Multistage enzyme function prediction method based on multi-view depth interactive learning

Abstract

The application belongs to the field of intelligent cell biological recognition, and relates to a multi-level enzyme function prediction method based on multi-view deep interaction learning. The main process of the method is as follows: firstly, enzyme sequences are converted into a combination view of solvent accessibility and secondary structure, a PSSM view containing evolutionary information and a domain view, so as to construct initial multi-view feature data; then, the above three original views are input into different deep specificity network modules to obtain deep specificity information of each view; further, a deep view interaction network is designed to extract interaction information between views; finally, the specificity information of each view and the interaction information thereof are introduced into a subsequent multi-view adaptive loss weighting classification module for collaborative training, so that a final prediction model is trained.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent cell biological recognition, and particularly relates to a multi-level enzyme function prediction method based on multi-view deep interaction learning. BACKGROUND

[0002] Enzymes are a class of important biological molecules with catalytic action, and are indispensable substances of life. Living organisms cannot be separated from replication and reproduction, metabolism, aging and death, and metabolism is accompanied throughout the life cycle and is the basis of all life activities, and the catalytic action of enzymes is indispensable for metabolism. The dysfunction of certain enzymes can cause serious metabolic diseases. For example, the deficiency of alpha-galactosidase can cause skin lesions and nervous system lesions. To study the causes of these diseases, the key is to understand the functions of related enzymes. The most rigorous enzyme function identification method is a biological experiment method, but the enzyme function identification through biological experiments requires a large time cost and cost expenditure, and this way cannot cope with the problem of rapid increase in the number of new enzymes. Therefore, rapid function prediction of unknown enzymes has become an important research direction in the field of bioinformatics.

[0003] According to the chemical reactions catalyzed by enzymes, enzymes are divided into seven categories: oxidoreductases, transferases, hydrolases, lyases, isomerases, synthetases and translocases. The Enzyme Commission has developed a set of enzyme number classification method, that is, EC number to represent different enzymes. The EC number uses four digits to specify the functions of four levels of enzymes. For example, the EC number of alanine racemase is 5.1.1.1, wherein 5 represents that it belongs to the isomerase category, 5.1 represents the 5 subcategory: racemase and epimerase, 5.1.1 represents the 5.1 subcategory: enzymes of amino acids and derivatives, and 5.1.1.1 represents the 5.1.1 subcategory: alanine racemase. Therefore, the unknown enzyme can be functionally annotated by predicting the EC number of the enzyme.

[0004] Existing multi-view enzyme function prediction methods show better performance, but these methods still have the following deficiencies: (1) In multi-view data, the distinguishable information contained in each view is not consistent, but existing multi-view methods usually treat each view equally when performing enzyme function prediction, which makes the adaptability of these methods need to be improved, that is, the influence of different views needs to be adaptively regulated according to the specific modeling task. (2) The interactive information contained in the multi-view data is very useful for subsequent identification tasks, but most of the existing multi-view enzyme function prediction methods pay insufficient attention to such information and need to be further strengthened. (3) For multi-level enzyme function prediction, existing methods mostly use a hierarchical architecture, which needs to train multiple classification models, which makes the errors in the current level label prediction flow to the next level label, thereby causing the accumulation problem of label prediction errors. SUMMARY

[0005] To solve the above technical problems, the present application provides a multi-level enzyme function prediction method based on multi-view deep interactive learning.

[0006] The technical solution of the present application is as follows:

[0007] The multi-level enzyme function prediction method based on multi-view deep interactive learning comprises a shallow multi-view data construction stage, a deep specificity network stage, a deep interactive network stage, and a multi-view adaptive loss weighting classifier stage, and the four stages are as follows:

[0008] The first stage is the shallow multi-view data construction stage. This stage includes three steps, namely, Property initial view construction, PSSM initial view construction, and Domain initial view construction, and the specific steps are as follows:

[0009] The function of an enzyme depends on its local structure and properties, such as the secondary structure of the enzyme and solvent accessibility. The present application uses RaptorX Property to predict the secondary structure and solvent accessibility. This tool can predict the probability of each amino acid folding into three helices (alpha helix, beta sheet, random coil) and the probability of being in three states of solvent accessibility (buried, intermediate and exposed). By combining the two properties, the enzyme sequence can be encoded into an Lx6 matrix.

