control nodes, and the set of shared downstream target nodes. S10. Generate the context correction score of the candidate directed control edge based on the directed control context features, and add the context correction score to the basic edge score to obtain the active control score of the candidate directed control edge under drought conditions. S11. Sort the candidate directed regulatory edges according to the active regulatory scores and output the prediction results of the regulatory relationships of candidate genes in soybean root drought response.

[0007] Furthermore, the construction of the directed regulatory graph of transcription factors pointing to target genes described in S2 includes: unifying the soybean gene numbering, removing duplicate edges, and directionally adjusting the transcription factor-target gene regulatory relationship in the prior transcription regulatory relationship data, so that the directed edges point from transcription factors to target genes.

[0008] Furthermore, the construction of the gene node feature matrix described in S3 includes: using the overall expression mean, overall expression standard deviation, and proportion of zero values ​​of each gene node as expression statistical features; using the expression mean of control samples, expression mean of stress samples, logarithmic fold change, and absolute logarithmic fold change as drought response features; and constructing the gene node feature matrix from the expression statistical features and drought response features.

[0009] Furthermore, the directed neighborhood aggregation of the gene node feature matrix based on the directed regulatory graph described in S4 includes: taking the transcription factor nodes connected by directed edges as source nodes, taking the target gene nodes connected by directed edges as target nodes, calculating the directed edge attention weights based on the projection representations of the source nodes and target nodes, and then performing node representation aggregation after normalization within the in-neighbor set of the target nodes.

[0010] Furthermore, the calculation of the dimension-wise gated vector described in S5 includes: performing layer normalization, linear mapping, and nonlinear activation processing on the gene node feature matrix to obtain the attribute representation, and then concatenating the topological context representation and the attribute representation and inputting them into the gated network to obtain the dimension-wise gated vector.

[0011] Furthermore, the basic edge score for generating candidate directed control edges as described in S8 includes: mapping the source node fusion node representation and the target node fusion node representation to the source node role representation and the target node role representation, respectively; constructing candidate edge combination features based on the source node role representation, the target node role representation, their element-wise product and absolute value difference; and generating basic edge scores based on the candidate edge combination features.

[0012] Furthermore, the directed control context features for constructing candidate directed control edges described in S9 include: determining nodes that simultaneously belong to the outgoing neighbor set of the source node and the incoming neighbor set of the target node as intermediate nodes in the two-hop control path node set; determining nodes that simultaneously belong to the outgoing neighbor set of the source node and the target node as nodes in the shared downstream target node set; and determining nodes that simultaneously belong to the incoming neighbor set of the source node and the target node as nodes in the shared upstream control node set.

[0013] Furthermore, the construction of the directed control context features of the candidate directed control edge described in S9 also includes: pooling the fusion node representations in the two-hop control path node set, the shared upstream control node set, and the shared downstream target node set to obtain three types of context pooling representations, and constructing directed control context features based on the three types of context pooling representations and the scale features of the three types of node sets.

[0014] Furthermore, the training method of the trained candidate edge scoring model described in S1 includes: constructing positive and negative candidate edge pairs based on general regulation edge training samples and drought active edge training samples, and jointly training the candidate edge scoring model according to the positive and negative candidate edge pairs, difficult negative candidate edges, weakly enhanced graph views and reconstructed node features.

[0015] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By using soybean root transcriptome expression data, transcription factor set, and prior transcriptional regulatory relationship data as inputs, and constructing a directed regulatory map of transcription factors pointing to target genes, the gene regulatory relationship prediction process retains the directional information between transcription factors and target genes, thereby avoiding the simplification of candidate regulatory relationships into undirected gene pair associations.

[0017] By extracting the expression statistics and drought response features of each gene node, and constructing a gene node feature matrix based on these features, the candidate edge scoring model can simultaneously utilize gene expression distribution information and expression change information under drought conditions, thereby improving the problem of insufficient condition specificity when relying solely on general expression data to predict gene regulatory relationships.

[0018] By performing directed neighborhood aggregation on the gene node feature matrix based on the directed regulatory graph, and then performing dimension-wise gated weighted fusion of the obtained topological context representation and attribute representation, the gene node representation can simultaneously contain prior regulatory topological information and its own expression response information, thereby reducing the limitation of single graph structure representation or single expression feature representation on the scoring of candidate regulatory edges.

[0019] By constructing a set of two-hop control path nodes, a set of shared upstream control nodes, and a set of shared downstream target nodes for the source and target nodes of candidate directed control edges, and constructing directed control context features accordingly, the candidate edge scoring process can utilize the local directed control relationships around the candidate control edge, thereby improving the problem that conventional link prediction only scores based on the representations of the nodes at both ends of the candidate edge.

[0020] By generating context-corrected scores based on directed regulatory context features and adding these scores to the base edge scores to obtain active regulatory scores, local directed regulatory contexts can participate in candidate edge scoring correction, thereby enabling the ranking and output of candidate directed regulatory edges under drought conditions. The "local directed regulatory context" refers to the local directed structure formed around candidate directed regulatory edges, consisting of two-hop mediating paths extracted from the directed regulatory graph, shared downstream target gene relationships, and shared upstream regulatory factor relationships.

[0021] This invention features the ability to predict gene regulatory relationships by combining soybean root transcriptome expression data, drought response characteristics, prior transcriptional regulatory relationships, and local context of candidate directed regulatory edges. It can score and rank candidate transcription factor-target gene regulatory relationships under drought conditions and is applicable to fields such as predicting gene regulatory relationships in soybean root drought response, plant transcriptome data analysis, and candidate gene regulatory network construction. Attached Figure Description

[0022] Figure 1 This is a flowchart of the soybean root drought gene regulation relationship prediction method described in the implementation method; Figure 2 This is a framework diagram of the candidate edge scoring model and training strategy described in the implementation method. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0024] Implementation Method 1 A method for predicting gene regulatory relationships in soybean root drought includes: S1. Obtain soybean root transcriptome expression data, transcription factor set, prior transcriptional regulatory relationship data, and trained candidate edge scoring model; The "prior transcriptional regulatory relationship data" refers to data obtained before the candidate edge scoring model runs, containing existing directed regulatory relationships between transcription factors and target genes. This prior transcriptional regulatory relationship data, after being standardized with soybean gene numbering, having duplicate edges removed, and having its direction adjusted, is used to construct a directed regulatory graph. Here, "prior" indicates that the data is the input basis for the candidate edge scoring model, not the prediction result generated by the model in this run. The "candidate edge scoring model" is a computational model that receives candidate directed gene pairs, gene node representations, and local directed regulatory context features, and outputs the strength of the regulatory association between candidate directed gene pairs. The candidate edge scoring model includes computational processes such as graph attention network encoding, gene feature fusion, basic edge scoring, and directed regulatory context correction.

