Circular RNA IRES prediction method

Through the DeepCIP computing framework combined with deep learning methods, RNA sequences and structural features are extracted, and multiple models are integrated for soft voting, solving the problem of insufficient prediction performance of circular RNA IRES in the existing technology, and achieving more efficient circular RNA IRES. Identify.

WO2023197718A9PCT designated stage expired Publication Date: 2024-11-28AOMING (HANGZHOU) BIOMEDICAL CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2023/072236
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2022-09-28
Filing Date
2023-01-16
Publication Date
2024-11-28

AI Technical Summary

Technical Problem

The existing IRES prediction methods mainly target linear mRNAs, and rely on hand-made features. They lack a deep learning framework and are difficult to effectively identify the features of circular RNA IRES, resulting in insufficient model performance.

Method used

Using the DeepCIP computing framework, feature embedding is combined with deep learning methods, RNA sequence and structural features are extracted through S-LSTM and GCN modules, and multiple models are integrated through soft voting strategies, and DeepIRESc tool is developed to predict circular RNA IRES.

Benefits of technology

The prediction performance of circular RNA IRES is significantly improved, which is better than traditional methods, especially when identifying circular RNA IRES, demonstrating the potential of deep learning in IRES prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2023072236_28112024_PF_FP_ABST
    Figure CN2023072236_28112024_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed is a method for realizing prediction of circular RNA IRESs by combining feature embedding and deep learning methods based on a DeepCPI computing framework. In a prediction process, the method uses an S-LSTM to learn a global sentence-level node to represent a whole RNA sequence, thereby modeling context information for a classification task. In addition, secondary structures of the RNA are simulated by means of a weighted RNA graph, and needed structural features are extracted on the basis of a GCN. Also, the method utilizes down-sampling and model integration to solve problems caused by imbalanced data. Experimental verification results show that compared with various existing deep learning methods, the method disclosed by the present invention has the best performance with respect to extracting features in sequences and structures, and can also effectively predict human and viral circular RNA IRESs of different lengths. Therefore, the method provided by the invention may be applied to research on the coding ability of circular RNAs and subsequent circular RNA design and application, such as development of circular RNA vaccines.
Need to check novelty before this filing date? Find Prior Art

Description

A method for predicting circular RNA IRES Technical Field

[0001] [Corrected 16.06.2023 according to Rule 91] The invention relates to the field of computer-aided biomolecular design, and in particular to a method for predicting circular RNA IRES by combining feature embedding with deep learning methods based on the DeepCIP computational framework. Background Art

[0002] [Corrected 16.06.2023 according to Rule 91] Using deep sequencing and computational analysis, researchers have discovered that circular RNAs (circRNAs) are a large class of RNAs with diverse functions, found not only in plants and animals but also in viruses (Nature 2013, 495:333–338; Front. Plant Sci. 2019, 10:379; Nat. Commun. 2019; 10:2300). Although circRNAs are generally considered a class of non-coding RNAs, with a covalently bound circular structure and no 5' cap or 3' poly A tail, increasing evidence suggests that circRNAs have protein-coding functions (JNCI J. Natl. Cancer Inst. 2018, 110:304–315; Oncogene 2018, 37:1805–1814; Mol. Cell 2017, 66:22-37.e9). Proteins encoded by circular RNAs may be prevalent in a variety of diseases, particularly in human cancers, where they play anti-tumor or tumor-promoting roles (Mol. Cancer 2021, 20:119; Mol. Cancer 2020, 19:71; Mol. Cancer 2021; 20:66). This discovery suggests the potential for the development and clinical application of circular RNA translation. Compared to linear mRNA, circular RNAs have improved thermal stability, longer efficacy, and more specific tissue expression, making them an ideal platform for developing next-generation mRNA drugs (Mol. Ther.-Nucleic Acids 2021, 25:355–371; J. Controlled Release 2022, 348:84–94). Recently, a research team developed a novel coronavirus circular RNA vaccine encoding the receptor-binding domain (RBD) of the spike protein of SARS-CoV-2 and its variants (Cell 2022, 185:1728-1744.e16). In addition, circular RNAs also have wide applications in preventing or treating infectious diseases and autoimmune diseases.

[0003] [Corrected 16.06.2023 according to Rule 91] Translation of circular RNAs requires internal ribosome entry sites (IRESs). IRESs are cis-acting RNA regulatory elements that recruit small ribosomal subunits to the translation start site in the absence of a 5' cap (Nat. Commun. 2018, 9:4475; Mol. Cell 2017, 66:9-21.e7). IRESs were first discovered in RNA viral genomes, such as those of poliovirus (PV) and encephalomyocarditis virus (EMCV) (Nature 1988, 334:320–325; J. Virol. 1988; 62:2636–2643). Using bicistronic analysis, IRESs have been widely discovered in viral and cellular mRNAs (J. Biol. Chem. 2005, 280:23425–23428). Compared to cap-dependent translation, the mechanisms of IRES-mediated translation are relatively unknown. However, researchers believe that primary sequence and RNA structure play important roles in IRES activity, either through direct recruitment of the structured RNA to the ribosome or through indirect interactions with ribosomal components assisted by canonical initiation factors and additional IRES trans-acting factors (ITAFs) (Oncogene 2004, 23:3200–3207; Front. Oncol. 2015, 5). Furthermore, compared to linear RNA IRESs, circular RNA IRESs have higher GC content, lower minimum free energy (MFE), and are generally more structurally sound (Mol. Cell 2021; 81:4300-4318.e13). These distinct properties contribute to distinct IRES activities in linear and circular RNA systems.

