Discovering dysfunction of multiple microRNAs cooperation in disease by a conserved microRNA co-expression network.

Abstract

MicroRNAs, a new class of key regulators of gene expression, have been shown to be involved in diverse biological processes and linked to many human diseases. To elucidate miRNA function from a global perspective, we constructed a conserved miRNA co-expression network by integrating multiple human and mouse miRNA expression data. We found that these conserved co-expressed miRNA pairs tend to reside in close genomic proximity, belong to common families, share common transcription factors, and regulate common biological processes by targeting common components of those processes based on miRNA targets and miRNA knockout/transfection expression data, suggesting their strong functional associations. We also identified several co-expressed miRNA sub-networks. Our analysis reveals that many miRNAs in the same sub-network are associated with the same diseases. By mapping known disease miRNAs to the network, we identified three cancer-related miRNA sub-networks. Functional analyses based on targets and miRNA knockout/transfection data consistently show that these sub-networks are significantly involved in cancer-related biological processes, such as apoptosis and cell cycle. Our results imply that multiple co-expressed miRNAs can cooperatively regulate a given biological process by targeting common components of that process, and the pathogenesis of disease may be associated with the abnormality of multiple functionally cooperative miRNAs rather than individual miRNAs. In addition, many of these co-expression relationships provide strong evidence for the involvement of new miRNAs in important biological processes, such as apoptosis, differentiation and cell cycle, indicating their potential disease links.

Full text

Page 1

Discovering Dysfunction of Multiple MicroRNAs
Cooperation in Disease by a Conserved MicroRNA
Co-Expression Network
Yun Xiao., Chaohan Xu., Jinxia Guan, Yanyan Ping, Huihui Fan, Yiqun Li, Hongying Zhao, Xia Li*
College of Bioinformatics Science and Technology, Harbin Medical University, Harbin, Heilongjiang, China
Abstract
MicroRNAs, a new class of key regulators of gene expression, have been shown to be involved in diverse biological
processes and linked to many human diseases. To elucidate miRNA function from a global perspective, we constructed a
conserved miRNA co-expression network by integrating multiple human and mouse miRNA expression data. We found that
these conserved co-expressed miRNA pairs tend to reside in close genomic proximity, belong to common families, share
common transcription factors, and regulate common biological processes by targeting common components of those
processes based on miRNA targets and miRNA knockout/transfection expression data, suggesting their strong functional
associations. We also identified several co-expressed miRNA sub-networks. Our analysis reveals that many miRNAs in the
same sub-network are associated with the same diseases. By mapping known disease miRNAs to the network, we identified
three cancer-related miRNA sub-networks. Functional analyses based on targets and miRNA knockout/transfection data
consistently show that these sub-networks are significantly involved in cancer-related biological processes, such as
apoptosis and cell cycle. Our results imply that multiple co-expressed miRNAs can cooperatively regulate a given biological
process by targeting common components of that process, and the pathogenesis of disease may be associated with the
abnormality of multiple functionally cooperative miRNAs rather than individual miRNAs. In addition, many of these co-
expression relationships provide strong evidence for the involvement of new miRNAs in important biological processes,
such as apoptosis, differentiation and cell cycle, indicating their potential disease links.
Citation: Xiao Y, Xu C, Guan J, Ping Y, Fan H, et al. (2012) Discovering Dysfunction of Multiple MicroRNAs Cooperation in Disease by a Conserved MicroRNA
Co-Expression Network. PLoS ONE 7(2): e32201. doi:10.1371/journal.pone.0032201
Editor: Cynthia Gibas, University of North Carolina at Charlotte, United States of America
Received May 31, 2011; Accepted January 24, 2012; Published February 22, 2012
Copyright: ? 2012 Xiao et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted
use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 30871394, 30370798 and 30571034), the National
High Tech Development Project of China, the 863 Program (Grant Nos. 2007AA02Z329), the National Basic Research Program of China, the 973 Program (Grant
Nos. 2008CB517302) and the National Science Foundation of Heilongjiang Province (Grant Nos. ZJG0501, 1055HG009, GB03C602-4, HCXB2009019, HCXS2010008,
BMFH060044 and 1155H012). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: lixia@hrbmu.edu.cn
. These authors contributed equally to this work.
Introduction
MicroRNAs (miRNAs) are a class of small non-coding RNA
molecules that ‘fine-tune’ gene expression on the posttranscrip-
tional level. MiRNAs can negatively regulate their target genes by
imperfect base pairing to the 39-untranslated region (UTR) of their
targets, which induce translational inhibition or deadenylation and
mRNA decay. A large number of studies have demonstrated that
miRNAs play important roles in a wide range of biological
processes, such as development, differentiation and apoptosis.
Furthermore, emerging evidence also indicates that miRNAs are
involved in the pathogenesis of many human diseases, such as
cancer and cardiovascular disease. Especially in cancer, miRNAs
can function as oncogenes or tumor suppressors. Nonetheless,
understanding of miRNA function and their roles in disease is still
in its infancy.
A systematic genetic mutation study discovered that the
majority of miRNA gene mutations in Caenorhabditis elegans do
not result in obviously abnormal phenotypes [1]. A recent study
further revealed that only few abnormal phenotypes are observed
in Caenorhabditis elegans strains that each lack of multiple or all
miRNA family members [2]. These results show that miRNAs
may function together with other miRNAs. Many recent studies
also found that some miRNAs can cooperatively control a variety
of biological processes, such as cell development [3] and
differentiation [4,5], apoptosis [6], cell cycle [7,8], and epithelial
cell polarity [9]. Moreover, the multiplicity of miRNA targets can
confer miRNAs the ability to cooperatively regulate a single
biological process by targeting common components of that
process. Using predicted targets, several bioinformatics studies have
discovered many miRNA-mRNA modules [10,11,12,13,14,15].
Our recent work also demonstrated potential functional relation-
ships between miRNAs based on common targets [16]. Thus, it is
reasonable to assume that miRNAs can function in a cooperative
manner, rather than in a separate way. Exploring functional
relationships between miRNAs may provide important clues about
their function and how miRNAs contribute to human disease.
Over the last decade, microarrays have emerged as a powerful
tool for comprehensively analyzing the expression levels for
thousands of genes, and many studies utilized gene expression
profiles to learn about gene functions [17,18,19,20]. Like genes,
miRNA microarrays have been widely used for exploring the roles
PLoS ONE | www.plosone.org1February 2012 | Volume 7 | Issue 2 | e32201

