RNA-protein binding site discrimination method based on convolutional block attention mechanism dense network

Abstract

The invention relates to an RNA-protein binding site discrimination method based on a convolutional block attention mechanism dense network, and belongs to the field of bioinformatics. Firstly, a public RNA-protein binding site data set is acquired, then, a feature space is constructed by adopting four feature extraction methods, and finally, a dense connection convolutional network improved based on a convolutional block attention mechanism is used for learning a long-term dependency relationship in a combination capture sequence of advanced local features, a bidirectional gating circulation network and a self-attention mechanism. The RNA-protein binding site is discriminatively analyzed by creatively using the dense network improved based on the convolutional block attention mechanism in combination with the idea of the bidirectional gating loop network and the self-attention mechanism, richer feature information can be learned and extracted, and then the model discrimination accuracy is improved.

Description

Technical Field

[0001] The present invention relates to an RNA-protein binding site discrimination method based on a convolutional block attention mechanism dense network, and belongs to the field of bioinformatics. Background Art

[0002] Circular RNA (circRNA) is a class of conservative single-stranded RNA molecules. Unlike classical linear RNA, circRNA forms a covalently closed, continuous and stable loop without a 5' end cap and a 3' poly(A) tail. However, circRNA was originally considered to be a byproduct of abnormal RNA splicing and has long been ignored. With the development of bioinformatics, a large number of circRNAs have been identified, and some of their characteristics have become increasingly clear.

[0003] Emerging evidence suggests that circRNAs play a role in various human diseases, especially cancer, and may serve as better discriminant biomarkers and therapeutic targets for cancer treatment. Considering their potential clinical significance, circRNAs have become a new research hotspot in the field of tumor pathology.

[0004] Given the increasing recognition of the critical roles of circRNAs in disease pathophysiology, their potential as therapeutic targets and diagnostic biomarkers warrants further investigation. Therefore, it is of great significance to investigate the interaction mechanisms between circRNAs and RBPs.

[0005] In recent years, many computational methods for discriminating circRNA-RBP interactions have gradually emerged due to the high cost of high-throughput sequencing technology in detecting the binding sites of each RNA and RBP pair. For example, CRIP uses a coding scheme based on stacked codons. CSCRSites use multiple convolutional layers to extract features from the original circRNA sequence. CRPBsites only use the CircRNA2Vec coding scheme to extract sequence features. In addition, circ-psBlA and circRB use the one-hot method as the encoding mechanism. In the field of bioinformatics, convolutional neural networks (CNNs) are often used to extract local feature information from the original sequence. Recurrent neural networks (RNNs) are often used to extract global feature information from the original sequence. However, CNNs have problems such as gradient disappearance and network degradation. Moreover, when processing long sequence features, RNNs have problems with gradient disappearance or gradient explosion. Therefore, researchers prefer to use residual neural networks (ResNet), dense convolutional networks (DenseNet), long short-term memory networks (LSTM), and gated recurrent units (GRU). ResNet prevents the problems of gradient disappearance and network degradation by short-circuiting connections. Based on ResNet, DenseNet proposed a dense block structure that uses a dense connection mechanism to reuse the feature information of all previous layers. Therefore, compared with ResNet, DenseNet is more effective. iCircRBP-DHN uses multiple feature subsets and a multi-scale residual network architecture to train a deep neural network for identifying circRNA-RBP binding sites. However, existing methods still have limitations: First, some studies have shown that using multiple feature encoding methods to analyze circRNA sequences can improve the accuracy of the model compared to models that only use a single encoding method. The basic principles of these methods are reflected in the literature: circRNA sequences contain rich short-range and long-range features. Therefore, it is crucial to choose a suitable sequence encoding method in order to effectively capture these short-range and long-range features. Reasonable combination of different sequence encoding methods is expected to further improve the accuracy of binding site discrimination. Second, for RBP binding site discrimination models, there are both CNN models that focus on local feature information and RNN models that focus on global feature information. However, studies have shown that these single models cannot fully capture the complex features in the sequence. In contrast, hybrid models can extract local and global feature information at the same time, thus showing better performance. By combining local and global features, these hybrid models are able to more accurately identify RBP binding sites. Therefore, the development of hybrid models represents a quest for more comprehensive sequence information utilization and is expected to promote further developments in the field of RBP binding site identification. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide an RNA-protein binding site discrimination method based on a convolutional block attention mechanism dense network. The traditional discrimination method has the problems of insufficient feature information extraction and inaccurate discrimination.