[0004] [Corrected 16.06.2023 according to Rule 91] It is estimated that approximately 10% of mRNAs may initiate translation through an IRES-mediated, cap-independent mechanism. Several studies have reported that IRESs play important roles in various cellular processes and diseases (Oncogene 2004, 23:3200–3207; Front. Oncol. 2015, 5). However, traditional experimental methods for identifying IRES elements are time-consuming and laborious, resulting in a very limited number of identified IRESs, hindering the study of common features and functions of IRESs. Data-driven computational methods, such as machine learning, are increasingly being applied to biological data. From a bioinformatics perspective, IRES prediction is difficult because there are no universally conserved features across all IRESs. However, several prediction tools have been developed. Kolekar et al. (Sci. Rep. 2016, 6:27436) developed a support vector machine (SVM)-based predictor, called IRESpred, that uses 35 features to predict viral and cellular IRESs. However, the negative samples in the training dataset used in IRESpred are not experimentally verified IRES-negative sequences. In 2016, Weingarten Gabbay et al. (Science 2016;351:aad4939) designed a high-throughput bicistronic analysis and identified thousands of new human and viral IRES sequences. The substantial increase in the number of new available IRES sequences has enabled machine learning methods to be better applied to IRES identification. Gritsenko et al. (Comput. Biol. 2017;13:e1005734) developed IRESpredictor, a stochastic gradient boosted random forest regression model based on the high-throughput analysis dataset of Weingarten Gabbay et al., to predict IRES activity using 6120 global and local sequence k-mer features. However, the large number of features may lead to problems with model overfitting and slow training time. Subsequently, Zhao et al. (J. Genet. Genomics 2018;45:403–406) proposed the IRESfinder model, which uses 19 carefully selected k-mer sequence features to predict cellular IRESs using only the human IRESs from the Weingarten Gabbay et al. dataset. Wang et al. (BMC Bioinformatics 2019;20:409) developed an XGBoost model called IRESpy, a state-of-the-art IRES prediction tool with improved performance and shorter training time, based on 340 global k-mer sequence features from Weingarten Gabbay et al. Studies in IRESpy also demonstrated that models based on mixed sequence / structure features slightly improved performance over sequence-based models. Technical issues

[0005] [Corrected 16.06.2023 according to Regulation 91] However, currently reported studies are limited to the use of handcrafted features, which are rarely optimal, and the role of structure in IRES prediction requires further exploration. Compared to traditional machine learning algorithms, deep neural network frameworks are still in their exploratory stages but may hold great promise for IRES prediction. Furthermore, current IRES prediction methods are designed for linear mRNAs. Given the differences in IRES activity between linear and circular RNAs, there is an urgent need to develop circular RNA IRES prediction methods. Summary of the Invention

[0006] [Corrected 16.06.2023 according to Rule 91] To address deficiencies in the prior art, the present invention provides a method for predicting circular RNA IRESs based on the DeepCIP computational framework by combining feature embedding with deep learning methods. This method achieves the prediction of circular RNA IRESs for the first time through sequence processing, RNA graph construction, feature extraction, and model integration. In the integration module, the DeepIRESc tool, developed using the methods described herein, employs a soft voting strategy to integrate three fusion models trained on different datasets. Each fusion model comprises four modules: an RNA processing module, an S-LSTM module, a GCN module, and a feature fusion module. First, the RNA processing module preprocesses the input RNA sequence for sequence encoding, structure prediction, and RNA graph construction. Second, the S-LSTM module and GCN module extract features of the RNA sequence and RNA secondary structure, respectively. Finally, the feature fusion module fuses the features extracted from the RNA sequence and structure. The fused features are passed through a fully connected layer and a sigmoid function, ultimately outputting a probability indicating the likelihood that the input RNA sequence is a circular RNA IRES. Specifically, sequence processing includes four steps: high-confidence data screening, independent test data set construction, three training subset construction, and RNA sequence representation method selection.

[0007] [Corrected 16.06.2023 according to Rule 91] To obtain sequence information, the present invention employs a sentence-state LSTM (S-LSTM) for the first time to automatically learn sequence features. Furthermore, a graph convolutional network (GCN) is used to construct a weighted RNA graph and extract secondary structure information from the RNA graph. A feature fusion module integrates sequence and structural features to identify circular RNA IRESs. To demonstrate the effectiveness of the DeepIRESc tool developed using the present invention's method, the present invention benchmarked DeepIRESc against an XGBoost model on a defined independent test set. The benchmark results demonstrate that DeepIRESc outperforms other compared methods in predicting circular RNA IRESs.

[0008] [Corrected 16.06.2023 according to Rule 91] The present invention is implemented as follows.

[0009] [Corrected 16.06.2023 according to Rule 91] In a first aspect, the present invention provides sequence processing steps. The corresponding steps are as follows: first, screening high-confidence data for model training; second, dividing the dataset screened in the first step into a training dataset and a test dataset; third, splitting the training and test datasets in the second step into three training subsets by combining downsampling and model ensemble methods.

[0010] [Corrected 16.06.2023 according to rule 91] Thanks to Weingarten-Gabbay et al. for identifying thousands of sequences with IRES activity from 55,000 oligonucleotides using high-throughput bicistronic analysis<Do Zotero Refresh:[@weingarten-gabbaySystematic2016]> . On this basis, Chen et al. identified 17,201 eGFP(+) oligonucleotides and 23,654 eGFP(-) oligonucleotides by constructing oligo-split-eGFP circular RNA reporters. Among them, the IRES of 1,639 oligonucleotides had linear-specific activity (i.e., linear IRES), while the IRES of 4,582 oligonucleotides had circular-specific activity (i.e., circular IRES). Since high-quality datasets are crucial to the predictive performance of the model, in the present invention, the first step in sequence processing in the method for predicting circular RNA IRES is to screen out high-confidence data for training the model.

[0011] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention selected cyclic IRESs from the study by Weingarten-Gabbay et al. as positive samples, and excluded linear sequences and sequences with IRES activity greater than background (excluding sequences with promoter activity greater than 0.2 and splicing activity less than -2.5) as negative samples. During the screening process, a total of 24,525 RNA sequences were obtained, including 4,582 positive samples and 19,943 negative samples. Since these data contain a portion of synthetic sequences used to test the effects of specific mutations on IRES activity, the present invention uses the method of Weingarten-Gabbay et al. to retain only the natural sequences labeled "CDS_screen", "Genome_Wide_Sceen_Elements", "High_Priority_Genes_Blocks", "Hight_Priurity_Virus_Block", "Human_5UTR_screen", "IRESite_blocks", "Viral_5UTR_Screen" and "rRNA_Matching_5UTRs" from the data set.

[0012] [Corrected 16.06.2023 according to Rule 91] Preferably, the present invention screened a dataset containing 4531 positive samples and 9616 negative samples. Furthermore, since the oligonucleotide library used for IRES activity assay was artificially constructed, the RNA sequence length in the present invention was selected to be 174 nt.

