A method for predicting RNA-disease correlation based on graph attention network

Abstract

The application belongs to the field of bioinformatics and relates to an RNA-disease correlation prediction method based on a graph attention network, aiming to improve the prediction performance of the graph attention network on RNA-disease correlation. First, the graph structure of the graph attention network is used as a training network, and a graph data structure is constructed based on a PiRNA dataset; second, the structure of the graph is changed, and the adjacency matrix of the undirected graph is converted into a weighted adjacency matrix, wherein the weight of the adjacency matrix is the number of times of gene occurrence, and the main diagonal of the adjacency matrix is allowed to have a value; finally, after all the features are sparse matrixed, they are input into the graph attention network, binary classification is performed, and the maximum probability is taken as the prediction result. piRNA is abnormally expressed in most cancer cells, and this RNA-disease correlation prediction method based on the graph attention network can improve the prediction accuracy and has important significance for preventing and treating diseases.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and relates to an RNA-disease correlation prediction method based on graph attention networks, including a method for transforming weighted adjacency matrices. Background Technology

[0002] Piwi protein, representing a branch of the Ago protein family, was first discovered in Drosophila and plays a role in regulating the maintenance of germline stem cells. RNAs that interact with Piwi protein are called PiRNAs. Piwi protein and piRNAs are commonly detected in the human reproductive system. Studies have found that PiRNAs are aberrantly expressed in various cancers, suggesting that aberrant PiRNA expression may contribute to human diseases. Traditional biological experimental techniques, such as ribosome assays and transcriptome sequencing, are generally reliable for identifying potential associations between piRNAs and human diseases. With the emergence of piRNA-disease association databases, such as piRNABank, piRBase, and piRNAQuest, these known association databases provide a solid foundation for deep learning prediction, thus partially addressing the time-consuming, expensive, and labor-intensive problems of traditional biological experimental methods. Currently, most computational prediction methods only consider the attribute features of piRNAs. However, in databases, one piRNA may be associated with multiple diseases, ignoring the connections between diseases. Therefore, it is necessary to design attribute features that can associate them. As the combinations of relationships between non-coding RNAs and diseases become increasingly complex, predicting disease associations can provide deeper insights into many biological mechanisms.

[0003] The PiRNAs database has been established, and many computational methods have been proposed. Related methods such as neural networks and support vector machines have been successfully applied to predict the correlation of RNA-related diseases. However, most neural network methods have the following drawbacks: they are simply feature mappings and do not establish connections between data.

[0004] The most significant innovation of this invention is the proposal of a novel graph data structure that decomposes the entire network graph structure into different sub-networks, thereby combining a weighted adjacency matrix with a graph attention network. The weighted adjacency matrix designed in this invention not only utilizes attribute features but also reflects the relationships between them through the adjacency matrix itself. Summary of the Invention

[0005] An RNA-disease correlation prediction method based on graph attention networks includes three processes: preprocessing node sequences, training the graph attention network, and testing the graph attention network. The specific steps are as follows:

[0006] Step 1: To extract more binding information from RNA-disease sequences, it is necessary to use... p indivual( p Indicates the number of nodes. p The size is determined by the number of RNA-disease sequences. Different nodes, each containing two pieces of information, one is the disease, and the other is the RNA.

[0007] Step 2: Construct an adjacency matrix between these nodes. When any two nodes contain the same RNA, a connection is established, and the value in the adjacency matrix is ​​the reciprocal of the frequency of that RNA. The RNA frequency is obtained by counting the nodes containing that RNA.

[0008] Step 3: The product generated in Step 2... p Each node serves as input to the graph attention network. The RNA data obtained from K-mer counting is divided into 5 parts for 5-fold cross-validation. K-mer refers to a segment of length... k The gene fragments are obtained by cutting a portion of sequencing reads. k It is an odd number;

[0009] Step 4: Save the trained graph attention network model and use the test set to evaluate the model.

[0010] An RNA-disease correlation prediction method based on graph attention networks, the implementation process of step 1 is as follows:

[0011] Determine the number of RNA-disease sequences to process. p , p The size is equal to the number of data points in the dataset. Each sequence is equivalent to a node. Once the number of nodes is determined, a system is established. p*p phalanx.

