Method for predicting disease lncrna based on layer refinement graph convolutional neural network

Abstract

The application discloses a method for predicting disease lncRNA based on layer refinement graph convolutional neural network, comprising the following steps: fusing various similarity information of nodes; constructing a heterogeneous network; using a graph convolutional neural network with a layered refinement mechanism to learn the feature aggregation information of each node in the heterogeneous network; integrating the feature information of each hidden layer to obtain the final feature expression of the node; applying matrix transpose multiplication to the final feature expression matrix to obtain the prediction correlation score of the lncRNA-disease pair. The layered refinement mechanism is integrated into the graph convolutional neural network, so that the similarity between each hidden layer and the initial feature layer of the node can be fully considered in the message propagation process, the weight of the layer more similar to the initial layer can be amplified, and the weight of the layer less similar to the initial layer can be reduced, so that the common over-smoothing problem in the deep neural network is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biological information processing, and particularly relates to a method for predicting disease lncRNA based on layer refinement graph convolutional neural network. BACKGROUND

[0002] Recent studies have found that various non-coding RNAs such as lncRNA play an important role in human complex diseases, and in most cases, lncRNA has been proven to be a major regulatory factor in gene expression process, and plays an important role in various biological functions and disease development. For example, lncRNA LUCATI causes the occurrence and development of breast cancer by reducing the expression level of microRNA-7-5p, and is considered as a potential therapeutic target for breast cancer. Therefore, identifying the potential association between lncRNA and disease helps to identify the function of lncRNA and understand the potential pathogenesis of disease at the molecular level of lncRNA.

[0003] In recent years, with the rapid development of high-throughput sequencing technology and the establishment of a large number of biological information databases, a large amount of biological data has been discovered, and biological big data has a high potential value in the field of intelligent medical treatment. The generation of a large amount of lncRNA sequence information and disease semantic information makes it possible to comprehensively analyze the pathogenesis of disease from the molecular level of lncRNA. However, it is expensive and time-consuming to find the association between lncRNA and disease by using traditional biological experiments, so it is necessary to develop an effective and efficient computational model to predict disease-related lncRNAs.

[0004] At present, the lncRNA-disease association prediction method based on graph neural network (GNN) is paid more and more attention, such as the computational model proposed by Xuan et al. based on graph convolutional neural network (GCN) and convolutional neural network (CNN) to predict disease-related lncRNAs, and the method proposed by Wu et al. based on graph auto-encoder (GAE) and random forest (RF) to predict the potential association between lncRNA and disease. However, these methods all have the problem of over-smoothing caused by too many stacked network layers, which reduces the prediction accuracy of the model. SUMMARY

[0005] In order to solve the above technical problems, the application provides a method for predicting disease lncRNA based on layer refinement graph convolutional neural network, which is simple in algorithm and high in prediction accuracy.

[0006] The technical scheme for solving the above problems is: a method for predicting disease lncRNA based on layer refinement graph convolutional neural network, comprising the following steps:

[0007] 1) fusing various similarity information of lncRNA, disease and miRNA nodes;

[0008] 2) constructing a heterogeneous network by using known lncRNA-disease association, known miRNA-disease association, lncRNA-miRNA interaction and the similarity information fused in step 1);

[0009] 3) using a graph convolutional neural network with a hierarchical refinement mechanism to learn the feature aggregation information of each node in the heterogeneous network;

[0010] 4) using a READOUT function to further integrate the feature information of each hidden layer refined in the message propagation process to obtain the final feature expression of the node;

[0011] 5) applying matrix transpose multiplication to the final feature expression matrix of the lncRNA and disease nodes to obtain the prediction association score of the lncRNA-disease pair.

[0012] The method for predicting disease lncRNA based on layer refinement graph convolutional neural network, wherein the step 1) specifically comprises the following steps:

[0013] 1-1) for disease nodes, disease similarity matrix is obtained by fusing disease semantic similarity and Gaussian kernel similarity information;

[0014] 1-2) for lncRNA nodes, lncRNA similarity matrix is obtained by integrating functional similarity and Gaussian kernel similarity information;

[0015] 1-3) for miRNA nodes, miRNA similarity matrix is obtained by integrating miRNA sequence similarity and Gaussian kernel similarity information.

