A new method for identifying RNA pseudouridine sites

Abstract

This solution discloses a new method for identifying RNA pseudouridine sites. This method proposes to use a variety of feature representation techniques to extract sequence features, and then uses the SVM-RFE method for feature selection to compress the feature space and optimize the feature subset. The best feature set after feature selection is input into the kernel method KeMRF based on polynomial random forest to identify pseudouridine sites in the sequence. As a newly proposed classification method, compared with the traditional random forest, KeMRF not only optimizes the discriminant criterion for node splitting, but also combines with an easy-to-interpret kernel method, making the classification performance more superior. This method reduces the training time of the model, improves the classification performance of the model, and further enhances the accuracy of identifying pseudouridine sites.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a novel method for identifying RNA pseudouridine sites. Background Technology

[0002] Pseudoruridin (pyrimidine) is one of the most common post-transcriptional modifications of RNA. Identification of pseudouridine (pyrimidine) sites is a crucial step in understanding RNA function, RNA structural stability, translation processes, and RNA stability. Currently, commonly used methods for pseudouridine site identification include random forest, support vector machine, and convolutional neural network methods. However, these methods generally suffer from limited biological sequence features, resulting in limited classification accuracy and efficiency. Therefore, this proposal suggests a method that combines multiple feature extraction methods to extract sequence features, thereby improving identification accuracy. Simultaneously, feature selection methods are used to optimize these multiple sequence features, improving the efficiency and accuracy of subsequent model identification. Finally, a kernel method based on multinomial random forest is proposed to identify pseudouridine sites in target sequences based on the selected features. Compared to traditional random forests, this method not only optimizes the node splitting criteria but also incorporates an easily interpretable kernel method, resulting in superior classification performance. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned problems by providing a novel method for identifying pseudouridine sites in RNA. This method employs multiple feature representation techniques to extract sequence features, then utilizes the SVM-RFE method for feature selection to compress the feature space and optimize the feature subset. The optimal feature set after feature selection is input into the KeMRF kernel method based on multinomial random forest to identify pseudouridine sites in the sequence. As a newly proposed classification method, KeMRF, compared with traditional random forest, not only optimizes the node splitting discrimination criterion but also combines an easily interpretable kernel method, resulting in superior classification performance. This method reduces model training time, improves model classification performance, and further enhances the accuracy of pseudouridine site identification.

[0004] To achieve the above objectives, the present invention adopts the following technical solutions:

[0005] A novel method for identifying RNA pseudouridine sites, comprising:

[0006] S1. Prepare a multi-species dataset for training, each dataset consisting of several sequences containing AUGGs;

[0007] S2. Multiple sequence feature extraction methods are used to extract features from the sequence data to obtain multiple feature subsets composed of feature vectors;

[0008] S3. Merge multiple feature subsets to obtain the original feature set, and then use feature selection methods to select features from the original feature set to obtain the optimal feature set;

[0009] S4. Using the KeMRF kernel method based on multinomial random forest as the classifier, train the classifier using the optimal feature set;

[0010] S5. Use the trained classifier to identify pseudouridine sites in the sequence to be identified.

[0011] In the aforementioned new method for identifying target sites in sequences, step S5 specifically includes:

[0012] S5-1. Use multiple sequence feature extraction methods to extract features from the sequence to be identified to obtain multiple feature subsets;

[0013] S5-2. Merge the obtained multiple feature subsets to obtain the original feature set, and use the feature selection method to select features from the original feature set to obtain the optimal feature set;

[0014] S5-3. Input the optimal feature set into the trained KeMRF classifier and output the identification result of the pseudouridine site of the sequence to be identified.

[0015] Among the novel methods for identifying target sites in the aforementioned sequences, various sequence feature extraction methods include binary extraction, PSTNPss extraction, NCP extraction, and PseKNC extraction.

[0016] Among the aforementioned new methods for identifying target sites in sequences, binary extraction, PSTNPss extraction, NCP extraction, and PseKNC extraction are all existing efficient sequence feature extraction methods.

