Protein AI enzyme digestion prediction method based on CNN-Bilstm model and application

Abstract

The invention discloses a protein AI enzyme digestion prediction method based on a CNN-Bilstm model and application, and belongs to the field of proteomics. The method comprises the following steps: 1, collecting and preprocessing protein sequence information, and carrying out characterization processing on data to obtain a standard data set; 2, designing a protein enzyme digestion prediction model based on a CNN-BiLstm architecture, and inputting the data set in the step 1 into the model for learning to obtain a trained model; and step 3, inputting to-be-predicted protein sequence data into the trained model in the step 2 to obtain a prediction result. According to the method, by combining a convolutional neural network (CNN) and a bidirectional long-short term memory network (BiLSTM), local and global features of a protein sequence can be captured more accurately, and the accuracy of enzyme digestion prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of proteomics, and in particular to a protein AI enzyme cleavage prediction method and application based on a CNN-Bilstm model. Background Art

[0002] Protein degradation occurs in a protease-catalyzed hydrolysis system, and this process has broad application value in many research fields such as biomedicine and food science. A deep understanding of the mechanism of action of enzymes in the process of protein hydrolysis and their impact on protein degradation characteristics is crucial to the advancement of proteomics research and spatial structure analysis. However, the complexity of the enzymatic hydrolysis process lies in the fact that it is affected by multiple factors, including the distribution of amino acid residues in the protein sequence, enzymatic hydrolysis conditions (such as pH, temperature, enzyme concentration, etc.), and the secondary and tertiary structures of the protein itself. These factors work together to lead to significant uncertainty in the formation of enzymatic hydrolysis sites and enzymatic hydrolysis products. Therefore, there is currently a lack of a model that can effectively predict the enzymatic hydrolysis process in this field. This limitation greatly restricts the quantitative study and precise control of protein degradation behavior.

[0003] At present, the prediction of protein enzymatic hydrolysis still faces significant challenges, and a method that can effectively predict the hydrolysis behavior of proteins under the action of enzymes has not yet been developed. In the research process of proteomics and structural biology, accurately predicting whether the enzyme cuts the protein sequence is of great significance for understanding protein function, studying protein-protein interactions, and developing new biopharmaceuticals. However, current research still lacks effective tools or models for predicting protein enzymatic hydrolysis sites. Although some methods based on sequence characteristics and statistical analysis have been proposed, these methods have limited application effects in most enzymatic hydrolysis systems. Therefore, developing an enzymatic hydrolysis prediction model with high prediction accuracy is a key technical challenge in the field of protein research. It has great research significance for promoting the development of related fields such as proteomics identification.

[0004] To this end, we proposed a protein AI enzyme cleavage prediction method and application based on the CNN-Bilstm model to solve the above problems. Summary of the invention

