A method for identifying multiple lysine modification sites based on EPDCC technology

Abstract

This invention belongs to the field of artificial intelligence algorithm applications—biological sequence recognition, and relates to a method for identifying multiple lysine modification sites based on EPDCC technology. This invention can simultaneously predict multiple lysine modification sites. It also specifically addresses the crosstalk problem between modifications, which previous techniques failed to adequately consider, thereby improving the accuracy and reliability of predictions. By fitting discrete data into functional data for clustering to balance the dataset, the extreme point bias compensation clustering undersampling algorithm employed in this invention effectively reduces the bias towards the majority class during model training. By constructing multi-scale data and combining sequence information features and amino acid physicochemical properties, the semantics of the data are enriched, significantly improving the model's predictive performance for lysine post-translational modification sites.

Description

Technical Field

[0001] This 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 Technology

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

[0003] Post-translational modifications of lysine residues include acetylation, methylation, and succinylation, among others. These modifications typically occur through enzymatic reactions and regulate many important cellular activities. However, traditional experimental methods, such as mass spectrometry and specific antibody detection, while providing relatively accurate information on modification sites, are costly, complex, and time-consuming. Furthermore, these methods are inefficient when processing large-scale data, failing to meet the demands of modern biological research for high-throughput and rapid prediction.

[0004] With the development of bioinformatics and artificial intelligence technologies, an increasing number of computational tools are being used to identify post-translational modification sites on amino acids such as lysine. Compared to traditional experimental methods, predictive identification models based on machine learning and deep learning offer higher efficiency and lower cost, enabling rapid analysis of large amounts of protein sequences and identification of potential modification sites. However, existing predictive models still face multiple challenges, such as imbalanced sample data distribution, insufficient consideration of interactions (crosstalk) between different modification types, and limited ability to capture complex dependencies in sequence information.

[0005] Therefore, developing a predictive tool that can comprehensively consider the interactions between multiple lysine modification types, effectively address the data imbalance problem, 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] To address the aforementioned issues, this invention proposes a method for simultaneously predicting four different types of lysine post-translational modification sites. This method utilizes a cross-scale convolutional neural network (CSCNN) and an extreme point bias-compensated 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 this invention comes from the CPLM4.0 database, which focuses on recording various post-translational modifications (PTMs) of lysine residue side chain amino groups at specific positions in proteins. This invention utilizes 18,978 human protein sequences provided by this database and, through a carefully designed data processing workflow, constructs a dataset suitable for model training and validation. Despite a significant imbalance problem in the dataset (ratio exceeding 100:1), this invention effectively addresses this challenge by proposing an innovative method of extreme point bias compensation clustering undersampling, improving the model's generalization ability under imbalanced data, thereby enhancing the accuracy and reliability of predictions. Furthermore, this invention employs two cross-scale feature extraction algorithms and combines them with a proposed cross-scale convolutional neural network for classification and recognition, finally developing a user-friendly web prediction tool.

[0009] The method for identifying multiple lysine modification sites based on EPDCC technology includes the following steps:

[0010] Step 1: Dataset Construction

[0011] To develop statistical prediction models, building reliable training and testing datasets is crucial. The specific steps for building such datasets are as follows:

[0012] 1.1 Sequence Extraction and Validation: For the 18,978 collected human protein sequences, extraction experiments were conducted to verify peptide fragments modified with lysine (K) as “acetyllysine”, “crotonyllysine”, “methyllysine”, or “succinyllysine” (with K as the center, the sliding window size was 24, i.e., the peptide fragment length was 49). Each peptide fragment can be represented in the form of Equation (1):

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

[0014] 1.2 Data Classification: The preprocessed lysine-modified data was classified into 15 categories. After removing categories with fewer than 60 sequences, the dataset was finally divided into 11 categories:

[0015]

