A non-coding RNA and disease association prediction method based on hypergraph contrastive learning
By constructing hypergraphs and heterogeneous graphs and introducing self-supervised contrastive learning, the problem of insufficient high-order relationship modeling in the prediction of non-coding RNA association with diseases in existing technologies is solved, and more efficient multi-task association prediction performance and stability are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-15
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to effectively model high-order complex relationships when predicting the association between non-coding RNA and diseases, and their feature learning and generalization capabilities are insufficient in multi-source heterogeneous data environments, resulting in poor model performance in practical applications.
We employ a hypergraph-based contrastive learning approach. By constructing a hypergraph and a heterogeneous graph, and combining a hypergraph attention encoder and a graph attention encoder, we extract high-order relation features. We also introduce a self-supervised contrastive learning module to enhance the discriminative ability and robustness of node embeddings.
It overcomes the limitations of single-task approaches, effectively utilizes the complex relationship between known non-coding RNAs and diseases, improves the performance and stability of multi-task association prediction, and enhances robustness and generalization performance under limited labeled data.
Smart Images
Figure CN122290689A_ABST
Abstract
Description
Technical Field
[0001] This invention pertains to biomolecular association prediction methods in the field of bioinformatics, specifically involving a method for predicting the association between non-coding RNA and diseases based on hypergraph contrastive learning. Background Technology
[0002] Non-coding RNAs (ncRNAs) are a class of RNA molecules that do not participate in protein translation but play important regulatory roles in biological activities. Based on length, ncRNAs are generally divided into two main categories: long non-coding RNAs (lncRNAs) and small non-coding RNAs. lncRNAs are generally longer than 200 nucleotides, while small non-coding RNAs are shorter than 200 nucleotides. Among small non-coding RNAs, microRNAs (miRNAs) are the most typical, typically less than 20 nucleotides in length. Numerous studies have shown that lncRNAs and miRNAs play crucial roles in the occurrence and development of various diseases (such as cancer, cardiovascular diseases, and neurological disorders) and interact within complex biological regulatory networks. However, relying on traditional biological experiments to discover novel associations between lncRNAs, miRNAs, and diseases is both time-consuming and costly. Therefore, developing efficient computational methods to assist experimental research and predict potential associations has become an important research direction.
[0003] In recent years, the rapid development of biotechnology has driven the construction of various computational models aimed at discovering potential lncRNA-disease associations (LDA) and miRNA-disease associations (MDA). Currently, mainstream LDA and MDA prediction models can be categorized into three types: network-based methods, matrix factorization-based methods, and machine learning-based methods. However, these methods are mostly optimized for single tasks and lack cross-task transfer capabilities; furthermore, they struggle to effectively model high-order complex relationships between lncRNAs, miRNAs, and diseases. Moreover, due to the sparsity of known association data, these models still face challenges in feature learning and generalization, limiting their performance in practical applications. Therefore, how to fully mine high-order structural information and alleviate data sparsity in multi-source heterogeneous data environments has become a critical problem that urgently needs to be solved. Summary of the Invention
[0004] To address the aforementioned challenges, this invention proposes a method for predicting the association between non-coding RNAs and diseases based on hypergraph contrastive learning. This method can effectively utilize the complex relationships between known lncRNAs, miRNAs, and diseases to infer potential LDA and MDA.
[0005] To address the shortcomings of existing technologies, this invention models biomolecular association networks from two levels: high-order cooperative relationships and pairwise dependencies, by constructing two types of structural views: hypergraphs and heterogeneous graphs. By combining a hypergraph attention encoder and a graph attention encoder, effective extraction of structural information from both types is achieved, and a hybrid attention mechanism promotes information interaction between different structural spaces. Simultaneously, a self-supervised contrastive learning module is introduced to improve the discriminative power and robustness of node embeddings by aligning node representations under different structural views, thereby further enhancing the overall performance and stability of multi-task association prediction. Finally, extensive experiments on two publicly available datasets demonstrate that the prediction performance of this invention outperforms existing methods.
[0006] This invention discloses a method for predicting the association between non-coding RNA and diseases based on hypergraph contrastive learning, the method comprising:
[0007] Step 1: Based on domain knowledge and known associations between lncRNAs, miRNAs, and diseases, construct hypergraphs and heterographs;
[0008] First, by combining Gaussian interaction spectrum kernel similarity and functional similarity information, the fusion similarity matrices of lncRNA, miRNA and disease are calculated respectively. Then, based on the fusion similarity matrices, the Top-k nearest neighbor strategy is used to construct hyperedges and similar edges. On this basis, the hypergraph structure of lncRNA, miRNA and disease is generated according to the constructed hyperedges, and the corresponding heterogeneous graph structure is constructed by combining known association relationships and similar edges.
