Circular RNA and drug association prediction method and system

Abstract

This invention discloses a method and system for predicting associations between circular RNA and drugs. The method includes: S1. Constructing a multi-source similarity network for circular RNA and a multi-source similarity network for drugs based on a known dataset of associations between circular RNA and drugs; S2. Inputting each similarity network into a shared constraint-driven collaborative feature learning module to output a collaborative feature matrix; S3. Inputting the collaborative feature matrix into a minimum entropy graph structure learning module to obtain a graph structure matrix; S4. Inputting the graph structure matrix and the collaborative feature matrix into a graph convolutional network for predicting the association between circular RNA and drugs. The collaborative feature learning method proposed in this scheme fully utilizes the different information between multi-source data, enabling full interaction of multi-source information to obtain more comprehensive features. In addition, the graph structure learning module uses the minimum entropy learning method to learn more discriminative features, enhances graph topological features, further optimizes feature representation, and effectively improves the prediction accuracy and generalization ability of the model.

Description

Technical Field

[0001] This invention belongs to the field of computer bioinformatics, and particularly relates to a method and system for predicting the association between circular RNA and drugs based on collaborative feature learning and minimum entropy graph structure learning. Background Technology

[0002] In the field of bioinformatics, predicting the association between circular RNA and drugs is a key technology for drug localization and identifying therapeutic targets. Traditional correlation prediction methods often rely on building models based on single biological features and integrating data using simple feature splicing or fixed-weight fusion, making it difficult to fully explore the complementary information hidden in multi-source data.

[0003] To predict the association between RNA and diseases, this applicant proposed a dual-view scheme [Publication No. CN119400252B] that constructs a similarity network and a meta-path network. This scheme introduces shared units, a multi-channel attention mechanism, and a contrastive learning strategy to achieve cross-view interaction of multi-view information and dynamic allocation of view weights. However, on the one hand, this scheme is used for predicting the association between circular RNA and diseases, and does not involve predicting the association between circular RNA and drugs. On the other hand, the scheme design focuses on multi-view information interaction and view weight optimization. When dealing with scenarios with sparse data and high graph structure noise, the insufficient integrity of the meta-path network makes it difficult for shared units to capture effective cross-view information, resulting in a significant drop in prediction performance. Therefore, it is not applicable to the scenario of predicting the association between circular RNA and drugs with few known associations. Summary of the Invention

[0004] The purpose of this invention is to propose a method and system for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, addressing the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solutions:

[0006] A method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, the method comprising:

[0007] S1. Based on the known association datasets between circular RNA and drugs, construct multi-source similarity networks for circular RNA and drugs respectively;

[0008] S2. Input the multi-source similarity network of the circular RNA and the multi-source similarity network of the drug into the shared constraint-driven collaborative feature learning module, and output the collaborative feature matrix;

[0009] S3. Input the aforementioned collaborative feature matrix into the minimum entropy graph structure learning module to obtain the graph structure matrix;

[0010] S4. Input the graph structure matrix and the collaborative feature matrix into a graph convolutional network for predicting the association between circular RNA and drugs.

[0011] In the above-mentioned method for predicting the association between circular RNA and drugs based on collaborative feature learning and minimum entropy graph structure learning, in step S1, the multi-source similarity network of circular RNA includes the sequence similarity, information entropy similarity, and Gaussian interaction kernel similarity of circular RNA.

[0012] The multi-source similarity network of the drug includes structural similarity, information entropy similarity, and Gaussian interaction kernel similarity.

[0013] In the above-mentioned circular RNA and drug association prediction methods based on collaborative feature learning and minimum entropy graph structure learning, three similarity networks for circular RNA are constructed in the following ways:

[0014] The host gene information of each circular RNA is obtained from a biological database, and the sequence similarity of the circular RNA is calculated based on the host gene information.

[0015] The information entropy of a single circular RNA is calculated based on the known association dataset between the circular RNA and drugs. Then, the information entropy similarity of the circular RNA is calculated based on the common association set of drugs between the circular RNA and other circular RNAs.

