 Research
 Open Access
 Published:
Double matrix completion for circRNAdisease association prediction
BMC Bioinformatics volume 22, Article number: 307 (2021)
Abstract
Background
Circular RNAs (circRNAs) are a class of singlestranded RNA molecules with a closedloop structure. A growing body of research has shown that circRNAs are closely related to the development of diseases. Because biological experiments to verify circRNAdisease associations are timeconsuming and wasteful of resources, it is necessary to propose a reliable computational method to predict the potential candidate circRNAdisease associations for biological experiments to make them more efficient.
Results
In this paper, we propose a double matrix completion method (DMCCDA) for predicting potential circRNAdisease associations. First, we constructed a similarity matrix of circRNA and disease according to circRNA sequence information and semantic disease information. We also built a Gauss interaction profile similarity matrix for circRNA and disease based on experimentally verified circRNAdisease associations. Then, the corresponding circRNA sequence similarity and semantic similarity of disease are used to update the association matrix from the perspective of circRNA and disease, respectively, by matrix multiplication. Finally, from the perspective of circRNA and disease, matrix completion is used to update the matrix block, which is formed by splicing the association matrix obtained in the previous step with the corresponding Gaussian similarity matrix. Compared with other approaches, the model of DMCCDA has a relatively good result in leaveoneout crossvalidation and fivefold crossvalidation. Additionally, the results of the case studies illustrate the effectiveness of the DMCCDA model.
Conclusion
The results show that our method works well for recommending the potential circRNAs for a disease for biological experiments.
Background
CircRNA is a circular singlestranded RNA molecule. Because circRNA does not have a free 5′ terminal cap and 3′ terminal tail, it is biologically stable and has a longer halflife than linear RNA molecules. It also has evolutionary conservatism and tissue specificity [1]. Existing studies have shown that circRNAs are enriched in exosomes [2], which means that they can be promising biomarkers in diagnosing disease. Identifying potential circRNAdisease associations can help understand the pathogenesis of disease at the molecular level and help identify biomarkers for diagnosing and treating disease [3].
However, biological experiments to confirm circRNAdisease associations are timeconsuming and wasteful of resources. Therefore, it is urgent to predict the potential circRNAdisease associations for biological experiments to make them more efficient. In recent years, many circRNAdisease interaction databases have been built whose data are manually curated from the publications, such as CircR2Disease [4], Circ2Disease [5], and circRNADisease [6] databases. The information on these verified circRNAdisease interactions provides us with an opportunity to develop computational methods to predict potential circRNAdisease associations. Until now, much effort has been made to combine available data with different methods to predict potential circRNAdisease associations. These methods can be broadly divided into two categories; the first is the networkbased method, and the second is the machine learningbased method. The networkbased method usually uses a network to obtain the final prediction result. In contrast, the machine learning method usually starts with a training set of balanced positive and negative samples. Then the features and labels of the training set are used to train the model and then use the prediction model. For example, Lei et al. [7] proposed a pathweighted method to predict the circRNAdisease association based on a heterogeneous network composed of a circRNA similarity network, a disease similarity network and a circRNAdisease association network. Fan et al. [8] introduced a KATZ method for predicting the potential circRNAdisease association based on the expression profile similarity of circRNAs, the phenotypic similarity of diseases, and known circRNAsdisease associations. Wei and Liu [9] proposed an iCircDAMF method that uses MF to predict all unknown associations. Yan et al. [10] designed a DWNNRLS model for predicting circRNAdisease associations by using the Kronecker product kernel, which is based on the regularized leastsquares method. Li et al. [11] utilized the NCPCDA method to identify potential circRNAdisease associations using multiview similarity data, including circRNA functional similarity, disease semantic similarity, and association profile similarity. Lei and Fang [12] proposed a gradient boosting decision tree algorithm to make the final prediction using multiple biological data on circRNAs and diseases. Wang et al. [13] developed a method in which a numerical descriptor was constructed according to the similarity of diseases and circRNAs, and a deep learning convolutional neural network algorithm was used to extract the deep features of circRNAdisease descriptors. Finally, an extreme learning machine was used as the final classifier. Lei and Bian [14] proposed an RWRKNN model, where the random walk algorithm with restart is used to weight the characteristics of circRNA and the disease, and KNN was used to make the final prediction. Wang et al. [15] constructed a model named GCNCDA, which extracts