Lncrna-disease prediction method and system based on multi-source biological information fusion and graph neural network

Abstract

The application discloses a LncRNA-disease prediction method and system based on multi-source biological information fusion and a graph neural network. The method comprises the following steps: constructing multi-source heterogeneous biological data, including a LncRNA-disease correlation matrix, a LncRNA-miRNA correlation matrix, a disease-miRNA correlation matrix and prior similarity information; performing consensus enhancement similarity fusion processing on the prior similarity information, introducing a fixed fusion factor for weighted balance, and utilizing the fusion similarity of the last iteration round for feedback updating to obtain a fusion similarity matrix; constructing a multi-view heterogeneous biological graph and inputting the graph into an attention self-encoding network to obtain optimized node representation; fusing the node representation with multi-source features, performing layer-by-layer nonlinear transformation on the features through a residual attention multilayer perception machine, and outputting a LncRNA-disease correlation probability result. The application realizes deep fusion of multi-source heterogeneous information, and improves the correlation prediction accuracy and stability.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and artificial intelligence, and relates to a method and system for predicting LncRNA-diseases using multi-source biological information fusion and graph neural networks. Background Technology

[0002] LncRNAs play a crucial role in gene expression regulation, tumorigenesis and development, and various complex disease mechanisms. However, traditional wet laboratory experiments for identifying the association between LncRNAs and diseases suffer from drawbacks such as long processing times, high costs, and low throughput.

[0003] Existing computational methods primarily rely on single-type biological information or shallow machine learning models, which have the following shortcomings: First, the fusion methods for multi-source biological data (such as lncRNA, miRNA, and disease information) are simple and fail to fully explore cross-modal associations. Second, traditional graph models treat neighbor nodes uniformly, ignoring the differences in the importance of different biological relationships. Third, under high-dimensional and heterogeneous feature conditions, traditional classifiers or shallow models struggle to model nonlinear high-dimensional features, leading to unstable prediction performance. Therefore, there is an urgent need for a deep learning method that can efficiently fuse multi-source biological information and possess strong expressive power. Summary of the Invention

[0004] This invention aims to provide a method and system for predicting LncRNA-related diseases using multi-source biological information fusion and graph neural networks, to address the problems of insufficient feature fusion, weak graph structure modeling ability, and low prediction accuracy in existing technologies. To achieve the above objectives, this invention adopts the following technical solution: A multi-source bioinformatics fusion and graph neural network-based method for LncRNA disease prediction, including: A multi-source heterogeneous biological dataset is constructed, which includes a LncRNA-disease association matrix, a LncRNA-miRNA association matrix, a disease-miRNA association matrix, and prior similarity information; the prior similarity information includes functional / semantic similarity and Gaussian interaction contour kernel similarity. Functional / semantic similarity includes: LncRNA functional similarity, miRNA functional similarity, and disease semantic similarity; Gaussian interaction contour kernel similarity includes: LncRNA similarity calculated based on Gaussian interaction contour kernel, miRNA similarity calculated based on Gaussian interaction contour kernel, and disease similarity calculated based on Gaussian interaction contour kernel; A consensus-enhanced similarity framework is constructed, and the functional / semantic similarity and Gaussian interaction contour kernel similarity are iteratively fused to obtain an initial consensus similarity matrix. In the i-th iteration, a fixed fusion factor is introduced to adjust the contribution ratio of functional / semantic similarity and Gaussian interaction contour kernel similarity in the iteration process. The consensus similarity matrix obtained in the previous iteration is used to update the current similarity matrix to obtain the fused similarity matrix. A multi-view heterogeneous biological graph containing LncRNA nodes, miRNA nodes, and disease nodes was constructed based on the fusion of similarity matrices and multi-source heterogeneous biological data. The multi-view heterogeneous biological graph is input into the graph attention autoencoder network. The encoder performs weighted aggregation of neighborhood node information based on the attention mechanism to obtain node embedding representation. The decoder reconstructs the graph structure and node features to jointly optimize structural and attribute information, thereby obtaining the optimized node representation. The optimized node representation is fused with multi-source biological features to form a joint feature representation, which is mapped to a unified feature space. Layer-by-layer nonlinear feature transformation is performed through a multilayer perceptron based on residual structure. The attention pooling mechanism is used to adaptively weight and fuse different features to obtain the final feature representation. The final feature representation is input into the prediction model to output the LncRNA-disease association probability result.

