Protein cysteine residue redox modification prediction method based on MusteDeep framework

Abstract

The invention relates to the technical field of protein redox, and particularly discloses a protein cysteine residue redox modification prediction method based on a MusteDeep framework, and the method comprises the following steps: data preprocessing, convolutional neural network construction, attention mechanism introduction, Bootstrap method introduction, model result output and the like. The extended model is used for predicting the redox reaction of the cysteine residue in the protein, the accuracy of the extended model in predicting the redox reaction of the cysteine residue is remarkably improved, the redox reaction can be learned and predicted from original protein sequence data, predefined feature engineering is not needed, and the accuracy of the extended model in predicting the redox reaction of the cysteine residue is greatly improved. Manual intervention and subjective judgment are greatly reduced, and the objectivity and efficiency of prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of protein redox modification, and in particular relates to a method for predicting redox modification of protein cysteine ​​residues based on the MusiteDeep framework. Background Art

[0002] The redox reaction of proteins is one of the important mechanisms for regulating the intracellular and extracellular environment. It plays a key role in signal transduction, metabolic regulation, and the regulation of protein structure and function. For example, the redox reaction of cysteine ​​residues involves the formation and cleavage of disulfide bonds, which directly affects the three-dimensional structure of proteins and thus their biological activity. Accurately predicting the redox state of these sites is crucial for understanding the function of proteins and designing targeted drugs.

[0003] Traditionally, methods for predicting redox reaction sites rely on feature engineering, a process that requires experts to manually design and extract features from protein sequences. However, this method has significant limitations: on the one hand, manual feature selection may introduce incompleteness or bias due to subjective judgment, making it difficult to fully capture the complex and subtle patterns in the sequence; on the other hand, the feature extraction step is cumbersome and time-consuming, and is highly dependent on specific data sets, resulting in limited performance when processing large-scale, diverse protein data, making it difficult to meet the current needs of high-throughput data analysis. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting redox modification of protein cysteine ​​residues based on the MusiteDeep framework to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The protein cysteine ​​residue redox modification prediction method based on the MusiteDeep framework includes the following steps:

[0007] S1. Extracting sequences from a protein sequence file in FASTA format, dividing the sequences into segments of fixed length, wherein the center position of each segment is the target predicted residue - cysteine, and the sequence segments are converted into a digital format that can be processed by a neural network through one-hot encoding;

[0008] S2, inputting the encoded sequence fragments into a multi-layer convolutional neural network, and using the multi-layer convolutional neural network (CNN) to extract features of the encoded sequence fragments;

[0009] S3, introduce a two-dimensional attention mechanism to weight the features, focusing on the sequence positions and features that are important for redox reaction prediction;

[0010] S4. The Bootstrap method is used to deal with the imbalance of positive and negative samples in the data set. The Bootstrap method divides the negative samples into multiple subsets. In each iteration, a subset with balanced positive and negative samples is selected for training. Multiple classifiers are generated through multiple iterations, and the average value is taken to determine the final prediction result.

[0011] S5. Input the weighted features into the fully connected layer and output the probability value representing the possibility of residue redox reaction.

[0012] Preferably, the fixed-length fragment is 33 residues, including 16 leading residues, 16 trailing residues and one central residue, and the central residue is the target predicted cysteine ​​residue.

[0013] Preferably, the MusiteDeep framework adopts a multi-layer CNN structure, and each layer of CNN gradually extracts high-level feature representations through convolution and pooling operations.

[0014] Preferably, the dual-dimensional attention mechanism weights features in both the sequence dimension and the feature map dimension.

[0015] Preferably, the probability value output by the fully connected layer is used to determine whether the target residue has undergone a redox reaction.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] The present invention is based on the deep learning model of the MusiteDeep framework, which is used to predict the redox reaction of cysteine ​​residues (Cysteine, C) in proteins. The accuracy of the extended model in predicting the redox reaction of cysteine ​​residues is significantly improved, and the redox reaction can be learned and predicted from the original protein sequence data without the need for predefined feature engineering, which greatly reduces manual intervention and subjective judgment, and improves the objectivity and efficiency of prediction. Provide new perspectives and tools for protein function research. For example, by predicting the key sites of redox reactions, researchers can understand the regulatory mechanism of proteins in vivo, and then reveal their relationship with the occurrence and development of diseases. By accurately predicting redox reactions, experimental schemes can be designed more effectively to accelerate the drug development process. At the same time, the model constructed by the present invention is applicable to the prediction of protein cysteine ​​redox modification in most species including but not limited to rice, corn, and humans. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the steps of the present invention.

