RNA N4-acetylcytidine modification site prediction method and system based on RNAErnie pre-training model

Abstract

The present invention relates to a method and system for predicting RNAN4-acetylcytidine modification sites based on an RNAErnie pre-training model. The method comprises: collecting an RNA sequence data set and inputting it into the RNAErnie pre-training model to capture context dependencies and extract global features; performing feature encoding by combining the RNAErnie pre-training model with six traditional feature encoding methods; performing feature dimensionality reduction in a deep neural network model and inputting the result into a soft voting integration model to integrate the prediction results of different classifiers, thereby obtaining RNAN4-acetylcytidine modification site prediction results. Multi-level masking is performed by the RNAErnie pre-training model, and the six traditional feature encoding methods are combined to capture the details and physicochemical properties of the sequence; feature dimensionality reduction is automatically performed using a deep neural network, which reduces computational complexity and retains key information; the reduced-dimensional features are input into the soft voting integration model, and the final prediction result is obtained by integrating multiple classifiers, which significantly improves the accuracy and robustness of the prediction.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method and system for predicting RNAN4-acetylcytidine modification sites based on an RNAErnie pre-training model. Background Art

[0002] To date, over 170 modified nucleosides have been discovered in RNA. Post-transcriptional chemical modifications of RNA, collectively referred to as the "epitranscriptome," have substantial impacts on gene expression and cellular processes, playing a crucial role in molecular interactions and intermolecular relationships. N4-acetylcytidine (ac4C), a common type of modification catalyzed by the enzyme NAT10, adds an acetyl group to the fourth nitrogen position of the cytidine base. Initially discovered in tRNA and rRNA of eukaryotes and prokaryotes, ac4C has been identified in human mRNA, where it can improve translation efficiency, enhance mRNA stability, and regulate gene expression. Furthermore, increasing evidence suggests that ac4C is associated with a variety of human diseases, including inflammation, metabolic disorders, autoimmune diseases, and cancer. In summary, ac4C modification, as a key post-transcriptional modification of RNA, plays an important role in cellular function and disease processes. Elucidating the functions and mechanisms of RNA-ac4C modification sites is crucial for elucidating their biological significance and advancing therapeutic strategies for related diseases.

[0003] Traditional RNA-ac4C modification site detection includes biological experimental methods, high-throughput sequencing technology, and computer-assisted analysis methods. Among them, traditional biological experimental methods are traditional methods for detecting ac4C modification in RNA, mainly including chemical analysis and immunoassay, which are widely used in laboratories, have high sensitivity and specificity, and are suitable for qualitative and quantitative analysis of specific modifications; high-throughput sequencing technology has been widely used in recent years for whole-genome level detection of ac4C modification, among which MeRIP-seq (mRNA immunoprecipitation sequencing) is the representative. This technology enriches the modification sites in RNA samples and then performs sequencing analysis, which can conduct large-scale exploration and analysis of ac4C modification across the whole genome; computer-assisted analysis methods are bioinformatics tools based on machine learning algorithms. By analyzing RNA sequence, structure, evolutionary conservation and other characteristics, a prediction model is constructed to help researchers quickly identify potential ac4C modification sites.

[0004] However, traditional biological experimental methods for detecting RNA-ac4C modification sites are often expensive and time-consuming, with high detection costs, complex operations, and low sensitivity and specificity. High-throughput sequencing technologies often rely on antibody enrichment, resulting in low resolution, and detection accuracy is limited by antibody quality and background noise. Computer-assisted analysis methods rely heavily on traditional feature encoding technologies, requiring complex feature engineering steps and lacking a comprehensive understanding of contextual semantic relationships. As a result, traditional methods for detecting RNA-ac4C modification sites often suffer from high costs and insufficient information mining, resulting in low accuracy in detecting RNA-ac4C modification sites. Summary of the Invention

[0005] Based on this, in order to solve the above technical problems, a method and system for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is provided, which can be fast, low-cost, and improve the accuracy of the model's RNAN4-acetylcytidine modification site detection.

[0006] A method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model, the method comprising:

[0007] Collecting an RNA sequence data set; the RNA sequence data set includes positive and negative samples;

[0008] Each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model for multi-level masking to capture context dependencies and extract global features; and the RNAErnie pre-training model is combined with six traditional feature encoding methods to perform feature encoding on each RNA sequence to obtain encoded high-dimensional features;

[0009] Inputting the encoded high-dimensional features into a deep neural network model to perform feature dimensionality reduction to obtain reduced-dimensional features;

[0010] Inputting the reduced-dimensional features into a soft voting ensemble model, integrating the prediction results of different classifiers through the soft voting ensemble model to obtain the RNAN4-acetylcytidine modification site prediction result;

[0011] The soft voting ensemble model is constructed by XGBoost, MLP, and CatBoost classifiers.

