A cancer gene identification method based on a multiplexing heterogeneous graph neural network

Abstract

The embodiment of the present application relates to a cancer gene identification method based on a multiplexing heterogeneous graph neural network, comprising: updating the initial features of each node in the cancer heterogeneous graph network, and outputting the updated features; generating a corresponding node embedding dictionary accordingly; the node embedding dictionary comprises the embedding representation of each node in the cancer heterogeneous graph network after learning under different relationship subgraphs, each node corresponds to multiple embedding vectors, and each embedding vector corresponds to a different semantic relationship; embedding vectors are taken as inputs of an attention fusion module, and a node representation dictionary obtained after attention fusion is output; each node in the node representation dictionary corresponds to a fused embedding vector; each node in the node representation dictionary of the last layer attention fusion module is input into a binary classifier with a corresponding comprehensive embedding, so as to obtain a probability value, and when the probability value is greater than a preset threshold, a cancer gene is predicted.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for cancer gene identification based on multiplexed heterogeneous graph neural networks. Background Technology

[0002] A model called Transformer-based Graph Representation Learning (TREE) is specifically designed for the classification and identification of cancer genes. This model innovatively combines multi-omics data with the topological structure of biological networks to improve the accuracy and interpretability of cancer gene prediction. In its implementation, TREE first constructs a normalized graph structure, integrating rich information from multi-omics gene data, including mutations, expression, methylation, and copy number variations, into node features. To address the computational challenges posed by large-scale graph inputs, TREE employs subgraph sampling techniques, extracting subgraphs containing a fixed number of nodes from the original graph through random walks, effectively reducing computational complexity. During gene representation learning, TREE fully leverages the advantages of the Transformer architecture, incorporating structural information from biological networks, such as node degree and network distance, into the model. It uses a self-attention mechanism to accurately capture long-distance dependencies between genes and employs a multi-channel aggregation module to fuse gene representations learned from different subgraphs, ultimately generating embedded gene representations. Ultimately, leveraging gene embedding representations, TREE achieves classification and prediction of whether a gene is a cancer gene through a fully connected layer containing a single neuron and an activation function, such as the sigmoid function. Furthermore, TREE significantly enhances its ability to model gene network structures by introducing positional encoding to integrate gene degree information and utilizing a co-attention mechanism to integrate network distance information. When dealing with complex heterogeneous networks, TREE can model the interdependencies between different molecules by designing meta-paths and fully utilize the semantic relationships of heterogeneous networks through random walk sampling of subgraphs, thereby achieving accurate identification of cancer genes.

[0003] The disadvantages of existing technologies are as follows:

[0004] 1. High computational resource consumption: When processing large-scale graph inputs, computational resources are consumed in large quantities, and it is difficult to effectively model the global structural information of the input graph.

[0005] 2. Strong dependence on background data: The explanatory model relies on background data. The selection and quality of the background data have a significant impact on the accuracy and reliability of the explanatory results. If the background data is not representative or differs greatly from the test data, the explanatory results may deviate from the actual decision-making logic of the model, affecting the understanding and trust in the model's behavior. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a cancer gene identification method based on multiplexed heterogeneous graph neural networks, thereby solving the problems existing in the prior art.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a cancer gene identification method based on a multiplexed heterogeneous graph neural network, comprising:

[0008] The initial features of each node in the cancer heterogeneous graph network are updated, and the updated features are output.

[0009] Based on the updated features of the nodes, a corresponding node embedding dictionary is generated; the node embedding dictionary includes the learning embedding representation of each node in the cancer heterogeneous graph network under different relation subgraphs, each node corresponds to multiple embedding vectors, and each embedding vector corresponds to a different semantic relation;

[0010] The embedding vector is used as input to the attention fusion module, and the output is a node representation dictionary after attention fusion. The node representation dictionary is a key-value pair geometry, where the key is the node identifier and the value is the comprehensive embedding representation of the node after attention fusion. Each node in the node representation dictionary corresponds to a fused embedding vector. The embedding vector represents the embedding representation of the node under different relational subgraphs.

[0011] Each node in the dictionary of the node representation of the previous attention fusion module is used to embed the corresponding comprehensive data into the binary classifier to obtain a probability value. When the probability value is greater than a preset threshold, it is predicted to be a cancer gene.

[0012] In one possible implementation, the method further includes the following preceding steps:

[0013] Extract node expression data and clinical data for various cancers from the TCGA dataset;

[0014] Based on clinical information, the samples were divided into experimental and control groups, and the relative changes in gene expression between the two groups were calculated.

[0015] Determine the node and edge types of the cancer heterogeneous graph network;

[0016] Using relative expression changes as the initial feature of genes, the network topology and clinical features are captured by propagating node features through heterogeneous graph neural networks.

[0017] For each gene, the fusion unit is embedded in different relational subgraphs to generate a comprehensive representation;

[0018] The training and test sets are divided according to cancer type or sample proportion, and a cancer heterogeneous graph network is obtained after training.

