Polypeptide immunocompetence prediction and generation method fusing polypeptide physicochemical properties, sequence characteristics and word vector embedding
By integrating the physical and chemical properties, sequence characteristics and word vector embedding of polypeptides, ResNet and LSTM networks are constructed to successfully predict and generate polypeptides with immunoactive activity, solving the problems of high cost and low efficiency of predicting and generating polypeptides in the prior art. , realize efficient and accurate polypeptide screening and generation.
Patent Information
- Application Number
- CN202411079363.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-12-03
AI Technical Summary
The prior art has failed to effectively predict and generate immunoactive polypeptides, resulting in high cost and time-consuming experiments for predicting polypeptide activity.
Using the method of fusion polypeptide physical and chemical properties, sequence characteristics and word vector embedding, the ResNet network is constructed for polypeptide feature extraction and prediction through One-Hot encoding, amino acid physicochemical properties encoding and Word2Vec embedding vectors, and the LSTM generator module is used to generate false amino acid sequences. , and finally generates an immune-active polypeptide.
The efficient prediction and generation of peptides was achieved, the experimental cost was reduced, the peptide screening efficiency was improved, the accuracy rate reached 93.7%, and the immune activity of the peptide was verified.
Smart Images

Figure CN119068997A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of bioinformatics, and specifically is a method for predicting and generating polypeptide immune activity by integrating the physicochemical properties, sequence characteristics and word vector embedding of polypeptides. Background Art
[0002] Peptides represent a unique class that has some properties of small-sized drugs as well as the properties of large proteins and other biologics. However, the discovery and design of peptides using traditional methods is a time-consuming and expensive process. Among peptide drugs, active peptides with immune functions can treat diseases such as liver cancer, lung cancer, leukemia, and osteosarcoma, and are an important class of peptide drugs. With the accumulation of protein and peptide data and the improvement of computer computing power, the trend of using machine learning to predict the activity of peptides such as antimicrobial peptides has gradually emerged. However, no generation and discrimination model for immune-active peptides has been developed. In this case, if the activity of immune-active peptides can be predicted using the peptide's own information or if immune-active peptides can be directly generated, it is expected to ensure large-scale screening of peptides before conducting peptide activity prediction experiments in order to save costs. Summary of the invention
[0003] The technical problem to be solved by the present invention is: how to accurately predict the activity of immunologically active peptides using the self-information of the polypeptides, and directly generate polypeptides with immunological activity.
[0004] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:
[0005] A method for predicting and generating polypeptide immunological activity by integrating polypeptide physicochemical properties, sequence characteristics and word vector embedding, comprising the following steps:
[0006] Step 1: Preprocess the data;
[0007] Step 2: Use One-Hot coding and amino acid physicochemical property coding to obtain the sequence characteristics of the peptide and vectorize the peptide sequence;
[0008] Step 3: Use the Skip-Gram model in Word2Vec to obtain the embedding vector of amino acids;
[0009] Step 4: Construct a peptide immune activity prediction network (ResNet) based on the feature descriptors to extract peptide features;
[0010] Step 5: Use the non-immune active peptide dataset to transfer the model and save the pre-trained model with better results;
[0011] Step 6: According to the obtained matrix integrating the above three features, use the pre-trained model saved in step 5 for training, and use the trained model for weight feature fusion prediction;
[0012] Step 7: Construct a generative model generator module network (LSTM) architecture, use high-dimensional random noise as the generator module input, and map it into a false amino acid sequence through the generator module;
[0013] Step 8: Perform feature extraction on the immune active peptide and the false peptide sequence generated in step 7, extract the One-Hot code of step 2 and step 3, the physicochemical property code of amino acids, and the Word2Vec word embedding code;
[0014] Step 9: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network. The features of the two types of peptides generated in step 8 are input into the discriminator module for discrimination, and finally immune active peptides are generated.
[0015] Furthermore, the step 1 is specifically as follows:
[0016] Step 1.1: Preprocess the peptide sequence data and filter out the sequences with outliers and missing amino acids;
[0017] Step 1.2: Combine the immunoreactive and non-immunoreactive peptide datasets.
[0018] Furthermore, the step 2 is specifically as follows:
[0019] Step 2.1: For the immune-active and non-immune-active peptide datasets, the peptide sequences were converted into One-Hot codes and amino acid physicochemical property codes by using the One-Hot code generation method and the amino acid physicochemical property code generation method. Finally, the two codes were concatenated into a one-dimensional feature vector. The formula is:
[0020] Y = Y.concat(seq i )i=1,2,3...length(seq)#(1)
[0021] where seq i is the One-Hot code or physicochemical property code corresponding to the ith amino acid in the polypeptide sequence, and Y is the characteristic code of this sequence.
