Circrna and mirna interaction prediction method and system fusing multi-scale features and two-stage contrast learning

Abstract

This invention relates to a method and system for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning. The method includes: extracting multi-dimensional heterogeneous features from circRNA and miRNA sequence data respectively; constructing circRNA and miRNA homogeneity maps based on the multi-dimensional heterogeneous features; generating circRNA and miRNA structure embedding vectors with fused structural information using a graph attention network; inputting the circRNA structure embedding vectors into a depthwise separable convolutional network to extract local motif features, and inputting the miRNA structure embedding vectors into a selective kernel attention network to extract multi-scale global features; performing single-molecule level contrastive learning and interaction level contrastive learning; fusing the final circRNA and miRNA representations; and inputting the fused representation data into a multilayer perceptron network to obtain the prediction result. This method can reduce prediction costs and improve the prediction accuracy and robustness of the model.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics processing technology, and in particular to a method and system for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage comparative learning. Background Technology

[0002] Non-coding RNAs (ncRNAs) participate in gene expression regulation and various cellular life processes. With the rapid development of transcriptomics research, increasing evidence suggests that ncRNAs are not merely transcriptional byproducts, but actively participate in the regulation of numerous key biological functions. Among them, circular RNAs (circRNAs) and microRNAs (miRNAs) have attracted widespread attention due to their good stability and important regulatory roles, playing crucial roles in development, metabolism, and disease progression. circRNAs possess a unique covalently closed circular structure, which gives them high stability and evolutionary conservation within cells. Existing studies have confirmed that circRNAs participate in gene regulation through multiple mechanisms, with their interaction with miRNAs being particularly typical. miRNAs are short-chain non-coding RNAs that primarily exert their regulatory functions by inhibiting the expression of target genes. circRNAs can bind to miRNAs and regulate their biological activity, playing a key role in the competitive endogenous RNA (ceRNA) regulatory network. This regulatory mechanism also participates in the regulation of various physiological processes and disease-related signaling pathways.

[0003] Currently, methods for identifying circRNA-miRNA interactions are mainly divided into traditional biological experimental methods and computer-aided prediction methods. Traditional biological experimental methods primarily verify circRNA-miRNA interactions through techniques such as molecular hybridization, luciferase reporter genes, and RNA immunoprecipitation. While the verification results are reliable, they have significant drawbacks: cumbersome and complex experimental procedures, high labor and time costs; low detection throughput, making large-scale, systematic interaction screening difficult; and long experimental cycles, failing to quickly provide candidate targets for mechanism research and drug development, and failing to meet the analytical needs of high-throughput omics data. Computer-aided analysis methods, relying on machine learning, graph neural networks, and other algorithms, integrate multi-source data such as RNA sequences, structures, and interaction networks to construct predictive models, achieving efficient mining of circRNA-miRNA interactions. They offer the advantages of low cost and high throughput, becoming an important supplementary means to experimental verification. However, existing computational models still have significant shortcomings in feature learning, structural modeling, and representation optimization: feature dimensions are singular or fusion strategies are unreasonable, often relying only on a few features such as sequence and structure, lacking the integration of multi-scale biological information such as semantics, network roles, context, and evolutionary conservation; multi-source feature fusion strategies lack optimization, making it difficult to fully leverage the complementary value of heterogeneous features; some models rely solely on data-driven modeling, resulting in poor interpretability of prediction results; and the models have limited generalization ability, performing poorly in predicting interactions between low-similarity and newly discovered RNA molecules.

[0004] Therefore, traditional methods for predicting circRNA and miRNA interactions are often costly and have poor prediction performance for low-similarity sequences and novel molecules due to the cumbersome nature of traditional biological experiments, low throughput, and high cost. Existing computational models also lack differentiated characterization of the two types of RNA molecules and have limited generalization ability. Summary of the Invention

[0005] Based on this, in order to solve the above-mentioned technical problems, a method and system for predicting circRNA and miRNA interactions that integrates multi-scale features and two-stage contrastive learning is provided, which can reduce prediction costs and improve the prediction accuracy and robustness of the model.

[0006] A method for predicting circRNA and miRNA interactions that integrates multi-scale features and two-stage contrastive learning, the method comprising:

[0007] Obtain the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA sequence data and miRNA sequence data respectively. The multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features.

[0008] Based on the sequence similarity features, construct circRNA homogeneity maps and miRNA homogeneity maps, and use the Doc2Vec feature, Role2Vec feature, TPCP feature and RNAErnie feature as node attributes. Generate circRNA structure embedding vectors and miRNA structure embedding vectors with fused structural information through graph attention network.

[0009] The circRNA structure embedding vector is input into a depthwise separable convolutional network to extract local motif features, and the miRNA structure embedding vector is input into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations.

[0010] Based on the enhanced circRNA and miRNA representations, single-molecule hierarchical contrastive learning is performed to optimize the consistency of single-molecule representations; interaction hierarchical contrastive learning is performed to optimize the discriminativeness of interaction patterns, resulting in the final circRNA and final miRNA representations.

[0011] The final circRNA characterization and the final miRNA characterization are fused, and the fused characterization data is input into a multilayer perceptron network to obtain a prediction result that there is an interaction between circRNA and miRNA.

[0012] In one embodiment, multi-dimensional heterogeneous features are extracted from the circRNA sequence data and miRNA sequence data, respectively, including:

[0013] The nucleotide sequences of the circRNA and miRNA sequence data are split into word sequences using k-mer, and the word sequences are input into the Doc2Vec model to obtain Doc2Vec semantic features.

[0014] Based on the sequence similarity network, an attribute random walk is performed, and the structural role embedding vector of each node is learned to obtain the Role2Vec structural role features;

[0015] Trinucleotide fragments are generated by traversing the circRNA sequence data and miRNA sequence data. Based on the physicochemical characteristics of the trinucleotide fragments, the physicochemical characteristics of TPCP trinucleotides are matched.

[0016] The circRNA and miRNA sequence data are input into a pre-trained Transformer model, and long-distance contextual dependencies in the sequences are captured through a multi-layer multi-head self-attention mechanism to obtain RNAErnie context-dependent features.

[0017] The minimum edit distance between two RNA sequences was calculated using the Levenstein distance, and after normalization, it was converted into a similarity score to obtain sequence similarity features.

[0018] In one embodiment, a circRNA homogeneity map and a miRNA homogeneity map are constructed based on the sequence similarity features. The Doc2Vec feature, Role2Vec feature, TPCP feature, and RNAErnie feature are used as node attributes. A graph attention network is used to generate circRNA structure embedding vectors and miRNA structure embedding vectors that fuse structural information, respectively, including:

[0019] Based on the sequence similarity features, pairwise similarity matrices of the circRNA molecule set and the miRNA molecule set are calculated respectively, and undirected edges are constructed based on the similarity matrices with a preset similarity threshold to obtain the circRNA homogeneity map and the miRNA homogeneity map respectively.

[0020] The Doc2Vec, Role2Vec, TPCP, and RNAErnie features are used as node attributes. A graph attention network is used to aggregate neighborhood information on the circRNA and miRNA homogeneity maps, thereby generating circRNA and miRNA structure embedding vectors that incorporate structural information.

[0021] In one embodiment, the circRNA structure embedding vector is input into a depthwise separable convolutional network to extract local motif features, and the miRNA structure embedding vector is input into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations, including:

[0022] The circRNA structure is embedded into a vector input into a depthwise separable convolutional network. Local motif features are extracted through depthwise convolution in the depthwise separable convolutional network, and cross-channel information is fused through pointwise convolution to obtain an enhanced circRNA representation.

[0023] The miRNA structure is embedded into a vector input selective kernel attention network. By splitting the network through scale convolution branches, feature maps of different spatial scales are extracted. After global information fusion and adaptive weight selection, the enhanced miRNA representation is obtained.

[0024] In one embodiment, based on the enhanced circRNA and miRNA representations, single-molecule hierarchical contrastive learning is performed to optimize the consistency of single-molecule representations, including:

[0025] The circRNA structure embedding vector and the enhanced circRNA representation are used as different feature views of the same circRNA molecule;

[0026] Different feature view pairs of the same circRNA molecule are used as the first positive sample pair, and feature view pairs of different circRNA molecules are used as the first negative sample pair;

[0027] Based on the first positive sample pair and the first negative sample pair, cosine similarity is used as the similarity measure, and a temperature parameter is introduced to regulate the smoothness of the similarity distribution. A single-molecule contrastive loss function for circRNA is constructed to complete the single-molecule hierarchical contrastive learning of circRNA representation.

[0028] The miRNA structure embedding vector and the enhanced miRNA characterization are used as different feature views of the same miRNA molecule;

[0029] Different feature view pairs of the same miRNA molecule are used as the second positive sample pairs, and feature view pairs of different miRNA molecules are used as the second negative sample pairs.

[0030] Based on the second positive sample pair and the second negative sample pair, cosine similarity is used as the similarity measure, and a temperature parameter is introduced to regulate the smoothness of the similarity distribution. A single-molecule contrastive loss function for miRNA is constructed to complete the single-molecule hierarchical contrastive learning of miRNA characterization.

[0031] In one embodiment, based on the enhanced circRNA and miRNA representations, interaction-level comparative learning is performed to optimize the discriminative power of interaction patterns, including:

[0032] Positive samples are selected from circRNA-miRNA pairs that have interaction, and negative samples are selected from circRNA-miRNA pairs that do not have interaction. Positive and negative samples are sampled evenly within a batch.

[0033] Cosine similarity was used to calculate the similarity of each circRNA-miRNA pair in the final representation space, and an interaction hierarchical contrastive loss function was constructed to complete the interaction hierarchical contrastive learning.