[0009] Step 2: Use the hypergraph attention encoder to learn the representation of the hypergraph and extract high-order relation features;
[0010] A dedicated hypergraph attention layer is introduced to encode higher-order structural information; this attention layer establishes collaborative relationships between multiple nodes during the node-hyperedge-node propagation process, obtaining node representations based on higher-order structures: ,in, , and These represent the embedding representations of lncRNA, disease, and miRNA nodes in the hypergraph structure, respectively.
[0011] Step 3: Use a graph attention encoder to learn representations of heterogeneous graphs and extract pairwise relation features;
[0012] A dedicated graph attention layer is introduced to model the pairwise dependencies between nodes; through graph attention encoding, a node embedding representation based on heterogeneous relationship modeling is obtained: ,in, , and These represent the embedding representations of lncRNA, disease, and miRNA nodes in the heterogeneous graph space, respectively.
[0013] Step 4: Fuse hypergraph and heterogeneous graph information using a hybrid attention encoder to obtain a unified node representation;
[0014] By using a hybrid attention encoder, common feature embeddings that fuse the two types of structural information are obtained. Correspondingly, the common embedding in the hypergraph space is represented as: ,in, , and The embeddings of lncRNA, disease, and miRNA in the hypergraph space are represented, respectively; the common embeddings in the heterogeneous graph space are represented as follows: , , and This represents the embedding representation of the corresponding node in the heterogeneous graph space;
[0015] Finally, the representations from the two spaces are fused to obtain a shared embedding: ;
[0016] Step 5: Adaptively fuse node representations under different structural views to further enhance feature representation capabilities;
[0017] Through a view-level attention mechanism, the above multi-view representations are fused into a final node embedding representation. For a node... Let its corresponding multi-view embedding representation be... ,in Includes two specific view embeddings , And a shared embedding ;
[0018] Final node representation Defined as a weighted sum of the embeddings of each view: ,in, This represents the view-level attention coefficient, used to measure the importance of different views to the final node representation;
[0019] Step 6: Introduce self-supervised contrastive learning to optimize node representations, and complete the prediction of non-coding RNA associations with diseases based on fusion representations.
[0020] Furthermore, the encoding method of the hypergraph attention encoder in step 2 is as follows:
[0021] During the stage of updating the hyperedge representation at the node, the contribution levels of different nodes in the same hyperedge are distinguished; nodes are defined. For superedge Attention weights for: ;
[0022] in, Indicates the superedge The set of nodes contained therein For a learnable parameter matrix, For trainable weight vectors, This indicates an attention mechanism using the scaled dot product form. Represents a node;
[0023] Introducing an attention mechanism from hyperedges to nodes, nodes Its associated superedge Attention score for: ;in, Indicates the presence of nodes The set of superedges and is a learnable parameter matrix.
[0024] Furthermore, the attention weights of the attention encoder in step 3... The calculation method is as follows:
[0025] node Its neighboring nodes Attention weights for:
[0026] ;
[0027] in, Represents a node The neighborhood group, This represents a vector concatenation operation. The weight matrix is a learnable matrix; , , Represents a node. This represents the activation function.
[0028] Furthermore, the specific method of the hybrid attention encoder in step 4 is as follows:
[0029] Step 4.1: Let the pairwise attention matrix learned by the graph attention encoding module be... , of which elements Represents a node With nodes The strength of pairwise dependence between them The number of nodes is represented; the hypergraph attention encoding module obtains the attention matrix from each node to the hyperedge. And the attention matrix from the hyperedge to the node ,in Measure node For superedge Important contributions Then characterize the super-edge For nodes Indicates the degree of impact of the update. Indicates the number of edges;
[0030] Step 4.2: To map the implicit higher-order collaborative relationships in the hypergraph structure back to the node pairs, the joint participation patterns of nodes in the hypergraph are modeled to construct the node association strength based on the higher-order structure, which is then used to enhance the original graph attention matrix; the updated graph attention matrix... for: Among them, the symbol " " indicates an element-wise addition operation;
[0031] Step 4.3: To feed back pairwise relationship information to the hypergraph attention propagation process, a high-order relationship attention matrix is constructed by aggregating pairwise attention scores between nodes within the same hyperedge. Its elements are defined as: ;in, Indicates the superedge The set of nodes contained therein This indicates the number of elements in the corresponding set. Reflecting nodes The average pairwise association strength between nodes and other nodes within the same hyperedge is used to characterize the aggregation features of pairwise relationships in higher-order structures. Represents a node The pairwise attention score with node k;
[0032] Step 4.4: Based on the matrix The original hypergraph attention weights are updated to obtain the node-to-hyperedge matrix under the hybrid attention form. Attention matrix from hyperedge to node : ; express transpose;
