Method for predicting 2OM site in RNA (Ribonucleic Acid) based on ensemble learning

Abstract

The invention discloses a method for predicting a 2OM site in RNA (Ribonucleic Acid) based on ensemble learning, and belongs to the technical field of bioinformatics. According to the method, specific task fine adjustment is carried out on the Promoter-BERT model according to the RNA sequence characteristics, so that the Promoter-BERT model can more effectively capture the complex mode of the RNA sequence in the specific task, and high-quality biological characteristic representation is obtained; the ANOVA technology is adopted to select extracted features, redundant features are removed, the most influential features are reserved, and in addition, the expression ability of the model is enhanced by combining traditional sequence features and embedded features obtained through a Word2Vec model; the prediction results of the lightweight gradient elevator and the deep learning model are combined, a final prediction model is formed through a soft voting mechanism, and the integration method not only improves the generalization ability of the model, but also improves the prediction stability.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for predicting 2OM sites in RNA based on ensemble learning. Background Art

[0002] 2'-O-methylation (2OM) is an important post-transcriptional modification of RNA, which exists in a variety of RNAs, including rRNA, tRNA, mRNA, snoRNA, miRNA, and piRNA. 2OM can stabilize the secondary structure of RNA, especially the helical structure, making it more stable and helping to maintain the three-dimensional conformation of RNA molecules. 2OM can also regulate the interaction between RNA and proteins or other RNA molecules, affecting cell signaling, gene expression, and regulatory processes. In addition, 2OM modification can help the innate immune system distinguish between endogenous and exogenous mRNA, and some viruses evade immune recognition through 2OM modification. Identifying 2OM sites is important for revealing their functions in biology and disease treatment.

[0003] Early experimental techniques such as HClO4 hydrolysis and high-throughput techniques such as Nm-seq have been used to identify 2OM sites. However, these methods have their own limitations. For example, the early experimental techniques are complicated to operate and may cause irreversible damage to the samples, while the high cost of high-throughput techniques limits their application in general laboratories. Computational methods, especially those based on artificial intelligence, can efficiently predict 2OM sites. Some researchers have developed a computational method using support vector machines (SVMs), others have built iRNA-2OM models based on SVMs, and some have developed iRNA-PseKNC (2methyl) models using convolutional neural networks (CNNs). Some researchers have developed an integrated model by combining seven optimal classifiers. In addition, there are prediction models developed based on other data sets, such as NmSEER and DeepOMe.

[0004] Traditional 2OM site prediction models rely on manually designed features, such as Nucleotide Composition and secondary structure prediction, which may not fully capture the complex biological properties of RNA sequences, resulting in limited prediction accuracy; existing computational models, such as support vector machines (SVM) and basic neural networks, are often trained on small-scale or single datasets. These models may perform poorly on new or diverse datasets and lack sufficient generalization capabilities; many high-throughput experimental methods, such as Nm-seq, although they can provide accurate modification site information, are complex and costly to operate, and are not suitable for large-scale or rapid biological research needs.

[0005] The above problems need to be solved urgently. To this end, the present invention proposes a method for predicting 2OM sites in RNA based on ensemble learning. Summary of the invention

[0006] The technical problem to be solved by the present invention is: how to solve the problems existing in the prior art of limited prediction accuracy, lack of sufficient generalization ability, complex operation and high cost, and provide a method for predicting 2OM sites in RNA based on ensemble learning.

[0007] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0008] S1: Data Preparation

[0009] 2OM site data were collected from public databases and randomly divided into training and test sets according to a set ratio;

[0010] S2: Feature Learning

[0011] Three manually designed features, K-mer, ANF and NCP, were extracted from RNA sequences, and the RNA sequences were vectorized using the Word2Vec model to obtain Word2Vec embedded features. ANOVA technology was used to select a subset of features highly correlated with 2OM site prediction from the manually designed features, and combined with the embedded features generated by the Word2Vec model.

