Automatic identification method of circular RNA with biological interpretability

Abstract

The invention discloses a biological interpretability circular RNA automatic identification method, and belongs to the technical field of bioinformatics and computers. According to the method, sequence data is obtained from a database and preprocessed, reverse complement matching and ALU repeated element features are extracted, after the features are screened through an MRMD2.0 algorithm, a depth model combining a convolutional neural network and a bidirectional long-short-term memory network is constructed for classification training, feature contribution is analyzed through an SHAP method, and finally the model is applied to multiple scenes. According to the method, circRNA cyclization key features can be accurately captured, high-order feature representation is automatically learned to improve the classification accuracy, meanwhile, a feature biological mechanism is disclosed, multi-task migration such as circRNA and PCG distinguishing and stem cell expression prediction is achieved, and the defects of an existing algorithm in the aspects of feature representation and interpretability are overcome.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and computer technology, and in particular to an automated method for identifying circular RNA with biological interpretability. Background Technology

[0002] Circular RNAs (circRNAs), a class of closed-loop RNAs formed by backsplicing, have overturned classical gene expression theories and are increasingly attracting attention in non-coding RNA research. With the development of RNA-seq technology, numerous algorithms for identifying circRNAs have emerged. In early research, biological experimental methods were gradually replaced by computational methods due to their time-consuming and laborious nature. In recent years, machine learning algorithms have been widely used in circRNA identification: Pan et al. extracted sequence features and used random forests to build the WebCircRNA server; PredcircRNA used a multi-kernel learning model to classify circRNAs and lncRNAs; H-ELM used extreme learning machines for feature selection; and circDeep used a deep learning framework to learn multimodal feature representations. These various algorithms, through different feature extraction and model building strategies, have driven the development of circRNA identification technology.

[0003] However, most current circRNA identification algorithms rely directly on manually designed features during feature extraction, failing to fully consider key factors influencing circRNA circularization (such as inverse complement matching structures and ALU repeat element pairing effects). This results in insufficient feature representation of the circularization mechanism, making it difficult to accurately capture the essential sequence features of circRNAs. Furthermore, existing deep learning models typically operate as "black boxes," lacking interpretable analysis of feature importance and failing to reveal the biological relationship between sequence features and circRNA identification results. Simultaneously, the models have limited generalization ability in extended application scenarios (such as distinguishing circRNAs from protein-coding genes and predicting expression in stem cells), making it difficult to meet diverse bioinformatics analysis needs. Summary of the Invention

[0004] The purpose of this invention is to overcome one or more shortcomings of the prior art and provide a biologically interpretable method for automatic identification of circular RNA.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A bio-interpretable automated method for identifying circular RNA includes the following sequential steps:

[0007] S1. Data preparation: Human circRNA and lncRNA sequence data were obtained from circAtlas and GENCODE. Sequences shorter than 200bp and repetitive or similar sequences were removed. The sequences were extracted into FASTA format and the coordinate system and orientation were standardized. Positive and negative sample sets were constructed based on the processed sequences.

[0008] S2. Feature Extraction: For the positive and negative sample sets constructed in S1, extract the inverse complement matching features and ALU repeating elements to characterize the circularization potential of the sequence.

[0009] S3. Feature Selection: The MRMD2.0 algorithm is used to select the most discriminative feature subset from the high-dimensional features extracted in S2, providing optimized feature input for subsequent model training;

[0010] S4. Model Building: Based on the feature subset selected in S3, a deep neural network model combining convolutional neural network and bidirectional long short-term memory network is constructed. The network structure automatically learns multi-dimensional feature representations and performs classification training.

[0011] S5. Model Evaluation: Using the model trained in S4, evaluate its performance through cross-validation and independent test sets to verify the model's ability to recognize circRNAs;

[0012] S6. Interpretability Analysis: For the model evaluated in S5, the SHAP method is used to analyze the contribution of each feature to the model prediction, revealing the biological association between the features and the prediction results;

[0013] S7. Extended Applications: Apply the model validated in S6 to distinguish between circRNA and PCG and to predict the expression of circRNA in stem cells, thus expanding the practical application scenarios of the model.