[0016] The method for predicting disease lncRNA based on layer refinement graph convolutional neural network, wherein the step 1-1) specifically comprises the following steps:

[0017] 1-1-1) calculating the semantic similarity of diseases by using the directed acyclic graph DAG of diseases in the MeSH database;

[0018] 1-1-2) For the Gaussian kernel similarity of diseases, it is assumed that similar diseases are related to similar lncRNAs, and vice versa; the Gaussian kernel similarity between diseases u and v is defined as follows:

[0019] KD(u,v)=exp(-α d ‖IP(u)-IP(v)‖ 2 ) (1)

[0020] Where IP(u) represents the interaction spectrum of disease u and lncRNAs, Exp is the exponential function with base e, and α d represents the kernel bandwidth control parameter, and the formula is as follows:

[0021]

[0022] Where n d is the number of disease nodes;

[0023] 1-1-3) The similarity fusion strategy of diseases is as follows:

[0024]

[0025] Where SMD u,v , DSS u,v and KD u,v represent the fusion similarity, Gaussian kernel similarity and semantic similarity of diseases u and v, respectively.

[0026] The above method for predicting lncRNA of diseases based on layer refinement graph convolutional neural network, the specific process of step 1-2) is as follows:

[0027] 1-2-1) Based on disease semantic similarity and miRNA-disease known association, the functional similarity of miRNA is calculated;

[0028] 1-2-2) Based on miRNA functional similarity and lncRNA-miRNA interaction, the functional similarity of lncRNA is calculated;

[0029] 1-2-3) The similarity fusion strategy of lncRNA is as follows:

[0030]

[0031] Where SML i,j , LFS i,j and KL i,j represent the fusion similarity, functional similarity and Gaussian kernel similarity between lncRNAs i and j, respectively, and the calculation formula of KL(i,j) is as follows:

[0032] KL(i,j)=exp(-αl ‖IP(i)-IP(j)‖ 2 ) (5)

[0033]

[0034] wherein, a l represents the kernel bandwidth control parameter, n l represents the number of lncRNAs.

[0035] The method for predicting disease lncRNAs based on the layer-refined graph convolutional neural network, the specific process of steps 1-3) is as follows:

[0036] 1-3-1) Obtain sequence similarity information corresponding to the miRNA node;

[0037] 1-3-2) Calculate the Gaussian kernel similarity matrix of the miRNA node, and the Gaussian kernel similarity calculation formula between miRNAs p and q is as follows:

[0038] KM(p,q)=exp(-α m ‖IP(p)-IP(q)‖ 2 ) (7)

[0039]

[0040] wherein, a m represents the kernel bandwidth control parameter, n m is the number of miRNAs;

[0041] 1-3-3) The similarity fusion of miRNAs is the same as that of lncRNAs and diseases, and the fusion similarity matrix SMM of miRNAs is obtained, and the formula is as follows:

[0042] wherein, SMM p,q , MSS p,q and KM p , respectively, represent the fusion similarity, functional similarity and Gaussian kernel similarity between lncRNAs i and j,

[0043] The method for predicting disease lncRNAs based on the layer-refined graph convolutional neural network, the specific process of steps 2) is as follows:

[0044] 2-1) Collect the data set in the lncRNA-disease association prediction research, including 240 lncRNAs, 405 diseases and 495 miRNAs, and the known association matrix LD of lncRNA-disease, the known association matrix MD of miRNA-disease and the lncRNA-miRNA interaction matrix LM;

[0045] 2-2) Construct a heterogeneous network using known association / interaction information and the fusion similarity information of the three nodes, and the formula definition of the adjacency matrix A of the heterogeneous network is as follows:

[0046]

[0047] wherein, LD T , LM T and MD T represent the transposition of LD, LM and MD matrices respectively.

