Circrna-mirna association prediction method, device and medium

Abstract

The application discloses a CircRNA-miRNA correlation prediction method and device and a medium. The method comprises the following steps: constructing a data set, wherein the data set comprises a plurality of CircRNAs and a plurality of miRNAs; based on the data set, biological attribute features are extracted from word embeddings of CircRNA sequences and miRNA sequences; behavior features are extracted based on isomorphic graph embedding; training features of a fusion carrier are obtained based on the biological attribute features and the behavior features, so that the interaction scores of circRNAs and miRNAs are learned, and potential CMAs are inferred. The application successfully predicts the complex relationship between circRNAs and miRNAs, the accuracy is 82.90%, and the AUC is 0.9075.

Description

Technical Field

[0001] This invention belongs to the technical field of biological intelligent processing methods, and more specifically, relates to a CircRNA-miRNA association prediction method, device and medium. Background Technology

[0002] Circular RNAs (circRNAs) were discovered in the early 1970s, but they were not considered key elements for RNA expression analysis in human physiology. Since 2013, a new batch of circRNAs has been discovered and quantified due to high-resolution and high-throughput RNA-seq data, particularly end-to-end reading and deep sequencing. According to published literature, there are three types of methods for identifying circRNAs: ab initio prediction methods, RNA-seq alignment tools (such as Segemehl), and algorithms and tools specifically designed for circRNA detection (such as CIRI). These studies have demonstrated that circRNAs are endogenous single-stranded non-coding RNA (ncRNA) molecules that covalently close circular structures at the downstream 5' splice site and upstream 3' site of the RNA transcript, possessing important biological functions. They play a crucial role in disease treatment by recognizing circRNAs and their association with diseases. Extensive experimental analysis shows that miRNAs are a type of ncRNA and an indispensable part of eukaryotic gene regulation. To date, a growing body of biological experiments has validated that circRNAs are miRNA sponges and can serve as a new class of biomarkers. To alleviate the labor-intensive nature of exploring circRNA-miRNA interactions through wet laboratory experiments, many computational methods have been used to identify circRNA-miRNA associations (CMAs).

[0003] Recent studies have shown that rapidly emerging computer algorithm models not only provide excellent solutions for CMA prediction, but also achieve higher accuracy while rapidly processing massive amounts of data in machine learning techniques. For example, Lan et al. developed an NECMA model that uses inner product and neighborhood-regularized logistic matrix factorization to predict CMAs. Qian et al. proposed a CMIVGSD model based on singular value decomposition and graph variational autoencoder to predict miRNA-related circRNAs. Guo et al. designed WSCD, which is based on structural deep neural network embedding and uses convolutional neural networks and deep neural networks to predict CMAs. Wang et al. inferred the association scores between miRNA-related circRNAs by integrating multimodal information and fusing node similarity features.

[0004] Furthermore, most existing CMA prediction models neglect the fully relevant biological information contained in their sequences, their impact on miRNA function, and their complex associations with circRNAs. Therefore, the following is a clearly worthwhile issue to address: (1) A reasonable feature fusion model integrates multimodal heterogeneous information and extracts the association between circRNAs and miRNAs; (2) Consider appropriate sequence and interaction information strategies to learn good features; (3) It is necessary to address the noise impact caused by cross-reactions in experiments. Summary of the Invention

[0005] The present invention is provided to solve the above-mentioned problems existing in the prior art. Therefore, a method, device and medium for predicting CircRNA-miRNA associations are needed, and a machine learning framework CMALF based on graph representation learning is needed to predict CMAs. This framework utilizes a deep learning isomorphic embedding fusion model and decomposition method. Specifically, we first extract potential biological attribute features from word embeddings based on circRNA sequences (a natural language processing method), word2vec

[24] . Secondly, the multi-source feature information contains its influence on the circRNA-miRNA interaction relationship. Heterogeneous graphs are obtained by verifying interaction pairs and the same number of unlabeled samples, and a reliable molecular association network is constructed. Therefore, the network is input into a fusion model of large-scale information network embedding and graph decomposition for low-dimensional embedding vector generation. Finally, the potential CMAs are effectively inferred using a gradient-enhanced decision tree classifier.

