CircRNA-disease association prediction method and system and medium

Abstract

The invention provides a circRNA-disease association prediction method and system and a medium, and the method comprises the steps: firstly obtaining a secondary structure of a target circRNA sequence, and carrying out the dynamic structure guide word segmentation of the sequence based on the secondary structure, so as to obtain a sequence fragment containing a structure attribute; then, deep semantic features of the sequence fragments are extracted through a trained deep semantic coding model, and meanwhile, deep semantic features of a target disease are extracted through a language model pre-trained on biomedical corpus; further fusing the circRNA and the deep semantic features of the disease to form a fusion feature vector; and finally inputting the fusion feature vector into a classification model, and outputting an association prediction result of the two. According to the method, by combining the secondary structure attribute of the circRNA and the pre-training semantic information in the biomedical field, the discrimination and relevance of the fusion features are improved, and the accuracy of circRNA-disease relevance prediction is effectively improved.

Description

A method, system, and medium for predicting circRNA-disease associations Technical Field

[0001] This invention relates to the field of bioinformatics processing technology, specifically providing a circRNA-disease association prediction method, system, and medium. Background Technology

[0002] CircRNAs are a special class of RNA molecules with a circular structure. In the study of circRNAs, scientists have discovered that they possess a variety of biological functions, including regulating transcription and splicing, and modulating the stability of cytoplasmic mRNA. Scientists have demonstrated that the secondary structure of circRNAs significantly influences their function in biological processes, thus playing a crucial role in disease development. To improve the efficiency and reliability of identifying circRNA-disease associations, computational biologists have developed various predictive models. These models can be broadly classified into two categories.

[0003] Predictions can be made based on past behavior of circRNA molecules. For example, Wang et al. developed MGRCDA, which uses a combination of metapath and recommendation systems to extract features from circRNA and disease networks to predict the association between circRNA and disease. Li et al. proposed the BiSGTAR model, which uses an encoding-regression-decoding (ERR) framework to analyze the circRNA and disease association network. These methods, using different techniques such as matrix factorization, graph neural networks, and machine learning, predict CDA (Cognitive Disorders Affected) on circRNA-related biological association networks (circRNA-miRNA-Disease or circRNA-Disease association networks) based on past known information. While these methods have high accuracy, they often fail to utilize information about the circRNA itself, thus limiting the model's practical performance.

[0004] Combining circRNA molecular ontology features and behavioral features for prediction: For example, Zhang et al. proposed the iGRLCDA model, which uses the Word2vec algorithm to process circRNA sequences to obtain sequence features and circRNA and disease behaviors on the circRNA-Disease association network to predict circRNA-Disease associations. The AEDNN model uses the shortest character editing algorithm to model circRNA sequences and uses Gaussian kernel functions to model the behavioral features of circRNA and diseases, finally using a deep neural network to predict the association between circRNA and disease.

[0005] The aforementioned methods utilize circRNA sequence information for preliminary applications; however, their utilization is rather crude. For example, they treat RNA fragments as traditional natural language strings to calculate information based on the shortest edit distance. However, in organisms, RNA sequences are unstable and undergo self-folding, pairing some of their internal bases. Clearly, replacing these bases should incur different computational costs than replacing unpaired bases. Furthermore, when using machine learning algorithms like Word2vec to calculate RNA semantics, the RNA sequence needs to be first divided into sentences composed of small fragments using the Kmer method, and then semantic extraction and modeling are performed using natural language processing techniques. Here, the Kmer method has limitations: its fragments are of fixed length and cannot flexibly represent RNA of arbitrary granularity. Moreover, it fails to consider the secondary structure of RNA, leading to bases in different regions being grouped into the same fragment, resulting in semantic information that deviates from the biological reality. Summary of the Invention

[0006] To address the problem of insufficient utilization of structural information in current circRNA representation, this invention provides a circRNA-disease association prediction method, comprising: obtaining the secondary structure of a target circRNA sequence, and performing dynamic structure-guided word segmentation on the sequence based on the secondary structure to obtain multiple sequence fragments with structural attributes; extracting deep semantic features of the target circRNA, wherein the deep semantic features are obtained by inputting the embedding representation of the sequence fragments into a trained deep semantic coding model; extracting deep semantic features of a target disease, wherein the deep semantic features of the disease are obtained by inputting the disease name into a language model pre-trained on a biomedical corpus; at least fusing the deep semantic features of the target circRNA and the deep semantic features of the target disease to form a fused feature vector; and inputting the fused feature vector into a classification model to obtain the association prediction result between the target circRNA and the target disease.

[0007] Preferably, obtaining the secondary structure of the target circRNA sequence includes: predicting the secondary structure of the target circRNA sequence based on the RNAfold algorithm to obtain the secondary structure of the sequence represented by dotted brackets.

