DeepAI
Log In Sign Up

Deep Belief Network based representation learning for lncRNA-disease association prediction

06/22/2020
by   Manu Madhavan, et al.
NITC
0

Background: The expanding research in the field of long non-coding RNAs(lncRNAs) showed abnormal expression of lncRNAs in many complex diseases. Accurately identifying lncRNA-disease association is essential in understanding lncRNA functionality and disease mechanism. There are many machine learning techniques involved in the prediction of lncRNA-disease association which use different biological interaction networks and associated features. Feature learning from the network structured data is one of the limiting factors of machine learning-based methods. Graph neural network based techniques solve this limitation by unsupervised feature learning. Deep belief networks (DBN) are recently used in biological network analysis to learn the latent representations of network features. Method: In this paper, we propose a DBN based lncRNA-disease association prediction model (DBNLDA) from lncRNA, disease and miRNA interactions. The architecture contains three major modules-network construction, DBN based feature learning and neural network-based prediction. First, we constructed three heterogeneous networks such as lncRNA-miRNA similarity (LMS), disease-miRNA similarity (DMS) and lncRNA-disease association (LDA) network. From the node embedding matrices of similarity networks, lncRNA-disease representations were learned separately by two DBN based subnetworks. The joint representation of lncRNA-disease was learned by a third DBN from outputs of the two subnetworks mentioned. This joint feature representation was used to predict the association score by an ANN classifier. Result: The proposed method obtained AUC of 0.96 and AUPR of 0.967 when tested against standard dataset used by the state-of-the-art methods. Analysis on breast, lung and stomach cancer cases also affirmed the effectiveness of DBNLDA in predicting significant lncRNA-disease associations.

READ FULL TEXT VIEW PDF

page 1

page 2

page 3

page 4

07/19/2020

Recent Advances in Network-based Methods for Disease Gene Prediction

Disease-gene association through Genome-wide association study (GWAS) is...
07/12/2019

Towards Probabilistic Generative Models Harnessing Graph Neural Networks for Disease-Gene Prediction

Disease-gene prediction (DGP) refers to the computational challenge of p...
05/15/2017

Representation learning of drug and disease terms for drug repositioning

Drug repositioning (DR) refers to identification of novel indications fo...
08/08/2021

MuCoMiD: A Multitask Convolutional Learning Framework for miRNA-Disease Association Prediction

Growing evidence from recent studies implies that microRNA or miRNA coul...
07/19/2020

EPGAT: Gene Essentiality Prediction With Graph Attention Networks

The identification of essential genes/proteins is a critical step toward...
03/11/2020

A deep belief network-based method to identify proteomic risk markers for Alzheimer disease

While a large body of research has formally identified apolipoprotein E ...

1 Introduction

Long non-coding RNAs (lncRNA) are RNA transcripts with more than 200 nucleotides in length and lack of protein-coding potential Goff and Rinn (2015). The studies revealed that lncRNAs were involved as a regulator in many biological processes such as ageing, cell differentiation, epigenetic mechanisms and protein synthesis Lee (2012). The aberrant expression of lncRNAs is also associated with many complex diseases like cancers Huarte (2015), alzheimer’s disease Luo and Chen (2016), and heart failure Wang et al. (2019). Therefore, identifying the role of lncRNAs in diseases will help to improve the understanding of the disease mechanisms and derive new insights on drug therapeutics Bhartiya et al. (2012).

An array of computational models to predict the lncRNA-disease association are available in the literature with varying performance. All these works utilise varieties of lncRNA features from known lncRNA-disease associations and their interactions with other molecules like micorRNAs (miRNA), proteins and messengerRNAs (mRNA). There are three major categories of works in lncRNA-disease association. The first category of works utilise the knowledge of lncRNA functional similarity under the assumption that functionally similar lncRNAs associate with similar diseases. Based on this assumption, an lncRNA-disease association network was constructed. Then, algorithms from machine learning and social network analysis were used to make the lncRNA-disease prediction. For example, methods such as RWRlncD Sun et al. (2014), IRWRLDA Chen et al. (2016), BRWLDA Yu et al. (2017) used various random walk algorithm and KATZLDA Chen (2015) applied Katz page ranking algorithm for analysing the similarity network. All these methods relied upon the network structure features and results were biased towards nodes having high degree and centrality.

