Method for predicting circrna-mirna interactions in zero-shot scenarios

Abstract

The application provides a method for predicting circRNA-miRNA interaction in a zero-shot scenario, comprising: obtaining circRNA sequences, miRNA sequences, and a disease correlation network of the two; respectively using different pre-training language models based on a Transformer architecture to extract deep semantic representations from the sequences; simultaneously, based on a diffusion-wave multi-hop topology modeling mechanism, extracting multi-hop topology features from the disease correlation network; fusing the semantic representations and the topology features to obtain a fusion feature vector; finally, using a classifier to predict the interaction. The application combines deep semantic coding with multi-hop structure modeling, and uses a zero-shot dataset and a negative sample construction strategy based on subcellular localization, effectively solving the problem of insufficient generalization ability of traditional methods in predicting new biological molecule interactions, significantly improving the prediction accuracy and reliability in a zero-shot scenario.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and computational biology, specifically providing a method for predicting circRNA-miRNA interactions in zero-sample scenarios. Background Technology

[0002] Ribonucleic acid (RNA), as one of the earliest molecular forms to emerge in life, participates in and regulates multidimensional fundamental life processes. Besides transcripts responsible for encoding proteins, a significant portion of the transcriptome consists of non-coding RNAs (ncRNAs). These molecules, while not translated into proteins, perform diverse regulatory functions within the cell. Among them, circular RNAs (circRNAs) possess a covalently closed circular structure, making them highly stable and resistant to exonuclease degradation; microRNAs (miRNAs), short-chain RNAs of approximately 20–24 nucleotides in length, primarily participate in gene expression regulation through post-transcriptional regulatory mechanisms. Notably, the 2024 Nobel Prize in Physiology or Medicine was awarded to Victor Ambros and Gary Ruvkun for their pioneering contributions to revealing the crucial role of miRNAs in gene regulation, further highlighting the profound biological and clinical significance of this type of RNA.

[0003] Mounting evidence suggests that circRNAs can act as "sponges" for competitive endogenous RNAs (ceRNAs) or miRNAs, thereby modulating the latter's activity. These circRNA–miRNA interactions (CMIs) play crucial roles in various pathological processes, including tumorigenesis, neurological disorders, and cardiovascular diseases. Therefore, accurate identification of CMIs is essential for elucidating disease mechanisms and offers significant potential for discovering novel therapeutic targets. However, traditional experimental methods are time-consuming, costly, and have limited scalability, making the development of efficient computational prediction methods an urgent need.

[0004] Despite progress in computational prediction of CMIs, several significant challenges remain. Most existing models assume that training and test data come from the same distribution, with the test set typically containing both known and unknown interactions. While this setup facilitates benchmark performance evaluation, it can lead to an overestimation of predictive power, as the model can "memorize" information from partially established associations, thus affecting its true generalization ability. This is particularly true for previously uncharacterized circRNAs or miRNAs, whose interaction discovery is a typical zero-shot prediction task, and traditional methods often fail to effectively extrapolate to such novel biological entities.

[0005] Secondly, many existing computational methods fail to fully utilize the inherent properties of biomolecules themselves. Both circRNAs and miRNAs carry rich biochemical and sequence-level information, serving as important domain priors. However, most current models over-rely on network connectivity patterns or structural features based on similarity, neglecting the semantic information contained within the sequence. This inefficient use of molecular properties limits the model's ability to make predictions beyond the observed network.

[0006] Furthermore, structured modeling of molecular interaction networks remains insufficient. Existing methods treat circRNA–miRNA bipartite graphs as simple adjacency structures, neglecting key structural features such as higher-order subgraph similarities and multi-hop association patterns. Local patterns, such as shared regulatory partners and community-like structures, can provide important clues for predicting unknown interactions. The lack of effective modeling of such graph structural information leads to insufficient utilization of network topology, thereby weakening the model's predictive ability in complex biological contexts. Summary of the Invention

[0007] To overcome the aforementioned shortcomings, in a first aspect, the present invention provides a method for predicting circRNA-miRNA interactions in zero-sample scenarios, comprising: acquiring circRNA sequences, miRNA sequences, and circRNA-disease association networks and miRNA-disease association networks; extracting deep semantic representations from circRNA and miRNA sequences based on a Transformer-based pre-trained language model; extracting multi-hop topological features from the circRNA-disease association networks and miRNA-disease association networks based on a diffusion-wavelet multi-hop topological modeling mechanism; fusing the deep semantic representations and the multi-hop topological features to obtain a fused feature vector; and inputting the fused feature vector into a classifier to predict whether there is an interaction between the circRNA and miRNA.