[0006] According to a first aspect of the present invention, a CircRNA-miRNA association prediction method is provided, the method comprising: Construct a dataset that includes multiple circRNAs and multiple miRNAs; Based on the dataset, biological attribute features were extracted from the word embeddings of CircRNA and miRNA sequences; Behavioral features are extracted based on isomorphic graph embedding; Training features of the fusion vector are obtained based on the biological attributes and behavioral features to learn the interaction scores between circRNA and miRNA and infer potential CMAs.

[0007] Furthermore, the constructed dataset is described by formula (1): (1) In the formula, D represents the union of elements. and These represent the sets of 20,208 positive samples and 4,091,280 negative samples, respectively.

[0008] Furthermore, construct a dimension of m × n The adjacency matrix of the dataset, where m Indicates the number of circRNAs. n Indicates the number of miRNAs; If miRNA ( ) and circRNA ( If they are unrelated, then In Set to 0; otherwise, In The value is 1, where miRNA ( ) represents the first element in the adjacency matrix. i miRNAs, circRNAs ( ) represents the first element in the adjacency matrix. j One circRNA, In the adjacency matrix, the th i The miRNA and the first j The correlation of individual circRNAs.

[0009] Furthermore, the extraction of biological attribute features from word embeddings of CircRNA and miRNA sequences based on the dataset specifically includes: Provide extraction models to extract biological attribute features; The objective function for extracting the output value of the model is expressed as: (2) In the formula, Represents the weight matrix. The parameters representing the target word and the context words are treated as... , T Indicates the number of samples. t This indicates the current sample number.

[0010] Furthermore, the extraction of behavioral features based on isomorphic graph embedding specifically includes: A multilayer perceptron is used to learn low-dimensional dense vectors while maintaining first-order and second-order similarity, where the first-order and second-order objective functions are expressed as follows: (3) (4) In the formula, O 1 represents a first-order objective function.O 2 represents the second-order objective function. Indicates the distance between two distributions. and It is considered a representation of the empirical distribution. and Let them represent the joint distribution and the contextual conditional distribution, respectively. i Represents the regularization parameter; Factorize the adjacency matrix of the graph. The objective function of factorization is expressed as: (5) In the formula, and These represent the weight adjacency matrix, factor matrix, and regularization parameter, respectively. E Indicates the size of the adjacency matrix. Z i Represents the factor vector in the i-th row. Y ij This represents the weight of the i-th row and j-th column. Z j This represents the factor vector of the j-th column.

[0011] Furthermore, the extraction of behavioral features based on isomorphic graph embedding specifically includes: It provides a fusion graph embedding model to extract behavioral features; The objective function of the fusion graph embedding model is expressed as: (6) In the formula, Y Represents graph embedding. and Let Y represent the embedding of nodes and the embedding of the graph, respectively. Indicates weight, L Describe the Laplace operator of graph G. Y T This represents the transpose of a graph embedding. tr This represents minimizing the objective function.

[0012] Furthermore, the step of obtaining training features for the fusion vector based on the biological attributes and behavioral characteristics to learn the interaction scores between circRNA and miRNA and infer potential CMAs specifically includes: A classifier is provided to learn the interaction scores between circRNA and miRNA and infer potential CMAs.

[0013] Furthermore, a weak classifier is provided, which is used for multiple iterations and the residuals of previous classifiers; The objective function of the weak classifier is expressed as: (7) In the formula, This represents the iteration result at this point. Represents the single-sample objective function. N Indicates the number of samples. y i Indicates a label, L () represents the loss function. T ( x i ; ) indicates error.

[0014] According to a second technical solution of the present invention, a CircRNA-miRNA association prediction device is provided, the device comprising: The building module is configured to build a dataset that includes multiple circRNAs and multiple miRNAs; The first extraction module is configured to extract biological attribute features from word embeddings of CircRNA and miRNA sequences based on the dataset. The second extraction module is configured to extract behavioral features based on isomorphic graph embedding. The prediction module is configured to obtain training features of the fusion vector based on the biological attribute features and the behavioral features, in order to learn the interaction scores of circRNA and miRNA and infer potential CMAs.