[0012] In one embodiment, the method further comprises:

[0013] Determining an evaluation indicator, and performing a performance evaluation on the soft voting ensemble model using a ten-fold cross validation method based on the evaluation indicator to obtain an evaluation result;

[0014] The evaluation indicators include sensitivity, specificity, accuracy, Matthews correlation coefficient, and area under the curve.

[0015] In one embodiment, the method further comprises:

[0016] Displaying a user interaction interface and obtaining the RNA sequence to be predicted through the user interaction interface;

[0017] Inputting the RNA sequence to be predicted into the soft voting integrated model, and outputting the RNAN4-acetylcytidine modification site prediction result corresponding to the RNA sequence to be predicted;

[0018] The prediction result of RNAN4-acetylcytidine modification site corresponding to the RNA sequence to be predicted is displayed in the user interaction interface.

[0019] In one embodiment, after collecting the RNA sequence dataset, the method further comprises:

[0020] Determine the data set division ratio;

[0021] Based on the data set division ratio, the RNA sequence data set is subjected to stratified sampling processing to obtain a divided training data set and a test data set.

[0022] In one embodiment, the RNAErnie pre-trained model is based on a framework for enhancing representation through knowledge integration, and combines Transformer layers and a multi-head self-attention mechanism; wherein:

[0023] Each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model, and based on the multi-head self-attention mechanism, the attention score corresponding to each RNA sequence is calculated by each attention head;

[0024] The obtained attention scores are connected, and each RNA sequence is mapped to a query, key and value matrix through a linear transformation matrix.

[0025] In one embodiment, each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model for multi-level masking, including:

[0026] The RNAErnie pre-training model adopts motif-level masking, subsequence-level masking, and motif-level random masking strategies, combined with coarse-grained RNA as vocabulary markers;

[0027] The RNAErnie pre-training model appends the vocabulary token to the last segment of each RNA sequence to enhance the RNA sequence representation.

[0028] In one embodiment, the high-dimensional features are input into a deep neural network model to perform feature dimensionality reduction to obtain the reduced features, including:

[0029] The encoded high-dimensional features are input into a deep neural network model, and the high-dimensional features are mapped from a high-dimensional space to a low-dimensional space through the multi-layer nonlinear mapping of the deep neural network model to obtain features after dimensionality reduction.

[0030] In one embodiment, the prediction results of different classifiers are integrated by the soft voting ensemble model to obtain the RNAN4-acetylcytidine modification site prediction results, including:

[0031] Determining the prediction probabilities of all classifiers by the soft voting ensemble model;

[0032] Performing weighted average calculation on each of the predicted probabilities to obtain the maximum value of the weighted average probability;

[0033] The maximum value is used as the RNAN4-acetylcytidine modification site prediction result.

[0034] A RNAN4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model, the system comprising:

[0035] A data set acquisition module is used to acquire an RNA sequence data set; the RNA sequence data set includes positive and negative samples;

[0036] A feature encoding module is used to input each RNA sequence in the RNA sequence dataset into the RNAErnie pre-training model for multi-level masking, capture context dependencies and extract global features; and combine the RNAErnie pre-training model with six traditional feature encoding methods to perform feature encoding on each RNA sequence to obtain encoded high-dimensional features;

[0037] A feature dimensionality reduction module is used to input the encoded high-dimensional features into a deep neural network model to perform feature dimensionality reduction and obtain features after dimensionality reduction;

[0038] A result prediction module is used to input the reduced-dimensional features into a soft voting ensemble model, and integrate the prediction results of different classifiers through the soft voting ensemble model to obtain the RNAN4-acetylcytidine modification site prediction result;

[0039] The soft voting ensemble model is constructed by XGBoost, MLP, and CatBoost classifiers.

[0040] The above-mentioned RNAN4-acetylcytidine modification site prediction method and system based on the RNAErnie pre-training model performs multi-level masking through the RNAErnie pre-training model, which can accurately capture contextual dependencies and extract global features, capture more comprehensive RNA sequence information, and combine six traditional feature encoding methods to capture sequence details and physical and chemical properties; use deep neural networks to automatically perform feature dimensionality reduction, and reduce computational complexity and retain key information by automatically learning and screening the most relevant features; input the reduced dimensionality features into the soft voting integration model, and obtain the final prediction results by integrating multiple classifiers, which significantly improves the accuracy and robustness of the prediction, and can quickly and cost-effectively improve the model's RNAN4-acetylcytidine modification site detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a diagram illustrating the application environment of a method for predicting RNAN4-acetylcytidine modification sites based on an RNAErnie pre-training model in one embodiment;

[0042] Figure 2 Schematic diagram of a process for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model in one embodiment;

[0043] Figure 3 Schematic diagram of the model framework of Voting-ac4C in one embodiment;