features by using the graph convolutional neural network and predicts the potential circRNAdisease associations by forest penalizing attributes (Forest PA) classifier. Wang et al. [16] used a deep generative adversarial network to draw features from multisource fusion information. They employed a logistic model tree classifier to infer the potential circRNAdisease association. Xiao et al. [17] exploited graph regularization and mixednorm constraint terms to improve their model prediction potential for circRNAdisease associations. Li et al. [18] proposed a method to predict potential circRNAdisease associations by inductive matrix completion based on the sequence similarity of circRNAs and semantic similarity of diseases. Zhao et al. [19] developed a method that integrates the KATZ approach and the bipartite network projection algorithm to perform the prediction. Xiao et al. [20] proposed a weighted lowrank approximation optimization algorithm that combined dualmanifold regularizations to predict the potential circRNAdisease association. Ge et al. [21] exploited localityconstrained linear coding to reconstruct similarity networks and developed a label propagation method to obtain the final score matrices. Moreover, efficient association prediction models for miRNAdisease, lncRNAdisease, drugdisease, and lncRNAmiRNA are all very helpful in the design of our models and the results analysis [22,23,24,25,26,27,28].
Because the machine learningbased method requires reliable negative samples to train the model while there are no reliable samples, we chose the networkbased method to build the model. Additionally, a model that can accurately predict is crucial for the model itself construction algorithm and essential for the selection of the dataset and its characteristics. Therefore, in this paper, we presented a double matrix completion for predicting the circRNAdisease association (DMCCDA) method, a networkbased method for circRNAdisease association prediction. First, we construct a twolayer network based on known circRNAdisease associations, so the association matrix was established. Additionally, we established a Gaussian similarity matrix for circRNA and disease, respectively, according to the association matrix. Second, we construct a sequence similarity network of circRNA and a semantic similarity network of diseases, and then the corresponding similarity matrix is established. Third, we used matrix multiplication to update the association matrix from the circRNA and disease aspects. Fourth, we use double matrix completion to update the matrix block, which is composed of the updated association matrix and the corresponding Gaussian similarity matrix from two aspects. Finally, we integrated the results as the final prediction score. After we have the final model, we use leaveoneout crossvalidation (LOOCV) and fivefold crossvalidation (FFCV) to evaluate the performance of the DMCCDA. The value of AUC was 0.9597 under LOOCV, the mean AUC was 0.9623, and the standard deviation was 0.0029 under 100 FFCV. A case study also demonstrated that DMCCDA could accurately predict potential circRNAdisease associations.
Result
Performance evaluation
In this study, to evaluate the performance of the model, we conducted global LOOCV and FFCV. The AUC value of LOOCV reached 0.9494, whereas the average AUC of 100 FFCV was 0.9623. First, all known circRNAdisease associations were treated as positive samples, and the other samples were considered candidate samples. In LOOCV, a known circRNAdisease association serves as a test sample. The remaining positive samples are used to train the model; then, we integrate the predicted scores of all the candidate samples with the predicted scores of each test sample. Finally, we drew a receiveroperating characteristic (ROC) curve and calculated the area under the ROC curve (AUC) based on the labels and the predicted scores. The corresponding ROC curves are shown in Fig. 1.
As for FFCV, we divide all known circRNAdisease samples into five parts; each part is considered as a test sample, whereas the other four parts serve as training sets, and we can integrate all known circRNAdisease sample prediction scores with all candidate samples' prediction scores and draw a ROC curve and calculate the AUC. To avoid the impact of positive sample partitioning on the performance of the evaluation model, we performed FFCV 100 times, each time selecting a different partition. Before we draw the ROC curve, we first rank the scores in descending order and set one score at a time as a threshold. If the score is greater than the threshold, the prediction is positive; if the score is less than the threshold, the prediction is negative. The true positive rate (TPR/sensitivity) and falsepositive rate (FPR/1 specificity) were calculated at different thresholds. Sensitivity means that the prediction of the sample is positive, and the actual label is true. In contrast, specificity means that the prediction of the sample is positive, and the actual label is false. The true positive and falsepositive rates were formulated as follows:
where TP indicates that the actual label of a sample is positive and the predicted result is positive, FP represents that the actual label of a sample is negative and the predicted result is positive, TN denotes that the actual label of a sample is positive and the predicted result is negative, and FN indicates that the actual label of a sample is positive, and the predicted result is negative.