[0012] S3: Model building

[0013] Based on 1mer, 3mer and 5mer word segmentation, texts of sequence units of different lengths, i.e. segmented sequences, are obtained. The segmented sequences are input into the pre-trained Promoter-BERT model for fine-tuning, and three fine-tuned models are obtained. Based on the obtained feature subset, a basic model is constructed using a lightweight gradient boosting machine, and it is integrated with the partial model obtained by fine-tuning the pre-trained Promoter-BERT model. The prediction probabilities of each basic model are integrated through a soft voting mechanism to form the final integrated model 2OMPro, in which the partial model obtained by fine-tuning the pre-trained Promoter-BERT model is also the basic model.

[0014] S4: Optimization and Evaluation

[0015] The performance of the integrated model 2OMPro was evaluated using cross-validation and independent test sets;

[0016] S5: Deployment and prediction

[0017] The integrated model 2OMPro is deployed in actual biomedical research to predict unknown 2OM sites.

[0018] Furthermore, in step S1, the specific processing process is as follows:

[0019] S11: Collect raw 2OM loci data from RMBase v2.0 and experimental datasets generated based on Nm-seq technology;

[0020] S12: For each 2OM site data, RNA fragments with a set length of nucleotides were cut from the upstream and downstream of the 2OM site as positive samples; at the same time, sequences of the same length were randomly selected from the upstream and downstream of each 2OM site as negative samples;

[0021] S13: Set the sequence similarity threshold, use CD-HIT software to perform sequence clustering to remove redundant sequences, and obtain a data set with equal positive samples and negative samples, that is, obtain an RNA sequence data set;

[0022] S14: Based on the modified base types of the 2OM site, the dataset was divided into four subsets, and each subset was randomly divided into a training set and a test set according to a set ratio.

[0023] Furthermore, in step S2, in order to cover different levels of sequence information, the RNA sequence is segmented into 1mer, 3mer, and 5mer as vocabulary, and then the RNA sequence is input as a sentence into a pre-trained Promoter-BERT model for fine-tuning to obtain a fine-tuned Promoter-BERT model based on 1mer, 3mer, and 5mer.

[0024] Furthermore, in step S2, the expression for encoding the RNA sequence by the fine-tuned Promoter-BERT model is as follows:

[0025] X BERT =BERT(S)

[0026] Where S represents the input RNA sequence, BERT represents the fine-tuned Promoter-BERT model, and X BERT Represents the encoding of the input sequence S by the fine-tuned Promoter-BERT model.

[0027] Furthermore, in step S2, the three manually designed features are nucleotide composition K-mer, nucleotide compound property NCP, and auto-correlation property ANF; wherein the nucleotide composition K-mer is used to calculate the frequency of k consecutive nucleotides to extract the short-range information of the RNA sequence, the nucleotide compound property NCP is used to use vector encoding (1,1,1), (0,0,1), (0,1,0) and (1,0,0) to represent A, U, C and G respectively, and the auto-correlation property ANF is used to describe the distribution of nucleotides in the RNA sequence and their position correlation in the sequence, by calculating the cumulative occurrence frequency d of each nucleotide from the beginning of the sequence to position i i to achieve.

[0028] Furthermore, in the nucleotide composition K-mer, using the k-mer frequency, each primary RNA sequence R can be converted into a vector with 4k elements as follows:

[0029]

[0030] in, is the normalized frequency of occurrence of the i-th k-tuple:

[0031]

[0032] Among them, n i is the number of nucleotide components in the i-th k-tuple in the RNA sequence, and L represents the length of the RNA sequence.

[0033] Furthermore, in the autocorrelation feature ANF, the cumulative occurrence frequency d of each nucleotide from the beginning of the sequence to position i in the RNA sequence is i Defined as:

[0034]

[0035] Among them, |N i | represents the length of the prefix subsequence from the beginning of the sequence to position i, that is, the number of nucleotides before the i-th position, f(n j ) is an indicator function that indicates when nucleotide n at position j j When the target nucleotide ni is the same, f(n j ) is equal to 1, otherwise it is 0.

