A deep learning prediction method for comprehensive classification of RNA based on graph representation
By using graph representation and multi-channel convolutional neural network models, the specificity problem of RNA classification tools has been solved, achieving universality and biological interpretability of RNA classification, and enhancing its auxiliary role in RNA function research and drug development.
Patent Information
- Application Number
- CN202411786736.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-06
- Publication Date
- 2025-12-16
- Estimated Expiration
- 2044-12-06
AI Technical Summary
Existing RNA classification tools are too specific to be applied to the entire RNA system and cannot capture long-distance nucleotide interactions, resulting in a lack of biological interpretability in predicting RNA coding potential.
We employ a graph-based deep learning prediction method for comprehensive RNA classification. By constructing a graph-represented RNA dataset and a multi-channel convolutional neural network model, combined with hyperparameter optimization, we generate an RNA classification prediction model that captures long-distance nucleotide interactions and performs comprehensive classification.
This has enabled the universality and biological interpretability of RNA classification, enhancing its auxiliary role in RNA function research and drug development.
Smart Images
Figure CN119920325B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of RNA comprehensive classification deep learning prediction, and particularly relates to a RNA comprehensive classification deep learning prediction method based on graph representation. BACKGROUND
[0002] RNA plays a vital role in organisms. According to whether it has coding potential, RNA can be divided into messenger RNA (mRNA) and non-coding RNA (ncRNA). As a coding template for proteins, the function of most mRNA has been widely studied. For a long time, researchers believed that ncRNA was a "garbage" fragment produced in the process of transcription. In recent years, it has been found that ncRNA accounts for more than 90% of coding RNA in the human genome, and ncRNA plays an important role in regulating the occurrence and development of diseases and life activities such as embryonic development. According to the sequence length of ncRNA, ncRNA can be divided into long non-coding RNA (lncRNA) and small non-coding RNA (sncRNA). The main problem of lncRNA classification is how to predict circular RNA (circRNA) and linear RNA, and sncRNA classification is a more complex 13-classification problem.
[0003] Since the existing tools are generally established for a certain RNA classification problem, the RNA representation method is often too specific and difficult to migrate to the classification prediction of the entire RNA system. In addition, due to the existing RNA features that cannot capture long-distance nucleotide interactions, the prediction of RNA coding potential lacks biological interpretability. The existence of the above problems leads to the inability of existing tools to comprehensively predict the RNA classification system, and hinders the auxiliary role of computer prediction in mining biological principles.
[0004] Therefore, providing a universal RNA representation and prediction method is helpful to comprehensively predict RNA classification and further mine the biological interpretability of RNA coding potential. SUMMARY
[0005] To solve the above technical problems, the present application provides a RNA comprehensive classification deep learning prediction method based on graph representation to solve the problems existing in the prior art.
[0006] To achieve the above purpose, the present application provides a RNA comprehensive classification deep learning prediction method based on graph representation, comprising the following steps:
[0007] S1: Select and process RNA data, construct a graph to represent the RNA data set and three data sets for RNA key classification prediction tasks;
[0008] S2: Adopt RNAImage graph representation strategy, construct template graph by graph representing RNA data set, and then map each RNA sequence in the three data sets for RNA key classification prediction tasks to feature using the template graph to obtain the representation graph, and all the obtained representation graphs constitute the data set after encoding processing;
[0009] S3: Adopting a double-line multi-channel convolutional neural network model, training through the data set after encoding processing, obtaining an RNA classification training model;
[0010] S4: Evaluate the RNA classification training model using an evaluation matrix, and optimize the hyperparameters to generate an RNA classification prediction model;
[0011] S5: Use the RNA classification prediction model to classify and predict the RNA needed to be predicted, and use the classification results for RNA function exploration and development of biological drugs.
[0012] In step S1, the three data sets for RNA key classification prediction tasks are the data set for RNA coding potential prediction, the data set for sncRNA sub-classification prediction, and the data set for linear and circular lncRNA prediction.
