A deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection

Abstract

The application discloses a kind of deep learning methods for identifying phosphorylation sites of SARS-CoV-2 infection, comprising: first, collect the phosphorylation site dataset of known SARS-CoV-2 infected human A549 cells;Then five feature encoding methods are used to extract the feature representation of peptide sequence respectively;The features extracted by the five feature encoding methods are vector splicing;Then a deep learning model is constructed based on convolutional neural network, gating mechanism and bidirectional gated recurrent unit;The encoding vector of positive and negative samples and its label are used to train the deep learning neural network model;The trained model is used to predict unknown peptide sequence, the application makes full use of the advantages of convolutional neural network, gating mechanism and bidirectional gated recurrent unit network, not only realizes the synergistic effect of feature extraction, information integration and time sequence processing, but also provides an innovative and accurate method for efficient identification of SARS-CoV-2 infection phosphorylation sites.

Description

Technical Field

[0001] This invention relates to the field of computational bioinformatics, and in particular to a deep learning method for identifying phosphorylation sites in SARS-CoV-2 infection. Background Technology

[0002] Phosphorylation is a common post-translational modification that affects many basic cellular processes. Studies have found that SARS-CoV-2 proteins are extensively phosphorylated in host cells. SARS-CoV-2 infection triggers host kinases, which are responsible for heavy phosphorylation in both the host and the virus.

[0003] With the continuous development of high-throughput sequencing technology, the functions of many phosphorylation sites have been well annotated. Bouhaddou et al. presented a quantitative mass spectrometry-based phosphorylated proteomics study investigating SARS-CoV-2 infection in cells, revealing recombinant phosphorylation of host and viral proteins. Hekman et al. conducted a quantitative phosphorylated proteomics study of SARS-CoV-2 infection to identify the connections. Due to the biological importance of protein kinases in cellular signaling and the steadily increasing number of reports identifying phosphorylation sites, it has become impractical for experimental molecular biologists to track all phosphorylation modifications of proteins in their field of study, as most experiments require expensive equipment and labor.

[0004] Furthermore, many machine learning methods have been developed to predict phosphorylation sites. For example, QUOKKA uses multiple sequence scoring functions combined with an optimized logistic regression algorithm to predict phosphorylation sites. PhosPred-RF and PhossphoSVM use only sequence-based features for prediction based on random forest (RF) and support vector machine (SVM), respectively. However, these machine learning-based methods rely too heavily on features. Although various feature and feature selection methods have been adopted to improve model performance, the generalization ability of these methods needs to be improved.

[0005] This patent, CN111696621B, discloses a method, system, device, and storage medium for identifying the relationship between protein phosphorylation modification sites and diseases. The method includes: acquiring information; constructing a relational dataset and a non-relational dataset; constructing relational feature vectors and non-relational feature vectors; and inputting the relational feature vectors and non-relational feature vectors to obtain the identification result. The system includes an acquisition module, a dataset module, a feature vector module, and an output module. The device includes a memory and a processor for executing the aforementioned method for identifying the relationship between protein phosphorylation modification sites and diseases. This patent is based on random forests for prediction; however, these machine learning-based methods rely too heavily on features. Although various feature selection methods have been employed to improve model performance, the generalization ability of this method needs improvement. Summary of the Invention

[0006] To address the problems mentioned in the background section, this invention provides a deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection, thereby resolving the issues of existing technologies.

[0007] The technical solution adopted in this invention is:

[0008] A deep learning method for identifying phosphorylation sites in SARS-CoV-2 infection includes the following steps:

[0009] S1: Collect a dataset of known phosphorylation sites in human A549 cells infected with SARS-CoV-2;

[0010] S2: Five feature encoding methods were used to extract feature representations of peptide sequences: one-hot encoding, amino acid substitution scoring matrix (BLOSUM62), new extended and updated principal attribute scale (ZScale), 5-bit binary type 1, and 5-bit binary type 2.

