IncRNA-miRNA association prediction system and method fused with hypergraph perspective

Abstract

The invention discloses an lncRNA-miRNA association prediction system and method fused with a hypergraph perspective, and belongs to the technical field of biological information. The LncRNA-miRNA association prediction method aims at solving the problem that in an existing LncRNA-miRNA association prediction method, the node representation capacity is insufficient, interaction relation modeling is not comprehensive, and consequently prediction performance is poor. According to the method, lncRNA and miRNA sequence data are obtained and subjected to data preprocessing, then K-mer frequency features, Doc2Vec semantic features, CTD structural features and Role2Vec topological features of each RNA are obtained, multi-level feature fusion is carried out to obtain unified fusion features of each RNA, then the unified fusion features are sent into GCN and HGCN networks to obtain respective corresponding features, optimization is carried out on the basis of comparative learning optimization, and the fusion features of each RNA are obtained. And then dynamic fusion is carried out, final fusion feature representation is obtained based on the gating factor, and lncRNA-miRNA association prediction is realized based on the final fusion feature representation.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and relates to a lncRNA-miRNA association prediction system and method. Background Art

[0002] Non-coding RNA (ncRNA) is a class of RNA molecules that do not encode proteins but possess diverse regulatory functions. They play important roles in gene expression, cell proliferation, differentiation, and the development and progression of various diseases. Long non-coding RNA (lncRNA) and microRNA (miRNA) are two of the most extensively studied classes of the ncRNA family, exhibiting key regulatory functions in biological processes such as cancer, neurological diseases, and immune regulation. Previous studies have demonstrated that lncRNAs and miRNAs interact through various mechanisms. For example, lncRNAs can act as miRNA precursors, ceRNAs (competing endogenous RNAs), or miRNA sponges, while miRNAs can also reversely target lncRNAs and mediate their degradation, forming a complex bidirectional regulatory network. This highly dynamic lncRNA-miRNA interaction is crucial for understanding biological regulatory mechanisms and disease signaling pathways. While experimental approaches can validate these molecular interactions, the process is often time-consuming and resource-intensive, making it difficult to rapidly mine large-scale bioinformatics data. Therefore, developing efficient and accurate computational models to predict potential lncRNA–miRNA associations has become a hot topic in bioinformatics research. In recent years, lncRNA–miRNA association prediction methods have primarily evolved into three categories: similarity-based computational methods, representation learning methods based on multi-view fusion, and structural modeling methods based on graph neural networks. While these methods have improved prediction efficiency and performance to a certain extent, they still face numerous challenges in practical application.

[0003] First, current lncRNA-miRNA association studies often employ graph-based modeling, such as constructing bipartite or heterogeneous graphs to simulate the interactions between lncRNAs and miRNAs. These approaches have achieved some success in modeling direct one-to-one or one-to-many interactions. However, actual biological regulation processes are often more complex. Many lncRNAs can simultaneously regulate multiple miRNAs through diverse mechanisms, forming cross-target, multi-pathway synergistic regulatory networks. These high-order, many-to-many interaction patterns are difficult to effectively represent using existing graph modeling paradigms. Limited by the binary nature of edges, traditional graph models can only capture point-to-point information transfer and struggle to capture the collective synergistic interactions inherent in the complex regulatory landscape mediated by lncRNAs. Therefore, there is an urgent need to introduce more flexible graph-based representations that can more naturally simulate the biological processes of lncRNA regulation of multiple miRNAs, thereby enhancing the model's ability to model multi-source information and high-order relationships.

[0004] Furthermore, due to the limited amount of directly verified lncRNA–miRNA interaction information in existing databases, modeling higher-order patterns in the absence of known interaction samples further exacerbates the sparsity of the feature space. On the one hand, traditional graph structures struggle to construct effective neighborhoods when some nodes lack direct connections, limiting their ability to express node features. On the other hand, existing deep learning models often rely on neighborhood aggregation mechanisms for representation learning. Given the sparse nature of high-order interactions, this not only makes it difficult to obtain stable propagation paths, but also easily leads to noise accumulation and feature degradation. This series of challenges severely restricts the scalability and robustness of lncRNA–miRNA association prediction models in complex regulatory networks. Therefore, there is an urgent need to introduce higher-order representation mechanisms that can mitigate structural sparsity and enhance the ability to model node associations, providing a stronger expressive foundation for subsequent feature fusion and prediction tasks. Summary of the Invention

[0005] The present invention aims to solve the problems in existing LncRNA-miRNA association prediction methods, such as insufficient node representation capability and incomplete interaction relationship modeling, which lead to poor prediction performance.

[0006] A lncRNA-miRNA association prediction method integrating a hypergraph perspective includes the following steps: LncRNA and miRNA sequence data were acquired and preprocessed. K-mer frequency features, Doc2Vec semantic features, and CTD structural features were then obtained for each RNA. A weighted undirected graph structure was constructed based on the similarity relationships between RNA fragments. Each node in the graph corresponds to an RNA fragment, and the edge weights represent the cosine similarity of the nodes. The Role2Vec model was introduced into this graph for structure-aware embedding to obtain Role2Vec topological features. Based on K-mer frequency features, Doc2Vec semantic features, CTD structural features and Role2Vec topological features, multi-level feature fusion is performed to obtain a unified fusion feature representation for each RNA , the fusion features Unified representation of RNA , as the initial embedded features of lncRNA and miRNA; The initial embedded features of lncRNA and miRNA are fed into the graph convolutional network GCN to obtain features , the initial embedding features of lncRNA and miRNA are fed into the hypergraph convolutional network HGCN to obtain ; Calculate the feature Z and The similarity between ;based on Perform comparative learning optimization; then, and Perform dynamic fusion and obtain the final fusion feature representation based on the gating factor : in, It is a hyperparameter used to balance the contribution of sequence features to the fusion results; Based on the final fusion feature representation Enable lncRNA–miRNA association prediction.

[0007] Furthermore, the cosine similarity of the nodes corresponding to the edge weights in the weighted undirected graph structure is as follows: in, , , is the feature weighting parameter, satisfying is the cosine similarity of RNA fragments in k-mer frequency features, is the cosine similarity of RNA fragments in CTD structural features, is the cosine similarity of RNA fragments in Doc2Vec semantic features.

[0008] Furthermore, based on K-mer frequency features, Doc2Vec semantic features, CTD structural features and Role2Vec topological features, a multi-layer perceptron combined with an attention mechanism is used to perform multi-level feature fusion.