[0019] In one possible implementation, updating the initial features of each node in the cancer heterogeneous graph network and outputting the updated features specifically includes:

[0020] For each type of relation Corresponding subgraph G (r) =(v, ε) (r) ), for all nodes initial features pass Update; among them, This represents the neighbors of node v under relation r; ∈ represents a learnable or predefined hyperparameter; MLP (r) (·) is a relation-specific multilayer perceptron.

[0021] In one possible implementation, each learning module consists of two fully connected layers, with each layer connected via... To prevent overfitting; among them, MLP (r) (·) represents a relation-specific multilayer perceptron; x is the input feature vector of the node; It is the weight matrix of the first fully connected layer, used for the r-th relation, mapping the input features from the input dimension to the hidden layer dimension; φ is the weight matrix of the second fully connected layer, which maps the hidden layer features to the output dimension; φ is a non-linear activation function used to introduce non-linearity between fully connected layers, enabling the model to learn more complex feature relationships; the Dropout layer is used to prevent overfitting.

[0022] In one possible implementation, the embedding vector is used as input to the attention fusion module, and the output is a dictionary representation of the nodes after attention fusion.

[0023] For node v under different relations, generate three sets of projections: in, It is a learnable parameter, d a For the attention dimension;

[0024] pass Calculate the attention score of a node across different semantics; where α (r) β represents the original attention score of node v under relation r. (r) It is the normalized attention weight of node v with respect to relation r, which represents the importance of this semantic in the final representation of the node;

[0025] The outputs from each header are concatenated and linearly transformed into a unified node representation, retaining key terms from various semantic categories. The final fused representation is as follows:

[0026] pass Residual linking and layer normalization; where, f res (·) represents the nonlinear residual path function, which consists of two fully connected layers and the GELU activation function; This represents the representation of node v after multi-head attention fusion; This represents the original input representation of node v.

[0027] In one possible implementation, the step of taking each node in the dictionary of the node representation of the previous layer attention fusion module, embedding the corresponding comprehensive data into the input binary classifier, and obtaining a probability value, and predicting it as a cancer gene when the probability value is greater than a preset threshold, specifically includes:

[0028] pass Calculate the predicted probability distribution of node v for each category; where the fused node is represented as... W c ∈R C×d C is the number of categories, and b is a learnable bias term;

[0029] The final prediction is in, Node v is predicted to be a cancer gene. A value of 0 indicates that node v is predicted to be a non-cancer gene;

[0030] pass Determine the loss function; where y v，c This is the encoding of the actual label. If node v belongs to category c, this item is 1; otherwise, it is 0. v，c It is the probability that node v is predicted to be of class c.

[0031] In one possible implementation, the key of the node embedding dictionary is the node identifier, and the value is the set of embedding vectors corresponding to the node under all relational subgraphs.

[0032] Secondly, the present invention provides a cancer gene identification device based on a multiplexed heterogeneous graph neural network, the device comprising:

[0033] Multiplexed GNNs are used to update the initial features of each node in the cancer heterogeneous graph network and output the updated features.

[0034] Based on the updated features of the nodes, a corresponding node embedding dictionary is generated; the node embedding dictionary includes the learning embedding representation of each node in the cancer heterogeneous graph network under different relation subgraphs, each node corresponds to multiple embedding vectors, and each embedding vector corresponds to a different semantic relation;

[0035] The attention fusion module takes the embedding vector as input and outputs a node representation dictionary after attention fusion. The node representation dictionary is a key-value pair geometry, where the key is the node identifier and the value is the comprehensive embedding representation of the node after attention fusion. Each node in the node representation dictionary corresponds to a fused embedding vector. The embedding vector represents the embedding representation of the node under different relational subgraphs.

[0036] The classifier module is used to take each node in the dictionary of the node representation of the previous attention fusion module, embed the corresponding comprehensive data into the binary classifier, and obtain the probability value. When the probability value is greater than the preset threshold, it is predicted to be a cancer gene.

[0037] Thirdly, the present invention provides a computer server, comprising: a memory, a processor, and a transceiver;

[0038] The processor is used to couple with the memory, read and execute instructions in the memory to implement the cancer gene identification method based on multiplexed heterogeneous graph neural network as described in any of the first aspects;

[0039] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0040] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, the computer program being executed by a processor of the cancer gene identification method based on multiplexed heterogeneous graph neural networks as described in any one of the first aspects.