[0010] Evolutionary information is information generated during the evolution process, which is of great significance to the function prediction of enzymes. PSSM is a common representation of evolutionary information and is used in many bioinformatics studies. In the present application, the PSI-BLAST program is used to perform multiple sequence alignment between each enzyme sequence and the Swiss-Prot database to obtain a PSSM matrix to encode the sequence. When the program is executed, the number of iterations is set to 3, and the E-value cutoff value for each iteration is set to 0.002. The final PSSM view-based feature matrix obtained is an Lx20 matrix, where L is the length of the enzyme sequence, and each column represents one kind of amino acid, a total of 20 kinds of amino acids.

[0011] Proteins are composed of domains, which are repeated fragments with highly similar sequences in different proteins, and are the units of protein structure and evolution. Pfam is a large protein domain and family database based on multiple sequence alignment and hidden Markov model, which currently covers 19179 common protein domains. Based on this fact, the application uses hmmer to search the domains of enzyme sequences in the Pfam database, and if it contains a certain domain, it is coded as 1; otherwise, it is coded as 0. Finally, for each enzyme sequence, a 19179-dimensional vector is used to encode its domain.

[0012] The specific steps of this stage are as follows:

[0013] Step 1: Convert the original enzyme sequence into Property initial view feature X using RaptorX Property 1 .

[0014] Step 2: Perform multiple sequence alignment of each sequence with the Swiss-Prot database using PSI-BLAST to obtain PSSM initial view feature X containing evolutionary information 2 ;

[0015] Step 3: Use hmmer to search the domains contained in each sequence in the Pfam database to obtain Domain initial view feature X 3 ; obtain the preliminary multi-view dataset D = {X 1 , X 2 , X 3 , y}.

[0016] Second stage: Since deep learning can dig out the nonlinear relationship in the data, therefore, using deep learning can obtain more discriminative features. For the shallow feature data obtained from the above three views, the application designs three different deep network models according to their respective characteristics, and obtains more discriminative multi-view specific features from the shallow view feature data through these deep networks.

[0017] Figure 2The CNN network architecture used for Property view depth-specific information extraction includes three two-dimensional convolution layers, three pooling layers, one flat layer, one dropout layer, and one fully connected layer. The first layer of the convolution layer is 16 3*3 convolution kernels, obtaining 16 998*4 feature maps. The pooling length of the second layer of the pooling layer is 2, obtaining 16 499*2 feature maps. The third layer is 16 3*3 convolution kernels, obtaining 16 501*4 feature maps. The pooling length of the fourth layer of the pooling layer is 2, obtaining 16 250*2 feature maps. The fifth layer is 16 3*3 convolution kernels, obtaining 16 252*4 feature maps. The pooling length of the sixth layer of the pooling layer is 2, obtaining 16 126*2 feature maps. The seventh layer is a flat layer, obtaining 1 1*4032 feature map. The eighth layer is a fully connected layer, converting the 1 1*4032 feature map into a 1*128 vector. The ninth layer is a dropout layer with a probability of 0.5, obtaining a 1*128 feature vector.

[0018] Figure 3 The CNN network architecture used for PSSM view depth-specific information extraction includes three parallel CNN architectures. The first CNN architecture is: first through 128 16*20 convolution kernels, obtaining 128 983*1 feature maps, then using 1-max pooling operation, obtaining 128 1*1 feature maps, and then using a flat layer to obtain a 1*128 feature vector. The second CNN architecture is: first through 128 24*20 convolution kernels, obtaining 128 975*1 feature maps, then using 1-max pooling operation, obtaining 128 1*1 feature maps, and then using a flat layer to obtain a 1*128 feature vector. The first CNN architecture is: first through 128 32*20 convolution kernels, obtaining 128 967*1 feature maps, then using 1-max pooling operation, obtaining 128 1*1 feature maps, and then using a flat layer to obtain a 1*128 feature vector. Then the vectors obtained by the three CNNs are spliced into a fully connected layer to obtain a 1*128 feature vector.

[0019] The multilayer perceptron network architecture used for domain view depth-specific information extraction consists of two dropout layers and three fully connected layers. The first fully connected layer transforms a 1*19179 feature map into a 1*1025 vector; the second layer is a dropout layer with a probability of 0.5, resulting in a 1*1025 feature map; the third layer is a fully connected layer, transforming the 1*1025 feature map into a 1*512 vector; the fourth layer is a dropout layer with a probability of 0.5, resulting in a 1*512 feature map; and the fifth layer is a fully connected layer, transforming the 1*512 feature map into a 1*128 vector.

