A method for efficiently predicting the multi-association of miRNA, lncRNA and diseases by using contrast hypergraph generation technology
By comparing hypergraph generation techniques, integrating heterogeneous networks and multi-task learning frameworks, a three-layer heterogeneous graph structure was constructed, robust representations were extracted, and the problem of high-order collaborative interactions between miRNA, lncRNA, and diseases was solved. This enabled efficient multivariate association prediction and improved the accuracy and robustness of disease association prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHIHEZI UNIVERSITY
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-10
AI Technical Summary
Existing methods are ineffective at capturing high-order synergistic interactions between miRNAs, lncRNAs, and diseases, and biomedical data suffer from small sample sizes and sparse labels, leading to inaccurate disease association predictions.
By employing contrastive hypergraph generation technology, a three-layer heterogeneous graph structure is constructed by integrating heterogeneous network construction, contrastive hypergraph generation, and a multi-task learning framework. Robust representations are extracted using topological view and attribute view encoders. Combined with multi-task learning and data augmentation modules, loss weights are dynamically adjusted and negative samples are generated to achieve multivariate association prediction.
It breaks through the binary relation limitations of traditional graph models, captures high-order synergistic regulatory patterns between molecules, improves the accuracy and robustness of disease association prediction, solves the noise interference problem of single data sources, and provides an efficient multivariate association prediction method.
Smart Images
Figure CN122369591A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of bioinformatics and artificial intelligence, and in particular to an efficient method for predicting the multivariate associations between miRNAs, lncRNAs and diseases using contrastive hypergraph generation technology. Background Technology
[0002] In the biomedical field, miRNAs and lncRNAs, as two key regulatory non-coding RNAs, are closely related to the occurrence and development of diseases (especially complex diseases such as cancer and neurodegenerative diseases) through various mechanisms (such as ceRNA competition, epigenetic silencing, and transcriptional regulation). Accurate prediction of the multivariate associations between miRNAs, lncRNAs, and diseases is of great significance for elucidating disease mechanisms, screening early diagnostic biomarkers, and discovering drug targets. Existing methods, based on graph models of random walks and CNN / LDA models of shallow neural networks, are mostly based on a single view (such as only the molecular-disease binary relationship) or shallow feature modeling, making it difficult to capture high-order synergistic interactions between miRNAs, lncRNAs, and diseases. Biomedical data, especially disease-related miRNA / lncRNA associations, generally suffer from small sample sizes and sparse labels. Therefore, this invention proposes an efficient method for predicting the multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology to address the problems existing in current technologies. Summary of the Invention
[0003] To address the aforementioned problems, the present invention aims to propose an efficient method for predicting the multivariate associations between miRNAs, lncRNAs, and diseases using contrast hypergraph generation technology. This method integrates heterogeneous network construction, contrast hypergraph generation, and a multi-task learning framework to achieve efficient prediction of the multivariate associations between miRNAs, lncRNAs, and diseases. Through a three-layer heterogeneous graph structure and a contrast hypergraph generation module, it overcomes the limitation of traditional graph models that can only model binary relationships, capturing higher-order association patterns of intermolecular synergistic regulation. Robust characterization is extracted through a dual-view encoder, solving the problem of noise interference from a single data source.
[0004] To achieve the objectives of this invention, the following technical solution is provided: a method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology, comprising a heterogeneous network construction module, a contrastive hypergraph generation module, a multivariate association prediction module, an interpretability module, and a data augmentation module. The heterogeneous network construction module integrates miRNA-lncRNA interactions, disease association networks, and multi-source gene expression data to construct a three-layer heterogeneous graph containing molecular nodes and disease nodes. The contrastive hypergraph generation module enhances representation capabilities through hypergraph structure construction, warning interaction, and contrastive learning. The multivariate association prediction module processes multi-source data uniformly, predicts different types of associations separately, and dynamically adjusts loss weights. The interpretability module performs attention weight clustering analysis on molecular nodes and disease nodes in the hypergraph. The data augmentation module generates synthetic negative samples from the topological features of high-risk nodes to adjust the training set distribution.
[0005] A further improvement is that the contrast hypergraph generation module includes a topology view encoder and an attribute view encoder. The topology view encoder extracts topological features based on a graph convolutional network, and the attribute view encoder is used to extract sequence features through word embeddings and convolutional neural networks.
[0006] A further improvement is made in that the calculation formula for the topology view encoder is: in, For nodes In the The hidden state of the layer For activation function, For nodes The neighborhood group, For the first Neighbor nodes in the layer For the target node Attention weights For the first The learnable weight matrix of the layer is used to linearly transform the features of neighboring nodes. For neighboring nodes exist The hidden state of the layer.
[0007] A further improvement is that the multivariate association prediction module adopts a multi-task learning framework, and the multivariate association prediction module adjusts the loss weight based on task complexity to balance the prediction bias of different association types.
[0008] Further improvements are made in that the multivariate association prediction module includes a shared coding unit and independent task units. The shared coding unit is used to uniformly process the hypergraph representation of miRNA, lncRNA, and disease. The independent task units are used to design independent decoders for miRNA-disease, lncRNA-disease, and lncRNA-miRNA association prediction, respectively, and balance the multi-task loss through dynamic weight allocation.