[0009] Furthermore, the construction process of the hypergraph convolutional network HGCN is as follows: Step c1: lncRNA and miRNA nodes are considered as two types of heterogeneous nodes in the hypergraph. If a lncRNA node has an interaction relationship with one or more miRNA nodes, a hyperedge is constructed to connect the lncRNA node with the corresponding multiple miRNA nodes to form a lncRNA-dominated hypergraph structure. ; For each hyperedge connected to multiple nodes, the characteristics of each node Add the sum by dimension and divide by the number of nodes to get the unified vector representation of the hyperedge; and If there is a known interaction between them, a connection is established in the hypergraph; Step c2: Based on the lncRNA and miRNA feature vectors output by the feature extraction module, set For the fused feature vector, the cosine similarity between similar nodes is calculated and normalized to the interval [0,1]. When the normalized value is greater than the first threshold, a new hyperedge connection is added. When the normalized value is less than the second threshold, the hyperedge connection is removed. In other cases, the initial structure remains unchanged. At the same time, the hyperedge weight matrix W is determined. e .

[0010] Furthermore, the formula for calculating the cosine similarity between similar nodes in step c2 and normalizing it to the interval [0,1] is as follows: in, 、 Represents two nodes of the same type .

[0011] Furthermore, the hyperedge weight matrix W is determined in step c2. e The Gaussian mixture model GMM is used to calculate the hyperedge weight matrix in the process, including: First, the fusion lncRNA and miRNA characteristics Perform GMM clustering; then, GMM calculates the probability of each RNA node belonging to different mixed components and uses this as the hyperedge weight matrix W e .

[0012] Furthermore, the initial embedded features of lncRNA and miRNA are fed into the graph convolutional network GCN to obtain the features The process includes: Graph convolutional encoding is performed on the local neighborhood connectivity between lncRNA and miRNA nodes to extract their local structural representation. The direct connectivity between lncRNA and miRNA is represented by the adjacency matrix 𝐴, which is treated as a bipartite graph structure. Graph convolutional neural network (GCN) is used for feature propagation. During this process, the adjacency matrix is ​​normalized and the normalized adjacency matrix is ​​calculated. Where D is the degree matrix of the adjacency matrix, which is used to represent the sparsity of the connection between lncRNA and miRNA; In the GCN layer, graph convolution operation is performed based on the normalized adjacency matrix: in, is the initial embedded feature of lncRNA and miRNA, To learn the parameters, is the nonlinear activation function ReLu, and Z represents the local topological features of lncRNA and miRNA after transmission through GCN.

[0013] Furthermore, the initial embedding features of lncRNA and miRNA are fed into the hypergraph convolutional network HGCN to obtain The process includes: The hypergraph structure between lncRNA and miRNA is modeled using the hypergraph convolutional neural network HGCN to further extract hypergraph information features; trainable parameter matrices are set for the hyperedge weights of lncRNA and miRNA nodes respectively. and , which are used to calculate the hyperedge weights of lncRNA and miRNA nodes respectively; among them, initially and Together they form the overall edge weight matrix W e , used for edge weight adjustment in the subsequent hypergraph convolution information propagation process; the initial feature embedding of lncRNA and miRNA is obtained respectively as well as , 、 That is, the unified vector representation of hyperedges; During the HGCN propagation process, the hypergraph aggregates the features between lncRNA and miRNA through hyperedges to obtain the aggregated features. ; Feature propagation based on hyperedge weights, the formula is as follows: Where H is the hyperedge adjacency matrix, and are the node degree matrix and the hyperedge degree matrix respectively, It is the node feature representation obtained after HGCN propagation, and BatchNorm is used for feature normalization.

[0014] Further, based on Contrastive loss function in the contrastive learning optimization process as follows: in, To control the temperature parameter of the contrast loss, Represents the negative sample set.

[0015] Furthermore, the gating factor ,in, is the weight matrix, is the bias term, Represents vector concatenation operation, is the Sigmoid activation function.

[0016] Beneficial Effects: This application proposes a lncRNA-miRNA association prediction method that integrates a hypergraph perspective, constructing high-order interactions to characterize the complex biological regulatory patterns in which lncRNAs simultaneously regulate multiple miRNAs. By constructing hyperedge structures between lncRNAs and multiple miRNAs, natural modeling of many-to-many regulatory patterns is achieved, effectively enhancing the expressive power of high-order relationships between nodes. To address the feature sparsity issue in hypergraphs, structural feature encoding methods such as Role2Vec and CTD are further introduced to enrich the initial feature representation of nodes from multiple perspectives. Sequence semantic information such as K-mer and Doc2Vec is integrated to construct a multi-source feature representation, improving the separability of lncRNAs and miRNAs in the feature space. A hyperedge enhancement mechanism and trainable hyperedge weights are introduced to effectively optimize the expressive power of the hypergraph structure. Local and global features are jointly extracted through GCN and HGCN. To align the embedding representations from the two perspectives, a cross-perspective comparative learning strategy based on InfoNCE loss is introduced to enhance the model's ability to consistently learn multimodal features. During the feature fusion stage, a gating mechanism is introduced to achieve an adaptive weighted combination of sequence and structural features, and the final prediction accuracy is improved through the KAN classifier. Experiments on multiple benchmark datasets have demonstrated that this method significantly outperforms existing advanced models in metrics such as AUC and AUPR, exhibiting good generalization and stability. It is particularly suitable for predicting LncRNAs involved in the coordinated regulation of multiple targets in complex biological environments, and has broad application prospects and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Predict a flow chart for this application; Figure 2 Comparison of ROC and AUPR prediction performance in a five-fold cross-validation experiment on the dataset with other comparison methods; Figure 3 This is a diagram showing the impact of the combination of multi-source features on the model prediction performance on the dataset; Figure 4 A graph comparing the impact of experimental ablation on model prediction performance on the dataset; Figure 5 The result graph of clustering visualization before and after training on the dataset. DETAILED DESCRIPTION

[0018] The present invention is mainly aimed at solving the following problems existing in existing lncRNA-miRNA association prediction methods: On the one hand, most current methods are based on bipartite graph structure modeling, which makes it difficult to effectively depict the complex coordinated regulatory pattern of LncRNAs regulating multiple miRNAs simultaneously, and lack the ability to model many-to-many interaction relationships; On the other hand, due to the sparsity of high-order relational data and limited node feature expression, the model still has shortcomings in multi-source information fusion and deep feature expression, which affects the prediction performance and generalization ability in complex regulatory contexts.

[0019] Existing lncRNA-miRNA association prediction methods suffer from insufficient node representation capabilities and incomplete interaction modeling. Especially in the case of many-to-many high-order regulatory relationships, traditional graph structures lack high precision in expressing biological information, resulting in limited feature aggregation and difficulty improving prediction performance. To address this, the present invention proposes a lncRNA-miRNA association prediction method that incorporates a hypergraph perspective. This is explained in detail below with reference to specific implementations.

