 Research
 Open access
 Published:
Predicting miRNAdisease associations using a hybrid feature representation in the heterogeneous network
BMC Medical Genomics volume 13, Article number: 153 (2020)
Abstract
Background
Studies have found that miRNAs play an important role in many biological activities involved in human diseases. Revealing the associations between miRNA and disease by biological experiments is timeconsuming and expensive. The computational approaches provide a new alternative. However, because of the limited knowledge of the associations between miRNAs and diseases, it is difficult to support the prediction model effectively.
Methods
In this work, we propose a model to predict miRNAdisease associations, MDAPCOM, in which protein information associated with miRNAs and diseases is introduced to build a global miRNAproteindisease network. Subsequently, diffusion features and HeteSim features, extracted from the global network, are combined to train the prediction model by eXtreme Gradient Boosting (XGBoost).
Results
The MDAPCOM model achieves AUC of 0.991 based on 10fold crossvalidation, which is significantly better than that of other two stateoftheart methods RWRMDA and PRINCE. Furthermore, the model performs well on three unbalanced data sets.
Conclusions
The results suggest that the information behind proteins associated with miRNAs and diseases is crucial to the prediction of the associations between miRNAs and diseases, and the hybrid feature representation in the heterogeneous network is very effective for improving predictive performance.
Background
MicroRNAs(miRNAs) are a kind of small singlestranded endogenous noncoding RNAs with a length about 22 nucleotides, which play an important role in regulating the gene expression during the posttranscriptional level [1, 2]. Many studies have shown that the dysregulation of miRNAs is involved in multiple human diseases like cancers [3], cardiovascular diseases [4] and Alzheimer’s diseases [5] etc., and the prediction of miRNAsdiseases associations is crucial to understand the diseases pathogenesis [6]. Furthermore, George Adrian, et al. found that the miR15 and miR16 are deleted in a lot B cell chronic lymphocytic leukemias (BCLL) [7], T. Sredni et al. demonstrated that miR129 and miR25 express abnormally in all pediatric brain tumor types [8]. Besides, Jun Lu et al. successfully classified poorly differentiated tumours using miRNA expression profiles [9], which demonstrated the potential of miRNAs as biomarkers. Therefore, Predicting miRNAdisease associations is very meaningful. However, a lot of miRNAdisease associations remain unknown and experimental approaches for predicting the associations are timeconsuming and expensive. Therefore, a lot computational methods have been developed to predict the miRNAdisease associations.
Computational methods can be grouped into two categories: networkbased methods and machine learningbased methods. Networkbased methods usually use similarity measurement to predict the associations. For example, Jiang et al. [10] presented a computational method to predict the associations between miRNAs and diseases by prioritizing entire human microRNAome according to the disease of interest. The higher the rank is, the more possibly the miRNA can associate with the disease. In 2010, the model was improved by introducing genomic data [11]. However, the performance of the model was still not satisfactory because the known target genes of miRNAs are too rare to support the methods effectively. Chen et al. develpoed a method called RWRMDA [12], the author ran random walk with restart algorithm on a miRNA functional similarity network to obtain a score for every miRNA, and the miRNA with a higher score is more likely to associate with a certain disease. Shi et al. [13] extended random walk with restart algorithm (RWR), they used proteins associated with diseases and miRNAs as seed nodes to calculate the ES score by RWR respectively, and then used the Pvalue to predict whether the disease and miRNA are related. PRINCE [14] is another algorithm optimized based on RWR, it proposed a novel method to initial probability of miRNAs. However, these methods, based on RWR, are dependent on known associations between miRNAs and a given disease, so it couldn’t be applied to predict the relationships between miRNAs and a new disease, without any associations with miNRAs. Furthermore, defining a proper similarity calculation model is challengeable in this category.