[0016] Data category description: This refers to a protein sequence where the lysine residue at the center position contains only acetylated modifications. This refers to protein sequences that contain both acetylation and crotonylation modifications. After removing duplicate peptide fragments from the data, 11 data categories were obtained, with the following sample sizes for each category: As mentioned above, the number of samples in the first class is 39,938, the number of samples in the second class is 2,463, and so on. To ensure that the performance evaluation of the multi-label prediction model is not overestimated due to sequence redundancy and homology, this invention uses the CD-HIT procedure and sets the threshold to 0.4 to eliminate these potentially homologous sequences and redundant samples. For the obtained 11 classes of data, 70% of each class is randomly selected as training data, and the remaining 30% is used as test data, resulting in the following deredundant segmented dataset:

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

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

[0019] Step 2: Feature Extraction

[0020] In order for the computer to recognize the data cleaned in the first stage, this stage performs feature extraction (expressing the data in mathematical form). This invention extracts features from the data based on two scales: lysine sequence information and the physicochemical properties of amino acids.

[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 step 1. Then calculate the frequency matrix F of each triplet amino acid at each position in the lysine sequence of class t. t The matrix size is 20 3 ×46.

[0023]

[0024] Where F t (TAA i / j) represents the TAA in the t-th type of lysine sequence. i The frequency of the triamino acid at the j-th position. (TAA) 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 triplet amino acid at each position in the lysine sequences of the other ten classes (excluding class k). t The matrix size is 20 3 ×46.

[0026]

[0027] Where F k (TAA i / j) represents the ten categories other than the kth category. The TAA in the lysine sequence i The frequency of the triamino acid at the j-th position.

[0028] 2.1.3: The 11 F values ​​obtained t Matrix and 11 FF k The average of the matrices is obtained by subtracting the matrices F and FF to obtain matrix F. 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: Feature encoding is performed by finding the corresponding value in the Fr matrix based on the type of triplet amino acid at each position in each sequence. For example, for a 48-bit sequence "AGAT……TGLR", encoding is performed continuously using a sliding window of size 3. First, the position of the triplet amino acid fragment in the sequence must be determined. For example, "AGA" is located at the first position in the sequence, and "GAT" is located at the second position. Then, according to the amino acid alphabet 'ACDEFGHIKLMNPQRSTVWY' (row position = position 1 × 20), feature encoding is performed. 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 located 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. Ultimately, the sequence is encoded by features, and the first position of the feature vector is: f 101,1 The second position is: f 2017,2 And so on, until a feature vector of length 46 is obtained.

[0032] 2.2 The physicochemical properties of amino acids were extracted as follows:

[0033] The molecular weight, isoelectric point, hydrophilicity, and hydrophobicity of amino acids have a significant impact on protein structure, which in turn affects the types of post-translational modification sites. Therefore, this invention uses these three physicochemical properties of amino acids as a feature extraction method. The physicochemical properties of 20 natural amino acids are listed in the table below:

[0034]

[0035] This invention encodes each amino acid in each protein sequence. Hydrophilic amino acids are labeled as 1, and hydrophobic amino acids as -1. Each amino acid is converted into a vector with three values; for example, the vector for glycine (G) is (75.07, 5.97, -1), and the vector for alanine (A) is (89.09, 6.01, 1). Thus, each protein sequence is ultimately converted into a 49-row, 3-column matrix, where each row represents the specific encoding of one amino acid.

[0036] Step 3: Handling Data Imbalance

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

[0038] 3.1 Function Fitting:

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

[0040] 3.1.1 Dimensionless Processing: To eliminate the order-of-magnitude differences between data, this invention employs a normalization method to perform dimensionless processing on the 46-dimensional data. This allows data with different dimensions to be compared on the same scale.