Parameter analysis
In the last step of predicting for all samples, we need to integrate two score matrices. To further improve the performance of the model, we adjusted the parameter α to integrate the two score matrices. We look for the most appropriate parameter α in the interval from 0 to 1 by setting the step size to 0.1 and calculating the AUC under LOOCV. Finally, we obtained different AUCs under different values of α; when the value of α was set to 0.7, the value of AUC under LOOCV reached the highest value of 0.9597. Therefore, we chose 0.7 as the final value of α. The values of AUC under LOOCV with different values of α are shown in the scatter diagram in Fig. 2. From Fig. 2, we can see that the values show a trend and get the highest value when α is 0.7, which means that the model's performance is better when we pay more attention to the scoring matrix of the circRNA space.
The importance of model components
We conducted the following two experiments to show that matrix multiplication and double matrix completion in the model are important for predicting the potential circRNAdisease associations. On the one hand, we exploit matrix multiplication to obtain the updated association matrix from the disease perspective and circRNA perspective and integrate the matrices as the final prediction score. The AUC value of this model for LOOCV was 0.8914. On the other hand, we use double matrix completion to update the matrix block, composed of an association matrix and a Gaussian matrix. We then integrate the two new association matrices as the final prediction score, whereas the matrix multiplication operation does not update the association matrix. The corresponding AUC was 0.7811 under the LOOCV. As shown in the two experiments above, when only part of the model is used to predict the potential correlation, the results are not competitive. Therefore, we conclude that both parts of the DMCCDA model are essential for predicting potential associations.
Prediction of a new node
A new node in the association network is a disease node with no known circRNA associated with the disease or a circRNA node with no known disease associated with this circRNA. To evaluate the performance of our model for predicting new nodes, we selected four diseases: stomach cancer, breast cancer, colorectal cancer, and malignant glioma cancer. The total number of circRNAs known to be associated with these three diseases was the highest. Additionally, the total number of circRNAs known to be associated with them was 76, 43, 38, 31, as shown in Fig. 3. Firstly, we suppose all associations between all circRNAs, and stomach cancer are unknown. Thus, we set all values of the column corresponding to stomach cancer in the association matrix to 0. Then, we calculated the Gaussian similarity matrix according to the new association matrix and exploited the DMCCDA model to predict all samples, including associations between all circRNAs and stomach cancer. Finally, all circRNAs and stomach cancer prediction scores were ranked in descending order. We also calculated the number of associations among the top 50 predicted outcomes as known associations. As a result, we find that the top 50 predicted associations are known. We also used the same method to evaluate the performance of our model for predicting breast cancer, colorectal cancer, and malignant glioma. As for breast cancer, 43 of the top 50 predicted associations are known, 35 of the top 50 predicted associations are known for colorectal cancer, and 28 of the top 50 predicted associations are known. We also look at the top 30, and the results are shown in Fig. 4.
As shown in Fig. 3, the number of circRNAs related to these four diseases showed a decreasing trend. The total number of experimentally validated associations among the top 50 predicted results for the four diseases also decreased (Fig. 4). However, the total number of experimentally validated associations among the top 30 predicted results for the four diseases all reached 30, which means that the DMCCDA model has high efficiency in predicting potential circRNAdisease associations.
The effect of the model on different datasets
We have shown above that our known circRNAdisease associations were collected from three databases. We built the association matrix after integrating the data, built the corresponding association matrix for the associated data in the three databases, and used our model to make predictions for the samples involved in the three association matrices. Data in the database CircR2Disease were processed into 445 known association pairs, involving 389 circRNAs and 61 diseases, and data in the database Circ2Disease were processed into 240 known associations, involving 215 circRNA and 46 diseases. The circRNADisease database was processed into 241 known associations, involving 223 circRNAs and 34 diseases. The LOOCV was used to evaluate the performance of our model on these datasets, and the AUCs were 0.9673, 0.9447, and 0.9568, respectively. From the high value of the AUC, we can conclude that our model can obtain good results on different datasets.
