CN111816255A - Fusion of multi-view and optimal multi-label chain learning for RNA-binding protein identification - Google Patents

Abstract

The invention belongs to the field of bioinformatics, and relates to RNA-binding protein recognition integrating multi-view and optimal multi-label chain learning. The method includes two parts: training phase and use phase. The training phase includes initial multi-view data construction, multi-view deep feature extraction model training, multi-label feature learning and optimal multi-label chain classifier training. Multiple perspectives include RNA sequence perspective, amino acid sequence perspective, multi-gap dipeptide component perspective and RNA sequence semantics perspective. In order to improve the effectiveness of multi-view features, the present invention constructs deep multi-view features by using CNN to perform deep learning based on the original multi-view data. In order to link multi-view features with multi-label learning, the present invention establishes a multi-label feature learning model, which is used to integrate the advantages of all perspectives, and uses the optimal CC chain classifier which is different from ordinary CC multi-label classifiers to learn labels. The correlation between them can improve the classification accuracy more effectively.

Description

Fusion of multi-view and optimal multi-label chain learning for RNA-binding protein identification

technical field

The invention belongs to the field of bioinformatics, and relates to RNA-binding protein recognition integrating multi-view and optimal multi-label chain learning.

Background technique

RNA, the full name of ribonucleic acid, exists in biological cells and the genetic information carrier in some viruses and viroids. It mainly plays a role in regulating the expression of coding genes in living organisms, and also acts as a template for protein synthesis after gene transcription. It is an indispensable ingredient in life. If an RNA wants to perform its function smoothly, it generally needs to be mediated by RNA-binding protein (RBP), so the lack of a certain RBP may cause a certain type of RNA to fail to perform its regulatory or translational functions, thus making life lack some important Abnormal proliferation of protein or certain proteins, affecting its own function.

RNA-binding proteins (RBPs) are key players in post-transcriptional events, and the versatility and structural flexibility of their domains enable RBPs to control the metabolism of a large number of transcripts. RBPs are involved in almost all steps of the post-transcriptional regulatory layer, and they establish highly dynamic interactions with other proteins as well as coding and noncoding RNAs to generate functional units called ribonucleoprotein complexes that regulate RNA splicing, polyadenylation, Stability, localization, translation and degradation. Studies have found that some specific RBPs have the efficacy of regulating RNA synthesis of oncoproteins and tumor suppressor proteins, so deciphering the intricate inter-binding network between RBPs and their cancer-associated RNA targets will provide a better understanding of tumor biology and may lead to therapeutic discoveries A new approach to cancer.

In the context of the high development of big data and sequencing technology, it is difficult for the biomedical industry to detect the binding of each pair of RNA and RBP, so many algorithms that use machine learning models to identify RBP binding sites from RNA sequences have emerged. For example, Maticzka et al. proposed the GraphProt method, which learned the binding preferences of RBP sequences and structures from high-throughput experimental data, and designed a unique computational framework; Corrado et al. proposed RNACommender, a method for predicting binding sites that can Recommend RNA targets to unexplored RBPs by taking into account the protein structure and the simulated secondary structure of RNA through available interaction information; HOCNNLB, proposed by Zhang et al., uses higher-order nucleotide codes as initial features to predict a certain segment Whether a given RNA is a binding site. The focus of these methods is to use the sequence features or structural features of the original RNA sequence to determine whether a given RNA sequence fragment is a binding site for a specific RBP, but few methods use the existing binding information of RNA and RBP to predict provide help. In response to this, Pan et al. proposed the iDeepM method, which uses multi-label classification and deep learning to find multiple RBPs that can be combined with one RNA, and successfully achieves the expected effect of multi-label classification. However, iDeepM also has the following shortcomings: although the single-view RNA sequence data it uses is effective for prediction and classification, it is limited by the insufficient amount of RNA sequence information, resulting in low accuracy; in addition, this method uses convolutional neural network. Using long and short-term memory network to perform multi-label classification, it fails to fully learn the relationship between labels, which also affects the prediction accuracy.

SUMMARY OF THE INVENTION

The invention realizes the recognition of RNA binding proteins that integrates multi-view and optimal multi-label chain learning. The method includes two parts: a training stage and a use stage. The training stage includes an initial multi-view feature construction model, a deep multi-view feature extraction model, and a multi-view feature extraction model. Label feature learning and optimal multi-label chain classifier training.

Training phase: The initial multi-view feature construction model uses molecular biology principles, statistical principles and Word2Vec (word turn vector) technology to convert the original RNA sequence into amino acid sequence, multi-gap dipeptide components and RNA sequence semantic matrix, obtain the sequence of The order, composition and semantic features are then constructed together with the original RNA sequence to form the initial multi-view feature, and the initial multi-view feature construction model is obtained; the deep multi-view feature extraction model constructs four convolutional neural networks, and the initial four views are Features are trained to obtain deep multi-view features with better classification ability, and a deep multi-view feature extraction model is obtained; the extracted deep features are used to train a multi-label feature learning model, and the labels learned by the multi-label feature learning model are related. The linked weighted feature vectors are used to train an optimal CC multi-label chain classifier to learn the associations between labels and obtain a model with the ability to identify RNA-binding proteins.