[0034] The weighted sum of the circRNA single-molecule contrast loss function, miRNA single-molecule contrast loss function, and interaction hierarchy contrast loss function is used to obtain the final loss function.

[0035] Based on the final loss function, the final circRNA characterization and the final miRNA characterization are obtained.

[0036] In one embodiment, the final circRNA characterization and the final miRNA characterization are fused, and the fused characterization data is input into a multilayer perceptron network to obtain a prediction result indicating an interaction between circRNA and miRNA, including:

[0037] The final circRNA representation and the final miRNA representation are concatenated to obtain the fused feature vector.

[0038] The fused feature vector is sequentially input into a multilayer perceptron network containing fully connected layers, nonlinear activation layers, batch normalization layers, and Dropout layers to extract deep nonlinear features.

[0039] The deep nonlinear features are mapped using the Sigmoid activation function to obtain the predicted probability of an interaction between circRNA and miRNA.

[0040] In one embodiment, among the multidimensional heterogeneous features, the Doc2Vec semantic feature is used to capture global-level semantic information of the RNA sequence; the Role2Vec structural role feature is used to encode the structural role information of the RNA molecule in the interaction network through attribute random walks; the TPCP trinucleotide physicochemical feature is used to extract biophysical information related to local structural stability and stacking interactions at the trinucleotide level; the RNAErnie context-dependent feature is used to capture multi-level contextual dependencies of the RNA sequence at the base level, subsequence level, and motif level based on the Transformer architecture; and the sequence similarity feature is based on the Levenstein distance calculation and is used to quantify evolutionary conservation information between sequences.

[0041] In one embodiment, the method further includes:

[0042] Obtain the dataset and divide it into a validation set and a training set;

[0043] Five-fold cross-validation is performed based on the validation set and training set, and the average of the five validation results is taken as the prediction performance result.

[0044] Determine the evaluation metrics and use the evaluation metrics to evaluate the prediction performance results.

[0045] A circRNA and miRNA interaction prediction system integrating multi-scale features and two-stage contrastive learning, the system comprising:

[0046] The feature encoding module is used to acquire the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA sequence data and miRNA sequence data respectively. The multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features.

[0047] The feature learning module is used to construct circRNA homogeneity maps and miRNA homogeneity maps based on the sequence similarity features. The Doc2Vec feature, Role2Vec feature, TPCP feature and RNAErnie feature are used as node attributes. The graph attention network is used to generate circRNA structure embedding vectors and miRNA structure embedding vectors that fuse structural information, respectively.

[0048] The feature enhancement module is used to input the circRNA structure embedding vector into a depthwise separable convolutional network to extract local motif features, and to input the miRNA structure embedding vector into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations.

[0049] The contrastive learning module is used to perform single-molecule-level contrastive learning based on the enhanced circRNA and miRNA representations to optimize the consistency of single-molecule representations; and to perform interaction-level contrastive learning to optimize the discriminativeness of interaction patterns, thereby obtaining the final circRNA and final miRNA representations.

[0050] The prediction module is used to fuse the final circRNA characterization and the final miRNA characterization, and input the fused characterization data into a multilayer perceptron network to obtain a prediction result that there is an interaction between circRNA and miRNA.

[0051] The aforementioned method and system for predicting circRNA and miRNA interactions, which integrates multi-scale features and two-stage comparative learning, can characterize RNA molecules from multiple levels by extracting multi-dimensional heterogeneous features, thereby improving the completeness and biological rationality of feature representation. By constructing a homogeneous graph based on sequence similarity and combining it with a graph attention network, it can adaptively aggregate key information from neighboring nodes, deeply fusing multi-scale features with topological information to generate more discriminative structural embedding vectors and improve the utilization rate of structural information. Addressing the differences in circRNA and miRNA molecule length and functional regions, it employs depthwise separable convolution to focus on local motifs and selective kernel attention to adaptively capture multi-scale global information, achieving differentiated and accurate representation. Through two-stage comparative learning at the single-molecule level and the interaction level, it first strengthens the consistency of single-molecule self-representation and then optimizes the interaction mode discrimination boundary, significantly improving the model's robustness, prediction accuracy, and generalization ability. End-to-end prediction using a multilayer perceptron enables efficient and stable output of interaction results, improving the model's prediction accuracy and robustness while reducing costs. Attached Figure Description

[0052] Figure 1 This is an application environment diagram of a circRNA and miRNA interaction prediction method that integrates multi-scale features and two-stage contrastive learning in one embodiment.

[0053] Figure 2 This is a flowchart illustrating a method for predicting circRNA and miRNA interactions that integrates multi-scale features and two-stage contrastive learning in one embodiment.

[0054] Figure 3 This is a schematic diagram of the model framework of MSCL-CMI in one embodiment;

[0055] Figure 4 The following diagrams illustrate the performance analysis of the MSCL-CMI model under different hyperparameter and feature combinations: (a) bar charts of indicators under different learning rates; (b) box plots of indicators under different batch sizes; (c) radar charts of indicators under different temperature coefficients; (d) radar charts of indicators for the CMI9589 dataset; (e) bar charts of indicators for the CMI9905 dataset; and (f) scatter plots of indicators for the CMI2028 dataset.

[0056] Figure 5This diagram illustrates the contribution of two-stage contrastive learning to model performance, comparing the model's performance with three ablation variants, including removing the first-stage contrastive learning (no first-stage CL), removing the second-stage contrastive learning (no second-stage CL), and removing both contrastive learning stages simultaneously (no CL); including: (a) box plot results on the CMI9589 dataset; (b) histogram results on the CMI9905 dataset; (c) polar histogram results on the CMI20208 dataset;

[0057] Figure 6 Heatmaps comparing the performance of the MSCL-CMI model on multiple evaluation metrics under different graph neural network architectures include: (a) CMI9589 dataset; (b) CMI9905 dataset; (c) CMI20208 dataset;

[0058] Figure 7 Visualizations of t-SNE for learning embedding representations of circRNA-miRNA pairings before and after training the MSCL-CMI model, including: before training (a) and after training (d) on the CMI9589 dataset; before training (b) and after training (e) on the CMI9905 dataset; and before training (c) and after training (f) on the CMI2008 dataset. Blue dots represent negative samples (no circRNA-miRNA interaction), and red dots represent positive samples (circRNA-miRNA interaction exists).

[0059] Figure 8 This is a block diagram of a circRNA and miRNA interaction prediction system that integrates multi-scale features and two-stage contrastive learning in one embodiment.

[0060] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe positive and negative sample pairs, but these positive and negative sample pairs are not limited by these terms. These terms are only used to distinguish the first positive and negative sample pair from another positive and negative sample pair. For example, without departing from the scope of this application, the first positive sample pair may be referred to as the second positive sample pair, and similarly, the second positive sample pair may be referred to as the first positive sample pair. Both the first and second positive sample pairs are positive sample pairs, but they are not the same positive sample pair.

[0063] The circRNA and miRNA interaction prediction method integrating multi-scale features and two-stage contrastive learning provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes computer device 110. Computer device 110 can acquire the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA and miRNA sequence data respectively. These multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features. Computer device 110 can construct circRNA and miRNA homogeneity maps based on the sequence similarity features, using Doc2Vec, Role2Vec, TPCP, and RNAErnie features as node attributes, and generating circRNA and miRNA structure embedding vectors with fused structural information through a graph attention network. 10 can embed the circRNA structure into a vector input depthwise separable convolutional network to extract local motif features, and embed the miRNA structure into a vector input selective kernel attention network to extract multi-scale global features, obtaining enhanced circRNA and miRNA representations. Computer device 110 can perform single-molecule hierarchical contrastive learning based on the enhanced circRNA and miRNA representations to optimize the consistency of single-molecule representations; it can also perform interaction hierarchical contrastive learning to optimize the discriminativeness of interaction patterns, obtaining final circRNA and final miRNA representations. Computer device 110 can fuse the final circRNA and final miRNA representations and input the fused representation data into a multilayer perceptron network to obtain a prediction result indicating the interaction between circRNA and miRNA. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, etc.

[0064] In one embodiment, such as Figure 2As shown, a method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning is provided, including the following steps:

[0065] Step 202: Obtain the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA sequence data and miRNA sequence data respectively. The multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features.

[0066] In this embodiment, three publicly available and experimentally validated benchmark datasets in the field of circRNA-miRNA interaction prediction—CMI9589, CMI9905, and CMI20208—were selected. CMI9589, sourced from the CircBank database, contains 2115 circRNAs, 821 miRNAs, and 9589 interaction pairs after redundancy removal. CMI9905 contains 2346 circRNAs, 962 miRNAs, and 9905 association pairs, offering both high-throughput and experimentally validated data, making it highly representative and reliable, and thus serving as the primary evaluation dataset. CMI20208, from the BGF-CMAP study, contains 3569 circRNAs, 1152 miRNAs, and 20208 experimentally validated association pairs, providing a larger dataset and allowing for a more comprehensive evaluation of the model's generalization ability. All datasets are used to construct circRNA-miRNA association matrices in adjacency matrix form: if circRNAi and miRNAj interact, then matrix element A... ij =1 otherwise 0. Simultaneously, an equal number of negative samples are randomly selected to supplement the association matrix, and 5-fold cross-validation is used to divide the dataset into training and test sets to minimize the impact of data partitioning bias on model evaluation.