[0013] [Corrected 16.06.2023 according to Rule 91] Due to the lack of an independent test dataset for cyclic IRES, the present invention divides the dataset obtained above into a training dataset and a test dataset for model training and evaluation.

[0014] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention collected human circular RNA sequences from CircAtlas (Do Zotero Refresh: [@wuCircAtlas2020]) and viral circular RNA sequences from Viruscircbase. Blastn was then used to map the RNA sequences in the dataset to the circular RNAs in the database. As a result, 582 positive samples with 100% identity were obtained with the datasets obtained above.

[0015] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention randomly selected 582 negative samples and constructed an independent test set together with the previously described positive samples for subsequent evaluation of the model constructed by the present invention.

[0016] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention removes the samples used for the independent test set from the above-mentioned data set, and the remaining 3949 positive samples and 9034 negative samples will be used as training sets. In order to solve the problem that the model is biased towards the class with more samples due to the imbalance in the number of positive and negative samples, thereby reducing the generalization ability of the model, the present invention adopts a method combining downsampling and model integration, and randomly samples negative samples to obtain three subsets of negative samples with the same number of positive samples. The requirements of random sampling are that data is not wasted and that data is not repeated between subsets. Then, all negative subsets are combined with the positive data set to obtain three training subsets with 3949 positive samples and 3949 negative samples. Finally, the models trained using the three training subsets are integrated into a final model.

[0017] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention represents RNA sequences from two different perspectives: sequence information and structural information.

[0018] [Corrected 16.06.2023 according to Rule 91] Preferably, the present invention uses hot encoding to represent bases a, T / U, C, and G as [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1], respectively. Each sequence is represented as a feature matrix of dimension [L×4], where L represents the length of the sequence. In the dataset of the present invention, L=174.

[0019] [Corrected 16.06.2023 according to Rule 91] Preferably, the present invention uses RNApfold in the ViennaRNA package (version 2.5.1) to capture dynamic RNA secondary structure. Simultaneously, the McCaskill algorithm is used to calculate local stable RNA secondary structures and output RNA base pairing probabilities. The parameters used in the RNApfold run were W = 150, c = 1e-3, and structures with isolated pairs (--noLP) were not allowed. All other parameters were set to default.

[0020] [Corrected 16.06.2023 according to Rule 91] In a second aspect, the present invention uses the base pairing probability obtained by prediction in the first aspect to construct an RNA weighted graph G = (V, E, W).

[0021] [Corrected 16.06.2023 according to Rule 91] Further, each basis in the constructed RNA weighted graph is represented as a node V in the graph, where the features of each node are similarly encoded using a one-hot vector.

[0022] [Corrected 16.06.2023 according to Rule 91] Furthermore, the edges E in the constructed RNA weighted graph contain information on two different types of chemical bonds: covalent bonds connecting consecutive nucleotides along the RNA backbone and hydrogen bonds connecting paired bases, where the weight W of a covalent bond is defined as 1, and the weight W of a hydrogen bond is equal to the probability of pairing the two bases it connects.

[0023] [Corrected 16.06.2023 according to Rule 91] In a third aspect, the present invention provides a method for RNA sequence feature extraction.

[0024] [Corrected 16.06.2023 according to Rule 91] Further, the model structure used for RNA sequence feature extraction is the sentence-state LSTM (S-LSTM). S-LSTM is an alternative recurrent neural network designed to address the limitations of BiLSTM. BiLSTM is a variant of recurrent neural network, consisting of forward and backward LSTM, and is widely used in natural language processing. S-LSTM uses a gate mechanism similar to BiLSTM to control the flow of information. The difference is that S-LSTM regards each word as a word-level node w and adds a sentence-level node g to represent a complete sentence. Although S-LSTM acts on sequences, the way of node aggregation and message passing is similar to the operation of GNN, and it can be classified as a text GNN. That is, at each time step t, information can be exchanged between the sentence-level node and each word-level node, and each word node will also exchange information with its context node. The state of the S-LSTM at each time step t can be defined as Formula-1: in is the sub-state of each word, g t is a substate of the sentence.

[0025] [Corrected 16.06.2023 according to Article 91] As the time step t increases, each h i Capturing larger and larger n-grams while exchanging information with g, the contextual information learned by g becomes increasingly rich. Eventually, g can be used for task classification. By default, each word node only exchanges information with adjacent word nodes, and its window size is 1. However, increasing the window size allows for more information exchange. In the present invention, the window size is 3 and the time step is 7. The update process of the word state and sentence state in S-LSTM is given by the following formula-2 to formula-7: Formula-2 Formula-3 Formula-4 Formula-5 Formula-6 Formula-7 in, is the concatenated vector of context windows, Indicates different doors. and The control information comes from the input x i ,Left context cell, Right context cell and sentence context cells is the output gate. is the actual input. W, U, V and b are trainable parameters. σ is the sigmoid function.

[0026] [Corrected 16.06.2023 according to Rule 91] At the same time, the S-LSTM sentence state g in the model t The update process can be described by the following formula-8 to formula-14: Formula-8 Formula-9 Formula-10 Formula-11 Formula-12 Formula-13 Formula-14 in, and Standardized doors for controlling and o t is the output gate. W, U, and b are trainable parameters.

[0027] [Corrected 16.06.2023 according to Rule 91] Furthermore, each RNA secondary structure is represented using a weighted RNA graph.

[0028] [Corrected 16.06.2023 according to Rule 91] Preferably, the present invention learns the feature vectors of nucleotide nodes by stacking multiple GCN layers. In order to facilitate the classification of RNA sequences, the present invention aggregates the nucleotide node features in each RNA graph into a graph embedding. Global sum, maximum and average pooling are the most commonly used strategies for aggregating node features. That is, the node features can be represented by the matrix X∈R N×D Indicates that the connectivity between nodes can be expressed as an adjacency matrix A∈R with weights N×NRepresented as , where N is the number of nodes and D is the dimension of each node’s feature vector. The matrices X and A are used as model inputs, and the GCN layer-by-layer propagation rule in Kipf and Welling (Semi-Supervised Classification with Graph Convolutional Networks.2017) is given by Formula-15: in, and I N is the identity matrix. yes The diagonal matrix H (l) is the activation matrix, W (l) is the layer l th The trainable matrix is ​​σ(·), and σ(·) is a nonlinear activation function.

