Generative adversarial network-based bioactive peptide sequence generation method

Abstract

The invention provides a bioactive peptide sequence generation method based on a generative adversarial network, and belongs to the field of biological calculation. Comprising the following steps: constructing four single-label data sets according to a generation target, and carrying out corresponding preprocessing operation on the data sets; constructing a bioactive peptide generation model BAP-GAN based on the generative adversarial network, wherein the model BAP-GAN is mainly composed of two parts: a generator network and a discriminator network; based on a multi-stage learning strategy, learning characteristic information contained in a bioactive peptide sequence by using BAP-GAN, and introducing a search tool BLAST based on a local alignment algorithm in a training process to monitor characteristic learning of the model; a plurality of sequences are generated by using BAP-GAN, the sequences generated by the BAP-GAN are subjected to characteristic accuracy analysis, amino acid distribution analysis and homology analysis, and the method can be used for simultaneously learning and generating bioactive peptide sequences with various characteristics.

Description

Technical Field

[0001] The present invention relates to a method for generating bioactive peptide sequences based on a generative adversarial network, which is particularly suitable for generating a large number of bioactive peptides with low toxicity and antibacterial properties, and belongs to the field of biological computing. Background Art

[0002] Traditionally, human protein design mainly relies on various biochemical experiments, which is not only time-consuming and labor-intensive, but also has certain cost losses and safety risks. However, with the integration and development of biotechnology and artificial intelligence technology, humans have gained powerful tools for understanding and transforming proteins. The reason why artificial intelligence technology can be used to analyze biological sequences is that there is significant homogeneity between natural language and biological sequences. For example, words form sentences with specific semantics through grammatical rules, and amino acids are spliced ​​into proteins with specific structures and functions through peptide bonds. Therefore, natural language processing technology can be transferred to the protein field to understand the functions and biochemical properties of these proteins from amino acid sequences. With the continuous advancement of artificial intelligence technologies such as deep learning and natural language processing, their applications in biopharmaceuticals, disease treatment and other fields are becoming more and more extensive. Using artificial intelligence technology to design and generate bioactive peptides with specific functions has become an inevitable choice for technological development, which can not only greatly shorten the drug development cycle, but also effectively reduce the cost and risk of drug experiments.

[0003] In recent years, generative models based on natural language processing technology have achieved remarkable results in protein sequence generation. For example, Greener et al. generated protein sequences based on the conditional variational autoencoder (CVAE) model with the desired characteristics as conditions, and added potential copper and calcium binding sites to non-metal binding proteins to generate sequences with new protein topological structures. Repecka et al. designed a generative adversarial network ProteinGAN, which uses malate dehydrogenase as the synthesis target, learns the evolutionary relationship of protein sequences from the multidimensional amino acid sequence space through the self-attention mechanism, and creates new highly diverse sequence variants with natural physical properties, demonstrating the potential of generative adversarial networks to quickly generate highly diverse functional proteins. Bitard-Feildel et al. proposed an adversarial autoencoder (AAE) based on autoencoders and adversarial mechanisms, and used it to generate and explore new protein sequences and their functions. Studies on three protein sequence families with known functions show that AAE can extract and encode biological features from the latent space, and then generate sequences with intermediate features, thus providing original solutions for protein sequence design and functional exploration.

[0004] At present, the mainstream generative models include generative adversarial networks, variational autoencoders, and diffusion models. By designing these generative models, problems related to protein sequence generation can be solved and the process of protein design and research can be accelerated. However, the existing generative models are insufficient in learning the characteristics of bioactive peptide sequences and cannot simultaneously learn and generate bioactive peptide sequences with multiple characteristics. In addition, in the generator and discriminator of the traditional generative adversarial network, only ordinary neural network layers such as linear layers and convolutional layers are usually used. The linear layer limits the model's ability to selectively focus on specific parts of the input, and the convolutional layer may cause the model to learn information in the sequence only in a fixed local area, and cannot capture the dependencies between amino acids that are far apart. In order to solve the above problems, the present invention proposes a method for generating bioactive peptide sequences based on a generative adversarial network. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that traditional models cannot generate bioactive peptide sequences with multiple characteristics and have insufficient learning ability, and propose a method for generating bioactive peptide sequences based on a generative adversarial network.

[0006] The method for generating bioactive peptide sequences based on a generative adversarial network proposed in the present invention is based on the BAP-GAN model. The bioactive peptide generation model BAP-GAN based on the generative adversarial network learns the characteristics of different data sets and trains a model that can generate multi-label sequences through a multi-label generation strategy. In addition, by introducing a multi-head self-attention mechanism, the model can comprehensively learn the correlation between amino acids in the sequence, thereby better understanding the meaning of the sequence and revealing its functional relationship.

[0007] The objectives of the present invention are achieved through the following technical solutions.

[0008] The present invention proposes a method for generating bioactive peptide sequences based on a generative adversarial network, comprising the following steps:

[0009] Step 1: According to the generation target of the present invention, four single-label data sets are constructed: an antibacterial positive data set, an antibacterial negative data set, a toxicity positive data set, and a toxicity negative data set, and the data sets are preprocessed accordingly, specifically:

[0010] Step 1.1: Map the sequence into a vector using the dictionary position information of the amino acid composition.

[0011] Protein sequences are usually composed of 20 standard amino acids, and may also contain some non-standard amino acids. The physicochemical properties of these non-standard amino acids are not significant compared to standard amino acids, so sequences containing non-standard amino acids are screened out. Further, a 20-letter dictionary is used to save the 20 standard amino acids, and the letter "X" is used to save the space symbol, and their position information in the dictionary is used as the basis for the sequence mapping vector.

[0012] Step 1.2: Make the length of each sequence vector equal to 30 by padding or truncation.

