A Deep Learning-Based Method for Predicting lncRNA m5C Methylation Sites

Abstract

The application discloses a deep learning-based lncRNA m5C methylation site prediction method, and relates to the technical field of m5C site prediction.The application is aimed at solving the problem of low prediction accuracy of the existing lncRNA m5C methylation site method.The application comprises the following steps: obtaining lncRNA sequence fragments corresponding to a methylation data file in a human gene library, and constructing a training set by using the lncRNA sequence fragments; performing sequence coding on the data in the training set to obtain an original feature vector of the lncRNA and obtain a coded training set; training an lncRNA m5C methylation site prediction model by using the coded training set to obtain a trained lncRNA m5C methylation site prediction model; obtaining the original features of a to-be-detected lncRNA sequence; and inputting the original features of the to-be-detected lncRNA sequence into the trained lncRNA m5C methylation site prediction model to obtain an m5C site prediction result.The application is used for predicting lncRNA m5C methylation sites.

Description

Technical Field

[0001] This invention relates to the field of m5C site prediction technology, and in particular to a method for predicting lncRNA m5C methylation sites based on deep learning. Background Technology

[0002] m5C is an important RNA modification that is widely involved in the stability, localization, and functional regulation of RNA. As an important class of regulatory molecules, lncRNAs not only affect their own structure and function through m5C modification, but also participate in biological processes such as gene expression regulation, chromatin remodeling, and signaling pathway activation through interactions with proteins, DNA, or other RNAs.

[0003] In recent years, the prediction of RNA m5C sites has sparked a widespread research boom both domestically and internationally, achieving a series of significant advancements. To date, numerous computational methods based on sequence-derived information have been developed to predict m5C sites in various species, including humans, mice, and *Saccharomyces cerevisiae*. For example, iRNAm5C-PseDNC uses random forests to predict m5C sites. This model is built on an imbalanced and redundant dataset containing 475 m5C sites and 1425 non-m5C sites. RNAm5Cfinder, a web server based on the random forest algorithm, collects data from the GEO database and can be used to predict m5C sites in eight cell or tissue types in mice and humans. Although it collects all m5C sites recorded in three GEO records and all other non-m5C sites in the genome to train the model, the redundancy of the dataset is not well addressed. During the development of the Deepm5C model, a new benchmark dataset was constructed. A mixture of three traditional feature encoding algorithms and features extracted from word embedding methods was studied. Four deep learning classifier variants and four commonly used traditional classifiers were adopted and trained using four encoding methods. Finally, 32 baseline models were obtained. The predicted outputs of the best baseline models were then integrated and trained with a one-dimensional convolutional neural network, effectively utilizing a stacking strategy to improve model performance.

[0004] While some progress has been made in RNA m5C site prediction research, current studies focus on ordinary RNA, especially mRNA, and lack specific research on lncRNA m5C prediction. lncRNA and mRNA m5Cs differ significantly in structure, function, and distribution. Applying current prediction models for ordinary RNA m5C sites to predict lncRNA m5C methylation sites leads to low prediction accuracy. Furthermore, due to the limited availability of lncRNA m5C site data, models built using traditional machine learning or deep learning methods suffer from overfitting, limiting their generalization ability and resulting in low prediction accuracy. Summary of the Invention

[0005] The purpose of this invention is to address the problem of low prediction accuracy in existing lncRNA m5C methylation site methods, and to propose a deep learning-based lncRNA m5C methylation site prediction method.

[0006] The deep learning-based method for predicting lncRNA m5C methylation sites is as follows:

[0007] Step 1: Obtain the original characteristics of the lncRNA sequence to be tested;

[0008] Step 2: Input the original features of the lncRNA sequence to be tested into the trained lncRNA m5C methylation site prediction model to obtain the m5C site prediction results;

[0009] If the predicted m5C site is greater than the preset threshold, it indicates that the m5C site exists in the lncRNA sequence to be tested; otherwise, it indicates that the m5C site does not exist in the lncRNA sequence to be tested.

[0010] The trained lncRNA m5C methylation site prediction model was obtained through the following method:

[0011] S1. Obtain the lncRNA sequence fragments corresponding to the methylation data files in the human gene bank, and construct a training set using the lncRNA sequence fragments;

[0012] S2. Sequence encoding is performed on the data in the training set to obtain the original feature vector of lncRNA, thus obtaining the encoded training set;

[0013] S3. Use the encoded training set to train the lncRNA m5C methylation site prediction model to obtain the trained lncRNA m5C methylation site prediction model.