[0005] Furthermore, functional / semantic similarity and Gaussian interaction contour kernel similarity are iteratively fused. The iterative fusion includes normalization fusion and setting diagonal elements to zero to obtain an initial consensus similarity matrix: ; in, For initial functional / semantic similarity; For the initial Gaussian interaction contour kernel similarity; This is the initial consensus similarity matrix; This is a normalized fusion operation; This indicates the operation of setting the diagonal elements to zero; In the i-th iteration, the update process of functional / semantic similarity and Gaussian interaction contour kernel similarity is expressed as: ; ; Where α is a fixed fusion factor used to adjust the contribution ratio of the two types of similarity information in the iteration process; For the i-th functional / semantic similarity; For the i-th Gaussian interaction contour kernel similarity; Let be the consensus similarity matrix for the (i-1)th iteration; after each iteration update, the fused similarity matrix is ​​normalized, and the diagonal elements of the fused similarity matrix are set to zero.

[0006] Furthermore, the graph neural network is a graph attention autoencoder network, which learns the representation of nodes in a multi-view heterogeneous biological graph through an encoder-decoder symmetric structure. The encoder performs adaptive weighted aggregation of neighboring node information based on an attention mechanism to obtain a low-dimensional embedding representation of nodes that integrates structural and attribute information. The decoder reconstructs the structure and features of the node embedding representation based on a reverse attention mechanism.

[0007] Furthermore, the low-dimensional embedding representation is obtained by minimizing the joint loss function of feature reconstruction error and structure preservation error.

[0008] Furthermore, the optimized node representation is fused with multi-source biological features to form a joint feature representation, and the fusion method is splicing fusion.

[0009] Furthermore, the residual attention multilayer perceptron includes a residual connection structure, a normalization layer, and an attention pooling module. The residual structure is used to enhance gradient propagation capability, and the attention module generates corresponding attention weights based on the importance of different feature representations. The features are then weighted and summed based on the attention weights to obtain the final fused representation.

[0010] Furthermore, the association probability is obtained by mapping through a fully connected output layer, and a binary cross-entropy loss function is used to constrain the difference between the prediction result and the true label, so as to achieve supervised optimization training of the association between LncRNA and disease.

[0011] This invention also provides a multi-source biological information fusion and graph neural network-based LncRNA disease prediction system, the system comprising: The multi-source bioinformatics construction module is used to acquire multi-source biological data of LncRNA, miRNA and diseases, calculate functional / semantic similarity and Gaussian interaction contour kernel similarity, and construct LncRNA-disease association matrix, LncRNA-miRNA association matrix and disease-miRNA association matrix to form a multi-view heterogeneous biological map; The consensus-enhancing feature fusion module is used to perform consensus-enhancing iterative fusion of functional / semantic similarity and Gaussian interaction contour kernel similarity to generate a fused similarity network containing multi-scale topological information as a unified basic feature representation. The graph attention representation learning module is used to input multi-view heterogeneous biological graphs into the graph attention autoencoder network to obtain low-dimensional embedding representations of LncRNAs and disease nodes; The joint feature fusion module is used to splice and fuse the low-dimensional embedding representation of LncRNA and disease nodes with the basic feature representation to construct the LncRNA-disease joint feature representation; The residual attention prediction module is used to perform a layer-by-layer nonlinear transformation of the joint feature representation of LncRNA and disease based on the residual structure of the multilayer perceptron. It combines the attention pooling mechanism to adaptively weight different feature dimensions and finally outputs the association probability between LncRNA and disease.

[0012] The beneficial effects of this invention are: This invention constructs a multi-source similarity network between LncRNA, miRNA and disease, and introduces a consistency enhancement fusion strategy to achieve collaborative modeling of different biological information sources, effectively improving the sufficiency of fusion and expression integrity of heterogeneous biological data.

[0013] This invention employs a graph attention autoencoder network to encode heterogeneous biological graphs with multiple views. Through an adaptive neighbor weight allocation mechanism, it effectively captures high-order topological relationships and nonlinear dependency structures in LncRNA-disease networks, thereby improving the expressive power of node embedding representations.

