Multi-view enzyme function prediction method considering molecular structure and sequence mining

Abstract

The present application belongs to the field of intelligent cell biological recognition, and particularly relates to a multi-view enzyme function prediction method considering molecular structure and sequence mining. The method comprises four stages of initial enzyme feature construction, deep enzyme sequence feature construction, deep enzyme structure feature construction, and model training and prediction based on a TSK fuzzy system. The method regards the structure features and sequence features of enzymes as two different perspectives, extracts, crosses and discriminates information from different modes through a newly constructed multi-view deep network, mines the complementary and consistent information between multi-view enzyme data, trains multi-view features using a multi-view TSK fuzzy system model, and thus realizes the final enzyme function prediction. The method considers the sequence features and structure features of enzymes, makes the prediction information more complete, and relearns the sequence and structure features of enzymes in the new network through the TSK fuzzy system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent cell biological recognition, and particularly relates to a multi-view enzyme function prediction method considering molecular structure and sequence mining. BACKGROUND

[0002] Enzyme is a protein or RNA produced by living cells, which has high specificity and high catalytic efficiency for its substrate. Enzyme can catalyze the substrate to produce biochemical reactions, thereby producing high-efficiency catalytic protein or RNA. The functional classification of enzyme plays an extremely important role in the application of enzyme in people's production and life and disease diagnosis. The enzyme commission (EC) number system usually divides enzymes into six categories according to their catalytic properties: oxidoreductases (EC1), transferases (EC2), hydrolases (EC3), lyases (EC4), isomerases (EC5) and ligases (EC6). The continuous increase of enzyme sequences in the database poses a serious challenge to the functional classification of enzyme. Since the number of enzymes with unknown functions is huge, and it is very time-consuming and expensive to determine the characteristics of enzymes by using biological experiments, it is urgent to develop some efficient and low-cost enzyme function prediction technology.

[0003] In recent years, more and more attention has been paid to the use of various computational models to predict the types of enzymes. At present, some researches have achieved encouraging results, which provide an important means for enzyme function annotation and enzyme-related drug design. With the continuous development of bioinformatics and deep learning, some effective feature extraction methods and machine learning methods based on sequence information of enzymes have been proposed to predict the types of enzymes, and these methods mostly use support vector machine (SVM), random forest (RF) and KNN, etc. However, the existing methods still have the following shortcomings:

[0004] (1) Most methods only use sequence information of enzymes, and often use one-hot encoding, position score matrix and other feature extraction methods for feature extraction, and then use traditional machine learning methods such as support vector machine, random forest and KNN for classification. These traditional feature extraction and classification methods have limited learning ability in the task of enzyme function classification, and new feature extraction and classification technologies need to be further developed;

[0005] (2) At present, most network models using three-dimensional structure information of enzymes for enzyme classification use processing vectors and matrices, ignoring the complex representation of enzyme graph structure, which leads to insufficient learning of enzyme structure information;

[0006] (3) Although a small number of methods have taken into account the sequence and structure information of enzymes to realize the function prediction of enzymes, the commonness and characteristics of the data of enzymes from multiple perspectives such as sequence and structure are not fully mined, and the ability of efficiently processing multi-modal feature fusion is lacked. Therefore, it is still a challenging work with important value to study an enzyme function prediction method which can fully utilize multi-modal information such as sequence features and structure features of enzymes. SUMMARY

[0007] Most of the enzyme function classification methods are based on single sequence features or single structure features. In the present application, a new enzyme function prediction method is adopted, which takes into account the structure features and sequence features of enzymes, performs deep sequence and structure feature learning, and adopts a multi-perspective TSK fuzzy system classifier. The method regards the structure features and sequence features of enzymes as two different perspectives, extracts, crosses and discriminates information from different modalities through a newly constructed multi-perspective deep network, and can mine the complementary and consistent information between multi-perspective enzyme data.