[0044] Figure 4 Schematic diagram of parameters used to verify the reliability of hybrid feature encoding in ablation experiments;

[0045] Figure 5 Schematic diagram of the parameters used to verify the reliability of the selected machine learning classifier for ablation experiments;

[0046] Figure 6 Schematic diagram of parameters used to verify the reliability of the soft voting method in ablation experiments;

[0047] Figure 7 1 is a structural block diagram of an RNAN4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model in one embodiment;

[0048] Figure 8 This is a structural block diagram of an RNAN4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model in another embodiment;

[0049] Figure 9 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0051] The RNAN4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Figure 1 As shown, the application environment includes a computer device 110. The computer device 110 can collect an RNA sequence dataset; the RNA sequence dataset includes positive and negative samples; the computer device 110 can input each RNA sequence in the RNA sequence dataset into the RNAErnie pre-trained model for multi-level masking, thereby capturing contextual dependencies and extracting global features; and the RNAErnie pre-trained model is combined with six traditional feature encoding methods to encode each RNA sequence to obtain encoded high-dimensional features; the computer device 110 can input the encoded high-dimensional features into a deep neural network model for feature dimensionality reduction to obtain reduced features; the computer device 110 can input the reduced features into a soft voting ensemble model, and the soft voting ensemble model integrates the prediction results of different classifiers to obtain RNAN4-acetylcytidine modification site prediction results; wherein the soft voting ensemble model is constructed by XGBoost, MLP, and CatBoost classifiers. The computer device 110 can be, but is not limited to, various personal computers, laptops, smart phones, robots, unmanned aerial vehicles, tablet computers, and other devices.

[0052] In one embodiment, Figure 2 As shown, a method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is provided, comprising the following steps:

[0053] Step 202: Collect an RNA sequence data set; the RNA sequence data set includes positive and negative samples.

[0054] Establishing a reliable benchmark dataset is fundamental to developing robust predictions of the underlying mechanisms of ac4C modification sites. In this example, a pre-designed and well-prepared RNA sequence dataset can be used, specifically including 2758 balanced positive and negative samples. The dataset is divided into a training set (2206 samples) and a test set (552 samples). Positive and negative samples are evenly distributed in the training and test datasets to ensure the stability and generalization ability of the model.

[0055] In one embodiment, a method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is provided, which may also include a process of dividing the data set ratio. The specific process includes: determining the data set division ratio; based on the data set division ratio, performing stratified sampling processing on the RNA sequence data set to obtain a divided training data set and a test data set.

[0056] Specifically, in this embodiment, all samples were stratified and divided into training and test datasets at a ratio of 4:1. The training dataset included 2206 positive and negative samples. In contrast, the independent test dataset consisted of 552 samples from each category, including both positive and negative.

[0057] In step 204, each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model for multi-level masking to capture context dependencies and extract global features; and the RNAErnie pre-training model is combined with six traditional feature encoding methods to perform feature encoding on each RNA sequence to obtain encoded high-dimensional features.

[0058] For each RNA sequence in the RNA sequence dataset, global contextual features are obtained through the RNAErnie pre-trained model. The RNAErnie pre-trained model is used to encode features of RNA sequences. RNAErnie is a pre-trained model based on a Transformer architecture designed specifically for RNA sequences. It is built on the Enhanced Representation through Knowledge Integration (ERNIE) framework and combines multiple Transformer layers and a multi-head self-attention mechanism. The hidden state dimension of each Transformer block is 768. These design choices enable the RNAErnie pre-trained model to capture complex patterns and deep biological information in RNA sequences.

[0059] In one embodiment, the RNAErnie pre-training model is based on a representation framework enhanced by knowledge integration, and is combined with a Transformer layer and a multi-head self-attention mechanism; a method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is provided, which can also include a process of feature processing using the RNAErnie pre-training model, the specific process including: inputting each RNA sequence in the RNA sequence data set into the RNAErnie pre-training model respectively, and calculating the attention score corresponding to each RNA sequence through each attention head based on the multi-head self-attention mechanism; connecting the obtained attention scores, and mapping each RNA sequence to a query, key, and value matrix through a linear transformation matrix.

[0060] In this embodiment, the RNAErnie pre-training model is based on the Transformer architecture, one of its core components is the multi-head attention mechanism, which captures different aspects of the input sequence by computing multiple attention heads in parallel, thereby enhancing the ability to understand RNA sequences. For each attention head, W represents a set of linear transformation matrices. Specifically, W Q 、W K and W V are the linear transformation matrices used to map the input sequence X to the query (Q), key (K), and value (V) matrices. The calculation formula involved in this process is: Q = XW Q ; K = XW K ; V = XW V ; Each head calculates the attention score through the self-attention mechanism to evaluate the relationship between the elements of the input sequence. The calculation formula is: Among them, the dimension of the key vector is d k The dot product of the query matrix Q and the key matrix K is used to scale to prevent gradient problems. This scaled dot product generates the original attention score, which is then converted into a probability distribution through the softmax function. This distribution is used to weight the vectors in the value matrix V to produce a contextual representation of the sequence. The outputs of multiple attention heads are connected and then linearly transformed to obtain the final result. The calculation formula involved in this process is: MultiHead(Q,K,V)=Concat(head1,head2,…,headh)W O , where the concatenation operator Concat merges the outputs of the various heads, and W O is a trainable transformation matrix that maps the combined features to the desired output space.