[0067] To comprehensively characterize the sequence and structural information of circRNA and miRNA, computer devices can extract features from the collected sequence data. In one embodiment, a method for predicting circRNA and miRNA interactions that integrates multi-scale features and two-stage contrastive learning may further include a feature extraction and encoding process. Specifically, this process includes: splitting the nucleotide sequences of circRNA and miRNA sequence data into word sequences using k-mer algorithms; inputting the word sequences into a Doc2Vec model to obtain Doc2Vec semantic features; performing an attribute random walk based on a sequence similarity network and learning the structural role embedding vectors of each node to obtain Role2Vec structural role features; traversing circRNA and miRNA sequence data to generate trinucleotide fragments; matching TPCP trinucleotide physicochemical features based on the physicochemical characteristics of the trinucleotide fragments; inputting circRNA and miRNA sequence data into a pre-trained Transformer model; capturing long-distance contextual dependencies in the sequences through a multi-layer multi-head self-attention mechanism to obtain RNAErnie contextual dependency features; calculating the minimum edit distance between two RNA sequences using Levenstein distance, and converting it into a similarity score after normalization to obtain sequence similarity features.

[0068] Computer equipment can extract five multi-scale complementary features: Doc2Vec, Role2Vec, TPCP, RNAErnie, and sequence similarity. These features characterize RNA molecules from dimensions such as graph structure, distributed semantics, physicochemical properties, sequence information, and sequence similarity, forming a multi-level feature system covering local motifs, global structure, semantic distribution, and physicochemical properties. This provides rich and complementary input information for subsequent models.

[0069] Specifically, to obtain distributed semantic representations of RNA sequences, the computer device employs the classic document embedding model Doc2Vec, which can simultaneously learn local contextual information and global sequence-level features. In this embodiment, circRNA and miRNA sequences can be treated as sentences, and the k-mers (or nucleotides) they compose can be used as word input models. The computer device can encode the structural roles of nodes through attribute random walks, providing vector representations to facilitate the effective acquisition of network structure information. Role2Vec can generate node embedding vectors containing information about the roles of nodes in the interactive network.

[0070] Trinucleotide physicochemical properties (TPCP) is a characterization method that extracts biophysical information related to local structural stability, stacking interactions, and conformational properties based on the physicochemical characteristics of trinucleotides. Compared to traditional coding methods that rely solely on sequence structure, TPCP provides more easily interpretable physicochemical background information, thereby enhancing the model's ability to identify potential functional regions. Computer devices can generate trinucleotide fragments by traversing circRNA and miRNA sequence data, and then match TPCP trinucleotide physicochemical characteristics based on the physicochemical characteristics of these fragments.

[0071] Based on the ERNIE (Knowledge Fusion Enhanced Representation) framework, RNAErnie can be constructed using the Transformer architecture. Through pre-training, it captures multi-level features of RNA sequences at the base, subsequence, and motif levels. A multi-head attention mechanism is used to process different information channels of the sequence, effectively modeling the global contextual relationships of the sequence and generating high-dimensional feature representations that simultaneously reflect biological characteristics and sequence context. In this embodiment, the feature vectors extracted by RNAErnie are used as input to construct a circRNA-miRNA association prediction model, providing rich and reliable representation information for downstream prediction tasks.

[0072] In this embodiment, the computer device can use the Levenshtein distance to describe the similarity of RNA sequences. Specifically, the degree of difference between two sequences can be quantified by calculating the minimum number of editing operations (such as nucleotide substitutions, insertions, and deletions) required to convert one sequence into another. By simultaneously capturing changes in sequence length and nucleotide composition, this distance value is further converted into sequence similarity features for subsequent model analysis.

[0073] Step 204: Construct circRNA homogeneity maps and miRNA homogeneity maps based on sequence similarity features. Use Doc2Vec features, Role2Vec features, TPCP features, and RNAErnie features as node attributes. Generate circRNA structure embedding vectors and miRNA structure embedding vectors with fused structural information through a graph attention network.

[0074] Computer devices can construct undirected connections based on sequence similarity features and set similarity thresholds to build isomorphic network graphs for circRNA and miRNA, respectively. Under the unified graph topology, four feature codes—Doc2Vec, Role2Vec, TPCP, and RNAErnie—are encoded as node attributes and input to the graph attention network (GAT). Through the attention mechanism, information of similar nodes in the neighborhood is adaptively aggregated to learn a structured embedding vector that integrates multi-scale features.

[0075] Specifically, in one embodiment, the provided method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage comparative learning may further include a feature learning process. The specific process includes: calculating pairwise similarity matrices of the circRNA molecule set and the miRNA molecule set based on sequence similarity features, and constructing undirected edges based on the similarity matrices with a preset similarity threshold to obtain circRNA homogeneity maps and miRNA homogeneity maps respectively; using Doc2Vec features, Role2Vec features, TPCP features, and RNAErnie features as node attributes, and performing neighborhood information aggregation on the circRNA homogeneity map and miRNA homogeneity map through a graph attention network to generate corresponding circRNA structure embedding vectors and miRNA structure embedding vectors that integrate structural information.

[0076] Computer equipment can calculate pairwise similarity scores between sets of circRNA and miRNA molecules based on obtained sequence similarity features, forming a symmetric similarity matrix. A preset similarity threshold is used. For any two molecular nodes i and j, if the similarity score is greater than 0.5, an undirected edge is established between nodes i and j; if it is less than or equal to 0.5, no edge is established. This yields homogeneous undirected graphs for circRNA and miRNA, containing only similar molecular nodes and edge relationships. Next, the computer device can use the Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, and RNAErnie context-dependent features corresponding to each circRNA and miRNA node as multidimensional node attributes, completing the binding of graph nodes with heterogeneous features.

[0077] Next, the computer device can confidently perform forward propagation and neighborhood information aggregation using a graph attention network (GAT). In this embodiment, taking a circRNA homogeneous graph as an example, for each node i, its neighboring nodes j are traversed, and the following calculations are performed:

[0078] Linear transformations are applied to the node attributes of nodes i and j to obtain feature mapping vectors. The two mapping vectors are concatenated, and the original attention coefficients are calculated using the attention vector and nonlinear activation. Softmax normalization is applied to all original attention coefficients in the neighborhood to obtain adaptive attention weights, achieving weighted focusing on key neighborhood nodes. The features of neighborhood nodes are weighted and aggregated according to the attention weights to obtain the preliminary structural embedding of node i. Residual connections and activation functions are added to output a stable and smooth circRNA structural embedding vector.

[0079] In this embodiment, the computer device can perform the exact same GAT encoding process on the miRNA homogeneous graph to obtain the miRNA structure embedding vector. The final circRNA structure embedding vector and miRNA structure embedding vector have been combined with the node's own multi-dimensional heterogeneous features and graph topology association information, and used as input for subsequent differential representation branches.

[0080] In other words, given a set of RNA molecules, a computer device can first calculate the pairwise sequence similarity scores to obtain a sequence similarity matrix. ,in, This is used to quantify the similarity between the i-th and j-th RNA molecules. Based on the similarity matrix, a computer device can construct an undirected graph using a thresholding strategy. Specifically, if the similarity scores of two nodes i and j satisfy... (In this embodiment, a preset threshold) If the similarity is positive, an undirected edge is created between the two; otherwise, no edge is added. This process can generate sparse graph structures that can preserve high-confidence sequence similarity relationships while reducing potential noise introduced by weak similarity.

[0081] The resulting graph topology is fixed and shared in all subsequent feature modeling processes. On the constructed sequence similarity graph, each RNA molecule corresponds to multiple heterogeneous feature representations generated by different coding strategies. Specifically, this study employs four methods—Doc2Vec, Role2Vec, TPCP, and RNAErnie—to generate four independent feature views for each RNA molecule, denoted as: '';in This represents the feature vector of the i-th RNA molecule under the k-th coding strategy.

[0082] To model the structural dependencies between RNA molecules, this embodiment applies a graph attention network independently to each feature view while maintaining the graph topology. For node i, its relationships with neighboring nodes... The attention coefficient between people is calculated as follows: Where W is a learnable linear transformation matrix, and a is the attention vector. This indicates a vector concatenation operation.

[0083] Next, the computer device can normalize the attention coefficients using the Softmax function to obtain normalized attention weights: Then, by aggregating the information of neighboring nodes, the updated node representation is obtained: ;in, It is a non-linear activation function.

[0084] On the same sequence similarity map, a graph attention network is applied to each feature encoding result, ultimately yielding four sets of embedding vectors that fuse structural information: By capturing structural association information between RNA molecules from different feature representations, a foundational representation of fused structural information is provided for subsequent feature fusion and two-stage contrastive learning.

[0085] Step 206: Input the circRNA structure embedding vector into a depthwise separable convolutional network to extract local motif features, and input the miRNA structure embedding vector into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations.

[0086] Computer devices can innovatively design differentiated feature characterization branches to address the essential differences between circRNA and miRNA in molecular length, functional mechanism, and key functional regions, overcoming the limitations of existing technologies that use a unified feature fusion strategy for the two types of RNA.

[0087] In one embodiment, the provided method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning may further include a feature extraction and enhancement process. The specific process includes: embedding the circRNA structure into a vector input into a depthwise separable convolutional network, extracting local motif features through depthwise convolution in the depthwise separable convolutional network, fusing cross-channel information through pointwise convolution, and obtaining an enhanced circRNA representation; embedding the miRNA structure into a vector input into a selective kernel attention network, splitting it through scale convolution branches, extracting feature maps of different spatial scales respectively, and obtaining an enhanced miRNA representation through global information fusion and adaptive weight selection.

[0088] Specifically, considering the short sequences and function-dependent key local motifs of circRNAs, a depthwise separable convolution module is used in the circRNA branch to perform deep modeling of five features, enhancing the ability to capture local feature patterns. In other words, circRNA sequences are relatively short (ranging from tens to hundreds of nucleotides in length), and their function and interaction with miRNAs are largely determined by local sequence motifs. Therefore, capturing local patterns is more valuable than modeling long-range global sequences. Based on this, computer devices can use depthwise separable convolution (DSC) for circRNA feature extraction to better capture local sequence information and reduce computational overhead. DSC decomposes standard convolution into depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations independently on each channel, enabling the extraction of local features from each channel. ;in, This represents the input of the c-th channel. For the corresponding convolution kernel, This represents a convolution operation, where C represents the total number of channels. Subsequently, pointwise convolutions (1×1 convolutions) are used to linearly combine cross-channel features to integrate multi-channel information. ;in, This represents the 1×1 convolution weight.