[0014] Furthermore, in the S1 data preparation, reliable circRNAs are selected from circAtlas2.0 as positive samples, and lncRNAs with the same length as the positive samples are selected from GENCODE as negative samples. Sequences with a length of less than 200bp and lncRNAs that overlap with the positive samples are deleted. Training and test sets are constructed while ensuring a balance between the number of positive and negative samples.

[0015] Furthermore, in the S2 feature extraction, the method for extracting the inverse complement matching feature is as follows: for the sequence processed by S1, a matching matrix is ​​constructed. When the bases at position i and position j are complementary pairs, the value is 1; otherwise, it is 0. The possible inverse complementary segments are calculated by traversing the sequence, and their length, number and distribution are statistically analyzed to form a numerical feature describing the potential of the ring structure.

[0016] Furthermore, in the S2 feature extraction, the method for extracting ALU repeating elements is as follows: considering the effect of pairing and looping of ALU elements in different directions, a composite scoring function is introduced:

[0017] ;

[0018] in , , , These represent the different combinations of forward and reverse Alu pairings in the upstream and downstream introns, respectively, and the scoring results are used as feature values.

[0019] Furthermore, in the S3 feature selection, the MRMD2.0 algorithm, which is an improvement on the PageRank algorithm, is used to rank the importance of the features extracted in S2 and obtain the optimal feature subset.

[0020] Furthermore, in the construction of the S4 model, the model receives the feature subset filtered by S3 as input, first extracts key patterns through convolution and pooling operations, then learns time series characteristics through a multi-layer convolutional neural network, and finally performs classification prediction through a fully connected layer. The output layer uses the Sigmoid activation function, and the training process uses the cross-entropy loss function and the Adam optimizer, with the learning rate set to 0.001.

[0021] Furthermore, in the S5 model evaluation, ten-fold cross-validation and independent test sets are used to train and test the model constructed in S4 to verify metrics such as accuracy, sensitivity, and AUC.

[0022] Furthermore, in the S6 interpretability analysis, the marginal contribution of the i-th feature is measured by the Shapley value to determine the sequence features that play a key role in the S5 classification result.

[0023] Furthermore, in the S7 extended application, the model is used to distinguish between circRNA and PCG, and classification is achieved based on the feature importance analyzed in S6.

[0024] Furthermore, in the S7 extended application, the model is used to predict whether circRNA is expressed in stem cells, and the network model constructed in S4 is used to analyze and predict the features.

[0025] The beneficial effects of this invention are:

[0026] (1) By integrating the inverse complement matching feature and the feature extraction technology of ALU repeat elements, and combining the MRMD2.0 algorithm for feature screening, the effect of accurately capturing the key sequence features of circRNA circularization and reducing the data dimensionality is achieved.

[0027] (2) By constructing a deep model that combines a convolutional neural network and a bidirectional long short-term memory network, the system can automatically learn high-order feature representations of sequences and improve the classification accuracy of circRNAs.

[0028] (3) By introducing the SHAP interpretability analysis method and expanding the application scenarios of the model, the effect of revealing the contribution mechanism of characteristic biology and realizing the multi-task transfer such as distinguishing circRNA from PCG and predicting stem cell expression is achieved. Attached Figure Description

[0029] Figure 1 This is a flowchart of a biologically interpretable automated method for identifying circular RNA.

[0030] Figure 2 This is a diagram of a deep learning model. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1

[0033] See Figure 1 This paper provides an automated method for identifying circular RNA with biological interpretability, comprising the following sequential steps:

[0034] S1. Data preparation: Human circRNA and lncRNA sequence data were obtained from circAtlas and GENCODE. Sequences shorter than 200bp and repetitive or similar sequences were removed. The sequences were extracted into FASTA format and the coordinate system and orientation were standardized. Positive and negative sample sets were constructed based on the processed sequences.

[0035] S2. Feature Extraction: For the positive and negative sample sets constructed in S1, extract the inverse complement matching features and ALU repeating elements to characterize the circularization potential of the sequence.