The advanced research on lncRNA mechanism revealed that the regulation of lncRNA is largely determined by co-expressed miRNAs Paraskevopoulou and Hatzigeorgiou (2016). The second category of works used expression levels of lncRNAs, genes and miRNAs in various diseases. The earlier methods Yao et al. (2020) in this category used experimentally validated disease associated genes/miRNAs and lncRNA co-expression data. These methods were not useful with lncRNAs which have no experimentally validated gene/miRNA interaction. Recent works under this category used mathematical models such as matrix completion Lu et al. (2018), matrix factorization Fu et al. (2018) and Graph-based algorithms (TPGLDA Ding et al. (2018), DisLncPri Wang et al. (2017)) to predict lncRNA-disease association.

The third category of works constructed a heterogeneous interaction network based on lncRNA, miRNA and mRNA functional similarity and associations. Machine learning techniques such as Random Walk Yao et al. (2020)

, Support Vector Machine (SVM)

Lan et al. (2017), and Laplacian regularised least square method Chen et al. (2015)

have found use in the analysis of these networks. Major challenge in the above methods was the effective representation of lncRNA-disease features. Introduction of deep learning models eliminated the need for feature extraction by enabling unsupervised representation learning. Among them, CNNLDA

Xuan et al. (2019a)

used convolutional neural networks and GCNLDA

Xuan et al. (2019b) used graph convolutional neural networks Kipf and Welling (2016) to learn global representations of lncRNA, miRNA and disease nodes. A recent work titled GAMCLDA Wu et al. (2020)

used graph autoencoder and matrix completion to predict lncRNA-disease association.

Deep learning algorithms such as Deep Belief Networks (DBN) were used recently in computer vision and text mining to learn the latent representation of the data

Hinton (2009)

. A DBN model has stacked layers of Restricted Boltzman Machines (RBM), which contains visible layers and hidden layers to compute probability distribution as latent representation. DBN based models were successfully used in Bioinformatics to predict drug-target

Wang and Zeng (2013), multiple types of miRNA-disease association Chen et al. (2015); Luo et al. (2019) and cancer sub-type prediction Liang et al. (2014), but not applied in the field of lncRNA-disease association prediction.

In this study, we propose a DBN based representation learning model (called DBNLDA) for lncRNA-disease association prediction. DBNLDA make use of heterogeneous information on functional similarity, co-expression and interactions between lncRNAs and diseases for making the prediction. The architecture of DBNLDA contains 3 modules namely-(i)network construction, (ii) DBN based feature learning and (iii) association prediction. Summary of the proposed architecture is depicted in Figure 1. Instead of using a single heterogeneous network of lncRNAs, miRNAs and diseases as in previous works Yao et al. (2020); Xuan et al. (2019b), DBNLDA constructs three functional similarity networks such as lncRNA-miRNA (LMS), disease-miRNA (DMS) and lncRNA-Disease association (LDA). Then, for each pair of lncRNA-disease, DBN subnetwork-1 learns lncRNA and disease representations from their functional associations with miRNAs. Similarly, DBN subnetwork-2 learns lncRNA-disease representations from LDA network. The representations learned by subnetworks are combined for learning higher-level representation using the third DBN (DBN-combined). In order to reduce the obstructive impact of features learned from sparse network, the features from DBN-combined are recomputed using an attention layer. In the final stage, a neural network-based classifier is used to predict the lncRNA-disease association based on the joint feature representation. A 5-fold cross-validation and case studies on cancer dataset showed that DBNLDA significantly improves the performance and potential lncRNA-disease association prediction.

Figure 1: Architecture of DBNLDA. LMS:LncRNA-miRNA similarity; DMS: Disease-miRNA similarity; LDA: LncRNA-Disease Association; L:embedding of lncRNAs from LMS; D: embedding of diseases from DMS; LD: embedding of lncRNAs and diseases from LDA; DBN: Deep Belief Network.

The remainder of the paper is organized as follows. Section 2 gives a detailed discussion of DBNLDA architecture. Section 3 discusses the results of the proposed method and Section 4 concludes the paper.