[0013] When the sequence length is greater than 30, the first 30 amino acid characters are truncated as the current sequence; when the sequence length is less than 30, the dictionary characters corresponding to "X" are used to fill in until the sequence length reaches 30.

[0014] Step 2: Construct a bioactive peptide generation model BAP-GAN based on a generative adversarial network. The BAP-GAN proposed in the present invention mainly consists of two parts: a generator network and a discriminator network. The generator network and the discriminator network also include a data processing module responsible for data generation and conversion. The data processing module in the generator network can generate noise data and convert it into a corresponding sequence vector. The data processing module in the discriminator network can convert the input sequence information and label information into a corresponding sequence vector, specifically:

[0015] Step 2.1: Build the generator network and the discriminator network.

[0016] The generator network designed by the present invention includes several Block basic units. In each Block basic unit, the vector will first perform a ConvTranspose2d inverse convolution operation to upsample the data, then perform a BatchNorm2d operation to normalize the data to prevent the network from being unstable due to excessive data, and finally activate it through the ReLU function. The data processing module in the generator network will generate a 90*1*1 Gaussian noise vector and a 1*1 label vector based on the input information. The label vector is converted into a 20*1 vector through the dictionary, and then converted into a 20*1*1 vector through the dimensionality increase operation. In the first dimension, the two vectors are combined into a 110*1*1 vector and input into the generator network, which is converted into a vector of dimension 1*1*110. After passing through the first Block basic unit, a 4*4*256 vector is obtained. After passing through the second Block basic unit, a 8*8*128 vector is obtained. After passing through the third Block basic unit, a 16*16*64 vector is obtained. After passing through the fourth Block basic unit, a 21*21*32 vector is obtained. After passing through the Attention self-attention module, the vector dimension does not change. After passing through the fifth Block basic unit, a 30*21*1 vector is finally obtained, and finally it is output after being activated by the Tanh function.

[0017] The discriminator network is essentially a classifier, whose task is to classify the input vector. The output of 0 represents that the input vector is a vector generated by the generator, and the output of 1 represents that the input vector is a vector converted from a real sample. The discriminator network designed by the present invention includes several Block basic units. In each Block basic unit, the vector will first undergo a Conv2d convolution operation to upsample the data, then undergo an InstanceNorm2d operation to normalize the data, and finally be activated by the LeakyReLU function. The discriminator network first inputs a 30*21 sequence vector and a 1*1 label vector. The label vector is converted into a 20*1 vector through a dictionary, and then converted into a 20*1*1 vector through a dimensionality increase operation, and the two vectors are merged into a 650*1*1 vector in the first dimension. The vector is then input into a fully connected unit to obtain a 630*1*1 vector. The dimension of the vector is converted to 30*21 and then input into a convolutional layer Conv2d to obtain a 16*21*21 vector. After passing through a LeakyReLU activation function and the first Block basic unit, a 32*16*16 vector is obtained. Then, after passing through the Attention self-attention module, the vector dimension remains unchanged. After passing through the second Block basic unit, a 64*8*8 vector is obtained. After passing through the third Block basic unit, a 128*4*4 vector is obtained. After passing through a convolutional layer Conv2d, a 1*1*1 vector is output.

[0018] When the data processing module is responsible for data generation, it will generate a 90*1*1 Gaussian noise vector. The elements in this vector are a set of random numbers that obey a certain distribution. The random number distribution used in this article is a standard normal distribution, that is, the mean of this set of data is 0 and the variance is 1. When the data processing module is responsible for data conversion, it will receive a numpy format file containing several arrays containing sequence information. The module converts the file into a vector of dimension N*30*21 through one-hot encoding, where N is the number of sequences contained in the file, 30 is the length of each sequence, and 21 represents the positions of the twenty standard amino acids and non-standard amino acids "X".

[0019] Step 2.2: Define the loss function of BAP-GAN. The loss function used by BAP-GAN contains two parts, one is the loss function of the discriminator, and the other is the loss function of the generator.

[0020] The loss function of the discriminator is designed based on WGAN-GP, and uses Wasserstein distance as a part of it. Regardless of whether the real sample distribution overlaps with the generated sample distribution, Wasserstein distance can reflect the distance between the two distributions. The calculation formula of Wasserstein distance between real samples and generated samples is expressed as formula (1):

[0021]

[0022] Among them, E[] represents the expectation; D() represents the score of the discriminator on the data; P g represents the distribution of the generated sequence; P r represents the distribution of the real sequence; represents the sample generated by the generator; x represents the real sample.

[0023] When dealing with Lipschitz continuity constraints, if the weight pruning strategy of WGAN is directly adopted, the weights learned by the discriminator may be concentrated at the maximum and minimum values, resulting in gradient explosion or model non-convergence. Therefore, the discriminator loss function adds an additional gradient penalty term based on the Wasserstein distance. The formula of the gradient penalty term is expressed as formula (2):

[0024]

[0025] Among them, ‖·‖2 represents the gradient norm of the discriminator at the interpolation point; represents the gradient operator; λ is the penalty term coefficient, which is used to control the influence of the gradient penalty term. The penalty term coefficient of the present invention is set to 10; represents the sampling distribution of the gradient penalty term; Represents the linear interpolation between the real sample data x and the generated data G(z).