[0013] In step S2, the RNAImage graph representation strategy is adopted to construct a template graph by graph representing the RNA data set, which specifically includes:
[0014] The graph representing the RNA data set generates a template graph after sequentially passing through four processes of matrix generation, similarity calculation, dimension reduction, and grid distribution. After the four processes of matrix generation, similarity calculation, dimension reduction, and grid distribution, a template image is generated. Then, each coding feature abundance of each RNA sequence entering the model is mapped to the corresponding position of the template graph, thereby completing the class image representation of the RNA sequence.
[0015] In step S2, the coding feature abundance of each RNA sequence in the three data sets for RNA key classification prediction tasks is mapped to the feature of the template graph to obtain the representation graph.
[0016] In step S3, the model adopts multi-line multi-channel convolution, the number of lines depends on the number of template graphs, and the number of channels depends on the number of feature categories. The double-line multi-channel convolutional neural network model adopts convolution kernel for feature extraction. The model adopts three size convolution kernels (1×1, 3×3 and 5×5) for feature extraction at the same time.
[0017] In step S4, the evaluation matrix includes:
[0018] AUC evaluation, Matthew correlation coefficient evaluation, specificity evaluation, sensitivity evaluation, precision evaluation, accuracy evaluation. The process of performance evaluation of the RNA classification prediction model comprises: comprehensive evaluation of the RNA classification prediction model according to six evaluation standards of AUC, Matthew correlation coefficient, specificity, sensitivity, precision and accuracy.
[0019] In step S4, hyperparameter optimization is performed, specifically including:
[0020] The grid search method is used to optimize the parameters and find the optimal hyperparameters. The hyperparameters include the learning rate and the batch size. The grid search method is used to optimize the parameters and find the optimal hyperparameter combination (including the learning rate and the batch size).
[0021] Further preferably, the RNA deep learning classification prediction method and system based on graph representation comprise:
[0022] The RNA data is selected and processed, the graph representation RNA data set is constructed, and the data set for three key classification prediction tasks (RNA coding potential prediction, sncRNA sub-classification prediction and linear and circular lncRNA prediction) is constructed;
[0023] The RNAImage graph representation strategy is adopted, two template graphs are constructed through the graph representation RNA data set, and each RNA sequence in the classification prediction task data set is mapped to obtain the corresponding two representation graphs, so as to obtain the data set after coding processing;
[0024] A double-line multi-channel convolutional neural network (CNN) model is adopted, the deep neural network model is trained through the processed data set, and the RNA classification training model is obtained;
[0025] The RNA classification training model is evaluated by using the evaluation matrix, and the hyperparameters are optimized to generate the RNA classification prediction model;
[0026] The RNA classification prediction model is used to classify and predict various RNAs, and further combined with the biological interpretability of the coding potential of the instance mining.
[0027] Compared with the prior art, the present application has the following advantages:
[0028] In order to reduce the specificity of the RNA characterization method and improve the biological interpretability, the application provides a comprehensive RNA classification deep learning prediction method based on graph representation, expands the RNA feature space and adopts a universal representation method. Compared with the existing RNA classification method, the method has the advantages of portability and continuously achieves the optimal level in multiple RNA classification tasks. In addition, the G-feature proposed by the method can capture long-distance nucleotide interactions and can be used to explore the biological interpretability of RNA coding potential. The results of the application can be applied to RNA function research and RNA-related drug research.
[0029] The RNA comprehensive classification deep learning prediction method based on graph representation of the application is helpful for comprehensively predicting RNA classification and further exploring the biological interpretability of RNA coding potential. BRIEF DESCRIPTION OF DRAWINGS
[0030] The present specification is further illustrated by way of example embodiments, and is further explained in detail with the aid of the accompanying drawings.
[0031] Figure 1 FIG. 1 is a schematic diagram of the generation of G-feature according to the embodiments of the present specification.
[0032] Figure 2 FIG. 3 is a method flow chart of the RNAImage graph representation strategy according to the embodiments of the present specification.
[0033] Figure 3 FIG. 4 is a flow chart of the RNA deep learning classification prediction method based on graph representation according to the embodiments of the present specification. DETAILED DESCRIPTION
[0034] It should be noted that the model training and evaluation processes of the prediction method in the embodiments are basically the same, and the main difference lies in the research problems and different data set constructions of the embodiments. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0035] It should be noted that the steps shown in the accompanying flow chart can be executed in a computer system, but in specific cases, the steps shown can be executed in different order or with some steps removed.