[0041] 3.1.2 Selecting the Fitting Basis Function. Function fitting reconstructs discrete data to obtain a continuous, smooth curve, i.e., the 46-dimensional discrete feature values ​​[y] extracted for a sample i. i1 ,y i2 ,…,y i46 The function curve x is obtained by fitting the curve. i (t), whose basic form is shown in equation (6):

[0042] y ij =x i (t j )+ε(t j ) (6)

[0043] Among them, t j To mark the x-coordinate corresponding to the current position (this invention uses 45 equidistant points to divide the interval [0,1], and the coordinate value of each point is used as the x-coordinate of the corresponding data point), j = 1, 2, ..., 46, ε(t j The error term is represented by (). To better capture the dynamic characteristics of the data, this invention utilizes B-spline basis functions suitable for aperiodic data to fit the data in data clustering analysis.

[0044] 3.1.3 Selecting Data Nodes: Appropriate nodes are selected based on the distribution of data points. Node selection can be achieved through equal-interval selection or adaptive selection based on data point density. Based on the research results of this invention, this invention selects 10 nodes as the data stage by uniformly generating them between the minimum and maximum values ​​of the data's horizontal coordinate using an equal-interval selection method.

[0045] 3.1.4 Construct the k-th order B-spline basis functions as shown in equation (7):

[0046]

[0047] Among them, c ik The coefficient to be estimated is the k-th term, φ. k (t) is the k-th order B-spline basis function. To make the fitted function smooth and accurate, it is necessary to choose an appropriate order. Based on the characteristics of the data in this invention and the research results, the order of the basis function is finally determined to be 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, It is x i The p-th derivative of (t). By solving the above optimization problem, the parameter c can be obtained. ik This leads to the fitted function curve x. i (t).

[0051] 3.2 Distance metric in cluster analysis:

[0052] After fitting the data into functional data, this invention uses an improved K-means++ algorithm for cluster analysis. To clearly define the similarities and differences between samples, this invention incorporates the idea of ​​penalizing extreme point deviations into the distance measurement. This not only measures the absolute level difference of the curves but also captures the morphological differences between the curves, thereby improving the effectiveness of cluster analysis. In cluster analysis, the calculation of two functional data x... i (t) and x j The distance steps for (t) are as follows:

[0053] 3.2.1 First, regarding x i (t) and x j (t) Calculate its first derivative, and then find the points where the derivative is zero, i.e., the extreme points. Let the set of extreme points be 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 jSince the sizes of the two extreme point sets are different, interpolation is used to make the sizes of the two extreme point sets equal, resulting in an adjusted extreme point set. Let... and

[0055] 3.2.3 Calculation of extreme point deviation distance. Extreme point deviation distance D ext Defined as the form of equation (9):

[0056]

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

[0058]

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

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

[0061] The similarity measurement method based on curve extreme point deviation compensation not only considers the differences in function values ​​themselves, but also incorporates the deviation of extreme point positions, thus improving the accuracy and reliability of similarity measurement.

[0062] 3.3 Cluster Analysis

[0063] After fitting the data to a function, this invention uses an improved K-means++ clustering analysis method for data clustering. The basic idea of ​​this improved method is: first, calculate the pairwise D between all samples. total For distance analysis, the two samples with the largest distance are selected as the initial cluster centers. Then, assuming n cluster centers have already been selected, when selecting the (n+1)th cluster center, the sample with the largest sum of distances to the current n samples is chosen. This method is recursively used to generate K (number of categories) samples as initial cluster centers. This improvement, compared to K-means++ which randomly selects an arbitrary point, completely eliminates randomness and speeds up clustering. The samples selected after undersampling are more representative of the features of the original samples.

[0064] The pseudocode for the Extreme Point Deviation Compensation Clustering Undersampling Algorithm (EPDCC) is as follows:

[0065]

[0066] 3.4 Confirm the number of samples and clusters for the first type of data.