[0014] This invention constructs a joint feature representation by fusing the low-dimensional embedding representation learned by the graph attention autoencoder with multi-source initial topological features, while preserving structural and semantic information, effectively improving the information richness and discriminativeness of the feature representation.

[0015] This invention employs a multilayer perceptron model based on residual structure and attention mechanism. It enhances the training stability of deep networks through residual connections and adaptively adjusts feature weights using the attention mechanism, effectively improving the modeling capability of complex nonlinear mapping relationships. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the consensus enhancement framework of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be described in detail and clearly below with reference to the embodiments.

[0019] like Figure 1 As shown, this invention provides a method for LncRNA-based disease prediction using multi-source biological information fusion and graph neural networks, specifically including the following steps: Construction of Multi-Source Heterogeneous Biological Data: Multi-source heterogeneous biological data is constructed, including LncRNA-disease association matrices, LncRNA-miRNA association matrices, disease-miRNA association matrices, and prior similarity information. Prior similarity information includes functional / semantic similarity and Gaussian interaction contour kernel similarity. Functional / semantic similarity includes LncRNA functional similarity, miRNA functional similarity, and disease semantic similarity. Gaussian interaction contour kernel similarity includes LncRNA similarity calculated based on the Gaussian interaction contour kernel, miRNA similarity calculated based on the Gaussian interaction contour kernel, and disease similarity calculated based on the Gaussian interaction contour kernel. Consensus-enhanced similarity fusion: A consensus-enhanced similarity framework is constructed to iteratively fuse the functional / semantic similarity and Gaussian interaction contour kernel similarity of three types of entities (lncRNA, miRNA, and disease) to obtain an initial consensus similarity matrix. In the i-th iteration, a fixed fusion factor is introduced to adjust the contribution ratio of functional / semantic similarity and Gaussian interaction contour kernel similarity in the iteration process. The consensus similarity matrix obtained in the previous iteration is used to update the current similarity matrix to obtain the fused similarity matrix. Construction of multi-view heterogeneous biological graphs: Based on the fusion of similarity matrices and multi-source heterogeneous biological data, a multi-view heterogeneous biological graph containing LncRNA nodes, miRNA nodes, and disease nodes is constructed. Graph attention autoencoder representation learning: Input a multi-view heterogeneous biological graph into a graph attention autoencoder network. The encoder performs weighted aggregation of neighborhood node information based on the attention mechanism to obtain node embedding representation. The decoder reconstructs the graph structure and node features to jointly optimize structural and attribute information and obtain optimized node representation. Joint Feature Fusion and Residual Attention Prediction: The optimized node representation is fused with multi-source biological features to form a joint feature representation, which is mapped to a unified feature space. Layer-by-layer nonlinear feature transformation is performed through a multilayer perceptron based on residual structure. The attention pooling mechanism is combined to adaptively weight and fuse different features to obtain the final feature representation. The final feature representation is input into the prediction model to output the LncRNA-disease association probability result.

[0020] Example 1: Multi-source heterogeneous biological data were acquired, including data on three types of biological entities: lncRNAs, miRNAs, and diseases. LncRNA-disease association data were obtained from the LncRNA Disease and Lnc2Cancer databases; miRNA-disease association data were obtained from the HMDD database; LncRNA-miRNA interaction data were obtained from the starBase database; and disease semantic information was obtained from the Disease Ontology database. The above data were uniformly numbered and normalized to construct a standardized set of biological entities, generating initial LncRNA-disease association matrices LD, LM, and MD. Based on the acquired biological entity association data, functional / semantic similarity and Gaussian interaction contour kernel similarity were calculated respectively. Functional / semantic similarity included: disease semantic similarity, LncRNA functional similarity and miRNA functional similarity. Gaussian interaction profile kernel similarity includes: LncRNA similarity calculated based on Gaussian interaction profile kernel, miRNA similarity calculated based on Gaussian interaction profile kernel, and disease similarity calculated based on Gaussian interaction profile kernel; Based on the calculated similarity information, three homogeneous networks of LncRNA, miRNA, and disease were constructed, generating corresponding LncRNA similarity matrices, miRNA similarity matrices, and disease similarity matrices, providing a foundation for subsequent multi-source heterogeneous network fusion.

