A circRNA-disease association prediction method based on graph convolutional neural network and node2vec

Abstract

A circRNA-disease association prediction method based on graph convolutional neural network and Node2vec, the steps are as follows: obtain circRNA-disease association matrix; calculate circRNA functional similarity, circRNA Gaussian interaction spectrum kernel similarity, disease Gaussian interaction spectrum kernel similarity, disease semantic similarity, construct circRNA integrated similarity and disease integrated similarity, and generate circRNA-disease heterogeneous graph; sparse autoencoder extracts features of circRNA (disease) integrated similarity and transforms it into a 64-dimensional feature vector, and fuses the circRNA feature vector and the disease feature vector into the final circRNA-disease feature vector; graph convolutional neural network extracts local structural information of nodes from circRNA-disease heterogeneous graph; Node2vec method extracts global structural information of nodes from circRNA-disease heterogeneous graph; the node information obtained in the first two steps is sent to random forest classifier to predict potential circRNA-disease association. Predicting disease-related circRNA by computational method saves time, helps to clarify the pathogenesis of the disease, and finds effective treatment plans.

Description

Technical Field

[0001] The present invention relates to the field of association prediction in bioinformatics, and in particular to a circRNA-disease association prediction method based on graph convolution neural network and Node2vec. Background Art

[0002] In 1976, the first circRNA was discovered in the study of RNA viruses. Due to the structural specificity, unknown function and low abundance of circRNA, it was considered to be an artifact or a product of mis-splicing. With the development of sequencing technology, more and more circRNAs have been identified in thousands of organisms, such as plants, animals and bacteria. People have found that circRNA has important molecular functions: participating in the regulation of gene expression, acting as a molecular sponge to absorb microRNA, inhibiting the activity of miRNA, regulating the expression of messenger RNA, etc. Mutations or abnormal functions of circRNA lead to the disorder of various life activities, thus causing diseases. To this end, exploring the mechanism of circRNA in the occurrence of diseases and its role in disease treatment, and understanding the relationship between circRNA and diseases are important contents of bioinformatics research, which are beneficial to the prognosis, diagnosis and treatment of diseases, and are new avenues for future research.

[0003] Traditional biological experimental verification methods require a lot of manpower and material resources, and have high prediction accuracy but are time-consuming. Using computational methods to mine the biological characteristics of data and predict the relationship between circRNA and disease is convenient and efficient. The current computational methods used to predict the relationship between circRNA and disease can be divided into two categories: network-based and machine learning-based.

[0004] The network propagation-based method uses circRNA-disease association data to construct a circRNA (disease) similarity network to predict potential circRNA-disease associations. Fan et al. developed a computational model KATZHCDA using the KATZ measure on a heterogeneous network constructed from circRNA expression profiles, disease phenotype similarities, and known circRNA-disease associations. This model can successfully predict circRNA-disease associations using a simple metric for heterogeneous networks, but it is not suitable for predicting new diseases without any known circRNA associations or isolated circRNAs without any known disease associations. Li et al. proposed a method for circRNA-disease association prediction DWNCPCDA using DeepWalk and network consistency projection methods. Its advantage is that it uses the network embedding method DeepWalk to learn the node embedding of the known circRNA-disease association network, and combines it with similarity-based methods to provide greater flexibility for circRNA-disease association prediction. In the future, more biomedical association data of circRNAs or diseases, such as circRNA-miRNA associations and miRNA-disease associations, will be integrated to further improve the prediction performance.