[0005] To solve the above technical problems, the primary purpose of the present invention is to provide a method and application for protein enzymatic cleavage prediction. The method combines convolutional neural networks and BiLstm networks to build and train artificial intelligence models after preprocessing protein data, thereby realizing protein sequence enzymatic cleavage prediction. The method no longer relies on the type of enzyme and the early Keil rules, and can also associate species sources by selecting training data sets to ensure the accuracy of protease cleavage prediction. In addition, the proposed technology can be calculated on CPU or GPU, has a wide range of applications and is lightweight, and is suitable for the analysis and processing of high-throughput omics data. The method has significant efficiency advantages and application potential in proteomics in the fields of biomedicine and food science.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A protein AI enzyme cleavage prediction method and application based on a CNN-Bilstm model, comprising the steps of collecting protein sequence information and preprocessing it, and characterizing the preprocessed protein sequence information to obtain a standard data set; The characterization process of the protein sequence information is to retain only the polypeptide data directly related to the enzyme cleavage, and clean and remove the rest of the irrelevant polypeptide information; specifically includes: The maximum number of missed cleavages of the peptide is set to N, that is, each peptide contains at most M sites. A sliding shearing window is used, and the window length is set to W, that is, the N-terminus and C-terminus of the potential enzyme cleavage site each contain W-1 / 2 amino acid residues; the screened peptides will be mapped back to the original protein sequence one by one, and the sequences of W-1 / 2 amino acids on both sides of the enzyme cleavage site will be extracted as the training data set of the model, denoted as S; according to the position of the enzyme cleavage site in the peptide, it will be marked as 0 or 1 (representing uncut / cut, MC / C, respectively) as the label of the training data, denoted as L Step 2: Establish a protein enzyme cleavage prediction model based on the CNN-BiLstm architecture, wherein the model includes a sequence feature extraction layer, a BiLstm layer, and an output layer; input a standard data set into the model for learning to obtain a trained model; Step 3: Input the protein sequence to be tested into the trained model to obtain the prediction result; The working steps of the trained model in step 2 are as follows: S21: The training data set S in step 1 is one-hot encoded to obtain S1; S22: Input the one-hot encoded data S1 obtained in step S21 into the sequence feature extraction layer for feature extraction to obtain a vector S containing the sequence data P , the vector S P Contains implicit information between amino acid residues at each position in the input sequence; S23: The vector S obtained in step S22 is P Input BiLstm layer to extract global information between sequences, and obtain local and global information S between amino acids in the input sequence. L ; S24: The hidden local and global information S obtained in step S23 is L Input to the output layer to get the predicted result L P , i.e. whether the enzyme is cut or not; S25: The prediction result L obtained in step S24 is P Calculate the loss with the training data label L. If the loss converges, save the model parameters and record them as M. K ; If the loss does not converge, repeat steps S22 to S24.

[0007] Preferably, in step 1, the standard data set is obtained by performing a one-to-one mapping between the training data set S and the training data label L.

[0008] Preferably, the one-hot encoding calculation steps of the training data set S in step S21 are as follows: , Where L is a categorical variable of different categories, i.e., a label feature. For a specific amino acid site in a sequence, whether it is cut or not, i.e., if the category is L k , where k is an integer from 1 to n, then the one-hot encoded vector is , Indicates whether the current position belongs to the category. If it does, it is 1; if not, it is 0.

[0009] Preferably, the sequence feature extraction layer in step S22 includes a convolution layer and a pooling layer, wherein the convolution layer is mainly used to extract local features between sequences, and the number of convolution layers is denoted as C. i , the specific calculation method is as follows: , in, is the input matrix, is the output after convolution, represents the weight coefficient; The number of pooling layers is recorded as Pk, and its specific calculation method is as follows: , in For the input feature map, the size of the pooling window is , the stride is s, is the output after pooling, denoted as S P ,In order to ensure the accuracy of prediction, batch normalization and dropout operations are performed after the convolutional layer.

[0010] Preferably, the number of convolutional layers Ci is 1, the number of pooling layers is denoted as Pk is 1, and the dropout is 0.5.

[0011] Preferably, the Bilstm layer in step S23 is mainly divided into a forward and a backward Lstm layer, wherein the specific calculation method of the forward Lstm is as follows: , in, is the hidden state, which depends on the current input Sp and the hidden state of the previous time step , is the recursive function of the forward Lstm; The specific calculation method of backward Lstm is as follows: , In the formula, the hidden state of the backward Lstm layer depends on the current input and the hidden state at the next time step ; Combine the forward and backward hidden states to get the final hidden state ht.

[0012] Preferably, in step S24, the output layer is a fully connected layer with a dimension of Nq, and the specific calculation formula is as follows: , in: Represents the hidden local and global information vector obtained in step S23. Represents the weight matrix of the output layer. Represents the bias vector of the output layer. represents the activation function, is the sigmoid function.