[0033] Step 4.5: Guided by the updated hybrid attention weights, calculate the node representations in the hypergraph space and heterogeneous graph space respectively; firstly, obtain the hyperedge representation by aggregating the feature information of nodes inside the hyperedge. : ;in, Indicates the superedge The set of nodes contained therein The weight matrix is a learnable matrix;
[0034] Step 4.6: Update the node representation based on the hybrid attention weights from the hyperedge to the node; Hybrid embedding representation in hypergraph space for: ;in, Represents nodes A set of connected hyperedges For the updated attention weights from the superedge to the node, The weight matrix is a learnable matrix. Represents a non-linear activation function;
[0035] In heterogeneous graph space, nodes The hybrid representation is updated by aggregating information from its neighboring nodes, and its form is as follows: ;in, Represents a node In the neighbor set of a heterogeneous graph, The updated pairwise attention weights;
[0036] Step 4.6: Through a hybrid attention encoder, the model can collaboratively model higher-order cooperative relationships and pairwise heterogeneous relationships within a unified framework, thereby obtaining common feature embeddings that fuse both types of structural information; correspondingly, the common embedding in the hypergraph space is represented as: ;in, , and The embeddings of lncRNA, disease, and miRNA in the hypergraph space are represented, respectively; the common embeddings in the heterogeneous graph space are represented as follows: ; , and This represents the embedding representation of the corresponding node in the heterogeneous graph space;
[0037] Step 4.7: Merge the common embeddings from the hypergraph space and the heterogeneous graph space, and define the shared node embedding representation of the two structural spaces as follows: .
[0038] Furthermore, in step 5, the view-level attention coefficient The calculation method is as follows:
[0039] First, for the nodes In the Embedded representation under each view Perform linear transformations and nonlinear mappings to obtain the corresponding attention scores. for: ;in, The weight matrix is a learnable matrix. For bias vectors, It is a non-linear activation function. Attention vectors shared across different views;
[0040] Subsequently, attention scores were calculated for all views. conduct Normalization is performed to obtain the final view-level attention coefficients: ;in, Indicates the total number of views.
[0041] Furthermore, the specific method for step 6 is as follows:
[0042] Step 6.1: Introduce a self-supervised contrastive learning task;
[0043] Overall Comparative Learning Loss for: , and Represents a node Representation in hypergraph and heterogeneous graph views For the number of nodes, This is a contrastive loss function at the single-node level; this loss function optimizes the bidirectional alignment between two views, enabling the model to learn node representations that maintain consistency across different views;
[0044] Step 6.2: Supervised Learning Loss for:
[0045] ,in, Indicates lncRNA With disease The real-world association tags between them Indicates lncRNA With disease The probability of association between them This is the training sample set for lncRNA-disease pairs. Indicates the number of training sample pairs;
[0046] The overall objective function is expressed as: ,in, These are hyperparameters used to balance the effects of different loss terms; the model is optimized through backpropagation under the guidance of labeled data, learning the embedding representations of lncRNA, miRNA, and disease for association prediction tasks.
[0047] Compared with existing technologies, the advantages of this invention are as follows:
[0048] 1. This invention overcomes the limitations of previous single-task methods by proposing a multi-task framework based on hypergraph learning. It effectively utilizes the complex relationships between known lncRNAs, miRNAs, and diseases to infer potential LDA and MDA.
[0049] 2. By constructing a hypergraph, the complex collaborative relationships between multiple nodes can be captured simultaneously, thereby more comprehensively characterizing the intrinsic association structure between lncRNA, miRNA and disease and improving feature expression capabilities.
[0050] 3. By introducing self-supervised contrastive learning tasks in hypergraph and heterogeneous graph spaces, the model's ability to discriminate node representations is enhanced, and its robustness and generalization performance under limited labeled data are improved. Attached Figure Description
[0051] Figure 1 This is a flowchart of the method of the present invention.
[0052] Figure 2 This is a comparative learning diagram of the present invention.
[0053] Figure 3 This is a diagram of the overall model architecture of the present invention. Detailed Implementation Plan
[0054] To make the purpose and innovation of this invention clearer, the following is combined with Figure 3 The present invention will be described in further detail below.
[0055] The specific embodiments of the present invention are as follows:
[0056] 1. Construction of Hypergraphs and Heterogeneous Graphs
[0057] First, by combining Gaussian interaction spectrum kernel similarity and functional similarity information, fusion similarity matrices between lncRNA, miRNA, and the disease were calculated. Then, based on the fusion similarity matrices, a Top-k nearest neighbor strategy was used to construct hyperedges and similar edges.
[0058] Based on this, a hypergraph structure of lncRNA, miRNA and disease is generated according to the constructed hyperedges, and a corresponding heterogeneous graph structure is constructed by combining known associations and similar edges.