[0061] In one embodiment, a method for predicting RNAN4-acetylcytidine modification sites based on an RNAErnie pre-training model is provided, which may also include a multi-level masking process, the specific process including: the RNAErnie pre-training model adopts motif-level masking, subsequence-level masking, and motif-level random masking strategies, combined with coarse-grained RNA as vocabulary tags; the RNAErnie pre-training model appends the vocabulary tags to the last segment of each RNA sequence to enhance the RNA sequence representation.

[0062] Among them, the RNAErnie pre-training model adopts three masking strategies to enhance RNA sequence representation: motif-level masking refers to masking single nucleotides to capture local features; subsequence-level masking targets continuous fragments of RNA sequences, aiming to capture long-term dependencies and global features; motif-level random masking introduces randomness by masking nucleotides without considering their position or continuity, thereby improving the robustness and generalization ability of the model.

[0063] In addition, the RNAErnie pre-training model incorporates coarse-grained RNA types (e.g., mRNA, miRNA, lncRNA) as special vocabulary tags, appending these type tags to the last segment of each RNA sequence during pre-training. This strategy enables the model to identify and utilize RNA type-specific features when processing various downstream tasks, thereby enhancing the model's domain adaptability and its ability to generalize across multiple tasks.

[0064] Among them, six traditional feature encoding methods are combined for feature encoding, namely: One-hot, ENAC, C2, ND, TPCP, and Ksnpf feature encoding methods. Specifically, the One-hot encoding method: used in bioinformatics to represent nucleotide sequences, representing the four RNA bases in RNA molecules, adenine (A), cytosine (C), guanine (G), and uracil (U), as binary vectors composed of 0 and 1. Specifically, this means that the nucleotides A, C, G, and U are represented by four vectors (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0), and (0, 0, 0, 1), respectively; the ENAC encoding method: using a sliding window method to calculate the nucleic acid composition within a window of fixed length, generating a feature vector for each window, which can capture local structural information in the sequence and provide useful feature representation for subsequent analysis and modeling; the C2 encoding method: converting the elements of the biological sequence into specific values ​​from the perspective of the global sequence, and converting the RNA bases in the nucleotide sequence of the RNA molecule into 2-bit binary For example, adenine (A) is encoded as (0, 0), cytosine (C) is encoded as (1, 1), guanine (G) is encoded as (1, 0), and uracil (U) is encoded as (0, 1); ND encoding method: the main principle is to regard one or more bases in the RNA sequence as a single element and calculate their frequency in the sequence; TPCP encoding method: 11 physicochemical properties are used. For each sequence window containing TPCP, a 1375-dimensional vector (i.e., 125 trinucleotide combinations × 11 physicochemical properties) is generated. Any trinucleotide containing nucleotide N is set to zero; Ksnpf encoding method: the occurrence of 16 nucleotide pairs separated by k arbitrary nucleotides in the sequence is quantified. By setting k to values ​​of 0, 1, 2, 3, and 4, the sequence is converted into various feature representations reflecting the frequency of these nucleotide pairs at different intervals.

[0065] In this embodiment, a multi-dimensional feature encoding method is adopted for each RNA sequence. The global context features are obtained through the RNAErnie pre-training model. Combined with six traditional encoding methods, including One-hot, ENAC, C2, ND, TPCP and Ksnpf, rich features of the RNA sequence are extracted from multiple angles such as physicochemical properties and position specificity. The hybrid feature representation of these multiple encoding methods can comprehensively reflect the local and global information of the sequence, effectively enhancing the feature expression ability of the model.

[0066] In step 206, the encoded high-dimensional features are input into a deep neural network model to perform feature dimensionality reduction to obtain reduced-dimensional features.

[0067] The generated high-dimensional features are input into the deep neural network model for feature dimensionality reduction. The deep neural network automatically learns and filters features, retains the most predictive features, reduces redundant information, and reduces computational complexity.

[0068] In one embodiment, a method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is provided, which can also include a process of feature dimensionality reduction. The specific process includes: inputting the encoded high-dimensional features into a deep neural network model, and mapping the high-dimensional features from the high-dimensional space to the low-dimensional space through the multi-layer nonlinear mapping of the deep neural network model to obtain the reduced-dimensional features.

