A Disease Similarity Prediction Method and System Based on Multivariate Data Fusion

Abstract

This invention relates to a disease similarity prediction method and system based on multivariate data fusion. The method includes: collecting gene network data, miRNA network data, disease-gene association matrices, and disease-miRNA association matrices to form a unified heterogeneous biological network; extracting a multimodal feature set through graph convolutional networks and improved graph attention networks; calculating the independent weights of each feature in the multimodal feature set using a multilayer perceptron, and combining the association matrices to generate disease-specific gene-perspective embeddings and miRNA-perspective embeddings, which are used as model inputs to train a lightweight bilinear tower model; optimizing the model parameters through an alternating training strategy and a ReconBoost-inspired regularization mechanism to obtain an optimized lightweight bilinear tower model, and obtaining the similarity prediction results for target disease pairs. This effectively overcomes the shortcomings of existing technologies in multivariate data integration, sparse data processing, and prediction stability.

Description

Technical Field

[0001] This invention belongs to the field of biomedical data analysis technology, and in particular relates to a method and system for predicting disease similarity based on multivariate data fusion. Background Technology

[0002] Disease similarity prediction is a crucial technique in biomedical research. By analyzing molecular and phenotypic associations between diseases, it reveals underlying pathological mechanisms, providing key support for drug repositioning, disease classification optimization, and personalized treatment. Its core value lies in identifying diseases with similar molecular mechanisms or clinical manifestations, thus providing a basis for drug repositioning. Drug repositioning accelerates the drug development process by exploring new indications for already approved drugs. Compared to traditional new drug development, utilizing the safety and pharmacokinetic data of existing drugs can significantly shorten the development cycle and reduce R&D costs and risks. This strategy has shown great potential in the treatment research of cancer, neurodegenerative diseases, cardiovascular diseases, and infectious diseases, bringing significant commercial value and social benefits to the pharmaceutical industry and medical research. For example, by identifying the molecular similarities between diabetes and cardiovascular disease, researchers can repurpose already approved hypoglycemic drugs for cardiovascular disease treatment, significantly improving drug development efficiency.

[0003] In recent years, the rapid development of high-throughput sequencing technology and bioinformatics has enabled researchers to construct complex biological networks using multidimensional biological data (e.g., gene expression, miRNA regulatory networks, and disease-molecular association data) to assess disease similarity. Deep learning technology, particularly the rise of graph neural networks, has provided new technical means for disease similarity prediction. Graph neural networks, through their powerful graph structure data processing capabilities, efficiently model the complex relationships between diseases, genes, and miRNAs. For example, graph convolutional networks aggregate node neighborhood information through multi-layer convolutional operations to achieve high-dimensional topological feature expression of gene networks; improved graph attention networks dynamically capture local structural differences in miRNA networks by adaptively allocating node weights. These methods have shown significant advantages in mining disease-molecular associations and have been widely applied in disease similarity prediction and drug repositioning research.

[0004] However, existing disease similarity prediction methods still face many challenges in practical applications. Multivariate biological data, such as gene networks and miRNA networks, are characterized by high dimensionality, heterogeneity, and sparsity. Existing methods struggle to effectively integrate the complementary information from both. For example, gene network data is massive but contains much redundant information, while miRNA network data is sparse and lacks complete associations, leading to low feature extraction efficiency. Furthermore, existing methods typically lack adaptive mechanisms to screen key biological features, making it difficult to highlight the genes or miRNAs most important for disease similarity prediction. Traditional methods may treat all genes equally, ignoring the significant contributions of key regulatory genes, thus reducing prediction accuracy. Simultaneously, complex graph neural network-based models are computationally expensive and time-consuming when processing multivariate data, making them difficult to apply in resource-constrained clinical or research settings. Moreover, when integrating gene and miRNA data, competition or inconsistency between modalities can lead to fluctuating prediction results. For instance, the richness of gene data may dominate predictions, while the sparsity of miRNA data is ignored, affecting the balance between modalities and the robustness of the model. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for predicting disease similarity based on multivariate data fusion.