[0059] 2. Hypergraph Attention Encoding
[0060] Based on the constructed hypergraph structure, a dedicated hypergraph attention layer is introduced to encode higher-order structural information. This attention layer models the collaborative relationships between multiple nodes during the node-hyperedge-node propagation process, thereby obtaining node representations based on higher-order structures.
[0061] During the stage of updating the hyperedge representation at a node, it is necessary to distinguish the contribution levels of different nodes within the same hyperedge. To this end, nodes are defined. For superedge Attention weights for: .
[0062] in, Indicates the superedge The set of nodes contained therein For a learnable parameter matrix, For trainable weight vectors, This indicates an attention mechanism using the scaled dot product form.
[0063] After obtaining the updated hyperedge representation, it is necessary to further pass higher-order information from the hyperedge back to the node. Therefore, an attention mechanism from the hyperedge to the node is introduced. Its associated superedge The attention score is defined as: .in, Indicates the presence of nodes The set of superedges and is a learnable parameter matrix.
[0064] Through the hypergraph attention encoding process described above, a node embedding representation based on high-order collaborative relationship modeling can be obtained: .in, , and These represent the embedding representations of lncRNA, disease, and miRNA nodes in the hypergraph structure, respectively.
[0065] 3. Graph Attention Encoding
[0066] Based on the heterogeneous graph structure, a dedicated graph attention layer is introduced to model the pairwise dependencies between nodes. This attention layer utilizes the heterogeneous graph structure information to assign adaptive weights to different neighbor nodes within a local neighborhood, thereby highlighting the contribution of key neighbors to the node representation update.
[0067] Specifically, nodes Its neighboring nodes Attention weights are defined as follows: .in, Represents a node The neighborhood group, This represents a vector concatenation operation. This is a learnable weight matrix.
[0068] Graph attention encoding can be used to obtain node embedding representations based on heterogeneous relationship modeling: .in, , and These represent the embedding representations of lncRNA, disease, and miRNA nodes in the heterogeneous graph space, respectively.
[0069] 4. Hybrid Attention Encoder
[0070] It should be noted that while the hypergraph attention mechanism and graph attention mechanism used in the hybrid attention encoder are mathematically consistent with the aforementioned dedicated attention layers, their functional positioning and modeling objectives differ fundamentally. The dedicated attention layers are used to independently model high-order collaborative relationships and paired heterogeneous relationships within their respective structural spaces. The hybrid attention encoder, however, acts as a cross-structural interaction enhancement module. By introducing attention information from another structural space, it adaptively reweights the original attention weights, thereby achieving collaborative modeling between high-order relationships and paired relationships. To avoid mutual interference between different modeling objectives, the attention layers in the hybrid attention encoder do not share parameters with the dedicated attention layers.
[0071] Let the pairwise attention matrix learned by the graph attention encoding module be... , of which elements Represents a node With nodes The pairwise dependency strength between nodes; the hypergraph attention encoding module obtains the attention matrix from each node to the hyperedge. And the attention matrix from the hyperedge to the node ,in Measure node For superedge Important contributions Then characterize the super-edge For nodes This indicates the degree of impact of the update.
[0072] First, to map the implicit high-order collaboration relationships in the hypergraph structure back to node pairs, we model the joint participation patterns of nodes in the hypergraph, construct the node association strength based on the high-order structure, and use this to enhance the original graph attention matrix. Specifically, the updated graph attention matrix is defined as: Among them, the symbol " "" indicates an element-wise addition operation. This update process allows graph attention to further integrate higher-order cooperative structure information on top of the existing pairwise dependencies.
[0073] Furthermore, to feed back pairwise relationship information to the hypergraph attention propagation process, a high-order relationship attention matrix is constructed by aggregating pairwise attention scores between nodes within the same hyperedge. Its elements are defined as: .in, Indicates the superedge The set of nodes it contains. Reflecting nodes The average pairwise association strength between the node and other nodes within the same hyperedge is used to characterize the aggregation features of pairwise relationships in higher-order structures.
[0074] Based on matrix The original hypergraph attention weights are updated to obtain the node-to-hyperedge matrix under the hybrid attention form. Attention matrix from hyperedge to node : .
[0075] Guided by the updated hybrid attention weights, the model computes node representations in both the hypergraph space and the heterogeneous graph space. Specifically, it first obtains the hyperedge representation by aggregating the feature information of nodes inside the hyperedge. : .in, Indicates the superedge The set of nodes contained therein This is a learnable weight matrix.
[0076] Subsequently, the node representation is updated based on the hybrid attention weights from the hyperedge to the node. Hybrid embedding representation in hypergraph space is defined as: .in, Represents nodes A set of connected hyperedges For the updated attention weights from the superedge to the node, The weight matrix is a learnable matrix. This represents a non-linear activation function.
[0077] Accordingly, in heterogeneous graph space, nodes The hybrid representation is updated by aggregating information from its neighboring nodes, and its form is as follows: .in, Represents a node In the neighbor set of a heterogeneous graph, This refers to the updated pairwise attention weights.