[0015] According to a third technical solution of the present invention, a readable storage medium is provided, wherein the readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method described above.

[0016] The present invention has at least the following beneficial effects: According to the present invention, CMALF can be constructed, which is a graph representation learning framework based on a fusion model that combines two isomorphic graph embedding methods to predict CMAs.

[0017] Overall evaluation of the constructed dataset confirms that CMALF outperforms other excellent models in predicting CMAs. To evaluate the best performance of this paper, we conducted several ablation experiments on the same dataset, including comparisons of the CMALF-LINE model, CMALF-GF and LAP models, different experimental classifier models, and the performance of the association prediction model compared with other state-of-the-art models. Overall evaluation of the constructed dataset confirms that CMALF outperforms other advanced models in predicting CMAs. Higher association scores were achieved in the prediction of a large number of experimental outcome pairs through case studies in recent literature. Careful analysis of the results shows that this invention achieves excellent prediction of potential CMAs in academic research.

[0018] This invention exhibits superior predictive performance. In particular, it considers the attribute and behavioral features provided by circRNA and miRNA sequences, which previous methods have paid less attention to fully utilizing. More specifically, 1. compared to previous methods that focused on CMA feature extraction, using sequence information can enhance feature expression. 2. It can perform whole-genome quantitative testing and analysis, with almost every reading fragment well-characterized by its associated RNA sequence, and without experimental cross-reactivity or noise interference. Furthermore, combining sequence information with interaction information yields highly reliable feature vectors of behavioral attributes, improving the quality of feature extraction. (See attached figures.) Figure 1 This is a flowchart of a CircRNA-miRNA association prediction method.

[0019] Figure 2 The flowchart shows the operation of the CMALF model, where: A. Nodes and node topology features are constructed by preprocessing the heterogeneous association graph; B. Low-dimensional representation features of circRNA and miRNA node sequences and topology are learned through word embedding and graph embedding models; C. Flowchart of GBDT prediction.

[0020] Figure 3 This is a flowchart of the operation of a homogeneous embedding fusion model.

[0021] Figure 4 This is a schematic diagram of AUC and AUPR values ​​implemented using CMALF, where: A. AUC is obtained by summing the areas under panel A where the ROC curve is plotted; B. AUPR refers to the area under the curves enclosed by precision and recall.

[0022] Figure 5 This diagram illustrates the comparison of prediction performance for different strategies.

[0023] Figure 6The graphs are: A. ROC curve generated using the LAP results on the dataset with five-fold cross-validation; B. AUPR curve generated using the results on the dataset with five-fold cross-validation.

[0024] Figure 7 The results show a 5x CV for different classifier models.

[0025] Figure 8 This is a structural diagram of a CircRNA-miRNA association prediction device. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples, but this is not intended to limit the present invention. If there is no necessary sequential relationship between the various steps described herein, the order in which they are described as examples should not be considered a limitation. Those skilled in the art should understand that the order can be adjusted, as long as it does not disrupt the logical consistency between them and render the entire process impossible.

[0027] Example 1: This invention provides a method for predicting circRNA-miRNA associations, such as... Figure 1 and Figure 2 As shown, the method includes the following steps: Step S100: Construct a dataset, which includes multiple circRNAs and multiple miRNAs.

[0028] For example, to collect high-reliability sequence information of circRNA and miRNA to evaluate the performance of the CMALF model, we utilize the experimentally validated CircBank and miRbase databases, which are currently available, as high-quality selection datasets to evaluate the quality of the aforementioned model. Therefore, the constructed dataset can be described as follows: (1) In this embodiment, 20208 pairs of experimentally validated CMAs were first selected from the constructed dataset, involving 3569 circRNAs and 1152 miRNAs. and These represent the sets of 20,208 positive samples and 4,091,280 negative samples, respectively. This represents the union of elements. Next, we construct the adjacency matrix of the dataset with dimension 3569×1152. When miRNA ( ) and circRNA ( When unrelated, of It is set to 0; otherwise, its value is 1.

[0029] Step S200: Based on the dataset, extract biological attribute features from the word embeddings of CircRNA and miRNA sequences.

[0030] For example, biological attribute features can be extracted based on Word2vec.