[0067] Since the ratio of the 11 training data categories in the training set of this invention is 9279:710:600:454:561:252:360:88:153:454:73, it is clear that the first category has a much larger number of samples than the other ten categories, accounting for approximately 71% of the total samples. Therefore, EPDCC undersampling was only performed on the first category of data. This invention was verified through experiments, selecting a cluster size of 4, resulting in a final undersampling ratio of 0.1, i.e., a sample size of 928. After successful sampling, the sample distribution ratio of the 11 categories in the training set is: 928:710:600:454:561:252:360:88:153:454:73.

[0068] 3.5 Combine the sampled data of the first class with the original data of classes 2-11 to form a new training dataset.

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

[0070] In this stage, the invention constructs a multi-label, cross-scale convolutional neural network based on the training set obtained in step four, 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 proposed Cross-Scale Convolutional Neural Networks (CSCNN) utilizes convolutional neural networks (CNNs) to extract hidden layer relationships between labels. By fusing features and weights at different scales, it enriches the dimensions of data analysis, thereby achieving superior classification results. This learning architecture mainly consists of the following parts:

[0073] (1) Unified Alignment of Multi-Source Feature Representation: By developing a unified alignment of multi-source feature representation, the homogeneity of lysine sequence information features and amino acid physicochemical properties features in dimensions is ensured, thereby generating potential feature representations.

[0074] (2) Construct cross-scale feature and weight extractors to fully extract the depth information of features and generate feature weights in the process. Both extractors contain CNN modules with the same kernel size to understand the contextual dependencies under different features, thereby making accurate interaction predictions. The extractors generate feature weights based on mid-level features, which improves the accuracy of subsequent label predictions while preventing overfitting caused by excessive feature extraction.

[0075] (3) Feature Fusion and Prediction: The model fuses latent information learned at different scales and combines it with weights to obtain prediction results, in order 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 multiplying the weights. The network output uses the Sigmoid activation function and sets a threshold of 0.5 to predict lysine post-translational modification sites. This process reduces the model's bias towards the majority class and can intuitively display the scores for predicting different modification sites.

[0076] 4.2 Model Training and Validation Testing

[0077] After building the model, it is trained using the training set obtained in step three. Once a mature model is obtained, it is tested on an independent test set. The model of this invention predicts lysine modification sites in protein samples by outputting a matrix containing four elements, where each element represents the probability of a modification type. The model is trained using the cross-entropy loss function and the Adam optimizer with a learning rate of 0.001. After training, the optimal model is selected using five-fold cross-validation and five evaluation metrics (Aiming, Coverage, Accuracy, Absolute_True, Absolute_False). Finally, the model is evaluated on an independent test set to verify its predictive performance.

[0078] The beneficial effects of this invention are:

[0079] 1. This invention overcomes the limitations of existing technologies, which typically focus only on a single type of protein post-translational modification (PTM) while neglecting the potential interactions between different PTMs. In contrast, this invention develops a multi-tag prediction tool capable of simultaneously predicting multiple lysine modification sites. Furthermore, this invention specifically addresses the crosstalk problem between modifications that previous technologies failed to adequately consider, thereby improving the accuracy and reliability of predictions.

[0080] 2. By fitting discrete data into functional data for clustering to balance the dataset, the extreme point bias compensation clustering undersampling algorithm (EPDCC) used in this 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 features and amino acid physicochemical properties, the semantics of the data were enriched, and the predictive performance of the model for lysine post-translational modification sites was significantly improved.

[0082] 4. To facilitate researchers in using this invention to improve experimental efficiency, a free web application has been developed: http: / / www.mlypredcsed.com. Attached Figure Description

[0083] Figure 1 This is a flowchart illustrating the overall technical structure of the present invention.

[0084] Figure 2 This is the multi-label cross-scale convolutional neural network diagram in this invention.

[0085] Figure 3 This is a comparison chart of the indicators corresponding to the number of clusters in the EPDCC algorithm of this invention.

[0086] Figure 4(a) is a comparison of the performance indicators of the EPDCC algorithm in this invention with other similar algorithms.