[0022] Furthermore, the step 3 is specifically as follows:
[0023] Step 3.1: Use a large number of sorted peptide sequences and remove redundant sequences. The remaining sequences are used as a corpus for training the Word2vec model to train the Word2vec model.
[0024] Step 3.2: Use the Skip-Gram model in the trained Word2vec to obtain the embedding vectors of 20 amino acids and obtain the word embedding vector of the immunoactive polypeptide. The Skip-Gram model is used to obtain the co-occurrence similarity of amino acids. The formula is as follows:
[0025]
[0026]
[0027] where c is the context size, a t is the first part of the t-th amino acid embedded, a t+j is the first partial embedding of the adjacent amino acids, n is the negative sampling size, a′ k is not in the same context as a t Co-occurring negative samples.
[0028] Furthermore, the step 4 is specifically as follows:
[0029] Step 4.1: Construct a peptide immune activity prediction network (ResNet) based on the feature descriptors. The ResNet network consists of multiple residual units, each of which can be expressed as:
[0030] y l =h(x l )+F(x l +w L )#(4)
[0031] x l+1 =f(y l )#(5)
[0032] where x l and x l+1 They represent the input and output of the lth residual unit respectively. Each residual unit generally contains a multi-layer structure. F is the residual function, which represents the learned residual, and h(x l )=x l represents the identity mapping, f is the ReLU activation function, and the learning feature formula from the shallow layer 1 to the deep layer L can be obtained from formula (4) and (5), as shown in formula (6).
[0033]
[0034] By performing reverse deduction according to the chain rule, we can derive the gradient formula of the reverse process, as shown in formula (7).
[0035]
[0036] In the formula: is the loss function in the Lth residual unit. The meaning of "1" in formula (7) is that the propagation gradient is carried out without loss, while the other residual gradients need to pass through the weighted layer and cannot be directly transmitted. Even if the value of the residual gradient becomes smaller, it will not all be -1. Therefore, the meaning of "1" is to prevent the gradient from disappearing, so that the residual neural network can be continuously accumulated and improved, reducing the difficulty of residual learning.
[0037] Furthermore, the step 5 is specifically as follows:
[0038] Step 5.1: Select eight types of non-immune active peptides from the data set, mainly including antibacterial peptides, antiviral peptides, antihypertensive peptides, hemolytic peptides, etc. The number of peptides of each active type is roughly equal. These eight types of peptides are combined into a transfer learning data set, and the feature descriptors of the peptide information in the data set are obtained according to the above three feature extraction methods.
[0039] Step 5.2: Use the ResNet network built in step 4 to perform transfer learning on the eight-category dataset, and save the pre-trained model with better results for subsequent peptide immune activity training.
[0040] Furthermore, the step 6 is specifically as follows:
[0041] Step 6.1: The 522 collected immune-active peptides are used as positive samples, and a variety of non-immune-active peptides selected from the database are used as negative samples. About 1,500 negative samples are selected, and then the positive and negative samples are spliced and shuffled, and then divided into training set and test set in a ratio of 2:1. The training set and test set are used according to the above three feature extraction methods to obtain the feature descriptors of the peptide information.
[0042] Step 6.2: Get the pre-trained model saved in step 5, train it using the training set, and save the model with better results.
[0043] Step 6.3: The model saved in step 6.2 is F. Model F can be divided into two parts. The first part is a convolutional neural network composed of residual blocks for extracting features. This module is named F1. The second part is a neural network composed of fully connected layers for obtaining the final result. This module is named F2.
[0044] Get the One-Hot code of the peptide sequence (E ot ), the physical and chemical properties of amino acids (E ph ) and Word2vec encoding (E wv ) and the overall features of the three codes (E all ) ; Input the above four codes into the F1 module to obtain four feature vectors Y with a length of 2048 ot F1, Y ph F1, Ywv F1, Y all F1, change Y ot F1, Y ph F1, Y wv F1 and Y all F1 concatenation is then passed through a convolutional layer to obtain a feature vector Y with a length of 2048 ot Y all , Y ph Y all , Y wv Y all ; Then E ot 、E ph 、E wv Input into model F to get three weight values u ot 、u ph 、u wv .
