Prediction method of virulence gene based on topology and biological feature fusion

Abstract

The invention provides a topology and biological feature fusion-based virulence gene prediction method, which comprises the following steps of: obtaining a to-be-detected gene; inputting the to-be-detected gene into a trained DAVGAE model, and predicting the correlation degree of the to-be-detected gene and the disease to obtain a disease gene correlation prediction conclusion; wherein the DAVGAE model comprises a data enhancement module, an encoder and an inner product decoder. The problems of data sparsity and heterogeneous data integration in gene-disease association prediction can be effectively solved at least through a DAVGAE model formed by a data enhancement module, an encoder and an inner product decoder.

Description

Technical Field

[0001] The present application relates to the field of biological gene technology, and in particular to a method for predicting pathogenic genes based on the fusion of topology and biological features. Background Art

[0002] Predicting disease-gene associations is crucial for gaining a deeper understanding of disease pathogenesis and guiding drug development, but it is also a challenging task. Constructing effective predictive models requires learning the complex relationships between genes and diseases and effectively characterizing them. Traditional computational approaches primarily rely on similarities in sequence, function, and interactions between candidate genes and known disease-causing genes. However, these approaches are often limited by the amount of known disease-gene association data and fail to fully exploit the inherent connections between biological data. Furthermore, while traditional statistical analysis-based methods, such as linkage analysis and genome-wide association studies, are fundamental to disease gene prediction and can provide predictive biomarkers by studying human genetic variation (such as single nucleotide polymorphisms), their statistical analysis and biological validation processes are time-consuming and costly, primarily due to the need for further analysis of a large number of false-positive results. Furthermore, these techniques rely solely on direct genotype-phenotype associations. However, biomolecules often function in a collaborative manner within specific biological pathways. Projecting their specific roles and collaborative relationships onto network or graph structures can reveal more useful knowledge and provide a more systematic perspective. When using molecular networks, disease-causing factors, such as gene mutations, epigenetic factors, and pathogens, can be more efficiently tracked by tracking perturbations (i.e., removal of edges or nodes) in molecular networks. Molecular networks are an efficient and effective data representation that can model complex interactions between molecules from a broader perspective and track potential disruptions in biological pathways caused by pathogenic factors. Therefore, computational methods have widely used molecular networks to supplement and enrich existing linkage analyses and genome-wide association studies.

[0003] Among existing methods for predicting disease-causing genes, random walk methods, while widely used in biological network analysis, are highly dependent on fine-tuning parameters and the quality of the underlying network structure. Small parameter changes or deviations in network topology can lead to unstable results and reduce prediction reliability. When the network structure is sparse or noisy, random walks are easily trapped in local regions, unable to effectively propagate information and obtain global insights, thus affecting prediction accuracy. This is directly related to the sparse and incomplete nature of disease-causing gene data, as sparse networks restrict the exploration range of random walks and weaken the propagation of useful signals. Traditional machine learning methods, such as those based on feature engineering, can improve prediction performance by integrating features from multiple data sources, but their effectiveness depends largely on the quality of manually designed features. Feature design and selection require extensive domain knowledge and manual intervention, which is not only time-consuming and labor-intensive, but also makes it difficult to ensure that the selected features fully capture the complex relationships underlying the data. When disease-causing gene data is sparse, the number of samples available for model training is limited, which can lead to model overfitting or insufficient generalization, further affecting prediction accuracy. Graph representation learning methods, including graph embedding and graph neural networks, aim to automatically learn node or graph representations from network structures, avoiding tedious manual feature engineering. However, the effectiveness of these methods heavily relies on the quality and completeness of the input network structure. When the network structure is sparse or there are missing connections, the model has difficulty learning effective representations, resulting in reduced prediction performance. For example, graph neural networks rely on information transfer between nodes to learn node representations, but sparse networks limit the propagation of information, resulting in the model being unable to effectively capture the relationship between nodes. This is closely related to the problem of sparse and incomplete pathogenic gene data, because sparse network structures hinder graph representation learning methods from extracting useful information, making the learned feature representations unreliable, which in turn affects the accuracy of predictions. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a prediction method for pathogenic genes based on the fusion of topology and biological features, which can at least effectively solve the problems of data sparsity and heterogeneous data integration in gene-disease association prediction by integrating the DAVGAE model formed by the data enhancement module, encoder and inner product decoder.

[0005] An embodiment of the present application provides a method for predicting pathogenic genes based on the fusion of topology and biological features, the method comprising: obtaining a gene to be detected; inputting the gene to be detected into a trained DAVGAE model, predicting the degree of association between the gene to be detected and the disease, and obtaining a disease gene association prediction conclusion; wherein the DAVGAE model comprises: a data enhancement module, an encoder, and an inner product decoder.

[0006] Optionally, the data augmentation module performs the following steps: establishing a gene-disease network based on a set of actually connected gene-disease node pairs, gene nodes, and disease nodes, and determining an adjacency matrix of the gene-disease network; obtaining topological features of each gene node and disease node using a Node2vec model on the gene-disease network; calculating the cosine similarity of all edges formed between each gene node and a disease node in the gene nodes and disease nodes; and adding a gene-disease pair to a positive training set if the cosine similarity of the gene-disease pair is greater than a preset threshold.

[0007] A set of new pairs with the same number is randomly selected from the positions where the elements in the adjacency matrix are 0, and the set is added to the negative training set.

[0008] Optionally, the Node2Vec model is used to feature capture relevant network information to compare different types of nodes in a shared topological space, where c(v p , v q ),

[0009]

[0010] Among them, Vp is the gene node, is the topological feature representation of Vp, Vq is the disease node, is the topological characteristic representation of Vq, c(v p , v q ) is the association probability between gene node Vp and disease node Vq, is the gene node vector with disease node vector The cosine similarity of .

[0011] Optionally, the association density between genes and diseases in the gene-disease network is calculated by the following formula:

[0012]

[0013] in, It's G gd The number of gene nodes in It's G gd The number of disease nodes in G gd is the gene-disease network, and AD represents the association density.

[0014] Optionally, the encoder includes a first graph convolution layer, a first GELU activation function, a second graph convolution layer, a second GELU activation function, a first graph convolutional neural network layer, and a second graph convolutional neural network layer;

[0015] Among them, a dual-headed graph attention mechanism is applied between the first and second graph convolutional layers, so that the model pays attention to the most relevant neighbors when aggregating features;

[0016] After applying the attention mechanism, the mean and log-variance outputs are obtained using the first and second graph convolutional neural network layers.

[0017] Optionally, the input data of the encoder includes first input data and second input data, the first input data is a gene-disease association network G' obtained by data augmentation, and the second input data is a node feature matrix X∈RN×f of nodes in G', where N is the number of nodes in G', f is the feature dimension, and R represents a set of real numbers.

[0018] In the encoder model, the node feature matrix passes through the GCN layer to obtain the hidden representation; GELU activation is applied to the output to obtain H1, and the dropout layer is applied for regularization; GELU is applied after the GCN layer to provide smooth gradient transition and capture complex nonlinear relationships.