[0036] Furthermore, in step S2, the specific process of selecting a feature subset highly correlated with 2OM site prediction from the manually designed features using the ANOVA technique is as follows:

[0037] Assuming there are m manually designed features, the feature subset F selected by ANOVA technologyselected It is expressed as:

[0038] F selected ={f i ∣p(f i )≤α}

[0039] Among them, f i represents the i-th manually designed feature, p(f i ) represents the p-value of the ANOVA test of the ith feature, and α represents the set significance level for feature screening.

[0040] Furthermore, in step S3, the partial models obtained by fine-tuning the pre-trained Promoter-BERT model as the basic model are respectively fine-tuned Promoter-BERT models based on 3mer and 5mer.

[0041] Furthermore, in step S3, if there are M basic models and N categories, the predicted probability of the m-th basic model for the n-th category is expressed as Pmn, then the predicted probability of the n-th category under soft voting is P soft (n) The calculation formula is:

[0042]

[0043] Among them, P mn Represents the predicted probability of the mth basic model for the nth class;

[0044] The soft voting mechanism averages the predicted probabilities of all base models for each category, and the final result C consists of the category with the highest average predicted probability as the final prediction result of the integrated model 2OMPro:

[0045] C = argmax n P soft (n)

[0046] Among them, argmax n Denotes that P soft (n) The maximum value of n, where n is the category number.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] 1. Based on the characteristics of RNA sequences, the Promoter-BERT model is fine-tuned for specific tasks, so that it can more effectively capture the complex patterns of RNA sequences in specific tasks, thereby obtaining high-quality biological feature representations;

[0049] 2. Use ANOVA technology to select the extracted features, eliminate redundant features, and retain the most influential features. In addition, combine traditional sequence features with embedded features obtained through the Word2Vec model to enhance the expressiveness of the model;

[0050] 3. Combining the prediction results of the lightweight gradient boosting machine and the deep learning model, the final prediction model is formed through a soft voting mechanism. This integration method not only improves the generalization ability of the model, but also increases the stability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the process of the present invention;

[0052] Figure 2 5-fold cross-validation result diagram of the Promoter-BERT model after fine-tuning on each subset in an embodiment of the present invention, where a is the Am subset, b is the Cm subset, c is the Gm subset, and d is the Um subset;

[0053] Figure 3 It is the ROC curve and AUC value of each model under different feature combinations on each subset in the embodiment of the present invention, where a is the Am subset, b is the Cm subset, c is the Gm subset, and d is the Um subset. DETAILED DESCRIPTION

[0054] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0055] like Figure 1 As shown, this embodiment provides a technical solution: a method for predicting 2OM sites in RNA based on ensemble learning, which mainly includes the following steps:

[0056] 1. Data preparation

[0057] 2OM loci data were collected from public databases and randomly divided into training and test sets in a ratio of 7:3.

[0058] 2. Feature Learning

[0059] Three manually designed features, namely K-mer, ANF and NCP, were extracted from RNA sequences, and the Word2Vec model was used to vectorize the RNA sequences to obtain Word2Vec embedded features. ANOVA technology was used to select a feature subset that was highly correlated with 2OM site prediction from the manually designed features and combine it with the embedded features generated by the Word2Vec model.

[0060] 3. Model building

[0061] Based on 1mer, 3mer and 5mer word segmentation, we get texts of sequence units of different lengths, i.e. segmented sequences. The segmented sequences are input into the pre-trained Promoter-BERT model for fine-tuning, and three fine-tuned models are obtained. Based on the obtained feature subset, a basic model is constructed using a lightweight gradient boosting machine, and it is integrated with the model obtained by fine-tuning the pre-trained Promoter-BERT model. The prediction probabilities of each basic model are integrated through a soft voting mechanism to form the final integrated model 2OMPro.

[0062] S4: Optimization and Evaluation

[0063] The performance of the integrated model 2OMPro was evaluated using cross-validation and independent test sets;

[0064] S5: Deployment and prediction

[0065] The integrated model 2OMPro is deployed in actual biomedical research work to predict unknown 2OM sites.