2 Materials and Methods

2.1 The Dataset

Datasets used in the previous reports Yao et al. (2020); Xuan et al. (2019b); Fu et al. (2018) were used for lncRNA-Disease associations, lncRNA-miRNA interaction, lncRNA function similarity and disease semantic similarity. LncRNA-disease associations were downloaded from two reference databases: LncRNADisease Chen et al. (2012) and lnc2cancer Ning et al. (2016). MiRNA-lncRNA interactions and miRNA-disease interactions were downloaded from miRNet Fan et al. (2016) and Starbase Li et al. (2014) databases respectively. All these downloaded associations and similarities were then compiled for 240 lncRNAs, 412 diseases, 495 miRNAs and 2697 known lncRNA- disease interactions. These known interaction pairs constituted the positive example for training the model. All other pairs between lncRNAs and diseases, which were not listed in reference databases, were considered to be negative interactions. We randomly selected 2967 samples from negative interactions to construct a balanced dataset. Summary of the dataset is given in Table 1.

Particular Count
LncRNAs 290
Diseases 412
MiRNAs 495
Positive pairs 2697
Negative pairs 2697
Table 1: Summary of dataset

2.2 Construction of LMS, DMS and LDA networks

The first step in the DBNLDA architecture was the construction of three similarity networks such as LMS, DMS and LDA as defined in Section 1. Let and be the number of lncRNAs, miRNAs and diseases in the dataset. LMS network (with lncRNAs and miRNAs) was constructed using lncRNA-lncRNA similarity and lncRNA-miRNA interactions. LncRNA functional similarities computed by Chen’s method Chen et al. (2015) was adapted here, and an edge was added if the similarity score was greater than 0. For the list of lncRNAs, the known miRNA targets downloaded from the miRNet were used as lncRNA-miRNA edges.

Similar to LMS, DMS network was constructed (with diseases and miRNAs) using disease-disease similarities and known disease-miRNA associations. LncRNA-Disease association (LDA) network contains lncRNAs and diseases, where an un-directed edge was used to represent the known association between lncRNA and disease.

2.3 Computing node embedding features from networks

Node2vec Grover and Leskovec (2016) algorithm was used to get node embedding from each network. Node2vec output a node representation based on the neighbourhood information of the node. First, LMS network was passed through node2vec embedding layer, and embedding matrix for lncRNA was obtained as , where is the embedding dimension. The embedding matrix for diseases from DMS was obtained as . Then, for each lncRNA-disease pair in the dataset (both positive and negative examples), and were concatenated to form vector . This resulted in a feature matrix , where is the total number of examples in the dataset. Finally, the embedding matrices of lncRNAs and diseases, represented as , were obtained from LDA network. Then as in the case of , for each pairs of lncRNA-disease, vectors from and were concatenated to form feature matrix .

2.4 DBN based feature learning

Following the Node2vec embedding layer, DBNLDA implements Deep Belief Network based feature learning. In this work, DBN was used to learn latent representation of lncRNA and disease nodes and encode them to a new dimension , where . The architecture consisted of two DBN subnetworks- one to learn lncRNA-disease representation from functional similarity networks (LMS and DMS) and other from LDA network. The DBN subnetwork-1 received embedded feature matrix as input, and produce (, the number of hidden units in DBN) as output. Similarly, the DBN subnetwork-2 received as input and produce as output. Then both and were concatenated to form combined feature representation, . The network DBN-combined accepted as input and learn the feature representation as , where is the number of hidden units in the DBN-combined and .

2.5 Feature attention layer

Since all the three networks used in the DBNLDA architecture were sparse in nature, there was a chance of losing feature importance in the latent representation learned by DBN. The attention mechanism in deep learning was used to solve this issue by recomputing the feature values from all available information so that their contributions could be different and unique. In this work, attention mechanism similar to the one used in GCNLDA Xuan et al. (2019b) was applied to features.

Let represents the feature vector of entry in the dataset learned by DBN-combined, where . The attention score for each element in was calculated by introducing attention weight parameters , and bias as in Equation 1.

(1)

Next, the attention enhanced feature values were recomputed as in Equation 2.

(2)

where represents pairwise multiplication. Finally, the matrix was used as input for association prediction.

