RNA comprehensive classification deep learning prediction method based on graph representation

Abstract

The invention discloses an RNA (Ribonucleic Acid) comprehensive classification deep learning prediction method based on graph representation, which comprises the following steps of: adopting an RNAImage graph representation strategy, constructing a template graph through a graph representation RNA data set, performing feature mapping on each RNA sequence in a data set for an RNA key classification prediction task by utilizing the template graph to obtain a representation graph, all the obtained representation graphs form a data set after coding processing; a double-line multi-channel convolutional neural network model is adopted, and through data set training after coding processing, an RNA classification training model is obtained; evaluating the RNA classification training model by using the evaluation matrix, and performing hyper-parameter optimization to generate an RNA classification prediction model; the RNA is classified and predicted, and the classification result is used for RNA function exploration and biological medicine development. The method provided by the invention is helpful to comprehensively predict RNA classification and further excavate the biological interpretability of the RNA coding potential.

Description

Technical Field

[0001] The present invention relates to the field of RNA comprehensive classification deep learning prediction, and in particular to a RNA comprehensive classification deep learning prediction method based on graph representation. Background Art

[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 functions of most mRNAs have been widely studied. For a long time in the past, researchers believed that ncRNA was a "junk" fragment produced during transcription. In recent years, studies have found that ncRNA accounts for more than 90% of the coding RNA in the human genome, and 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 short non-coding RNA (sncRNA). The main problem of lncRNA classification is how to predict circular RNA (circRNA) and linear RNA, while sncRNA classification is a more complex classification problem.

[0003] Since existing tools are generally built for a certain type of RNA classification problem, their RNA characterization methods are often too specific and difficult to transfer to the classification prediction of the entire RNA system; in addition, because the existing RNA features cannot capture long-range nucleotide interactions, the prediction of RNA coding potential lacks biological interpretability. The existence of the above problems makes it impossible for existing tools to comprehensively predict RNA classification systems and hinders the auxiliary role of computer prediction in mining biological principles.

[0004] Therefore, providing a universal RNA characterization and prediction method will help to comprehensively predict RNA classification and further explore the biological interpretability of RNA coding potential. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a deep learning prediction method for comprehensive RNA classification based on graph representation to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a method for RNA comprehensive classification deep learning prediction based on graph representation, comprising the following steps:

[0007] S1: Select and process RNA data, construct a graph representation RNA dataset and three datasets for RNA key classification prediction tasks;

[0008] S2: Using the RNAImage graph representation strategy, the template graph is constructed by graph representation of RNA datasets. Then, each RNA sequence in the three datasets used for RNA key classification prediction tasks is feature mapped using the template graph to obtain a representation graph. All the obtained representation graphs constitute the encoded dataset.

[0009] S3: A dual-route multi-channel convolutional neural network model is used to train the encoded data set to obtain an RNA classification training model;

[0010] S4: Use the evaluation matrix to evaluate the RNA classification training model and perform hyperparameter optimization to generate an RNA classification prediction model;

[0011] S5: Use the RNA classification prediction model to classify and predict the various types of RNA that need to be predicted, and use the classification results for RNA function exploration and development of biological drugs.

[0012] In step S1, three datasets used for RNA key classification prediction tasks are a dataset for RNA coding potential prediction, a dataset for sncRNA subclassification prediction, and a dataset for linear and circular lncRNA prediction.

[0013] In step S2, the RNAImage graph representation strategy is adopted to construct a template graph by graph representation of the RNA dataset, specifically including:

[0014] The template image is generated after the RNA dataset is sequentially processed through four processes in series: matrix generation, similarity calculation, dimensionality reduction, and grid distribution. The template image is generated after four processes in series (matrix generation, similarity calculation, dimensionality reduction, and grid distribution), and then the abundance of each encoding feature of each RNA sequence to be included in the model is mapped to the corresponding position of the template image, thus completing the image-like representation of the RNA sequence.

[0015] In step S2, the abundance of each coding feature of each RNA sequence in the three data sets for RNA key classification prediction tasks is feature mapped using the template graph to obtain a representation graph.