[0005] Machine learning-based methods use supervised or unsupervised methods to mine deep features of circRNA and disease data, iterate learning to gradually optimize model parameters, and design classifiers to identify circRNAs associated with diseases. Lei et al. proposed a computational method RWRKNN, which applies the restarted random walk algorithm to weighted features with global network topology information, and uses the K nearest neighbor algorithm to classify according to features to improve prediction performance. However, RWRKNN is slightly insufficient in revealing the association between diseases and new circRNAs that have no association or between circRNAs and new diseases that have no association. Ding et al. developed a computational model based on random walks and logistic regression to predict the association between circRNA and disease RWLR. The restarted random walk method obtains global structural information for each circRNA, which is better than methods based only on similarity. RWLR can predict new circRNAs that are associated with no known diseases. However, RWLR only considers circRNA similarity and does not contain enough disease information, resulting in low prediction accuracy. Zhang et al. proposed a computational method based on graph representation learning to identify circRNA-disease associations, iGRLCDA, for predicting the potential association between circRNA and disease. This method uses a deep learning model of graph convolutional neural network and graph decomposition to effectively mine higher-level circRNA and disease information. However, iGRLCDA relies on the properties or features of known circRNA-disease associations and has low sensitivity to new circRNA-disease associations.

[0006] In view of this, it is of great significance to study the prediction method of circRNA-disease association relationship. The present invention proposes a circRNA-disease association prediction method based on graph convolutional neural network and Node2vec to predict potential circRNA-disease association. Summary of the invention

[0007] The purpose of the present invention is to solve the problems of low prediction accuracy, time-consuming and performance-consuming training of existing circRNA-disease association prediction models, and propose a circRNA-disease association prediction method based on graph convolutional neural network and node2vec to improve prediction accuracy and reduce training costs.

[0008] The technical solution of the present invention specifically comprises the following steps:

[0009] Step 1: Obtain the circRNA-disease association matrix.

[0010] Experimentally verified circRNA-disease association data were obtained from the circR2Disease database, redundant data were deleted, and only known association data related to human complex diseases were selected as the circRNA-disease association matrix.

[0011] Step 2: Calculate disease semantic similarity, disease Gaussian interaction spectrum core similarity, circRNA functional similarity, circRNA Gaussian interaction spectrum core similarity, construct circRNA integrated similarity and disease integrated similarity, and generate a circRNA-disease heterogeneous graph.

[0012] The relevant annotation words of each disease were obtained from the MESH database, and the semantic similarity between diseases was calculated using a directed acyclic graph (DAG) to obtain the disease semantic similarity; the circRNA (disease) Gaussian interaction spectrum kernel similarity was calculated based on the circRNA-disease association matrix; the circRNA functional similarity was calculated based on the disease semantic similarity and the circRNA-disease association matrix. By integrating complementary information from multiple data sources and different representation methods, integrated similarity was used to quantify the similarity of each pair of diseases to overcome inherent sparsity, and the circRNA integrated similarity matrix and disease integrated similarity matrix were obtained.

[0013] Step 3: The sparse autoencoder extracts and transforms the circRNA (disease) integrated similarity into a 64-dimensional feature vector, and then fuses the circRNA feature vector and the disease feature vector into the final circRNA-disease feature vector.

[0014] Sparse autoencoders can not only automatically learn features, but also give better feature descriptions than raw data. The features learned by sparse autoencoders replace the raw data, and the model prediction performance can be improved to a certain extent. For this reason, the present invention uses sparse autoencoders to integrate similarities of circRNA (disease) respectively, minimize the error between input and output through the back propagation algorithm, extract and transform features, and obtain 64-dimensional circRNA (disease) feature vectors. Finally, the circRNA (disease) feature vectors are combined to obtain the final circRNA-disease feature vector.

[0015] Step 4: Graph convolutional neural network extracts local structural information of nodes from the circRNA-disease heterogeneous graph.

[0016] Local structural information describes the local similarity between nodes in the graph. Specifically, if there is an edge connection between two nodes, the two nodes will be connected in the embedding space; if there is no edge connection between two nodes, their first-order proximity is 0. The graph convolutional neural network inputs the structure of the circRNA-disease heterogeneous graph and the circRNA (disease) node features, outputs the node pooling information and graph structure information, and obtains local structural information.

[0017] Step 5: The Node2vec method extracts the global structural information of nodes from the circRNA-disease heterogeneous graph.