2.6 Prediction Layer

The final module of DBNLDA architecture is a neural network classifier to predict the association between lncRNA and disease. The classifier network followed a 1-2-1 structure, where the input layer has nodes to receive input from

and output layer has one node to predict the association score. The number of nodes in hidden layers kept as hyperparameter. Activation function used in all layers except the output was Relu. In order to reduce the overfitting, a dropout of 0.02 probability was added between the hidden layers. Output layer used a sigmoid function to compute the association score. The network used ADAM optimizer with learning rate 0.01 and binary cross-entropy loss function.

2.7 Hyperparameters

Various hyperparameters determine the performance of DBNLDA in different modules. The values of these hyperparameters were set empirically following performance evaluation. The dimension of node2vec embedding, is set to be 64 for all networks with keeping other parameters to default values in Grover and Leskovec (2016). Following the implementation of Luo et al. (2019), the DBN subnetworks in this work uses three stacked layers of RBM, with nodes in all layers. The combined DBN has dimension . It is found that very low values and degrade the performance and high values have no effect on the performance (see Figure  2

for details). For the classification module, the number of neurons in both hidden layers set as 128. The classifier iterate over 30 epochs, since the model performance became stable after

epoch.

Figure 2: Comparison of accuracy of DBNLDA model on different number of nodes in DBN subnetworks () and DBN combined (). Points are annotated with (, ) values

A detailed description DBNLDA steps and implementation details are available in SupplementaryFile-S1.pdf.

3 Results and Discussion

3.1 Performance evaluation metrics

We used 5-fold cross-validation to measure and compare the performance of the model. The dataset consist of 5394 lncRNA-disease pairs (2697 positive associations and 2697 negative associations). The area under the receiver operating characteristic (AUC-ROC) curve was used to compare the global performance of the prediction model. The area under the precision-recall curve (AUPR) and average accuracy were also used to measure the prediction performance.

3.2 Overall performance

The DBNLDA was trained over 30 epochs and the learning curve (refer Figure 3) showed consistent characteristics in all folds of cross validation. The model gave average AUC of 0.96 over cross validations and ROC curve is shown in Figure 4. The model reported an accuracy of 0.957 and AUPR value of 0.968.

Figure 3: Learning Curve for different folds of 5-fold cross-validation
Figure 4: ROC curve for different folds of 5-fold cross-validation

In order to analyse how DBN based features improve the performance of the prediction model, we have repeated the experiments with different levels of feature combinations. It is clear from the Table 2 that introduction of DBN based learning significantly improved the accuracy of the prediction model.

Experiment Feature Accuracy
Exp 1 only Node2vec featues 0.817
Exp 2 Node2vec, DBN subnetwork-1 and 0.896
DBN subnetwork-2
Exp 3 Node2vec, DBN subnetwork-1, 0.956
DBN subnetwork-2 and DBN-combined
Table 2: Accuracy of the model based on feature combinations

3.3 Comparison with other methods

We compared the performance of DBNLDA with other state-of-the-art methods such as RFLDA Yao et al. (2020), GCNLDA Xuan et al. (2019b), SIMCLDALu et al. (2018), Ping’s Method Ping et al. (2018), MFLDA Fu et al. (2018), LDAP Lan et al. (2017), GAMCLDA Wu et al. (2020) and CNNLDA Xuan et al. (2019a), based on area under ROC curve (AUC) and Area under Precision-Recall curve (AUPR). The above methods used knowledge from heterogeneous information from different data sources to predict lncRNA-disease association and not consider network structure features. AUC and AUPR values of all LDA prediction models were given in Table 3. The AUC and AUPR values of all methods except DBNLDA were taken from Yao et al. (2020) and Wu et al. (2020). It was evident from the table that DBNLDA reported second best AUC (0.96) which is closer (1.6% less) to the highest AUC value reported by RFLDA. Moreover, DBNLDA outperforms all other methods in terms of AUPR values, where DBNLDA have 0.968 which is 18.9% better than the second highest value (RFLDA). These results show that DBNLDA predicts lncRNA-disease associations effectively.