[0017] In the above-mentioned new method for identifying target sites in sequences, in step S3, the SVM-RFE algorithm is used to select features from the original feature set;

[0018] The SVM-RFE algorithm uses a model trained with training samples constructed from the original feature set.

[0019] In the novel method for identifying target sites in the aforementioned sequence, the classifier is constructed based on the kernel method KeRF of random forest and the multinomial random forest method MRF.

[0020] In the aforementioned novel method for identifying target sites in sequences, the method for constructing the classifier includes:

[0021] S41. First, construct a multinomial random forest;

[0022] S42. For unclassified sample points, calculate the posterior probability that the unclassified sample point and all training sample points fall on the same leaf node in the multinomial random forest.

[0023] S43. Multiply the original label of the corresponding training sample by the sum of the posterior probabilities, and divide by the sum of the posterior probabilities to obtain an approximate prediction value;

[0024] S44. Based on the obtained approximate prediction values, use the sign function to predict the label values.

[0025] In the novel method for identifying target sites in the aforementioned sequence, step S41, the method for constructing the multinomial random forest, includes randomly dividing the training set into two non-overlapping subsets: a structure point set for constructing the tree structure and an estimation point set for predicting leaf labels.

[0026] (1) Selection of node splitting features

[0027] First, normalize the eigenvectors, let G = {g ij} represents the set of all possible split points of a node, where g ij This represents the j-th eigenvalue under the i-th eigenvalue, and a vector G is obtained based on G. m ={g1, g2, ..., g n}, where gi is the maximum impurity reduction of the i-th feature, and n is the number of features. G m Normalized to G′ m , represented as

[0028]

[0029] Second, calculate the probability.

[0030]

[0031] Where B1 (B1≥0) is a parameter related to the privacy budget, and the splitting feature is randomly selected according to a multinomial distribution;

[0032] (2) Selection of splitting value corresponding to splitting feature

[0033] After determining the splitting characteristics of the nodes, select the corresponding splitting values ​​to construct two child nodes. First, normalize the splitting value vector, then calculate the probability, and randomly select the splitting value of the node according to the multinomial distribution.

[0034] Repeat the two-step node splitting process (1) and (2) until the criterion for stopping splitting is met;

[0035] (3) Determine leaf label information

[0036] After determining the overall structure of the tree, the label information of its leaf nodes is determined based on the estimated point set. First, the predicted probability vector ρ is re-determined based on the structural point set of each leaf γ.

[0037] ρ=(ρ1,ρ2,...,ρ k )

[0038]

[0039] Where γ e Let k be the set of estimated points in leaf γ, and k be the number of class labels. Calculate the number of labels in the structure point set that are the same as the current category, and then calculate the probability γ for each leaf.

[0040]

[0041] B3 (B3≥0) is also a parameter related to the privacy budget, and leaf label information is randomly selected for each leaf γ according to a multinomial distribution.

[0042] In the novel method for identifying target sites in the aforementioned sequence, during the KeMRF classifier construction process in S4, the classifier hyperparameters are optimized in the following way:

[0043] S4-1. Initialize the hyperparameters B1, B2, and B3 in the KeMRF kernel method based on multinomial random forest;

[0044] S4-2. Calculate the accuracy of KeMRF in classifying pseudouridine site sequences, and use the accuracy as the basis for hyperparameter selection;

[0045] S4-3. Based on the obtained classification accuracy, continuously update the hyperparameters B1, B2, and B3 related to the privacy budget and other hyperparameters;

[0046] S4-4. Determine if the grid search is complete. If it is, proceed to S4-5; otherwise, proceed to S4-2 and continue iterating.

[0047] S4-5. Input the obtained optimal hyperparameters B1, B2, and B3, along with other hyperparameters, into the KeMRF classifier to construct the KeMRF classification model with the highest accuracy.

