Recently, more and more research has shown that microRNAs (miRNAs) play critical roles in
the development and progression of various diseases, but it is not easy to predict potential human
miRNA–disease associations from the vast amount of biological data. Computational methods
for predicting potential disease–miRNA associations have gained a lot of attention based on their
feasibility, guidance and effectiveness. Differing from traditional local network similarity
measures, we adopted global network similarity measures and developed Random Walk with
Restart for MiRNA–Disease Association (RWRMDA) to infer potential miRNA–disease interactions
by implementing random walk on the miRNA–miRNA functional similarity network. We tested
RWRMDA on 1616 known miRNA–disease associations based on leave-one-out cross-validation,
and achieved an area under the ROC curve of 86.17%, which significantly improves previous
methods. The method was also applied to three cancers for accuracy evaluation. As a result, 98%
(Breast cancer), 74% (Colon cancer), and 88% (Lung cancer) of top 50 predicted miRNAs are
confirmed by published experiments. These results suggest that RWRMDA will represent an
important bioinformatics resource in biomedical research of both miRNAs and diseases.
Introduction
MicroRNAs (miRNAs) are a class of small (B22 nt) non-coding
regulatory RNAs, normally suppressing the expression of the
target mRNA at post-transcriptional level by binding to the-UTRs of the target mRNA through sequence-specific
base pairing.
However, some reports have pointed out that
miRNAs may also function as positive regulators in some
cases.
Caenorhabditis elegans (C. elegans)lin-4andlet-7,the
first two known miRNAs, were identified by conventional
forward genetic screens.
Over the past few years, thousands
of miRNAs have been discovered in eukaryotic organisms ranging
from nematodes to humans.ThenewestversionofmiRBasehas
contained 16772 entries and more than 1000 miRNAs have been
discovered in human (miRBase, Release 17).
Plenty of studies reveal miRNA to be one of the most
important components in the cell and it plays critical roles in
diverse fundamental biological processes, such as cell develop-
ment, proliferation, differentiation, apoptosis, signal transduction,
viral infection, and so on.
Therefore, the miRNA-related
dysfunction is associated with various diseases.
An exciting
example is mir-375, which can regulate insulin secretion.
Additionally, dysregulation of numerous miRNAs is associated
Nowadays, miRNAs have taken centre stage in the field of
human molecular oncology.
28
Therefore, a large scale search
for the relationship between miRNAs and diseases has become an
important goal of biomedical research,
29
which will accelerate the
understanding of disease pathogenesis at the molecular level, but
more importantly will benefit the prognosis, diagnosis, evaluation,
treatment and prevention of disease and promote the improve-
ment of human medicine.
18,21,30–33
However, current knowledge
about the relationship between miRNAs and diseases is relatively
limited. Experimental identification of disease-related miRNAs by
existing techniques is expensive and time-consuming.Fortunately
vast amount of biological data about miRNAs has been generated,
so there is a strong motivation to develop powerful computational
methods that can effectively uncover potential disease–miRNA
associations in a large scale. Computational methods can select
most promising miRNAs for further analysis and hence decrease
the number of the experiments, benefit the understanding of
miRNAs function. However, the difficulty of prediction task lies
in the rarity of known disease–miRNA interactions.
Some important conclusions and computational methods
aboutdisease-related miRNAs prediction havebeenproposed.
Lu et al.
analyzed the human microRNA–disease association
data and proposed many important patterns between miRNAs
National Center for Mathematics and Interdisciplinary Sciences,
Chinese Academy of Sciences, Beijing 100190, China
Academy of Mathematics and Systems Science, Chinese Academy of
Sciences, Beijing 100190, China. E-mail: ,
Graduate University of Chinese Academy of Sciences,
Beijing 100190, China
w Electronic supplementary information (ESI) available. See DOI:
10.1039/c2mb25180a
z Author contributions: XC conceived and developed the prediction
method, conceived, designed and implemented the experiments,
analyzed the result, and wrote the paper. MXL analyzed the result.
GYY conceived the prediction method, analyzed the result, wrote the
paper, and provided guidance and supervision. All authors read and
approved the final manuscript.
Molecular
BioSystems
Dynamic Article Links
www.rsc.org/molecularbiosystems PAPER
This journal is
c
The Royal Society of Chemistry 2012 Mol. BioSyst., 2012, 8, 2792–2798 2793
and human diseases, which laid a solid foundation for current
disease-related miRNA research and provided powerful support
to the research about the diseases at the miRNA level. Based on
the assumption that phenotypical similar diseases tend to be
associated with functional related miRNAs proposed by Lu
et al.,
3
Zhang et al.
34
developed the first miRNA–disease associa-
tion prediction method, which identified potential cardiovascular
disease relatedmiRNAs by miRNAs set, family analysis andGene
Ontology. However, the fact that this method strongly depend on
miRNA sets has limited its application. Jiang et al.
