Calculation method for predicting interaction between circular RNA and micro RNA

Abstract

The invention provides a method and a system TGrKCMI for predicting interaction between circRNA (Ribonucleic Acid) and miRNA (Micro Ribonucleic Acid). According to the method, a pre-training model is used for extracting sequence features of circRNA and miRNA, and redundant information is reduced through PCA dimension reduction; then, a multi-head attention module with a gating mechanism is introduced to encode sequence features, and the expression ability of key information is enhanced; on the basis, a circRNA-miRNA interaction diagram is constructed, and robust diagram feature learning is realized in combination with diagram attention convolution of random feature masks. And finally, a kernel-based adaptive nonlinear classifier is adopted to carry out modeling on the fusion features, and high-precision and high-robustness interaction prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of RNA molecule interaction prediction technology in bioinformatics, specifically a method for predicting circular RNA (circRNA)–microRNA (miRNA) interactions. Background Technology

[0002] Circular RNA (circRNA) and microRNA (miRNA) are two classes of non-coding RNAs (ncRNAs) that have received widespread attention in gene regulation and complex disease research in recent years. miRNAs are short-chain molecules approximately 21–23 nt in length that can regulate gene expression by binding to the 3′ untranslated region of target genes, playing important roles in tumor monitoring, diagnosis, and treatment. Although circRNAs were discovered earlier, their biological functions remained largely unrecognized due to limitations in detection methods. With the development of high-throughput sequencing and computational biology, circRNAs have been confirmed to be ubiquitous in various eukaryotes and possess high stability and a long half-life due to their closed-circular structure. Existing research indicates that circRNAs can bind to miRNAs through their miRNA response elements, acting as a "miRNA sponge" to influence biological processes such as cell proliferation, migration, differentiation, and tumorigenesis. Further studies show that circRNA–miRNA interaction (CMI) plays a crucial regulatory role in various cancers, cardiovascular diseases, and metabolic disorders. Therefore, developing effective CMI recognition methods is of significant value for pathological mechanism research, disease biomarker development, and the formulation of related treatment strategies. However, traditional wet experimental methods for validating circRNA typically involve complex sample preparation processes, long experimental cycles, and require expensive reagents, instruments, and related equipment. Furthermore, these experiments demand high levels of operational skills and expertise from researchers, have limited overall throughput, and are insufficient to meet the needs of large-scale, systematic circRNA discovery. This significantly restricts circRNA-related research in practice. With the development of big data analytics, various computational methods for inferring biomolecular associations have emerged, capable of building models based on existing interaction data and predicting potential unknown associations. These methods can provide candidate object screening before subsequent experimental validation, thereby reducing experimental input and improving research efficiency. Although these techniques alleviate the limitations of low-throughput experiments to some extent, problems such as insufficient prediction accuracy and feature representation still exist. Summary of the Invention

[0003] The purpose of this invention is to address the problems of insufficient semantic feature expression, inadequate utilization of structural information, and limited robustness and generalization ability in existing circRNA–miRNA interaction prediction models. This invention provides a circRNA–miRNA interaction prediction method (TGrKCMI) based on gated multi-head attention and graph structure learning. This method can efficiently extract discriminative features of RNA sequences, adapt to topological relationship modeling under conditions of feature noise and missing features, and improve the recognition ability of complex interaction patterns through a nonlinear classifier, thereby achieving more reliable CMI prediction. The technical solution for achieving the purpose of this invention includes the following steps: The following is the technical solution for achieving the objective of this invention, comprising the following steps: Dataset Collection and Preprocessing: To systematically evaluate algorithm performance and fairly compare it with existing models, this study used three publicly available experimental validation sets: CMI-9905, CMI-9589, and CMI-20208, containing 9905, 9589, and 20208 circRNA-miRNA interactions, respectively. Each dataset was randomly paired to generate an equal number of negative samples, which were then divided into training, validation, and test sets in an 8:1:1 ratio. Five-fold cross-validation was performed during training. Gated multi-head attention feature enhancement: The dimensionality-reduced sequence features are input into the gated multi-head attention module, key semantic information is extracted through the multi-head attention mechanism, and the feature channels that contribute to the prediction task are selectively enhanced through the gating unit, thereby obtaining more discriminative circRNA and miRNA representations; Construction of a KAN-based nonlinear classifier: The nodes embedded from the graph convolution output are concatenated or combined and then input into the Kernel-based Adaptive Nonlinearity (KAN) classification module. The adaptive kernel function is used to enhance the nonlinear representation ability of the feature space, thereby improving the recognition performance of complex interaction patterns. Model optimization methods: Adaptive optimization algorithms are used to train the model parameters, and the classification performance is optimized through the cross-entropy loss function; during the training process, batch normalization and an appropriate proportion of Dropout are used to prevent overfitting and improve model stability. Validation system establishment: Cross-validation and independent test set are used for dual evaluation, and the area under the ROC curve (AUC) and the area under the PR curve (AUPR) are used to measure the effectiveness and generalization ability of the model. Beneficial effects

[0004] The prediction framework of this model includes: first, extracting high-dimensional RNA features through a pre-trained model and using PCA for dimensionality reduction to reduce redundant information; then, leveraging the Gated Multi-Head Attention (GMHA) module to enhance key semantic information. Next, combining Graph Attention Convolution (GAT) and Random Feature Masking (RFM) improves robustness to noise and missing features, and enhancing the expressive power of the feature space through a Kernel-based Adaptive Nonlinearity (KAN) classifier, ultimately achieving accurate nonlinear discrimination. Attached Figure Description

[0005] Figure 1 This is a flowchart of the drug-drug interaction prediction method based on pre-trained drug features proposed in this invention.