[0014] Furthermore, the acquisition of the original characteristics of the lncRNA sequence to be tested in step one specifically includes:

[0015] Step 11: Encode the physicochemical properties and cumulative frequency information of each nucleotide in the lncRNA sequence to be tested to obtain the first coding feature matrix of the lncRNA sequence to be tested;

[0016] Step 1 and Step 2: Encode the dinucleotide chemical properties of the lncRNA sequence to be tested to obtain the dinucleotide chemical property feature vector of the lncRNA sequence to be tested;

[0017] Step 13: Concatenate the first coding feature vector and the dinucleotide chemical property feature vector of the lncRNA sequence to be tested to obtain the original feature vector of the lncRNA sequence to be tested.

[0018] Further, in step one, the physicochemical properties and cumulative frequency information of each nucleotide in the lncRNA sequence to be tested are encoded to obtain the first coding feature matrix of the lncRNA sequence to be tested, specifically as follows:

[0019] First, the physicochemical properties of each nucleotide in the lncRNA sequence to be tested are encoded, specifically as follows:

[0020] If the nucleotide is adenine A, the physicochemical property code is (1,1,1); if the nucleotide is cytosine C, the physicochemical property code is (0,1,0); if the nucleotide is guanine G, the physicochemical property code is (1,0,0); if the nucleotide is uracil U, the physicochemical property code is (0,0,1).

[0021] Then, the cumulative base frequency of the lncRNA sequence to be tested is obtained.

[0022] Among them, f i' d is the frequency of the i'th nucleotide appearing before the i'th nucleotide position. i' It represents the number of times the i'th nucleotide appears before the position of the i'th nucleotide, where i' is the nucleotide number in the lncRNA to be tested;

[0023] Finally, the physicochemical properties of nucleotides and the corresponding cumulative base frequencies in the lncRNA sequence to be tested are combined to form the first coding feature vector of nucleotides, thereby obtaining the first coding feature matrix of the lncRNA sequence to be tested.

[0024] Further, in steps one and two, the dinucleotide chemical properties of the lncRNA sequence to be tested are encoded to obtain a feature vector of the dinucleotide chemical properties of the lncRNA sequence to be tested, specifically as follows:

[0025] First, obtain the PCP matrix of the lncRNA sequence to be tested:

[0026]

[0027] Where p∈[1,10], p is the PCP label, PCP p (P j P j+1 ) is a dinucleotide P j P j+1 The p-th PCP value, where L is the length of the lncRNA sequence to be tested, j∈[1,L-1], and j is the nucleotide label;

[0028] Then, the PCP matrix is ​​normalized to obtain the normalized PCP matrix, and the dinucleotide autocovariance matrix A' and cross-covariance matrix C' are obtained using the normalized PCP matrix.

[0029] Finally, matrices A' and C' are concatenated to obtain the dinucleotide chemical property feature vector of the lncRNA sequence to be tested.

[0030] Furthermore, the step of obtaining the dinucleotide autocovariance matrix A' and cross-covariance matrix C' using the normalized PCP matrix specifically involves:

[0031]

[0032] Where s = 1, 2, 3, 4, λ1 ≠ λ2, λ1 and λ2 are integers, λ1 = 1, 2, ..., 10, λ2 = 1, 2, ..., 10. It is the average value of the data in the λ1-th row of the normalized PCP matrix. It is the average value of the data in the λ2-th row of the normalized PCP matrix, PCP' p (P j P j+1 ) is the normalized dinucleotide P j P j+1 The p-th PCP value, It is the average value of the data in the p-th row of the normalized PCP matrix.

[0033] Further, in step S1, obtaining the lncRNA sequence fragment corresponding to the methylation data file in the human gene bank and constructing a training set using the lncRNA sequence fragment specifically involves:

[0034] S11. Obtain the lncRNA m5C site information from the RMBase v3.0 database or the m5C-Atlas database, and merge the lncRNA m5C site information from the RMBase v3.0 database with the lncRNA m5C site information from the corresponding m5C-Atlas database to obtain the total set of lncRNA m5C site information.

[0035] The site information of the lncRNA m5C includes: chromosome number, start position, end position, and positive / negative strand identifier;

[0036] S12. Based on the lncRNA m5C site information, extract multiple lncRNA sequence fragments from the human genome, then cut out sequence fragments of a preset length L, use the CD-HIT tool to remove sequence fragments that are larger than a preset similarity threshold, and use the remaining sequence fragment set as the positive sample set.