[0008] The technical scheme of the present application is as follows:

[0009] The multi-perspective enzyme function prediction method taking into account the molecular structure and sequence mining comprises four stages of initial enzyme feature construction, deep enzyme sequence feature construction, deep enzyme structure feature construction, and model training and prediction based on TSK fuzzy system as shown below:

[0010] The first stage is the initial feature construction stage. This stage includes two steps, namely enzyme sequence initial feature construction and enzyme structure initial feature construction, and the specific steps are as follows:

[0011] Step 1: Initial sequence feature extraction

[0012] In this work, the BioVec method is used to extract the initial features of the amino acid sequence of the enzyme. BioVec is a method for biological sequence representation and feature extraction. BioVec regards the sequence as a long sentence, and uses natural language processing to process it. Each biological sequence is embedded in an n-dimensional vector, which uses a neural network to represent the biological and biochemical properties of the sequence. The Skip-gram neural network model in Word2Vec is used to train the embedding of the corpus data composed of 3-gram sequences, so that the probability of observing each word sequence is maximized. This model only needs to be trained once. Finally, BioVec is used to represent each enzyme amino acid sequence as a 3*100 vector, which is used for deep sequence feature extraction.

[0013] Step 2: Initial structure feature extraction

[0014] The three-dimensional coordinates (x, y, z) of the carbon atoms are extracted from the amino acids in each chain in order from the PDB file of each enzyme, while each amino acid residue is represented by a one-hot encoding of 21 dimensions (20 standard amino acids and others), so that each amino acid residue can finally be represented by a vector of 24. Finally, a (n*24) -dimensional matrix is obtained to represent the enzyme features of length n. By default, the amino acid sequence length n in this paper is 1000, and if the sequence length is less than 1000, it is supplemented with 0. The above (n*24) -dimensional matrix will be used as the initial structure feature for subsequent point cloud network learning to extract deep structure features.

[0015] The second stage: deep enzyme sequence feature construction. This stage contains 4 steps, which are: SMOTE data oversampling processing, BBA residual module processing, Bio-CS attention module processing and full connection module processing.

[0016] The specific steps are as follows:

[0017] The third step: SMOTE data oversampling processing.

[0018] In the process of extracting sequence features, due to the uneven distribution of the categories of the 6 enzymes in the PDB data set adopted, the SMOTE oversampling method is used to process the unbalanced data after BioVec representation. SMOTE is an oversampling technique for synthesizing minority class samples, which is an improved scheme of random oversampling algorithm. Since random oversampling takes the strategy of simply duplicating samples to increase minority class samples, it is easy to produce the problem of model overfitting, which leads to the information learned by the model being too specific and not general enough. SMOTE sampling is to interpolate between adjacent minority class samples. Therefore, it can increase the number of minority class samples by constructing new minority class samples in the neighborhood of existing samples, thereby helping the classifier to improve its generalization ability

[23] . The flowchart of the SMOTE algorithm is as follows:

[0019] (1) For each sample x in the minority class, calculate the distance from all samples in the minority class sample set according to the Euclidean distance to get its k-nearest neighbors.

[0020] (2) Set a sampling ratio according to the sample imbalance ratio to determine the sampling rate N. For each minority class sample x, randomly select several samples from its k-nearest neighbors, assuming that the selected neighbor is x0.

[0021] (3) For each randomly selected neighbor x0, construct a new sample x_new according to formula 1 with the original sample x.

[0022] x_new = x + rand(0, 1) x (x0-x)

[0023] The SMOTE processed data will be input into the BBA residual module for subsequent feature extraction.

[0024] Step 4: Deep sequence features are extracted using the BBCNet neural network.