[0016] A circular RNA-drug association matrix is ​​constructed based on the known associated dataset. The Gaussian interaction kernel similarity of the corresponding circular RNAs is calculated based on the column vectors corresponding to the circular RNAs in the circular RNA-drug association matrix.

[0017] In the above-mentioned method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, three drug similarity networks are constructed in the following ways:

[0018] The SMILES structure data of each drug are searched in the bioinformatics database, and the topological fingerprint spectrum of each drug is obtained using relevant tools. The drug structure similarity is calculated based on the topological fingerprint spectrum.

[0019] The information entropy of a single drug is calculated based on the known associated dataset, and then the drug information entropy similarity is calculated based on the common associated circular RNA set of the drug and other drugs.

[0020] Based on the row vectors corresponding to drugs in the circular RNA-drug association matrix, the Gaussian interaction kernel similarity of the corresponding drugs is calculated.

[0021] In the above-described method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, step S2 includes a collaborative feature learning module comprising a first autoencoder, a shared constraint module, and a second autoencoder, and step S2 further includes:

[0022] The multi-source similarity network is input into the first autoencoder to generate the node feature matrix corresponding to each similarity network, which is then input into the shared constraint module.

[0023] The shared constraint module uses the Pearson correlation coefficient to calculate the similarity between any two nodes, filters out node pairs with significant similarity, and constructs a constraint matrix containing information on the significantly similar node pairs.

[0024] The common constraint set of circular RNAs is obtained by performing an intersection operation on the constraint matrices corresponding to all similarity networks of circular RNAs.

[0025] The intersection operation is performed on the constraint matrices corresponding to all drug similarity networks to obtain the shared constraint set of the drugs;

[0026] The shared constraint set of circular RNA and the shared constraint set of drugs are input together with the low-dimensional node features output by the corresponding first autoencoder into the subsequent second autoencoder, and finally the enhanced similarity features of circular RNA and drugs are obtained.

[0027] By splicing the similarity features enhanced by circular RNA and the similarity features enhanced by drugs, the synergistic feature matrix is ​​obtained.

[0028] In the above-mentioned method for predicting circular RNA and drug association based on collaborative feature learning and minimum entropy graph structure learning, the minimum entropy graph structure learning module in step S3 includes an optimization view and an inference view.

[0029] The optimized view is constructed based on the initial adjacency matrix A and the cooperative feature matrix X;

[0030] The inference view is based on the adjacency matrix. And the collaborative feature matrix X is constructed as follows;

[0031] The adjacency matrix The initial adjacency matrix A is obtained by processing the graph learner with a nonlinear activation function and then normalizing and symmetricizing it;

[0032] The initial adjacency matrix A is constructed based on the circular RNA-drug association matrix.

[0033] In the above-mentioned method for predicting circular RNA and drug association based on collaborative feature learning and minimum entropy graph structure learning, the minimum entropy graph structure learning module introduces edge dropping and feature masking strategies, and adopts different feature masking probabilities for the optimized view and the inference view, while adopting the same edge dropping probability.

[0034] After enhancing the view and the inference view by edge dropping and feature masking, the node representations of the two enhanced views are extracted using a graph encoder and mapped to the final feature vectors by a multilayer perceptron.

[0035] In the aforementioned method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, in step S3, the minimum entropy graph structure learning module learns by minimizing the entropy value of the output, and the minimum entropy loss function... as follows:

[0036]

[0037] These are the final feature vectors of the optimized view and the inference view, respectively;

[0038] They are respectively , The dimension in which the maximum value is located;

[0039] , These are the normalized features of the final feature vectors of the optimized view and the inference view, respectively.

[0040] This represents the cross-entropy loss function.

[0041] In the above-described method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, step S4 includes:

[0042] The graph structure matrix obtained in step S3 and the collaborative feature matrix output in step S2 are input together into the graph convolutional network;

[0043] The training method uses the difference between the true and predicted values ​​of circular RNA-drug association as the objective and the mean squared error as the loss function.