[0011] S3: Concatenate the vectors of the features extracted by these five feature encoding methods;

[0012] S4: A one-dimensional convolutional neural network is concatenated with three parallel one-dimensional convolutional layers of the same scale, which are respectively connected to a pooling layer, a batch normalization layer, a dropout layer, and a sigmoid activation layer. The activation results are multiplied and flattened. The gating mechanism controls the output information of the convolutional layer through the sigmoid activation function. The convolutional neural network is also concatenated with a bidirectional gated recurrent unit network to flatten the output feature vector again. The results of the two flattening are concatenated to form a vector containing multiple feature representations. Finally, this vector is passed to a fully connected layer for classification.

[0013] S5: Peptide sequence fragments with phosphorylated S / T sites at the center are defined as positive samples and labeled as 1; otherwise, they are considered negative samples and labeled as 0. The shuffled training set is used to supervise the training of the deep learning neural network model. The encoding vectors of positive and negative samples and their labels are used to train the deep learning neural network model.

[0014] S6: Use a trained model to predict unknown peptide sequences.

[0015] Furthermore, the one-hot encoding method is based on a given protein sequence S = {a1, a2, ..., a...} i}, where i represents the length of the sequence, a iThe residue at position i is represented as a 20-dimensional vector, where only one dimension has a value of 1, representing the amino acid type at that position, and the other dimensions have values ​​of 0.

[0016] Furthermore, the amino acid substitution scoring matrix method includes: first, constructing a substitution matrix based on known protein sequence comparison data, which describes the frequency at which one amino acid is substituted by another in a protein sequence; then, calculating a similarity score. For the comparison between two protein sequences, the similarity score can be calculated based on the substitution matrix. The similarity score between different positions of the two protein sequences can be expressed by the following formula: Score = M(s1[i], s2[j]), where M is a BLOSUM62 matrix, M ∈ R m×d Where m is the number of amino acid types, d is the dimension of the embedding vector, and Score represents the similarity score between position i of protein sequence 1 and position j of protein sequence 2. Finally, the scores are accumulated. For the alignment of the entire protein sequence, the similarity scores between each pair of amino acids are accumulated to obtain a total similarity score. This score can be used to measure the similarity between two protein sequences. The higher the score, the more similar the two sequences are.

[0017] Furthermore, the newly extended and updated main attribute scaling method characterizes each amino acid using five physicochemical descriptors. By using these descriptor variables, the differences and similarities between different amino acids in the protein sequence are captured, thereby analyzing and predicting the structure and function of the protein. The specific representation of the amino acids is shown in Table 1.

[0018] Table 1

[0019]

[0020]

[0021] Furthermore, the 5-bit binary type method uses a set of amino acid groups containing 5 letters {a1,a2,a3,a4,a5}. Each amino acid group has a corresponding five-dimensional binary vector representation. Then, each group is represented by a five-dimensional binary vector. a1 is encoded using binary vector (10000), a2 is encoded using binary vector (01000), ..., a5 is encoded using binary vector (00001). During the encoding process, each binary bit corresponds to a specific amino acid. When an amino acid belongs to the corresponding group, the corresponding binary bit is set to 1, otherwise it is 0.

[0022] Furthermore, the 5-bit binary type 2 method is based on all possible ways to combine 1s and 0s in a 5-bit unit. There are 32 possible representations for 20 amino acids. When the representations with no 1s or all 1s, as well as the representations with one or four 1s, are deleted, there are exactly 20 representations. The specific encoding is shown in Table 2 below:

[0023] Table 2

[0024]

[0025]

[0026] Furthermore, in one-dimensional convolution, the convolution operation can be represented as: Where N is the length of the convolution kernel, x represents the input data, w represents the convolution kernel, and Conv(x, w) represents the output of the convolution operation.

[0027] Furthermore, the Gating mechanism uses the sigmoid activation function to control information flow, which can be expressed by the formula: O = sigmoid(c1) × sigmoid(c2), where c1 and c2 represent the outputs of the previous layer, respectively. By introducing the Gating mechanism, the model can learn the correlations between different channels in the input features, enabling the network to better understand the structural and semantic information of the data. This integration of relevant information helps improve the model's performance, allowing it to better capture patterns and features in the data.