[0025] 4.1 BBA residual module processing

[0026] The BBA residual module mainly consists of two bidirectional long short-term memory models Bi-LSTM and an additional attention layer implemented by a Keras self-attention package. The BBA residual module can be regarded as an improved version of the traditional BBA module after introducing the residual mechanism. For the Bi-LSTM in the BBA residual module, the design concept is to make the feature data obtained at any time have information between the past and the future. The model is divided into two independent LSTMs, and the input sequence is input to the two LSTM neural networks in forward and reverse order for feature extraction. The word vector formed by concatenating the two output vectors (i.e., the extracted feature vector) is used as the final feature representation of the word. Generally, a sequential structure is used, but as the number of network layers increases, the training set loss gradually decreases and tends to be saturated. When the network depth is increased again, the training set loss will increase instead. Therefore, in order to avoid the phenomenon of network degradation, and to better utilize local features, first, the two layers of Bi-LSTM are concatenated to fuse the past and future information. At the same time, the fused features and the features obtained after self-attention are fused again. This kind of skip connection avoids the loss of local features, and the two times of fusion finally form the proposed BBA residual module. The module contains two layers of Bi-LSTM, and the one-way output of each Bi-LSTM contains 128 nodes using the hyperbolic tangent activation function; the third layer is a sequence self-attention layer, also using the hyperbolic tangent activation function.

[0027] 4.2 Bio-CS attention module processing.

[0028] In order to pay attention to the relationship between different dimensional features, the Bio-CS attention module is proposed, hoping that the model can automatically learn the importance of different dimensional channel features.

[0029] The structure of the Bio-CS attention module is shown in FIG. 2(b). The Bio-CS module first performs a global average pooling operation on the convolutional feature map to obtain a global feature at the channel level; then performs an activation operation on the global feature to learn the relationship between the channels and obtain the weights of different channels; and finally multiplies the weights by the original feature map to obtain the final feature. Essentially, the Bio-CS module is a form of attention mechanism operation in the channel dimension. This attention mechanism allows the model to pay more attention to the channel features with the largest amount of information and suppress the unimportant channel features. In this work, both the channel Bio-CS module and the spatial Bio-CS module are used to obtain new features from the channel and spatial directions, and the two features are combined to obtain channel and spatially fused feature information.

[0030] (1) Channel Bio-CS Block

[0031] The sequence feature U e R D×C is input into the BBA residual module, where D is the spatial dimension and C is the channel dimension. The global average pooling is used to compress and convert the calculation F s to obtain a statistical quantity z, thereby compressing the global spatial information into a channel descriptor. The formula for calculating the cth element of z is:

[0032]

[0033] In the channel module, z avg is obtained by global average pooling, which can be expressed as:

[0034] C avg : z avg = AvgPool(U)

[0035] Then, in order to further obtain the dependency relationship between the channel features, two fully connected layers are used for non-linear parameterization. The bottleneck structure formed by the two fully connected layers can reduce the model complexity and improve the generalization ability. The first fully connected layer uses the ReLU activation function for non-linearization, and the second fully connected layer uses the sigmoid activation function for non-linearization:

[0036] s = F ex (z, W) = δ(g(z, W)) = δ(W2σ(z, W1))

[0037] where σ denotes the ReLU activation function, δ denotes the sigmoid activation function, and s denotes the channel weight. W1 e R C×C , W2 e R C×CTo prevent overfitting of the model, dropout layers are added after both fully connected layers. Experiments show that the two dropout layers improve the accuracy of the network and effectively enhance the stability of the model. The new features are obtained by calculating the product of the channel weight s and the initial input u

[0038]

[0039] where F scale (u c ,s c ) denotes the channel multiplication between the scalar s c and the feature map u c ∈R C It can be seen that the excitation operator maps the input specific descriptor to a set of channel weights, which can be regarded as a self-attention function in the channel direction.

[0040] (2) Spatial Bio-CS Block

[0041] In the same way as in the channel direction, the spatial direction takes the sequence feature U∈R D×C processed by the BBA residual module as input, and through a one-dimensional convolution, the network depth increases by one layer, while introducing a spatial weight scalar s, but does not change the size of the feature map. The final output is:

[0042]

[0043] (3) Merge Attention Module

[0044] The channel direction feature and the spatial direction feature are added together to perform a skip connection operation to obtain X, thereby ensuring the reusability of the features and more effectively combining the channel and spatial direction features:

[0045]

[0046] 4.3 Fully Connected Module Processing

[0047] The features X processed by the Bio-CS attention module are then processed by the global average pooling and fully connected module (with a Softmax activation function) to obtain the final deep sequence features.