Page 2

of different miRNAs in various pathophysiological states. Many
miRNA microarray studies have demonstrated that miRNAs can
be used for disease diagnosis, prognosis and treatment [21,22].
These large number of available miRNA expression profiles have
been used to predict miRNA targets and analyze functional
relationships between miRNAs. For example, Ritchie et al. [23]
combined expression data from human and mouse to predict
putative miRNA targets. A recent study completed by Volinia et
al. [24] constructed miRNA networks in normal tissues and cancer
using miRNA expression, and identified important miRNA cliques
in cancer.
In this study, we performed a large-scale bioinformatics analysis
of conserved miRNA co-expression relationships to systematically
investigate functional links between miRNAs. By integrating
human and mouse miRNA expression data, a conserved miRNA
co-expression network was built. We confirmed that these
conserved co-expressed miRNA pairs in the network are more
likely to be functionally relevant. By mapping known disease
miRNAs to the network, we identified three miRNA sub-networks
that are highly related to cancer, and further explored their
functions based on predicted targets and miRNA knockout/
transfection expression data. Our results suggest that the
pathogenesis of human disease may be associated with the
impairment of functional cooperation between miRNAs.
Results
Construction of a conserved miRNA co-expression
network
We collected 16 human and 8 mouse miRNA expression data
sets respectively including 611 and 107 samples (Figure 1A). All
expression data sets were generated using Agilent arrays. After
normalization and probes mapping, 702 and 490 mature miRNAs
were consistently present in human and mouse miRNA expression
data sets, respectively. To identify miRNAs that are co-expressed
across human and mouse, we identified 285 human-mouse
orthologous miRNAs by all-against-all alignment of precursor
miRNA (pre-miRNA) sequences with 11 bp flanking regions.
Because all expression data sets used in this study are specific for
mature miRNAs, we then linked mature miRNAs in human with
their corresponding mature miRNAs in mouse according to these
285 orthologous miRNAs. Finally, 341 human-mouse orthologous
mature miRNAs were identified. Of these, 253 with both members
having expression measurements were used in the following
analysis (Table S1).
Using the method described in [20], we identified 182
significant conserved co-expression relationships (corrected p-
value,0.05). Based on these relationships, a conserved miRNA
co-expression network with 163 nodes and 182 edges was
constructed. In order to verify the significance of these conserved
co-expressed relationships, we randomly permuted the ortholo-
gous mature miRNAs (i.e. random selections of miRNAs from
human and mouse as putative human-mouse orthologous mature
miRNAs), and constructed a conserved miRNA co-expression
network using the permuted orthologous mature miRNAs. By
repeating the permutation procedure 1000 times, we found
significantly more co-expression relationships in the real conserved
co-expression network than in the random networks (p-value,0.001,
Figure 1B).
Because samples in the human and mouse miRNA expression
data sets are derived from different tissues with different
phenotypes, these conserved co-expression relationships may be
influenced by sample heterogeneity. To determine whether these
co-expression relationships are robust to the choice of samples, we
randomly selected 80% of samples from human and mouse
expression data sets for identifying conserved co-expression
relationships. We repeated the procedure 1000 times. An average
of 184 relationships was identified. Furthermore, we found that an
average of 89.1% of relationships is also included in the network
constructed using all samples. These results suggest that the
majority of co-expression relationships are widespread in a variety
of tissues and disease states, representing a general set of co-
expression relationships.
To further evaluate the conserved co-expression relationships,
we sought to use the receiving operator characteristic (ROC)
curve, which provide a way of measuring sensitivity and specificity,
to quantify the significant of these relationships. The area under
the ROC curve (AUC) was used as a measure for the overall
accuracy. Due to the absence of validated co-expression
relationships, we constructed three sets of co-expression relation-
ships according to different Pearson correlation thresholds using a
PCR-based miRNA expression data (GSE23024). Our results
illustrated in Figure 1C show that the conserved co-expression
relationships achieve high AUC scores (.85.0%), which are
significantly higher than random (p-value,0.001).
Functional relationships between conserved co-
expressed miRNA pairs
In order to further understand these conserved co-expression
relationships, we retrieved the human network from the human-
mouse conserved miRNA co-expression network by simply
extracting human miRNAs in each node. We sought to analyze
some characteristics of conserved co-expressed miRNA pairs in
the human network. To determine the significance of these
characteristics, we generated 1000 random sets with the same
number of co-expression relationships as in the real network. For
each randomization, we generated 182 miRNA pairs by randomly
choosing two miRNAs and linking them together.
Firstly, we examined genomic distances between conserved co-
expressed miRNAs. Genomic distances of 64 pairs of miRNAs
located on the same chromosome were computed. We found 32
pairs of miRNAs (50%) within 1 Kb of each other, 58 (90.6%)
within 50 Kb (Figure 2A). These conserved co-expressed miRNA
pairs are significantly closer to each other than non co-expressed
miRNA pairs (Wilcoxon rank sum test, p-value,2.2e-16). Similarly,
50 pairs of miRNAs belonging to the same miRNA cluster were
observed. We found that the real conserved co-expressed miRNA
pairs have significantly more pairs of miRNAs within the same
cluster than randomly selected miRNA pairs (p-value,0.001,
Figure 2B). These results, consistent with those reported earlier
by Baskerville et al. [25], suggest that these co-expressed miRNAs
within a short distance may derive from common transcriptions.
Regulation of miRNA gene transcription is similar to that of
protein-coding gene transcription, which is controlled by many
TFs. In order to investigate whether these co-expressed miRNA
pairs are regulated by common TFs, we obtained 260 experi-
mentally validated TF-miRNA regulation relationships from the
TransmiR database [26]. We observed that 47 pairs of miRNAs
are co-regulated by at least one common TF, significantly more
than randomly selected miRNA pairs (p-value,0.001, Figure 2C).
This reflects that the co-expression of miRNAs may be resulted
from common regulation, indicating their potential functional
relevance.
Emerging evidence shows that members of a miRNA family
(such as the let-7 family) have similar functions since an abundance
of overlapping targets resulted from their common seed sequences.
We observed 44 pairs of miRNAs belonging to the same family,
Dysfunction of MicroRNAs Cooperation in Disease
PLoS ONE | www.plosone.org2 February 2012 | Volume 7 | Issue 2 | e32201