[0008] Preferably, dynamic structure-guided word segmentation of the sequence based on the secondary structure includes: parsing the dotted bracket representation of the secondary structure, identifying regions with consecutive paired base pairs as stem regions, and regions with consecutive unpaired base pairs as loop regions; dividing consecutive bases in the sequence corresponding to the same stem region or the same loop region into an independent sequence segment; and labeling each sequence segment with a corresponding structure type attribute tag, the tag including at least stem region segments and loop region segments.

[0009] Preferably, inputting the embedded representation of the sequence fragment into a trained deep semantic coding model includes: converting each sequence fragment into a fragment word vector through a word vector model; performing one-hot encoding on the structural attribute label corresponding to each fragment to obtain a structural attribute vector; and concatenating the fragment word vector with the structural attribute vector to form the embedded representation of each fragment.

[0010] Preferably, the deep semantic encoding model is an attention-based LSTM autoencoder; the encoder part of the LSTM autoencoder is used to receive the embedding representation sequence of the sequence fragment and output the deep semantic features of the target circRNA; the decoder part of the LSTM autoencoder is used to reconstruct the embedding representation sequence of the sequence fragment based on the deep semantic features; the training objective of the model is to minimize the error between the reconstructed sequence and the original input sequence.

[0011] Preferably, extracting deep semantic features of the target disease includes: adding a classification marker before the disease name and inputting it into the language model, and extracting the output vector corresponding to the classification marker as the deep semantic features.

[0012] Preferably, the method further includes the step of extracting network topology behavior features: based on known circRNA-disease association data and circRNA-miRNA association data, a heterogeneous network containing circRNA nodes, disease nodes, and miRNA nodes is constructed; a second-order random walk is performed on the heterogeneous network using the node2vec algorithm to generate a node sequence, and a network embedding vector for each node is learned using the Skip-gram model; from the network embedding vectors, vectors corresponding to the target circRNA and the target disease are obtained respectively as their respective network topology behavior features.

[0013] Preferably, forming the fusion feature vector includes: normalizing the deep semantic features of the target circRNA, the deep semantic features of the target disease, and the network topology behavior features of the target circRNA and the target disease respectively; concatenating all the normalized features of the target circRNA to obtain the circRNA fusion feature; concatenating all the normalized features of the target disease to obtain the disease fusion feature; and concatenating the circRNA fusion feature and the disease fusion feature to form the final fusion feature vector.

[0014] Secondly, the present invention provides a circRNA-disease association prediction system, comprising: a first acquisition module, configured to acquire the secondary structure of a target circRNA sequence and perform dynamic structure-guided word segmentation on the sequence based on the secondary structure to obtain multiple sequence fragments with structural attributes; a second acquisition module, configured to extract deep semantic features of the target circRNA, wherein the deep semantic features are obtained by inputting the embedding representation of the sequence fragments into a trained deep semantic coding model; a third acquisition module, configured to extract deep semantic features of a target disease, wherein the deep semantic features of the disease are obtained by inputting the disease name into a language model pre-trained on a biomedical corpus; a fusion module, configured to fuse at least the deep semantic features of the target circRNA and the deep semantic features of the target disease to form a fused feature vector; and a prediction module, configured to input the fused feature vector into a classification model to obtain the association prediction result between the target circRNA and the target disease.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the circRNA-disease association prediction method.

[0016] The beneficial effects of this invention are: by introducing additional information about the "paired base pairs" existing inside the CircRNA through the secondary structure of the CircRNA, the structural feature representation of the CircRNA is optimized.

[0017] Structure-guided segmentation divides sequences into segments that possess structural characteristics. These segments can be treated as their corresponding parts of speech and fed into the LSTM model, achieving a joint semantic-structural representation of functional regions. Attached Figure Description

[0018] 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 figures are used to denote similar components, wherein: Figure 1 is a schematic flowchart of a circRNA-disease association prediction method according to an embodiment of the present invention; Figure 2 is a schematic diagram of the secondary structure of circRNA obtained by RNAfold calculation according to an embodiment of the present invention; Figure 3 is a five-fold cross-validation ROC curve obtained by a model according to an embodiment of the present invention; Figure 4 is a five-fold cross-validation PR curve obtained by a model according to an embodiment of the present invention; Figure 5 is a comparison result of the secondary structure features and primary structure features of circular RNA according to an embodiment of the present invention; Figure 6 is a comparison result of different word segmentation methods according to an embodiment of the present invention; Figure 7 is an experimental result of different embedding dimensions of a model according to an embodiment of the present invention; Figure 8 is a comparison result of different disease semantic extraction methods according to an embodiment of the present invention; Figure 9 is a comparison result of different attribute feature models according to an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] As shown in Figures 1-9, Example 1 provides a circRNA-disease association prediction method, including: Step S1, obtaining the secondary structure of the target circRNA sequence, and performing dynamic structure-guided word segmentation on the sequence based on the secondary structure to obtain multiple sequence fragments with structural attributes.