[0041] By applying the cancer gene identification method based on multiplexed heterogeneous graph neural networks provided in this invention, a multi-path heterogeneous graph modeling mechanism based on relation type division is proposed for cancer biological networks with multiple node types, such as mRNA and lncRNA. Each relation type is constructed as an independent "subgraph channel," preserving the independent semantics of various molecular interactions and effectively avoiding relation confounding. An attention fusion mechanism is designed to fuse node representations under various semantic subgraphs in multiple channels. This mechanism dynamically models the importance of different semantic paths to node representations. It learns a unique query, key, and value representation for each relation type, calculates its semantic contribution score, and uses this score as weight to perform weighted fusion of multi-path embeddings, thereby achieving context-aware modeling of nodes across semantic perspectives. Compared to traditional averaging or static weighting strategies, this mechanism has stronger adaptive modeling capabilities and can accurately extract the most discriminative semantic features. Meanwhile, the attention fusion module also introduces residual connections and layer normalization mechanisms. By using the residual paths formed by two layers of nonlinear MLPs, it enhances information flow capabilities, effectively mitigating gradient vanishing and feature degradation issues when multiple layers are stacked, and improving the model's nonlinear modeling ability and generalization robustness. Layer normalization further stabilizes the model training process, suppresses feature drift, and strengthens the collaborative expression of multi-channel semantic information. The EmhGNN model proposed in this study has made some progress in the field of heterogeneous graph cancer gene recognition, providing a new tool for accurate prediction of cancer genes and a deeper understanding of cancer molecular mechanisms. By constructing and analyzing heterogeneous graphs containing various biomolecules and their interactions, EmhGNN can effectively integrate multi-source heterogeneous data, uncover complex relationships between biomolecules, and thus improve the accuracy and efficiency of cancer gene recognition. The EmhGNN model employs a multiplexing structure and attention mechanism, which can not only model different types of nodes and relationships separately, but also capture complex interactions between biomolecules through feature fusion. Attached Figure Description

[0042] Figure 1 A flowchart for constructing a heterogeneous graph neural network provided in an embodiment of the present invention;

[0043] Figure 2 This is an architecture diagram of cancer gene recognition based on multiplexed heterogeneous graph neural networks provided in an embodiment of the present invention;

[0044] Figure 3 The flowchart illustrates the cancer gene identification method based on a multiplexed heterogeneous graph neural network provided in this embodiment of the invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0046] Before implementing the cancer gene identification method based on multiplexed heterogeneous graph neural networks proposed in this application, it is necessary to construct a cancer heterogeneous graph network. A cancer graph network refers to a network integrating PPIs and non-coding RNAs, containing different node types such as coding genes, lncRNAs, and miRNAs. Heterogeneous refers to the graph being composed of nodes of different types. Figure 1 The flowchart for constructing a heterogeneous graph neural network provided in the embodiments of the present invention is as follows: Figure 1 As shown, constructing a cancer heterogeneous graph network includes the following steps:

[0047] Step 110: Extract node expression data and clinical data for various cancers from the TCGA dataset;

[0048] Specifically, node expression data and clinical data from 5236 samples of 15 types of cancer, such as breast cancer, pancreatic cancer, gastric cancer, and thyroid cancer, were extracted from the TCGA dataset. Based on 10 clinical information items—age, sex, survival status, overall survival, cancer stage, tumor classification, regional lymph node lesions, metastasis status, treatment method, and health status—the samples were grouped. The relative expression change of each gene in each clinical feature group was calculated to construct biomarker identification features for cancer. The constructed biomarker identification features are the 10 clinical information items: age, sex, survival status, overall survival, cancer stage, tumor classification, regional lymph node lesions, metastasis status, treatment method, and health status. Each piece of information is ultimately a number, and each gene represents one of the 15 types of cancer, with 10 clinical information features for each type of cancer, resulting in a total of 150 dimensions of features.

[0049] Step 120: Based on clinical information, the samples are divided into experimental and control groups, and the relative expression changes of genes between the two groups are calculated.

[0050] Specifically, node characteristics were calculated based on relative gene expression changes rather than absolute values. The samples were stratified according to ten key clinical factors to determine the relative expression changes of each gene: (1) Age: 68 years old was the dividing line, with younger patients in the experimental group and older patients in the control group. (2) Gender: Males were in the experimental group and females in the control group. (3) Survival status: Deceased patients were in the experimental group and survivors in the control group. (4) Overall survival time: ≤426 days in the experimental group and >426 days in the control group. (5) Cancer stage: Late stage (III and IV) characterized by local invasion and distant metastasis was in the experimental group, and early stage (I and II) was in the control group. (6) Tumor classification (T): T3 and T4 tumors that have penetrated the serosa or invaded adjacent structures were in the experimental group, and T1 and T2 tumors were in the control group. (7) Regional lymph node metastasis (N): N1 to N3 were in the experimental group, and N0 was in the control group. (8) Metastasis (M) status: M1 (distant metastasis) was in the experimental group, and M0 (no distant metastasis) was in the control group. (9) Treatment method: Radiotherapy was used as the experimental method. (10) Health status: The tumor group was the experimental group, and the normal group was the control group. After grouping the samples, the relative expression change of each gene in the clinical characteristic group was calculated.

[0051] Model testing: By inputting the heterogeneous graph network and node features into the model, training and testing can be performed. Specifically, the heterogeneous graph and node features are packaged into a Python class called `dgl.heterogaph`, which is then input into the model for training and testing.