[0006] The technical solution adopted in this invention is:

[0007] Firstly, a disease similarity prediction method based on multivariate data fusion is provided, including:

[0008] Collect gene network data, miRNA network data, disease-gene association matrix, and disease-miRNA association matrix to form a unified heterogeneous biological network;

[0009] Gene embedding features and miRNA embedding features were extracted from heterogeneous biological networks using graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set;

[0010] The independent weights of each feature in the multimodal feature set are calculated using a multilayer perceptron. The disease-gene association matrix and the disease-miRNA association matrix are combined to generate disease-gene-perspective embeddings and miRNA-perspective embeddings.

[0011] A lightweight bilinear tower model was obtained by training the model using gene-perspective embeddings and miRNA-perspective embeddings as model inputs.

[0012] The model parameters of the lightweight bilinear tower model are optimized by using an alternating training strategy and a regularization mechanism inspired by ReconBoost, resulting in an optimized lightweight bilinear tower model.

[0013] By optimizing a lightweight bilinear tower model, the similarity prediction results of the target disease pairs are obtained.

[0014] Furthermore, gene networks Contains a set of gene nodes and edge sets based on gene co-expression or functional association The graph structure of the gene network is as follows: ;

[0015] miRNA network Includes miRNA node set and edge sets based on miRNA similarity or regulatory relationships The graph structure of the miRNA network is as follows: .

[0016] Furthermore, gene network data, miRNA network data, disease-gene association matrices, and disease-miRNA association matrices are collected to form a unified heterogeneous biological network, including:

[0017] Collecting gene networks Gene network data and miRNA network miRNA network data;

[0018] Obtaining the disease-gene association matrix Disease-miRNA Association Matrix Disease-gene association matrix Disease-miRNA Association Matrix These include the associations between known diseases and genes and miRNAs, respectively.

[0019] Through disease-gene association matrix Disease-miRNA Association Matrix Gene networks and miRNA network Gene nodes and miRNA nodes are mapped to disease nodes, forming a unified heterogeneous biological network.

[0020] Furthermore, gene embedding features and miRNA embedding features are extracted from heterogeneous biological networks using graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set, including:

[0021] Gene networks of heterogeneous biological networks based on graph convolutional networks (GCNs) Feature extraction is performed to obtain gene embedding features. ;

[0022] miRNA networks in heterogeneous biological networks based on the improved graph attention network GATv2 Feature extraction was performed to obtain miRNA embedding features. ;

[0023] Collective gene embedding features and miRNA embedding features Generate a multimodal feature set.

[0024] Furthermore, the independent weights of each feature in the multimodal feature set are calculated using a multilayer perceptron, and combined with the disease-gene association matrix and the disease-miRNA association matrix to generate disease-gene-perspective embeddings and miRNA-perspective embeddings, including:

[0025] Two multilayer perceptrons are applied to each feature in the multimodal feature set to calculate independent weights;

[0026] Based on independent weights and disease-gene association matrices and disease-miRNA association matrices, the gene-perspective embedding of the disease is calculated by substituting them into the disease characterization calculation formula. Embedding from the perspective of miRNA The formula for calculating disease characteristics is:

[0027] ;

[0028] in, Embedding from a gene perspective Or miRNA perspective embedding h represents g or m, and A represents the disease-gene association matrix or the disease-miRNA association matrix. Indicates independent weights.

[0029] Furthermore, by using gene-perspective embeddings and miRNA-perspective embeddings as model inputs, a lightweight bilinear tower model is obtained through training, including:

[0030] Embedding the genetic perspective Embedding from the miRNA perspective As input to the model;

[0031] Disease pairs for known disease i and known disease j Embedded from a genetic perspective Embedding from the miRNA perspective Characterization was extracted from the data to construct gene feature vectors and miRNA feature vectors;

[0032] The expression for the gene feature vector is:

[0033] ;

[0034] The expression for the miRNA feature vector is:

[0035] ;

[0036] After dimensionality reduction via linear projection, the gene tower score and miRNA tower score are calculated using the bilinear transformation formula, which is:

[0037] ;

[0038] The weight matrix is ​​a learnable bilinear weight matrix. Gene Tower Score or miRNA tower score ;

[0039] Calculate the gene tower score. and miRNA tower score The average of these values ​​is used to obtain the final similarity score.