[0078] Through a hybrid attention encoder, the model can collaboratively model high-order cooperative relationships and pairwise heterogeneous relationships within a unified framework, thereby obtaining common feature embeddings that fuse both types of structural information. Correspondingly, the common embedding in the hypergraph space is represented as: .in, , and These represent the embeddings of lncRNAs, diseases, and miRNAs in the hypergraph space, respectively. The common embeddings of the heterogeneous hypergraph space are: . , and This represents the embedding representation of the corresponding node in the heterogeneous graph space.
[0079] Finally, the common embeddings from the hypergraph space and the heterogeneous graph space are merged, and the shared node embedding representation of the two structural spaces is defined as follows: .
[0080] 5. Integration of view-level attention mechanism
[0081] Considering that LDA and MDA prediction tasks may rely on information from a single view or a combination of multiple views, a view-level attention mechanism is introduced to adaptively fuse node representations from different views. Specifically, for nodes... Let its corresponding multi-view embedding representation be... ,in Includes two specific view embeddings , And a shared embedding The above multi-view representations are merged into a final node embedding representation through a view-level attention mechanism.
[0082] Final node representation Defined as a weighted sum of the embeddings of each view: .in, This represents the view-level attention coefficient, used to measure the importance of different views to the final node representation. This coefficient is adaptively learned through a shared attention scoring function, and its calculation process consists of two stages: attention scoring and normalization.
[0083] First, for the nodes In the Embedded representation under each view Perform linear transformations and nonlinear mappings to obtain the corresponding attention scores. for: .in, The weight matrix is a learnable matrix. For bias vectors, It is a non-linear activation function. Attention vectors shared across different views.
[0084] Subsequently, attention scores were calculated for all views. conduct Normalization is performed to obtain the final view-level attention coefficients: .in, Indicates the total number of views.
[0085] Through the above normalization process, it can be ensured that the attention coefficients of each view satisfy the following conditions: This enables adaptive weighted fusion of information contributions from different views.
[0086] 6. Objective Function and Inference
[0087] (1) Comparative learning loss
[0088] For specific encoder modules, considering the generally sparse nature of hypergraphs and heterogeneous graph structures, a self-supervised contrastive learning task is introduced into the corresponding hypergraph and heterogeneous graph encoding modules to enhance the model's ability to represent and utilize sparse structural information. This task guides the model to learn consistent and more discriminative representations at the structural level by maximizing the mutual information between node representations under different structural views, thereby mitigating the adverse effects of sparsity on representation learning.
[0089] Specifically, contrastive learning is performed between two views generated by different attention encoders, corresponding to a hypergraph structure and a heterogeneous graph structure, respectively. Representations of the same node in both views are considered positive sample pairs, while representations of different nodes are considered negative sample pairs. Under this setting, the objective function of contrastive learning is defined as the average of the losses for all positive sample pairs, thereby enabling the model to learn semantically consistent node embeddings across different views.
[0090] Numerous studies have demonstrated that mapping learned representations to new latent spaces and applying nonlinear projections can improve representation quality within contrastive learning frameworks. Therefore, a projection head is introduced after the outputs of the hypergraph attention layer and the graph attention layer to map the representations to another latent space. This projection head consists of two multilayer perceptron (MLP) layers and an exponential linear unit (ELU) activation function to enhance the discriminative power and generalization performance of the representations. The hypergraph node representation obtained after the projection head... and heterogeneous graph node representation They are respectively: , .in, and is a learnable parameter matrix.
[0091] Specifically, nodes The representations in the two views are denoted as follows: and The corresponding contrastive learning loss is defined as: .in, This represents the similarity measurement function. For temperature coefficient, Indicates the size of the set of nodes.
[0092] Based on this, the overall contrastive learning loss function is defined as: This loss function optimizes bidirectional alignment between two views, enabling the model to learn consistent node representations across different views.
[0093] (2) Monitoring loss
[0094] Previously, the fused node embedding representation was obtained through a view-level attention mechanism: ,in, , and These represent the node embeddings for lncRNA, disease, and miRNA, respectively.
[0095] For different prediction tasks, a multi-relationship decoder is used to infer whether potential interactions exist between lncRNA-disease pairs, miRNA-disease pairs, and lncRNA-miRNA pairs. Taking the LDA prediction task as an example, lncRNA... With disease The probability of association between them is defined as: .in Represents the lncRNA embedding matrix The OK, Represents the disease embedding matrix The Line, symbol " "and" "" represents element-wise multiplication and vector concatenation, respectively. It consists of two fully connected networks.
[0096] Accordingly, the supervised learning loss function is defined as: .in, Indicates lncRNA With disease The real-world association tags between them This is the training sample set for lncRNA-disease pairs. This indicates the number of training sample pairs.