[0052] Step 130: Determine the node types and edge types of the cancer heterogeneous graph network;

[0053] Specifically, node types are constructed, including core nodes and extended nodes. Core nodes include gene nodes and disease nodes. Gene nodes include protein-coding genes, miRNAs, and lncRNAs. Disease nodes include cancer types, such as BRCA and LUAD. Extended nodes include drug nodes, such as targeted drugs, and pathway nodes, such as the KEGG pathway.

[0054] We model edge types and relationships, including biomolecular interaction edges, clinical association edges, and co-expression edges.

[0055] Biomolecular interactions include PPI edges and lncRNA-miRNA edges. PPI edges are protein-protein interactions obtained from the STRING database with a confidence level >0.7. lncRNA-miRNA interactions are predicted regulatory relationships using starBase or lncRNA2Target. miRNA-mRNA edges are also included, representing targeting relationships validated experimentally based on miRTarBase. Clinically relevant edges include gene-disease edges, connecting differentially expressed genes to corresponding cancer types. Co-expression edges are calculated based on the correlation coefficient between genes, such as |r|>0.6 and p-value<0.05.

[0056] The initial feature encoding of gene nodes uses relative expression changes as input features, such as a logFC matrix, with dimensions of gene number × cancer type number. Extended features include concatenable gene function annotations, such as GO term embeddings or epigenetic features, such as DNA methylation. Disease nodes use one-hot encoding or clinical feature embeddings, such as numerical encoding of TNM staging. Independent graph convolutional layers are designed for each edge type, such as PPI and lncRNA-miRNA.

[0057] Step 140: Using relative expression changes as the initial feature of the gene, the network topology and clinical features are captured by propagating node features through a heterogeneous graph neural network.

[0058] Specifically, clinical features, such as stage and age, are mapped to the same dimension as gene embedding through MLP, and then spliced ​​together and input into the next layer.

[0059] Step 150: For each gene, the embedding of the fusion device in different relational subgraphs generates a comprehensive representation;

[0060] For each gene, its embedding vector under different relational subgraphs, such as PPI and lncRNA-miRNA, is extracted and fused using gating mechanisms or Transformer.

[0061] Step 160: Divide the training set and test set according to cancer type or sample ratio, and obtain the cancer heterogeneous graph network after training.

[0062] Specifically, when dividing by cancer type, 1-2 cancer types, such as LUAD and COAD, are retained as independent test sets, and the rest are used for training. When dividing by sample proportion, 70% is used for training, 15% for validation, and 15% for testing to ensure that samples from the same patient do not cross sets. After training, a cancer heterogeneous graph network can be obtained.

[0063] Figure 2 This is an architecture diagram of cancer gene recognition based on multiplexed heterogeneous graph neural networks provided in an embodiment of the present invention; as shown below. Figure 2As shown, this application includes a multiplexed GNN learning module and an attention fusion module. In the multiplexed GNN learning module, for cancer heterogeneous graph networks, multi-semantic subgraph modeling based on node type pairs is performed, including sub-... Figure 1 :IncRNA-protein, sub- Figure 2 The model is structured as follows: miRNA-protein, up to subgraph n, where different type pairs are represented as a single subgraph. After learning through GIN node embedding, the data is input into the attention fusion module. Finally, the output of the attention fusion module is passed through a classifier to obtain the classifier result.

[0064] Figure 3 This is a flowchart illustrating a cancer gene identification method based on a multiplexed heterogeneous graph neural network, provided in an embodiment of the present invention. The following is in conjunction with... Figure 2 The technical solution of the present invention will be described with reference to specific embodiments. Figure 3 As shown, this application includes the following steps:

[0065] Step 310: Update the initial features of each node in the cancer heterogeneous graph network and output the updated features;

[0066] Specifically, for each node in the cancer heterogeneous graph network, based on at least two of the different relational subgraphs to which it belongs, including protein-protein interaction subgraphs, lncRNA-miRNA regulation subgraphs, and miRNA-mRNA regulation subgraphs, initial feature updates are performed using relation-specific graph neural network layers to generate independent embedding representations of each node under each relational subgraph, forming a node embedding dictionary; the key of the node embedding dictionary is the node identifier, and the value is the set of embedding vectors corresponding to the node under all relational subgraphs.

[0067] In step 310, a multiplexed GNN representation learning module is used. This module's design is based on the "multiplexing" mechanism, which constructs independent subgraph channels for different types of node pairs in the graph, with each subgraph channel modeled using a separate graph neural network. This application uniformly uses a Graph Isomorphism Network (GIN) as the graph convolutional unit in this module, which possesses strong graph discriminative ability and structural sensitivity, enabling it to better capture higher-order semantics in local subgraphs. Different types of node pairs refer to node pairs in the input heterogeneous graph between gene-coding RNA, lncRNA, and miRNA types, such as gene-coding gene, gene-lncRNA, etc. Specifically, step 310 includes the following:

[0068] For each type of relation Corresponding subgraph G (r) =(ν, ε) (r) ), for all nodes initial features To perform an update, the GIN update rules are as follows:

[0069]

[0070] The formula represents the generation of a new representation of node v in the next layer and relation r by performing a nonlinear transformation on the above information using MLP(r). This represents the representation of node v at the (k+1)th level and under relation r, which is the embedding vector. This represents the current representation of node v at level k and under relation r. This represents the representation of neighbor node u at the k-th level and under relation r. This represents the neighbors of node v under relation r; ∈ is a learnable or predefined hyperparameter, which is a constant; MLP (r) (·) is a relation-specific multilayer perceptron used to perform nonlinear transformations on aggregated information.