[0044] A circular RNA and drug association pre-system based on collaborative feature learning and minimum entropy graph structure learning is used to perform the above methods.

[0045] The advantages of this invention are:

[0046] This scheme designs a collaborative feature extraction method to deeply mine the shared constraints of the multi-source similarity network between circular RNA and drugs. By screening the necessary link pairs, a global shared constraint matrix is ​​constructed, which enables the multi-source data to fully interact and complement each other. Even in scenarios where there are few known associations and the data is sparse, the collaborative information from multiple sources can be used to make up for the deficiencies, improve the completeness of feature representation, and thus improve the performance of the model in predicting the association between circular RNA and drugs.

[0047] This invention introduces a graph structure learning module, which uses the minimum entropy learning method to learn more discriminative features, enhances graph topological features, further optimizes feature representation, and effectively improves the model's prediction accuracy and generalization ability. Attached Figure Description

[0048] Figure 1 The diagram shown is a flowchart of a circular RNA and drug association prediction based on collaborative feature learning and minimum entropy graph structure learning provided in an embodiment of the present invention.

[0049] Figure 2 The diagram shown is a schematic of the principle of circular RNA and drug association prediction based on collaborative feature learning and minimum entropy graph structure learning provided in an embodiment of the present invention.

[0050] Figure 3 The figure shows the predictive ability of the model evaluated by the present invention under 5-fold and 10-fold cross-validation.

[0051] Figure 4 The figure shows a comparison of the AUC results of this invention with other methods;

[0052] Figure 5 The diagram shows a comparison of the ablation results of the present invention on a dataset. Detailed Implementation

[0053] like Figure 1 and Figure 2 As shown, this scheme provides a circular RNA-drug association prediction method based on collaborative feature learning and minimum entropy graph structure learning. By constructing multi-source information on circular RNA and drugs, collaborative learning is used to learn the collaborative feature matrix from the multi-source information, and then graph structure learning is used to obtain the graph structure matrix features, thereby more accurately predicting the potential association between circular RNA and drugs. The specific technical solution is as follows:

[0054] S1. Collect and organize known association datasets between circular RNA and drugs from public databases;

[0055] S2. Based on the acquired known associated datasets, construct a multi-source similarity network for circular RNA: sequence similarity, information entropy similarity, and Gaussian interaction kernel similarity; and a multi-source similarity network for drugs: structural similarity, information entropy similarity, and Gaussian interaction kernel similarity.

[0056] S3. Input the multi-source similarity networks of circular RNA and drugs into the shared constraint-driven collaborative feature learning module, and the collaborative feature learning module learns the multi-source information to obtain the collaborative feature matrix;

[0057] S4. Input the obtained collaborative feature matrix into the minimum entropy graph structure learning module to obtain the graph structure matrix through graph structure learning;

[0058] S5. Use the obtained graph structure matrix and co-feature matrix to train the graph convolutional network GCN for predicting the association between circular RNA and drugs.

[0059] Specifically, the construction process of the three multi-source similarity networks of circular RNA and drug in step S2 includes the following steps:

[0060] (1) Constructing circular RNA sequence similarity CSS

[0061] First, host gene information is searched from the NCBI bioinformatics database. Then, the Levenshtein distance between host genes is calculated, and finally, the circular RNA sequence similarity (CSS) is obtained. The calculation formula is as follows:

[0062]

[0063] in, These represent the circular RNA sequences being compared;

[0064] This indicates the specific position of the current nucleotide residue in the circular RNA sequence a;

[0065] This indicates the specific position of the current nucleotide residue in the circular RNA sequence b;

[0066] The Levenshtein distance is calculated from the first i-1 nucleotide residues of circular RNA sequence a and the first j nucleotide residues of circular RNA sequence b, and so on.

[0067] (2) Analyze the structure of drugs and construct a structural similarity DSS between drugs.