[0018] Global structural information describes the relationship between two nodes that are not directly connected. The Node2vec method is a targeted improvement on DeepWalk. It samples the graph based on random walks and maps the node adjacency structure to a sequence structure. The sampled sequence is then used to train the Skip-gram model to capture the connectivity between nodes and obtain global structural information.

[0019] Step 6: The node information obtained in the first two steps is fed into the random forest classifier to predict potential circRNA-disease associations.

[0020] The node information obtained in the first two steps was sent to the random forest classifier to predict the potential circRNA-disease association relationship, and the AUC value and AUPR value of the present invention were obtained by five-fold cross validation to obtain the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a schematic diagram of the process of predicting circRNA-disease association relationship based on graph convolutional neural network and node2vec according to the present invention.

[0022] Figure 2 It is the ROC curve diagram of the implementation method of the present invention.

[0023] Figure 3 It is a PR curve diagram of the implementation method of the present invention. DETAILED DESCRIPTION

[0024] The present invention is a circRNA-disease association prediction method based on graph convolutional neural network and node2vec. The present invention is further described in detail below in conjunction with specific implementation methods and simulation experiments. It should be understood by professionals in the field that these implementation methods are only used to explain the technical principles of the present invention and are not intended to limit the scope of evidence collection of the present invention.

[0025] like Figure 1 As shown, a circRNA-disease association prediction method based on graph convolutional neural network and node2vec specifically includes the following steps:

[0026] Preferably, the obtaining of the association matrix in step 1 is specifically as follows:

[0027] 739 experimentally verified circRNA-disease known associations (involving 661 circRNAs and 100 diseases) were obtained from the circR2Disease database. After removing redundant data, only 650 known association data related to human complex diseases (involving 585 circRNAs and 88 diseases) were selected as the known association matrix nc and nd represent the number of circRNA and disease, respectively. i With disease j There is a known correlation verified by experiment, so define the matrix element A(c i ,d j )=1; if any circRNA c i With disease j There is no known experimentally verified correlation, so define the matrix element A(c i ,d j )=0.

[0028] Preferably, the calculation of disease semantic similarity in step 2 is specifically as follows:

[0029] By downloading disease-related data, we can build a disease semantic similarity matrix Any disease t For disease i The semantic contribution value of It is calculated as follows:

[0030]

[0031] Where σ represents the attenuation coefficient of semantic contribution.

[0032] The matrix element DS(d i ,d j ) indicates disease i With disease j The disease semantic similarity between them is calculated as follows:

[0033]

[0034] Preferably, the calculation of circRNA functional similarity described in step 2 is specifically as follows:

[0035] The functional similarity of circRNAs is measured based on the fact that circRNAs with similar functions are often associated with diseases with similar phenotypes. Matrix element CS(c i ,c j ) indicates circRNA c i With c j The functional similarity between is calculated as follows:

[0036]

[0037] In the formula, the set D i Indicates circRNA c i Set of related diseases; set D j Indicates circRNA c j Related disease groups; |D i | and |D j | respectively represent the set D i and D j The number of diseases.

[0038] Preferably, the calculation of circRNA (disease) Gaussian interaction spectrum core similarity described in step 2 is specifically:

[0039] Combine the association matrix and disease semantic similarity to calculate the circRNA (disease) Gaussian interaction spectrum kernel similarity. The Gaussian interaction spectrum kernel similarity of the disease is represented by the matrix element DK(d i ,d j ) indicates disease i With disease j The Gaussian interaction spectrum kernel similarity is calculated as follows:

[0040]

[0041] In the formula, the parameter μ d Indicates the control kernel bandwidth for GIP affinity.

