ARTICLE
Received 9 Aug 2015
| Accepted 3 Mar 2016 | Published 6 Apr 2016
Circular RNA profiling reveals an abundant
circHIPK3 that regulates cell growth by sponging
multiple miRNAs
Qiupeng Zheng
1,2,
*, Chunyang Bao
1,
*, Weijie Guo
1,2
, Shuyi Li
1,2
, Jie Chen
1
, Bing Chen
1
, Yanting Luo
1
,
Dongbin Lyu
1
, Yan Li
1
, Guohai Shi
1
, Linhui Liang
1
, Jianren Gu
2
, Xianghuo He
1
& Shenglin Huang
1
Circular RNAs (circRNAs) represent a class of widespread and diverse endogenous RNAs
that may regulate gene expression in eukaryotes. However, the regulation and function of
human circRNAs remain largely unknown. Here we generate ribosomal-depleted RNA
sequencing data from six normal tissues and seven cancers, and detect at least 27,000
circRNA candidates. Many of these circRNAs are differently expressed between the normal
and cancerous tissues. We further characterize one abundant circRNA derived from Exon2 of
the HIPK3 gene, termed circHIPK3. The silencing of circHIPK3 but not HIPK3 mRNA sig-
nificantly inhibits human cell growth. Via a luciferase screening assay, circHIPK3 is observed
to sponge to 9 miRNAs with 18 potential binding sites. Specifically, we show that circHIPK3
directly binds to miR-124 and inhibits miR-124 activity. Our results provide
evidence that circular RNA produced from precursor mRNA may have a regulatory role in
human cells.
DOI: 10.1038/ncomms11215
OPEN
1
Fudan University Shanghai Cancer Center and Institutes of Biomedical Sciences, Shanghai Medical College, Fudan University; Department of Oncology,
Shanghai Medical College, Fudan University, Shanghai 200032, China.
2
State Key Laboratory of Oncogenes and Related Genes, Shanghai Cancer Institute,
Renji Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai 200032, China. * These authors contributed equally to this work. Correspondence
and requests for materials should be addressed to X.H. (email: xhhe@fudan.edu.cn) or to S.H. (email: slhuang@fudan.edu.cn).
NATURE COMMUNICATIONS | 7:11215 | DOI: 10.1038/ncomms11215 | www.nature.com/naturecommunications 1
C
ircular RNAs from back-spliced exons (circRNAs) are
recently identified as a naturally occurring family of
noncoding RNAs that is highly represented in the
eukaryotic transcriptome
1,2
. The formation of circRNAs had
occasionally been identified more than 20 years ago from a few
transcribed genes
3–5
. Nevertheless, these species had generally
been considered to be of low abundance and likely representing
errors in splicing. Thus, their widespread and substantial presence
within transcriptomes has only recently been demonstrated via
the high-throughput sequencing and novel computational
approaches
1,6–8
. Specifically, a large number of circRNAs have
been successfully identified in various cell lines and across
different species
9–13
.
circRNAs are characterized by covalently closed loop structures
with neither 5
0
to 3
0
polarity nor a polyadenylated tail. They are
highly stable in vivo compared with their linear counterparts,
and are predominantly in the cytoplasm and can be sorted
into exosomes
14
. Two mechanisms, ‘exon skipping’ and ‘direct
back-splicing’, have been proposed to form mammalian exonic
circRNA
1,6,7
. Exon skipping leads to a lariat whose restricted
structure promotes circularization, whereas direct back-splicing
refers to the pairing of a downstream splice donor with an
unspliced upstream splice acceptor, which results in the
circularization of the intervening RNA. Both mechanisms
involve back-splicing being formed by the canonical
spliceosome
15
. Recent studies have shown that exon
circularization is facilitated by complementary sequences
16,17
and regulated by specific protein factors
18–20
. It is becoming
increasingly evident that circRNAs are not simply by-products of
mis-splicing or splicing errors, rather, they are the products of
regulated back-splicing with distinct sets of cis-elements and/or
trans-factors
21
. Accordingly, many circRNAs have been found to
be upregulated during mouse neural development and human
epithelial–mesenchymal transition
10,20
. Recently, circRNAs have
been shown to act as microRNA (miRNA) sponges to regulate
gene expression
8,22
. Specifically, the circRNA ciRS-7 (also
termed CDR1as), which harbours more than 70 conventional
miR-7-binding sites, has been identified as a miRNA inhibitor.
However, only a few such circRNAs contain multiple binding
sites to trap one particular miRNA
11
, and the function of
circRNA remains largely unknown.