[0071] In the implementation, each GIN module consists of two fully connected layers, with non-linear activations such as ReLU or GELU and Dropout used between each layer to prevent overfitting.

[0072]

[0073] Among them, MLP (r) (·) represents a relation-specific multilayer perceptron. x is the input feature vector of the node. It can be the original feature or the feature after processing by the previous layer. This is the weight matrix of the first fully connected layer, used for the r-th relation. It maps the input features from the input dimension to the hidden layer dimension. This is the weight matrix of the second fully connected layer, which maps the hidden layer features to the output dimension. φ is a non-linear activation function, such as ReLU or GELU. It is used to introduce non-linearity between fully connected layers, enabling the model to learn more complex feature relationships. The Dropout layer is used to prevent overfitting. It randomly discards the output of some neurons during training, thereby improving the model's generalization ability.

[0074] This module can be stacked in multiple layers, supporting the acquisition of structure-aware representations in deeper semantic spaces. For each relation type... The output embedding is denoted as: These features will be integrated in the next module. Among them, H (r) H is a set representing the final output embeddings of all nodes v under relation type r. In other words, H (r) Let v be the set of all node embeddings under relation r, and v be the nodes in the graph. V is the set of all nodes in the graph. It includes the embedding of each node v in the graph under the last layer and relation r.

[0075] Step 220: Generate a corresponding node embedding dictionary based on the updated features of the nodes; the node embedding dictionary includes the learning embedding representation of each node in the cancer heterogeneous graph network under different relation subgraphs, each node corresponds to multiple embedding vectors, and each embedding vector corresponds to a different semantic relation;

[0076] Step 230: The embedding vector is used as the input to the attention fusion module, and the output is a node representation dictionary after attention fusion; the node representation dictionary is a key-value pair geometry, where the key is the node identifier and the value is the comprehensive embedding representation of the node after attention mechanism fusion; each node in the node representation dictionary corresponds to a fused embedding vector; the embedding vector represents the embedding representation of the node under different relation subgraphs;

[0077] Specifically, the attention fusion module, preceding it, multiplexes the node embedding dictionary output by the GNN representation learning module. This dictionary contains the embedding representations of each node in the heterogeneous graph after learning by the GNN under different relation subgraphs. Each node corresponds to multiple embedding vectors, each corresponding to a different semantic relation. For example, the embedding representation of a node under relation r1 is hr1, the embedding representation under relation r2 is hr2, and so on. These embedding vectors serve as the input to the attention fusion module.

[0078] Because different types of edges in cancer biological networks, such as lncRNA-miRNA, miRNA-miRNA, and protein-protein, carry drastically different biological semantic information, simply summing or averaging the node representations corresponding to each type of edge will lead to semantic confusion and information loss, failing to accurately depict the true dependencies between biomolecules. To address this challenge, this application introduces a multi-head attention-driven attention fusion module to dynamically model semantic importance between different relationship channels and achieve adaptive fusion of multiple semantic representations of nodes in heterogeneous graphs. In cancer biological heterogeneous graphs, different relationship types reflect various regulatory mechanisms and functional semantics between biomolecules. For example, lncRNA and miRNA may form a ceRNA regulatory pair, while protein-protein relationships often reflect pathway synergy and signal transduction. To avoid the loss of semantic information caused by averaging different relationship channels, this invention introduces a multi-head attention mechanism to dynamically model the semantic importance of each type of relationship in the node representation, achieving semantic context-aware node fusion expression. For example, when a protein node is connected to multiple key cancer proteins through PPI relationships, the attention mechanism will assign higher weights to the PPI channels, thereby strengthening the role of the relationship in the final representation and achieving semantically driven embedding expression.

[0079] The attention fusion module introduces a multi-head attention mechanism to generate three sets of representations—query (Q), key (K), and value (V)—for each type of relation, and uses self-attention to calculate the relevance scores between different semantics. Each attention head focuses on different dimensions in the semantic space, thereby achieving parallel modeling and weighted aggregation of information for multi-relation representations, improving representation accuracy and generalization ability. Let node v be in each relation... The following is represented as Aggregation is performed using multi-head attention. For each head, three sets of projections are constructed as follows:

[0080]

[0081] in, It is a learnable parameter, W q The learnable weight matrix W used to generate the query vector. k The learnable weight matrix W used to generate the key vectors. v Learnable weight matrices are used to generate value vectors. These weight matrices are optimized during training via backpropagation to learn how to extract useful information from node embeddings. v (r) d represents the embedding vector of node v under relation r. a For the attention dimension.