[0048] Third stage: Deep enzyme structure feature construction

[0049] Step 5: Extract deep structure features using Pointnet++ point cloud network

[0050] The CA (calcium) atoms extracted from amino acids in the PDB file are used as points in the point cloud. The three-dimensional coordinates (x, y, z) of the calcium atoms are used as the coordinate features of the points in the point cloud, and the one-hot encoding of the amino acid residue containing the calcium atom is used as the sequence feature of the points in the point cloud. Therefore, the input to the PointNet++ network is a point set of size N×3 and a sequence feature matrix of size N×21, where N is the number of points in the point cloud. PointNet++ consists of three parts: a Sampling & Grouping & PointNet hierarchical structure, fully connected layers, and a Softmax layer. Figures 3(a)-3(d) As shown, the Sampling, Grouping, and PointNet hierarchical structure consists of multiple set abstraction units. The input to each set abstraction level is an N×(d+C) matrix, and the output is an N′×(d+C′) matrix, where N is the number of input points, d=3 is the coordinate dimension of the points, C is the feature dimension of the input points, N' is the number of output points, and C′ is the feature dimension of the output points. Each set abstraction unit mainly includes three parts: a Sampling layer, a Grouping layer, and a PointNet layer. The Sampling layer primarily samples the input points and selects several center points from these points; the Grouping layer uses the center points obtained from the Sampling layer to divide the point set into several regions; the PointNet layer encodes each of the obtained regions to obtain a new feature vector. The specific steps of this stage are as follows:

[0051] 5.1 Sampling layer.

[0052] Farthest point sampling (FPS) is used to sample points, selecting N′ points from N points. Compared to random sampling, this method can better cover the entire point set. The specific algorithm of FPS is as follows: First, randomly select a point x0 from the point set S with N points; then, select the point x1 that is farthest from point x0 using the distance formula; then find the point x2 that is farthest from point x1 in the remaining point set after removing x0 and x1, and so on until N′ sampling points are found.

[0053] 5.1 Grouping layer.

[0054] The input of this layer is a point set S with size N x (d+C) and a coordinate matrix of N' sampling points with size N' x 3, where N is the number of points, d=3 is the coordinate dimension, and C is the feature dimension. This layer uses the Ball query method to find K points in the input N points with the N' sampling points extracted by the Sampling layer as the center and a radius of R. The K points constitute a local region, and finally N' local regions are generated. The output of this layer is a coordinate and feature matrix N' x K x (d+C) containing N' sampling points.

[0055] 5.3 PointNet layer.

[0056] The input of this layer is N' x K x (d+C), and the output is N' x (d+C'), where C' is the new feature dimension. First, the point coordinates in the region are changed to relative coordinates around the center point before inputting into the network, which can better obtain the relationship between points and points. Then, the unordered point set is encoded through a multi-layer perceptron (MLP) network.

[0057] 5.4 Final extraction of structural features.

[0058] After the Sampling & Grouping & Pointnet layer, the PointNet++ network obtains the feature F e R N′×(d+C′) , and then integrates the feature F into a new vector F' through two fully connected layers with a Relu activation function. Finally, the output of the Softmax layer is used as the deep structural feature of the enzyme. Through end-to-end learning, the PointNet++ network can effectively extract features from enzyme structure data.

[0059] Fourth stage: model training and prediction based on TSK fuzzy system.

[0060] Through the above method, the deep features based on enzyme sequence data and the deep features based on enzyme structure data are obtained. How to make full use of these multi-view features of enzymes to efficiently predict the function of enzymes is also a challenging task. Traditional methods usually use a simple late fusion method, for example, splicing features of different perspectives. Although such a method is operable, it is difficult to fully realize the efficient cooperation of different perspective features. In view of this, a rule-based multi-view fuzzy system classifier is introduced to fully learn the multi-view deep feature data obtained in the previous step, and to realize more effective classification.