[0019] Figure 2 The test set verifies the accuracy of the model.

[0020] Figure 3 Predicting protein cysteine ​​redox modifications in rice.

[0021] Figure 4 Predicting protein cysteine ​​redox modifications in maize. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] Example:

[0024] refer to Figure 1 As shown, the protein cysteine ​​residue redox modification prediction method based on the MusiteDeep framework includes the following steps:

[0025] S1. Extracting sequences from a protein sequence file in FASTA format, dividing the sequences into segments of fixed length, wherein the center position of each segment is the target predicted residue - cysteine, and the sequence segments are converted into a digital format that can be processed by a neural network through one-hot encoding;

[0026] S2, inputting the encoded sequence fragments into a multi-layer convolutional neural network, and using the multi-layer convolutional neural network (CNN) to extract features of the encoded sequence fragments;

[0027] S3, introduce a two-dimensional attention mechanism to weight the features, focusing on the sequence positions and features that are important for redox reaction prediction;

[0028] S4. The Bootstrap method is used to deal with the imbalance of positive and negative samples in the data set. The Bootstrap method divides the negative samples into multiple subsets. In each iteration, a subset with balanced positive and negative samples is selected for training. Multiple classifiers are generated through multiple iterations, and the average value is taken to determine the final prediction result.

[0029] S5. Input the weighted features into the fully connected layer and output the probability value representing the possibility of residue redox reaction.

[0030] Specifically, detailed parameter settings based on the MusiteDeep framework and reference training and prediction codes.

[0031] Input files and data formats

[0032] Input file (-input): The input data file should be a protein sequence file in FASTA format;

[0033] Training process: . / new_training_data_2 / Train.fasta;

[0034] Prediction process: . / new_training_data_2 / new_task.fasta is used to batch predict the redox reactions of protein sequences.

[0035] Output file and prefix settings

[0036] Output prefix (-output-prefix): The prefix of the output file is used to identify the output results of different models;

[0037] Training: new_custom_general_multi_spec_2_OXI_C generates output files related to training;

[0038] Prediction: 8_16_new_task_predict, used to specify the output file of the prediction results.

[0039] Residue type and classification number settings

[0040] -residue-types C: This model focuses on predicting the redox reactions of cysteine ​​(C); this parameter ensures that the model only focuses on the prediction of the specified residue;

[0041] Number of classes (-nclass=5): The prediction results are divided into five categories, representing different redox states; this multi-classification setting helps capture subtle changes in redox states in protein sequences.

[0042] Training and prediction types

[0043] Training Type:

[0044] Use the train_general.py script and specify -residue-types C to train a model for a specific residue type.

[0045] Prediction type (-predict-type custom):

[0046] Use the predict.py script and specify -predict-type custom to call the trained model for prediction;

[0047] Model file settings

[0048] Model file prefix (-model-prefix): specifies the model file prefix to be used during the prediction process. For example, -model-prefix custom_general_multi_spec_2_OXI_C is used to call the five trained classification models to predict the input sequence;

[0049] Other parameter settings

[0050] Window size and sequence fragment: A window size of 33 residues was used, with 16 residues located before the central residue and 16 residues located after it. The central residue was the target predicted redox site.

[0051] Encoding:

[0052] The sequence fragments are encoded as one-hot vectors and converted into a digital format that can be processed by the neural network. The 20 common amino acids are encoded as 20-dimensional vectors, while unknown or non-standard amino acids (such as "X") use a special encoding strategy to prevent negative impact on model training.

[0053] Optimization method:

[0054] During the training process, the Adam optimizer was used, combining the momentum method with an adaptive learning rate to ensure fast convergence and reduce the dependence on parameter adjustments during training.