[0097] Intuitively, the optimization objective consists of two parts: supervised learning loss. Comparative learning loss Therefore, the overall objective function can be expressed as: .in, These are hyperparameters used to balance the effects of different loss terms. The model is optimized via backpropagation, guided by labeled data, to learn embeddings of lncRNA, miRNA, and disease for association prediction tasks.
[0098] 7. Evaluation Model
[0099] This invention selects the area under the receiver operating feature curve (AUC) and the area under the precision-recall curve (AUPR) as evaluation metrics for model prediction performance, and employs 5-fold cross-validation (5-cv) to evaluate the model's performance.
[0100] Specifically, for each dataset, all known LDAs and MDAs are first randomly divided into five subsets of similar size. In each round of cross-validation, one subset is selected as the test set, and the remaining four subsets are used for model training.
[0101] During the training phase, the same number of negative samples as the number of training positive samples are randomly selected from the candidate negative sample set, and these negative samples, together with the training positive samples, constitute the training dataset. During the testing phase, the positive samples from the test subset are used together with all candidate negative samples for model performance evaluation. This strategy mitigates the impact of the imbalance between the number of positive and negative samples on model training while more realistically evaluating the model's predictive ability in large-scale, unknown correlation scenarios.
[0102] Two publicly available datasets used in previous studies were selected for the experiment. Dataset 1 includes 240 lncRNAs, 405 diseases, 495 miRNAs, 2687 LDAs, and 13559 MDAs. Dataset 2 contains 665 lncRNAs, 316 diseases, 295 miRNAs, 3833 LDAs, and 8540 MDAs.
[0103] The following frontier association prediction models were selected as benchmarks for comparison:
[0104] GCLMTP: A multi-task prediction method based on graph contrastive learning. This method constructs a three-layer heterogeneous network by integrating similarity and association information between lncRNAs, miRNAs, and diseases, and uses a graph contrastive learning model combined with a graph convolutional network to learn the latent representations of nodes. Subsequently, a classifier is used to predict the latent LDA and MDA.
[0105] GAERF: A method for predicting LDA based on graph autoencoders and random forests. This method first constructs a lncRNA-miRNA-disease heterogeneous network and utilizes a graph autoencoder to learn low-dimensional representations of nodes from the network, thereby reducing the dimensionality and heterogeneity of biological data. Subsequently, the learned node representations are used as features input to a random forest classifier to predict potential LDA.
[0106] GCN (Graph Convergence Neighborhood Network): By aggregating neighborhood information on a graph structure and fusing the features of a node's own features with those of its neighbors, GCN learns a low-dimensional representation that includes topological information. Because it can effectively model structural relationships in graph data, GCN has been widely used in association prediction tasks in bioinformatics.
[0107] MLGCNET: A method for LDA prediction based on multi-layer graph convolutional networks. This method constructs a lncRNA and disease similarity network using Top-k similarity information and combines it with known associations to build a heterogeneous network. Subsequently, it uses a multi-layer graph convolutional network to learn the latent representations of nodes and uses an ExtraTrees classifier to predict the potential associations between lncRNAs and diseases.
[0108] IPCARF: An ensemble learning-based method for predicting LDA. This method constructs feature vectors by integrating disease semantic similarity, lncRNA similarity, and Gaussian kernel similarity, performs feature dimensionality reduction using incremental principal component analysis, and finally employs a random forest model to predict potential LDA.
[0109] LDA-VGHB: An LDA prediction method based on matrix factorization and graph autoencoder. This method first extracts features through singular value decomposition and uses a variational graph autoencoder to learn the latent representation of nodes. Then, a heterogeneous Newton booster is used to predict the potential association between lncRNAs and diseases.
[0110] GAEMDA: A graph autoencoder-based MDA prediction method. This method utilizes a graph neural network encoder to aggregate node neighborhood information to learn node embedding representations, and reconstructs the potential association between miRNAs and diseases through a bilinear decoder, thereby achieving association prediction.
[0111] PMDFI: An ensemble learning-based method for predicting MDA. This method utilizes stacked autoencoders to extract high-order features from the similarity matrix, further enhances feature representation through feature interaction learning, and finally employs an ensemble model consisting of random forest and logistic regression to predict MDA.
[0112] PGCNMDA is an MDA prediction method based on graph convolutional networks. This method improves the graph convolution process by learning path-based graph space operators, thereby obtaining more efficient node representations for predicting potential MDAs.