[0048] The method for predicting lncRNA of diseases based on the layer refinement graph convolutional neural network, the step 3) comprises the following steps:

[0049] 3-1) Adopting affine transformation to adjust the feature dimension of lncRNA, disease and miRNA to be consistent, and constructing initial feature representation of lncRNA, disease and miRNA nodes, respectively represented as and The formula of affine transformation is as follows:

[0050]

[0051] wherein (·) represents matrix multiplication, and represent trainable weight matrix, and R represents dimension, and represent trainable bias parameters, and d represents the feature dimension of the three nodes after affine transformation;

[0052] 3-2) Constructing initial feature representation for training model The formula is as follows:

[0053]

[0054] 3-3) Aggregating node features by using hierarchical refinement mechanism, dynamically extracting feature information of nodes from the initial layer in the message propagation process, and the formula definition of message propagation with hierarchical refinement mechanism is as follows:

[0055]

[0056] wherein, I is a unit matrix, is the diagonal matrix of (A+I);

[0057]

[0058] c l+1 = Sim (X l+1 , X 0 ) (14)

[0059] X l+1 = (c l+1 + ε)X l+1 (15)

[0060] where X l+1 represents the node feature of the l+1 layer, W is a trainable weight matrix, Sim(·) represents a similarity function, represents a cosine similarity vector between the initial node features of the l+1 layer, and ε is a vector of the same type as c l +1 to prevent c l+1 from being a 0 vector.

[0061] The method for predicting disease lncRNA based on the layer-refined graph convolutional neural network, the process of step 4) is as follows:

[0062] 4-1) When calculating the contribution of each hidden layer to the final node feature embedding, the node feature embedding of the initial layer has been considered, therefore, the initial layer is no longer considered when calculating the final node feature embedding X, and the formula is defined as follows:

[0063] X = READOUT(X 1 , X 2 ,..., X L ) (16)

[0064] where READOUT(·) adopts an average aggregation function, and L represents the number of network layers.

[0065] The method for predicting disease lncRNA based on the layer-refined graph convolutional neural network, the process of step 5) is as follows:

[0066] 5-1) Extract the node feature embedding X l of the lncRNA and the node feature embedding X d of the disease from the obtained final node feature embedding X respectively, and the formula is as follows:

[0067] X l = X(0:n l ,0:d) (17)

[0068] X d = X(n l :n l +n d ,0:d) (18)

[0069] The above two formulas respectively represent all columns of X from 0 to (n l -1) rows and X from n l to (n l +n d-1) all columns of the row;

[0070] 5-2) using matrix transpose multiplication to obtain the prediction correlation score Pred of the lncRNA-disease pair score , as follows:

[0071]

[0072] wherein, denotes the transpose of X d , (·) denotes matrix multiplication.

[0073] The present application has the beneficial effects that:

[0074] The present application integrates the hierarchical refinement mechanism into the graph convolutional neural network, so that it can fully consider the similarity between each hidden layer and the initial feature layer of the node in the message propagation process, can amplify the weight of the layer more similar to the initial layer, and reduce the weight of the layer less similar to the initial layer, thereby solving the common over-smoothing problem in deep neural networks; and when constructing a heterogeneous network, the miRNA closely related to the lncRNA and the disease is also added, thereby constructing an lncRNA-disease-miRNA heterogeneous network, so that the learned feature information of the lncRNA and disease nodes is more rich. The present application has high accuracy and has certain market value in the field of intelligent medical treatment. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 is the overall flowchart of the present application. DETAILED DESCRIPTION

[0076] The present application will be further described below in combination with the drawings and examples.

[0077] As shown in Figure 1 , a method for predicting disease lncRNA based on layer refinement graph convolutional neural network comprises the following steps:

[0078] 1) Fusion of multiple similarity information of lncRNA, disease and miRNA nodes. Step 1) specifically comprises the following steps:

[0079] 1-1) For disease nodes, disease semantic similarity and Gaussian kernel similarity information are fused to obtain a disease similarity matrix;

[0080] 1-2) For lncRNA nodes, integrate their functional similarity and Gaussian kernel similarity information to obtain an lncRNA similarity matrix;

[0081] 1-3) For miRNA nodes, integrate miRNA sequence similarity and Gaussian kernel similarity information to obtain a miRNA similarity matrix.