[0040] A lightweight bilinear tower model is trained based on the model input and the final similarity score.

[0041] Furthermore, the model parameters of the lightweight bilinear tower model are optimized through an alternating training strategy and a ReconBoost-inspired regularization mechanism, resulting in an optimized lightweight bilinear tower model, including:

[0042] Using an alternating training strategy, in each round of training, the encoder parameters and bilinear tower parameters of one modality (gene modality and miRNA modality) are fixed, while the encoder parameters and bilinear tower parameters of the other modality are updated. The dynamic loss function is defined as:

[0043] ;

[0044] in, For cross-entropy loss, For current mode prediction, The relative entropy KL divergence, For the weighting factor;

[0045] Based on a regularization mechanism inspired by ReconBoost, all encoder parameters and bilinear tower parameters are jointly optimized every preset fixed number of rounds. The global loss is defined as:

[0046] ;

[0047] For weighted average prediction, The learning rate;

[0048] An optimized lightweight bilinear tower model is obtained based on dynamic loss and global loss.

[0049] Secondly, a disease similarity prediction system based on multivariate data fusion is provided, including:

[0050] The multivariate data fusion module is used to collect gene network data, miRNA network data, disease-gene association matrix, and disease-miRNA association matrix to form a unified heterogeneous biological network.

[0051] The feature extraction module is used to extract gene embedding features and miRNA embedding features from heterogeneous biological networks through graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set;

[0052] The dynamic weighted embedding module is used to calculate the independent weights of each feature in the multimodal feature set using a multilayer perceptron, and combine the disease-gene association matrix and the disease-miRNA association matrix to generate disease-gene-perspective embedding and miRNA-perspective embedding.

[0053] A lightweight bilinear tower prediction module is used to train a lightweight bilinear tower model by using gene-perspective embeddings and miRNA-perspective embeddings as model inputs.

[0054] The alternating training and optimization module is used to optimize the model parameters of the lightweight bilinear tower model through an alternating training strategy and a regularization mechanism inspired by ReconBoost, resulting in an optimized lightweight bilinear tower model.

[0055] The disease similarity prediction module is used to predict the similarity of target disease pairs by optimizing a lightweight bilinear tower model, and obtain the similarity prediction results of the target disease pairs.

[0056] The beneficial effects achieved by this invention are as follows:

[0057] Gene network data, miRNA network data, disease-gene association matrices, and disease-miRNA association matrices were collected to form a unified heterogeneous biological network. Gene embedding features and miRNA embedding features were extracted from the heterogeneous biological network using graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set. The independent weights of each feature in the multimodal feature set were calculated using a multilayer perceptron, and combined with the disease-gene association matrix and the disease-miRNA association matrix to generate disease-specific gene embeddings and miRNA-specific embeddings. The gene-specific embeddings and miRNA-specific embeddings were used as model inputs to train a lightweight bilinear tower model. The model parameters of the lightweight bilinear tower model were optimized using an alternating training strategy and a regularization mechanism inspired by ReconBoost to obtain an optimized lightweight bilinear tower model. The optimized lightweight bilinear tower model was used to predict the similarity of target disease pairs to obtain the similarity prediction results of the target disease pairs. By deeply integrating gene networks, miRNA networks, and disease-related data, and combining graph neural networks, dynamic weight embedding, lightweight bilinear tower prediction, and alternating training optimization strategies, high-precision and low-computational-cost disease similarity prediction is achieved. This effectively compensates for the shortcomings of existing technologies in multivariate data integration, sparse data processing, and prediction stability, providing strong support for pathological mechanism research, disease classification optimization, and drug repositioning. Attached Figure Description

[0058] Figure 1 This is a flowchart of the disease similarity prediction method based on multivariate data fusion according to the present invention;