[0008] Preferably, the pre-trained language model based on Transformer extracts deep semantic representations from circRNA and miRNA sequences, specifically including: using a first semantic encoder based on the Transformer architecture to encode the circRNA sequence to obtain a circRNA semantic embedding vector; using a second semantic encoder based on the Transformer architecture to encode the miRNA sequence to obtain a miRNA semantic embedding vector; wherein the first semantic encoder and the second semantic encoder are different pre-trained language models.

[0009] Preferably, the first semantic encoder is a BERT model, used to capture long-range dependent semantics in long circRNA sequences; the second semantic encoder is an ELECTRA model, used to capture key motif semantics in short miRNA sequences.

[0010] Preferably, based on the diffusion-wavelet multi-hop topology modeling mechanism, multi-hop topological features are extracted from the circRNA-disease association network and the miRNA-disease association network. Specifically, this includes: constructing a heterogeneous association graph based on the circRNA-disease association network and the miRNA-disease association network and calculating its normalized Laplacian matrix; calculating the diffusion wavelet coefficient matrix based on the hot kernel filter and the eigenvalue decomposition of the Laplacian matrix to capture the multi-scale topological structure; calculating the multi-hop topological similarity between nodes based on the diffusion wavelet coefficient matrix using the minimum pairwise difference metric; and weighting and aggregating the local features of nodes within the k-hop neighborhood subgraph based on the topological similarity to generate the multi-hop topological features.

[0011] Preferably, the local features of the nodes include the node degree after logarithmic transformation and the clustering coefficient.

[0012] Preferably, during the model training phase, the training data used satisfies the zero-sample setting, that is, the circRNA and miRNA nodes in the training set have no overlap with the circRNA and miRNA nodes in the test set.

[0013] Preferably, the construction of negative samples in the training data includes screening using the predicted subcellular localization information of circRNA and miRNA to exclude candidate negative sample pairs located in the same subcellular compartment.

[0014] Preferably, the classifier is a multilayer perceptron.

[0015] In a second aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described thereon.

[0016] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described thereon.

[0017] The beneficial effects of this invention are as follows: 1. Overcoming the generalization bottleneck of zero-shot prediction: By employing a strict "zero-shot" data strategy (the molecules in the training and test sets have no overlap) and constructing an evaluation dataset specifically for this scenario, the model is forced to learn generalizable deep patterns rather than memorizing known associations. This enables the invention to reliably predict the most valuable and novel circRNA-miRNA interactions in biological research, extending the applicability of computational models from "association reasoning" to true "discovery prediction".

[0018] 2. Achieved deep complementarity and synergistic enhancement of semantic and structural information: The model does not simply concatenate two types of features, but extracts high-order structural dependencies of the network through an original "diffusion-wavelet multi-hop topology modeling" module, organically fusing them with Transformer-based deep sequence semantic features. This fusion mechanism enables the model to possess both "molecular ontology understanding ability" (derived from the sequence) and "network context reasoning ability" (derived from the structure). In zero-shot scenarios, semantic information provides strong priors on inherent molecular properties; while in scenarios with rich associations, topological structural information can reveal more complex regulatory patterns. The two work synergistically, significantly improving the model's robustness and accuracy under different data conditions.

[0019] 3. A targeted molecular semantic encoding optimization strategy was proposed: Addressing the biological characteristics of long circRNA sequences and short, conserved miRNA sequences, different pre-trained Transformer architectures (BERT and ELECTRA, respectively) were innovatively adapted. This heterogeneous encoding strategy fully leverages the technical advantages of different models, achieving efficient and accurate capture of contextual dependencies in long sequences and key motifs in short sequences, thus improving the quality of semantic features from the source.

[0020] 4. Improved the biological rationality and interpretability of computational predictions: At the data level, "subcellular localization consistency" was introduced as a constraint for negative sample selection, preventing biologically impossible molecular pairs from being used as negative examples, reducing data noise, and making the patterns learned by the model closer to real biological laws. Simultaneously, the interactions predicted by the model can be traced back to their source through semantic features (such as sequence motifs) and structural features (such as network similarity), providing valuable clues and directions for subsequent biological experimental verification. Attached Figure Description

[0021] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1This is a schematic diagram of an architecture for predicting circRNA-miRNA interactions in a zero-sample scenario according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of a method for predicting circRNA-miRNA interactions in a zero-sample scenario according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the BERT structure according to an embodiment of the present invention; Figure 4 This is a diagram of the ELECTRA architecture according to an embodiment of the present invention; Figure 5 The ROC and PR curves of the method according to one embodiment of the present invention under five-fold cross-validation are shown. (a) Performance on the CMI-Zero dataset; (b) Performance on the CMI-9589 dataset; (c) Performance on the CMI-9905 dataset; Figure 6 This is a performance comparison of the method of this invention and its ablation variant on three benchmark datasets using five-fold cross-validation, according to one embodiment of the invention. Figure 7 This invention provides a performance comparison of the method of this embodiment with the Cross-Semantics variant on the CMI-Zero, CMI-9589, and CMI-9905 datasets using five-fold cross-validation. Figure 8 This invention provides a performance comparison of the method described in one embodiment with a baseline classifier. Detailed Implementation

[0022] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] Example 1 like Figure 1-2 As shown, this invention provides a method for predicting circRNA-miRNA interactions in a zero-sample scenario, comprising: Step S1: Obtain circRNA sequences, miRNA sequences, and circRNA-disease association networks and miRNA-disease association networks.