[0009] Further improvements are made in that the data augmentation module adopts progressive training, first pre-training the shared encoder unit and then fine-tuning the independent task units, and enhancing the robustness of the model through adversarial training.
[0010] Further improvements are made in that the prediction method includes the following steps: Step 1: Construct a heterogeneous network containing miRNAs, lncRNAs, and disease nodes to integrate multi-source heterogeneous data; Step 2: Extract node contrast features to generate an enhanced hypergraph representation; Step 3: Predict multivariate associations between miRNA and disease, lncRNA and disease, and lncRNA and miRNA based on a multi-task learning framework; Step 4: Analyze key association paths and output prediction results and confidence scores.
[0011] A further improvement is that the multi-source heterogeneous data sources in step one include miRNA-lncRNA interaction data, miRNA disease association data, lncRNA disease association data, and gene expression data.
[0012] A further improvement is made in the following: the confidence score formula in step four is: in, Score the confidence level. The raw logits values output by the model. The model predicts the probability of this association existing. The variance of the predicted logits values for all samples on the test set.
[0013] The beneficial effects of this invention are as follows: By integrating heterogeneous network construction, contrastive hypergraph generation, and multi-task learning framework, this invention achieves efficient prediction of multivariate associations between miRNA, lncRNA, and diseases. Through a three-layer heterogeneous graph structure and a contrastive hypergraph generation module, it breaks through the limitation of traditional graph models that can only model binary relationships, captures high-order association patterns of intermolecular synergistic regulation, and extracts robust characterizations through a dual-view encoder, thus solving the problem of noise interference from a single data source. Attached Figure Description
[0014] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a diagram of the architecture of the supergraph generation module of the present invention; Figure 3 This is a diagram of the multivariate correlation prediction module architecture of the present invention; Figure 4 This is a flowchart of the steps of the present invention. Detailed Implementation
[0015] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.
[0016] miRNAs and lncRNAs are two important classes of non-coding RNA molecules that play a crucial role in gene expression regulation. miRNAs are endogenous non-coding single-stranded RNAs approximately 20-25 nucleotides in length that regulate target gene expression through base pairing. Their high conservation across species underscores their functional importance. lncRNAs are non-coding RNAs exceeding 200 nt in length, possessing mRNA-like structures but lacking coding function. Currently, miRNA / lncRNA-based disease diagnosis has seen preliminary clinical application, while therapeutic applications remain in clinical trials, though they have already demonstrated the potential to break through traditional therapies. With advancements in delivery technologies and target research, they are expected to become an important component of precision medicine in the future.
[0017] Based on this, according to Figure 1 , Figure 2 , Figure 3 , Figure 4 As shown, this embodiment provides an efficient method for predicting multiple associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology. The method includes a heterogeneous network construction module, a contrastive hypergraph generation module, a multiple association prediction module, an interpretability module, and a data augmentation module. The heterogeneous network construction module integrates miRNA-lncRNA interactions, disease association networks, and multi-source gene expression data to construct a three-layer heterogeneous graph containing molecular nodes and disease nodes. This three-layer heterogeneous graph includes a molecular node layer, a hyperedge layer, and a disease node layer, containing topological structure and attribute features to provide high-quality input for subsequent modeling. Multimodal data fusion technology is introduced to integrate heterogeneous features. Through the three-layer structure of "molecular node-hyperedge-disease node," higher-order synergistic effects are captured, overcoming the limitation of traditional graph models that can only model binary relationships. Edge weights are dynamically adjusted based on node functional similarity to reflect changes in the strength of associations between nodes, enhancing the biological rationality of the graph structure.
[0018] The contrastive hypergraph generation module is used to enhance representation capabilities through hypergraph structure construction, interaction, and contrastive learning. By combining contrastive learning with the hypergraph structure, it enhances node representation capabilities, solves the generalization problem under low-label data, and provides high-quality features for multivariate association prediction. The contrastive hypergraph generation module includes a topology view encoder and an attribute view encoder. The topology view encoder extracts topological features based on graph convolutional networks, while the attribute view encoder extracts sequence features through word embeddings and convolutional neural networks.
[0019] The calculation formula for the topology view encoder is: in, For nodes In the The hidden state of the layer For activation function, For nodes The neighborhood group, For the first Neighbor nodes in the layer For the target node Attention weights For the first The learnable weight matrix of the layer is used to linearly transform the features of neighboring nodes. For neighboring nodes exist The hidden state of the layer.
[0020] The multivariate association prediction module is used to uniformly process multi-source data, predict different types of associations separately, dynamically adjust loss weights, and transform the node representations generated by the comparison hypergraph into specific association prediction results, supporting collaborative inference of multiple types of associations.
[0021] The multivariate association prediction module adopts a multi-task learning framework. The module adjusts the loss weights based on task complexity to balance the prediction bias of different association types.
[0022] The multivariate association prediction module includes a shared coding unit and independent task units. The shared coding unit is used to uniformly process the hypergraph representation of miRNA, lncRNA, and disease. The independent task units are used to design independent decoders for miRNA-disease, lncRNA-disease, and lncRNA-miRNA association prediction, respectively, and the multi-task loss is balanced through dynamic weight allocation.