[0020] The specific steps for this stage are as follows:

[0021] Step 1: For X 1 Design a depth-specific network to extract view depth features S with higher property discriminative power. 1 ;

[0022] Step 2: For X 2 Design a depth-specific network to extract more discriminative view depth features S from the PSSM. 2 ;

[0023] Step 3: For X 3 Design a depth-specific network to extract view depth features S with higher property discriminative power. 3 ;

[0024] The third stage: The depth-specific networks for the three views described above only mine the specific information of the respective views. To further mine the interaction information between views, this invention proposes a deep interaction network, the framework of which is shown in the figure below. Figure 4 As shown, the network consists of two stages: Stage 1 is the feature learning stage, which uses fully connected layers and dilated convolutions to extract deep features from the feature vectors and transform them into the same common space; Stage 2 is the view interaction stage, which imports the deep features extracted from each view into the view interaction network to extract interaction information. The interaction information is generated by the interrelationships between different views, explicitly modeling the relationships between views and exploring potential consistent information.

[0025] Compared to traditional convolutions, dilated convolutions can acquire multi-scale contextual information by setting different dilation rates, and support exponential expansion of the receptive field without sacrificing resolution and coverage. We employ dilated convolutions to extract long-range interactions in the shallow Property view. This dilated convolution consists of five layers, each using a 3×3 kernel with dilation rates of 1, 2, 4, 8, and 16. The feature maps obtained from the dilated convolutions are then fed into max-pooling layers to further extract salient features, yielding the depth features of the Property view. Next, we use the same dilated convolution architecture as the Property view to learn the PSSM shallow view. Finally, for the Domain shallow view, a three-layer multilayer perceptron network architecture is employed.

[0026] In the view interaction phase, we first construct the interaction matrix C. i This indicates the interaction between views, specifically:

[0027] C i =v k v l T

[0028] Where k represents the k-th view, l represents another view, and v k C represents the value of the k-th view. i The result of the association matrix between the k-th view and the l-th view is saved.

[0029] Then we use a CNN model with pooling layers to learn local interaction features, and then use the interaction matrix C i We flatten the graph to learn global features. We concatenate local and global features into [a] il ||a ig (|| represents the join operator that connects local and global features), and this is put into the MLP as the final interaction feature a. i This involves extracting deeper interactive information, further identifying the interactivity between views, and maximizing the relevance between them. The specific process can be described as follows:

[0030] a il =Pooling(ReLU(W c *C i +b c ))

[0031] a ig =Flatten(C i )

[0032] a i =MLP([a il ||aig ])

[0033] Among them W c * represents a filter, * represents a convolution operation, and b c This represents the bias vector. Here, we use ReLU and maxpooling as the nonlinear activation and pooling functions, respectively.

[0034] Finally, we associate the interaction information a with each view. i The data is merged to obtain the final interactive information I.

[0035] The specific steps for this stage are as follows:

[0036] Step 1: Extract deep features from the feature vectors using fully connected layers and dilated convolutions, and transform them to the same common space.

[0037] Step 2: Import the depth features extracted from each view into the view interaction network to extract interaction information.

[0038] The fourth stage: This invention uses multi-view adaptive loss weighted classification for model training and function prediction. Through deep specificity networks and interaction networks, deep specificity information and deep interaction information for each view are obtained. These two types of information are combined and fed into the downstream multi-view classifier module for final decision-making.

[0039] Since each view contains unequal amounts of discriminative information, treating each view equally does not yield optimal prediction performance. Therefore, in the multi-view classifier module of our proposed method, we employ an adaptive weighting mechanism to fuse the losses from multiple views. The loss function of the entire model can be described as follows:

[0040]

[0041] Where M is the number of views, Z v Let L be the predicted value for the v-th view, and label be the true sample label. v Let α be the cross-entropy loss for the v-th view. v Let r be the weight of the v-th view, where r>1 is a constant, representing the weight of the v-th view raised to the power of r.