[0024] Step S2: Extract deep semantic representations from circRNA and miRNA sequences using a pre-trained language model based on Transformer.

[0025] In this embodiment, to fully utilize the domain prior knowledge inherent in molecules, this method introduces a Transformer-based pre-trained language model to extract deep semantic representations from circRNA and miRNA sequences. Such models were originally used in natural language processing, and by treating nucleotides as basic tokens in "molecular language," they have been extended to the field of biological sequence analysis. By contextually encoding nucleotide sequences, the model can not only identify local motif-like features but also uncover higher-order semantic dependencies reflecting molecular function.

[0026] Step S3: Based on the diffusion-wavelet multi-hop topology modeling mechanism, extract multi-hop topology features from the circRNA-disease association network and miRNA-disease association network.

[0027] In this embodiment, although semantic embedding can capture the intrinsic biological information of molecules from nucleotide sequences, relying solely on semantic features is still insufficient to fully characterize the higher-order structural dependencies in the circRNA–disease and miRNA–disease association networks. Therefore, this method introduces a diffusion-wavelet technique inspired by spectral theory to model multi-hop topological features.

[0028] Furthermore, in the diffusion-wavelet modeling process, the thermal diffusion time was set to 2.5, the number of feature evaluation points was set to 15, and the maximum θ value was set to 2.5. The hop count for adjacency matrix propagation was set to 1 by default; this parameter is crucial to the model, and its impact will be discussed further in the subsequent section on multi-hop propagation analysis. The input graph was constructed from the association data of circRNA-disease and miRNA-disease, and each node was characterized using a logarithmically transformed degree and clustering coefficient.

[0029] Step S4: Fuse the deep semantic representation with the multi-hop topological features to obtain a fused feature vector.

[0030] Step S5: Input the fused feature vector into the classifier to predict whether there is an interaction between the circRNA and miRNA.

[0031] In one embodiment, a Transformer-based pre-trained language model extracts deep semantic representations from circRNA and miRNA sequences. Specifically, this includes: using a first semantic encoder based on a Transformer architecture to encode the circRNA sequence to obtain a circRNA semantic embedding vector; and using a second semantic encoder based on a Transformer architecture to encode the miRNA sequence to obtain a miRNA semantic embedding vector. The first and second semantic encoders are different pre-trained language models. Specifically, the first semantic encoder is a BERT model used to capture long-range dependency semantics in long circRNA sequences; and the second semantic encoder is an ELECTRA model used to capture key motif semantics in short miRNA sequences.

[0032] Specifically, semantic encoding of circRNA sequences based on BERT.

[0033] circRNAs are typically long, ranging from hundreds to thousands of nucleotides, forming long-range dependencies and complex secondary structures, making it difficult for traditional similarity measures to effectively capture their sequence features. To address this, BERT is employed, which uses a masked language model (MLM) pre-training mechanism to achieve bidirectional context-aware encoding. Specifically, BERT employs a multi-layer bidirectional Transformer architecture, with each layer consisting of a self-attention mechanism and a feedforward neural network to extract contextual dependencies within the sequence.

[0034] like Figure 3 As shown, the circRNA sequence is first converted into tokens at the input stage and represented by three types of embeddings: Token Embedding maps each nucleotide to a dense vector; Position Embedding encodes the sequence information of the token; and Segment Embedding is used to distinguish different sequence fragments. The sum of these three types of embeddings constitutes the input representation of the model, which is then fed into stacked Transformer layers to generate rich semantic features.

[0035] BERT structure: Formally, given a segmented circRNA sequence as input... BERT maps tokens into embedding sequences using token embedding matrices, segment embeddings, and positional encoding. : (1) in, This represents the one-hot encoded form of a nucleotide. Nucleotide vocabulary The embedding matrix, This indicates positional embedding.