[0013] Preferably, in step S25, the loss function adopts the BCE loss function, and the specific calculation method is as follows: , in It is the real label, indicating whether the enzyme is actually cut or not (L=1 means enzyme cutting, L=0 means no enzyme cutting). is the probability value predicted by the model, Represents the probability of enzyme cutting predicted by the model.

[0014] Preferably, the method is used for proteomic identification and analysis in the fields of biomedicine and food.

[0015] In summary, the beneficial effects of the present invention are: 1. The protein enzyme cleavage prediction method of the present invention can more accurately capture the local and global characteristics of protein sequences and improve the accuracy of enzyme cleavage prediction by combining convolutional neural networks (CNN) and bidirectional long short-term memory networks (BiLSTM). This method no longer depends on the type of specific enzymes, and thus has wider applicability to more types of protein sequences.

[0016] 2. The protein enzyme cleavage prediction method of the present invention can be efficiently run on a CPU or GPU and is suitable for large-scale proteomics data analysis. Its lightweight design ensures efficiency and savings in computing resources when processing large data sets, and is suitable for high-throughput omics research.

[0017] 3. The protein enzyme cleavage prediction method of the present invention can be trained for data from different species by adjusting the training data set, thereby achieving more targeted enzyme cleavage prediction. Different data set ratios can be selected during model training, and the ratio of training and test sets can be flexibly set, thereby enhancing the adaptability of the method.

[0018] In summary, the present invention proposes a method and application for protein enzyme cleavage prediction, which is proposed under the premise of fully considering the problems faced by current protein enzyme cleavage prediction. The method has the ability to process high-throughput omics data and is suitable for large-scale proteomics research. By flexibly adjusting the training data set and the standardized data preprocessing process, the prediction accuracy and robustness of the model are ensured, and the scope of application is wide, and it has important application potential and development prospects in the fields of biomedicine and food science. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the protein enzyme cleavage prediction method of the CNN-BiLstm model in the present invention; Figure 2 It is a schematic diagram of the characterization process in the present invention; Figure 3 Schematic diagram of the protein enzyme cleavage prediction method of the CNN-BiLstm model in the present invention; Figure 4 This is the prediction result diagram of Example 1. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0021] See also Figure 1 , a protein enzyme cleavage prediction method based on CNN-BiLstm model, comprising the following steps: Step 1: Collection and preprocessing of protein sequence information, and characterization of the data to obtain a standard data set; Step 2, design a protein enzyme cleavage prediction model based on the CNN-BiLstm architecture, and input the data set described in step 1 into the model for learning to obtain a trained model; Step 3: Input the protein sequence data to be predicted into the trained model described in step 2 to obtain the prediction result.

[0022] The specific steps of step 1 are: 1.1 Collection of protein sequence information: Mass spectrometry instruments can be used to collect sample data covering proteins from multiple species or integrate open source protein mass spectrometry data on the Internet; 1.2 Dataset preprocessing: Use relevant methods of proteomic identification to identify and analyze mass spectrometry data, and save the identified peptide sequences and protein database information of the original species: 1.3 Characterization Processing: See Figure 2 To ensure the prediction reliability of the model, only the peptide data directly related to the enzyme cleavage are retained, and the rest of the irrelevant peptide information is cleaned and eliminated. The maximum number of missed peptides is set to N, that is, each peptide contains a maximum of M sites, and a sliding shearing window is used, and the window length is set to W, that is, the N-terminus and C-terminus of the potential enzyme cleavage site each contain W-1 / 2 amino acid residues. Since the length of the identified peptides is at least W-1 / 2 amino acids, only those peptides with W-1 / 2 amino acid residues at the N-terminus of the potential enzyme cleavage site are retained, and the peptides with at least 1 amino acid at the C-terminus of the potential enzyme cleavage site are retained. The screened peptides will be mapped back to the original protein sequence one by one, and the sequences of W-1 / 2 amino acids on both sides of the enzyme cleavage site will be extracted as the training data set of the model, denoted as S. According to the position of the enzyme cleavage site in the peptide, it will be marked as 0 or 1 (representing uncut / cut, MC / C, respectively) as the label of the training data, denoted as L.

