Method for predicting binding site of RNA binding protein and circular RNA
By using variable-length RNA binding site sequences and a hybrid neural network architecture, the accuracy problem of predicting RNA-binding protein binding sites and circular RNA binding sites was solved, achieving higher-precision prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2022-11-10
- Publication Date
- 2026-07-24
Smart Images
Figure CN115881215B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics, and in particular to a method for predicting the binding site of RNA-binding proteins and circular RNA. Background Technology
[0002] RNA-binding proteins (RBPs) are a general term for ubiquitous proteins that can bind to RNA. RBPs regulate RNA functions directly or indirectly by specifically binding to RNA, such as mediating RNA maturation, transport, localization, and translation. Circular RNAs (circRNAs) are a special class of non-coding RNA molecules and a recent research hotspot in the field of RNA. Circular RNAs play an important role in various biological processes and have become an emerging hotspot in RNA research in recent years. They influence transcription by binding to specific proteins, thereby participating in the regulation of biological processes. In recent years, various deep learning frameworks have been used to predict RBP binding sites on circular RNAs.
[0003] However, these methods typically only perform single-level feature extraction on sequence information, which leads to insufficient information acquisition. Furthermore, current common networks for predicting the binding sites of circular RNA and proteins require fixed-length nucleotide sequences as input and cannot accept variable-length inputs. Fixed-length inputs may introduce additional interference, thereby affecting the accuracy of binding site prediction. Therefore, this invention proposes a method for predicting the binding sites of RNA-binding proteins and circular RNA to solve the problems existing in the prior art. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to propose a method for predicting the binding sites of RNA-binding proteins and circular RNA, thereby solving the problem of low accuracy in predicting the binding sites of existing RNA-binding proteins and circular RNA.
[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for predicting the binding site of RNA-binding proteins and circular RNA, comprising the following steps:
[0006] Step 1: First, in a network that combines deep and shallow features, use the variable-length RNA binding site sequence as the input feature. Then, use the word2vec model and the skip word algorithm to encode the RNA binding site sequence and convert the input feature into an embedded feature vector. Finally, expand the shallow features in the network in a multi-scale manner.
[0007] Step 2: First, use a hybrid architecture of convolutional neural network and bidirectional long short-term memory network to process the transformed embedded feature vector. Then, use the input feature signal to adaptively remember, forget and update the memory of information. Use a multi-scale downsampling structure to expand deeper features. Decode the features before entering the recurrent neural network to expand the deep features.
[0008] Step 3: After the expanded deep and shallow features are fused, they are fed into the fully connected layer of the classification network for classification. Finally, the classification results are used to determine whether they are binding sites, thus enabling the prediction of the binding sites of RNA-binding proteins and circular RNA.
[0009] A further improvement is made in the following steps: In step one, the specific steps for converting the input features into embedded feature vectors are as follows: a window of size k=3 is slid across a circular RNA sequence of length L with a sliding step size s=1 to divide the sequence into L-k+1 words of length k-mer, and these words are projected into a multidimensional space to generate a vocabulary. Then, the words are input into the embedding matrix to obtain the corresponding vector representation. The embedding matrix is obtained by training and fine-tuning a pre-trained word2vec model.
[0010] A further improvement is that, in step one, the encoding method for the RNA binding site sequence includes electron-ion interaction pseudopotential, nucleotide chemical properties, and k-spacer nucleotide pair frequency. Specifically, the electron-ion interaction pseudopotential encoding method encodes the RNA sequence into a digital carrier, the nucleotide chemical property encoding method converts the RNA sequence into a digital carrier, and the k-spacer nucleotide pair frequency encoding method converts the RNA sequence into a digital carrier.
[0011] Further improvements involve combining the word2vec encoding method with feature encoding methods including electron-ion interaction pseudopotentials, nucleotide chemical properties, and k-spaced nucleotide pair frequencies, and aggregating the expanded features.
[0012] A further improvement is made in step two, where, before entering the bidirectional long short-term memory network, the embedded feature vector is first processed by a convolutional filter to obtain features of different scales and concatenate them to obtain a concatenated feature matrix.
[0013] A further improvement is that, in step two, each long short-term memory unit in the bidirectional long short-term memory network uses a gating mechanism to adaptively memorize, forget, and update information based on input signals.
[0014] Further improvements are made in step two, where the hybrid architecture of the convolutional neural network and the bidirectional long short-term memory network uses batch normalization layers, swish activation functions, and Avepool pooling layers to reduce the network burden. At the same time, random deactivation layers are used to prevent overfitting of features. The features are then fed into the bidirectional long short-term memory network, and the weights of key features are enhanced through an attention mechanism.