[0025] S2. Construct a directed regulatory graph of transcription factors pointing to target genes based on prior transcriptional regulation relationship data, with genes as nodes in the directed regulatory graph and candidate regulatory relationships of transcription factors to target genes as directed edges in the directed regulatory graph. The "directed regulatory graph" refers to a graph structure with genes as nodes and directed edges representing the regulatory relationship between transcription factors and target genes, where the direction of the directed edges points from the transcription factors to the target genes.

[0026] S3. Extract the expression statistical features and drought response features of each gene node based on the soybean root transcriptome expression data, and construct a gene node feature matrix based on the expression statistical features and drought response features of each gene node. The “expression statistical features” refer to the statistical quantities calculated based on the expression values ​​of genes in soybean root transcriptome samples, used to describe the overall expression distribution of genes; the “drought response features” refer to the features calculated based on the expression data of control samples and drought stress samples, used to describe the changes in gene expression under drought conditions; the “gene node feature matrix” refers to the matrix formed by arranging the expression statistical features and drought response features of each gene node in a uniform order.

[0027] S4. Input the directed regulation graph and gene node feature matrix into the trained candidate edge scoring model, and perform directed neighborhood aggregation on the gene node feature matrix based on the directed regulation graph to obtain the topological context representation of each gene node. The "directed neighborhood aggregation" refers to the process of receiving and weighting the node representation from its incoming neighbor nodes according to the edge direction in the directed control graph; the "topological context representation" refers to the node representation obtained by the gene node after directed neighborhood aggregation and inter-layer update of the graph attention network. This representation contains the information obtained by the gene node from the upstream control neighborhood of the directed control graph, rather than the directed control graph itself or the original expression features of the gene node.

[0028] S5. Perform attribute mapping on the gene node feature matrix to obtain the attribute representation of each gene node, and calculate the dimension-wise gate vector based on the topological context representation and attribute representation. The "attribute mapping" refers to the process of sequentially performing layer normalization, linear transformation, nonlinear activation, and random inactivation on the original feature vectors of gene nodes, mapping them to the same feature dimension as the topological context representation. The "attribute representation" refers to the node representation obtained after attribute mapping of the gene node feature vectors. This representation mainly includes the gene's own expression statistics and drought response information, but does not include graph neighborhood information obtained from aggregation of other gene nodes. The "dimensional gating vector" refers to the weight vector calculated for each feature dimension of the fused node representation based on the topological context representation and the attribute representation, with each component ranging from 0 to 1.

[0029] S6. The topological context representation and the attribute representation of each gene node are weighted and fused according to the dimension-wise gating vector to obtain the fused node representation of each gene node. The “weighted fusion” refers to the computational process of assigning weights to the topological context representation and attribute representation on the corresponding feature dimensions and then adding them together using a dimension-wise gating vector; the “fusion node representation” refers to the node representation formed after the topological context representation and attribute representation are weighted and fused, which simultaneously contains upstream regulatory information in the directed regulation graph as well as gene expression statistics and drought response information.

[0030] S7. Construct candidate directed regulatory edges. The source nodes of the candidate directed regulatory edges belong to the set of transcription factors, and the target nodes of the candidate directed regulatory edges are gene nodes. The “candidate directed regulatory edge” refers to the ordered node pair from candidate transcription factor nodes to candidate target gene nodes, which is judged and sorted by the candidate edge scoring model. The term “candidate” indicates that the directed regulatory edge belongs to the object to be evaluated, but does not mean that it has been verified by wet experiments or that there is necessarily a real regulatory relationship.

[0031] S8. Generate the basic edge score of the candidate directed control edge based on the source node fusion node representation and the target node fusion node representation of the candidate directed control edge. The "basic edge score" refers to the real value calculated based on the source node representation and target node representation of the candidate directed control edge without sigmoid transformation, in the absence of explicit introduction of local directed control context correction. It is used to characterize the initial control association strength of the candidate directed control edge.

[0032] S9. For the source node and target node of the candidate directed control edge, determine the set of two-hop control path nodes that point from the source node to the target node through the intermediate node, the set of shared upstream control nodes that point to both the source node and the target node, and the set of shared downstream target nodes that both the source node and the target node point to. Construct the directed control context features of the candidate directed control edge based on the set of two-hop control path nodes, the set of shared upstream control nodes, and the set of shared downstream target nodes. The “two-hop control path node set” refers to the set of intermediate nodes in a two-hop directed path formed by the source node of a candidate directed control edge pointing to an intermediate node, and the intermediate node continuing to point to the target node. The “shared downstream target node set” refers to the set of downstream gene nodes that are simultaneously pointed to by the source node and the target node of the candidate directed regulatory edge. The “shared upstream control node set” refers to the set of upstream control nodes that simultaneously point to the source node and the target node of the candidate directed control edge. The "directed control context feature" refers to the candidate edge feature formed by splicing the pooled representations and corresponding scale features of the above three types of node sets, which is used to describe the local directed control structure around the candidate directed control edge.

[0033] S10. Generate the context correction score of the candidate directed control edge based on the directed control context features, and add the context correction score to the basic edge score to obtain the active control score of the candidate directed control edge under drought conditions. The “context correction score” refers to the residual value calculated based on the fusion node representations at both ends of the candidate directed control edge and the directed control context features, which is used to supplement and correct the basic edge score. The context correction score can be positive or negative, representing an increase or decrease in the basic edge score, respectively. The “active control score” refers to the final edge log probability obtained by adding the basic edge score and the context correction score in the drought active edge prediction task, without sigmoid transformation, which is used to rank candidate directed control edges. The active probability of the control edge under drought conditions is obtained by performing a Sigmoid transformation on the active control score.

[0034] S11. Sort the candidate directed regulatory edges according to the active regulatory scores and output the prediction results of the regulatory relationships of candidate genes in soybean root drought response.