[0020] Specific implementation method 1: Combination Figure 1 To explain this embodiment,

[0021] This embodiment is a lncRNA-miRNA association prediction system and prediction method that integrates a hypergraph perspective. The prediction system includes a data acquisition and preprocessing unit, a feature extraction module, a hypergraph construction module, a structure information fusion module, and a relationship prediction module; The data acquisition and preprocessing module is used to collect lncRNA and miRNA sequence data, lncRNA-miRNA association data, and perform data preprocessing to obtain the initial training and prediction data set for the model. The specific process includes: Step a1: Obtain lncRNA sequence data from the GENCODE v38 database; at the same time, obtain miRNA sequence data from the miRBase v22.1 database; Step a2: Combine the lncRNASNP database to obtain known lncRNA-miRNA association data as LMI positive samples, and match the corresponding entities based on the sequence information obtained in step a1; Step a3: RNA sequences were normalized, such as by uniformly replacing bases T with U, removing duplicate interaction pairs, and filtering invalid or missing sequence data. Furthermore, a sample balancing strategy was employed to construct an interaction sample dataset containing 17,369 lncRNA–miRNA interaction samples. Negative interaction samples were randomly sampled from lncRNA–miRNA pairs with unannotated interaction relationships, ultimately forming a dataset with a balanced ratio of positive and negative samples. The lncRNA and miRNA feature extraction module specifically performs the following steps: Step b1: Each RNA sequence According to fixed length Divide into overlapping k-mer segments ,in ; Then, count the occurrences of all k-mer segments to generate a k-mer count vector v (k-mer frequency feature): (1) in, For the summation symbol, is the indicator function. When the i-th k-mer segment is equal to the j-th k-mer type, the indicator function value is 1, otherwise it is 0. vⱼ is the j-th element of the k-mer count vector, which indicates the total number of times the k-mer of type j appears in the sequence.

[0022] Step b2: Treat each RNA sequence as a document and the k-mer fragment as a word in the document. Use the distributed memory (DM) method to train the Doc2vec model to generate semantic features for each lncRNA and miRNA sequence. Step b3: Based on the composition, transition and distribution characteristics of nucleotides, the CTD descriptor is introduced. The CTD feature consists of three parts: composition, transition and distribution. Composition characteristics , used to calculate nucleotides Composition features theoretically include the proportions of nucleotides A, U, C, and G in a sequence. However, since the sum of these four features is always 1, a linear dependency exists. To reduce redundancy and improve feature efficiency, this paper only retains the three-dimensional composition features as a representation, ensuring sufficient information expression while avoiding noise caused by redundant dimensions. Convert Features , used to calculate the frequency of adjacent occurrences of different nucleotides; Distribution characteristics , and count the distribution of specific nucleotides at five key positions in the RNA sequence: 0%, 25%, 50%, 75%, and 100%.

[0023] The concatenation of the three parts of the CTD feature actually results in a 21-dimensional vector: Composition 3-dimensional, Transition 3-dimensional, and Distribution 15-dimensional (5 quantiles × 3 categories).

[0024] To improve the robustness of feature extraction, the present invention uses a sliding window method to calculate features of RNA sequences. The window size w is set and slides across the sequence with a step size s. Each time, the CTD features of the nucleotides within the window are extracted and their normalized representation in the complete sequence is calculated. Step b4: A weighted undirected graph is constructed based on the similarity relationships between RNA fragments. Each node in the graph represents an RNA fragment, and the edge weights represent their cosine similarity (calculated based on k-mer frequency features, CTD features, and Doc2Vec representations). To further characterize the roles of different nodes in this graph, the Role2Vec model is introduced for structure-aware embedding. This extracts the graph role representation (i.e., structural features) of each RNA node, which serves as a supplementary feature in the downstream network construction and fusion process. This embedding, as part of the structural features, is combined with other features for downstream graph neural network modeling.

[0025] To construct RNA similarity graph, RNA fragments were calculated and The similarity score between them is as shown in the formula: in, , , is the feature weighting parameter, satisfying . is the cosine similarity of RNA fragments in the k-mer frequency feature space, It is the cosine similarity of two RNA fragments in the 21-dimensional CTD structural features composed of Composition, Transition, and Distribution. is the cosine similarity of RNA fragments in the Doc2Vec semantic feature space.

[0026] Step b5: extract the multi-source heterogeneous features from steps b1 to b4, including K-mer frequency features , Doc2Vec semantic features, CTD structural features, and Role2Vec topological features; multi-layer perceptron (MLP) combined with attention mechanism is used for multi-level feature fusion: For each feature, MLP is first used to perform nonlinear transformation on each feature so that each feature is mapped to the same dimension d. Each MLP consists of two layers of fully connected neural networks and uses ReLU as the activation function to introduce nonlinear expression capabilities.

[0027] After the transformation, all features are mapped to the same dimension d to ensure that different features can be calculated uniformly. In order to further optimize the feature fusion process, the attention mechanism is used to automatically learn the importance of different features and calculate the weighted fusion RNA feature representation. Specifically, a learnable importance weight is first calculated for each feature. , which is normalized by Softmax to ensure that the sum of the importance weights of all features is 1. The calculation process is shown in the formula; Among them, w is the trainable weight vector, Represents the features obtained after different features are transformed by MLP, Represents the features obtained after different features are transformed by MLP.

[0028] After MLP alignment and attention mechanism fusion, all RNA features are mapped to the same feature space and weighted merged according to the learned importance weights to obtain a unified fusion feature representation for each RNA. , m refers to the total number of multi-source features; the fusion feature It is used as input for subsequent graph construction and node representation learning, i.e., a unified representation of all RNAs , as the initial embedded features of lncRNA and miRNA.

[0029] The hypergraph construction module specifically performs the following steps: Step c1: Construct a hypergraph structure based on known lncRNA-miRNA interaction data to express the high-order interaction relationship between lncRNA and multiple miRNAs, including the following steps: Step c11: lncRNA and miRNA nodes are considered as two types of heterogeneous nodes in the hypergraph. If a lncRNA node has an interaction relationship with one or more miRNA nodes, a hyperedge is constructed to connect the lncRNA node with the corresponding multiple miRNA nodes to form a lncRNA-dominated hypergraph structure. ; Where V is the node set composed of lncRNA and miRNA, E is the hyperedge set, and W represents the weight of the hyperedge. Connecting a lncRNA to multiple miRNAs.

[0030] For each hyperedge in the actual hypergraph, Inconsistent numbers of connected miRNA nodes (e.g. some hyperedges connect two nodes, some connect three or even more nodes) may lead to inconsistent feature dimensions, affecting the calculation and propagation of subsequent models. To solve this problem, a unified node feature mean aggregation strategy is adopted, that is, for multiple nodes connected by each hyperedge, the feature of each node is averaged. The unified vector representation of the hyperedge is obtained by summing the dimensions and dividing by the number of nodes. This method ensures that all hyperedge feature dimensions are consistent, facilitating unbiased input into the hypergraph neural network model and improving the stability and robustness of feature information propagation.

[0031] Step c12: Construct a hyperedge connection matrix in the form of a Boolean matrix ,Right now and If there is a known interaction between them, a connection is established in the hypergraph, otherwise it is set to 0.