[0007] The technical solution of the present invention is: a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanism, characterized in that it includes the following steps:

[0008] Step 1: Obtain a public RNA-protein binding site dataset and divide it into training set, test set and validation set in proportion, extract sequence features, connect sequence features, and obtain a new feature matrix;

[0009] Step 2: Build a deep learning network based on the convolutional block attention mechanism dense network, including a dense network improved based on the convolutional block attention mechanism, a bidirectional gated recurrent unit, and a self-attention mechanism;

[0010] Step 3: Use the constructed deep learning network to extract local feature information and global feature information from the new feature matrix, and output the combined feature information;

[0011] Step 4: Set class weights;

[0012] Step 5: Input the combined feature information into the classification module for model training;

[0013] Step 6: Input the data to be tested into the trained model to identify the protein binding sites on the RNA sequence.

[0014] The step 1 is specifically as follows:

[0015] According to the physicochemical properties of RNA sequences, four encoding methods were used to extract sequence features. The length of RNA sequences in the dataset was 101. The sequence features with a feature matrix of 101×41 were extracted by combining the hybridization nucleotide frequency with the encoding of mononucleotides, dinucleotides and trinucleotides. The position and frequency information of nucleotides were obtained by nucleotide density, and the sequence features with a feature matrix of 101×1 were extracted. The four nucleotides A, G, C and U were encoded as (1,1,1), (1,0,0), (0,1,0) and (0,0,1) by the chemical properties of nucleotides, respectively, to obtain a sequence feature of size 101×3. A feature matrix of size 101×11 was obtained by the physicochemical properties of dinucleotides. The four sequence features were connected to obtain a new feature matrix of size 101×56.

[0016] The step 2 is specifically as follows:

[0017] The deep learning network based on the dense network of the convolutional block attention mechanism includes a dense network improved based on the convolutional block attention mechanism, a bidirectional gated recurrent unit and a self-attention mechanism. The dense network structure improved based on the convolutional block attention mechanism introduces the convolutional block attention mechanism in the dense block, evaluates the feature map of each convolution layer in the dense block from the channel and spatial dimensions, sets the number of dense blocks to 4, and in the dense network, in order to solve the problem of a large number of channels in the feature map and parameter explosion, a transition layer is used to optimize this problem. The transition layer consists of 1×1 convolution and 2×2 average pooling. Through 1×1 convolution, the number of channels in the feature map can be reduced, and the complexity of the model can be reduced. At the same time, 2×2 average pooling reduces the size of the feature map, further reducing the number of parameters and calculations, and then the bidirectional gated recurrent unit and the self-attention mechanism are connected after the improved dense network;

[0018] In the dense block, X l The input of a layer is related to the output of all previous layers, X l It is expressed as:

[0019] X l =H l ([X0,X1,X2,…,X l-1 ])

[0020] Where [] indicates a connection operation, H l (·) is a nonlinear transformation function, which consists of batch normalization, ReLU activation function and 3×3 convolution; where l represents a layer in the network (l=0,1,2,…), X l is the output feature of the lth layer, [X0,X1,X2,…,X l-1 ] means that the output features from layer 0 to layer l-1 will be concatenated in the channel dimension, H l (·) is a nonlinear transformation function, which consists of batch normalization, ReLU activation function and 3×3 convolution, and finally obtains the output of the lth layer;

[0021] The convolutional block attention module is an attention module that can combine space and channels at the same time. The channel attention module uses global maximum pooling and global average pooling to compress the original feature map to obtain feature vectors of two channel dimensions. These two feature vectors are trained by a shared fully connected neural network, and then the two feature vectors are added with the corresponding elements to finally obtain the weight coefficient of the channel attention after sigmoid:

[0022] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0023] Among them, M c(F) is the output of the channel attention module, F is the input feature map, AvgPool and MaxPool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents the sigmoid activation function.

[0024] Finally, the normalized weights are multiplied by the input features to obtain the new features

[0025] The spatial attention module focuses on the location information of the feature map to solve the problem of focusing on the location. The spatial attention module first performs channel dimension compression by using global maximum pooling and global average pooling operations on the input feature map, then connects the two feature matrices in the channel dimension, and finally uses convolution operation to obtain the weight coefficient of spatial attention:

[0026] M s (F') = σ(f 7×7 [AvgPool(F');MaxPool(F')])

[0027] Among them, M s (F') is the output of the spatial attention module, f 7×7 Represents a 7x7 convolution operation, [AvgPool(F'); MaxPool(F')] means concatenating the average pooling and maximum pooling results along the channel axis.