[0036] Embodiment one
[0037] As shown in FIG. 1, a graph-based RNA representation method RNAImage is constructed in the embodiment, which includes: Figure 1
[0038] S1. RNA feature generation:
[0039] The method adopts two types of RNA features: one is the gap feature (G-feature) based on nucleotide interval, and the other is the descriptor-based feature (D-feature). The G-feature emphasizes the distribution of nucleotides in the RNA sequence (especially the relationship between nucleotide pairs), and the D-feature reflects the intrinsic, structural and physicochemical properties of the RNA sequence.
[0040] The G-feature is the RNA feature proposed by the method, as shown in the formula (1), and the generation method of the G-feature is provided by taking G-C and C-G nucleotide pairs as examples. The G-feature is calculated by counting the frequency of nucleotide pairs in different interval ranges (from 0 to 40). For example, 2-GC represents the frequency of G-C nucleotide pairs separated by two nucleotides in the RNA sequence. Figure 1
[0041] In addition, the present application generates a total of 753 D-features to encode RNA sequences, of which 314 are classical encoding features reported previously (such as open reading frame, codon-related features, and secondary structure features, etc.). The remaining 439 features are newly generated based on specific physicochemical properties or structural properties (such as hydrogen bond basicity, topological polarity surface area, and atomic ratio path length, etc.). The 439 D-features are calculated using the Composition, Transition and Distribution (CTD) strategy
[0042] S2. RNAImage chart representation strategy:
[0043] The method proposes an RNAImage strategy to realize the image-like representation of RNA. As shown in the formula (2), first, two widely used RNA databases (ENCODE and RefSeq) are used to obtain a total of 788,576 RNA sequences, and two template images (D-template and G-template) are generated through four processes in series (matrix generation, similarity calculation, dimension reduction, and grid distribution), and the sizes of the template images obtained are 28x27 (D-template) and 27x27 (G-template), respectively. Figure 2
[0044] The specific steps for generating the template images (D-template and G-template) are as follows: First, all 788,576 RNAs are transformed into two different feature matrices: 753×788,576 for D-features and 724×788,576 for G-features. Next, by calculating the cosine similarity between features, similarity matrices of 753×753 and 724×724 are obtained from the feature matrices. Then, UMAP is used to reduce the dimensionality of the features based on the similarity matrices. Finally, the JV algorithm is used to calculate the minimum cost distance, and the reduced-dimensionality images are distributed in a grid, thus obtaining the two template images.
[0045] Finally, the abundance of each coding feature of each RNA sequence to be entered into the model is mapped to the corresponding position of the two template maps, thereby converting each RNA sequence into two RNA maps, thus completing the image-like representation of the RNA sequence.
[0046] Example 2
[0047] like Figure 3 As shown, this embodiment provides an RNA deep learning classification and prediction method, IntRNA, based on graph representation, which includes:
[0048] S1. Dataset Construction:
[0049] The first step in constructing the dataset in this embodiment is to select high-quality, usable data, followed by data partitioning and utilization. The data used in this embodiment can be divided into two categories: one is graph representation RNA data used as model input, and the other is RNA data used for model training and evaluation.
[0050] The RNA data used for model input was collected from the ENCODE and Refseq databases. After deduplication and other processing, a total of 788,576 RNA data were used for template graph generation in the graph representation.
[0051] The RNA data used for model training and evaluation mainly consist of three datasets: RNA coding potential prediction (from the study CPCred by Xiaoxue Tong et al.), sncRNA subclassification prediction (from the study nRC by Antonino-Fiannaca et al.), and linear and circular lncRNA prediction (from the study circDeep by Mohamed-Chaabane et al.).
[0052] The RNA encoding potential prediction dataset contains RNA sequences of five species, in which: 74,321 RNA sequences from human (41,917 mRNA sequences, 32,404 ncRNA sequences), 51,032 RNA sequences from mouse (31,102 mRNA sequences, 19,930 ncRNA sequences), 26,256 RNA sequences from zebrafish (15,594 mRNA sequences, 10,662 ncRNA sequences), 26,256 RNA sequences from fruit fly (17,400 mRNA sequences, 4,098 ncRNA sequences), 7,126 RNA sequences from yeast (6,713 mRNA sequences, 413 ncRNA sequences). The RNA sequences of human are divided into training set and test set in the ratio of 4:1, and the data of other species are all as test set.