[0066] The above steps are further described below.

[0067] Step 1: Data Preparation

[0068] In order to ensure a fair comparison with other advanced models, the present invention selected 7597 original 2OM site data collected from RMBase v2.0 and experimental data sets generated based on Nm-seq technology, which have been stored in the GEO database (GSE90164). These 2OM sites are widely distributed in various regions of RNA, including coding sequences (CDS), 3'-UTR, 5'-UTR, introns, exons and intergenic regions, covering almost all types of RNA such as tRNA, rRNA, scRNA, scaRNA, snRNA, etc. In order to maintain the consistency of data processing, for each 2OM site, a 20-nucleotide RNA fragment was cut from the upstream and downstream of the site as a positive sample. At the same time, sequences of the same length (41 nucleotides) were randomly selected upstream and downstream of each 2OM site (the central position nucleotides of these sequences were not methylated) as negative samples, and the ratio of positive samples to negative samples was approximately 1:3.

[0069] In addition, in order to reduce data redundancy, CD-HIT software is used in the present invention for sequence clustering, and the cutoff rate (sequence similarity threshold) is set to 80%. After reprocessing, a data set with an equal number of positive and negative samples is finally obtained, totaling 6091 samples. Based on the modified base type of the 2OM site, the data set is further divided into four subsets: Am, Um, Cm, and Gm (as shown in Table 1). Each subset is randomly divided into a training set and a test set at a ratio of 7:3. In order to pre-train the BERT model, the UCSC data set from the UCSC Genome Browser is used in the present invention, which is one of the widely used databases in the field of biology.

[0070] Table 1 Detailed information of the dataset

[0071]

[0072] Step 2: Feature Learning

[0073] 2.1 Fine-tuning of Promoter-BERT model:

[0074] In the present invention, the Promoter-BERT model pre-trained to understand the language characteristics of DNA sequences is first fine-tuned for specific RNA sequence characteristics. The Promoter-BERT model was originally based on the Transformer architecture and was pre-trained to understand the language characteristics of DNA sequences. In the present invention, the model is retrained to adapt to the feature recognition of 2OM sites. In order to cover different levels of sequence information, three Promoter-BERT models based on 1mer, 3mer and 5mer vocabularies are fine-tuned respectively, and the weights of the models are adjusted to better adapt to specific biological tasks. Complex relationships from single nucleotides to longer sequence fragments can be captured.

[0075] X BERT =BERT(S)

[0076] Where S represents the input RNA sequence, BERT represents the fine-tuned Promoter-BERT model, and X BERT Represents the encoding of the input sequence S by the fine-tuned Promoter-BERT model.

[0077] 2.2 Feature extraction and fusion:

[0078] The present invention extracts three manually designed features from RNA sequences: nucleotide composition (K-mer), nucleotide compound properties (NCP) and autocorrelation properties (ANF). These features describe the basic physical and chemical properties of RNA sequences and provide the model with preliminary information about sequence structure and function. In order to enhance the model's understanding of RNA sequence context information, the present invention also uses the Word2Vec model to vectorize sequence data. This method can convert sequence information into a dense vector form, thereby helping the model capture and utilize local and global context dependencies in the sequence. Then, the above three manually designed features (K-mer, NCP and ANF) are screened by variance analysis (ANOVA) technology to select the most informative feature subset. The screening and fusion of these manually designed features enhance the diversity and richness of feature expression and optimize the efficiency of feature use during model training. Finally, these selected and fused feature sets together with the embedded features generated by Word2Vec form the basis for model training, providing solid data support for the high-accuracy prediction of 2OM sites.

[0079] Assuming there are m manually designed features (K-mer, NCP and ANF), the feature subset F after ANOVA screening selected It can be expressed as:

[0080] F selected ={f i ∣p(f i )≤α}

[0081] Among them, f i represents the i-th manually designed feature, p(f i ) represents the p-value of the ANOVA test of the ith feature, and α represents the set significance level for feature screening.