Compare with other methods
To prove the effectiveness of our method, we compared it with five stateoftheart methods, that is, NCPCDA [11], PWCDA [7], iCircDAMF [9], RWRKNN [14], and GCNCDA [15]. Among them, three methods (NCPCDA, PWCDA, and iCircDAMF) are networkbased approaches, and the rest are machine learningbased methods. We compared our DMCCDA method with these five methods in terms of the AUC under FFCV. The corresponding AUC values of all these methods are listed in Table 1. We can see that the AUC value in our method reached 0.9623, whereas the AUC values of NCPCDA, PWCDA, iCircDAMF, RWRKNN, and GCNCDA were 0.9201, 0.890, 0.9178, 0.9333, and 0.9090, respectively. Therefore, we can conclude that our method is superior to other methods, and our method can be used to predict potential circRNAdisease associations.
Case study
To further verify the efficiency of our model, we selected the top ten samples from all unknown circRNAdisease pairs according to the score produced by our model. We used the published literature to verify the prediction results (Table 2). We can see that six of the top 10 prediction samples are verified in the literature. Based on the existing literature, we can verify that circRNA hsa_circ_0001649 and circRNA hsa_circ_0001141 are related to liver cancer [29, 30]. In addition, a recently published paper showed that circRNA hsa_circ_0000284 promotes gemcitabine sensitivity in bladder cancer [31] and has also been verified to be related to esophageal cancer via microRNA [32]. A published study also showed that circRNA hsa_circ_0001141 is associated with esophageal cancer [33]. Another study showed that circRNA has_circ_0001141 inhibits stomach cancer migration, invasion, and proliferation by regulating the Wnt/βcatenin pathway [34].
Conclusion
Identifying circRNAs associated with the disease can provide a better understanding of the pathogenesis of the disease at the molecular level and help identify biomarkers of the disease and the design of drugs. In this paper, we propose a novel method, DMCCDA, to predict potential circRNAdisease associations for biological experiments to promote its efficiency and reduce resource consumption. First, we integrated circRNAdisease associations from three databases collected by circRNAdisease associations manually from published papers and constructed an association matrix to obtain as many experimentally verified circRNAdisease associations as we can. Second, we calculated the circRNA sequence similarity based on the circRNA sequence information and disease semantic similarity according to the disease ontology. Additionally, we calculated the Gaussian similarity matrices according to the association matrix for circRNA and disease. Finally, we exploit the matrix multiplication operation to update the association matrix from circRNA and disease respect by circRNA sequence similarity and disease semantic similarity. Then, we used matrix completion to predict all the unknown samples. We mainly evaluated the performance of the model using the AUC of LOOCV and FFCV. The experimental results and case study demonstrate the high efficiency of the model.
Although many models have been developed to predict potential circRNAdisease associations, there are still many problems in this field. For instance, we lack reliable negative samples to develop machine learningbased methods. Additionally, some models cannot make predictions for new nodes, such as a new disease that has no known related circRNAs or a circRNA that has no known related diseases. In this study, we developed a semisupervised method to predict potential circRNAdisease associations, which means that we do not need negative samples. Additionally, our model can be used to predict the new nodes. However, our model had some limitations. First, the method used to calculate the similarity between circRNAs and diseases is insufficient. Second, matrix completion is often used to complete the missing values. However, in this study, matrix completion is used to update the new association matrices that integrate the similarity in circRNA and disease spaces, which may introduce some noise. In the future, we will consider these problems and design a better method to predict potential circRNAdisease associations. In addition, as the association prediction, the circRNAdisease association prediction has a close relationship with the microRNAdisease association prediction, the lncRNAdisease association prediction, and drug reposition. We will pay attention to these kinds of association prediction, which belong to association prediction and are essential for diagnosing and treating complex human diseases. As for these types of association prediction methods, there are many innovative methods based on deep learning to learn the feature representation and achieve a good result. In the future, we will design a model that considers the characteristics and reduces the impact of unverified negative samples on the model.