The prediction models in another category are based on machine learning. For example, Xu et al. [15] extracted features from a miRNAdisease network, and then used the features to train a prediction model by support vector machine (SVM), the method can discover positive samples from massive negative samples. Chen et al. [16] presented a semisupervised and global method RLSMDA, the method calculated possibilities of being associated with a given disease for each miRNA by a continuous classification function, and it could predict the associations of diseases and miRNAs without any known association between them. However, the method didn’t integrate the information related to miRNAs and diseases completely since the continuous classification function is established for the miRNA network and the disease network separately.
Recently, more computational methods are proposed. Zheng et al. [17] developed a machine leaningbased method MLMDA, which used a variety of information including miRNA sequence information, miRNA functional similarity, disease semantic similarity and Gaussian interaction profile kernel similarity information to train their model by applying random forest classifier. The classifier achieved promising performance, but it might take a lot of effort to prepare the required data. What’s more, the knowledge of deep learning was also applied in this field. Peng et al. [18] utilized a convolutional neural network to predict miRNAdisease association, input data was reduced miRNAdisease interaction features which were captured from a threelayer network. The similarity metric is essential in order to predict associations between miRNAs and diseases, where Yang et al. [19] used a novel method miRGOFS to measure functional similarities of miRNAs, and the method considered both common ancestors and descendants of GO terms when it was used to calculate similarities among GO sets in an asymmetric manner, so it can help predict the miRNAdisease associations. Chen et al. [20] presented the first decision tree, learningbased model, whose informative feature vectors were extracted from miRNA functional similarities, the disease semantic similarities, and known miRNAdisease associations. Yin et al. [21] put forward LWPCMF, they used weighted profile (WP), collaborative matrix factorization (CMF) and logistic function to optimize their model.
In this work, we present a computational method named MDAPCOM to predict the associations between miRNAs and diseases by combined features. First, we construct a miRNAproteindisease global network by merging six subnetworks, which are miRNAmiRNA Similarity Network, ProteinProtein Interaction Network, DiseaseDisease Similarity Network, miRNATarget Interaction Network, miRNADisease Relationship Network and ProteinDisease Association Network respectively. Subsequently, we extract diffusion features for each node and a 39dimensional HeteSim feature for each miRNAdisease pair in the global network. The diffusion features are extracted by random walk with restart algorithm and then reduced in dimension using the singular value decomposition algorithm (SVD). Finally, we integrate these two features to train the miRNAdisease association prediction model using eXtreme Gradient Boosting (XGBoost) algorithm. We apply the MDAPCOM method under 10fold crossvalidation and achieve an AUC of 0.991. MDAPCOM also performs better when compared with other two previous methods RWRMDA [12] and PRINCE [14], which also used network features for prediction. Furthermore, our method performs well on three unbalanced data sets with positive and negative samples ratios 1:2, 1:5 and 1:10, respectively.
Results
Data sources
We collect six different types of data from the Internet, which are the miRNAmiRNA similarity data, miRNAProtein interactions, miRNADisease relationships, PPI data (ProteinProtein interactions), ProteinDisease association data, DiseaseDisease similarity data, respectively, containing 2588 miRNAs, 18143 proteins and 5080 different kinds of diseases totally.
miRNAmiRNA similarity network
We obtain miRNA expression data from miRmine database [22]. In this database, the researchers analyzed overall expression profile of human, obtained from different miRNAseq databases. It contains 2822 different precursor miRNAs where more than two of them consist one mature miRNA, so we can derive the expression values of every mature miRNA from the average values of its precursors’. In this way, we obtain 2588 miRNA expression profiles. Moreover, the Pearson Correlation Coefficient (PCC) scores are calculated to preform similarities of the expression profiles between two miRNAs [23]. The higher the PPC score is, the more likely these two miRNAs are similar. The miRNAmiRNA Similarity network is also built. In the network, every miRNA is a node and the PPC scores present the edges, and the negative edges are cut down.