[0089] Given shorter sequence lengths, compared to standard convolutional neural networks (CNNs), the depthwise convolutions in DSC can focus more on extracting local motifs. Pointwise convolutions can efficiently compensate for insufficient information exchange between channels, while significantly reducing the total number of model parameters and computational cost, thus mitigating the risk of overfitting. Therefore, DSC is suitable for short circRNA sequences, facilitating the targeted extraction of local feature information and providing high-quality representations for downstream circRNA-miRNA association (CMI) prediction.

[0090] In this embodiment, considering the characteristics of long miRNA sequences and their functional dependence on multi-scale global information, a Selective Kernel Attention (SKAttention) mechanism is introduced into the miRNA branch to adaptively adjust and dynamically allocate the weights of multi-scale feature channels, thereby improving feature fusion efficiency and representation capabilities. In other words, because miRNA sequences are relatively long, and their key functional features involve not only local nucleotide patterns but also long-distance contextual information, a Selective Kernel Attention (SKAttention) mechanism is introduced in miRNA feature extraction. SKAttention is an adaptive attention mechanism suitable for convolutional neural networks, designed to overcome the limitation of a fixed receptive field in standard convolution operations. This mechanism integrates parallel convolutional branches with channel attention, enabling the network to dynamically select an appropriate convolutional kernel scale based on the input feature content, thereby adaptively adjusting the network's effective receptive field. SKAttention typically includes three consecutive stages: Split, Fuse, and Select.

[0091] Specifically, the splitting stage mainly involves multi-scale feature extraction, specifically, given the input feature map... Features at different spatial scales are extracted through multiple parallel convolutional branches: Where Fi represents the transformation function of the i-th branch, typically composed of depthwise convolution, batch normalization, and ReLU activation. The fusion stage mainly involves global information aggregation, specifically by summing the outputs of all branches element by element: To obtain the channel-level descriptor, perform a global average pooling operation: The descriptor is processed by a fully connected layer with nonlinear activation to obtain a compact feature representation. ;in, Represents the ReLU function. These are learnable parameters. The selection phase mainly involves adaptive kernel selection, specifically projecting the compact vector z to generate attention weights for each convolutional branch, and then normalizing them using the Softmax function. The final feature map is obtained by weighted aggregation of the outputs of each branch. This enables the network to adaptively integrate information from different convolutional scales based on input features. Therefore, the MSCL-CMI model can incorporate both local feature details and global contextual information in miRNA sequence representation, providing support for downstream circRNA-miRNA association (CMI) prediction.

[0092] In this embodiment, the molecular characteristic-adaptive bibranch structure described above enables precise matching between feature representation and molecular function, significantly improving the targeting and effectiveness of feature learning.

[0093] Step 208: Based on the enhanced circRNA and miRNA representations, perform single-molecule hierarchical contrastive learning to optimize the consistency of single-molecule representations; perform interaction hierarchical contrastive learning to optimize the discriminativeness of interaction patterns, and obtain the final circRNA and final miRNA representations.

[0094] A two-stage multimodal contrastive learning framework (MSC) can be set up in the computer equipment for two-stage contrastive learning of circRNA and miRNA characterization.

[0095] In one embodiment, the provided method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning may further include a process of performing single-molecule-level contrastive learning. Specifically, this process includes: using the circRNA structure embedding vector and the enhanced circRNA representation as different feature views of the same circRNA molecule; using the different feature view pairs of the same circRNA molecule as the first positive sample pair and the feature view pairs of different circRNA molecules as the first negative sample pair; based on the first positive sample pair and the first negative sample pair, using cosine similarity as a similarity metric and introducing a temperature parameter to regulate the smoothness of the similarity distribution, constructing a circRNA single-molecule contrastive loss function to complete the single-molecule-level contrastive learning of circRNA representation; using the miRNA structure embedding vector and the enhanced miRNA representation as different feature views of the same miRNA molecule; using the different feature view pairs of the same miRNA molecule as the second positive sample pair and the feature view pairs of different miRNA molecules as the second negative sample pair; based on the second positive sample pair and the second negative sample pair, using cosine similarity as a similarity metric and introducing a temperature parameter to regulate the smoothness of the similarity distribution, constructing a miRNA single-molecule contrastive loss function to complete the single-molecule-level contrastive learning of miRNA representation.

[0096] Computer devices can use the characterization of different modal features of the same RNA molecule as positive sample pairs and the characterization of different RNA molecules as negative sample pairs. Through the contrast constraint at the single-molecule level, the consistent expression of the multimodal features of RNA itself is strengthened, and the robustness and discriminativeness of single-molecule feature characterization are improved.

[0097] Specifically, in the first stage of single-molecule level contrastive learning, for each circRNA molecule, its multi-scale feature representation is comprehensively integrated, covering features obtained from four encoding strategies: Doc2Vec, Role2Vec, TPCP, and RNAErnie. Based on these heterogeneous features, in this embodiment, the computer device adopts two complementary circRNA feature representations from different modeling perspectives, using Graph Attention Network (GAT) and Depthically Separable Convolution (DSC) respectively.

[0098] To achieve consistent alignment of circRNA representations across different modeling perspectives, this embodiment introduces a contrastive learning strategy. The core principle is that the representations of the same circRNA from different perspectives should have high similarity, while the representations of different circRNAs should have low similarity. Specifically, in a mini-batch containing B circRNAs, the cross-perspective representations obtained from the same circRNA are aligned. The first positive sample pair is considered as the first positive sample pair, while the cross-perspective characterization pairs obtained from different circRNAs (i.e., The first negative sample pair is considered as such. Next, the computer device can use cosine similarity to measure the similarity between different feature representations and introduce a temperature parameter τ to regulate the smoothness of the similarity distribution. Accordingly, the contrastive learning loss function for circRNA is defined as: Where N represents all negative samples in the same mini-batch. After comparison and optimization, the circRNA representations obtained from the two perspectives are concatenated and mapped to a unified feature space through a linear projection layer to obtain the final circRNA feature representation: ;in, Let be a trainable linear projection matrix, and || denote the vector concatenation operation.

[0099] For miRNAs, this embodiment employs the same approach to construct their multi-scale heterogeneous feature representations and completes modeling within the same contrastive learning framework. Specifically, the computer device can jointly encode different types of miRNA features using a graph attention network (GAT) and a selective kernel attention mechanism (SKAttention): from the GAT perspective, the heterogeneous feature encodings of miRNAs are aggregated through the graph attention mechanism to obtain miRNA representations that integrate correlation information. Simultaneously, the SKAttention module is used to adaptively adjust the weights and fuse channels for multiple feature encodings of miRNAs, resulting in a miRNA representation based on SKAttention. After obtaining miRNA representations from two modeling perspectives, a contrastive learning strategy identical to that used for circRNAs was employed: the cross-perspective representations of the same miRNA were compared. Defined as the second positive sample pair, a cross-perspective characterization pair of different miRNAs. Defined as the second negative sample pair; using cosine similarity as the similarity metric, the distribution of similarity scores is adjusted by the temperature parameter τ. The contrastive learning loss of miRNA is denoted as... Its formula is exactly the same as that for the contrast loss of circRNA.

[0100] After comparison and optimization, the miRNA representations from the two perspectives are concatenated and mapped to a unified feature space through a linear projection layer to obtain the final miRNA feature representation: ;in, is a trainable linear projection matrix.

[0101] Final characterization of the obtained circRNA With miRNA final characterization This constitutes the output of first-order contrastive learning. By aligning cross-modal features, the feature representation capabilities of circRNA and miRNA are enhanced, thereby providing high-quality single-molecule feature representations for subsequent interaction modeling.

[0102] In one embodiment, the provided method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning may further include a process of performing interaction-level contrastive learning. Specifically, this process includes: designating interacting circRNA-miRNA pairs as positive samples and non-interacting circRNA-miRNA pairs as negative samples; uniformly sampling positive and negative samples within a batch; calculating the similarity of each sampled circRNA-miRNA pair in the final representation space using cosine similarity, constructing an interaction-level contrastive loss function, and completing the interaction-level contrastive learning; weighted summing of the circRNA single-molecule contrastive loss function, the miRNA single-molecule contrastive loss function, and the interaction-level contrastive loss function to obtain the final loss function; and obtaining the final circRNA representation and the final miRNA representation based on the final loss function.

[0103] Computer devices can use experimentally verified circRNA-miRNA pairs as positive samples and pairs with no interaction or unknown interaction relationships as negative samples to construct a contrastive learning task at the interaction pair level. By optimizing the distribution of interaction feature space through similarity measurement, the distribution of real interaction pairs is made closer in the feature space and non-interaction pairs are made farther apart.

[0104] Specifically, in the second phase of interaction-level contrastive learning, the computer device can further introduce contrastive learning at the circRNA-miRNA association (CMI) level, explicitly incorporating known interaction information into the optimization objective. Specifically, the computer device can experimentally verify that circRNA-miRNA pairs exhibiting positive interactions are considered positive samples, while RNA pairs with no interaction or unclear interaction relationships are considered negative samples. This encourages RNA pairs with positive interactions to have closer embedding representations in the latent space, while simultaneously increasing the representation distance of RNA pairs with no interaction or unclear interaction relationships.