[0037] S13. Extract sequence fragments of length L centered on cytosine from upstream and downstream of the positive sample site. Then, use the CD-HIT tool to remove sequence fragments larger than the preset similarity threshold and fragments that overlap with the positive sample site. The remaining sequence fragment set is used as the negative sample set.

[0038] S14. Label the positive and negative samples, and combine the labeled positive and negative sample sets to obtain the training set.

[0039] In this study, the label for a positive sample was 1, and the label for a negative sample was 0.

[0040] Further, in step S2, sequence encoding is performed on the data in the training set to obtain the original feature vector of the lncRNA, thus obtaining the encoded training set, specifically as follows:

[0041] S21. Encode the nucleotide physicochemical properties and cumulative frequency information of the sequence fragments in the training set to obtain the first coding feature matrix;

[0042] S22. Encode the dinucleotide chemical properties of the sequence fragment to obtain the dinucleotide chemical property feature vector of the sequence fragment;

[0043] S23. Concatenate the first coding feature vector obtained in S21 and the dinucleotide chemical property feature vector obtained in S22 to obtain the original feature vector of the lncRNA sequence. Use the original feature vector of the lncRNA sequence and the label to form the encoded training set.

[0044] Furthermore, the lncRNA m5C methylation site prediction model includes: a CNN network module, a Bi-LSTM network module, a feature fusion module, and a fully connected module;

[0045] The CNN network module is used to perform preliminary feature extraction on the original feature vector of the lncRNA sequence to obtain local feature information F1 of the sequence fragment;

[0046] The Bi-LSTM network module uses feature F1 to obtain sequence feature F2;

[0047] The feature fusion module is used to concatenate features F1 and F2 to obtain the optimal feature vector F of the lncRNA sequence;

[0048] The fully connected layer uses the optimal feature vector F of the lncRNA sequence to output the m5C site prediction result.

[0049] Furthermore, the CNN network module includes: three convolutional units;

[0050] Each convolutional unit includes, in sequence: a convolutional layer, an activation function layer, and a pooling layer;

[0051] The convolutional layer in the first convolutional unit is specifically as follows:

[0052]

[0053] Among them, X u+m,n is the original feature vector of lncRNA, u is the output position index, k is the kernel index, and conv(X) is the kernel index. uk It is the output of the convolutional layer. It is an M*N weight matrix, where M is the window size and N is the number of input channels;

[0054] The convolutional layers in other convolutional units follow the same formula as above, with the input being the output of the previous convolutional unit;

[0055] The activation function layer is as follows:

[0056]

[0057] Here, x is the input to the activation function; the input to each activation function is x, which is the output of the previous convolutional layer.

[0058] Furthermore, the pooling layer specifically comprises:

[0059] pooling(x') u' =max(x'1,x'2,x'3,...,x' M' ) u'

[0060] Where x′ represents the input feature, the input x′ of each pooling layer is the output of the previous activation function, u' represents the output position index, M' represents the pooling window size, and x'1, x'2, x'3, ..., x' M' It is the sequence output by the previous activation function.

[0061] The beneficial effects of this invention are as follows:

[0062] This invention encodes RNA sequences using nucleotide physicochemical properties and frequency encoding methods and dinucleotide chemical properties encoding methods, and then inputs them into a lncRNA m5C methylation site prediction model. The lncRNA m5C methylation site prediction model first uses a CNN convolutional module to capture local features, then uses Bi-LSTM to capture context-dependent features, and finally combines local features and context-dependent features to obtain combined features. The classification of lncRNA sequences is completed through a fully connected network. This invention improves the accuracy of lncRNA m5C methylation site prediction. Attached Figure Description

[0063] Figure 1 This is a flowchart of the present invention;

[0064] Figure 2 A performance comparison chart for single features and combined features;

[0065] Figure 3 The ROC curves of the model of this invention compared with those of other learning methods are shown below.

[0066] Figure 4 This is a PR curve diagram comparing the model of this invention with other learning methods;

[0067] Figure 5 The ROC curve is predicted by performing a five-fold cross-validation experiment on the model of this invention and other comparison methods.