[0021] In this embodiment, to train the model, it is first necessary to collect the sequence information of circRNAs and the confirmed association information between circRNAs and diseases, and between circRNAs and miRNAs.

[0022] Duplicate data were removed from the CircR2Disease dataset. 607 confirmed association pairs were obtained between 561 circRNAs and 100 diseases. To construct a balanced dataset, 607 pairs were randomly sampled from the remaining circRNA-Disease pairs as negative samples for subsequent training.

[0023] The corresponding circRNA sequence was found in CircBase. CircRNA-miRNA interaction pairs were obtained from the starBase biomolecular interaction network database.

[0024] In one embodiment, the RNAfold algorithm is used to calculate the secondary structure of the circRNA based on the circRNA sequence collected in the previous step and the principle of minimum binding energy of RNA. Since circRNA does not have a closed loop structure at the 3' and 5' ends, RNAfold performs a different state transition equation than linear RNA when calculating its external loop.

[0025] Specifically, initialization: Initialize the dynamic programming table to store the minimum free energy of the substructure.

[0026] Recursive calculation: Fill the dynamic programming table by recursively calculating the free energy of the substructure.

[0027] Backtracking: By backtracking the dynamic programming table, the free energy of different pairing methods is compared to construct the optimal RNA secondary structure.

[0028] Different structures have different free energies. The specific state transition equations are as follows.

[0029] Single-chain interval: Hairpin ring: Inner ring: Multi-branch ring: The outer loop suitable for circRNA: This algorithm utilizes dynamic programming to predict the minimum free energy (MFE) structure and incorporates adaptive adjustments to the closed-loop topology of circRNAs. Key features include: forced terminal base pairing, ensuring terminal base pairing of the circular RNA to simulate its closed-loop characteristics; and an improved outer loop energy calculation (Equation 5), which adjusts the outer loop energy calculation to account for the continuity of the circular RNA.

[0030] The algorithm above was used to calculate the secondary structures corresponding to different CircRNAs on all CircRNA sequences.

[0031] For example, the circRNA molecule sequence is: GAAGAAGAUGCUCGGAAAGCAAAGCAUUUGAAAAAAUGUAACUCAAGAGAGAAACCAAAACCUGAUUUCUAUAUUCAAGAUAUUAAG. The secondary structure after RNAFOLD folding is shown in Figure 2.

[0032] In one embodiment, to more accurately measure the structural differences between different CircRNA molecules, the RNAalifold module in the ViennaRNA Package is used to calculate the secondary structure reconstruction distance.

[0033] RNAalifold uses dynamic programming to calculate the minimum number of operations required to convert two RNA molecules into each other's secondary structures by inputting the dot-bracket notation of RNA secondary structures. Compared to sequence alignment distance, this method more accurately reflects the structural differences of RNA molecules after folding.

[0034] For all CircRNA molecules, calculate the structural reconstruction distance pairwise, construct an n×n distance matrix (n is the number of CircRNAs), and use cosine similarity to convert the distance matrix into the final structural feature representation.

[0035] Step S2: Extract the deep semantic features of the target circRNA. The deep semantic features are obtained by inputting the embedding representation of the sequence fragment into a trained deep semantic coding model.

[0036] In this embodiment, since the base sequence of RNA molecules can be regarded as corpus, natural language processing computation has fully demonstrated its effectiveness in bioinformatics analysis. To capture the sequence semantics related to circRNA function, a structure-guided segmentation-embedding-coding framework is proposed, integrating RNA secondary structure information with a deep learning model.

[0037] First, the CircRNA molecule is segmented according to its secondary structure, and then fed into an LSTM model with an attention layer to train and obtain the deep semantic embedding of the CircRNA molecule.

[0038] Specifically, a dynamic segmentation method based on RNA secondary structure is proposed to segment circRNA sequences. This method no longer divides RNA molecules into several segments of length k, because the functional segments of RNA are often related to its secondary structure.

[0039] Based on the secondary structure of the circRNA obtained in the previous step, each circRNA molecule is divided into several fragments. Then, the Skim-Gram algorithm provided by Word2Vec is used to train these fragments and extract their features. This yields a 128-dimensional word vector representation for each fragment.

[0040] The following is a detailed process of feature extraction.

[0041] The goal of the Skip-gram model is to maximize the value of a given center word. Context The conditional probability. Specifically, for each center word... The model attempts to predict its neighboring nodes within its context window.