[0021] Based on the Disease Ontology database, a directed acyclic graph corresponding to each disease is constructed. The semantic contribution value of each disease is calculated according to the hierarchical relationship between disease nodes, and the semantic similarity matrix (DSS) is further obtained to characterize the degree of semantic association between diseases.

[0022] Based on the biological hypothesis that functionally similar LncRNAs or miRNAs tend to be associated with semantically similar diseases, and combining known LncRNA-disease associations and disease semantic similarity information, the LncRNA functional similarity matrix and the miRNA functional similarity matrix are calculated respectively.

[0023] To address the issues of incomplete semantic information about diseases and the lack of functional annotations for some entities, entity interaction spectral vectors are constructed based on the association matrix. The Gaussian kernel similarity matrix between LncRNA, miRNA, and disease is calculated using the LncRNA similarity method based on Gaussian interaction contour kernels, which is used to supplement the deficiencies of traditional similarity information.

[0024] like Figure 2As shown, a consensus-enhanced similarity framework (CAME) is constructed. First, the initial functional / semantic similarity and the initial Gaussian interaction contour kernel similarity are normalized and fused, and the diagonal elements are set to zero to obtain the initial consensus similarity matrix. In the i-th iteration, the two types of similarity are updated using the consensus similarity matrix from the previous round. The updated functional / semantic similarity and the Gaussian interaction contour kernel similarity are normalized and fused again, and the diagonal elements are set to zero to obtain the current consensus similarity matrix. Convergence is then judged. If convergence is achieved, the fused similarity matrix is ​​output; otherwise, the iteration continues.

[0025] Specifically, the iterative fusion of the functional / semantic similarity and the Gaussian interaction contour kernel similarity includes normalization fusion and setting the diagonal elements to zero to obtain an initial consensus similarity matrix: ; in, For initial functional / semantic similarity; For the initial Gaussian interaction contour kernel similarity; This is the initial consensus similarity matrix; This is a normalized fusion operation; This indicates the operation of setting the diagonal elements to zero; In the i-th iteration, the update process of functional / semantic similarity and Gaussian interaction contour kernel similarity is expressed as: ; ; Where α is a fixed fusion factor used to adjust the contribution ratio of the two types of similarity information in the iteration process; For the i-th functional / semantic similarity; For the i-th Gaussian interaction contour kernel similarity; Let be the consensus similarity matrix for the (i-1)th iteration; after each iteration, the fused similarity matrix is ​​normalized, and the diagonal elements of the fused similarity matrix are set to zero. In this embodiment, by traversing four iteration rounds, the round with the best combined AUC and AUPR on the validation set is selected as the model convergence point, and the consensus similarity matrix corresponding to this round is the final output, thus obtaining a fusion similarity network containing multi-scale topological information.

[0026] Based on the LncRNA homogeneous network, miRNA homogeneous network, and disease homogeneous network, and combining the original association relationships between LncRNA-disease, LncRNA-miRNA, and miRNA-disease, a unified multi-view heterogeneous biological map is constructed. This network is represented in block matrix form. ; Wherein, LD is the initial LncRNA-disease association matrix, LM is the LncRNA-miRNA association matrix, and DM is the miRNA-disease association matrix; The weighted fusion matrix of LncRNA consensus is the i-th layer. Let i be the disease consensus similarity matrix after fusion at layer i; Let i be the consensus similarity matrix of the fused miRNAs at the i-th layer; For the transpose of LD; This is the transpose of LM; This is the transpose of DM; the diagonal submatrix represents the similarity relationship within each biological entity, and the off-diagonal submatrix represents the interaction relationship between different biological entities in the same layer.

[0027] The obtained multi-view heterogeneous biological graph input graph attention autoencoder is used for nonlinear feature extraction to fully explore the high-order topological associations between LncRNA, miRNA and disease nodes.