[0031] Word2Vec was proposed by Google. The popular word2vec is a word embedding model that uses relationships between words appearing in a sentence to obtain word vectors in a high-dimensional to low-dimensional space in machine learning. Similarly, nodes are used as words, and sequences of nodes are used as sentences, as described in the sequence analysis of circRNA and miRNA node representations by the word2vec algorithm.

[0032] Here, word2vec utilizes the extraction of circRNA and miRNA sequence features in the vector space dependent on the CBOW (bag-of-words) model, which is used to characterize the node features in this study. The optimal target word determined by CBOW is used for prediction before and after the target word, where the objective function of the CBOW output value is as follows: (2) here It is a weight matrix. The parameters representing the target word and the context words are treated as... .

[0033] In this embodiment, by using the gensim-python package of the word2vec model to train circRNA and miRNA sequence features, 64-dimensional target features can be obtained.

[0034] Step S300: Extract behavioral features based on isomorphic graph embedding.

[0035] Isomorphic graph embedding is a graph representation learning method that aims to preserve graph topology while learning low-dimensional representations of vertices. It is also widely recognized as network embedding or non-attribute graph embedding, encompassing 11 methods based on random walks, deep learning, and matrix factorization. Therefore, this embodiment selects GF and LAP from matrix factorization, and LINE from deep learning.

[0036] In the constructed dataset, multi-source feature information includes its impact on circRNA-miRNA interactions. Heterogeneity maps were obtained using validated interaction pairs and an equal number of unlabeled samples, enabling the construction of a reliable molecular association network, such as... Figure 2 As shown in the image in the upper right corner, behavioral features are extracted using isomorphic graph embedding based on this molecular association network.

[0037] Specifically, a low-dimensional dense vector is learned by using a multilayer perceptron to maintain first-order and second-order similarity. The first-order and second-order objective functions are shown below. (3) (4) Here, set and Representing vertices and edges respectively, this means that in a graph middle, from arrive . It represents the distance between two distributions. and It is considered a representation of the empirical distribution. and These represent the joint distribution and the contextual conditional distribution, respectively.

[0038] To address the storage and construction issues of the downstream classifier, GF is chosen to perform factorization on the adjacency matrix of the graph. The objective function is described as follows: (5) in and These represent the weight adjacency matrix, factor matrix, and regularization parameter, respectively.

[0039] In some embodiments, during the behavioral feature extraction process, a fusion graph embedding model (LPA) based on the integration of LINE and GF is proposed, combining the advantages of deep learning and factorization, such as... Figure 3 As shown.

[0040] This embodiment proposes a LAP model, a matrix factorization method designed to decompose high-dimensional matrices to obtain embeddings. Specifically, the obtained objective function is minimized using Laplacian eigenmaps.

[0041] (6) in and It is the embedding of nodes and the embedding of the graph represented by Y. and They are respectively considered as the Laplace operators representing the weights and graph G.

[0042] This paper applies the established open-NE library to behavioral feature extraction in graph embedding models. After training the model, a vector with 64 dimensions for each node is obtained.

[0043] Step S400: Based on the biological attribute characteristics and the behavioral characteristics, obtain the training characteristics of the fusion vector to learn the interaction scores of circRNA and miRNA and infer potential CMAs.

[0044] This step is a prediction process that can be implemented based on the decision tree method.

[0045] For example, a classifier is proposed that can be used for classification, regression, or feature filtering. Under this strategy, a weak classifier is then set up, and the loss function is calculated by iterating over multiple iterations and using the residuals of the previous classifier. In this embodiment, the loss function of the weak classifier is expressed as... (7) here This represents the iteration result at this point. We set the weak classifier (a decision tree) as... The parameter value; the number of iterations, which we denote as m.

[0046] Example 2: Predictive performance of the CMALF model The CMALF model is a model constructed based on the method described in Example 1. It can achieve CircRNA-miRNA association prediction. The purpose of this example is to illustrate the predictive performance of the model, so as to demonstrate the feasibility and progress of the present invention.