[0059] Figure 2 This is a structural diagram of the disease similarity prediction system based on multivariate data fusion according to the present invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0061] like Figure 1 As shown, this embodiment of the invention provides a disease similarity prediction method based on multivariate data fusion, including:

[0062] 101. Collect gene network data, miRNA network data, disease-gene association matrix, and disease-miRNA association matrix to form a unified heterogeneous biological network;

[0063] Gene Network Contains a set of gene nodes and edge sets based on gene co-expression or functional association The graph structure of the gene network is as follows: ;

[0064] miRNA network Includes miRNA node set and edge sets based on miRNA similarity or regulatory relationships The graph structure of the miRNA network is as follows: ;

[0065] Collecting gene networks Gene network data and miRNA network miRNA network data;

[0066] Obtaining the disease-gene association matrix Disease-miRNA Association Matrix In a matrix, an element is 1 if an association exists, and 0 otherwise.

[0067] Disease-Gene Association Matrix Disease-miRNA Association Matrix These include the associations between known diseases and genes and miRNAs, respectively.

[0068] For example, the gene network data contains 17,247 gene nodes and 371,501 edges, sourced from the HumanNet database; the miRNA network data contains 4,798 miRNA nodes and 126,264 edges, sourced from the RNADisease database; and the disease association data contains 1,030,082 disease-gene associations and 355,196 disease-miRNA associations, sourced from the DisGeNET and RNADisease databases. A pre-set training set contains 6,182 disease pairs (3,091 of which are positive samples), and a test set contains 1,330 disease pairs (665 of which are positive samples). and miRNA network It is stored in the form of a sparse adjacency matrix, and the edge weights are normalized based on the biological association strength (such as co-expression coefficient or regulatory strength); the association matrix is ​​binarized to reduce computational complexity;

[0069] Through disease-gene association matrix Disease-miRNA Association Matrix Gene networks and miRNA network Gene nodes and miRNA nodes are mapped to disease nodes, forming a unified heterogeneous biological network.

[0070] 102. Gene embedding features and miRNA embedding features were extracted from heterogeneous biological networks using graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set;

[0071] Gene networks based on graph convolutional networks (GCN) for heterogeneous biological networks Feature extraction is performed to obtain gene embedding features. The formula for updating gene network node features is:

[0072] ;

[0073] in, This is the feature matrix of the gene nodes in the l-th layer. For the normalized adjacency matrix, It is an adjacency matrix. For degree matrix, For learnable weight matrix, As the activation function, the first layer of GCN maps the initial features (based on gene expression levels) to a 128-dimensional hidden layer, and the second layer further aggregates neighborhood information to generate gene embedding features. (Dimension 128), Dropout rate 0.4, weight matrix randomly initialized, optimized during training via backpropagation;

[0074] miRNA networks in heterogeneous biological networks based on the improved graph attention network GATv2 Feature extraction was performed to obtain miRNA embedding features. Node features are updated through a multi-head attention mechanism, and the attention coefficient is calculated using the following formula:

[0075] ;

[0076] in, and For node features, This is the weight matrix. For learnable attention vectors, Indicating the splicing operation, GATv2 uses a two-head attention mechanism to generate miRNA embedding features. The number of attention heads is 2, the dropout rate is 0.4, the weight matrix and attention vector are randomly initialized, and optimization is performed through backpropagation during training.

[0077] Collective gene embedding features and miRNA embedding features Generate a multimodal feature set;

[0078] In this step, gene node features are based on expression levels (extracted from RNA sequencing data), and miRNA node features are based on regulatory relationships (extracted from RNA disease). The initial dimensions are determined based on the dataset (default 512 dimensions). Both GCN and GATv2 apply Dropout (default 0.4) and L2 regularization (coefficient 0.0001) to enhance the model's generalization ability. Sparse matrix operations are implemented using PyTorch Geometric to reduce memory usage and accelerate feature extraction.

[0079] 103. Utilize a multilayer perceptron to calculate the independent weights of each feature in the multimodal feature set, and combine the disease-gene association matrix and the disease-miRNA association matrix to generate disease-gene-perspective embeddings and miRNA-perspective embeddings.

