IncRNA-miRNA association prediction system and method based on GCN and comparative learning

Abstract

The invention discloses an lncRNA-miRNA association prediction system and method based on GCN and comparative learning, and belongs to the technical field of biological information. The method aims at solving the problem that an existing lncRNA-miRNA association prediction method is limited in node expression capacity. According to the method, k-mer counting vectors are generated on the basis of all lncRNA and miRNA sequences, a Doc2Vec model is further introduced into k-mer sequences to obtain sequence semantic features, and then a combined feature vector is obtained on the basis of the counting vectors and the sequence semantic features; according to the method, initial representation of nodes in a graph is obtained based on an lncRNA and miRNA heterogeneous graph and a joint feature vector, then neighbor node information aggregation is carried out by using a node-level attention mechanism, information aggregation is carried out by using a hierarchical attention mechanism to obtain final node features of lncRNA and miRNA, weight fusion is carried out to obtain LMI features, and a prediction score is obtained through prediction.

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 RNAs (NRAs) are a class of functional RNA molecules transcribed from the genome. Although they cannot encode proteins, they play key roles in gene regulation, cellular metabolism, and the development and progression of diseases such as cancer. Long non-coding RNAs (lncRNAs) and microRNAs (miRNAs) are two important types of non-coding RNAs (ncRNAs) that have been shown to be involved in numerous biological processes, such as gene regulation, cellular metabolism, and cancer development. Furthermore, there is ample evidence that lncRNAs and miRNAs can regulate each other. LncRNAs primarily regulate miRNAs by acting as miRNA precursors, competing endogenous RNAs (ceRNAs), and miRNA sponges. MiRNAs can also target lncRNAs and promote their degradation. Therefore, lncRNAs and miRNAs can form complex regulatory networks that control the physiological processes of many human diseases. Existing studies have demonstrated a complex regulatory relationship between lncRNAs and miRNAs.

[0003] However, validating these interactions through traditional biological experiments is time-consuming and costly. Therefore, computational methods offer an efficient, large-scale screening tool, making lncRNA-miRNA interaction (LMI) prediction particularly important. Existing LMI prediction methods can be broadly categorized into three categories: similarity-based, multi-view-based, and graph-based. First, lncRNA-miRNA association data are generally sparse in practical applications, leaving a large number of potential interactions unobserved. Most existing methods overly rely on known interaction samples and lack the ability to effectively model underlying interaction patterns, resulting in limited generalization performance. Second, when constructing heterogeneous graphs, existing methods often treat lncRNA and miRNA nodes as structural equivalents, failing to fully consider their essential differences in functional attributes, biological characteristics, and expression patterns. This results in excessive mixing of node interaction features during modeling, weakening semantic discriminability and compromising prediction accuracy. Furthermore, many methods utilize only raw sequence information during feature extraction, neglecting the multi-scale fusion of local frequency features and contextual semantic information, resulting in insufficient node representation capabilities. Furthermore, during the multi-layer propagation of graph neural networks, the importance of neighborhood information at different levels is not modeled differently. This leads to oversmoothing of high-order neighborhood features during the accumulation process, obscuring the original local structural information and making it difficult to fully explore deep-level interaction patterns. These issues have, to varying degrees, limited the breadth of application and depth of prediction of existing methods in complex biological networks. Summary of the Invention

[0004] The present invention aims to solve the problem that existing lncRNA-miRNA association prediction methods have limited node representation capabilities, which in turn affects the prediction accuracy.

[0005] A lncRNA-miRNA association prediction system based on GCN and contrastive learning includes a prediction unit that performs lncRNA-miRNA association prediction based on a prediction network model; the prediction network model includes a lncRNA and miRNA feature extraction module, a neighbor information aggregation module, and a lncRNA-miRNA potential relationship prediction module;

[0006] The lncRNA and miRNA feature extraction module applies k-mer counting technology to all lncRNA and miRNA sequences to generate k-mer count vectors, introduces the Doc2Vec model into the k-mer sequences to obtain sequence semantic features, and then obtains a joint feature vector based on the count vectors and sequence semantic features;

[0007] The neighbor information aggregation module obtains the initial representation of the nodes in the graph based on the lncRNA and miRNA heterogeneous graph and the joint feature vector, then uses the node-level attention mechanism to aggregate the neighbor node information, and uses the hierarchical attention mechanism to aggregate the information to obtain the final node features of the lncRNA and the final node features of the miRNA;

[0008] The lncRNA-miRNA potential relationship prediction module: weighted fusion of the final node features of lncRNA and miRNA obtained by the neighbor information aggregation module to obtain the LMI feature f LMI , make predictions and get prediction scores Among them, MLP is a multi-layer perceptron, Sigmoid is the activation function, is the prediction score of the LMI.

[0009] Furthermore, the prediction network model is pre-trained, and during the training process, the network parameters are trained by constructing positive and negative sample pairs for comparison. The process of constructing the positive and negative sample pairs includes:

[0010] The LMI feature f is obtained in the lncRNA-miRNA potential relationship prediction module LMI Finally, the training samples were divided into positive and negative groups according to the similarity between LMI features. The lncRNA-miRNA pairs whose feature vector cosine similarity was higher than the first threshold were classified as the positive group, and the lncRNA-miRNA pairs whose similarity was lower than the second threshold were classified as the negative group.