[0036] BERT's hidden layers consist of multiple Transformer encoders, capturing contextual dependencies through a self-attention mechanism. For the... Layers, their hidden representations Updated using Multi-Head Self-Attention (MHSA) and Feedforward Network (FFN): (2) in, The multi-head self-attention (MHSA) mechanism calculates attention as follows: (3) The query, key, and value matrices are all obtained through linear mapping: (4) By stacking Layer encoders can obtain contextualized representations. For the semantic representation of circRNAs, average pooling is used to aggregate sequence embeddings: (5) ELECTRA-based miRNA sequence semantic encoding: Compared to circRNAs, miRNAs have shorter sequences, but their structures are concise and highly conserved, containing key sequence motifs that determine their regulatory functions. To efficiently capture these features, the ELECTRA model is employed, which introduces a generator-discriminator pre-training paradigm, where the discriminator's task is to distinguish between real nucleotides and substituted nucleotides. Figure 4 As shown, ELECTRA's architecture consists of a lightweight generator. and discriminator Components: The generator predicts reasonable alternatives for the masked tokens, while the discriminator determines whether each token in the sequence is the original token. Through this adversarial training mechanism, ELECTRA can learn fine-grained contextual representations more efficiently than traditional masked language models, thus maintaining high semantic sensitivity when dealing with relatively short sequences.

[0037] The ELECTRA architecture is as follows: given a segmented miRNA sequence The generator first samples its "destructive" sequence. In this case, tokens at some locations were replaced. For each location... The discriminator outputs the probability that it is a real token: (6) in, This represents the context representation generated by the Transformer encoder. is a trainable projection vector. It is the logical sigmoid function.

[0038] The discriminator is trained based on binary cross-entropy loss: (7) Among them, when (i.e., when the position is the original token) ,otherwise .

[0039] After pre-training, the encoder generates contextualized embedding representations for miRNA sequences. To obtain a fixed-dimensional semantic representation, the embeddings of all tokens are averaged and pooled. (8) In one embodiment, multi-hop topological features are extracted from the circRNA-disease association network and miRNA-disease association network based on a diffusion-wavelet multi-hop topological modeling mechanism. Specifically, this includes: constructing a heterogeneous association graph based on the circRNA-disease association network and miRNA-disease association network and calculating its normalized Laplacian matrix; calculating a diffusion wavelet coefficient matrix based on a hot kernel filter and the eigenvalue decomposition of the Laplacian matrix to capture multi-scale topological structures; calculating multi-hop topological similarity between nodes based on the diffusion wavelet coefficient matrix using a minimum pairwise difference metric; and weighting and aggregating the local features of nodes within the k-hop neighborhood subgraph based on the topological similarity to generate the multi-hop topological features. The local features of the nodes include the logarithmically transformed node degree and clustering coefficient.

[0040] In this embodiment, a given association graph is provided. ,in A set of nodes representing circRNA or miRNA. Represents a set of disease nodes. This represents the known interacting edges. The adjacency matrix of the graph is... When circRNA or miRNA With disease When there is a relationship, let Otherwise, it is 0.

[0041] More specifically, a) Construction of the deterministic feature matrix. A deterministic feature matrix is ​​constructed to characterize the local topological statistical features. For each node... Calculate its degree and clustering coefficient after logarithmic transformation: (9) (10) in, For nodes The neighborhood group, This represents the set of edges in the subgraph induced by neighboring nodes. To mitigate differences at different scales, these features are stabilized using a logarithmic transformation: (11) Then, by summing up the feature vectors of all nodes, we can obtain the overall feature matrix: (12) Among them, matrix The Rows correspond to vectors .matrix This signal is then used as the input signal for the diffusion-wavelet transform.

[0042] b) Calculation of diffusion-wavelet coefficients To capture multi-scale topological similarity, a spreading wavelet is applied to the graph Laplacian operator. It is defined as follows: in The degree matrix. Laplace operator. It can be further broken down into: (13) in, This represents the orthogonal normalized matrix composed of eigenvectors. Let be the corresponding eigenvalue diagonal matrix, satisfying: (14) To construct the diffusion wavelet, the thermonuclear filter is defined as: (15) in, The diffusion scale parameter controls the degree of spectral smoothing. This kernel function attenuates high-frequency components in the Laplace spectrum, thereby generating a set of wavelet basis functions with locality and scale dependence.

[0043] Based on this, the wavelet coefficient matrix It can be represented as: (16) For nodes The wavelet coefficient matrix of the th List It describes the diffusion pattern of the node in the graph, thereby effectively encoding its structural role and associated context in the network topology.

[0044] c) Topological similarity measurement based on MDPA: node With nodes The topological similarity between them is quantified by minimum pairwise difference (MDPA). Let... and They are respectively and Given the ascending sort result, the MDPA of the two nodes is defined as follows: (17) Then, the node With nodes Topological similarity is defined as: (18) To further reflect the local neighborhood structure, similarity is normalized within the k-hop subgraph, taking the following form: (19) in, Represented by node A k-jump subgraph constructed around the center.