[0015] A further improvement is made by using a parallel convolutional neural network to expand the deep features after the features have passed through the random deactivation layer, in order to avoid the limitations of the neural recurrent network.
[0016] The beneficial effects of this invention are as follows: This invention extracts and expands features from information at different levels. After the expanded deep and shallow features are fused, they are fed into a classification network. Finally, it is determined whether it is a binding site. Through multi-branch learning, feature expansion is performed hierarchically, and information is extracted a second time, which improves the prediction accuracy of the binding site of RNA binding protein and circular RNA. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a diagram of the bidirectional convolutional neural network of the present invention;
[0019] Figure 2 This is the overall network diagram of CRBP-HHFF of the present invention;
[0020] Figure 3 This is a single circuit diagram of different scale features connected in parallel according to the present invention;
[0021] Figure 4 This is a performance comparison chart of the CRBP-HFEF method of the present invention and the CRPBsites method for creating this dataset;
[0022] Figure 5 This is a scatter plot of the network before and after classification according to the present invention;
[0023] Figure 6 This is a diagram of the hierarchical feature extraction structure of the present invention. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] See Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 This embodiment provides a method for predicting the binding site of RNA-binding proteins and circular RNA, including the following steps:
[0026] Step 1: First, in the network (CRBP-HHFF) that combines deep and shallow features, the RNA binding site sequence of variable length is used as the input feature. Then, the word2vec model and skip-gram algorithm are used to encode the RNA binding site sequence and convert the input feature into an embedded feature vector. Then, the shallow features in the network are expanded in a multi-scale manner.
[0027] Extensive use of one-hot encoding leads to extremely sparse sample feature vectors, while the structural characteristics of deep learning are not conducive to processing sparse feature vectors. Word2vec, as an effective embedding method, is used in this embodiment with the word2vec model and skip-gram algorithm to encode the sequence. First, a window of size k=3 slides across a circular RNA sequence of length L with a step size s=1, segmenting the sequence into L-k+1 words of length k-mer. These words are then projected into a multi-dimensional space to generate a vocabulary. The words are then input into an embedding matrix to obtain corresponding vector representations. The embedding matrix is obtained by fine-tuning a pre-trained word2vec model. Finally, the vector representation is converted by combining site information.
[0028] X 1:L-k+1 =[X1,X2,...,X L-k+1 ]
[0029] Encoding methods for RNA binding site sequences include electron-ion interaction pseudopotential (EIIP), nucleotide chemical properties (NCP), and k-spacer nucleotide pair frequency (CKSNAP). For ease of encoding, the base "U" in RNA is replaced with the base "T" in DNA. This base substitution does not affect experimental accuracy; it is only to facilitate the use of some known properties of DNA for encoding. The encoding principle of nucleotide chemical properties is as follows: each nucleotide of A, C, G, and T has a different chemical structure and chemical binding characteristics. Each nucleotide has three chemical properties: ring structure, function, and hydrogen bond.
[0030] In the ring structure, purines "A" and "G" contain two rings, which are encoded as '1', while pyrimidines "C" and "T" contain one ring, which is encoded as '0'.
[0031] Based on chemical function, "A" and "C" contain amino groups and are coded with 1, while "G" and "T" belong to ketone groups and are coded with 0.
[0032] Based on the fact that hydrogen bonds 'A' and 'T' contain weak hydrogen bonds, they are coded as '1'; and "C" and "G" contain strong hydrogen bonds, they are coded as "0";
[0033] Therefore, “A”, “C”, “G”, “T” and “-” are encoded as (1,1,1), (0,1,0), (1,0,0), (0,0,1) and (0,0,0) respectively;
[0034] In k-spacer nucleotide pair frequency, the RNA sequence is converted into a digital vector by calculating the frequency of occurrence of all possible k-spacer nucleotide pairs on the circular RNA sequence. A k-spacer nucleotide pair indicates that there are k nucleotides between two nucleotides. For example, in the sequence “AXXXCXXT”, “AC” is a nucleotide pair with 3 spaces and “CT” is a nucleotide pair with 2 spaces. CKSNAP contains 16 values {AA,AC,AG,…,TG,TT} corresponding to nucleic acid pairs.
[0035] In electron-ion interaction pseudopotentials, the EIIP of a nucleotide describes the distribution of free electron energy along the DNA sequence. The EIIP values of nucleotides A, G, C, T, and "-" are 0.1260, 0.0806, 0.1340, 0.1335, and 0, respectively. Therefore, an EIIP encoding scheme can be used to encode RNA sequences into digital vectors. For example, "GTGTACT" is encoded as a digital vector consisting of (0.0806, 0.1335, 0.0806, 0.1335, 0.1260, 0.1340, 0.1335).