[0105] Computer devices, after defining positive and negative samples, are better able to capture biologically meaningful interaction patterns. Given a mini-batch of samples of size B (where positive and negative circRNA-miRNA pairs are uniformly sampled), the positive and negative sample sets are denoted as follows: and ,satisfy The contrastive loss function at the interaction level is defined as: Where CosSim is the cosine similarity function, Temperature parameters used to regulate the smoothness of similarity distribution.

[0106] The ultimate training objective is to simultaneously optimize the interactive prediction loss and the two-stage contrastive learning loss: in, and The comparison of circRNA and miRNA losses in the first stage are shown. For the comparative loss at the CMI level in the second phase; and is a hyperparameter used to balance the contribution weights of the prediction loss and different contrastive learning objectives.

[0107] The two-stage contrastive learning is not simply an additive process. The first stage addresses the redundancy and inconsistency in the representation of multimodal features in single molecules; the second stage addresses the ambiguity of the interaction relationship discrimination boundary. Together, they form a hierarchical optimization mechanism from the inside out and from the molecule to the interaction, resulting in overall performance that is significantly better than methods that only employ single-stage contrastive learning.

[0108] Step 210: Fuse the final circRNA characterization and the final miRNA characterization, and input the fused characterization data into a multilayer perceptron network to obtain a prediction result that there is an interaction between circRNA and miRNA.

[0109] After completing two-stage comparative learning, the computer device fuses the learned circRNA and miRNA representations to construct a feature with fused interactive information.

[0110] In one embodiment, a circRNA-miRNA interaction prediction method integrating multi-scale features and two-stage contrastive learning is provided and can be applied to the circRNA-miRNA interaction (CMI) prediction model. MSCL-CMI is a first-ever prediction model that integrates multi-scale heterogeneous features, differentiated representation branches, and a two-stage contrastive learning strategy. Its core lies in overcoming the limitations of existing models such as "single feature dimension, hierarchical contrastive learning, and homogeneous representation methods," achieving a dual improvement in prediction performance and generalization ability. The overall architecture of the model is as follows: Figure 3 As shown.

[0111] like Figure 3 As shown, in this embodiment, a five-dimensional heterogeneous feature system is constructed to address the differences in molecular characteristics between circular RNA (circRNA) and microRNA (miRNA), including Doc2Vec (semantic representation), Role2Vec (structural role representation), TPCP (trinucleotide physicochemical representation), RNAErnie (context-dependent representation), and sequence similarity (evolutionary conserved representation). This system comprehensively covers the core information of RNA molecules in terms of "semantics-structure-physicochemical-context-evolution," solving the problem that existing technologies only focus on three or fewer types of features and lack key biological context and structural role information.

[0112] Next, undirected edges can be constructed based on a threshold set according to the sequence similarity matrix. After generating homogeneous graphs for circRNA and miRNA respectively, the Graph Attention Network (GAT) is combined with multi-feature encoding depth: four features, Doc2Vec, Role2Vec, TPCP, and RNAErnie, are input into the GAT as node attributes. Through adaptive neighborhood attention weight allocation, key node information in the homogeneous graph is accurately aggregated to generate a structural embedding vector that integrates multi-scale features. This design differs from existing graph models that use "unified weight aggregation" or "fail to explicitly model node importance," significantly improving the utilization rate of structural information.

[0113] The MSCL-CMI model addresses the differences in characteristics between short circRNA sequences and long miRNA sequences by designing differentiated feature representation branches: the circRNA branch employs a depthwise separable convolution module to focus on deep modeling of key local motifs in short sequences; the miRNA branch introduces a selective kernel attention (SKAttention) mechanism to dynamically adapt to multi-scale global information in long sequences, achieving adaptive weighting of cross-channel features. This differentiated design solves the problem that existing technologies' "uniform feature fusion method" cannot adapt to the characteristics of the two types of RNA molecules, significantly improving the targeting and effectiveness of feature representation.

[0114] Finally, this embodiment proposes a two-stage contrastive learning strategy to achieve synergistic optimization of "single-molecule representation enhancement and precise capture of interaction patterns": The first stage is single-molecule hierarchical contrastive learning, which defines different feature views of the same RNA molecule as positive sample pairs and feature views of different RNA molecules as negative sample pairs, thereby enhancing the robustness and consistency of single-molecule representation and solving the limitation of existing contrastive learning that only focuses on cross-modal redundancy removal; The second stage is CMI interaction hierarchical contrastive constraint, which uses known interaction pairs as positive samples and pairs with no interaction or unclear relationship as negative samples, thereby enhancing the discriminativeness of interaction patterns through similarity modeling and making up for the lack of targeted optimization at the interaction level in existing models.

[0115] Ultimately, the circRNA-miRNA interaction features optimized by integrating structural embedding, differential branching representation, and two-stage comparative learning are input into a multilayer perceptron (MLP) to generate prediction results, comprehensively improving the model's accuracy and generalization ability in predicting interactions with conventional data, low-similarity RNAs, and newly discovered RNA molecules.

[0116] In one embodiment, the provided method for predicting circRNA and miRNA interactions by fusing multi-scale features and two-stage contrastive learning may further include a feature fusion and interaction prediction process. The specific process includes: concatenating the final circRNA representation and the final miRNA representation into vectors to obtain a fused feature vector; sequentially inputting the fused feature vector into a multilayer perceptron network containing fully connected layers, nonlinear activation layers, batch normalization layers, and Dropout layers to extract deep nonlinear features; and mapping the deep nonlinear features through a Sigmoid activation function to obtain the predicted probability that there is an interaction between circRNA and miRNA.

[0117] Specifically, computer devices can concatenate the embedding vectors of circRNA and miRNA and input them into a multilayer perceptron (MLP) to predict the interaction probability of each circRNA-miRNA pair: Where Wp represents the trainable parameters of the MLP, and σ represents the sigmoid activation function. Next, the interactive prediction task is trained under supervised conditions using binary cross-entropy loss: ;in, The actual label representing the interaction relationship (1 indicates that there is interaction, 0 indicates that there is no interaction).

[0118] In one embodiment, among the multidimensional heterogeneous features, Doc2Vec semantic features are used to capture global-level semantic information of RNA sequences; Role2Vec structural role features are used to encode the structural role information of RNA molecules in the interaction network through attribute random walks; TPCP trinucleotide physicochemical features are used to extract biophysical information related to local structural stability and stacking interactions at the trinucleotide level; RNAErnie context-dependent features are used to capture multi-level contextual dependencies of RNA sequences at the base level, subsequence level, and motif level based on the Transformer architecture; and sequence similarity features are calculated based on the Levenstein distance and used to quantify evolutionary conservation information between sequences.

[0119] In one embodiment, the provided method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning may further include a performance evaluation process, which specifically includes: acquiring a dataset and dividing the dataset into a validation set and a training set; performing five-fold cross-validation based on the validation set and the training set, and taking the average of the five validation results as the prediction performance result; determining the evaluation index, and using the evaluation index to evaluate the prediction performance result.

[0120] Computer equipment can be used to comprehensively evaluate the performance of the MSCL-CMI model using five-fold cross-validation and independent test sets. Evaluation metrics include area under the curve (AUC), area under the precision-recall curve (AUPR), accuracy (ACC), F1 score, recall, and precision, ensuring that the model has good generalization performance and robust predictive ability.

[0121] Specifically, computer equipment can use five-fold cross-validation to evaluate the predictive performance of circRNA-miRNA association (CMI). First, the dataset is divided into five equally sized subsets, each serving as the validation set sequentially, while the remaining four subsets are used as training data. The final performance is obtained by averaging the results of the five validations. The study uses the area under the curve (AUC), area under the precision-recall curve (AUPR), precision (ACC), F1 score, recall, and precision as evaluation metrics. The corresponding formulas for precision (ACC), precision, recall, and F1 score are defined as follows: ; ; ; The area under the receiver operating characteristic (AUC) curve measures the overall discriminative ability of a model at different classification thresholds. The area under the precision-recall curve (AUPR) focuses on the trade-off between precision and recall, and is particularly suitable for evaluating model performance on imbalanced datasets. Here, TP (true positive), TN (true negative), FP (false positive), and FN (false negative) represent the number of true positive, true negative, false positive, and false negative samples, respectively.

[0122] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0123] In one embodiment, to systematically and comprehensively evaluate the performance of the MSCL-CMI model, comparative experiments were conducted on three publicly available circRNA-miRNA interaction (CMI) datasets (CMI9589, CMI9905, and CMI20208). The MSCL-CMI model was systematically compared with six representative state-of-the-art methods (CA-CMA, SPGNN, GCNCMI, BGF-CMAP, NECMA, and MFERL). All methods were evaluated under uniform data partitioning and experimental settings to ensure the consistency and reproducibility of the comparative experiments.

[0124] The results of five-fold cross-validation for the seven methods on three datasets involve six core evaluation metrics: area under the curve (AUC), area under the precision-recall curve (AUPR), accuracy (ACC), F1 score, recall, and precision, as shown in the table below:

[0125]

[0126] On the CMI9589 dataset, the MSCL-CMI model performs excellently across all evaluation metrics, achieving an AUC of 97.00%, an AUPR of 96.97%, an ACC of 92.26%, an F1 score of 92.25%, and an accuracy of 92.34%, demonstrating a significant advantage in overall performance. Furthermore, experimental results on the CMI9905 and CMI20208 datasets also show that the MSCL-CMI model maintains stable and excellent performance across multiple key evaluation metrics, fully demonstrating its robustness and strong generalization ability.

[0127] Comparative experiments conducted on multiple datasets and evaluation metrics demonstrate that the MSCL-CMI model can more efficiently capture key features in circRNA-miRNA interactions, and its predictive performance and stability outperform existing mainstream methods. These results quantitatively validate the technological breakthrough achieved by the proposed model in the circRNA-miRNA interaction prediction task, providing a reliable methodological basis for subsequent bioinformatics research.