[0045] d) Characterization of subgraph feature distribution: Although the characteristics of each node are deterministically defined at the single-node level, in the k-hop neighborhood... Within this subgraph, these node features are treated as random variables to characterize the statistical distribution of features within the subgraph. Specifically, nodes... The characteristic function is defined as follows: (20) in, These are the evaluation points for the feature function. After summarizing all nodes, the graph-level feature function representation can be obtained: (twenty one) exist sampling points By sampling the feature function and concatenating its real and imaginary parts, a k-hop embedding can be formed. Subsequently, from arrive Cascaded combination of multi-hop embeddings: (twenty two) Simultaneously, using the normalized node degree as input, a set of parallel embeddings is constructed in the same manner and then combined with... The components are spliced ​​together to generate the final structural embedding representations for circRNA, miRNA, and disease nodes.

[0046] In one embodiment, the construction of negative samples in the training data includes screening using predicted subcellular localization information of circRNA and miRNA to exclude candidate negative sample pairs located in the same subcellular compartment.

[0047] In this embodiment, a novel dataset, CMI-Zero, was constructed specifically for zero-shot prediction tasks. This dataset was formed by systematically integrating circRNA-disease and miRNA-disease associations from multiple authoritative biological databases, including circAtlas 3.0, circRNADisease 2.0, and HMDD 4.0. After rigorous screening, the final CMI-Zero dataset contains 2,014 diseases, 442 circRNAs, and 310 miRNAs, involving a total of 1,812 circRNA-disease associations and 17,329 miRNA-disease associations. Based on existing experimental validation records in the databases and literature, 721 CMIs were identified and defined as positive samples.

[0048] To mitigate bias during negative sample construction, lncLocator and miRNALoc were used to predict molecular subcellular localization. Candidate negative samples located in the same subcellular compartment were discarded, and an equal number of negative samples were randomly sampled from different compartments to maintain biological plausibility. The final evaluation data included 721 experimentally validated CMIs and an equal number of constraint negative samples, ensuring that model inferences were based entirely on unseen associations and that no information leakage occurred.

[0049] To comprehensively evaluate the generalization ability of the proposed model, two additional benchmark datasets were used: CMI-9589 and CMI-9905. CMI-9589, derived solely from CircBank, contains 2,115 circRNAs, 821 miRNAs, and 9,589 high-confidence interactions. CMI-9905 integrates data from CircBank and circR2Cancer, covering 2,346 circRNAs, 962 miRNAs, and 9,905 interactions, including high-throughput prediction results and experimental validation records.

[0050] To comprehensively evaluate the performance of this method, five standard evaluation metrics were used: accuracy (Acc), F1 score, Matthews correlation coefficient (MCC), area under the ROC curve (AUC), and area under the PR curve (AUPR). To ensure the reliability and objectivity of the results, all experiments were conducted using five-fold cross-validation (5-fold CV). Specifically, the dataset was randomly divided into five subsets, with four subsets selected for training and the remaining subset used for testing in each iteration. This process was repeated five times, and the final performance was the average of the five experimental results to reduce variance fluctuations during the evaluation process.

[0051] In one embodiment, the classifier is a multilayer perceptron.

[0052] To utilize the fused semantic embeddings and multi-hop structural embeddings for CMI prediction, this method employs a multilayer perceptron (MLP) as the classifier. This classifier consists of an input layer, multiple hidden layers with non-linear activation functions, and an output layer for outputting interaction probabilities.

[0053] Specifically: a) Feedforward propagation For each pair of circRNA–miRNA, its corresponding eigenvector is denoted as . This vector serves as the input to the MLP and is passed sequentially through... Layer propagation. For the first layer... The activation values ​​of the hidden layers are calculated as follows: (twenty three) in, and The first Layer weight matrix and bias vector, This is a non-linear activation function. The input layer is defined as follows: ,and Then it serves as the input for the next layer. .

[0054] In the output layer, a logistic sigmoid function is used to generate the probability of the circRNA–miRNA pair interacting: (twenty four) in, This represents the probability predicted by the model that the circRNA–miRNA pair will interact.

[0055] b) Loss Function and Regularization: By minimizing the prediction probability With real labels The binary cross-entropy loss between them is used for training: (25) To mitigate the overfitting problem, an L2 regularization term was added to the loss function: (26) in, Used to control the severity of penalties for high-weight items.

[0056] c) Optimization and Prediction Parameter set Update via Adam optimizer: (27) in, For learning rate, This represents the set of trainable parameters.