[0048] In the new method for identifying target sites in the aforementioned sequence, the classification accuracy in S4-2 is represented by ACC, and its calculation method is as follows:

[0049]

[0050] Where ACC represents the accuracy of the KeMRF algorithm in classifying pseudouridine sites, TP represents the number of correctly predicted positive samples, FP represents the number of incorrectly predicted positive samples, TN represents the number of correctly predicted negative samples, and FN represents the number of incorrectly predicted negative samples. In the task of identifying pseudouridine sites, samples containing pseudouridine sites are considered positive samples, and samples not containing pseudouridine sites are considered negative samples.

[0051] The advantages of this invention are as follows:

[0052] This paper proposes to use multiple feature extraction methods to extract features from sequences, thereby improving the accuracy of recognition through various sequence features;

[0053] Based on the aforementioned feature extraction methods, we further propose using the SVM-RFE method for feature selection to optimize the feature space, which not only reduces the training time of the model but also improves the classification performance of the model.

[0054] A novel classification method, KeMRF, is proposed. It is an improvement on the advantages of the kernel method KeRF and the multinomial random forest method. Compared with the traditional random forest, it not only optimizes the discrimination criterion for node splitting, but also combines an easily interpretable kernel method to improve the performance of the model classification, thereby further improving the accuracy of pseudouridine site identification.

[0055] Experiments have shown that the proposed method has a significant advantage in recognition accuracy compared to existing methods. Attached Figure Description

[0056] Figure 1 The diagram shown is a flowchart of a novel method for identifying RNA pseudouridine sites provided by an embodiment of the present invention;

[0057] Figure 2 The diagram shown is a comparison of feature encoding schemes provided in embodiments of the present invention.

[0058] Figure 3 The diagram shown is a comparative illustration of the feature selection methods provided in the embodiments of the present invention.

[0059] Figure 4 The diagram shown is a schematic of the KeMRF kernel method based on multinomial random forest provided in an embodiment of the present invention.

[0060] Figure 5 The diagram shows a comparison of the performance of KeMRF and other classifiers on the training set according to an embodiment of the present invention.

[0061] Figure 6 The diagram shows a comparison of the performance of KeMRF and other classifiers on an independent dataset, as provided in this embodiment of the invention.

[0062] Figure 7 The diagram shows a comparison of the performance of the model provided in this embodiment of the invention with other models on the same independent dataset. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0064] This embodiment provides a novel method for identifying pseudouridine sites in RNA. Combining multiple RNA feature extraction methods and efficient feature selection techniques, and using the newly proposed KeMRF method as the classifier in the model, it can more accurately identify pseudouridine sites in RNA sequences. Figure 1 As shown, the specific method is as follows:

[0065] S1. Collect and download existing datasets provided by researchers. The datasets contain training sets and independent test sets. Each dataset consists of several sequences containing AUGG.

[0066] S2. Based on the acquired multi-species dataset, multiple sequence feature extraction methods are used to transform the sequence data into feature vectors to obtain multiple feature subsets;

[0067] S3. The feature vectors generated by different feature extraction methods are fused to obtain the original feature set. Then, feature selection is performed to optimize the feature space and obtain the optimal feature set.

[0068] S4. The KeMRF kernel method based on multinomial random forest is used as the classifier of the model. The optimal hyperparameters of the model are determined by grid search and 10-fold cross-validation. The classifier is then trained using the best feature set obtained from the training set.

[0069] S5. Construct a pseudouridine site recognizer based on the trained classification model, and input the optimal feature set obtained from the test set into the recognizer to complete the recognition of pseudouridine sites. After training and testing, the obtained pseudouridine site recognizer can be used for the recognition of pseudouridine sites.

[0070] Furthermore, in step S1, before obtaining the original sequence data, it is necessary to perform format judgment on the downloaded multi-species pseudouridine site sequence data file. The specific method for format judgment is as follows: when a line of the read-in sequence data file containing pseudouridine sites and not containing pseudouridine sites starts with the character ">", then the next line is taken as sequence text data.