[0080] For each feature in the multimodal feature set Independent weights were calculated by applying two layers of Multilayer Perceptron (MLP) respectively:

[0081] ;

[0082] The MLP consists of two linear transformations: the first layer reduces the dimensionality to 64 dimensions, and the second layer outputs a single weight value. The activation function is Sigmoid; the first layer of the MLP has a dimension of 128→64, and the second layer has a dimension of 64→1. The activation functions are ReLU (first layer) and Sigmoid (second layer), and the weights are randomly initialized.

[0083] Based on independent weights and disease-gene association matrix Disease-miRNA Association Matrix Substituting the data into the disease characterization calculation formula yields the gene-perspective embedding of the disease. Embedding from the miRNA perspective The formula for calculating disease characteristics is:

[0084] ;

[0085] in, Embedding from a gene perspective Or miRNA perspective embedding h represents g or m, and A represents the disease-gene association matrix or the disease-miRNA association matrix. Independent weights are represented; the aggregation process is implemented through matrix multiplication to ensure computational efficiency. Weight normalization avoids numerical instability.

[0086] 104. A lightweight bilinear tower model was obtained by using gene-perspective embeddings and miRNA-perspective embeddings as model inputs for training.

[0087] Embedding the genetic perspective Embedding from the miRNA perspective As input to the model;

[0088] Disease pairs for known disease i and known disease j Embedded from a genetic perspective Embedding from the miRNA perspective Characterization was extracted from the data to construct gene feature vectors and miRNA feature vectors;

[0089] The expression for the gene feature vector is:

[0090] ;

[0091] The expression for the miRNA feature vector is:

[0092] ;

[0093] The feature vectors include concatenation, interpolation, and element-wise multiplication, with a dimension of 128×4=512, ensuring the capture of multiple interaction patterns of disease pairs;

[0094] Dimensionality was reduced to a hidden dimension of 64 through linear projection. The gene tower score and miRNA tower score were then calculated using the bilinear transformation formula:

[0095] ;

[0096] The weight matrix is ​​a learnable bilinear weight matrix. Gene Tower Score or miRNA tower score Linear projection layer dimension increases from 512 to 64, bilinear matrix The dimensions are 64×64, randomly initialized, and optimized during training via backpropagation.

[0097] Calculate the gene tower score. and miRNA tower score The average of these values ​​is used to obtain the final similarity score. Average fusion ensures a balance between the gene perspective and the miRNA perspective;

[0098] Based on the model input and the final similarity score, a lightweight bilinear tower model is trained. The linear projection and bilinear transformation in this step are implemented through matrix operations, with a single prediction time of approximately 0.05 seconds, making it suitable for real-time applications. The lightweight design (approximately 100,000 parameters) reduces the computational cost by 50% compared to traditional fully connected networks, making it suitable for resource-constrained scenarios. Overfitting is avoided through Dropout (rate 0.3) and weight regularization (coefficient 0.0001).

[0099] 105. The model parameters of the lightweight bilinear tower model are optimized by using an alternating training strategy and a regularization mechanism inspired by ReconBoost, resulting in an optimized lightweight bilinear tower model.

[0100] The process of the alternating training strategy is as follows:

[0101] In each training round, the encoder parameters and bilinear tower parameters of one modality (gene modality and miRNA modality) are fixed, while the encoder parameters and bilinear tower parameters of the other modality are updated. The dynamic loss function is defined as:

[0102] ;

[0103] in, For cross-entropy loss, For current mode prediction, The relative entropy KL divergence, The weighting factor is 0.1 (default). In each iteration, the gene modality (GCN+MLP+gene tower) and the miRNA modality (GATv2+MLP+miRNA tower) are updated alternately, and each modality is updated 10 times in sub-iterations.

[0104] The global calibration process is as follows:

[0105] Based on a regularization mechanism inspired by ReconBoost, all encoder parameters and bilinear tower parameters are jointly optimized every preset fixed number of rounds (specifically every 3 rounds). The global loss is defined as:

[0106] ;

[0107] For weighted average prediction, The learning rate is used; global correction is performed via the SGD optimizer (learning rate 0.005) to ensure modal consistency.