Use stage: Obtain the RNA sequence to be tested, and use the principles of molecular biology, statistics and Word2Vec (word turn vector) technology to construct the initial multi-view feature of the sequence; then use the four trained convolutional neural networks to extract the The depth features of 4 perspectives; then use the trained multi-label feature learning model to perform weighting operations on the depth features of the spliced 4 perspectives to obtain the weighted feature vectors related to labels; then use the trained optimal CC multi-label chain The classifier predicts the weighted feature vector related to the label to obtain the final prediction result.

The described fusion multi-view and optimal multi-label chain learning RNA-binding protein identification set multi-view deep feature learning technology, multi-label feature learning technology and multi-label learning technology, deep learning deep structure optimization feature representation, multi-label The feature learning technology fuses and corrects the depth features of each view, makes full use of the advantages of each view feature, and establishes a weighted feature vector related to the label, and the multi-label technology effectively utilizes the independence of each label and between labels. correlation. The effective combination of multi-view deep learning technology, multi-label feature learning technology and multi-label learning technology can fully extract effective information in RNA sequences and improve the generalization ability of the classifier.

An RNA sequence is a piece of biological genetic material described by a text sequence. The deep convolution model cannot process text information, so it is necessary to preprocess the RNA text sequence and convert it into a numerical form acceptable to the program. One-hot (one-hot encoding) is a relatively popular encoding technology at present. Its principle is to construct a text sequence of length m composed of n elements into an n*m matrix, in which each element is converted into n Orthogonal basis vectors of dimension are padded to corresponding positions in length m. In terms of RNA sequence, one-hot (one-hot encoding) will construct an initial 4*m size blank matrix for an RNA sequence of length m, convert each base into a 4-dimensional orthogonal basis vector, fill in to the corresponding position of the sequence, as shown in Figure 7. The row header is a specific RNA sequence with an actual length of 2700. To compare the position of the base in the column, the base A in the sequence can be represented as a vector (1,0,0,0) ^T , the base C as a vector (0,1,0,0) ^T , the base G is represented by (0,0,1,0) ^T , base U is represented by (0,0,0,1) ^T , and so on.

Although the initial feature matrix constructed by the above method is helpful for feature extraction, the disadvantage is that the amount of information is less. The amino acid sequence is composed of 20 kinds of amino acids, and its information is much richer than that of the RNA sequence, so the one-hot (one-hot encoding) encoding matrix obtained by converting the amino acid sequence will provide better results for feature extraction. The translation of RNA sequence into amino acid sequence is unidirectional and unique, but because one amino acid can correspond to multiple base combinations, the resulting amino acid sequence cannot be restored to the original RNA sequence, which will cause information loss and misinterpretation of information. . For example, the base combination GCA can be translated to give a fixed amino acid A, but the amino acid A can be expressed as GCA, GCC, GCG, GCU. To deal with this problem, three modes of translation of RNA sequences to amino acid sequences are used, namely the first form of translation from scratch, the second form of translation to start translation by skipping the first base, and the skipping of the first and second bases Begin the third form of translation. Using this method, the above-mentioned RNA sequence with a length of m can be converted into three amino acid sequences with a length of 1/3 m, and the amino acid sequences of these three forms can restore the original RNA sequence information by complementing the sequence information. As described above, the base combination GCA can be uniquely determined by using the amino acids R, A, and H at the corresponding positions of the three morphological sequences. Therefore, the amino acid sequences of the three forms are spliced together to obtain a long amino acid chain with a length of m, which can completely inherit the sequence information of the original RNA sequence and has a richer expression. One-hot coding is performed on this long chain, the principle is the same as the RNA sequence, and an initial feature matrix of 20*m size can be obtained, as shown in Figure 8, which is the amino acid perspective data proposed by the present invention. The row title is a specific amino acid sequence, the actual length is 2700. All amino acids in the row sequence can be represented as a 20-dimensional standard orthonormal basis vector by comparing the positions of the amino acids in the column headings.

Both the RNA-view and amino-acid-view data mentioned above are biased towards extracting features for sequence order, and the composition of a sequence is equally important in addition to its sequence. Because 0-gap dipeptide is biased towards the composition of two-dimensional sequences, while 1-gap dipeptide has three-dimensional structural component information, so using 0-gap dipeptide and 1-gap dipeptide to extract RNA sequence composition information is the best, The present invention adopts their combined form to extract sequence components to form the perspective of multi-gap dipeptide components. Because dipeptides are sensitive to the arrangement of left and right amino acids, for the 21 amino acids in the present invention (20 natural amino acids and the temporary amino acid O added in the present invention), there are 21*21*2 multi-gap dipeptide species, due to OO and O The combination of *O is of little significance to our research, so it was discarded. Count the occurrences of these 880 kinds of multi-gap dipeptides to obtain feature vectors, which can effectively capture the component information of this amino acid sequence and RNA sequence and the information of amino acid space components. Since the 880-dimensional feature vector is one-dimensional, the effect of extracting deep features is not ideal, so we convert it into a two-dimensional histogram, which can more effectively use machine learning models to extract deep features, as shown in Figure 9 . The abscissa of the table in the upper part of the figure is the multi-gap dipeptide species, in which "AA" represents the 0-gap dipeptide with alanine on the left and right, 18 represents its number in the sample amino acid sequence; "A*D" represents The left side is alanine with any amino acid in the middle, and the right side is the 1-gap dipeptide of aspartic acid. The figure below lists only 12 multi-gap dipeptides, and the actual number is 880. The lower part of the chart is the converted histogram, and the upper limit of the number of each multi-gap dipeptide is set to 30, so we take a 30*880 matrix as the initial data of the multi-gap dipeptide for this amino acid sequence.

