Identification and prediction method for 2OM site in RNA sequence
By constructing a lightweight machine learning model and combining multiple feature extraction methods and feature selection algorithms, the problems of low computational efficiency and poor prediction performance of existing 2OM site prediction models are solved, and efficient and accurate RNA sequence 2OM site identification is achieved, which is suitable for large-scale biological research.
Patent Information
- Application Number
- CN202510784387.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-12
- Publication Date
- 2025-09-19
AI Technical Summary
Existing 2OM site prediction models are computationally inefficient, have poor prediction performance, perform poorly on new or diverse datasets, lack generalization capabilities, and high-throughput experimental methods are costly and unsuitable for large-scale studies.
An RNA sequence dataset was constructed, and K-mer, ENAC, PS2 and DPCP_2 feature extraction methods were used. Combined with random forest and forward feature selection methods, the model was trained using LGBM, GBDT, XGBoost, RF, and Adaboost algorithms. The optimal prediction model was selected for different bases, and a lightweight machine learning algorithm was constructed to improve prediction accuracy.
It improves the accuracy and generalization ability of 2OM site prediction, reduces computational complexity, provides an efficient and interpretable prediction solution, is suitable for large-scale data processing, and reduces cost and time costs.
Smart Images
Figure CN120673841A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for identifying and predicting 2OM sites in RNA sequences. Background Art
[0002] The most critical post-transcriptional modification of RNA is 2'-O-methylation (2OM). It is present in various types of RNA, including ribosomal RNA (rRNA), transfer RNA (tRNA), messenger RNA (mRNA), small nucleolar RNA (snoRNA), microRNA (miRNA), and piwi-interacting RNA (piRNA). 2OM modification of RNA plays important roles in both structure and function. For example, 2OM stabilizes RNA secondary structure, particularly its helical structure, making it more rigid and stable, which helps maintain the three-dimensional conformation of RNA molecules. Furthermore, it regulates interactions between RNA and proteins or other RNA molecules, influencing cell signaling, gene expression, and regulatory processes. Furthermore, it helps the innate immune system distinguish between endogenous and exogenous messenger RNA. Some viruses recruit the 2OM transferase complex to introduce host-like 2OM modifications to disguise their RNA, thereby evading detection by pattern recognition receptors (PRRs). For example, HIV uses 2OM to evade recognition by MDA5. Naturally occurring 2OM RNA can also suppress inflammatory responses. Therefore, elucidating the exact functions of 2OM modification sites will be of great benefit to both biology and disease treatment, and identifying 2OM sites is an important prerequisite for revealing their mechanisms.
[0003] Current 2OM site prediction models rely not only on manually designed features such as nucleotide composition and secondary structure prediction, but also employ large deep learning-based language models such as Word2Vec to extract RNA sequence features. These methods automatically learn sequence patterns by training on large amounts of data, but often overlook the biological significance of RNA sequences. While these models can identify some statistical regularities and underlying patterns, they lack a deep understanding of the functions and mechanisms of RNA sequences, limiting their prediction accuracy and the depth of their biological applications.
[0004] Current 2OM site prediction models, particularly those using support vector machines (SVMs) and basic neural networks, are typically trained on small or single datasets, which can lead to poor performance on new or diverse datasets and a lack of sufficient generalization. Due to overfitting or limitations in data diversity, these models are unable to effectively address a wider range of biological variability, resulting in insufficient prediction accuracy.
[0005] Current 2OM site prediction models include traditional machine learning and deep learning methods. While some deep learning models have achieved good prediction accuracy on specific test sets—for example, the H2Opred model achieved an AUC of 0.946 on its own independent test set—they are computationally complex and require significant computational resources during training and inference, resulting in low efficiency. Furthermore, while many high-throughput experimental methods (such as Nm-seq) can provide precise information on modification sites, the experimental procedures are cumbersome and costly, making them unsuitable for large-scale or rapid biological research.