[0082] 2.2.1 NCP

[0083] According to their properties, A, U, C, and G are divided into multiple categories, as shown in Table 2. This descriptor has been widely used in the prediction of various nucleotide modification sites. Specifically, A and G are double-ring structures, C and U are single-ring structures; A and C belong to amino groups, G and U belong to keto groups; A and U form weak hydrogen bonds, and C and G can form strong hydrogen bonds. Therefore, based on these chemical properties, A, U, C, and G can be represented by coordinates (vector encoding) (1,1,1), (0,0,1), (0,1,0), and (1,0,0), respectively.

[0084] Table 2 Classification information and corresponding coordinate representation

[0085]

[0086] 2.2.2ANF

[0087] In the analysis of RNA sequences, the autocorrelation feature (ANF) is used to describe the distribution of nucleotides in a sequence and their position correlation in the sequence by calculating the cumulative frequency d of each nucleotide from the beginning of the sequence to position i. i Specifically, the density d of the nucleotide at position i in the RNA sequence i is defined as:

[0088]

[0089] Among them, |N i | represents the length of the prefix subsequence from the beginning of the sequence to position i, that is, the number of nucleotides before the i-th position. f(n j ) is an indicator function that indicates when nucleotide n at position j j With target nucleotide n i When the same, f(n j ) is equal to 1, otherwise it is 0.

[0090] 2.2.3K-mer

[0091] k-mer calculates the frequency of consecutive k nucleotides to extract short-range information of RNA sequences. Using k-mer frequencies, each primary RNA sequence R can be converted into a vector with 4k elements as follows:

[0092]

[0093] in, is the normalized frequency of occurrence of the i-th k-tuple:

[0094]

[0095] Among them, n i is the number of nucleotide components in the i-th k-tuple in the RNA sequence, and L represents the length of the RNA sequence.

[0096] Step 3: Model building

[0097] During the feature encoding process, the Promoter-BERT model has been fine-tuned according to the characteristics of RNA sequences. Figure 2 The five-fold cross validation results of the fine-tuned model are shown. Figure 2 It can be seen that no matter what the data set is, the prediction results of 3mer and 5mer are very similar, and all indicators are much better than 1mer. Therefore, we choose the models of 3mer and 5mer as the basic models.

[0098] Based on the above features, a lightweight gradient boosting machine (LGBM) is used to build a basic model. In order to improve the accuracy and stability of the prediction, the present invention integrates two fine-tuning models based on Promoter-BERT with a feature model based on LGBM, and integrates the prediction probabilities of each model through a soft voting mechanism to improve the stability and accuracy of the prediction. In soft voting, the prediction probability of each basic model is used to calculate the final integrated output. Specifically, the prediction probability of each category is the weighted average of the prediction probabilities of all basic models for that category.

[0099] If there are M base models and N categories, the predicted probability of the mth base model for the nth category is denoted as Pmn, then the predicted probability of the nth category under soft voting is P soft (n) The calculation formula is:

[0100]

[0101] Among them, P mn Represents the predicted probability of the mth basic model for the nth class;

[0102] The soft voting mechanism averages the predicted probabilities of all base models for each category, and the final result C consists of the category with the highest average predicted probability as the final prediction result of the integrated model 2OMPro:

[0103] C = argmax n P soft (n).

[0104] Among them, argmax n Denotes that P soft (n) The maximum value of n, where n is the category number.

[0105] Step 4: Model evaluation and results analysis

[0106] In the experiment, we used the accuracy (ACC), Matthews correlation coefficient (MCC), precision and recall to evaluate the performance of the model (2OMPro). Their definitions are as follows:

[0107]

[0108] Among them, TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives, respectively. The area under the receiver operating characteristic curve (AUROC) was also used, which comprehensively evaluates the performance of the model by balancing the true positive rate (also known as sensitivity) and the false positive rate. Its value ranges from 0 to 1, and the closer it is to 1, the better the model performance. AUROC is not affected by class imbalance and data distribution, so it is very robust. AUROC is a valuable indicator for comparing the performance of different models and helps to determine the optimal classification threshold that balances sensitivity and specificity. The results are shown in Figure 3 shown.

