Method for identifying various lysine modification sites based on EPDCC technology

Abstract

The invention belongs to the field of artificial intelligence algorithm application-biological sequence recognition, and relates to a method for recognizing various lysine modification sites based on an EPDCC technology. According to the invention, multiple lysine modification sites can be predicted at the same time. According to the method, the problem of crosstalk between modifications which is not fully considered in the prior art is particularly solved, so that the accuracy and the reliability of prediction are improved. Discrete data is fitted into function type data for clustering to balance a data set. The extreme point deviation compensation clustering under-sampling algorithm adopted by the invention effectively reduces the deviation of a model training process to majority classes. By constructing multi-scale data and combining sequence information characteristics and amino acid physicochemical property characteristics, data semantics are enriched, and the prediction performance of the model on lysine post-translational modification sites is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence algorithm application-biological sequence recognition, and relates to a method for identifying multiple lysine modification sites based on EPDCC technology. Background Art

[0002] Post-translational modifications (PTMs) on protein lysine significantly change the physical and chemical properties of proteins by forming covalent bonds or binding chemical groups. These modifications have a profound impact on the structure, function and stability of proteins, and play a key role in various biological processes such as gene expression regulation, signal transduction, protein degradation and cell metabolism. Therefore, identifying and predicting various lysine post-translational modification sites is of great significance for a deeper understanding of biological processes and promoting drug development.

[0003] Post-translational modifications of lysine include acetylation, methylation, succinylation and other forms. These modifications usually occur through enzymatic reactions and regulate many important cellular activities. However, traditional experimental methods, such as mass spectrometry and specific antibody detection, although they can provide more accurate modification site information, are costly, complex and time-consuming. In addition, these methods are inefficient in processing large-scale data and are difficult to meet the needs of modern biological research for high-throughput and rapid prediction.

[0004] With the development of bioinformatics and artificial intelligence technologies, more and more computational tools are being used to identify post-translational modification sites on amino acids such as lysine. Compared with traditional experimental methods, predictive recognition models based on machine learning and deep learning have higher efficiency and lower cost, and can quickly analyze a large number of protein sequences and identify potential modification sites. However, existing prediction models still face multiple challenges, such as unbalanced sample data distribution, insufficient consideration of interactions (crosstalk) between different modification types, and limited ability of the model to capture complex dependencies of sequence information.

[0005] Therefore, developing a prediction tool that can comprehensively consider the interactions between multiple lysine modification types, effectively deal with data imbalance problems, and accurately capture sequence dependencies is of great significance for the study of protein post-translational modifications and the understanding and treatment development of related diseases. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a method for simultaneously predicting four different types of lysine post-translational modification sites. The method uses a method based on a cross-scale convolutional neural network (CSCNN) and an extreme point deviation compensation clustering undersampling algorithm (EPDCC) to identify multiple lysine modification sites.

[0007] The technical solution of the present invention is as follows:

[0008] The data used in the present invention is derived from the CPLM4.0 database, which focuses on recording various post-translational modifications (PTMs) of the side chain amino groups of lysine residues in proteins at specific positions. The present invention uses the 18,978 human protein sequences provided by the database and constructs a data set suitable for model training and verification through a carefully designed data processing process. Although there is a significant imbalance problem in the data set, with a ratio of more than 100:1, the present invention effectively responds to this challenge by proposing an innovative method of extreme point deviation compensation clustering undersampling technology, which improves the generalization ability of the model under imbalanced data, thereby improving the accuracy and reliability of the prediction. In addition, the present invention adopts two cross-scale feature extraction algorithms, and combines the proposed cross-scale convolutional neural network for classification and recognition, and finally develops a user-friendly web prediction tool.

[0009] The method for identifying various lysine modification sites based on EPDCC technology is as follows:

[0010] Step 1: Dataset Construction

[0011] In order to develop statistical prediction models, it is essential to build reliable training and test datasets. The specific steps are as follows:

[0012] 1.1 Sequence extraction and verification: For the 18,978 human protein sequences collected, peptide fragments whose lysine (K) was modified by "acetyllysine" or "crotonyllysine" or "methyllysine" or "succinyllysine" were extracted (centered on K, with a sliding window size of 24, i.e., a peptide fragment length of 49). Each peptide fragment can be expressed in the form of formula (1):

[0013] P=R -24 R -23 …R -2 R -1 R1R2…R 23 R 24 (1)

[0014] 1.2 Data classification: After preprocessing, the lysine modification data can be classified into 15 categories in total. After removing the categories with less than 60 sequences, the data set is finally divided into 11 categories:

[0015]