[0006] Therefore, existing methods still have significant limitations in terms of computational efficiency and practical application. To this end, a method for identifying and predicting 2OM sites in RNA sequences is proposed. Summary of the Invention
[0007] The technical problem to be solved by the present invention is how to solve the problems of low computational efficiency and poor prediction performance in the prior art, and provides a method for identifying and predicting 2OM sites in RNA sequences.
[0008] The present invention solves the above technical problems through the following technical solutions, which include the following steps:
[0009] S1: Dataset Construction
[0010] 2OM site data were obtained from public databases and relevant experimental data to construct an RNA sequence dataset, which was then divided into four subsets based on the modified base types of the 2OM sites.
[0011] S2: Feature Extraction
[0012] Four feature extraction methods, K-mer, ENAC, PS2, and DPCP_2, were used to encode the RNA sequences in each subset. For a single subset, after encoding the features using the four feature extraction methods, the encoded features were horizontally spliced to obtain spliced features, forming a feature set for the single subset. Among them, K-mer represents nucleotide composition, ENAC represents enhanced nucleic acid composition, PS2 represents the positional specificity of two nucleotides, and DPCP_2 represents the physical and chemical properties of dinucleotides.
[0013] S3: Feature Selection
[0014] The feature sets of each subset obtained in step S2 are respectively screened out using the random forest combined with the forward feature selection method to obtain the optimal feature subset;
[0015] S4: Build a predictive model
[0016] Based on the optimal feature subset obtained by screening, five algorithms, including LGBM, GBDT, XGBoost, RF, and Adaboost, were used for model training and comparison, and finally the optimal prediction model combination was selected;
[0017] S5: Identification Prediction
[0018] The constructed prediction model was used to identify and predict 2OM sites in a test set of different base types.
[0019] Furthermore, in step S1, the specific processing process is as follows:
[0020] S11: Obtain RNA fragments from public databases and experimental data generated by related technologies, adjust the length of all obtained RNA fragments to a set length, and randomly select unmethylated RNA fragments of equal length in the upstream and downstream regions of each 2OM site as negative samples. Public databases include RMBase v3.0 and RMBase v2.0, and experimental data generated by related technologies are experimental data generated by the Nm-seq method.
[0021] S12: Set a similarity threshold and use CD-HIT software to perform sequence clustering on all positive and negative samples based on the similarity threshold;
[0022] S13: After data cleaning and deduplication, an RNA sequence dataset with a balanced number of positive and negative samples is constructed.
[0023] Furthermore, in step S13, based on the modified base type of the 2OM site, the RNA sequence dataset is further divided into four subsets: A, C, G, and U, wherein subset A only considers whether the target base A in the RNA sequence can be 2OM modified; subset C only considers whether the target base C in the RNA sequence can be 2OM modified; subset G only considers whether the target base G in the RNA sequence can be 2OM modified; and subset U only considers whether the target base U in the RNA sequence can be 2OM modified.
[0024] Furthermore, in step S2, K-mer is used to calculate the frequency of k consecutive nucleotides to extract the short-range information of the RNA sequence, and each RNA sequence R is converted into a vector with 4k elements by using the k-mer frequency, as shown below:
[0025]
[0026] in, is the normalized frequency of occurrence of the i-th k-tuple, expressed as follows:
[0027]
[0028] Among them, n i is the number of nucleotide components in the i-th k-tuple of RNA sequence R, and L is the length of the RNA sequence.
[0029] Furthermore, in step S2, ENAC calculates the nucleotide composition based on a fixed-length window that slides continuously from the 5' to the 3' end of each RNA sequence, and the calculation formula is as follows:
[0030]
[0031] Among them, S represents the size of the sliding window, represents the number of nucleotide types t in the sliding window r, t∈{A,C,G,U}, r=1,2,…,L-S+1.
[0032] Furthermore, in step S2, PS2 encodes the paired adjacent nucleotides in the RNA sequence through binary coding.