Method AUC AUPR
MFLDA Fu et al. (2018) 0.626 0.066
SIMCLDA Lu et al. (2018) 0.746 0.095
LDAP Lan et al. (2017) 0.863 0.166
Ping’s Method Ping et al. (2018) 0.871 0.219
GAMCLDA Wu et al. (2020) 0.907 0.037
CNNLDA Xuan et al. (2019a) 0.952 0.251
GCNLDA Xuan et al. (2019b) 0.959 0.223
RFLDA Yao et al. (2020) 0.976 0.779
DBNLDA 0.960 0.968
Table 3: Comparison of performance with state-of-the-art methods

3.4 Case studies

lncRNA Rank  Evidence
GAS5 1 Lnc2Cancer
DLEU2 2 LncRNADisease
HCP5 3 Literature (PMID: 31864836) Wu et al. (2019)
HOTAIR 4 LncRNADisease, Lnc2Cancer
MEG3 5 LncRNADisease, Lnc2Cancer
HULC 6 Lnc2Cancer
BCYRN1 7 LncRNADisease
HOTTIP 8 Lnc2Cancer
UCA1 9 LncRNADisease, Lnc2Cancer
CDKN2B-AS1 10 LncRNADisease
NEAT1 11 LncRNADisease, Lnc2Cancer
TUG1 12 LncRNADisease, Lnc2Cancer
AFAP1-AS1 13 Lnc2Cancer
MIR100HG 14 Literature (PMID:30042378) Wang et al. (2018)
TINCR 15 Lnc2Cancer
Table 4: Top 15 DBNLDA predicted lncRNAs associated with breast cancer
lncRNA Rank  Evidence
TUG1 1 Literature (PMID:31532756) Guo et al. (2019)
PVT1 2 Lnc2Cancer
AFAP1-AS1 3 LncRNADiease, Lnc2Cancer
XIST 4 Literature (PMID: 28448993) Wang et al. (2017)
CCAT2 5 LncRNADisease
MALAT1 6 LncRNADiease, Lnc2Cancer
HOTTIP 7 LncRNADiease, Lnc2Cancer
SOX2-OT 8 LncRNADiease
HULC 9 Literature (PMID:30575912) Liu et al. (2018)
MIR155HG 10 Literature (PMID:32129458) Song et al. (2020)
CDKN2B-AS1 11 Literature (PMID:29541247) Du et al. (2018)
BANCR 12 LncRNADiease, Lnc2Cancer
BCYRN1 13 LncRNADiease
UCA1 14 LncRNADiease, Lnc2Cancer
H19 15 LncRNADiease, Lnc2Cancer
Table 5: Top 15 DBNLDA predicted lncRNAs associated with lung cancer
lncRNA Rank  Evidence
MIR17HG 1 Literature (PMID:26837962) Bahari et al. (2015)
BCYRN1 2 Literature (PMID:29435146) Ren et al. (2018)
BANCR 3 LncRNADisease
HCP5 4 *
AFAP1-AS1 5 LncRNADisease
HNF1A-AS1 6 LncRNADisease, Lnc2Cancer
NALT1 7 Literature (PMID:31802831) Piao et al. (2019)
DANCR 8 Lnc2Cancer
MIR99AHG 9 *
GAS5 10 LncRNADisease, Lnc2Cancer
HULC 11 LncRNADisease, Lnc2Cancer
HCG4 12 *
XIST 13 LncRNADisease, Lnc2Cancer
HOTTIP 14 LncRNADisease, Lnc2Cancer
UCA1 15 LncRNADisease, Lnc2Cancer
Table 6: Top 15 DBNLDA predicted lncRNAs associated with stomach cancer.

To further investigate the ability of DBNLDA in predicting significant lncRNA-disease associations, case studies on breast cancer, lung cancer and stomach cancer were conducted. For this study, first trained the DBNLDA model on a dataset containing all lncRNA-disease associations except the validated association between lncRNAs and the disease of interest (breast/lung/stomach cancer). Then the association score for all lncRNAs to the particular disease was calculated using the trained model and analysed the top 15 candidate lncRNAs for each disease. Tables 4,5, and 6 show the top 15 candidate lncRNAs in breast/lung/stomach cancer, predicted by DBNLDA. The evidence column shows the reference to the associations either from reference databases or literature.