Natural Language Processing (NLP) is an important direction in the field of computer science and artificial intelligence. From the perspective of initial data, bioinformatics and NLP have the same form of initial research data. Therefore, NLP methods can be used to solve the encoding of text and initial feature construction in bioinformatics. The present invention uses 6-mer RNA as the lexicon for training the semantic model, and the 6-mer RNA is a structure composed of 6 consecutive bases, so the lexicon is composed of 46 kinds of 6-mer RNA. The present invention uses the popular Word2Vec technology to construct a semantic model, and its principle is shown in FIG. 10 . Based on the 92102 RNA sequences in the data set used in the present invention, the following operations were performed on them one by one: 1) Using a sliding window with a size of 6 bases, obtain the arrangement order of 6-mer RNA in the RNA sequence; 2) For each 6-mer RNA The polymer RNA is encoded, that is, its position in the 4096 forms (with 'AAAAAA' as 1 and 'UUUUUU' as the rule of 4096); 3) The adjacent 2 6-mer RNAs are used as feature X and label respectively Y, put into training in the semantic model; 4) Extract the word vector results of 4096 kinds of 6-mer RNAs from the trained semantic model; 5) Use the word vector to replace each 6-mer RNA in the RNA sequence to construct the RNA sequence Semantic Matrix. The RNA sequence semantic matrix composed of 6-mer RNA word vectors not only has a small dimension, but also contains RNA sequence order and context structure information with 6 bases as motifs, which can be better used for deep feature learning.

The specific steps in this part are as follows:

Step 1: Use the one-hot transformation matrix of the original RNA sequence as the RNA initial feature X ¹ .

The second step: using the principles of molecular biology and the one-hot method to convert the original RNA sequence into the initial feature X ² of the amino acid sequence.

Step 3: Use statistical principles to convert the amino acid sequence into an initial feature ^X3 of the multi-gap dipeptide component.

The fourth step: use Word2Vec technology to train the RNA sequence semantic model, obtain 6-mer RNA word vectors, and form an RNA sequence semantic matrix as the initial feature X ⁴ of RNA sequence semantics. From this, the preliminary multi-view data set D={X ¹ , X ² , X ³ , X ⁴ , y} is obtained

The deep multi-view feature extraction part of the present invention uses a convolutional neural network to automatically extract each view feature of the RNA sequence. For the original RNA sequence, RNA sequence features, amino acid sequence features, multi-gap dipeptide component features, and RNA sequence semantic features can be obtained after preprocessing. Four different convolutional neural networks are constructed for features from four different perspectives. To automatically extract the depth of features from different perspectives.

During training, the CNN network uses the result of the last output layer to calculate the error and backpropagates, so as to learn the network. Because the feature vector calculated by the penultimate layer goes through only one fully connected layer to the output layer, it can be considered that while optimizing the network structure according to the training of the network output layer, the expression of the output feature vector of the penultimate layer is also optimized, that is, the network A better feature representation is also learned during training, so the output of the penultimate layer of the network is selected as the feature learned by the network. The features obtained through the automatic learning of the convolutional neural network have smaller dimensions than the original features, and the obtained features are features with better division ability after nonlinear combination, which can make the subsequent classification model have better generalization ability. effect.

Figure 11, Figure 12, Figure 13, and Figure 14 are the CNN network architecture diagrams used for deep feature extraction from four perspectives. Use k@m*n to represent the feature map of each layer of the network, k to represent the number of feature maps of this layer, and m*n to represent the size of the feature map. The two-dimensional convolution kernel of the network is represented by k*m*n, k is the number of convolution kernels, and m*n is the size of the convolution kernel. The stride of the convolution kernel is 1 by default. The input of the network is each perspective feature, and the output is a vector whose length is equal to 68 (that is, the combination of this RNA sequence and 68 RBPs). The first 67 dimensions of the result indicate that if the sample can be combined with the RBP of this dimension, it is equal to 1, otherwise it is equal to 0; the 68th dimension of the result indicates that if the sample RNA sequence cannot be combined with any of the first 67 RBPs, 1 otherwise, 0.

Figure 11 shows the CNN network architecture used for deep feature extraction from RNA perspective, including 1 2D convolutional layer, 1 pooling layer, 1 flattening layer, 2 dropout layers and 2 fully connected layers. The input of the network is a two-dimensional matrix of 4*2710. CNN network architecture The first convolutional layer is 101 convolution kernels of 4*10, and 101 feature maps of 1*2701 are obtained; the pooling length of the second pooling layer is 3, and 101 1*900 are obtained. The third layer is a flat layer, and a feature map of 1*90900 is obtained; the fourth layer is a dropout layer with a probability of 0.5, and a feature map of 1*90900 is obtained; the fifth is a fully connected layer. A 1*90900 feature map is converted into a 1*202 vector; the sixth layer is a dropout layer with a probability of 0.5, and a 1*202 feature map is obtained; the fifth is a fully connected layer, a 1*202 The feature map is converted into a 1*68 vector.