[0128] In one embodiment, to evaluate the model's sensitivity to key hyperparameter settings and further verify the stability of the MSCL-CMI method, the CMI9589 dataset was used as the experimental subject. The impact of different batch sizes, learning rates, and temperature parameters on the model's predictive performance in contrastive learning was systematically analyzed. Specifically, the batch sizes were set to 32, 64, 128, and 256; the learning rates were set to four gradients: 0.1, 0.01, 0.001, and 0.0001; and the contrastive learning temperature parameters were set to three values: 0.05, 0.1, and 0.15. All experiments were conducted under the same settings, with only the target parameter adjusted while all other parameters remained fixed.

[0129] like Figure 4 As shown in (a), different learning rates lead to differences in model performance. The model achieves optimal performance when the learning rate is set to 0.01. A learning rate that is too high will cause unstable model training, while a learning rate that is too low will reduce the convergence speed and thus affect the final performance of the model. Figure 4 (b) The results show that the performance of the MSCL-CMI model first improves and then stabilizes under different batch size settings. When the batch size is set to 128, the model achieves the best performance in all four evaluation metrics. Too large a batch size will lead to unstable gradient estimation, while too large a batch size will weaken the generalization ability of the model. Figure 4(c) shows that the temperature parameter in contrastive learning has a direct impact on model performance: the model has the best prediction effect when the temperature parameter is set to 0.1. This result indicates that a suitable temperature parameter helps to balance the smoothness of the similarity distribution between samples, thereby improving the quality of representation learning more efficiently; conversely, if the temperature parameter is too high or too low, it will limit the role of the contrastive learning module.

[0130] Parameter sensitivity analysis results show that the MSCL-CMI model exhibits good stability within a reasonable parameter range. The model achieves optimal overall prediction performance when the batch size is 128, the learning rate is 0.01, and the temperature parameter is 0.1. Based on these experimental results, this parameter combination will be used as the default configuration for all subsequent experiments.

[0131] In one embodiment, to systematically verify the unique contribution and complementary value of each feature in the five-dimensional heterogeneous feature system (Doc2Vec, Role2Vec, TPCP, RNAErnie, and sequence similarity), and to clarify its breakthrough advantages compared to existing feature systems, rigorous feature ablation experiments were conducted. The experiment was designed as follows: while maintaining the model architecture (GAT + differentiated representation branch + two-stage comparative learning) and the sequence similarity feature unchanged, the Doc2Vec, Role2Vec, TPCP, and RNAErnie features were removed one by one. The model's performance on the CMI9589, CMI9905, and CMI20208 datasets was systematically evaluated using six core metrics: area under the curve (AUC), area under the precision-recall curve (AUPR), accuracy (ACC), F1 score, recall, and precision. This comprehensively verified the necessity and superiority of multi-scale heterogeneous feature fusion.

[0132] like Figure 4 As shown, removing any single feature leads to a decline in all evaluation metrics to varying degrees, and the irreplaceability of each feature contrasts sharply with the limitations of existing technologies.

[0133] The most significant performance degradation of the model occurred after the removal of the RNAErnie feature. This feature, pre-trained based on the Transformer architecture, can capture the multi-level contextual dependencies of RNA sequences at the base, subsequence, and motif levels, and incorporate implicit biological prior information. However, most existing technologies do not introduce such contextual features and rely only on the surface information of the sequence itself, which makes it impossible to characterize the deep properties of RNA molecules.

[0134] After removing the Role2Vec feature, the model performance also showed a significant decline. This feature encodes the structural role of RNA molecules in the interaction network through attribute random walks, which can uncover potential functional associations between molecules. However, existing feature systems mostly focus on the sequence, structure and simple similarity of molecules themselves, lacking the capture of the key dimension of "network role", and are difficult to reflect the actual functional positioning of molecules in the interaction system.

[0135] After removing Doc2Vec or TPCP features, the model performance decreases relatively little, but there is still a clear attenuation. Doc2Vec provides global sequence-level semantic information, and TPCP extracts biophysical information related to the physicochemical properties of trinucleotides. The two supplement the feature dimensions from the semantic expression and physicochemical basis levels, respectively, complementing RNAErnie and Role2Vec. However, the feature systems of existing technologies generally have missing dimensions and cannot achieve comprehensive coverage of such multi-dimensional information.

[0136] Crucially, the complete model integrating all five features consistently outperformed existing technologies across all evaluation metrics across the three datasets, significantly surpassing the feature combination effects of current technologies. This result clearly demonstrates the innovation and superiority of the five-dimensional heterogeneous feature system of this invention: through a comprehensive feature design encompassing semantics, structure, physicochemical properties, context, and evolution, it not only compensates for the dimensional deficiencies of existing technologies but also avoids information redundancy through the complementarity of each feature, achieving a fusion effect of "1+1>2". Furthermore, this feature system deeply synergizes with the differentiated representation branch and two-stage comparative learning of this application—multi-dimensional features provide a rich and crucial input foundation for subsequent model learning, the differentiated branch achieves accurate feature modeling for the characteristics of two types of RNA molecules, and the two-stage comparative learning further strengthens the consistency of feature representation and the discriminative nature of interaction patterns. The synergy of these three elements enables the model to overcome the limitations of existing technologies' "single comparative learning + unified feature fusion," significantly improving prediction performance and generalization ability.

[0137] In one embodiment, to thoroughly verify the core contribution of the two-stage contrastive learning strategy to model performance and clarify its essential difference from existing single-dimensional contrastive learning (such as existing multimodal cross-feature contrastive learning), three rigorous ablation experiment settings were designed: while keeping all other components such as multi-scale feature encoding, differential representation branches (depth separable convolution / SKAttention), and graph attention network (GAT) unchanged, the first-stage single-molecule level contrastive learning, the second-stage CMI interaction level contrastive learning, and the contrastive learning of both stages were removed respectively. The unique value and synergistic effect of each stage were comprehensively evaluated through multiple datasets and multiple indicators.

[0138] like Figure 5As shown, the complete two-stage contrastive learning model achieves optimal results across all evaluation metrics (AUC, AUPR, ACC, F1-score, etc.) on the CMI9589, CMI9905, and CMI20208 datasets, significantly outperforming the three ablation variants. Specifically:

[0139] When the first stage of contrastive learning was removed, the model performance declined to some extent, with a more significant drop in metrics reflecting sample discrimination ability; however, the overall predictive performance remained at a high level. In contrast, removing the second stage of contrastive learning resulted in a more significant performance decline across multiple evaluation metrics, indicating that the second stage has a greater impact on model performance and plays a more crucial role in discriminative representation learning. When both stages of contrastive learning were removed simultaneously, the model performance declined further, reaching its lowest level across all evaluation metrics, fully demonstrating that the two-stage contrastive learning is the key innovation behind the model's performance breakthrough.

[0140] Therefore, the first stage addresses the issues of ambiguous single-molecule representation and insufficient robustness by comparing multiple feature views of the same RNA molecule. The second stage specifically enhances the discriminative power of interaction patterns by comparing known interacting pairs with non-interacting pairs, creating a complementary gain. This design is the first to deeply integrate molecular-level feature optimization with interaction-level discriminative learning, enabling the model to simultaneously capture both "single-molecule core features" and "interaction-specific patterns," ultimately achieving a dual improvement in predictive performance and generalization ability.

[0141] In one embodiment, to verify the effectiveness of introducing Graph Attention Networks (GAT) into the MSCL-CMI model, the GAT-based architecture is systematically compared with several representative graph neural network methods, including Graph Convolutional Networks (GCN), Graph Sampling Aggregation Networks (GraphSAGE), and Graph Isomorphism Networks (GIN), which correspond to different neighborhood information aggregation strategies. All models are evaluated under consistent experimental settings, differing only in the type of graph neural network architecture used for structural feature learning. Furthermore, to evaluate the contribution of graph structural information to the CMI prediction task, a baseline model without graph modeling (denoted as the non-graph model) is introduced.

[0142] Figure 6The performance comparison of the MSCL-CMI model across three datasets with different graph neural network architectures is presented, visualized using heatmaps. All graph-based models significantly outperform the non-graph baseline model, indicating that constructing graph structures based on the similarity of circular RNA (circRNA) and microRNA (miRNA) can capture potential molecular associations, thereby improving interaction prediction performance. Compared to other comparative methods, the GAT-based MSCL-CMI model performs best on all evaluation metrics. Compared to GCN and GraphSAGE, GAT introduces an attention mechanism, which can assign adaptive importance weights to different neighborhood nodes, enabling the model to focus on neighborhood nodes that are more critical to the representation of the target node. In contrast, GCN aggregates neighborhood information with uniform weights, making it difficult to distinguish the relative importance of different neighborhood nodes; while GraphSAGE, although enhancing inductive ability through sampling and aggregation strategies, usually relies on fixed or simple aggregation functions, lacking fine-grained modeling of neighborhood heterogeneity. While GIN demonstrates strong representational capabilities through its multilayer perceptron (MLP)-based aggregation approach, it also fails to explicitly model the importance of neighboring nodes, which limits its ability to capture heterogeneous neighborhood information. Therefore, in similarity-based circRNA and miRNA maps, GAT can more effectively model the differentiated contributions of neighboring nodes to representation learning, thereby achieving superior prediction performance.

[0143] In summary, the graph attention mechanism demonstrates its advantages in capturing information from heterogeneous neighborhoods, and confirms that the GAT-based graph modeling design makes a significant contribution to improving the prediction performance of the MSCL-CMI model.

[0144] In one embodiment, imbalanced distribution of positive and negative samples in the training dataset is a common and unavoidable problem in circRNA-miRNA interaction (CMI) prediction tasks. Due to the high cost of experimentally validating CMIs, the number of known interaction pairs remains limited; furthermore, a large number of potential non-interacting samples are difficult to validate experimentally, leading to significant imbalance in positive and negative samples in real-world applications. Therefore, the generalization ability and robustness of the model under different positive and negative sample ratios deserve close attention.