Proteinprotein interaction network
We derive data from the STRING database V10.0 [24]. The database offers data which is obtained from the results of biochemical experiments, biophysical or genetic techniques. We get 7,866,428 PPI entries from 18,143 proteins in the database and use them to build our ProteinProtein Interaction Network. In PPI network, each of the entry comprises a protein node A, a protein node B, and the predicted relationship’s score between them. The highest score means the two proteins can interact with each other with the biggest possibility and vice versa. Last, we utilize the predicted score to present the value of each edge between two protein nodes to construct our ProteinProtein Interaction Network.
Diseasedisease similarity network
To build the DiseaseDisease Similarity Network, we obtain data from the MimMine database. [25] It is mapped from OMIM database, containing more than 5000 human genetic disease phenotypes. It is worthy to point out that we normalize diseasedisease similarities’ values into [0,1] in MimMiner database. Subsequently, we receive 5080 kinds of diseases and get the similarities between them. Finally, we construct the DiseaseDisease Similarity Network where each node presents a kind of disease, and the weight is similarity between them.
miRNAtarget interactions network
We download miRNAtarget interactions from the miRTarBase database of release 7.0 [26], miRNATarget Interaction Network can be built. It should be point out all data is validated in this database. Moreover, we map the genes onto protein entries, and remove invalid entries (miRNA or protein), which are repeated and outofrange. Finally, we extract miRNAtarget interactions between 2,588 miRNAs and 18,143 proteins. Then, miRNATarget Interaction network is constructed based on these data.
miRNAdisease relationship network
We get miRNAdisease data from HMDD v3.0 database [27], which is a reliable online database containing 1102 gene on miRNA, 850 different types of diseases and 32281 associations between miRNA and disease, and they are all based on literature. Furthermore, we receive the relationships between 2588 miRNAs and 5080 diseases which are mentioned above. Lastly, we build the miRNADisease Relationship network using these validated data.
Proteindisease association network
We obtain data from DisGeNET database [25] which collects data on genotypephenotype relationships. In this work, we map genes into proteins and unify the name of diseases, so 18,143 proteins,5080 diseases and the associations between them are extracted. Then, we construct a ProteinDisease Association Network from these data.
Global heterogeneous network
We integrate the aforementioned networks to build the global heterogeneous network:
where T represents our global heterogeneous network, M, P, D present similarity of miRNAmiRNA, proteinprotein and diseasedisease respectively, B presents the miRNATarget Interaction Network, C indicates miRNADisease Relationship Network, and W shows the ProteinDisease Association Network. Obviously, the B^{T},C^{T} and W^{T} are transposed matrices of B, C and W, and the edges with value less than 0.5 are removed from the network.
There are 2588 miRNAs and 5080 diseases in our miRNAproteindisease global network, so we can get a total of 13147040 (2588 ×5080) miRNAdisease pairs. We extract a 639dimensional combined feature vector for each miNRAdisease pair in the global network, in which 11824 feature vectors are positive samples while the other 13135216 feature vectors are negative samples. We randomly select 11824 feature vectors from 13135216 negative samples to construct a standard dataset together with 11824 positive samples, subsequently, we execute 10fold crossvalidation on the standard dataset. The positive and negative samples are randomly divided into 10 subsamples equalled in size(the size of the tenth subsample is 1186 because 11824 is’t divisible by 10), one of which is retained as the validation set and the other 9 subsamples are regarded as the training set. Then the procedure iterates 10 times with each one in the 10 subsamples as the validation set, before each iteration, the associations occurred in the validation set are removed from the original global network, and then all feature vectors are reextracted from the new global network. Furthermore, another three unbalanced data sets are obtained in the same way except the size of the selected negative samples, and the size of negative samples in three unbalanced data sets is 23648, 59120 and 118240, respectively.
