Protein post-translational modification data enhancement method based on generative adversarial network

Abstract

The invention is suitable for the technical field of proteomics, and provides a generative adversarial network-based protein post-translational modification data enhancement method. According to the method, class imbalance can be effectively relieved, the positive sample recognition capability can be enhanced, minority class pseudo samples are generated through an improved clustering enhancement conditional generative adversarial network (RP-CGAN), features are extracted in combination with an ESM-2 pre-training protein language model, and the positive sample recognition capability and generalization capability are improved; eSM-2 feature extraction and RP-CGAN data enhancement technologies are fused, so that the overall prediction performance and classification stability are improved, and key indexes of multiple classifiers are remarkably improved; through multi-dimensional index screening and self-adaptive regulation and control, it is ensured that enhanced data is close to real distribution; by means of a lightweight architecture and a convergence strategy, the training efficiency and the model stability are optimized; the method has strong stability and cross-domain potential, can be migrated to other protein modification prediction tasks, and has application potential in cross-domain unbalanced classification scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of proteomics, and particularly relates to a method for enhancing post-translational modification data of proteins based on a generative adversarial network. Background Art

[0002] S-sulfenylation is a type of protein post-translational modification (PTMs), which defines the covalent and chemical modification of protein residues and plays an important role in regulating various biological functions such as cardiovascular balance. Detecting S-sulfenylation sites is an important prerequisite for deeply exploring its regulatory functions. In the past few decades, proteomics techniques such as chromatin immunoprecipitation, liquid chromatography, and mass spectrometry have been proven effective in detecting S-sulfenylation sites. However, these laboratory techniques have unavoidable drawbacks, such as high time costs and expensive experimental expenses. Therefore, it is urgent to develop effective computer methods to improve the efficiency of identifying S-sulfenylation sites in proteins, reduce personnel and time costs, and accelerate the research process of its potential effects. There is still much room for development in deep learning algorithms for predicting S-sulfenylation sites.

[0003] Currently, in the prediction research of protein S-sulfenylation sites, some methods have introduced deep learning and natural language processing (NLP) related technologies to mine the hidden context information in the sequence. Compared with traditional machine learning models based on manual features, such methods have improved in modeling ability and prediction effect. However, the existing deep learning methods still have the following obvious defects: First, in terms of model selection, existing methods mostly adopt lightweight embedding methods or simplified language models, rather than using high-performance protein pre-training models such as ESM-2. As a large-scale protein language model, ESM-2 can extract richer context semantics and structural features from amino acid sequences, and its feature expression ability far exceeds traditional NLP or shallow network embedding methods. However, in the current field of S-sulfenylation site prediction, ESM-2 has not been effectively applied.

[0004] Second, at the data level, although some studies have used the publicly available iSulf-Cys dataset, which is one of the most comprehensive S-sulfenylation site datasets currently collected, and screened for modified sites in human proteins from the NCBI database. However, this dataset has problems of scarce samples and serious imbalance between positive and negative classes. Among the 8169 sequence information, there are only 1045 S-sulfenylation site samples. Such a data distribution situation is extremely likely to lead to overfitting of the model to negative class samples during training, thereby causing a decline in prediction performance.

[0005] Thirdly, in terms of technical application, the existing technology has not combined generative models such as generative adversarial networks (GANs) to effectively augment biological sequence data. In the case of a limited number of positive samples (sequences containing S-sulfonylation sites), the lack of reasonable data amplification means will limit the learning ability of the model. Especially when facing unknown samples in real scenarios, the generalization ability of the model will further decline.

[0006] To solve the above problems, the present invention proposes a method for augmenting post-translational modification data of proteins based on generative adversarial networks. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for augmenting post-translational modification data of proteins based on generative adversarial networks, aiming to solve the problems raised in the above background technology.