[0068] Figure 6 The PR curve is a prediction plot of the model of this invention in a five-fold cross-validation experiment with other comparison methods. Detailed Implementation

[0069] Specific implementation method one: as follows Figure 1 As shown, the specific process of the deep learning-based lncRNA m5C methylation site prediction method in this embodiment is as follows:

[0070] Step 1: Obtain the original characteristics of the lncRNA sequence to be tested;

[0071] Step 11: Encode the physicochemical properties and cumulative frequency information of each nucleotide in the lncRNA sequence to be tested, and obtain the first coding feature matrix of the lncRNA sequence to be tested, specifically as follows:

[0072] First, the physicochemical properties of each nucleotide in the lncRNA sequence to be tested are encoded:

[0073] Adenine A, cytosine C, guanine G, and uracil U are represented by coordinates (1,1,1), (0,1,0), (1,0,0), and (0,0,1), respectively, to encode the physicochemical properties of nucleotides.

[0074] Then, the cumulative base frequency of the lncRNA sequence to be tested is obtained.

[0075] Among them, f i' d is the frequency of the i'th nucleotide appearing before the i'th nucleotide position. i' It represents the number of times the i'th nucleotide appears before the position of the i'th nucleotide, where i' is the nucleotide number in the lncRNA to be tested;

[0076] Finally, the physicochemical properties of each nucleotide in the lncRNA sequence to be tested and the corresponding cumulative base frequency are combined to form the first coding feature vector of the nucleotide, thereby obtaining the first coding feature matrix of the lncRNA sequence to be tested;

[0077] Steps 1 and 2: Encode the dinucleotide chemical properties of the lncRNA sequence to be tested to obtain the dinucleotide chemical property feature vector of the lncRNA sequence to be tested:

[0078] First, obtain the PCP matrix of the lncRNA sequence to be tested;

[0079] Then, the PCP matrix is ​​normalized to obtain the normalized PCP matrix, and the autocovariance matrix A' and crosscovariance matrix C' are obtained using the normalized PCP matrix:

[0080] Finally, matrices A' and C' are concatenated to obtain the dinucleotide chemical property feature vector of the lncRNA sequence to be tested;

[0081] Step 13: Concatenate the first coding feature vector and the dinucleotide chemical property feature vector of the lncRNA sequence to be tested to obtain the original feature vector of the lncRNA sequence to be tested.

[0082] Step 2: Input the original features of the lncRNA sequence to be tested into the trained lncRNA m5C methylation site prediction model to obtain the m5C site prediction results;

[0083] If the predicted value of the m5C site is greater than 0.5, it indicates that the m5C site exists in the lncRNA sequence to be tested; otherwise, it indicates that the m5C site does not exist.

[0084] The trained lncRNA m5C methylation site prediction model was obtained through the following method:

[0085] S1. Obtain the lncRNA sequence fragments corresponding to the methylation data files in the human gene bank, and construct training and testing sets using the lncRNA sequence fragments, specifically as follows:

[0086] S11. Obtain the lncRNA m5C site information from the RMBase v3.0 database or the m5C-Atlas database, and merge the lncRNA m5C site information from the RMBase v3.0 database with the lncRNA m5C site information from the corresponding m5C-Atlas database to obtain the total set of lncRNA m5C site information.

[0087] The site information of the lncRNA m5C includes: chromosome number, start position, end position, and positive / negative strand identifier;

[0088] S12. Based on the lncRNA m5C site information, multiple lncRNA sequence fragments were extracted from the human genome. Then, a 101nt sequence fragment was extracted and the CD-HIT tool was used to remove sequence fragments that had high similarity to other sequences (in this invention, the threshold in the CD-HIT tool is set to 0.9, and sequence fragments exceeding the set threshold are considered highly similar). The final set of sequence fragments was used as the positive sample set.

[0089] S13. Extract a 101nt sequence fragment centered on cytosine (C) upstream and downstream of the positive sample site. Then, use the CD-HIT tool to remove sequence fragments that have high similarity to other sequences (in this invention, the threshold in the CD-HIT tool is set to 0.9, and sequence fragments exceeding the set threshold are considered highly similar) as well as fragments that overlap with the positive sample site. The final set of sequence fragments is used as the negative sample set.

[0090] S14. Use 0 as the label for negative samples and 1 as the label for positive samples. Then combine the positive sample set and the negative sample set after setting the labels to obtain the dataset. Divide the dataset into training set and test set.

[0091] In this context, tag 0 indicates that the lncRNA sequence does not contain the m5C site; tag 1 indicates that the lncRNA sequence contains the m5C site.

[0092] In this step, taking into account the imbalance of positive and negative samples in actual RNA sequences, the ratio of positive to negative samples in the experimental dataset is approximately 1:3, and a five-fold cross-validation method is used for training on the training dataset.