[0019] Optionally, the inner product decoder of DAVGAE calculates the inner product between latent variables to produce a new adjacency matrix by the following formula:

[0020] A new =Sigmoid(ZZ Τ )

[0021] Among them, A new ∈R N×N , A new The value in is the probability of association between the disease and the gene pair, R N×N represents an N×N matrix, and Z represents the potential embedding matrix of the nodes in the network G'.

[0022] Optionally, train the DAVGAE model by following these steps:

[0023] The loss function of DAVGAE during training consists of three parts: reconstruction loss unit, KL divergence loss unit and L2 regularization unit;

[0024] The reconstruction loss unit quantifies the error between the predicted adjacency matrix and the original adjacency matrix calculated for the positive training edges, so that the DAVGAE model reconstructs the adjacency matrix. The reconstruction loss unit calculates the reconstruction error of each node pair and takes the expected value to determine the loss function of the DAVGAE model;

[0025] The KL divergence loss unit aligns the learned latent variable distribution q(Z|X,A) with the prior distribution p(Z), which is usually assumed to be a standard normal distribution to ensure that the latent space remains well-structured and prevent overfitting;

[0026] The L2 regularization unit adds a penalty for excessively large model parameters to prevent overfitting and improve generalization ability.

[0027] Optionally, the total loss function of the DAVGAE model is calculated by the following formula:

[0028] L=E q(Z|X,A) (logp(A|Z))+KL(q(Z|X,A)||p(Z))+λ||Θ|| 2 ;

[0029] Where p(A|Z) is the generative marginal probability distribution, which quantifies the likelihood of generating A using the latent variable Z, q(Z|X,A) is the encoder that outputs the conditional distribution of the latent variable, p(Z) is the prior distribution of the latent variable, λ is the regularization coefficient, and Θ contains all the model parameters;

[0030] The probability distribution of each element in the adjacency matrix is ​​generated by the following formula:

[0031]

[0032] where Apq indicates whether there is a connection between gene node p and disease node q, zvp and zvq are the embedding vectors of the nodes, σ(·) is the sigmoid activation function, and A is the adjacency matrix.

[0033] Optionally, the method further comprises: using a grid search strategy to explore the hyperparameter space and optimize the performance of DAVGAE, adjusting key parameters, including the similarity threshold of gene-disease pairing, the dropout rate, the number of GAT heads, and the regularization coefficient in the loss function.

[0034] The pathogenic gene prediction method based on the fusion of topology and biological features provided in the embodiment of the present application can at least effectively solve the problems of data sparsity and heterogeneous data integration in gene-disease association prediction by using the DAVGAE model formed by the data enhancement module, encoder and inner product decoder.

[0035] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 A schematic diagram of a process for predicting pathogenic genes based on the fusion of topology and biological features provided in an embodiment of the present application;

[0038] Figure 2 This is a schematic diagram of the framework of the DAVGAE model provided in the embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for which protection is claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work falls within the scope of protection of the present application.

[0040] First, the application scenarios to which this application is applicable are introduced. This application can be applied in the field of biological gene technology.

[0041] Research has found that predicting disease-gene associations is crucial for gaining a deeper understanding of disease pathogenesis and guiding drug development, yet it is also a challenging task. Constructing effective predictive models requires learning the complex relationships between genes and diseases and effectively characterizing them. Traditional computational methods primarily rely on similarities in sequence, function, and interactions between candidate genes and known disease-causing genes. However, these methods are often limited by the amount of known disease-gene association data and fail to fully exploit the inherent connections between biological data. Furthermore, while traditional statistical analysis-based methods, such as linkage analysis and genome-wide association studies, are fundamental to disease gene prediction and can provide predictive biomarkers by studying human genetic variation (such as single nucleotide polymorphisms), their statistical analysis and biological validation processes are time-consuming and costly, primarily due to the need for further analysis of a large number of false-positive results. Furthermore, these techniques rely solely on direct genotype-phenotype associations. However, biomolecules often function in a collaborative manner within specific biological pathways. Projecting their specific roles and collaborative relationships onto network or graph structures can reveal more useful knowledge and provide a more systematic perspective. When using molecular networks, disease-causing factors, such as gene mutations, epigenetic factors, and pathogens, can be more efficiently tracked by tracking perturbations (i.e., removal of edges or nodes) in molecular networks. Molecular networks are an efficient and effective data representation that can model complex interactions between molecules from a broader perspective and track potential disruptions in biological pathways caused by pathogenic factors. Therefore, computational methods have widely used molecular networks to supplement and enrich existing linkage analyses and genome-wide association studies.

[0042] Among existing methods for predicting disease-causing genes, random walk methods, while widely used in biological network analysis, are highly dependent on fine-tuning parameters and the quality of the underlying network structure. Small parameter changes or deviations in network topology can lead to unstable results and reduce prediction reliability. When the network structure is sparse or noisy, random walks are easily trapped in local regions, unable to effectively propagate information and obtain global insights, thus affecting prediction accuracy. This is directly related to the sparse and incomplete nature of disease-causing gene data, as sparse networks restrict the exploration range of random walks and weaken the propagation of useful signals. Traditional machine learning methods, such as those based on feature engineering, can improve prediction performance by integrating features from multiple data sources, but their effectiveness depends largely on the quality of manually designed features. Feature design and selection require extensive domain knowledge and manual intervention, which is not only time-consuming and labor-intensive, but also makes it difficult to ensure that the selected features fully capture the complex relationships underlying the data. When disease-causing gene data is sparse, the number of samples available for model training is limited, which can lead to model overfitting or insufficient generalization, further affecting prediction accuracy. Graph representation learning methods, including graph embedding and graph neural networks, aim to automatically learn node or graph representations from network structures, avoiding tedious manual feature engineering. However, the effectiveness of these methods heavily relies on the quality and completeness of the input network structure. When the network structure is sparse or there are missing connections, the model has difficulty learning effective representations, resulting in reduced prediction performance. For example, graph neural networks rely on information transfer between nodes to learn node representations, but sparse networks limit the propagation of information, resulting in the model being unable to effectively capture the relationship between nodes. This is closely related to the problem of sparse and incomplete pathogenic gene data, because sparse network structures hinder graph representation learning methods from extracting useful information, making the learned feature representations unreliable, which in turn affects the accuracy of predictions.

[0043] It is important to note that several molecular networks are currently available to describe gene relationships, such as protein-protein interaction networks, gene regulatory networks, gene co-expression networks, and metabolic interaction networks. Among these networks, protein-protein interaction networks are one of the most widely used for predicting disease-gene associations. A key reason for this is that proteins perform numerous critical functions to sustain biological activity, including biochemical reactions, metabolic catalysis, DNA replication, intercellular signaling, and maintaining cellular and tissue structure. More specifically, proteins perform biological functions through collaboration and interaction, and these interactions can be integrated and modeled as network data structures. Given a protein interaction, a mutation in the corresponding gene of one of the two proteins may render the interaction inoperable, leading to loss of certain important biological functions and disease. Another reason is that many studies using protein-protein interaction networks embody the assumption that protein positions are not random. Proteins associated with a common set of biological properties tend to share common topological properties in the network, such as node degree and centrality. Therefore, protein-protein interaction networks can be used to reveal protein-disease associations by deriving useful network-based features from proteins.