[0108] An optimized lightweight bilinear tower model is obtained based on dynamic loss and global loss.

[0109] In this step, RMSprop (learning rate 0.001) is used for gene modalities, RMSprop (learning rate 0.002) is used for miRNA modalities, and SGD (learning rate 0.005) is used for global optimization; a total of 60 iterations are performed, with a batch size of 128, and the training time is approximately 15 seconds (in practice, based on an NVIDIA RTX 4090 GPU); ReconBoost uses KL divergence to constrain the predicted distribution between modalities, balancing the loss and regularization terms (λ=0.1) to prevent overfitting. Alternating training reduces modality competition.

[0110] 106. By optimizing the lightweight bilinear tower model, the similarity prediction results of the target disease pairs are obtained.

[0111] After obtaining the optimized lightweight bilinear tower model through the above training, disease similarity prediction can be performed on the target disease pairs to obtain the similarity prediction results of the target disease pairs.

[0112] The beneficial effects achieved by the embodiments of the present invention are as follows:

[0113] Gene network data, miRNA network data, disease-gene association matrices, and disease-miRNA association matrices were collected to form a unified heterogeneous biological network. Gene embedding features and miRNA embedding features were extracted from the heterogeneous biological network using graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set. The independent weights of each feature in the multimodal feature set were calculated using a multilayer perceptron, and combined with the disease-gene association matrix and the disease-miRNA association matrix to generate disease-specific gene embeddings and miRNA-specific embeddings. The gene-specific embeddings and miRNA-specific embeddings were used as model inputs to train a lightweight bilinear tower model. The model parameters of the lightweight bilinear tower model were optimized using an alternating training strategy and a regularization mechanism inspired by ReconBoost to obtain an optimized lightweight bilinear tower model. The optimized lightweight bilinear tower model was used to predict the similarity of target disease pairs to obtain the similarity prediction results of the target disease pairs. By deeply integrating gene networks, miRNA networks, and disease-related data, and combining graph neural networks, dynamic weight embedding, lightweight bilinear tower prediction, and alternating training optimization strategies, high-precision and low-computational-cost disease similarity prediction is achieved. This effectively compensates for the shortcomings of existing technologies in multivariate data integration, sparse data processing, and prediction stability, providing strong support for pathological mechanism research, disease classification optimization, and drug repositioning.

[0114] Based on the disease similarity prediction method based on multivariate data fusion described in the above embodiments, the following embodiments illustrate a disease similarity prediction system based on multivariate data fusion.

[0115] like Figure 2 As shown, this embodiment of the invention provides a disease similarity prediction system based on multivariate data fusion, comprising:

[0116] The multivariate data fusion module 201 is used to collect gene network data, miRNA network data, disease-gene association matrix and disease-miRNA association matrix to form a unified heterogeneous biological network.

[0117] The feature extraction module 202 is used to extract gene embedding features and miRNA embedding features from heterogeneous biological networks through graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set;

[0118] The dynamic weight embedding module 203 is used to calculate the independent weights of each feature in the multimodal feature set using a multilayer perceptron, and combine the disease-gene association matrix and the disease-miRNA association matrix to generate disease-gene-perspective embedding and miRNA-perspective embedding.

[0119] The lightweight bilinear tower prediction module 204 is used to train a lightweight bilinear tower model by using gene-perspective embeddings and miRNA-perspective embeddings as model inputs.

[0120] Alternating training and optimization module 205 is used to optimize the model parameters of the lightweight bilinear tower model through an alternating training strategy and a regularization mechanism inspired by ReconBoost, so as to obtain an optimized lightweight bilinear tower model.

[0121] The disease similarity prediction module 206 is used to predict the similarity of target disease pairs by optimizing a lightweight bilinear tower model, and obtain the similarity prediction results of the target disease pairs.