[0045] Then the concatenated feature vector Y ot Y all , Y ph Y all , Y wv Y all With weight u ot 、u ph 、u wv Corresponding multiplications give three eigenvectors Y ot , Y ph , Y wv , then physically add the three feature vectors to obtain a feature vector Y with a length of 2048, input Y into the feature extraction module F1, and finally output the probability value P of the polypeptide sequence having immune activity. The formula of step 6.3 is as follows:
[0046] Y i F1=F1(E i )i=ot,wv,ph#(8)
[0047] Y i Y all =concat(Y all F1, Y i F1)i=ot,wv,ph#(9)
[0048] Y i =muti(Y i Y all , Y i Y all )i=ot,wv,ph#(10)
[0049] Y=add(Y ot , Y ph , Y wv)#(11)
[0050] P=F2(Y)#(12)
[0051] Furthermore, the step 7 is specifically as follows:
[0052] Step 7.1: First, build the generative model generator module. The generator module is built using the LSTM network. The LSTM network consists of three parts: input gate, forget gate, and output gate. The input gate is responsible for adding useful features to the cell storage unit. This process consists of two parts. First, input the data into the tanh layer to get the new At the same time, a sigmoid layer is also required to obtain the data feature i t , and then the tanh layer to get Combined input into the cell storage unit. The formula is as follows
[0053] i t =σ(W i [h t-1 ,x t ]+b i )# (13)
[0054]
[0055]
[0056] Among them, W i and W C represents the corresponding weight value, b i and b C represents the corresponding bias, C t Represents the status value of the current cell.
[0057] Finally, LSTM selects the feature data that can be used as output from the current active peptide data features and outputs it through the output gate. t1 ,x t ] Get the new weight value o through a sigmoid layer t , and then store the C of the cell at the current moment t After a tanh layer, and then o + Multiply them together to get the formula:
[0058] o t =σ(W o [h t-1 ,x t ]+b o )# (16)
[0059] h t=o + *tanh(C t )# (17)
[0060] Among them, W o Represents the weight value of the output gate, b o represents the bias of the output gate, h t Represents the output value of the cell storage space at this moment;
[0061] Step 7.2: Use Gaussian random noise as the input of the generator module, and map it into an amino acid sequence after being processed by the generator module.
[0062] Furthermore, the step 8 is specifically as follows:
[0063] Step 8.1: Perform feature extraction on the immune active peptide and the false peptide sequence generated in step 7, extract the One-Hot code of step 2 and step 3, the physicochemical property code of amino acids, and the Word2Vec word embedding code.
[0064] Furthermore, the step 9 is specifically as follows:
[0065] Step 9.1: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network constructed in step 4. The features of the three types of peptides generated in step 8 are input into the discriminator module for discrimination, and finally immune active peptides are generated.
[0066] The present invention provides a method for predicting and generating polypeptide immunological activity by integrating the physicochemical properties, sequence characteristics and word vector embedding of polypeptides. This technical solution uses deep learning methods to predict and generate immunologically active polypeptides, and has the following beneficial effects compared with the prior art:
[0067] 1. The present invention takes into account the available features of polypeptides in different aspects, integrates the physicochemical information of polypeptides, word vector embedding information and polypeptide sequence features, and increases the effectiveness of polypeptide feature extraction.
[0068] 2. The ResNet network model is used to model the peptide sequence and word vector embedding and physical and chemical property characteristics, effectively learning the multi-faceted characteristics of the peptide, and using the multi-feature weight feature fusion method to obtain the prediction results, so both the prediction and generation models have good effects. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 It is a flow chart of the prediction model of immune active peptides.
[0070] Figure 2 It is the ROC curve of the performance comparison between the embodiment of the present invention and other methods under the test set.
[0071] Figure 3 It is a flow chart of the model for generating immune active peptides.
[0072] Figure 4 It is a comparison chart of the BLEU scores of peptide sequences generated by the generative model and non-immune active peptides and true immune active peptides. DETAILED DESCRIPTION
[0073] The specific embodiments of the present invention will be described in detail below. In order to avoid too much unnecessary details, the well-known structures or functions will not be described in detail in the following examples. The approximate language used in the following examples can be used for quantitative expression, indicating that the quantity can be allowed to have a certain change without changing the basic function. Unless otherwise defined, the technical and scientific terms used in the following examples have the same meaning as those generally understood by those skilled in the art to which the present invention belongs.
[0074] Example 1
[0075] like Figure 1 As shown, the specific embodiment process of the present invention is:
[0076] Step 1: First, 522 immune active peptides and 1600 non-immune active peptides with lengths ranging from 5 to 150 amino acids were obtained from databases such as DBAASP, APD3, and PlantPepDB. After removing duplicate, abnormal, and missing data, 500 immune active peptides and 1500 non-immune active peptides remained.
[0077] Step 2: Use the One-Hot encoding generation method and the amino acid physicochemical property encoding generation method to extract feature vectors from the immune active peptide dataset compiled in step 1, and obtain the sequence features and physicochemical property encoding of the peptides as feature inputs for the prediction model.