[0036] S3. Feature Selection: The MRMD2.0 algorithm is used to select the most discriminative feature subset from the high-dimensional features extracted in S2, providing optimized feature input for subsequent model training;

[0037] S4. Model Building: Based on the feature subset selected in S3, a deep neural network model combining convolutional neural network and bidirectional long short-term memory network is constructed. The network structure automatically learns multi-dimensional feature representations and performs classification training.

[0038] S5. Model Evaluation: Using the model trained in S4, evaluate its performance through cross-validation and independent test sets to verify the model's ability to recognize circRNAs;

[0039] S6. Interpretability Analysis: For the model evaluated in S5, the SHAP method is used to analyze the contribution of each feature to the model prediction, revealing the biological association between the features and the prediction results;

[0040] S7. Extended Applications: Apply the model validated in S6 to distinguish between circRNA and PCG and to predict the expression of circRNA in stem cells, thus expanding the practical application scenarios of the model.

[0041] In the S1 data preparation, reliable circRNAs were selected from circAtlas2.0 as positive samples, and lncRNAs with the same length as the positive samples were selected from GENCODE as negative samples. Sequences with a length of less than 200bp and lncRNAs that overlapped with the positive samples were deleted. The training set and test set were constructed while ensuring a balance between the number of positive and negative samples.

[0042] In S2 feature extraction, the method for extracting the inverse complement matching feature is as follows: For the sequence processed by S1, a matching matrix is ​​constructed. When the bases at position i and position j are complementary pairs, the value is 1; otherwise, it is 0. The possible inverse complementary segments are calculated by traversing the sequence, and their length, number and distribution are statistically analyzed to form a numerical feature describing the potential of the ring structure.

[0043] In S2 feature extraction, the method for extracting ALU repeating elements is as follows: considering the effect of pairing and looping of ALU elements in different directions, a composite scoring function is introduced:

[0044] ;

[0045] in , , , These represent the different combinations of forward and reverse Alu pairings in the upstream and downstream introns, respectively, and the scoring results are used as feature values.

[0046] In feature selection S3, the MRMD2.0 algorithm, which is an improvement on the PageRank algorithm, is used to rank the importance of the features extracted in S2 and obtain the optimal feature subset.

[0047] In the S4 model construction, the model receives the feature subset filtered by S3 as input, first extracts key patterns through convolution and pooling operations, then learns time series characteristics through a multi-layer convolutional neural network, and finally performs classification prediction through a fully connected layer. The output layer uses the Sigmoid activation function, and the training process uses the cross-entropy loss function and the Adam optimizer, with the learning rate set to 0.001.

[0048] In the evaluation of the S5 model, ten-fold cross-validation and independent test sets were used to train and test the model built by S4 to verify metrics such as accuracy, sensitivity, and AUC.

[0049] In the S6 interpretability analysis, the Shapley value is used to measure the marginal contribution of the i-th feature and to determine the sequence features that play a key role in the S5 classification results.

[0050] In the S7 extended application, the model is used to distinguish between circRNA and PCG, and classification is achieved based on the feature importance analyzed in S6.

[0051] In the S7 extended application, the model is used to predict whether circRNA is expressed in stem cells, and the network model built in S4 is used to analyze and predict the features.

[0052] Example 2

[0053] Data preparation and preprocessing implementation details:

[0054] (I) Data Acquisition and Screening Process: Experimentally validated human circular RNA sequences were selected as positive samples from the circAtlas2.0 database. This database is validated through both high-throughput sequencing and biological experiments to ensure sequence reliability. Simultaneously, long non-coding RNAs were extracted from the GENCODEv42 database as negative samples. During extraction, the length distribution range of the positive samples was strictly matched to avoid length deviations interfering with subsequent classification. From the initially acquired 18,762 negative sample sequences, the following screening steps were used to improve data quality: First, short sequences less than 200 bp were batch filtered using a Python script. This threshold was set based on the minimum structural requirements for circular RNA formation. Second, sequence alignment was performed using the BLASTN tool to remove negative samples with an overlap of more than 50 bp with the positive samples, avoiding sequence homology interference between the positive and negative sample sets.