[0122] The beneficial effects achieved by the embodiments of the present invention are as follows:

[0123] The multivariate data fusion module 201 collects gene network data, miRNA network data, disease-gene association matrices, and disease-miRNA association matrices to form a unified heterogeneous biological network. The feature extraction module 202 extracts gene embedding features and miRNA embedding features from the heterogeneous biological network using graph convolutional networks and improved graph attention networks, respectively, to generate a multimodal feature set. The dynamic weight embedding module 203 calculates the independent weights of each feature in the multimodal feature set using a multilayer perceptron, and combines the disease-gene association matrix and the disease-miRNA association matrix to generate the disease-specific data. The system includes gene-perspective embedding and miRNA-perspective embedding; a lightweight bilinear tower prediction module 204, used to train a lightweight bilinear tower model by using gene-perspective embedding and miRNA-perspective embedding as model inputs; an alternating training and optimization module 205, used to optimize the model parameters of the lightweight bilinear tower model through an alternating training strategy and a ReconBoost-inspired regularization mechanism, resulting in an optimized lightweight bilinear tower model; and a disease similarity prediction module 206, used to predict the similarity of target disease pairs by optimizing the lightweight bilinear tower model, obtaining the similarity prediction results for the target disease pairs. By deeply integrating gene networks, miRNA networks, and disease-related data, and combining graph neural networks, dynamic weight embedding, lightweight bilinear tower prediction, and alternating training optimization strategies, high-precision and low-computational-cost disease similarity prediction is achieved. This effectively compensates for the shortcomings of existing technologies in multivariate data integration, sparse data processing, and prediction stability, providing strong support for pathological mechanism research, disease classification optimization, and drug repositioning.

[0124] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A disease similarity prediction method based on multi-element data fusion, characterized in that, The method comprises the following steps: Collecting gene network data, miRNA network data, disease-gene association matrix and disease-miRNA association matrix to form a unified heterogeneous biological network; Extracting gene embedding features and miRNA embedding features from the heterogeneous biological network through a graph convolution network and an improved graph attention network to generate a multi-modal feature set; Calculating independent weights of each feature in the multi-modal feature set using a multi-layer perceptron, combining the disease-gene association matrix and the disease-miRNA association matrix to generate gene perspective embedding and miRNA perspective embedding of the disease; Training a lightweight bilinear tower model by taking the gene perspective embedding and the miRNA perspective embedding as model inputs; Optimizing model parameters of the lightweight bilinear tower model through an alternating training strategy and a regularization mechanism inspired by ReconBoost to obtain an optimized lightweight bilinear tower model; Performing disease similarity prediction on a target disease pair through the optimized lightweight bilinear tower model to obtain a similarity prediction result of the target disease pair.

2. The disease similarity prediction method based on multi-element data fusion according to claim 1, wherein Gene network comprising a set of gene nodes and a set of edges based on co-expression or functional association of genes ; the graph structure of the gene network is ; miRNA network comprising a set of miRNA nodes and a set of edges based on miRNA similarity or regulatory relationship ; the graph structure of the miRNA network is . 3.The disease similarity prediction method based on multi-element data fusion according to claim 2, characterized in that, The collecting gene network data, miRNA network data, disease-gene association matrix and disease-miRNA association matrix to form a unified heterogeneous biological network comprises: collecting gene network data of the gene network and miRNA network data of the miRNA network collecting gene network data of the gene network and miRNA network data of the miRNA network obtaining a disease-gene association matrix and a disease-miRNA association matrix , the disease-gene association matrix and the disease-miRNA association matrix respectively comprising known associations of diseases with genes and miRNAs; mapping the gene nodes and miRNA nodes in the disease-gene association matrix and the disease-miRNA association matrix to disease nodes, forming a unified heterogeneous biological network. and the miRNA network ​ 4. The disease similarity prediction method based on multi-element data fusion according to claim 3, characterized in that, The extracting gene embedding features and miRNA embedding features from the heterogeneous biological network through a graph convolution network and an improved graph attention network to generate a multi-modal feature set comprises: based on a graph convolutional network, GCN, on the gene network of the heterogeneous biological network perform feature extraction to obtain gene embedding features ; based on an improved graph attention network GATv2 on the miRNA network of the heterogeneous biological network feature extraction is performed to obtain miRNA embedding features ; collecting the gene embedding features and the miRNA embedding features to generate a multimodal feature set.