[0028] Furthermore, a Bidirectional Gated Recurrent Unit (BiGRU) consists of two GRUs operating in opposite directions: one scans the sequence from left to right, and the other from right to left. The input sequence of the BiGRU passes through both GRUs, producing two output sequences. These two sequences are then concatenated to obtain a combined output sequence. This combined output sequence better captures long-term dependencies in the input sequence, improving the model's predictive performance. The structural formula of the BiGRU is as follows:

[0029] z t =σ(W z [h t-1 x t ]),

[0030] r t =σ(W r [h t-1 x t ]),

[0031]

[0032]

[0033] Where z t r t , h t These are the outputs for updating the gate, resetting the gate, candidate hidden states, and hidden states, respectively; W z W r W h These are the weight matrices for the update gate, reset gate, and candidate hidden states, respectively; x t h is the input vector at time step t; t-1 It saves the information from the previous time step; σ is the sigmoid function.

[0034] Further, step S6 specifically involves converting the unknown peptide sequence into a feature vector through step 2, then concatenating the obtained feature vectors in step 3, and then feeding them into the deep learning network model trained in step 5. If the output result is 1, then the central S / T of the predicted peptide sequence fragment is phosphorylated, which is a phosphorylation site of SARS-CoV-2 infection; otherwise, the central S / T of the predicted peptide sequence fragment is not phosphorylated, and therefore it is not a phosphorylation site of SARS-CoV-2 infection.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention employs convolutional neural networks, a gating mechanism, and a bidirectional gated recurrent unit network to construct a powerful neural network model for extracting phosphorylation information and semantic features from the central S / T sites of peptide sequence fragments, thereby achieving efficient identification of SARS-CoV-2 infection phosphorylation sites. First, through convolutional neural networks, this invention can efficiently extract key semantic features from large-scale protein sequence data without requiring time and manpower for complex manual feature engineering. Next, the introduction of a gating mechanism allows the model to selectively control the output information of convolutional layers, thereby enhancing the information representation capability. Furthermore, through a bidirectional gated recurrent unit network, this invention effectively processes temporal information in protein sequences, considering the semantic relationships between context and allowing for a more comprehensive understanding of the contextual information of phosphorylation sites. In summary, the deep learning model of this invention fully utilizes the advantages of convolutional neural networks, gating mechanisms, and bidirectional gated recurrent unit networks, achieving not only the synergistic effect of feature extraction, information integration, and temporal processing, but also providing an innovative and accurate method for efficiently identifying phosphorylation sites of SARS-CoV-2 infection. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the present invention;

[0038] Figure 2 This is a performance comparison chart of models using single feature encoding and their combinations;

[0039] Figure 3 This is a diagram showing the impact of the BiGRU layer on model performance in a deep learning model architecture.

[0040] Figure 4 It is a graph showing how convolutional layers and gating layers affect the performance of a deep learning model.

[0041] Figure 5 It is a UMAP plot that can reflect the overall predictive performance of the model. Detailed Implementation

[0042] To clearly illustrate the technical features of this solution, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing specific embodiments only and are not intended to limit the scope of protection of the present invention.

[0043] Example 1

[0044] Please refer to Figure 1 The present invention provides an embodiment of a deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection, comprising the following steps:

[0045] S1: Collect a dataset of known phosphorylation sites in human A549 cells infected with SARS-CoV-2;

[0046] S2: Five feature encoding methods are used to extract the feature representation of peptide sequences, namely one-hot encoding, amino acid replacement scoring matrix, new extended and updated principal attribute scale, 5-bit binary type one, and 5-bit binary type two.

[0047] S3: Concatenate the vectors of the features extracted by these five feature encoding methods;

[0048] S4: A deep learning model is constructed based on convolutional neural networks, the Gating mechanism, and bidirectional gated recurrent units. The deep learning model includes: a convolutional neural network connected in series with three parallel convolutional layers of the same scale, which are respectively connected to a pooling layer, a batch normalization layer, a dropout layer, and a sigmoid activation layer. The activation results are multiplied and flattened. The Gating mechanism controls the output information of the convolutional layers through the sigmoid activation function. The convolutional neural network is also connected in series with a bidirectional gated recurrent unit network to flatten the output feature vector again. The results of the two flattening are concatenated to form a vector containing multiple feature representations. Finally, this vector is passed to a fully connected layer for classification.