[0011] Furthermore, the loss function used in the training process of the prediction network model is as follows:

[0012]

[0013] in, Represents the LMI training sample set Each pair (lcn, mi) represents a set of known lncRNA–miRNA interaction samples; p lnc,mi is the true label of the interaction relationship, where 1 indicates the existence of interaction and 0 indicates the absence of interaction; Predicted probability.

[0014] Furthermore, the processing process of the lncRNA and miRNA feature extraction module includes:

[0015] Step b1: Each RNA sequence S = s1s2s3...s n Divide into overlapping k-mer segments of fixed length k {k-mer i}, where k-mer i =si s i+1 ...s i+k-1 ;

[0016] Count the occurrences of all k-mer segments and generate a k-mer count vector v:

[0017]

[0018] Among them, ∑ is 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.

[0019] Step b2: Treat each RNA sequence as a document and the k-mer fragment as a word in the document. Train the Doc2vec model using the distributed memory (DM) method to generate semantic features for each lncRNA and miRNA sequence.

[0020] The sequence semantic features obtained by the Doc2Vec model are concatenated or fused with the k-mer count vector v in step b1 to form a joint feature vector containing local frequency and global semantics.

[0021] Furthermore, the processing process of the neighbor information aggregation module includes:

[0022] Step c1: construct a lncRNA and miRNA heterogeneity graph using the positive samples in the known LMI, and use the joint feature vectors corresponding to the lncRNA and miRNA obtained by training in step b2 as the initial representation of the nodes in the graph;

[0023] Step c2: Use the node-level attention mechanism to aggregate neighbor node information, including the following steps:

[0024] Define node v i and its neighbor node v j The hidden representations at layer l-1 are and Then calculate the attention coefficient between nodes i and j at layer l Where a is the weight vector, σ(·) represents the activation function; ‖ represents concatenation; when l is equal to 1, it is the initial representation of the node;

[0025] Using the softmax function Normalize and get v i Attention coefficient to its neighbor nodes By the corresponding attention coefficient For the target node vi The neighbor nodes of represent weighted summation to complete the aggregation of node features;

[0026] Step c3: Use the hierarchical attention mechanism to aggregate information, including the following steps:

[0027] For node v i , the nodes in each layer are represented as l represents the layer number of the graph neural network, and L is the total number of layers of the graph neural network; the importance coefficient corresponding to the lth layer is as follows:

[0028]

[0029] in, is the node v in the lth layer i The attention score, (q (l) ) T is the transpose of the weight vector of the lth layer, W (l) is the weight matrix of the lth layer, is the node v in the lth layer i The transpose of the feature representation, b (l) is the bias vector of the lth layer;

[0030] Based on Determine the node v in layer l i The normalization coefficient of

[0031] Multi-layer feature fusion is performed on lncRNA nodes and miRNA nodes respectively, and the final node feature z i The weighted sum is obtained by the following formula:

[0032]

[0033] in, is the node v in layer l i The feature representation of is the node v in the lth layer i The normalized attention coefficient of .

[0034] Furthermore, the system also includes a data acquisition and preprocessing unit: used to acquire lncRNA and miRNA sequence data, as well as lncRNA and miRNA association data, and perform data preprocessing, wherein the preprocessing process includes the step of replacing the base T in all RNA sequences with U.

[0035] A lncRNA-miRNA association prediction method based on GCN and contrastive learning, including:

[0036] First, lncRNA and miRNA sequence data, as well as lncRNA and miRNA association data, are obtained; then, a prediction network model is used to predict lncRNA-miRNA associations; the prediction network model includes a lncRNA and miRNA feature extraction module, a neighbor information aggregation module, and a lncRNA-miRNA potential relationship prediction module;

[0037] The lncRNA and miRNA feature extraction module applies k-mer counting technology to all lncRNA and miRNA sequences to generate k-mer count vectors, introduces the Doc2Vec model into the k-mer sequences to obtain sequence semantic features, and then obtains a joint feature vector based on the count vectors and sequence semantic features;

[0038] The neighbor information aggregation module obtains the initial representation of the nodes in the graph based on the lncRNA and miRNA heterogeneous graph and the joint feature vector, then uses the node-level attention mechanism to aggregate the neighbor node information, and uses the hierarchical attention mechanism to aggregate the information to obtain the final node features of the lncRNA and the final node features of the miRNA;

[0039] The lncRNA-miRNA potential relationship prediction module: weighted fusion of the final node features of lncRNA and miRNA obtained by the neighbor information aggregation module to obtain the LMI feature f LMI , make predictions and get prediction scores Among them, MLP is a multi-layer perceptron, Sigmoid is the activation function, is the prediction score of the LMI.

[0040] Furthermore, the processing process of the lncRNA and miRNA feature extraction module includes:

[0041] Step b1: Each RNA sequence S = s1s2s3...s n Divide into overlapping k-mer segments of fixed length k {k-mer i}, where k-mer i =s i s i+1 ...s i+k-1 ;

[0042] Count the occurrences of all k-mer segments and generate a k-mer count vector v:

[0043]

[0044] Among them, ∑ is 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.