[0028] During the encoding process, an attention mechanism is first used to calculate the association weights between a node and its neighboring nodes. Through adaptive learning of the importance of different neighboring nodes, the model can assign different weights based on different biological relationships. In the k-th layer encoder, LncRNA nodes... Adjacent disease nodes The correlation strength is defined by the following formula: ; in, For nodes The feature vector output by the encoder at layer k-1; For nodes The feature vector output by the encoder at layer k-1; They represent the first time. The three learnable parameters of the layer; for transpose; for transpose, This represents the ReLU activation function.

[0029] Subsequently, the association weights between neighboring nodes are normalized to obtain the attention weights between nodes; therefore, in the k-th layer encoder, the node The representation of is obtained by the following formula: ; in, For nodes The feature vector output by the encoder at the k-th layer; Represents a node The neighboring nodes; The attention weights between nodes; For nodes The set of neighboring nodes; After multi-layer graph attention encoding, a low-dimensional node embedding representation is obtained. At the same time, the input network structure and node features are reconstructed by the decoder, so that the generated low-dimensional representation can maintain the topological structure information in the original heterogeneous network and obtain a biological entity feature representation with stronger expressive power.

[0030] The model is trained using a joint optimization strategy of feature reconstruction loss and structure preservation loss, and its total loss function is defined as: ; in, For feature reconstruction loss, mean squared error is used to measure the difference between the reconstructed node features and the original features; The structure preservation loss is calculated using the binary cross-entropy loss function; λ is the weight of the structure preservation loss; in this embodiment, the model achieves optimal performance when λ is 0.1.

[0031] The calculation method is as follows: ; in, Represents the original node features. This represents the reconstructed node features, where N is the number of samples.

[0032] The reconstruction error used to measure the structural relationships between nodes is defined as follows: ; Where M is the number of edges. Let be the cosine similarity between node pairs. This represents the Sigmoid activation function. The label represents the positive or negative sample label. A sample label of 1 indicates that the input edge is a positive sample, and a sample label of 0 indicates that the input edge is a negative sample.

[0033] The obtained topological feature representation is fused with the graph embedding feature representation to construct the final joint feature representation: The initial topological feature representation is constructed, and the adjacency structure is represented as follows: ; in, This represents the initial topological features; This indicates that the first step of the consensus-enhanced similarity framework... Features of layered LncRNA homogeneous networks This indicates that the first step of the consensus-enhanced similarity framework... Characteristics of homogeneous disease networks at different levels , , DM represent the association matrices between LncRNA-disease, LncRNA-miRNA, and disease-miRNA, respectively; express The matrix transpose.

[0034] Subsequently, the graph embedding representation and the topological feature representation are fused, specifically using a feature concatenation method: ; ; in, and The graph embedding representations of LncRNA and disease are respectively represented. Specifically, the multi-view heterogeneous biological graph is input into the graph attention autoencoder network. After the encoder attention aggregation and decoder reconstruction optimization, the graph embedding representations of LncRNA and disease are extracted. for and The image embedding representation obtained by vertical splicing; This represents the initial topological feature representation; Indicates feature concatenation operation; This represents the final feature representation after fusion.

[0035] A residual attention-based multi-layer perceptron (RAMLP) is constructed to predict lncRNA-disease associations. The fused feature representations are input into the RAMLP as lncRNA-disease pairs to predict the embedding associations between lncRNAs and diseases.

[0036] The high-dimensional fusion features of the input are transformed into a unified latent space representation through a linear mapping layer. The process is as follows: ; Where X represents the input feature. and These represent the weight matrix and bias term of the linear transformation, respectively; This represents the initial feature representation after mapping.

[0037] The initial features are input into a multilayer perceptron network composed of multiple residual blocks. Residual connections enhance the model's ability to learn deep nonlinear features, avoiding information loss caused by increasing network depth. In each residual module, the feature update process is represented as follows: ; in, This represents a nonlinear transformation function composed of a multilayer sensing mechanism; Represents the features in the i-th residual module; This represents the feature in the (i-1)th residual module.