In humans, circRNAs have been characterized in several cell
lines and brain tissue
9–12
. In this study, we generated ribominus
RNA sequencing data from six human normal tissues and seven
human cancers, and identified B27,000 circRNA candidates
(at least two unique back-spliced reads). Analysis of these
circRNAs revealed that there is often a predominately expressed
circRNA isoform from a given gene locus and a number
of circRNAs are highly abundant. We further characterize one
abundant circRNA produced from the HIPK3 gene, termed
circHIPK3. The formation of circHIPK3 is due to the long
intronic complementary repeat elements. Importantly, we found
that circHIPK3 RNA, but not HIPK3 mRNA, functions as a cell
growth modulator in human cells. We further performed a
luciferase screening and observed that circHIPK3 could bind to
multiple miRNAs, including a well-known tumor suppressor
miRNA miR-124. Our findings indicate that protein-coding
exons may exert additional regulatory functions when expressed
within circRNAs in human cells.
Results
Profiling of circRNAs in human normal and cancerous tissues.
First, we characterized circRNA transcripts using RNA-
sequencing (RNA-seq) analyses of ribosomal RNA-depleted total
RNA from six normal tissues (brain, colon, heart, liver, lung and
stomach) and seven cancerous tissues (bladder urothelial
carcinoma (BLCA), breast cancer, colorectal cancer (CRC),
hepacellular carcinoma (HCC), gastric cancer (GC), kidney clear
cell carcinoma (KCA) and prostate adenocarcinoma (PRAD)).
Each sample was sequenced on an Illumina HiSeq yielding B60
million reads, which were mapped to the human reference gen-
ome (GRCh37/hg19) by TopHat2 (ref. 23). A detailed summary
for each sample is provided in Supplementary Table 1. A com-
putational pipeline based on the anchor alignment of unmapped
reads was used to identify circRNAs without relying on gene
annotations
8
(Supplementary Fig. 1). In total, 67,358 distinct
circRNA candidates were found in these tissues and 27,296 of
these circRNAs contained at least two unique back-spliced reads
(Fig. 1a, Supplementary Data 1). Compared with previously
published databases obtained from circBase
9
(92,061 human
circRNAs) and a most recent study
10
(65,731 human circRNAs
identified mainly from human brain tissues), we found that there
are 19,071 overlapped circRNAs and 8,225 novel circRNAs
identified in our study (Supplementary Fig. 2 and Supplementary
Data 1). Notably, there are totally 148,701 unique human
circRNA candidates from all the studies, indicating that
circRNAs may contain one of the largest RNA families in
human transcription.
We annotated these circRNA candidates using the RefSeq
database
24
. More than 80% of the circRNAs consisted of protein-
coding exons, whereas smaller fractions aligned with introns, long
noncoding RNAs, unannotated regions of the genome and
antisense regions to known transcripts (Fig. 1b). The length of
most exonic circRNAs (n ¼ 20,553; only known splice lengths
without introns) was less than 1,500 nucleotide (nt), and the
median length was B500 nt (Fig. 1c, Supplementary Data 1). We
normalized the back-spliced reads (support for circRNA) by read
length and number of mapped reads (spliced reads per billion
mapping, denoted as SRPBM), which permits quantitative
comparisons between back splicing from different RNA-seq
data
7
. The expression analysis of these circRNA transcripts
revealed that numerous circRNAs seem to be specifically
expressed across various tissues (Fig. 1d, minimum value of
specificity score 0.5 and minimum level of mean þ 2 s.d., Methods
section). Analogously, the profile of circRNAs (SRPBM41) in
cancer often differs from that of normal tissue (Fig. 1e). The
differences of circRNAs expression were calculated using
Wilcoxon rank-sum test by comparing the cancerous sample
with matched normal sample for each tissue. We found that
circRNAs are significantly downregulated in BRCA (P ¼ 4.71e–
32), CRC (P ¼ 9.50e–35), GC (P ¼ 5.09e–10), HCC (P ¼ 1.86e–
32) and PRAD (P ¼ 1.52e–08) tumours, and upregulated in
BLCA (P ¼ 1.13e–09) and KCA (P ¼ 3.65e–79). However, the
parent genes are significantly upregulated in BLCA (P ¼ 8.15e–
49), BRCA (P ¼ 5.07e–204), GC (P ¼ 5.49e–206) and HCC
(P ¼ 1.47e–32) and downregulated in CRC (P ¼ 1.94e–4) and
KCA (P ¼ 4.51e–07), but unchanged in PRAD (P ¼ 0.4320).