[0082] Calculate the attention score of a node across different semantics:

[0083]

[0084] Where, α (r) It is the raw attention score of node v under relation r, which represents its importance under the current semantics and is usually unnormalized; Let v be the query vector for node v under relation r. This represents the key vector of node v under relation r. The scaling factor is the dimension of the query vector and the key vector. Scaling is typically used to prevent the dot product from becoming too large, which can lead to vanishing or exploding gradients. β (r) ∑ is the normalized attention weight of node v with respect to relation r, representing the importance of that semantic meaning in the final representation of the node. r′ exp(α (r′) ) represents the exp(α) for all relations r′. (r′) Summing them up gives the result as the normalized denominator.

[0085] The outputs from each header are concatenated and linearly transformed into a unified node representation, preserving key information from various relational semantics. The final fused representation is as follows:

[0086]

[0087] in, This represents the attention-weighted representation, or integrated embedding, of node v. It is obtained by summing the value vectors under different relations according to the attention weight β(r). It integrates the information of node v across all relations. r∈R indicates that relation type r belongs to the relation set R, where R is the set of all possible relation types in the graph. β (r) This represents the attention weight of relation r. It is a scalar value indicating the relative importance of relation r in generating the comprehensive representation of node v. β (r) Typically, the values ​​are normalized using the softmax function to ensure that the sum of the weights of all relationships is 1. This represents the value vector of node v under relation r. It is generated from the original embedding of node v through some transformation (such as a linear transformation) and is specifically designed to represent the features of node v under relation r.

[0088] To further improve the stability of model training and its nonlinear modeling capabilities, the attention fusion module also introduces residual connections and layer normalization mechanisms:

[0089]

[0090] in, This represents the final output representation of node v, or the normalized embedding. This is the node representation after layer normalization and residual connection processing, combining attention-weighted information and the original input information. res (·) represents the nonlinear residual path function, which consists of two fully connected layers and the GELU activation function. This represents the representation of node v after multi-head attention fusion. This represents the original input representation of node v (from the previous layer).

[0091] The residual path consists of two fully connected layers with activation functions, such as GELU. This not only helps alleviate the gradient vanishing problem caused by deep structures but also enhances the model's ability to model nonlinear complex relationships. The normalized structure effectively suppresses feature drift and enhances the ability to achieve cross-channel semantic alignment. Finally, the attention fusion module outputs a unified node representation that perceives semantic context, providing more expressive and interpretable embedded features for downstream cancer gene recognition tasks.

[0092] As one of the core innovations of the EmhGNN model, this attention fusion module breaks through the static processing bottleneck of traditional heterogeneous graph GNN in multi-relation fusion, realizes multi-relation modeling of heterogeneous biological information, and is particularly suitable for challenging scenarios in actual cancer networks where there are asymmetric relationships and complex semantics, thus improving the accuracy and robustness of the model in cancer gene identification tasks.

[0093] In one example, before generating the node representation dictionary, the multiplexed GNN first computes an embedding vector for each node in each relation subgraph. For example, the embedding of BRCA1 in the PPI subgraph is [0.1, 0.6, -0.2], and the embedding of BRCA1 in the miRNA-mRNA subgraph is [-0.4, 0.7, 0.2]. The fused embedding is generated by the attention fusion module, which combines these two vectors to generate a new vector, such as [0.2, -0.5, 1.0]. This vector contains the key features of BRCA1 in both relations but removes redundant information.

[0094] Step 340: For each node in the dictionary of the node representation of the previous attention fusion module, embed the corresponding comprehensive data into the binary classifier to obtain the probability value. When the probability value is greater than the preset threshold, it is predicted as a cancer gene.

[0095] Specifically, the fused integrated embedding representation is input into a binary classifier, which outputs the probability value of each node being a cancer gene; when the probability value exceeds a preset threshold, the gene corresponding to that node is determined to be a cancer gene.

[0096] The classifier module takes as input the node embedding representation from the previous attention fusion module, which has already incorporated features from multiple semantic channels and undergone residual and layer normalization processing. The output is the final binary classification result, indicating whether node v is a cancer gene.

[0097] Specifically, the fused node representation The input is fed into the classifier module to perform the target task, such as identifying cancer genes. This application employs a binary logistic regression layer, which can be configured with cross-entropy loss or weighted loss depending on the specific task.

[0098]

[0099] Where, p v W is the predicted probability distribution of node v for each category, such as cancer genes and non-cancer genes. c ∈R C×d C is the number of categories, b is a learnable bias term, and the final prediction is:

[0100]

[0101] The argmax function selects the class with the highest probability as the prediction result. A value of 1 indicates that node v is predicted to be a cancer gene, or A value of 0 indicates that node v is predicted to be a non-cancer gene.