[0016] Data category description: Refers to protein sequences containing only acetylated lysine residues at the central position. Refers to protein sequences that contain both acetylation and crotonylation modifications. After removing the repeated peptide fragments in the data, 11 types of data were obtained, and the number of samples for each type is as follows: As mentioned above, the number of samples in the first category is 39938, the number of samples in the second category is 2463, and so on. In order to ensure that the performance evaluation is not biased high due to sequence redundancy and homology when evaluating the multi-label prediction model, the present invention uses the CD-HIT program and sets the threshold to 0.4 to eliminate these potential homologous sequences and redundant samples. For the 11 types of data obtained, 70% of each class is randomly selected as training data, and the remaining 30% is used as test data. The segmented data set after de-redundancy is obtained as follows:

[0017] Training set: Category (1): 9279 items; Category (2): 710 items; Category (3): 600 items; Category (4): 454 items; Category (5): 561 items; Category (6): 252 items; Category (7): 360 items; Category (8): 88 items; Category (9): 153 items; Category (10): 454 items; Category (11): 73 items.

[0018] Independent test set: Category (1): 4062 items; Category (2): 304 items; Category (3): 257 items; Category (4): 194 items; Category (5): 240 items; Category (6): 107 items; Category (7): 154 items; Category (8): 42 items; Category (9): 73 items; Category (10): 191 items; Category (11): 36 items. These steps ensure the comprehensiveness and accuracy of the data, providing a solid foundation for subsequent model training and verification.

[0019] Step 2: Feature Extraction

[0020] In order to allow the computer to recognize the data cleaned in the first stage, the data is feature extracted in this stage (the data is expressed in mathematical form). The present invention extracts features from the two dimensions of lysine sequence information and amino acid physicochemical properties.

[0021] 2.1 The steps for extracting lysine sequence information features are as follows:

[0022] 2.1.1: Remove the central lysine (K) from the sequence obtained in the first step. Then calculate the frequency matrix F of each triple amino acid at each position in the lysine sequence of the tth class t , the matrix size is 20 3 ×46.

[0023]

[0024] where F t (TAA i / j) indicates the first TAA in the t-th lysine sequence i The frequency of the occurrence of the triad of amino acids at the jth position in the i ∈{AAA,AAC,AAD,…YYY}, that is, TAA1=AAA, TAA2=AAC,…, i=1,2,3,…,20 3 , j=1,2,3,…,46.

[0025] 2.1.2: Calculate the frequency matrix FF of each triple amino acid at each position in the ten lysine sequences except the kth class t , the matrix size is 20 3 ×46.

[0026]

[0027] where F k (TAA i / j) indicates the other ten categories except category k TAA in the lysine sequence i The frequency of occurrence of the triplet amino acid at the jth position in the sequence.

[0028] 2.1.3: The 11 Fs obtained t The Matrix and 11 FFs k The matrix F, FF is obtained by averaging the matrices, and the matrix F is obtained by subtracting them. r ,Right now

[0029]

[0030] where f i,j =F t (TAA i / j)-FF t (TAA i / j),F=(F1+F2+…+F 11 ) / 11,FF=(FF1+FF2+…+FF 11 ) / 11.

[0031] 2.1.4: According to the type of triple amino acid at each position in each sequence, find the corresponding value in the Fr matrix for feature encoding. For example: for the "AGAT...TGLR" sequence with a length of 48, use a sliding window of size 3 to continuously perform encoding. First, the position of the three amino acid fragments in the sequence must be determined. For example, "AGA" is at the first position of the sequence, and "GAT" is at the second position. Then, according to the amino acid alphabet 'ACDEFGHIKLMNPQRSTVWY' (row position = position 1×20 2 +Position 2×20 1 + Position 3×20 0 +1) Calculate the row position of each triamino acid in the matrix. In this alphabet, A, G, and T are at positions 0, 5, and 16, respectively. Therefore, the row position of "AGA" is determined by the formula: 0×20 2 +5×20 1 +0×20 0 +1=101; Similarly, the row position of "GAT" is calculated as: 5×20 2 +0×20 1 +16×20 0 +1=2017. Finally, the sequence is feature encoded and the first position of the feature vector is: f 101,1 , the second position is: f 2017,2 , and so on, we end up with a feature vector of length 46.