[0033] Furthermore, in step S2, the calculation formula of DPCP_2 is as follows:
[0034]
[0035] Among them, NA x ,NA y ∈{AA,AC,AG,AU,CA,…,UU}, m=1,2,…,(L-1), n=(m+1),(m+2),…,L, Represents dinucleotide The i-th physicochemical property.
[0036] Furthermore, in step S3, the specific processing process is as follows:
[0037] S31: For each subset of feature sets, use the random forest model to calculate the Gini importance score of each feature and rank each feature according to the score;
[0038] S32: Select feature subsets step by step according to the order of feature importance, adding a set number of features each time, and evaluate the impact of different feature subsets on the performance of the prediction model through 5-fold cross validation;
[0039] S33: When multiple rounds of feature selection are performed continuously and the model performance fails to improve by 0.001 compared with the current optimal AUC index value, the feature selection process in step S32 is stopped, and the optimal feature subset corresponding to each feature set is obtained.
[0040] Furthermore, in step S4, the model training process uses 5-fold cross validation to select the model, and the following indicators are used to evaluate the model performance:
[0041]
[0042] Among them, TP is true positive, FP is false positive, TN is true negative, and FN is false negative;
[0043] The area under the receiver operating characteristic curve was also used to evaluate the model performance.
[0044] Furthermore, in step S4, the optimal prediction model combination is specifically as follows:
[0045] For RNA sequences where the modified base type at the 2OM site is base U, the LGBM model is used as the prediction model for identification and prediction;
[0046] For RNA sequences whose modified base types at the 2OM site are bases A, C, or G, the GBDT model is used as the prediction model for identification and prediction.
[0047] Compared with the prior art, the present invention has the following advantages:
[0048] 1. This paper introduces the RMBase v3.0 dataset, which covers more 2′-O-methylation (2OM) site information. The data volume and quality are improved compared with previous versions, providing more comprehensive and accurate data support for model training, which is conducive to improving the generalization ability of the model.
[0049] 2. The present invention adopts four complementary and efficient feature extraction methods (K-mer, ENAC, PS2 and DPCP_2), each of which deeply mines the biological information of RNA sequences from different levels. By integrating multidimensional features such as nucleotide composition, sequence position specificity, physicochemical properties and nucleotide dependence, this combination can comprehensively and accurately capture various key information of RNA sequences, making up for the limitations of single feature extraction methods. The K-mer method can effectively capture the information of short sequence fragments, ENAC enhances position-specific features, PS2 combines nucleotide dependence, and DPCP_2 provides more biological associations through dinucleotide combination patterns. The combination of the four methods significantly improves the model's ability to predict 2′-O-methylation sites, which can not only comprehensively describe the characteristics of RNA sequences, but also ensure the efficiency and accuracy of the model.
[0050] 3. Compared with complex deep learning models, the present invention adopts lightweight machine learning algorithms (GBDT and LGBM). These algorithms have low computational overhead and high training efficiency, and can provide good interpretability, which is convenient for analyzing the prediction basis and optimization direction of the model. Compared with traditional deep learning methods, the model of the present invention has stronger understandability and flexibility while ensuring efficient processing of large-scale data, is suitable for a wide range of application scenarios, and has strong universality and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 2OM sites in RNA sequences according to an embodiment of the present invention;
[0052] Figure 2 Detailed framework diagram of the method for identifying and predicting 2OM sites in RNA sequences according to an embodiment of the present invention;
[0053] Figure 3 This is a comparison chart of the AUC scores of the prediction models on the four bases (A, C, G, U) in the embodiment of the present invention (training set);
[0054] Figure 4 3. This is a graph comparing the prediction performance of the model of the present invention and the H2Opred model at the 2OM site on base A in an embodiment of the present invention;
[0055] Figure 5 3. This is a graph comparing the prediction performance of the model of the present invention and the H2Opred model for 2OM sites on base C in an embodiment of the present invention;
[0056] Figure 6 3. This is a graph comparing the prediction performance of the model of the present invention and the H2Opred model at the 2OM site on base G in an embodiment of the present invention;
[0057] Figure 7 3 is a graph comparing the prediction performance of the model of the present invention and the H2Opred model at the 2OM site on base U in the embodiments of the present invention. DETAILED DESCRIPTION
[0058] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.