Only BLCA showed the similar expression patterns, indicating
that circRNAs and/or parent genes are post-transcriptionally
regulated. Moreover, a number of circRNAs seem to be
specifically expressed in cancer and matched normal tissues
(Fig. 1f, at least four reads as the specificity cut-off, Methods
section).
To verify that the back-spliced events were indicative of true
circular, and not linear, trans-splicing products, we examined
the physical properties of these products. Outward-facing
primers were designed against 36 commonly expressed circRNA
transcript candidates (Supplementary Table 2). Each primer pair
amplified a single, distinct product of the expected size from
HEK-293T cDNA (Supplementary Fig. 3A). The enrichment of
all 36 back-spliced events was apparent following RNase R
treatment, whereas the abundance of linear RNAs decreased
ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11215
2 NATURE COMMUNICATIONS | 7:11215 | DOI: 10.1038/ncomms11215 | www.nature.com/naturecommunications
(Supplementary Fig. 3B). We also validated 20 novel circRNAs
(out of 26 circRNA candidates) in original tissue samples by
quantitative reverse transcription-PCR(qRT–PCR) with RNase R
treatment and RT–PCR with Sanger sequencing (Supplementary
Fig. 4).
The characteristics of circRNA abundance in human cells.
Analysis of the number of circRNAs from their host genes
revealed that one gene could produce multiple circRNAs (Fig. 2a,
20,530 circRNAs from 5,955 host genes), which is consistent with
previous report
16
. A striking example is the oncogene PTK2
that may generate 47 distinct circRNAs (at least two unique
back-spliced reads). We further investigated the abundance of the
circRNAs within one gene locus, and found that about 50% of the
host genes (1,835/3,687, single circRNA from one gene not
included) produced a significantly higher expressed circRNAs
(at least twofold higher than other circRNAs, Fig. 2b). This result
indicates that there is often a predominantly expressed circRNA
isoform from one gene locus. Anchor alignment also allows us to
quantify the abundance of each circRNA with respect to its
alternative linear isoform in the ribosomal RNA-depleted
RNA-seq data (Fig. 2c and Methods section). For each of these
circRNAs, we estimated the circular ratio (CR) of a circRNA at
the 5
0
end or 3
0
end according to the number of reads spanning
a
c
d
e
f
b
Length of exonic circRNA
Liver
Brain
Lung
Heart
Stomach
Colon
BLCA
BRCA
GC
CRC
KCA
PRAD
HCC
Normal specific
Cancer specific
Backspliced reads
67,358 distinct circRNAs
(27,296 > 1 backspliced reads)
Number of circRNA
Number of circRNA
Number of circRNA
50
0
100
150
200
200
2,000
1
10
100
1,000
10,000
100,000
Intergenic
Antisense
lincRNA
170
3′UTR
5′UTR
Intronic
CDS
exons
16,959
5,090
1,730
153
381
845
circRNA genomic loci (27,296)
1,471
497
−2
−1
0
1
2
400
800
1,200
1,600
2,000
0
0
700
1,400
2,100
2,800
3,500
0
100
200
300
400
500
•
•
•
•
•
•
•
BLCA
BRCA
GC
CRC
KCA
PRAD
HCC
Relative abundance
4
2
0
–2
–4
•
Figure 1 | Profiling of circular RNAs in human normal and cancerous tissues. (a) The number of circRNAs and back-spliced reads identified in six human
normal tissues and seven human cancerous tissues. (b) Genomic origin of human circRNAs. (c) The length distribution for exonic circRNAs (n ¼ 20,533,
only known spliced length was considered). (d) Clustered heatmap for tissue-specific circRNAs from six human normal tissues, with rows representing
circRNAs and columns representing tissues. The circRNAs were classified according to the Pearson correlation. The numerical data represented
log10-transformed mean SRPBM of two replicates. (e) Violin plot of relative abundance of circRNAs in seven cancer tissues compared with the paired
normal tissues. Data are expressed as the log
2
foldchange of SRPBM. The white dot represents the median. (f) Numbers of specific circRNAs identified in
seven cancerous and matched normal tissues. Cancer-specific circRNAs are shown in red. Normal specific circRNAs are shown in green.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11215 ARTICLE
NATURE COMMUNICATIONS | 7:11215 | DOI: 10.1038/ncomms11215 | www.nature.com/naturecommunications 3
the back-spliced junctions and the number of reads spanning the
linearly spliced junctions. The CR for these sites considerably
varied, and linearly spliced products were absent in some cases
(Fig. 2d and Supplementary Data 2). When using a stringency of
CR and SRPBM cut-off (5
0
CR40.2; 3
0
CR40.2; SRPBM41), we
observed 990 high-abundance circRNAs, many of which were
more highly expressed than their linear isoforms (Fig. 2d).