[0032] 2.2 The physical and chemical properties of amino acids are extracted as follows:

[0033] The molecular weight, isoelectric point, and hydrophilicity and hydrophobicity of amino acids have an important influence on protein structure, which in turn affects the type of protein post-translational modification sites. Therefore, the present invention is based on these three amino acid physicochemical properties as a feature extraction method. The following table lists the physicochemical properties of 20 natural amino acids:

[0034]

[0035] The present invention encodes each amino acid in each protein sequence. The hydrophilic amino acids in the table are marked as 1, and the hydrophobic amino acids are marked as -1. Each amino acid is converted into a vector with three values, for example, the vector of glycine (G) is (75.07, 5.97, -1), and the vector of alanine (A) is (89.09, 6.01, 1). In this way, each protein sequence will eventually be converted into a matrix with 49 rows and 3 columns, in which each row represents a specific code for an amino acid.

[0036] Step 3: Data imbalance processing

[0037] After the data feature extraction is completed, data of two dimensions are obtained: (1) lysine sequence information features with a dimension of 46. (2) Amino acid physicochemical properties features with dimensions of (49, 3). In this stage, in order to make the training data category samples tend to be balanced as a whole, based on the 46-dimensional lysine sequence information features, the present invention provides an extreme point deviation compensation clustering undersampling algorithm (EPDCC) to solve the sample imbalance problem, which is specifically described as follows:

[0038] 3.1 Function fitting:

[0039] In order to capture the complex structure and potential dynamic characteristics of the data itself, the present invention first fits the 46-dimensional lysine sequence information features into continuous function features. The specific steps are as follows:

[0040] 3.1.1 Dimensionless processing: In order to eliminate the order of magnitude differences between data, the present invention uses a normalization method to perform dimensionless processing on the 46-dimensional data, so that data of different dimensions can be compared on the same scale.

[0041] 3.1.2 Select the fitting basis function. Function fitting reconstructs the discrete data to obtain a continuous smooth curve, that is, the 46-dimensional discrete feature value [y i1 ,y i2 ,…,y i46 ] to get the function curve x i (t), its basic form is shown in formula (6):

[0042] y ij =x i (t j )+ε(t j ) (6)

[0043] Among them, t j To mark the horizontal coordinate corresponding to the current position (the present invention uses 45 equidistant points to divide the [0,1] interval, and the coordinate value of each point is used as the horizontal coordinate of the corresponding data point), j = 1, 2, ..., 46, ε (t j ) is the error term. In order to better capture the dynamic characteristics of the data, in the data cluster analysis, the present invention uses the B-spline basis function suitable for non-periodic data to fit the data.

[0044] 3.1.3 Select data nodes: Select appropriate nodes according to the distribution of data points. The selection of nodes can be done by selecting them at equal intervals or adaptively selecting them according to the density of data points. According to the research results of the present invention, the present invention selects 10 nodes uniformly generated between the minimum and maximum values ​​of the data horizontal coordinate using the method of selecting them at equal intervals as the data stage.

[0045] 3.1.4 Construct the k-order B-spline basis function as shown in formula (7):

[0046]

[0047] Among them, c ik is the coefficient of the k-order term to be estimated, φ k (t) is the k-order B-spline basis function. In order to make the fitting function smooth and accurate, it is necessary to select an appropriate order. According to the characteristics of the data of the present invention and the research results, the present invention finally determines that the order of the basis function is 3.

[0048] 3.1.5 Estimate the parameter c by minimizing the objective function shown in equation (8) using the least squares method ik :

[0049]

[0050] Where λ is the smoothing factor, is x i By solving the above optimization problem, we can obtain the parameter c ik , and then get the fitting function curve x i (t).

[0051] 3.2 Cluster analysis distance metric:

[0052] After fitting the data into function data, the present invention uses an improved K-means++ algorithm for cluster analysis. In order to clarify the similarities and differences between samples, the present invention incorporates the idea of ​​penalizing extreme point deviations into the distance measurement, which can not only measure the absolute horizontal difference of the curve, but also capture the morphological difference between the curves, thereby improving the effect of cluster analysis. i (t) and x j The distance steps of (t) are as follows:

[0053] 3.2.1 First, x i (t) and x j (t) Calculate its first-order derivative, and then find the point where the derivative is zero, that is, the extreme point. Suppose the extreme point set is E i ={t i1 ,t i2 ,…,t im} and E j ={t j1 ,t j2 ,…,t jn}.

[0054] 3.2.2 Extreme point set interpolation. If E i and E jThe sizes of the two extreme point sets are different. The interpolation method is used to make the sizes of the two extreme point sets equal to obtain the adjusted extreme point set. and

[0055] 3.2.3 Calculation of extreme point deviation distance. Extreme point deviation distance D ext It is defined as the form of formula (9):

[0056]

[0057] 3.2.4 Calculate two functions x i (t) and x j The Euclidean distance of (t) is calculated as shown in formula (10):

[0058]

[0059] 3.2.5 Total distance calculation, final similarity measure D total Combines the numerical distance and the extreme point deviation distance:

[0060] D total =D num +D ext (11)

[0061] The measurement method based on the deviation compensation similarity of the curve extreme points not only considers the difference in the function value itself, but also combines the deviation of the extreme point position, which improves the accuracy and reliability of the similarity measurement.