[0035] The “prediction results of candidate gene regulatory relationships in soybean root drought response” refers to the set of candidate directed regulatory edges formed after sorting by active regulatory scores. The prediction results include at least candidate transcription factor nodes, candidate target gene nodes and corresponding active regulatory scores, which are used to indicate the priority of candidate regulatory relationships, but do not indicate that the corresponding candidate regulatory relationships have been confirmed as direct regulatory relationships.

[0036] Through the above steps, soybean root transcriptome expression data, prior transcriptional regulatory relationships, and the local context of candidate directed regulatory edges are involved in the same scoring process, forming a continuous processing chain from data input, node representation, edge scoring to sorting output.

[0037] Furthermore, the construction of the directed regulatory graph of transcription factors pointing to target genes described in S2 includes: unifying the soybean gene numbering, removing duplicate edges, and directionally adjusting the transcription factor-target gene regulatory relationship in the prior transcription regulatory relationship data, so that the directed edges point from transcription factors to target genes.

[0038] By standardizing the numbering, removing duplicate edges, and aligning the directions of prior transcriptional regulatory relationship data, the input regulatory relationships can participate in subsequent calculations according to unified gene nodes and unified edge directions, reducing the impact of inconsistent data representation on the construction of directed regulatory graphs.

[0039] Furthermore, the construction of the gene node feature matrix described in S3 includes: using the overall expression mean, overall expression standard deviation, and proportion of zero values ​​of each gene node as expression statistical features; using the expression mean of control samples, expression mean of stress samples, logarithmic fold change, and absolute logarithmic fold change as drought response features; and constructing the gene node feature matrix from the expression statistical features and drought response features.

[0040] By using the overall expression mean, overall expression standard deviation, and proportion of zero values ​​as expression statistical features, and incorporating the expression mean of control samples, expression mean of stress samples, logarithmic fold change, and absolute logarithmic fold change as drought response features into the gene node feature matrix, each gene node retains both expression distribution information and drought-related expression change information during subsequent encoding.

[0041] Furthermore, the directed neighborhood aggregation of the gene node feature matrix based on the directed regulatory graph described in S4 includes: taking the transcription factor nodes connected by directed edges as source nodes, taking the target gene nodes connected by directed edges as target nodes, calculating the directed edge attention weights based on the projection representations of the source nodes and target nodes, and then performing node representation aggregation after normalization within the in-neighbor set of the target nodes.

[0042] By normalizing and aggregating the target node using its in-neighbor set as the range, the target node primarily receives representation information from upstream regulatory nodes, ensuring that the topological context representation matches the regulatory direction of transcription factors towards the target gene.

[0043] Furthermore, the calculation of the dimension-wise gated vector described in S5 includes: performing layer normalization, linear mapping, and nonlinear activation processing on the gene node feature matrix to obtain the attribute representation, and then concatenating the topological context representation and the attribute representation and inputting them into the gated network to obtain the dimension-wise gated vector.

[0044] By jointly calculating the dimension-wise gated vector using topological context representation and attribute representation, topological and attribute information in different dimensions can be adjusted separately, making it easier to obtain node representations that are more suitable for candidate edge scoring.

[0045] Furthermore, the basic edge score for generating candidate directed control edges as described in S8 includes: mapping the source node fusion node representation and the target node fusion node representation to the source node role representation and the target node role representation, respectively; constructing candidate edge combination features based on the source node role representation, the target node role representation, their element-wise product and absolute value difference; and generating basic edge scores based on the candidate edge combination features.

[0046] By constructing source node role representations and target node role representations separately, and combining element-wise product and absolute value difference to form candidate edge combination features, the basic edge scores can reflect the role differences and representation matching relationships of the two ends of the candidate directed control edge.

[0047] Furthermore, the directed control context features for constructing candidate directed control edges described in S9 include: determining nodes that simultaneously belong to the outgoing neighbor set of the source node and the incoming neighbor set of the target node as intermediate nodes in the two-hop control path node set; determining nodes that simultaneously belong to the outgoing neighbor set of the source node and the target node as nodes in the shared downstream target node set; and determining nodes that simultaneously belong to the incoming neighbor set of the source node and the target node as nodes in the shared upstream control node set.

[0048] By determining three types of node sets according to the outgoing neighbor set and incoming neighbor set of the directed edge, the two-hop control path node set, the shared downstream target node set, and the shared upstream control node set can respectively reflect the indirect control relationship, common target relationship, and common upstream control relationship around the candidate directed control edge.

[0049] Furthermore, the construction of the directed control context features of the candidate directed control edge described in S9 also includes: pooling the fusion node representations in the two-hop control path node set, the shared upstream control node set, and the shared downstream target node set to obtain three types of context pooling representations, and constructing directed control context features based on the three types of context pooling representations and the scale features of the three types of node sets.

[0050] By pooling the fused node representations in the three types of node sets and combining them with scale features, the directed control context features can simultaneously reflect the local context node representations and context scale information.

[0051] Furthermore, the training method of the trained candidate edge scoring model described in S1 includes: constructing positive and negative candidate edge pairs based on general regulation edge training samples and drought active edge training samples, and jointly training the candidate edge scoring model according to the positive and negative candidate edge pairs, difficult negative candidate edges, weakly enhanced graph views and reconstructed node features.

[0052] The “difficult negative candidate edge” refers to a negative candidate edge that is not marked as a positive control edge in the general control edge training candidate set, but the candidate edge scoring model gives a high control connection score and is therefore easily confused with a positive control edge. The difficult negative candidate edge is used for candidate edge ranking training and is not added as a positive control edge to the information propagation structure of the directed control graph. The “weakly enhanced graph view” refers to a view with slight graph structure perturbation formed by randomly discarding edges from the same directed control graph; The “reconstructed node features” refer to the node features obtained by feature decoding based on the fused node representation and used for comparison with the original gene node features.

[0053] By using positive and negative candidate edge pairs, difficult negative candidate edges, weakly enhanced graph views, and reconstructed node features to jointly train the candidate edge scoring model, the candidate edge scoring model can simultaneously learn edge classification, edge ranking, representation consistency under graph perturbation, and node feature preservation during the training phase.

[0054] The method described in this embodiment can be executed by a processor calling a computer program, which can be stored in a computer storage medium. When the computer program is executed by the processor, the above-mentioned method for predicting the gene regulatory relationship of soybean root drought can be implemented.