[0078] Step 3: Use non-redundant peptide sequences with a length of 5 to 150 amino acids from databases such as DBAASP and APD3, remove redundant sequences, and use the remaining sequences as training Word2 vec The corpus of the model finally obtained 52525 peptide sequence information, and these sequences were used to train Word2 vec Model: Use the Skip-Gram model in the trained Word2vec to obtain the embedding vectors of 20 amino acids, and obtain the sequence vectors of immune active polypeptides and non-immune active polypeptides, where the Skip-Gram model obtains the co-occurrence similarity of amino acids.
[0079] Step 4: Construct a peptide immune activity prediction network (ResNet) based on the feature descriptors. The ResNet network consists of multiple residual blocks, and the output of each residual block is a nonlinear transformation of the input. The residual block generally contains a multi-layer structure. The input x passes through two convolutional layers to obtain F(x). F(x) is added to the input x, that is, x+F(x). The final output is x+F(x), which is passed to the next residual block. Therefore, the residual block allows the ResNet network to learn the identity mapping by directly adding the input to the output of the nonlinear transformation. Finally, the probability of the prediction result is obtained.
[0080] Step 5: Select eight types of non-immune active peptides from the data set, mainly including antibacterial peptides, antiviral peptides, antihypertensive peptides, hemolytic peptides, etc. The number of peptides of each active type is limited to about 1000. These eight types of peptides are combined into a transfer learning data set, and the feature descriptors of the peptide information in the data set are obtained according to the above three feature extraction methods; then use the ResNet network built in step 4 to perform transfer learning on the eight-category data set, and save the pre-trained model with better results.
[0081] Step 6: The 522 collected immune-active peptides are used as positive samples, and a variety of non-immune-active peptides selected from the database are used as negative samples. About 1,500 negative samples are selected, and then the positive and negative samples are spliced and shuffled, and then divided into training sets and test sets in a 1:1 ratio. The training set and the test set are used according to the above three feature extraction methods to obtain the feature descriptors of the peptide information.
[0082] Get the pre-trained model saved in step 5, train it using the training set, and save the model with better results.
[0083] The saved model is set to F. Model F can be divided into two parts. The first part is a convolutional neural network composed of residual blocks for extracting features. This module is named F1. The second part is a neural network composed of fully connected layers for obtaining the final result. This module is named F2.
[0084] Get the One-Hot code of the peptide sequence (E ot ), the physical and chemical properties of amino acids (E ph ) and Word2vec encoding (E wv ) and the overall features of the three codes (E all ) ; Input the above four codes into the F1 module to obtain four feature vectors Y with a length of 2048 ot F1, Y ph F1, Y wv F1, Y all F1, change Yot F1, Y ph F1, Y wv F1 and Y all F1 concatenation is then passed through a convolutional layer to obtain a feature vector Y with a length of 2048 ot Y all , Y ph Y all , Y wv Y all ; Then E ot 、E ph 、E wv Input into model F to get three weight values u ot 、u ph 、u wv .
[0085] Then the concatenated feature vector Y ot Y all , Y ph Y all , Y wv Y all With weight u ot 、u ph 、u wv Corresponding multiplications give three eigenvectors Y ot , Y ph , Y wv , and then physically add the three feature vectors to obtain a feature vector Y with a length of 2048, input Y into the feature extraction module F1, and finally output the probability value P that the polypeptide sequence has immune activity.
[0086] Step 7: The generative model adopts the generative adversarial network architecture, which includes a generator module and a discriminator module. First, the generative model generator module is constructed. The generator module is built using an LSTM network. The LSTM network contains three gated units: input gate, forget gate, and output gate. The input gate is responsible for adding features useful for generating active peptides to the cell storage unit. This process consists of two parts. First, the data is input into the tanh layer to obtain a new At the same time, a sigmoid layer is also required to obtain the data feature i t , and then the tanh layer to get Combined input into the cell storage unit. Finally, LSTM selects the feature data that can be used as output from the current active peptide data features and outputs it through the output gate. It passes the spliced features through a sigmoid layer to obtain a new weight value o t , and then store the C of the cell at the current moment t After a tanh layer, and then tThen, Gaussian random noise is used as the input of the generator module, and is mapped into an amino acid sequence after being processed by the generator module.
[0087] Step 8: Perform feature extraction on the immune active peptide and the false peptide sequence generated in step 7, extract the One-Hot coding of steps 2 and 3, the physicochemical property coding of amino acids, and the Word2Vec word embedding coding.
[0088] Step 9: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network constructed in step 4. The features of the three types of peptides generated in step 8 are input into the discriminator module for discrimination, and finally immune active peptides are generated.