[0145] Using the CMI9589 dataset as the experimental subject, five-fold cross-validation experiments were conducted on training sets with different positive-to-negative sample ratios (1:1, 1:2, 1:5, and 1:10). All experiments used the same feature encoding method and model parameter settings to ensure fairness in the comparison. As shown in the table below, the model exhibited the smallest overall fluctuation in the AUC metric under different positive-to-negative sample ratios, indicating that the model can maintain stable classification performance even when the class distribution changes.

[0146]

[0147] As the proportion of negative samples gradually increases, the model's performance on evaluation metrics, which are highly sensitive to class imbalance, declines to some extent, especially in highly imbalanced scenarios (1:5 and 1:10), where its ability to identify positive samples slightly diminishes. When the positive-to-negative sample ratio is 1:1, given the parameter settings, the model exhibits optimal or near-optimal performance on all evaluation metrics, demonstrating not only good generalization ability but also strong predictive performance. In the current CMI prediction task, maintaining a relatively balanced distribution of positive and negative samples is more conducive to the model fully learning interaction patterns, thereby improving overall predictive performance. This experiment ultimately adopts a 1:1 positive-to-negative sample ratio as the default core setting for model training and evaluation.

[0148] In one embodiment, to further verify the feature visualization separability achieved by the MSCL-CMI model in the circRNA-miRNA interaction (CMI) prediction task, the t-distributed random neighborhood embedding (t-SNE) algorithm is used to project the high-dimensional representation of circRNA-miRNA pairings into a low-dimensional space, thereby enabling a comparative analysis of feature distributions before and after model training. As a widely used nonlinear dimensionality reduction technique, t-SNE can effectively preserve the local structural features of samples in high-dimensional space, thus intuitively reflecting the distribution of different categories of samples in the feature space.

[0149] Specifically, for the three datasets, t-SNE dimensionality reduction was performed on the original feature representations before model training and the high-level feature representations learned after training, and the positive and negative samples were visualized in a two-dimensional space. Figure 7 As can be seen from (a), (b), and (c), before model training, positive and negative samples are mixed together in the feature space, lacking clear class boundaries. This phenomenon indicates that relying solely on initial feature representations, the distinguishing power between interacting and non-interacting pairs in circRNA-miRNA interaction (CMI) is low, the discriminative power of the features themselves is limited, and it is difficult to effectively capture the essential differences between the two classes of samples. In contrast, after model training (e.g., Figure 7 As shown in (d), (e), and (f), the distribution patterns of positive and negative samples in the low-dimensional space change significantly: the two classes of samples gradually exhibit a more clustered and relatively separated distribution trend, and the boundaries between categories become clearer. Although some overlap still exists in local areas, overall, positive and negative samples can be distinguished more effectively. The MSCL-CMI model effectively integrates multi-scale feature information during training and gradually extracts discriminative patterns related to CMI through deep representation learning, thereby significantly improving the discriminative ability of feature representation.

[0150] In summary, the t-SNE visualization results provide an intuitive validation of the effectiveness of the MSCL-CMI model in feature representation learning, which is consistent with the performance improvement observed in the quantitative evaluation metrics.

[0151] In one embodiment, such as Figure 8 As shown, a circRNA and miRNA interaction prediction system integrating multi-scale features and two-stage contrastive learning is provided, including: a feature encoding module 810, a feature learning module 820, a feature enhancement module 830, a contrastive learning module 840, and a prediction module 850, wherein:

[0152] The feature encoding module 810 is used to acquire the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA sequence data and miRNA sequence data respectively. The multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features.

[0153] The feature learning module 820 is used to construct circRNA homogeneity maps and miRNA homogeneity maps based on sequence similarity features. It uses Doc2Vec features, Role2Vec features, TPCP features and RNAErnie features as node attributes, and generates circRNA structure embedding vectors and miRNA structure embedding vectors with fused structural information through a graph attention network.

[0154] The feature enhancement module 830 is used to input the circRNA structure embedding vector into a depthwise separable convolutional network to extract local motif features, and to input the miRNA structure embedding vector into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations.

[0155] The contrastive learning module 840 is used to perform single-molecule-level contrastive learning based on the enhanced circRNA and miRNA representations to optimize the consistency of single-molecule representations; and to perform interaction-level contrastive learning to optimize the discriminativeness of interaction patterns, thereby obtaining the final circRNA and final miRNA representations.

[0156] The prediction module 850 is used to fuse the final circRNA characterization and the final miRNA characterization, and input the fused characterization data into the multilayer perceptron network to obtain the prediction result that there is an interaction between circRNA and miRNA.

[0157] In one embodiment, the feature encoding module 810 is further configured to convert the nucleotide sequences of circRNA and miRNA sequence data into word sequences by k-mer segmentation, input the word sequences into the Doc2Vec model to obtain Doc2Vec semantic features; perform attribute random walks based on sequence similarity networks and learn the structural role embedding vectors of each node to obtain Role2Vec structural role features; traverse circRNA and miRNA sequence data to generate trinucleotide fragments, and match TPCP trinucleotide physicochemical features based on the physicochemical features of the trinucleotide fragments; input circRNA and miRNA sequence data into a pre-trained Transformer model, capture long-distance contextual dependencies in the sequences through a multi-layer multi-head self-attention mechanism to obtain RNAErnie contextual dependency features; calculate the minimum edit distance between two RNA sequences using Levenstein distance, and convert it into a similarity score after normalization to obtain sequence similarity features.

[0158] In one embodiment, the feature learning module 820 is further configured to calculate pairwise similarity matrices of the circRNA molecule set and the miRNA molecule set based on sequence similarity features, and construct undirected edges based on the similarity matrices with a preset similarity threshold to obtain circRNA homogeneous graphs and miRNA homogeneous graphs respectively; using Doc2Vec features, Role2Vec features, TPCP features and RNAErnie features as node attributes, and performing neighborhood information aggregation on the circRNA homogeneous graph and miRNA homogeneous graph through a graph attention network to generate corresponding circRNA structure embedding vectors and miRNA structure embedding vectors that integrate structural information.

[0159] In one embodiment, the feature enhancement module 830 is further configured to embed the circRNA structure into a vector input depthwise separable convolutional network, extract local motif features through depthwise convolution in the depthwise separable convolutional network, fuse cross-channel information through pointwise convolution, and obtain an enhanced circRNA representation; embed the miRNA structure into a vector input selective kernel attention network, extract feature maps of different spatial scales through scale convolution branch splitting, and obtain an enhanced miRNA representation through global information fusion and adaptive weight selection.

[0160] In one embodiment, the contrast learning module 840 is further configured to use the circRNA structure embedding vector and the enhanced circRNA representation as different feature views of the same circRNA molecule; use the different feature view pairs of the same circRNA molecule as the first positive sample pair and the feature view pairs of different circRNA molecules as the first negative sample pair; based on the first positive sample pair and the first negative sample pair, use cosine similarity as a similarity measure and introduce a temperature parameter to regulate the smoothness of the similarity distribution, construct a circRNA single-molecule contrast loss function, and complete the single-molecule hierarchical contrast learning of circRNA representation; use the miRNA structure embedding vector and the enhanced miRNA representation as different feature views of the same miRNA molecule; use the different feature view pairs of the same miRNA molecule as the second positive sample pair and the feature view pairs of different miRNA molecules as the second negative sample pair; based on the second positive sample pair and the second negative sample pair, use cosine similarity as a similarity measure and introduce a temperature parameter to regulate the smoothness of the similarity distribution, construct a miRNA single-molecule contrast loss function, and complete the single-molecule hierarchical contrast learning of miRNA representation.

[0161] In one embodiment, the contrastive learning module 840 is further configured to treat interacting circRNA-miRNA pairs as positive samples and non-interacting circRNA-miRNA pairs as negative samples, and uniformly sample positive and negative samples within a batch; calculate the similarity of each sampled circRNA-miRNA pair in the final representation space using cosine similarity, construct an interaction-level contrastive loss function, and complete interaction-level contrastive learning; perform a weighted summation of the circRNA single-molecule contrastive loss function, the miRNA single-molecule contrastive loss function, and the interaction-level contrastive loss function to obtain the final loss function; and obtain the final circRNA representation and the final miRNA representation based on the final loss function.

[0162] In one embodiment, the prediction module 850 is further configured to concatenate the final circRNA representation and the final miRNA representation to obtain a fused feature vector; input the fused feature vector sequentially into a multilayer perceptron network containing a fully connected layer, a nonlinear activation layer, a batch normalization layer and a Dropout layer to extract deep nonlinear features; and map the deep nonlinear features through a Sigmoid activation function to obtain the predicted probability that there is an interaction between circRNA and miRNA.

[0163] In one embodiment, among the multidimensional heterogeneous features, Doc2Vec semantic features are used to capture global-level semantic information of RNA sequences; Role2Vec structural role features are used to encode the structural role information of RNA molecules in the interaction network through attribute random walks; TPCP trinucleotide physicochemical features are used to extract biophysical information related to local structural stability and stacking interactions at the trinucleotide level; RNAErnie context-dependent features are used to capture multi-level contextual dependencies of RNA sequences at the base level, subsequence level, and motif level based on the Transformer architecture; and sequence similarity features are calculated based on the Levenstein distance and used to quantify evolutionary conservation information between sequences.

[0164] In one embodiment, the system further includes an evaluation module for acquiring a dataset and dividing the dataset into a validation set and a training set; performing five-fold cross-validation based on the validation set and the training set, taking the average of the five validation results as the prediction performance result; determining an evaluation metric, and using the evaluation metric to evaluate the prediction performance result.