[0042] Given a base sequence ,in Represents the first in the sequence For each base, the objective function of the Skip-gram model is: in, It is the size of the context window. Indicates at a given central node Under the condition, context node The probability of occurrence.

[0043] In addition, one-hot encoding is used to represent different structural attributes, and these representations are appended to the end of the word vectors.

[0044] To construct a deep semantic representation of circRNA, a deep semantic extraction framework based on an LSTM autoencoder and attention mechanism was designed. By mapping RNA fragments generated by dynamic segmentation of secondary structures to "semantic words," and combining their structural attributes (analogous to part-of-speech) with word vector representations generated by Word2vec, serialized semantic units are constructed.

[0045] Model Structure: The model adopts an encoder-decoder architecture: In the encoding stage, LSTM is used to capture the contextual dependencies of the sequence and a self-attention mechanism is introduced to weight key segments; in the decoding stage, the encoder's representation ability is optimized by reconstructing the original sequence, forcing the model to extract essential semantic features under the constraint of information bottleneck.

[0046] LSTM Encoder: The input sequence consists of variable-length RNA fragment vectors, and LSTM layers (256D hidden units) are used for feature extraction. To address the differences in sequence length, zero-padding is implemented to unify the input dimension, and a packed sequence is used to instruct the model to ignore invalid padding, thus ensuring training stability. LSTM regulates information flow through gating mechanisms (input gate, forget gate, output gate), effectively solving the long-term dependency problem of traditional RNNs.

[0047] Attention Mechanism: To enhance the model's ability to recognize different regions, the output of the LSTM layer is connected to an attention layer. For the input word vector sequence... The attention output is: in: , , These are the query, key, and value matrices, respectively. It is the dimension of the key vector. , , It consists of 128-dimensional DNNs. Finally, a fully connected layer is added to output the embedding.

[0048] Model Training: To train the CircRNA deep semantic extraction model, an LSTM decoder network with the same dimension as the encoder but in the opposite direction was constructed.

[0049] The input to the LSTM decoder is the sentence vector output from the final fully connected layer of the attention mechanism. The goal of the decoder is to reconstruct the original input sequence using an LSTM network.

[0050] The decoder's computation process is similar to the encoder's, but in the opposite direction. The decoder's output is the reconstructed sequence. .

[0051] To train the model, the reconstructed sequence is calculated. With the original input sequence The mean squared error (MSE) between the two sides is used as the loss function: The model parameters are optimized using the backpropagation algorithm to minimize the loss function. During training, gradient descent is used to update the model parameters.

[0052] Step S3: Extract the deep semantic features of the target disease. The deep semantic features of the disease are obtained by inputting the disease name into a language model pre-trained on a biomedical corpus.

[0053] In this embodiment, to address the limitations of traditional MeSH semantic analysis in complex disease association modeling, a context-aware disease semantic embedding method based on BioBERT is proposed.

[0054] Traditional DAG-based methods are limited by the static representation of the hierarchical structure of medical terminology, focusing on the hierarchical information in disease taxonomy and ignoring the rich common-sense knowledge and interdisciplinary contextual information in the biomedical field. They are difficult to capture the pathological mechanism connections and interdisciplinary knowledge implicit in disease descriptions. To this end, BioBERT

[18] is used as a semantic encoder. Through its deep language understanding capabilities pre-trained on a large-scale biomedical corpus (covering 30 million PubMed documents), a dynamically differentiable disease semantic space is constructed.

[0055] BioBERT is a Transformer-based encoder-only architecture consisting of multiple layers of self-attention mechanisms and feedforward neural networks. The key components of the encoder are: input representation, self-attention mechanism, and feedforward neural network.

[0056] The input text is first converted into a combination of word embeddings, position embeddings, and segment embeddings. The self-attention mechanism captures contextual information by calculating attention weights between words. For the input sequence... The self-attention output is: in: , , These are the query, key, and value matrices, respectively. It is the dimension of the key vector. , , It is a learnable weight matrix.

[0057] The self-attention output is then further processed by a feedforward neural network (FFN) to finally obtain the encoded output.

[0058] For target disease After adding the [CLS] marker to the beginning of the name, input a BioBERT encoder consisting of multiple stacked Transformer blocks to obtain a context-sensitive word vector sequence. By extracting the first vector corresponding to the [CLS] marker. As a global semantic embedding for the disease.

[0059] in, This represents the overall embedding of the current disease name sequence. This represents the i-th disease in the disease list. The embedding not only encodes the literal meaning of the disease term, but also includes its functional annotations in the biomedical context (such as pathogenic pathways, phenotypic associations).

[0060] The cosine similarity between these embeddings is then calculated for subsequent use in CircRNA disease association prediction.