[0055] (II) Coordinate System and Orientation Standardization: The SAMtools tool was used to map the sequences to the human reference genome GRCh38, and the chromosome coordinate offset problem was corrected by comparison results. For sequences mapped to the negative strand, a reverse complementation transformation operation was performed to ensure that they were stored in the forward strand orientation. In practice, the sort function in the BedTools suite was used to sort the coordinates, and the merge command was used to process overlapping regions to ensure that each sequence was uniquely positioned on the genome and had a consistent orientation. This standardization step solved the problem of inconsistent coordinate systems from different data sources and provided a unified benchmark for subsequent feature extraction.

[0056] (III) Sample Set Construction Strategy: Stratified sampling was used to construct the training and test sets. First, positive and negative samples were divided in an 8:2 ratio, and then random undersampling was used to balance the number of samples in each class. Specifically, the StratifiedShuffleSplit module of the scikit-learn library was used to ensure that the training and test sets had consistent distributions in basic features such as sequence length and GC content. To verify the model's generalization ability, an additional 1500 circular RNAs and 1500 long non-coding RNAs not used in training were extracted from the GSE152641 dataset in the GEO database to form an independent test set. This dataset came from different experimental platforms and could effectively evaluate the model's applicability across data scenarios.

[0057] Specific implementation of feature extraction and selection:

[0058] (I) Inverse Complement Matching Feature Extraction: A complement pair matrix is ​​constructed for each sequence. When two base positions in the sequence form an AT (or AU) or CG pair, the corresponding position in the matrix is ​​marked as 1; otherwise, it is marked as 0. The matrix is ​​traversed by a sliding window with a window size of 20 bp and a step size of 5 bp to identify inverse complementary regions that meet the following conditions: the proportion of complementary pairs within the region is not less than 70%, and the length is not less than 15 bp. During the statistical process, the length, number, and distribution position of each region in the sequence are recorded to generate a feature vector containing twelve indicators, such as the total number of complementary regions, the length of the longest region, and the distribution ratio of regions at the 5' end / 3' end / middle. This feature extraction method is based on the inverse splicing mechanism of circular RNA formation and characterizes the circularization probability by quantifying the complementation potential within the sequence.

[0059] (II) ALU Repeat Element Feature Analysis: The RepeatMasker tool was used to annotate the 5kb intron regions upstream and downstream of the sequence. This tool, based on the human repetitive sequence database, can accurately identify ALU elements and label their orientation. The identified ALU elements were divided into two categories: forward (consistent with the transcription direction) and reverse (opposite to the transcription direction). The ALU pairing relationships in the upstream and downstream introns were further divided into four groups: upstream forward and downstream forward, upstream forward and downstream reverse, upstream reverse and downstream forward, and upstream reverse and downstream reverse. A composite scoring function was used to quantify the impact of different pairing combinations on circularization. The function's design logic is as follows: forward pairing combinations (such as upstream forward and downstream forward) promote circularization, while reverse pairing combinations (such as upstream forward and downstream reverse) inhibit circularization. A weighted comprehensive score was obtained, which was used as the feature input model.

[0060] (III) MRMD2.0 Feature Selection Algorithm: This feature selection algorithm is an improvement on the webpage ranking concept. First, a feature similarity graph is constructed, using mutual information between features as edge weights. Mutual information is calculated using the `mutual_info_classif` function from Python's scikit-learn library. Feature scores are iteratively updated, with a damping coefficient set to 0.85 to simulate the probability of page jumps during browsing. Finally, the top 20% of features are selected in descending order of scores to form the optimal feature subset. This step effectively reduces feature dimensionality, removes redundant information, and improves model training efficiency. Experimental results show that the selected 8-dimensional feature subset improves classification accuracy by approximately 7.3% compared to the original 32-dimensional feature set.