[0087] Figure 4(b) shows the prediction accuracy results of the EPDCC algorithm in this invention compared with other similar algorithms.

[0088] Figure 5 This is a comparison chart of the proposed method MlyPredCSED with other methods, the purpose of which is to verify the effectiveness of the proposed method. Detailed Implementation

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

[0090] This invention proposes a method based on EPDCC technology to identify multiple lysine modification sites, and a multi-label classification model for predicting eleven types of lysine modification sites in protein sequences. Its architecture is as follows: Figure 1 As shown, the model first uses lysine sequence information and the physicochemical properties of three amino acids for feature encoding. Then, it uses the Extreme Point Bias Compensation Clustering Undersampling Algorithm (EPDCC) to undersampling the majority class (Class 1) to alleviate class imbalance and improve the model's generalization ability to classes with fewer samples. Furthermore, this invention constructs a cross-scale feature and weight extractor by fusing two types of features, and then designs a multi-label cross-scale convolutional neural network architecture, such as... Figure 2As shown in the diagram, a multi-label binary cross-entropy loss function was employed during model training, and the network parameters were optimized using a backpropagation algorithm to improve the model's prediction accuracy for various lysine modification sites. This invention also comprehensively evaluated the model's performance using six evaluation metrics—aiming rate, coverage rate, accuracy, absolute accuracy, absolute false rate, and j-level matching rate—through five-fold cross-validation and independent testing experiments. Experimental results show that the proposed method, MlyPredCSED, significantly outperforms existing models such as iMul-kSite, predML-Site, and PreMLS on the task of predicting various lysine modification sites. Furthermore, to facilitate user access to the multi-label prediction model constructed in this invention, a publicly available web prediction tool was developed.

[0091] In summary, the multi-tag lysine modification site prediction model of this invention achieves efficient identification of multiple lysine modification sites in protein sequences by employing an extreme point bias-compensated clustering undersampling algorithm (EPDCC) and a cross-scale convolutional neural network architecture. This model solves the class imbalance problem and significantly improves prediction accuracy. Furthermore, comprehensive testing across a series of evaluation metrics verifies the model's superior performance, ensuring its reliability and effectiveness in practical applications. This invention also develops a user-friendly web-based prediction tool, making the prediction model more accessible and usable, providing a powerful scientific tool for biological research and drug development.

[0092] Example 1: Data Preprocessing and Feature Extraction

[0093] In this embodiment, multiple datasets for predicting lysine modification sites were constructed. First, 18,978 sequences underwent meticulous truncation, classification, and optimization to select datasets for 11 different categories. To ensure dataset quality, the CD-HIT procedure was used to remove redundant samples, thereby improving dataset purity.

[0094] Next, we performed feature extraction on lysine sequence information and the physicochemical properties of amino acids. This step was crucial because it not only ensured the diversity of the dataset but also guaranteed its reliability. These feature extraction efforts laid a solid foundation for subsequent model training and evaluation.

[0095] This series of data preprocessing and feature extraction steps provides high-quality input data for the lysine modification site prediction model, thereby improving the model's prediction accuracy and reliability.

[0096] Example 2: Sample Balancing Technique

[0097] In bioinformatics and protein sequence analysis, the EPDCC undersampling algorithm is used to address the imbalanced dataset problem. This method involves function fitting followed by cluster analysis of the function data to select representative samples from the majority class, thereby alleviating the sample imbalance problem while maintaining data diversity and representativeness.

[0098] A comparative experiment was conducted to determine the undersampling ratio and the number of clusters. Figure 3 Comparing the results of five metrics on the independent test set with cluster numbers ranging from 1 to 9, the results show that the method performs best when the number of clusters is 4. Table 1 shows the comparison results of metrics on the independent test set under different undersampling ratios. As shown in the table, when the undersampling ratio is 0.1, the model's generalization performance for four classes of samples containing only acetyllysine modification, only crotonyllysine modification, only methyllysine modification, and both acetyllysine and methyllysine modification is significantly better than other ratios. Furthermore, with an undersampling ratio of 0.1, the model exhibits a certain degree of generalization across all classes, avoiding an ineffective model that only generalizes to a few classes. Therefore, this invention selects 0.1 as the final undersampling ratio.