[0042] By introducing `r`, the view's weight can be adaptively adjusted based on the view's loss. Below, we explain how to update the weight `α`. v :

[0043] The optimization objective of model training is to minimize the loss function, which can be described as the following optimization problem:

[0044]

[0045] According to the Lagrange multiplier method, we can obtain:

[0046]

[0047] Where λ is a Lagrange multiplier, J(α,λ) is calculated with respect to α. v And by taking the derivative of λ and setting it to zero, we can obtain the weight α. v The update equation:

[0048] The above rules enable adaptive adjustment of the weights of each view during network training.

[0049] The specific steps for this stage are as follows:

[0050] Step 1: Based on the generated depth-specific information and interaction information T = {S} 1 ,S 2 ,S 3 ,I 1 ,I 2 ,I 3 Each view is categorized using a fully connected layer.

[0051] Step 2: Use the multi-view adaptive loss weighted classification module to fuse the classification results of multiple views.

[0052] During the use of this invention, the specific steps are as follows:

[0053] Step 1: Use the shallow multi-view data construction module to build the initial multi-view test dataset.

[0054] Step 2: Use a depth-specific network module to obtain feature information of the deep multi-view test dataset.

[0055] Step 3: Use the deep interactive network module to obtain the interaction information of the multi-view test dataset.

[0056] Step 4: Based on the generated feature information and interaction information, use multi-view adaptive loss weighted classification to train and calculate the prediction results.

[0057] The beneficial effects of this invention are:

[0058] 1) A multi-level enzyme function prediction model based on multi-view interactive learning is proposed. This model can not only mine the specific information of each view, but also effectively mine the interactive information between views.

[0059] 2) The proposed method introduces an adaptive loss weighted classification module into the model, which realizes adaptive learning of the importance of each view in decision-making.

[0060] 3) In order to effectively mine the interactive information between views, a new mechanism for interaction between views is proposed in the model; Attached Figure Description

[0061] Figure 1 This is a flowchart of the algorithm method framework of the present invention.

[0062] Figure 2 This is a depth-specific network framework diagram of the Property view of the present invention.

[0063] Figure 3 This is a depth-specific network framework diagram of the PSSM view of the present invention.

[0064] Figure 4 This is a diagram of the multi-view deep interactive network framework of the present invention.

[0065] Figure 5(a) shows a comparison of the present invention with three baseline methods on the enzyme dataset using ACC.

[0066] Figure 5(b) shows the F1 score of the present invention compared with three baseline methods on an enzyme dataset.

[0067] Figure 5(c) shows a comparison of the present invention with three baseline methods on the isozyme dataset ACC.

[0068] Figure 5(d) shows the F1 score of the present invention compared with three baseline methods on the isozyme dataset.

[0069] Figure 6(a) is an ACC comparison of the effectiveness of deep interaction and multi-view adaptive weighting in this invention.

[0070] Figure 6(b) is an F1 comparison of the effectiveness of deep interaction and multi-view adaptive weighting in this invention.

[0071] Figure 6(c) shows the ACC comparison of the effectiveness of multi-view learning in this invention.

[0072] Figure 6(d) shows the F1 comparison of the effectiveness of multi-view learning in this invention. Detailed Implementation

[0073] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0074] like Figures 1 to 4 As shown, this invention implements a multi-level enzyme function prediction model based on multi-view deep interactive learning, the architecture of which is as follows: Figure 1As shown, for the input enzyme sequence, the method first determines whether it is an enzyme or not, and then performs specific enzyme function prediction. This method specifically includes four modules: a shallow multi-view data construction module, a deep specific network module, a deep interaction network module, and a multi-view adaptive loss weighted classifier module. The shallow multi-view data construction module constructs features from three views; in the deep specific network module, different networks are used to extract deep specific features from different views; in the deep interaction network module, different views are associated to generate interaction information, maximizing the correlation between views. The multi-view adaptive loss weighted classification module combines the results from multiple views to construct the final enzyme function prediction classifier.

[0075] Example 1

[0076] The performance of our proposed method was evaluated using 5-fold cross-validation, and the final results were generated by averaging. We experimentally compared our invention with two existing representative enzyme function prediction methods. For the first type of method, since no code was available, the results were directly derived from the literature. For the second type of method, code was provided, and we re-run the relevant methods on the dataset used in this paper for testing, and then compared and analyzed them with our proposed method.