Page 3

significantly more than randomly selected pairs of miRNAs (p-
value,0.001, Figure 2D).
To validate whether co-expressed miRNAs have functional
associations, we obtained their predicted targets from the
TargetScan5.1 database [27]. Considering a large number of
targets predicted by TargetScan5.1, we used a ‘context score’
(20.3), which corresponds to a certain magnitude of regression, as
a cutoff to screen targets regardless of conservation status of
binding sites. Among 182 pairs of miRNAs, 149 pairs in which
both members have targets were analyzed. We found that 132
pairs of miRNAs have significantly more overlapping targets than
expected by chance (hypergeometric distribution, FDR-corrected
p-value,0.05). When compared with randomly selected miRNA
pairs, more pairs of miRNAs having significant overlapping targets
were identified in these real conserved co-expressed miRNA pairs
(p-value=0.001, Figure 2E). Furthermore, functional enrichment
analyses identified 88 miRNA pairs with common targets
significantly involved in at least one biological process, significantly
more than randomly selected pairs of miRNAs (p-value=0.022,
Figure 2F).
Subsequently, we obtained an additional paired miRNA and
mRNA expression dataset (GSE25692), and then calculated
Pearson correlation coefficients (PCCs) between miRNAs and
genes. For each miRNA, genes with absolute PCC.0.5 were
regarded as its expression-related genes. Using the hypergeometric
distribution, we identified 81 miRNA pairs with significantly
overlapping expression-related genes, significantly more than
randomly selected pairs of miRNAs (p-value=0.034, Figure 2G).
Figure 1. Evaluation of the conserved co-expression relationships. (A) Pie charts of miRNA expression data from human (top) and mouse
(bottom) included in the analysis. Colors represent different tissues. (B) Probability density of the number of co-expression links identified through the
permutation of orthologous miRNAs. The permutation experiment was repeated 100 times. (C) ROC curves used to quantify the significant of these
relationships using a PCR-based miRNA expression data (GSE23024).
doi:10.1371/journal.pone.0032201.g001
Dysfunction of MicroRNAs Cooperation in Disease
PLoS ONE | www.plosone.org3February 2012 | Volume 7 | Issue 2 | e32201

Page 4

Dysfunction of MicroRNAs Cooperation in Disease
PLoS ONE | www.plosone.org4 February 2012 | Volume 7 | Issue 2 | e32201