[0062] 3.3 Cluster analysis

[0063] After fitting the function to the data, the present invention uses an improved K-means++ clustering analysis method to cluster the data. The basic idea of ​​the improved method is: first calculate the D total Distance, select the two samples with the largest distance as the initial cluster centers. Then, assuming that n cluster centers have been selected, when selecting the n+1th cluster center, select the sample with the largest sum of distances from the current n samples, and recursively use this method to obtain K (number of categories) samples as the initial cluster centers. Compared with K-means++ randomly selecting any point, this improvement completely eliminates randomness and speeds up clustering. The samples selected after undersampling are more able to replace the characteristics of the original samples.

[0064] The pseudo code of the extreme point deviation compensation clustering undersampling algorithm (EPDCC) is as follows:

[0065]

[0066] 3.4 Confirm the sampling number and cluster number of the first type of data

[0067] Since the ratio of the 11 categories of training data in the training set of the present invention is 9279:710:600:454:561:252:360:88:153:454:73, it is obvious that the number of the first category is much larger than the other ten categories, accounting for about 71% of the total samples, so only the first category of data is undersampled by EPDCC. The present invention is verified by experiments, and the number of clusters is selected as 4, and the final undersampling ratio is 0.1, that is, the number of samples is 928. After the sampling is successfully completed, the 11 categories in the training set are selected, and the sample distribution ratio is: 928:710:600:454:561:252:360:88:153:454:73.

[0068] 3.5 The sampled first category data and the original 2-11 category data are combined into a new training data set.

[0069] Step 4: Multi-label cross-scale convolutional neural network classification model

[0070] In this stage, the present invention constructs a multi-label cross-scale convolutional neural network training based on the training set obtained in the fourth step, and tests it on an independent test set. The specific steps of this stage are as follows:

[0071] 4.1 Construction of multi-label cross-scale convolutional neural network classification model

[0072] The multi-label cross-scale convolutional neural network (CSCNN) proposed in this paper uses convolutional neural network (CNN) to extract the hidden relationship between labels, and enriches the dimension of data analysis by fusing features and weights of different scales, thereby achieving better classification results. The learning architecture mainly consists of the following parts:

[0073] (1) Unified aligned multi-source feature representation: By developing a unified aligned multi-source feature representation, the dimensional homogeneity of lysine sequence information features and amino acid physicochemical property features is ensured, thereby generating potential feature representations.

[0074] (2) Construct a cross-scale feature and weight extractor to fully extract the deep information of the features and generate feature weights in the process. The two extractors contain CNN modules with the same kernel size to understand the contextual dependencies under different features, so as to make accurate interaction predictions. The extractor generates feature weights based on mid-level features, which prevents the overfitting problem caused by excessive feature extraction and improves the accuracy of subsequent label prediction.

[0075] (3) Feature fusion and prediction: The model fuses the learned potential information at different scales and combines the weights to obtain the prediction results to study the association between sequence, amino acid physicochemical properties and lysine post-translational modification sites. After integrating all the extracted features and normalizing them, the network output is obtained by multiplying the weights. The network output uses the Sigmoid activation function and sets the threshold to 0.5 to predict lysine post-translational modification sites. This process reduces the model's bias towards majority class samples and can display the scores for predicting different modification sites in an intuitive way.

[0076] 4.2 Model training and validation testing

[0077] After the model is built, the model is trained based on the training set obtained in the third step. After obtaining a mature model, it is tested on an independent test set. The model of the present invention predicts the lysine modification sites in the protein sample by outputting a matrix containing four elements, each element representing the probability of a modification type. The cross entropy loss function is used to train the model, and the Adam optimizer is used, and the learning rate is set to 0.001. After the model is trained, the optimal model is selected by five-fold cross validation and five evaluation indicators (Aiming, Coverage, Accuracy, Absolute_True, Absolute_False). Finally, an evaluation is performed on an independent test set to verify the prediction effect of the model.

[0078] Beneficial effects of the present invention:

[0079] 1. The present invention breaks through the limitations of the prior art, which usually only focuses on a single type of protein post-translational modification (PTM) and ignores the possible interactions between different PTMs. In contrast, the present invention develops a multi-tag prediction tool that can simultaneously predict multiple lysine modification sites. In addition, the present invention also specifically solves the crosstalk problem between modifications that the prior art has not fully considered, thereby improving the accuracy and reliability of the prediction.