[0032] Step c2: Introducing a feature similarity calculation mechanism to construct auxiliary edge connections to enhance the expressiveness of the graph structure and the information dissemination effect, including the following steps: Step c21: First, according to the lncRNA and miRNA feature vectors output by the feature extraction module, set For the fused feature vector, the cosine similarity between similar nodes (lncRNA-lncRNA, miRNA-miRNA) is calculated and normalized to the [0,1] interval. The calculation process is as follows: in, 、 Represents two nodes of the same type ; Used to represent two nodes of the same type.

[0033] Step c22: Based on the similarity score, set the threshold screening rule: when S(r,s)>0.8, add a new hyperedge connection; when S(r,s)<0.2, remove the weakly related hyperedge connection; in other cases, keep the initial structure unchanged. Finally, the hyperedge set E after hyperedge enhancement forms the final hypergraph structure; In order to more accurately model the many-to-many interaction relationship between lncRNA and miRNA and improve the feature expression ability between different nodes in the hypergraph construction process, the present invention uses Gaussian Mixture Model (GMM) to calculate the hyperedge weight matrix. Specifically, first, the fused lncRNA and miRNA features are GMM clustering is performed to capture the potential class distribution information. Subsequently, GMM calculates the probability of each RNA node belonging to different mixture components and uses this as the hyperedge weight matrix W e The initial value of is used to measure the likelihood of different lncRNA–miRNA associations. During GMM training, the expectation maximization (EM) algorithm is used for parameter estimation, and the Bayesian Information Criterion (BIC) is used to automatically select the optimal number of clusters K to ensure good generalization ability of the model.

[0034] The lncRNA-miRNA structure information fusion module specifically performs the following steps: Step d1: Perform graph convolution encoding on the local neighborhood connectivity between lncRNA and miRNA nodes to extract their local structural representation; The first-order neighbor structure formed by a node (lncRNA node) and its first-order neighbor nodes (i.e., miRNA nodes directly connected to it by structural edges) is the local neighborhood.

[0035] In step d2, the adjacency matrix 𝐴 is used to represent the direct connections between lncRNAs and miRNAs. This is treated as a bipartite graph and features are propagated using a graph convolutional neural network (GCN). The adjacency matrix 𝐴 is a graph constructed from known lncRNA-miRNA interactions and is used to model local structures within the GCN. This graph is a bipartite graph whose edges connect only interacting pairs. This structure excludes non-verified pairings, ensuring that the GCN module only enhances features based on highly reliable edge information during propagation, thereby ensuring the stability and accuracy of downstream predictions.

[0036] To stabilize the training process, the adjacency matrix is ​​first normalized and the normalized adjacency matrix is ​​calculated; Where D is the degree matrix of the adjacency matrix, which is used to represent the sparsity of the connection between lncRNA and miRNA; In the GCN layer, graph convolution operation is performed based on the normalized adjacency matrix. The specific calculation process is shown in the formula: in, is the initial embedded feature of lncRNA and miRNA, are the learnable parameters for GCN training, is the nonlinear activation function ReLu, and Z represents the local topological features of lncRNA and miRNA after transmission through GCN; Step d3: Use a hypergraph convolutional neural network (HGCN) to model the hypergraph structure (high-order multi-node interaction structure) between lncRNA and miRNA to further extract hypergraph information features. Let the hyperedge adjacency matrix of the hypergraph structure be H, which is used to represent the connection relationship between each node and the high-order hyperedge in the hypergraph structure. Next, set the trainable parameter matrix for the hyperedge weights of lncRNA and miRNA nodes respectively. and , which are used to calculate the hyperedge weights of lncRNA and miRNA nodes respectively. and Together they form the overall edge weight matrix W e , which is used for edge weight adjustment in the subsequent hypergraph convolution information propagation process. The initial feature embeddings of lncRNA and miRNA are obtained respectively. as well as , 、 That is, the unified vector representation of hyperedges.

[0037] During the HGCN propagation process, the hypergraph aggregates the features between lncRNA and miRNA through hyperedges to obtain the aggregated features. , and feature propagation is performed based on hyperedge weights. First, the adjustment of hyperedge weights directly affects the hypergraph Laplacian matrix The calculation of , which in turn affects the information propagation process of HGCN, enabling it to more accurately capture the lncRNA-miRNA interaction pattern when updating features. The calculation process is shown in the formula: Where H is the hyperedge adjacency matrix, and are the node degree matrix and the hyperedge degree matrix respectively, It is the node feature representation obtained after HGCN propagation, and BatchNorm is used for feature normalization.

[0038] Step d4: Introduce cross-view contrast learning mechanism to compare the local structural features Z extracted by GCN and HGCN and the local structural features Z extracted by HGCN perspective. , use cosine similarity to calculate the similarity S(Z, ), and construct the contrast loss function , the specific calculation process is shown in the formula: in, To control the temperature parameter of the contrast loss, Represents the negative sample set.

[0039] Step d5: Introduce a gating mechanism to dynamically fuse the GCN and HGCN representations to adapt to the differences in the dependence of different nodes on local or global information. Define the gating factor g, and the calculation process is as follows: in, is the learnable weight matrix, is the bias term, Represents vector concatenation operation, is the Sigmoid activation function, so that the gating factor g ranges from 0 to 1.

[0040] Based on the calculation results of the gating factor, the final fusion feature representation As shown in the formula: in, The representative representation is a fusion of multi-source features such as k-mer, CTD, Doc2Vec and Role2Vec, and the RNA node is finally embedded through the weighted aggregation of the attention mechanism. It is a hyperparameter used to balance the contribution of sequence features to the fusion results.

[0041] The relationship prediction module specifically performs the following steps:

[0042] Step e1: The fusion feature obtained in step d5 is represented as Input to the KAN classifier to predict the interaction relationship; Step e2, KAN classifier outputs predicted probability value , represents the probability that the current lncRNA–miRNA pair is associated. Its calculation formula is: In step e3, we construct a training objective function to optimize the model output through supervised learning. We use the cross entropy loss function and the L2 regularization loss to form the total loss function 𝐿, which is expressed as follows: in, and is a hyperparameter used to balance the weights of the two loss terms, represents the cross entropy loss, represents the L2 regularization loss.

[0043] Example:

[0044] The lncRNA-miRNA association prediction method based on GCN and contrastive learning described in this embodiment includes the following steps: S1: Collect lncRNA and miRNA sequence data, lncRNA and miRNA association data, preprocess the data, and obtain the initial data set of the model.

[0045] In this embodiment, three parts of data need to be obtained: S11: Obtain lncRNA sequence data; the present invention collects lncRNA sequence data from several classic databases, including LNCipedia2 (LNCipedia) and GENCODE3 (GENCODE - Home page (gencodegenes.org)), and obtains sequence data for a total of 1786 lncRNAs.

[0046] S12: Obtaining miRNA sequence data; Based on previous work, the present invention collected miRNA sequence data from the GENCODE-v38 (GENCODE - Human Release 38 (gencodegenes.org) database, and obtained sequence data for a total of 396 miRNAs.