[0012] An RNA-disease correlation prediction method based on graph attention networks, the implementation process of step 2 is as follows:

[0013] Use the product generated in step 1 p*p square array A ;set up i These are the row numbers of the square array. j It is the column number of the square matrix. f For the first i Line 1 j The weight of a column position is determined by the number of genes. The weight of a node relative to itself is set to 1, meaning all diagonal elements are set to 1, while other unrelated elements are set to 0.

[0014] An RNA-disease correlation prediction method based on graph attention networks, the implementation process of step 3 is as follows:

[0015] The labels are one-hot encoded, and then the weighted adjacency matrix and RNA extraction information from step 2 are input into the graph attention network for five-fold cross-validation. The graph attention network has two layers. The first layer uses the ReLU activation function, with 609 graph data nodes, 2424 data samples, and 64 output nodes. The second layer is the output layer, using the Softmax activation function, and has 2 output nodes.

[0016] An RNA-disease association prediction method based on graph attention network is proposed. The implementation process of step 4 is as follows: save the graph attention network model trained in step 3 for predicting the input RNA sequence data. Attached Figure Description

[0017] Figure 1 This is a flowchart of a graph attention-based method.

[0018] Figure 2 This is a flowchart of data node processing.

[0019] Figure 3 It is a method for representing the relationships between nodes and converting them into an adjacency matrix.

[0020] Figure 4 This is an introduction to K-mer counting.

[0021] Figure 5 This is a diagram of the network architecture. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0023] The purpose of this invention is to propose an RNA-disease correlation prediction method based on graph attention networks, comprising three processes: preprocessing RNA-disease sequences, training the graph attention network, and testing the graph attention network. The flowchart of this method is as follows: Figure 1 As shown, the specific process includes the following steps:

[0024] Step 1: Determine the matrix size to be used for processing RNA-disease sequences:

[0025] p It is the number of nodes; for example... Figure 2 As shown, each square represents a node, containing two pieces of information: RNA and disease; there are a total of p indivual ， This leads to the construction of a matrix whose number of rows and columns equals the total number of nodes. p， Therefore, the size of the matrix is p*p .

[0026] Step 2: Establish connections between different nodes. If any two nodes share the same RNA, it means they belong to the same subnetwork. Then, fill in the corresponding position in their adjacency matrix with the reciprocal of the frequency of this RNA. For example... Figure 3 As shown, if nodes 1 and 2 have the same RNA, then the positions (1, 2) and (2, 1) in the adjacency matrix are set to f1. Because it is an undirected graph, the positions in row 1, column 2 and row 2, column 1 of the adjacency matrix are both set to the reciprocal of the frequency f1. Nodes 3, 4, 5, and 6 have the same RNA, so they belong to the same network. The positions (3, 4) and (4, 3), (3, 5) and (5, 3), (3, 6) and (6, 3), (4, 5) and (5, 4), (4, 6) and (6, 4), (5, 6) and (6, 5) in the adjacency matrix are all set to the reciprocal of the frequency of this RNA, f2. In graph terms, this means that there are direct connections between each of them.

[0027] Step 3: Using the weighted adjacency matrix generated in Step 2 as the network input, divide the RNA data into 5 parts and perform 5-fold cross-validation. The RNA data is obtained by K-mer counting, such as... Figure 4 As shown, the 3-mer counting method involves dividing the sequence into three parts and counting the number of occurrences in each part. The attention network structure is shown in the figure. Figure 5As shown, the Tensorflow framework, Keras, Sklearn, NumPy libraries, and the GraphAttention package are used. Several functions in Keras are called: first, the categorical_crossentropy cross-entropy loss function is used, Adam is used as the training optimizer, the EarlyStopping function is used for early stopping to prevent overfitting due to continuous training, the ModelCheckpoint function saves the optimal model, the evaluate function evaluates the model, and plot_model plots the network model. Then, functions such as roc_curve and auc in sklearn are called to calculate the model's AUC. The model consists of two attention layers. First, graph nodes are randomly dropped with a probability of 0.01 (Dropout operation). The first graph attention layer consists of an input of (609, 609) graph nodes and (2424, 2424) RNA data, using ReLU activation and an output of randomly dropped nodes with a probability of 0.02 (Dropout operation). It uses eight attention heads and L2 regularization. The second layer consists of an input of a (64, 64) feature map from the first layer and (2424, 2424) RNA data, using one attention head, L2 regularization, and a Softmax activation function. It outputs two nodes for binary classification. The optimal parameters for the model are shown below. The probability of randomly dropping nodes is 0.01, primarily used in the input layer to randomly drop data, ensuring slight differences in the training data for each iteration. This parameter should not be set too high, otherwise, the network may not converge.