[0026] In summary, the discriminator loss consists of two parts, one is the Wasserstein distance between the real sample and the generated sample, and the other is the penalty term of the discriminator. The discriminator loss of BAP-GAN is defined as formula (3):

[0027]

[0028] The generator loss of BAP-GAN is defined as the negative value of the mean of the generated samples classified as real samples by the discriminator, reflecting the difference between the generated samples and the real samples. The generator loss calculation formula can be expressed as formula (4):

[0029]

[0030] Step 3: Based on the multi-stage learning strategy, BAP-GAN is used to learn the characteristic information contained in the bioactive peptide sequence, and a search tool based on the local alignment algorithm (Basic Local Alignment Search Tool, BLAST) is introduced during the training process to monitor the feature learning of the model. The multi-stage learning strategy proposed in the present invention is divided into three stages: the first stage is the single-label characteristic learning stage, in which a single-label BAP-GAN is trained through the information in the antibacterial data and toxicity data, and a part of the single-label sequence is generated; the second stage is the single-label data integration stage, in which the single-label sequences generated in the first stage are screened and integrated using the characteristic classifiers of antibacterial and toxicity to construct a multi-label data set; the third stage is the multi-label characteristic learning stage, in which a multi-label BAP-GAN is trained through the multi-label data set constructed in the second stage, and finally multi-label data with both antibacterial positive and toxicity positive labels are generated, as follows:

[0031] Step 3.1: Single-label feature learning phase.

[0032] A single-label BAP-GAN is constructed based on the Conditional Generative Adversarial Network (CGAN), and single-label BAP-GAN is used to generate antibacterial positive and toxicity positive data. It is worth noting that the training set in this stage not only contains positive samples of antibacterial and toxicity data, but also adds negative samples and corresponds to separate labels, which will help to better learn the relevant information of the two characteristics, thereby improving the generation accuracy of the model. After the model training is completed, each single-label data is generated, and each label generates several sequences as input data for the single-label data integration stage. The objective function of the single-label feature learning stage can be expressed as formula (5):

[0033] m G inm D aX s ing le (D,G)

[0034] =E x～Pdata(x) [logD(x|y single-label )]+E z～Pz(z) [log(1-D(G(z|y single-label )))] (5)

[0035] This formula can reflect the training process of the model in a single label dataset, where x represents the real sample; z represents the noise generated by the generator; P data represents the distribution of real data; P z represents the distribution of generated noise; y single-labelRepresents the label information corresponding to the single-label data, which serves as a condition to control the direction of model generation.

[0036] At the same time, BLAST was introduced in the single-label BAP-GAN training process to evaluate the consistency of the generated sequences to determine whether the model is overfitting.

[0037] First, the complete training set was merged and CD-HIT was used to remove redundancy with default parameters to obtain a non-redundant data set, which was used as the target database for BLAST search of the generated sequence. Secondly, the sequences in the non-redundant data set were randomly shuffled and divided according to a certain ratio, one part was used as the target database of the training set, and the other part was used as the query sequence for BLAST search. Then, the BLAST results output by each sequence were grouped, the highest consistency of each sequence in the target database was retained, and the median of the consistency was calculated as the consistency between natural sequences. After each round of iteration, a certain number of sequences of each tag were generated in turn, and BLAST was used to perform multiple sequence alignment using the generated sequences as query sequences, and the consistency in each round of alignment results was recorded. After the training, the BLAST search results were grouped in the same way, the highest consistency of each sequence was retained, and the consistency of all sequences was averaged, and the average value was used as the consistency of the sequence generated by the model after the round of training. Finally, the consistency results of the generated sequences after each round of training were plotted to determine the consistency change trend of the generated sequences.

[0038] Step 3.2: Single-label data integration phase.

[0039] The present invention selects ToxIBTL as a toxicity classification model and CAMP as an antibacterial classification model, and classifies all data including the training samples and generated samples in the first stage through these two characteristic classification models, and screens out data with both positive toxicity and positive antibacterial properties. These two classification models are existing high-reliability classification models that can directly output classification results.

[0040] Two multi-label datasets were constructed from the screened data: the toxicity & antibacterial positive dataset and the toxicity & antibacterial negative dataset. It should be noted that toxicity & antibacterial negative refers to not satisfying toxicity and antibacterial properties at the same time, so there may be a situation where only one characteristic is satisfied but not the other. These datasets eventually constitute the complete multi-label BAP-GAN training set.

[0041] Step 3.3: Multi-label feature learning phase.

[0042] The model is trained based on the multi-label BAP-GAN training set to obtain a multi-label BAP-GAN model. At this time, the multi-label BAP-GAN model can generate corresponding multi-label data after specifying the corresponding label, and can also generate corresponding single-label data when other labels are specified for the model, thereby realizing the unification of single-label and multi-label generation models. The objective function of the multi-label data learning stage is expressed as formula (6):

[0043] m G inm D aX mu l t i(D,G)

[0044] =E x～Pdata(x) [logD(x|y multi-label )]+E z～Pz(z) [log(1-D(G(z|y multi-label )))] (6)

[0045] This formula can reflect the training process of the model in a multi-label dataset, where x represents the real sample; z represents the noise generated by the generator; P data represents the distribution of real data; P z represents the distribution of generated noise; y multi-label Represents the label information corresponding to the multi-label data, which serves as a condition to control the direction of model generation.

[0046] At the same time, BLAST is introduced in the multi-label BAP-GAN training process to evaluate the consistency of the generated sequences to determine whether the model is overfitting. The introduction strategy is consistent with step 3.1 of the invention.

[0047] Step 4: Use BAP-GAN to generate several sequences, and perform characteristic accuracy analysis, amino acid distribution analysis, and homology analysis on the sequences generated by BAP-GAN.

[0048] Step 4.1: The present invention uses ACC_single and ACC_multi as evaluation indicators of the model to evaluate the characteristic accuracy. The specific formula is shown in formula (7):

[0049]

[0050] Among them, TP is the number of true positive samples, FN is the number of false positive samples, TN is the number of true negative samples, and FP is the number of false negative samples. True positive samples refer to the generated positive samples that are indeed positive after testing, true negative samples refer to the generated negative samples that are indeed negative after testing, false positive samples refer to the generated positive samples that are negative after testing, and false negative samples refer to the generated negative samples that are positive after testing.

[0051] Step 4.2: Taking the antibacterial positive as an example, the generated sequence is compared with the natural sequence. The multi-label BAP-GAN model is used to generate several antibacterial positive single-label sequences, and the antibacterial positive sample set in the initial data set is selected to count the amino acid frequency information.