[0093] S2. Sequence encoding is performed on the data in the training and test sets to obtain the original feature vectors of lncRNAs, resulting in the encoded training and test sets, specifically:

[0094] S21. Encode the nucleotide physicochemical properties and cumulative frequency information of the sequence fragments in the training and test sets to obtain the first coding feature matrix:

[0095] First, obtain the physicochemical property codes for each nucleotide in the sequence fragments in the training and test sets;

[0096] Adenine A, cytosine C, guanine G, and uracil U are represented by coordinates (1,1,1), (0,1,0), (1,0,0), and (0,0,1), respectively, to encode the physicochemical properties of nucleotides.

[0097] In this step, RNA is composed of four nucleotides: A, C, G, and U. Based on their chemical properties, they can be divided into three categories. In terms of the number of rings, adenine and guanine have two rings, while cytosine and uracil have one ring. In terms of chemical function, adenine and cytosine contain amino groups, while guanine and uracil contain ketone groups. For the formation of secondary structures, guanine and cytosine have strong hydrogen bonds, while adenine and uracil have weak hydrogen bonds. This invention uses three coordinates (x, y, z) to represent the chemical properties of the four nucleotides, where the x-coordinate represents the ring structure, the y-coordinate represents the functional group, and the z-coordinate represents the hydrogen bond, and these three coordinates are assigned values ​​of 0 and 1; the nucleotide S at the i-th position in the RNA sequence... i Use S i =(x i ,y i ,z i )coding;

[0098] Then, the cumulative base frequencies of sequence fragments in the training and test sets are obtained.

[0099] Among them, f i d is the frequency of the i-th nucleotide appearing before the position of the i-th nucleotide. i It represents the number of times the i-th nucleotide appears before the position of the i-th nucleotide;

[0100] This step encodes the nucleotide accumulation frequency, which is defined as the frequency at which the nucleotide at position i appears before position i; the accumulation frequency f of the nucleotide at position i. i =d i / i, where d i It is the sum of the number of times the i-th nucleotide appears in the first i nucleotides;

[0101] Finally, the physicochemical properties of each nucleotide and the corresponding cumulative base frequency are combined to form the first coding feature vector of the current nucleotide, thereby obtaining the first coding feature matrix corresponding to the sequence fragment;

[0102] If the sequence segment length is L, then the first coding feature matrix is ​​obtained as a 4*L dimensional matrix;

[0103] S22. Encode the dinucleotide chemical properties of the sequence fragment to obtain the dinucleotide chemical property feature vector of the sequence fragment:

[0104] First, obtain the PCP matrix of the sequence fragment:

[0105]

[0106] Where p∈[1,10], p is the PCP label, PCP p (Pj P j+1 ) is a dinucleotide P j P j+1 The p-th PCP value, where L is the sequence fragment length, j∈[1,L-1], and j is the nucleotide label;

[0107] Then, the PCP matrix is ​​normalized to obtain the normalized PCP matrix. Using the normalized PCP matrix, a 10*4 dinucleotide autocovariance matrix A and a 10*9*4 cross-covariance matrix C are obtained.

[0108]

[0109] Where s = 1, 2, 3, 4, and λ1 and λ2 are integers. It is the average value of the data in the λ1-th row of the normalized PCP matrix. It is the average value of the data in the λ2-th row of the normalized PCP matrix, PCP' p (P j P j+1 ) is the normalized dinucleotide P j P j+1 The p-th PCP value, It is the average value of the data in the p-th row of the normalized PCP matrix;

[0110] Finally, matrices A and C are concatenated to obtain the eigenvectors of dinucleotide chemical properties.

[0111] S23. Concatenate the first coding feature vector obtained in S21 and the dinucleotide chemical property feature vector obtained in S22 to obtain the original feature vector of the lncRNA sequence. Use the original feature vector of the lncRNA sequence and the label to form the encoded training set and test set.

[0112] S3. Use the encoded training set to train the lncRNA m5C methylation site prediction model to obtain the trained lncRNA m5C methylation site prediction model.

[0113] The lncRNA m5C methylation site prediction model includes: a CNN network module, a Bi-LSTM network module, a feature fusion module, and a fully connected module;

[0114] The CNN network module is used to perform preliminary feature extraction on the original feature vector of the lncRNA sequence, ignoring unimportant features, and obtaining local feature information F1 of the sequence fragment;

[0115] The CNN network module includes three convolutional units in sequence; each convolutional unit includes: a convolutional layer, an activation function, and a pooling layer;

[0116] The convolutional layer of the first convolutional unit is specifically as follows:

[0117]

[0118] Among them, X u+m,n is the original feature vector of lncRNA, u is the output position index, k is the kernel index, and conv(X) is the kernel index. uk It is the output of the convolutional layer. It is an M*N weight matrix, where M is the window size and N is the number of input channels;

[0119] The convolutional layers in other convolutional units follow the same formula as above, with the input being the output of the previous convolutional unit;

[0120] The activation function ReLU is expressed as:

[0121]

[0122] Where x is the input to the activation function; the input to each activation function is the output of the previous convolutional layer.