[0165] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a circRNA and miRNA interaction prediction method that integrates multi-scale features and two-stage contrastive learning. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0166] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0167] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a circRNA and miRNA interaction prediction method that integrates multi-scale features with two-stage contrastive learning.

[0168] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement steps of a circRNA and miRNA interaction prediction method that integrates multi-scale features and two-stage contrastive learning.

[0169] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0170] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0171] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A circRNA and miRNA interaction prediction method fusing multi-scale features and two-stage contrastive learning, characterized in that, The method includes: Obtain the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA sequence data and miRNA sequence data respectively. The multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features. Based on the sequence similarity features, construct circRNA homogeneity maps and miRNA homogeneity maps, and use the Doc2Vec feature, Role2Vec feature, TPCP feature and RNAErnie feature as node attributes. Generate circRNA structure embedding vectors and miRNA structure embedding vectors with fused structural information through graph attention network. The circRNA structure embedding vector is input into a depthwise separable convolutional network to extract local motif features, and the miRNA structure embedding vector is input into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations. Based on the enhanced circRNA and miRNA representations, single-molecule hierarchical contrastive learning is performed to optimize the consistency of single-molecule representations; interaction hierarchical contrastive learning is performed to optimize the discriminativeness of interaction patterns, resulting in the final circRNA and final miRNA representations. The final circRNA characterization and the final miRNA characterization are fused, and the fused characterization data is input into a multilayer perceptron network to obtain a prediction result that there is an interaction between circRNA and miRNA.

2. The method for predicting circRNA-miRNA interaction by fusing multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, Multidimensional heterogeneous features were extracted from the circRNA sequence data and miRNA sequence data, respectively, including: The nucleotide sequences of the circRNA and miRNA sequence data are split into word sequences using k-mer, and the word sequences are input into the Doc2Vec model to obtain Doc2Vec semantic features. Based on the sequence similarity network, an attribute random walk is performed, and the structural role embedding vector of each node is learned to obtain the Role2Vec structural role features; Trinucleotide fragments are generated by traversing the circRNA sequence data and miRNA sequence data. Based on the physicochemical characteristics of the trinucleotide fragments, the physicochemical characteristics of TPCP trinucleotides are matched. The circRNA and miRNA sequence data are input into a pre-trained Transformer model, and long-distance contextual dependencies in the sequences are captured through a multi-layer multi-head self-attention mechanism to obtain RNAErnie context-dependent features. The minimum edit distance between two RNA sequences was calculated using the Levenstein distance, and after normalization, it was converted into a similarity score to obtain sequence similarity features. 3.The circRNA-miRNA interaction prediction method of fusing multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, Based on the sequence similarity features, circRNA and miRNA homogeneity maps are constructed. The Doc2Vec, Role2Vec, TPCP, and RNAErnie features are used as node attributes. A graph attention network is then used to generate circRNA and miRNA structure embedding vectors that fuse structural information, including: Based on the sequence similarity features, pairwise similarity matrices of the circRNA molecule set and the miRNA molecule set are calculated respectively, and undirected edges are constructed based on the similarity matrices with a preset similarity threshold to obtain the circRNA homogeneity map and the miRNA homogeneity map respectively. The Doc2Vec, Role2Vec, TPCP, and RNAErnie features are used as node attributes. A graph attention network is used to aggregate neighborhood information on the circRNA and miRNA homogeneity maps, thereby generating circRNA and miRNA structure embedding vectors that incorporate structural information.

4. The method for predicting circRNA and miRNA interactions by integrating multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, The circRNA structure embedding vector is input into a depthwise separable convolutional network to extract local motif features, and the miRNA structure embedding vector is input into a selective kernel attention network to extract multi-scale global features, thus obtaining enhanced circRNA and miRNA representations, including: The circRNA structure is embedded into a vector input into a depthwise separable convolutional network. Local motif features are extracted through depthwise convolution in the depthwise separable convolutional network, and cross-channel information is fused through pointwise convolution to obtain an enhanced circRNA representation. The miRNA structure is embedded into a vector input selective kernel attention network. By splitting the network through scale convolution branches, feature maps of different spatial scales are extracted. After global information fusion and adaptive weight selection, the enhanced miRNA representation is obtained.

5. The method for predicting circRNA and miRNA interactions by fusing multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, Based on the enhanced circRNA and miRNA characterizations, single-molecule hierarchical contrastive learning is performed to optimize the consistency of single-molecule characterizations, including: The circRNA structure embedding vector and the enhanced circRNA representation are used as different feature views of the same circRNA molecule; Different feature view pairs of the same circRNA molecule are used as the first positive sample pair, and feature view pairs of different circRNA molecules are used as the first negative sample pair; Based on the first positive sample pair and the first negative sample pair, cosine similarity is used as the similarity measure, and a temperature parameter is introduced to regulate the smoothness of the similarity distribution. A single-molecule contrastive loss function for circRNA is constructed to complete the single-molecule hierarchical contrastive learning of circRNA representation. The miRNA structure embedding vector and the enhanced miRNA characterization are used as different feature views of the same miRNA molecule; Different feature view pairs of the same miRNA molecule are used as the second positive sample pairs, and feature view pairs of different miRNA molecules are used as the second negative sample pairs. Based on the second positive sample pair and the second negative sample pair, cosine similarity is used as the similarity measure, and a temperature parameter is introduced to regulate the smoothness of the similarity distribution. A single-molecule contrastive loss function for miRNA is constructed to complete the single-molecule hierarchical contrastive learning of miRNA characterization.

6. The method for predicting circRNA and miRNA interactions by fusing multi-scale features and two-stage contrastive learning according to claim 5, characterized in that, Based on the enhanced circRNA and miRNA characterizations, interaction hierarchy contrastive learning is performed to optimize the discriminative power of interaction patterns, including: Positive samples are selected from circRNA-miRNA pairs that have interaction, and negative samples are selected from circRNA-miRNA pairs that do not have interaction. Positive and negative samples are sampled evenly within a batch. Cosine similarity was used to calculate the similarity of each circRNA-miRNA pair in the final representation space, and an interaction-level contrastive loss function was constructed to complete the interaction-level contrastive learning. The weighted sum of the circRNA single-molecule contrast loss function, miRNA single-molecule contrast loss function, and interaction hierarchy contrast loss function is used to obtain the final loss function. Based on the final loss function, the final circRNA characterization and the final miRNA characterization are obtained.

7. The method for predicting circRNA and miRNA interactions by fusing multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, The final circRNA characterization and the final miRNA characterization are fused, and the fused characterization data is input into a multilayer perceptron network to obtain prediction results of interactions between circRNAs and miRNAs, including: The final circRNA representation and the final miRNA representation are concatenated to obtain the fused feature vector. The fused feature vector is sequentially input into a multilayer perceptron network containing fully connected layers, nonlinear activation layers, batch normalization layers, and Dropout layers to extract deep nonlinear features. The deep nonlinear features are mapped using the Sigmoid activation function to obtain the predicted probability of an interaction between circRNA and miRNA.

8. The method for predicting circRNA and miRNA interactions by fusing multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, Among the multidimensional heterogeneous features, the Doc2Vec semantic features are used to capture global-level semantic information of RNA sequences; the Role2Vec structural role features are used to encode the structural role information of RNA molecules in the interaction network through attribute random walks; the TPCP trinucleotide physicochemical features are used to extract biophysical information related to local structural stability and stacking interactions at the trinucleotide level; the RNAErnie context-dependent features are used to capture multi-level contextual dependencies of RNA sequences at the base, subsequence, and motif levels based on the Transformer architecture; and the sequence similarity features are based on Levenstein distance calculations and are used to quantify evolutionary conservation information between sequences.

9. The method for predicting circRNA and miRNA interactions by fusing multi-scale features and two-stage contrastive learning according to claim 1, characterized in that, The method further includes: Obtain the dataset and divide it into a validation set and a training set; Five-fold cross-validation is performed based on the validation set and training set, and the average of the five validation results is taken as the prediction performance result. Determine the evaluation metrics and use the evaluation metrics to evaluate the prediction performance results.

10. A circRNA and miRNA interaction prediction system integrating multi-scale features and two-stage contrastive learning, characterized in that, The system includes: The feature encoding module is used to acquire the circRNA sequence data and miRNA sequence data to be predicted, and extract multi-dimensional heterogeneous features from the circRNA sequence data and miRNA sequence data respectively. The multi-dimensional heterogeneous features include Doc2Vec semantic features, Role2Vec structural role features, TPCP trinucleotide physicochemical features, RNAErnie context-dependent features, and sequence similarity features. The feature learning module is used to construct circRNA homogeneity maps and miRNA homogeneity maps based on the sequence similarity features. The Doc2Vec feature, Role2Vec feature, TPCP feature and RNAErnie feature are used as node attributes. The graph attention network is used to generate circRNA structure embedding vectors and miRNA structure embedding vectors that fuse structural information, respectively. The feature enhancement module is used to input the circRNA structure embedding vector into a depthwise separable convolutional network to extract local motif features, and to input the miRNA structure embedding vector into a selective kernel attention network to extract multi-scale global features, thereby obtaining enhanced circRNA and miRNA representations. The contrastive learning module is used to perform single-molecule-level contrastive learning based on the enhanced circRNA and miRNA representations to optimize the consistency of single-molecule representations; and to perform interaction-level contrastive learning to optimize the discriminativeness of interaction patterns, thereby obtaining the final circRNA and final miRNA representations. The prediction module is used to fuse the final circRNA characterization and the final miRNA characterization, and input the fused characterization data into a multilayer perceptron network to obtain a prediction result that there is an interaction between circRNA and miRNA.

Images