[0057] Results: First, the proposed method was evaluated using five-fold cross-validation on the CMI-Zero dataset. As shown in Table 1, the model achieved an average Acc of 94.24%, an F1 score of 93.99%, and an MCC of 88.87% on this dataset, with an AUC of 0.9697 and an AUPR of 0.9801. The low standard deviations of all metrics indicate that the model has stable and reliable predictive performance.

[0058] Table 1 shows the performance of this method on the CMI-Zero dataset using five-fold cross-validation.

[0059] Table 2 shows the performance of this method on the CMI-9589 dataset using five-fold cross-validation.

[0060] Table 3 shows the performance of this method on the CMI-9905 dataset using five-fold cross-validation.

[0061] On the CMI-9589 dataset, the model's average Acc, F1, and MCC were 94.27%, 94.41%, and 88.66%, respectively, with AUC and AUPR reaching 0.9780 and 0.9675, respectively (see Table 2). On the CMI-9905 dataset, the corresponding metrics were 94.66%, 94.85%, 89.55%, 0.9838, and 0.9764, respectively (see Table 3). The ROC and PR curves for the three datasets are shown below. Figure 5 As shown, this further verifies that the proposed method maintains strong discriminative ability and good precision-recall performance under different data conditions.

[0062] Multi-hop propagation analysis: To investigate the impact of multi-hop propagation on this method, the model performance was tested on three benchmark datasets with different k-hop settings. Table 4 presents the evaluation results of metrics such as Acc, F1, MCC, AUC, and AUPR.

[0063] On the CMI-Zero dataset, the model performs best under one-hop propagation, achieving the highest performance across all evaluation metrics: Acc 94.24%, F1 93.99%, MCC 88.87%, AUC 0.9697, and AUPR 0.9801. As the propagation depth exceeds one hop, the model performance gradually declines, indicating that the immediate neighborhood provides the most discriminative topological signal for zero-shot prediction, while higher-order neighborhoods may introduce noise, thus weakening the model's effectiveness.

[0064] In contrast, on the CMI-9589 and CMI-9905 datasets, the best performance was achieved in the three-hop and four-hop propagation settings, respectively. On the CMI-9589 dataset, three-hop propagation achieved the highest Acc (94.27%), F1 (94.41%), and MCC (88.66%), while also showing the best performance in AUC and AUPR. For the CMI-9905 dataset, four-hop propagation achieved peak performance with Acc of 94.66%, F1 of 94.85%, MCC of 89.55%, AUC of 0.9838, and AUPR of 0.9764. These results demonstrate that on larger, denser datasets, multi-hop propagation can effectively capture richer structural information and improve prediction performance by aggregating features from more distant neighbors.

[0065] Table 4. Performance of this method under different k-hop settings and five-fold cross-validation.

[0066] To evaluate the contribution of each module in this method, ablation experiments were conducted. Specifically, the multi-hop structure modeling module (without Multi-hop) and the semantic representation learning module (without Semantic) were removed from the model. The relevant results are as follows: Figure 6 As shown.

[0067] On the CMI-Zero dataset, removing semantic features leads to a significant drop in all evaluation metrics. The AUC decreases from 0.9697 to 0.8632, and the MCC decreases from 0.8887 to 0.6058, highlighting the indispensable role of semantic representation in capturing molecular features. In contrast, removing multi-hop modeling only results in a modest decrease, with an AUC of 0.9686, not significantly different from the original model's 0.9697, indicating that semantic information dominates in zero-shot scenarios. Since the test molecules do not contain any known interactions in the training graph, their structural dependencies are greatly restricted. Therefore, semantic embeddings derived from nucleotide sequences become an important source of prior knowledge for the model to generalize to unobserved associations.

[0068] On the CMI-9589 dataset, removing multi-hop structure modeling significantly degrades model performance, with the AUC dropping from 0.9780 to 0.8158. However, removing semantic features allows the model to maintain relatively strong performance, with an AUC of 0.9764. These results indicate that when the dataset contains sufficient CMI samples, the topological dependencies captured through multi-hop diffusion-wavelet modeling contribute more than molecular semantic information.

[0069] On the CMI-9905 dataset, our method still achieves state-of-the-art performance, with an AUC of 0.9838 and an AUPR of 0.9764. Removing semantic features also leads to a significant performance drop (AUC = 0.9645), while removing multi-hop structural information has an even more severe impact (AUC = 0.8086). These results indicate that the semantic module and the structural module are functionally complementary, and their synergistic fusion is crucial for achieving stable and reliable prediction performance.

[0070] To validate the rationale for matching a specific Transformer architecture with the sequence characteristics it is suited for, a cross-semantics analysis was conducted. In this experiment, a cross-semantics variant was constructed, applying BERT to miRNA sequences and ELECTRA to circRNA sequences, and its performance was compared with the original model on three benchmark datasets. The comparison results are as follows: Figure 7 As shown.