[0049] S5: Peptide sequence fragments with phosphorylated S / T sites at the center are defined as positive samples and labeled as 1; otherwise, they are considered negative samples and labeled as 0. The shuffled training set is used to supervise the training of the deep learning neural network model. The encoding vectors of positive and negative samples and their labels are used to train the deep learning neural network model.

[0050] S6: Use a trained model to predict unknown peptide sequences.

[0051] Example 2

[0052] Please refer to Figure 1 The present invention provides an embodiment of a deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection, comprising the following steps:

[0053] S1: Collect a dataset of known phosphorylation sites in human A549 cells infected with SARS-CoV-2.

[0054] We used experimentally validated phosphorylation sites from human A549 cells infected with SARS-CoV-2, a dataset containing 14,119 phosphorylation sites. To avoid data homology and redundancy, the CD-HIT procedure was used to set a sequence homology threshold of 30%, ensuring a high-quality dataset and preventing model overfitting. For ease of comparison with other phosphorylation site prediction methods, we truncated the processed data to sequences of 33 residues; here, we selected samples centered on S / T sites.

[0055] S2: Five feature encoding methods were used to extract feature representations of peptide sequences: one-hot encoding, amino acid substitution scoring matrix (BLOSUM62), new extended and updated principal attribute scale (ZScale), 5-bit binary type 1 (Binary_5bit_type1), and 5-bit binary type 2 (Binary_5bit_type2). Proteins consist of 20 amino acids. One-hot, Binary_5bit_type1, and Binary_5bit_type2 represented amino acids as binary vectors according to different encoding rules. BLOSUM62 used a predefined amino acid substitution scoring matrix that reflected the relative frequency and biological similarity of substitutions between amino acids. Each amino acid in the protein sequence was represented by a vector extracted from a row or column of the BLOSUM62 matrix. ZScale used a set of new extended and updated principal attribute scales, including physicochemical descriptors such as polarity, hydrophobicity, volume, and hydrogen bonding ability, to characterize amino acids.

[0056] S2.1 One-hot

[0057] Given a protein sequence S = {a1, a2, ..., a...} i}, where i represents the length of the sequence, a i This represents the residue at position i. Each residue is represented as a 20-dimensional vector, where only one dimension has a value of 1, representing the amino acid type at that position, and the other dimensions have values ​​of 0.

[0058] S2.2 BLOSUM62

[0059] BLOSUM62 is a statistical method for encoding protein sequences, used to describe the substitution relationships between different amino acids. First, a substitution matrix is ​​constructed based on known protein sequence comparison data. This matrix describes the frequency with which one amino acid is substituted by another in a protein sequence. Then, a similarity score is calculated. For the alignment of two protein sequences, the similarity score can be calculated based on the substitution matrix. The similarity score between different positions of the two protein sequences can be expressed by the following formula: Score = M(s1[i], s2[j]), where M is the BLOSUM62 matrix, M ∈ R. m×dWhere m is the number of amino acid types and d is the dimension of the embedding vector. The score represents the similarity score between position i in protein sequence 1 and position j in protein sequence 2. Finally, the scores are accumulated. For the alignment of the entire protein sequence, the similarity scores between each pair of amino acids are summed to obtain a total similarity score. This score can be used to measure the similarity between two protein sequences. The higher the score, the more similar the two sequences are.

[0060] S2.3ZScale

[0061] This method characterizes each amino acid using five physicochemical descriptors. By using these descriptor variables, the differences and similarities between different amino acids in a protein sequence can be captured, thereby enabling the analysis and prediction of the protein's structure and function.

[0062] S2.4Binary_5bit_type l

[0063] This method uses a set of five-letter amino acid groups {a1, a2, a3, a4, a5}, each represented by a corresponding five-dimensional binary vector. Specifically, a1 ∈ {G, A, V, L, M, I}, a2 ∈ {F, Y, W}, a3 ∈ {K, R, H}, a4 ∈ {D, E}, and a5 ∈ {S, T, C, P, N, Q}. Each group is then represented by a five-dimensional binary vector; for example, a1 is encoded using a binary vector (10000), a2 using a binary vector (01 000), ..., a5 using a binary vector (00001). During encoding, each binary bit corresponds to a specific amino acid; when an amino acid belongs to its group, the corresponding binary bit is set to 1, otherwise it is 0.