[0082] Step 1-1) The specific process is as follows:

[0083] 1-1-1) Calculate the semantic similarity of diseases through the directed acyclic graph DAG of diseases in the MeSH database;

[0084] 1-1-2) For the Gaussian kernel similarity of diseases, it is assumed that similar diseases are related to functionally similar lncRNAs, and vice versa; the Gaussian kernel similarity between diseases u and v is defined as follows:

[0085] KD(u,v) = exp(-a d ‖IP(u)-IP(v)‖ 2 ) (1)

[0086] where IP(u) represents the interaction spectrum of disease u and lncRNAs, Exp is the exponential function with base e, a d represents the kernel bandwidth control parameter, and the formula is as follows:

[0087]

[0088] where n d is the number of disease nodes;

[0089] 1-1-3) The similarity fusion strategy of diseases is as follows:

[0090]

[0091] where SMD u,v , DSS u,v and KD u,v represent the fusion similarity, Gaussian kernel similarity and semantic similarity of diseases u and v, respectively.

[0092] Step 1-2) The specific process is as follows:

[0093] 1-2-1) Calculate the functional similarity of miRNAs based on disease semantic similarity and known miRNA-disease associations;

[0094] 1-2-2) Calculate the functional similarity of lncRNAs based on miRNA functional similarity and lncRNA-miRNA interaction;

[0095] 1-2-3) The similarity fusion strategy of lncRNAs is as follows:

[0096]

[0097] where SML i,j , LFS i,j and KL i,jrespectively, the fusion similarity, the functional similarity and the Gaussian kernel similarity between lncRNAs i and j, and KL(i,j) is calculated as follows:

[0098] KL(i,j) = exp(-a l ‖IP(i)-IP(j)‖ 2 ) (5)

[0099]

[0100] where a l represents the kernel bandwidth control parameter, n l represents the number of lncRNAs.

[0101] The specific process of step 1-3) is as follows:

[0102] 1-3-1) Obtain the sequence similarity information of the corresponding miRNA node;

[0103] 1-3-2) Calculate the Gaussian kernel similarity matrix of the miRNA node, and the Gaussian kernel similarity between miRNAs p and q is calculated as follows:

[0104] KM(p,q) = exp(-a m ‖IP(p)-IP(q)‖ 2 ) (7)

[0105]

[0106] where a m represents the kernel bandwidth control parameter, n m is the number of miRNAs.

[0107] 1-3-3) The similarity fusion of miRNA is the same as that of lncRNA and disease, and the fusion similarity matrix SMM of miRNA is obtained, and the formula is as follows:

[0108] where SMM p,q , MSS p,q and KM p , respectively, represent the fusion similarity, the functional similarity and the Gaussian kernel similarity between lncRNAs i and j,

[0109] 2) Use the known lncRNA-disease association, miRNA-disease known association, lncRNA-miRNA interaction and the fused similarity information of step 1) to construct a heterogeneous network.

[0110] The specific process of step 2) is as follows:

[0111] 2-1) Collect the dataset about lncRNA-disease association prediction research, including 240 lncRNAs, 405 diseases and 495 miRNAs, and the known association matrix of lncRNA-disease LD, miRNA-disease MD and lncRNA-miRNA interaction LM;

[0112] 2-2) Use the known association / interaction information and the fusion similarity information of the three nodes to construct a heterogeneous network, the formula definition of the adjacency matrix A of which is as follows:

[0113]

[0114] Wherein, LD T , LM T and MD T represent the transposition of LD, LM and MD matrices respectively.

[0115] 3) Use the graph convolutional neural network with hierarchical refinement mechanism to learn the feature aggregation information of each node in the heterogeneous network.

[0116] Step 3) includes the following steps:

[0117] 3-1) Adopt affine transformation to adjust the feature dimensions of lncRNA, disease and miRNA to be consistent, and construct the initial feature representation of lncRNA, disease and miRNA nodes, represented as and The formula of affine transformation is as follows:

[0118]

[0119] Wherein (·) represents matrix multiplication, and represent trainable weight matrices, and R represents the dimension, and represent trainable bias parameters, and d represents the feature dimension of the three nodes after affine transformation;

[0120] 3-2) Construct the initial feature representation for training the model, the formula of which is as follows:

[0121]