[0061] In one embodiment, to capture the topological behavior characteristics of circRNAs in biological networks, a node2vec algorithm based on second-order random walks is used to perform characterization learning on heterogeneous circRNA-disease and circRNA-miRNA networks. This method balances the extraction of local neighborhood information and global structural features through a parameterized walk strategy, specifically implemented as follows: Given a circRNA node... Step size is Random walk to generate sequences ,in The transition probability at each step is determined by the following formula: in, It is the current node. It is the next candidate node. It is a node To the node The nonnormalized transition probability. It is a normalized constant, ensuring that the sum of probabilities is 1.

[0062] Transition probability Calculated using the following formula: in, It is the edge The weight (if the edge exists) ;otherwise ). It is the bias coefficient, determined by the parameter and The control determines whether the resulting sequence is more biased towards global or local information. By convention, we set both p and q to 1. This results in a walk sequence that balances global and local information.

[0063] Subsequently, the generated wandering sequences are treated as structured corpus and input into the Skip-gram model for distributed representation learning. Skip-gram maps each circRNA node to a 128-dimensional continuous vector space by maximizing the co-occurrence probability within the node's context window.

[0064] Finally, the 128-dimensional embedding representation of disease nodes is generated on the CircRNA-Disease heterogeneous network using the same process, forming a unified topological feature space. Experiments show that this embedding method significantly enhances the model's ability to capture the functional consistency of circRNAs across networks while preserving the high-order structural properties of the network, laying the foundation for multimodal feature fusion.

[0065] Step S4: At least the deep semantic features of the target circRNA and the deep semantic features of the target disease are fused to form a fused feature vector.

[0066] In this embodiment, in order to achieve the association prediction task, the previously extracted CircRNA secondary structure features are... Deep semantic features of CircRNA CircRNA behavioral characteristics Deep semantic features of disease and Disease behavioral characteristics Integration was performed, and CircRNA fusion features were constructed separately. Features of fusion with disease .

[0067] First, each feature class is normalized separately to prevent larger-scale features from dominating the model training process, ensuring they contribute equally during fusion. Then, these features are concatenated to fully utilize the complementary information between them. The concatenation operation preserves the original information of each feature and allows the model to automatically learn the importance of different features. Therefore, CircRNA feature fusion... Features of fusion with disease They are represented as follows: in, This indicates a feature splicing operation.

[0068] In this way, multi-source features can be effectively integrated, providing a richer and more comprehensive feature representation for subsequent association prediction tasks.

[0069] Step S5: Input the fused feature vector into the classification model to obtain the association prediction result between the target circRNA and the target disease.

[0070] To comprehensively evaluate the performance of this method on the gold-standard dataset using 5-fold cross-validation, a series of evaluation metrics were employed. These metrics include accuracy (ACC), F1 score (F1-score), precision (PRE), and recall (REC). Furthermore, receiver operating characteristic (ROC) curves and their corresponding area under the curves (AUC values) were calculated and plotted to comprehensively assess the overall performance of the model. Additionally, the area under the precision-recall curve (AUPR) was calculated, which provides a more detailed evaluation of model performance by comprehensively considering precision and recall, especially in cases of imbalanced data.

[0071] in, Indicates a real example, Indicates a true negative example. Indicates a false positive example. This indicates a false negative.

[0072] Using the above metrics and curves, we can comprehensively evaluate the model's performance, especially its performance when dealing with complex data.

[0073] Model Performance Evaluation: To evaluate the performance of the LMSSNCDA model, five-fold cross-validation was performed on the CircR2Disease dataset. The AUC values ​​for the five cross-validations were 0.9787, 0.9807, 0.9884, 0.9907, and 0.9977, respectively, with an average AUC of 0.9877, significantly higher than the random guessing baseline (gray diagonal line). This indicates that the model has extremely high sensitivity and specificity in distinguishing between positive and negative samples, making it suitable for high-precision biomarker screening scenarios.

[0074] To evaluate the effectiveness of RNA secondary structure-based features in circRNA molecular characterization, a comparative experiment was designed. The experimental group used the RNAalifold algorithm to calculate the secondary structure reconstruction distance between circRNA molecules, while the control group used the traditional Levenshtein algorithm, calculating the shortest edit distance based solely on RNA sequences. To focus more on assessing the impact of RNA secondary structure, a 64-dimensional random vector was used in disease characterization, and behavioral features of circRNAs and diseases were not introduced. Through performance comparison on the CircR2Disease dataset using five-fold cross-validation, we aimed to verify whether secondary structure-based computational methods can provide more biologically meaningful RNA molecular characterization, particularly in the analysis of circRNA differences and similarities.