[0042] Similarly, the matrix The matrix element CK(c i ,c j ) indicates circRNA c i With c j The Gaussian interaction spectrum kernel similarity is calculated as follows:

[0043] CK(c i ,c j )=exp(-μ c ||A(c i ,d j )-A(c j ,d j )|| 2 )

[0044] In the formula, the parameter μ c Control kernel bandwidth indicating GIP similarity

[0045] Preferably, the calculation of circRNA (disease) integrated similarity described in step 2 is specifically as follows:

[0046] Considering the inherent sparsity of disease semantic similarity and circRNA functional similarity, we integrate complementary information from multiple data sources and different representation methods, and use integrated similarity to quantify the similarity of each pair of circRNA (disease) to overcome the inherent sparsity. Represents the disease ensemble similarity, matrix element X d (d i ,d j ) is calculated as follows:

[0047]

[0048] CircRNA integration similarity matrix Indicates that the matrix element X c (c i ,c j ) is calculated as follows:

[0049]

[0050] Preferably, the sparse autoencoder described in step 3 extracts and transforms the circRNA (disease) integrated similarity into a 64-dimensional feature vector, and fuses the circRNA feature vector and the disease feature vector into the final circRNA-disease feature vector, specifically:

[0051] The sparse autoencoder encodes the original input features and reduces the dimensions to find the potential correlation between the input features and extract expressive high-order features. The sparse autoencoder consists of an encoder and a decoder, which is a three-layer neural network, including an input layer, a hidden layer, and an output layer, where the input layer x is mapped to the hidden layer y one by one. The encoder is calculated as follows:

[0052] y=sigmoid(W1x(i)+a1)

[0053] In the formula, sigmoid represents the activation function; W1 represents the connection parameter between the input layer x and the hidden layer y; a1 represents the offset.

[0054] The decoder calculates as follows:

[0055] z=sigmoids(W2y+a2)

[0056] Where W2 represents the connection parameter between the hidden layer y and the output layer z, and a2 represents the offset.

[0057] The circRNA (disease) integrated similarity is input into the sparse autoencoder, and the 64-dimensional feature vectors are extracted and transformed by minimizing the error between input and output through the back propagation algorithm, which are Z c and Z d , the two are combined to obtain the final circRNA-disease feature vector Z cd , calculated as follows:

[0058]

[0059] Preferably, the graph convolutional neural network described in step 4 extracts local structural information of nodes from the circRNA-disease heterogeneous graph, specifically:

[0060] The graph convolutional neural network inputs the structure of the graph and the characteristics of each node, and can output the pooling information of the nodes and the information of the graph (node) structure to obtain the local structural information of the graph. To this end, the circRNA-disease known association matrix A is converted into an adjacency matrix by calculation The spatial method of graph convolutional neural network is used to obtain local structural information, which is calculated as follows:

[0061]

[0062] In the formula, ReLU(*) represents the activation function of the two layers of the neural network; express The measurement matrix of ; W represents the weight matrix; represents the adjacency matrix with self-loops added, which is calculated as in, Represents the identity matrix.

[0063] Preferably, the Node2vec method described in step 5 extracts the global structural information of nodes from the circRNA-disease heterogeneous graph, specifically:

[0064] Node2vec is a semi-supervised learning method for scalable feature learning in networks, which can maximize the possibility of retaining the network domain of nodes in the d-dimensional feature space. First, the graph is sampled based on random walks, and the node adjacency structure is mapped to a sequence structure. Then, the sampled sequence is used to train the Skip-gram model to capture the connectivity between nodes and obtain global structural information.

[0065] Preferably, the step 6 of sending to the random forest classifier is specifically:

[0066] The node information obtained in the first two steps is sent to the random forest classifier to predict the potential circRNA-disease association and obtain the prediction results.

[0067] The technical effects of the present invention are further illustrated by experimental verification as follows:

[0068] 1. Experimental conditions and contents:

[0069] The experiments of the present invention were all completed on an AMD 1.80 GHz CPU and a Windows 10 operating system.

[0070] 2. Experimental results analysis:

[0071] The results showed that the prediction accuracy of circRNA-disease association relationship was evaluated by five-fold cross validation, and the evaluation indicators were ROC and PR. Among them, ROC is the area under the ROC curve with FPR as the horizontal axis and TPR as the vertical axis, and PR is the area under the Pre-Recall curve with recall as the horizontal axis and precision as the vertical axis. The larger the ROC value and PR value, the higher the accuracy.