[0052] Step 4.3: The present invention uses an output parameter "E value" in the BLAST search results to evaluate the homology of the final generated sequence. The E value refers to the expected value of the actual sequence comparison score, which indicates the statistical significance of the match between the query sequence and the hit sequence in a database of a specific size. If the E value is large, it means that the probability that the query sequence and the hit sequence have no homology is high. When the UniProt non-redundant database containing about 200 million sequences is used as the BLAST search target library, the 10 -3 As the threshold for determining homology. -3 , indicating that the query sequence has no homology with the hit sequence; when the E value is less than 10 -3 , indicating that the query sequence has homology with the hit sequence. Since the training dataset of BAP-GAN contains about 10,000 sequences after CD-HIT redundancy removal with default parameters, it is about 10 smaller than the Uniprot database. 4 times, so the threshold should be reduced by 10 4 times, that is, 10 -7 is the E value threshold for judging homology.

[0053] The multi-label BAP-GAN model was used to generate several sequences for each label, and the BLAST search was performed using the de-redundant training set as the target database. The E value with the highest consistency in each query sequence was recorded, and the E value distribution in the generated sequence was statistically analyzed. If the E value of the generated sequence was basically greater than 10 -3 , it means that the generated sequence has low homology and good biological significance.

[0054] Beneficial Effects

[0055] 1. The present invention discloses a method for generating bioactive peptide sequences based on a generative adversarial network, which greatly improves the possibility of BAP-GAN generating multi-label data through a multi-stage generation strategy. The present invention uses three stages: a single-label feature learning stage, a single-label data integration stage, and a multi-label feature learning stage, which avoids the problem of a small number of multi-label samples and takes the common characteristics of antibacterial and toxicity as an independent characteristic, so that the model can learn the relevant information of the two characteristics at the same time.

[0056] 2. The present invention discloses a method for generating bioactive peptide sequences based on a generative adversarial network, which selects antibacterial properties and toxicity as target characteristics, and can generate a large number of bioactive peptide sequences that meet the requirements in a short period of time. The accuracy of generating samples with both antibacterial and low toxicity reached 85.3%, the accuracy of generating antibacterial samples alone reached a maximum of 86.9%, and the accuracy of generating toxic samples alone reached a maximum of 92.7%. The above results demonstrate the effectiveness of BAP-GAN based on the generative adversarial network architecture in the generation of bioactive peptide sequences. This method not only speeds up the progress of bioactive peptide-related research, but also reduces the cost loss and safety risks of traditional biochemical experiments, which is of great significance for the development of bioactive peptides.

[0057] 3. The present invention discloses a method for generating bioactive peptide sequences based on a generative adversarial network, which introduces a self-attention mechanism module, improves the model's ability to capture information about long-distance amino acid pairs in the sequence, comprehensively learns the correlation between amino acids in the sequence, and enhances the ability of BAP-GAN to learn characteristic information.

[0058] 4. The present invention discloses a method for generating bioactive peptide sequences based on a generative adversarial network, which effectively avoids the overfitting problem of the model by continuously generating sequences and performing BLAST searches during model training. After each round of training, several sequences are generated and BLAST analyzed with natural sequences in the training set to determine whether the model is overfitting, thereby avoiding saving and loading overfitted BAP-GAN. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a technical architecture diagram of a method for generating bioactive peptide sequences based on a generative adversarial network according to the present invention.

[0060] Figure 2 It is a detailed structural diagram of the generator network and the discriminator network of the generative model BAP-GAN designed by the present invention.

[0061] Figure 3 This is a diagram of the verification results of the generation sequence of the designed generation model BAP-GAN in the present invention. DETAILED DESCRIPTION

[0062] In order to better illustrate the purpose and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0063] Example

[0064] like Figure 1 As shown, the present invention proposes a method for generating bioactive peptide sequences based on a generative adversarial network, and the specific implementation steps are as follows:

[0065] This experiment was conducted on a computer and a server. The specific configuration of the computer is: Intel(R) i7-14650HX, CPU 2.20GHz, memory 16GB, operating system is Windows 11, 64-bit; the specific configuration of the server is: Intel(R) Core(TM) i7-10700K, CPU 3.80GHz, GPU NVIDIA GeForce RTX 3090, operating system is Ubuntu 20.04.4. The deep learning framework is Pytorch 2.1.2, and the Python version is Python 3.8.18.

[0066] Step 1: Construct four single-label datasets according to the model generation target: antibacterial positive dataset, antibacterial negative dataset, toxicity positive dataset, and toxicity negative dataset.

[0067] The antibacterial data set was constructed by combining four public databases: DBAASP, AVPdb, SATPdb, and UniProt. Sequences annotated with antibacterial function in the SATPdb database or peptide sequences that showed a minimum inhibitory concentration (MIC) of less than 25 μg / mL against any target species in the DBAASP database were classified as antibacterial positive samples, and peptide sequences with an average activity of more than 100 μg / mL against any target species in the DBAASP database and peptide sequences annotated as "AMP", "membrane", "toxic", "secretory", "defensive", "antibiotic", "anticancer", "antiviral", and "antifungal" in the UniProt database were classified as antibacterial negative samples. Referring to the method of constructing a toxicity dataset by Wei et al. (Wei L, Ye X, Xue Y, et al. ATSE: a peptide toxicity predictor by exploiting structural and evolutionary information based on graph neural network and attention mechanism [J]. Brief Bioinform, 2021, 22 (5).), the keywords "NOTKW-0800" and "NOT KW-0020" were used to search for manually annotated and checked non-toxic peptides in the SwissProt database as positive samples; experimentally verified toxic peptides were collected from three public databases ConoServer, ArachnoServer and SwissProt as negative samples. After removing some repeated peptides, the dataset consists of 3864 samples, with 1932 positive samples and 1932 negative samples, of which the positive samples are manually annotated non-toxic peptide sequences, and the negative samples are experimentally verified toxic peptide sequences. After removing the overly long sequences and invalid sequences in the dataset, the toxicity positive dataset and toxicity negative dataset are obtained. The constructed antibacterial data set and toxicity data set were preprocessed, and the sequences were mapped into vectors through the position information of the dictionary composed of standard amino acids. Then, the length of each sequence vector was made equal to 30 by padding or truncation.