Figure 12 shows the CNN network architecture used for deep feature extraction from the amino acid perspective, which includes a total of 1 2D convolutional layer, 1 pooling layer, 1 flattening layer, 2 dropout layers and 2 fully connected layers. The input is a two-dimensional matrix of 20*2710. CNN network architecture The first convolutional layer is 101 convolution kernels of 20*10, and 101 feature maps of 1*2701 are obtained; the pooling length of the second pooling layer is 3, and 101 1*900 are obtained. The third layer is a flat layer, and a feature map of 1*90900 is obtained; the fourth layer is a dropout layer with a probability of 0.5, and a feature map of 1*90900 is obtained; the fifth is a fully connected layer. A 1*90900 feature map is converted into a 1*202 vector; the sixth layer is a dropout layer with a probability of 0.5, and a 1*202 feature map is obtained; the fifth is a fully connected layer, a 1*202 The feature map is converted into a 1*68 vector.

Figure 13 shows the CNN network architecture used for multi-gap dipeptide view depth feature extraction, which includes a total of 1 2D convolutional layer, 1 flat layer, 2 dropout layers and 2 fully connected layers. The input to the network is a 30*440 two-dimensional matrix. CNN network architecture The first convolutional layer is 101 30*10 convolution kernels, and 101 1*871 feature maps are obtained; the second layer is a flat layer, and 1 1*87971 feature map is obtained; the third The layer is a dropout layer with a probability of 0.5, and a feature map of 1*87971 is obtained; the fourth is a fully connected layer, which converts a feature map of 1*87971 into a vector of 1*202; the fifth layer is a feature map with a probability of 0.5 The dropout layer obtains a 1*202 feature map; the sixth is the fully connected layer, which converts a 1*202 feature map into a 1*68 vector.

Figure 14 shows the CNN network architecture used for deep feature extraction from the RNA sequence semantic perspective, which includes a total of 1 2D convolutional layer, 1 pooling layer, 1 flattening layer, 2 dropout layers and 2 fully connected layers. The input is a 25*2710 2D matrix. CNN network architecture The first convolutional layer is 101 convolution kernels of 25*10, and 101 feature maps of 1*2701 are obtained; the pooling length of the second pooling layer is 3, and 101 1*900 are obtained. The third layer is a flat layer, and a feature map of 1*90900 is obtained; the fourth layer is a dropout layer with a probability of 0.5, and a feature map of 1*90900 is obtained; the fifth is a fully connected layer. A 1*90900 feature map is converted into a 1*202 vector; the sixth layer is a dropout layer with a probability of 0.5, and a 1*202 feature map is obtained; the fifth is a fully connected layer, a 1*202 The feature map is converted into a 1*68 vector.

The last layer of the four networks uses the sigmoid function as the activation function to introduce nonlinear transformation. The expression of the sigmoid function is as follows:

The rest of the layers use the relu function as the activation function. The expression of the relu function is as follows:

R(x)=max(0,x)

The loss function of the network adopts the binary cross entropy (binary_crossentropy) loss function, which is defined as follows.

where p(x _i ) and q( _xi ) both represent the membership of the sequence x to category i, p represents the true label value, that is, 1 or 0, and q represents the predicted value. Here, because it is activated by the Sigmoid function, q∈ (0,1).

The specific steps in this part are as follows:

The first step: use X ¹ , y to train the RNA sequence deep feature extraction network, and take the penultimate layer of the CNN network architecture used for RNA perspective deep feature extraction as the RNA sequence deep feature

Step 2: Use X ² , y to train the amino acid sequence depth feature extraction network, and take the penultimate layer of the CNN network architecture used for amino acid perspective depth feature extraction as the amino acid sequence depth feature

Step 3: Use X ³ , y to train the multi-gap dipeptide component depth feature extraction network, and take the penultimate layer of the CNN network architecture used for multi-gap dipeptide perspective depth feature extraction as the multi-gap dipeptide component depth feature

得到多视角数据集

Step 4: Use X ⁴ , y to train the RNA sequence semantic depth feature extraction network, and take the penultimate layer of the CNN network architecture used for RNA sequence semantic depth feature extraction as the RNA sequence semantic depth feature