[0055] Implementation Method 2 This implementation method uses soybean root transcriptome expression data, transcription factor set, and prior transcriptional regulatory relationship data to predict candidate gene regulatory relationships under drought conditions as a specific example. Combined with... Figure 1The method described in this embodiment is executed in the following order: data acquisition, directed regulatory graph construction, gene node feature construction, node representation generation, candidate edge scoring, context correction, and sorting output. Figure 2 The trained candidate edge scoring model includes a gene feature fusion module, a directed regulatory context module, and a curriculum-based ranking enhancement and regularization training strategy.

[0056] The "gene feature fusion module" refers to a model component that extracts topological context representation and attribute representation from the directed regulatory graph and gene node feature matrix, respectively, and then uses a one-dimensional gating vector to weight and fuse the two representations to output the fused node representation.

[0057] The aforementioned "curriculum-based ranking enhancement and regularization training strategy" refers to incorporating the following into the training process of the candidate edge scoring model: regulation of connection classification loss, drought condition activity classification loss, node feature reconstruction loss, positive and negative candidate edge ranking loss, hard negative sample ranking loss, and graph view. Figure 1 The training method involves jointly optimizing the consistency loss and gradually increasing the weight of the hard negative sample ranking loss with each training round. The term "course-based" means that the hard negative sample ranking loss does not use the full weight from the first training round, but gradually increases to the target weight according to a preset function after the preset starting round.

[0058] In S1, soybean root transcriptome expression data, transcription factor set, prior transcriptional regulatory relationship data, and a trained candidate edge scoring model were acquired. Specifically, soybean root transcriptome expression data characterizes gene expression status in root samples; the transcription factor set limits the source node range of candidate directed regulatory edges; prior transcriptional regulatory relationship data provides existing regulatory relationships between transcription factors and target genes; and the trained candidate edge scoring model outputs basic scores, contextual corrections, and active regulatory scores for candidate directed regulatory edges.

[0059] In S2, a directed regulatory graph of transcription factors pointing to target genes is constructed based on prior transcriptional regulatory relationship data. Genes are treated as nodes in the directed regulatory graph, and candidate regulatory relationships between transcription factors and target genes are treated as directed edges. Specifically, the transcription factor-target gene regulatory relationships in the prior transcriptional regulatory relationship data are standardized by soybean gene numbering, duplicate edges are removed, and directions are adjusted so that directed edges point from transcription factors to target genes. The resulting directed regulatory graph is used for subsequent directed neighborhood aggregation, candidate directed regulatory edge construction, and directed regulatory context feature construction.

[0060] In S3, expression statistical features and drought response features of each gene node are extracted from soybean root transcriptome expression data, and a gene node feature matrix is ​​constructed from these features. Specifically, expression statistical features characterize the expression distribution of genes in the transcriptome expression data, while drought response features characterize the expression changes of genes in drought-related samples. Each row in the gene node feature matrix corresponds to a gene node, and each column corresponds to either an expression statistical feature or a drought response feature.

[0061] In this embodiment, the expression statistical features include the overall expression mean, overall expression standard deviation, overall expression median, lower quartile, upper quartile, minimum, maximum, and proportion of zero values. The drought response features include the expression mean of the control sample, the expression mean of the stressed sample, the expression standard deviation of the control sample, the expression standard deviation of the stressed sample, the logarithmic fold change, and the absolute logarithmic fold change. These eight expression statistical features and six drought response features together constitute a 14-dimensional gene node feature matrix.

[0062] Let the prior directed regulatory graph, the gene node feature matrix, and the original feature vector of the i-th gene node be respectively:

[0063] in, This represents a priori directed control graph. Represents a set of gene nodes. This represents the set of directed edges pointing from prior transcription factors to target genes. Represents the feature matrix of gene nodes. Represents the set of real numbers. Indicates the number of gene nodes. This represents the original node feature dimension, which is 14 in this embodiment. Indicates the first The original feature vectors of each gene node. Representation matrix The All rows and columns This represents the gene node index.

[0064] In S4, the directed regulatory graph and gene node feature matrix are input into the trained candidate edge scoring model. Based on the directed regulatory graph, directed neighborhood aggregation is performed on the gene node feature matrix to obtain the topological context representation of each gene node. Combined with... Figure 2 In Part A of the candidate edge scoring model, the gene feature fusion module receives gene expression features and prior directed regulatory graphs, and obtains the topological context representation through directed neighborhood aggregation.

[0065] The aforementioned “prior directed regulatory graph” refers to the directed regulatory graph constructed based on prior transcriptional regulatory relationship data, which is used to define the direction of information propagation between gene nodes.

[0066] Specifically, directed edge-connected transcription factor nodes are used as source nodes, and directed edge-connected target gene nodes are used as target nodes. The directed edge attention weights are calculated based on the projected representations of the source and target nodes, and then normalized within the in-neighbor set of the target node before node representation aggregation. Through this process, the topological context representation of the target node is obtained by aggregating the representations of its upstream regulatory nodes, thus aligning the topological context representation with the regulatory direction of the transcription factor towards the target gene.

[0067] For the directed graph encoding process, the first... Layer Each attention point will first focus on the node The input representation is mapped to the hidden space:

[0068] in, Represents a node In the Layer Projection representation under each attention head Indicates the first Layer The linear transformation matrix of each attention head. Represents a node In the The input representation of the layer, Indicates the node index. Indicates the index of the graph attention network layer. This indicates the attention head index.

[0069] Subsequently, based on the projected representations of the source and target nodes, the unnormalized attention score for the directed edge i→j is calculated:

[0070] in, Represents a directed edge In the Layer Unnormalized attention scores in each attention head and These represent the indexes of the source node and the target node, respectively. and Let these represent the learnable attention parameter vectors at the source node and the target node, respectively, with subscripts... and These represent the source node and the target node, respectively. and Represent the projected representations of the source node and the target node, respectively, with superscripts. This represents the transpose of a vector. This represents the activation function of the linear unit with leakage correction.

[0071] When using prior edge weights At that time, the unnormalized attention score is represented as:

[0072] in, Represents a priori directed edge The right to the border, This embodiment indicates that positive numbers are prevented from being logarithmic with respect to zero. , Represents the natural logarithm function. This indicates taking the larger of two numbers.

[0073] When prior edge weights are not used, the edge weight term in the above formula is omitted.

[0074] For the same target node All incoming edges are defined in their set of incoming neighbors. Perform softmax normalization on top:

[0075] in, Represents a directed edge In the Layer Normalized attention weights in each attention head Represents the target node The set of neighbors, This represents the index of the node being traversed within the current summation range. Indicates that by node Point to target node The directed edge in the first Layer Unnormalized attention scores in each attention head This represents the natural exponential function. This represents summing all incoming neighbors of the target node.