[0045] Step b2: Treat each RNA sequence as a document and the k-mer fragment as a word in the document. Train the Doc2vec model using the distributed memory (DM) method to generate semantic features for each lncRNA and miRNA sequence.

[0046] The sequence semantic features obtained by the Doc2Vec model are concatenated or fused with the k-mer count vector v in step b1 to form a joint feature vector containing local frequency and global semantics.

[0047] Furthermore, the processing process of the neighbor information aggregation module includes:

[0048] Step c1: construct a lncRNA and miRNA heterogeneity graph using the positive samples in the known LMI, and use the joint feature vectors corresponding to the lncRNA and miRNA obtained by training in step b2 as the initial representation of the nodes in the graph;

[0049] Step c2: Use the node-level attention mechanism to aggregate neighbor node information, including the following steps:

[0050] Define node v i and its neighbor node v j The hidden representations at layer l-1 are and Then calculate the attention coefficient between nodes i and j at layer l Where a is the weight vector, σ(·) represents the activation function; ‖ represents concatenation; when l is equal to 1, it is the initial representation of the node;

[0051] Using the softmax function Normalize and get v i Attention coefficient to its neighbor nodes By the corresponding attention coefficient For the target node v i The neighbor nodes of represent weighted summation to complete the aggregation of node features;

[0052] Step c3: Use the hierarchical attention mechanism to aggregate information, including the following steps:

[0053] For node v i , the nodes in each layer are represented as l represents the layer number of the graph neural network, and L is the total number of layers of the graph neural network; the importance coefficient corresponding to the lth layer is as follows:

[0054]

[0055] in, is the node v in the lth layer i The attention score, (q (l) ) T is the transpose of the weight vector of the lth layer, W (l) is the weight matrix of the lth layer, is the node v in the lth layer i The transpose of the feature representation, b (l) is the bias vector of the lth layer;

[0056] Based on Determine the node v in layer l i The normalization coefficient of

[0057] Multi-layer feature fusion is performed on lncRNA nodes and miRNA nodes respectively, and the final node feature z i The weighted sum is obtained by the following formula:

[0058]

[0059] in, is the node v in layer l i The feature representation of is the node v in the lth layer i The normalized attention coefficient of .

[0060] Furthermore, after obtaining lncRNA and miRNA sequence data, as well as lncRNA and miRNA association data, and before using the prediction network model for prediction, the obtained data needs to be preprocessed. The preprocessing process includes the step of replacing the base T in all RNA sequences with U.

[0061] Beneficial effects:

[0062] The present invention uses the Doc2vec method to pre-train RNA sequences, making full use of sequence information, thereby capturing local and global features of RNA sequences and providing high-quality representation for subsequent LMI prediction. Node-level and hierarchical attention mechanisms are used to finely model different neighbor nodes and multi-level structural information in heterogeneous graphs, effectively enhancing the structural expression capability. An InfoNCE-based comparative learning optimization strategy is also introduced to improve the discriminability and robustness of the embedding space, thereby improving the prediction accuracy while enhancing the stability and generalization ability of the model in sparse data scenarios. The prediction method of the present invention makes full use of known lncRNA and miRNA data to improve the accuracy of lncRNA-miRNA association prediction. A large number of experiments conducted on benchmark datasets have verified that the present invention is superior to several of the most advanced LMI prediction methods under different experimental settings, and the present invention has good robustness and generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Flowchart for lncRNA–miRNA association prediction;

[0064] Figure 2 Comparison curves of ROC and AUPR prediction performance in a five-fold cross-validation experiment on the dataset with other comparison methods;

[0065] Figure 3 The impact of different GCN layers on the model prediction performance on the dataset;

[0066] Figure 4 The impact of lncRNA and miRNA features of different dimensions on the model prediction performance in the dataset;

[0067] Figure 5 The impact of different numbers of multi-scale attention heads on the model prediction performance on the dataset;

[0068] Figure 6 Results of visualizing the level-by-level attention weights on the dataset. DETAILED DESCRIPTION

[0069] The present invention is mainly aimed at solving the following problems existing in existing lncRNA-miRNA association prediction methods:

[0070] On the one hand, existing methods usually only utilize raw sequence information in the feature extraction stage, ignoring multi-scale feature fusion and contextual semantic expression, resulting in limited node representation capabilities;

[0071] On the other hand, when graph neural networks perform structural learning, they fail to effectively distinguish the importance of different neighbor nodes and information at different graph convolution levels, which affects the accuracy and expression effect of feature aggregation, thereby reducing the prediction performance and model generalization ability.

[0072] To this end, the present invention proposes a lncRNA-miRNA association prediction method based on a graph convolutional network and a contrastive learning mechanism, which is described in detail below in conjunction with specific embodiments.

[0073] Specific implementation method 1: Combination Figure 1 To explain this embodiment,

[0074] This embodiment is a lncRNA-miRNA association prediction system and prediction method based on GCN and contrastive learning. The prediction system includes a data acquisition and preprocessing unit and a prediction network model. The prediction network model includes a lncRNA and miRNA feature extraction module, a neighbor information aggregation module, and a lncRNA-miRNA potential relationship prediction module.