Get a multi-view dataset

The present invention uses an optimal multi-label chain learning algorithm based on multiple perspectives, which includes multi-label feature learning and optimal multi-label chain learning. The CC algorithm is a multi-label classification algorithm that can efficiently learn the association between labels. The principle is to build several binary classifiers to predict the corresponding labels. After training a binary classifier, the algorithm will predict the classifier After the corresponding label result is appended to the initial feature, it is used as the input feature for the next binary classifier training until all classifiers are trained. Different from the existing methods, the present invention improves the single-view CC algorithm, applies it to multi-view scenarios, and adds the advantages of multi-view data to the CC algorithm, so that it can better learn the association between tags. The specific principle As shown in Figure 15. The algorithm is divided into two parts: multi-label feature learning and multi-label learning. First, we obtain the deep feature vectors of each perspective from the upstream CNN model, stitch them together, and put them into the training of the multi-label feature learning model. The input size of this model is an 808-dimensional vector, and the output is a 68-dimensional result, corresponding to 68 labels. Through the learning of this model, we can obtain 68 groups of 808-dimensional weight coefficients, which correspond to the contribution weight of each dimension of the input vector to predicting each label. Multiply the 808-dimensional feature vector with these 68 sets of weight coefficients in turn to obtain 68 sets of weighted feature vectors, which are used to train the downstream CC multi-label classifier. The CC multi-label classifier in this experiment consists of 68 binary classifiers, which predict the membership of an RNA to 68 labels. First, we obtain the weighted feature vector x ₁ from the multi-label feature learning module and use it as the input feature to start training the first binary classifier. The first label value predicted by it is appended to the end of the weighted feature vector x ₂ to train the second binary classifier. This process is repeated until the last binary classifier is trained. Different from the traditional CC multi-label classifier, the optimal CC multi-label classifier proposed by the present invention is characterized in that after training the i-th binary classifier, all the currently predicted label values are appended to the next label. At the end of the associated weighted feature vector xi+1, the training of the i+1th binary classifier is performed. This not only retains the CC algorithm's ability to learn label associations, but also reflects the advantages of multi-view data in the process of training sub-classifiers, combining the advantages of multi-view and multi-label algorithms. Training the optimal CC multi-label classifier and prediction algorithm is shown in Algorithms 1, 2.

The specific steps in this part are as follows:

形成

使用

和y训练多标签特征学习模型，获得68个标签相关的加权特征向量

Step 1: Splicing

form

use

Train a multi-label feature learning model with y to obtain 68 label-related weighted feature vectors

y¹训练最优CC链式多标签分类器的第一个二分类器；Step 2: Use

y ¹ trains the first binary classifier of the optimal CC chained multi-label classifier;

后，使用附加标签后的

和y²训练最优CC链式多标签分类器的第二个二分类器；Step 2: Append the labels predicted by the above steps to

after using the additional tag

and y ² to train the second binary classifier of the optimal CC chained multi-label classifier;

后，使用附加标签后的

和y³训练最优CC链式多标签分类器的第三个二分类器，以此类推，直至第68个二分类器训练完毕。Step 3: Append the labels predicted by the above steps to

after using the additional tag

and y ³ to train the third binary classifier of the optimal CC chained multi-label classifier, and so on, until the 68th binary classifier is trained.

In the use stage of this method, the specific steps are as follows:

Step 1: Build a preliminary multi-view test dataset using the initial multi-view feature building model on the test data

Step 2: Use a deep multi-view feature extraction model to obtain a deep multi-view test dataset

形成

输入至训练好的多标签特征模型，得到加权特征向量

Step 3: Splicing

form

Input to the trained multi-label feature model to get the weighted feature vector

输入至训练好的最优CC多标签链式分类器中，获取预测到的的所有标签值

Step 4: put

Input into the trained optimal CC multi-label chain classifier to obtain all predicted label values

The advantages of the present invention include the following points:

1) Construction of initial multi-view RNA sequence features: There are many methods for constructing features in RNA sequences, and features constructed in different ways have certain effects and have their own advantages and disadvantages. The use of multi-view features for feature extraction of RNA sequences and identification of RNA-binding proteins that can bind to them can well combine the advantages of different methods for constructing features. The present invention cites amino acid sequence representation that can well express RNA sequence order and context features, multi-gap dipeptide data that can well represent RNA sequence component information and structural information, and RNA sequence semantics that can well represent RNA sequence semantic information Data to construct multi-perspective initial features, which can complement perspective information from multiple different aspects.

2) Construction of deep multi-view features: In order to improve the effectiveness of multi-view features, based on the original multi-view data, CNN is used for deep learning to construct deep multi-view features. Compared with the original multi-view features, the multi-view features extracted by deep features have smaller data dimensions and higher classification effects;

3) Construction of multi-label feature learning model: Use multi-label feature learning technology to integrate the learned in-depth features from multiple perspectives, and use the principle of logistic regression to make feature corrections for different labels, resulting in better training capabilities. Multi-label features for label classifiers;

4) Construction of the optimal chained multi-label classifier: improve the CC multi-label classifier, use the modified weighted feature vector obtained from the above multi-label special learning model to perform multi-label learning, and obtain more generalization capabilities. Tag classifiers for RNA-binding protein identification.

Description of drawings

FIG. 1 is a frame diagram of an algorithm method of the present invention.

FIG. 2 is a frame diagram of an algorithm for obtaining initial feature data from different perspectives of the present invention.

FIG. 3 is a frame diagram of a multi-view deep feature learning algorithm of the present invention.

FIG. 4 is a frame diagram of the multi-label feature learning algorithm of the present invention.

FIG. 5 is a frame diagram of the multi-view learning algorithm of the present invention.