[0044] Researchers have proposed many computational methods to predict disease-gene associations, including traditional random walk methods, machine learning methods, and graph representation learning methods.

[0045] Random walk methods: These methods propagate information on homogeneous or heterogeneous biological networks (such as protein-protein interaction networks and phenotype-gene networks) to predict disease-gene associations. For example, DriverRWH[1] ranks disease-causing genes by performing random walks on a gene mutation network that includes somatic mutations and interactions. PSNE[2] uses a random walk method to identify potential disease genes by propagating information on a network that combines disease-gene connections, protein interactions, and disease similarities. MultiXrank[3] predicts gene-disease associations by applying random walks with restarts on a multi-layer network and using proximity scores to train binary classifiers for association prediction. However, the random walk method has a strong dependence on parameter adjustment and network structure, resulting in its lack of stability, which has prompted researchers to explore machine learning methods as an alternative.

[0046] Machine learning methods: Feature learning plays a key role in many machine learning tasks, and can extract meaningful representations from complex data to enhance model performance. Recent studies have emphasized its importance in biological data analysis, including disease-gene association prediction. For example, BRIDGE[4] integrates features from multiple sources of data. Metagraph+[5] uses protein-protein interaction networks and biological annotations to construct features. The dgMDL[6] method first trains two multimodal deep belief networks. One of the multimodal deep belief networks uses a protein-protein interaction network and the other uses a gene ontology similarity network. Then, dgMDL uses the outputs from the two initial DBNs to train the final joint DBN for prediction. However, such methods require manual feature design and feature selection, and rely on a lot of work and domain knowledge. These challenges have shifted attention to graph representation learning that can automatically extract features.

[0047] Graph representation learning methods: Neural networks have been widely used in biomedical research, demonstrating their potential in processing complex data structures. However, traditional neural networks designed for independent and identically distributed data have difficulty in effectively capturing such graph structure information. Graph representation learning methods, including graph embedding, matrix factorization, and graph neural networks, can directly learn embedded representations of nodes from complex biological networks. Predicting disease-gene associations requires modeling the complex relational structures inherent in these networks. Graph neural networks can effectively integrate node features and edge relationships, providing a natural way to capture local and contextual information to predict disease-gene associations. For example, HerGePred[7] developed a heterogeneous network embedding method to represent disease genes for prediction. It integrates disease similarity, symptom similarity, and protein-protein interaction networks into a heterogeneous graph for embedding. MGREL is an ensemble learning-based method that integrates multi-graph representation learning for predicting gene-disease associations. It combines genetic, treatment, and network topology information through two feature generation channels. However, graph embedding methods require high-quality network structures to learn effective representations. IMC uses inductive matrix completion to predict gene-disease associations. It formulates the prediction problem as a matrix completion problem and inductively expands the matrix using unlabeled gene-disease pairs. GeneHound uses Bayesian matrix factorization with genomic and phenotypic side information to prioritize genes. It integrates multiple genomic data sources with phenotypic similarity into a joint matrix factorization framework. However, matrix factorization methods require many known gene-disease associations to train latent factors. These methods may not achieve good results if the input gene-disease matrix is ​​very sparse. HNMME applies graph embedding representations and graph convolutional networks to heterogeneous gene-disease graphs to predict gene-disease associations. PGCN embeds diseases and genes using graph convolutional networks for disease gene prediction. It constructs a disease-gene association graph and applies graph convolution to learn the latent representation. LUPI [Likely referring to a LUPI-like network] uses privileged information and heteroscedastic dropout in graph convolutional networks to predict disease genes. It incorporates gene expression data as privileged information to improve prediction performance. Graph convolutional networks can be used to learn to extract topological features of nodes in a network and have been shown to be powerful in predicting disease-gene associations. DGP-PGTN utilizes a parallel graph attention network to integrate heterogeneous information from diseases, genes, phenotypes, and ontologies. It can automatically capture multiple potential interactions between diseases and genes. However, graph convolutional networks rely on rich graph structure and node features to propagate valuable signals. Noisy or incomplete network structures can mislead the model. Therefore, sufficient and high-quality training data is crucial to fully realize its potential. Furthermore, advanced methods in multi-label learning and network clustering have shown great potential for extracting latent relationships from complex biological data, which can provide valuable insights for enhancing disease-gene prediction models.

[0048] Based on this, the embodiment of the present application provides a prediction method for pathogenic genes based on the fusion of topology and biological features, which can at least effectively solve the problems of data sparsity and heterogeneous data integration in gene-disease association prediction by integrating the DAVGAE model formed by the data enhancement module, encoder and inner product decoder.

[0049] It should be noted here that the DAVGAE method proposed in this application integrates the biological and topological features of genes and diseases, and uses variational graph autoencoders for modeling. Through a four-layer graph neural network encoder, DAVGAE can effectively learn robust and discriminative feature representations of genes and diseases in association networks. This modeling method that integrates multiple information enables the model to understand the complex relationship between genes and diseases more comprehensively, overcoming the limitations of a single data source. DAVGAE uses an inner product decoder to predict the association scores of all gene-disease pairs. Compared with other complex decoder structures, the inner product decoder is concise and efficient, and can directly calculate the association strength between them based on the learned gene and disease representations. This design simplifies the complexity of the model while ensuring the effectiveness of the prediction.

[0050] To address the problem of incomplete and sparse gene-disease association data, the DAVGAE method proposes a novel data augmentation strategy. Rather than randomly generating new associations, this strategy calculates cosine similarities between genes and diseases based on their biological and topological characteristics. Gene-disease pairs with high similarity scores are then selected to augment the association data, effectively expanding the training set and more specifically supplementing potential true associations, mitigating the negative impact of data sparsity on model performance. This is one of the core innovations of this patent, as it avoids the noise that can be introduced by blind data augmentation and significantly improves data quality.

[0051] Experimental results demonstrate that this data augmentation strategy significantly improves DAVGAE's performance on sparse datasets, validating its effectiveness. Comparisons with other pathogenic gene prediction models on three gene-disease association datasets show that DAVGAE outperforms other models in terms of prediction accuracy, area under the receiver operating characteristic (ROC) curve (AUC), and area under the precision-recall curve (AUPR), demonstrating the effectiveness of integrating multiple features. These experimental results strongly demonstrate the effectiveness and superiority of the DAVGAE approach.

[0052] The DAVGAE model effectively solves the problems of data sparsity and heterogeneous data integration in gene-disease association prediction by combining a similarity-based data augmentation strategy, a variational graph autoencoder model that integrates multiple features, and an efficient inner product decoder. It outperforms other comparative models in comprehensive experiments on three gene-disease association datasets.

[0053] See also Figure 1 , Figure 1 This is a flow chart of a method for predicting pathogenic genes based on the fusion of topology and biological features provided in the embodiment of the present application. Figure 1 As shown in , the prediction method of pathogenic genes based on the fusion of topology and biological features provided in the embodiments of the present application includes:

[0054] S101. Obtain the gene to be detected.