[0075] The data acquisition and preprocessing unit is used to collect lncRNA and miRNA sequence data, lncRNA and miRNA association data, and perform data preprocessing to obtain the initial data set of the model. The specific process includes:

[0076] Step a1: lncRNA sequence data were collected from the LNCipedia2, GENCODE3, and GENCODE-v38 databases. Simultaneously, miRNA sequence data were collected from the miRbase-v22.1 database.

[0077] Step a2: Obtain LMI data from the lncRNA SNP database, and then screen the obtained LMIs based on the lncRNA and miRNA in step a1. Obtain LMI data. Then, use all known LMIs as positive samples and the remaining unknown LMIs as negative samples.

[0078] Step a3: First, replace base T with U in all RNA sequences. Next, 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, a balanced dataset is constructed.

[0079] The lncRNA and miRNA feature extraction module specifically performs the following steps:

[0080] Apply k-mer counting technology to all lncRNA and miRNA sequences to generate corresponding k-mer feature vectors, which serve as the initial input features in subsequent model training. The following steps are included:

[0081] Step b1: Each RNA sequence S = s1s2s3...s n Divide into overlapping k-mer segments of fixed length k {k-mer i}, where k-mer i =s i s i+1 ...s i+k-1 ;

[0082] Then, count the occurrences of all k-mer segments to generate a k-mer count vector v:

[0083]

[0084] Among them, ∑ is 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.

[0085] Step b2: To enhance the semantic information and contextual relevance of the sequence expression, the k-mer sequence obtained in step b1 is pre-trained using the Doc2Vec model to obtain a joint feature vector. This specifically includes the following steps:

[0086] First, each RNA sequence is regarded as a document and the k-mer fragment is regarded as a word in the document. The Doc2vec model is trained by the distributed memory (DM) method to generate semantic features for each lncRNA and miRNA sequence.

[0087] To further improve the feature expression capability, the semantic features of the sequence obtained by the Doc2Vec model are concatenated or fused with the k-mer count vector v in step b1 to form a joint feature vector representation that includes local frequency and global semantics.

[0088] The neighbor information aggregation module specifically performs the following steps:

[0089] Step c1: construct a lncRNA and miRNA heterogeneity graph using the positive samples in the known LMI, and use the joint feature vectors corresponding to the lncRNA and miRNA obtained by training in step b2 as the initial representation of the nodes in the graph;

[0090] Step c2: Use the node-level attention mechanism to aggregate neighbor node information, including the following steps:

[0091] Step c21, define node v i and its neighbor node v j The hidden representations at layer l-1 are and The specific formula for calculating the attention coefficient between nodes i and j at layer l is as follows:

[0092]

[0093] Among them, a is the trainable weight vector, σ(·) represents the activation function; ‖ represents splicing; when l is equal to 1, it is the initial representation of the node.

[0094] Step c22, use the softmax function to Normalize and get v i Attention coefficient to its neighbor nodes

[0095]

[0096] Then, through the corresponding attention coefficient For the target node v i The neighbor nodes of the target node v are weighted summed to complete the aggregation of node features. i Feature representation at layer l The calculation formula is:

[0097]

[0098] Where W represents the trainable weight matrix, For node v i The set of neighbor nodes.

[0099] Step c3: Use the hierarchical attention mechanism to aggregate information, including the following steps:

[0100] Step c31: For node v i , the nodes in each layer are represented as l represents the layer number of the graph neural network, L is the total number of graph neural network layers in the model, and is a set hyperparameter that controls the aggregation depth of node representation. Its corresponding importance coefficient is calculated by the following formula:

[0101]

[0102] in, is the node v in the lth layer i The attention score, (q (l)) T is the transpose of the trainable weight vector of layer l, W (l) is the trainable weight matrix of layer l, is the node v in the lth layer i The transpose of the feature representation, b (l) is the bias vector of the lth layer.

[0103] Then determine the node v in the lth layer i The normalization coefficient of :

[0104]

[0105] in, is the node v in the lth layer i The normalized attention coefficient of Score for attention Normalize, It is an exponential function used to perform attention scoring, which is used to amplify the score difference.

[0106] Step c32: perform multi-layer feature fusion on lncRNA nodes and miRNA nodes respectively, and the final node feature z i The weighted sum is obtained by the following formula:

[0107]

[0108] in, is the node v in layer l i The feature representation of is the node v in the lth layer i The normalized attention coefficient of .

[0109] The lncRNA-miRNA potential relationship prediction module specifically performs the following steps:

[0110] Step d1: weighted fusion of the final node feature of lncRNA and the final node feature representation of miRNA obtained by the neighbor information aggregation module to obtain the LMI feature f LMI ;

[0111] Step d2: To enhance the discriminative ability of feature representation, the LMI features obtained by weighted fusion in step d1 are optimized through comparative learning, and positive and negative sample pairs are constructed. By maximizing the feature similarity between positive sample pairs and minimizing the similarity between negative sample pairs, the structural discrimination ability of LMI representation is further enhanced.