[0071] Experiments revealed that Cross-Semantics outperformed our method across all datasets and evaluation metrics. This performance gap was most significant on the CMI-Zero dataset, while it was relatively smaller on CMI-9589 and CMI-9905. This phenomenon is consistent with the findings of ablation experiments, indicating that in zero-shot scenarios, the model primarily relies on prior semantic knowledge carried by the molecular sequence. Across semantic variants, the matching relationship between the Transformer architecture and sequence characteristics is broken, weakening the effectiveness of semantic representation and leading to decreased prediction accuracy. In contrast, the other two datasets are less sensitive to sequence semantic modeling, resulting in relatively smaller observed performance degradation.

[0072] To comprehensively evaluate the effectiveness of this method, it is compared with several representative graph embedding methods under five-fold cross-validation: DeepWalk: A network embedding method inspired by natural language modeling that captures potential node co-occurrence patterns in a graph through random walks.

[0073] LINE aims to preserve both first-order and second-order proximity by learning representations of large-scale networks through modeling local and neighborhood structural dependencies between nodes.

[0074] SDNE: A framework based on deep autoencoders that can model highly nonlinear topological relationships while maintaining structural proximity, thereby achieving robust feature extraction on complex graph structures.

[0075] MetaGC: A GNN-based clustering method that introduces a meta-weighting mechanism to adaptively reduce the influence of noisy edges, thereby enhancing meaningful structural patterns in the graph.

[0076] MUSAE: A multi-scale, attribute-aware embedding method that combines Skip-gram style random walks to capture attribute-neighborhood relationships across multiple scales, enabling joint modeling of node features and structural information.

[0077] The comparison results are shown in Table 5. Our proposed method achieves the best performance across all metrics and datasets, significantly outperforming shallow embedding methods such as DeepWalk and LINE. These methods primarily capture local co-occurrence statistics but fail to fully utilize higher-order structural dependencies. Although SDNE, MetaGC, and MUSAE employ deep architectures or attribute-aware mechanisms, their predictive performance remains lower than our method. The advantage of our proposed method stems from its multi-hop propagation mechanism based on diffusion-wavelet, which enables the model to progressively integrate information from local neighborhoods and distant nodes, thereby obtaining a more comprehensive and discriminative structural representation.

[0078] Table 5. Performance comparison of different graph embedding methods under five-fold cross-validation.

[0079] To evaluate the effectiveness of our method in capturing discriminative features for CMI prediction, it was compared with several typical classifiers, including Support Vector Machine (SVM), Random Forest, Logistic Regression, and LightGBM. All models underwent five-fold cross-validation on three benchmark datasets, and the results are shown below. Figure 8 As shown.

[0080] Experimental results demonstrate that our proposed method exhibits a consistent performance advantage across all metrics and datasets. Particularly noteworthy is the significant performance improvement on the CMI-Zero dataset, which only includes experimentally validated interactions, highlighting its robustness in real-world biological scenarios. Furthermore, while LightGBM shows superior performance in accuracy and AUC, it still falls short of our proposed method overall. On the CMI-9589 and CMI-9905 datasets, SVM performs relatively better among the evaluated traditional classifiers, but its predictive ability remains significantly inferior to our proposed method.

[0081] Table 6 compares the performance of the five-fold cross-validation with the latest model.

[0082] To further validate the predictive power of our method, a comprehensive comparison was conducted with several state-of-the-art methods on three benchmark datasets, and the results are shown in Table 6. The baseline models included in the comparison are EGATCMI, BioKG-CMI, ADA-GAD, JSNDCMI, BEROLECMI, MRHRL, LSAGHI, GraCMI, DGCLCMI, and MFERL. Evaluation metrics encompassed three standard measures: Acc, AUC, and AUPR, to comprehensively reflect the predictive performance of each method.

[0083] On the CMI-Zero dataset, our method significantly outperforms all baseline methods. Its Acc reaches 94.24%, a 10.47% improvement over the second-best method, ADA-GAD, while also improving AUC and AUPR by 0.0605 and 0.0548, respectively. These results demonstrate our method's strong generalization ability in scenarios without known interactions. For the CMI-9589 dataset, our method also maintains a leading performance, achieving Acc and AUC of 94.27% and 0.9780, respectively. Although the advantage over the top-performing baseline model, MFERL, is slightly smaller, it still exhibits stable consistency across all metrics, indicating that our method remains robust in more diverse network architectures.