[0055] S102: Input the gene to be detected into the trained DAVGAE model, predict the correlation between the gene to be detected and the disease, and obtain a disease gene correlation prediction conclusion.

[0056] The DAVGAE model includes: a data enhancement module, an encoder and an inner product decoder

[0057] Specifically, the data augmentation module performs the following steps: establishing a gene-disease network based on a set of actually connected gene-disease node pairs, gene nodes, and disease nodes, and determining an adjacency matrix of the gene-disease network; obtaining topological features of each gene node and disease node using a Node2vec model on the gene-disease network; and calculating the cosine similarity of all edges formed between each gene node and a disease node in the gene nodes and disease nodes.

[0058] If the cosine similarity of a gene-disease pair is greater than a preset threshold, it is added to the positive training set;

[0059] A set of new pairs with the same number is randomly selected from the positions where the elements in the adjacency matrix are 0, and the set is added to the negative training set.

[0060] The Node2vec model is used to capture relevant network information, thereby comparing different types of nodes in a shared topological space. The following formula is used to calculate c(v p , v q ),

[0061]

[0062] Among them, Vp is the gene node, is the topological feature representation of Vp, Vq is the disease node, is the topological characteristic representation of Vq, c(v p , v q ) is the association probability between gene node Vp and disease node Vq, is the gene node vector with disease node vector The cosine similarity of .

[0063] Specifically, the association density between genes and diseases in the gene-disease network is calculated using the following formula:

[0064]

[0065] in, It's G gd The number of gene nodes in It's G gd The number of disease nodes in G gd is the gene-disease network, and AD represents the association density.

[0066] If there are isolated nodes in a disease-gene network, it is difficult to determine their associations with other nodes. To address this issue, we collected features of disease and gene nodes. We calculated the cosine similarity of all disease-gene pairs, which all contained isolated nodes, as their probabilistic similarity. We augmented associations with high cosine similarities into the positive training set and randomly sampled the same number of negative associations.

[0067] Disease-gene association prediction is a highly imbalanced problem, with negative samples far outnumbering positive samples. To address this, we employed a 1:1 sampling strategy, pairing each positive sample with a randomly selected negative sample. This approach prevents the model from being biased toward abundant negative samples while ensuring that irrelevant pairs (not necessarily true negative samples) do not introduce excessive noise. Balanced sampling improves the model's ability to accurately identify meaningful associations. Specifically, let Ggd = (Vgd, Egd) be the gene-disease network, and its adjacency matrix be Agd, where Vgd contains both gene and disease nodes, and Egd contains only gene-disease associations. The adjacency matrix Agd represents the connectivity between gene and disease nodes in Ggd, with entries representing connections (e.g., 1 for connected and 0 for unconnected). Egd is the set of gene-disease node pairs in Ggd that are actually connected. If vi is a gene node and the degree of vi∈Vgd is 0, the cosine similarity between vi and all disease nodes is calculated, followed by the cosine similarity between vi and all disease nodes. If vi is a disease node, its cosine similarity with all gene nodes is calculated. For each node with degree 0, we select the disease-gene pair with the highest similarity that contains the node and add the pair to the positive training sample. Then, we randomly select a set from the position 0 in Agd. The number of samples in this set matches the number of new positive training samples, and this set is used as a new negative training set. In addition, zvi represents the topological feature of vi, which is obtained by applying Node2Vec to Ggd. The cosine similarity between gene nodes and disease nodes is based on their topological features. and These features encode the structural relationships within Ggd. Node2Vec ensures that the features capture relevant network information, enabling the comparison of different types of nodes (e.g., genes and diseases) in a shared topological space.

[0068] Known disease-gene associations are often limited, especially when the association density (AD), which measures the connectivity between genes and diseases in a network, is very low (e.g., AD ≤ 0.006).

[0069] To address this issue, we use data augmentation to enhance associations. First, we use the Node2vec method on Ggd to obtain the representation zvi of each node vi in ​​Vgd. Then, we calculate the cosine similarity of all edges formed by any gene node and a disease node in Vgd. If the similarity of a gene-disease pair exceeds a threshold r, we add it to the positive training set. Next, we randomly select a set of new pairs with the same number of pairs from the positions where the elements in Agd are zero. Finally, we add this set to the negative training set.

[0070] DAVGAE addresses the challenge of limited disease-gene associations by using data augmentation to expand the training set and enhance the network topology, thereby achieving more accurate disease-gene association prediction. DAVGAE then uses a variational graph autoencoder (VGAE)

[18] with an attention mechanism and an inner product decoder to predict disease-gene associations. DAVGAE can better learn the embedding representations of nodes in the new association network obtained through data augmentation. It then uses the obtained node embeddings as input to the disease-gene association prediction decoding model.

[0071] The encoder includes a first graph convolution layer, a first GELU activation function, a second graph convolution layer, a second GELU activation function, a first graph convolutional neural network layer, and a second graph convolutional neural network layer.

[0072] Among them, a dual-headed graph attention mechanism is applied between the first graph convolutional layer and the second graph convolutional layer, so that the model focuses on the most relevant neighbors when aggregating features; after applying the attention mechanism, the first graph convolutional neural network layer and the second graph convolutional neural network layer are used to obtain the mean and logarithmic variance outputs.

[0073] Specifically, the input data of the encoder includes first input data and second input data, the first input data is the gene-disease association network G' obtained by data augmentation, and the second input data is the node feature matrix X∈RN×f of the nodes in G', where N is the number of nodes in G', f is the feature dimension, and R represents a set of real numbers.

[0074] In the encoder model, the node feature matrix passes through the GCN layer to obtain the hidden representation; GELU activation is applied to the output to obtain H1, and the dropout layer is applied for regularization; GELU is applied after the GCN layer to provide smooth gradient transition and capture complex nonlinear relationships.

[0075] DAVGAE's encoder primarily consists of two graph convolutional layers, each followed by a GELU activation function and a dropout layer (to prevent overfitting). Between these graph convolutional layers, we apply a dual-headed graph attention mechanism, enabling the model to focus on the most relevant neighbors when aggregating features. After applying the attention mechanism, we use two more graph convolutional layers to obtain mean and log-variance outputs.

[0076] Specifically, the encoder has two input data. The first is the gene-disease association network G' obtained through data augmentation. The second is the node feature matrix of the nodes in G' Where N is the number of nodes in G' and f is the feature dimension. In the encoder model, X first passes through the GCN layer to obtain the hidden representation. Then, GELU activation is applied to the output to obtain H1. A dropout layer is applied for regularization. After GELU is applied to the GCN layer, it provides smooth gradient transitions and captures complex nonlinear relationships, which is conducive to learning fine-grained representations in the early stages of the model. Then, a GAT layer with dual-headed self-attention is applied to the output to obtain an enhanced representation H2. GAT connects the outputs of the two attention heads. In addition, the attention mechanism helps the model focus on key nodes and learn related features. Next, ReLU activation is applied here to implement nonlinearity and output H2. After ReLU is applied to the GAT layer, it is computationally efficient and introduces sparsity, helping the model focus on the key features learned through the attention mechanism. Finally, two independent GCN layers are used to predict the mean μ and logarithmic standard deviation logσ from the obtained representation Z, which parameterize the multivariate Gaussian distribution of the latent variable Z. The above layer-by-layer propagation rule can be written as:

[0077]

[0078] Where A is the adjacency matrix of G', I is the identity matrix, and is the diagonal matrix of (A+I). is the normalized adjacency matrix with self-connections added, and W1 is the weight matrix.