[0036] Step 2: First, use a hybrid architecture of Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory Network (Bi-LSTM) to process the transformed embedded feature vector. Before entering the Bi-LSTM network, use a convolutional filter to process the embedded feature vector, obtain features at different scales, and concatenate them to obtain a concatenated feature matrix. Then, use the input feature signal to adaptively remember, forget, and update the memory of information. Use a multi-scale downsampling structure to expand deeper features. Decode the features before entering the recurrent neural network to expand the deep features.
[0037] In Bi-LSTM, each LSTM unit can adaptively memorize, forget, and update information based on the input signal through a gating mechanism. Regardless of the start and end positions, the RBP binding sites in both directions are crucial. Therefore, in Bi-LSTM, the hidden layers of a bidirectional recurrent neural network must store two values, such as... Figure 1 As shown, A participates in the forward calculation, A' participates in the reverse calculation, and the final output value y depends on A and A'.
[0038] After obtaining the output s of the Bi-LSTM, an attention layer is applied to the feature vector to enhance the weights of key features. Specifically, the initial weight matrix is obtained by transposing the feature matrix s, and the feature weight matrix is calculated from the initial weight matrix through a fully connected layer and a softmax function.
[0039] S = (S1, S1, ... S1) i' )
[0040] A i' =softmax(Dense(S) T ))*S
[0041] Where S is the output of Bi-LSTM, ST is the transpose of S, Dense is a fully connected layer, and finally, the fully connected neural network is used to determine whether the sequence contains a binding site:
[0042] Out=softmax(Dense(flatten(A)))
[0043] The hybrid architecture of convolutional neural networks and bidirectional long short-term memory networks uses batch normalization (BN) layers, swish activation functions, and Avepool pooling layers to reduce the network load. At the same time, random dropout layers are used to prevent features from overfitting. The features are then fed into the bidirectional long short-term memory network, and the weights of key features are enhanced through an attention mechanism. The features after the random dropout layers are expanded into deep features using a parallel convolutional neural network to avoid the limitations of recurrent neural networks (RNNs).
[0044] Step 3: After the expanded deep and shallow features are fused, they are fed into the fully connected layer of the classification network for classification. Finally, the classification results are used to determine whether they are binding sites, thus enabling the prediction of the binding sites of RNA-binding proteins and circular RNA.
[0045] The overall network diagram of CRBP-HHFF is as follows: Figure 2 As shown, the input RNA sequence is transformed into feature descriptors by the level 1 module. The features extracted in level 1 in two ways are processed by convolutional layers. After processing and fusion, the feature vector is sent to Bi-LSTM. At the same time, key features are enhanced by attention mechanisms before entering the classification network. The second-level module expands the shallow features in a multi-scale manner by encoding and decoding the features of the first layer. The third-level module expands the deep features by decoding the features before entering the recurrent neural network. Finally, the features processed by different modules are fused and then entered into the classification.
[0046] The specific expansion of multi-layer features is as follows:
[0047] First-level feature expansion:
[0048] Feature extraction was performed using multiple encoding methods. The word2vec encoding method was combined with features encoding electron-ion interaction pseudopotential (EIIP), nucleotide chemical properties (NCP), and k-spaced nucleotide pair frequency (CKSNAP), and these features were aggregated.
[0049] Second-level feature expansion:
[0050] A multi-scale parallel network is used to fully extract the encoded features. To enrich local feature information, filter kernels of different sizes are used to obtain features at different scales, and these features are then connected in parallel. A single line is shown in the diagram. Figure 3 As shown, a convolutional BN Swish sequence is used to reduce the training challenge and alleviate the burden of propagating large feature maps to subsequent layers. Then, a dropout layer is added after the activation layer to avoid overfitting. After that, the output is passed through a max pooling layer to obtain the maximum value of the spatially adjacent overlapping window to ensure that the most prominent features of the output are output. After the above series of processing, four feature maps are generated. Finally, the feature maps are fused and input into the next level network.
[0051] Third-level feature expansion
[0052] For deeper features, I obtain additional features by incorporating a downsampling branch, and then expand the features through feature fusion.
[0053] Experimental comparison
[0054] (This dataset consists of 4 independent datasets. First, the network is compared with the other network on the 4 datasets, and then compared with more networks on one of the datasets, LIN28B.)
[0055] To compare the performance of the network in this embodiment, we compared the performance of CRBP-HFEF with that of the CRPBsites method used to create this dataset. All samples were divided into training and test sets in an 8:2 ratio. To evaluate the performance of the method proposed in this embodiment and other methods, several traditional evaluation metrics were used. The results are shown in Table 1 below:
[0056] Table 1
[0057]
[0058] Furthermore, this embodiment was compared with other different network methods on the LIN28B dataset, where precision, recall, and F1-score were positive indicators. Figure 4 As shown, the network proposed in this embodiment achieves the best performance in almost all metrics. Specific data is shown in Table 2.