[0122] 3-3) Aggregate node features by using hierarchical refinement mechanism, dynamically extract feature information of nodes from the initial layer in the message propagation process, and the formula definition of message propagation with hierarchical refinement mechanism is as follows:

[0123]

[0124] where I is the identity matrix, is a diagonal matrix of (A + I);

[0125]

[0126] c l+1 = Sim(X l+1 , x 0 ) (14)

[0127] x l+1 = (c l+1 +s)X l+1 (15)

[0128] where X l+1 denotes the node features of the (l+1)-th layer, W is a trainable weight matrix, and Sim(·) denotes a similarity function, denotes the cosine similarity vector between the initial node features and the node features of the (l+1)-th layer, and ε is a vector of the same type as c l +1 to prevent c l+1 from being a zero vector.

[0129] 4) Use the READOUT function to further integrate the feature information of each hidden layer refined during the message propagation process to obtain the final feature representation of the node.

[0130] The process of step 4) is as follows:

[0131] 4-1) When calculating the contribution of each hidden layer to the final node feature embedding, the node feature embedding of the initial layer has already been considered, therefore, the initial layer is no longer considered when calculating the final node feature embedding X, which is defined as follows:

[0132] X = READOUT(X 1 , X 2 , …, X L ) (16)

[0133] where READOUT(·) adopts an average aggregation function, and L denotes the number of network layers.

[0134] 5) Apply matrix transpose multiplication to the final feature representation matrices of the lncRNA and disease nodes to obtain the predicted association scores of the lncRNA-disease pairs.

[0135] The process of step 5) is as follows:

[0136] 5-1) Extract the node feature embedding X l of the lncRNA and the node feature embedding X d of the disease from the obtained final node feature embedding X, which is defined as follows:

[0137] X l = X(0:n l ,0:d) (17)

[0138] X d = X(n l :n l +n d ,0:d) (18)

[0139] The above two formulas respectively represent all columns of X from 0 to (n l -1) rows and all columns of X from n l to (n l +n d -1) rows.

[0140] 5-2) The predicted correlation score Pred score of the lncRNA-disease pair is obtained by using matrix transpose multiplication, and the formula is as follows:

[0141]

[0142] Wherein, represents the transpose of X d , and (·) represents matrix multiplication.

[0143] Examples

[0144] [1] Lu C, Yang M, Luo F, et al. Prediction of lncRNA-disease associations based on inductive matrix completion. Bioinformatics. 2018; 34(19): 3357-3364.

[0145] The literature [1] discloses a method for predicting lncRNA-disease associations based on inductive matrix completion. The method extracts the principal vector features of lncRNA and disease nodes using principal component analysis method, and for new lncRNA or disease nodes, the interaction spectrum of its neighbors is used instead.

[0146] [2] Wu X, Lan W, Chen Q, Dong Y, Liu J, Peng W. Inferring LncRNA-disease associations based on graph autoencoder matrix completion [published online ahead of print, 2020 May 20]. Comput Biol Chem. 2020; 87: 107282.

[0147] The document [2] firstly uses a graph convolutional neural network to encode the local graph structure and node features, and learns the latent feature vectors of lncRNA and diseases; then uses the inner product of the feature vectors of lncRNA and diseases as a decoder to reconstruct the lncRNA-disease association matrix.

[0148] [3] Zeng M, Lu C, Fei Z, et al. DMFLDA: A Deep Learning Framework for Predicting lncRNA-Disease Associations. IEEE / ACM Trans Comput Biol Bioinform. 2021; 18(6): 2353-2363.

[0149] The model in document [3] uses a cascade of nonlinear hidden layers to learn the latent feature representation of lncRNA and diseases, which can learn more complex and nonlinear relationships between nodes to improve prediction accuracy.

[0150] In order to verify the prediction accuracy of the method of the present application, the performance of the method of the present application is evaluated on six indicators of area under the receiver operating characteristic curve (AUC), area under the PR curve (AP), accuracy (Accuracy), precision (Precision), recall (Recall) and F1, and compared with the above-mentioned three documents, and the comparison results are shown in Table 1:

[0151] Table 1 Comparison of performance indicators with other methods

[0152]

[0153] Conclusion: Compared with other methods for predicting lncRNA-disease associations, the method of the present application achieves the best performance in most performance indicators, and achieves the second best in other indicators, and has the advantages of high accuracy, strong reliability, etc.