Performance measures
We apply 10fold crossvalidation, and obtain the average performance of our model through the performance evaluation. In terms of performance evaluation, we select precision(PRE), recall(REC), Fscore(FSC), accuracy(ACC) and the area under the receiver operating characteristic curve(AUC):
TP and FP are the amount of correctly predicted positive and negative samples, FP and FN are the numbers of positive and negative samples predicted by mistakes. Simultaneously, we calculate the area under ROC curve (AUC) to measure the overall performance.
Excellent combined feature
In our method, we extract two different features from a global heterogeneous network, a global matrix of nine different data, and combine them to construct our training dataset. Firstly, with the help of random walk with restart algorithm, we extract diffusion feature of each node from our global network, so we can get a 20588*20588 feature matrix, where a row represents a feature vector of one node. For example, the first row shows the miRNA1’s feature vector, the 2589 th row is the protein1’s feature vector, and the 20732 th row is the disease1’s feature vector. In the next step, we apply SVD algorithm on this feature matrix to reduce the dimension of it from 20588 to 300, here our feature matrix is 20588*300. After obtaining reduced feature vectors of each node, we combine each miRNA’s feature vector with each disease’s, so we get a (2588*5080) * 600 miRNAdisease feature matrix, where a row shows the feature vector of a pair of miRNAdisease. Secondly, we calculate HeteSim scores of each miRNAdisease pair, and get a (2588*5080) * 39 HeteSim matrix. Finally, in order to construct our training data, we joint the SVD feature and HeteSim score, so we get a (2588*5080) * 639 feature vector, where a row is the combined feature vector of a miRNAdisease pair. To show excellent performance of our method, we use diffusion features, the HeteSim feature and the combined feature to train the prediction model using 10fold crossvalidation under the standard data set, respectively, and the result shows in Fig. 1. The AUC value of training model with the diffusion feature and the HeteSim feature reach 0.970 and 0.986, respectively, and we get an AUC of 0.991 using combined feature.
Superiority of XGBoost algorithm
In this work, we apply eXtreme Gradient Boosting(XGBoost) [28] algorithm to train our model. We compare XGBoost algorithm with other machine learning algorithm to present that the eXtreme Gradient Boosting(XGBoost) algorithm is the most suitable method for us. To achieve the goal, we obtain other classifiers from python toolkits scikitlearn and apply 10fold crossvalidation. We compare XGBoost algorithm with random forest (RF) [29], support vector machine (SVM) [30]and gradient tree boosting (GTB) [31] algorithm. In random forest algorithm, we set the minimize samples split to 42, maximize depth of tree to 11 and the resting parameter values to default. In the support vector machine algorithm, we use RBF kernel setting the C value to 100, gamma value to 0.0001. In gradient tree boosting algorithm, we set the minimize samples split to 110, the maximize depth of tree to 9. The results perform in Fig. 2.
Performance comparison with existing methods
We implement RWRMDA [12] and PRINCE [14] under a standard dataset and three unbalanced datasets, applying 10fold crossvalidation to calculate their AUC values and compare theirs with MDAPCOM’s. For PRINCE, we set α=0.95, d=log (9999), c=15 and then apply the random walk with restart 10 times. The probability of restarting in RWRMDA is set to 0.5. To visually describe and compare the performance of the three methods, we plot the Receiver Operating Characteristic (ROC) curve with its horizontal axis representing false positive rate (FPR) and the vertical axis representing true positive rate (TPR). Subsequently, we use the area under the ROC curve (AUC) to accurately compare the performance of the three methods. Figures 3, 4, 5 and 6 show the performance of the three methods under four datasets with different positive and negative ratios, respectively. Among three methods, MDAPCOM significantly outperforms the other two methods, achieving an amazing AUC score 0.99. Furthermore, the AUC scores of our method are all above 0.99 under four data sets, which proves its stability.
Conclusions