[0071] Furthermore, such as Figure 2As shown, six common RNA feature encoding schemes were compared during the construction process, including Binary, NCP, PSTNPss, Kmer, ANF, and PseKNC. These were input as individual features into the KeMRF classifier, and their performance was evaluated based on the accuracy of 10-fold cross-validation. According to the results in the figure, the PSTNPss feature outperformed the other features in all three species. Considering that the predictive performance of training a model using a single type of feature is usually limited, this scheme selected four features—PSTNPss, Binary, NCP, and PseKNC—to be combined to construct the model's feature space based on the test accuracy. Therefore, the feature extraction methods in step S2 include four RNA feature extraction methods: binary extraction, PSTNPss extraction algorithm, NCP extraction algorithm, and PseKNC extraction method. Using the aforementioned feature extraction methods, the original sequence data was transformed into numerical vectors, resulting in multiple feature subsets.

[0072] Furthermore, such as Figure 3 As shown, during the construction process, three feature selection methods were compared: Support Vector Machine Recursive Feature Elimination (SVM-RFE), Maximum Relevance Minimum Redundancy (mRMR), and Chi Square test. This embodiment applies each method to the model for testing and then compares their performance within the model. Figure 3 The test accuracy of these methods on three species training datasets is shown. The test results indicate that the SVM-RFE method outperforms the other two methods on these three species datasets. Therefore, this approach selects the SVM-RFE method for feature selection, optimizes the feature space, and further improves the accuracy of pseudouridine site identification. Specifically, in step S3, the multiple feature subsets obtained in step S2 are merged to obtain the original feature set, and then the SVM-RFE algorithm is used to compress the feature vector space.

[0073] Binary extraction, PSTNPss extraction, NCP extraction, and PseKNC extraction are all methods for extracting features from sequence data by encoding features based on the different chemical properties of nucleotides. The specific implementation methods are similar to those of the corresponding general methods and will not be elaborated here.

[0074] The SVM-RFE algorithm trains a model using training samples formed from the original feature set obtained through the aforementioned feature extraction methods. Its principle is as follows: each feature is sorted by score, the feature with the lowest score is removed, and the model is then trained again using the remaining features. This process is repeated for the next iteration, and finally, the required number of features is selected. Similarly, the specific training method is similar to that of general implementations of the corresponding method and will not be elaborated upon here.

[0075] Furthermore, the construction process of the KeMRF kernel method based on multinomial random forest in step S4 is divided into the following steps:

[0076] This method is proposed based on the kernel method KeRF of random forest and the multinomial random forest method MRF;

[0077] First, the training set is divided. MRF replaces the random sampling of the training set used in RF with training set partitions. For example... Figure 4 As shown, before building the tree, the training set is randomly divided into two non-overlapping subsets. One subset is called the structure point set, used to construct the tree structure. The other subset is called the estimation point set, used to predict leaf labels. The two subsets are divided according to a set ratio, and the training set needs to be re-divided before each tree construction.

[0078] When constructing the tree, two impurity-based multinomial distributions are used to split nodes: one for selecting splitting features and the other for selecting splitting values. A split point consists of both a splitting feature and a splitting value. The node feature selection method and its corresponding splitting value selection method are divided into the following steps:

[0079] (1) Selection of node splitting features

[0080] First, normalize the eigenvectors, let G = {g ij} represents the set of all possible split points of a node, where g ij This represents the j-th eigenvalue under the i-th eigenvalue. A vector G is obtained from G. m ={g1, g2, ..., g n}, where gi is the maximum impurity reduction of the i-th feature, and n is the number of features. G m Normalized to G′ m , represented as

[0081]

[0082] Second, calculate the probability.

[0083]

[0084] B1 (B1≥0) is a parameter related to the privacy budget, and the splitting feature is randomly selected according to a multinomial distribution.