[0154] The present method is analyzed in the case of predicting lncRNAs potentially associated with diseases, as shown in Tables 2 and 3. The top 10 predicted associated lncRNAs of two common cancers in humans, breast cancer and lung cancer, are shown in the table, which are confirmed in the public database LncRNADisease v2.0, Lnc2cancer or published literature. (PMID is the ID in PubMed, a free biomedical literature retrieval system)

[0155] Table 2 lncRNAs potentially associated with breast cancer (top 10 predicted)

[0156]

[0157] Analysis of Table 2 shows that among the top 10 candidate lncRNAs associated with breast cancer, 7 have been confirmed, of which 4 have been confirmed to be closely related to breast cancer among the top 5 candidate lncRNAs. For example, lncRNA MIR17HG can inhibit the proliferation and migration of breast cancer cells by competing endogenous RNAs (ceRNA) mechanism through sponge miR-454-3p, indicating that targeting MIR17HG can be a feasible method for screening candidate drugs for breast cancer.

[0158] Table 3 lncRNAs potentially associated with lung cancer (top 10 predicted)

[0159]

[0160] Analysis of Table 3 shows that among the top 10 candidate lncRNAs potentially associated with lung cancer, 8 have been confirmed to be associated with lung cancer, of which 4 have been confirmed by public data or published articles among the top 5 candidate lncRNAs.

[0161] Conclusion: Through case analysis of two closely related cancers in humans, it is found that most of the lncRNAs predicted to be potentially associated with the related diseases have been confirmed to be associated with the diseases, further indicating the reliability and practicality of the present method. At the same time, these lncRNAs predicted to be potentially associated with the related diseases but not yet confirmed to be actually associated with the diseases can be closely observed by biomedical researchers, and detailed biomedical experiments can be conducted on them, greatly improving the efficiency and accuracy of scientific research.

Claims

1. A method for predicting disease lncRNA based on layer refinement graph convolutional neural network, characterized in that, The method comprises the following steps: 1) fusing multiple similarity information of lncRNA nodes, fusing multiple similarity information of disease nodes, and fusing multiple similarity information of miRNA nodes; The step 1) specifically comprises the following steps: 1-1) for disease nodes, fusing disease semantic similarity and Gaussian kernel similarity information to obtain a disease similarity matrix; 1-2) for lncRNA nodes, integrating functional similarity and Gaussian kernel similarity information to obtain an lncRNA similarity matrix; 1-3) for miRNA nodes, integrating miRNA sequence similarity and Gaussian kernel similarity information to obtain a miRNA similarity matrix; 2) constructing a heterogeneous network by using lncRNA-disease known associations, miRNA-disease known associations, lncRNA-miRNA interactions and the similarity information fused in step 1); 3) learning feature aggregation information of each node in the heterogeneous network by using a graph convolutional neural network with a hierarchical refinement mechanism; The step 3) comprises the following steps: 3-1) Affine transformation is used to adjust the feature dimensions of lncRNA, disease and miRNA to be consistent, and the initial feature representations of lncRNA, disease and miRNA nodes are constructed, respectively represented as , , , and the affine transformation formula is as follows: ； wherein denotes matrix multiplication, , and denotes a trainable weight matrix, R denotes a dimension, , and denotes a trainable bias parameter, denotes the feature dimension after affine transformation of the three kinds of nodes; 3-2) Constructing an initial feature representation for training a model The formula is as follows: ； 3-3) aggregating node features by using a hierarchical refinement mechanism, and dynamically extracting feature information of nodes from an initial layer in a message propagation process, and a message propagation formula with a hierarchical refinement mechanism is defined as follows: ； wherein is the identity matrix, is a diagonal matrix; ； ； ； wherein, represents the node feature of the layer, is a trainable weight matrix, represents a similarity function, represents a cosine similarity vector between the initial node feature of the node feature of the layer, is a vector of the same type as to prevent is a vector; 4) further integrating feature information of each hidden layer refined in the message propagation process by using a READOUT function to obtain final feature expression of nodes; 5) applying matrix transposition multiplication to the final feature expression matrix of lncRNA and disease nodes to obtain a prediction association score of lncRNA-disease pairs.