[0068] First, SMILES structural data were obtained from the PubChem bioinformatics database. Then, topological fingerprints were obtained using RDKit, and finally, the Tanimoto method was used to calculate structural similarity. The formula is as follows:

[0069]

[0070] in, and These represent the topological fingerprint vectors of drugs A and B, respectively.

[0071] (3) Constructing the information entropy similarity CES of circular RNA and the information entropy similarity DES of drugs.

[0072] Taking circular RNA as an example, the formula for calculating the information entropy of a single circular RNA is:

[0073]

[0074]

[0075] AN represents the number of circular RNA-drug association pairs. This refers to all circular RNAs that are associated with drugs. The number of related quantities.

[0076] The formula for calculating the information entropy similarity of circular RNAs is:

[0077]

[0078] This represents a set of drugs that are associated with both circular RNA A and circular RNA B.

[0079] Based on the same calculation process, the information entropy similarity (DE) of the drugs can be obtained, which will not be elaborated here.

[0080] (4) The Gaussian interaction kernel similarity (CGS) between circular RNAs is calculated using the Gaussian interaction kernel. The formula is:

[0081]

[0082] in: and Drugs and Column vectors in the circular RNA-drug association matrix;

[0083] The normalized bandwidth parameter is calculated using the following formula:

[0084]

[0085] in, The number of columns in the incidence matrix. Let be the norm of each column vector.

[0086] (5) The Gaussian interaction kernel similarity (DGS) between drugs is calculated using the Gaussian interaction kernel. The formula is:

[0087]

[0088] and Circular RNA and Row vectors in the incidence matrix;

[0089] The normalized bandwidth parameter is calculated using the following formula:

[0090]

[0091] in, The number of rows in the correlation matrix. Let be the norm of each row vector.

[0092] The above yielded a multi-source similarity network for circular RNAs: sequence similarity, information entropy similarity, and Gaussian interaction kernel similarity; and a multi-source similarity network for drugs: structural similarity, information entropy similarity, and Gaussian interaction kernel similarity.

[0093] Specifically, step S3 includes the following processes:

[0094] (1) Use a multi-source similarity network as the input features of an autoencoder.

[0095] Using circular RNA similarity networks For example, a graph encoder is used to encode it to generate a representation vector for the node.

[0096]

[0097] in, Indicates the first A circular RNA network The node feature matrix, Indicates the first The encoder of the network.

[0098] (2) Construct shared constraints based on the similarity between nodes. Specifically, the Pearson Correlation Coefficient (PCC) is used to calculate the similarity between any two nodes. and Similarity between them:

[0099]

[0100] in, and Representing nodes respectively and exist The corresponding row in The dimension of the feature space. and This represents the mean of the corresponding vector.

[0101] (3) After obtaining the PCC similarity of all circular RNA networks, select some node pairs with significant similarity, set the corresponding positions of these node pairs to 1, and set the remaining positions to 0, thereby constructing the constraint matrix. .

[0102]

[0103] (4) Perform an intersection operation on the constraint matrices (MLMs) corresponding to all circular RNA networks to obtain a global constraint set shared by all networks, denoted as . ,Right now

[0104]

[0105] Here This represents the intersection operation of sets. The set of shared constraints. It includes node pairs that are considered to have strong similarity across all networks—pairs that must be linked.

[0106] (5) The shared constraint and the low-dimensional features output by the first autoencoder are input together into the subsequent second autoencoder to further guide the feature learning process.

[0107]

[0108] in, This represents the k-th decoder.

[0109] (6) After the above-mentioned collaborative feature learning, enhanced similarity features between circular RNA and drugs were obtained. Finally, these features were integrated to form a comprehensive collaborative feature matrix. Its structure is as follows:

[0110]

[0111] The feature matrix is ​​obtained by co-feature learning of circular RNA sequence similarity.