[0113] As shown in Table 1, the experimental results demonstrate that the present invention outperforms other comparative methods in both LDA and MDA prediction tasks on both datasets. In the LDA prediction task, the present invention achieves the best results on both datasets. On Dataset 1, the AUC and AUPR of the present invention are 0.9461 and 0.3597, respectively. Compared to the suboptimal method LDA-VGHB (AUC of 0.9358), the AUC is improved by approximately 1.10%; its AUPR is also slightly higher than GCLMTP and MLGCNET. On Dataset 2, the AUC and AUPR of the present invention reach 0.9455 and 0.1963, respectively, representing improvements of approximately 0.84% and 7.50% compared to the suboptimal method GCLMTP (AUC of 0.9371, AUPR of 0.1826).
[0114] In the MDA prediction task, this invention also performed optimally. On Dataset 1, the AUC and AUPR of this invention were 0.9491 and 0.4133, respectively, both higher than GCLMTP (AUC 0.9416, AUPR 0.4093) and PGCMNMDA (AUC 0.9404, AUPR 0.3889). On Dataset 2, the AUC and AUPR of this invention reached 0.9219 and 0.2994, respectively, representing improvements of approximately 0.77% and 3.00% compared to the suboptimal method PGCMNMDA (AUC 0.9148, AUPR 0.2907).
[0115] Table 1. Prediction results of this invention on two public datasets.
[0116]
Claims
1. A method for predicting the association between non-coding RNA and disease based on hypergraph contrastive learning, the method comprising: Step 1: Based on domain knowledge and known associations between lncRNAs, miRNAs, and diseases, construct hypergraphs and heterographs; First, by combining Gaussian interaction spectrum kernel similarity and functional similarity information, the fusion similarity matrices of lncRNA, miRNA and disease are calculated respectively. Then, based on the fusion similarity matrix, the Top-k nearest neighbor strategy is used to construct hyperedges and similar edges. On this basis, the hypergraph structure of lncRNA, miRNA and disease is generated according to the constructed hyperedges, and the corresponding heterogeneous graph structure is constructed by combining the known association relationships and similar edges. Step 2: Use the hypergraph attention encoder to learn the representation of the hypergraph and extract high-order relation features; A dedicated hypergraph attention layer is introduced to encode higher-order structural information; this attention layer establishes collaborative relationships between multiple nodes during the node-hyperedge-node propagation process, obtaining node representations based on higher-order structures: ,in, , and These represent the embedding representations of lncRNA, disease, and miRNA nodes in the hypergraph structure, respectively. Step 3: Use a graph attention encoder to learn representations of heterogeneous graphs and extract pairwise relation features; A dedicated graph attention layer is introduced to model the pairwise dependencies between nodes; through graph attention encoding, a node embedding representation based on heterogeneous relationship modeling is obtained: ,in, , and These represent the embedding representations of lncRNA, disease, and miRNA nodes in the heterogeneous graph space, respectively. Step 4: Fuse hypergraph and heterogeneous graph information using a hybrid attention encoder to obtain a unified node representation; By using a hybrid attention encoder, common feature embeddings that fuse the two types of structural information are obtained. Correspondingly, the common embedding in the hypergraph space is represented as: ,in, , and The embeddings of lncRNA, disease, and miRNA in the hypergraph space are represented, respectively; the common embeddings in the heterogeneous graph space are represented as follows: , , and This represents the embedding representation of the corresponding node in the heterogeneous graph space; Finally, the representations from the two spaces are fused to obtain a shared embedding: ; Step 5: Adaptively fuse node representations under different structural views to further enhance feature representation capabilities; Through a view-level attention mechanism, the above multi-view representations are fused into a final node embedding representation. For a node... Let its corresponding multi-view embedding representation be... ,in Includes two specific view embeddings , And a shared embedding ; Final node representation Defined as a weighted sum of the embeddings of each view: ,in, This represents the view-level attention coefficient, used to measure the importance of different views to the final node representation; Step 6: Introduce self-supervised contrastive learning to optimize node representations, and complete the prediction of non-coding RNA associations with diseases based on fusion representations.
2. The method for predicting the association between non-coding RNA and disease based on hypergraph contrastive learning as described in claim 1, characterized in that, The encoding method of the hypergraph attention encoder in step 2 is as follows: During the stage of updating the hyperedge representation at the node, the contribution levels of different nodes in the same hyperedge are distinguished; nodes are defined. For superedge Attention weights for: ; in, Indicates the superedge The set of nodes contained therein For a learnable parameter matrix, For trainable weight vectors, This indicates an attention mechanism using the scaled dot product form. Represents a node; Introducing an attention mechanism from hyperedges to nodes, nodes Its associated superedge Attention score for: ;in, Indicates the presence of nodes The set of superedges and is a learnable parameter matrix.
3. The method for predicting the association between non-coding RNA and disease based on hypergraph contrastive learning as described in claim 2, characterized in that, The attention weights of the attention encoder in step 3 are shown. The calculation method is as follows: node Its neighboring nodes Attention weights for: ； in, Represents a node The neighborhood group, This represents a vector concatenation operation. The weight matrix is a learnable matrix; , , Represents a node. This represents the activation function.