[0089] Verification of the effectiveness of the present invention:
[0090] For the prediction model of immune active polypeptide, the performance of the present invention was evaluated by validating a separate test set. In the test set validation, the area under the ROC curve (AUC) and the accuracy were used to quantitatively evaluate the performance of the present invention. To prevent randomness, the validation results were the average of 100 test validations, and the accuracy of the prediction model was about 93.7%. The present invention was compared with other methods, and the results are as follows: Figure 2 As shown, the present invention achieved the highest AUC value, AUC = 0.955. The verification results show that the performance of the present invention is better than other methods.
[0091] For the immune active peptide generation model, the performance of the present invention was evaluated by verifying the BLEU scores of the generated peptides and the non-immune active peptides and immune active peptides. To prevent randomness, the verification results are the average of 100 test verifications. The results are shown in Figure 4 As shown, the BLEU scores of the generated polypeptides and the immune active polypeptides are higher than the BLEU scores of the non-immune active polypeptides and the immune active polypeptides, which means that the generated polypeptides are more similar to the immune active polypeptides, and therefore the two have more similar functions. In addition, the present invention randomly selected multiple groups of generated polypeptides, each group of 1000 polypeptides, and used the constructed immune active polypeptide prediction model for prediction. After multiple verifications, the prediction model judged that 68.3% of the 1000 generated polypeptides were immune active. The verification results show that the immune active polypeptide generation model has a good effect.
[0092] Example 2
[0093] A method for predicting and generating polypeptide immunological activity by integrating polypeptide physicochemical properties, sequence characteristics and word vector embedding, comprising the following steps:
[0094] Step 1: Preprocess the data.
[0095] Step 2: Use One-Hot coding and amino acid physicochemical property coding to obtain the sequence characteristics of the peptide and vectorize the peptide sequence.
[0096] Step 3: Use the Skip-Gram model in Word2Vec to obtain the embedding vector of amino acids.
[0097] Step 4: Construct a peptide immune activity prediction network (ResNet) based on the feature descriptors to extract peptide features.
[0098] Step 5: Use the non-immune active peptide dataset to transfer the model and save the pre-trained model with better results.
[0099] Step 6: Based on the obtained matrix that integrates the above three features, use the pre-trained model saved in step 5 for training, and use the trained model for weight feature fusion prediction.
[0100] Step 7: Construct a generative model generator module network (LSTM) architecture, use high-dimensional random noise as the generator module input, and map it into a false amino acid sequence through the generator module.
[0101] Step 8: Perform feature extraction on the immune active peptide and the false peptide sequence generated in step 7, extract the One-Hot coding of steps 2 and 3, the physicochemical property coding of amino acids, and the Word2Vec word embedding coding.
[0102] Step 9: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network. The features of the two types of peptides generated in step 8 are input into the discriminator module for discrimination, and finally immune active peptides are generated.
[0103] Wherein, the step 1 is specifically as follows:
[0104] Step 1.1: Preprocess the peptide data to remove outliers and missing values of amino acids in the sequence;
[0105] Step 1.2: Store the peptide data according to the peptide activity classification and combine the immune active peptide data set.
[0106] The step 2 is specifically as follows:
[0107] Step 2.1: For the immune active peptide and non-immune active peptide datasets, use the One-Hot code generation method and the amino acid physicochemical property code generation method to generate the One-Hot code and amino acid physicochemical property code for the peptide sequence, and then concatenate the two codes into a one-dimensional feature vector. The formula is:
[0108] Y = Y.concat(seq i)i=1,2,3...length(seq)#(1)
[0109] where seq i is the One-Hot code or physicochemical property code corresponding to the ith amino acid in the polypeptide sequence, and Y is the characteristic code of this sequence.
[0110] The step 3 is specifically as follows:
[0111] Step 3.1: Use a large number of peptide sequences as a corpus for training the Word2vec model to train the Word2vec model;
[0112] Step 3.2: Use the Skip-Gram model in the trained Word2vec to obtain the embedding vectors of 20 amino acids and obtain the word embedding vectors of the peptide dataset. The Skip-Gram model is used to obtain the co-occurrence similarity of amino acids. The formula is as follows:
[0113]
[0114]
[0115] where c is the context size, a t is the first part of the t-th amino acid embedded, a t+j is the first partial embedding of the adjacent amino acids, n is the negative sampling size, a′ k is not in the same context as a t Co-occurring negative samples.