[0080] 2. The data set is balanced by fitting discrete data into functional data for clustering. The extreme point deviation compensation clustering undersampling algorithm (EPDCC) adopted in the present invention effectively reduces the bias towards the majority class in the model training process.

[0081] 3. By constructing multi-scale data and combining sequence information characteristics with the physical and chemical properties of amino acids, the data semantics is enriched and the model's prediction performance for lysine post-translational modification sites is significantly improved.

[0082] 4. In order to facilitate researchers to use the present invention to improve experimental efficiency, a free and available web site was developed: http: / / www.mlypredcsed.com. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is the overall technical structure flow chart of the present invention.

[0084] Figure 2 It is a multi-label cross-scale convolutional neural network diagram in the present invention.

[0085] Figure 3 It is a comparison chart of indicators corresponding to each clustering number in the EPDCC algorithm of the present invention.

[0086] FIG. 4( a ) is a graph showing the comparison of the indicators of the EPDCC algorithm in the present invention and other similar algorithms.

[0087] FIG. 4( b ) is a graph showing the prediction accuracy of the EPDCC algorithm of the present invention and other similar algorithms.

[0088] Figure 5 It is a comparison chart between the method MlyPredCSED proposed in the present invention and other methods, the purpose of which is to verify the effectiveness of the proposed method. DETAILED DESCRIPTION

[0089] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0090] The present invention proposes a method for identifying multiple lysine modification sites based on EPDCC technology, which is used to predict a multi-label classification model of eleven types of lysine modification sites in protein sequences. Figure 1 As shown. The model first uses lysine sequence information and the physicochemical properties of three amino acids for feature encoding. Then, the extreme point deviation compensation clustering undersampling algorithm (EPDCC) is used to undersample the majority class samples (the first class) to alleviate the class imbalance problem and improve the model's generalization ability for classes with fewer samples. Furthermore, the present invention constructs a cross-scale feature and weight extractor by fusing the two features, and then designs a multi-label cross-scale convolutional neural network architecture, such as Figure 2As shown. During the model training process, a multi-label binary cross entropy loss function was used, and the network parameters were optimized by the back propagation algorithm to improve the model's prediction accuracy for a variety of lysine modification sites. The present invention also comprehensively evaluated the model performance using six evaluation indicators, namely, targeting rate, coverage rate, accuracy rate, absolute accuracy rate, absolute false rate, and j-level matching rate, through five-fold cross validation and independent test experiments. The experimental results show that the method MlyPredCSED proposed in the present invention is significantly superior to the existing iMul-kSite, predML-Site, and PreMLS models in the prediction tasks of a variety of lysine modification sites. In addition, in order to facilitate users to use the multi-label prediction model constructed by the present invention, a publicly available web prediction tool has been developed.

[0091] In summary, the multi-label lysine modification site prediction model of the present invention realizes efficient identification of multiple lysine modification sites in protein sequences by adopting the extreme point deviation compensation clustering undersampling algorithm (EPDCC) and the cross-scale convolutional neural network architecture. The model solves the problem of class imbalance and significantly improves the accuracy of prediction. In addition, the present invention also verifies the superior performance of the model through comprehensive testing of a series of evaluation indicators, ensuring its reliability and effectiveness in practical applications. The present invention also develops a user-friendly Web prediction tool, which makes the prediction model easier to obtain and use, and provides a powerful scientific tool for the fields of biological research and drug development.

[0092] Example 1: Data preprocessing and feature extraction

[0093] In this embodiment, multiple data sets for lysine modification site prediction were constructed. First, 18,978 sequences were finely intercepted, classified and optimized to screen out 11 different categories of data sets. In order to ensure the quality of the data set, the CD-HIT program was used to remove redundant samples, thereby improving the purity of the data set.

[0094] Next, we extracted features of lysine sequence information and physicochemical properties of amino acids. This step is crucial because it not only ensures the diversity of the data set, but also ensures the reliability of the data. These feature extractions lay a solid foundation for subsequent model training and evaluation.

[0095] Through this series of data preprocessing and feature extraction steps, high-quality input data can be provided for the lysine modification site prediction model, thereby improving the prediction accuracy and reliability of the model.

[0096] Example 2: Sample balancing technique

[0097] In bioinformatics and protein sequence analysis, the EPDCC undersampling algorithm is used to solve the problem of unbalanced data sets. This method selects representative samples from the majority class samples by fitting a function and then performing cluster analysis on the function data, thereby alleviating the problem of sample imbalance while maintaining data diversity and representativeness.