[0061] Compared with single-view classifiers, multi-view classifiers can mine more commonalities and characteristics between different view features, and get better prediction results. Multi-view classification technology has been widely studied, and researchers have proposed a variety of effective algorithms. Among them, fuzzy sets and fuzzy logic systems are gradually applied to multi-view classification. Multi-view fuzzy system classification method is a characteristic multi-view classification method, which not only can effectively realize the effective learning of multi-view data, but also has good transparency and is easy to explain, so it has shown its advantages in various modeling tasks. For example, TSK-FS-CVH, as a representative multi-view fuzzy system classification method, has been effectively applied to circRNA binding protein site prediction. In this paper, another multi-view fuzzy classifier MV-TSK-FS is introduced to construct the classifier based on enzyme multi-view deep features. MV-TSK-FS is developed based on the classic TSK-FS. In addition to the interpretability and data-driven learning ability of TSK-FS, MV-TSK-FS also has efficient multi-view collaborative learning ability, which can fully mine the consistency and complementarity between the multi-view features of enzymes. On the one hand, MV-TSK-FS can make full use of the differences of different views and more fully utilize the complementary information of different view data, thereby enhancing the generalization ability of the model. On the other hand, MV-TSK-FS can also make full use of the consistency between different views to guide and constrain the efficient learning of the classification model. Therefore, the adopted MV-TSK-FS method is very suitable for learning based on enzyme multi-view data to realize the classification of enzyme function.

[0062] The specific steps of this stage are as follows:

[0063] Step 6: 5-fold cross-validation is performed using TSK fuzzy system, and deep sequence features and deep structure features are trained respectively to learn the independent information of each view.

[0064] Step 7: 5-fold cross-validation is performed using multi-view TSK fuzzy system, and deep sequence features F s3 and deep structure features F t2 are retrained, and through the learned information, the importance between different views is adjusted using information entropy, and the sample is classified and tested.

[0065] The beneficial effects of the present application are:

[0066] (1) Other methods mostly only use sequence features or only use structure features of enzymes. This method takes into account both sequence features and structure features of enzymes, making the prediction information more complete.

[0067] (2) In the feature extraction process, the existing method is difficult to consider the two different perspectives of enzyme features, while in the present method, through the TSK fuzzy system, the sequence and structure features of the enzyme can be well re-learned in the new network. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is the algorithm method framework of the present application;

[0069] Fig. 2(a) is a graph of the accuracy of the model in the PDB database five-fold cross-validation result;

[0070] Fig. 2(b) is a graph of the precision of the model in the PDB database five-fold cross-validation result;

[0071] Fig. 2(c) is a graph of the regression rate of the model in the PDB database five-fold cross-validation result;

[0072] Fig. 2(d) is a graph of the F1 score of the model in the PDB database five-fold cross-validation result;

[0073] Fig. 3(a) is a BBCNet neural network;

[0074] Fig. 3(b) NBBA is a sequential network with Bio-CS attention module without BBA residual module;

[0075] Fig. 3(c) NBCS is a network with BBA residual module without Bio-CS attention module;

[0076] Fig. 3(d) NBBA-NBCS is a network without BBA residual module and without Bio-CS attention module. DETAILED DESCRIPTION

[0077] The present application will be described in detail below in conjunction with the drawings and examples.

[0078] As Figure 1 shown, the present application realizes a model for enzyme function prediction combining sequence features and structure features of enzymes, and the architecture is as shown in Figure 1 . First, the model constructs the initial features of the enzyme through the PDB file. Second, the model uses the BBC neural network to extract the deep features of the sequence information of the enzyme. Then, the Pointnet++ network is used to extract the deep features of the structure information of the enzyme. Finally, the TSK classification system based on fuzzy rules is used to train and classify the deep features.