[0116] The step 4 is specifically as follows:
[0117] Step 4.1: Construct a peptide immune activity prediction network (ResNet) based on the feature descriptors. The ResNet network consists of multiple residual units, each of which can be expressed as:
[0118] y l =h(x l )+F(x l +w L )#(4)
[0119] x l+1 =f(y l )#(5)
[0120] where x l and x l+1 They represent the input and output of the lth residual unit respectively. Each residual unit generally contains a multi-layer structure. F is the residual function, which represents the learned residual, and h(x l )=x lrepresents the identity mapping, f is the ReLU activation function, and the learning feature formula from the shallow layer l to the deep layer L can be obtained from formula (4) and (5), as shown in formula (6).
[0121]
[0122] By performing reverse deduction according to the chain rule, we can derive the gradient formula of the reverse process, as shown in formula (7).
[0123]
[0124] In the formula: is the loss function in the Lth residual unit. The meaning of "1" in formula (7) is that the propagation gradient is carried out without loss, while the other residual gradients need to pass through the weighted layer and cannot be directly transmitted. Even if the value of the residual gradient becomes smaller, it will not all be -1. Therefore, the meaning of "1" is to prevent the gradient from disappearing, so that the residual neural network can be continuously accumulated and improved, reducing the difficulty of residual learning.
[0125] The step 5 is specifically as follows:
[0126] Step 5.1: Select eight types of non-immune active peptides from the data set, mainly including antimicrobial peptides, antiviral peptides, antihypertensive peptides, hemolytic peptides, etc. The number of peptides of each active type is limited to about 1000. These eight types of peptides are combined into a transfer learning data set, and the feature descriptors of the peptide information in the data set are obtained according to the above three feature extraction methods;
[0127] Step 5.2: Use the ResNet network built in step 4 to perform transfer learning on the eight-category dataset, and save the pre-trained model with better results for subsequent peptide immune activity training.
[0128] The step 6 is specifically as follows:
[0129] Step 6.1: The 522 collected immune-active peptides are used as positive samples, and a variety of non-immune-active peptides selected from the database are used as negative samples. About 1,500 negative samples are selected, and then the positive and negative samples are spliced and shuffled, and then divided into training sets and test sets in a ratio of 2:1. The training set and the test set are used to obtain feature descriptors of peptide information according to the above three feature extraction methods;
[0130] Step 6.2: Get the pre-trained model saved in step 5, train it using the training set, and save the model with better results;
[0131] Step 6.3: The model saved in step 6.2 is F. Model F can be divided into two parts. The first part is a convolutional neural network composed of residual blocks for extracting features. This module is named F1. The second part is a neural network composed of fully connected layers for obtaining the final result. This module is named F2.
[0132] Get the One-Hot code of the peptide sequence (E ot ), the physical and chemical properties of amino acids (E ph ) and Word2vec encoding (E wv ) and the overall features of the three codes (E all ) ; Input the above four codes into the F1 module to obtain four feature vectors Y with a length of 2048 ot F1, Y ph F1, Y wv F1, Y all F1, change Y ot F1, Y ph F1, Y wv F1 and Y all F1 concatenation is then passed through a convolutional layer to obtain a feature vector Y with a length of 2048 ot Y all , Y ph Y all , Y wv Y all ; Then E ot 、E ph 、E wv Input into model F to get three weight values u ot 、u ph 、u wv .
[0133] Then the concatenated feature vector Y ot Y all , Y ph Y all , Y wv Y all With weight u ot 、u ph 、u wv Adaptively multiply to obtain three eigenvectors Y ot , Y ph , Y wv , then physically add the three feature vectors to obtain a feature vector Y with a length of 2048, input Y into the feature extraction module F1, and finally output the probability value P of the polypeptide sequence having immune activity. The formula of step 6.3 is as follows:
[0134] Y i F1=F1(E i)i=ot,wv,ph#(8)
[0135] Y i Y all =concat(Y all F1, Y i F1)i=ot,wv,ph#(9)
[0136] Y i =muti(Y i Y all , Y i Y all )i=ot,wv,ph#(10)
[0137] Y=add(Y ot , Y ph , Y wv )#(11)
[0138] P=F2(Y)#(12)
[0139] The step 7 is specifically as follows:
[0140] Step 7.1: First, build the generative model generator module. The generator module is built using the LSTM network. The LSTM network consists of three parts: input gate, forget gate, and output gate. The input gate is responsible for adding useful features to the cell storage unit. This process consists of two parts. First, input the data into the tanh layer to get the new At the same time, a sigmoid layer is also required to obtain the data feature i t , and then the tanh layer to get Combined input into the cell storage unit. The formula is as follows:
[0141] i t =σ(W i [h t-1 ,x t ]+b i )#(13)
[0142]
[0143]
[0144] Among them, W i and W C represents the corresponding weight value, b i and b C represents the corresponding bias, C t Represents the status value of the current cell.