[0069] In the model construction process of this embodiment, the hybrid features that integrate the RNAErnie pre-training model and six traditional feature encoding methods are input into the deep neural network (DNN) for feature dimensionality reduction. Through the multi-layer nonlinear mapping of the DNN, the features are mapped from high-dimensional space to low-dimensional space, which effectively reduces the computational complexity and storage requirements while retaining the main features that are critical to prediction. In addition, DNN can capture the complex interactions between different features and achieve a higher level of feature expression. This feature dimensionality reduction method not only avoids the interference of redundant information, but also improves the model's understanding of the importance of features, thereby significantly improving the accuracy of predictions.

[0070] In step 208 , the reduced-dimensional features are input into a soft voting ensemble model, and the prediction results of different classifiers are integrated through the soft voting ensemble model to obtain the RNAN4-acetylcytidine modification site prediction result.

[0071] The soft voting ensemble model is constructed using XGBoost, MLP, and CatBoost classifiers. The reduced features are fed into the soft voting ensemble model, which integrates the prediction results of these different classifiers to improve the model's predictive performance and stability.

[0072] In one embodiment, a method for predicting RNAN4-acetylcytidine modification sites based on an RNAErnie pre-training model is provided, which may also include a prediction process using a soft voting integration model, the specific process including: determining the prediction probability of all classifiers using a soft voting integration model; performing a weighted average calculation on each prediction probability to obtain the maximum value of the weighted average probability; and using the maximum value as the RNAN4-acetylcytidine modification site prediction result.

[0073] Soft voting is an ensemble learning method that calculates the final prediction probability by taking a weighted average of the prediction probabilities of all classifiers. The core formula of the soft voting model is: In the formula, represents the final prediction, c represents the category, n is the number of base models, w i is the weight of the i-th model, and P i (y=c|x) is the probability that the sample predicted by the i-th model belongs to class c. The final class prediction is based on the maximum value of the weighted average probability, that is, the class with the highest predicted probability is selected as the final prediction result.

[0074] In one embodiment, a method for predicting RNAN4-acetylcytidine modification sites based on an RNAErnie pre-training model is provided, which may also include a process for performing performance evaluation on a soft voting ensemble model. The specific process includes: determining an evaluation indicator, and performing a performance evaluation on the soft voting ensemble model using a ten-fold cross-validation method based on the evaluation indicator to obtain an evaluation result; wherein the evaluation indicator includes sensitivity, specificity, accuracy, Matthews correlation coefficient, and area under the curve.

[0075] Among them, ten-fold cross validation and independent test sets were used to comprehensively evaluate the performance of the model. The evaluation indicators included sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC) and area under the curve (AUC). The evaluation method ensured that the model had good generalization performance and robust predictive ability.

[0076] In order to evaluate the performance of the model, in this embodiment, five commonly used evaluation indicators are used, including sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC) and area under the curve (AUC). The formulas of these indicators are expressed as follows: Among them, TP, TN, FN, and FP represent the number of true positives, true negatives, false negatives, and false positives, respectively; SN represents the proportion of correctly identified positive samples; SP represents the proportion of correctly identified negative samples; ACC represents the proportion of all samples correctly classified; and MCC measures the correlation between true and predicted values, ranging from -1 to 1. Furthermore, to comprehensively compare the performance of different models, it is necessary to calculate evaluation metrics based on the area under the receiver operating characteristic (ROC) curve, which shows the ratio of true positives to false positives. AUC values ​​range from 0 to 1, with higher AUC values ​​indicating better predictive performance of the underlying model. Higher values ​​for these five metrics indicate better model performance.

[0077] In one embodiment, a method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is provided, which may also include a process of establishing a website for user interaction, the specific process including: displaying a user interaction interface, and obtaining the RNA sequence to be predicted through the user interaction interface; inputting the RNA sequence to be predicted into the soft voting integration model, and outputting the RNAN4-acetylcytidine modification site prediction result corresponding to the RNA sequence to be predicted; and displaying the RNA N4-acetylcytidine modification site prediction result corresponding to the RNA sequence to be predicted in the user interaction interface.

[0078] A user-friendly online platform can be established on a computer device to provide an intuitive interface, where users can easily input RNA sequences and quickly obtain prediction results for ac4C modification sites. The design of the platform makes the application of the RNAN4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model provided in this application more convenient and popular.

[0079] In one embodiment, Figure 3 As shown, an application model framework for the RNAN4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model is provided, which mainly includes five parts: A. data acquisition; B. feature encoding; C. feature dimensionality reduction; D. model evaluation; and E. website service.