Methods
Human circRNAdisease associations
To make full use of the information available, we collected experimentally verified associations from three databases, including CircR2Disease, Circ2Disease, and circRNADisease. All–RNA the disease associations of these three databases were manually collected from published articles. The CircR2Disease database collected 739 circRNAdisease associations involving 512 circRNAs and 71 diseases from articles published before March 31, 2018. The Circ2Disease database collected 273 circRNAdisease associations involving 237 circRNAs and 54 diseases from articles published before November 1, 2017. The circRNADisease database collected 354 circRNAdisease associations involving 330 circRNAs and 48 diseases from articles published before November 2017. First, we extracted human circRNAdisease associations, integrated circRNAdisease associations from three databases and removed duplicate associations. Then, we deleted a part of these circRNAdisease associations that include the circRNA, which has no circRNA sequence information in circBase or contains the disease that has no disease ontology identity (DOID) information in the Disease Ontology (DO) database [35]. Finally, we obtained 609 circRNAdisease associations involving 512 circRNAs and 71 diseases. Additionally, we constructed an adjacency matrix A that has nc rows and nd columns, whereas nc represents the total number of circRNAs, and nd denotes the total number of diseases that are involved in the known circRNAdisease associations. If a certain circRNA is experimentally verified to be related to a certain disease, the element in the corresponding position in matrix A is 1; otherwise, it is 0.
circRNA sequence similarity
To calculate circRNA sequence similarity, we first downloaded 140,790 circRNA sequence information from the database circBase [36] and then extracted the circRNA sequence information from the known associations involved. Next, we utilized the Levenshtein distance [18] to measure the similarity between any two circRNAs. The Levenshtein distance represents the minimum number of operands required to convert string A sequence to string B, which means that the shorter the distance, the greater the similarity between the two circRNAs. Formula (3) was used to calculate the sequence similarity between the two circRNAs.
where dis describes the number of operands needed to convert circRNA circ_{i} sequence to circRNA circ_{j} sequence, and len represents the sequence length of some circRNAs.
Disease semantic similarity
As for disease semantic similarity, we first collected disease DOID information from the database Disease Ontology. Then we used the DOSim [37] function to calculate disease semantic similarity based on Wang's method [38]. Because there is an R package DOSE, we can easily obtain the disease semantic similarity by inputting the disease DOID. Wang's method was based on the following formula:
where Tdi represents disease di and all ancestor node of disease di in the directed acyclic graph of disease, and SDI (t) indicates the contribution from all nodes in the set Tdi to disease di. The details are shown in the following formula:
Gauss interaction profile kernel similarity
The Gaussian interaction profile kernel similarity is another algorithm constructed to measure disease similarity and circRNA similarity based on the known association matrix. As for the association matrix, the ith row IP (i) represents the associations between the ith circRNA and all diseases. The jth column IP (j) denotes the associations between the jth disease and all circRNAs. Based on the assumption that similar circRNAs are more likely to be associated with similar diseases and vice versa, we calculated the Gaussian interaction profile kernel similarity for circRNA and disease as follows:
where β_{d} and β_{c} are the kernel bandwidths, which can be calculated as follows:
where β_{d}′ and β_{c}′ are the original bandwidths, and according to previous research, we assign the value of the initial bandwidth to 1. Finally, we can obtain the Gaussian interaction profile kernel similarity matrix KD, KC for disease, and circRNA, respectively.
Model construction
To make full use of the known information and find the potential associations, we use the following steps to predict potential circRNAdisease associations in this paper. First, we collected the known circRNAdisease associations, sequence information about circRNAs, and semantic information about diseases. Second, we construct association matrix A, sequence similarity matrix CC, and semantic similarity matrix DD based on the information gathered in the step above.