[0023] Preferably, the maximum number of missed cuts N of the polypeptide is 2, the site M is 3, and the window length W is 15.

[0024] 1.4 Construction of standard data set: Map the sequence data S in 1.3 to the label data L one-to-one to obtain a standardized data set, which can set the ratio of training and test sets according to training needs.

[0025] Preferably, the sequence identification method is sequence library search, and the ratio of training set to test set is 8 / 2 or 9 / 1.

[0026] The specific steps of step 2 are: 2.1 Design of a protein enzyme cleavage prediction model based on the CNN-BiLstm architecture, see Figure 3 , including sequence feature extraction layer, BiLstm layer, output layer, where: Sequence feature extraction layer: used to extract local features between peptide sequences to prepare for subsequent enzyme cutting prediction; BiLstm layer: used to identify complex sequence compositions in protein sequences and learn long-term dependencies between previous and next sequences; Output layer: used to output the cutting probability predicted by the model and convert it into a label feature indicating whether the site is cut or not; 2.2 The working steps of the protein enzyme cleavage prediction model of the CNN-BiLstm architecture described in step 2.1 are as follows: Step 2.2.1, perform one-hot encoding on the sequence data S obtained in step 1.4 and record it as S1; Step 2.2.2, input the one-hot encoded data S1 obtained in step 2.2.1 into the sequence feature extraction layer, perform feature extraction, and obtain a vector SP containing sequence data, wherein the vector contains implicit information between amino acid residues at positions in the input sequence; Step 2.2.3, input the implicit information SP obtained in step 2.2.2 into the BiLstm layer to extract the global information between sequences, and obtain the local and global information SL between the amino acids in the input sequence; Step 2.2.4, input the hidden local and global information SL obtained in step 2.2.3 into the output layer to obtain the predicted result LP, i.e., whether the enzyme is cut or not; Step 2.2.5, calculate the loss between the result LP of whether the enzyme is cut or not obtained in step 2.2.4 and the real label L; Step 2.2.6, according to the result of loss calculation in step 2.2.5, repeat steps 2.2.2-2.2.4 until the loss converges, and save the model parameters as Mk; In step 2.2.1, the one-hot encoding calculation process of sequence data is as follows: , Where L is a categorical variable of different categories, i.e., a label feature. For a specific amino acid site in a sequence, whether it is cut or not, i.e., if the category is Lk, where k is an integer from 1 to n, then the one-hot encoding vector is , Si is whether the current position belongs to this category, if yes, it is 1, if not, it is 0; In step 2.2.2, the sequence feature extraction layer mainly includes a convolution layer and a pooling layer, wherein the convolution layer is mainly used to extract local features between sequences. The number of convolution layers is denoted as Ci, and the specific calculation method is as follows: , Where Xi+m,j is the input matrix, Yi,j,k is the output after convolution, and Wm,0,k represents the weight coefficient.

[0027] The number of pooling layers is recorded as Pk, and its specific calculation method is as follows: , in For the input feature map, the size of the pooling window is p×p, the stride is s, The output after pooling is denoted as Sp. In order to ensure the accuracy of prediction, batch normalization and dropout operations are performed after the convolutional layer.

[0028] Preferably, the number of convolutional layers Ci is 1, the number of pooling layers is denoted as Pk is 1, and the dropout is 0.5.

[0029] In step 2.2.3, the Bilstm layer is mainly divided into forward and backward Lstm layers, wherein the specific calculation method of the forward Lstm is as follows: , In the formula is the hidden state, which depends on the current input Sp and the hidden state of the previous time step , is the recursive function of the forward Lstm.