4. The method for predicting the association between non-coding RNA and disease based on hypergraph contrastive learning as described in claim 3, characterized in that, The specific method of the hybrid attention encoder in step 4 is as follows: Step 4.1: Let the pairwise attention matrix learned by the graph attention encoding module be... , of which elements Represents a node With nodes The strength of pairwise dependence between them The number of nodes is represented; the hypergraph attention encoding module obtains the attention matrix from each node to the hyperedge. And the attention matrix from the hyperedge to the node ,in Measure node For superedge Important contributions Then characterize the super-edge For nodes Indicates the degree of impact of the update. Indicates the number of edges; Step 4.2: To map the implicit higher-order collaborative relationships in the hypergraph structure back to the node pairs, the joint participation patterns of nodes in the hypergraph are modeled to construct the node association strength based on the higher-order structure, which is then used to enhance the original graph attention matrix; the updated graph attention matrix... for: Among them, the symbol " " indicates an element-wise addition operation; Step 4.3: To feed back pairwise relationship information to the hypergraph attention propagation process, a high-order relationship attention matrix is constructed by aggregating pairwise attention scores between nodes within the same hyperedge. Its elements are defined as: ;in, Indicates the superedge The set of nodes contained therein This indicates the number of elements in the corresponding set. Reflecting nodes The average pairwise association strength between nodes and other nodes within the same hyperedge is used to characterize the aggregation features of pairwise relationships in higher-order structures. Represents a node The pairwise attention score with node k; Step 4.4: Based on the matrix The original hypergraph attention weights are updated to obtain the node-to-hyperedge matrix under the hybrid attention form. Attention matrix from hyperedge to node : ; express transpose; Step 4.5: Guided by the updated hybrid attention weights, calculate the node representations in the hypergraph space and heterogeneous graph space respectively; firstly, obtain the hyperedge representation by aggregating the feature information of nodes inside the hyperedge. : ;in, Indicates the superedge The set of nodes contained therein The weight matrix is a learnable matrix; Step 4.6: Update the node representation based on the hybrid attention weights from the hyperedge to the node; Hybrid embedding representation in hypergraph space for: ;in, Represents nodes A set of connected hyperedges For the updated attention weights from the superedge to the node, The weight matrix is a learnable matrix. Represents a non-linear activation function; In heterogeneous graph space, nodes The hybrid representation is updated by aggregating information from its neighboring nodes, and its form is as follows: ;in, Represents a node In the neighbor set of a heterogeneous graph, The updated pairwise attention weights; Step 4.6: Through a hybrid attention encoder, the model can collaboratively model higher-order cooperative relationships and pairwise heterogeneous relationships within a unified framework, thereby obtaining common feature embeddings that fuse both types of structural information; correspondingly, the common embedding in the hypergraph space is represented as: ;in, , and The embeddings of lncRNA, disease, and miRNA in the hypergraph space are represented, respectively; the common embeddings in the heterogeneous graph space are represented as follows: ; , and This represents the embedding representation of the corresponding node in the heterogeneous graph space; Step 4.7: Merge the common embeddings from the hypergraph space and the heterogeneous graph space, and define the shared node embedding representation of the two structural spaces as follows: .
5. The method for predicting the association between non-coding RNA and disease based on hypergraph contrastive learning as described in claim 4, characterized in that, View-level attention coefficient in step 5 The calculation method is as follows: First, for the nodes In the Embedded representation under each view Perform linear transformations and nonlinear mappings to obtain the corresponding attention scores. for: ;in, The weight matrix is a learnable matrix. For bias vectors, It is a non-linear activation function. Attention vectors shared across different views; Subsequently, attention scores were calculated for all views. conduct Normalization is performed to obtain the final view-level attention coefficients: ;in, Indicates the total number of views.
6. The method for predicting the association between non-coding RNA and disease based on hypergraph contrastive learning as described in claim 5, characterized in that, The specific method for step 6 is as follows: Step 6.1: Introduce a self-supervised contrastive learning task; Overall Comparative Learning Loss for: , and Represents a node Representation in hypergraph and heterogeneous graph views For the number of nodes, This is a contrastive loss function at the single-node level; this loss function optimizes the bidirectional alignment between two views, enabling the model to learn node representations that maintain consistency across different views; Step 6.2: Supervised Learning Loss for: ,in, Indicates lncRNA With disease The real-world association tags between them Indicates lncRNA With disease The probability of association between them This is the training sample set for lncRNA-disease pairs. Indicates the number of training sample pairs; The overall objective function is expressed as: ,in, These are hyperparameters used to balance the effects of different loss terms; the model is optimized through backpropagation under the guidance of labeled data, learning the embedding representations of lncRNA, miRNA, and disease for association prediction tasks.