[0112] The feature matrix is ​​obtained after collaborative feature learning of drug structure similarity;

[0113] The feature matrix is ​​obtained by co-feature learning of the information entropy similarity of circular RNAs.

[0114] The feature matrix is ​​obtained by collaborative feature learning of drug entropy similarity.

[0115] The feature matrix is ​​obtained by co-feature learning of Gaussian-interacting nuclear similarity of circular RNA.

[0116] The feature matrix is ​​obtained by co-feature learning of the Gaussian similarity of the drug action kernel.

[0117] Furthermore, step S4, which involves inputting the obtained collaborative features into the minimum entropy graph structure learning module, specifically includes:

[0118] (1) In this embodiment, a graph learner (FGP) is used to learn the adjacency matrix. The model is constructed as follows:

[0119]

[0120] Where σ represents a nonlinear activation function;

[0121] A represents the initial adjacency matrix, constructed based on the circular RNA-drug association matrix:

[0122]

[0123] This represents the circular RNA-drug association matrix.

[0124] (2) Normalize and symmetricize the adjacency matrix S to obtain the processed adjacency matrix. .

[0125]

[0126] Where D is a matrix The degree matrix is ​​used to achieve symmetric normalization; It is a non-linear activation function.

[0127] (3) The minimum entropy graph structure learning module includes an inference view and an optimization view as a dual-view mechanism. The optimization view is constructed based on the initial adjacency matrix A and the collaborative feature matrix X extracted by the collaborative feature learning module:

[0128]

[0129] The inference view is composed of the adjacency matrix Together with the collaborative feature matrix X output by the collaborative feature learning module, it constitutes:

[0130]

[0131] (4) The minimum entropy graph structure learning module introduces edge discarding and feature masking strategies, which helps the model extract richer semantic information.

[0132] Specifically, different feature masking probabilities are applied to the two views: Optimize view Using masking probability Reasoning view Using masking probability .

[0133]

[0134] in, and Represent the feature matrices of the optimized view and the inference view after feature masking, respectively; and represent the original feature matrix. The Middle Transpose of row features and The masking vector is generated according to the Bernoulli distribution, and its generation probabilities are respectively and .

[0135] (5) Apply the same edge discard probability to both views. This further enhances diversity. The adjacency matrix after edge discarding is defined as follows:

[0136] in, It is based on the Bernoulli distribution, with the probability of dropping. The generated discard vector;

[0137] Finally, the two enhanced views are defined as follows:

[0138]

[0139] (6) For the optimized view and the inference view, the graph encoder is used to extract node representations from the enhanced view. and Then, it is mapped to the final feature vector through a multilayer perceptron (MLP). and The specific process is as follows:

[0140]

[0141] in, This represents the parameters of the graph convolution encoder. This represents the parameters of the MLP layer.

[0142] (7) By minimizing the entropy of the output, the model is prompted to generate more discriminative and focused feature representations, thereby improving the overall performance.

[0143] The minimum entropy learning method of the minimum entropy graph structure learning module consists of three steps:

[0144] First, normalize the features.

[0145]

[0146] Then, determine the dimension in which the maximum value of the standardized feature is located.

[0147]

[0148] Finally, a minimum entropy loss function is constructed based on cross-entropy loss. .

[0149]

[0150] Furthermore, in step S5, the obtained graph structure matrix is ​​used to train a graph convolutional network (GCN) for predicting the association between circular RNA and drugs, specifically including the following processes:

[0151] (1) The optimized graph structure matrix obtained in the previous step Based on this, a graph convolutional network is used for association prediction:

[0152]

[0153] (2) The mean squared error is used as the loss function for the prediction task, and its definition is as follows:

[0154]

[0155] Among them, set The Middle The predicted value for each sample is denoted as , and set The Middle The true value of each sample is represented as .