[0029] [Corrected 16.06.2023 according to Rule 91] Therefore, after several GCN layers, the matrix X is transformed into formula-16: (F is the number of filters), each node can be expressed as Equation-17.

[0030] [Corrected 16.06.2023 in accordance with Rule 91] Formula-16 Formula-17 Here, Θ represents the filter parameter matrix, where e j,i is the edge weight from node j to node i.

[0031] [Corrected 16.06.2023 according to Rule 91] We choose to learn node-level embeddings by stacking three GCN layers, because too few GCN layers will result in a small receptive field, while too many GCN layers may lead to over-smoothing. In addition, we add a global and pooling layer after the last GCN layer to extract graph-level representation for task classification, because such processing has relatively better performance.

[0032] [Corrected 16.06.2023 according to Rule 91] In order to combine the information sequence features and structural features in the S-LSTM and GCN modules, the present invention concatenates them in the last dimension. Then, a classifier with a sigmoid function is used to output the predicted probability.

[0033] [Corrected 16.06.2023 according to Rule 91] All training subsets are used for training separately to obtain three models. The present invention adopts a soft voting strategy to integrate the three models. The voting mechanism is the most typical model integration method, and its basic idea is to output the class with the most votes among all classifiers. The classifier directly gives the final prediction label called hard voting, and gives the predicted probability of the output label called soft voting. The present invention uses soft voting to calculate the weighted sum of the probabilities of the three models, and then determines the predicted label as 0 or 1 according to the default threshold of 0.5. Preferably, the present invention sets the weights of the three models to be equal.

[0034] [Corrected 16.06.2023 according to Rule 91] In the fourth aspect, the present invention uses nested cross validation (CV) for hyperparameter tuning and model selection. Errica et al. compared different GNN models using nested cross validation (CV). Simply put, nested CV contains outer CV and inner CV, where CV can choose k-fold or holdout techniques. K-fold CV means randomly dividing the dataset into k out Non-overlapping subsets, where each subset is used as the test set in turn, and the remaining subsets are used for training. Maintaining CV means directly dividing the dataset into two mutually exclusive subsets, one for training and the other for validation. In the present invention, the validation is performed using any of the three training subsets described above.

[0035] [Corrected 16.06.2023 in accordance with Article 91]

[0036] [Corrected 16.06.2023 according to Rule 91] Further, the present invention uses k-fold technique for external CV and k out = 5, using a 90% training split and a 10% validation split for the inner CV. That is, the present invention trains each outer training fold and uses 10% random data as a validation set. If performance does not improve after n epochs, execution is stopped early and then tested in the test fold. The final hyperparameter evaluation score is the average of the scores of all test folds.

[0037] [Corrected 16.06.2023 according to Rule 91] Furthermore, after completing the hyperparameter selection, the present invention uses 10-fold cross validation to retrain and validate the model on all data, and finally evaluates the model performance on an independent test set constructed by the present invention. In order to evaluate the effectiveness of the proposed model and its various submodules, the present invention compares it with other baseline methods using nested CV. The description of different baseline methods is as follows: TextCNN for sequences: TextCNN is a convolutional neural network for text classification, which consists of an embedding layer, a convolution layer, a max pooling layer, and a fully connected layer. The sequence is encoded as a one-hot vector in the embedding layer, and n filters = 64 and filters_size = [2, 3, 4] are set in the convolution layer.

[0038] [Corrected 16.06.2023 according to Rule 91] BiLSTM: BiLSTM is used to learn effective features from sequences and can be stacked in multiple layers. Single-layer, two-layer, and three-layer BiLSTMs are used for model comparison, with hidden_size set to 64.

[0039] [Corrected 16.06.2023 according to rule 91] TextCNN + Annotated Secondary Structure: Here, RNAfold (Mol. Biol. AMB 2011;6:26) is used to predict RNA secondary structure and output a dot-bracket sequence representing the base-pairing pattern. bpRNA (Nucleic Acids Res. 2018;46:5381–5394) annotates the predicted RNA secondary structure, parsing the base-pairing information output by RNAfold into a detailed structure and providing relevant contextual annotation information, including stem (S), hairpin loop (H), polyloop (M), internal loop (I), bulge (B), and end (E). This allows RNA secondary structure to be represented as a sequence and fed into a TextCNN model for training. The previous section describes the TextCNN setup.

[0040] [Corrected 16.06.2023 according to Rule 91] Graph Multilayer Perceptron (MLP): This method applies a three-layer MLP with ReLU activation on the node features of the RNA graph, and then uses global and pooling layers to learn graph-level embeddings. It differs from the GCN module in that it does not use the relationships between nodes when learning node features and is a graph topology-agnostic model.

[0041] [Corrected 16.06.2023 according to rule 91] XGBoost: This model was used to develop the IRESpy tool (BMC Bioinformatics 2019;20:409). The present invention uses 340 sequence k-mer (1-mer, 2-mer, 3-mer and 4-mer) features to train the XGBoost model and sets the same XGBoost hyperparameters as IRESpy. The scale_pos_weight parameter is provided in the XGBoost model and is useful for unbalanced classes by controlling the balance between positive and negative weights. The present invention uses the circular RNA IRES training dataset to train two XGBoost models, called XGBoost_weight for circ and XGBoost for cirk, and sets the scale_pos_weight parameter to 3 and scale_pos_weight to 1 respectively (the same as IRESpy). The present invention uses 10-fold CV and early stopping method to obtain the optimal num_boost_round parameter, and selects the best model for comparison with the final model.

[0042] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention adopts five evaluation indicators, including accuracy (Acc), sensitivity (Sn), specificity (Sp), precision and Matthews correlation coefficient (MCC). The indicator definitions are described as follows: Formula-18 to Formula-22: Formula-18 Formula-19 Formula-20 Formula-21 Formula-22 where TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives.