[0059] like Figure 1 、 2 As shown, this embodiment provides a technical solution: a method for identifying and predicting 2OM sites in RNA sequences, comprising the following steps:
[0060] 1. Build a benchmark dataset: Collect 2OM site data from the public databases RMBase v3.0, RMBase v2.0, and experimental data generated by Nm-seq technology to build an initial RNA sequence dataset.
[0061] 2. Feature extraction: Four feature extraction methods, including nucleotide composition (K-mer), enhanced nucleic acid composition (ENAC), position specificity of two nucleotides (PS2), and dinucleotide physicochemical property characteristics (DPCP_2), were used to encode the features of RNA sequences.
[0062] 3. Feature selection: The four feature sets obtained in step 2 (features for the four bases A, C, G, and U) are respectively selected using random forest combined with forward feature selection method (RF_SFS) to select the most informative feature subsets.
[0063] 4. Construct a prediction model: Based on the feature subset selected in step 3, five algorithms, including LGBM, GBDT, XGBoost, RF, and Adaboost, were used for model training and comparison. Finally, the optimal model combination was selected. Among them, a GBDT model was constructed for base A, a GBDT model was constructed for base C, a GBDT model was constructed for base G, and a LGBM model was constructed for base U.
[0064] 5. Model Performance Evaluation: 5-fold cross-validation was used to evaluate the model's performance on the training set. The model's generalization ability was further evaluated by testing it on an independent test set. Finally, the model's performance was compared with the existing, leading prediction model, H2Opred.
[0065] In this embodiment, the specific processing process of the above steps is further explained:
[0066] Step 1: Build a benchmark dataset
[0067] Previous studies on 2′-O-methylation (2OM) sites mostly relied on the RMBase v2.0 database and combined with experimental data generated by Nm-seq technology. However, with the upgrade of the RMBase database to v3.0, its data volume and data quality have been improved, providing more comprehensive information for the study of 2OM sites. The present invention introduces the RMBasev3.0 dataset into the 2OM site prediction model and optimizes the dataset construction process to ensure the comprehensiveness and consistency of the data. First, in order to ensure the uniformity of the samples, the present invention adjusts the length of all extracted RNA fragments to 41bp, and randomly selects non-methylated RNA fragments of equal length in the upstream and downstream regions of each 2OM site as negative samples to ensure that the number of positive and negative samples is equal. In addition, in order to further enrich the data source, we also integrated the relevant 2OM site data previously extracted from other databases, including data generated by Nm-seq technology, and de-redundantly processed the merged dataset. Specifically, the present invention uses CD-HIT software to perform sequence clustering on all positive and negative samples, and sets the similarity threshold to 80% to remove redundant sequences and improve the representativeness of the data. After data cleaning and deduplication, a balanced dataset consisting of 10,475 positive and 10,475 negative samples was constructed. Based on the type of modified base at the 2′-O-methylation site, the dataset was further divided into four subsets: A, C, G, and U (as shown in Table 1). Each subset was randomly divided into training and test sets in a 7:3 ratio. Subset A only considered whether the target base A in the RNA sequence could undergo 2OM modification; subset C only considered whether the target base C in the RNA sequence could undergo 2OM modification; subset G only considered whether the target base G in the RNA sequence could undergo 2OM modification; and subset U only considered whether the target base U in the RNA sequence could undergo 2OM modification.
[0068] Table 1 Dataset details
[0069] datasets train_positive train_negative test_positive test_negative A 1701 1701 730 730 C 1701 1701 730 730 G 1803 1803 774 774 U 2125 2125 911 911 Total 7330 7330 3145 3145
[0070] Step 2: Feature Extraction
[0071] When building a 2OM site recognition model based on machine learning, the primary task is to achieve effective digital encoding of RNA sequences on the basis of retaining the original RNA information to the greatest extent, which is the key to ensuring model accuracy and robustness. To this end, the present invention selects four feature extraction methods, including nucleotide composition (K-mer), enhanced nucleic acid composition (ENAC), positional specificity of two nucleotides (PS2) and dinucleotide physicochemical property characteristics (DPCP_2). These feature extraction methods can extract RNA sequence information from different angles, covering nucleotide composition, positional information and physicochemical properties, thereby providing comprehensive and efficient feature representation for the construction of the model. These feature extraction methods will be described in detail below.