[0080] Specifically, such as Figure 3 As shown, first, an RNA sequence dataset is collected, specifically, a dataset of each RNA sequence is collected from the database, and then divided into a training set and a test set; then, the RNAErnie pre-training model and six traditional feature encoding methods are used to encode the RNA sequence to obtain high-dimensional features; the high-dimensional features are input into a deep neural network model for feature dimensionality reduction; the reduced features are input into a soft voting integration model; the prediction results of different classifiers are integrated through the soft voting integration model to obtain the RNA N4-acetylcytidine modification site prediction results.

[0081] In one embodiment, in order to verify the superiority of the soft voting integrated model provided in this application in predicting ac4C, namely the Voting-ac4C model, the experimental results are compared with the existing ten-fold cross validation results as follows:

[0082] The Voting-ac4C model was deeply compared with several typical existing models, including PACES, XG-ac4C, iRNA-ac4C, Auto-ac4C, and ac4C-AFL. The comparison results are shown in the table:

[0083]

[0084] The 10-fold cross-validation results in the table show that Voting-ac4C outperforms all existing models across multiple key performance metrics, including sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC), and area under the curve (AUC). This demonstrates that Voting-ac4C not only significantly improves accuracy but also balances predictive power across different categories while maintaining high sensitivity and specificity. Furthermore, Voting-ac4C combines a pre-trained large language model with multiple traditional feature encoding methods to capture more comprehensive RNA sequence information, achieving breakthrough performance and further validating its effectiveness and reliability as a tool for predicting ac4C modification sites.

[0085] Next, the experimental results of the Voting-ac4C model and the existing models on independent dataset tests are compared as follows:

[0086] The test results on independent datasets show that the Voting-ac4C model performs well in all performance indicators. The comparison results are shown in the following table:

[0087]

[0088]

[0089] As shown in the table, the Voting-ac4C model achieved sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC), and area under the curve (AUC) of 85.14%, 81.15%, 83.15%, 66.35%, and 88.73%, respectively. These results demonstrate that the Voting-ac4C model not only exhibits high accuracy in the ac4C modification site prediction task, but also possesses strong sensitivity and specificity, effectively distinguishing positive and negative samples. Compared to existing models, Voting-ac4C significantly outperforms in overall performance, particularly in accuracy and MCC, demonstrating its ability to achieve balanced performance across different prediction categories. This further validates the robustness and practicality of the Voting-ac4C model for widespread application in ac4C modification site prediction.

[0090] Next, the performance comparison of the RNAErnie pre-trained model combined with other encoding methods is as follows:

[0091] First, we evaluated the model performance of RNAErnie combined with each traditional feature encoding method separately. The results are shown in the following table:

[0092]

[0093] It can be seen that the performance of single encoding methods reveals their respective advantages and disadvantages, among which Ksnpf and ENAC have superior performance, reaching 78.07% and 75.54% accuracy (ACC) respectively, but they still do not achieve high prediction performance.

[0094] After evaluating individual methods, this example explored the impact of feature encoding combinations on model performance. By gradually combining different encoding methods, it was found that the model performance improved significantly, as shown in the following table:

[0095]

[0096]

[0097] As shown in the table above, combining Ksnpf and ENAC encoding increased the accuracy (ACC) to 80.34%, a significant improvement over the results of using either encoding method alone (78.07% and 75.54%). Further adding C2 encoding increased the ACC to 80.71%, demonstrating that the inclusion of diverse features helps capture more sequence information. When one-hot, ND, and TPCP encodings were added, the final combined accuracy reached 83.15%, significantly higher than RNAErnie or any of the traditional encoding methods alone (the best result was 78.07%). Furthermore, the Matthews correlation coefficient (MCC) increased from a peak of 56.16% to 66.35%, and the AUC improved from 86.47% to 88.73%. These results demonstrate the significant advantages of feature combination, demonstrating that by integrating different encoding methods, the model can better utilize the multidimensional information of RNA sequences, thereby improving prediction performance and achieving a "1+1>2" effect.

[0098] In this embodiment, if Figure 4 As shown, Figure 4 It was demonstrated that the performance of the model after the combination of multiple features was significantly better than that of a single encoding method. The results verified that the RNAErnie pre-training model proposed in this application combined with a hybrid strategy of six traditional feature encodings can more comprehensively characterize the characteristics of RNA sequences, thereby improving the accuracy and stability of predictions, demonstrating the innovation and advantages of this application.

[0099] Next, the comparison between the soft voting ensemble model in this application and other machine learning classifiers is as follows:

[0100] When constructing the RNA-ac4C modification site prediction model, XGBoost, CatBoost and MLP were selected as the base classifiers of the soft voting ensemble learning model. In order to further verify the effectiveness of these classifiers, the performance of six different machine learning classifiers was compared, such as Figure 5 As shown in the figure, XGBoost, CatBoost, and MLP performed well across multiple metrics, including sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC), and area under the curve (AUC), significantly outperforming other classifiers. Therefore, these three classifiers are ideal for building soft voting ensemble learning models, effectively improving the overall performance of the prediction model.