[0047] For evaluation in machine learning, different commonly used metrics are used to measure the predictive performance of the proposed CMALF model, including Specificity, Precision, Sensitivity, Matthews correlation coefficient (MCC), and Accuracy. These evaluation metrics are defined as follows: (8) (9) (10) (11) (12) Here, the abbreviations for the evaluation criteria include TP for true positive, TN for true negative, FP for false positive, and FN for false negative. Furthermore, we visualized the ROC (Receiver-Operating-Characteristic) curves by calculating the TP and FP rates generated by CMALF, and calculated their mean AUC (Area Under Precision-Recall (PR)) to account for imbalance. Additionally, reliable 5-fold cross-validation (CV) was performed.

[0048] In the evaluation, we used our constructed dataset and employed a 5-fold cross-validation approach to examine the CMALF model's ability to predict latent CMAs in terms of Acc., Sen., Spe., Pre., MCC, AUC, and AUPR. All experimental results and the mean predictions (highlighted in bold) are shown in Table 1. CMALF achieved a mean AUC of 0.9075 and a standard deviation of 0.0016, with AUCs of 0.9067, 0.9103, 0.9076, 0.9070, and 0.9061 for the 5-fold cross-validation experiments. Figure 4 In the diagram, AUC can be obtained by summing the areas under the ROC curves in panel A, while AUPR refers to the area under the PR curves enclosed by precision and recall in panel B, with the larger areas reaching the top left and top right corners of the image, respectively. Overall, the above statistics demonstrate that the model has state-of-the-art performance, effectively predicting potential CMAs and providing strong evidence for a deeper understanding of the circRNA-miRNA relationship.

[0049] Table 1. Results of the five-fold cross-validation experiment conducted by CMALF.

[0050] Example 3: Comparison of different feature extraction strategies To confirm whether the two different graph embedding methods outperformed the proposed model, we compared the test LINE and GF models with CMALF, naming them CMALF-LINE and CMALF-GF, respectively. Results from the same five-fold cross-validation experiments conducted on the same dataset with CMALF-LINE and CMALF-GF show… Figure 3The specific values ​​in the histogram are shown in Tables 2 and 3. Based on the data in Tables 1 and 2, the average values ​​of Acc., Sen., Spe., Pre., MCC, and AUC obtained from the CMALF-LINE model are 3.65%, 3.32%, 5.77%, 5.33%, 9.1%, and 0.0342 lower than those of the model proposed in this paper, respectively. Compared with Table 1, Table 3 shows that the average value of the CMALF-GF model is lower than that of the constructed model. Similarly, Figure 5 The comparison of prediction performance also demonstrates the advantages of CMALF. In summary, CMALF's feature extraction performance is superior to the two unilateral information feature extraction strategies.

[0051] Table 2 Results obtained by CMALF-LINE through five-fold cross-validation experiment

[0052] Table 3 Results of CMALF-GF obtained through five-fold cross-validation experiment

[0053] Example 4: Comparison with the Laplace eigenmap model To infer the effectiveness of the proposed model in optimizing its performance by utilizing biological attribute features and multi-source behavioral features as model attributes, we compared it with behavioral topological feature vectors generated by the LAP method. For fairness and consistency, we used low-dimensional embedding vectors generated by LAP instead of the graph embeddings of the proposed fusion model during the testing process, while keeping other parts of the model unchanged. Using our constructed dataset, we trained the model with the LAP method using a five-fold cross-validation experiment, and the results are shown in Table 4. As can be seen from Table 4, CMALF achieved better results, with its predictions of Acc., Spe., Pre., MCC, and AUC being 8.91%, 22.6%, 14.4%, 16.18%, and 0.0814 higher than the LAP model, respectively. This result demonstrates that the fusion model combining LINE and GF used in the proposed model can effectively build vector features and train the computer model, which helps improve the model and achieve the most promising predictive performance. The advantages of CMALF's AUPR and ROC curves can be seen from... Figure 5 and Figure 6 This can be seen from the comparison.