[0068] The data set finally used in this example is shown in Table 1:

[0069] Table 1 Experimental data for generating bioactive peptide sequences

[0070]

[0071] Step 2: Construct a bioactive peptide generation model BAP-GAN based on generative adversarial network.

[0072] The BAP-GAN proposed in this invention mainly consists of two parts: a generator network and a discriminator network. The architecture of the generator network and the discriminator network is as follows: Figure 2 As shown in the figure, the generator network and the discriminator network also include a data processing module responsible for data generation and conversion. The data processing module of the generator network can generate noise data and convert it into a corresponding sequence vector. The data processing module of the discriminator network can convert the input sequence information and label information into a corresponding sequence vector.

[0073] The specific implementation process includes the construction of the generator network and the discriminator network and the definition of the loss function, which is consistent with step 2 in the invention content.

[0074] Step 3: Based on a multi-stage learning strategy, BAP-GAN is used to learn the characteristic information contained in the bioactive peptide sequence, and a search tool based on the local alignment algorithm (Basic Local Alignment Search Tool, BLAST) is introduced during the training process to monitor the feature learning of the model.

[0075] The multi-stage learning strategy proposed in the present invention is divided into three stages: the first stage is the single-label characteristic learning stage, in which a single-label BAP-GAN is trained through the information in the antibacterial data and toxicity data, and a part of the single-label sequence is generated; the second stage is the single-label data integration stage, in which the single-label sequences generated in the first stage are screened and integrated using the characteristic classifiers of antibacterial and toxicity to construct a multi-label data set; the third stage is the multi-label characteristic learning stage, in which a multi-label BAP-GAN is trained through the multi-label data set constructed in the second stage, and finally multi-label data with both antibacterial positive and toxicity positive labels are generated. The details are as follows:

[0076] Step 3.1: Single-label feature learning phase.

[0077] A single-label BAP-GAN was constructed based on the Conditional Generative Adversarial Network (CGAN), and single-label BAP-GAN was used to generate antibacterial positive and toxicity positive data. It is worth noting that the training set at this stage not only contains positive samples of antibacterial and toxicity data, but also adds negative samples and corresponds to separate labels, which will help to better learn the relevant information of the two characteristics, thereby improving the generation accuracy of the model. After the model training is completed, each single-label data is generated, and 3,000 sequences are generated for each label as input data for the single-label data integration stage.

[0078] At the same time, BLAST was introduced in the single-label BAP-GAN training process to evaluate the consistency of the generated sequences to determine whether the model is overfitting.

[0079] First, the complete training set was merged to obtain a redundant data set containing 15,835 sequences. CD-HIT was used to remove redundancy with default parameters to obtain a non-redundant data set containing 9,180 sequences, which was used as the target database for BLAST search of generated sequences. Secondly, the sequences in the non-redundant data set were randomly shuffled and divided at a ratio of 1:9. 918 sequences were used as the target database of the training set, and the remaining 8,262 sequences were used as query sequences for BLAST search. Then, the BLAST results output by each sequence were grouped, the highest consistency of each sequence in the target database was retained, and the median consistency of 64.007% was calculated, which was the consistency between natural sequences. After each round of iteration, 64 sequences of each tag were generated in turn, and BLAST was used to perform multiple sequence alignment using the generated sequences as query sequences, and the consistency in each round of alignment results was recorded. After the training, the BLAST search results were grouped in the same way, the highest consistency of each sequence was retained, and the consistency of all sequences was averaged. The average value was used as the consistency of the sequence generated by the model after this round of training. Finally, the consistency results of the generated sequences after each round of training are plotted to determine the consistency change trend of the generated sequences.

[0080] Step 3.2: Single-label data integration stage: The specific implementation process is consistent with step 3.2 in the invention content.

[0081] Step 3.3: Multi-label feature learning phase.

[0082] The model is trained based on the multi-label BAP-GAN training set to obtain a multi-label BAP-GAN model. At this time, the multi-label BAP-GAN model can generate corresponding multi-label data after specifying the corresponding label, and can also generate corresponding single-label data when other labels are specified for the model, thereby realizing the unification of single-label and multi-label generation models.

[0083] At the same time, BLAST is introduced in the multi-label BAP-GAN training process to evaluate the consistency of the generated sequences to determine whether the model is overfitting. The introduction strategy is consistent with step 3.1 of the specific implementation method.

[0084] Step 4: Perform feature accuracy analysis, amino acid distribution analysis, and homology analysis on the multi-label data generated by BAP-GAN.

[0085] Step 4.1: Evaluate the feature accuracy using the indicators mentioned in step 4.1 in the invention content.

[0086] After the first stage, the single-label BAP-GAN model saved after the 60th iteration was selected to generate peptide sequences. 3000 sequences were generated for each label. Table 2 shows the characteristic accuracy of these samples.

[0087] Table 2 Single label model feature accuracy

[0088]

[0089] As can be seen from Table 2, the toxicity accuracy in the generated sequences reached 92.7%, and the antibacterial accuracy reached 84.3%, indicating that the accuracy of the features generated by the single-label BAP-GAN is high.

[0090] After the third stage, the multi-label BAP-GAN model saved after the 80th iteration was selected to generate peptide sequences. 3000 sequences were generated for each label. Table 3 shows the characteristic accuracy of these samples.