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

[0100]

[0101] To further verify the effectiveness of the EPDCC method, comparative experiments were conducted, as shown in Figure 4. Figure 4(a) compares the performance of the models without undersampling, random undersampling, undersampling using ClusterCentroids, and the model using the EPDCC algorithm on the multi-label prediction task. Experimental results show that the model using the EPDCC algorithm outperforms the other three methods in all five metrics. Figure 4(b) compares the prediction accuracy of each method on 11 categories of data. The results show that the EPDCC sampling algorithm outperforms other methods in prediction accuracy in all categories except for categories 5, 9, 10, and 11, and achieves completely correct predictions in each category. These results fully verify the effectiveness of the sampling method EPDCC proposed in this invention.

[0102] Example 3: Model Performance Evaluation and Comparison

[0103] like Figure 5As shown, the model MlyPredCSED constructed in this invention is compared with predML-Site, iMul-kSite, and PreMLS on independent test datasets. Existing prediction tools iMul-kSite and predML-Site both decompose the original multi-label prediction problem into multiple single-label binary classification problems. Specifically, these tools divide the dataset into positive and negative classes for each label, then use an improved algorithm based on Support Vector Machine (SVM) for binary classification, and finally integrate the individual binary classification models to construct a multi-label prediction model. PreMLS, on the other hand, directly constructs a multi-label prediction tool using a deep learning convolutional neural network.

[0104] Experimental results show that MlyPredCSED improves performance by approximately 10% on Aiming and Absolute-True metrics, approximately 7% on Accuracy and 4% on Coverage metrics, while reducing performance on Absolute-False metrics by approximately 6%. Clearly, MlyPredCSED outperforms the three existing models mentioned above on the test set.

[0105] To further validate the effectiveness of MlyPredCSED, as shown in Table 2, MlyPredCSED was compared with the three models mentioned above on the independent test set in terms of four metrics: MR1, MR2, MR3, and MR4. The results show that MlyPredCSED outperforms the other models in MR1 and MR2, exceeding the worst-performing model by approximately 40% and 50% in these two metrics, respectively. However, its performance on the MR4 metric is significantly worse than the PreMLS model. This is because PreMLS achieves 100% accuracy in predicting the class with the label (1,1,1,1), which represents a very small percentage of the entire dataset. This clearly demonstrates that MlyPredCSED outperforms both ensemble binary classification methods and the PreMLS model.