[0043] [Corrected 16.06.2023 according to Rule 91] Further, the present invention uses ROC (Receiver Operating Characteristic) curve and PR (Precision Recall) curve to intuitively evaluate the overall performance of the classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] [Corrected on 16.06.2023 according to Rule 91][Corrected on 16.06.2023 according to Rule 91] Figure 1 is a flow chart of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] [Corrected 16.06.2023 in accordance with Article 91] DETAILED DESCRIPTION

[0046] [Corrected 16.06.2023 in accordance with Article 91]

[0047] [Corrected 16.06.2023 in accordance with Article 91]

[0048] [Corrected 16.06.2023 in accordance with Article 91] Modes for Carrying Out the Invention

[0049] [Corrected 16.06.2023 according to Rule 91] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and do not limit the scope of protection of the present invention. Figure 1 is a flow chart of the present invention.

[0050] [Corrected 16.06.2023 according to Rule 91] This embodiment provides a multimodal deep learning method based on the DeepCIP computational framework that uses RNNs and GNNs to extract features from RNA sequences and structures. These two features are then combined for circular RNA-IRES prediction, comprising four stages: sequence processing, RNA graph construction, feature extraction, and model integration. Specifically, in the integration module, the DeepIRESc tool developed using the present invention employs a soft voting strategy to integrate three fusion models trained on different datasets. Each fusion model comprises four modules: an RNA processing module, an S-LSTM module, a GCN module, and a feature fusion module. First, the RNA processing module preprocesses the input RNA sequence for sequence encoding, structure prediction, and RNA graph construction. Second, the S-LSTM module and GCN module extract features from the RNA sequence and RNA secondary structure, respectively. Finally, in the feature fusion module, the features extracted from the RNA sequence and structure are fused. The fused features are passed through a fully connected layer and a sigmoid function, ultimately outputting a probability indicating the likelihood that the input RNA sequence is a circular RNA IRES. More specifically, sequence processing includes four steps: high-confidence data screening, independent test dataset construction, three training subset construction, and RNA sequence representation method selection.

[0051] [Corrected 16.06.2023 according to rule 91] For sequence processing: the first step is to filter out high-confidence data for training the model; first, the sequences with IRES activity identified by Weingarten-Gabbay et al.<Do Zotero Refresh:[@weingarten-gabbaySystematic2016]> , circular IRES was selected as the positive sample, and the eGFP(-) sample after excluding the linear sequence and the sequence with IRES activity greater than the background activity (excluding the sequence with promoter activity greater than 0.2 and splicing activity less than -2.5) was selected as the negative sample. As a result, 24,525 RNA sequences were obtained, including 4,582 positive samples and 19,943 negative samples. To remove synthetic sequences collected from the screened samples to test the effects of specific mutations on IRES activity, the dataset was further optimized using the method of Weingarten-Gabbay et al., retaining only natural sequences labeled "CDS_screen," "Genome_Wide_Sceen_Elements," "High_Priority_Genes_Blocks," "Hight_Priurity_Virus_Block," "Human_5UTR_screen," "IRESite_blocks," "Viral_5UTR_Screen," and "rRNA_Matching_5UTRs." As a result, a high-confidence dataset containing 4,531 positive samples and 9,616 negative samples was selected for model training. Given that the oligonucleotide library for IRES activity assays was artificially constructed, the RNA sequence length was selected to be 174 nt.

[0052] [Corrected 16.06.2023 according to Rule 91] Second, to construct an independent test dataset for circular IRESs, the dataset obtained in the first step was divided into a training dataset and a test dataset. First, human circular RNA sequences were collected from CircAtlas (Do Zotero Refresh: [@wuCircAtlas2020]) and viral circular RNA sequences were collected from Viruscircbase. Blastn was then used to map the RNA sequences in the dataset to the circular RNAs in the database. As a result, 582 positive samples with 100% identity to the dataset were obtained. Subsequently, 582 negative samples were randomly selected and used with the positive samples to construct an independent test set for subsequent evaluation of the constructed model. Furthermore, after removing the 582 positive and 582 negative samples used for the independent test set from the dataset obtained in the first step, the remaining 3949 positive and 9034 negative samples were used as the training set.

[0053] [Corrected 16.06.2023 according to Rule 91] In the third step, the training and test sets from the second step are split into three training subsets using a combination of downsampling and model ensembling. To address the problem of the imbalance in the number of positive and negative samples, which causes the model to favor the class with more samples and thus reduces the model's generalization ability, a method combining downsampling and model ensembling is used to randomly sample the negative samples to obtain three subsets with the same number of negative samples as the positive samples. The requirements for random sampling are that no data is wasted and that data is not repeated between subsets. Then, all negative subsets are combined with the positive dataset to obtain three training subsets with 3949 positive samples and 3949 negative samples.

[0054] [Corrected 16.06.2023 according to Rule 91] Step 4: Select a representation method for RNA sequences: First, use sequence information to represent RNA sequences. Specifically, use hot encoding to represent bases a, T / U, C, and G as [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1], respectively. Each sequence is represented as a feature matrix of dimension [L×4], where L represents the length of the sequence; in this case, L=174. Second, use structural information to represent RNA sequences. Specifically, RNApfold (Bioinformatics 2006, 22:614–615) in the ViennaRNA package (version 2.5.1) is used to capture dynamic RNA secondary structure. Simultaneously, the McCaskill algorithm (Biopolymers 1990; 29:1105–1119) is used to calculate the local stable secondary structure of RNA and output the probability of RNA base pairing. The parameters during the RNApfold run were set to W = 150, c = 1e-3, and structures with isolated pairs (--noLP) were not allowed to be generated. Other parameters were set by default.

[0055] [Corrected 16.06.2023 according to Rule 91] For RNA graph construction: the base pairing probability obtained in the fourth step of the sequence processing stage is used to construct an RNA weighted graph G = (V, E, W): each base in the constructed RNA weighted graph is represented as a node V in the graph, where the characteristics of each node are similarly encoded using a one-hot vector; the edge E in the constructed RNA weighted graph contains two different types of chemical bond information: covalent bonds connecting consecutive nucleotides along the RNA backbone and hydrogen bonds connecting paired bases, where the weight W of the covalent bond is defined as 1, and the weight W of the hydrogen bond is equal to the probability of pairing the two bases it connects.

[0056] [Corrected 16.06.2023 according to Rule 91] Regarding feature extraction: First, the model structure for RNA sequence feature extraction adopts sentence state LSTM (S-LSTM). The S-LSTM state at each time step t is defined as: Formula-1: in is the sub-state of each word, g t is a substate of the sentence.