[0079] Example 1

[0080] The performance of the method is evaluated using a 5-fold cross-validation experiment, and the final results are generated in an average manner. To evaluate the performance of the method, the proposed method is compared with several representative methods. Among them, ABLE is a method proposed in 2020 that only uses sequence information of enzymes, uses a bidirectional LSTM model based on attention to classify enzymes; EnzyNet is a method proposed in 2017 that only uses structural information of enzymes, uses a three-dimensional convolutional network to classify enzymes; DeepFri is a method proposed in 2019 that combines sequence information and structural information of enzymes, uses a graph convolutional network to classify enzymes using enzyme contact maps. The indicators are shown in Table 1.

[0081] As can be seen, the Accuracy, Precision, Recall and F1-score of the method are 0.9161, 0.9387, 0.8544 and 0.8946 respectively, which are the best among the four methods. This is because: 1) compared with the ABLE method which only uses sequence information and the EnzyNet method which only uses structural information, the multi-view method combines the sequence and structural features of enzymes, which can more comprehensively use the information of enzymes and obtain better prediction results. 2) Compared with the EnzyNet method which only uses structural information, the method extracts sequence information while more fully mining structural information, thereby obtaining better performance. 3) Compared with the DeepFri method, although both methods take into account sequence and structural information and make more in-depth mining of the two types of information, especially the multi-view classification technology used further improves the performance of the method.

[0082] Table 1 Comparison of performance of different prediction methods in predicting six types of enzyme functions.

[0083]

[0084] Example 2

[0085] The proposed method is compared with the version that only uses sequence information (denoted as view 1) and the version that only uses structural information (view 2). Specifically, BBCNet is taken as a method based on sequence as view 1, and PointNet++ is taken as a method based on structure as view 2. Five-fold cross-validation experiments are performed under the same data set and experimental environment, and the experimental results are shown in Table 2 and Figures 2(a)-2(d)

[0086] Table 2 Experimental results for evaluating the effectiveness of the multi-view learning mechanism

[0087]

[0088] Table 2 and Figures 2(a)-2(d) ​The average values of the five-fold cross-validation experiments obtained by the three methods are given, from which it can be seen that the performance of the method in this paper after comprehensively using sequence and structural information is effectively improved compared with the performance of using only sequence information (perspective 1) and only structural information (perspective 2). This also shows that the multi-perspective learning mechanism adopted in this paper is very effective.

[0089] According to Figures 2(a)-2(d) It can be found that the column chart of gray (ours) is obviously higher than the column charts of green (BBCNet) and red (PointNet++) in the five-fold cross-validation results of accuracy, precision, F1 score and regression rate.

[0090] The ablation experiment proves that by mining the complementary and consistent information between perspectives, multi-perspective learning is more suitable for complex data scenarios than single-perspective learning.

[0091] Embodiment 3

[0092] The ablation experiment is used to verify the effectiveness of the BBA residual module and the Bio-CS attention module. Figures 3(a)-3(d) The network structure diagrams designed for the ablation experiment of the enzyme sequence deep feature extraction network are shown in FIG. 3.

[0093] According to the experimental results in Table 3, it can be seen that the accuracy, precision, F1 score and regression rate of the NBBA network with the Bio-CS attention module and without the BBA residual module and the NBCS network with the BBA residual module and without the Bio-CS attention module are better than the NBBA-NBCS network without the BBA residual module and the Bio-CS attention module. It can be seen that the BBA residual module and the Bio-CS attention module can effectively improve the efficiency of the network. Further, the BBCNet network with the BBA residual module and the Bio-CS attention module obtains the best results. Therefore, the experiment proves that the BBCNet network combining the BBA residual module and the Bio-CS attention module can greatly improve the effectiveness of the constructed sequence deep feature extraction network.