[0085] (2) Selection of splitting value corresponding to splitting feature

[0086] After determining the splitting characteristics of a node, it is necessary to select the corresponding splitting values ​​to construct two child nodes. Similar to the steps for determining the splitting characteristics, first normalize the splitting value vector, then calculate the probability, and randomly select the node splitting value according to the multinomial distribution.

[0087] Repeat steps (1) and (2) to split the node until the criterion for stopping splitting is met. The criterion for stopping splitting is related to the minimum number of leaf nodes specified.

[0088] (3) Determine leaf label information

[0089] After determining the overall structure of the tree, it is necessary to determine the label information of its leaf nodes based on the estimated point set. First, the predicted probability vector ρ is re-determined based on the structural point set of each leaf γ.

[0090] ρ=(ρ1,ρ2,…,ρ k )

[0091]

[0092] Where γ e Let k be the set of estimated points in leaf γ, and k be the number of class labels. Calculate the number of labels in the structure point set that match the current category. Then, calculate the probability γ for each leaf.

[0093]

[0094] B3 (B3≥0) is also a parameter related to the privacy budget. A leaf label is randomly selected for each leaf γ according to a multinomial distribution.

[0095] After the multinomial random forest is constructed, for unclassified sample points, unlike traditional prediction methods that directly input the data into the random forest and output corresponding labels, this method will provide corresponding predicted labels based on the KeRF kernel method of random forest.

[0096] First, the posterior probability that an unclassified sample point and all training sample points fall on the same leaf node in the multinomial random forest is calculated. Then, the original label of the corresponding training sample is multiplied by the sum of the posterior probabilities, and divided by the sum of the posterior probabilities to obtain an approximate predicted value. Finally, based on the obtained approximate predicted value, the sign function is used to predict the label value. The sign function in the kernel method based on random forest is a judgment based on the approximate predicted value. When the approximate predicted value is less than a certain threshold, the unclassified sample is considered a negative sample; when the approximate predicted value is greater than a certain threshold, the unclassified sample is considered a positive sample, thus obtaining the label value of the unclassified sample.

[0097] Furthermore, the principle behind using grid search and 10-fold cross-validation in step S4 to optimize the model's classification performance is as follows:

[0098] Hyperparameters are crucial for classification models in the KeMRF kernel method based on multinomial random forests; appropriate hyperparameters can ensure the model's classification accuracy. A 10-fold cross-validation method is used to evaluate the model's classification performance during modeling, as detailed below:

[0099] First, the training data is divided into 10 subsets. One subset is selected as the validation set, and the remaining nine subsets are used as the training set. This process is repeated 10 times, and the average accuracy is calculated. This result is used to evaluate the model performance. Based on the above model evaluation method, a grid search method is used to optimize the parameters and select the optimal hyperparameters. By iterating through all candidate parameters and comparing the average accuracy results, the hyperparameter with the highest average accuracy is selected and incorporated into the model.

[0100] Furthermore, step S4 can be specifically divided into the following steps:

[0101] S4-1. Initialize the hyperparameters B1, B2, and B3 in the KeMRF kernel method based on multinomial random forest;

[0102] S4-2. Calculate the accuracy of KeMRF in classifying pseudouridine site sequences, and use the accuracy as the basis for hyperparameter selection;

[0103] S4-3. Based on the obtained classification accuracy, continuously update the hyperparameters B1, B2, and B3 related to the privacy budget and other hyperparameters;

[0104] S4-4. Determine if the grid search is complete. If it is, proceed to S4-5; otherwise, proceed to S4-2 and continue iterating.

[0105] S4-5. Input the obtained optimal hyperparameters B1, B2, and B3, along with other hyperparameters, into the KeMRF classifier to construct the KeMRF classification model with the highest accuracy.