[0008] The purpose of the present invention is achieved through the following technical solutions: A method for augmenting post-translational modification data of proteins based on generative adversarial networks, comprising the following steps: Step S1: Data preprocessing and feature extraction; Read the protein sequence file, divide the training set and the test set, and use the ESM-2 pre-trained protein language model for feature encoding to generate standardized feature vectors; Step S2: Sample augmentation of the clustering-enhanced conditional generative adversarial network; Construct an RP-CGAN model composed of a generator and a discriminator; the generator inputs a random noise vector, a target class label, and a class center vector generated by K-means clustering, and outputs a pseudo-sample of the target class; the discriminator adopts a dual-output structure to respectively judge the authenticity of the sample and the class label, and combines Softplus the relative adversarial loss, binary cross-entropy loss, and gradient regularization term to optimize the training stability; Step S3: Pseudo-sample screening; Apply the trained RP-CGAN model to generate more pseudo-samples than the target number, and select the pseudo-samples closest to the real sample distribution through a screening mechanism with multiple indicators; Step S4: Classifier training and evaluation; Merge the real training data and the screened pseudo-samples, input them into the classifier for training, and optimize the classification performance; Step S5: Prediction output; Use the trained classifier to predict new protein sequences and output the probability scores of modification sites.

[0009] Furthermore, the data preprocessing and feature extraction step includes: Centered on the target residue, extract 10 amino acid residues upstream and downstream respectively to form a window sequence of 21 residues; Load the ESM-2 pre-trained protein language model and the alphabet; Use the get_batch_converter() method to convert the protein sequence into an input format recognizable by the model; Extract the token-level embedding representation of the 33rd layer of the model as the feature of each residue; Retain the original-dimensional embedding for each sequence to form a structure-aware feature matrix; Calculate the mean of the feature matrix for each sequence to obtain the final feature vector of the sequence.

[0010] Furthermore, the generator design of the RP-CGAN model includes: The input conditions are a random noise vector , a class label and the class center vector obtained by K-means clustering , and generate samples approximating the true distribution through a joint input splicing mechanism ; The generation steps of the class center vector include: Perform K-means clustering on the minority class and majority class samples respectively, and determine the optimal number of clustering clusters through the DBI index; Extract the optimal clustering centers of the minority class and majority class as the prior condition input of the generator.

[0011] Furthermore, the loss function of the RP-CGAN model includes: Softplus The relative adversarial loss of the variant: ; Among them, is the adversarial loss function; represents the expectation of the samples under the true sample distribution ; represents the expectation of the samples under the generated sample distribution ; is the true sample feature vector; is the pseudo-sample generated by the generator; Softplus The function is defined as ; is the authenticity output of the discriminator; The binary cross-entropy loss: ; Among them, is the classification loss function; Represents the joint distribution of sample features and labels Find the expectation; Is the sample feature vector; Is the true class label; Is the discriminator's class prediction output value for the sample ; Gradient regularization term: R1 regularization term: ; R2 regularization term: ; Among them, Is the gradient regularization term for real samples; Is the gradient regularization term for generated samples; Represents the gradient of the discriminator output with respect to the real sample input ; Represents the gradient of the discriminator output with respect to the generated sample input ; Represents the square of the L2 norm of the vector; Total loss function: Discriminator loss function: ; Generator loss function: ; Among them, Is the total loss function of the discriminator; Is the total loss function of the generator; Is the adversarial loss term of the discriminator; Is the adversarial loss term of the generator; Is the classification loss term of the discriminator; Is the classification loss term of the generator; Is the weight of the classification loss; And Are respectively Regularization term and The weight of the regularization term.

[0012] Furthermore, the pseudo-sample screening step includes: Calculate sensitivity and specificity through the confusion matrix, and dynamically determine the number of pseudo-sample augmentations ; Use the generator to generate pseudo-samples; construct a feature space using an AutoEncoder, calculate three distance metrics, namely the Pearson correlation coefficient, Euclidean distance, and mean squared error, between the pseudo-samples and real samples, normalize and weight-fuse the three distance metrics to obtain a comprehensive score; select the top pseudo-samples to form the final enhanced sample set .

[0013] Furthermore, the classifier training and evaluation steps include: Train using Random Forest, Support Vector Machine, XGBoost, and LightGBM; evaluate the classification performance through accuracy, F1 score, AUC, MCC, and G-mean values.