[0047] S13: Obtain lncRNA and miRNA association data; obtain LMI data from the lncRNASNP v3 (hzau.edu.cn) database. Then, filter the obtained LMIs based on the 1786 lncRNAs collected in S11 and the 396 miRNAs collected in S12. The resulting LMI data contains 17,369 interaction pairs, with a dataset sparsity of 93.31%. All known LMIs are considered positive samples, and the remaining unknown LMIs are considered negative samples. S14: Data preprocessing: First, we replace base T with U in all RNA sequences. Next, we remove duplicate lncRNA-miRNA interaction pairs and filter out lncRNAs and miRNAs with unavailable sequences. After removing duplicate LMIs and filtering out lncRNAs / miRNAs with unavailable sequences, we construct a balanced dataset containing 24,990 lncRNA and miRNA association data, including 12,495 positive and 12,495 negative samples. S2: Kmer count features, Doc2Vec semantic features, CTD structural features, and Role2Vec graph embedding features are extracted from the lncRNA and miRNA sequences collected in S1, and weighted fusion is performed using the attention mechanism.

[0048] S21: Apply k-mer counting to 1786 lncRNA and 396 miRNA sequences to generate k-mer count vectors, which are used as the initial features of lncRNA and miRNA (dimension = 128); in, is the indicator function, For the summation symbol, is the indicator function, is the k-mer count vector elements.

[0049] S22: The present invention uses the Doc2Vec method to extract the global distributed features of lncRNA and miRNA (corresponding to step b2), wherein each RNA sequence is regarded as a document and the k-mer fragment is regarded as a word in the document. The semantic feature vector is obtained through distributed representation learning. The specific formula is: in, is the output vector used to predict the next k-mer segment, softmax(⋅) is the normalization function, b is the bias term, Is the first A vector representation of k-mer fragments, The first The weight of each k-mer segment, V is the weight of the entire RNA sequence, and D is a vector representing the features of the RNA sequence being processed.

[0050] S23: Extract CTD structural features to encode the physicochemical properties of RNA sequences. CTD features consist of three parts: composition, transition, and distribution, which are used to describe the base class ratio, adjacent relationship, and spatial distribution. , used to calculate nucleotides The proportion in the sequence. Conversion features , used to calculate the frequency of adjacent appearance of different nucleotides. Distribution characteristics , the distribution of specific nucleotides at five key positions in the RNA sequence (0%, 25%, 50%, 75%, and 100%) is counted. To improve the robustness of feature extraction, the present invention uses a sliding window method to calculate features of RNA sequences. The window size w is set and slides across the sequence with a step size s. Each time, the CTD features of the nucleotides in the window are extracted, and their normalized representation in the complete sequence is calculated. Finally, after each RNA is processed through the sliding window, a 30-dimensional structural feature vector is generated for lncRNA-miRNA interaction modeling.

[0051] S24: The Role2Vec algorithm is used to extract the role embedding features of lncRNA and miRNA nodes in the graph structure, reflecting the structural position and functional role information of the nodes in the graph. The specific approach is to calculate the similarity between RNAs based on k-mer, CTD and Doc2Vec, and construct lncRNA-lncRNA and miRNA-miRNA interaction graphs. Then, the Role2Vec algorithm is used to perform graph embedding to obtain the network structure features of RNA. First, to construct the RNA similarity graph, the RNA fragments are calculated. and The similarity score between them is as shown in the formula: in, , , is the feature weighting parameter, satisfying .

[0052] Based on the calculated similarity scores, the local nearest neighbor search method was used to construct the lncRNA-lncRNA and miRNA-miRNA interaction graphs, that is, for each lncRNA (e.g. ), retaining the top k lncRNAs with the highest similarity as interaction nodes, where k=15 was set based on experimental experience. Based on this strategy, highly homologous lncRNAs were connected to , forming a lncRNA-lncRNA interaction graph, and the miRNA-miRNA interaction graph is constructed in a similar way. After generation, the Role2Vec method is used to embed the RNA graph. Perform a random walk on the network to obtain the local neighborhood information of RNA in the network, that is, get the node Random walk sequences on a graph , step length Set to 1 to ensure the capture of local topological information and reduce the information dilution problem caused by long random walk paths. Then, the node embedding is trained by Role2Vec ,Finally, each lncRNA and miRNA is mapped into a 128-dimensional embedding vector to characterize the topological role of RNA in the network.

[0053] S25: To unify the expression space of multi-source heterogeneous features such as K-mer, Doc2Vec, CTD, and Role2Vec, and improve the representation capability after feature fusion, this embodiment introduces a multi-layer perceptron (MLP) and an attention mechanism for weighted feature fusion to obtain a unified node input representation. First, each type of feature vector is input into a set of independent MLP networks for nonlinear transformation, so that it is mapped to the same dimension d. The MLP contains two layers of fully connected networks and uses the ReLU activation function to improve its nonlinear representation capability. The transformed feature vector is recorded as Subsequently, in order to model the importance of different types of features, an attention mechanism is introduced to automatically learn the attention weight of each feature. The formula is as follows: Among them, w is the trainable weight vector, Represents the representation of the i-th category feature after MLP encoding.

[0054] Finally, all features are weighted and summed to obtain a unified fusion feature representation Finally, each lncRNA and miRNA was encoded into a 128-dimensional feature vector, integrating sequence, structure, and topology information.

[0055] S3: To support structural information modeling, this example constructs a lncRNA–miRNA heterogeneous graph and a hypergraph structure for use as graph structure input in the GCN and HGCN modules. This specifically involves the following three steps.

[0056] S31: Based on the known lncRNA-miRNA interaction information, construct a bipartite heterogeneous graph structure. Let the lncRNA node set be , the miRNA node set is . Then construct the adjacency matrix of the graph .in Indicates the presence of lncRNA–miRNA interaction edges.

[0057] S32: Using the high-order interaction information of lncRNA acting on multiple miRNAs simultaneously, a hypergraph structure is constructed. Define hypergraph , where V is the node set composed of lncRNA and miRNA, E is the hyperedge set, and W represents the weight of the hyperedge. Connecting a lncRNA to multiple miRNAs to characterize the miRNA group that the lncRNA may regulate. First, the present invention generates an initial hyperedge matrix based on the known lncRNA-miRNA interaction database. ,Right now and If there is a known interaction between them, a connection is established in the hypergraph, otherwise it is set to 0. Based on the initial hyperedge, in order to further improve the model's ability to learn unlabeled but potentially related lncRNA-miRNA relationships, the present invention uses the RNA fusion feature similarity matrix obtained in the above summary to perform hyperedge enhancement, setting The cosine similarity is used to calculate the similarity score between lncRNA and miRNA, and normalized to the interval [0,1]. The calculation process is shown in the formula: In order to improve the model's ability to express node interaction differences, the hyperedge weight matrix is ​​constructed. In this example, the Gaussian Mixture Model (GMM) is introduced to perform unsupervised clustering of lncRNA and miRNA features. First, the feature vectors of lncRNA and miRNA nodes are fused, and then GMM clustering is used to fit the node embedding to capture its mixed feature space under multimodal distribution. GMM outputs the posterior probability distribution of each node belonging to different Gaussian components, which is denoted as Used as the initial weight matrix of the hyperedge To determine the optimal number of clusters K, the Bayesian Information Criterion (BIC) is used to automatically select the optimal model structure, and finally the hypergraph structure is constructed. Enter the HGCN module.