In this work, we present a prediction method based on machine learning to predict the associations between miRNAs and diseases, MDAPCOM. We build a miRNAproteindisease global network, then extract dimensional reduced RWR diffusion feature and HeteSim feature from the network, the diffusion feature reflects the node topological information in the heterogeneous network and the HeteSim feature extracts the correlation of node pairs. Subsequently, the two features are combined to train the miRNAdisease association prediction model using 10fold crossvalidation by eXtreme Gradient Boosting (XGBoost). The MDAPCOM shows better performance than other two previous methods, based on network feature. The excellent performance suggests that the information behind proteins which are associated with miNRAs and diseases is crucial to predict associations between miRNAs and diseases. Furthermore, the two features extract network information from different perspectives and the combination of them integrates network information effectively, which also contributes to the excellent performance of the method.
Methods
Overview of MDAPCOM
Our method is displayed in Fig. 7, which is built through following steps: (A) Collect six types of data sources and remove invalid and repeated data. (B) Merge the six networks to build a global miRNAproteindisease heterogeneous network. (C) Run random walks with restart (RWR) algorithm in the global network to calculate a diffusion feature for every node, which reflects the relevance of one node with all other nodes (miRNAs, proteins and diseases) in the network (D) Run the singular value decomposition (SVD) algorithm to reduce dimension of the diffusion feature, obtaining a 300dimensional feature vector for every node. (E) Use HeteSim measure to estimate the correlation between two nodes and get a 39dimensional HeteSim feature for each miRNAdisease pair. (F) Integrate the 600dimension diffusion feature(300dimensional for miRNA and 300dimensional for disease) and 39dimensional HeteSim feature to train a miRNAdisease association prediction model by eXtreme Gradient Boosting (XGBoost).
Diffusion feature of reduced dimension
To predict the miRNAdisease associations, we transform the problem to obtain possibility that a miRNA can associate with a disease. The Random Walk with Restart algorithm can capture the relationships between two nodes and the global topological information of nodes in the network [32–34]. In this study, we run RWR algorithm on the global heterogeneous network and get a highdimensional(25,811) vector for each node. The vector reveals the topological properties of the node in the network, which includes a set of possibilities that a node can access to other nodes. We use D to represent the adjacency matrix of our global heterogeneous network, and T, a normalized matrix, represents the transition probability from the node i to the node j, T is defined as
If a node i is connected with a node j, the value of D_{ij} is 1, otherwise the value is 0. The RWR can be regarded as an iterative process, which is expressed as
Where α is the restart rate of random walker which is in the range of [0,1], P_{0} is the initial probability of the heterogeneous network, P_{t} is the state of the heterogeneous network when the process is in the tth.
Here, we get a 25,811dimensional feature for every node which reveals the topological relevance of a node to other nodes(2,588 miRNAs, 18,143 proteins and 5,080 diseases) in the network. Using such tremendous features directly to train the model is pretty timeconsuming and unnecessary, since they contain some noise. Therefore, we reduce the 25,811dimensional diffusion feature to 300dimension by singular value decomposition (SVD) algorithm [35, 36].
HeteSim measure
The HeteSim measure performs well in measuring the correlation of nodes in the heterogeneous biological network [37]. It’s a selfmaximum and symmetric measure, using an uniform framework to measure the correlation of nodes based on specified path [38]. In this paper, we use HeteSim scores of miNRAdisease pairs to extract network information.
Definition 1
(Transition probability matrix [38]) A and B are two types of nodes in the heterogeneous network. (M_{AB})_{m∗n} is an adjacency matrix indicating the relation between A and B, if there is an association between a node i belonging to A and a node j belonging to B, M_{AB}(i,j)=1, otherwise M_{AB}(i,j)=0. The transition probability matrix T_{AB} is defined as follows
Definition 2
(Reachable probability matrix [38]) R_{ρ} represents the reachable probability matrix based on the path \( \rho = P_{1}P_{2}P_{3} \dots P_{n + 1} \), where P_{i} represents any types of nodes of the heterogeneous network. R_{ρ} can be calculated as
Based on the above 2 definitions, we can calculate the HeteSim score in 3 steps [38].