[0109] For the Am dataset, the AUC value of the 3mer model is 0.9300, the AUC value of the 5mer model is 0.9325, the AUC value of the 3-5mer combination is 0.9365, the AUC value of the 3mer and LGBM model is 0.9418, the AUC value of the 5mer and LGBM model is 0.9437, and the AUC value of the 3-5mer and LGBM model is 0.9451. This trend is consistent in different datasets (including Cm, Gm, and Um), among which the combination of the 3mer and 5mer models with the LGBM model shows the best performance. This improvement is attributed to the combination of the popular large language model BERT with traditional feature extraction methods, which enables the model to effectively capture RNA sequence information, thereby obtaining more efficient prediction results.

[0110] In this paper, the integrated model 2OMPro model consists of three key sub-models: two base models are fine-tuned by the internal DNA language model Promoter-BERT; the third base model is built using lightweight gradient boosting machine (LGBM) based on features extracted from RNA sequences and fused with Word2Vec embedding. These sub-models are integrated into an integrated model that combines their prediction probabilities.

[0111] This invention marks the first application of a large language model (Promoter-BERT) to 2OM site prediction, significantly improving performance by leveraging its ability to extract subtle information from RNA sequences. The ensemble algorithm design integrates the predictions of three base models, mitigating the prediction errors of each model, resulting in more stable and reliable predictions. Results on four different datasets show that this embodiment consistently outperforms existing state-of-the-art models, demonstrating its robustness and effectiveness.

[0112] The model of the present invention is not only applicable to human RNA, but can also be extended to RNA analysis of other species and is applicable to various types of RNA modification studies, such as methylation, acetylation, etc. Compared with traditional experimental methods, the present invention provides a cost-effective, fast and accurate prediction method, which is particularly suitable for processing large-scale data, greatly saving time and economic costs.

[0113] In summary, the technical solution of the present invention is innovative in theory and shows 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, which helps to promote research and application development in related biomedical fields.

[0114] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for predicting 2OM sites in RNA based on ensemble learning, characterized in that: The following steps are involved: S1: Data Preparation 2OM site data were collected from public databases and randomly divided into training and test sets according to a set ratio; S2: Feature Learning Three manually designed features, K-mer, ANF and NCP, were extracted from RNA sequences, and the RNA sequences were vectorized using the Word2Vec model to obtain Word2Vec embedded features. ANOVA technology was used to select a subset of features highly correlated with 2OM site prediction from the manually designed features, and combined with the embedded features generated by the Word2Vec model. S3: Model building Based on 1mer, 3mer and 5mer word segmentation, texts of sequence units of different lengths, i.e. segmented sequences, are obtained. The segmented sequences are input into the pre-trained Promoter-BERT model for fine-tuning, and three fine-tuned models are obtained. Based on the obtained feature subset, a basic model is constructed using a lightweight gradient boosting machine, and it is integrated with the partial model obtained by fine-tuning the pre-trained Promoter-BERT model. The prediction probabilities of each basic model are integrated through a soft voting mechanism to form the final integrated model 2OMPro, in which the partial model obtained by fine-tuning the pre-trained Promoter-BERT model is also the basic model. S4: Optimization and Evaluation The performance of the integrated model 2OMPro was evaluated using cross-validation and independent test sets; S5: Deployment and prediction The integrated model 2OMPro is deployed in actual biomedical research to predict unknown 2OM sites.

2. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 1, characterized in that: In step S1, the specific processing process is as follows: S11: Collect raw 2OM loci data from RMBase v2.0 and experimental datasets generated based on Nm-seq technology; S12: For each 2OM site data, RNA fragments with a set length of nucleotides were cut from the upstream and downstream of the 2OM site as positive samples; at the same time, sequences of the same length were randomly selected from the upstream and downstream of each 2OM site as negative samples; S13: Set the sequence similarity threshold, use CD-HIT software to perform sequence clustering to remove redundant sequences, and obtain a data set with equal positive samples and negative samples, that is, obtain an RNA sequence data set; S14: Based on the modified base types of the 2OM site, the dataset was divided into four subsets, and each subset was randomly divided into a training set and a test set according to a set ratio.

3. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 1, characterized in that: In step S2, in order to cover different levels of sequence information, the RNA sequence is segmented into 1mer, 3mer, and 5mer as vocabulary, and then the RNA sequence is input as a sentence into a pre-trained Promoter-BERT model for fine-tuning to obtain a fine-tuned Promoter-BERT model based on 1mer, 3mer, and 5mer.

4. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 1, characterized in that: In step S2, the expression for encoding the RNA sequence by the fine-tuned Promoter-BERT model is as follows: XBERT=BERT(S) Where S represents the input RNA sequence, BERT represents the fine-tuned Promoter-BERT model, and X BERT Represents the encoding of the input sequence S by the fine-tuned Promoter-BERT model.

5. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 1, characterized in that: In step S2, the three manually designed features are nucleotide composition K-mer, nucleotide compound property NCP, and auto-correlation property ANF; wherein the nucleotide composition K-mer is used to calculate the frequency of k consecutive nucleotides to extract the short-range information of the RNA sequence, the nucleotide compound property NCP is used to use vector encoding (1,1,1), (0,0,1), (0,1,0) and (1,0,0) to represent A, U, C and G respectively, and the auto-correlation property ANF is used to describe the distribution of nucleotides in the RNA sequence and their position correlation in the sequence, by calculating the cumulative occurrence frequency d of each nucleotide from the beginning of the sequence to position i i to achieve.

6. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 5, characterized in that: In the nucleotide composition K-mer, using the k-mer frequency, each primary RNA sequence R can be converted into a vector with 4k elements as follows: Among them, k -mer is the normalized frequency of occurrence of the i-th k-tuple: Among them, n i is the number of nucleotide components in the i-th k-tuple in the RNA sequence, and L represents the length of the RNA sequence.

7. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 5, characterized in that: In the autocorrelation feature ANF, the cumulative occurrence frequency d of each nucleotide in the RNA sequence from the beginning of the sequence to position i i Defined as: Among them, |N i | represents the length of the prefix subsequence from the beginning of the sequence to position i, that is, the number of nucleotides before the i-th position, f(n j ) is an indicator function that indicates when nucleotide n at position j j When the target nucleotide ni is the same, f(n j ) is equal to 1, otherwise it is 0.

8. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 5, characterized in that: In step S2, the specific process of selecting a feature subset highly correlated with 2OM site prediction from the manually designed features using the ANOVA technique is as follows: Assuming there are m manually designed features, the feature subset F selected by ANOVA technology selected It is expressed as: F selected ={f i ∣p(f i )≤α} Among them, f i represents the i-th manually designed feature, p(f i ) represents the p-value of the ANOVA test of the ith feature, and α represents the set significance level for feature screening.

9. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 3 or 8, characterized in that: In step S3, the partial models obtained by fine-tuning the pre-trained Promoter-BERT model as the basic model are respectively fine-tuned Promoter-BERT models based on 3mer and 5mer.

10. A method for predicting 2OM sites in RNA based on ensemble learning according to claim 9, characterized in that: In step S3, if there are M basic models and N categories, the predicted probability of the mth basic model for the nth category is denoted as Pmn, then the predicted probability of the nth category under soft voting is P soft (n) The calculation formula is: Among them, P mn Represents the predicted probability of the mth basic model for the nth class; The soft voting mechanism averages the predicted probabilities of all base models for each category, and the final result C consists of the category with the highest average predicted probability as the final prediction result of the integrated model 2OMPro: C=argmax n P soft (n) Among them, argmax n Denotes that P soft (n) The maximum value of n, where n is the category number.

Images