[0094] Table 3 Ablation experiment results based on BBA residual module and Bio-CS attention module

[0095]

Claims

1. A multi-perspective enzyme function prediction method considering molecular structure and sequence mining, comprising the following steps: First step: initial feature extraction is performed on the amino acid sequences of the enzymes using the Biovec biological sequence processing method, and the amino acid sequence of each enzyme is represented as a vector as the initial sequence feature F of the enzyme s1 ; Second step: Extract the amino acid sequence from the PDB file of each enzyme in order, encode it with one-hot, and then extract the three-dimensional coordinates (x, y, z) of the carbon atoms from the amino acid sequence. Take the (n * 24)-dimensional matrix as the initial structural features F of the enzyme t1 ; Third step: initial sequence feature F of enzyme s1 SMOTE data oversampling processing is adopted to obtain feature F s2 ; Step 4: Based on feature F s2 , a deep sequence feature F is extracted using the BBCNet neural network s3 , which contains three modules: BBA residual module, Bio-CS attention module, and fully connected module; Step 5: initial structure feature F of enzyme t1 Pointnet++ point cloud network is used to extract deep structure feature F t2 This step includes Sampling layer, Grouping layer, PointNet layer and structure feature final extraction module In the Sampling layer, FPS is used to sample points, and N' points are selected from N points. Compared with random sampling, this method can better cover the entire point set. The specific algorithm of FPS is as follows: first, a point x0 is randomly selected from the point set S with N points; then, the point x1 farthest from the point x0 is selected by using the distance formula; then, the point x2 farthest from the point x1 is found in the remaining point set after removing x0 and x1, and so on until N' sampling points are found; In the Grouping layer, the Ball query method is used to find K points in a ball with a radius R and a center of N' sampling points extracted by the Sampling layer in the input N points. The K points constitute a local region, and finally N' local regions are generated. The output of this layer is a coordinate and feature matrix N'xKx(d+C) containing N' sampling points; In the PointNet layer, the point coordinates in the region are changed to relative coordinates around the center point before inputting into the network, which can better obtain the relationship between points. Then, the unordered point set is encoded through a multi-layer perceptron (MLP) network. Step 6: 5-fold cross-validation is used to train the deep sequence feature F s3 and the deep structure feature F t2 , respectively, to learn the independent information of each view. Step 7: Use multi-view TSK fuzzy system to perform 5-fold cross-validation, retrain deep sequence feature F s3 and deep structure feature F t2 , through the information learned before, adjust the importance between different views using information entropy, and test the sample classification.

2. The multi-view enzyme function prediction method that gives consideration to both molecular structure and sequence mining according to claim 1, wherein: The BBA residual module in the fourth step includes two Bi-LSTM layers and one self-attention layer. In order to better utilize local features, first, the two Bi-LSTM layers are concatenated to fuse past and future information. At the same time, the fused features and the features obtained after self-attention are fused again. Such a skip connection avoids the loss of local features, and the twice fusion finally forms the proposed BBA residual module. The module includes two Bi-LSTM layers, and the one-way output of each Bi-LSTM layer includes 128 nodes using the hyperbolic tangent activation function. The self-attention layer uses the hyperbolic tangent activation function.

3. The multi-view enzyme function prediction method that gives consideration to both molecular structure and sequence mining according to claim 2, characterized in that: In the Bio-CS attention module in the fourth step, u c is the initial embedding matrix, and is denoted as U ∈ R D×C The channel-level features are subjected to self-attention feature extraction by the channel Bio-CS module to obtain features The channel-level features are subjected to self-attention feature extraction by the spatial Bio-CS module to obtain features Finally, the channel-directional features and the spatial-directional features are added to perform a skip connection operation to obtain X, thereby ensuring the reusability of the features and more effectively combining the channel and spatial directional features, which is denoted as:

4. The multi-view enzyme function prediction method considering both molecular structure and sequence mining of claim 3, wherein: The fully connected module of the fourth step processes the features X processed by the Bio-CS attention module through global average pooling and a fully connected module to obtain the final deep sequence features.

5. The multi-view enzyme function prediction method that takes into account both molecular structure and sequence mining of claim 4, wherein, In the structural feature final extraction of the fifth step, the PointNet++ network obtains a feature F after the Sampling & Grouping & Pointnet layer N′×(d+C′) , then integrates the feature F into a new vector F' by using two full connection layers with a Relu activation function, and finally uses the output of the Softmax layer as the deep structural feature of the enzyme.

Images