[0064] S2.5Binary_5bit_type2

[0065] This encoding scheme is based on all possible ways to combine 1s and 0s in a 5-bit unit. There are 32 possible representations for 20 amino acids. When the representations with no 1s or all 1s, as well as the representations with one or four 1s, are removed, there are exactly 20 representations.

[0066] S3: Concatenate the vectors of the features extracted by these five feature encoding methods;

[0067] S4: A deep learning model is constructed based on convolutional neural networks, the Gating mechanism, and bidirectional gated recurrent units. The deep learning model includes: a convolutional neural network connected in series with three parallel convolutional layers of the same scale, which are respectively connected to a pooling layer, a batch normalization layer, a dropout layer, and a sigmoid activation layer. The activation results are multiplied and flattened. The Gating mechanism controls the output information of the convolutional layers through the sigmoid activation function. The convolutional neural network is also connected in series with a bidirectional gated recurrent unit network to flatten the output feature vector again. The results of the two flattening are concatenated to form a vector containing multiple feature representations. Finally, this vector is passed to a fully connected layer for classification.

[0068] S4.1 In one-dimensional convolution, the convolution operation can be represented as: Where N is the length of the convolution kernel, x represents the input data, w represents the convolution kernel, and Conv(x, w) represents the output of the convolution operation.

[0069] The S4.2 Gating mechanism uses the sigmoid activation function to control information flow, which can be expressed by the formula: O = sigmoid(c1) × sigmoid(c2), where c1 and c2 represent the outputs of the previous layer, respectively. By introducing the gating mechanism, the model can learn the correlations between different channels in the input features, enabling the network to better understand the structural and semantic information of the data. This integration of relevant information helps improve the model's performance, allowing it to better capture patterns and features in the data.

[0070] The S4.3 Bidirectional Gated Recurrent Unit (BiGRU) consists of two GRUs operating in opposite directions: one scans the sequence from left to right, and the other from right to left. The input sequence of the BiGRU passes through both GRUs, producing two output sequences. These two sequences are then concatenated to obtain a combined output sequence. This combined output sequence better captures long-term dependencies in the input sequence, improving the model's predictive performance. The structure of the BiGRU is as follows:

[0071] z t =σ(W z [h t-1 x t ]),

[0072] r t =σ(W r [h t-1 x t ]),

[0073]

[0074]

[0075] Where z t r t , h t These are the outputs for updating the gate, resetting the gate, candidate hidden states, and hidden states, respectively; W z W r W h These are the weight matrices for the update gate, reset gate, and candidate hidden states, respectively; x t h is the input vector at time step t; t-1 It saves the information from the previous time step; σ is the sigmoid function.

[0076] S5: Peptide sequence fragments with phosphorylated S / T sites at the center are defined as positive samples and labeled as 1; otherwise, they are considered negative samples and labeled as 0. The shuffled training set is used to supervise the training of the deep learning neural network model. The encoding vectors of positive and negative samples and their labels are used to train the deep learning neural network model.

[0077] S6: Predict unknown peptide sequences using a trained model.

[0078] After converting the unknown peptide sequence into a feature vector in step 1, the resulting feature vectors are concatenated in step 2 and then fed into the deep learning network model trained in step 4. If the output result is 1, the central S / T of the predicted peptide sequence fragment is phosphorylated, which is a phosphorylation site of SARS-CoV-2 infection. Otherwise, the central S / T of the predicted peptide sequence fragment is not phosphorylated, and therefore it is not a phosphorylation site of SARS-CoV-2 infection.

[0079] Example 3

[0080] Please refer to Figure 1 The present invention provides an embodiment of a deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection, which includes all the steps of embodiment 2 and also includes a performance evaluation method.

[0081] Independent testing was used to evaluate the method proposed in this patent. In the independent testing, all datasets were randomly divided into an 8:2 ratio: 80% training set and 20% test set. The 80% training set was used for model training, and the 20% test set was used for model performance testing. To ensure model reproducibility, a random seed was used in this experiment. To evaluate the model's performance, this experiment used sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC), receiver operating characteristic curve (ROC), and area under the ROC curve (AUC) as evaluation metrics.