[0079]

[0080] Where Watt is the attention coefficient matrix, is the original attention matrix with unactivated attention scores, W2 is the weight matrix, and LeakyReLU is the activation function.

[0081]

[0082] where μ, logσ∈R N×F , μ and logσ are the mean and logarithmic standard deviation of the latent embedding Z of each node in G′, respectively, and F is the dimension of Z.

[0083] The decoder of DAVGAE is the inner product between latent variables to produce a new adjacency matrix:

[0084]

[0085] Among them, the inner product decoder of DAVGAE calculates the inner product between latent variables through the following formula to generate a new adjacency matrix:

[0086] A new =Sigmoid(ZZ Τ )

[0087] Among them, A new ∈R N×N , A new The value in is the probability of association between the disease and the gene pair, R N×N represents an N×N matrix, and Z represents the potential embedding matrix of the nodes in the network G'.

[0088] Optionally, the DAVGAE model can be trained by the following steps: The loss function of DAVGAE during training consists of three parts: a reconstruction loss unit, a KL divergence loss unit, and an L2 regularization term unit; wherein the reconstruction loss unit quantifies the error between the predicted adjacency matrix and the original adjacency matrix calculated for the positive training edges, so that the DAVGAE model reconstructs the adjacency matrix. The reconstruction loss unit determines the loss function of the DAVGAE model by calculating the reconstruction error of each node pair and taking the expected value; the KL divergence loss unit aligns the learned latent variable distribution q(Z|X,A) with the prior distribution p(Z), which is usually assumed to be a standard normal distribution to ensure that the latent space maintains a good structure and prevent overfitting; the L2 regularization unit adds a penalty for excessively large model parameters to prevent overfitting and improve generalization ability.

[0089] Specifically, the total loss function of the DAVGAE model can be calculated by the following formula:

[0090] L=E q(Z|X,A) (logp(A|Z))+KL(q(Z|X,A)||p(Z))+λ||Θ|| 2 ;

[0091] where p(A|Z) is the generative marginal probability distribution, which quantifies the likelihood of generating A using the latent variable Z, q(Z|X,A) is the encoder that outputs the conditional distribution of the latent variable, p(Z) is the prior distribution of the latent variable, λ is the regularization coefficient, and Θ contains all the model parameters.

[0092] The probability distribution of each element in the adjacency matrix is ​​generated by the following formula:

[0093]

[0094] where Apq indicates whether there is a connection between gene node p and disease node q, zvp and zvq are the embedding vectors of the nodes, σ(·) is the sigmoid activation function, and A is the adjacency matrix.

[0095] Optionally, the method further comprises: using a grid search strategy to explore the hyperparameter space and optimize the performance of DAVGAE, adjusting key parameters, including the similarity threshold of gene-disease pairing, the dropout rate, the number of GAT heads, and the regularization coefficient in the loss function.

[0096] As an example, see Figure 2 , DAVGAE's loss function during training consists of three parts: reconstruction loss, KL divergence loss, and an optional L2 regularization term on the model parameters. The reconstruction loss quantifies the error between the predicted adjacency matrix Anew and the original adjacency matrix A calculated only for the positive training edges. It helps the model learn key patterns in the input data so that it can reconstruct the adjacency matrix A. This loss is calculated by calculating the reconstruction error for each node pair (vp, vq) and then taking the expected value. The KL divergence loss aligns the learned latent variable distribution q(Z|X,A) with the prior distribution p(Z), which is usually assumed to be a standard normal distribution. This regularization ensures that the latent space remains well-structured and prevents overfitting. The L2 regularization term adds a penalty for excessively large model parameters, which helps prevent overfitting and improve generalization ability. The total loss function is given by the combination of the following terms:

[0097] L=E q(Z|X,A) (logp(A|Z))+KL(q(Z|X,A)||p(Z))+λ||Θ|| 2 ;

[0098] where p(A|Z) is the generative marginal probability distribution, which quantifies the likelihood of generating A using the latent variable Z, q(Z|X,A) is the encoder that outputs the conditional distribution of the latent variable, p(Z) is the prior distribution of the latent variable, λ is the regularization coefficient, and Θ contains all the model parameters.

[0099] The generation process of the adjacency matrix is ​​given by the following formula:

[0100]

[0101] where Apq indicates whether there is a connection between the gene node p and the disease node q, zvp and zvq are the embedding vectors of the nodes, and σ(·) is the sigmoid activation function.

[0102] The model is trained using an optimizer to minimize the loss L. To stabilize the training process and achieve faster convergence, we adopt an adaptive learning rate decay strategy. Specifically, the learning rate at training round number t is updated according to the following formula:

[0103] lr t+1 =max(lr t *(1-t / T),lr min )

[0104] Where t / T is a hyperparameter controlling the decay rate, T represents the total number of training epochs, and lrmin is a lower bound to prevent the learning rate from decreasing to zero. Compared to a fixed learning rate scheme, this strategy ensures a gradual decrease in the learning rate, avoiding instability while achieving faster convergence. Furthermore, if the training data does not contain biometric features, the input features X are initialized to an all-one vector to ensure a uniform representation for all samples.

[0105] During training, we implemented several techniques to address potential overfitting. Specifically, we added dropout layers to prevent neurons from co-adapting, ensuring that the model does not become overly dependent on any single neuron. Additionally, we applied L2 regularization to penalize large weights in the model, encouraging simpler, more robust representations.

[0106] Here, we selected three disease-gene association datasets from different databases. We conducted ablation studies and hyperparameter tuning experiments to analyze the performance of our model. In addition, we evaluated DAVGAE and eight other algorithms for predicting disease-gene associations.

[0107] To mitigate the risk of overfitting due to the complexity of the model and the limited dataset size, we employed 10-fold cross-validation. This approach splits the dataset into ten subsets, nine for training and one for validation, repeated ten times, with each subset used as a validation set. This approach provides a more reliable estimate of the model's performance on different data subsets.

[0108] We evaluated our disease-gene association prediction algorithm on three datasets. Each dataset contains known disease-gene associations, which serve as the ground truth for model comparison. Table 1 below summarizes these datasets used in the experiments.