[0053] The sncRNA subcategory prediction dataset contains 8,920 human sncRNA sequences of 13 subcategories (“miRNA”, “5S rRNA”, “5.8S rRNA”, “ribozymes”, “CD-box”, “HACA-box”, “scaRNA”, “tRNA”, “Intron gpI”, “Intron gpII”, “IRES”, “leader”, “riboswitch”), in which each subcategory of RNA obtains 700 sequences, and “IRES” only obtains 520 sequences because of limited quantity. When dividing the dataset, the training set contains 400 sncRNA sequences of each subcategory (“IRES” only 256), the validation set contains 100 sncRNA sequences of each subcategory (“IRES” only 64), and the test set contains 200 sncRNA sequences of each subcategory.
[0054] The linear and circular lncRNA prediction dataset contains 51,625 human lncRNA sequences (19,683 linear lncRNA sequences, 31,942 circular lncRNA sequences), which are divided into 38,722 lncRNA sequences as training set, 5,161 lncRNA sequences as validation set, and 7,742 lncRNA sequences as test set in proportion.
[0055] S2. Model training
[0056] A. Model input
[0057] The input of the present application is a sequence, which is further transformed into two RNA representation maps by mapping the abundance of different features to the corresponding positions of two template maps (D-template and G-template), with the normalized abundance value of each feature ranging between 0 and 1. Both D-template and G-template are generated by four sequential steps ("matrix generation", "similarity calculation", "dimension reduction" and "grid distribution"), with D-template having a dimension of 28x27 and covering 753 descriptor features, and G-template having a size of 27x27 and covering 724 gap features.
[0058] Each RNA sequence entering the network is characterized by two RNA maps by mapping the abundance of encoded features to the corresponding positions of the two template maps.
[0059] B. Model framework
[0060] The model framework of the present method comprises a two-line CNN encoder and a classifier module.
[0061] The CNN encoder used in the present method adopts two lines of D-template and G-template as a whole, and divides the RNA map into multiple channels according to the number of different feature categories as input. Then, these channels are input into a convolutional layer containing a single convolutional kernel in a double-path manner, and information is extracted by convolutional blocks, one convolutional block including a max-pooling layer, three convolutional layers with kernel sizes of 1x1, 3x3 and 5x5, and a concatenation layer, with the convolutional block repeated twice to further extract embedding features, resulting in 192 matrices with a dimension of 10x10 on each path; then, a global max-pooling layer is applied on each path to convert the 192 matrices into a 192x1 vector, and the vectors of the two paths are concatenated to generate a new 384x1 vector; finally, the newly generated vector enters the classifier for RNA classification prediction.
[0062] C. Hyperparameter optimization
[0063] The present method uses grid search to adjust parameters to find suitable hyperparameter combinations (including learning rate and batch size). The hyperparameter optimization is based on the results of MCC, with the learning rate taking 10 -5 , 10 -4 , 2x10 -4 and 5x10 -4 , and the batch size taking 32, 64 and 96, respectively.
[0064] S3. Model evaluation
[0065] The application comprehensively considers six evaluation indexes of AUC (area under curve), Matthews correlation coefficient, specificity, sensitivity, precision, and accuracy.
[0066] AUC (area under curve): represents the area surrounded by the ROC curve on the coordinate axis, and the ROC curve is full name receiver operating characteristic curve, the abscissa is the false positive rate, and the ordinate is the true positive rate.
[0067] Matthews correlation coefficient (MCC): a comprehensive model performance evaluation index, which provides a more comprehensive perspective compared with other evaluation indexes. All categories (true positive, true negative, false positive, and false negative) in the confusion matrix are considered, and the value is-1 to 1, wherein 1 represents completely correct prediction, 0 represents random prediction, and-1 represents completely wrong prediction.
[0068]
[0069] Specificity: represents the proportion of samples actually negative that are judged as negative, and focuses on the ability of the model to correctly identify negative samples.
[0070]
[0071] Sensitivity: represents the proportion of samples actually positive that are judged as positive, and focuses on the ability of the model to correctly identify positive samples.