[0016] In step S3, the model uses multi-path multi-channel convolution, the number of paths depends on the number of template images, and the number of channels depends on the number of feature categories. The dual-path multi-channel convolutional neural network model uses convolution kernels for feature extraction. The model uses three convolution kernels of different sizes (1×1, 3×3, and 5×5) to extract features simultaneously.

[0017] In step S4, the evaluation matrix includes:

[0018] AUC evaluation, Matthews correlation coefficient evaluation, specificity evaluation, sensitivity evaluation, precision evaluation, and accuracy evaluation. The process of evaluating the performance of the RNA classification prediction model includes: evaluating the RNA classification prediction model based on six evaluation criteria: AUC, Matthews 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 perform parameter optimization and find the optimal hyperparameters. The hyperparameters include learning rate and batch size. The present invention uses a grid search method to perform parameter optimization and find a suitable hyperparameter combination (including learning rate and batch size).

[0021] Further preferably, a method and system for RNA deep learning classification prediction based on graph representation comprises:

[0022] Select and process RNA data, construct a graph representation RNA dataset and a dataset for three key classification prediction tasks (RNA coding potential prediction, sncRNA subclassification prediction, and linear and circular lncRNA prediction);

[0023] The RNAImage graph representation strategy is adopted to construct two template graphs by graph representation of RNA dataset. Then, each RNA sequence in the classification prediction task dataset is subjected to feature mapping to obtain two corresponding representation graphs, thereby obtaining the encoded dataset.

[0024] A dual-route multi-channel convolutional neural networks (CNN) model is used to train the deep neural network model by processing the data set, thereby obtaining an RNA classification training model;

[0025] The RNA classification training model was evaluated using the evaluation matrix, and hyperparameters were optimized to generate an RNA classification prediction model.

[0026] The RNA splitting prediction model is used to classify and predict various types of RNA, and further combined with examples to explore the biological interpretability of the coding potential.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] In order to reduce the specificity of RNA characterization methods and improve biological interpretability, the present invention proposes a deep learning prediction method for comprehensive RNA classification based on graph representation, which expands the RNA feature space and adopts a universal representation method. Compared with existing RNA classification methods, this method shows the advantage of transferability and continues to achieve the optimal level in multiple RNA classification tasks; in addition, the G-feature proposed by this method can capture long-range nucleotide interactions and can be used to explore the biological interpretability of RNA coding potential. The results of the present invention can be subsequently applied to RNA function research and RNA-related drug development.

[0029] The RNA comprehensive classification deep learning prediction method based on graph representation of the present invention helps to comprehensively predict RNA classification and further explore the biological interpretability of RNA coding potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present specification is further described in the form of exemplary embodiments and is further explained in detail with the help of the accompanying drawings.

[0031] Figure 1 It is a schematic diagram of the method for generating the G-feature according to the embodiment of this specification.

[0032] Figure 2 It is a method flow chart of the RNAImage image characterization strategy shown in the embodiments of this specification.

[0033] Figure 3 It is a flow chart of the RNA deep learning classification prediction method based on graph representation shown in the embodiments of this specification. DETAILED DESCRIPTION

[0034] It should be noted that the model training and evaluation processes adopted by the prediction method in each embodiment are basically the same, and the main difference lies in the research problems and different data set constructions targeted by 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 attached flowchart can be executed in a computer system, but in certain circumstances, the steps shown can be executed in a different order or by removing some steps.

[0036] Embodiment 1

[0037] like Figure 1 As shown, in this embodiment, a graph-based RNA characterization method RNAImage is constructed, including:

[0038] S1. RNA signature generation:

[0039] This method uses two types of RNA features: one is the gap feature based on nucleotide spacing (G-feature), and the other is the feature based on descriptors (D-feature). G-feature emphasizes the distribution of nucleotides in RNA sequences (especially the relationship between nucleotide pairs), and D-feature reflects the intrinsic sequence properties, structural properties, and physical and chemical properties of RNA.