[0102] The loss function used is the binary cross-entropy loss function:

[0103]

[0104] Among them, y v，c This is the encoding of the actual label (if node v belongs to category c, this item is 1; otherwise, it is 0), p v，c It is the probability that node v is predicted to be of class c. By repeating the process multiple times to reduce the loss function, the classification of cancer genes is achieved.

[0105] In one example, assuming we are studying genes related to breast cancer, we constructed a heterogeneous graph containing the following relationships:

[0106] Node types: genes, such as BRCA1, TP53; miRNAs, such as miR-21; diseases, such as breast cancer; Relationship types: gene-gene (PPI interaction), gene-miRNA (regulatory relationship), gene-disease (association relationship).

[0107] By using multiplexed GNNs, embedding vectors that fuse multiple semantic relationships are generated for each node. For example, the embedding vector of BRCA1 has a length of 128 dimensions.

[0108] The comprehensive embedding vector of the target gene node is extracted from the node representation dictionary output by the previous attention fusion module. This vector is then input into a binary classifier. The classifier structure uses a logistic regression layer, a fully connected layer, and a sigmoid activation. Assuming the model is already trained, the current parameters are: w = [0.8, -1.2, 0.5, ..., 0.3], length 128, b = -0.1. For the BRCA1 node, the calculated probability of BRCA1 being predicted as a cancer gene is 90%. The preset threshold can be set to 0.5; 0.90 ≥ 0.5, thus confirming its prediction as a cancer gene.

[0109] By applying the cancer gene identification method based on multiplexed heterogeneous graph neural networks provided in this invention, a multi-path heterogeneous graph modeling mechanism based on relation type division is proposed for cancer biological networks with multiple node types, such as mRNA and lncRNA. Each relation type is constructed as an independent "subgraph channel," preserving the independent semantics of various molecular interactions and effectively avoiding relation confounding. An attention fusion mechanism is designed to fuse node representations under various semantic subgraphs in multiple channels. This mechanism dynamically models the importance of different semantic paths to node representations. It learns a unique query, key, and value representation for each relation type, calculates its semantic contribution score, and uses this score as weight to perform weighted fusion of multi-path embeddings, thereby achieving context-aware modeling of nodes across semantic perspectives. Compared to traditional averaging or static weighting strategies, this mechanism has stronger adaptive modeling capabilities and can accurately extract the most discriminative semantic features. Meanwhile, the attention fusion module also introduces residual connections and layer normalization mechanisms. By using the residual paths formed by two layers of nonlinear MLPs, it enhances information flow capabilities, effectively mitigating gradient vanishing and feature degradation issues when multiple layers are stacked, and improving the model's nonlinear modeling ability and generalization robustness. Layer normalization further stabilizes the model training process, suppresses feature drift, and strengthens the collaborative expression of multi-channel semantic information. The EmhGNN model proposed in this study has made some progress in the field of heterogeneous graph cancer gene recognition, providing a new tool for accurate prediction of cancer genes and a deeper understanding of cancer molecular mechanisms. By constructing and analyzing heterogeneous graphs containing various biomolecules and their interactions, EmhGNN can effectively integrate multi-source heterogeneous data, uncover complex relationships between biomolecules, and thus improve the accuracy and efficiency of cancer gene recognition. The EmhGNN model employs a multiplexing structure and attention mechanism, which can not only model different types of nodes and relationships separately, but also capture complex interactions between biomolecules through feature fusion.

[0110] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0111] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0112] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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.

Claims

1. A method for cancer gene identification based on multiplexed heterogeneous graph neural networks, characterized in that, The method includes: The initial features of each node in the cancer heterogeneous graph network are updated, and the updated features are output. Based on the updated features of the nodes, a corresponding node embedding dictionary is generated; the node embedding dictionary includes the learning embedding representation of each node in the cancer heterogeneous graph network under different relation subgraphs, each node corresponds to multiple embedding vectors, and each embedding vector corresponds to a different semantic relation; The embedding vector is used as input to the attention fusion module, and the output is a node representation dictionary after attention fusion. The node representation dictionary is a key-value pair geometry, where the key is the node identifier and the value is the comprehensive embedding representation of the node after attention fusion. Each node in the node representation dictionary corresponds to a fused embedding vector. The embedding vector represents the embedding representation of the node under different relational subgraphs. Each node in the dictionary is represented by a node in the attention fusion module of the previous layer. The corresponding comprehensive embedding is input into the binary classifier to obtain a probability value. When the probability value is greater than a preset threshold, it is predicted to be a cancer gene. The method further includes the following steps prior to the above: Extract node expression data and clinical data for various cancers from the TCGA dataset; Based on clinical information, the samples were divided into experimental and control groups, and the relative changes in gene expression between the two groups were calculated. Determine the node and edge types of the cancer heterogeneous graph network; Using relative expression changes as the initial feature of genes, the network topology and clinical features are captured by propagating node features through heterogeneous graph neural networks. For each gene, the fusion unit is embedded in different relational subgraphs to generate a comprehensive representation; The training set and the test set are divided according to cancer type or sample ratio, and a cancer heterogeneous graph network is obtained after training. The process of updating the initial features of each node in the cancer heterogeneous graph network and outputting the updated features specifically includes: For each type of relation Corresponding subgraph For all nodes initial features ,pass Update; among them, This represents the neighbors of node v under relation r; These are learnable or predefined hyperparameters; It is a relation-specific multilayer perceptron.