[0076] During the training phase, random deactivation is applied to the normalized attention weights to obtain the actual aggregate weights:

[0077] in, This represents the attention weights that actually participate in the aggregation after random deactivation during the training phase. The tilde above indicates that this value has undergone random deactivation. This indicates a random deactivation operation, which is enabled during the training phase and disabled during the inference phase.

[0078] Subsequently, the target node In the The neighborhood aggregation under each attention head is represented as follows:

[0079] in, Represents the target node In the Layer The neighborhood aggregation representation under each attention head. The aggregation results of each attention head are concatenated along the feature dimension, and the 1st attention head is added. Learnable bias terms of a layer:

[0080] in, Represents a node In the The layer completes the multi-head splicing and adds the offset representation. Indicates the attention head index, Indicates the total number of attention heads. This indicates concatenation along the feature dimension. Indicates the first Learnable bias vectors of a layer.

[0081] Subsequently, ELU activation and layer normalization are performed sequentially on the multi-head splicing results:

[0082] in, Represents a node After activation and layer normalization, the first layer that has not yet been added with residual connections In layer representation, ELU(·) denotes the exponential linear unit activation function, and LN(·) denotes the layer normalization operation. This indicates the next network layer.

[0083] When the dimensions of the input representation of the previous layer are the same as those of the output representation of the current layer, a residual connection is added after layer normalization:

[0084] in, Represents a node In the The final output of the layer represents, Represents a node In the The layer's input representation, with the plus sign indicating element-wise residual addition when the input and output dimensions are consistent.

[0085] When the dimensions of the input representation of the previous layer are inconsistent with the dimensions of the output representation of the current layer, residual addition is not performed:

[0086] Based on the above node feature projection, edge attention scoring, in-neighbor normalization aggregation, and inter-layer update process, the candidate edge scoring model obtains the topological context representation of each gene node in the directed regulatory graph.

[0087] This implementation employs a two-layer, dual-head GAT, with each attention head having a hidden dimension of 64, resulting in a node representation dimension of 128 after multi-head concatenation. The first layer maps 14-dimensional node features to 128 dimensions; since the input and output dimensions differ, residual addition is not performed. The second layer has 128-dimensional input and output, thus incorporating residual connections. This design preserves the node representation of the previous layer while propagating upstream control information, mitigating information attenuation caused by prior network noise, missing edges, and multi-layer neighborhood aggregation.

[0088] In S5, attribute mapping is performed on the gene node feature matrix to obtain the attribute representation of each gene node, and a dimensional gating vector is calculated based on the topological context representation and the attribute representation. Specifically, layer normalization, linear mapping, and nonlinear activation are performed on the gene node feature matrix to obtain the attribute representation. The topological context representation and the attribute representation are concatenated and input into the gating network to obtain the dimensional gating vector. The dimensional gating vector has a corresponding dimension to the topological context representation and the attribute representation, and is used to adjust the contribution of the topological context representation and the attribute representation on different feature dimensions.

[0089] For gene nodes The original feature vector of the gene node Layer normalization is performed, and the layer is projected onto the same feature space as the topological context representation through linear and nonlinear mappings:

[0090] in, Represents a node The attribute is represented by the superscript. This indicates that the representation originates from the characteristics of the gene node itself. Represents a node The original feature vector, and These represent the weight matrix and bias vector of the attribute feature projection layer, respectively, with subscripts... Indicates the attribute feature projection branch, This represents a nonlinear mapping consisting of Gaussian error linear unit activation and random deactivation.

[0091] The nonlinear mapping can be expanded as follows:

[0092] Where GELU(·) represents the Gaussian error linear unit activation function, and Dropout(·) represents the random deactivation operation.

[0093] Subsequently, the topological context representation and attribute representation are concatenated along the feature dimension and input into the gating network to compute the dimension-wise gating vector:

[0094] in, Represents a node Dimensionally gated vectors, Represents a node The topological context representation, Represents a node The attribute representation, and These represent the weight matrix and bias vector of the gated layer, respectively, with subscripts... Indicates the gate layer. Represents an sigmoid function, such that Each component is between 0 and 1. ,in, This implementation method represents the dimension of the fused node representation. , It represents the set of real numbers.

[0095] In S6, the topological context representation and the attribute representation of each gene node are weighted and fused according to the dimension-wise gating vector to obtain the fused node representation of each gene node. The fused node representation simultaneously carries the topological context information of the directed regulatory graph and the expression attribute information of the gene node itself, and serves as the basis for the node representation construction of the subsequent candidate directed regulatory edge base score and directed regulatory context features.

[0096] The dimension-wise weighted fusion is represented as:

[0097] in, Represents a node The fusion node is indicated by the superscript. This indicates that the representation is obtained by fusing the representations of the two types of nodes. Represents the complementary weights of the dimension-wise gated vectors. This represents element-wise multiplication. When the value of a certain dimension of the gating vector is close to 1, that dimension focuses more on preserving the topological context information learned by GAT; when its value is close to 0, it focuses more on preserving the gene's own expression statistics and drought response characteristics.

[0098] Combination Figure 2In part A, the attribute representation is gated and then fused with the topological context representation dimension by dimension to obtain the fused node representation of each gene node.

[0099] In S7, candidate directed regulatory edges are constructed. The source nodes of the candidate directed regulatory edges belong to the set of transcription factors, and the target nodes are gene nodes. The candidate directed regulatory edges are constructed according to the direction from the transcription factors to the gene nodes, so that the source end of the candidate edge corresponds to the potential regulatory end, and the target end of the candidate edge corresponds to the potential regulated end.

[0100] Candidate directed control edges are defined as follows:

[0101] in, Indicates that by node Pointing to node Candidate directed regulatory edges, Indicates the candidate source node. Indicates the candidate target node. Represents a set of transcription factor nodes. Represents a set of gene nodes. This indicates that the set of transcription factor nodes is a subset of the set of gene nodes. This definition restricts the source node of a candidate edge to belong to the set of transcription factors, while the target node can be any gene node in the prior network.

[0102] For the task The predicted log odds of candidate directed control edges are output by the edge prediction function corresponding to the task.

[0103] in, Indicates task For candidate edges The side log odds of the output sigmoid function before transformation. Indicates task The corresponding edge prediction function, Indicates the task type. This indicates a general control and forecasting task. This indicates a drought-prone area forecasting task. Indicates a candidate directed control edge. and These represent the indexes of the candidate edge source node and the target node, respectively.