[0109] Table 1 Summary of three disease-genenetwork datasets:

[0110]

[0111] The graph structures of the datasets vary, including the number of disease-gene edges, gene nodes, and disease nodes, which directly impact the model's learning and generalization capabilities. For example, Dataset B's edges come from two independent sources (DisGeNet and MalaCards), ensuring robust association prediction by providing a more diverse set of associations. Dataset C, derived from the Online Mendelian Inheritance in Man (OMIM) database, highlights the challenges of sparse networks, as evidenced by the presence of 9,122 gene nodes and 636 disease nodes with degree 0. This sparsity necessitates the use of node features to strengthen the network and mitigate the impact of disconnected nodes on embedding quality. In contrast, the interconnected nature of nodes in the other two datasets reduces reliance on node-specific features, enabling the model to focus on edge-driven learning. In Dataset C, we leveraged the gene and disease features described in the paper. Gene features encompass both gene expression and functional associations. First, we used gene expression data from the BioGPS and Connectivity Map databases. To reduce redundancy, we used principal component analysis (PCA) to select the top 100 eigenvectors as features. Second, we used functional association features from eight different species. The species were worms, plants, fruit flies, E. coli, yeast, zebrafish, chicken, and mice. Each had 17,480 genetic signatures.

[0112] Disease features include disease ontology annotations and clinical features. The first set consists of ontology annotations obtained from the Human Disease Ontology (DO) database. The second set consists of clinical features extracted from the OMIM database. These two sets of features collectively contribute to our disease analysis. Each disease has 16,592 disease features. To facilitate efficient information transfer between different nodes, we apply PCA to process the features of disease and gene nodes into a unified dimension.

[0113] Here, model performance is evaluated by calculating the probability of known disease-gene associations and comparing them with the true associations. We use a set of evaluation metrics, including precision, recall, accuracy, F1 score, area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPR), and Matthews correlation coefficient (MCC). These metrics can provide a clearer understanding of its ability to predict new disease-gene associations. Values ​​closer to 1 indicate a better model.

[0114] Precision = TP / (TP+FP)

[0115] Recall = TP / (TP+FN)

[0116] Accuracy=(TN+TP) / (FN+FP+TN+TP)

[0117] F1-score=2*Recall*Precision / (Recall+Precision)

[0118]

[0119] Among them, TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

[0120] For gene-disease association models, AUC and AUPR are important evaluation metrics. AUC quantifies the model's ability to distinguish between true positive (TP) and false positive (FP) predictions at different classification thresholds. AUC values ​​closer to 1 indicate a strong separability between TP and FP predictions, which is crucial for identifying relevant gene-disease pairs. AUPR captures the balance between precision and recall, providing a more targeted measure of the model's effectiveness in predicting TP gene-disease associations over FP predictions.

[0121] Optionally, this application also designs ablation experiments, in which we modify only specific components to evaluate their impact on model performance, while keeping all other model settings consistent. We first compare the Graph Autoencoder (GAE) and VGAE. Both encoders use the same layers. Next, we experiment with various encoder layers, including linear, GCN, and GAT layers. We then evaluate different activation functions, such as ReLU and GELU. After comparing the framework and autoencoder architectures, we explore loss function combinations: reconstruction loss plus Kullback-Leibler (KL) divergence loss and reconstruction loss plus KL divergence loss with an additional L2 regularization term. We then analyze different optimizers, including Adam, AdamW, RAdam, and AMSGrad, all with a fixed learning rate of 0.001. An optimizer is an algorithm that adjusts model parameters during training to minimize the loss function. AdamW is characterized by separating weight decay from gradient updates, which improves regularization and enhances the model's generalization ability. Next, we explored the performance of fixed learning rates (0.01, 0.001, and 0.0001) and adaptive learning rates used in our model. Finally, we compared the impact of the data augmentation module on our model.

[0122] In ablation experiments, we evaluated the results based on two key metrics: Area Under the Circumstances (AUC) and Average Percentage (AUPR). These experiments used Datasets A and B. Each result represents the average obtained through a 10-fold cross-validation process. First, we compared the GAE and VGAE frameworks. The VGAE framework outperformed GAE in terms of both AUPR and AUC. We then investigated various encoder layer depths consisting of linear, GCN, and GAT layers. Experimental results showed that encoding with a combination of three GCN layers and one GAT layer achieved the best performance. The three GCN layers effectively captured local structural features, while the GAT layer improved performance by incorporating contextual information using an attention mechanism, focusing on key nodes and edges. Shallow models (e.g., two layers) cannot learn complex representations, while deeper models suffer from oversmoothing. Therefore, this combination offers the best balance between local and contextual information. This combination achieves a better balance between capturing local structure and incorporating relevant contextual information into the autoencoder, resulting in improved AUPR and AUC scores. The best performance was observed using the sum of the reconstruction loss, KL divergence loss, and L2 regularization. Including an L2 regularization term is beneficial because it helps mitigate overfitting and improves the model's generalization. We conducted experiments to evaluate the impact of different activation functions on encoder performance, comparing the use of ReLU, GELU, and a combination of GELU and ReLU in the encoder. The combination of GELU and ReLU achieved the best performance on both datasets, with higher AUC and AUPR scores than using either activation function alone. The results demonstrate that GELU effectively captures fine-grained nonlinear relationships, while ReLU introduces sparsity and computational efficiency.

[0123] When we compared various optimization algorithms, the model using AdamW achieved the best metrics and became the most suitable optimizer for this problem. By balancing weight decay and regularization properties, AdamW exhibited consistent convergence properties and robustness during training. We then found that using an adaptive learning rate could optimize the model's performance. Although several fixed learning rates (0.01, 0.001, and 0.0001) were explored, the adaptive learning rate strategy showed more favorable results. This strategy enhanced responsiveness and adaptability to evolving training dynamics. Finally, we investigated the impact of the data augmentation module on model performance. DAVGAE without data augmentation outperformed other compared methods, but its performance was weaker than the full DAVGAE model.

[0124] Based on ablation studies, each component of DAVGAE plays a crucial role in achieving the best AUPR and AUC scores for predicting disease-gene associations. The VGAE framework outperforms GAE by better capturing the network structure. The combination of three GCN layers and one GAT layer provides the best balance between local and contextual information. Adding an L2 regularization term to the loss function can reduce overfitting and improve generalization. AdamW proves to be the most effective optimizer, and adaptive learning rate further improves performance compared to a fixed learning rate. Finally, the data augmentation module significantly improves the model's ability to predict disease-gene associations, demonstrating its importance in the DAVGAE framework.

[0125] Specifically, a grid search strategy was employed to systematically explore the hyperparameter space and optimize the performance of DAVGAE. This process involved adjusting key parameters, including the similarity threshold for gene-disease pairings, the dropout rate, the number of GAT heads, and the regularization coefficient in the loss function. These hyperparameters were carefully selected to balance model performance and computational efficiency.

[0126] The similarity threshold is crucial for identifying associations between gene-disease pairs. Through grid search, we determined the optimal threshold to be 0.91 for Dataset A and 0.98 for Dataset B. These values ​​were chosen to improve the model's ability to distinguish relevant from irrelevant associations. A smaller similarity threshold retains more data pairs, but the excessive number of augmented samples may introduce noise and reduce model performance. Conversely, a larger threshold reduces the number of augmented pairs, which may limit the diversity of the training data and reduce the model's ability to generalize effectively. To prevent overfitting and improve model generalization, we employed a dropout rate. We evaluated model performance with various dropout rates ranging from 0.1 to 0.5. Performance on both datasets initially improved with increasing dropout rates, but performance declined when the dropout rate was too high. This observation highlights the importance of adjusting the dropout rate to maintain an ideal balance between overfitting and underfitting.