[0030] The specific calculation method of backward Lstm is as follows: , The hidden state of the backward Lstm layer depends on the current input Spt and the hidden state of the next time step. ; The forward and backward hidden states are combined to obtain the final hidden state ht. The specific calculation method is as follows: , Therefore, ht contains the bidirectional dependency information of the sequence, including both the information before time step t and the information after time step t.

[0031] After processing all time steps t, we get the hidden state matrix SL of the entire sequence, where each row is the bidirectional Lstm hidden state representation of a time step: , Where SL is a matrix of dimension T×2d, which contains the local information of the input sequence (capturing short-term dependencies through Lstm) and global information (capturing bidirectional long-distance dependencies through BiLSTM); The above formula can finally be expressed as follows: ,

[0032] In step 2.2.4, the output layer is mainly a fully connected layer with a dimension of Nq. The specific calculation formula is as follows: , Where: SL represents the hidden local and global information vector obtained in step 2.2.3. WL represents the weight matrix of the output layer. bL represents the bias vector of the output layer. f represents the activation function, is the sigmoid function.

[0033] Preferably, the dimension Nq of the fully connected layer is 512; Step 2.2.5, the loss function uses the BCE loss function, and the specific calculation method is as follows: , Where Li is the true label, indicating whether the enzyme is actually cut or not (L=1 indicates cut, L=0 indicates no cut). Li is the probability value predicted by the model, and Lpi indicates the probability of the model predicting cut. Repeat steps 2.2.1-2.2.5 until the loss value converges, and save the parameters of the model at this moment as Mk; Preferably, other relevant parameters in step 2 are set as follows: activation function: Relu, dropout rate: 0-1, Batch_size: 0-2048; the network uses Adam optimizer, training 0-100 rounds, learning rate: 0-1.

[0034] The specific steps of step 3 are: inputting the protein sequence data to be predicted into the model Mk trained in step 2 to obtain the prediction result.

[0035] The data to be measured in step 3 can be a protein mass spectrum collected by any mass spectrometer.

[0036] The application effects of the present invention are described in detail below in conjunction with the embodiments.

[0037] The data set used in the example is the protein data of three species: H. sapiens, D. melanogaster and M. musculus, and the active enzyme is trypsin. The detailed information is as follows: Table 1 Sources of public datasets

[0038] Some important model parameters are set as follows: the maximum number of missed cleavages of polypeptides N is 2, the site M is 3, the window length W is 15, the number of convolution modules: 1, the number of pooling layers: 1, the discard rate: 0.1, the learning rate: 0.001, the batch size: 128, the dimension of the fully connected layer: 512, the network uses the Adam optimizer, the activation function: Relu; the network uses the Adam optimizer, and is trained for 10 rounds. In the implementation example, the weights of each layer of the network are updated by the loss function to obtain the optimal network structure, thereby realizing the prediction of whether the protein sequence is cleaved or not. The prediction results are as follows: Figure 4 shown.