[0061] Deep Neural Network Model Construction and Training:

[0062] (I) Model Architecture Design: A CNN-BLSTM hybrid network structure is constructed, with the input layer receiving an 8-dimensional normalized feature vector. The convolutional layer module contains two convolutional operations: the first layer uses 16 3×1 convolutional kernels, coupled with the ReLU activation function and a batch normalization layer, followed by a max pooling layer; the second layer uses 32 3×1 convolutional kernels, followed by an average pooling layer. This design can effectively extract local patterns from the features. The bidirectional long short-term memory layer consists of forward and backward LSTMs, each with 128 neurons. Overfitting is prevented by a dropout rate of 0.2. The bidirectional outputs are concatenated to form a 256-dimensional feature vector, capturing the temporal dependencies of sequence features. The fully connected layer contains two layers, with 128 dimensions and 64 dimensions respectively. The output layer uses the Sigmoid activation function, and the output probability value between 0 and 1 represents the probability of circular RNA.

[0063] See Figure 2This deep learning model takes a subset of features as input, first extracts local features through convolutional layers and batch normalization and pooling operations, then captures the temporal dependencies of features through bidirectional long short-term memory layers, and then further learns through flattening and two fully connected layers. Finally, the classification probability of circular RNA is output by the sigmoid activation function. Its architecture reflects the hybrid advantages of convolutional neural networks extracting local patterns and bidirectional long short-term memory networks modeling temporal features.

[0064] (II) Model Training Parameter Configuration: Binary cross-entropy is used as the loss function, which is suitable for binary classification problems and can effectively measure the difference between the predicted probability and the true label. The Adam algorithm is selected as the optimizer, and the initial learning rate is set to 0.001. This parameter is determined through grid search to balance convergence speed and accuracy. The batch size is set to 32 to achieve a balance between GPU memory utilization and training stability. An early stopping strategy is implemented during training. Training is automatically terminated when the validation set loss does not decrease for 5 consecutive rounds to avoid overfitting. The model is implemented based on the TensorFlow framework and trained on a GPU. The training time per round is approximately 12 minutes, and the total training time is approximately 10 hours.

[0065] (III) Training Process Optimization: Xavier initialization was used when initializing weights to ensure consistent variance of activation values ​​across layers. Batch normalization layers were added after convolutional layers to reduce internal covariate bias and stabilize training. The training process was visualized using TensorBoard, observing the loss function and accuracy curves. It was found that the model converged after 30 rounds, achieving a final training set accuracy of 93.2% and a validation set accuracy of 90.8%, demonstrating good generalization ability. Comparative experiments showed that this hybrid model improved accuracy by approximately 5.6% compared to single CNN or LSTM models.

[0066] Model evaluation and performance validation:

[0067] (I) Implementation of 10-fold cross-validation: The training set was divided into 10 equal parts. Nine parts were used as the training set and one part as the test set in each iteration, repeated 10 times. The accuracy, sensitivity, and specificity of each test were calculated, and the average value was used as the model performance evaluation result. Specifically, the StratifiedKFold module of scikit-learn was used to ensure a consistent ratio of positive to negative samples in each fold. Experimental results showed an average accuracy of 90.5% ± 1.2%, sensitivity of 88.7% ± 1.8%, and specificity of 92.3% ± 1.5%, demonstrating the model's stability under different data partitions.

[0068] (II) Independent Test Set Validation: On the independent test set, the model achieved an accuracy of 89.3%, a sensitivity of 87.6%, a specificity of 91.1%, and an AUC of 0.945. This test set data, sourced from different experimental platforms and containing RNA sequences from various tissues, validated the model's effectiveness in real-world applications. Compared to mainstream algorithms such as PredcircRNA and circDeep, this method showed significant improvements across all metrics, particularly in sensitivity, which increased by approximately 9.2%, indicating a stronger ability to identify low-expression circular RNAs.

[0069] (III) Robustness and Adaptability Testing: By randomly replacing 10% of the bases in the sequence to simulate a sequencing noise environment, the model accuracy remained at 85.7%, demonstrating strong anti-interference ability. Analysis of the classification performance for sequences of different lengths showed that the accuracy was 89.1% for 200-500bp sequences, 91.2% for 500-1000bp sequences, and 90.8% for sequences exceeding 1000bp, indicating that the model has good adaptability to RNA sequences of different lengths. This characteristic solves the problem of decreased accuracy in identifying long sequences in existing algorithms.