[0054] Table 4 Results of 5x CV obtained through LAP and CMALF

[0055] Example 5: Comparison of different classifier models To ensure the optimal classification method was selected through feature extraction, we compared different classifier models to evaluate the impact of CMALF on features and performance in this study. Specifically, we retained the invariant biological attribute features and behavioral topology feature extraction methods, replacing the GBDT model with only five different classifiers: KNN (K-Nearest Neighbour), LR (Logistic Regression), RF (Rotation Forest), SVM (Support Vector Machine), and AdaBoost. Table 5 lists the average values ​​of the five-fold cross-validation experiments achieved by the above models on the same dataset, as shown in the bars. Figure 5 As shown in Table 5, AdaBoost ranks second in Acc., Sen., Spe., Pre., MCC, and AUC, but is 1.54%, 1.26%, 1.82%, 1.76%, 3.09%, and 0.0189% lower than the best results, respectively. Figure 7 The comparisons also show that the CMALF model yields the best results. In short, these results demonstrate that our GFBDT classifier, the CMALF model, outperforms other classifier models.

[0056] Table 5. Average results obtained using different classifier models.

[0057] Example 6: Comparison with other advanced methods With the deepening of research on CMAs in recent years, many famous scholars have proposed different CMAs prediction methods. To evaluate its prediction performance, CMALF is compared more fairly with the above methods. Since there is a precise comparison consideration here, we calculated the AUC and AUPR scores generated in the previous models. These results are listed in Table 6, which includes our model and the CMIVGSD, WSCD, KGDCMI

[20] , and SGCNCMI models of several newly published papers in the new research field of CMAs prediction. The table shows that CMALF achieved the highest AUC and AUPR scores, which are 0.0133 and 0.0233 better than the second best SGCNCMI model, respectively, and about 0.0198 and 0.0372 higher than the average of the other three methods. Therefore, the above comparison can conclude that CMALF can provide the most competitive theoretical guidance for further academic research.

[0058] Table 6. AUC and AUPR scores obtained by various models

[0059] Example 7: Case Study To further investigate the effectiveness of CMALF in identifying novel miRNA candidate circRNAs, we conducted a case study using the trained model with known miRNA-associated circRNAs and predicted all unknown CMAs using the trained model. Then, candidates for unknown association pairs were ranked according to their highest scores, and the validity and accuracy of the predictions were ensured by searching relevant research literature or experiments. The model prediction results are shown in Table 7. From this, we can see that only 7 out of the top 30 miRNA-associated circRNA pairs in recent literature were not validated. Overall, this case study demonstrates that CMALF has superior predictive performance in predicting potential CMAs, and these valuable circRNA candidates for miRNA research may be selected for additional wet laboratory experiments to reduce the shortcomings of manual errors.

[0060] Table 7 Top 30 CMALF Predictions

[0061] Based on Examples 1-7, this paper proposes CMALF, a graph representation learning framework based on a fusion model that combines two isomorphic graph embedding methods to predict CMAs. We first construct low-dimensional embedding vectors based on sequence information using word embeddings, and then construct a low-dimensional representation based on a fusion model combining LINE and GF, while preserving the graph's topology and node attributes. Next, we obtain training features for the fusion vector by combining biological attribute features and multi-source behavioral features to learn the interaction scores between circRNA and miRNA. Overall evaluation of the constructed dataset confirms that CMALF outperforms other excellent models in CMA prediction. To evaluate the best performance of our paper, we conducted several ablation experiments on the same dataset, including comparisons of the CMALF-LINE model, CMALF-GF, and LAP models, different experimental classifier models, and the performance of the association prediction model compared with other state-of-the-art models. Overall evaluation of the constructed dataset confirms that CMALF outperforms other advanced models in CMA prediction. Through case studies in recent literature, higher association scores were obtained for the prediction of a large number of experimental outcome pairs. Careful analysis of the results shows that our proposed model achieves excellent prediction of potential CMAs in academic research.

[0062] Example 8: Apparatus for CircRNA-miRNA Association Prediction Method This invention also provides an apparatus for a CircRNA-miRNA association prediction method, such as... Figure 8 As shown, the device 800 includes: Construction module 801 is configured to construct a dataset that includes multiple circRNAs and multiple miRNAs; The first extraction module 802 is configured to extract biological attribute features from word embeddings of CircRNA and miRNA sequences based on the dataset. The second extraction module 803 is configured to extract behavioral features based on isomorphic graph embedding. The prediction module 804 is configured to obtain training features of the fusion vector based on the biological attribute features and the behavioral features, in order to learn the interaction scores of circRNA and miRNA and infer potential CMAs.