[0014] Furthermore, the prediction output step includes: Extract features from the new protein sequence and input them into the trained classifier; output the probability score of each residue being a modification site, and judge the modification site according to the threshold.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Effectively alleviate class imbalance and enhance the ability to identify positive samples: Generate minority class (positive) pseudo-samples through an improved clustering-enhanced conditional generative adversarial network (RP-CGAN), capture protein sequence context and structural information by combining high-dimensional semantic features extracted by the ESM-2 pre-trained protein language model to supplement scarce positive data. The generator introduces class labels and the clustering center of real samples as conditional inputs to fit the target class distribution, and the discriminator's dual-output structure guides the generation of samples close to the real distribution. The data of the embodiments shows that the balance indicators such as sensitivity (Sen), F1 score, MCC, and G-mean of the enhanced classifier are significantly improved, effectively improving the model bias caused by class imbalance and enhancing the ability to identify positive samples and generalization ability.

[0016] 2. Significantly improve the overall prediction performance and classification stability: Integrate ESM-2 feature extraction and RP-CGAN data augmentation techniques to provide rich and balanced training data for the classifier. The generated samples are optimized for stability through Softplus relative adversarial loss and gradient regularization terms (R1 / R2) to avoid the problem of gradient vanishing in traditional GANs. On mainstream classifiers such as RF, SVM, XGBoost, and LightGBM, the key indicators such as enhanced accuracy, AUC, F1 score, and G-mean are significantly improved, verifying the prediction stability and accuracy of the model under imbalanced data.

[0017] 3. Ensure that the enhanced data is similar to the real distribution: Through a multi-dimensional similarity index screening mechanism (Pearson correlation coefficient, Euclidean distance, autoencoder latent space error), comprehensively evaluate the generated samples, and combine adaptive sample quantity regulation (dynamically determine the enhancement scale based on the performance of the training set) to retain pseudo-samples that are highly similar to real samples, avoiding distribution drift. This mechanism improves the credibility of the enhanced data, ensures that no noise is introduced after the generated samples are integrated into the real distribution, provides reliable input for the subsequent classifier training, and indirectly supports the improvement of performance indicators.

[0018] 4. Optimize training efficiency and model stability: Through lightweight architecture design (such as the generator joint input splicing mechanism) and convergence optimization strategies (Early Stopping, GradScaler mixed-precision training, gradient regularization), improve training stability and accelerate convergence. The model shows a more robust parameter update trend during training, and the enhanced samples can quickly connect to the classifier training process, reflecting the efficiency advantages in engineering practice.

[0019] 5. Have strong stability and cross-domain expansion potential: RP-CGAN Softplus relieves the training instability of traditional GANs through relative adversarial loss and gradient regularization. In the embodiments, the performance of the classifier is steadily improved without crashing, proving the reliability of the generation model. Moreover, the method is based on a general unbalanced data augmentation framework, can be transferred to other protein modification prediction tasks, and has the application potential in cross-domain unbalanced classification scenarios such as financial fraud detection and medical image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the flowchart of the method of the present invention.

[0021] Figure 2 is the structural diagram of RP-CGAN. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] For a clearer understanding of the technical features, objectives, and beneficial effects of the present invention, the technical solutions of the present invention are described in detail below, but it should not be construed as a limitation on the scope of implementation of the present invention.

[0023] The present invention provides a method for enhancing post-translational modification data of proteins based on a generative adversarial network, and its flowchart is as Figure 1 shown. The method includes the following steps: Step S1: Data preprocessing and feature extraction; Input Standardization: Read the protein sequence file in FASTA format, which contains positive samples (labeled 1, representing modification) and negative samples (labeled 0, representing unmodified). Taking the target residue as the center, extract 10 amino acid residues upstream and downstream respectively to form a window sequence of 21 residues. Perform operations such as capitalizing the sequence, removing redundancy, and eliminating low-quality sequences. Each sequence is accompanied by a corresponding label (1 represents modification, 0 represents unmodified), providing standardized input for model training. The dataset partitioning is shown in Table 1.

[0024] Table 1 Dataset Partitioning Number of positive samples Number of negative samples Total Training set 900 6856 7756 Test set 145 268 413 Total 1045 7124 8169 Sequence Encoding: Use the ESM-2 pre-trained protein language model (with 650M parameters) open-sourced by Facebook to perform embedding encoding on the window sequences. The model is based on the Transformer architecture and is pre-trained on millions of protein sequences, capable of capturing complex context relationships and implicit structural information between amino acid residues.

[0025] The specific steps are as follows: Load the ESM-2 pre-trained protein language model and the alphabet; Use the get_batch_converter() method to convert the protein sequences into an input format recognizable by the model; Extract the token-level embedding representation of the 33rd layer of the model as the feature of each residue; Retain the original-dimensional embedding (1280 dimensions) for each sequence to form a structure-aware feature matrix; Calculate the mean of the feature matrix of each sequence to obtain the final feature vector of the sequence.