[0098] In order to determine the undersampling ratio and the number of clusters, a comparative experiment was conducted. Figure 3 Comparing the results of 5 indicators under the independent test set with the number of clusters from 1 to 9, the results show that the method has the best effect when the number of clusters is 4. Table 1 shows the comparison results of indicators on the independent test level under different undersampling ratios. As shown in the table, when the undersampling ratio is 0.1, the generalization performance of the model for 4 types of samples including only acetyllysine modification, only crotonyllysine modification, only methyllysine modification and acetyllysine and methyllysine modification is significantly better than other ratios. In addition, when the undersampling ratio is 0.1, the model shows a certain degree of generalization in all categories, avoiding obtaining an invalid model that only has generalization ability for several categories. Therefore, the present invention selects 0.1 as the final undersampling ratio.

[0099] Table 1. Absolute accuracy of eleven types of samples under different undersampling rates for independent testing (%)

[0100]

[0101] In order to further verify the effectiveness of the EPDCC method, a comparative experiment was conducted, as shown in Figure 4. Figure 4 (a) compares the performance of the model in the multi-label prediction task without undersampling, random undersampling, undersampling using ClusterCentroids, and using the EPDCC algorithm. The experimental results show that the performance of the EPDCC algorithm in five indicators is better than the other three. Figure 4 (b) compares the prediction accuracy of each method on 11 categories of data. The results show that the prediction accuracy of the EPDCC sampling algorithm is better than other methods in all categories except the 5th, 9th, 10th and 11th categories, and each category achieves completely correct predictions. These results fully verify the effectiveness of the sampling method EPDCC proposed in the present invention.

[0102] Example 3: Model performance evaluation and comparison

[0103] like Figure 5As shown, the model MlyPredCSED constructed by the present invention is compared with predML-Site, iMul-kSite and PreMLS on an independent test data set. The existing prediction tools iMul-kSite and predML-Site both process the original multi-label prediction problem by decomposing it into multiple single-label binary classification problems. Specifically, these tools divide the data set into positive and negative samples containing the label and excluding the label for each label, and then use an improved algorithm based on support vector machine (SVM) for binary classification, and finally integrate the various binary classification models to construct a multi-label prediction model. PreMLS uses a deep learning convolutional neural network to directly construct a multi-label prediction tool.

[0104] The experimental results show that MlyPredCSED improves Aiming and Absolute-True by about 10%, Accuracy and Coverage by about 7% and 4% respectively, and reduces Absolute-False by about 6%. Obviously, the overall performance of MlyPredCSED on the test set is better than the above three existing models.

[0105] To further verify the effectiveness of MlyPredCSED, as shown in Table 2, MlyPredCSED is also compared with the above three models in the four indicators of MR1, MR2, MR3, and MR4 on the independent test set. The results show that MlyPredCSED outperforms other models in MR1 and MR2, and is about 40% and 50% higher than the model with the worst results. The performance on the MR4 indicator is far worse than that of the PreMLS model, because PreMLS has a 100% prediction accuracy for the category labeled (1,1,1,1), while this category accounts for a very small proportion of the entire data set. This fully demonstrates that MlyPredCSED outperforms the integrated binary classification-based method and the PreMLS model in performance.

[0106] Table 2. Absolute accuracy of eleven types of samples under different undersampling rates for independent testing (%)

[0107]