[0057] [Corrected 16.06.2023 according to Article 91] As the time step t increases, each h i By capturing increasingly larger n-grams and exchanging information with g, the contextual information learned by g becomes increasingly rich. Ultimately, g can be used for task classification. By default, each word node exchanges information only with its neighbors, with a window size of 1. However, increasing the window size allows for more information exchange.

[0058] [Corrected 16.06.2023 according to Rule 91] The performance of sequence models is affected by many hyperparameters, such as the number of sentence-level nodes, window size, time step, and hidden layer size in the S-LSTM module. In the design of the experiments, the present invention focuses on the effects of window size and time step in the S-LSTM model, where the window size ranges from {1, 2, 3} and the time step ranges from {5, 7, 9}. By comparing and analyzing the experimental results of various S-LSTM settings, it is found that the best value of average AUC is obtained when the window size is set to 3 and the time step is set to 9. The performance of S-LSTM, TextCNN, and BiLSTM with different numbers of layers is compared, and it is found that S-LSTM outperforms other models in terms of average AUC value. The above experimental results show that S-LSTM can more effectively extract circular RNA IRES sequence features. In the present invention, the window size is 3 and the time step is 9.

[0059] [Corrected 16.06.2023 according to Rule 91] Meanwhile, the updating process of word state and sentence state in S-LSTM is given by the following formula-2 to formula-7: Formula-2 Formula-3 Formula-4 Formula-5 Formula-6 Formula-7 in, is the concatenated vector of context windows, Indicates different doors. and The control information comes from the input x i ,Left context cell, Right context cell and sentence context cells is the output gate. is the actual input. W, U, V and b are trainable parameters. σ is the sigmoid function.

[0060] [Corrected 16.06.2023 according to Rule 91] At the same time, the S-LSTM sentence state g in the model t The update process can be described by the following formula-8 to formula-14: Formula-8 Formula-9 Formula-10 Formula-11 Formula-12 Formula-13 Formula-14 in, and Standardized doors for controlling and o t is the output gate. W, U, and b are trainable parameters.

[0061] [Corrected 16.06.2023 according to Rule 91] Secondly, each RNA secondary structure is represented by a weighted RNA graph. The feature vector of the nucleotide node is learned by stacking multiple GCN layers. In order to facilitate the classification of RNA sequences, the nucleotide node features in each RNA graph are aggregated into the graph embedding. The node features are represented by the matrix X∈R N×D The connectivity between nodes is represented by the adjacency matrix A∈R with weights N×N Represented as , where N is the number of nodes and D is the dimension of each node’s feature vector. The matrices X and A are used as model inputs, and the GCN layer-by-layer propagation rule in Kipf and Welling (Semi-Supervised Classification with Graph Convolutional Networks.2017) is given by Formula-15: in, and I N is the identity matrix. yes The diagonal matrix H (l) is the activation matrix, W (l) is the layer l th The trainable matrix is ​​σ(·), and σ(·) is a nonlinear activation function.

[0062] [Corrected 16.06.2023 according to Rule 91] Therefore, after several GCN layers, the matrix X is transformed into formula-16: (F is the number of filters), each node can be expressed as Equation-17.

[0063] [Corrected 16.06.2023 in accordance with Rule 91] Formula-16 Formula-17 Here, Θ represents the filter parameter matrix, where e j,i is the edge weight from node j to node i.

[0064] [Corrected 16.06.2023 according to Rule 91] The number of GCN layers and readout functions are crucial for extracting RNA features from graphs. Here, we construct a GCN module by varying the number of GCN layers and selecting different readout functions during the experimental design process to investigate the impact of different configurations on model performance. The number of GCN layers ranges from {1, 2, 3}, and the readout strategies are selected from {global sum pooling, global max pooling, global average pooling}. Experimental results show that the three GCN layers and the global sum pooling layer achieve the best performance in terms of AUC scores.

[0065] [Corrected 16.06.2023 according to Rule 91] Therefore, the present invention subsequently chooses to learn node-level embeddings by stacking three GCN layers, because too few GCN layers will result in a small receptive field, while too many GCN layers may lead to over-smoothing. In addition, a global and pooling layer is added after the last GCN layer to extract graph-level representation for task classification, because such processing has relatively better performance.

[0066] [Corrected 16.06.2023 according to Rule 91] To investigate whether using a weighted RNA graph in the model of the present invention can improve the representation of circular RNA-IRES structural features, this example compares the performance of a weighted RNA graph + GCN, TextCNN + annotated secondary structure (TextCNN_nannotate_struct), and graph MLP. The comparison results show that the GCN module achieves the best performance. All of the above shows that the weighted RNA graph can better represent the secondary structure of RNA than the sequential annotated structure (among which GCN performs better than TextCNN_nannotate_struct). At the same time, using different chemical bond relationships between nucleotides is a better choice, and the results show that GCN performs better than MLP in terms of graphs.

[0067] [Corrected 16.06.2023 according to Rule 91] In addition, in order to combine the information sequence features and structural features in the S-LSTM and GCN modules, the present invention connects them in the last dimension. Then, a classifier with a sigmoid function is used to output the predicted probability.

[0068] [Corrected 16.06.2023 according to Rule 91] For model ensemble: First, and k out = 5, using a 90% training split and a 10% validation split for the internal CV. Specifically, the present invention trains each outer training fold and uses 10% random data as a validation set. If performance does not improve after n epochs, execution is stopped early and then tested on the test fold. The final hyperparameter evaluation score is the average of the scores across all test folds. After completing hyperparameter selection, the model is retrained and validated on all data using 10-fold cross-validation, and finally, model performance is evaluated on the constructed independent test set.