[0104] Convert the predicted log odds to predicted probabilities using the Sigmoid function:

[0105] in, Indicates task For candidate edges The output predicted probability, when hour, This indicates that the existence of a general control edge is probable, when hour, This represents the probability of a candidate regulatory edge being active under drought conditions. Specifically, This represents the probability of the existence of a candidate general control edge. This indicates the probability of the regulatory edge being active under drought conditions.

[0106] In S8, basic edge scores for candidate directed control edges are generated based on the source node fusion node representation and the target node fusion node representation. Specifically, the source node fusion node representation and the target node fusion node representation are mapped to the source node role representation and the target node role representation, respectively. Candidate edge combination features are constructed based on the source node role representation, the target node role representation, their element-wise product, and the difference in absolute value. Basic edge scores are then generated based on these candidate edge combination features. The source node role representation characterizes the representation state of the source node as the control end in the candidate directed control edge, and the target node role representation characterizes the representation state of the target node as the controlled end.

[0107] For the task The merged node representing the source and target nodes represents the input task-specific role projection branch:

[0108] in, Indicates task Next source node The role of transcription factors is indicated. Indicates task Next target node The target gene role indicates, Indicates task The corresponding transcription factor role nonlinear projection branch, Indicates task The corresponding target gene role nonlinear projection branch, subscript and These represent the roles of transcription factors and target genes, respectively. Candidate edge combination features are constructed based on the source node role representation, the target node role representation, their element-wise product, and the absolute value difference between the two.

[0109] in, Indicates task Next candidate edge Candidate edge combination features, This represents the role of the source node. This represents the role of the target node. Indicates feature splicing, Represents element-wise product. The two types of roles are represented by the absolute value of the element-wise subtraction. The basic edge score of the candidate directed control edge is calculated by the task-specific basic edge scoring network.

[0110] in, Indicates task The candidate edge basic log odds have not yet been explicitly introduced into the local directed control context. (superscript) Indicates the basic edge score. Indicates task The corresponding basic edge multilayer perceptron scoring network. In S9, for the source node and target node of the candidate directed control edge, the set of two-hop control path nodes pointing from the source node to the target node via intermediate nodes, the set of shared upstream control nodes pointing to both the source node and the target node, and the set of shared downstream target nodes jointly pointed to by the source node and the target node are determined respectively. The directed control context features of the candidate directed control edge are constructed based on the set of two-hop control path nodes, the set of shared upstream control nodes, and the set of shared downstream target nodes. Combined with... Figure 2 In Part B, the directed regulation context module extracts a set of two-hop regulation path nodes, a set of shared upstream regulation nodes, and a set of shared downstream target nodes around the candidate directed regulation edges.

[0111] For candidate directed control edges The set of nodes for the two-hop control path is represented as:

[0112] in, Indicate candidate edges The set of nodes in the two-hop control path, index Indicates the type of two-hop control path. Represents a node The neighbors gathered. Represents a node The set of neighbors, This represents the intersection of sets, and the nodes that belong to that intersection. Able to form The two-hop control path. The shared downstream target node set is represented as:

[0113] in, Represents a node and nodes A shared set of downstream target nodes that point to each other, with indexes... Indicates shared downstream target type. and Representing nodes respectively and nodes The neighbors gathered. and These represent the indexes of the candidate edge source node and the target node, respectively.

[0114] The shared set of upstream control nodes is represented as:

[0115] in, Indicates pointing to the same node and nodes A shared set of upstream control nodes, index Indicates shared upstream regulation type, and Representing nodes respectively and nodes The set of neighbors.

[0116] To limit the size of the context set, three types of context sets are processed:

[0117] in, Indicates a context type index. , and These represent three types of contexts: two-hop regulatory pathways, shared downstream targets, and shared upstream regulation. Indicate candidate edges The The original context node collection of the class. Indicates the number of samples after processing. Class context node collection, This indicates that when the number of nodes in the set exceeds 8, 8 nodes are sampled; otherwise, all nodes are retained. The tilde above indicates that the set has already been sampled.

[0118] Pooling is performed on the fused node representations in the processed context set to obtain the first... Class context pooling representation:

[0119] in, Indicate candidate edges The Class context pooling representation, Indicates a context type index. Indicates the first The class context collection is the processed collection of context nodes. This indicates the number of nodes in the set. This represents the index of the node being traversed within the current summation range. Represents a node The fusion node is represented as follows: This is used to avoid the denominator being zero when the set is empty. When the set is empty, the summation result is the zero vector.

[0120] The number of nodes after logarithmic transformation is used to describe the size characteristics of various context sets:

[0121] in, Indicate candidate edges The Class context size characteristics This represents the natural logarithm function, with the constant 1 used to ensure that the input to the logarithm function is 1 when the set is empty.

[0122] The three types of context pooling representations and their scale features are concatenated into a directed control context feature:

[0123] in, Indicate candidate edges The directional regulatory context features, , and These represent three types of context pooling representations: two-hop regulation path, shared downstream target, and shared upstream regulation, respectively. , and These represent the scale characteristics of the three types of contexts mentioned above.

[0124] In S10, context correction scores for candidate directed control edges are generated based on the directed control context features. These context correction scores are then added to the base edge scores to obtain the active control score of the candidate directed control edge under drought conditions. Combined with... Figure 2 The context correction process involves the directed control context features entering the candidate edge scoring model. The context correction score is used to supplement and correct the basic edge score. The basic edge score reflects the matching relationship between the node representations at both ends of the candidate directed control edge, while the context correction score reflects the corrective effect of the local directed control relationship around the candidate directed control edge on the edge score.

[0125] The context correction score is calculated using a task-specific context residual network:

[0126] in, Indicates task The following is the context-corrected log-probability generated by the locally oriented control context. Indicates task The corresponding context-sensitive multilayer perceptron residual network, The expression "merging the two ends" indicates that the absolute value is taken after subtracting each element.

[0127] The task-specific score of a candidate directed control edge is obtained by adding the basic edge score and the context-corrected score:

[0128] in, Indicates task Next candidate edge The final edge logarithmic odds, Indicates task The candidate edge basic log odds have not yet been explicitly introduced into the local directed control context. (superscript) Indicates the final edge score. Indicates task The context-corrected log-probability is generated by the local directed control context. In this embodiment, the active control score is the score of the candidate directed control edge before the sigmoid transformation, and the predicted probability is the probability value obtained after performing a sigmoid transformation on the active control score. When sorting the candidate directed control edges, each candidate directed control edge adopts the same scoring caliber.