[0123] The pooling layer retains the extracted content while reducing its feature dimension, specifically:

[0124] pooling(x') u' =max(x'1,x'2,x'3,...,x' M' ) u'

[0125] Where x′ represents the input feature, the input x′ of each pooling layer is the output of the previous activation function, u' represents the output position index, M' represents the pooling window size, and x'1, x'2, x'3, ..., x' M' It is the sequence output by the previous activation function.

[0126] The Bi-LSTM network module uses feature F1 to obtain sequence features F2 that are context-dependent between sequence nucleotides, which contain sequence order dependency information;

[0127] Bi-LSTM's two-directional feature extraction makes dependency information extraction more complete. F1 is further processed by the Bi-LSTM network layer to obtain the sequence features F2.

[0128] The feature fusion module is used to concatenate features F1 and F2 to obtain the optimal feature vector F of the lncRNA sequence;

[0129] The fully connected layer performs a comprehensive analysis and nonlinear transformation on F through a series of weight matrices, maps it to the classification space, and outputs the predicted probability of whether the m5C site exists, thereby determining whether the input sequence has the m5C site.

[0130] If the predicted probability of the m5C site is greater than 0.5, it means that the m5C site exists in the current sequence fragment; otherwise, it means that it does not exist.

[0131] S4. The trained lncRNA m5C methylation site prediction model was evaluated using the test set. Precision, F1 score, recall, accuracy, and Matthews correlation coefficient (MCC) were used as evaluation metrics. The specific calculation formulas are as follows:

[0132]

[0133] In F1-Score, TP stands for True Positive, meaning both the actual and predicted labels of the sample are positive; TN stands for True Negative, meaning both the actual and predicted labels of the sample are negative; FP stands for False Positive, meaning both the actual and predicted labels of the sample are negative; and FN stands for False Negative, meaning both the actual and predicted labels of the sample are positive. In F1-Score, P stands for Precision and R stands for Recall. The Matthews correlation coefficient (MCC) represents the correlation between the prediction and the label (MCC = 0 indicates random guessing, and 1 indicates a perfect model).

[0134] Example: To verify the beneficial effects of the present invention, the following experiments were conducted:

[0135] A total of 2435 human lncRNA m5C methylation sites were collected from the RMBase v3.0 database and the m5C-Atlas database as positive samples. Based on the site information, non-m5C sites were extracted from the human genome as negative samples. Considering that there are far more non-m5C sites than m5C sites in the actual sequence, more non-m5C site sequences were extracted for the experiment, so 5703 data samples were screened.

[0136] Taking a specific lncRNA sequence as an example, 50 nt lengths of the start and end ends of the RNA sequence are taken. The sequence is encoded using nucleotide physicochemical properties, cumulative frequency, and dinucleotide chemical properties. The encoded results are then input into a convolutional neural network to capture local information. Next, a bidirectional long short-term memory network (Bi-LSTM) is used to obtain contextual features. The two are then combined for the prediction of the lncRNA m5C site.

[0137] To delve deeper into the impact of the information contained in features on experimental results, this embodiment conducts a detailed comparative analysis of single features and combined features. The experimental results are clearly presented in... Figure 2 From. Figure 2 As can be intuitively observed, although the two feature encoding methods each have their advantages and disadvantages in these performance metrics, and both achieve good performance, the combined feature method shows a significant improvement in all performance metrics, including ACC, Recall, Pre, F1, and MCC, compared to the single feature method. This series of specific data fully demonstrates that by combining different types of features, information from various aspects can be gathered, thereby obtaining more comprehensive and accurate results.

[0138] To fully and deeply reveal the significant advantages of this invention, a comprehensive comparison is made with other learning algorithms, including logistic regression, Naive Bayes, and support vector machines, as well as CNN and Bi-LSTM deep learning methods. To more intuitively illustrate the balance between accuracy and recall among different learning methods, ROC curves and PR curves are plotted, as shown in the figure. Figure 3 and Figure 4 As shown, it is immediately apparent that other learning algorithms performed relatively poorly, achieving lower experimental results. Among all the learning methods compared, this invention achieved a high level in both AUROC and AUPR, demonstrating the best prediction results and enabling more accurate classification of various types of samples.