[0026] Step S2: Sample Enhancement of the Recursive Prototype-based Conditional Generative Adversarial Network (RP-CGAN); I. In the case where the number of S-sulfenylation samples is limited and the distribution is highly imbalanced, traditional supervised models are difficult to construct effective classification boundaries in high-dimensional and non-linear tasks. Data augmentation is an important means to improve model performance. Data augmentation methods based on Generative Adversarial Networks (GANs) can be used to generate data samples in the minority class.

[0027] Goodfellow et al. proposed Generative Adversarial Networks (GANs) in 2014, which consist of two parts: a generative model G and a discriminative model D. GANs utilize the idea of a zero-sum game and train the generative model through the adversarial process between the generative model G and the discriminative model D. The probability distribution of the real data x is , and the prior distribution of the input noise z is . The generative model G is a parameter-containing The multi-layer perceptron can be represented as a mapping function that maps the input noise z to the generated data . The discriminative model D is also a multi-layer perceptron with parameters , and its discriminative result is expressed as . The objective function of GANs can be defined as follows: ; where represents the distribution of real data; represents the distribution of generated data; indicates that x comes from real data rather than generated data; denotes the expectation. During the training process, mainly by training the discriminative model D to enhance the discriminative ability of the discriminator to maximize whether the recognition label comes from the training samples or the G distribution, that is, to maximize , while training the generative model G to minimize , which can be described as the min-max game problem of the objective function . Even though the generative adversarial network has powerful generative capabilities and high generation quality, it still has well-known deficiencies such as difficult network training, vanishing gradients, and model instability.

[0028] Conditional generative adversarial network (CGAN) is one of the extended forms of the classical generative adversarial network (GANs), first proposed by Mirza and Osindero in 2014. Its core idea is to add prior conditional information c (class label, text description, image feature, etc.) to the inputs of the generator G and the discriminator D to control the attributes or categories of the generated samples.

[0029] ; Compared with the original GANs that can only generate "unconditional" samples, CGAN has the ability to "generate on demand", so it is widely used in tasks such as image generation, image-to-image translation, sequence modeling, and molecular design.

[0030] In 2024, Huang et al. proposed a relative loss enhancement (RpGAN) that can be centered at zero gradient penalty objective, improving stability. Mathematically, it shows that RpGAN with gradient penalty has the same local convergence guarantee as the regularized classical GANs, and the regularization method ensures adversariality.

[0031] Second, based on the traditional CGAN structure, the present invention designs a generative adversarial network with classification output and regularization constraints to enhance the quantity and diversity of positive samples, named the clustering-enhanced conditional generative adversarial network (RP-CGAN), and its structure is as Figure 2As shown in the figure. RP-CGAN consists of a generator G and a discriminator D.

[0032] Generator G: Input: Random noise vector , the class label to be constructed (1 for positive), the class center vector obtained by clustering class samples ( d is the sample feature dimension).

[0033] To obtain the class center vector , for the set of feature vectors of real samples of the minority class and the majority class and , perform K-means (KMeans) clustering, and the objective function is: ; Among them, K is the number of clusters; is the th clustering cluster; is the sample feature vector belonging to the kth cluster ; is the center of this cluster.

[0034] For a given sample , where is the sample feature vector, is the true label, and the optimal class center extraction algorithm is shown in Algorithm 1. Suppose a certain class is clustered into k clusters by KMeans, and the corresponding set of class centers is , where each center vector is consistent with the original sample feature dimension.

[0035]

[0036] In the optimal class center extraction algorithm, is the number of clusters that minimizes the DBI of the clustering result of minority class samples (i.e., the optimal clustering number of the minority class); The number of clusters that minimizes the DBI for the clustering results of the majority class samples (i.e., the optimal number of clusters for the majority class); best_dbi_min is the current optimal DBI value for the minority class clustering; best_dbi_maj is the current optimal DBI value for the majority class clustering; kmeans_min_final is the final KMeans clustering model that performs optimally on the minority class samples among all the attempted numbers of clusters; kmeans_maj_final is the final KMeans clustering model that performs optimally on the majority class samples among all the attempted numbers of clusters; kmeans_min_final.cluster_centers_ is the set of cluster centers corresponding to the optimal clustering model for the minority class; kmeans_maj_final.cluster_centers_ is the set of cluster centers corresponding to the optimal clustering model for the majority class.