Page 5

Moreover, we sought to use miRNA-affected genes generated by
knockout/transfection of miRNAs to further determine function
associations of these conserved co-expressed miRNA pairs. Multiple
gene expression data sets referring to knockout/transfection of
different miRNAs were obtained from Gene Expression Omnibus
(GEO). Sixteen miRNA pairs in which both members have
knockout/transfection data were used. Genes with FC.1.2 were
regarded as miRNA-affected genes. Using the hypergeometric
distribution, we found that these pairs have significantly more
overlapping affected genes except one (Table S2).
Together, our findings suggest strong functional relationships
between conserved co-expressed miRNAs, which may be derived
from common clusters and/or common families, and can be
regulated by common TFs. More importantly, miRNAs can
cooperatively regulate a single biological process by targeting
common components of that process, which may be the crucial
mechanism underlying the ‘fine-tuning’ of gene expression.
Dysfunction of miRNA cooperation in disease
Currently, many miRNAs have been identified as disease
miRNAs, which play important roles in the development and
progression of various diseases. Is the functional cooperation
between miRNAs associated with disease? We obtained disease
miRNAs from the miR2Disease database [28], and then 204
miRNAs with ‘causal’ relations were gained. Among them, 73
disease miRNAs can be mapped to the network. We identified 54
miRNA pairs sharing a common disease. We verified the
significance by means of permuting co-expressed miRNA pairs
and disease miRNAs separately. First, we generated 1000 random
sets, each of which is composed of 182 randomly selected miRNA
pairs. We found that the real conserved co-expressed miRNA pairs
have significantly more pairs sharing a common disease than
randomly selected miRNA pairs (p-value,0.001, Figure 3A). We
next evaluated the significance by constructing 1000 sets of
randomly selected disease miRNAs. We compared the observed
number with the distribution of the number of miRNA pairs
sharing a common disease seen in the 1000 randomly generated
disease miRNA sets. The similar result was observed (p-
value,0.001, Figure 3B). These results indicate that dysfunction
of function-related miRNAs may be associated with the patho-
genesis of disease.
Subsequently, we identified several highly connected sub-
networks using random walks [29]. We observed that some sub-
networks are obviously enriched by disease miRNAs (Figure 4).
The three sub-networks containing the most disease miRNAs were
identified. The sub-network I, II and III contain 16, 14 and 8
miRNAs, respectively. To our surprise, we found that multiple
miRNAs in each sub-network are significantly involved in the
same diseases (Figure 3C). For example, 14 miRNAs in the sub-
network I are disease miRNAs. Of these, 9 miRNAs have been
identified as important factors contributing to the development of
hepatocellular carcinoma. Furthermore, the remaining 5 disease
miRNAs in this sub-network are also associated with various other
cancers. In the sub-network III, 6 of 8 miRNAs are breast cancer-
related miRNAs. These findings indicate that carcinogenesis may
be associated with dysfunction of multiple function-related
miRNAs in these sub-networks rather than single miRNAs.
To determine whether these miRNA sub-networks are involved
in cancer-related biological processes, we obtained their predicted
targets using TargetScan5.1 (without the ‘context score’ limita-
tion). In the sub-network I and II, 64 and 90 common targets
shared by at least 13 and 12 miRNAs were identified, respectively.
In the sub-network III, 54 common targets regulated by all 8
miRNAs were identified. Through functional enrichment analyses
of these sets of common targets, we found that the sub-network I,
II and III is significantly involved in ‘induction of apoptosis by
intracellular signals’, ‘regulation of Ras protein signal transduc-