[0077] The first category of methods compared includes HECNet, DEEPre, ECPred, and EzyPred. Since HECNet, DEEPre, and EzyPred lack source code, to ensure the accuracy of the test results, we directly compare the results from the literature with our proposed method using the aforementioned comparison methods. Although ECPred provides code, for the sake of fairness, we also directly present the results reported in HECNet. Because HECNet and our proposed method use the same dataset and experimental settings, the comparison results are fair.

[0078] Table 1 shows the results of the five models (the prediction results for class 7 enzymes were removed from HECNet; for fairness in the comparison, our method also removed the prediction results for class 7 enzymes). Among the five methods, EzyPred has the lowest performance because its features are not deeply learned, resulting in poor generalization ability compared to other methods. DEEPre and HECNet's prediction performance is lower than MVDINET because: although these two methods have used multi-view data, they treat different views equally without considering the differences in the impact of each view on the prediction results; furthermore, these two methods fail to explore and utilize the interactions between different views. Compared to the above methods, MVDINET's performance is significantly improved after optimizing the network structure and learning view interaction information. In particular, MVDINET's F1 score in the third layer is 91.88%, while HECNet and DEEPre are 79.4% and 68.4%, respectively; similarly, in the prediction of the fourth layer, MVDINET's F1 score is 90.51%, while HECNet is only 81.9%, showing that our proposed model has a significant performance advantage.

[0079] Table 1: Performance Comparison with Existing Methods and Tools

[0080]

[0081] The second category of methods compared includes DeepEC, DETECT V2, and EFICAz. Since these methods provide pre-trained models, we tested them on our dataset. Figure 5a Figures 5b and 5b show the experimental results of MVDINET and three baseline methods of this invention. These results demonstrate that MVDINET outperforms other models in both accuracy and F1 score. These results also indicate that MVDINET has varying degrees of advantage over existing representative methods.

[0082] Example 2

[0083] Isoenzymes are special enzymes that catalyze the same reaction in an organism but have different molecular structures. For example, a protein can have multiple isoforms produced by the same gene. These isoforms usually perform the same function, but they can differ significantly in sequence length. One of the most difficult challenges in enzyme function prediction is accurately predicting the function of isoenzymes of different lengths. To address this, we collected isoenzymes from Swiss-Prot 2019. In our experiments, we removed isoforms with sequence lengths less than 50 or greater than or equal to 1000, as well as isoforms belonging to multiple categories.

[0084] Figure 5cFigure 5b shows the prediction results of MVDINET and several other representative methods on isoenzymes. It can be seen that MVDINET outperforms several other representative methods in almost all levels of enzyme function prediction, including accuracy and F1 scores, which also verifies the high adaptability of the proposed method MVDINET.

[0085] Example 3

[0086] To verify the effectiveness of the novel mechanism employed in this invention, several ablation experiments were conducted. First, we verify the effectiveness of the multi-view depth interaction component used in the proposed method. Specifically, we tested the prediction performance of the method with and without the interaction component. Here, we denote the model without the multi-view depth interaction component as MVDIENT_without_Interative. Figure 6a As shown in Figure 6b, the addition of interactive methods improves the model's prediction performance, especially at higher levels of enzyme function prediction. This demonstrates that the interaction information extracted from different views through the multi-view deep interaction module can fully uncover the potential common discriminative information of each view, resulting in a significant improvement in classification performance.

[0087] Furthermore, we validated the multi-view adaptive weighted loss classification component used in our method. Specifically, we tested the classification performance of directly concatenating information from multiple views, as well as the classification performance of fusing views using multi-view adaptive weighted classification. Here, we denote the model that directly concatenates views as MVDINET_without_Weight. Figure 6a As can be seen in 6b, the use of a multi-view adaptive weighted classification component enables adaptive learning of importance in decision-making, which greatly improves prediction performance.

[0088] Finally, we verified the effectiveness of the multi-view learning mechanism. We superimposed three views sequentially to form a comparative experiment, verifying the effectiveness of each view in multi-view learning. Specifically, we use MVDINET_V1 to represent the prediction model built based on PSSM single-view features; MVDINET_V12 to represent the prediction model built based on the collaboration of PSSM view and Domain view; and MVDINET_V23 to represent the prediction model built based on the collaboration of Domain view and Property view. We compared our proposed method MVDINET with the above three methods. The performance of the four methods is shown in Figure 6. Figure 6c As can be seen from 6d, the proposed method improves both performance metrics as the number of views increases. This indicates that representing the features of each view is valuable.