Additionally, we calculated the Gaussian similarity matrix KC and KD for circRNA and disease according to the association matrix. The preparation materials are shown in Fig. 5. Third, we exploit the similarity of circRNA and diseases based on the sequence and semantic information of circRNAs and diseases to update the association matrix and obtain two updated association matrices Ac and Ad. Fourth, we use double matrix completion to update the matrix block, which is composed of the updated association matrix and the corresponding Gaussian similarity matrix from two aspects. Finally, we extract the association matrix parts Ac* and Ad* from the two matrix blocks after using the matrix completion algorithm and integrated the two matrices as the predicted score of each sample. The corresponding flowchart is shown in Fig. 6, in which we introduce the association matrix (A), sequence similarity matrix (CC), and Gaussian similarity matrix (KC) of circRNA, and introduced the semantic similarity matrix (DD) and Gaussian similarity matrix (KD) of the disease. Then, we use matrix multiplication to update the association matrix; thus, even if we obtain a new disease with no known circRNA association, we can predict this disease based on the semantic similarity between this disease and other diseases. The formulae are as follows:
To control the values of Ac and Ad within a specific range, we carried out the following processing for Ac and Ad [11].
Then, we can obtain two updated association matrices Ac and Ad, which integrate the similarity of circRNAs and diseases, respectively. Next, we splice the updated matrix Ac and the Gaussian matrix KC to form the first matrix block M1. Similarly, we splice Ad and the Gaussian matrix KD to form the second matrix block M2. To make the best use of the circRNAdisease association matrix information and similarity information, we used two matrix complements to update the two matrix blocks obtained in the previous step in circRNA and disease space, respectively, which was also inspired by Yang et al. [39]. First, in the circRNA space, we integrate the bounded nuclear norm regularization to the nuclear norm minimization problem [40], as follows:
where M1* represents the nuclear norm of M1, P is the projection operation, and Ω is the universal set. Additionally, α is a harmonic parameter, and the initial value of M is M1. W is a new matrix that the following formula can represent:
Then, we can get the model's augmented Lagrangian function as follows,
where Y is the Lagrange multiplier, and β is the penalty coefficient. Then, the closed result is obtained as follows:
We can also get the value of M1 and Y by iterating,
Therefore, by iterating to convergence, we obtain the final recovery matrix W. We similarly conduct matrix M2 in the disease space, and we can obtain the updated matrix Ad*. Finally, we obtain two scoring matrices and integrate them using the following equation:
where α is the integration parameter. After the relevant experiments, we set the parameter value to 0.7. Then, we can obtain all the samples' scoring matrix Ascore.
Availability of data and materials
All data and code underlying this study are available in an online archive https://github.com/zzl1996zzl/DMCCDA.
Abbreviations
 circRNA:

Circular RNA
 DMCCDA:

Double matrix completion for predicting the circRNAdisease association
 LOOCV:

Leaveoneout cross validation
 FFCV:

Fivefold cross validation
 DOID:

Disease ontology identities
 DO:

Disease ontology
 ROC:

Receiveroperating characteristics
 AUC:

Area under the ROC curve
 TPR:

True positive rate
 FPR:

False positive rate
 TP:

True positives
 FP:

False positives
 TN:

True negatives
 FN:

False negatives
References
 1.
Guo Y, Yang J, Huang Q, Hsueh C, Zheng J, Wu C, Chen H, Zhou L. Circular RNAs and their roles in head and neck cancers. Mol Cancer. 2019;18(1):44.
 2.
Li Y, Zheng Q, Bao C, Li S, Guo W, Zhao J, Chen D, Gu J, He X, Huang S. Circular RNA is enriched and stable in exosomes: a promising biomarker for cancer diagnosis. Cell Res. 2015;25(8):981–4.
 3.
Guo W, Zhu L, Deng S, Zhao X, Huang D. Understanding tissuespecificity with human tissuespecific regulatory networks. Sci China Inf Sci. 2016;59(7):070105.
 4.
Fan C, Lei X, Fang Z, Jiang Q, Wu FX. CircR2Disease: a manually curated database for experimentally supported circular RNAs associated with various diseases. Database. 2018. https://doi.org/10.1093/database/bay044.
 5.
Yao D, Zhang L, Zheng M, Sun X, Lu Y, Liu P. Circ2Disease: a manually curated database of experimentally validated circRNAs in human disease. Sci Rep. 2018;8(1):11018.
 6.