[0038] A feature-level attention mechanism is introduced to adaptively weight different feature representations. First, the attention score for each feature is calculated: ; in, The attention score is the value corresponding to the z-th feature. and This is a learnable attention parameter matrix; For bias terms; This represents the z-th eigenvalue in the eigenvector. The activation function for the Gaussian error linear unit; Subsequently, the scores are normalized along the feature dimension using the Softmax function to obtain the attention weights. Based on the obtained attention weights, the feature embeddings are weighted to obtain the fused LncRNA-disease pair feature vector; the LncRNA-disease pair feature vector is fed into the prediction head consisting of a Dropout layer, a linear transformation layer, and a Sigmoid activation function to calculate the prediction probability that there is an association between the current LncRNA and the disease. The training objective is to minimize the binary cross-entropy: ; Where y is the true label, and if the LncRNA-disease pair is associated, then y=1, otherwise y=0; This indicates the predicted probability that there is an association between the current LncRNA and the disease.

[0039] This embodiment employs a five-fold cross-validation strategy for model training and optimization. Specifically, known LncRNA-disease associations are used as positive samples, while an equal number of samples are randomly sampled from unknown association pairs as negative samples to construct the training dataset. This embodiment is implemented on two datasets: Dataset 1 contains 240 lncRNAs, 495 miRNAs, and 412 diseases, as well as 2,697 experimentally validated lncRNA-disease associations (LDAs), 13,562 miRNA-disease associations (MDAs), and 1,002 lncRNA-miRNA interactions. Dataset 2 contains 665 LncRNAs, 316 diseases, 295 miRNAs, 3833 LDAs, 2108 LncRNA-miRNA interactions, and 8540 MDAs; Among them, LncRNA-disease associations (LDAs) were obtained from the LncRNA Disease database and the Lnc2Cancer database; miRNA-disease associations (MDAs) were obtained from the HMDD database; and LncRNA-miRNA interactions were obtained from the starBase database.

[0040] During model training, the Adam optimizer is used to iteratively update the model parameters, and the model is evaluated and tuned by the loss function performance on the validation set, thereby determining the optimal model parameters.

[0041] To comprehensively evaluate the predictive performance of the constructed model, this embodiment evaluates the model's performance from multiple dimensions. Specifically, multiple standard evaluation indicators are used to quantitatively analyze the LncRNA-disease association prediction results, including AUC, AUPR (area under the precision-recall curve), Matthews Correlation Coefficient (MCC), accuracy, precision, recall, and F1 score, to comprehensively measure the model's predictive ability and stability under different evaluation dimensions.

[0042] Table 1 shows the performance metrics of the model of this invention on the two selected datasets: Table 1 Comparison of performance metrics of this method on the two selected datasets.

[0043] As shown in Table 1, on Dataset 1, the AUC and AUPR reached 0.9908 and 0.9906 respectively, indicating that the model has good ranking and discrimination abilities for LncRNA-disease associations; the Precision was 0.9682, Recall was 0.9481, and F1-score was 0.9580, indicating that the model has good positive sample recall ability while maintaining high accuracy. On Dataset 2, the AUC and AUPR reached 0.9448 and 0.9446 respectively, slightly lower than Dataset 1, but still maintained a high level overall. In addition, to evaluate the effectiveness of this method in predicting LncRNA-disease associations, it was compared with four existing methods: DPFELDA (Dual-Path Feature Extraction Network for LncRNA-Disease Associations) employs a dual-path feature extraction network (Yao D, Zhang B, Zhan X, et al. Predicting lncRNA–Disease Associations Based on a Dual-Path Feature Extraction Network with Multiple Sources of Information Integration[J]. ACSomega, 2024, 9(32): 35100.). LDA-GARB (LncRNA-Disease Associations-Graph Autoencoder and noise-robust gradient Boosting) employs a graph autoencoder and noise-robust gradient boosting (Tang L, Huang L, Yuan Y. Predicting lncRNA and disease associations with graph autoencoder and noise robust gradient boosting[J]. Scientific Reports, 2025, 15(1): 19178.). GCLMTP (Graph Contrastive Learning for Multi-Task Prediction) employs multi-task prediction-based graph contrastive learning (Sheng N, Wang Y, Huang L, et al. Multi-task prediction-based graph contrastive learning for inferring the relationship among lncRNAs, miRNAs and diseases[J]. Briefings in bioinformatics, 2023, 24(5):bbad276.). VGAELDA (Variational Graph Autoencoder for LncRNA-Disease Associations) employs a variational graph autoencoder. (Shi Z, Zhang H, Jin C, et al. A presentation learning model based on variational inference and graph autoencoder for predicting lncRNA-disease associations[J]. BMCbioinformatics, 2021, 22(1): 136.) To ensure the fairness of the experimental comparison, all four comparison methods used the default parameter configuration in the paper, and all methods were trained and tested in the same experimental environment, including the same dataset partitioning method (5-fold cross-validation), the same positive and negative sample ratio, and the same evaluation metric calculation method.