[0006] Figure 2 The ROC curve and AUPR curve of the model under 5-fold cross-validation are shown. Detailed Implementation

[0007] The specific embodiments of the present invention are illustrated below with reference to the accompanying drawings. The drawings are for illustrative purposes only and should not be construed as limiting the present invention. The drawings are for reference and illustration only and do not constitute a limitation on the scope of protection of the present invention. As long as various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims, all inventions utilizing the concept of the present invention are within the scope of protection. like Figure 1 As shown in the figure, this illustrates the overall workflow of this study. High-dimensional RNA features are extracted using a pre-trained model, and PCA is employed for dimensionality reduction to decrease redundant information. Then, the Gated Multi-Head Attention (GMHA) module is used to enhance key semantic information. Next, Graph Attention Convolution (GAT) and Random Feature Masking (RFM) are combined to improve robustness to noise and missing features, and a Kernel-based Adaptive Nonlinearity (KAN) classifier is used to enhance the expressive power of the feature space. To construct and evaluate the proposed model, we used three datasets to test its circRNA–miRNA interaction capabilities. The models employed the same network structure and parameters, differing only in the input data. Detailed information about the datasets is provided in Table 1.

[0008] Table 1 Standard Dataset Dataset CircRNA MiRNA CMI CMI-9589 2115 821 9589 CMI-9905 2346 962 9905 CMI-20208 3569 1152 20208 We evaluated the effectiveness of the proposed model using five-fold cross-validation on the training set and further validated its performance using an independent test set. To comprehensively measure the predictor's performance, we selected the area under the ROC curve (AUC) and AUPR as evaluation metrics. Finally, we compared the proposed method with state-of-the-art methods on three datasets, and the comparison results on the independent test set are summarized in Tables 2, 3, and 4, respectively.

[0009] Table 2 Performance comparison of other state-of-the-art models based on the CMI-9589 dataset Method WSCD KS-CMI BioDGW-CMI BEROLECMI Hither-CMI MuGNet-CMI Ours AUC 0.8923 0.9086 0.9026 0.9104 0.9134 0.9248 0.9740 AUPR 0.8935 0.9144 0.8962 0.9086 0.9127 0.9278 0.9734 Table 3 Performance comparison of other state-of-the-art models based on the CMI-9905 dataset Method CACMA KS-CMI BioDGW-CMI BEROLECMI Hither-CMI MuGNet-CMI Ours AUC 0.9156 0.9179 0.9476 0.9491 0.9517 0.9568 0.9653 AUPR 0.9086 0.9181 0.9416 0.9431 0.9469 0.9577 0.9687 Table 4. Performance comparison of other state-of-the-art models based on the CMI-20208 dataset. Method WSCD CACMA Hither-CMI MuGNet-CMI Ours AUC 0.8898 0.9170 0.9182 0.9221 0.9728 AUPR 0.8847 0.9131 0.9164 0.9221 0.9708

Claims

1. A computational method for predicting interactions between circular RNA and microRNA, comprising the following steps: 1) RNA data sample set construction: Collect sequence data of circRNA and miRNA and their known interaction information to construct positive and negative sample pairs; divide the data into training set, validation set and independent test set according to the preset ratio for model training and performance evaluation; 2) High-dimensional sequence feature extraction and dimensionality reduction: High-dimensional sequence embedding features of circRNA and miRNA are extracted using a pre-trained RNA characterization model; Principal component analysis (PCA) is used to reduce the dimensionality of the embedding vectors to reduce redundant information and improve the compactness of feature expression; 3) Enhanced gated multi-head attention features: The dimensionality-reduced sequence features are input into the gated multi-head attention module, key semantic information is extracted through the multi-head attention mechanism, and the feature channels that contribute to the prediction task are selectively enhanced through the gating unit, thereby obtaining more discriminative circRNA and miRNA representations; 4) Construction of a KAN-based nonlinear classifier: The nodes embedded from the graph convolution output are concatenated or combined and then input into the Kernel-based Adaptive Nonlinearity (KAN) classification module. The adaptive kernel function is used to enhance the nonlinear representation ability of the feature space, thereby improving the recognition performance of complex interaction patterns. 5) Model optimization methods: Adaptive optimization algorithms are used to train the model parameters, and the classification performance is optimized through the cross-entropy loss function; during the training process, batch normalization and an appropriate proportion of Dropout are used to prevent overfitting and improve model stability. 6) Establishment of validation system: The model performance is evaluated through cross-validation and independent test sets. The accuracy, sensitivity, specificity and Matthews correlation coefficient (MCC) are used to verify the effectiveness and generalization ability of the model.

2. The prediction framework of this model includes: First, high-dimensional RNA features are extracted using a pre-trained model, and PCA is used for dimensionality reduction to reduce redundant information. Then, the Gated Multi-Head Attention (GMHA) module is used to enhance key semantic information. Next, Graph Attention Convolution (GAT) and Random Feature Masking (RFM) are combined to improve robustness to noise and missing features. Finally, the Kernel-based Adaptive Nonlinearity (KAN) classifier is used to enhance the expressive power of the feature space, ultimately achieving accurate discrimination.

3. The computational method for predicting the interaction between circular RNA and microRNA according to claim 1, characterized in that: The feature encoding schemes and model building methods used in steps 2), 3), and 4) achieve multi-source feature fusion and hierarchical feature representation through pre-trained sequence representation, principal component analysis dimensionality reduction, gated multi-head attention feature enhancement, and collaborative modeling using a Kernel-based Adaptive Nonlinearity (KAN) classifier. This significantly improves the ability to resolve complex sequence semantics and nonlinear interaction patterns in circRNA–miRNA interactions, thereby enhancing prediction accuracy and model generalization performance.

Images