[0040] G-feature is the RNA feature proposed by this method, such as Figure 1 As shown, a method for generating G-feature is provided by taking GC and CG nucleotide pairs as an example, and G-feature is calculated by counting the frequencies of nucleotide pairs in different intervals (from 0 to 40). For example, 2-GC represents the frequency of GC nucleotide pairs separated by two nucleotides in the RNA sequence.

[0041] In addition, the present invention generated a total of 753 D-features to encode RNA sequences, of which 314 are previously reported classic coding features (such as open reading frames, codon-related features, and secondary structure features). The remaining 439 features were newly generated in this study based on specific physicochemical properties or structural properties (such as hydrogen bond alkalinity, topological polar surface area, and atomic ratio path length). These 439 D-features were calculated using the Composition, Transition and Distribution (CTD) strategy.

[0042] S2. RNAImage image characterization strategy:

[0043] This method proposes an RNAImage strategy to achieve image-like representation of RNA. Figure 2 As shown, two widely used RNA databases (ENCODE and RefSeq) were first used to obtain a total of 788,576 RNA sequences. After four serial processes (matrix generation, similarity calculation, dimensionality reduction, and grid distribution), two template images (D-template and G-template) were generated, and the template images were 28×27 (D-template) and 27×27 (G-template) in size, respectively.

[0044] The specific steps of generating template graphs (D-template and G-template) are as follows: First, all 788,576 RNAs were converted into two different feature matrices, D-feature was 753×788,576, and G-feature was 724×788,576. Then, by calculating the cosine similarity between the features, the above feature matrices were converted into similarity matrices with dimensions of 753×753 and 724×724, respectively. Subsequently, UMAP was used to reduce the dimensions of the features based on the similarity matrix. Finally, the JV algorithm was used to calculate the minimum cost distance, and the reduced dimension graph was grid-distributed to obtain two template graphs.

[0045] Finally, the abundance of each encoding feature of each RNA sequence to be included in the model is mapped to the corresponding positions of the two template graphs, so that each RNA sequence is converted into two RNA graphs, thereby completing the image-like representation of the RNA sequence.

[0046] Embodiment 2

[0047] like Figure 3 As shown, this embodiment provides an RNA deep learning classification prediction method IntRNA based on graph representation, including:

[0048] S1. Dataset construction:

[0049] The first step in constructing the data set of this embodiment is to screen out high-quality available data, and then divide and use the data. The data used in this embodiment can be divided into two categories, one is the graph representation RNA data used for model input, and the other is the RNA data used for model training and evaluation.

[0050] The graph representation RNA data used for model input was collected from the ENCODE and Refseq databases. After deduplication and other sorting, a total of 788,576 RNA data were used for template graph generation in graph representation.

[0051] The RNA data used for model training and evaluation mainly include three datasets: RNA coding potential prediction (from CPPred in the study of Xiaoxue Tong et al.), sncRNA subclassification prediction (from nRC in the study of Antonino-Fiannaca et al.), and linear and circular lncRNA prediction (from circDeep in the study of Mohamed-Chaabane et al.).

[0052] The RNA coding potential prediction dataset contains RNA sequences from five species, including: 74,321 RNA sequences from humans (including 41,917 mRNA sequences and 32,404 ncRNA sequences), 51,032 RNA sequences from mice (including 31,102 mRNA sequences and 19,930 ncRNA sequences), 26,256 RNA sequences from zebrafish (including 15,594 mRNA sequences and 10,662 ncRNA sequences), 26,256 RNA sequences from fruit flies (including 17,400 mRNA sequences and 4,098 ncRNA sequences), and 7,126 RNA sequences from yeast (including 6,713 mRNA sequences and 413 ncRNA sequences). Human RNA sequences were divided into training and test sets in a ratio of 4:1, and data from other species were used as test sets.