[0091] Table 3. Accuracy of multi-label model features

[0092]

[0093] As shown in Table 3, the accuracy of toxicity in the generated sequence reached 91.2%, the accuracy of antibacterial property reached 86.9%, and the accuracy of toxicity and antibacterial property reached 85.3%, indicating that the accuracy of the features generated by multi-label BAP-GAN is high.

[0094] Step 4.2: Take the antibacterial positive as an example to compare the generated sequence with the natural sequence. Use the multi-label BAP-GAN model to generate 3000 antibacterial positive single-label sequences, and select the antibacterial positive sample set in the initial data set to count their amino acid frequency information.

[0095] Amino acid frequency information such as Figure 3 (a) and Figure 3 (b) is shown. Figure 3 (a) It can be seen that the generated antibacterial positive sequence is highly similar to the original antibacterial positive sequence in terms of amino acid distribution, which shows that BAP-GAN has a good learning ability for the distribution of data sets at the amino acid level. In addition, the present invention further calculates the relative difference between the generated sequence and the natural sequence at the amino acid level. Figure 3 (b) It can be seen that the frequencies of alanine "A", serine "S", lysine "K", leucine "L", and arginine "R" in the generated sequence are relatively high, while the frequencies of glutamic acid "E", aspartic acid "D", cysteine ​​"C", tyrosine "Y", and isoleucine "I" in the natural sequence are higher. Overall, the frequency difference of most amino acids does not exceed 1%, indicating that the generated antibacterial positive data and the actual antibacterial positive data have a strong correlation in amino acid distribution.

[0096] Step 4.3: The evaluation method of the homology of the final generated sequence is consistent with step 4.3 in the invention content. The E value is the expected value of the actual sequence comparison score, which indicates the statistical significance of the match between the query sequence and the hit sequence in a database of a specific size. If the E value is large, it means that there is a high probability that the query sequence and the hit sequence have no homology. When the UniProt non-redundant database containing about 200 million sequences is used as the BLAST search target library, the 10 -3 The threshold for inferring homology is when the E value is less than 10 -3 , indicating that the query sequence has no homology with the hit sequence. When the E value is greater than 10 -3 , indicating that the query sequence has homology with the hit sequence. Since the training dataset of BAP-GAN contains about 10,000 sequences after CD-HIT redundancy removal with default parameters, it is about 10 smaller than the Uniprot database. 4 times, so the threshold should be reduced by 10 4 times, that is, 10 -7 is the E value threshold for judging homology.

[0097] The multi-label BAP-GAN model was used to generate 3000 sequences for each label, and the BLAST search was performed using the de-redundant training set as the target database. For each query sequence, the E value with the highest consistency was recorded, and the distribution of the E values ​​in the generated sequences was statistically analyzed, as shown in Table 4.

[0098] Table 4 E-value distribution of generated sequences

[0099]

[0100] As shown in Table 4, observing the E-value distribution of the generated sequences, it can be found that the E-values ​​of most sequences are between 1 and 10, and the E-values ​​of very few sequences are less than or equal to 0.001, indicating that the probability of homology between the generated sequences and the original sequences is very low.

[0101] Final note: The above specific implementation methods are intended to better illustrate the concept of the present invention and are by no means a limitation of the present invention. Any modifications, equivalent substitutions or improvements made based on the essential content of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating bioactive peptide sequences based on generative adversarial networks, characterized in that: The following steps are involved: Step 1: Construct four single-label datasets according to the generation target: antibacterial positive dataset, antibacterial negative dataset, toxicity positive dataset and toxicity negative dataset, and perform corresponding preprocessing operations on the datasets; Step 2: Construct a bioactive peptide generation model BAP-GAN based on generative adversarial network. The model BAP-GAN mainly consists of two parts: a generator network and a discriminator network. The generator network and the discriminator network also include a data processing module responsible for data generation and conversion. The data processing module in the generator network can generate noise data and convert it into a corresponding sequence vector. The data processing module in the discriminator network can convert the input sequence information and label information into a corresponding sequence vector. Step 3: Based on the multi-stage learning strategy, BAP-GAN is used to learn the characteristic information contained in the bioactive peptide sequence, and the search tool BLAST based on the local alignment algorithm is introduced during the training process to monitor the feature learning of the model. The multi-stage learning strategy is divided into three stages: the first stage is the single-label characteristic learning stage, in which the single-label BAP-GAN is trained through the information in the antibacterial data and toxicity data, and a part of the single-label sequence is generated; the second stage is the single-label data integration stage, in which the single-label sequences generated in the first stage are screened and integrated using the characteristic classifiers of antibacterial and toxicity to construct a multi-label data set; the third stage is the multi-label characteristic learning stage, in which the multi-label BAP-GAN is trained through the multi-label data set constructed in the second stage, and finally multi-label data with both antibacterial positive and toxicity positive labels are generated; Step 4: Use BAP-GAN to generate several sequences, and perform characteristic accuracy analysis, amino acid distribution analysis, and homology analysis on the sequences generated by BAP-GAN.

2. A method for generating bioactive peptide sequences based on generative adversarial networks according to claim 1, characterized in that: Step 1 is as follows: Step 1.1: Map the sequence into a vector using the dictionary position information of the amino acid composition; Protein sequences are usually composed of 20 standard amino acids. In addition, they may contain some non-standard amino acids. The physical and chemical properties of these non-standard amino acids are not significant compared with those of standard amino acids. Therefore, sequences containing non-standard amino acids are screened out. Furthermore, a 20-letter dictionary is used to save the 20 standard amino acids, and the letter "X" is used to save the space symbol. Their position information in the dictionary is used as the basis for the sequence mapping vector. Step 1.2: Make the length of each sequence vector equal to 30 by padding or truncating; When the sequence length is greater than 30, the first 30 amino acid characters are taken as the sequence; when the sequence length is less than 30, the dictionary characters corresponding to "X" are used to fill in until the sequence length reaches 30.