[0101] The comparison between the soft voting ensemble model and other ensemble learning methods is as follows:

[0102] In order to further verify the stability and reliability of the model, several common ensemble learning methods were compared, including blending, stacking, bagging, hard voting, and soft voting. Figure 6 As shown in the figure, Soft Voting outperforms other methods in all evaluation indicators, proving that it has significant advantages in improving model prediction performance.

[0103] It should be understood that, although the various steps in the above flow chart are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the above flow chart may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0104] In one embodiment, Figure 7 As shown, a RNA N4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model is provided, including: a data set acquisition module 710, a feature encoding module 720, a feature dimensionality reduction module 730 and a result prediction module 740, wherein:

[0105] The data set acquisition module 710 is used to acquire RNA sequence data sets; the RNA sequence data sets include positive and negative samples;

[0106] Feature encoding module 720 is used to input each RNA sequence in the RNA sequence dataset into the RNAErnie pre-trained model for multi-level masking to capture context dependencies and extract global features; and combine the RNAErnie pre-trained model with six traditional feature encoding methods to perform feature encoding on each RNA sequence to obtain encoded high-dimensional features;

[0107] A feature dimensionality reduction module 730 is used to input the encoded high-dimensional features into a deep neural network model to perform feature dimensionality reduction and obtain reduced-dimensional features;

[0108] The result prediction module 740 is used to input the reduced-dimensional features into a soft voting ensemble model, integrate the prediction results of different classifiers through the soft voting ensemble model, and obtain the RNAN4-acetylcytidine modification site prediction result; wherein the soft voting ensemble model is constructed by XGBoost, MLP, and CatBoost classifiers.

[0109] like Figure 8 As shown, in one embodiment, a RNA N4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model is provided, which also includes a model evaluation module 750 for determining evaluation indicators and performing a performance evaluation on the soft voting ensemble model using a ten-fold cross-validation method according to the evaluation indicators to obtain an evaluation result; wherein the evaluation indicators include sensitivity, specificity, accuracy, Matthews correlation coefficient, area under the curve,

[0110] like Figure 8 As shown, in one embodiment, a RNA N4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model is provided, which also includes a user interaction module 760 for displaying a user interaction interface and obtaining an RNA sequence to be predicted through the user interaction interface; inputting the RNA sequence to be predicted into the soft voting integration model, and outputting the RNAN4-acetylcytidine modification site prediction result corresponding to the RNA sequence to be predicted; and displaying the RNAN4-acetylcytidine modification site prediction result corresponding to the RNA sequence to be predicted in the user interaction interface.

[0111] In one embodiment, the dataset collection module 710 is further configured to determine a dataset division ratio; based on the dataset division ratio, the RNA sequence dataset is subjected to stratified sampling to obtain a divided training dataset and a test dataset.

[0112] In one embodiment, the RNAErnie pre-training model is based on a representation framework enhanced by knowledge integration, and is combined with a Transformer layer and a multi-head self-attention mechanism; the feature encoding module 730 is also used to input each RNA sequence in the RNA sequence dataset into the RNAErnie pre-training model separately, and based on the multi-head self-attention mechanism, the attention score corresponding to each RNA sequence is calculated by each attention head; the obtained attention scores are connected, and each RNA sequence is mapped to a query, key, and value matrix through a linear transformation matrix.

[0113] In one embodiment, the feature encoding module 720 is also used in the RNAErnie pre-training model to adopt motif-level masking, subsequence-level masking, and motif-level random masking strategies, combined with coarse-grained RNA as vocabulary tags; the RNAErnie pre-training model appends vocabulary tags to the last segment of each RNA sequence to enhance the RNA sequence representation.

[0114] In one embodiment, the feature dimensionality reduction module 730 is also used to input the encoded high-dimensional features into the deep neural network model, and through the multi-layer nonlinear mapping of the deep neural network model, the high-dimensional features are mapped from the high-dimensional space to the low-dimensional space to obtain the reduced-dimensional features.

[0115] In one embodiment, the result prediction module 740 is further used to determine the prediction probability of all classifiers through a soft voting integration model; perform weighted average calculation on each prediction probability to obtain the maximum value of the weighted average probability; and use the maximum value as the RNAN4-acetylcytidine modification site prediction result.

[0116] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model is implemented. The display screen of the computer device can be a liquid crystal display or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0117] Those skilled in the art will understand that Figure 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0118] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the RNAN4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model are implemented.

[0119] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for predicting RNAN4-acetylcytidine modification sites based on the RNAErnie pre-training model are implemented.