Claims

1. A method for identifying multiple lysine modification sites based on EPDCC technology, characterized in that: Here are the steps: Step 1: Dataset Construction Construct training and testing datasets; Step 2: Feature Extraction In this stage, the data is feature extracted from two dimensions: lysine sequence information and amino acid physicochemical properties. Step 3: Data imbalance processing After the data features are extracted, data of two dimensions are obtained: (1) Obtaining lysine sequence information features with a dimension of 46; (2) Physical and chemical properties of amino acids with a dimension of (49, 3); In this stage, in order to make the training data category samples tend to be balanced as a whole, based on the 46-dimensional lysine sequence information features, the specific operations are as follows: 3.1 Function fitting: First, the 46-dimensional lysine sequence information features are fitted into continuous function features. The specific steps are as follows: 3.1.1 Dimensionless processing: The 46-dimensional data is dimensionless processed by normalization method, so that data of different dimensions can be compared on the same scale; 3.1.2 Select the fitting basis function; function fitting reconstructs the discrete data to obtain a continuous smooth curve, that is, the 46-dimensional discrete eigenvalues ​​[y i1 ,y i2 ,…,y i46 ] to get the function curve x i (t), its basic form is shown in formula (6): y ij =x i (t j )+ε(t j ) (6) in, t j is the horizontal coordinate corresponding to the current position, j = 1, 2, · · ·, 46, ε (t j ) is the error term; 3.1.3 Select data nodes: Select appropriate nodes according to the distribution of data points; nodes are selected by selecting them at equal intervals or adaptively selecting them according to the density of data points; 3.1.4 Construct the k-order B-spline basis function as shown in formula (7): in, c ik is the coefficient of the k-order term to be estimated, φ k (t) is the k-order B-spline basis function; the order of the basis function is 3; 3.1.5 Estimate the parameter c by minimizing the objective function shown in equation (8) using the least squares method ik : in, λ is the smoothing factor, is x i (t) is the p-order derivative; by solving the above optimization problem, we obtain the parameter c ik , and then get the fitting function curve x i (t); 3.2 Cluster analysis distance metric: Calculate two function data x in cluster analysis i (t) and x j The distance steps of (t) are as follows: 3.2.1 First, x i (t) and x j (t) Calculate its first-order derivative, and then find the point where the derivative is zero, that is, the extreme point; let the extreme point set be E i ={t i1 ,t i2 ,…,t im } and E j ={t j1 ,t j2 ,…,t jn }; 3.2.2 Extreme point set interpolation; if E i and E j The sizes of the two extreme point sets are different. The sizes of the two extreme point sets are made equal through interpolation to obtain the adjusted extreme point set. and 3.2.3 Calculation of extreme point deviation distance; Extreme point deviation distance D ext It is defined as the form of formula (9): 3.2.4 Calculate two functions x i (t) and x j The Euclidean distance of (t) is calculated as shown in formula (10): 3.2.5 Total distance calculation, final similarity measure D total Combines the numerical distance and the extreme point deviation distance: D total =D num +D ext (11) 3.3 Cluster analysis Calculate the D between all samples total Distance, select the two samples with the largest distance as the initial cluster centers; Then, assuming that n cluster centers have been selected, when selecting the n+1th cluster center, select the sample with the largest sum of distances from the current n samples, and recursively obtain K (number of categories) samples as the initial cluster center in this way; 3.4 Confirm the sampling number and cluster number of the first type of data; 3.5 The sampled first category data and the original 2-11 category data are combined into a new training data set; Step 4: Multi-label cross-scale convolutional neural network classification model A multi-label cross-scale convolutional neural network training is constructed based on the obtained training set pairs and tested on an independent test set.

2. The method for identifying multiple lysine modification sites based on EPDCC technology according to claim 1, characterized in that the first step: data set construction, the specific steps of construction are as follows: 1.1 Sequence extraction and verification: For the 18,978 human protein sequences collected, peptide fragments with experimental verification that lysine (K) is modified by "acetyllysine" or "crotonyllysine" or "methyllysine" or "succinyllysine" were extracted. The length of the peptide fragment is 49, and each peptide fragment is expressed in the form of formula (1): P=R -24 R -23 …R -2 R -1 R1R2…R 23 R 24 (1) 1.2 Data classification: After preprocessing, the lysine modification data can be classified into 15 categories in total. After removing the categories with less than 60 sequences, the data set is finally divided into 11 categories: Data category description: Refers to protein sequences in which the central lysine contains only acetylation modifications; Refers to protein sequences that contain both acetylation and crotonylation modifications; repeated peptide fragments in the data were removed, resulting in 11 types of data, each with the following sample numbers 3. The method for identifying multiple lysine modification sites based on EPDCC technology according to claim 1 or 2, characterized in that the second step: feature extraction, is specifically performed as follows: 2.1 The steps for extracting lysine sequence information features are as follows: 2.1.1: Remove the central lysine (K) from the sequence obtained in the first step; then calculate the frequency matrix F of each triplet amino acid at each position in the lysine sequence of the tth class t , the matrix size is 20 3 ×46; where F t (TAA i / j) indicates the first TAA in the t-th lysine sequence i The frequency of the occurrence of the triad of amino acids at the jth position; TAA i ∈{AAA,AAC,AAD,…YYY}, that is, TAA1=AAA, TAA2=AAC,…, i=1,2,3,…,20 3 , j = 1, 2, 3, ..., 46; 2.1.2: Calculate the frequency matrix FF of each triple amino acid at each position in the ten lysine sequences except the kth class t , the matrix size is 20 3 ×46; where F k (TAA i / j) indicates the other ten categories except category k TAA in the lysine sequence i The frequency of occurrence of the amino acid triad at the jth position; 2.1.3: The 11 Fs obtained t The Matrix and 11 FFs k The matrix F and FF are obtained by averaging the matrices, and the matrix F is obtained by subtracting them. r ,Right now Part f i,j = F t (TAA i / j)-FF t (TAA i / j), F = (F1 + f2 + ... + f 11 ) / 11, FF=(FF1+FF2+…+FF 11 ) / 11; 2.1.4: According to the type of triplet amino acid at each position in each sequence, find the corresponding value in the Fr matrix for feature encoding; 2.2 The physical and chemical properties of amino acids are extracted as follows: Each amino acid in each protein sequence is encoded; the hydrophilic amino acids in the table are marked as 1, and the hydrophobic amino acids are marked as -1; each amino acid is converted into a vector with three values, and each protein sequence is finally converted into a matrix with 49 rows and 3 columns, where each row represents the specific code of an amino acid.