Figure 6 is a framework diagram of the RNA-binding protein identification algorithm of the present invention.

Figure 7 is RNA-seq one-hot matrix data.

Fig. 8 is the one-hot matrix data of amino acid sequence obtained by transforming the RNA sequence of Fig. 7.

FIG. 9 is the columnar data of the multi-gap dipeptide composition obtained by converting the amino acid sequence of FIG. 8 .

Fig. 10 is the semantic matrix data obtained by transforming the RNA sequence of Fig. 7 after being trained by the semantic model.

Figure 11 is a deep feature extraction network for RNA sequences.

Figure 12 is an amino acid sequence deep feature extraction network.

Figure 13 is a deep feature extraction network for multi-gap dipeptide components.

Figure 14 is an RNA sequence semantic deep feature extraction network.

Figure 15 is a flow chart of the optimal multi-label chain learning algorithm integrating multi-label feature learning and multi-label learning.

FIG. 16 is a line graph comparing the performance of the algorithm used in the present invention and the existing algorithm on a single class.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

As shown in Figures 1 to 6, the present invention realizes the recognition of RNA-binding proteins that integrates multi-view and optimal multi-label chain learning. The method includes initial multi-view feature construction, deep multi-view feature extraction, and multi-label feature model training. and the optimal multi-label chain classifier training in four parts. The initial multi-view feature construction part obtains the initial multi-view features of the original RNA sequence; the deep multi-view feature extraction part performs deep feature learning on the initial multi-view features to obtain multi-view depth features; the multi-label feature model training part uses multi-view depth features , construct the weighted feature vector related to the label; the training part of the optimal multi-label chain classifier uses the weighted feature vector to learn the CC classifier associated with the label, and obtain the final prediction result.

Specific steps in the training phase. The initial multi-view feature construction part of the method first extracts four features from the original RNA sequence: RNA sequence, amino acid sequence, multi-gap dipeptide component and RNA sequence semantic matrix, and constructs multi-view data with a total of 4 views.

The original RNA sequence is a text sequence, and its numerical matrix expression can be obtained by transforming it using one-hot encoding technology. This algorithm utilizes RNA-seq data as features of RNA perspective. Figure 7 plots the RNA sequence features after one-hot encoding, where the horizontal axis represents a specific RNA sequence, and the vertical axis represents one-hot encoding rules.

Example 1

In accordance with the implementation of the training phase, the examples were completed for the RNA-RBP binding data of the AURA2 dataset. The dataset contains 67 RBPs and 73,681 RNA sequences and their 550,386 binding site information, as shown in Table 1. The amount of sample RNA that can be bound by each RBP varies widely. The length of each RNA sequence is different, so we uniformly stipulate a length of 2700, and the lack is filled with base B. Table 2 shows the comparison results between the method RRMVL used in the present invention and the current advanced methods in this field.

The performance index of this algorithm in table 2 embodiment 1

It includes the decision tree classifier model without deep learning, the current advanced iDeepM method in this field, and the prediction performance indicators of each sub-perspective model and the overall model under the RRMVL model. As can be seen from the above table, the iDeepM model using deep learning and any single-view model under RRMVL are better than the decision tree model, which proves that deep learning has obvious advantages in extracting longer sample features. At the same time, the AUC value and F1 value of the RRMVL method under the integration of all perspective models are higher than those of any single-view model, which reflects the information complementarity between multi-view data, and also shows that the multi-view data is important in biology. In the field of informatics, better results can be achieved. From a single perspective, the multi-gap dipeptide component perspective achieves the best results, because the multi-gap dipeptide contains not only sequence order information, but also sequence component and structural information, which is the most informative of all perspectives. . The effect of RNA sequence semantic single perspective is slightly lower than that of the initial RNA sequence perspective. This is because training a good semantic model usually requires millions of sample data, and the dataset of the present invention only contains 92,102 RNA sequences, which is not enough for training. The 6-mer RNA word vector with ideal effect is obtained, so the experimental effect is not good. Overall, among the three comparison algorithms, the RRMVL proposed in the present invention achieved the best results in 3 AUCs and 3 F1s, which proves that the optimal multi-label chain learning method based on multiple perspectives is effective in identifying RNA-binding proteins. The expected effect has been achieved.

Example 2

In order to test the multi-label feature learning and the optimal multi-label chain learning effect used in the present invention, the present invention conducts two comparison experiments on RRMVL and its variant methods on the AURA data set, respectively using the integration based on multi-view voting. Learning RRMVL versus learning RRMVL using multi-label features, and comparing RRMVL without multi-label learning with RRMVL using optimal multi-label chain learning. Since the ensemble learning model based on multi-view voting is not a classifier, there is no AUC indicator, and the five-fold cross-validation results of the remaining methods are shown in the following table.