[0072] 2.1 Nucleotide composition (K-mer)
[0073] K-mer is used to calculate the frequency of consecutive k nucleotides to extract short-range information of RNA sequences. Each primary RNA sequence R can be converted into a vector with 4k elements by using the k-mer frequency as follows:
[0074]
[0075] in, is the normalized frequency of occurrence of the i-th k-tuple, which can be expressed as:
[0076]
[0077] Among them, n i is the number of the i-th k-tuple nucleotide components in the RNA sequence R, and L is the length of the RNA sequence. In this embodiment, k is set to 2, 3.
[0078] 2.2 Enhanced Nucleic Acid Composition (ENAC)
[0079] ENAC calculates nucleotide composition based on a fixed-length window that slides continuously from the 5' to the 3' end of each RNA sequence. This method is commonly used to encode RNA sequences of equal length and is suitable for our data. ENAC is calculated as follows:
[0080]
[0081] Among them, S represents the size of the sliding window, represents the number of nucleotide types t in the sliding window r, t∈{A, C, G, U} and r=1, 2,…, L−S+1. In this embodiment, the size of the sliding window is 5.
[0082] 2.3. Positional specificity of two nucleotides (PS2)
[0083] PS2 represents pairs of adjacent nucleotides, including 16 different nucleotide pairs: AA, AC, AG, AU, CA, CC, CG, CU, GA, GC, GG, GU, UA, UC, UG, and UU. These 16 different nucleotide pairs are encoded as 16 binary bits, each of which is in the set {0, 1}, with only the corresponding position set to 1 and the rest set to 0. For example, AG is represented as (00100000000000000), with AG at the third position, while GC is represented as (0000000001000000), with GC at the 10th position. Therefore, an RNA sequence like AGGU is encoded as (00100000000000000000000000100000000000000010000).
[0084] 2.4 Physicochemical properties of dinucleotide (DPCP_2)
[0085] There are 22 different physicochemical properties related to dinucleotides in RNA sequences, detailed in Table 2. When considering an RNA sequence of length L, DPCP_2 can be expressed as:
[0086]
[0087] Among them, NA x ,NA y ∈{AA,AC,AG,AU,CA,…,UU}, m=1,2,…,(L-1), n=(m+1),(m+2),…,L, Represents dinucleotide The i-th physicochemical property.
[0088] Table 2 22 physicochemical properties of dinucleotides
[0089]
[0090]
[0091] Step 3: Feature Selection
[0092] Feature selection is an important step in biological sequence analysis, which can effectively remove redundant features and improve the stability and robustness of the model. The present invention is directed to the sequence data of four bases (A, C, G, U), and adopts four feature extraction methods of K-mer, ENAC, PS2 and DPCP_2 to extract features, and splices them horizontally to obtain high-dimensional features (splicing features). In order to avoid the impact of noise features on model construction, the present invention has adopted RF_SFS method to carry out optimal feature selection for the feature set of each base. Specifically, for each base (A, C, G, U), first use random forest model to calculate the Gini importance score of each feature (each independent feature in the high-dimensional feature) corresponding to it, and sort each feature according to score. Then, according to the importance order of feature, progressively select feature subsets, each increase a certain number of features, and evaluate the impact of different feature subsets on model performance by 5-fold cross validation (using AUC as evaluation index in the present invention). In the feature selection process, the present invention introduces an early stopping mechanism, that is, when multiple rounds of feature selection are performed continuously and the model performance fails to improve significantly (when the model performance fails to absolutely improve by 0.001 compared to the current optimal AUC index value), the feature selection process is automatically stopped to avoid excessive feature selection or overfitting. Finally, the present invention uses the RF_SFS method to independently screen the feature sets of four bases (A, C, G, U) and obtain the corresponding optimal feature subsets. Feature selection can effectively reduce the interference of redundant features on the model and improve the classification performance of the model.