[0139] To demonstrate the superior performance of the model constructed in this invention, it was compared with other existing cutting-edge prediction models, including iRNAm5C-PseDNC, iRNA-PseTNC, RNAm5CFinder, and m5CPred-SVM. The significant advantages of this invention are illustrated by comparison with known methods, and the comparison results are shown in Table 1. The accuracy (ACC) of this invention is 0.8457, precision (Pre) is 0.8260, recall reaches 0.6427, F1 score is 0.7189, and Matthews correlation coefficient (MCC) is 0.6186. Compared with existing models, this invention achieves higher performance in these metrics by 2.82%, 6.76%, 11.09%, 8.12%, and 8.11%, respectively, demonstrating significant progress. In addition to the above metrics used to evaluate model robustness, ROC and PR curves of this invention and other prediction methods were plotted to further verify the prediction performance, such as... Figure 5 and Figure 6As shown in the figure, by comparing the AUROC and AUPR values ​​in the corresponding regions under the ROC and PR curves, it can be seen that the prediction results of this invention on the training dataset are higher than those of the other four prediction methods, indicating that this invention can improve the robustness of the model.

[0140] Table 1 compares the performance of other related models.

[0141]

[0142] The detailed description of the embodiments of the present invention provided above is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A method for predicting lncRNA m5C methylation sites based on deep learning, characterized in that... The specific process of the method is as follows: Step 1: Obtain the original characteristics of the lncRNA sequence to be tested; Step 2: Input the original features of the lncRNA sequence to be tested into the trained lncRNA m5C methylation site prediction model to obtain the m5C site prediction results; If the predicted m5C site is greater than the preset threshold, it indicates that the m5C site exists in the lncRNA sequence to be tested; otherwise, it indicates that the m5C site does not exist in the lncRNA sequence to be tested. The trained lncRNA m5C methylation site prediction model was obtained through the following method: S1. Obtain the lncRNA sequence fragments corresponding to the methylation data files in the human gene bank, and construct a training set using the lncRNA sequence fragments, specifically: S11. Obtain the lncRNA m5C site information from the RMBase v3.0 database or the m5C-Atlas database, and merge the lncRNA m5C site information from the RMBase v3.0 database with the lncRNA m5C site information from the corresponding m5C-Atlas database to obtain the total set of lncRNA m5C site information. The site information of the lncRNA m5C includes: chromosome number, start position, end position, and positive / negative strand identifier; S12. Extract multiple lncRNA sequence fragments from the human genome based on the lncRNA m5C site information, and then truncate them to a predetermined length. The sequence fragments were processed using the CD-HIT tool to remove those with a similarity greater than a preset threshold, and the remaining set of sequence fragments was used as the positive sample set. S13. Extract sequence fragments of length L centered on cytosine C from upstream and downstream of the positive sample site. Then, use the CD-HIT tool to remove sequence fragments larger than the preset similarity threshold and fragments that overlap with the positive sample site. The remaining sequence fragment set is used as the negative sample set. S14. Label the positive and negative samples, and combine the labeled positive and negative sample sets to obtain the training set. In this study, the label for a positive sample was 1, and the label for a negative sample was 0. S2. Sequence encoding is performed on the data in the training set to obtain the original feature vector of lncRNA, thereby obtaining the encoded training set; S3. Use the encoded training set to train the lncRNA m5C methylation site prediction model to obtain the trained lncRNA m5C methylation site prediction model. The lncRNA m5C methylation site prediction model includes: a CNN network module, a Bi-LSTM network module, a feature fusion module, and a fully connected module; The CNN network module is used to perform preliminary feature extraction on the original feature vector of the lncRNA sequence to obtain local feature information F1 of the sequence fragment; The Bi-LSTM network module uses feature F1 to obtain sequence feature F2; The feature fusion module is used to concatenate features F1 and F2 to obtain the optimal feature vector F of the lncRNA sequence; The fully connected layer uses the optimal feature vector F of the lncRNA sequence to output the m5C site prediction result.

2. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 1, characterized in that: The process of obtaining the original characteristics of the lncRNA sequence to be tested in step one specifically includes: Step 11: Encode the physicochemical properties and cumulative frequency information of each nucleotide in the lncRNA sequence to be tested to obtain the first coding feature matrix of the lncRNA sequence to be tested; Step 1 and Step 2: Encode the dinucleotide chemical properties of the lncRNA sequence to be tested to obtain the dinucleotide chemical property feature vector of the lncRNA sequence to be tested; Step 13: Concatenate the first coding feature vector and the dinucleotide chemical property feature vector of the lncRNA sequence to be tested to obtain the original feature vector of the lncRNA sequence to be tested.

3. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 2, characterized in that: The encoding of the physicochemical properties and cumulative frequency information of each nucleotide in the lncRNA sequence to be tested in step one-to-one, to obtain the first coding feature matrix of the lncRNA sequence to be tested, is specifically as follows: First, the physicochemical properties of each nucleotide in the lncRNA sequence to be tested are encoded, specifically as follows: If the nucleotide is adenine A, the physicochemical property code is (1,1,1); if the nucleotide is cytosine C, the physicochemical property code is (0,1,0); if the nucleotide is guanine G, the physicochemical property code is (1,0,0); if the nucleotide is uracil U, the physicochemical property code is (0,0,1). Then, the cumulative base frequency of the lncRNA sequence to be tested is obtained. ; Among them, It is the first The nucleotide in the first The frequency of occurrence before each nucleotide position, It is the first The nucleotide in the first The number of times it appears before a nucleotide position, These are the nucleotide labels in the lncRNA to be tested; Finally, the physicochemical properties of nucleotides and the corresponding cumulative base frequencies in the lncRNA sequence to be tested are combined to form the first coding feature vector of nucleotides, thereby obtaining the first coding feature matrix of the lncRNA sequence to be tested.

4. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 3, characterized in that: In steps one and two, the dinucleotide chemical properties of the lncRNA sequence to be tested are encoded to obtain a feature vector of the dinucleotide chemical properties of the lncRNA sequence to be tested, specifically as follows: First, obtain the PCP matrix of the lncRNA sequence to be tested: in, , It is the PCP designation. It is a dinucleotide The p-th PCP value, It is the length of the lncRNA sequence to be tested. , These are nucleotide labels; Then, the PCP matrix is ​​normalized to obtain the normalized PCP matrix, and the dinucleotide autocovariance matrix is ​​obtained using the normalized PCP matrix. and cross-covariance matrix ; Finally, the matrix sum matrix The dinucleotide chemical property feature vector of the lncRNA sequence to be tested is obtained by tandem.

5. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 4, characterized in that: The dinucleotide autocovariance matrix is ​​obtained using the normalized PCP matrix. and cross-covariance matrix Specifically: Where s = 1, 2, 3, 4, , , It is an integer. , , It is the th in the normalized PCP matrix The average value of the row data It is the th in the normalized PCP matrix The average value of the row data It is a normalized dinucleotide. The p-th PCP value, It is the th in the normalized PCP matrix The average value of the row data.

6. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 5, characterized in that: In step S2, sequence encoding is performed on the data in the training set to obtain the original feature vector of lncRNA, thus obtaining the encoded training set. Specifically: S21. Encode the nucleotide physicochemical properties and cumulative frequency information of the sequence fragments in the training set to obtain the first coding feature matrix; S22. Encode the dinucleotide chemical properties of the sequence fragment to obtain the dinucleotide chemical property feature vector of the sequence fragment; S23. Concatenate the first coding feature vector obtained in S21 and the dinucleotide chemical property feature vector obtained in S22 to obtain the original feature vector of the lncRNA sequence. Use the original feature vector of the lncRNA sequence and the label to form the encoded training set.

7. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 6, characterized in that: The CNN network module includes: three convolutional units; Each convolutional unit includes, in sequence: a convolutional layer, an activation function layer, and a pooling layer; The convolutional layer in the first convolutional unit is specifically as follows: in, This is the original feature vector of the lncRNA, where u is the output position index and k is the kernel index. It is the output of the convolutional layer. It is an M * N weight matrix, where M is the window size and N is the number of input channels; The convolutional layers in other convolutional units follow the same formula as above, with the input being the output of the previous convolutional unit; The activation function layer is as follows: Here, x is the input to the activation function; the input to each activation function is x, which is the output of the previous convolutional layer.

8. The method for predicting lncRNA m5C methylation sites based on deep learning according to claim 7, characterized in that: The pooling layer is specifically: in, This represents the input features, the input of each pooling layer. This is the output of the previous activation function, where u' represents the output position index. Indicates the size of the pool window. It is the sequence output by the previous activation function.

Images