[0028] The weight coefficient M s (F') is multiplied by F to get the final output feature, the formula is as follows:

[0029]

[0030] Among them, F” is the final output feature.

[0031] The step 3 is specifically as follows:

[0032] The new feature matrix is ​​input into the constructed deep learning network based on the dense network with convolutional block attention mechanism, wherein the dense network improved based on the convolutional block attention mechanism extracts local feature information from the new feature matrix, and the bidirectional gated recurrent unit and the self-attention mechanism extract global feature information, and finally outputs the feature information which is a combination of the local feature information and the global feature information.

[0033] The step 4 is specifically as follows:

[0034] Setting class weights automatically adjusts the weights of unbalanced samples to balance the impact of unbalanced samples in training. In this study, the ratio of positive samples to negative samples in the baseline dataset is 1:1. Therefore, for the positive sample weight, we set the class weight parameter to 1, and for the negative sample weight, we set the class weight parameter to 1. The dataset used this time is a balanced dataset. When the dataset is unbalanced, just adjust the ratio appropriately.

[0035] The step 5 is specifically as follows:

[0036] First, set the "early stopping method" trigger mechanism and the maximum number of model training rounds threshold, then input the combined feature information into the classification module, use the training set combined with label data, cross entropy loss function and back propagation algorithm to train the model, and determine the parameters of each branch of the model. Finally, after each round of training, input the test set for testing. If the early stopping condition is triggered or the maximum number of model training rounds threshold is reached, the model training is completed, otherwise return to step 3.

[0037] The trigger mechanism of the "early stopping method" is specifically:

[0038] The volume of RNA-protein dataset is quite large. If the number of training rounds is too small, insufficient feature extraction may occur; if the number of rounds is too large, overfitting may occur. Therefore, while setting a reasonable number of training rounds for the model, it is also necessary to set the trigger mechanism of the "early stopping method"; when this mechanism is triggered during model training, or when the number of model training rounds reaches the maximum round threshold, the model training should be stopped; in addition, when the model accuracy reaches the expected standard, the current model needs to be saved; the triggering condition for early stopping is the custom training round, and when the loss of the validation set does not decrease within the custom training round, the "early stopping method" mechanism is triggered.

[0039] The classification module is specifically:

[0040] The classification module is responsible for judging whether the RNA base sequence is a protein binding site based on the input features and outputting the results; the classification module consists of 1 flattening layer, 3 fully connected layers and 2 random dropout layers, where the 3 fully connected layers have 240, 40 and 2 neurons respectively, and the dropout rates of the 2 random dropout layers are 0.5 and 0.2 respectively, which are used to classify the input data.

[0041] The beneficial effects of the present invention are as follows: compared with the prior art, the present invention creatively uses a dense network based on an improved convolutional block attention mechanism combined with a bidirectional gated recurrent network and the idea of ​​a self-attention mechanism to discriminate and analyze RNA-protein binding sites, which can learn to extract richer feature information and improve the model's discrimination accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0043] Figure 2 It is a schematic diagram of the physical and chemical properties of RNA sequences using four encoding methods;

[0044] Figure 3 is the AUC value of the model on 13 RNA-protein binding site datasets. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0046] Example 1: Figure 1 As shown, a method for distinguishing RNA-protein binding sites based on a dense network with a convolutional block attention mechanism comprises the following steps:

[0047] Step 1: The circRNA sequences in each sub-dataset are from the CircRNA interactome database. CD-HIT with a threshold of 0.8 is used to eliminate redundant sequences in the sample to reduce the errors caused by high similarity sequences between different methods. For the obtained public RNA-protein binding site datasets, such as Figure 2 , considering the physical and chemical properties of RNA sequences, four encoding methods are used to extract sequence features, and the four encoding features are connected to obtain a new feature matrix; the RNA sequence length is 101, and four encoding schemes are used to obtain four feature vectors. (1) Hybridization nucleotide frequency (HNF) combines the encoding of mononucleotides, dinucleotides and trinucleotides to extract sequence features with a feature matrix of 101×41. (2) Nucleotide density (ND) obtains the position and frequency information of nucleotides to extract sequence features with a feature matrix of 101×1. (3) Nucleotide chemical properties (NCP) can encode A, G, C and U as (1,1,1), (1,0,0), (0,1,0) and (0,0,1) respectively to obtain a feature matrix of 101×3. (4) Dinucleotide physical and chemical properties (DPCP) obtains a feature matrix of 101×11. The four encoding features are connected to obtain a new feature matrix of 101×56.