[0075] Experimental results show that the secondary structure features based on the RNAalifold algorithm outperform traditional sequence alignment methods on all evaluation metrics. Specifically, as shown in Table 1, the improvement in AUPR (Area Under Precision-Recall Curve) is the most significant, reaching 0.06; Specificity is improved by 6%, AUC (Area Under ROC Curve) by 3%, while the improvement in Sensitivity is relatively small, only 1%. In addition, Accuracy is also improved by 0.03.

[0076] Table 1

[0077] Experimental results show that the secondary structure reconstruction distance based on the RNAalifold algorithm more sensitively captures structural differences between molecules, especially among highly similar circRNA molecules. The RNAalifold method provides more accurate discrimination, while the Levenshtein algorithm, by ignoring secondary structure, fails to effectively distinguish these molecules. Although the RNAalifold algorithm has a relatively high computational complexity, its ability to provide more accurate secondary structure information makes secondary structure characterization methods more biologically significant and valuable for circRNA molecule analysis.

[0078] To evaluate the performance improvement of structure-guided segmentation (Secondarymer) over traditional k-mer segmentation in deep semantic extraction of RNA sequences, we designed a comparative experiment.

[0079] The experimental group used structure-guided segmentation (Secondarymer) to divide RNA molecules into segments with "loop" and "stem" attributes; the control group used 1-mer, 2-mer, 3-mer, and 4-mer segmentation methods to divide RNA sequences into fixed-length segments. Since the k-mer segmentation method cannot capture the secondary structure information of RNA molecules, zero-vector padding with the same dimension as the experimental group was added to the end of the k-mer features to avoid bias introduced by the inconsistency in feature dimensions.

[0080] After feature extraction, the two sets of features were fed into an LSTM model for training, and their performance was evaluated using five-fold cross-validation on the CircR2Disease dataset. To focus more on evaluating the impact of RNA secondary segmentation and structural information on the semantic extraction of CircRNA molecules, we did not introduce disease semantic features or behavioral features of CircRNA and disease; instead, we used a 64-dimensional random vector in the disease representation. The experiments aimed to verify whether the secondary structure-based segmentation method can provide more biologically meaningful RNA molecular representations, especially in the analysis of CircRNA differences and similarities.

[0081] As shown in Table 2, the Secondary-mer method based on secondary structure achieved an AUC of 0.74. Among the k-mer methods, 1-mer performed best with an AUC of 0.709. However, compared to Secondarymer, the 1-mer method still lags significantly behind, indicating that k-mer segmentation based solely on sequence information cannot adequately capture the complex secondary structural features of RNA molecules. The significant advantage of secondary structure-based segmentation methods may be attributed to their ability to more rationally segment RNA fragments and capture information on whether fragments are self-paired. For example, the 'loop' and 'stem' structures of RNA molecules have key biological functions, but k-mer segmentation methods fail to integrate such structural information. Furthermore, k-mer methods only segment RNA sequences into fixed-length fragments, failing to adapt to the dynamic structural changes of RNA molecules. In contrast, Secondarymer can dynamically segment fragments according to the secondary structure of RNA, thus more rationally representing RNA features. Therefore, the Secondarymer approach can more realistically reflect the characteristics of circRNA, thus performing excellently in deep semantic extraction of RNA sequences.

[0082] Table 2 LSTM Model Optimization Analysis: The deep semantic extraction module of the CircRNA in this application utilizes an LSTM model. An important parameter of the LSTM model is its hidden layer dimension d, which is the representation dimension of the hidden states ht during model training. To find the optimal operating parameters for this model, five-fold cross-validation was performed on the model's performance while only adjusting the size of d.

[0083] Experimental data show significant performance differences across different dimensions, confirming the strict dependence of circRNA pattern learning on dimensionality parameters. The optimal performance window is observed to be around 200 dimensions. Lower dimensions result in lower performance and insufficient feature abstraction, while excessively high dimensions, such as an 8.9% decrease in AUC from d=200 to 512, exhibit performance degradation.

[0084] As shown in Table 3, based on the circular structure characteristics of circRNA, the 200-dimensional representation space may form a spatial mapping relationship with its average loop length (approximately 300 nt) and secondary structure complexity. However, excessively high dimensionality can introduce pseudo-sequence pattern learning. Experiments show that the 200-dimensional representation achieves a golden balance between complexity and efficiency: its AUC is 0.3% higher than the second-best 150-dimensional representation, while the number of parameters only increases by 28%.

[0085] Table 3 Comparison of Disease Semantic Extraction Methods: To evaluate the impact of different disease semantic extraction methods on model performance, the following comparative experiments were designed: Main Experiment Group: Disease Semantic Extraction Based on BioBERT. The main experiment group used the BioBERT model for disease semantic embedding extraction. BioBERT is a pre-trained model based on the BERT architecture, specifically optimized for biomedical text. It can capture deep semantic information in disease descriptions, providing high-quality embeddings for disease representation.