[0037] Finally, the generator will with concatenated into a joint input vector to generate a sample: ; The concatenation of the joint input vector guides it to generate samples that are more in line with the true distribution, effectively enhancing the generator's ability to model the target class and improving the pertinence and accuracy of sample generation. This design not only alleviates the common problem of single sample distribution in traditional generative adversarial networks but also accelerates the network convergence process by introducing prior structural information, enhancing the training stability and generation quality.

[0038] Discriminator D: has dual outputs and , which respectively judge "real / fake" and "predicted class" to assist the generator in generating target class samples specifically.

[0039] III. Loss Function and Training Strategy; Based on the loss of the original generative adversarial network, introduce: Softplus The relative adversarial loss of the variant: ; where is the adversarial loss function; is the feature vector of the real sample; is the fake sample generated by the generator; comes from the real sample distribution; is the distribution of the fake samples generated by the generator; represents the expectation of the samples under the real sample distribution ; represents the expectation of the samples under the generated sample distribution Find the expectation; Softplus The function is defined as ; is the authenticity output of the discriminator.

[0040] The adversarial loss adopts Softplus in the form to replace the traditional and , which can avoid the problem of vanishing gradient of the log function and enhance the training stability. This loss term simultaneously punishes the two situations of misjudging real samples as fake and misjudging fake samples as real, thereby guiding the discriminator to learn a more accurate real-fake boundary.

[0041] Binary cross-entropy loss (BCE): ; Among them, is the classification loss function; represents the expectation of the joint distribution of sample features and labels Find the expectation; is the sample feature vector; is the true class label; is the class prediction output value of the discriminator for the sample ; In addition to outputting the adversarial score, the discriminator also outputs the class prediction probability at the same time. For this reason, the present invention introduces a binary cross-entropy loss function to measure the deviation between the predicted label of the discriminator and the true label. This loss term can effectively strengthen the ability to utilize conditional information, guide the discriminator to distinguish between positive and negative sample classes, and also prompt the generator to synthesize samples that more conform to the target class distribution.

[0042] Gradient regularization term (R1 / R2): R1 regularization term (for real samples): ; R2 regularization term (for generated samples): ; Among them, is the gradient regularization term for real samples; is the gradient regularization term for generated samples; represents the gradient of the discriminator output with respect to the real sample input ; represents the gradient of the discriminator output with respect to the generated sample input ; represents the square of the L2 norm of the vector.

[0043] The present invention introduces a gradient regularization mechanism in GAN training. Among them, the R1 regularization term imposes a penalty on the gradient of real samples to prevent the discriminator from overfitting the training data; the R2 regularization term is used to limit the gradient fluctuation of generated samples, effectively alleviating the problems of unstable training and gradient explosion.

[0044] Total loss function: Discriminator loss function: ; Generator loss function: ; Among them, is the total loss function of the discriminator; is the total loss function of the generator; is the adversarial loss term of the discriminator; is the adversarial loss term of the generator; is the classification loss term of the discriminator; is the classification loss term of the generator; is the weight of the classification loss; and are the weights of the R1 regularization term and the R2 regularization term respectively, and take the same value in this task.

[0045] Training control mechanism: Set Early Stopping to prevent overfitting; use mixed-precision training through GradScaler to accelerate convergence.

[0046] After the RP-CGAN training is completed, save the weight parameters of the generator G for subsequent expansion of the training set.

[0047] Step S3: Pseudo-sample screening; Determine the number of enhanced pseudo-samples according to the training difficulty. Apply the trained RP-CGAN model to generate more pseudo-samples than the target number. Screen high-quality pseudo-samples through the discriminator-guided and multi-index-constrained pseudo-sample screening algorithm (as shown in Algorithm 2) to enhance the rationality of the training data and reduce the risk of distribution drift. The specific process includes: determining the number of pseudo-sample enhancements based on the classifier performance, constructing a feature space using the AutoEncoder, comprehensively evaluating the quality of pseudo-samples through three indicators: Pearson correlation coefficient (PCC), Euclidean distance (ED), and minimum mean square error (MSE), and finally selecting the pseudo-samples most similar to the real samples as the enhanced data to provide reliable input for subsequent classifier training.