[0129] Convert the final edge score into a predicted probability:

[0130] in, Indicates task Next candidate edge The final predicted probability, Represents an sigmoid function, when The output generally controls the probability of the edge existing, when The probability of active regulation edge under drought conditions is output.

[0131] In S11, candidate directed regulatory edges are sorted according to their active regulatory scores, and the predicted results of candidate gene regulatory relationships in soybean root drought response are output. The candidate directed regulatory edges in the sorting results are arranged according to their active regulatory scores, which are used to represent the candidate priority of different transcription factor-target gene candidate regulatory relationships under drought conditions.

[0132] Combination Figure 2In Part C of this embodiment, the training method for the trained candidate edge scoring model includes: constructing positive and negative candidate edge pairs based on general regulatory edge training samples and drought-active edge training samples; and jointly training the candidate edge scoring model based on the positive and negative candidate edge pairs, difficult negative candidate edges, weakly enhanced graph views, and reconstructed node features. General regulatory edge training samples are used to train the candidate edge scoring model to identify basic regulatory relationships; drought-active edge training samples are used to train the candidate edge scoring model to identify active candidate regulatory relationships under drought conditions; difficult negative candidate edges are used to participate in candidate edge ranking training; weakly enhanced graph views are used to participate in representation consistency training under graph structure perturbation; and reconstructed node features are used to constrain the fusion node representation to retain gene node attribute information.

[0133] For the edge-pair odds vector output by the model and its binary label vector The standard binary cross-entropy loss can be expressed as:

[0134] in, Represents the standard binary cross-entropy loss, subscript Represents the binary cross-entropy. Represents the edge logarithmic probability vector; This represents the corresponding binary label vector. This indicates the number of training edges involved in the loss calculation. This represents the training edge sample index. This represents the logarithmic probability of the nth training edge. This represents the binary label of the nth training edge. .

[0135] Standard binary cross-entropy loss is typically used for edge prediction tasks.

[0136] in, This represents the classification loss for a typical control edge prediction task. This represents the edge log-probability vector for a general edge prediction task. This represents a binary label vector for a general control edge prediction task.

[0137] The drought active edge prediction branch uses a weighted binary cross-entropy loss:

[0138] in, Represents the weighted binary cross-entropy loss, with subscript […]. This represents the weighted binary cross-entropy. Represents the log-odds vector of the edges; This represents the corresponding binary label vector. This indicates the number of training edges involved in the loss calculation. This represents the training edge sample index. Indicates the weight of positive samples, subscript This represents a positive sample, used to adjust the relative contribution of positive and negative classes to the loss when needed. At that time, the loss degenerates into the standard binary cross-entropy loss.

[0139] The classification loss for drought-active edges is expressed as:

[0140] in, This represents the classification loss of the drought active edge prediction task. This represents the edge log probability vector for the drought active edge prediction task. This represents the binary label vector for the drought active edge prediction task.

[0141] To preserve gene node features in the constraint node representation, the candidate edge scoring model reconstructs the input features through a feature decoder. The node feature reconstruction loss is expressed as:

[0142] in, Represents the node feature reconstruction loss. This represents the original node feature dimension, which is 14 in this embodiment. Represents a node The original feature vector, The feature decoder is a node. The reconstructed feature vector, Represents a set of gene nodes. Indicates the number of gene nodes. Let L2 norm be denoted, and squared represent the reconstruction error using the square.

[0143] For the task Positive and negative edges are extracted from the corresponding training edges to form... The loss for pairwise sorting of edge pairs is expressed as:

[0144] in, Indicates task The pairwise sorting loss, Indicates task The number of positive and negative edge pairs involved in the sorting. Indicates the index of positive and negative edge pairs. and They represent the first One positive edge and one negative edge task The logarithmic odds below, superscript " "and" "Represents positive and negative edges respectively, The sorting interval is 0.2 in this embodiment. Represents a soft positive function. Represented by natural constant as the base, with It is an exponential function of an exponent.

[0145] The hard negative sample ranking loss is expressed as:

[0146] in, This represents the loss for difficult negative sample ranking. This indicates the number of positive edges paired with difficult negative edges that are involved in the loss calculation. Indicates the first A general regulation of the logarithmic probability of the positive side, Indicates the first The log-odds of a difficult negative candidate edge.

[0147] The difficult negative sample ranking loss uses course weights. Represented as:

[0148] in, This represents the course weights of the difficult negative sample ranking loss corresponding to the t-th training round. Indicates the current training round. This indicates the training epoch in which the ranking of difficult negative samples begins to take effect, in this implementation method. , This indicates the number of warm-up rounds used from the start of the course weighting until it increases to 1, in this implementation method. , This indicates taking the smaller value so that the course weight does not exceed 1.

[0149] In this embodiment, , Therefore, the difficulty ranking weight at the start of round 10 is... The value increases linearly during subsequent warm-up rounds, reaching 1 by the 39th round. This strategy allows the model to first learn a relatively stable discrimination space using basic classification and random sorting objectives, and then gradually focus on the indistinguishable negative edges, thus balancing training stability and sorting accuracy of high-scoring candidate regions.

[0150] Considering the potential presence of missing and noisy edges in the prior graph, the curriculum-based ranking enhancement and regularization training strategy constructs two weakly enhanced views by slightly perturbing the graph structure. In this implementation, the edge drop rate is applied independently to the same prior graph. Random edges are discarded and input into an encoder with shared parameters to obtain node representations. and First, L2 normalization is performed on both sets of representations:

[0151] in, and Representing nodes respectively In weak enhancement map view and weak enhancement graph view The nodes below represent the superscript. and This represents the graph view obtained by discarding two independent random edges. and These represent the results of L2 normalization for the corresponding nodes.

[0152] In the set of nodes participating in consistency constraints Compute weak graph view Figure 1 Induced loss:

[0153] in, Indicates weak graph view Figure 1 Induced loss, subscript This represents a consistency constraint. This represents the set of nodes participating in the consistency constraints. This indicates the number of nodes in the set. This represents the inner product of two normalized node representations.