[0070] Interpretability analysis implementation:

[0071] (I) SHAP value calculation and feature importance assessment:

[0072] This method analyzes the contribution of each feature to the model's prediction based on Shapley value theory. It quantifies the impact of each feature on the classification result by calculating its marginal contribution. Specifically, the DeepExplainer module of the shap library was used to perform feature attribution analysis on 1000 test samples. The results show that the longest complementary segment length had the highest mean Shapley value (0.32), followed by the ALU score function value (0.28) and the proportion of 5' complementary segments (0.21), demonstrating that the complementary structure within the sequence and ALU element pairing are key factors influencing circular RNA formation.

[0073] (II) Feature-Prediction Relationship Visualization: SHAP force-directed graphs are used to illustrate the direction and strength of each feature's influence in single-sample prediction. For example, when the longest complementary region of a sample is 35 bp (12 bp higher than the mean), its SHAP value is +0.42, significantly shifting the prediction probability towards circular RNA; while when the ALU score is 5 (positive), the SHAP value is +0.35, similarly enhancing the circularization probability. Conversely, when the GC content is below the mean, the SHAP value is -0.21, inhibiting circular RNA prediction. This visualization method transforms the decision-making process of the black-box model into an understandable biological explanation, helping researchers understand the model's prediction logic.

[0074] (III) Biological Experimental Validation: Experimental validation was conducted on the key features identified by SHAP analysis. The longest complementary region of a circular RNA in HEK293 cells was knocked out using CRISPR-Cas9 technology, and qPCR analysis showed a 68% decrease in circularization efficiency. A sequence with a high ALU score was artificially synthesized and transfected into cells; quantitative real-time PCR results showed a 42% increase in circularization efficiency. The experimental results were highly consistent with the model analysis, confirming the biological rationality of the feature extraction method and providing computational support for the study of circular RNA formation mechanisms.

[0075] Extended implementation:

[0076] Differentiating circRNAs from protein-coding genes: The model was applied to a classification task of 1000 circular RNAs and 1000 protein-coding genes (PCGs), adding open reading frame length as a feature to the existing features. The classification accuracy reached 92.7%, with an AUC of 0.968. The longest complementary region length and ALU score remained the most critical features, accounting for 35% and 28% of feature importance, respectively. This application helps researchers quickly distinguish between non-coding circular RNAs and coding genes, laying the foundation for functional studies, and is particularly suitable for the initial screening of transcriptome data.

[0077] circRNA expression prediction in stem cells: Based on the GSE115472 dataset, which includes RNA sequencing data from embryonic stem cells (ESCs) and fibroblasts, a model was used to predict cell-specific expression of circular RNAs. In ESC-specific expression prediction, the accuracy reached 88.3%, successfully identifying 127 high-confidence circRNAs, of which 93 were validated by qPCR to have significantly higher expression levels in ESCs than in fibroblasts (p<0.01). This application provides candidate molecules for studying the regulatory role of circular RNAs during stem cell differentiation; for example, the prediction results showed that circRNA-123 was highly expressed in ESCs, and functional experiments confirmed its involvement in maintaining stem cell pluripotency.

[0078] This method achieves high-precision circular RNA identification through multi-dimensional technological innovation: at the data level, rigorous screening and standardization ensure input quality; at the feature level, a targeted feature extraction scheme is designed by combining the biological mechanisms of inverse complementary structures and ALU element pairing; at the model level, the local feature extraction capabilities of CNNs and the temporal modeling capabilities of BLSTMs are integrated to improve classification accuracy; and at the analysis level, the SHAP interpretability method is introduced to bridge the gap between computational models and biological mechanisms. Extended applications validate the method's generalization ability, providing a complete computational solution for circular RNA research, and it can be widely applied to biological scenarios such as transcriptome data analysis and disease-related circRNA screening.