[0145] Finally, LSTM selects the feature data that can be used as output from the current active peptide data features and outputs it through the output gate. t1 ,x t ] Get the new weight value o through a sigmoid layer + , and then store the C of the cell at the current moment t After a tanh layer, and then t Multiply them together to get the formula:
[0146] o t =σ(W o [h t-1 ,x t ]+b o )# (16)
[0147] h t =o t *tanh(C t )# (17)
[0148] Among them, W o Represents the weight value of the output gate, b o represents the bias of the output gate, h t Represents the output value of the cell storage space at this moment;
[0149] Step 7.2: Use Gaussian random noise as the input of the generator module, and map it into an amino acid sequence after being processed by the generator module.
[0150] The step 8 is specifically as follows:
[0151] Step 8.1: Perform feature extraction on the immune active peptide and the false peptide sequence generated in step 7, extract the One-Hot code of step 2 and step 3, the physicochemical property code of amino acids, and the Word2Vec word embedding code.
[0152] The step 9 is specifically as follows:
[0153] Step 9.1: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network constructed in step 4. The features of the three types of peptides generated in step 8 are input into the discriminator module for discrimination, and finally immune active peptides are generated.
[0154] The embodiments of the present invention are described in detail above, but the contents are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the scope of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for predicting and generating polypeptide immune activity by integrating the physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide, characterized in that: The following steps are involved: Step 1: Preprocess the data; Step 2: Use One-Hot coding and amino acid physicochemical property coding to obtain the sequence characteristics of the peptide and vectorize the peptide sequence; Step 3: Use the Skip-Gram model in Word2Vec to obtain the embedding vector of amino acids; Step 4: Construct a peptide immune activity prediction network ResNet based on the feature descriptors to extract peptide features; Step 5: Use the non-immune active peptide dataset to transfer the model and save the pre-trained model; Step 6: According to the obtained matrix integrating the above three features, use the pre-trained model saved in step 5 for training, and use the trained model for weight feature fusion prediction; Step 7: Construct the LSTM architecture of the generative model generator module network, use high-dimensional random noise as the generator module input, and map it into a false amino acid sequence through the generator module; Step 8: Extract features from the immune active peptides and the false peptide sequences generated in step 7, extract the One-Hot codes of steps 2 and 3, the physicochemical property codes of amino acids, and Word2Ve c Word embedding encoding; Step 9: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network. The features of the two types of peptides generated in step 8 are input into the discriminator module for discrimination to generate immune active peptides.
2. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 1 includes: Step 1.1: Preprocess the peptide data to remove outliers and missing values of amino acids in the sequence; Step 1.2: Store the peptide data according to the peptide activity classification and combine the immune active peptide data set.
3. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 2 includes: Step 2.1: For the immune active peptide and non-immune active peptide datasets, use the One-Hot code generation method and the amino acid physicochemical property code generation method to generate the One-Hot code and amino acid physicochemical property code for the peptide sequence, and then concatenate the two codes into a one-dimensional feature vector. The formula is: Y=Y.concat(seq i )i=1,2,3...length(seq)#(1) where seq i is the One-Hot code or physicochemical property code corresponding to the ith amino acid in the polypeptide sequence, and Y is the characteristic code of this sequence.
4. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 3 includes: Step 3.1: Use several peptide sequences as a corpus for training the Word2vec model to train the Word2vec model; Step 3.2: Use the Skip-Gram model in the trained Word2vec to obtain the embedding vectors of several amino acids and obtain the word embedding vectors of the peptide dataset. The Skip-Gram model obtains the co-occurrence similarity of amino acids. The formula is as follows: where c is the context size, a t is the first part of the t-th amino acid embedded, a t+j is the first partial embedding of the adjacent amino acids, n is the negative sampling size, a′ k is not in the same context as a t Co-occurring negative samples.
5. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 4 includes: Step 4.1: Construct a peptide immune activity prediction network ResNet based on the feature descriptor; the ResNet network includes multiple residual units, each of which is represented as: y l =h(x l )+F(x l +W L )#(4) x l+1 =f(y l )#(5) where x l and x l+1 They represent the input and output of the lth residual unit respectively. Each residual unit contains a multi-layer structure. F is the residual function, which represents the learned residual. h(x l )=x l represents the identity mapping, f is the ReLU activation function, and the learning feature formula from the shallow layer l to the deep layer L is obtained from formula (4) and (5), as shown in formula (6): According to the chain rule, reverse deduction is performed to obtain the gradient formula of the reverse process, as shown in formula (7): In the formula: is the loss function in the Lth residual unit.
6. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 5 includes: Step 5.1: Select non-immune active peptides from the data set, the non-immune active peptides include antimicrobial peptides, antiviral peptides, antihypertensive peptides, and hemolytic peptides, and the number of peptides of each active type is limited to 800 to 1200, and the non-immune active peptides are used to form a transfer learning data set, and obtain feature descriptors of the peptide information in the data set; Step 5.2: Use the ResNet network built in step 4 to perform transfer learning on the transfer learning dataset described in step 5.1 and save the pre-trained model.
7. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 6 includes: Step 6.1: The collected immune active peptides are used as positive samples, and a variety of non-immune active peptides selected from the database are used as negative samples. 1200 to 1800 negative samples are selected, and then the positive and negative samples are spliced and shuffled, and then divided into training sets and test sets in a certain ratio, and feature descriptors of peptide information are obtained from the training sets and test sets; Step 6.2: Get the pre-trained model saved in step 5, train it using the training set, and save the model; Step 6.3: The model saved in step 6.2 is F. Model F is divided into two parts. The first part is a convolutional neural network composed of residual blocks. This module is named F1. The second part is a neural network composed of fully connected layers. This module is named F2. Get the One-Hot code of the peptide sequence ot 、The physicochemical properties of amino acids are coded ph and Word2vec encoding E wv And by E ot 、E ph 、E wv The overall feature E all ; The above code E ot 、E ph 、E wv 、E all Input into F1 module respectively to obtain four feature vectors Y of length H ot F1, Y ph F1, Y wv F1, Y all F1, change Y ot F1, Y ph F1, Y wv F1 and Y all F1 concatenation is then passed through a convolutional layer to obtain a feature vector Y of length H ot Y all , Y ph Y all , Y wv Y all ; Then E ot 、E ph 、E wv Input into model F to get three weight values u ot 、u ph 、u wv ; Then the concatenated feature vector Y ot Y all , Y ph Y all , Y wv Y all With weight u ot 、u ph 、u wv Adaptively multiply to obtain three eigenvectors Y ot , Y ph , Y wv , and then physically add the three feature vectors to obtain a feature vector Y with a length of H, input Y into the feature extraction module F1, and output the probability value P of the polypeptide sequence having immune activity; the formula of step 6.3 is as follows: Y i F1=F1(E i )i=ot,wv,ph#(8) Y i Y all =concat(Y all F1,Y i F1)i=ot,wv,ph#(9) Y i =muti(Y i Y all ,Y i Y all )i=ot,wv,ph#(10) Y=add(Y ot ,Y ph ,Y wv )#(11) P=F2(Y)#(12).
8. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 7 includes: Step 7.1: First, build a generative model generator module. The generator module is built using an LSTM network. The LSTM network consists of three parts: an input gate, a forget gate, and an output gate. The input gate adds features to the cell storage unit. Adding features to the cell storage unit includes: first inputting data into the tanh layer to obtain a new At the same time, a sigmoid layer is used to obtain the data feature i t , and then the tanh layer to get Combined with the input into the cell storage unit, the formula is as follows: i t =σ(W i [h t-1 ,x t ]+b i )#(13) Among them, W i and W C represents the corresponding weight value, b i and b C represents the corresponding bias, C t Represents the status value of the current cell; LSTM selects the feature data in the current active peptide data feature and outputs it through the output gate; it first concatenates [h t1 ,x t ] Get the new weight value o through a sigmoid layer t , and then store the C of the cell at the current moment t After a tanh layer, and then t Multiplying together, the formula is as follows: the t =σ(W o [h t-1 ,x t ]+b o )#(16) h t =o t *tanh(C t )#(17) Among them, W o Represents the weight value of the output gate, b o represents the bias of the output gate, h t Represents the output value of the cell storage space at this moment; Step 7.2: Use Gaussian random noise as the input of the generator module, and map it into an amino acid sequence after being processed by the generator module.
9. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 8 includes: Step 8.1: Perform feature extraction on the immune active peptide and the false peptide sequence generated in step 7, extract the One-Hot code of step 2 and step 3, the physicochemical property code of amino acids, and the Word2Vec word embedding code.
10. A method for predicting and generating polypeptide immune activity based on physicochemical properties, sequence characteristics and word vector embedding of a fusion polypeptide according to claim 1, characterized in that: Step 9 includes: Step 9.1: Build the network architecture of the generative model discriminator module. The discriminator module uses the ResNet network constructed in step 4. The features of the three types of peptides generated in step 8 are input into the discriminator module for discrimination to generate immune active peptides.
Citation Information
Cited By
Taste peptide design method based on deep learning
CN119943158A
Polypeptide drug toxicity prediction method and system based on artificial intelligence
CN120183537A
A collagen peptide composition and its efficacy and use in improving immunity
CN120607629B
Polypeptide target spot prediction method based on sequence and structural characteristics
CN121528292A