[0112] First, training samples were divided into positive and negative groups based on the similarity between LMI features. Specifically, lncRNA-miRNA pairs with high cosine similarity of their feature vectors were classified as positive, while lncRNA-miRNA pairs with low similarity were classified as negative. Specifically, cosine similarity was used as the feature distance metric, with thresholds set at 0.8 and 0.2. When the cosine similarity between two sets of LMI features was greater than 0.8, the pair was considered positive; when the cosine similarity between the two sets of LMI features was less than 0.2, the pair was considered negative. By constructing positive and negative pairs and comparing them to train network parameters, the embedding distribution structure of the LMI features was optimized.

[0113] Step d3: Using LMI features as a basis, MLP and Sigmoid function are used to predict the potential relationship between lncRNA and miRNA:

[0114]

[0115] Among them, MLP is a multi-layer perceptron, Sigmoid is the activation function, is the predicted score of the LMI.

[0116] Step d4: Use the binary cross entropy loss function to optimize the deviation between the predicted output and the true label;

[0117] The binary cross entropy loss function is:

[0118]

[0119] in, Represents the LMI training sample set Each pair (lcn, mi) represents a set of known lncRNA–miRNA interaction samples; p lnc,mi is the true label of the interaction relationship, where 1 indicates the existence of interaction and 0 indicates the absence of interaction; is the predicted probability calculated by MLP and Sigmoid function, which represents the predicted probability value of the interaction between the lncRNA–miRNA pair.

[0120] Based on the loss function, a trained prediction network model is obtained, and lncRNA-miRNA association prediction is realized based on the prediction network model to determine whether new LMI pairs have potential interaction relationships.

[0121] Example:

[0122] The lncRNA-miRNA association prediction method based on GCN and contrastive learning described in this embodiment includes the following steps:

[0123] S1: Collect lncRNA and miRNA sequence data, lncRNA and miRNA association data, preprocess the data, and obtain the initial data set of the model.

[0124] In this embodiment, three parts of data need to be obtained:

[0125] S11: Obtain lncRNA sequence data; the present invention collects lncRNA sequence data from several classic databases, LNCipedia2 (LNCipedia), and GENCODE3 (GENCODE-Home page (gencodegenes.org) databases, and obtains sequence data of 1663 lncRNAs in total.

[0126] S12: Obtaining miRNA sequence data; Based on the work of predecessors, the present invention collected miRNA sequence data from the GENCODE-v38 (GENCODE-Human Release 38 (gencodegenes.org) database, and obtained sequence data of 258 miRNAs in total.

[0127] 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 1663 lncRNAs collected in S11 and the 258 miRNAs collected in S12. The resulting LMI data contains 15,386 interaction pairs, with a dataset sparsity of 96.41%. All known LMIs are considered positive samples, and the remaining unknown LMIs are considered negative samples.

[0128] S14: Data preprocessing: First, the present invention replaces base T with U in all RNA sequences. Next, duplicate lncRNA-miRNA interaction pairs are removed, and lncRNAs and miRNAs with unavailable sequences are filtered out. After removing duplicate LMIs and filtering out lncRNAs / miRNAs with unavailable sequences, a balanced dataset is constructed, which contains 24,990 lncRNA and miRNA association data, including 12,495 positive samples and 12,495 negative samples.

[0129] S2: K-mer feature extraction is performed on the lncRNA and miRNA sequences collected in S1, and the Doc2vec method is used to pre-train lncRNA and miRNA.

[0130] S21: Apply k-mer counting to 418 lncRNA and 262 miRNA sequences to generate k-mer count vectors, which are used as the initial features of lncRNA and miRNA (dimension = 128);

[0131]

[0132] in, is the indicator function, ∑ is the summation symbol, is the indicator function, v j is the j-th element of the k-mer count vector.

[0133] 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, and the semantic feature vector is obtained through distributed representation learning.

[0134] The k-mer local statistical features obtained in step S21 are weightedly fused with the Doc2Vec features to generate a joint feature vector representation.

[0135] S3: Using the lncRNA and miRNA features learned in S2 as the initial features, a graph neural network with a multi-scale attention mechanism is used to aggregate neighbor information on the lncRNA and miRNA heterogeneous graph.

[0136] S31: Use node-level attention mechanism to aggregate neighbor node information; first define node v i and its neighbor node v j The hidden representations at layer l-1 are and The specific formula for calculating the attention coefficient between nodes i and j at layer l is as follows:

[0137]

[0138] Among them, a is the trainable weight vector, σ(·) represents the activation function. Then the softmax function is used to Normalize and get v i Attention coefficient to its neighbor nodes

[0139]

[0140] Finally, through the corresponding attention coefficient For the target node v i The neighbor nodes of the target node v are weighted summed to complete the aggregation of node features. iFeature representation at layer l The calculation formula is:

[0141]

[0142] Where W represents the trainable weight matrix, For node v i The set of neighbor nodes.

[0143] S32: Use hierarchical attention mechanism for information aggregation; First, for node v i , the nodes in each layer are represented as The corresponding importance coefficient is calculated by the following formula:

[0144]

[0145] in, is the node v in the lth layer i The attention score, (q (l) ) T is the transpose of the trainable weight vector of layer l, W (l) is the trainable weight matrix of layer l, is the node v in the lth layer i The transpose of the feature representation, b (l) is the bias vector of the lth layer.