[0072] The ROC curve and PR curve obtained by the five-fold cross validation of the present invention are as follows: Figure 2-3 shown.

[0073] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in the field, understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention, which should be included in the protection scope of the claims of the present invention.

Claims

1. A circRNA-disease association prediction method based on graph convolutional neural network and Node2vec, characterized in that: The following steps are involved: Step 1: Obtain circRNA-disease association matrix; Step 2: Calculate disease semantic similarity, disease Gaussian interaction spectrum core similarity, circRNA functional similarity, circRNA Gaussian interaction spectrum core similarity, construct circRNA integrated similarity and disease integrated similarity, and generate circRNA-disease heterogeneous graph A; Step 3: Input the circRNA integrated similarity and disease integrated similarity into the sparse autoencoder respectively, and extract and transform the 64-dimensional feature vectors by minimizing the error between input and output through the back propagation algorithm, which are Z c and Z d , the two are combined to obtain the final circRNA-disease feature vector Z cd , calculated as follows: The sparse autoencoder is composed of an encoder and a decoder. Let it be a three-layer neural network, including an input layer, a hidden layer and an output layer, where the input layer x is mapped one by one to the hidden layer y, and the encoder is calculated as follows: y = sigmoid(W1x(i)+a1); In the formula, sigmoid represents the activation function; W1 represents the connection parameter between the input layer x and the hidden layer y; a1 represents the offset; The decoder calculates as follows: z=sigmoids(W2y+a2) Where W2 represents the connection parameter between the hidden layer y and the output layer z, and a2 represents the offset; Step 4: The circRNA-disease heterogeneous graph A is converted into an adjacency matrix A′ through calculation, and the local structural information of the nodes is extracted using the spatial method of graph convolutional neural network. The calculation is as follows: In the formula, ReLU(*) represents the activation function of the two layers of the neural network; express The measurement matrix of ; w represents the weight matrix; represents the adjacency matrix with self-loops added, which is calculated as Where i represents the identity matrix; Step 5: The Node2vec method first samples the circRNA-disease heterogeneous graph a based on random walks, maps the node adjacency structure to a sequence structure, and then uses the sampled sequence to train the Skip-gram model to capture the connectivity between nodes and extract the global structural information of the nodes; Step 6: The node information obtained in the first two steps is fed into the random forest classifier to predict potential circRNA-disease associations.

2. The circRNA-disease association prediction method based on graph convolutional neural network and Node2vec according to claim 1, characterized in that: In step 1, specifically: Experimentally verified circRNA-disease association data were obtained from the circR2Disease database, redundant data were deleted, and only known association data related to human complex diseases were selected as the circRNA-disease association matrix.

3. The circRNA-disease association prediction method based on graph convolutional neural network and Node2vec according to claim 1, characterized in that: In step 2, specifically: The relevant annotation words for each disease were obtained from the MESH database, and the semantic similarity between diseases was calculated using the directed acyclic graph (DAG) to obtain the disease semantic similarity. According to the circRNA-disease association matrix, the circRNA Gaussian interaction spectrum kernel similarity and the disease Gaussian interaction spectrum kernel similarity were calculated. According to the disease semantic similarity and the circRNA-disease association matrix, the circRNA functional similarity was calculated. By integrating complementary information from multiple data sources and different representation methods, integrated similarity was used to quantify the similarity of each pair of diseases to overcome inherent sparsity, and the circRNA integrated similarity matrix and the disease integrated similarity matrix were obtained.

4. The circRNA-disease association prediction method based on graph convolutional neural network and Node2vec according to claim 1, characterized in that: In step 6, specifically: The node information obtained in the first two steps was sent to the random forest classifier to predict the potential circRNA-disease association relationship, and the AUC value and AUPR value of the present invention were obtained by five-fold cross validation to obtain the prediction results.

Images