[0106] Furthermore, the classification accuracy in S4-2 is represented by ACC, and its calculation method is as follows:

[0107]

[0108] Where ACC represents the accuracy of the KeMRF algorithm in classifying pseudouridine sites, TP represents the number of correctly predicted positive samples, FP represents the number of incorrectly predicted positive samples, TN represents the number of correctly predicted negative samples, and FN represents the number of incorrectly predicted negative samples. In the task of identifying pseudouridine sites, samples containing pseudouridine sites are considered positive samples, and samples not containing pseudouridine sites are considered negative samples.

[0109] Step S5 specifically includes the following steps:

[0110] S5-1. Use multiple feature extraction methods, including binary, NCP, PSTNPss and PseKNC, to extract features from the independent test set and obtain multiple feature subsets;

[0111] S5-2. The feature subsets obtained above are fused, and then the SVM-RFE method is used for feature selection;

[0112] S5-3. Input the obtained optimal feature set into the KeMRF classifier to obtain the classification results of the samples, and evaluate the classification effect.

[0113] This approach yields a model capable of identifying pseudouridine sites on RNA. When faced with an RNA sequence to be predicted, similar to testing with a test set, the same multiple sequence feature extraction techniques are used to convert the RNA sequence into a feature vector, resulting in multiple feature subsets. Subsequently, feature selection methods are used to select features, optimize the feature space, and obtain the optimal feature set for the RNA sequence. Finally, the optimal feature set is input into the trained classification model, which then outputs the pseudouridine site identification result for the sequence to be identified.

[0114] like Figure 5 The diagram illustrates a comparison of the performance of KeMRF provided in this embodiment of the invention with other classifiers on the training set. To demonstrate the prediction accuracy of the KeMRF method, we compare its performance with other commonly used tree model methods on the same dataset. The experimental steps involve performing steps S5-1, S5-2, and S5-3 for each different method to statistically analyze its classification performance. Based on the accuracy comparison results in the figure, it can be seen that KeMRF's classification performance is superior to the other two methods, demonstrating the effectiveness of the method of this invention.

[0115] Figure 6 The diagram illustrates a comparison of the performance of KeMRF provided in this embodiment of the invention with other classifiers on an independent dataset. To further illustrate the generalization performance of this method, this embodiment compares its performance with other commonly used tree model methods on the same dataset. The experimental steps involve training different methods on the training set using selected hyperparameters to obtain trained classification models, and then inputting data from the test set into the trained models to obtain the final results. Figure 6 As can be seen, KeMRF outperforms the other two methods in terms of test accuracy on the independent test set, further demonstrating that the model in this invention has better generalization performance and can improve the accuracy of pseudouridine site identification.

[0116] Figure 7The diagram illustrates a comparison of the performance of the model provided in this embodiment of the invention with other models on the same independent dataset. The experiments used independent test datasets to evaluate and validate the trained model, assessing and comparing its predictive performance with other existing methods. Other existing methods include iRNA-PseU, PseUI, iPseU-CNN, RF-PseU and Porpoise; PseU-KeMRF is the method of this invention. As can be seen from the results, PseU-KeMRF outperforms other models in terms of test accuracy on both independent test datasets, achieving highly competitive performance on the independent test datasets.