[0146]

[0147] in is the node v in the lth layer i The normalized attention coefficient of Score for attention Normalize, is an exponential function used to perform attention scoring, which is used to amplify the score difference. The lncRNA and miRNA node features z obtained by aggregating neighbor information i The weighted sum is obtained by the following formula:

[0148]

[0149] in is the node v in layer l i The lncRNA-miRNA potential relationship prediction module specifically performs the following steps:

[0150] S4: Concatenate the final feature representations of lncRNA and miRNA learned in S3 and use them to predict LMI. Use MLP and Sigmoid function to predict scores

[0151] S41: concatenate the final feature representations of lncRNA and miRNA learned in S3 to obtain LMI features (dimension = 256);

[0152] S42: All obtained LMI features are optimized using contrastive learning. Specifically, a contrastive learning model is trained to further obtain more discriminative LMI feature representations. First, the training samples are divided into positive and negative groups based on the similarity between the LMI features. Specifically, lncRNA-miRNA pairs with high feature vector cosine similarity (greater than 0.8) are classified as positive, while lncRNA-miRNA pairs with low similarity (less than 0.2) are classified as negative.

[0153]

[0154] in, is the positive sample set of lncRNA and miRNA pairs, is the negative sample set of lncRNA and miRNA pairs, exp is the exponential function, τ is the temperature parameter, and * is the dot product operation.

[0155] S43: LMI features were used as the basis, and MLP and Sigmoid functions were used to predict the potential relationship between lncRNA and miRNA.

[0156]

[0157] Among them, MLP is a multi-layer perceptron, Sigmoid is the activation function, is the predicted score of the LMI.

[0158] S44: Use binary cross entropy loss function to optimize the prediction results:

[0159]

[0160] in, Represents an LMI collection The size of p lnc,mi is the true label value, is the prediction score of LMI.

[0161] S5: Test and evaluate the model performance through various experiments on the dataset.

[0162] In order to test and evaluate the performance of the present invention, the present invention was evaluated on a data set 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, the horizontal axis of the ROC curve is FP, and the vertical axis is TP. AUPR is the area under the PR curve, the horizontal axis of the PR curve is Precision, and the vertical axis is Recall. The specific calculation formula is as follows:

[0163]

[0164] 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.

[0165] 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) in the figure is the ROC curve of different methods, and (b) is the AUPR curve of different methods.

[0166] To verify the prediction performance of the method of the present invention, three representative data sets were selected for experimental evaluation: Dataset 1 contains 15,000 positive and negative samples, with a positive-to-negative sample ratio of 1:1; Dataset 2 contains 15,000 positive samples and 30,000 negative samples, with a positive-to-negative sample ratio of 1:2; Dataset 3 contains 15,000 positive samples and 75,000 negative samples, with a positive-to-negative sample ratio of 1:5. The experiment adopted a five-fold cross-validation strategy to systematically compare the performance of the method of the present invention with that of various mainstream LMI prediction methods including PmliPred, PmliPEMG, LncMirNet, RNAI-FRID, NDALMA and preMLI on the above data sets. The results are shown in Tables 1 and 2. In all data sets, the method of the present invention showed an overall prediction ability that was superior to that of existing methods, especially maintaining a high accuracy and robustness under unbalanced data distribution.

[0167] Table 1 Comparison of prediction performance of other indicators in five-fold cross validation experiments with other methods on the dataset

[0168]

[0169] Table 2 Comparison of prediction performance of other indicators in five-fold cross validation experiments with other methods on different sampling data sets

[0170]

[0171] In order to evaluate the stability and adaptability of the method of the present invention under the setting of key structural parameters, parameter sensitivity experiments were further carried out, focusing on the impact of the number of graph convolution layers, the number of attention heads and the node representation dimension on the model performance. During the experiment, the remaining parameters were kept unchanged, and different numbers of graph convolution layers (L = 1, 2, 3, 4, 5), the number of attention heads (K = 1, 2, 3, 4) and the node dimensions (D = 16, 64, 128, 256, 512) were set respectively. Five-fold cross-validation training was performed on the benchmark dataset, and the model performance under each group of settings was systematically compared. The experimental results are shown in the figure below. Figures 3 to 5 As shown in the figure, the method of the present invention shows strong stability under various parameter settings, and the best performance is achieved when the L value is 3, the number of attention heads is 4, and the node representation dimension is 128, indicating that this parameter combination can ensure the model's expressive ability while taking into account both prediction accuracy and computational efficiency, further reflecting the robustness and adjustability of the present invention in practical application scenarios.

[0172] In order to further explore the actual role of the hierarchical attention mechanism introduced in this invention in the multi-layer structure of graph convolution, a box plot of the attention weights of each GCN layer is drawn, as shown in the figure below. Figure 6 As shown. By statistically analyzing the attention weight distribution 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 a higher weight concentration in the shallower layers (such as the first layer), indicating that the low-level neighborhood information has a stronger influence on the basic feature expression; while in the deep structure, the weight distribution is relatively dispersed, indicating that the model can flexibly allocate the weights of high-order neighborhood information to achieve hierarchical expression and adaptive aggregation of information. The visualization results verify that the hierarchical attention mechanism introduced in the present invention has good interpretability, which helps to improve the model's ability to model key information in multi-layer graph structures.