[0058] S33: Input the aforementioned heterogeneous graph structure and hypergraph structure into GCN and HGCN modules respectively for structural information encoding. Both use the same node feature input matrix, but the graph structure connection method is different. Input to GCN, hypergraph Input to HGCN.

[0059] S4: To fully utilize the lncRNA and miRNA interaction patterns contained in the graph structure, this example uses a dual-channel structural information encoding method to construct GCN and HGCN modules respectively to learn structural features of heterogeneous graphs and hypergraph structures. Specifically, the following steps are included: S41: This step is to build the heterogeneous graph structure As input, a graph convolutional network (GCN) is used to encode the adjacency structure information of lncRNA and miRNA nodes. To improve the training stability, the adjacency matrix A is first symmetrically normalized and the normalized adjacency matrix is ​​calculated. , as shown in the formula: Where D is the degree matrix of the adjacency matrix, and then the standard graph convolution propagation rule is used to update the node features. The calculation formula is as follows: in, is the initial embedding feature of lncRNA and miRNA, W is the learnable parameter for GCN training, is the nonlinear activation function ReLu, and Z represents the local topological features of lncRNA and miRNA after transmission through GCN.

[0060] S42: To capture the complex high-order interactions between lncRNA and multiple miRNAs, a trainable hyperedge weight optimization method is used. This allows hyperedges to not only serve as fixed connection structures, but also dynamically adjust weights during training, thereby optimizing the information propagation process. The hyperedge weights of lncRNA and miRNA nodes are modeled separately. and are two sets of trainable parameter matrices, which are used to calculate the hyperedge weights of lncRNA and miRNA nodes, respectively, to obtain the initial feature embeddings of lncRNA and miRNA, respectively. as well as During the HGCN propagation process, the hypergraph aggregates the features between lncRNA and miRNA through hyperedges and propagates the features based on the hyperedge weights. First, the adjustment of the hyperedge weights directly affects the hypergraph Laplacian matrix. The calculation of , which in turn affects the information propagation process of HGCN, enabling it to more accurately capture the lncRNA-miRNA interaction pattern when updating features. The calculation process is shown in the formula: Where H is the hyperedge adjacency matrix, and are the node degree matrix and the hyperedge degree matrix respectively, It is the node feature representation obtained after HGCN propagation, and BatchNorm is used for feature normalization.

[0061] S43: To improve the feature consistency between the local structure representation (GCN) and the high-level structure representation (HGCN), this embodiment introduces a cross-view contrast learning mechanism to minimize the difference in the representation of the same node under the two structural perspectives. Based on the lncRNA feature Z extracted from the GCN perspective and the lncRNA feature Z extracted from the HGCN perspective obtained above, , use cosine similarity to calculate the similarity between them , and construct the contrast loss function , so that the feature vectors of the same lncRNA in the GCN and HGCN perspectives are as close as possible, while the feature vectors of different lncRNAs remain distinguishable. The specific calculation process for each node i is shown in the formula: in, To control the temperature parameter of the contrast loss, Represents the negative sample set.

[0062] The final contrast loss is the average of all node pair losses: By minimizing , the model can learn to express consistent node embeddings under two structural perspectives, thereby improving the stability and robustness of the structural representation.

[0063] S44: After completing the cross-view alignment, in order to achieve effective integration of the two structural pathways, this embodiment introduces a gating mechanism to perform weighted fusion of the GCN and HGCN output results to generate a unified node representation. This gating mechanism allows each node to automatically allocate the fusion ratio from the GCN or HGCN channel based on its structural characteristics, solving the problem of expressing node preferences under different structures, and helping to improve the discriminability of the fusion expression and the generalization ability of the model. Let g be the gating factor, and its calculation process is shown in the formula in, is the learnable weight matrix, is the bias term, Represents vector concatenation operation, is the Sigmoid activation function, so that the gating factor g ranges from 0 to 1.

[0064] Based on the calculation results of the gating factor, the final fusion feature representation As shown in the formula: in, represents the initial sequence characteristics, It is a hyperparameter used to balance the contribution of sequence features to the fusion results.

[0065] S5: To achieve accurate prediction of potential interactions between lncRNA and miRNA, this example constructs matching feature vectors based on structural information encoding and fusion, uses the KAN classifier with B-spline activation function for prediction, and introduces a dual loss function to jointly optimize model training.

[0066] S51: This step combines the node feature vector obtained by S44 with the structure As the model input, the KAN classifier is used to predict whether there is an interaction relationship between lncRNA-miRNA pairs. The KAN model introduces a learnable B-spline activation function to replace the traditional ReLU activation method, which has a more flexible nonlinear mapping capability and can improve the model's expressiveness and training stability under complex data distribution. For any pair of lncRNA and miRNA, the interaction probability prediction value is .

[0067] S52: To further improve the model's prediction accuracy and generalization ability, this step designs a joint loss function for model optimization training. The loss function includes cross entropy loss and L2 regularization term, which together constitute the total loss function L, expressed as follows: in, and is a hyperparameter used to balance the weights of the two loss terms, represents the cross entropy loss, represents the L2 regularization loss.

[0068] S6: Test and evaluate the model performance through various experiments on the dataset.

[0069] In order to test and evaluate the performance of this application, this application was evaluated on a dataset and used AUC (area under the ROC curve), AUPR (area under the PR curve), precision, recall, F1 score, etc. as evaluation indicators. AUROC is the area under the ROC (receiver operating characteristic) curve, with the horizontal axis of the ROC curve being FP and the vertical axis being TP. AUPR is the area under the PR curve, with the horizontal axis of the PR curve being Precision and the vertical axis being Recall. The specific calculation formula is as follows: Where TP represents the number of correctly predicted LMIs, FP represents the number of incorrectly predicted LMIs, TN represents the number of correctly predicted non-LMIs, and FN represents the number of incorrectly predicted non-LMIs.

[0070] The ROC and AUPR prediction performance comparisons were performed on the dataset with other comparison methods using a five-fold cross validation experiment. Figure 2 As shown, Figure 2 (a) and (b) correspond to the ROC and AUPR curves, respectively.