[0063] In some embodiments, the building module is further configured to describe the constructed dataset using formula (1): (1) In the formula, D represents the union of elements. and These represent the sets of 20,208 positive samples and 4,091,280 negative samples, respectively.

[0064] In some embodiments, the dimension is constructed as follows: m × n The adjacency matrix of the dataset, where m Indicates the number of circRNAs. n Indicates the number of miRNAs; If miRNA ( ) and circRNA ( If they are unrelated, then In Set to 0; otherwise, In The value is 1, where miRNA ( ) represents the first element in the adjacency matrix. i miRNAs, circRNAs ( ) represents the first element in the adjacency matrix. j One circRNA, In the adjacency matrix, the th i The miRNA and the first j The correlation of individual circRNAs.

[0065] In some embodiments, the first extraction module is further configured to: Provide extraction models to extract biological attribute features; The objective function for extracting the output value of the model is expressed as: (2) In the formula, Represents the weight matrix. The parameters representing the target word and the context words are treated as... , T Indicates the number of samples. t This indicates the current sample number.

[0066] In some embodiments, the second extraction module is further configured to: A multilayer perceptron is used to learn low-dimensional dense vectors while maintaining first-order and second-order similarity, where the first-order and second-order objective functions are expressed as follows: (3) (4) In the formula, O 1 represents a first-order objective function. O 2 represents the second-order objective function. Indicates the distance between two distributions. and It is considered a representation of the empirical distribution. and Let them represent the joint distribution and the contextual conditional distribution, respectively. i Represents the regularization parameter; Factorize the adjacency matrix of the graph. The objective function of factorization is expressed as: (5) In the formula, and These represent the weight adjacency matrix, factor matrix, and regularization parameter, respectively. E Indicates the size of the adjacency matrix. Z i Represents the factor vector in the i-th row. Y ij This represents the weight of the i-th row and j-th column. Z j This represents the factor vector of the j-th column.

[0067] In some embodiments, the second extraction module is further configured to: It provides a fusion graph embedding model to extract behavioral features; The objective function of the fusion graph embedding model is expressed as: (6) In the formula, Y Represents graph embedding. and Let Y represent the embedding of nodes and the embedding of the graph, respectively. Indicates weight, L Describe the Laplace operator of graph G. Y T This represents the transpose of a graph embedding. tr This represents minimizing the objective function.

[0068] In some embodiments, the prediction module is further configured to: A classifier is provided to learn the interaction scores between circRNA and miRNA and infer potential CMAs.

[0069] In some embodiments, the prediction module is further configured to provide a weak classifier for multiple iterations and the residuals of previous classifiers; The objective function of the weak classifier is expressed as: (7) In the formula, This represents the iteration result at this point. Represents the single-sample objective function. N Indicates the number of samples. y i Indicates a label, L () represents the loss function. T ( x i ; ) indicates error.

[0070] It should be noted that the device proposed in this embodiment and the method described earlier belong to the same technical concept, have the same working principle, and can achieve the same beneficial effect, which will not be elaborated here.

[0071] Finally, this embodiment of the invention also provides a readable storage medium storing one or more programs, which can be executed by one or more processors to implement the CircRNA-miRNA association prediction method as described in any of the above embodiments.

[0072] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on the invention that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, and such examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0073] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments may be used by those skilled in the art upon reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the invention. This should not be construed as an intention that a feature of an unclaimed invention is necessary for any claim. Rather, the subject matter of the invention may be less than all the features of a particular embodiment of the invention. Thus, the following claims are incorporated herein by reference as examples or embodiments, wherein each claim is independently considered as a separate embodiment, and these embodiments are contemplated as being able to be combined with each other in various combinations or arrangements. The scope of the invention should be determined by reference to the appended claims and the full scope of their equivalents.