[0173] To validate the effectiveness of this method in practical applications, experiments were conducted to predict miRNA-associated associations of specific lncRNAs and lncRNA-associated associations of specific miRNAs. As shown in Table 5, nine of the top 10 miRNAs predicted to have potential associations with the lncRNA NONHSAT137542.2 using this method have been validated experimentally or in the literature. As shown in Table 3, nine of the top 10 lncRNAs predicted to have potential associations with the miRNA hsa-miR-590-3p have been validated. These results demonstrate that this method possesses strong predictive power and excellent practical application value in practical lncRNA-miRNA association identification tasks.

[0174] Table 3 Prediction results

[0175]

[0176] 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 system based on GCN and contrastive learning, characterized by: The method comprises a prediction unit, which performs lncRNA-miRNA association prediction based on a prediction network model; the prediction network model comprises a lncRNA and miRNA feature extraction module, a neighbor information aggregation module, and a lncRNA-miRNA potential relationship prediction module; The lncRNA and miRNA feature extraction module applies k-mer counting technology to all lncRNA and miRNA sequences to generate k-mer count vectors, introduces the Doc2Vec model into the k-mer sequences to obtain sequence semantic features, and then obtains a joint feature vector based on the count vectors and sequence semantic features; The neighbor information aggregation module obtains the initial representation of the nodes in the graph based on the lncRNA and miRNA heterogeneous graph and the joint feature vector, then uses the node-level attention mechanism to aggregate the neighbor node information, and uses the hierarchical attention mechanism to aggregate the information to obtain the final node features of the lncRNA and the final node features of the miRNA; The lncRNA-miRNA potential relationship prediction module: weighted fusion of the final node features of lncRNA and miRNA obtained by the neighbor information aggregation module to obtain the LMI feature f LMI , make predictions and get prediction scores Among them, MLP is a multi-layer perceptron, Sigmoid is the activation function, is the prediction score of the LMI.

2. The lncRNA-miRNA association prediction system based on GCN and contrastive learning according to claim 1, characterized in that: The prediction network model is pre-trained. During the training process, the network parameters are trained by constructing positive and negative sample pairs for comparison. The process of constructing positive and negative sample pairs includes: The LMI feature f is obtained in the lncRNA-miRNA potential relationship prediction module LMI Finally, the training samples were divided into positive and negative groups according to the similarity between LMI features. The lncRNA-miRNA pairs whose feature vector cosine similarity was higher than the first threshold were classified as the positive group, and the lncRNA-miRNA pairs whose similarity was lower than the second threshold were classified as the negative group.

3. The lncRNA-miRNA association prediction system based on GCN and contrastive learning according to claim 2, characterized in that: The loss function used in the training process of the prediction network model is as follows: in, Represents the LMI training sample set Each pair (lcn, mi) represents a set of known lncRNA–miRNA interaction samples; p lnc,mi is the true label of the interaction relationship, where 1 indicates the existence of interaction and 0 indicates the absence of interaction; Predicted probability.

4. The lncRNA-miRNA association prediction system based on GCN and contrastive learning according to claim 1, characterized in that: The processing process of the lncRNA and miRNA feature extraction module includes: Step b1: Each RNA sequence S = s1s2s3...s n Divide into overlapping k-mer segments of fixed length k {k-mer i }, where k-mer i =s i s i+1 ...s i+k-1 ; Count the occurrences of all k-mer segments and generate a k-mer count vector v: Among them, ∑ is 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; v j is the jth element of the k-mer count vector, indicating the total number of times the k-mer of type j appears in the sequence; Step b2: Treat each RNA sequence as a document and the k-mer fragment as a word in the document. Train the Doc2vec model using the distributed memory (DM) method to generate semantic features for each lncRNA and miRNA sequence. The sequence semantic features obtained by the Doc2Vec model are concatenated or fused with the k-mer count vector v in step b1 to form a joint feature vector containing local frequency and global semantics.

5. The lncRNA-miRNA association prediction system based on GCN and contrastive learning according to claim 1, characterized in that: The processing process of the neighbor information aggregation module includes: Step c1: construct a lncRNA and miRNA heterogeneity graph using the positive samples in the known LMI, and use the joint feature vectors corresponding to the lncRNA and miRNA obtained by training in step b2 as the initial representation of the nodes in the graph; Step c2: Use the node-level attention mechanism to aggregate neighbor node information, including the following steps: Define node v i and its neighbor node v j The hidden representations at layer l-1 are and Then calculate the attention coefficient between nodes i and j at layer l Where a is the weight vector, σ(·) represents the activation function; ‖ represents concatenation; when l is equal to 1, it is the initial representation of the node; Using the softmax function Normalize and get v i Attention coefficient to its neighbor nodes By the corresponding attention coefficient For the target node v i The neighbor nodes of represent weighted summation to complete the aggregation of node features; Step c3: Use the hierarchical attention mechanism to aggregate information, including the following steps: For node v i , the nodes in each layer are represented as l represents the layer number of the graph neural network, and L is the total number of layers of the graph neural network; the importance coefficient corresponding to the lth layer is as follows: in, is the node v in the lth layer i The attention score, (q (l) ) T is the transpose of the weight vector of the lth layer, W (l) is the weight matrix of the lth layer, is the node v in the lth layer i The transpose of the feature representation, b (l) is the bias vector of the lth layer; Based on Determine the node v in layer l i The normalization coefficient of Multi-layer feature fusion is performed on lncRNA nodes and miRNA nodes respectively, and the final node feature z i The weighted sum is obtained by the following formula: in, is the node v in layer l i The feature representation of is the node v in the lth layer i The normalized attention coefficient of .