[0093] Step 4: Build a predictive model
[0094] After screening and obtaining the optimal feature subsets of each of the four bases (A, C, G, U), the present invention adopts five widely used machine learning algorithms, LGBM, GBDT, XGBoost, RF and Adaboost, to perform model training. The performance of each model is evaluated by 5-fold cross-validation, and the optimal model is selected for different bases to build an accurate and efficient prediction model. Finally, the present invention determines the optimal model combination, wherein base A corresponds to the GBDT model, base C corresponds to the GBDT model, base G corresponds to the GBDT model, and base U corresponds to the LGBM model, thereby achieving the optimal prediction strategy for different bases.
[0095] 4.1 Cross-validation and evaluation metrics
[0096] During model training, we use 5-fold cross-validation for model selection. To measure the performance of the model, we use the following indicators to evaluate the model:
[0097]
[0098] Where TP is true positive, FP is false positive, TN is true negative, and FN is false negative. The area under the receiver operating characteristic curve (AUROC, or AUC) is also used. It comprehensively evaluates model performance by balancing the true positive rate (also known as sensitivity) and the false positive rate. Its value ranges from 0 to 1, with the closer it is to 1, the better the model performance. The AUC is unaffected by class imbalance and data distribution, making it highly robust. The AUC is a valuable metric for comparing the performance of different models and helps determine the optimal classification threshold that balances sensitivity and specificity.
[0099] 4.2. Optimal model selection for four bases
[0100] The present invention uses the five widely used machine learning algorithms, LGBM, GBDT, XGBoost, RF and Adaboost, to train the model for the optimal feature subsets of the four bases (A, C, G, U). The training results are shown in the attached figure. Figure 3 . As can be seen from the figure, for the prediction of 2'-O-methylation (2OM) sites of base A in RNA sequences, the AUC value of the model trained using GBDT is 0.9572, which has the best prediction performance compared to LGBM (0.05% higher), XGBoost (0.16% higher), RF (0.63% higher), and Adaboost (2.01% higher). Therefore, GBDT was finally selected as the 2OM site prediction model for base A. Similarly, for bases C, G, and U, the model with the highest evaluation index AUC was selected as the final prediction model to ensure the optimal prediction performance.
[0101] Step 5: Use the constructed model to make predictions
[0102] 5.1 Prediction on an Independent Test Set
[0103] In order to evaluate the predictive performance of the model constructed by the present invention, the trained model was verified on an independent test set. The experimental results (as shown in Table 3) show that for the 2'-O-methylation (2OM) site of base A in the RNA sequence, the constructed GBDT model has an AUC value of 0.9615 on the independent test set, showing excellent predictive ability. Similarly, the optimal model selected for bases C, G, and U also showed excellent predictive performance on the independent test set, further verifying the effectiveness and generalization ability of the method of the present invention. The model constructed by the present invention can accurately predict RNA 2OM sites, has high stability and generalization ability, provides a powerful support tool for the identification and functional research of RNA 2OM sites, and can also lay the foundation for the related research and application of RNA methylation modification.
[0104] Table 3 Prediction results of the constructed model on the independent test set
[0105] model AUC ACC MCC Pre Rec F1 A-GBDT 0.9615 0.9014 0.8038 0.9234 0.8753 0.8987 C-GBDT 0.9421 0.8610 0.7223 0.8738 0.8438 0.8585 G-GBDT 0.9704 0.9128 0.8269 0.9371 0.8850 0.9103 U-LGBM 0.9285 0.8589 0.7180 0.8641 0.8518 0.8579
[0106] 5.2 Comparison of prediction performance with the existing superior model H2Opred
[0107] To further validate the superiority of our proposed model, we compared it with H2Opred, a leading existing model for 2OM site prediction. Specifically, we input our independent test set of newly constructed data into the H2Opred online prediction platform, derived relevant evaluation metrics, and compared the performance of our model on the independent test set. The experimental results, shown in Table 4, show that our model outperforms H2Opred across all key metrics for each base. For example, for base A, the AUC and ACC values of our model are 5.57% and 5.76% higher than those of H2Opred, further demonstrating the superior performance and reliability of our model in 2OM site prediction.