[0117] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for identifying RNA pseudouridine sites, characterized in that, The method includes: S1. Prepare a multi-species dataset for training, each dataset consisting of several sequences containing AUGGs; S2. Four sequence feature extraction methods, namely binary, PSTNPss, NCP and PseKNC, are used to extract features from the sequence data to obtain multiple feature subsets composed of feature vectors; S3. Merge multiple feature subsets to obtain the original feature set, and then use the SVM-RFE algorithm to select features from the original feature set to obtain the optimal feature set; S4. Using the KeMRF kernel method based on multinomial random forest as the classifier, train the classifier using the optimal feature set; The classifier is constructed based on the kernel method KeRF of random forest and the multinomial random forest method MRF, and the construction method includes: S41. First, construct a multinomial random forest; S42. For unclassified sample points, calculate the posterior probability that the unclassified sample point and all training sample points fall on the same leaf node in the multinomial random forest. S43. Multiply the original label of the corresponding training sample by the sum of the posterior probabilities, and divide by the sum of the posterior probabilities to obtain an approximate prediction value; S44. Based on the obtained approximate prediction values, use the sign function to predict the label values; During the construction process, the classifier hyperparameters were optimized in the following ways: S4-1. Initializing Hyperparameters in the KeMRF Kernel Method Based on Multinomial Random Forest ; S4-2. Calculate the accuracy of KeMRF in classifying pseudouridine site sequences, and use the accuracy as the basis for hyperparameter selection; S4-3. Continuously update the hyperparameters related to the privacy budget based on the obtained classification accuracy. and Other hyperparameters; S4-4. Determine if the grid search is complete. If it is, proceed to S4-5; otherwise, proceed to S4-2 and continue iterating. S4-5. Obtain the optimal hyperparameters and These, along with other hyperparameters, are fed into the KeMRF classifier to build the KeMRF classification model with the highest accuracy. S5. Use the trained classifier to identify pseudouridine sites in the sequence to be identified: S5-1. Use four sequence feature extraction methods, namely binary, PSTNPss, NCP, and PseKNC, to extract features from the sequence to be identified and obtain multiple feature subsets; S5-2. The obtained feature subsets are merged to obtain the original feature set. The SVM-RFE algorithm is used to select features from the original feature set to obtain the optimal feature set. S5-3. Input the optimal feature set into the trained KeMRF classifier and output the identification result of the pseudouridine site of the sequence to be identified.

2. The method for identifying RNA pseudouridine sites according to claim 1, characterized in that, The binary extraction method, PSTNPss extraction method, NCP extraction method, and PseKNC extraction method all use extraction methods trained on sequence datasets.

3. The method for identifying RNA pseudouridine sites according to claim 1, characterized in that, In step S3, The SVM-RFE algorithm uses a model trained with training samples constructed from the original feature set.

4. The method for identifying RNA pseudouridine sites according to claim 1, characterized in that, In step S41, the method for constructing a multinomial random forest includes randomly dividing the training set into two non-overlapping subsets: a structure point set for constructing the tree structure and an estimation point set for predicting leaf labels. (1) Selection of node splitting features First, normalize the feature vectors, let Let represent the set of all possible split points of a node, where Indicates the first The first feature under the first characteristic Each feature value, based on Get a vector ,in For the first The maximum impurity reduction of each characteristic The number of features. Normalization , represented as ， Second, calculate the probability. ， in These are parameters related to the privacy budget, and the splitting features are randomly selected based on a multinomial distribution; (2) Selection of splitting values ​​corresponding to splitting features After determining the splitting characteristics of the nodes, select the corresponding splitting values ​​to construct two child nodes. First, normalize the splitting value vector, then calculate the probability, and randomly select the splitting value of the node according to the multinomial distribution. Repeat the two-step node splitting process (1) and (2) until the criterion for stopping splitting is met; (3) Determine leaf label information After determining the overall structure of the tree, the label information of its leaf nodes is determined based on the estimated point set. First, based on each leaf node... The structure point set redetermines the predicted probability vector. , ， ， in It is a leaf The set of estimated points, It is the number of category labels. Calculate the number of labels in the structure point set that are the same as the current category, and then calculate the number of labels for each leaf. probability ， in This is also a parameter related to privacy budgets, randomly assigned to each leaf according to a multinomial distribution. Select a leaf label.

5. The method for identifying RNA pseudouridine sites according to claim 4, characterized in that, In S4-2, the classification accuracy is represented by ACC, and it is calculated as follows: Where ACC represents the accuracy of the KeMRF algorithm in classifying pseudouridine sites, TP represents the number of correctly predicted positive samples, FP represents the number of incorrectly predicted positive samples, TN represents the number of correctly predicted negative samples, and FN represents the number of incorrectly predicted negative samples.

Images