[0082] Appendix Figure 2 This document presents a comparison of the model performance of single feature encoding and its combinations in this patent; Appendix Figure 3 This shows a performance comparison of models with and without BiGRU layers, and models where BiGRU layers are replaced with BiLSTM layers; Appendix Figure 4 This illustrates the impact of having or not having one-dimensional convolution Conv1, the kernel size of Conv1, having or not having convolution Conv2 of the same scale, and the kernel size of Conv2 on model performance. (Appendix) Figure 5 The diagram shows the prediction graphs of the method proposed in this patent on the training and test sets, respectively.

[0083] (1) This invention studies the impact of single feature encoding methods and their combinations on model performance. The experimental results are shown in Table 3-6 below (the best results are highlighted in bold). The experimental results show that the ACC of the five feature encoding combinations of onehot, BLOSUM62, Zscale, Binary_5bit_type1 and Binary_5bit_type2 is better than that of any other combination of feature encodings, which means that such feature combinations can enable the model to learn better sequence information.

[0084] Table 3

[0085]

[0086]

[0087] Table 4

[0088]

[0089] Table 5

[0090]

[0091]

[0092] Table 6

[0093]

[0094]

[0095] (2) This invention studies the performance comparison of the model with and without BiGRU layers, as well as the performance comparison when the BiGRU layer is replaced with a similar network, such as a Bidirectional Long Short-Term Memory Network (BiLSTM). The BiGRU layer is mainly able to better capture long-term dependencies in the input sequence. Table 7 below shows the performance comparison with and without BiGRU layers and the performance comparison with BiLSTM. The experimental results show that the BiGRU layer significantly improves the model performance (the best results are highlighted in bold).

[0096] Table 7

[0097]

[0098] This invention investigates the performance comparison of the model with and without the first one-dimensional convolutional layer Conv1, and with and without the second one-dimensional convolutional layer Conv2 of the same scale. Furthermore, to demonstrate the superiority of the proposed method, the impact of the convolution kernel setting on model performance is studied when Conv1 is included, and the impact of different scale convolution kernels on model performance when Conv2 of the same scale is included. Finally, this invention also verifies the performance of the model structure without a gating mechanism. The results are shown in Table 8 below (the best results are highlighted in bold). The proposed method achieves the best performance on the S / T dataset. These results demonstrate the rationality of the proposed method architecture.

[0099] Table 8

[0100]

[0101]

[0102] This invention compares with currently widely used machine learning algorithms, such as logistic regression (LR), decision tree (DT), support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), and lightweight gradient boosting machine (LGBNM). The results of independent tests for identifying phosphorylation sites are shown in Table 9 (the best results are highlighted in bold).

[0103] Table 9

[0104]

[0105] This patent also compares the results with several representative deep learning methods (IPs-GRUAtt, DeepIPs, Adapt-ST, MusiteDeep2017, MusiteDeep2020, DeepP SP). The comparison results are shown in Table 10 below (the best results are highlighted in bold).

[0106] Table 10

[0107]

[0108]