[0108] Table 4 Comparison of prediction performance between the constructed model and the existing better model H2Opred
[0109]
[0110] The present invention introduces the RMBase v3.0 dataset into the 2′-O-methylation (2OM) site prediction model. Compared with previous methods that mostly rely on experimental data generated by RMBase v2.0 and Nm-seq technology, we fully utilize the larger data scale and higher data quality of RMBase v3.0 to construct a more representative 2OM dataset. In terms of feature extraction, the present invention introduces four feature extraction methods: K-mer, ENAC, PS2 and DPCP_2 to comprehensively characterize RNA sequence information. The K-mer method retains the information of short sequence fragments, which helps to capture local sequence patterns; ENAC emphasizes the local position characteristics of the sequence and improves the precision of recognition; PS2 combines nucleotide dependency information and enhances the ability to analyze sequence structure; DPCP_2 further explores dinucleotide combination patterns and strengthens the biological relevance of features. Through random forest combined with forward feature selection (RF_SFS), we screened out the most informative feature subset to improve the stability and generalization ability of the model. In terms of model selection, the present invention adopts a strategy optimized for different bases, and ultimately determines that the bases A, C, and G use the GBDT model, and the base U uses the LGBM model, ensuring the efficiency and accuracy of the prediction.
[0111] The experimental results show that the model of the present invention is always better than the existing advanced methods in the tests on four different data sets, which verifies its robustness and effectiveness. Figure 4-Figure 7. Compared with previous deep learning-based methods, the model of the present invention has significant advantages. It has much lower computational overhead than deep learning models, has higher training efficiency, and can complete efficient training in a shorter time. In addition, this method has good interpretability and can intuitively reveal the key factors affecting the prediction of RNA 2′-O-methylation sites. Compared with the "black box" nature of deep learning models, it can be easier to analyze and optimize the results. Most importantly, this method is not only suitable for the study of human RNA, but can also be extended to the analysis of RNA of other species. It has wide applicability and is suitable for the study of various types of RNA modifications, such as methylation, acetylation, etc. Compared with traditional experimental methods, the present invention provides a cost-effective, fast and accurate prediction solution, which is particularly suitable for the processing of large-scale data, greatly saving time and economic costs.
[0112] In summary, the method for identifying and predicting 2OM sites in RNA sequences in the above embodiment is innovative in theory and has demonstrated significant effects and advantages in practical applications. Through precise feature processing and efficient model design, it effectively improves the prediction performance of RNA 2′-O-methylation sites, helping to promote research and application development in related biomedical fields.
[0113] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and deform the above embodiments within the scope of the present invention.
Claims
1. A method for identifying and predicting 2OM sites in RNA sequences, characterized in that: The following steps are involved: S1: Dataset Construction 2OM site data were obtained from public databases and relevant experimental data to construct an RNA sequence dataset, which was then divided into four subsets based on the modified base types of the 2OM sites. S2: Feature Extraction Four feature extraction methods, K-mer, ENAC, PS2, and DPCP_2, were used to encode the RNA sequences in each subset. For a single subset, after encoding the features using the four feature extraction methods, the encoded features were horizontally spliced to obtain spliced features, forming a feature set for the single subset. Among them, K-mer represents nucleotide composition, ENAC represents enhanced nucleic acid composition, PS2 represents the positional specificity of two nucleotides, and DPCP_2 represents the physical and chemical properties of dinucleotides. S3: Feature Selection The feature sets of each subset obtained in step S2 are respectively screened out using the random forest combined with the forward feature selection method to obtain the optimal feature subset; S4: Build a predictive model Based on the optimal feature subset obtained by screening, five algorithms, including LGBM, GBDT, XGBoost, RF, and Adaboost, were used for model training and comparison, and finally the optimal prediction model combination was selected; S5: Identification Prediction The constructed prediction model was used to identify and predict 2OM sites in a test set of different base types.
2. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 1, wherein In step S1, the specific processing process is as follows: S11: Obtain RNA fragments from public databases and experimental data generated by related technologies, adjust the length of all obtained RNA fragments to a set length, and randomly select unmethylated RNA fragments of equal length in the upstream and downstream regions of each 2OM site as negative samples. Public databases include RMBase v3.0 and RMBase v2.0, and experimental data generated by related technologies are experimental data generated by the Nm-seq method. S12: Set a similarity threshold and use CD-HIT software to perform sequence clustering on all positive and negative samples based on the similarity threshold; S13: After data cleaning and deduplication, an RNA sequence dataset with a balanced number of positive and negative samples is constructed.
3. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 2, wherein: In step S13, based on the modified base type of the 2OM site, the RNA sequence dataset is further divided into four subsets: A, C, G, and U, wherein subset A only considers whether the target base A in the RNA sequence can be 2OM modified; subset C only considers whether the target base C in the RNA sequence can be 2OM modified; subset G only considers whether the target base G in the RNA sequence can be 2OM modified; and subset U only considers whether the target base U in the RNA sequence can be 2OM modified.
4. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 3, wherein: In step S2, K-mer is used to calculate the frequency of consecutive k nucleotides to extract the short-range information of the RNA sequence. Each RNA sequence R is converted into a vector with 4k elements by using the k-mer frequency, as shown below: in, is the normalized frequency of occurrence of the i-th k-tuple, expressed as follows: Among them, n i is the number of nucleotide components in the i-th k-tuple of RNA sequence R, and L is the length of the RNA sequence.
5. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 4, wherein In step S2, ENAC calculates the nucleotide composition based on a fixed-length window that slides continuously from the 5' to the 3' end of each RNA sequence. The calculation formula is as follows: Among them, S represents the size of the sliding window, represents the number of nucleotide types t in the sliding window r, t∈{A,C,G,U}, r=1,2,…,L-S+1.
6. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 5, wherein: In step S2, PS2 encodes the adjacent nucleotide pairs in the RNA sequence through binary coding.
7. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 6, wherein: In step S2, the calculation formula of DPCP_2 is as follows: Among them, NA x ,NA y ∈{AA,AC,AG,AU,CA,…,UU}, m=1,2,…,(L-1), n=(m+1),(m+2),…,L, Represents dinucleotide The i-th physicochemical property.
8. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 1, wherein In step S3, the specific processing process is as follows: S31: For each subset of feature sets, use the random forest model to calculate the Gini importance score of each feature and rank each feature according to the score; S32: Select feature subsets step by step according to the order of feature importance, adding a set number of features each time, and evaluate the impact of different feature subsets on the performance of the prediction model through 5-fold cross validation; S33: When multiple rounds of feature selection are performed continuously and the model performance fails to improve by 0.001 compared with the current optimal AUC index value, the feature selection process in step S32 is stopped, and the optimal feature subset corresponding to each feature set is obtained.
9. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 1, wherein In step S4, the model training process uses 5-fold cross validation to select the model, and the following indicators are used to evaluate the model performance: Among them, TP is true positive, FP is false positive, TN is true negative, and FN is false negative; The area under the receiver operating characteristic curve was also used to evaluate the model performance.
10. The method for identifying and predicting 2OM sites in an RNA sequence according to claim 1, wherein: In step S4, the optimal prediction model combination is specifically as follows: For RNA sequences where the modified base type at the 2OM site is base U, the LGBM model is used as the prediction model for identification and prediction; For RNA sequences whose modified base types at the 2OM site are bases A, C, or G, the GBDT model is used as the prediction model for identification and prediction.