[0053] The sncRNA subclassification prediction dataset contains a total of 8,920 human sncRNA sequences of 13 subclasses ("miRNA", "5S rRNA", "5.8S rRNA", "ribozymes", "CD-box", "HACA-box", "scaRNA", "tRNA", "Intron gpI", "Intron gpII", "IRES", "leader", "riboswitch"), of which 700 sequences were obtained for each subclass of RNA, while only 520 sequences were obtained for "IRES" due to its limited number. When dividing the dataset, the training set contains 400 sncRNAs of each subclass (only 256 for "IRES"), the validation set contains 100 sncRNAs of each subclass (only 64 for "IRES"), and the test set contains 200 sncRNAs of each subclass.

[0054] The linear and circular lncRNA prediction dataset contains 51,625 human lncRNA sequences (including 19,683 linear lncRNA sequences and 31,942 circular lncRNA sequences), which are divided into 38,722 lncRNA sequences as training sets, 5161 lncRNA sequences as validation sets, and 7742 lncRNA sequences as test sets.

[0055] S2. Model training

[0056] A. Model Input

[0057] The input of the present invention is a sequence, and each sequence is further converted into two RNA characterization maps by mapping the abundance of different features to the corresponding positions of two template maps (D-template and G-template), and the abundance value of each feature is normalized to be between 0 and 1. Both D-template and G-template are generated by four serial steps ("matrix generation", "similarity calculation", "dimensionality reduction" and "grid distribution"), where the dimension of D-template is 28×27, covering 753 descriptor features; the scale of G-template is 27×27, covering 724 gap features.

[0058] Each RNA sequence that enters the network is represented as two corresponding RNA graphs by mapping the abundance of the encoded features to the corresponding positions of the two template graphs.

[0059] B. Model Framework

[0060] The model framework of this method includes two modules: a dual-route CNN encoder and a classifier.

[0061] The CNN encoder used in this method adopts two lines, D-template and G-template, as a whole. The RNA graph is divided into multiple channels as input according to the number of different feature categories. Then, these channels are input into the convolution layer containing a single convolution kernel in a dual-path manner, and convolution blocks are used to extract information. A convolution block includes a maximum pooling layer, a convolution layer with three convolution kernels of different sizes (1×1, 3×3 and 5×5), and a concatenation layer. The convolution block is repeated twice to further extract embedded features, and finally 192 matrices of 10×10 dimensions are obtained on each path; then, a global maximum pooling layer is applied on each path to convert the 192 matrices into a 192×1 vector, and the vectors of the two paths are concatenated to generate a new vector of 384×1; finally, the newly generated vector enters the classifier for RNA classification prediction.

[0062] C. Hyperparameter Optimization

[0063] This method uses grid search to adjust parameters to find the appropriate hyperparameter combination (including learning rate and batch size). Hyperparameter optimization is based on the results of MCC, where the learning rate is 10 -5 , 10 -4 , 2×10 -4 and 5×10 -4 , the batch sizes were 32, 64, and 96 respectively.

[0064] S3. Model Evaluation

[0065] The present invention comprehensively considers six evaluation indicators: AUC (area under curve), Matthews correlation coefficient, specificity, sensitivity, precision, and accuracy.

[0066] AUC (area under curve): represents the area enclosed by the ROC curve on the coordinate axis. The full name of the ROC curve is the receiver operating characteristic curve (receiver operating characteristic curve). The horizontal axis is the false positive rate and the vertical axis is the true positive rate.

[0067] Matthews correlation coefficient (MCC): A comprehensive model performance evaluation metric that provides a more comprehensive perspective than other evaluation metrics. It takes into account the number of all categories (true positives, true negatives, false positives, false negatives) in the confusion matrix and takes a value from -1 to 1, where 1 represents a completely correct prediction, 0 represents a random prediction, and -1 represents a completely wrong prediction.

[0068]

[0069] Specificity: It indicates the proportion of samples that are actually negative that are judged to be negative, focusing on the ability of the model to correctly identify negative samples.

[0070]

[0071] Sensitivity: It indicates the proportion of samples that are actually positive that are judged to be positive, focusing on the ability of the model to correctly identify positive samples.

[0072]

[0073] Precision: It indicates the proportion of samples predicted to be positive that are actually positive.