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

[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 Build training and testing datasets; 1.1 Sequence extraction and verification: For the 18,978 human protein sequences collected, the extraction experiment verified that the lysine residue was modified by "acetyllysine", "crotonyllysine", "methyllysine", or "succinyllysine", and the peptide fragment length was 49. 1.2 Data Classification: The preprocessed lysine-modified data was classified into 15 categories. After removing categories with fewer than 60 sequences, the dataset was finally divided into 11 categories: the category with the largest number of samples was defined as the first category. Step 2: Feature Extraction This stage involves feature extraction from the data, using lysine sequence information and amino acid physicochemical properties as two dimensions. Step 3: Handling Data Imbalance After the data feature extraction was completed, two dimensions of data were obtained: (1) Obtain lysine sequence information features with a dimension of 46; (2) Physicochemical properties of amino acids with dimensions (49, 3); In this stage, to ensure that the training data categories are generally balanced, the following steps are taken based on 46-dimensional lysine sequence information features: 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 processed to be dimensionless using a normalization method, so that data with different dimensions can be compared on the same scale; 3.1.2 Selecting the fitting basis function; function fitting reconstructs the discrete data to obtain a continuous smooth curve, that is, the 46-dimensional discrete feature values ​​[y] extracted for a sample i. i1 ,y i2 ,…,y i46 The function curve x is obtained by fitting the curve. i (t), whose basic form is shown in equation (6): y ij =x i (t j )+ε(t j ) (6) in, t j To mark the x-coordinate corresponding to the current position, j = 1, 2, ..., 46, ε(t) j ) represents the error term; 3.1.3 Selecting Data Nodes: Select appropriate nodes based on the distribution of data points; node selection can be achieved through equal spacing or adaptive selection based on data point density. 3.1.4 Construct the k-th order B-spline basis functions as shown in equation (7): Among them, c ik The coefficient to be estimated is the k-th term, φ. k (t) is a k-th degree B-spline basis function; the order of the basis function K is 3; 3.1.5 Estimate parameter c by minimizing the objective function shown in equation (8) using the least squares method. ik : Where λ is the smoothing factor. It is x i Find the p-th derivative of (t); obtain the parameter c by solving. ik This leads to the fitted function curve x. i (t); 3.2 Distance metric in cluster analysis: Calculating two function data x in cluster analysis i (t) and x j The distance steps for (t) are as follows: 3.2.1 First, regarding x i (t) and x j (t) Calculate its first derivative, and then find the points where the derivative is zero, i.e., the extreme points; let the set of extreme points 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 Since the sizes of the two extreme point sets are different, interpolation is used to make the sizes of the two extreme point sets equal, resulting in an adjusted extreme point set. Let... and 3.2.3 Calculation of extreme point deviation distance; extreme point deviation distance D ext Defined as the form of equation (9): 3.2.4 Calculate two functions x i (t) and x j The Euclidean distance (t) is calculated using the formula shown in equation (10): 3.2.5 Total distance calculation, final similarity measure D total It combines numerical distance and extreme point deviation distance: D total =D num +D ext (11) 3.3 Cluster Analysis Calculate D between all pairs of samples total Distance: Select the two samples with the largest distance between them as the initial cluster centers; Then, assuming that n cluster centers have been selected, when selecting the (n+1)th cluster center, select the sample with the largest sum of distances to the current n samples, and recursively select multiple samples as initial cluster centers in this way. 3.4 Confirm the number of samples and clusters for the first type of data; 3.5 Combine the sampled data of the first class with the original data of classes 2-11 to form a new training dataset; Step 4: Multi-label cross-scale convolutional neural network classification model The multi-label, cross-scale convolutional neural network was trained based on the obtained training set and tested on an independent test set. 4.1 Construction of Multi-Label Cross-Scale Convolutional Neural Network Classification Model The proposed multi-label cross-scale convolutional neural network (CSCNN) utilizes convolutional neural networks (CNNs) to extract hidden layer relationships between labels and consists of the following parts: (1) Unified alignment of multi-source feature representation; (2) Construct cross-scale feature and weight extractors, with two extractors containing CNN modules with the same kernel size; the extractors generate feature weights based on the mid-layer features; (3) Feature fusion and prediction: The latent information learned at different scales is fused and combined with weights to obtain the prediction result; after integrating all extracted features and normalizing them, the network output is obtained by multiplying the weights. 4.2 Model training and validation testing.

2. The method for identifying multiple lysine modification sites based on EPDCC technology as described in claim 1, characterized in that, The first step, dataset construction, involves the following specific steps: 1.1 Sequence Extraction and Verification: For the 18,978 human protein sequences collected, extraction experiments were conducted to verify peptide fragments modified with lysine by "acetyllysine", "crotonyllysine", "methyllysine", or "succinyllysine". The peptide fragment length was 49, and each peptide fragment was represented 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: The preprocessed lysine-modified data was classified into 15 categories. After removing categories with fewer than 60 sequences, the dataset was finally divided into 11 categories: Data category description: This refers to a protein sequence where the lysine residue at the center position contains only acetylated modifications. This refers to protein sequences that contain both acetylation and crotonylation modifications; after removing repetitive peptide fragments from the data, 11 data categories were obtained, with the following sample sizes for each category:

3. The method for identifying multiple lysine modification sites based on EPDCC technology as described in claim 1 or 2, characterized in that the second step: feature extraction, is specifically operated as follows: 2.1 The steps for extracting lysine sequence information features are as follows: 2.1.1: Remove the central lysine from the sequence obtained in step 1; then calculate the frequency matrix F of each triplet amino acid at each position in the lysine sequence of class t. t The matrix size is 20 3 ×46; Where F t (TAA i / j) represents the TAA in the t-th type of lysine sequence. i The frequency of the triamino acid at the j-th position; TAA i ∈{AAA,AAC,AAD,…YYY}, that is, TAA1=AAA, i = 1, 2, 3, ..., 20 3 j = 1, 2, 3, ..., 46; 2.1.2: Calculate the frequency matrix FF of each triplet amino acid at each position in the lysine sequences of the other ten classes (excluding class k). t The matrix size is 20 3 ×46; Where F k (TAA i / j) represents the ten categories other than the kth category. The TAA in the lysine sequence i The frequency of the triamino acid at the j-th position; 2.1.3: The 11 F values ​​obtained t Matrix and 11 FF k The average of the matrices is obtained as matrices F and FF, and the difference is obtained as matrix F. 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: Based on 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 physicochemical properties of amino acids were extracted as follows: Each amino acid in each protein sequence is encoded; hydrophilic amino acids are marked as 1 and hydrophobic amino acids as -1; each amino acid is converted into a vector with three values, and each protein sequence is finally converted into a 49-row, 3-column matrix, where each row represents a specific code for an amino acid.

4. The method for identifying multiple lysine modification sites based on EPDCC technology as described in claim 1 or 2, characterized in that the fourth step comprises the following steps: 4.1 Construction of Multi-Label Cross-Scale Convolutional Neural Network Classification Model The proposed multi-label cross-scale convolutional neural network (CSCNN) utilizes convolutional neural networks (CNNs) to extract hidden layer relationships between labels and consists of the following parts: (1) Unified Alignment of Multi-Source Feature Representation: By developing a unified alignment of multi-source feature representation, the homogeneity of lysine sequence information features and amino acid physicochemical properties 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 depth 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 extractors generate feature weights based on the mid-level features, which can prevent overfitting caused by excessive feature extraction and improve the accuracy of subsequent label predictions. (3) Feature fusion and prediction: The model fuses the potential information learned at different scales and combines the weights to obtain the prediction results in order 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 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. 4.2 Model Training and Validation Testing After building the model, it is trained on the training set obtained in step three. Once a mature model is obtained, it is tested on an independent test set. A four-element matrix is ​​output to predict lysine modification sites in protein samples, with each element representing the probability of a modification type. The model is trained using the cross-entropy loss function and the Adam optimizer with a learning rate of 0.

001. After training, the optimal model is selected using five-fold cross-validation and five evaluation metrics: Aiming, Coverage, Accuracy, Absolute_True, and Absolute_False. Finally, the model is evaluated on an independent test set to verify its predictive performance.

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

6. The method for identifying multiple lysine modification sites based on EPDCC technology as described in claim 1, characterized in that, In step 3.4, the specific operations are as follows: The proportion of the 11 training data categories 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 undersampling ratio is 0.1, that is, the number of samples is 928. After successful sampling, the sample distribution of the 11 categories in the training set is 928:710:600:454:561:252:360:88:153:454:

73.

7. The method for identifying multiple lysine modification sites based on EPDCC technology as described in claim 1, characterized in that, In step 3.12, during data clustering analysis, B-spline basis functions suitable for non-periodic data are used to fit the data.

Images