class="head no_bottom_margin" id="sec1title">IntroductionMicroRNAs (miRNAs) are a type of small non-coding RNAs with a size of about 22 nucleotides, and they interact with other RNAs to play important roles in transcriptional and post-transcriptional gene regulation (). It is estimated that over 60% of all human protein-coding genes (PCGs) are regulated by miRNAs (), and these miRNAs have been implicated in diseases. To date, the associations between diseases and PCGs are well investigated; many disease-PCG associations have been discovered and collected in public databases, e.g., DISEASES (), OUGene (), and DisGeNET (). Compared with PCG's well-known important roles in diseases, the studies of effects of miRNAs are increasing. With increasing high-throughput sequencing data generated, more and more miRNAs are being discovered, and experimentally identifying their functions is costly and time consuming. Thus, it is imperative to develop computational methods to identify functional miRNA biomarkers associated with diseases, especially using rich information buried in disease-associated PCGs.Some miRNAs are mainly expressed in certain tissues and show tissue specificity (), which have certain tissue-specific expression patterns associated with diseases (). They are expected to behave similarly to other disease-associated genes like PCGs or long non-coding RNAs (lncRNAs). Thus, several existing computational methods have used tissue expression data to infer gene-disease associations. For instance, GeneTIER makes use of disease-tissue associations to prioritize disease candidate genes (). NetWAS identifies disease-associated genes by combining tissue-specific interaction networks and genome-wide association studies (). Especially, some methods use tissue expression profiles with machine learning models to infer disease-associated lncRNAs. For example, DislncRF trains machine learning models on tissue expression profiles of disease-associated PCGs and further applies the trained models to infer disease-associated lncRNAs (). All the above-mentioned studies demonstrated that tissue expression profiles indeed can facilitate the detection of disease-gene associations.On the other hand, interaction networks contain rich clues for linking miRNAs to diseases. Many computational methods have been developed under the context of gene-gene networks (). For example, Jiang el al. integrate miRNA and disease similarity network and miRNA-disease association to prioritize disease candidate miRNAs using a network-based approach (); midp applies random walk on the interaction network to infer disease-associated miRNAs (). Similarly, RWRMDA implements random walk on the miRNA functional similarity network to link miRNAs to diseases (); the MDHGI integrates the predicted association score based on sparse learning method to infer disease-associated miRNAs (). More closely related studies are as follows: DRMDA applies stacked autoencoder to learn deep representation for predicting miRNA-disease association (href="#bib8" rid="bib8" class=" bibr popnode">Chen et al., 2018a), LRSSLMDA and PBMDA use Laplacian regularized sparse subspace learning and path-based computational model for miRNA-disease association prediction (href="#bib7" rid="bib7" class=" bibr popnode">Chen and Huang, 2017, href="#bib43" rid="bib43" class=" bibr popnode">You et al., 2017), BNPMDA uses Bipartite Network Projection based on the known miRNA-disease associations (href="#bib9" rid="bib9" class=" bibr popnode">Chen et al., 2018b), and KBMF-MDI employs kernelized Bayesian matrix factorization to score miRNA-disease associations by integrating disease and miRNA similarity (href="#bib22" rid="bib22" class=" bibr popnode">Lan et al., 2018). Similarly, DLRMC infers disease-associated miRNAs using dual Laplacian regularized matrix completion (href="#bib37" rid="bib37" class=" bibr popnode">Tang et al., 2019). FamCluRank applies non-negative matrix factorization on the heterogeneous network with node attributes to predict disease-associated miRNAs (href="#bib41" rid="bib41" class=" bibr popnode">Xuan et al., 2018b). Especially deep learning has been utilized to extract deep representation for disease-miRNA association prediction (href="#bib40" rid="bib40" class=" bibr popnode">Xuan et al., 2018a).A common hypothesis for the above methods is they assume that similar miRNAs can be associated with the same disease and similar diseases would be associated with the same miRNA. Thus, they commonly train and evaluate the models with representations of miRNAs and diseases as inputs on verified disease-miRNA associations through cross-validation approach.However, as pointed out in href="#bib24" rid="bib24" class=" bibr popnode">Lehtinen et al. (2015), in the context of gene function prediction, cross-validation may be problematic because some gene-function associations are not independent in the benchmark set. There exists the same issue for disease-miRNA associations due to the following: (1) miRNAs from the same family may be associated with the same disease, (2) disease-associated miRNAs from miRNA-target assay may be derived from the targets that these miRNAs interact with, and (3) the associated miRNAs of child diseases are related to the miRNAs of parent diseases in disease ontology. When training and evaluating the models using cross-validation, randomly dividing the disease-miRNA associations may cause dependent associations to be separated into the training and test sets, potentially leading to an overestimated predictive performance. Cross-validating miRNA-disease associations may not actually reflect the method's ability to predict new miRNA-disease associations, but rather which information is dissipated in the benchmark set. In addition, as reported in href="#bib33" rid="bib33" class=" bibr popnode">Park and Marcotte (2012), there may exist flaws in cross-validation for computational pair-input prediction. One disease or one miRNA may be associated with multiple miRNAs or diseases, so randomly dividing disease-miRNA pairs into training and test sets will make some pairs in the test set share either the miRNA or the disease with the pairs in the training set, which causes the trained models to not generalize well to unseen