21
developed a
computational method based on the hypergeometric distribution
to infer disease-related miRNAs by integrating miRNAs
functional interactions network, disease similarity network
and known phenome–microRNAome network consisting of
270 experimentally verified disease–miRNAs associations
obtained from miR2Disease.
35
Although miRNA functional
network has been constructed in this paper, only the neighbor
information of each miRNA was used in the scoring system.
Making full use of the global network similarity information
would improve the accuracy of the algorithm. Another limitation
is that this method highly depends on the predicted miRNA
target. It is known that the current in silico prediction tool for
miRNA target prediction has a high rate of false-positive and
high false-negative results. As a result, the prediction accuracy of
this method is not high. Jiang et al.
22
further proposed an
approach for prioritizing candidate miRNAs based on genomic
data integration by Naı¨ve Bayes model, which strongly relies on
datasets of disease–gene associations and miRNA–target inter-
actions. The molecular bases for about 60% of human diseases
are still unknown.
36
Also, the problem of high false-positive and
high false-negative in the miRNA–target interactions predicted
by current different algorithms still exists in this method. Jiang
et al.
29
proposed an approach for distinguishing positive disease
miRNAs from negative disease miRNAs based on Support
Vector Machine by extracting the features based on micro-
RNA–target data and phenotype similarity data. Under the
assumption that miRNAs implicated in a specific disease will
show aberrant regulations of their target mRNAs, Xu et al.
28
introduced a network-centric method to prioritize candidate
disease miRNAs by constructing four topological features to
distinguish prostate cancer (PC) miRNAs and non-PC miRNAs.
The common problem of the aforementioned two methods is the
selection of negative samples as there are no verified negative
miRNA–disease associations. Compilinga list ofnegative disease
miRNAs is currently difficult or even impossible.
28
Besides the
above methods, which arre based on similar train of thought,
Rossi et al. proposed a method called OMiR to identify potential
relationships between miRNAs and OMIM diseases, which
achieved their aim through calculating the significance of
the overlap between miRNA loci and OMIM disease loci,
without utilizing known miRNA–disease associations and
other information, such as miRNA target, disease pathogeny
information and so on.
37
Taken together, the above mentioned methods for miRNA–
disease association prediction have various limitations. There-
fore, novel methods are urgently needed. In this paper, we
have investigated the hypothesis that global network similarity
measures are better suited to capture the associations between
diseases and miRNAs than traditional local network similarity
measures, such as neighbor information. Based on the global
network similarity measure and the assumption that functionally
related miRNAs tend to be associated with phenotypical similar
diseases,
21
the method of Random Walk with Restart for
MiRNA–Disease Association (RWRMDA) has been developed
to infer potential miRNA–disease associations by implementing
random walk on the miRNA functional similarity network to
prioritize candidate miRNAs for disease of interest. Random
walk has been widely applied in bioinformatics, especially for
disease gene identification and drug target interactions predic-
tion.
38–41
Cross validation and case studies about three kinds of
cancers have illustrated RWRMDA is superior to previous
predictive method based on local network similarity measure.
Methods
The human miRNA–disease association data
Considering many studies have produced a large number of
miRNA–disease associations, Lu et al.
3
collected the miRNA–
disease associations and constructed a human miRNA–associated
disease database (HMDD), which contains 3760 miRNA–disease
associations and the information of 493 miRNAs and 295 diseases
from 1688 publications. Jiang et al.
35
also constructed a manually
curated miRNA–disease relationships database (miR2Disease),
which aims to provide a comprehensive resource of experimentally
confirmed miRNA–disease associations. After recent updates,
3273 miRNA–disease associations about 349 miRNAs and
163 diseases have been collected in the database. Yang et al.
42
constructed a publicly available database of Differentially
Expressed MiRNAs in human Cancers (dbDEMC) with
the aim to provide potential cancer-related miRNAs by
in silco computing. The current version of dbDEMC includes
607 differentially expressed miRNAs (590 mature miRNAs
and 17 precursor miRNAs) in 14 cancers from 48 microarray
experiments in peer-reviewed publications.
The human miRNA–disease association data used for
prediction accuracy evaluation was downloaded from the
supplementary data of Ref. 43. This dataset consists of
1616 distinct high-quality experimentally verified human
miRNA–disease associations, which were obtained from
HMDD in September, 2009. The operation of merging the
records of different miRNA copies that produce the same
mature miRNA into one group, unifying the name of different
mature miRNAs as one miRNA gene, and curating the disease
name based on standard MeSH disease terms were implemented.
These associations were used as the benchmark dataset for the
performance evaluation of our model in the cross validation
schema and the seed dataset for predicting potential human
miRNA–disease associations (see Table S1, ESIw). We did not
use the latest version of the HMDD data because potential
human miRNA–disease associations predicted by our
model can be evaluated by new associations introduced to
HMDD after September, 2009. MiRNA–Disease Association
Network (MDAN) was constructed, where vertices set M =