[0084] On the CMI-9905 dataset, our method achieves the best overall prediction performance, with an Acc of 94.66%, an AUC of 0.9838, and an AUPR of 0.9764. Compared with the best-performing competing models (such as MFERL and DGCLCMI), our method improves Acc by 3.82% and AUC by 0.0193. These results demonstrate that the multi-hop diffusion-wavelet structure representation employed in our method effectively balances local interactive details with global structural dependencies, thereby significantly improving prediction accuracy.

[0085] Case Studies Lung cancer, especially non-small cell lung cancer (NSCLC), remains one of the leading causes of cancer-related deaths worldwide, accounting for approximately 85% of all diagnosed cases. Its complex molecular mechanisms and high heterogeneity pose significant challenges to early diagnosis and effective treatment. Similarly, osteosarcoma is the most common malignant bone tumor among adolescents and young adults, characterized by its aggressive growth, rapid metastasis, and poor prognosis. Therefore, identifying reliable cancer molecular mechanisms (CMIs) associated with these two diseases is crucial for understanding their pathogenesis and exploring new therapeutic targets.

[0086] To further evaluate the practical application value of this method, two case studies were conducted focusing on NSCLC and osteosarcoma. Specifically, the predicted CMIs were ranked, and the top 10 interaction pairs associated with each disease were selected, as shown in Tables 7 and 8. Of the top 10 predictions associated with NSCLC, 9 pairs of interactions have been validated in wet experiments, with all relevant evidence derived from published literature (the corresponding PMIDs are given in the tables). For osteosarcoma, the model's predictive performance was equally outstanding, with 9 of the top 10 predictions also being experimentally validated and reported in the literature. Only one pair in each disease case lacked direct experimental validation, indicating that the proposed model can reliably predict biologically significant potential associations.

[0087] Table 7 Top 10 CMIs Related to NSCLC

[0088] Table 8 Top 10 CMIs Associated with Osteosarcoma

[0089] This method effectively connects molecular sequence information with the topological structure of heterogeneous biological networks by jointly integrating deep semantic representations from the Transformer architecture with multi-hop diffusion-wavelet modeling. Extensive experimental results on multiple benchmark datasets demonstrate that this model significantly outperforms state-of-the-art methods, especially in zero-shot prediction scenarios, where predictions of previously unobserved interactions are required. Furthermore, case studies further validate the method's ability to reveal biologically significant CMIs with potential disease relevance. Overall, this study establishes a unified modeling paradigm that fully leverages the complementary advantages of semantic encoding and structural modeling, thereby improving prediction accuracy and biological interpretability.

[0090] Example 2 The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described thereon.

[0091] Example 3 The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0092] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predicting circRNA-miRNA interactions in a zero-sample scenario, characterized in that, include: Obtain circRNA sequences, miRNA sequences, and circRNA-disease association networks and miRNA-disease association networks; This paper describes a pre-trained language model based on Transformer to extract deep semantic representations from circRNA and miRNA sequences. Specifically, it includes: using a first semantic encoder based on the Transformer architecture to encode the circRNA sequence, obtaining circRNA semantic embedding vectors; and using a second semantic encoder based on the Transformer architecture to encode the miRNA sequence, obtaining miRNA semantic embedding vectors. The first and second semantic encoders are different pre-trained language models. The first semantic encoder is a BERT model, used to capture long-range dependency semantics in long circRNA sequences; the second semantic encoder is an ELECTRA model, used to capture key motif semantics in short miRNA sequences. Based on the diffusion-wavelet multi-hop topology modeling mechanism, multi-hop topological features are extracted from the circRNA-disease association network and miRNA-disease association network. Specifically, this includes: constructing a heterogeneous association graph and calculating its normalized Laplacian matrix based on the circRNA-disease association network and miRNA-disease association network; calculating the diffusion wavelet coefficient matrix based on the heat kernel filter and the eigenvalue decomposition of the Laplacian matrix to capture multi-scale topological structures; calculating the multi-hop topological similarity between nodes using the minimum pairwise difference metric based on the diffusion wavelet coefficient matrix; and weighting and aggregating the local features of nodes within the k-hop neighborhood subgraph based on the topological similarity to generate the multi-hop topological features. The local features of the nodes include the logarithmically transformed node degree and clustering coefficient. By fusing the deep semantic representation with the multi-hop topological features, a fused feature vector is obtained; The fused feature vector is input into a classifier to predict whether there is an interaction between the circRNA and miRNA. During the model training phase, the training data used satisfies the zero-sample setting, that is, the circRNA and miRNA nodes in the training set have no overlap with the circRNA and miRNA nodes in the test set.

2. The method according to claim 1, characterized in that, The construction of negative samples in the training data includes screening using the predicted subcellular localization information of circRNA and miRNA to exclude candidate negative sample pairs located in the same subcellular compartment.

3. The method according to claim 1, characterized in that, The classifier is a multilayer perceptron.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 3.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 3.

Images