disease-miRNA associations.Thus, during cross-validation, complicated steps are required to make sure that dependent samples are divided into the same training set or the same test set and that pairs in the training and test sets do not share the miRNA or disease. It is almost impossible to construct a completely independent test set. An alternative strategy is that we do not use disease-miRNA associations for model training. For instance, instead of using miRNA-disease associations, the miRPD approach combines PCG-disease associations and miRNA-PCG network to score miRNAs and diseases (href="#bib27" rid="bib27" class=" bibr popnode">Mork et al., 2014). This has triggered us to further investigate disease-miRNA associations based on an interaction network. To date, there exist many high-confidence disease-PCG associations, and one miRNA may share the same disease with its PCG targets; we will be capable of transferring PCG-associated diseases to miRNAs on an interaction network under a new semi-supervised framework.Recently, deep learning has achieved remarkable results in computational biology (href="#bib3" rid="bib3" class=" bibr popnode">Angermueller et al., 2016, href="#bib13" rid="bib13" class=" bibr popnode">Ching et al., 2018), especially convolutional neural networks (CNNs) (href="#bib23" rid="bib23" class=" bibr popnode">Lecun et al., 1998). CNNs can capture local correlation buried in data and mainly consist of convolutional layers, pooling layers, and fully connected layers. Many studies have demonstrated that the CNN networks are powerful in learning the hidden patterns from complicated biological data. For example, DeepBind (href="#bib1" rid="bib1" class=" bibr popnode">Alipanahi et al., 2015) and DeepSEA (href="#bib45" rid="bib45" class=" bibr popnode">Zhou and Troyanskaya, 2015) apply CNNs to predict preference of DNA/RNA-binding proteins and the impact of non-coding variants, respectively. iDeep (href="#bib29" rid="bib29" class=" bibr popnode">Pan and Shen, 2017) and iDeepE (href="#bib30" rid="bib30" class=" bibr popnode">Pan and Shen, 2018) further improve the performance of predicting RNA-binding protein (RBP)-binding sites and motifs using hybrid CNNs. The iDeepS (href="#bib31" rid="bib31" class=" bibr popnode">Pan et al., 2018) identifies binding sequence and structure preferences of RBPs simultaneously using CNNs and long short-term memory network.Although the CNN has shown its power, it cannot handle structured datasets, like gene-gene networks. To analyze these types of network data, graph convolutional networks (GCNs) have been developed (href="#bib15" rid="bib15" class=" bibr popnode">Defferrard et al., 2016, href="#bib18" rid="bib18" class=" bibr popnode">Hamilton et al., 2017, href="#bib21" rid="bib21" class=" bibr popnode">Kipf and Welling, 2017). Under the framework of spectral graph convolutions, it encodes both local graph structure and features of nodes. The GCNs have been used on the graph data to predict polypharmacy side effects, where the graph is a multimodal graph constructed from protein-protein interactions, drug-protein interactions, and the polypharmacy side effects (href="#bib46" rid="bib46" class=" bibr popnode">Zitnik et al., 2018). The GCN is a graph-based semi-supervised learning method that does not require labels for all nodes. This setting is especially powerful for inferring miRNA-associated diseases, because many miRNAs are not well investigated about their associations with diseases and many disease-PCG associations are available. Compared with traditional semi-supervised methods (href="#bib19" rid="bib19" class=" bibr popnode">Jia et al., 2016, href="#bib38" rid="bib38" class=" bibr popnode">Wan and Wang, 2019, href="#bib44" rid="bib44" class=" bibr popnode">Zhang et al., 2018, href="#bib47" rid="bib47" class=" bibr popnode">Zoidi et al., 2018), GCNs can capture the structural information within the node's local network, similar to CNNs in images. In addition, one PCG or miRNA can be associated with multiple diseases. Thus, we can formulate the prediction of disease-miRNA associations as a multi-label classification problem.In this study, we present a new semi-supervised multi-label learning method, DimiG, based on GCNs to integrate multiple networks of PCG-PCG interactions, PCG-miRNA interactions, PCG-disease associations, and tissue expression profiles to infer miRNA-associated diseases. The DimiG does not require the disease-miRNA associations, and it is trained on the graph consisting of PCG-PCG and miRNA-PCG interactions, where only PCGs have labeled diseases. Then DimiG is further used to score associations between diseases and miRNAs.This study has made the following four major contributions for understanding disease-miRNA associations. (1) We further demonstrate that cross-validation performance of methods trained on known disease-miRNA associations could be overestimated and may not be able to reflect the method's actual ability to predict new disease-miRNA associations. We have proposed a network-based knowledge transfer approach for this problem. Considering that an miRNA may share the same disease with its PCG targets and there exist many high-confidence disease-PCG associations, we will be able to transfer the PCG-associated diseases to miRNAs in an interaction network framework. (2) We have formulated disease-miRNA association prediction as a semi-supervised multi-label node classification in a graph, which can help learn the complex networks composed of unlabeled miRNAs and labeled PCGs and the multi-label associations. This is a new prediction protocol for this problem. (3) We use semi-supervised GCN to learn patterns from PCG-associated diseases on an interaction network, which are further used to score diseases and miRNAs. This GCN-based approach combines the advantages of deep learning for representation learning and network-based methods. (4) We have further incorporated the domain knowledge into our model construction. Considering that miRNAs are often expressed in a tissue-specific way, we integrate the expression profiles across tissues into our GCN framework. Our results demonstrate that informative signals in more tissues can be captured for aiding the inference of disease-associated miRNAs.
展开▼