Table 3 The performance test of the algorithm in Embodiment 2 about the multi-label feature learning model and the optimal multi-label chain learning model

As can be seen from the above table, for multi-view data, after using multi-label feature learning on it, the prediction performance of the model is always better than voting-based ensemble learning, indicating that multi-label feature learning takes full advantage of multi-view data. . On the other hand, in dealing with multi-label classification problems, methods using multi-label classifiers consistently outperform methods without multi-label techniques, proving that the association between labels has a non-negligible effect on prediction. It is worth noting that the AUC metric of RRMVL decreased after multi-label learning, which is due to a slight gap between the classification performance of the multi-label CC classifier and the classification ability of the “Sigmoid” network in the last layer of the neural network. For the three F1 indicators, the multi-label learning method RRMVL based on multiple perspectives achieved the best results, again proving that the method proposed in the present invention can more accurately identify which RBPs an unexplored RNA can bind to.

Example 3

In order to study the influence of the number of class samples on the experimental effect, the present invention uses RRMVL to conduct a separate experiment on 68 class data sets, and the experimental results of the iDeepM method are shown in the following table.

Table 3. Different RBP prediction effects

The line chart of prediction accuracy is shown in Figure 16. As can be seen from Figure 16, among the two comparison algorithms, RRMVL has achieved the best results in the prediction accuracy of most classes. With the gradual increase of the number of class samples, the indicators of the two methods gradually improve and tend to be flat. trend. It is noted that when the number of samples is less than 5000, the fluctuations of various indicators are large, because the number of samples of some classes is too small, so that the model cannot learn the deep features of these samples well. And from the comparison of the two curves, the learning ability of the iDeepM method in the low-sample environment is not as good as that of the RRMVL, and the volatility is more severe, which indirectly reflects the advantages of multi-view data in small-sample learning. In general, the method proposed in the present invention achieves the expected effect on various data sets.

Claims (9)

1. an RNA-binding protein identification of fusion multi-view and optimal multi-label chain learning, is characterized in that, the step of training stage is: The first step: use the one-hot encoding technology to encode the original RNA sequence into a numerical matrix as the initial RNA sequence feature X ¹ ; The second step: using the principle of molecular biology to convert the original RNA sequence into an amino acid sequence, and then convert it into a numerical matrix using one-hot encoding technology, as the initial amino acid sequence feature X ² ; The third step: using statistical principles to convert the amino acid sequence into a multi-gap dipeptide columnar numerical matrix, as the initial dipeptide component feature X ³ ; Step 4: Use the Word2Vec technology to build a model, learn the word vector with 6-mer RNA as the vocabulary, convert the RNA sequence into an RNA sequence semantic matrix, and use it as the initial RNA sequence semantic feature X ⁴ ; Obtain the initial multi-view data set D= {X ¹ , X ² , X ³ , X ⁴ , y}; Step 5: Use X ¹ , y to train the RNA sequence depth feature extraction network, and take the penultimate layer of the CNN network architecture used for RNA sequence depth feature extraction as the RNA sequence depth feature

Step 6: Use X ² , y to train the amino acid sequence depth feature extraction network, and take the penultimate layer of the CNN network architecture used for amino acid perspective depth feature extraction as the amino acid sequence depth feature

Step 7: Use X ³ , y to train the multi-gap dipeptide component depth feature extraction network, and take the penultimate layer of the CNN network architecture used for multi-gap dipeptide perspective depth feature extraction as the multi-gap dipeptide component depth feature

Step 8: Use X ⁴ , y to train the RNA sequence semantic depth feature extraction network, and take the penultimate layer of the CNN network architecture used for RNA sequence semantic depth feature extraction as the RNA sequence semantic depth feature

Step 9: Splicing

form

use

Train a multi-label feature learning model with y to obtain the weighted feature vector associated with each label

Step 10: Use

y trains the optimal chained CC multi-label classifier model; The described optimal chained CC multi-label classifier model uses the weighted feature vector corresponding to each label to train each sub-classifier in the CC multi-label classifier, thereby obtaining a classifier model with better classification effect; Step 11: Use the initial multi-view feature building model on the test data to build a preliminary multi-view test dataset

Step 12: Use a deep multi-view feature extraction model to obtain a deep multi-view test dataset

Step Thirteen: Splicing

form

Input to the trained multi-label feature model to get the weighted feature vector

Step Fourteen: Put the

Input into the trained optimal chained CC multi-label classifier to obtain all predicted label values