[0048] In the pseudo-sample screening algorithm with discrimination guidance and multi-index constraints, TP is the number of true positives, that is, the number of positive classes correctly predicted as positive by the model; TN is the number of true negatives, that is, the number of negative classes correctly predicted as negative by the model; FP is the number of false positives, that is, the number of negative classes incorrectly predicted as positive by the model; FN is the number of false negatives, that is, the number of positive classes incorrectly predicted as negative by the model; is the indicator function, which is 1 when the condition in the parentheses is true and 0 otherwise; N is the total number of negative samples; P is the total number of positive samples; is the basic enhancement coefficient, is a tiny constant to prevent the denominator from being zero; is the set of Pearson maximum correlation coefficients for all pseudo-samples; is the maximum value of the Pearson correlation coefficient between the is the set of minimum Euclidean distances between all pseudo-samples and real samples; is the minimum Euclidean distance between the is the set of minimum feature space errors for all pseudo-samples; is the minimum mean square error between the is the standardized score of the Pearson correlation coefficient of the is the standardized score of the Euclidean distance of the is the standardized score of the latent space MSE of the is the comprehensive score of the

[0049] Step S4: Classifier training and evaluation; After sample enhancement, the real training data and the screened pseudo-samples are merged and input into multiple classic classifiers for supervised learning training, including Random Forest (RF), Support Vector Machine (SVM), XGBoost, and LightGBM. Multiple evaluation metrics such as Accuracy, F1-Score, Area Under the Curve (AUC), Matthews Correlation Coefficient (MCC), and Geometric Mean (G-mean value) are used to evaluate the performance of the classifier, and the classification threshold is automatically optimized through grid search within the range of 0.1 - 0.9 to evaluate the recognition ability of the S-sulfenylation site of the classifier.

[0050] Step S5: Prediction output; Use the trained classifier to predict new protein sequences, output the probability score that each residue is an S-sulfenylation site, and determine the modification site according to the probability score to achieve automatic recognition.

[0051] The following describes the specific implementation of the present invention in detail in conjunction with specific embodiments.

[0052] Example 1: To verify the effectiveness of the proposed RP-CGAN and its pseudo-sample screening mechanism in the S-sulfenylation site prediction task, four mainstream classifiers (RF, SVM, XGBoost, LightGBM) were selected for comparative testing. The performance indicators include accuracy (Acc), sensitivity (Sen), specificity (Spe), precision (Pre), Matthews correlation coefficient (MCC), F1 score, area under the curve (AUC), and G-mean value. Table 2 shows the comparison of classification performance before and after enhancement: Table 2 Comparison of classification performance before and after enhancement

[0053] It can be seen from the data in the table that among all classifiers, the performance indicators after enhancement have been improved to varying degrees, especially in the balance indicators such as Sen, F1 score, MCC, and G-mean value, indicating that the generated samples effectively improve the class imbalance problem. In addition, the improvement of RF&SVM is the most obvious. The F1 of RF increases from 0.5312 to 0.5674, and the G-mean increases from 0.5984 to 0.6408, proving that the enhanced data improves the ability to identify positive samples; XGBoost maintains a high Spe and makes up for the low Sen: Although the Sen value is still lower than other models, both F1 and AUC increase, indicating that the pseudo-samples have an effect on improving the recall rate; LightGBM is the most balanced, maintaining a high overall performance before and after enhancement, and is currently the best classifier option in terms of comprehensive indicators.

[0054] The above is only the preferred implementation manner of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent.

Claims

1. A method for post - translational modification data augmentation of proteins based on generative adversarial networks, characterized in that, It includes the following steps: Step S1: Data preprocessing and feature extraction; Read the protein sequence file, divide the training set and the test set, and use the ESM-2 pre-trained protein language model for feature encoding to generate standardized feature vectors; Step S2: Sample augmentation of the clustering-enhanced conditional generative adversarial network; Construct an RP-CGAN model consisting of a generator and a discriminator; the generator takes as input a random noise vector, a target class label, and a class center vector generated by K-means clustering, and outputs a pseudo-sample of the target class; the discriminator adopts a dual-output structure to respectively judge the authenticity of the sample and the class label, and combines Softplus the relative adversarial loss, binary cross-entropy loss, and gradient regularization term to optimize the training stability; Step S3: Pseudo-sample screening; Apply the trained RP-CGAN model to generate pseudo-samples exceeding the target quantity, and select the pseudo-samples closest to the real sample distribution through a screening mechanism with multiple metric constraints; Step S4: Classifier training and evaluation; Merge the real training data and the screened pseudo-samples, input them into the classifier for training, and optimize the classification performance; Step S5: Prediction output; Use the trained classifier to predict new protein sequences and output the probability scores of modification sites.