[0086] Control Group A: Disease Semantic Similarity Based on MeSH Lexicon. Control Group A calculates the semantic similarity of diseases based on a traditional MeSH (Medical Subject Headings) lexicon. This method measures semantic similarity by calculating the path distance of diseases within the MeSH tree structure. Although the MeSH method has explicit biological interpretability, its reliance on a manually constructed lexicon structure may not fully capture the complex semantic information of disease descriptions.

[0087] Control Group B: Disease Semantic Extraction Based on a General-Purpose Text Model. Control Group B also uses a large-scale model approach, but it is a general-purpose model not specifically trained in biomedical contexts. The nomic-embed-text-v1.5 model is used for disease semantic embedding extraction. This model is based on the BERT architecture, with a maximum context window of 2048, and its pre-training corpus includes BooksCorpus and a 2023 Wikipedia dump.

[0088] After obtaining the features, similar to the other experimental setups mentioned above, but considering that the semantic features obtained by Biobert and nomic are different in dimension from the semantic features obtained by Mesh, we uniformly reduced all features to 100 dimensions using PCA, set the RNA feature representation to 64-dimensional random numbers and no longer introduced other features, and then used 5-fold cross-validation to obtain the model's evaluation score on the XGBoost classifier.

[0089] As shown in Table 4, the experimental data demonstrates that BioBERT outperforms the other two methods across all evaluation metrics. For instance, BioBERT's AUC Score and AUPR Score reach 0.6984 and 0.6849, respectively, higher than nomic-embed-text-v1.5's 0.6858 and 0.6582, and MeSH's 0.5890 and 0.6182.

[0090] The nomic-embed-text-v1.5 outperforms the MeSH method on most metrics, but slightly underperforms BioBERT. For example, its accuracy is 0.6376, higher than MeSH's 0.5610, but lower than BioBERT's 0.6499.

[0091] The MeSH method performed the worst across all metrics. For example, its AUC score was only 0.5890, significantly lower than BioBERT and nomic-embed-text-v1.5.

[0092] Table 4

[0093] Experimental results show that the BioBERT-based disease semantic extraction method performs best in the CircRNA-disease association prediction task. It can more accurately capture the semantic information in disease descriptions, thus providing the model with a more biologically meaningful disease representation. In contrast, the traditional MeSH method relies on a manually constructed vocabulary tree structure, which cannot fully capture the complex semantic information of disease descriptions and performs poorly. Furthermore, although nomic-embed-text-v1.5 performs excellently in general domains, its performance in the biomedical field is still slightly lower than that of BioBERT, which is specifically trained on biomedical corpora.

[0094] Ablation experiments were conducted to evaluate the individual and joint contributions of different feature types in the CircRNA-disease association prediction model. The following ablation experiments were designed: First, the model of our method was divided into two large sub-modules, LMSSNCDA-A and LMSSNCDA-B. LMSSNCDA-A includes ontology feature models of secondary structure, deep semantics, and disease semantics, while LMSSNCDA-B includes behavioral feature models of CircRNA and the Disease graph embedding method. Furthermore, to verify the influence of different components within our ontology feature model, we also conducted separate ablation experiments within LMSSNCDA-A to determine the complementarity between our structural and semantic features. Specifically, the following models were constructed: CSt+DS (CircRNA structural features and disease semantic features), CSe + DS (CircRNA semantic features and disease semantic features), and CSe+CSt+DS (CircRNA structural features and disease semantic features).

[0095] As shown in Table 5, for each different experimental group, after obtaining these features, the features were normalized in order to balance the contribution of all features in the classifier, and then they were concatenated and fed into the XGboost classifier for five-fold cross-validation on the CircR2Disease dataset.

[0096] Table 5

[0097] Table 6 shows comparisons with other models. To more intuitively demonstrate the advantages of the LMSSNCDA model's predictive performance, we compared it with other mainstream models such as iCircDA-MF, NMFCDA, GGAECDA, and MNMDCDA on the CircR2Disease dataset. Some models did not provide certain data, so we labeled them with N / A. The results are shown in the table below. This model achieved optimal performance on several key metrics, providing a more reliable computational tool for related biomedical research.

[0098] Table 6

[0099] Example 2: This invention provides a circRNA-disease association prediction system, comprising: a first acquisition module, used to acquire the secondary structure of a target circRNA sequence and perform dynamic structure-guided word segmentation on the sequence based on the secondary structure to obtain multiple sequence fragments with structural attributes; a second acquisition module, used to extract deep semantic features of the target circRNA, wherein the deep semantic features are obtained by inputting the embedding representation of the sequence fragments into a trained deep semantic coding model; a third acquisition module, used to extract deep semantic features of a target disease, wherein the deep semantic features of the disease are obtained by inputting the disease name into a language model pre-trained on a biomedical corpus; a fusion module, used to fuse at least the deep semantic features of the target circRNA and the deep semantic features of the target disease to form a fused feature vector; and a prediction module, used to input the fused feature vector into a classification model to obtain the association prediction result between the target circRNA and the target disease.