3. A method for generating bioactive peptide sequences based on generative adversarial networks according to claim 1, characterized in that: Step 2 is as follows: Step 2.1: Build the generator network and the discriminator network The generator network contains several Block basic units. In each Block basic unit, the vector will first undergo a ConvTranspose2d inverse convolution operation to upsample the data, then undergo a BatchNorm2d operation to normalize the data to prevent the network from being unstable due to excessive data size, and finally be activated by the ReLU function. The data processing module in the generator network will generate a 90*1*1 Gaussian noise vector and a 1*1 label vector based on the input information, convert the label vector into a 20*1 vector through the dictionary, and convert it into a 20*1*1 vector through the dimensionality increase operation. In the first dimension, the two vectors are transformed into Combined into a 110*1*1 vector and input into the generator network, it is converted into a vector with a dimension of 1*1*110. After passing through the first Block basic unit, a 4*4*256 vector is obtained. After passing through the second Block basic unit, a 8*8*128 vector is obtained. After passing through the third Block basic unit, a 16*16*64 vector is obtained. After passing through the fourth Block basic unit, a 21*21*32 vector is obtained. After passing through the Attention self-attention module, the vector dimension does not change. After passing through the fifth Block basic unit, a 30*21*1 vector is finally obtained. Finally, it is output after being activated by the Tanh function. The discriminator network is essentially a classifier, whose task is to classify the input vector. The output of 0 represents that the input vector is the vector generated by the generator, and the output of 1 represents that the input vector is the vector converted from the real sample. The discriminator network designed by the present invention includes several Block basic units. In each Block basic unit, the vector will first undergo a Conv2d convolution operation to upsample the data, and then undergo an InstanceNorm2d operation to normalize the data, and finally be activated by the LeakyReLU function. The discriminator network first inputs a 30*21 sequence vector and a 1*1 label vector, converts the label vector into a 20*1 vector through a dictionary, and converts it into a 20*1*1 vector through a dimensionality increase operation. In the first dimension, the two vectors are merged into a 650*1*1 vector, and then the vector is input into a fully connected unit to obtain a 630*1*1 vector. The dimension of the vector is converted to 30*21 and then input into a convolutional layer Conv2d to obtain a 16*21*21 vector. After passing through a LeakyReLU activation function, it passes through the first Block basic unit to obtain a 32*16*16 vector. Then it passes through the Attention self-attention module, and the vector dimension remains unchanged. After passing through the second Block basic unit, a 64*8*8 vector is obtained. After passing through the third Block basic unit, a 128*4*4 vector is obtained. After passing through a convolutional layer Conv2d, a 1*1*1 vector is output; When the data processing module is responsible for data generation, a 90*1*1 Gaussian noise vector is generated. The elements in the vector are a group of random numbers that obey a certain distribution. The random number distribution used in the present invention is a standard normal distribution, that is, the mean of the group of data is 0 and the variance is 1. When the data processing module is responsible for data conversion, a numpy format file is received. The file contains several arrays containing sequence information. The module converts the file into a vector with a dimension of N*30*21 through one-hot encoding, where N is the number of sequences contained in the file, 30 is the length of each sequence, and 21 represents the positions of twenty standard amino acids and non-standard amino acids "X"; Step 2.2: Define the loss function of BAP-GAN. The loss function used by BAP-GAN contains two parts, one is the loss function of the discriminator, and the other is the loss function of the generator.

4. A method for generating bioactive peptide sequences based on generative adversarial networks as claimed in claim 3, characterized in that: Step 2.2 is as follows: The loss function of the discriminator is designed based on WGAN-GP, using Wasserstein distance as part of it. Regardless of whether the real sample distribution overlaps with the generated sample distribution, Wasserstein distance can reflect the distance between the two distributions. The calculation formula of Wasserstein distance between real samples and generated samples is expressed as: Among them, E[] represents the expectation, D() represents the score of the discriminator on the data, and P g represents the distribution of generated sequences, P r represents the distribution of the real sequence, represents the sample generated by the generator, and x represents the real sample; When dealing with the Lipschitz continuity constraint, if the WGAN weight pruning strategy is directly adopted, the weights learned by the discriminator may be concentrated at the maximum and minimum values, resulting in gradient explosion or model non-convergence. Therefore, the discriminator loss function adds an additional gradient penalty term on the basis of the Wasserstein distance. The formula of the gradient penalty term is expressed as: Among them, ‖·‖2 represents the gradient norm of the discriminator at the interpolation point, represents the gradient operator, λ is the penalty term coefficient, which is used to control the influence of the gradient penalty term. represents the sampling distribution of the gradient penalty term, Represents the linear interpolation between the real sample data x and the generated data G(z); In summary, the discriminator loss consists of two parts, one is the Wasserstein distance between the real sample and the generated sample, and the other is the penalty term of the discriminator. The discriminator loss of BAP-GAN is defined as: The generator loss of BAP-GAN is defined as the negative value of the mean of the generated samples classified as real samples by the discriminator, reflecting the difference between the generated samples and the real samples. The generator loss calculation formula can be expressed as:

5. The method for generating bioactive peptide sequences based on generative adversarial networks according to claim 1, characterized in that: Step 3 is as follows: Step 3.1: Single-label feature learning phase A single-label BAP-GAN is constructed based on the Conditional Generative Adversarial Network (CGAN), and single-label BAP-GAN is used to generate antibacterial positive and toxicity positive data. It is worth noting that the training set in this stage not only contains positive samples of antibacterial and toxicity data, but also adds negative samples and corresponds to separate labels, which will help to better learn the relevant information of the two characteristics, thereby improving the generation accuracy of the model. After the model training is completed, each single-label data is generated, and each label generates several sequences as the input data of the single-label data integration stage. The objective function of the single-label feature learning stage can be expressed as: m G inm D axV s ing le (D,G) =E x～Pdata(x) [logD(x|y single-label )]+E z～Pz(z) [log(1-D(G(z|y single-label )))] (5) This formula can reflect the training process of the model in a single label dataset, where x represents the real sample; z represents the noise generated by the generator, P data represents the distribution of real data, P z represents the distribution of generated noise, y single-label Represents the label information corresponding to the single-label data, which serves as a condition to control the direction of model generation; At the same time, BLAST was introduced in the single-label BAP-GAN training process to evaluate the consistency of the generated sequences to determine whether the model is overfitting; Step 3.2: Single-label data integration phase ToxIBTL was selected as the toxicity classification model and CAMP as the antibacterial classification model. All data including the training samples and generated samples in the first stage were classified through these two characteristic classification models to screen out data with both positive toxicity and positive antibacterial properties. These two classification models are existing high-reliability classification models and can directly output classification results. Two multi-label datasets are constructed through the screened data: the toxicity & antibacterial positive dataset and the toxicity & antibacterial negative dataset. It should be noted that toxicity & antibacterial negative refers to not satisfying toxicity and antibacterial properties at the same time, so there may be a situation where only one characteristic is satisfied but not the other. These datasets finally constitute the complete multi-label BAP-GAN training set; Step 3.3: Multi-label feature learning phase Based on the multi-label BAP-GAN training set training model, a multi-label BAP-GAN model is obtained. At this time, the multi-label BAP-GAN model can generate corresponding multi-label data after specifying the corresponding labels, and can also generate corresponding single-label data when other labels are specified for the model, thereby realizing the unification of single-label and multi-label generation models. The objective function of the multi-label data learning stage is expressed as: m G inm D axV mu l t i(D,G) =E x～Pdata(x) [logD(x|y multi-label )]+E z～Pz(z) [log(1-D(G(z|y multi-label )))] (6) This formula can reflect the training process of the model in a multi-label dataset, where x represents the real sample, z represents the noise generated by the generator, and P data represents the distribution of real data, P z represents the distribution of generated noise, y multi-label Represents the label information corresponding to the multi-label data, which serves as a condition to control the direction of model generation; At the same time, BLAST is introduced in the multi-label BAP-GAN training process to evaluate the consistency of the generated sequences to determine whether the model is overfitting. The introduction strategy is consistent with step 3.1 of the invention.

6. A method for generating bioactive peptide sequences based on generative adversarial networks as claimed in claim 5, characterized in that: The specific operations for evaluating the consistency of the generated sequences in step 3.1 are as follows: First, the complete training set was merged and CD-HIT was used to remove redundancy with default parameters to obtain a non-redundant data set, which was used as the target database for BLAST search of generated sequences. Secondly, the sequences in the non-redundant data set were randomly shuffled and divided according to a certain ratio. One part was used as the target database of the training set, and the other part was used as the query sequence for BLAST search. Then, the BLAST results output for each sequence were grouped, and the highest consistency of each sequence in the target database was retained, and the median of the consistency was calculated as the consistency between natural sequences. After each round of iteration, a certain number of sequences of each tag were generated in turn, and BLAST was used to perform multiple sequence alignment using the generated sequences as query sequences, and the consistency in each round of alignment results was recorded. After the training, the BLAST search results were grouped in the same way, the highest consistency of each sequence was retained, and the consistency of all sequences was averaged. The average value was used as the consistency of the sequence generated by the model after this round of training. Finally, the consistency results of the generated sequences after each round of training were plotted to determine the consistency change trend of the generated sequences.

7. The method for generating bioactive peptide sequences based on generative adversarial networks according to claim 1, characterized in that: Step 4 is as follows: Step 4.1: Use ACC_single and ACC_multi as the evaluation indicators of the model to evaluate the feature accuracy. The specific formula is as follows: Among them, TP is the number of true positive samples, FN is the number of false positive samples, TN is the number of true negative samples, FP is the number of false negative samples, true positive samples refer to the generated positive samples that are indeed positive after testing, true negative samples refer to the generated negative samples that are indeed negative after testing, false positive samples refer to the generated positive samples that are negative after testing, and false negative samples refer to the generated negative samples that are positive after testing; Step 4.2: Taking the antibacterial positive as an example, the generated sequence is compared with the natural sequence, and the multi-label BAP-GAN model is used to generate several antibacterial positive single-label sequences. At the same time, the antibacterial positive sample set in the initial data set is selected, and its amino acid frequency information is counted respectively; Step 4.3: The present invention uses an output parameter "E value" in the BLAST search results to evaluate the homology of the final generated sequence. The E value refers to the expected value of the actual sequence comparison score, which indicates the statistical significance of the match between the query sequence and the hit sequence in a database of a specific size. If the E value is large, it means that the probability that the query sequence and the hit sequence have no homology is high. When the UniProt non-redundant database containing about 200 million sequences is used as the BLAST search target library, the E value is 10 -3 As the threshold for judging homology, when the E value is greater than 10 -3 , indicating that the query sequence has no homology with the hit sequence; when the E value is less than 10 -3 , indicating that the query sequence has homology with the hit sequence; since the training data set of BAP-GAN contains about 10,000 sequences after CD-HIT redundancy removal with default parameters, it is about 10 smaller than the Uniprot database 4 times, so the threshold should be reduced by 10 4 times, that is, 10 -7 is the E value threshold for judging homology; The multi-label BAP-GAN model was used to generate several sequences for each label, and the BLAST search was performed using the de-redundant training set as the target database. The E value with the highest consistency in each query sequence was recorded, and the E value distribution in the generated sequence was statistically analyzed. If the E value of the generated sequence was basically greater than 10 -3 , it means that the generated sequence has low homology and good biological significance.

Images