[0044] Paired t-tests were used to test the significance of pairwise differences. Specifically, at each fold of the five-fold cross-validation, the differences in AUC and AUPR between our method and other comparative methods were calculated, resulting in five difference samples. The overall mean of this difference sequence was then tested to see if it was significantly greater than zero.

[0045] As shown in Tables 2 and 3, each P-value in the tables independently represents the pairwise test result between the method of the present invention and the single comparison method of the corresponding column: Table 2 compares the significance of the difference between our method and other comparison methods on Dataset 1.

[0046] Table 3 compares the significance of the difference between our method and other comparison methods on dataset 2.

[0047] As shown in Tables 2 and 3, the performance improvement of the method of the present invention compared with the four comparative methods mentioned above is statistically significant (p<0.05), among which the improvement of the method of the present invention compared with the LDA-GARB method is the most significant.

[0048] This invention also provides a multi-source biological information fusion and graph neural network-based LncRNA disease prediction system, the system comprising: The multi-source bioinformatics construction module is used to acquire multi-source biological data of LncRNA, miRNA and diseases, calculate functional / semantic similarity and Gaussian interaction contour kernel similarity, and construct LncRNA-disease association matrix, LncRNA-miRNA association matrix and disease-miRNA association matrix to form a multi-view heterogeneous biological map; The consensus-enhancing feature fusion module is used to perform consensus-enhancing iterative fusion of functional / semantic similarity and Gaussian interaction contour kernel similarity to generate a fused similarity network containing multi-scale topological information as a unified basic feature representation. The graph attention representation learning module is used to input multi-view heterogeneous biological graphs into the graph attention autoencoder network to obtain low-dimensional embedding representations of LncRNAs and disease nodes; The joint feature fusion module is used to splice and fuse the low-dimensional embedding representation of LncRNA and disease nodes with the basic feature representation to construct the LncRNA-disease joint feature representation; The residual attention prediction module is used to perform a layer-by-layer nonlinear transformation of the joint feature representation of LncRNA and disease based on the residual structure of the multilayer perceptron. It combines the attention pooling mechanism to adaptively weight different feature dimensions and finally outputs the association probability between LncRNA and disease.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source biological information fusion and graph neural network-based LncRNA disease prediction method, characterized in that, The methods include: A multi-source heterogeneous biological dataset is constructed, which includes a LncRNA-disease association matrix, a LncRNA-miRNA association matrix, a disease-miRNA association matrix, and prior similarity information; the prior similarity information includes functional / semantic similarity and Gaussian interaction contour kernel similarity. A consensus-enhanced similarity framework is constructed, and the functional / semantic similarity and Gaussian interaction contour kernel similarity are iteratively fused to obtain an initial consensus similarity matrix. In the i-th iteration, a fixed fusion factor is introduced to adjust the contribution ratio of functional / semantic similarity and Gaussian interaction contour kernel similarity in the iteration process. The consensus similarity matrix obtained in the previous iteration is used to update the current similarity matrix to obtain the fused similarity matrix. A multi-view heterogeneous biological graph containing LncRNA nodes, miRNA nodes, and disease nodes was constructed based on the fusion of similarity matrices and multi-source heterogeneous biological data. The multi-view heterogeneous biological graph is input into the graph attention autoencoder network. The encoder performs weighted aggregation of neighborhood node information based on the attention mechanism to obtain node embedding representation. The decoder reconstructs the graph structure and node features to jointly optimize structural and attribute information, thereby obtaining the optimized node representation. The optimized node representation is fused with multi-source biological features to form a joint feature representation, which is mapped to a unified feature space. Layer-by-layer nonlinear feature transformation is performed through a multilayer perceptron based on residual structure. The attention pooling mechanism is used to adaptively weight and fuse different features to obtain the final feature representation. The final feature representation is input into the prediction model to output the LncRNA-disease association probability result.