It was found that 13 (86.67%) lncRNAs associated with breast cancer predicted by DBNLDA were also confirmed by lnc2cancer or LncRNADisease database. For the two unconfirmed predictions, we could find the evidence from recent publications. In the case of lung cancer, among the top 15 predicted lncRNAs, 10 were confirmed by reference databases and remaining were reported in recent literature. In case of stomach cancer, DBNLDA could predict 12 associations reported either by reference databases or literature. The three new associations (HCP5, HCG4 and MIR99AHG, indicated by ’*’ in Table 6) which could be considered as new suggestions for further laboratory validations. The detailed comparison of LDA prediction by DBNLDA is available in SupplementaryTable-S2.xls.

4 Conclusion

Identifying disease-associated lncRNA is a necessary step in discerning the functional roles of lncRNAs in the disease mechanism. In this work, we developed DBNLDA, a deep belief network-based model for lncRNA-disease association prediction. This work integrates information of lncRNA, miRNA, disease interactions, and functional similarities to construct heterogeneous networks. Then DBN based latent representations of lncRNAs and diseases are used to predict the lncRNA-disease association accurately. The cross-validation confirmed that DBNLDA has comparable performance in terms of AUC and significant improvement in terms of AUPR. Case studies on breast cancer, lung cancer and stomach cancer show the ability of DBNLDA to predict potential disease-associated lncRNAs. The model could be extended further with multi-modal data such as lncRNA drug-target interactions and lncRNA-epigenetic-disease interactions.

Acknowledgement

This research work is an outcome of the R& D work under the Visvesvaraya PhD Scheme of Ministry of Electronics and Information Technology, Government of India

References

References

  • Goff and Rinn (2015) L. A. Goff, J. L. Rinn, Linking rna biology to lncrnas, Genome research 25 (2015) 1456–1465.
  • Lee (2012) J. T. Lee, Epigenetic regulation by long noncoding rnas, Science 338 (2012) 1435–1439.
  • Huarte (2015) M. Huarte, The emerging role of lncrnas in cancer, Nature medicine 21 (2015) 1253.
  • Luo and Chen (2016) Q. Luo, Y. Chen, Long noncoding rnas and alzheimer’s disease, Clinical interventions in aging 11 (2016) 867.
  • Wang et al. (2019) G. Wang, X. Zheng, Y. Zheng, R. Cao, M. Zhang, Y. Sun, J. Wu, Construction and analysis of the lncrna-mirna-mrna network based on competitive endogenous rna reveals functional genes in heart failure, Molecular medicine reports 19 (2019) 994–1003.
  • Bhartiya et al. (2012) D. Bhartiya, S. Kapoor, S. Jalali, S. Sati, K. Kaushik, C. Sachidanandan, S. Sivasubbu, V. Scaria, Conceptual approaches for lncrna drug discovery and future strategies, Expert opinion on drug discovery 7 (2012) 503–513.
  • Sun et al. (2014) J. Sun, H. Shi, Z. Wang, C. Zhang, L. Liu, L. Wang, W. He, D. Hao, S. Liu, M. Zhou, Inferring novel lncrna–disease associations based on a random walk model of a lncrna functional similarity network, Molecular BioSystems 10 (2014) 2074–2081.
  • Chen et al. (2016) X. Chen, Z.-H. You, G.-Y. Yan, D.-W. Gong, Irwrlda: improved random walk with restart for lncrna-disease association prediction, Oncotarget 7 (2016) 57919.
  • Yu et al. (2017) G. Yu, G. Fu, C. Lu, Y. Ren, J. Wang, Brwlda: bi-random walks for predicting lncrna-disease associations, Oncotarget 8 (2017) 60429.
  • Chen (2015) X. Chen, Katzlda: Katz measure for the lncrna-disease association prediction, Scientific reports 5 (2015) 16840.
  • Paraskevopoulou and Hatzigeorgiou (2016) M. D. Paraskevopoulou, A. G. Hatzigeorgiou, Analyzing mirna–lncrna interactions, in: long non-coding RNAs, Springer, 2016, pp. 271–286.
  • Yao et al. (2020) D. Yao, X. Zhan, X. Zhan, C. K. Kwoh, P. Li, J. Wang,

    A random forest based computational model for predicting novel lncrna-disease associations,