[0028] Table 1. Optimal model parameters.

[0029] Parameter name Parameter value Number of iterations 200 Batch size 128 Probability of randomly discarding nodes 0.01 The first layer of the graph's attention network outputs feature numbers. 8 First-layer graph attention network, number of attention heads 8 Learning rate <![CDATA[5×10 -2 ]]> Regularization parameters <![CDATA[1.25×10 -4 ]]>

[0030] Step 4: Save the model from Step 3 using ModelCheckpoint, saving only the model parameters as an .h5 file. Then, test it with new data. The input data consists of new RNA data and the previously created adjacency matrix. AUC and ACC are used as evaluation metrics, and the final prediction result is the average of the 5-fold cross-validation results, achieving efficient and accurate predictions. The model proposed in this invention achieves an average AUC of 0.91 and an ACC of 0.865 on the 5-fold cross-validation dataset. SVM's AUC is 0.856 and ACC is 0.815, while Att-based methods have an AUC of 0.873 and an ACC of 0.8265. Compared to these methods, experiments demonstrate that the model proposed in this invention achieves the best prediction results.

[0031] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An RNA-disease correlation prediction method based on graph attention networks, comprising three processes: preprocessing RNA-disease association pairs, training the graph attention network, and testing the graph attention network. The specific steps are as follows: Step 1: Obtain p RNA-disease association pairs, each association pair containing an RNA and a disease information; determine p nodes based on the RNA-disease association pairs, each node corresponding to an RNA-disease association pair; Step 2: Construct the adjacency matrix of these nodes, where, If two nodes contain the same RNA, a weight is assigned to the matrix position corresponding to the two nodes, and the weight is the reciprocal of the total number of nodes containing the RNA; the weight between a node itself is set to 1, and the weight between unrelated nodes is set to 0. Step 3: Input the weighted adjacency matrix and RNA-extracted information from Step 2 into the graph attention network and perform 5-fold cross-validation; Step 4: Save the trained graph attention network model and use the saved model to predict the input RNA-disease association pairs data.

2. The RNA-disease correlation prediction method based on graph attention network according to claim 1, the implementation process of step 1 is as follows: Determine the number of RNA-disease association pairs processed. p Each sequence is equivalent to a node. Once the number of nodes is determined, a system is established. p * p phalanx.

3. The method for predicting RNA-disease correlation based on graph attention networks according to claim 1, the implementation process of step 2 is as follows: Use the product generated in step 1 p * p square array A ;set up i These are the row numbers of the square array. j It is the column number of the square matrix. f For the first i Line 1 j The weight of each column position is determined by the statistical analysis of the number of genes. f It equals the reciprocal of the gene frequency; the weight of a node relative to itself is set to 1, that is, all elements on the diagonal are set to 1, and other unrelated parts are set to 0.

4. The method for predicting RNA-disease correlation based on graph attention networks according to claim 1, the implementation process of step 3 is as follows: The labels are encoded one-hot, and then the weighted adjacency matrix and RNA-extracted information from step 2 are input into the graph attention network. The data is divided into 10 parts, 5 for training and 5 for validation. The attention network has 2 layers. The first layer uses the ReLU activation function, the number of graph data nodes is 609, the number of data samples is 2424, and the output is 64 nodes. The second layer is the output layer, which uses the Softmax activation function and outputs 2 nodes.

Images