[0039] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A protein AI enzyme cleavage prediction method based on CNN-Bilstm model, characterized in that: The method comprises: Step 1: Collect protein sequence information and perform preprocessing, and perform characterization processing on the preprocessed protein sequence information to obtain a standard data set; Among them, the pre-processed protein sequence information is characterized to retain only the peptide data directly related to the enzyme cleavage, and clean and eliminate the rest of the irrelevant peptide information; specifically including: The maximum number of missed cleavages of a peptide is set to N, that is, each peptide contains at most M sites. A sliding shearing window is used, and the window length is set to W, that is, the N-terminus and C-terminus of the potential enzyme cleavage site each contain W-1 / 2 amino acid residues; the screened peptides will be mapped back to the original protein sequence one by one, and the sequences of W-1 / 2 amino acids on both sides of the enzyme cleavage site will be extracted as the training data set of the model, denoted as S; according to the position of the enzyme cleavage site in the peptide, it will be marked as 0 or 1, specifically, 0 means no enzyme cleavage, and 1 means enzyme cleavage, which is used as the label of the training data, denoted as L; Step 2: Establish a protein enzyme cleavage prediction model based on the CNN-BiLstm architecture, wherein the model includes a sequence feature extraction layer, a BiLstm layer, and an output layer; input a standard data set into the model for learning to obtain a trained model; Step 3: Input the protein sequence to be tested into the trained model to obtain the prediction result; The steps for obtaining the trained model in step 2 are as follows: S21: The training data set S in step 1 is one-hot encoded to obtain S1; S22: Input the one-hot encoded data S1 obtained in step S21 into the sequence feature extraction layer for feature extraction to obtain a vector S containing the sequence data P , the vector S P Contains implicit information between amino acid residues at each position in the input sequence; S23: The vector S obtained in step S22 is P Input BiLstm layer to extract global information between sequences, and obtain local and global information S between amino acids in the input sequence. L ; S24: The hidden local and global information S obtained in step S23 is L Input to the output layer to get the predicted result L P , i.e. whether the enzyme is cut or not; S25: The prediction result L obtained in step S24 is P Calculate the loss with the training data label L. If the loss converges, save the model parameters and record them as M. K ; If the loss does not converge, repeat steps S22 to S24.

2. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 1, characterized in that: In step 1, the standard data set is obtained by performing a one-to-one mapping between the training data set S and the training data label L.

3. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 1, characterized in that: The one-hot encoding calculation steps of the training data set S in step S21 are as follows: ， Where L is a categorical variable of different categories, i.e., a label feature. For a specific amino acid site in a sequence, whether it is cut or not, i.e., if the category is L k , where k is an integer from 1 to n, then the one-hot encoded vector is , Indicates whether the current position belongs to the category. If it does, it is 1; if not, it is 0.

4. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 1, characterized in that: The sequence feature extraction layer described in step S22 includes a convolution layer and a pooling layer, wherein the convolution layer is mainly used to extract local features between sequences. The number of convolution layers is denoted as C. i , the specific calculation method is as follows: ， in, is the input matrix, is the output after convolution, represents the weight coefficient; The number of pooling layers is recorded as Pk, and its specific calculation method is as follows: ， in For the input feature map, the size of the pooling window is , the stride is s, is the output after pooling, denoted as S P ,In order to ensure the accuracy of prediction, batch normalization and dropout operations are performed after the convolutional layer.

5. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 4, characterized in that: The number of convolutional layers Ci is 1, the number of pooling layers is denoted as Pk is 1, and the dropout is 0.

5.

6. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 1, characterized in that: The Bilstm layer in step S23 is mainly divided into a forward Lstm layer and a backward Lstm layer, wherein the specific calculation method of the forward Lstm is as follows: ， in, is the hidden state, which depends on the current input Sp and the hidden state of the previous time step , is the recursive function of the forward Lstm; The specific calculation method of backward Lstm is as follows: ， In the formula, the hidden state of the backward Lstm layer depends on the current input and the hidden state at the next time step ; Combine the forward and backward hidden states to get the final hidden state ht.

7. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 1, characterized in that: The output layer in step S24 is a fully connected layer with a dimension of Nq, and the specific calculation formula is as follows: ， in: represents the hidden local and global information vector obtained in step S23, represents the weight matrix of the output layer, represents the bias vector of the output layer, represents the activation function, is the sigmoid function.

8. The protein AI enzyme cleavage prediction method based on the CNN-Bilstm model according to claim 1, characterized in that: In step S25, the loss function adopts the BCE loss function, and the specific calculation method is as follows: ， in It is the real label, indicating whether the enzyme has actually been cut or not. Specifically, L=1 means it has been cut, and L=0 means it has not been cut. is the probability value predicted by the model, Represents the probability of enzyme cutting predicted by the model.

9. The method according to any one of claims 1 to 8, characterized in that: The method is used for proteomic identification and analysis in the fields of biomedicine and food.

Images