Claims

1. A method for predicting circRNA-miRNA association, characterized in that, The method includes: Construct a dataset that includes multiple circRNAs and multiple miRNAs; Based on the dataset, biological attribute features were extracted from the word embeddings of CircRNA and miRNA sequences; Behavioral features are extracted based on isomorphic graph embedding; Training features of the fusion vector are obtained based on the biological attributes and behavioral features to learn the interaction scores between circRNA and miRNA and infer potential CMAs. The extraction of behavioral features based on isomorphic graph embedding specifically includes: A multilayer perceptron is used to learn low-dimensional dense vectors while maintaining first-order and second-order similarity, where the first-order and second-order objective functions are expressed as follows: （3） （4） In the formula, O 1 represents a first-order objective function. O 2 represents the second-order objective function. Indicates the distance between two distributions. and It is considered a representation of the empirical distribution. and Let them represent the joint distribution and the contextual conditional distribution, respectively. i Represents the regularization parameter; Factorize the adjacency matrix of the graph. The objective function of factorization is expressed as: （5） In the formula, and These represent the weight adjacency matrix, factor matrix, and regularization parameter, respectively. E Indicates the size of the adjacency matrix. Z i Represents the factor vector in the i-th row. Y ij This represents the weight of the i-th row and j-th column. Z j Represents the factor vector of the j-th column; The extraction of behavioral features based on isomorphic graph embedding specifically includes: It provides a fusion graph embedding model to extract behavioral features; The objective function of the fusion graph embedding model is expressed as: （6） In the formula, Y Represents graph embedding. and Let Y represent the embedding of nodes and the embedding of the graph, respectively. Indicates weight, L Describe the Laplace operator of graph G. Y T This represents the transpose of a graph embedding. tr This represents minimizing the objective function.

2. The method according to claim 1, characterized in that, The constructed dataset is described by formula (1): （1） In the formula, D represents the union of elements. and These represent the sets of 20,208 positive samples and 4,091,280 negative samples, respectively.

3. The method according to claim 2, characterized in that, Construct a dimension of m × n The adjacency matrix of the dataset, where m Indicates the number of circRNAs. n Indicates the number of miRNAs; If miRNA ( ) and circRNA ( If they are unrelated, then In Set to 0; otherwise, In The value is 1, where miRNA ( ) represents the first element in the adjacency matrix. i miRNAs, circRNAs ( ) represents the first element in the adjacency matrix. j One circRNA, In the adjacency matrix, the th i The miRNA and the first j The correlation of individual circRNAs.

4. The method according to claim 1, characterized in that, The extraction of biological attribute features from word embeddings of CircRNA and miRNA sequences based on the dataset specifically includes: Provide extraction models to extract biological attribute features; The objective function for extracting the output value of the model is expressed as: （2） In the formula, Represents the weight matrix. The parameters representing the target word and the context words are treated as... , T Indicates the number of samples. t This indicates the current sample number.

5. The method according to claim 1, characterized in that, The process of obtaining training features for the fusion vector based on the biological attributes and behavioral characteristics, in order to learn the interaction scores between circRNA and miRNA and infer potential CMAs, specifically includes: A classifier is provided to learn the interaction scores between circRNA and miRNA and infer potential CMAs.

6. The method according to claim 5, characterized in that, Provide a weak classifier, which is used for multiple iterations and the residual of previous classifiers; The objective function of the weak classifier is expressed as: （7） In the formula, This represents the iteration result at this point. Represents the single-sample objective function. N Indicates the number of samples. y i Indicates a label, L () represents the loss function. T ( x i ; ) indicates error.

7. A CircRNA-miRNA association prediction device for implementing the method as described in any one of claims 1 to 6, characterized in that, The device includes: The building module is configured to build a dataset that includes multiple circRNAs and multiple miRNAs; The first extraction module is configured to extract biological attribute features from word embeddings of CircRNA and miRNA sequences based on the dataset. The second extraction module is configured to extract behavioral features based on isomorphic graph embedding. The prediction module is configured to obtain training features of the fusion vector based on the biological attribute features and the behavioral features, in order to learn the interaction scores of circRNA and miRNA and infer potential CMAs.

8. A readable storage medium, characterized in that, The readable storage medium stores one or more programs, which can be executed by one or more processors to implement the method as described in any one of claims 1 to 6.

Images