tion’ and ‘regulation of cell differentiation’ respectively, suggesting
that miRNAs in each sub-network can cooperatively regulate
important cancer-related biological processes by targeting com-
mon transcripts.
Especially, many miRNAs in the sub-networks are from the
same family. This might make the prediction of their functions
based on their targets biased, because miRNAs in the same family
have similar targets. Therefore, we re-predicted the functions of
these sub-networks by eliminating the affection of miRNAs from
the same family. In each sub-network, common targets of miRNAs
from the same family were obtained, and then these family-based
target sets together with target sets of other miRNAs not belonging
to the same family were used to determine the functions of the sub-
network. The sub-network I contains a total of 16 miRNAs, with
15 having predicted targets. In the sub-network I, 12 miRNAs are
from 5 miRNA families. For each miRNA family, their common
targets were determined. Then, the 5 family-based target sets and
3 target sets from the remaining 3 miRNAs were obtained. We
identified 158 common targets shared by at least 6 of these 8 target
sets. By function enrichment, we found ‘induction of apoptosis by
intracellular signals’ significantly overrepresented in the common
targets. Similarly, we identified ‘regulation of Ras protein signal
transduction’ and ‘regulation of cell differentiation’ significantly
overrepresented in sub-network II and III, respectively. These
results are obviously consistent with the above results determined
using all predicted targets.
In addition, we determined the functions of these sub-networks
based on a paired miRNA and mRNA expression data set
(GSE25692). For each miRNA, genes with absolute PCC.0.5
were identified as its expression-related genes. In the sub-network
I, 13 miRNAs have expression values in the miRNA expression
dataset. 64 common expression-related genes shared by at least 11
miRNAs were identified. We found that these common genes are
Figure 2. Functional relationships of 182 conserved co-expressed miRNA pairs. (A) Genomic distances of the observed miRNA pairs, which
are significantly shorter than the distances of non co-expressed miRNA pairs (Wilcoxon rank sum test, p-value,2.2e-16). (B) Probability density of the
number of miRNA pairs that belong to the same cluster from randomly selected miRNA pairs. The count observed in the real co-expressed miRNA
pairs (50, located by the blue arrow) is significantly higher than those in the random pairs (p-value,0.001). (C) Probability density of the number of
miRNA pairs that share common TFs from randomly selected miRNA pairs. The count observed in the real co-expressed miRNA pairs (47, located by
the blue arrow) is significantly higher than those in the random pairs (p-value,0.001). (D) Probability density of the number of miRNA pairs belonging
to the same family from randomly selected miRNA pairs. The count observed in the real co-expressed miRNA pairs (44, located by the blue arrow) is
significantly higher than those in the random pairs (p-value,0.001). (E) The number of miRNA pairs with significantly overlapping targets in the real
conserved co-expression pairs (132, located by the blue arrow) is significantly higher than those in the randomly selected miRNA pairs (p-
value=0.001). (F) The number of miRNA pairs with common targets significantly involved in at least one biological process in the real conserved co-
expression pairs (88, located by the blue arrow) is significantly higher than those in the randomly selected miRNA pairs (p-value=0.022). (G) The
number of miRNA pairs with significantly overlapping expression-related genes in the real conserved co-expression pairs (81, located by the blue
arrow) is significantly higher than those in the randomly selected miRNA pairs (p-value=0.034).
doi:10.1371/journal.pone.0032201.g002
Dysfunction of MicroRNAs Cooperation in Disease
PLoS ONE | www.plosone.org5 February 2012 | Volume 7 | Issue 2 | e32201