2. The method of predicting disease lncRNA based on layer refinement graph convolutional neural network according to claim 1, characterized in that, The specific process of the step 1-1) is as follows: 1-1-1) calculating semantic similarity of diseases by using a directed acyclic graph (DAG) of diseases in a MeSH database; 1-1-2) for Gaussian kernel similarity of diseases, it is assumed that similar diseases are related to functionally similar lncRNAs, and vice versa; Disease , The Gaussian kernel similarity definition between and is as follows: ； wherein indicates a disease interaction profile with lncRNAs, is an exponential function with base e, where e is the natural number, indicates a nuclear bandwidth control parameter, and is expressed by the following formula: ； wherein, is the number of disease nodes; 1-1-3) the similarity fusion strategy of diseases is as follows: ； wherein , and represent fusion similarity, Gaussian kernel similarity and semantic similarity of diseases and respectively.

3. The method of predicting disease lncRNA based on layer refinement graph convolutional neural network according to claim 2, characterized in that, The specific process of the step 1-2) is as follows: 1-2-1) calculating functional similarity of miRNAs based on disease semantic similarity and miRNA-disease known associations; 1-2-2) calculating functional similarity of lncRNAs based on miRNA functional similarity and lncRNA-miRNA interactions; 1-2-3) the similarity fusion strategy of lncRNAs is as follows: ； wherein, , and represent fusion similarity, functional similarity and Gaussian kernel similarity between lncRNA and respectively, The calculation formula is as follows: ； ； wherein represents a core bandwidth control parameter, represents the number of IncRNAs.

4. The method of predicting disease lncRNA based on layer refinement graph convolutional neural network according to claim 3, characterized in that, The specific process of the step 1-3) is as follows: 1-3-1) obtaining sequence similarity information of corresponding miRNA nodes; 1-3-2) Calculate the Gaussian kernel similarity matrix of miRNA nodes, miRNA and The Gaussian kernel similarity calculation formula between them is as follows: ； ； wherein, represents a core bandwidth control parameter, is the number of miRNAs; 1-3-3) Similarity fusion of miRNA and lncRNA and disease, get the fusion similarity matrix of miRNA The formula is as follows: wherein, , and represent fusion similarity, functional similarity and Gaussian kernel similarity between lncRNAs and respectively, .

5. The method of predicting disease lncRNA based on layer refinement graph convolutional neural network according to claim 4, characterized in that, The specific process of the step 2) is as follows: 2-1) Collecting datasets about lncRNA-disease association prediction studies, including K lncRNAs, K diseases and K miRNAs, and known association matrices of lncRNA-disease , known association matrices of miRNA-disease and lncRNA-miRNA interaction matrices ; 2-2) Constructing a heterogeneous network using known association / interaction information and fusion similarity information of three nodes, whose adjacency matrix The formula is defined as follows: ； wherein , and denote the transpose of the matrices , and , respectively.

6. The method of predicting disease lncRNA based on layer refinement graph convolutional neural network according to claim 5, characterized in that, The process of the step 4) is as follows: 4-1) When computing the contribution of each hidden layer to the final node feature embedding, the node feature embedding of the initial layer has already been taken into account, therefore, the final node feature embedding is computed without taking into account the initial layer, defined by the following formula: ​ ； where READOUT ) employs an average aggregation function, denotes the number of network layers.

7. The method of predicting disease lncRNA based on layer refinement graph convolutional neural network according to claim 1, characterized in that, The process of the step 5) is as follows: 5-1) from the resulting final node feature embedding the node feature embedding of the lncRNA and the node feature embedding of the disease are extracted, respectively the node feature embedding of the lncRNA and the node feature embedding of the disease are extracted, respectively , as follows: ； ； The above two formulas represent, respectively, the extraction of all columns from rows to and the extraction of all columns from rows to ;​​ 5-2) Using matrix transpose multiplication to obtain the predicted correlation score of lncRNA-disease pairs , as follows: ； wherein denotes the transpose of denotes matrix multiplication.

Images