[0048] Step 2: The dense network structure improved based on the convolutional block attention mechanism has been changed from the original dense network, mainly by introducing the convolutional block attention mechanism in the dense block, evaluating the feature map of each convolutional layer in the dense block from the channel and spatial dimensions, setting the number of dense blocks to 4, and in the dense network, in order to solve the problem of large number of feature map channels and parameter explosion, we use the Transition layer to optimize this problem. The transition layer consists of 1×1 convolution and 2×2 average pooling. Through 1×1 convolution, the number of channels in the feature map can be reduced, reducing the complexity of the model. At the same time, 2×2 average pooling reduces the size of the feature map, further reducing the number of parameters and calculations, and then the improved dense network is followed by a bidirectional gated recurrent unit and a self-attention mechanism;

[0049] In the dense block, X l The input of a layer is related to the output of all previous layers, X l It is expressed as:

[0050] X l =H l ([X0,X1,X2,…,X l-1 ])

[0051] Where [] indicates a connection operation, H l (·) is a nonlinear transformation function, which consists of batch normalization, ReLU activation function and 3×3 convolution;

[0052] The convolutional block attention module is an attention module that can combine both space and channels. The channel attention module uses global maximum pooling and global average pooling to compress the original feature map to obtain feature vectors of two channel dimensions. These two feature vectors are trained by a shared fully connected neural network. Then the two feature vectors are added to their corresponding elements, and finally the weight coefficient of the channel attention after sigmoid is obtained:

[0053] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0054] Finally, the normalized weights are multiplied by the input features to obtain the new features

[0055] The spatial attention module focuses on the location information of the feature map to solve the problem of focusing on the location. The spatial attention module first performs channel dimension compression by using global maximum pooling and global average pooling operations on the input feature map, then connects the two feature matrices in the channel dimension, and finally uses convolution operation to obtain the weight coefficient of spatial attention:

[0056] M s(F') = σ(f 7×7 [AvgPool(F');MaxPool(F')])

[0057] Finally, the weight is multiplied by the input feature matrix to obtain the final feature

[0058] Step 3: Connect the four encoding features to obtain a new 101×56 feature matrix, which is input into the deep learning network based on the dense network with convolutional block attention mechanism. The dense network improved based on the convolutional block attention mechanism extracts more advanced local feature information from the new feature matrix, and the bidirectional gated recurrent unit and self-attention mechanism are used to extract global feature information. The final output feature information is a combination of these two features, resulting in richer feature information.

[0059] Step 4: Set the class weight to automatically adjust the weight of the unbalanced samples to balance the impact of the unbalanced samples in training. In this study, the ratio of positive samples to negative samples in the baseline dataset is 1:1. Therefore, for the positive sample weight, we set the class weight parameter to 1, and for the negative sample weight, we set the class weight parameter to 1. The dataset used this time is a balanced dataset. When the dataset is unbalanced, you can adjust the ratio appropriately. This setting is to reduce the impact of unbalanced samples during model training by simply adjusting the class weight ratio when the model training dataset is an unbalanced dataset;

[0060] Step 5: Input the feature vector obtained by the deep learning network based on the convolutional block attention mechanism dense network into the classification module, train the model with the label data, cross entropy loss function and back propagation algorithm, optimize the parameters of each branch of the model, and input the test set for testing after each round of training;

[0061] Step 6: Set the "early stopping method" trigger mechanism to prevent the model from overfitting, repeat steps 3-5 until the early stopping condition is triggered or the number of training rounds reaches the set threshold, and finally save the model;

[0062] Step 7: Call the model saved in step 6, input the test data to identify the protein binding sites on the RNA sequence, and experimentally verify the model on the RNA-protein binding site dataset, and follow the evaluation indicators widely used in the field of RNA-protein interaction: area under the ROC curve (AUC), Matthews correlation coefficient (MCC), accuracy (ACC) and F1 score. The most critical indicator in this study is AUC, which ranges from 0.5 to 1 and is positively correlated with performance. These indicators are defined as follows:

[0063]

[0064]

[0065] Among them, TP is a true positive example, FN is a false negative example, FP is a false positive example, and TN is a true negative example.