[0127] The number of heads in a GAT layer directly affects how the model aggregates information from neighboring nodes. We tested configurations with 2, 4, and 8 heads to evaluate their impact. Increasing the number of heads initially improves performance due to better representation learning. However, beyond a certain number, additional heads can introduce redundancy and computational overhead, leading to diminishing returns. The regularization coefficient in the loss function plays a crucial role in balancing the trade-off between model complexity and overfitting. We evaluated regularization coefficient values ​​between 0.00001 and 0.01, with a coefficient of 0.0001 achieving optimal performance, while values ​​that are too small or too large can lead to suboptimal results due to under- or over-regularization.

[0128] After multiple experimental adjustments, the final hyperparameter configuration of DAVGAE is: 100 training rounds, latent dimension and output dimension are 128 and 32 units respectively, random activation rate is 0.2, 2 GAT heads, regularization coefficient is 0.0001, and similarity thresholds of datasets A and B are 0.91 and 0.98 respectively. These hyperparameters combine to achieve a balance between performance and computational efficiency.

[0129] This application compares DAVGAE with traditional methods such as GeneHound, dgMDL, and MultiXrank, as well as state-of-the-art graph neural network models such as PGCN, LUPI, HNMME, DGP-PGTN, and MGREL on seven evaluation metrics, including precision and recall. To ensure a fair comparison, we also added new edges obtained through our data augmentation method to the training data of the compared methods. DAVGAE significantly outperformed the competing methods across all metrics on Datasets A and B. On Dataset C, where the gene-disease network connectivity is relatively poor, DAVGAE demonstrated overall superiority, but its recall and precision values ​​were not balanced. This indicates that DAVGAE excels at correctly identifying positive instances, but may miss a significant number, resulting in lower recall. PGCN ranked second overall. This method achieves a good balance between precision and recall, demonstrating its ability to effectively distinguish positive and negative instances. LUPI and HNMME performed moderately well, each with strengths and weaknesses across different metrics. On the other hand, GeneHound and dgMDL perform the lowest in all metrics, indicating that these methods may have difficulty in coping with the challenges posed by Dataset C.

[0130] To more fully understand the error distribution and performance characteristics of each method, we calculated a confusion matrix for all methods. This confusion matrix allows us to analyze each model's TP, TN, FP, and FN. By examining these components, we can better assess each method's ability to identify true gene-disease associations and understand where the most common errors lie, particularly when distinguishing between positive and negative instances. DAVGAE's TP values ​​significantly improve, especially with data augmentation, indicating a stronger ability to correctly identify positive examples. DAVGAE's FN values ​​are relatively low, indicating that the method effectively reduces the number of missed positive predictions. This enhancement effect is most pronounced when data augmentation is applied, further improving the model's ability to identify true positive gene-disease associations. In contrast, methods such as PGCN and LUPI perform more balanced between precision and recall, but they capture fewer positive instances (lower TP values) than DAVGAE, resulting in higher FN values. This suggests that while these methods make more conservative predictions, they often miss some true positive cases. On the other hand, GeneHound and dgMDL showed lower performance on all metrics, with higher false negatives and true positives, indicating that these methods have difficulty in accurately identifying true gene-disease relationships.

[0131] Comparing DAVGAE with and without data augmentation revealed that data augmentation significantly improved model performance. While DAVGAE without data augmentation already performed well, the model with data augmentation was even more robust, correctly identifying more true positives without a significant increase in false positives. This demonstrates the effectiveness of data augmentation in improving recall and overall prediction accuracy, particularly in gene-disease association prediction tasks, where positive examples are often underrepresented.

[0132] DAVGAE outperforms other methods on three datasets. This performance advantage demonstrates that our proposed method can accurately predict gene-disease interactions even under challenging conditions. Our future work will focus on improving DAVGAE's recall, which will further enhance its overall performance and practical utility in real-world applications.

[0133] To demonstrate the practical effectiveness of DAVGAE, we designed a case study experiment to discover the causal relationships of a specific disease. Specifically, we trained DAVGAE using all confirmed gene-disease associations in the dataset and used this model to predict all unknown gene-disease associations. We then selected the top 10 probable causal genes for a specific disease and searched for relevant literature evidence to verify the reliability of these predictions. Lung neoplasms were used as a case study and literature evidence was searched for associations with these 10 predicted genes. The top 10 genes with the highest association probabilities for this disease were TLR2, CST3, NR3C1, GUK1, GRPR, GSK3A, GLE1, GATA4, FGFR3, and PPARG. Eight of these top 10 predicted associations could be directly confirmed by literature evidence. For the unconfirmed gene, GUK1, prior studies have demonstrated differential expression of GUK1 in low-grade diffuse astrocytomas and pituitary tumors, suggesting a potential association between GUK1 and lung cancer. Similarly, for the unidentified gene GLE1, related studies have pointed out that GLE1 is a valuable therapeutic target for the treatment of giant cell tumor of bone, which is associated with lung metastasis.

[0134] In addition, to further verify the correlation between these top 10 predicted genes and lung tumors, we used the transcriptome data of lung squamous cell carcinoma (the most common primary tumor in lung tumors) from The Cancer Genome Atlas (TCGA, TCGA-LUSC), which included 52 normal samples and 503 primary tumor samples. We then used the Student's t-test to calculate the statistical results of the difference in gene expression levels between normal samples and primary tumor samples. As shown in Table 1, the results showed that the top 10 genes predicted by DAVGAE had significant expression differences between normal samples and tumor samples (p value < 0.05), which means that these genes may be involved in the occurrence and development of lung tumors.

[0135] In this application, DAVGAE is based on a variational graph autoencoder, integrating the biological and topological features of genes and diseases. It learns robust and discriminative feature representations through a four-layer graph neural network encoder, overcoming the limitations of a single data source. It uses an inner product decoder to predict gene-disease association scores. The decoder is concise and efficient, directly calculating the association strength based on features, simplifying the model complexity and ensuring prediction validity. To address the problem of incomplete and sparse gene-disease association data, DAVGAE proposes a data augmentation strategy based on cosine similarity. This strategy calculates the similarity score of genes and diseases, selects gene-disease pairs with high similarity for data augmentation, expands the training set and supplements potential true associations, effectively alleviating the data sparsity problem, avoiding the introduction of noise by blind data augmentation, and significantly improving data quality. Experimental results show that DAVGAE accurately predicts disease-related genes and outperforms other models in seven evaluation metrics, including precision and recall. We used lung cancer as a case disease, predicted the top 10 candidate causative genes for this disease, and searched for literature evidence associated with these 10 genes. Of these predictions, eight of the top ten predicted association genes could be directly confirmed through literature evidence. For the unconfirmed gene GUK1, studies have demonstrated differential expression of GUK1 in low-grade diffuse astrocytomas and pituitary tumors, suggesting a potential correlation between GUK1 and lung tumors. Similarly, for the unconfirmed gene GLE1, related studies have indicated that GLE1 is a valuable therapeutic target for giant cell tumors of bone, which are associated with lung metastasis. The correlation between these top ten predicted genes and lung tumors was further verified using transcriptome data from lung tumors in the Cancer Genome Atlas database. The results showed that the top ten genes predicted by DAVGAE showed significant expression differences between normal and tumor samples, suggesting that these genes may be involved in the occurrence and development of lung tumors. These findings demonstrate the effectiveness of DAVGAE in discovering potential pathogenic genes and can be used to explore therapeutic targets for diseases.