2. the RNA-binding protein identification of fusion multi-view and optimal multi-label chain learning as claimed in claim 1, it is characterized in that, the CNN network architecture that the RNA-view depth feature extraction in the described 5th step uses, comprises 1 2-dimensional convolutional layers, 1 pooling layer, 1 flattening layer, 2 dropout layers and 2 fully connected layers; the first convolutional layer of the CNN network architecture is 101 4*10 convolution kernels, obtaining 101 feature maps of 1*2701; the pooling length of the second pooling layer is 3, and 101 feature maps of 1*900 are obtained; the third layer is a flat layer, and one feature map of 1*90900 is obtained; The fourth layer is a dropout layer with a probability of 0.5, and a 1*90900 feature map is obtained; the fifth is a fully connected layer, which converts a 1*90900 feature map into a 1*202 vector; the sixth layer is a probability The dropout layer of 0.5 obtains a 1*202 feature map; the fifth is the fully connected layer, which converts a 1*202 feature map into a 1*68 vector. 3. The RNA-binding protein identification of fusion multi-perspective and optimal multi-label chain learning as claimed in claim 1, wherein the CNN network architecture used in the amino acid perspective depth feature extraction in the sixth step comprises 1 2-dimensional convolutional layers, 1 pooling layer, 1 flattening layer, 2 dropout layers and 2 fully connected layers; the first convolutional layer of the CNN network architecture is 101 20*10 convolution kernels, obtaining 101 feature maps of 1*2701; the pooling length of the second pooling layer is 3, and 101 feature maps of 1*900 are obtained; the third layer is a flat layer, and one feature map of 1*90900 is obtained; The fourth layer is a dropout layer with a probability of 0.5, and a 1*90900 feature map is obtained; the fifth is a fully connected layer, which converts a 1*90900 feature map into a 1*202 vector; the sixth layer is a probability The dropout layer of 0.5 obtains a 1*202 feature map; the fifth is the fully connected layer, which converts a 1*202 feature map into a 1*68 vector. 4. The RNA-binding protein identification of fusion multi-view and optimal multi-label chain learning as claimed in claim 1, wherein the CNN network architecture used in the multi-gap dipeptide angle depth feature extraction in the seventh step , including 1 two-dimensional convolutional layer, 1 flat layer, 2 dropout layers and 2 fully connected layers; the first convolutional layer of the CNN network architecture is 101 30*10 convolution kernels, resulting in 101 The feature map of 1*871; the second layer is a flat layer, and a feature map of 1*87971 is obtained; the third layer is a dropout layer with a probability of 0.5, and a feature map of 1*87971 is obtained; the fourth is a fully connected layer. , convert a 1*87971 feature map into a 1*202 vector; the fifth layer is a dropout layer with a probability of 0.5, and a 1*202 feature map is obtained; the sixth is a fully connected layer, which converts a 1 The feature map of *202 is converted into a 1*68 vector. 5. The RNA-binding protein identification of fusion multi-view and optimal multi-label chain learning as claimed in claim 1, characterized in that, the CNN network architecture used in the RNA sequence semantic depth feature extraction in the eighth step, comprising 1 two-dimensional convolutional layer, 1 pooling layer, 1 flat layer, 2 dropout layers and 2 fully connected layers; the first convolutional layer of the CNN network architecture is 101 25*10 convolution kernels, The obtained 101 feature maps of 1*2701; the pooling length of the second pooling layer is 3, and 101 feature maps of 1*900 are obtained; the third layer is a flat layer, and 1 feature map of 1*90900 is obtained ; The fourth layer is a dropout layer with a probability of 0.5, and a 1*90900 feature map is obtained; the fifth is a fully connected layer, which converts a 1*90900 feature map into a 1*202 vector; The sixth layer is The dropout layer with a probability of 0.5 obtains a 1*202 feature map; the fifth is a fully connected layer, which converts a 1*202 feature map into a 1*68 vector. 6. The RNA-binding protein identification of fusion multi-view and optimal multi-label chain learning as claimed in claim 2, characterized in that, the CNN network architecture used in the deep feature extraction from the RNA perspective, and the deep feature extraction from the amino acid perspective are used. The last layer of the CNN network architecture, the multi-gap dipeptide perspective depth feature extraction, and the CNN network architecture used for RNA sequence semantic perspective depth feature extraction all use the sigmoid function as the activation function to introduce nonlinear transformations, and the remaining layers The relu function is used as the activation function, and the loss function of the four networks adopts the Binary cross-entropy binary cross entropy loss function. 7. The RNA-binding protein identification of fusion multi-view and optimal multi-label chain learning as claimed in claim 3, it is characterized in that, described RNA view depth feature extraction uses CNN network architecture, amino acid view depth feature extraction and uses The last layer of the CNN network architecture, the multi-gap dipeptide perspective depth feature extraction, and the CNN network architecture used for RNA sequence semantic perspective depth feature extraction all use the sigmoid function as the activation function to introduce nonlinear transformations, and the remaining layers The relu function is used as the activation function, and the loss function of the four networks adopts the Binary cross-entropy binary cross entropy loss function. 8. The RNA-binding protein identification of fusion multi-perspective and optimal multi-label chain learning as claimed in claim 4, characterized in that, the CNN network architecture used for deep feature extraction from described RNA perspective, and the deep feature extraction from amino acid perspective are used. The last layer of the CNN network architecture, the multi-gap dipeptide perspective depth feature extraction, and the CNN network architecture used for RNA sequence semantic perspective depth feature extraction all use the sigmoid function as the activation function to introduce nonlinear transformations, and the remaining layers The relu function is used as the activation function, and the loss function of the four networks adopts the Binary cross-entropy binary cross entropy loss function. 9. The RNA-binding protein identification of fusion multi-perspective and optimal multi-label chain learning as claimed in claim 5, wherein the CNN network architecture used for deep feature extraction from described RNA perspective, and the deep feature extraction from amino acid perspective are used. The last layer of the CNN network architecture, the multi-gap dipeptide perspective depth feature extraction, and the CNN network architecture used for RNA sequence semantic perspective depth feature extraction all use the sigmoid function as the activation function to introduce nonlinear transformations, and the remaining layers The relu function is used as the activation function, and the loss function of the four networks adopts the Binary cross-entropy binary cross entropy loss function.

Images