Zhao Z, Wang K, Wu F, Wang W, Zhang K, Hu H, Liu Y, Jiang T. circRNA disease: a manually curated database of experimentally supported circRNAdisease associations. Cell Death Dis. 2018;9(5):475.
 7.
Lei X, Fang Z, Chen L, Wu FX. PWCDA: Path weighted method for predicting circRNAdisease associations. Int J Mol Sci. 2018;19(11):3410.
 8.
Fan C, Lei X, Wu FX. Prediction of CircRNAdisease associations using KATZ model based on heterogeneous networks. Int J Biol Sci. 2018;14(14):1950–9.
 9.
Wei H, Liu B. iCircDAMF: identification of circRNAdisease associations based on matrix factorization. Brief Bioinform. 2019;21(4):1356–67.
 10.
Yan C, Wang J, Wu FX. DWNNRLS: regularized least squares method for predicting circRNAdisease associations. BMC Bioinform. 2018;19(Suppl 19):520.
 11.
Li G, Yue Y, Liang C, Xiao Q, Ding P, Luo J. NCPCDA: network consistency projection for circRNA–disease association prediction. RSC Adv. 2019;9:33222–8.
 12.
Lei X, Fang Z. GBDTCDA: predicting circRNAdisease associations based on gradient boosting decision tree with multiple biological data fusion. Int J Biol Sci. 2019;15(13):2911–24.
 13.
Wang L, You ZH, Huang YA, Huang DS, Chan KCC. An efficient approach based on multisources information to predict circRNA–disease associations using deep convolutional neural network. Bioinformatics. 2019;36(13):4038–46.
 14.
Lei X, Bian C. Integrating random walk with restart and kNearest Neighbor to identify novel circRNAdisease association. Sci Rep. 2020;10(1):1943.
 15.
Wang L, You ZH, Li YM, Zheng K, Huang YA. GCNCDA: A new method for predicting circRNAdisease associations based on Graph Convolutional Network Algorithm. PLoS Comput Biol. 2020;16(5):e1007568.
 16.
Wang L, You Z, Li L, Zheng K, Wang Y. Predicting circRNAdisease associations using deep generative adversarial network based on multisource fusion information. In: 2019 IEEE international conference on bioinformatics and biomedicine (BIBM): 18–21 Nov. 2019 2019. p. 145–152.
 17.
Xiao Q, Yu H, Zhong J, Liang C, Li G, Ding P, Luo J. An insilico method with graphbased multilabel learning for largescale prediction of circRNAdisease associations. Genomics. 2020;112(5):3407–15.
 18.
Li M, Liu M, Bin Y, Xia J. Prediction of circRNAdisease associations based on inductive matrix completion. BMC Med Genomics. 2020;13(Suppl 5):42.
 19.
Zhao Q, Yang Y, Ren G, Ge E, Fan C. Integrating bipartite network projection and KATZ measure to identify novel CircRNAdisease associations. IEEE Trans Nanobiosci. 2019;18(4):578–84.
 20.
Xiao Q, Luo J, Dai J. Computational prediction of human disease associated circRNAs based on manifold regularization learning framework. IEEE J Biomed Health Inform. 2019;23(6):2661–9.
 21.
Ge E, Yang Y, Gang M, Fan C, Zhao Q. Predicting human diseaseassociated circRNAs based on localityconstrained linear coding. Genomics. 2020;112(2):1335–42.
 22.
Zhao Y, Chen X, Yin J. Adaptive boostingbased computational model for predicting potential miRNAdisease associations. Bioinformatics. 2019;35(22):4730–8.
 23.
Li J, Zhang S, Liu T, Ning C, Zhang Z, Zhou W. Neural inductive matrix completion with graph convolutional networks for miRNAdisease association prediction. Bioinformatics. 2020;36(8):2538–46.
 24.
Chen X, Xie D, Wang L, Zhao Q, You ZH, Liu H. BNPMDA: bipartite network projection for MiRNAdisease association prediction. Bioinformatics. 2018;34(18):3178–86.
 25.
Ding L, Wang M, Sun D, Li A. TPGLDA: Novel prediction of associations between lncRNAs and diseases via lncRNAdiseasegene tripartite graph. Sci Rep. 2018;8(1):1065.