[0071] In order to verify the predictive performance of the method of the present invention and to comprehensively evaluate the stability and robustness of the HCLMP model under different data partitioning conditions, the present invention adopts a five-fold cross-validation strategy to conduct experiments. Specifically, the original data set is evenly divided into five subsets, one of which is used as the test set in turn, and the remaining four subsets are used to train the model. By averaging the performance indicators obtained from the five rounds of experiments, the overall performance of the model under different training-testing partitions is evaluated, thereby verifying the stability and generalization ability of its predictive ability. The experimental results are as follows Figure 2As shown. The present invention has competitive advantages in terms of overall prediction accuracy, precision and recall, and can effectively balance the errors in positive and negative sample predictions to ensure high-quality prediction results. In particular, in terms of the F1 indicator, the present invention also achieved the highest score, which further verifies the method's ability to balance precision and recall, helps to reduce false positives and false negatives, and improves the practical applicability of the model. At the same time, compared with other methods, HCLMP can still maintain high stability and generalization ability in complex biological backgrounds, and is suitable for large-scale lncRNA-miRNA interaction prediction tasks. Table 1 Comparison of prediction performance of other indicators in a five-fold cross-validation experiment on the dataset with other methods.

[0072] Table 1 Results of five-fold cross validation of different methods on the benchmark dataset The stability and adaptability of the proposed method under key feature combination settings were evaluated, and further multi-view feature effectiveness experiments were conducted. The experiment was conducted on a standard dataset, keeping all other parameters unchanged, and setting different feature combinations to evaluate the performance trends of the model under each combination. Specifically, this embodiment set four feature types, including K-mer (K), Doc2Vec (D), CTD (C), and Role2Vec (R). Six input settings were constructed by combining them, namely: To further explore the practical role of the hierarchical attention mechanism introduced in this invention in the multi-layer graph convolution architecture, box plots of the attention weights of each GCN layer were plotted. By statistically analyzing the distribution of attention weights of different GCN layers during model training, the relative contribution and stability of each layer in the feature aggregation process can be intuitively reflected. The experimental results show that the model exhibits higher weight concentration in shallower layers (such as the first layer), indicating that low-level neighborhood information has a stronger influence on the basic feature expression; while in deeper structures, the weight distribution is relatively dispersed, indicating that the model can flexibly allocate weights for high-order neighborhood information, achieving hierarchical expression and adaptive aggregation of information. The visualization results verify that the hierarchical attention mechanism introduced by this invention has good interpretability and helps improve the model's ability to model key information in multi-layer graph structures. Figure 3 And shown in Table 2.

[0073] Table 2 Experimental results on the effectiveness of multi-source feature combination Methods ACC PRE REC F1 K 0.9589 0.9482 0.9517 0.9479 K+R 0.9593 0.9492 0.9617 0.9576 K+D+R 0.9612 0.9517 0.9673 0.9602 C+K+D+R 0.9620 0.9520 0.9674 0.9622 In order to verify the effect of the cross-perspective contrastive learning mechanism introduced in the method of the present invention on the quality of expression of model fusion features, this section designs a comparative experiment on a standard dataset for verification. Specifically, under the premise of keeping the graph structure construction method consistent with the node features, a variant model HCLMP-w / o-CL with the contrastive learning module removed is constructed and compared with the complete model HCLMP. During the experiment, a five-fold cross-validation method was used to calculate the average performance of the two models under indicators such as AUC, AUPR, ACC, and F1, respectively, to evaluate the impact of the contrastive learning mechanism on the model's discrimination ability and stability. The experimental results are shown in the figure below. Figure 4 As shown in Table 3, Figure 4 (a) ROC curve for the comparison between HCLMP and HCLMP-w / o-CL, (b) PR curve for the comparison between HCLMP and HCLMP-w / o-CL.

[0074] Table 3 Experimental results on the effectiveness of comparative learning mechanisms Methods ACC PRE REC F1 HCLMP 0.9620 0.9570 0.9674 0.9622 HCLMP w / o CL 0.9120 0.8778 0.9085 0.8969 The results show that the complete HCLMP model outperforms the control group across all evaluation metrics, with particularly significant improvements in metrics such as AUPR and F1-score, which reflect the model's predictive ability for minority class samples. This demonstrates that the contrastive learning mechanism helps improve the consistency and expression quality of multi-view feature fusion. Furthermore, this mechanism can compensate for inconsistent feature distributions, enhancing the discriminative power and robustness of the fused representation.

[0075] To verify the effectiveness of the method of the present invention in the feature extraction and fusion process, and to observe the ability of different feature processing stages to distinguish lncRNA from miRNA, this example uses the t-SNE (t-distributed Stochastic Neighbor Embedding) method to reduce the dimension of the feature representation of each stage and visualize its distribution in two-dimensional space, as shown in the following figure: Figure 5 As shown, Figure 5(a) and (b) correspond to the initial feature visualization results and the visualization results after feature fusion, respectively. The specific experimental process is as follows: First, the K-mer, Doc2Vec, CTD, and Role2Vec features from the original feature stage are embedded and extracted as the initial node representation input; then, the features are subjected to dimensionality reduction mapping during the fusion stage, the post-fusion MLP encoding stage, and the final post-fusion output stage. The experimental results show that in the initial stage, there is significant overlap between samples of different categories, and the category distinction is weak. After the MLP fusion stage, the cluster boundaries between different categories gradually become clearer, indicating that redundant information between features is effectively compressed and key features are preserved. In the final fusion output stage, the sample distribution of lncRNA and miRNA is more concentrated, and the category separation boundaries are more distinct, verifying the good discriminatory ability of the proposed method during the feature fusion process. The above results indicate that the multi-perspective feature extraction and fusion strategy proposed in this paper can effectively guide the high-dimensional feature vectors of lncRNA and miRNA to present a clear category clustering trend in low-dimensional space, verifying the model's discriminative ability and expression rationality from the distribution structure level, which will help improve the representation accuracy and result credibility in subsequent LMI prediction tasks.

[0076] To validate the applicability and practical value of the present method for the task of discovering unknown associations, this example further evaluated the model's predictive ability in identifying potential LMIs (lncRNA-miRNA interactions) in real-world scenarios. The specific method involved: Based on the constructed HCLMP model, a comprehensive prediction and scoring of unannotated lncRNA-miRNA pairs in a standard dataset was performed, obtaining association probabilities for all potential interaction pairs and sorting the predictions in descending order. The top 10 lncRNA-miRNA pairs with the highest prediction scores were then selected as candidate results and cross-referenced with public literature and databases for validation. The results are shown in Table 4. Of the top 10 predicted pairs, 9 have been experimentally validated in databases such as miRTarBase and LncBase, or in relevant literature. The remaining pair has not yet been clearly annotated, but its prediction score is high and therefore worthy of further research. The experimental results show that the method of the present invention has strong practical applicability and scalability, and can accurately identify lncRNA-miRNA associations with potential functional significance in real biological environments without annotations, providing feasible candidate targets for subsequent experimental research and mechanism exploration.