2. The method for enhancing post-translational modification data of proteins based on a generative adversarial network according to claim 1, wherein The data preprocessing and feature extraction steps include: Centered on the target residue, extract 10 amino acid residues upstream and downstream respectively to form a window sequence of 21 residues; Load the ESM-2 pre-trained protein language model and the alphabet; Use the get_batch_converter() method to convert the protein sequence into an input format recognizable by the model; Extract the token-level embedding representation of the 33rd layer of the model as the feature of each residue; Retain the original-dimensional embedding for each sequence to form a structure-aware feature matrix; Calculate the mean of the feature matrix of each sequence to obtain the final feature vector of the sequence.

3. The method for enhancing post-translational modification data of proteins based on a generative adversarial network according to claim 1, wherein The generator design of the RP-CGAN model includes: The input conditions are a random noise vector , a class label and a class center vector obtained by K-means clustering , and samples approximating the true distribution are generated through a joint input splicing mechanism ; The generation steps of the class center vector include: Perform K-means clustering on the minority-class and majority-class samples respectively, determine the optimal number of clustering clusters through the DBI index; extract the optimal clustering centers of the minority class and the majority class as the prior condition input of the generator.

4. The method for enhancing post-translational modification data of proteins based on a generative adversarial network according to claim 1, wherein The loss function of the RP-CGAN model includes: Softplus Relative adversarial loss of the variant: ； Among them, is the adversarial loss function; represents taking the expectation of samples under the true sample distribution ; represents taking the expectation of samples under the generated sample distribution ; is the feature vector of the true sample; is the pseudo-sample generated by the generator; Softplus The function is defined as ; The authenticity output of the discriminator; Binary cross-entropy loss: ； Among them, is the classification loss function; represents the joint distribution of sample features and labels to calculate the expectation; is the sample feature vector; is the true class label; is the discriminator's class prediction output value for the sample ; Gradient regularization term: R1 regularization term: ； R2 regularization term: ； Among them, is the gradient regularization term for real samples; is the gradient regularization term for generated samples; represents the gradient of the discriminator output with respect to the real sample input ; represents the gradient of the discriminator output with respect to the generated sample input ; represents the square of the L2 norm of the vector. Total loss function: Discriminator loss function: ； Generator loss function: ； Among them, is the total loss function of the discriminator; is the total loss function of the generator; is the adversarial loss term of the discriminator; is the adversarial loss term of the generator; is the classification loss term of the discriminator; is the classification loss term of the generator; is the weight of the classification loss; and are respectively the regularization term and the weight of the regularization term.

5. The method for enhancing post-translational modification data of proteins based on a generative adversarial network according to claim 1, wherein The pseudo-sample screening steps include: Calculate sensitivity and specificity through the confusion matrix, and dynamically determine the number of pseudo-sample augmentations ; Use the generator to generate pseudo-samples; Use the AutoEncoder to construct the feature space, calculate three types of distance metrics, namely the Pearson correlation coefficient, Euclidean distance, and least mean square error between the pseudo-samples and the real samples, normalize and weight-fuse the three types of distance metrics to obtain a comprehensive score; Select the top pseudo-samples with the lowest comprehensive scores to form the final augmented sample set .

6. The method for enhancing post-translational modification data of proteins based on a generative adversarial network according to claim 1, wherein The classifier training and evaluation steps include: Use random forest, support vector machine, XGBoost and LightGBM for training; evaluate the classification performance through accuracy, F1 score, AUC, MCC and G-mean values.

7. The method for enhancing post - translational modification data of proteins based on generative adversarial networks according to claim 1, wherein The prediction output steps include: Extract features from the new protein sequence and input them into the trained classifier; output the probability scores of each residue as the modification site, and judge the modification site according to the threshold.

Images