[0156] Thus, after construction and training using the above methods, a model for predicting the association between circular RNA and drugs is obtained. For ease of description, the model constructed in this scheme is referred to as CFGSCDSA. When predicting the association between a target drug and a target circular RNA, three multi-source similarity networks for the target drug and three multi-source similarity networks for the target circular RNA are constructed using a known association dataset of circular RNA and drugs. The multi-source similarity networks for the target drug and the drugs are respectively input into the collaborative feature learning module. After integrating the multi-source information, the collaborative feature matrix X is output. The collaborative feature matrix X is input into the minimum entropy graph structure learning module. After processing by the minimum entropy graph structure learning module, an optimized graph structure matrix S' containing discriminative topological features is obtained. Finally, the graph structure matrix S' and the collaborative feature matrix X are input into the GCN. The GCN predicts the association between the target drug and the target circular RNA based on the graph structure matrix S' and the collaborative feature matrix X.

[0157] To verify the effectiveness and prediction accuracy of the proposed method—CFGSCDSA model—5-fold and 10-fold cross-validation were performed. The method was also compared with seven state-of-the-art methods, namely MHCDA, DHANMKF, DGATCCDA, MNGACDA, MNCLCDA, GATECDA and MHGTCDA, on a dataset. Ablation experiments were also conducted.

[0158] Figure 3 This study evaluates the predictive ability of the model using 5-fold and 10-fold cross-validation. As can be seen, under 5-fold cross-validation, the CFGSCDSA model achieved an AUC of 0.9354 and an AUPR of 0.9336. Under 10-fold cross-validation, the model achieved an AUC of 0.9358 and an AUPR of 0.9341. These test results demonstrate the excellent and stable performance of the CFGSCDSA model. The AUC and AUPR values ​​are both close to 0.94, indicating a strong ability to distinguish between positive and negative samples. The near-difference between the 5-fold and 10-fold cross-validation results indicates that the model performs stably under different data partitions and has reliable generalization ability.

[0159] Figure 4 The figure compares the AUC results of our proposed method with those of seven advanced methods—MHCDA, DHANMKF, DGATCCDA, MNGACDA, MNCLCDA, GATECDA, and MHGTCDA—on various datasets. As can be seen from the figure, CFGSCDSA achieves the highest AUC value across all datasets, outperforming the other methods and demonstrating the advantages of our proposed method.

[0160] Figure 5 This is a schematic diagram showing the comparison of ablation experiment results conducted on a dataset according to the present invention.

[0161] The models for ablation experiments include:

[0162] CFGSCDSA: A circular RNA-drug association prediction model that employs a collaborative feature learning module and a minimum entropy strategy.

[0163] CFGSCDSA-noCFE means that the collaborative feature learning module has been removed, and the input consists only of directly concatenating similar features.

[0164] CFGSCDSA-noloss means that the minimum entropy strategy is not used in graph structure learning.

[0165] Depend on Figure 5 It is evident that CFGSCDSA outperforms other methods in all its metrics.

[0166] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning, characterized in that, The method includes: S1. Based on the known association datasets between circular RNA and drugs, construct multi-source similarity networks for circular RNA and drugs respectively; S2. Input the multi-source similarity network of the circular RNA and the multi-source similarity network of the drug into the shared constraint-driven collaborative feature learning module, and output the collaborative feature matrix; S3. Input the aforementioned collaborative feature matrix into the minimum entropy graph structure learning module to obtain the graph structure matrix; S4. Input the graph structure matrix and the collaborative feature matrix into the graph convolutional network for predicting the association between circular RNA and drugs; In step S2, the collaborative feature learning module includes a first autoencoder, a shared constraint module, and a second autoencoder, and step S2 includes: The multi-source similarity network is input into the first autoencoder to generate the node feature matrix corresponding to each similarity network, which is then input into the shared constraint module. The shared constraint module uses the Pearson correlation coefficient to calculate the similarity between any two nodes, filters out node pairs with significant similarity, and constructs a constraint matrix containing information on the significantly similar node pairs. The common constraint set of circular RNAs is obtained by performing an intersection operation on the constraint matrices corresponding to all similarity networks of circular RNAs. The intersection operation is performed on the constraint matrices corresponding to all drug similarity networks to obtain the shared constraint set of the drugs; The shared constraint set of circular RNA and the shared constraint set of drugs are input together with the low-dimensional node features output by the corresponding first autoencoder into the subsequent second autoencoder, and finally the enhanced similarity features of circular RNA and drugs are obtained. By splicing the circular RNA enhanced similarity features with the drug enhanced similarity features, the aforementioned synergistic feature matrix is ​​obtained; The minimum entropy graph structure learning module includes an optimization view and an inference view; The optimized view is constructed based on the initial adjacency matrix A and the cooperative feature matrix X; The inference view is based on the adjacency matrix. And the collaborative feature matrix X is constructed as follows; The adjacency matrix The initial adjacency matrix A is obtained by processing the graph learner with a nonlinear activation function and then normalizing and symmetricizing it; The initial adjacency matrix A is constructed based on the circular RNA-drug association matrix.