1
Separate the path ρ from the middle into ρ_{L} and ρ_{R}. When the path length is even, ρ_{L} and ρ_{R} are equal in length, and \( R_{\rho _{L}} \) and \( R_{\rho _{R}} \) can be directly calculated. When the path length is odd, there are two intermediate nodes, take each one of them as intermediate node respectively to obtain \( \rho _{L_{1}} \), \( \rho _{L_{2}} \), \( \rho _{R_{1}} \) and \( \rho _{R_{2}} \), then \( R_{\rho _{L}} \), \( R_{\rho _{R}} \) can be calculated as
$$R_{\rho_{L}} = \frac{R_{\rho_{L_{1}}} + R_{\rho_{L_{2}}}}{2} $$$$R_{\rho_{R}} = \frac{R_{\rho_{R_{1}}} + R_{\rho_{R_{2}}}}{2} $$ 
2
Calculate the \( R_{\rho _{_{L}}} \) and \( R_{\rho _{R}^{1}} \), where \( \rho _{R}^{1} \) represents the reverse of ρ_{R}, for example, if ρ_{R}=ABC, then \( \rho _{R}^{1} = CBA \).

3
Achieve the HeteSim measure as
$$ HeteSim\left(a,b\rho\right) = \frac{ R_{\rho_{L}}(a, :)\left(R_{\rho_{R}^{1}}(b,:)\right)^{T} } { {\lVert R_{\rho_{L}}(a, :) \rVert}_{2} \times {\lVert R_{\rho_{R}^{1}}(b,:) \rVert}_{2} } $$(5)
Using the above method, we can derive 39 HeteSim scores for each miRNAdisease pair(i.e. a 39dimensional HeteSim feature vector for each miRNAdisease pair) based on all paths less than 5 in length starting at miRNA and ending at disease. The detailed paths are listed in Table 1.
The eXtreme gradient boosting (XGBoost) algorithm
The eXtreme Gradient Boosting is an endtoend system extended by tree boosting, and it’s used widely in machine learning [28]. The algorithm can be obtained from python toolkits scikitlearn. In this study, a 600dimensional diffusion feature(300dimensional for miRNA and 300dimensional for disease) and a 39dimensional HeteSim feature are extracted for each miNRAdisease pair in the global network. Subsequently, the two features are combined, forming a 639dimensional feature, to train the prediction model by XGBoost, where the optimal learning rate is 0.15, the number of iterations is 650, the max depth of tree is 4 and default values set for the other parameters.
Abbreviations
 XGBoost:

eXtreme Gradient Boosting
 miRNA:

MicroRNA
 RWR:

random walk with restart algorithm
 SVM:

support vector machine
 GO:

gene ontology
 WP:

weighted profile
 CMF:

collaborative matrix factorization
 SVD:

singular value decomposition
 PPI:

ProteinProtein interaction
 PCC:

Pearson correlation coefficient
 ROC:

receiver operating characteristic curve
 AUC:

area under the receiver operating characteristic curve
 PRE:

precision
 REC:

recall
 FSC:

Fscore
 ACC:

accuracy
 RF:

random forest
 GTB:

gradient tree boosting
 FPR:

false positive rate
 TPR:

true positive rate
 RBF:

Radial Basis Function
References
Ambros V. The functions of animal micrornas. Nature. 2004; 431(7006):350.
Liu H, Zhang W, Zou B, Wang J, Deng Y, Deng L. DrugCombDB: a comprehensive database of drug combinations toward the discovery of combinatorial therapy. Nucleic Acids Res. 2020; 48(D1):D871–D881. https://doi.org/10.1093/nar/gkz1007.
Nagaraja AK, Creighton CJ, Yu Z, Zhu H, Gunaratne PH, Reid JG, Olokpa E, Itamochi H, Ueno NT, Hawkins SM, et al. A link between mir100 and frap1/mtor in clear cell ovarian cancer. Mol Endocrinol. 2010; 24(2):447–63.
Latronico MV, Catalucci D, Condorelli G. Emerging role of micrornas in cardiovascular biology. Circ Res. 2007; 101(12):1225–36.
NunezIglesias J, Liu CC, Morgan TE, Finch CE, Zhou XJ. Joint genomewide profiling of mirna and mrna expression in alzheimer’s disease cortex reveals altered mirna regulation. PloS ONE. 2010; 5(2):8898.
Jopling CL, Yi M, Lancaster AM, Lemon SM, Sarnow P. Modulation of hepatitis c virus rna abundance by a liverspecific microrna. Science. 2005; 309(5740):1577–81.
Calin GA, Dumitru CD, Shimizu M, Bichi R, Zupo S, Noch E, Aldler H, Rattan S, Keating M, Rai K, et al. Frequent deletions and downregulation of microrna genes mir15 and mir16 at 13q14 in chronic lymphocytic leukemia. Proc Natl Acad Sci. 2002; 99(24):15524–9.
Sredni ST, Huang CC, Bonaldo MdF, Tomita T. Microrna expression profiling for molecular classification of pediatric brain tumors. Pediatr Blood Cancer. 2011; 57(1):183–4.
Lu J, Getz G, Miska EA, AlvarezSaavedra E, Lamb J, Peck D, SweetCordero A, Ebert BL, Mak RH, Ferrando AA, et al. Microrna expression profiles classify human cancers. Nature. 2005; 435(7043):834.
Jiang Q, Hao Y, Wang G, Juan L, Zhang T, Teng M, Liu Y, Wang Y. Prioritization of disease micrornas through a human phenomemicrornaome network. BMC Syst Biol. 2010; 4(1):2.
Jiang Q, Wang G, Wang Y. An approach for prioritizing diseaserelated micrornas based on genomic data integration. In: 2010 3rd International Conference on Biomedical Engineering and Informatics, vol. 6. Yantai: IEEE: 2010. p. 2270–4.
Chen X, Liu MX, Yan GY. Rwrmda: predicting novel human microrna–disease associations. Mol BioSyst. 2012; 8(10):2792–8.
Shi H, Xu J, Zhang G, Xu L, Li C, Wang L, Zhao Z, Jiang W, Guo Z, Li X. Walking the interactome to identify human mirnadisease associations through the functional link between mirna targets and disease genes. BMC Syst Biol. 2013; 7(1):101.
Vanunu O, Magger O, Ruppin E, Shlomi T, Sharan R. Associating genes and protein complexes with disease via network propagation. PLoS Comput Biol. 2010; 6(1):1000641.
Xu J, Li CX, Lv JY, Li YS, Xiao Y, Shao TT, Huo X, Li X, Zou Y, Han QL, et al. Prioritizing candidate disease mirnas by topological features in the mirna target–dysregulated network: Case study of prostate cancer. Mol Cancer Ther. 2011; 10(10):1857–66.
Chen X, Yan GY. Semisupervised learning for potential human micrornadisease associations inference. Sci Rep. 2014; 4:5501.
Zheng K, You ZH, Wang L, Zhou Y, Li LP, Li ZW. Mlmda: a machine learning approach to predict and validate microrna–disease associations by integrating of heterogenous information sources. J Transl Med. 2019; 17(1):260.