5. The disease similarity prediction method based on multi-element data fusion according to claim 1, characterized in that, The calculating independent weights of each feature in the multi-modal feature set using a multi-layer perceptron, combining the disease-gene association matrix and the disease-miRNA association matrix to generate gene perspective embedding and miRNA perspective embedding of the disease comprises: Applying two layers of multi-layer perceptron to each feature in the multi-modal feature set to calculate independent weights; based on the independent weight and the disease-gene association matrix and the disease-miRNA association matrix, a disease gene perspective embedding is calculated by substituting into a disease representation calculation formula and a miRNA perspective embedding , and the disease representation calculation formula is: ； wherein the is a gene perspective embedding or a miRNA perspective embedding , the h is g or m, the A represents the disease-gene association matrix or the disease-miRNA association matrix, the represents an independent weight.

6. The disease similarity prediction method based on multi-element data fusion according to claim 5, characterized in that, The training a lightweight bilinear tower model by taking the gene perspective embedding and the miRNA perspective embedding as model inputs comprises: embedding the gene perspective and the miRNA perspective as model input; Disease pair for known disease i and known disease j from the gene perspective embedding and the miRNA perspective embedding extracting the representation, and constructing a gene feature vector and a miRNA feature vector; The expression of the gene feature vector is: ； The expression of the miRNA feature vector is: ； After dimensionality reduction by linear projection, gene tower score values and miRNA tower score values are calculated using a bilinear transformation formula, wherein the bilinear transformation formula is: ； The is a learnable bilinear weight matrix, the is a gene tower score value or miRNA tower score value ; averaging the gene tower score value and the miRNA tower score value results in a final similarity score value; The training a lightweight bilinear tower model by taking the model input and the final similarity score value as inputs.

7. The disease similarity prediction method based on multi-element data fusion according to claim 1, characterized in that, The optimizing model parameters of the lightweight bilinear tower model through an alternating training strategy and a regularization mechanism inspired by ReconBoost to obtain an optimized lightweight bilinear tower model comprises: Through the alternating training strategy, the encoder parameters and the bilinear tower parameters of one of the gene modality and the miRNA modality are fixed in each round of training, and the encoder parameters and the bilinear tower parameters of the other modality are updated, and the dynamic loss function is defined as: ； wherein the is a cross-entropy loss, the is a current modal prediction, the is a relative entropy KL divergence, the is a weighting coefficient; Based on the regularization mechanism inspired by ReconBoost, all encoder parameters and bilinear tower parameters are jointly optimized every preset fixed number of rounds, and the global loss is defined as: ； The For weighted average prediction, the is a learning rate; An optimized lightweight bilinear tower model is obtained according to the dynamic loss and the global loss. 8.A disease similarity prediction system based on multi-element data fusion, characterized in that, Comprise: A multi-source data fusion module is configured to collect gene network data, miRNA network data, a disease-gene association matrix and a disease-miRNA association matrix, and form a unified heterogeneous biological network; A feature extraction module is configured to extract gene embedding features and miRNA embedding features from the heterogeneous biological network respectively by using a graph convolution network and an improved graph attention network, and generate a multi-modal feature set; A dynamic weight embedding module is configured to calculate independent weights of features in the multi-modal feature set by using a multi-layer perception machine, combine the disease-gene association matrix and the disease-miRNA association matrix, and generate gene perspective embedding and miRNA perspective embedding of a disease; A lightweight bilinear tower prediction module is configured to train a lightweight bilinear tower model by taking the gene perspective embedding and the miRNA perspective embedding as model inputs; An alternating training and optimization module is configured to optimize model parameters of the lightweight bilinear tower model by using an alternating training strategy and a regularization mechanism inspired by ReconBoost, and obtain an optimized lightweight bilinear tower model; A disease similarity prediction module is configured to perform disease similarity prediction on a target disease pair by using the optimized lightweight bilinear tower model, and obtain a similarity prediction result of the target disease pair.

Images