[0069] [Corrected 16.06.2023 according to Rule 91][Corrected 16.06.2023 according to Rule 91] To evaluate the effectiveness of the proposed model and its various submodules, we compared it with other baseline methods using nested CV. The description of the different baseline methods is as follows: [0069.1] [Corrected 16.06.2023 according to Rule 91] [Corrected 16.06.2023 according to Rule 91] [Corrected 16.06.2023 according to Rule 91] Sequence TextCNN: TextCNN is a convolutional neural network for text classification, which consists of an embedding layer, a convolution layer, a max pooling layer, and a fully connected layer. The sequence is encoded as a one-hot vector in the embedding layer, and n filters = 64 and filters_size = [2, 3, 4] are set in the convolution layer. [0069.1.1] [Corrected 16.06.2023 according to Rule 91] BiLSTM: BiLSTM is used to learn effective features from sequences and can be stacked in multiple layers. Single-layer, double-layer, and triple-layer BiLSTMs are used for model comparison, with hidden_size set to 64.

[0070] [Corrected 16.06.2023 according to rule 91] TextCNN + Annotated Secondary Structure: Here, RNAfold (Mol. Biol. AMB 2011;6:26) is used to predict RNA secondary structure and output a dot-bracket sequence representing the base-pairing pattern. bpRNA (Nucleic Acids Res. 2018;46:5381–5394) annotates the predicted RNA secondary structure, parsing the base-pairing information output by RNAfold into a detailed structure and providing relevant contextual annotation information, including stem (S), hairpin loop (H), polyloop (M), internal loop (I), bulge (B), and end (E). This allows RNA secondary structure to be represented as a sequence and fed into a TextCNN model for training. The TextCNN setup has been described previously.

[0071] [Corrected 16.06.2023 according to Rule 91] Graph Multilayer Perceptron (MLP): This method applies a three-layer MLP with ReLU activation on the node features of the RNA graph, and then uses global and pooling layers to learn graph-level embeddings. It differs from the GCN module in that it does not use the relationships between nodes when learning node features and is a graph topology-agnostic model.

[0072] [Corrected 16.06.2023 according to rule 91] XGBoost: This model was used to develop the IRESpy tool (BMC Bioinformatics 2019;20:409). The present invention uses 340 sequence k-mer (1-mer, 2-mer, 3-mer and 4-mer) features to train the XGBoost model and sets the same XGBoost hyperparameters as IRESpy. The scale_pos_weight parameter is provided in the XGBoost model and is useful for unbalanced classes by controlling the balance between positive and negative weights. The present invention uses the circular RNA IRES training dataset to train two XGBoost models, called XGBoost_weight for circ and XGBoost for cirk, and sets the scale_pos_weight parameter to 3 and scale_pos_weight to 1 respectively (the same as IRESpy). The present invention uses 10-fold CV and early stopping method to obtain the optimal num_boost_round parameter, and selects the best model for comparison with the final model.

[0073] [Corrected 16.06.2023 according to Rule 91] Specifically, after model training is complete, the model parameters are fixed, and the mean squared error between the predicted and true values ​​is used as the optimal target. This is repeated five times, with a different validation set used each time. During training, if the current model's accuracy or mean squared error on the validation set is better than the previous model, the previous model is saved and replaced. This process is repeated 100 times.

[0074] [Corrected 16.06.2023 according to Rule 91] After training the model, the obtained model is tested, specifically using a test set. The area under the receiver operating characteristic curve (AUC) is the primary basis for selecting model parameters. Using the test set, the accuracy and precision of the XGBOOST model reported in the literature were compared, and it was found that the results of the model constructed by the present invention outperformed existing methods.

[0075] [Corrected 16.06.2023 according to Rule 91] Five evaluation indicators are used, including accuracy (Acc), sensitivity (Sn), specificity (Sp), precision and Matthews correlation coefficient (MCC). The indicator definitions are described in the following formula-18 to formula-22: Formula-18 Formula-19 Formula-20 Formula-21 Formula-22 where TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives.

[0076] [Corrected 16.06.2023 according to Rule 91] The ROC (Receiver Operating Characteristic) curve and PR (Precision Recall) curve are used to intuitively evaluate the overall performance of the classification model; all training subsets are used for training separately to obtain three models. The present invention adopts a soft voting strategy to integrate these three models. The voting mechanism is the most typical model integration method, and its basic idea is to output the class with the most votes among all classifiers. The classifier directly gives the final prediction label called hard voting, and gives the predicted probability of the output label called soft voting. The present invention uses soft voting to calculate the weighted sum of the probabilities of the three models, and then determines the predicted label as 0 or 1 according to the default threshold of 0.5. Preferably, the present invention sets the weights of the three models to be equal.

[0077] [Corrected 16.06.2023 according to Rule 91] Furthermore, this example demonstrates that feature fusion and a soft voting mechanism can improve the prediction performance of the DeepIRESc tool developed using the methods of the present invention for circular RNA IRESs. To examine the importance of RNA structure in circular RNA IRES identification, the fusion model was compared with a model based solely on sequence features on an independent test set. The results showed improved performance in AUROC, AUPRC, ACC, SN, Precision, and MCC. This also demonstrates the importance of structural features in circular RNA IRES prediction.

[0078] [Corrected 16.06.2023 according to Rule 91] Subsequently, this embodiment tested an independent test dataset to verify whether the soft voting mechanism could improve prediction performance. In this embodiment, the soft voting method was applied to three ensemble models trained with three training subsets. Comparison of the prediction performance of the three individual models with the ensemble model showed that the ensemble model achieved the best performance, with an AUROC of 0.723 and an AUPRC of 0.697. Due to the small number of models used in the ensemble and the partial overlap between the training subsets, the improvement in prediction performance of the ensemble model was not particularly significant.

[0079] [Corrected 16.06.2023 according to Rule 91] Further, in order to evaluate the predictive performance of the model of the present invention, in this example, an independent test data set was first used to experimentally compare the tool developed using the method of the present invention and the XGBoost model. The comparison results showed that the tool developed using the method of the present invention had higher performance than the XGBoost_weight model, with an AUC of 0.723 (vs. 0.624), an ACC of 0.646 (vs. 0.624), an SP of 0.512 (vs. 0.503), an SN of 0.780 (vs. 0.744), an accuracy of 0.615 (vs. 0.600), and an MCC of 0.303 (vs. 0.255). This is attributed to the fact that the model of the present invention better represents the RNA structure. In addition, compared with the manually drawn k-mer features, the features automatically extracted by deep learning have better representation capabilities. The above further indicates that deep learning may be a beneficial choice to solve the current situation where the common features of IRES are not yet fully clarified.