Peng J, Hui W, Li Q, Chen B, Hao J, Jiang Q, Shang X, Wei Z. A learningbased framework for mirnadisease association identification using neural networks. Bioinformatics. 2019; 35(21):4364–71.
Yang Y, Fu X, Qu W, Xiao Y, Shen HB. Mirgofs: a gobased functional similarity measurement for mirnas, with applications to the prediction of mirna subcellular localization and mirna–disease association. Bioinformatics. 2018; 34(20):3547–56.
Chen X, Huang L, Xie D, Zhao Q. Egbmmda: extreme gradient boosting machine for mirnadisease association prediction. Cell death Dis. 2018; 9(1):3.
Yin MM, Cui Z, Gao MM, Liu JX, Gao YL. Lwpcmf: Logistic weighted profilebased collaborative matrix factorization for predicting mirnadisease associations. IEEE/ACM Trans Comput Biol Bioinforma. 2019. https://doi.org/10.1109/TCBB.2019.2937774.
Panwar B, Omenn GS, Guan Y. mirmine: a database of human mirna expression profiles. Bioinformatics. 2017; 33(10):1554–60.
Zhang J, Zhang Z, Chen Z, Deng L. Integrating multiple heterogeneous networks for novel lncrnadisease association inference. IEEE/ACM Trans Comput Biol Bioinforma. 2017; 16(2):396–406.
Szklarczyk D, Franceschini A, Wyder S, Forslund K, Heller D, HuertaCepas J, Simonovic M, Roth A, Santos A, Tsafou KP, et al. String v10: protein–protein interaction networks, integrated over the tree of life. Nucleic Acids Res. 2014; 43(D1):447–52.
Van Driel MA, Bruggeman J, Vriend G, Brunner HG, Leunissen JA. A textmining analysis of the human phenome. Eur J Hum Genet. 2006; 14(5):535.
Chou CH, Shrestha S, Yang CD, Chang NW, Lin YL, Liao KW, Huang WC, Sun TH, Tu SJ, Lee WH, et al. mirtarbase update 2018: a resource for experimentally validated micrornatarget interactions. Nucleic Acids Res. 2017; 46(D1):296–302.
Huang Z, Shi J, Gao Y, Cui C, Zhang S, Li J, Zhou Y, Cui Q. Hmdd v3. 0: a database for experimentally supported human microrna–disease associations. Nucleic Acids Res. 2018; 47(D1):1013–17.
Chen T, Guestrin C. Xgboost: A scalable tree boosting system. In: Proceedings of the 22nd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining. San Francisco: ACM: 2016. p. 785–94.
Liaw A, Wiener M, et al. Classification and regression by randomforest. R News. 2002; 2(3):18–22.
Burges CJ. A tutorial on support vector machines for pattern recognition. Data Min Knowl Disc. 1998; 2(2):121–67.
Friedman JH. Greedy function approximation: a gradient boosting machine. Ann Stat. 2001; 29(5):1189–232.
Wang F, Landau D. Determining the density of states for classical statistical models: A random walk algorithm to produce a flat histogram. Phys Rev E. 2001; 64(5):056101.
Liu Y, Zeng X, He Z, Zou Q. Inferring micrornadisease associations by random walk on a heterogeneous network with multiple data sources. IEEE/ACM Trans Comput Biol Bioinforma. 2016; 14(4):905–15.
Shang H, Liu ZP. Prioritizing type 2 diabetes genes by weighted pagerank on bilayer heterogeneous networks. IEEE/ACM Trans Comput Biol Bioinforma. 2019. https://doi.org/10.1109/TCBB.2019.2917190.
Golub GH, Reinsch C. Singular value decomposition and least squares solutions. In: Linear Algebra. Berlin, Heidelberg: Springer: 1971. p. 134–151.
Wang S, Cho H, Zhai C, Berger B, Peng J. Exploiting ontology graph for predicting sparsely annotated gene function. Bioinformatics. 2015; 31(12):357–64.
Zeng X, Liao Y, Liu Y, Zou Q. IEEE/ACM Trans Comput Biol Bioinforma (TCBB). 2017; 14(3):687–95.
Shi C, Kong X, Huang Y, Philip SY, Wu B. Hetesim: A general framework for relevance measure in heterogeneous networks. IEEE Trans Knowl Data Eng. 2014; 26(10):2479–92.
Author information
Authors and Affiliations
Corresponding author
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
Liu, M., Yang, J., Wang, J. et al. Predicting miRNAdisease associations using a hybrid feature representation in the heterogeneous network. BMC Med Genomics 13 (Suppl 10), 153 (2020). https://doi.org/10.1186/s12920020007830
Published:
DOI: https://doi.org/10.1186/s12920020007830