2. The LncRNA-disease prediction method based on multi-source biological information fusion and graph neural networks according to claim 1, characterized in that, The functional / semantic similarity and Gaussian interaction contour kernel similarity are iteratively fused. The iterative fusion includes normalization fusion and setting the diagonal elements to zero to obtain an initial consensus similarity matrix. ； in, For initial functional / semantic similarity; For the initial Gaussian interaction contour kernel similarity; This is the initial consensus similarity matrix; This is a normalization fusion operation; This indicates the operation of setting the diagonal elements to zero; In the i-th iteration, the update process of functional / semantic similarity and Gaussian interaction contour kernel similarity is expressed as: ； ； in, A fixed fusion factor is used to adjust the contribution ratio of the two types of similarity information during the iteration process; For the i-th functional / semantic similarity; For the i-th Gaussian interaction contour kernel similarity; Let be the consensus similarity matrix for the (i-1)th iteration; after each iteration update, the fused similarity matrix is ​​normalized, and the diagonal elements of the fused similarity matrix are set to zero.

3. The LncRNA-disease prediction method based on multi-source biological information fusion and graph neural networks according to claim 1, characterized in that, The graph neural network is a graph attention autoencoder network. It learns the representation of nodes in a multi-view heterogeneous biological graph through an encoder-decoder symmetric structure. The encoder performs adaptive weighted aggregation of neighboring node information based on an attention mechanism to obtain a low-dimensional embedding representation of nodes that integrates structural and attribute information. The decoder reconstructs the structure and features of the node embedding representation based on a reverse attention mechanism.

4. The LncRNA-disease prediction method based on multi-source biological information fusion and graph neural networks according to claim 3, characterized in that, The low-dimensional embedding representation is obtained by minimizing the joint loss function of feature reconstruction error and structure preservation error.

5. The LncRNA-disease prediction method based on multi-source biological information fusion and graph neural networks according to claim 1, characterized in that, The optimized node representation is fused with multi-source biological features to form a joint feature representation, and the fusion method is splicing fusion.

6. The LncRNA-disease prediction method based on multi-source biological information fusion and graph neural networks according to claim 1, characterized in that, The residual attention multilayer perceptron includes a residual connection structure, a normalization layer, and an attention pooling module. The residual structure is used to enhance gradient propagation capability. The attention module generates corresponding attention weights based on the importance of different feature representations and performs weighted summation on the features based on the attention weights to obtain the final fused representation.

7. The LncRNA-disease prediction method based on multi-source biological information fusion and graph neural networks according to claim 1, characterized in that, The association probability is obtained by mapping through a fully connected output layer, and a binary cross-entropy loss function is used to constrain the difference between the prediction result and the true label, so as to achieve supervised optimization training of the association between LncRNA and disease.

8. A multi-source biological information fusion and graph neural network-based LncRNA disease prediction system, used to implement the multi-source biological information fusion and graph neural network-based LncRNA disease prediction method as described in any one of claims 1-7, characterized in that the system... include: The multi-source bioinformatics construction module is used to acquire multi-source biological data of LncRNA, miRNA and diseases, calculate functional / semantic similarity and Gaussian interaction contour kernel similarity, and construct LncRNA-disease association matrix, LncRNA-miRNA association matrix and disease-miRNA association matrix to form a multi-view heterogeneous biological map; The consensus-enhancing feature fusion module is used to perform consensus-enhancing iterative fusion of functional / semantic similarity and Gaussian interaction contour kernel similarity to generate a fused similarity network containing multi-scale topological information as the basic feature representation. The graph attention representation learning module is used to input multi-view heterogeneous biological graphs into the graph attention autoencoder network to obtain low-dimensional embedding representations of LncRNAs and disease nodes; The joint feature fusion module is used to fuse the low-dimensional embedding representation of LncRNA and disease nodes with the basic feature representation to construct the LncRNA-disease joint feature representation; The residual attention prediction module is used to perform a layer-by-layer nonlinear transformation of the joint feature representation of LncRNA and disease based on the residual structure of the multilayer perceptron. It combines the attention pooling mechanism to adaptively weight different feature dimensions and finally outputs the association probability between LncRNA and disease.

Images