[0074]

[0075] Accuracy: It indicates the proportion of correctly classified samples among all samples.

[0076]

[0077] Where TP represents true positives (the number of positive samples predicted correctly), FP represents false positives (the number of negative samples wrongly predicted as positive), TN represents true negatives (the number of negative samples predicted correctly), and FN represents false negatives (the number of positive samples wrongly predicted as negative).

[0078] To evaluate the transferability of IntRNA in this method on different classification tasks, IntRNA and the classification methods CPPred, circDeep and nRC corresponding to the three datasets were predicted on each dataset, and the evaluation matrix was used to compare the model performance. To evaluate the transferability of IntRNA in this method on different species, IntRNA and six existing RNA coding potential prediction methods (CPPred, CPAT, CPC2, PLEK, RNAmining, CPE-SLD) were predicted on the cross-species data of the RNA coding potential prediction dataset, and the evaluation matrix was used to compare the model performance.

[0079] The evaluation results are shown in the following table:

[0080] Prediction of RNA coding potential

[0081]

[0082] sncRNA subclassification prediction

[0083]

[0084] Linear and circular lncRNA prediction

[0085]

[0086] Prediction of RNA coding potential across species

[0087]

[0088]

[0089] In addition, in order to evaluate the contribution of different IntRNA modules of this method, ablation experiments were carried out on the three modules of G-feature, D-feature, and RNAImage on the human RNA coding potential prediction data, and the comprehensive indicator MCC was adopted for comparison.

[0090] The evaluation results are shown in the following table:

[0091]

[0092] The above is only the best specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or modifications that can be easily thought of by any technician familiar with the technical field within the technical scope of this patent should be included in the protection scope of this patent application. Therefore, the protection scope of this patent application should be based on the protection scope of the claims.

Claims

1. A deep learning prediction method for RNA comprehensive classification based on graph representation, characterized in that: The following steps are involved: S1: Select and process RNA data, construct a graph representation RNA dataset and three datasets for RNA key classification prediction tasks; S2: Using the RNAImage graph representation strategy, the template graph is constructed by graph representation of RNA datasets. Then, each RNA sequence in the three datasets used for RNA key classification prediction tasks is feature mapped using the template graph to obtain a representation graph. All the obtained representation graphs constitute the encoded dataset. S3: A dual-route multi-channel convolutional neural network model is used to train the encoded data set to obtain an RNA classification training model; S4: Use the evaluation matrix to evaluate the RNA classification training model and perform hyperparameter optimization to generate an RNA classification prediction model; S5: Use the RNA classification prediction model to classify and predict the various types of RNA that need to be predicted, and use the classification results for RNA function exploration and development of biological drugs.

2. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 1 is characterized in that: In step S1, three datasets used for RNA key classification prediction tasks are a dataset for RNA coding potential prediction, a dataset for sncRNA subclassification prediction, and a dataset for linear and circular lncRNA prediction.

3. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 1 is characterized in that: In step S2, the RNAImage graph representation strategy is adopted to construct a template graph by graph representation of the RNA dataset, specifically including: Graph representation RNA dataset generates a template graph after four serial processes of matrix generation, similarity calculation, dimensionality reduction and grid distribution.

4. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 1 is characterized in that: In step S2, the abundance of each coding feature of each RNA sequence in the three data sets for RNA key classification prediction tasks is feature mapped using the template graph to obtain a representation graph.

5. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 1 is characterized in that: In step S3, the dual-path multi-channel convolutional neural network model adopts convolution kernel for feature extraction.

6. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 1, characterized in that: In step S4, the evaluation matrix includes: AUC evaluation, Matthews correlation coefficient evaluation, specificity evaluation, sensitivity evaluation, precision evaluation, and accuracy evaluation.

7. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 1 is characterized in that: In step S4, hyperparameter optimization is performed, specifically including: The grid search method is used to optimize parameters and find the optimal hyperparameters.

8. The RNA comprehensive classification deep learning prediction method based on graph representation according to claim 7 is characterized in that: In step S4, the hyperparameters include learning rate and batch size.

Images