[0079] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for automated identification of circular RNA with biological interpretability, characterized in that, Includes the following sequential steps: S1. Data preparation: Human circRNA and lncRNA sequence data were obtained from circAtlas and GENCODE. Sequences shorter than 200bp and repetitive or similar sequences were removed. The sequences were extracted into FASTA format and the coordinate system and orientation were standardized. Positive and negative sample sets were constructed based on the processed sequences. S2. Feature Extraction: For the positive and negative sample sets constructed in S1, extract the inverse complement matching features and ALU repeating elements to characterize the circularization potential of the sequence. S3. Feature Selection: The MRMD2.0 algorithm is used to select the most discriminative feature subset from the high-dimensional features extracted in S2, providing optimized feature input for subsequent model training; S4. Model Building: Based on the feature subset selected in S3, a deep neural network model combining convolutional neural network and bidirectional long short-term memory network is constructed. The network structure automatically learns multi-dimensional feature representations and performs classification training. S5. Model Evaluation: Using the model trained in S4, evaluate its performance through cross-validation and independent test sets to verify the model's ability to recognize circRNAs; S6. Interpretability Analysis: For the model evaluated in S5, the SHAP method is used to analyze the contribution of each feature to the model prediction, revealing the biological association between the features and the prediction results; S7. Extended Applications: Apply the model validated in S6 to distinguish between circRNA and PCG and to predict the expression of circRNA in stem cells, thus expanding the practical application scenarios of the model.

2. The method according to claim 1, characterized in that, In the S1 data preparation, reliable circRNAs are selected from circAtlas2.0 as positive samples, and lncRNAs with the same length as the positive samples are selected from GENCODE as negative samples. Sequences with a length of less than 200bp and lncRNAs that overlap with the positive samples are deleted. Training and test sets are constructed while ensuring a balance between the number of positive and negative samples.

3. The method according to claim 1, characterized in that, In the S2 feature extraction, the method for extracting the inverse complement matching feature is as follows: For the sequence processed by S1, a matching matrix is ​​constructed. When the bases at position i and position j are complementary pairs, the value is 1; otherwise, it is 0. The possible inverse complementary segments are calculated by traversing the sequence, and their length, number and distribution are statistically analyzed to form a numerical feature describing the potential of the ring structure.

4. The method according to claim 1, characterized in that, In the S2 feature extraction, the method for extracting ALU repeating elements is as follows: considering the effect of pairing and looping of ALU elements in different directions, a composite scoring function is introduced: ； in , , , These represent the different combinations of forward and reverse Alu pairings in the upstream and downstream introns, respectively, and the scoring results are used as feature values.

5. The method according to claim 1, characterized in that, In the S3 feature selection, the MRMD2.0 algorithm, which is an improvement on the PageRank algorithm, is used to rank the importance of the features extracted in S2 and obtain the optimal feature subset.

6. The method according to claim 1, characterized in that, In the construction of the S4 model, the model receives the feature subset filtered by S3 as input, first extracts key patterns through convolution and pooling operations, then learns time series characteristics through a multi-layer convolutional neural network, and finally performs classification prediction through a fully connected layer. The output layer uses the Sigmoid activation function, and the training process uses the cross-entropy loss function and the Adam optimizer, with the learning rate set to 0.

001.

7. The method according to claim 1, characterized in that, In the evaluation of the S5 model, ten-fold cross-validation and independent test sets are used to train and test the model built by S4 to verify metrics such as accuracy, sensitivity, and AUC.

8. The method according to claim 1, characterized in that, In the S6 interpretability analysis, the marginal contribution of the i-th feature is measured by the Shapley value to determine the sequence features that play a key role in the S5 classification result.

9. The method according to claim 1, characterized in that, In the S7 extended application, the model is used to distinguish between circRNA and PCG, and classification is achieved based on the feature importance analyzed in S6.

10. The method according to claim 1, characterized in that, In the S7 extended application, the model is used to predict whether circRNA is expressed in stem cells, and the network model constructed in S4 is used to analyze and predict the features.

Images