[0059] Table 2
[0060]
[0061] CircRBP-DNH achieved a recall of 0.95 for positive samples. In addition, circSLNN achieved higher accuracy than other methods but had the lowest AUC. This may be because circSLNN received imbalanced positive and negative instances during model training after treating the site prediction problem as a sequence labeling task. The higher precision, recall and F1 score for negative samples also illustrate this point.
[0062] The scatter distribution of the network before and after classification in this embodiment is visualized using t-SNE, as shown below. Figure 5 As can be seen, the network in this embodiment achieved a fairly good classification effect.
[0063] This invention proposes a three-layer hierarchical feature extraction structure. To verify that the modules proposed in this invention are effective at each layer, ablation experiments were conducted on these modules, and the results are as follows: Figure 6As shown, BaseLine is the word2vec feature extraction network used only. Level 1 is the word2vec feature extraction method added to the feature extraction combination. Level 2 is the network after adding a multi-scale feature extraction module, and Level 3 is a downsampling module. Since Level 2 needs to use the features processed by Level 1, Levels 1+2, 1+3, and 1+2+3 were compared and combined. Obviously, by adding the extension of Level 1, the network performance was significantly improved. Subsequently, Levels 2 and 3 were added on the basis of Level 1, and the performance was further improved. This shows that Levels 2 and 3 both have an impact on improving network performance. Finally, the best result was obtained by summing the three levels. This shows that each feature extension added in this invention contributes to the network performance.
[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting the binding site of RNA-binding proteins to circular RNA, characterized in that, Includes the following steps: Step 1: First, in a network that combines deep and shallow features, use the variable-length RNA binding site sequence as the input feature. Then, use the word2vec model and the skip word algorithm to encode the RNA binding site sequence and convert the input feature into an embedded feature vector. Finally, expand the shallow features in the network in a multi-scale manner. The specific steps to transform input features into embedded feature vectors are as follows: slide a window of size k=3 across a circular RNA sequence of length L with a sliding step size s=1 to segment the sequence into L-k+1 words of length k-mer, project them into a multidimensional space to generate a vocabulary, and then input the words into the embedding matrix to obtain the corresponding vector representation. The embedding matrix is obtained by training and fine-tuning a pre-trained word2vec model. The encoding methods for RNA binding site sequences include electron-ion interaction pseudopotential, nucleotide chemical properties, and k-spacer nucleotide pair frequency. Specifically, the electron-ion interaction pseudopotential encoding method encodes RNA sequences into digital vectors, the nucleotide chemical property encoding method converts RNA sequences into digital vectors, and the k-spacer nucleotide pair frequency encoding method converts RNA sequences into digital vectors. The word2vec encoding method is combined with feature encoding methods including electron-ion interaction pseudopotentials, nucleotide chemical properties, and k-spaced nucleotide pair frequencies, and the expanded features are aggregated. Step 2: First, use a hybrid architecture of convolutional neural network and bidirectional long short-term memory network to process the transformed embedded feature vector. Then, use the input feature signal to adaptively remember, forget and update the memory of information. Use a multi-scale downsampling structure to expand deeper features. Decode the features before entering the recurrent neural network to expand the deep features. The hybrid architecture of convolutional neural networks and bidirectional long short-term memory networks uses batch normalization layers, swish activation functions, and Avepool pooling layers to reduce the network burden, while random deactivation layers are used to prevent overfitting of features. The features are then fed into the bidirectional long short-term memory network, and the weights of key features are enhanced through an attention mechanism. Step 3: After the expanded deep and shallow features are fused, they are fed into the fully connected layer of the classification network for classification. Finally, the classification results are used to determine whether they are binding sites, thus enabling the prediction of the binding sites of RNA-binding proteins and circular RNA.
2. The method for predicting the binding site of RNA-binding proteins and circular RNA according to claim 1, characterized in that: In step two, before entering the bidirectional long short-term memory network, the embedded feature vector is first processed by a convolutional filter to obtain features of different scales and concatenate them to obtain a concatenated feature matrix.
3. The method for predicting the binding site of RNA-binding proteins and circular RNA according to claim 1, characterized in that: In step two, each long short-term memory unit in the bidirectional long short-term memory network uses a gating mechanism to adaptively memorize, forget, and update information based on input signals.
4. The method for predicting the binding site of RNA-binding proteins and circular RNA according to claim 1, characterized in that: After the features of the random deactivation layer are processed, a parallel convolutional neural network is used to expand the deep features to avoid the limitations of neural recurrent networks.