[0100] Example 3: This invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the circRNA-disease association prediction method.

[0101] 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 circRNA-disease association prediction method, characterized in that, include: The secondary structure of the target circRNA sequence is obtained, and the sequence is dynamically structure-guided segmented based on the secondary structure to obtain multiple sequence fragments with structural attributes. Deep semantic features of the target circRNA are extracted, which are obtained by inputting the embedding representation of the sequence fragment into a trained deep semantic coding model; deep semantic features of the target disease are extracted, which are obtained by inputting the disease name into a language model pre-trained on a biomedical corpus; at least the deep semantic features of the target circRNA and the deep semantic features of the target disease are fused to form a fused feature vector; the fused feature vector is input into a classification model to obtain the association prediction result between the target circRNA and the target disease.

2. The method according to claim 1, characterized in that, Obtaining the secondary structure of the target circRNA sequence includes: predicting the secondary structure of the target circRNA sequence based on the RNAfold algorithm, and obtaining the secondary structure of the sequence in bracket representation.

3. The method according to claim 2, characterized in that, Dynamic structure-guided word segmentation of the sequence based on the secondary structure includes: parsing the bracket representation of the secondary structure, identifying regions with consecutive paired base pairs as stem regions, and regions with consecutive unpaired base pairs as loop regions; dividing consecutive bases in the sequence corresponding to the same stem region or the same loop region into an independent sequence segment; and labeling each sequence segment with a corresponding structure type attribute tag, the tag including at least stem region segments and loop region segments.

4. The method according to claim 3, characterized in that, The embedding representation of the sequence segments is input into a trained deep semantic coding model, including: converting each sequence segment into a segment word vector through a word vector model; performing one-hot encoding on the structural attribute label corresponding to each segment to obtain a structural attribute vector; and concatenating the segment word vector with the structural attribute vector to form the embedding representation of each segment.

5. The method according to claim 4, characterized in that, The deep semantic coding model is an attention-based LSTM autoencoder; the encoder part of the LSTM autoencoder is used to receive the embedding representation sequence of the sequence fragment and output the deep semantic features of the target circRNA; the decoder part of the LSTM autoencoder is used to reconstruct the embedding representation sequence of the sequence fragment based on the deep semantic features; the training objective of the model is to minimize the error between the reconstructed sequence and the original input sequence.

6. The method according to claim 1, characterized in that, Extracting deep semantic features of the target disease includes: adding a classification marker before the disease name and inputting it into the language model, and extracting the output vector corresponding to the classification marker as the deep semantic features.

7. The method according to claim 1, characterized in that, The method further includes the step of extracting network topology behavior features: based on known circRNA-disease association data and circRNA-miRNA association data, a heterogeneous network containing circRNA nodes, disease nodes, and miRNA nodes is constructed; a second-order random walk is performed on the heterogeneous network using the node2vec algorithm to generate node sequences, and the network embedding vector of each node is learned using the Skip-gram model; from the network embedding vectors, vectors corresponding to the target circRNA and the target disease are obtained respectively as their respective network topology behavior features.

8. The method according to claim 7, characterized in that, The process of forming the fusion feature vector includes: normalizing the deep semantic features of the target circRNA, the deep semantic features of the target disease, and the network topology behavior features of the target circRNA and the target disease respectively; concatenating all the normalized features of the target circRNA to obtain the circRNA fusion feature; concatenating all the normalized features of the target disease to obtain the disease fusion feature; and concatenating the circRNA fusion feature and the disease fusion feature to form the final fusion feature vector.

9. A circRNA-disease association prediction system, characterized in that, include: The first acquisition module is used to acquire the secondary structure of the target circRNA sequence and perform dynamic structure-guided word segmentation on the sequence based on the secondary structure to obtain multiple sequence fragments with structural attributes; the second acquisition module is used to extract the deep semantic features of the target circRNA, which are obtained by inputting the embedding representation of the sequence fragments into a trained deep semantic coding model; the third acquisition module is used to extract the deep semantic features of the target disease, which are obtained by inputting the disease name into a language model pre-trained on a biomedical corpus; the fusion module is used to fuse at least the deep semantic features of the target circRNA and the deep semantic features of the target disease to form a fused feature vector; the prediction module is used to input the fused feature vector into a classification model to obtain the association prediction result between the target circRNA and the target disease.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the circRNA-disease association prediction method as described in any one of claims 1 to 8.

Images