[0154] The overall training objective of the candidate edge scoring model is expressed as:

[0155] in, Indicates the overall training objective. This represents the general control edge classification loss. Indicates the classification loss of drought-active edges. Represents the node feature reconstruction loss. This represents the loss for difficult negative sample ranking. Indicates weak graph view Figure 1 Sexual damage, The course weights represent the difficulty of ranking negative samples. This represents the loss from pairwise sorting of general control edges. represents the pairwise sorting loss of active edges in drought, and 0.5, 1, 0.02, 0.05, 0.02, 0.10 and 0.01 represent the preset weight coefficients of the corresponding loss terms in the overall training objective.

[0156] Based on the above general control edge classification loss, drought active edge classification loss, node feature reconstruction loss, pairwise ranking loss, difficult negative sample ranking loss, and weak graph view... Figure 1 Based on consistency loss and overall training objectives, the candidate edge scoring model updates its parameters during the training phase.

[0157] The embodiments of the present invention have been described above. These embodiments are used to help understand the technical solutions of the present invention and should not be construed as limiting the scope of the technical solutions of the present invention. Equivalent substitutions or conventional modifications made by those skilled in the art to the embodiments without departing from the technical concept of the present invention should all fall within the scope of the technical solutions of the present invention.

Claims

1. A method for predicting gene regulatory relationships in soybean root drought, characterized in that, include: S1. Obtain soybean root transcriptome expression data, transcription factor set, prior transcriptional regulatory relationship data, and trained candidate edge scoring model; S2. Construct a directed regulatory graph of transcription factors pointing to target genes based on prior transcriptional regulation relationship data, with genes as nodes in the directed regulatory graph and candidate regulatory relationships of transcription factors to target genes as directed edges in the directed regulatory graph. S3. Extract the expression statistical features and drought response features of each gene node based on the soybean root transcriptome expression data, and construct a gene node feature matrix based on the expression statistical features and drought response features of each gene node. S4. Input the directed regulation graph and gene node feature matrix into the trained candidate edge scoring model, and perform directed neighborhood aggregation on the gene node feature matrix based on the directed regulation graph to obtain the topological context representation of each gene node. S5. Perform attribute mapping on the gene node feature matrix to obtain the attribute representation of each gene node, and calculate the dimension-wise gate vector based on the topological context representation and attribute representation. S6. The topological context representation and the attribute representation of each gene node are weighted and fused according to the dimension-wise gating vector to obtain the fused node representation of each gene node. S7. Construct candidate directed regulatory edges. The source nodes of the candidate directed regulatory edges belong to the set of transcription factors, and the target nodes of the candidate directed regulatory edges are gene nodes. S8. Generate the basic edge score of the candidate directed control edge based on the source node fusion node representation and the target node fusion node representation of the candidate directed control edge. S9. For the source node and target node of the candidate directed control edge, determine the set of two-hop control path nodes that point from the source node to the target node through the intermediate node, the set of shared upstream control nodes that point to both the source node and the target node, and the set of shared downstream target nodes that both the source node and the target node point to. Construct the directed control context features of the candidate directed control edge based on the set of two-hop control path nodes, the set of shared upstream control nodes, and the set of shared downstream target nodes. S10. Generate the context correction score of the candidate directed control edge based on the directed control context features, and add the context correction score to the basic edge score to obtain the active control score of the candidate directed control edge under drought conditions. S11. Sort the candidate directed regulatory edges according to the active regulatory scores and output the prediction results of the regulatory relationships of candidate genes in soybean root drought response.

2. The method according to claim 1, characterized in that, S2 describes the construction of a directed regulatory graph of transcription factors pointing to target genes, which includes: unifying soybean gene numbering, removing duplicate edges, and directional adjustment of the transcription factor-target gene regulatory relationship in the prior transcription regulatory relationship data, so that the directed edges point from transcription factors to target genes.

3. The method according to claim 1, characterized in that, The construction of the gene node feature matrix described in S3 includes: using the overall expression mean, overall expression standard deviation, and proportion of zero values ​​of each gene node as expression statistical features; using the expression mean of control samples, expression mean of stress samples, logarithmic fold change, and absolute logarithmic fold change as drought response features; and constructing the gene node feature matrix from the expression statistical features and drought response features.

4. The method according to claim 1, characterized in that, S4 describes the directed neighborhood aggregation of gene node feature matrices based on directed regulatory graphs, which includes: taking transcription factor nodes connected by directed edges as source nodes, target gene nodes connected by directed edges as target nodes, calculating the directed edge attention weights based on the projection representations of source nodes and target nodes, and then normalizing the node representations within the in-neighbor set of the target nodes before performing node representation aggregation.

5. The method according to claim 1, characterized in that, The calculation of the dimension-wise gated vector described in S5 includes: performing layer normalization, linear mapping, and nonlinear activation on the gene node feature matrix to obtain the attribute representation; concatenating the topological context representation and the attribute representation and inputting them into the gated network to obtain the dimension-wise gated vector.

6. The method according to claim 1, characterized in that, S8 describes generating the basic edge score for candidate directed control edges, which includes: mapping the source node fusion node representation and the target node fusion node representation to the source node role representation and the target node role representation, respectively; constructing candidate edge combination features based on the source node role representation, the target node role representation, their element-wise product and absolute value difference; and generating basic edge scores based on the candidate edge combination features.

7. The method according to claim 1, characterized in that, The directed control context features for constructing candidate directed control edges as described in S9 include: determining nodes that simultaneously belong to the outgoing neighbor set of the source node and the incoming neighbor set of the target node as intermediate nodes in the two-hop control path node set; determining nodes that simultaneously belong to the outgoing neighbor set of the source node and the target node as nodes in the shared downstream target node set; and determining nodes that simultaneously belong to the incoming neighbor set of the source node and the target node as nodes in the shared upstream control node set.

8. The method according to claim 7, characterized in that, The construction of the directed control context features of the candidate directed control edge described in S9 further includes: pooling the fusion node representations in the two-hop control path node set, the shared upstream control node set, and the shared downstream target node set to obtain three types of context pooling representations, and constructing directed control context features based on the three types of context pooling representations and the scale features of the three types of node sets.

9. The method according to claim 1, characterized in that, The training method of the trained candidate edge scoring model described in S1 includes: constructing positive and negative candidate edge pairs based on general regulation edge training samples and drought active edge training samples; and jointly training the candidate edge scoring model based on positive and negative candidate edge pairs, difficult negative candidate edges, weakly enhanced graph views, and reconstructed node features.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 9.

Images