[0077] Table 4 Top 10 lncRNA–miRNA interactions predicted by HCLMP and their rankings Rank LncRNA name MiRNA name Confrmation 1 NONHSAT017458.2 hsa-miR-130b-3p Confrmed 2 NONHSAT017461.2 hsa-miR-20a-5p Confrmed 3 NONHSAT017461.2 hsa-miR-20b-5p Confrmed 4 NONHSAT017526.2 hsa-miR-30d-5p Unconfrmed 5 NONHSAT016903.2 hsa-miR-424-5p Confrmed 6 NONHSAT017774.2 hsa-miR-195-5p Confrmed 7 NONHSAT017458.2 hsa-miR-138-5p Confrmed 8 NONHSAT017458.2 hsa-miR-130a-3p Confrmed 9 NONHSAT017772.2 hsa-miR-424-5p Confrmed 10 NONHSAT017458.2 hsa-miR-148b-3p Confrmed The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A lncRNA-miRNA association prediction method integrating a hypergraph perspective, characterized by: The following steps are involved: LncRNA and miRNA sequence data were acquired and preprocessed. K-mer frequency features, Doc2Vec semantic features, and CTD structural features were then obtained for each RNA. A weighted undirected graph structure was constructed based on the similarity relationships between RNA fragments. Each node in the graph corresponds to an RNA fragment, and the edge weights represent the cosine similarity of the nodes. The Role2Vec model was introduced into this graph for structure-aware embedding to obtain Role2Vec topological features. Based on K-mer frequency features, Doc2Vec semantic features, CTD structural features and Role2Vec topological features, multi-level feature fusion is performed to obtain a unified fusion feature representation for each RNA , the fusion features Unified representation of RNA , as the initial embedded features of lncRNA and miRNA; The initial embedded features of lncRNA and miRNA are fed into the graph convolutional network GCN to obtain features , the initial embedding features of lncRNA and miRNA are fed into the hypergraph convolutional network HGCN to obtain ; Calculate the feature Z and The similarity between ;based on Perform comparative learning optimization; then, and Perform dynamic fusion and obtain the final fusion feature representation based on the gating factor : in, It is a hyperparameter used to balance the contribution of sequence features to the fusion results; Based on the final fusion feature representation Enable lncRNA–miRNA association prediction.

2. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 1, characterized in that: The cosine similarity of the nodes corresponding to the edge weights in the weighted undirected graph structure is as follows: in, , , is the feature weighting parameter, satisfying is the cosine similarity of RNA fragments in k-mer frequency features, is the cosine similarity of RNA fragments in CTD structural features, is the cosine similarity of RNA fragments in Doc2Vec semantic features.

3. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 1, characterized in that: Based on K-mer frequency features, Doc2Vec semantic features, CTD structural features and Role2Vec topological features, a multi-layer perceptron combined with an attention mechanism is used to perform multi-level feature fusion.

4. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 1, characterized in that: The construction process of the hypergraph convolutional network HGCN is as follows: Step c1: lncRNA and miRNA nodes are considered as two types of heterogeneous nodes in the hypergraph. If a lncRNA node has an interaction relationship with one or more miRNA nodes, a hyperedge is constructed to connect the lncRNA node with the corresponding multiple miRNA nodes to form a lncRNA-dominated hypergraph structure. ; For each hyperedge connected to multiple nodes, the characteristics of each node Add the sum by dimension and divide by the number of nodes to get the unified vector representation of the hyperedge; and If there is a known interaction between them, a connection is established in the hypergraph; Step c2: Based on the lncRNA and miRNA feature vectors output by the feature extraction module, set For the fused feature vector, the cosine similarity between similar nodes is calculated and normalized to the interval [0,1]. When the normalized value is greater than the first threshold, a new hyperedge connection is added. When the normalized value is less than the second threshold, the hyperedge connection is removed. In other cases, the initial structure remains unchanged. At the same time, the hyperedge weight matrix W is determined. e .

5. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 4, characterized in that: The formula for calculating the cosine similarity between similar nodes in step c2 and normalizing it to the interval [0,1] is as follows: in, 、 Represents two nodes of the same type .

6. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 4, characterized in that: Step c2 determines the hyperedge weight matrix W e The Gaussian mixture model GMM is used to calculate the hyperedge weight matrix in the process, including: First, the fusion lncRNA and miRNA characteristics Perform GMM clustering; then, GMM calculates the probability of each RNA node belonging to different mixed components and uses this as the hyperedge weight matrix W e .

7. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 1, characterized in that: The initial embedded features of lncRNA and miRNA are fed into the graph convolutional network GCN to obtain features The process includes: Graph convolutional encoding is performed on the local neighborhood connectivity between lncRNA and miRNA nodes to extract their local structural representation. The direct connectivity between lncRNA and miRNA is represented by the adjacency matrix 𝐴, which is treated as a bipartite graph structure. Graph convolutional neural network (GCN) is used for feature propagation. During this process, the adjacency matrix is ​​normalized and the normalized adjacency matrix is ​​calculated. Where D is the degree matrix of the adjacency matrix, which is used to represent the sparsity of the connection between lncRNA and miRNA; In the GCN layer, graph convolution operation is performed based on the normalized adjacency matrix: in, is the initial embedded feature of lncRNA and miRNA, To learn the parameters, is the nonlinear activation function ReLu, and Z represents the local topological features of lncRNA and miRNA after transmission through GCN.

8. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 1, characterized in that: The initial embedding features of lncRNA and miRNA are fed into the hypergraph convolutional network HGCN to obtain The process includes: The hypergraph structure between lncRNA and miRNA is modeled using the hypergraph convolutional neural network HGCN to further extract hypergraph information features; trainable parameter matrices are set for the hyperedge weights of lncRNA and miRNA nodes respectively. and , which are used to calculate the hyperedge weights of lncRNA and miRNA nodes respectively; among them, initially and Together they form the overall edge weight matrix W e , used for edge weight adjustment in the subsequent hypergraph convolution information propagation process; the initial feature embedding of lncRNA and miRNA is obtained respectively as well as , 、 That is, the unified vector representation of hyperedges; During the HGCN propagation process, the hypergraph aggregates the features between lncRNA and miRNA through hyperedges to obtain the aggregated features. ; Feature propagation based on hyperedge weights, the formula is as follows: Where H is the hyperedge adjacency matrix, and are the node degree matrix and the hyperedge degree matrix respectively, It is the node feature representation obtained after HGCN propagation, and BatchNorm is used for feature normalization.

9. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to any one of claims 1 to 8, characterized in that: based on Contrastive loss function in the contrastive learning optimization process as follows: in, To control the temperature parameter of the contrast loss, Represents the negative sample set.

10. The lncRNA-miRNA association prediction method integrating a hypergraph perspective according to claim 9, characterized in that: The gating factor ,in, is the weight matrix, is the bias term, Represents vector concatenation operation, is the Sigmoid activation function.

Images