2. The method of claim 1, wherein, Each learning module consists of two fully connected layers, which are connected via... To prevent overfitting; among which, It is a multilayer perceptron with specific relationships; It is the input feature vector of the node; It is the weight matrix of the first fully connected layer, used for the r-th relation, mapping the input features from the input dimension to the hidden layer dimension; It is the weight matrix of the second fully connected layer, which maps the hidden layer features to the output dimension; It is a non-linear activation function used to introduce non-linearity between fully connected layers, enabling the model to learn more complex feature relationships; the Dropout layer is used to prevent overfitting.

3. The method according to claim 1, characterized in that, The embedding vector is used as input to the attention fusion module, and the output is a dictionary of node representations after attention fusion. For node v in different relations The following is represented as Multi-head attention is used for aggregation, and for each head, three sets of projections are constructed: ;in, These are learnable parameters. For the attention dimension; pass Calculate the attention score of a node across different semantics; where, The original attention score for node v under relation r is represented. It is the normalized attention weight of node v with respect to relation r, which represents the importance of this semantic in the final representation of the node; The outputs from each header are concatenated and linearly transformed into a unified node representation, retaining key terms from various semantic categories. The final fused representation is as follows: ; pass Residual linking and layer normalization; among which, The nonlinear residual path function is represented by two fully connected layers and the GELU activation function. This represents the representation of node v after multi-head attention fusion; This represents the original input representation of node v.

4. The method according to claim 1, characterized in that, The step of taking each node in the dictionary of the node representation of the previous attention fusion module, embedding the corresponding comprehensive data into the input binary classifier, and obtaining a probability value, and predicting it as a cancer gene when the probability value is greater than a preset threshold, specifically includes: pass Calculate the predicted probability distribution of node v for each category; where the fused node is represented as... , C is the number of categories, and b is a learnable bias term; The final prediction is ; wherein, is 1 node v is predicted to be a cancer gene, is 0 indicates node v is predicted to be a non-cancer gene; By determining a loss function; wherein, is an encoding of the true label, this term is 1 if the node v belongs to class c, and 0 otherwise, is the probability that the node v is predicted to be of class c.

5. The method of claim 1, wherein, The key of the node embedding dictionary is the node identifier, and the value is the set of embedding vectors corresponding to the node under all relational subgraphs. 6.A cancer gene recognition device based on a multiplexed heterogeneous graph neural network, characterized by, The device includes: Multiplexed GNNs are used to update the initial features of each node in the cancer heterogeneous graph network and output the updated features. Based on the updated features of the nodes, a corresponding node embedding dictionary is generated; the node embedding dictionary includes the learning embedding representation of each node in the cancer heterogeneous graph network under different relation subgraphs, each node corresponds to multiple embedding vectors, and each embedding vector corresponds to a different semantic relation; The attention fusion module takes the embedding vector as input and outputs a node representation dictionary after attention fusion. The node representation dictionary is a key-value pair geometry, where the key is the node identifier and the value is the comprehensive embedding representation of the node after attention fusion. Each node in the node representation dictionary corresponds to a fused embedding vector. The embedding vector represents the embedding representation of the node under different relational subgraphs. The classifier module is used to take each node in the dictionary of the node representation of the previous attention fusion module and embed the corresponding comprehensive input into the binary classifier to obtain a probability value. When the probability value is greater than a preset threshold, it is predicted to be a cancer gene. The cancer gene recognition device is also used for: Extract node expression data and clinical data for various cancers from the TCGA dataset; Based on clinical information, the samples were divided into experimental and control groups, and the relative changes in gene expression between the two groups were calculated. Determine the node and edge types of the cancer heterogeneous graph network; Using relative expression changes as the initial feature of genes, the network topology and clinical features are captured by propagating node features through heterogeneous graph neural networks. For each gene, the fusion unit is embedded in different relational subgraphs to generate a comprehensive representation; The training set and the test set are divided according to cancer type or sample ratio, and a cancer heterogeneous graph network is obtained after training. The process of updating the initial features of each node in the cancer heterogeneous graph network and outputting the updated features specifically includes: For each type of relation Corresponding subgraph For all nodes initial features ,pass Update; among them, This represents the neighbors of node v under relation r; These are learnable or predefined hyperparameters; It is a relation-specific multilayer perceptron.

7. A computer server, characterized by include: Memory, processor, and transceiver; The processor is used to couple with the memory, read and execute instructions in the memory to implement the cancer gene identification method based on multiplexed heterogeneous graph neural network as described in any one of claims 1-5; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is executed by a processor according to any one of claims 1-5: the cancer gene identification method based on multiplexed heterogeneous graph neural networks.

Images