[0055] Learning rate setting:

[0056] The initial learning rate is set to 0.001, combined with a gradual decay mechanism. The learning rate determines the update speed of the model parameters. A moderate learning rate ensures stable convergence of the model and avoids skipping the global optimal solution.

[0057] Bootstrap strategy for training:

[0058] In order to deal with the problem of imbalance between positive and negative samples, the Bootstrap strategy is adopted to divide the negative samples into multiple subsets, and use the subsets with balanced positive and negative samples in each training.

[0059] The prediction result is determined by the average value of multiple classifiers to improve the stability of the model and the prediction accuracy.

[0060] The parameter settings ensure that the model based on the MusiteDeep framework can accurately predict the redox reactions of cysteine ​​residues in proteins, providing flexible training and prediction options to meet different research needs.

[0061] The method of the present invention automatically mines deep patterns in protein sequence data through deep learning technology, especially convolutional neural network (CNN) and attention mechanism, so as to accurately predict the redox state of cysteine ​​residues.

[0062] In summary, first, sequences are extracted from protein sequence files in FASTA format, and these sequences are segmented into fixed-length segments containing the target predicted residue, cysteine ​​(C). In the present invention, a window size of 33 residues is selected to ensure that sufficient contextual information is incorporated into the model. Next, the sequence segments are converted into a digital format that can be processed by a neural network through one-hot encoding.

[0063] Then, a multi-layer convolutional neural network (CNN) is used to extract features from the encoded sequence fragments. The advantage of CNN is that it can capture local patterns in the sequence and gradually extract high-level feature representations through convolution and pooling operations.

[0064] To further improve prediction accuracy, the present invention introduces a dual-dimensional attention mechanism. This mechanism weights features in both the sequence dimension and the feature graph dimension, thereby automatically focusing on those sequence positions and features that are most important for redox reaction prediction. This mechanism significantly improves the prediction accuracy of the model and enhances the interpretability of the model output.

[0065] To solve the problem of serious imbalance in the number of positive and negative samples in the data set, the present invention adopts the Bootstrap method. This method divides the negative samples into multiple subsets, and selects a subset with balanced positive and negative samples for training in each iteration. Multiple classifiers are generated through multiple iterations, and the average value is taken to determine the final prediction result, thereby improving the stability and prediction accuracy of the model.

[0066] Finally, the weighted features are input into the fully connected layer, and after further nonlinear transformation, a probability value representing the possibility of residue redox reaction is output. This probability value is used to determine whether the target residue has undergone redox reaction, and the prediction results are divided into five categories, representing different redox states.

[0067] With the rapid development of deep learning technology, especially the introduction of convolutional neural networks (CNN) and attention mechanisms, researchers in the field of computational biology have begun to use these technologies to automatically mine deep patterns in sequence data. The biggest advantage of deep learning is that it can directly process raw data without the need for predefined features, thereby reducing manual intervention and being able to more comprehensively and accurately predict key sites in biological processes.

[0068] In this context, a deep learning model based on the MusiteDeep framework is used to predict the redox reaction of cysteine ​​residues (Cysteine, C) in proteins. The MusiteDeep framework was originally developed to predict phosphorylation sites. It can effectively capture complex patterns in protein sequences by using convolutional neural networks combined with a two-dimensional attention mechanism. On this basis, its functionality has been expanded to enable it to predict the redox reaction of specific residues. This extension not only retains the advantages of the MusiteDeep framework, but also provides efficient prediction capabilities for redox reactions, providing a powerful tool for studying protein functions, designing experiments, and developing new drugs.