    BMC bioinformatics 21 (2020) 1–18.
  • Lu et al. (2018) C. Lu, M. Yang, F. Luo, F.-X. Wu, M. Li, Y. Pan, Y. Li, J. Wang, Prediction of lncrna–disease associations based on inductive matrix completion, Bioinformatics 34 (2018) 3357–3364.
  • Fu et al. (2018) G. Fu, J. Wang, C. Domeniconi, G. Yu, Matrix factorization-based data fusion for the prediction of lncrna–disease associations, Bioinformatics 34 (2018) 1529–1537.
  • Ding et al. (2018) L. Ding, M. Wang, D. Sun, A. Li, Tpglda: Novel prediction of associations between lncrnas and diseases via lncrna-disease-gene tripartite graph, Scientific reports 8 (2018) 1–11.
  • Wang et al. (2017) P. Wang, Q. Guo, Y. Gao, H. Zhi, Y. Zhang, Y. Liu, J. Zhang, M. Yue, M. Guo, S. Ning, et al., Improved method for prioritization of disease associated lncrnas based on cerna theory and functional genomics data, Oncotarget 8 (2017) 4642.
  • Lan et al. (2017) W. Lan, M. Li, K. Zhao, J. Liu, F.-X. Wu, Y. Pan, J. Wang, Ldap: a web server for lncrna-disease association prediction, Bioinformatics 33 (2017) 458–460.
  • Chen et al. (2015) X. Chen, C. C. Yan, C. Luo, W. Ji, Y. Zhang, Q. Dai, Constructing lncrna functional similarity network based on lncrna-disease associations and disease semantic similarity, Scientific reports 5 (2015) 11338.
  • Xuan et al. (2019a) P. Xuan, Y. Cao, T. Zhang, R. Kong, Z. Zhang, Dual convolutional neural networks with attention mechanisms based method for predicting disease - related lncrna genes, Frontiers in genetics 10 (2019a) 416.
  • Xuan et al. (2019b) P. Xuan, S. Pan, T. Zhang, Y. Liu, H. Sun, Graph convolutional network and convolutional neural network based method for predicting lncrna-disease associations, Cells 8 (2019b) 1012.
  • Kipf and Welling (2016) T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, arXiv preprint arXiv:1609.02907 (2016).
  • Wu et al. (2020) X. Wu, W. Lan, Q. Chen, Y. Dong, J. Liu, W. Peng, Inferring lncrna-disease associations based on graph autoencoder matrix completion, Computational Biology and Chemistry (2020) 107282.
  • Hinton (2009) G. E. Hinton, Deep belief networks, Scholarpedia 4 (2009) 5947.
  • Wang and Zeng (2013) Y. Wang, J. Zeng,