[0080] [Corrected 16.06.2023 according to Rule 91] To further determine the effectiveness of the DeepIRESc tool developed using the methods of the present invention for predicting circular RNA IRESs, this example explored the correlation between predicted circular RNA IRES probabilities and experimental circular RNA IRES activity. Circular RNA IRES activity was defined by eGFP expression in the Chen et al. experiment. In an independent test set, circular RNA IRES activity ranged from 0 to 6.9327, with background eGFP expression of 3.466387. The results showed that circular RNA IRES with higher activity generally had higher predicted probabilities.

[0081] [Corrected 16.06.2023 according to Rule 91] Furthermore, the present invention visualizes the prediction probabilities of the DeepIRESc tool developed using the present invention for human and viral sequences from an independent test set. Generally speaking, the prediction probabilities for human and viral circular RNA IRESs are higher than those for non-circular RNA IRESs. This further demonstrates the superior ability of the DeepIRESc tool developed using the present invention to predict human and viral circular RNA IRESs.

[0082] [Corrected 16.06.2023 according to Rule 91] For prediction applications: We collected experimentally validated circular RNA IRES sequences of varying lengths from the literature. Positive samples formed a ground truth dataset containing 10 IRES sequences from Homo sapiens circular RNAs and 4 IRES sequences from Drosophila melanogaster circular RNAs. Details of the ground truth data are shown in Table 1.

[0083] [Corrected 16.06.2023 in accordance with Article 91] [Corrected 16.06.2023 in accordance with Article 91] Table 1 [0083.1][According to detailed rules 91, corrected 16.06.2023][According to detailed rules 91, corrected 16.06.2023][According to detailed rules 91, corrected 16.06.2023]Meanwhile, the DeepIRESc tool prediction result that utilizes the inventive method to develop is compared with other models (comprising XGBoost_weight, XGBoost and IRESpy).The comparison of prediction result is shown in Table 2.For the model of the present invention, if the prediction threshold is set to 0.5, then 6 out of 10 RNA sequences of Homo sapiens are predicted to be circular RNA IRES, and 1 out of 4 RNA sequences of Drosophila is predicted to be circular RNA-IRES, which is better than the model performance of other methods.Wherein, the poor performance of the DeepIRESc tool that utilizes the inventive method to develop Drosophila melanogaster RNA sequence may be due to the presence of differences in the circular RNA-IRES features of different species.Because the training data set adopted by the present invention only comprises the RNA sequence of humans and viruses. [0083.1.1] [Corrected 16.06.2023 according to Rule 91]

[0084] [Corrected 16.06.2023 in accordance with Article 91]

[0085] Table 2 [XXXX] , [XXXX] In addition, the performance of different methods for predicting the probability distribution of human circular IRESs was compared using kernel density estimation (KDE). The results showed that the DeepIRESc tool developed using the methods of the present invention outperformed XGBoost in prediction performance. Furthermore, the prediction performance of IRESpy (for linear RNA IRES identification) was lower than that of other methods for circular RNA IRES prediction using actual circular RNA IRES data. This further demonstrates the significance of the present invention in developing an efficient method for predicting circular RNA IRESs.

[0086] It can be seen that the method provided by the present invention, combined with the trained model, can quickly and accurately predict circular RNA IRES sequences with translation-mediating function.

[0087] [Corrected 16.06.2023 in accordance with Rule 91] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention. Industrial Applicability

[0088] [Corrected 16.06.2023 in accordance with Article 91] Sequence Listing Free Content

[0089] [Corrected 16.06.2023 in accordance with Article 91]

Claims

1. A method for predicting circular RNA IRES. It is characterized in that The method predicts circular RNA IRES based on the sequence and structural features of the extracted RNA in combination with four stages of sequence processing, RNA graph construction, feature extraction and model integration. The sequence and structural features of the RNA are extracted based on the DeepCIP computing framework combined with RNN and GNN multi-mode deep learning methods. The tool for extracting the sequence and structural features of the RNA is the DeepIRESc tool. The DeepIRESc tool adopts a soft voting strategy to integrate three fusion models trained by different data sets. The fusion model includes an RNA processing module, an S-LSTM module, a GCN module and a feature fusion module. The RNA processing module is used to preprocess the input RNA sequence, the S-LSTM module is used to extract RNA sequence features, the GCN module is used to extract RNA secondary structure features, the feature fusion module is used to fuse the extracted RNA sequence and RNA secondary structure features, and the fusion outputs a probability through a fully connected layer and a sigmoid function, and the probability is the possibility that the input RNA sequence is a circular RNA IRES.

2. A method for predicting circular RNA IRES according to claim 1, It is characterized in that The sequence processing stage in the method for predicting circular RNA IRES includes the following steps: the first step is to screen out high-confidence data for training the model; the second step is to divide the data set screened in the first step into a training data set and a test data set; the third step is to split the training and test sets in the second step into three training subsets by combining downsampling and model integration.

3. A method for predicting circular RNA IRES according to claim 1, It is characterized in that The sequence processing in the method for predicting circular RNA IRES includes a method for representing RNA sequences using sequence information and structural information; the method for representing the sequence information is: using hot coding to represent bases a, T / U, C and G as [1,0,0,0], [0,1,0,0], [0,0,1,0] and [0,0,0,1] respectively, and each sequence is represented as a feature matrix with a dimension of [L×4], where L represents the length of the sequence; the method for representing the structural information is: using RNA folding to capture dynamic RNA secondary structures, and outputting the probability of RNA base pairing by calculating the local stable secondary structure of RNA.

4. A method for predicting circular RNA IRES according to claim 1, It is characterized in that The sentence state of the model structure in the feature extraction stage of the circular RNA IRES prediction method is S-LSTM, the node aggregation and message passing mode of the model structure is text GNN; the feature vector of the nucleotide node of the model structure is learned by stacking multiple GCN layers.

5. A method for predicting circular RNA IRES according to claim 1, It is characterized in that The model integration stage in the method for predicting circular RNA IRES adopts a nested CV method for hyperparameter adjustment and model selection. The hyperparameter adjustment includes using the k-fold technique for external CV, using a 90% training split and a 10% validation process for internal CV, and the model selection includes using 10-fold cross validation to retrain on all data, validate the model, and evaluate the model performance on an independent test set.