2. The method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning according to claim 1, characterized in that, In step S1, the multi-source similarity network of the circular RNA includes sequence similarity, information entropy similarity, and Gaussian interaction kernel similarity of the circular RNA. The multi-source similarity network of the drug includes structural similarity, information entropy similarity, and Gaussian interaction kernel similarity.

3. The method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning according to claim 2, characterized in that, Three similarity networks for circular RNA were constructed using the following methods: The host gene information of each circular RNA is obtained from a biological database, and the sequence similarity of the circular RNA is calculated based on the host gene information. The information entropy of a single circular RNA is calculated based on the known association dataset between the circular RNA and drugs. Then, the information entropy similarity of the circular RNA is calculated based on the common association set of drugs between the circular RNA and other circular RNAs. A circular RNA-drug association matrix is ​​constructed based on the known associated dataset. The Gaussian interaction kernel similarity of the corresponding circular RNAs is calculated based on the column vectors corresponding to the circular RNAs in the circular RNA-drug association matrix.

4. The method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning according to claim 3, characterized in that, Three drug similarity networks were constructed using the following methods: The SMILES structure data of each drug are searched in the bioinformatics database, and the topological fingerprint spectrum of each drug is obtained using relevant tools. The drug structure similarity is calculated based on the topological fingerprint spectrum. The information entropy of a single drug is calculated based on the known associated dataset, and then the drug information entropy similarity is calculated based on the common associated circular RNA set of the drug and other drugs. Based on the row vectors corresponding to drugs in the circular RNA-drug association matrix, the Gaussian interaction kernel similarity of the corresponding drugs is calculated.

5. The method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning according to claim 1, characterized in that, The minimum entropy graph structure learning module introduces edge dropping and feature masking strategies, and uses different feature masking probabilities for the optimized view and the inference view, while using the same edge dropping probability. After enhancing the view and the inference view by edge dropping and feature masking, the node representations of the two enhanced views are extracted using a graph encoder and mapped to the final feature vectors by a multilayer perceptron.

6. The method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning according to claim 5, characterized in that, In step S3, the minimum entropy graph structure learning module learns by minimizing the entropy value of the output, using the minimum entropy loss function. as follows: , These are the final feature vectors of the optimized view and the inference view, respectively; They are respectively , The dimension in which the maximum value is located; , These are the normalized features of the final feature vectors of the optimized view and the inference view, respectively. This represents the cross-entropy loss function.

7. The method for predicting circular RNA and drug associations based on collaborative feature learning and minimum entropy graph structure learning according to claim 1, characterized in that, Step S4 includes: The graph structure matrix obtained in step S3 and the collaborative feature matrix output in step S2 are input together into the graph convolutional network; The training method uses the difference between the true and predicted values ​​of circular RNA-drug association as the objective and the mean squared error as the loss function.

8. A circular RNA and drug association prediction system based on collaborative feature learning and minimum entropy graph structure learning, characterized in that, Used to perform the method according to any one of claims 1 to 7.

Images