6. The lncRNA-miRNA association prediction system based on GCN and contrastive learning according to any one of claims 1 to 5, characterized in that: The system also includes a data acquisition and preprocessing unit: used to acquire lncRNA and miRNA sequence data, as well as lncRNA and miRNA association data, and perform data preprocessing. The preprocessing process includes the step of replacing the base T in all RNA sequences with U.

7. A lncRNA-miRNA association prediction method based on GCN and contrastive learning, characterized by: include: First, lncRNA and miRNA sequence data, as well as lncRNA and miRNA association data, are obtained; then, a prediction network model is used to predict lncRNA-miRNA associations; the prediction network model includes a lncRNA and miRNA feature extraction module, a neighbor information aggregation module, and a lncRNA-miRNA potential relationship prediction module; The lncRNA and miRNA feature extraction module applies k-mer counting technology to all lncRNA and miRNA sequences to generate k-mer count vectors, introduces the Doc2Vec model into the k-mer sequences to obtain sequence semantic features, and then obtains a joint feature vector based on the count vectors and sequence semantic features; The neighbor information aggregation module obtains the initial representation of the nodes in the graph based on the lncRNA and miRNA heterogeneous graph and the joint feature vector, then uses the node-level attention mechanism to aggregate the neighbor node information, and uses the hierarchical attention mechanism to aggregate the information to obtain the final node features of the lncRNA and the final node features of the miRNA; The lncRNA-miRNA potential relationship prediction module: weighted fusion of the final node features of lncRNA and miRNA obtained by the neighbor information aggregation module to obtain the LMI feature f LMI , make predictions and get prediction scores Among them, MLP is a multi-layer perceptron, Sigmoid is the activation function, is the prediction score of the LMI.

8. The lncRNA-miRNA association prediction method based on GCN and contrastive learning according to claim 7, characterized in that: The processing process of the lncRNA and miRNA feature extraction module includes: Step b1: Each RNA sequence S = s1s2s3...s n Divide into overlapping k-mer segments of fixed length k {k-mer i }, where k-mer i =s i s i+1 ...s i+k-1 ; Count the occurrences of all k-mer segments and generate a k-mer count vector v: Among them, ∑ is 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. Step b2: Treat each RNA sequence as a document and the k-mer fragment as a word in the document. Train the Doc2vec model using the distributed memory (DM) method to generate semantic features for each lncRNA and miRNA sequence. The sequence semantic features obtained by the Doc2Vec model are concatenated or fused with the k-mer count vector v in step b1 to form a joint feature vector containing local frequency and global semantics.

9. The lncRNA-miRNA association prediction method based on GCN and contrastive learning according to claim 7, characterized in that: The processing process of the neighbor information aggregation module includes: Step c1: construct a lncRNA and miRNA heterogeneity graph using the positive samples in the known LMI, and use the joint feature vectors corresponding to the lncRNA and miRNA obtained by training in step b2 as the initial representation of the nodes in the graph; Step c2: Use the node-level attention mechanism to aggregate neighbor node information, including the following steps: Define node v i and its neighbor node v j The hidden representations at layer l-1 are and Then calculate the attention coefficient between nodes i and j at layer l Where a is the weight vector, σ(·) represents the activation function; ‖ represents concatenation; when l is equal to 1, it is the initial representation of the node; Using the softmax function Normalize and get v i Attention coefficient to its neighbor nodes By the corresponding attention coefficient For the target node v i The neighbor nodes of represent weighted summation to complete the aggregation of node features; Step c3: Use the hierarchical attention mechanism to aggregate information, including the following steps: For node v i , the nodes in each layer are represented as l represents the layer number of the graph neural network, and L is the total number of layers of the graph neural network; the importance coefficient corresponding to the lth layer is as follows: in, is the node v in the lth layer i The attention score, (q (l) ) T is the transpose of the weight vector of the lth layer, W (l) is the weight matrix of the lth layer, is the node v in the lth layer i The transpose of the feature representation, b (l) is the bias vector of the lth layer; Based on Determine the node v in layer l i The normalization coefficient of Multi-layer feature fusion is performed on lncRNA nodes and miRNA nodes respectively, and the final node feature z i The weighted sum is obtained by the following formula: in, is the node v in layer l i The feature representation of is the node v in the lth layer i The normalized attention coefficient of .

10. The lncRNA-miRNA association prediction method based on GCN and contrastive learning according to any one of claims 7 to 9, characterized in that: After obtaining lncRNA and miRNA sequence data, as well as lncRNA and miRNA association data, and before using the prediction network model for prediction, the obtained data needs to be preprocessed. The preprocessing process includes the step of replacing base T in all RNA sequences with U.

Images