[0109] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection, characterized in that, Includes the following steps: S1: Collect a dataset of known phosphorylation sites in human A549 cells infected with SARS-CoV-2; S2: Feature representations of peptide sequences are extracted using five feature encoding methods: one-hot encoding, amino acid replacement scoring matrix, new and updated principal attribute scales, 5-bit binary type I, and 5-bit binary type II; where: The amino acid substitution scoring matrix method includes: first, constructing a substitution matrix based on known protein sequence alignment data, which describes the frequency with which one amino acid is substituted by another in a protein sequence; then, calculating a similarity score, for an alignment between two protein sequences, calculating a similarity score based on the substitution matrix, using the following formula to represent the similarity score between different positions of the two protein sequences: Where M is a BLOSUM62 matrix, Where m is the number of amino acid types, d is the dimension of the embedding vector, and Score represents the similarity score between position i of protein sequence 1 and position j of protein sequence 2. Finally, the scores are accumulated. For the alignment of the entire protein sequence, the similarity scores between each pair of amino acids are accumulated to obtain a total similarity score. This score can be used to measure the similarity between two protein sequences. The higher the score, the more similar the two sequences are. The new extended and updated main attribute scaling method is to characterize each amino acid with five physicochemical descriptors. By using these descriptor variables, the differences and similarities between different amino acids in the protein sequence are captured, thereby analyzing and predicting the structure and function of the protein. S3: Concatenate the vectors of the features extracted by these five feature encoding methods; S4: A deep learning model is constructed based on a one-dimensional convolutional neural network, the Gating mechanism, and a bidirectional gated recurrent unit. The deep learning model includes: a one-dimensional convolutional neural network connected in series with three parallel one-dimensional convolutions of the same scale, which are respectively connected to a pooling layer, a batch normalization layer, a dropout layer, and a sigmoid activation layer. The activation results are multiplied and flattened. The Gating mechanism controls the output information of the convolutional layers through the sigmoid activation function. The convolutional neural network is also connected in series with a bidirectional gated recurrent unit network to flatten the output feature vector again. The results of the two flattenings are concatenated to form a vector containing multiple feature representations. Finally, this vector is passed to a fully connected layer for classification. The Gating mechanism uses the sigmoid activation function to control the information flow, and the formula is expressed as: ,in and These represent the outputs of the previous layer; S5: Peptide sequence fragments with phosphorylated S / T sites at the center are defined as positive samples and labeled as 1; otherwise, they are considered negative samples and labeled as 0. The deep learning neural network model is trained in a supervised manner using the shuffled training set, and the deep learning neural network model is trained using the encoding vectors of positive and negative samples and their labels. S6: Use a trained model to predict unknown peptide sequences.

2. The deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection according to claim 1, wherein, The one-hot encoding method in step S2 is to give a protein sequence wherein denotes the length of the sequence, denotes the position of the residue, each residue is represented as a 20-dimensional vector, where only one dimension has a value of 1, representing the amino acid type at that position, and the other dimensions have a value of 0.

3. The deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection according to claim 1, wherein, The method of step S2 is to use a set of 5-letter amino acid groups For each amino acid group, there is a corresponding 5-dimensional binary vector representation, then each group is represented by a 5-dimensional binary vector, Encoded using binary vector 10000, Encoded using binary vector 01000,... Encoded using binary vector 00001, in the encoding process, each binary bit corresponds to a specific amino acid, when the amino acid belongs to the corresponding group, the corresponding binary bit is set to 1, otherwise 0.

4. The deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection according to claim 1, wherein, The 5-bit binary type 2 method in step S2 is based on all possible ways to combine 1 and 0 in a 5-bit unit. There are 32 possible representations for 20 amino acids. When the representation with no 1 or all 1s and the representation with one or four 1s are deleted, there are exactly 20 representations.

5. The deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection according to claim 1, wherein, In the one-dimensional convolution in the step S4, the convolution operation can be expressed as: where N is the length of the convolution kernel, x represents the input data, w represents the convolution kernel, represents the output result of the convolution operation.

6. The deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection according to claim 1, wherein, In step S4, the bidirectional gated loop unit consists of two gated loop units (GRUs) operating in opposite directions. One scans the sequence from left to right, and the other scans from right to left. The input sequence of the bidirectional gated loop unit passes through the GRUs in both directions, generating two output sequences. These two sequences are then concatenated to obtain a comprehensive output sequence. The structural formula of the bidirectional gated loop unit is as follows: ， where , , , are the update gate, the reset gate, the candidate hidden state, the output of the hidden state, respectively; , , are the weight matrices of the update gate, the reset gate, the candidate hidden state, respectively; is the input vector at the t-th time step; holds the information from the previous time step; is the sigmoid function.

7. The deep learning method for identifying phosphorylation sites of SARS-CoV-2 infection according to claim 1, wherein, Specifically, step S6 in step S4 involves converting the unknown peptide sequence into a feature vector in step 2, then concatenating the resulting feature vectors in step 3, and finally feeding them into the deep learning network model trained in step 5. If the output result is 1, then the central S / T of the predicted peptide sequence fragment is phosphorylated, which is a phosphorylation site for SARS-CoV-2 infection; otherwise, the central S / T of the predicted peptide sequence fragment is not phosphorylated, and therefore it is not a phosphorylation site for SARS-CoV-2 infection.

Images