[0023] The interpretability module is used for attention weight clustering analysis of molecular nodes and disease nodes in the hypergraph. By enhancing the transparency of model decision-making, it reveals the biological basis of prediction results, assists in mechanism research and clinical validation, and displays high-frequency associated pathways through attention weight clustering and heatmaps, verifying their consistency with known biological pathways.
[0024] The data augmentation module is used to generate synthetic negative samples from the topological features of high-risk nodes to adjust the distribution of the training set. The data augmentation module adopts progressive training, first pre-training the shared encoder unit and then fine-tuning the independent task units. It enhances the robustness of the model through adversarial training, effectively improving the model's robustness and generalization ability to complex biological data, and alleviating the problems of data sparsity, noise and class imbalance.
[0025] The prediction method includes the following steps: Step 1: Construct a heterogeneous network containing miRNAs, lncRNAs, and disease nodes to integrate multi-source heterogeneous data; Step 2: Extract node contrast features to generate an enhanced hypergraph representation; Step 3: Predict multivariate associations between miRNA and disease, lncRNA and disease, and lncRNA and miRNA based on a multi-task learning framework; Step 4: Analyze key association paths and output prediction results and confidence scores.
[0026] The multi-source heterogeneous data sources in step one include miRNA-lncRNA interaction data, miRNA disease association data, lncRNA disease association data, and gene expression data.
[0027] The confidence score formula in step four is: in, Score the confidence level. The raw logits values output by the model. The model predicts the probability of this association existing. The variance of the predicted logits values for all samples on the test set.
[0028] Each module works collaboratively through the process of "data integration → feature enhancement → task prediction → interpretation and verification". The heterogeneous network provides structural input for the contrast hypergraph, which supports multivariate association prediction through feature enhancement. The interpretability and data enhancement modules respectively improve the credibility and robustness of the model.
[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for efficiently predicting the multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology, characterized in that: The system includes a heterogeneous network construction module, a contrastive hypergraph generation module, a multivariate association prediction module, an interpretability module, and a data augmentation module. The heterogeneous network construction module integrates multi-source data on miRNA-lncRNA interactions, disease association networks, and gene expression to construct a three-layer heterogeneous graph containing molecular nodes and disease nodes. The contrastive hypergraph generation module enhances representation capabilities through hypergraph structure construction, interaction, and contrastive learning. The multivariate association prediction module processes multi-source data uniformly, predicts different types of associations separately, and dynamically adjusts loss weights. The interpretability module performs attention weight clustering analysis on molecular nodes and disease nodes in the hypergraph. The data augmentation module generates synthetic negative samples from the topological features of high-risk nodes to adjust the training set distribution.
2. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 1, characterized in that: The contrast hypergraph generation module includes a topology view encoder and an attribute view encoder. The topology view encoder extracts topological features based on a graph convolutional network, and the attribute view encoder is used to extract sequence features through word embeddings and a convolutional neural network.
3. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 2, characterized in that: The calculation formula for the topology view encoder is: in, For nodes In the The hidden state of the layer For activation function, For nodes The neighborhood group, For the first Neighbor nodes in the layer For the target node Attention weights For the first The learnable weight matrix of the layer is used to linearly transform the features of neighboring nodes. Neighboring nodes exist The hidden state of the layer.
4. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 1, characterized in that: The multivariate association prediction module adopts a multi-task learning framework, and adjusts the loss weights based on task complexity to balance the prediction bias of different association types.
5. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 1, characterized in that: The multivariate association prediction module includes a shared coding unit and independent task units. The shared coding unit is used to uniformly process the hypergraph representation of miRNA, lncRNA, and disease. The independent task units are used to design independent decoders for miRNA-disease, lncRNA-disease, and lncRNA-miRNA association prediction, respectively, and balance the multi-task loss through dynamic weight allocation.
6. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 1, characterized in that: The data augmentation module employs progressive training, first pre-training the shared encoder unit and then fine-tuning the independent task units, and enhancing the model's robustness through adversarial training.
7. A method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology as described in claim 1, comprising the following steps: Step 1: Construct a heterogeneous network containing miRNAs, lncRNAs, and disease nodes to integrate multi-source heterogeneous data; Step 2: Extract node contrast features to generate an enhanced hypergraph representation; Step 3: Predict multivariate associations between miRNA and disease, lncRNA and disease, and lncRNA and miRNA based on a multi-task learning framework; Step 4: Analyze key association paths and output prediction results and confidence scores.
8. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 7, characterized in that: The multi-source heterogeneous data sources in step one include miRNA-lncRNA interaction data, miRNA disease association data, lncRNA disease association data, and gene expression data.
9. The method for efficient prediction of multivariate associations between miRNAs, lncRNAs, and diseases using contrastive hypergraph generation technology according to claim 7, characterized in that: The confidence score formula in step four is as follows: in, Score the confidence level. The raw logits values output by the model. The model predicts the probability of this association existing. The variance of the predicted logits values for all samples on the test set.