[0120] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0121] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for predicting RNA N4-acetylcytidine modification sites based on the RNAErnie pre-training model, characterized in that: The method comprises: Collecting an RNA sequence data set; the RNA sequence data set includes positive and negative samples; Each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model for multi-level masking to capture context dependencies and extract global features; the RNAErnie pre-training model is combined with six traditional feature encoding methods to perform feature encoding on each RNA sequence to obtain encoded high-dimensional features; the six feature encoding methods are: One-hot, ENAC, C2, ND, TPCP, and Ksnpf feature encoding methods; Inputting the encoded high-dimensional features into a deep neural network model to perform feature dimensionality reduction to obtain reduced-dimensional features; Inputting the reduced-dimensional features into a soft voting ensemble model, integrating the prediction results of different classifiers through the soft voting ensemble model to obtain the RNA N4-acetylcytidine modification site prediction result; The soft voting ensemble model is constructed by XGBoost, MLP, and CatBoost classifiers.

2. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that The method further comprises: Determining an evaluation indicator, and performing a performance evaluation on the soft voting ensemble model using a ten-fold cross validation method based on the evaluation indicator to obtain an evaluation result; The evaluation indicators include sensitivity, specificity, accuracy, Matthews correlation coefficient, and area under the curve.

3. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that The method further comprises: Displaying a user interaction interface and obtaining the RNA sequence to be predicted through the user interaction interface; Inputting the RNA sequence to be predicted into the soft voting integrated model, and outputting the RNA N4-acetylcytidine modification site prediction result corresponding to the RNA sequence to be predicted; The prediction result of the RNA N4-acetylcytidine modification site corresponding to the RNA sequence to be predicted is displayed in the user interaction interface.

4. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that After acquiring the RNA sequence dataset, the method further comprises: Determine the data set division ratio; Based on the data set division ratio, the RNA sequence data set is subjected to stratified sampling processing to obtain a divided training data set and a test data set.

5. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that The RNAErnie pre-trained model is based on a framework for enhancing representation through knowledge integration, and combines Transformer layers and a multi-head self-attention mechanism; Each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model, and based on the multi-head self-attention mechanism, the attention score corresponding to each RNA sequence is calculated by each attention head; The obtained attention scores are connected, and each RNA sequence is mapped to a query, key and value matrix through a linear transformation matrix.

6. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that Each RNA sequence in the RNA sequence dataset is input into the RNAErnie pre-training model for multi-level masking, including: The RNAErnie pre-training model adopts motif-level masking, subsequence-level masking, and motif-level random masking strategies, combined with coarse-grained RNA as vocabulary markers; The RNAErnie pre-training model appends the vocabulary token to the last segment of each RNA sequence to enhance the RNA sequence representation.

7. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that The encoded high-dimensional features are input into a deep neural network model for feature dimensionality reduction to obtain the reduced-dimensional features, including: The encoded high-dimensional features are input into a deep neural network model, and the high-dimensional features are mapped from a high-dimensional space to a low-dimensional space through the multi-layer nonlinear mapping of the deep neural network model to obtain features after dimensionality reduction.

8. The RNA N4-acetylcytidine modification site prediction method based on the RNAErnie pre-training model according to claim 1, characterized in that The prediction results of different classifiers are integrated through the soft voting ensemble model to obtain the RNA N4-acetylcytidine modification site prediction results, including: Determining the prediction probabilities of all classifiers by the soft voting ensemble model; Performing weighted average calculation on each of the predicted probabilities to obtain the maximum value of the weighted average probability; The maximum value is used as the RNA N4-acetylcytidine modification site prediction result.

9. A RNA N4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model, characterized in that: The system comprises: A data set acquisition module is used to acquire an RNA sequence data set; the RNA sequence data set includes positive and negative samples; A feature encoding module is used to input each RNA sequence in the RNA sequence dataset into the RNAErnie pre-trained model for multi-level masking, capture context dependencies and extract global features; and combine the RNAErnie pre-trained model with six traditional feature encoding methods to perform feature encoding on each RNA sequence to obtain encoded high-dimensional features; wherein the six feature encoding methods are: One-hot, ENAC, C2, ND, TPCP, and Ksnpf feature encoding methods; A feature dimensionality reduction module is used to perform feature dimensionality reduction on the high-dimensional features input into a deep neural network model to obtain features after dimensionality reduction; A result prediction module is used to input the reduced-dimensional features into a soft voting ensemble model, and integrate the prediction results of different classifiers through the soft voting ensemble model to obtain the RNA N4-acetylcytidine modification site prediction result; The soft voting ensemble model is constructed by XGBoost, MLP, and CatBoost classifiers.

10. The RNA N4-acetylcytidine modification site prediction system based on the RNAErnie pre-training model according to claim 9, characterized in that The system also includes a model evaluation module for determining evaluation indicators and performing performance evaluation on the soft voting ensemble model using a ten-fold cross-validation method based on the evaluation indicators to obtain evaluation results; wherein the evaluation indicators include sensitivity, specificity, accuracy, Matthews correlation coefficient, and area under the curve.

Images