[0066] The experimental results are shown in Table 1 and Figure 3 As shown:

[0067] Table 1 Effect of the model on 13 RNA-protein binding site datasets

[0068]

[0069] The specific implementation modes of the present invention are described in detail above (combined with the accompanying drawings), but the present invention is not limited to the above implementation modes. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for RNA-protein binding site discrimination based on a convolutional block attention mechanism dense network, characterized in that: The following steps are involved: Step 1: Obtain a public RNA-protein binding site dataset and divide it into training set, test set and validation set in proportion, extract sequence features, connect sequence features, and obtain a new feature matrix; Step 2: Build a deep learning network based on a dense network of convolutional block attention mechanisms; Step 3: Use the constructed deep learning network to extract local feature information and global feature information from the new feature matrix, and output the combined feature information; Step 4: Set class weights; Step 5: Input the combined feature information into the classification module for model training; Step 6: Input the data to be tested into the trained model to identify the protein binding sites on the RNA sequence.

2. According to claim 1, a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanism, characterized in that: The step 1 is specifically as follows: according to the physicochemical properties of the RNA sequence, four encoding methods are used to extract sequence features; the length of the RNA sequence in the data set is 101, and the sequence features with a feature matrix of 101×41 are extracted by combining the encoding of mononucleotides, dinucleotides and trinucleotides with the frequency of hybridized nucleotides, and the position and frequency information of the nucleotides are obtained by the nucleotide density, and the sequence features with a feature matrix of 101×1 are extracted; the four nucleotides A, G, C and U are encoded as (1,1,1), (1,0,0), (0,1,0) and (0,0,1) respectively by the chemical properties of the nucleotides to obtain a sequence feature with a size of 101×3; a feature matrix with a size of 101×11 is obtained by the physicochemical properties of dinucleotides, and the four sequence features are connected to obtain a new feature matrix of 101×56.

3. According to claim 1, a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanism, characterized in that: The step 2 is specifically as follows: Step 2.1: The new feature matrix F is processed by the channel attention module, and the feature matrix F is compressed by global maximum pooling and average pooling to obtain two feature vectors. After learning with the same fully connected neural network, they are added and the weight parameter M is obtained by the sigmoid activation function. c (F), calculated as: M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) Among them, AvgPool and MaxPool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents sigmoid activation function; Step 2.2: Set the weight parameter M c (F) is multiplied by F to obtain the combined feature F', which is calculated as: Step 2.3: The combined feature F' is processed by the spatial attention module, compressed and connected using global maximum pooling and average pooling, and the weight coefficient M is obtained by convolution s (F'), calculated as: M s (F')=σ(f 7×7 [AvgPool(F');MaxPool(F')]) Among them, f 7×7 Represents a 7x7 convolution operation, [AvgPool(F'); MaxPool(F')] means concatenating the average pooling and maximum pooling results along the channel axis; Step 2.4: Set the weight coefficient M s (F') is multiplied by F to get the final output feature, and the calculation formula is: Among them, F” is the final output feature.

4. According to claim 1, a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanism, characterized in that: The step 3 is specifically as follows: inputting the new feature matrix into the constructed deep learning network based on the dense network with convolutional block attention mechanism, wherein the dense network improved based on the convolutional block attention mechanism extracts local feature information from the new feature matrix, and the bidirectional gated recurrent unit and the self-attention mechanism extract global feature information, and finally outputs the feature information which is a combination of the local feature information and the global feature information.

5. According to claim 1, a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanism, characterized in that: The setting of the class weight specifically includes: setting the class weight to automatically adjust the weight of the unbalanced sample to balance the impact of the unbalanced sample in training, and the class weight value is adjusted according to the data set type.

6. According to claim 1, a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanism, characterized in that: The step 5 is specifically as follows: first, the "early stopping method" trigger mechanism and the maximum number of model training rounds threshold are set, and then the combined feature information is input into the classification module, and the model is trained using the training set combined with the label data, the cross entropy loss function and the back propagation algorithm to determine the parameters of each branch of the model. Finally, after each round of training is completed, the test set is input for testing. If the early stopping condition is triggered or the maximum number of model training rounds threshold is reached, the model training is completed, otherwise, return to step 3.

7. According to claim 6, a method for distinguishing RNA-protein binding sites based on a dense network of convolutional block attention mechanisms, characterized in that: The triggering early stopping condition is specifically: the triggering early stopping condition is a custom training round, and when the loss of the validation set does not decrease within the custom training round, the "early stopping method" mechanism is triggered.

8. According to claim 6, a method for distinguishing RNA-protein binding sites based on a convolutional block attention mechanism dense network is characterized in that: The classification module is specifically: the classification module judges whether the RNA base sequence is a protein binding site based on the input feature information and outputs the result. The classification module consists of a Flatten layer, a fully connected layer and a Dropout layer, and is used to classify the input feature information.

Images