Claims

1. A multi-level enzyme function prediction method based on multi-view deep interactive learning, characterized in that: The steps are as follows: Step 1: Use the RaptorX Property to calculate the probability of the secondary structure of each amino acid in the sequence, as well as the probability of being in one of the three solvent-accessible states, as the first initial view feature X. 1 ; Step 2: Use PSI-BLAST to perform multiple sequence alignment with the Swiss-Prot database to obtain the PSSM matrix containing evolutionary information, which serves as the second type of initial view feature X. 2 ; Step 3: Use Hmmer to search for the structural domains contained in each sequence in the Pfam database, encode the results, and use them as the third type of initial view feature X. 3 ; A preliminary multi-view dataset D = {X} is obtained. 1 ,X 2 ,X 3 ,y}; Step 4: For X 1 Design a depth-specific network to extract view depth features S with higher property discriminative power. 1 ; Step 5: For X 2 Design a depth-specific network to extract more discriminative view depth features S from the PSSM. 2 ; Step 6: For X 3 Design a depth-specific network to extract view depth features S with higher property discriminative power. 3 ; Step 7: Using X 1 X 2 X 3 Design a deep interaction network to extract interaction information between views, explicitly model the relationships between views and explore potential consistent information to obtain interaction information I. 1 I 2 I 3 ; Step 8: Based on the generated depth-specific information and interaction information T = {S} 1 ,S 2 ,S 3 ,I 1 ,I 2 ,I 3 The prediction result Y is obtained by training using multi-view adaptive loss weighted classification.

2. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 1, characterized in that: The CNN network architecture used for Property view depth-specific information extraction in the fourth step includes three 2D convolutional layers, three pooling layers, one flattening layer, one dropout layer, and one fully connected layer. The first convolutional layer of the CNN network has 16 3x3 convolutional kernels, resulting in 16 998x4 feature maps; the second pooling layer has a pooling length of 2, resulting in 16 499x2 feature maps; the third layer has 16 3x3 convolutional kernels, resulting in 16 501x4 feature maps; the fourth layer… The pooling length of the first pooling layer is 2, resulting in 16 feature maps of 250*2. The fifth layer has 16 convolutional kernels of 3*3, resulting in 16 feature maps of 252*4. The sixth pooling layer has a pooling length of 2, resulting in 16 feature maps of 126*2. The seventh layer is a flattened layer, resulting in a feature map of 1*4032. The eighth layer is a fully connected layer, which converts the 1*4032 feature map into a 1*128 vector. The ninth layer is a dropout layer with a probability of 0.5, resulting in a 1*128 feature vector.

3. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 1 or 2, characterized in that: The CNN network architecture used for PSSM view depth-specific information extraction in step five comprises three parallel CNN architectures. The first CNN architecture first uses 128 16*20 convolutional kernels to obtain 128 feature maps of 983*1 pixels each, then performs 1-max pooling to obtain 128 feature maps of 1*1 pixels each, and finally uses a flattening layer to obtain a single 1*128 feature vector. The second CNN architecture first uses 128 24*20 convolutional kernels to obtain 128 feature maps of 975*1 pixels each. Then, 1-max pooling is used to obtain 128 1*1 feature maps, and then a flattening layer is used to obtain a 1*128 feature vector. The first CNN architecture is as follows: first, it goes through 128 32*20 convolutional kernels to obtain 128 967*1 feature maps, then 1-max pooling is used to obtain 128 1*1 feature maps, and then a flattening layer is used to obtain a 1*128 feature vector. Then, the vectors obtained from the three CNNs are concatenated and put into a fully connected layer to obtain a 1*128 feature vector.

4. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 1 or 2, characterized in that: The domain view depth-specific information extraction in step six uses a multilayer perceptron network architecture, which includes two dropout layers and three fully connected layers. The first fully connected layer converts a 1*19179 feature map into a 1*1025 vector; the second layer is a dropout layer with a probability of 0.5, resulting in a 1*1025 feature map; the third layer is a fully connected layer, converting a 1*1025 feature map into a 1*512 vector; the fourth layer is a dropout layer with a probability of 0.5, resulting in a 1*512 feature map; and the fifth layer is a fully connected layer, converting a 1*512 feature map into a 1*128 vector.

5. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 3, characterized in that: The domain view depth-specific information extraction in step six uses a multilayer perceptron network architecture, which includes two dropout layers and three fully connected layers. The first fully connected layer converts a 1*19179 feature map into a 1*1025 vector; the second layer is a dropout layer with a probability of 0.5, resulting in a 1*1025 feature map; the third layer is a fully connected layer, converting a 1*1025 feature map into a 1*512 vector; the fourth layer is a dropout layer with a probability of 0.5, resulting in a 1*512 feature map; and the fifth layer is a fully connected layer, converting a 1*512 feature map into a 1*128 vector.

6. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 1, 2, or 5, characterized in that: The seventh step of the multi-view deep interactive network architecture employs a dilated convolutional architecture to learn the global features of the Property view. This dilated convolutional architecture consists of five layers, each using a 3×3 convolutional kernel with dilation rates of 1, 2, 4, 8, and 16, respectively. The feature maps obtained from the dilated convolutions are then fed into a max-pooling layer to obtain a 1*128 vector. The same architecture is used for the PSSM view, resulting in a 1*128 vector. A three-layer perceptron architecture is used for the Domain view, also yielding a 1*128 vector. C... i =v k v l T Construct an interaction matrix for each view vector, where C i The association matrix between the k-th view and the l-th view is saved. Then, a CNN model with pooling layers is used to learn the local interaction features in the interaction matrix, and the interaction matrix C is then processed. i Flattening the surface to learn global features; Ultimately, we obtained three 1*128 vectors containing interactive information.

7. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 3, characterized in that: The seventh step of the multi-view deep interactive network architecture employs a dilated convolutional architecture to learn the global features of the Property view. This dilated convolutional architecture consists of five layers, each using a 3×3 convolutional kernel with dilation rates of 1, 2, 4, 8, and 16, respectively. The feature maps obtained from the dilated convolutions are then fed into a max-pooling layer to obtain a 1*128 vector. The same architecture is used for the PSSM view, resulting in a 1*128 vector. A three-layer perceptron architecture is used for the Domain view, also yielding a 1*128 vector. C... i =v k v l T Construct an interaction matrix for each view vector, where C i The association matrix between the k-th view and the l-th view is saved. Then, a CNN model with pooling layers is used to learn the local interaction features in the interaction matrix, and the interaction matrix C is then processed. i Flattening the surface to learn global features; Ultimately, we obtained three 1*128 vectors containing interactive information.

8. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 4, characterized in that: The seventh step of the multi-view deep interactive network architecture employs a dilated convolutional architecture to learn the global features of the Property view. This dilated convolutional architecture consists of five layers, each using a 3×3 convolutional kernel with dilation rates of 1, 2, 4, 8, and 16, respectively. The feature maps obtained from the dilated convolutions are then fed into a max-pooling layer to obtain a 1*128 vector. The same architecture is used for the PSSM view, resulting in a 1*128 vector. A three-layer perceptron architecture is used for the Domain view, also yielding a 1*128 vector. C... i =v k v l T Construct an interaction matrix for each view vector, where C i The association matrix between the k-th view and the l-th view is saved. Then, a CNN model with pooling layers is used to learn the local interaction features in the interaction matrix, and the interaction matrix C is then processed. i Flattening the surface to learn global features; Ultimately, we obtained three 1*128 vectors containing interactive information.

9. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 1, 2, 5, 7, or 8, characterized in that: The eighth step, which involves multi-view adaptive loss weighted classification, CNN network architecture used for Property view depth-specific information extraction, CNN network architecture used for PSSM view depth-specific information extraction, multi-layer perceptron network architecture used for Domain view depth-specific information extraction, and multi-view depth interaction network architecture, uses adaptive loss weighting combined with multiple view classification results based on the generated depth-specific information and interaction information. The loss function is the cross-entropy multi-class cross-entropy loss function.

10. The multi-level enzyme function prediction method based on multi-view deep interactive learning as described in claim 3, characterized in that: The eighth step, which involves multi-view adaptive loss weighted classification, CNN network architecture used for Property view depth-specific information extraction, CNN network architecture used for PSSM view depth-specific information extraction, multi-layer perceptron network architecture used for Domain view depth-specific information extraction, and multi-view depth interaction network architecture, uses adaptive loss weighting combined with multiple view classification results based on the generated depth-specific information and interaction information. The loss function is the cross-entropy multi-class cross-entropy loss function.

Images