[0136] In the prediction method of pathogenic genes based on the fusion of topology and biological features provided in the embodiment of the present application, the DAVGAE model integrates the biological and topological features of genes and diseases, and uses a variational graph autoencoder for modeling. Through a four-layer graph neural network encoder, DAVGAE can effectively learn robust and discriminative feature representations of genes and diseases in the association network. This modeling method that integrates multiple information enables the model to understand the complex relationship between genes and diseases more comprehensively, overcoming the limitations of a single data source. DAVGAE uses an inner product decoder to predict the association scores of all gene-disease pairs. Compared with other complex decoder structures, the inner product decoder is concise and efficient, and can directly calculate the strength of the association between them based on the learned gene and disease representations. This design simplifies the complexity of the model while ensuring the effectiveness of the prediction.

[0137] To address the problem of incomplete and sparse gene-disease association data, the DAVGAE method proposes a novel data augmentation strategy. This strategy does not randomly generate new associations, but instead calculates the cosine similarity between genes and diseases based on their biological and topological characteristics. Gene-disease pairs with higher similarity scores are then selected to enhance the association data, thereby effectively expanding the training set and supplementing potential real associations in a more targeted manner, alleviating the negative impact of data sparsity on model performance. This is one of the core innovations of this patent, which avoids the noise that may be introduced by blind data augmentation and significantly improves the quality of the data. DAVGAE effectively solves the problems of data sparsity and heterogeneous data integration in gene-disease association prediction by combining a similarity-based data augmentation strategy, a variational graph autoencoder model that integrates multiple features, and an efficient inner product decoder, and performs better than other comparative models in comprehensive experiments on three gene-disease association datasets.

[0138] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0139] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for predicting pathogenic genes based on the fusion of topology and biological features, characterized in that: The method comprises: Obtain the gene to be tested; The gene to be detected is input into the trained DAVGAE model to predict the correlation between the gene to be detected and the disease, and obtain the disease gene correlation prediction conclusion; The DAVGAE model includes: a data enhancement module, an encoder and an inner product decoder.

2. The method according to claim 1, characterized in that The data enhancement module performs the following steps: Establishing a gene-disease network based on the set of actually connected gene-disease node pairs, gene nodes, and disease nodes, and determining an adjacency matrix of the gene-disease network; The Node2vec model is used on the gene-disease network to obtain the topological characteristics of each node in the gene node and disease node; Calculate the cosine similarity of all edges formed between each gene node and a disease node in the gene node and disease node; If the cosine similarity of a gene-disease pair is greater than a preset threshold, it is added to the positive training set; A set of new pairs with the same number is randomly selected from the positions where the elements in the adjacency matrix are 0, and the set is added to the negative training set.

3. The method according to claim 2, characterized in that The Node2vec model is used to capture relevant network information and compare different types of nodes in a shared topological space. The following formula is used to calculate c(v p , v q ), Among them, Vp is the gene node, is the topological feature representation of Vp, Vq is the disease node, is the topological characteristic representation of Vq, c(v p , v q ) is the association probability between gene node Vp and disease node Vq, is the gene node vector with disease node vector The cosine similarity of .

4. The method according to claim 3, characterized in that The association density between genes and diseases in the gene-disease network is calculated using the following formula: in, It's G gd The number of gene nodes in It's G gd The number of disease nodes in G gd is the gene-disease network, and AD represents the association density.

5. The method according to claim 1, wherein The encoder includes a first graph convolution layer, a first GELU activation function, a second graph convolution layer, a second GELU activation function, a first graph convolutional neural network layer, and a second graph convolutional neural network layer; Among them, a dual-headed graph attention mechanism is applied between the first and second graph convolutional layers, so that the model pays attention to the most relevant neighbors when aggregating features; After applying the attention mechanism, the mean and log-variance outputs are obtained using the first and second graph convolutional neural network layers.

6. The method according to claim 5, characterized in that The input data of the encoder includes first input data and second input data, wherein the first input data is a gene-disease association network G' obtained by data augmentation, and the second input data is a node feature matrix X∈RN×f of the nodes in G', where N is the number of nodes in G', f is the feature dimension, and R represents a set of real numbers. In the encoder model, the node feature matrix passes through the GCN layer to obtain the hidden representation; GELU activation is applied to the output to obtain H1, and the dropout layer is applied for regularization; GELU is applied after the GCN layer to provide smooth gradient transition and capture complex nonlinear relationships.

7. The method according to claim 1, characterized in that The inner product decoder of DAVGAE calculates the inner products between latent variables to produce a new adjacency matrix by the following formula: A new =Sigmoid(ZZ Τ ); Among them, A new ∈R N×N , A new The value in is the probability of association between the disease and the gene pair, R N×N represents an N×N matrix, and Z represents the potential embedding matrix of the nodes in the network G'.

8. The method according to claim 1, characterized in that The DAVGAE model is trained by the following steps: The loss function of DAVGAE during training consists of three parts: reconstruction loss unit, KL divergence loss unit and L2 regularization unit; The reconstruction loss unit quantifies the error between the predicted adjacency matrix and the original adjacency matrix calculated for the positive training edges, so that the DAVGAE model reconstructs the adjacency matrix. The reconstruction loss unit calculates the reconstruction error of each node pair and takes the expected value to determine the loss function of the DAVGAE model; The KL divergence loss unit aligns the learned latent variable distribution q(Z|X,A) with the prior distribution p(Z), which is usually assumed to be a standard normal distribution to ensure that the latent space remains well-structured and prevent overfitting; The L2 regularization unit adds a penalty for excessively large model parameters to prevent overfitting and improve generalization ability.

9. The method according to claim 8, characterized in that The total loss function of the DAVGAE model is calculated by the following formula: L=E q(Z|X,A) (logp(A|Z))+KL(q(Z|X,A)||p(Z))+λ||Θ|| 2 ; Where p(A|Z) is the generative marginal probability distribution, which quantifies the likelihood of generating A using the latent variable Z, q(Z|X,A) is the encoder that outputs the conditional distribution of the latent variable, p(Z) is the prior distribution of the latent variable, λ is the regularization coefficient, and Θ contains all the model parameters; The probability distribution of each element in the adjacency matrix is ​​generated by the following formula: where Apq indicates whether there is a connection between gene node p and disease node q, zvp and zvq are the embedding vectors of the nodes, σ(·) is the sigmoid activation function, and A is the adjacency matrix.

10. The method according to claim 1, characterized in that The method further comprises: A grid search strategy was adopted to explore the hyperparameter space and optimize the performance of DAVGAE, tuning key parameters including the similarity threshold of gene-disease pairings, dropout rate, number of GAT heads, and regularization coefficient in the loss function.

Images