    Predicting drug-target interactions using restricted boltzmann machines,

    Bioinformatics 29 (2013) i126–i134.
  • Chen et al. (2015) X. Chen, C. C. Yan, X. Zhang, Z. Li, L. Deng, Y. Zhang, Q. Dai, Rbmmmda: predicting multiple types of disease-microrna associations, Scientific reports 5 (2015) 13877.
  • Luo et al. (2019) P. Luo, Y. Li, L.-P. Tian, F.-X. Wu, Enhancing the prediction of disease–gene associations with multimodal deep learning, Bioinformatics 35 (2019) 3735–3742.
  • Liang et al. (2014) M. Liang, Z. Li, T. Chen, J. Zeng, Integrative data analysis of multi-platform cancer data with a multimodal deep learning approach, IEEE/ACM transactions on computational biology and bioinformatics 12 (2014) 928–937.
  • Chen et al. (2012) G. Chen, Z. Wang, D. Wang, C. Qiu, M. Liu, X. Chen, Q. Zhang, G. Yan, Q. Cui, Lncrnadisease: a database for long-non-coding rna-associated diseases, Nucleic acids research 41 (2012) D983–D986.
  • Ning et al. (2016) S. Ning, J. Zhang, P. Wang, H. Zhi, J. Wang, Y. Liu, Y. Gao, M. Guo, M. Yue, L. Wang, et al., Lnc2cancer: a manually curated database of experimentally supported lncrnas associated with various human cancers, Nucleic acids research 44 (2016) D980–D985.
  • Fan et al. (2016) Y. Fan, K. Siklenka, S. K. Arora, P. Ribeiro, S. Kimmins, J. Xia, mirnet-dissecting mirna-target interactions and functional associations through network-based visual analysis, Nucleic acids research 44 (2016) W135–W141.
  • Li et al. (2014) J.-H. Li, S. Liu, H. Zhou, L.-H. Qu, J.-H. Yang, starbase v2. 0: decoding mirna-cerna, mirna-ncrna and protein–rna interaction networks from large-scale clip-seq data, Nucleic acids research 42 (2014) D92–D97.
  • Chen et al. (2015) X. Chen, C. C. Yan, C. Luo, W. Ji, Y. Zhang, Q. Dai, Constructing lncrna functional similarity network based on lncrna-disease associations and disease semantic similarity, Scientific reports 5 (2015) 11338.
  • Grover and Leskovec (2016) A. Grover, J. Leskovec, node2vec: Scalable feature learning for networks, in: Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, 2016, pp. 855–864.
  • Ping et al. (2018) P. Ping, L. Wang, L. Kuang, S. Ye, M. F. B. Iqbal, T. Pei, A novel method for lncrna-disease association prediction based on an lncrna-disease association network, IEEE/ACM transactions on computational biology and bioinformatics 16 (2018) 688–693.
  • Wu et al. (2019) J. Wu, H. Chen, M. Ye, B. Wang, Y. Zhang, J. Sheng, T. Meng, H. Chen, Downregulation of long noncoding rna hcp5 contributes to cisplatin resistance in human triple-negative breast cancer via regulation of pten expression, Biomedicine & Pharmacotherapy 115 (2019) 108869.
  • Wang et al. (2018) S. Wang, H. Ke, H. Zhang, Y. Ma, L. Ao, L. Zou, Q. Yang, H. Zhu, J. Nie, C. Wu, et al., Lncrna mir100hg promotes cell proliferation in triple-negative breast cancer through triplex formation with p27 loci, Cell death & disease 9 (2018) 1–11.
  • Guo et al. (2019) S. Guo, L. Zhang, Y. Zhang, Z. Wu, D. He, X. Li, Z. Wang, Long non-coding rna tug1 enhances chemosensitivity in non-small cell lung cancer by impairing microrna-221-dependent pten inhibition, Aging (Albany NY) 11 (2019) 7553.
  • Wang et al. (2017) H. Wang, Q. Shen, X. Zhang, C. Yang, S. Cui, Y. Sun, L. Wang, X. Fan, S. Xu, The long non-coding rna xist controls non-small cell lung cancer proliferation and invasion by modulating mir-186-5p, Cellular Physiology and Biochemistry 41 (2017) 2221–2229.
  • Liu et al. (2018) L. Liu, X. Zhou, J. Zhang, G. Wang, J. He, Y. Chen, C. Huang, L. Li, S. Li, Lncrna hulc promotes non-small cell lung cancer cell proliferation and inhibits the apoptosis by up-regulating sphingosine kinase 1 (sphk1) and its downstream pi3k/akt pathway, Eur. Rev. Med. Pharmacol. Sci 22 (2018) 8722–8730.
  • Song et al. (2020) J. Song, Q. Wang, L. Zong, Lncrna mir155hg contributes to smoke-related chronic obstructive pulmonary disease by targeting mir-128-5p/brd4 axis, Bioscience Reports 40 (2020).
  • Du et al. (2018) Y. Du, X. Hao, X. Liu, Low expression of long noncoding rna cdkn2b-as1 in patients with idiopathic pulmonary fibrosis predicts lung cancer by regulating the p53-signaling pathway, Oncology letters 15 (2018) 4912–4918.
  • Bahari et al. (2015) F. Bahari, M. Emadi-Baygi, P. Nikpour, et al., mir-17-92 host gene, uderexpressed in gastric cancer and its expression was negatively correlated with the metastasis, Indian journal of cancer 52 (2015) 22.
  • Ren et al. (2018) H. Ren, X. Yang, Y. Yang, X. Zhang, R. Zhao, R. Wei, X. Zhang, Y. Zhang, Upregulation of lncrna bcyrn1 promotes tumor progression and enhances epcam expression in gastric carcinoma, Oncotarget 9 (2018) 4851.
  • Piao et al. (2019) H.-Y. Piao, S. Guo, Y. Wang, J. Zhang, Long noncoding rna nalt1-induced gastric cancer invasion and metastasis via notch signaling pathway, World Journal of Gastroenterology 25 (2019) 6508.