[0069] The datasets used to train the model come from literature search datasets and public databases, including data from corn (Zea_mays), rats (Rattus norvegicus), poplar plants (Populus), mice (Mus musculus), Caenorhabditis elegans (C.elegans), rapeseed (Brassica napus), humans (Homo sapiens), tomatoes (Solanumlycopersicum), Bacillus cereus, etc. These datasets contain protein sequences of thousands of different organisms and their corresponding redox states. Specifically, the main data sources include Uniprot and PlantPTM Viewer2.0 (https: / / www.psb.ugent.be / webtools / ptm-viewer / index.php) biological databases. The dataset splitting method during the model training process is as follows: collect data, perform data cleaning, and obtain the total dataset S. 1. Split S into S1 (training set) and S2 (external validation set) at a ratio of 0.85:0.15. 2. Model training: 1) Input S1 into the model and perform 10X cross validation, which will automatically split it into training set and test set during the process; 2) According to the various indicators obtained from the training, the optimal model L is obtained. 3. External validation: Use the L model to predict S2 and obtain various indicators of the predicted results and the actual results (accuracy, recall, specificity, etc.).

[0070] The data reported in the literature supplement the redox sites that may be missed in the database. In order to ensure the quality of the data, all sequences were strictly screened and standardized, low-quality sequences and redundant sequences were removed, and the label information of each sequence was ensured to be accurate. The sequences were also compared using the BLAST tool to ensure that there were no highly similar sequences in the training set and the test set, thereby avoiding the performance degradation of the model due to overfitting.

[0071] in Figure 2 The model based on the present invention is shown, and 14324 Cys in the test set data are used for prediction, and the results show that 4111 Cys among them can undergo redox modification, which is consistent with 4111 Cys sites among 5078 Cys sites identified by redox modification mass spectrometry, with an accuracy rate of 80.96%.

[0072] in Figure 3 The results show that the redox modification prediction of 14023 Cys in rice based on the model of the present invention shows that 5024 Cys can undergo redox modification, which is consistent with 5024 Cys sites among 5389 Cys sites identified by redox modification mass spectrometry, with an accuracy rate of 93.23%.

[0073] in Figure 4 The results show that the redox modification prediction of 440,159 Cys in corn was performed based on the model of the present invention, and the results show that 5,011 Cys can undergo redox modification, which is consistent with 5,011 Cys sites among 5,967 Cys sites identified by redox modification mass spectrometry, with an accuracy rate of 83.98%.

[0074] It will be appreciated that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but will be a routine task of design, fabrication, and production for those of ordinary skill having the benefit of this disclosure without undue experimentation.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting redox modification of protein cysteine ​​residues based on the MusiteDeep framework, characterized in that: The following steps are involved: S1. Extracting sequences from a protein sequence file in FASTA format, dividing the sequences into segments of fixed length, wherein the center position of each segment is the target predicted residue - cysteine, and the sequence segments are converted into a digital format that can be processed by a neural network through one-hot encoding; S2, inputting the encoded sequence fragments into a multi-layer convolutional neural network, and using the multi-layer convolutional neural network (CNN) to extract features of the encoded sequence fragments; S3, introduce a two-dimensional attention mechanism to weight the features, focusing on the sequence positions and features that are important for redox reaction prediction; S4. The Bootstrap method is used to deal with the imbalance of positive and negative samples in the data set. The Bootstrap method divides the negative samples into multiple subsets. In each iteration, a subset with balanced positive and negative samples is selected for training. Multiple classifiers are generated through multiple iterations, and the average value is taken to determine the final prediction result. S5. The weighted features are input into the fully connected layer, and after nonlinear transformation, the probability value representing the possibility of residue redox reaction is output.

2. The method for predicting protein cysteine ​​residue redox modification based on the MusiteDeep framework according to claim 1, characterized in that: The fixed length fragment is 33 residues, including 16 leading residues, 16 trailing residues and one central residue, and the central residue is the target predicted cysteine ​​residue.

3. The method for predicting redox modification of protein cysteine ​​residues based on the MusiteDeep framework according to claim 1, characterized in that: The MusiteDeep framework adopts a multi-layer CNN structure, and each layer of CNN gradually extracts high-level feature representations through convolution and pooling operations.

4. The method for predicting protein cysteine ​​residue redox modification based on the MusiteDeep framework according to claim 1, characterized in that: The dual-dimensional attention mechanism weights features in both the sequence dimension and the feature map dimension.

5. The method for predicting protein cysteine ​​residue redox modification based on the MusiteDeep framework according to claim 1, characterized in that: The probability value output by the fully connected layer is used to determine whether the target residue has undergone a redox reaction.

Images