 26.
Xuan P, Cao Y, Zhang T, Wang X, Pan S, Shen T. Drug repositioning through integration of prior knowledge and projections of drugs and diseases. Bioinformatics. 2019;35(20):4108–19.
 27.
Zhang L, Liu T, Chen H, Zhao Q, Liu H. Predicting lncRNAmiRNA interactions based on interactome network and graphlet interaction. Genomics. 2021;113(3):874–80.
 28.
Liu H, Ren G, Chen H, Liu Q, Yang Y, Zhao Q. Predicting lncRNA–miRNA interactions based on logistic matrix factorization with neighborhood regularized. Knowl Based Syst. 2020;191:105261.
 29.
Yao T, Chen Q, Fu L, Guo J. Circular RNAs: biogenesis, properties, roles, and their relationships with liver diseases. Hepatol Res. 2017;47(6):497–504.
 30.
Qiu LP, Wu YH, Yu XF, Tang Q, Chen L, Chen KP. The emerging role of circular RNAs in hepatocellular carcinoma. J Cancer. 2018;9(9):1548–59.
 31.
Xie F, Zhao N, Zhang H, Xie D. Circular RNA CircHIPK3 promotes gemcitabine sensitivity in bladder cancer. J Cancer. 2020;11(7):1907–12.
 32.
Ba Y, Liu Y, Li C, Zhu Y, Xing W. HIPK3 promotes growth and metastasis of esophageal squamous cell carcinoma via regulation of miR599/cMYC axis. Onco Targets Ther. 2020;13:1967–78.
 33.
Han C, Seebacher NA, Hornicek FJ, Kan Q, Duan Z. Regulation of microRNAs function by circular RNAs in human cancer. Oncotarget. 2017;8(38):64622–37.
 34.
Peng Y, Wang HH. CirITCH inhibits gastric cancer migration, invasion and proliferation by regulating the Wnt/βcatenin pathway. Sci Rep. 2020;10(1):17443.
 35.
Schriml LM, Mitraka E, Munro J, Tauber B, Schor M, Nickle L, Felix V, Jeng L, Bearer C, Lichenstein R, et al. Human disease ontology 2018 update: classification, content and workflow expansion. Nucleic Acids Res. 2019;47(D1):D955–62.
 36.
Glažar P, Papavasileiou P, Rajewsky N. circBase: a database for circular RNAs. RNA. 2014;20(11):1666–70.
 37.
Li J, Gong B, Chen X, Liu T, Wu C, Zhang F, Li C, Li X, Rao S, Li X. DOSim: An R package for similarity between diseases based on disease ontology. BMC Bioinform. 2011;12(1):266.
 38.
Wang JZ, Du Z, Payattakool R, Yu PS, Chen CF. A new method to measure the semantic similarity of GO terms. Bioinformatics. 2007;23(10):1274–81.
 39.
Yang M, Luo H, Li Y, Wu FX, Wang J. Overlap matrix completion for predicting drugassociated indications. PLoS Comput Biol. 2019;15(12):e1007541.
 40.
Yang M, Luo H, Li Y, Wang J. Drug repositioning based on bounded nuclear norm regularization. Bioinformatics. 2019;35(14):i455–63.
Acknowledgements
We thank all the teachers and students in the lab for their valuable advice on this work.
Funding
This work was supported by grants from the National Natural Science Foundation of China (Nos. U19A2064 61873001), and the Open Foundation of Engineering Research Center of Big Data Application in Private Health Medicine, Fujian Province University (KF2020006).
Author information
Affiliations
Contributions
ZZ was mainly responsible for the prediction model and drafting the manuscript, CR was mainly responsible for the result analysis, WP was responsible for the result analysis and drafting the manuscript, JF was mainly responsible for the data collection and ZC was mainly responsible for the final verification and the model design. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Zuo, ZL., Cao, RF., Wei, PJ. et al. Double matrix completion for circRNAdisease association prediction. BMC Bioinformatics 22, 307 (2021). https://doi.org/10.1186/s12859021042313
Received:
Accepted:
Published:
Keywords
 circRNAdisease associations
 Similarity matrix
 Matrix completion