[0072]
[0073] Precision: represents the proportion of samples actually positive among samples predicted as positive.
[0074]
[0075] Accuracy: represents the proportion of correctly classified samples among all samples.
[0076]
[0077] where TP represents true positive (the number of correctly predicted positive samples), FP represents false positive (the number of incorrectly predicted negative samples as positive), TN represents true negative (the number of correctly predicted negative samples), and FN represents false negative (the number of incorrectly predicted positive samples as negative).
[0078] To evaluate the transferability of the method IntRNA on different classification tasks, the IntRNA and the corresponding classification methods CPPred, circDeep and nRC of the three data sets were used to predict on each data set, and the evaluation matrix was used to compare the model performance. To evaluate the transferability of the method IntRNA on different species, the IntRNA and six existing RNA coding potential prediction methods (CPPred, CPAT, CPC2, PLEK, RNAmining, CPE-SLD) were used to predict on the cross-species data of the RNA coding potential prediction data set, and the evaluation matrix was used to compare the model performance.
[0079] The evaluation results are as follows:
[0080] RNA coding potential prediction
[0081]
[0082] sncRNA sub-category prediction
[0083]
[0084] Linear and circular lncRNA prediction
[0085]
[0086] Cross-species RNA coding potential prediction
[0087]
[0088]
[0089] In addition, to evaluate the contribution of different modules of the method IntRNA, ablation experiments were performed on the human RNA coding potential prediction data for the G-feature, D-feature and RNAImage modules, and a comprehensive index MCC was used for comparison.
[0090] The evaluation results are as follows:
[0091]
[0092] The above merely provides the best mode of the present application, and the protection scope of the present application is not limited thereto. Any changes or modifications made by those skilled in the art within the technical scope of the present patent should be covered within the protection scope of the present patent application. Therefore, the protection scope of the present patent application should be subject to the protection scope of the claims.
Claims
1. A deep learning prediction method for comprehensive RNA classification based on graph representation, characterized in that, Includes the following steps: S1: Select and process RNA data, construct a graph representation of the RNA dataset and three datasets for RNA key classification prediction tasks; S2: The RNAImage graph representation strategy is adopted. Two template graphs are constructed by graph representation of the RNA dataset. The two template graphs are D-template and G-template. Then, each RNA sequence in the three datasets used for RNA key classification prediction tasks is used to obtain two corresponding representation graphs through feature mapping. All the obtained representation graphs constitute the encoded dataset. S3: A dual-path, multi-channel convolutional neural network model is used, and the RNA classification training model is obtained by training the encoded dataset. S4: Use the evaluation matrix to evaluate the RNA classification training model and optimize the hyperparameters to generate an RNA classification prediction model; S5: Use RNA classification prediction models to classify and predict the RNAs that need to be predicted for each category, and use the classification results for RNA function research and biopharmaceutical development.
2. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 1, characterized in that, In step S1, the three datasets used for the RNA key classification prediction task are a dataset for predicting RNA coding potential, a dataset for predicting sncRNA subclassification, and a dataset for predicting linear and circular lncRNAs.
3. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 1, characterized in that, In step S2, the RNAImage graph representation strategy is used to construct a template graph from the RNA dataset, specifically including: The RNA dataset is represented by a template graph after passing through four sequential processes: matrix generation, similarity calculation, dimensionality reduction, and grid distribution.
4. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 1, characterized in that, In step S2, the abundance of each coding feature of each RNA sequence in the three datasets used for RNA key classification prediction tasks is used to perform feature mapping on two template maps to obtain two corresponding representation maps.
5. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 1, characterized in that, In step S3, the dual-path multi-channel convolutional neural network model uses convolutional kernels for feature extraction.
6. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 1, characterized in that, In step S4, the evaluation matrix includes: AUC assessment, Matthews correlation coefficient assessment, specificity assessment, sensitivity assessment, accuracy assessment, and precision assessment.
7. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 1, characterized in that, In step S4, hyperparameter optimization is performed, specifically including: A grid search approach was used to optimize the parameters and find the optimal hyperparameters.
8. The deep learning prediction method for comprehensive RNA classification based on graph representation according to claim 7, characterized in that, In step S4, the hyperparameters include the learning rate and the batch size.