4. The method for identifying multiple lysine modification sites based on EPDCC technology according to claim 1 or 2, characterized in that the fourth step: the specific steps are as follows: 4.1 Construction of multi-label cross-scale convolutional neural network classification model The proposed multi-label cross-scale convolutional neural network Cross-Scale Convolutional Neural Networks (CSCNN) uses convolutional neural network (CNN) to extract the hidden relationship between labels, which consists of the following parts: (1) Unified aligned multi-source feature representation: By developing a unified aligned multi-source feature representation, the homogeneity of lysine sequence information features and amino acid physicochemical property features in terms of dimension is ensured, thereby generating potential feature representations; (2) Construct a cross-scale feature and weight extractor to fully extract the deep information of the features and generate feature weights in the process; the two extractors contain CNN modules with the same kernel size to understand the contextual dependencies under different features, so as to make accurate interaction predictions; the extractor generates feature weights based on the mid-level features, which prevents the overfitting problem caused by excessive feature extraction and improves the accuracy of subsequent label predictions; (3) Feature fusion and prediction: The model fuses the learned potential information at different scales and combines the weights to obtain prediction results to study the association between sequence, amino acid physicochemical properties and lysine post-translational modification sites. After integrating all extracted features and normalizing them, the network output is obtained by weight multiplication. The network output uses the Sigmoid activation function and sets the threshold to 0.5 to predict lysine post-translational modification sites. 4.2 Model training and validation testing After the model is built, it is trained based on the training set obtained in the third step. After the mature model is obtained, it is tested on an independent test set. The lysine modification sites in the protein sample are predicted by outputting a matrix containing four elements, each element representing the probability of a modification type. The cross-entropy loss function is used to train the model, and the Adam optimizer is used with a learning rate set to 0.

001. After the model is trained, the optimal model is selected through five-fold cross-validation and five evaluation indicators (Aiming, Coverage, Accuracy, Absolute_True, Absolute_False). Finally, an evaluation is performed on an independent test set to verify the prediction effect of the model.

5. The method for identifying multiple lysine modification sites based on EPDCC technology according to claim 2, characterized in that in the step 1.2, the CD-HIT program is used, and the threshold is set to 0.4 to eliminate these potential homologous sequences and redundant samples; for the 11 types of data obtained, 70% of each class is randomly selected as training data, and the remaining 30% is used as test data, and the segmented data set after de-redundancy is obtained as follows: Training set: Category (1): 9279 items; Category (2): 710 items; Category (3): 600 items; Category (4): 454 items; Category (5): 561 items; Category (6): 252 items; Category (7): 360 items; Category (8): 88 items; Category (9): 153 items; Category (10): 454 items; Category (11): 73 items; Independent test set: Category (1): 4062 items; Category (2): 304 items; Category (3): 257 items; Category (4): 194 items; Category (5): 240 items; Category (6): 107 items; Category (7): 154 items; Category (8): 42 items; Category (9): 73 items; Category (10): 191 items; Category (11): 36 items.

6. The method for identifying multiple lysine modification sites based on EPDCC technology according to claim 1, characterized in that: In the step 3.4, the specific operations are as follows: The ratio of the 11 categories of training data in the training set is 9279:710:600:454:561:252:360:88:153:454:

73. The number of clusters is selected as 4, and the final under-sampling ratio is 0.1, that is, the sampling number is 928. After the sampling is successfully completed, the 11 categories in the training set have a sample distribution ratio of 928:710:600:454:561:252:360:88:153:454:

73.

7. The method for identifying multiple lysine modification sites based on EPDCC technology according to claim 1, characterized in that: In the step 3.12, in the data cluster analysis, the B-spline basis function suitable for non-periodic data is used to fit the data.

Images