Analysing these circRNAs revealed that flanking introns were
markedly longer than other circRNAs (Fig. 2e, P ¼ 1.20e–21,
Mann–Whitney U-test).
We further noted that one of these circRNAs, derived from the
HIPK3 gene Exon2 (termed circHIPK3), was particularly
abundant and featured a high back-spliced ratio (Fig. 2d, blue
dot). The high abundance of circHIPK3 was also observed by
previous reports
7,17
. Besides circHIPK3, there are another four
circRNA isoforms identified in our study in HIPK3 gene locus
(We termed them as circHIPK3.2, circHIPK3.3, circHIPK3.4,
circHIPK3.5, respectively; Fig. 3a). circHIPK3 is the predominant
circRNA isoform as evidence from the high-supported
back-spliced unique reads (circHIPK3, 1880; circHIPK3.2, 30;
Number of circRNAs
produced from one gene
a
c
d
e
b
0.0
0.2
0.4
0.6
0.8
23456
Intron length (Log
10
scaled)
Density
SRPBM (Log
10
scaled)
0.2
0.4
0.6
0.8
1.0
2.0
1.5
1.0
0.5
1.0
0.8
0.6
0.4
0.2
5′ CR
3′ CR
circHIPK3
Control
Enriched circRNAs
0
C1
C2
L1
L3 L4
L2
L5 L6
C2
L4
L5
C1
L4
5′ 3′
c
1
c
1
c
1
+ c
2
+ l
2
+ l
3
c
1
+ l
5
5′ CR =
5′ CR
3′ CR
3′ CR =
circGene
circGene.2
circGene.3
circGene.4
circGene.5
Relative abundance
0
2
4
6
1
2&3
4&5
6&7
8&9
10&11
12&13
14&15
16+
Figure 2 | The characteristics of circular RNA abundance in human cells. (a) Number of circRNAs produced from one gene (20,530 circRNAs from 5,955
host genes). (b) The box plots describe the comparison of the levels of most abundantly expressed circRNA isoform (circGene) and other circRNAs
(circGene.x) from one gene locus (n ¼ 3,687). The ends of the boxes define the 25th and 75th centiles, a line indicates the median, and bars define the 5th
and 95th centiles. The first five circRNAs were presented. (c) Schematic illustration of the methodology to estimate either the circular ratio at the 5
0
end
(5
0
CR) or at 3
0
end (3
0
CR) for circRNAs. C
i
and c
i
represent the back-spliced junctions (support for circRNA) and number of reads spanning these
junctions, respectively; L
i
and l
i
represented linear spliced junctions and the number of reads spanning these junctions, respectively. Solid squares, exons;
broken lines, linear spliced junctions; arc lines, back-spliced junctions. (d) Multidimensional scaling screen for highly abundant circRNAs (n ¼ 27,293,
three outlines were not shown). Red dots and black dots represented highly abundant circRNAs and low-abundance circRNAs, respectively. circHIPK3 is
highlighted in blue. The cut-off for highly abundant circRNAs is shown in grey. The pseudocount of 1 was added to SRPBM to avoid log
10
-transform issues.
Thus, the cut-off of log10-transformed SRPBM is 0.3010. (e) A density plot of the flanking intron length of highly abundant circRNAs (red) and
low-abundance circRNAs (green).
ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11215
4 NATURE COMMUNICATIONS | 7:11215 | DOI: 10.1038/ncomms11215 | www.nature.com/naturecommunications
circHIPK3.3, 7; circHIPK3.4, 6; circHIPK3.5, 2). The genomic
structure shows that circHIPK3 contains a large second exon
(1,099 bp) from the HIPK3 gene flanked by long introns on either
side (Fig. 3a). The distinct product of the expected size was
amplified using outward-facing primers and confirmed by Sanger
sequencing (Fig. 3a). The circular expression levels were
quantified by qRT–PCR with divergent primers calibrated by
standard curves. Consistent with the RNA-seq results, circHIPK3
was significantly more abundant in various tissues (except for
liver tissue) than the linear form as indicated by qRT–PCR
analysis (Fig. 3b). circHIPK3 is commonly expressed in various
tissues (100–600 copies per cell, assuming 20 pg RNA per cell)
and particularly enriched in the brain (Fig. 3b). We then
investigated the stability and localization of this circRNA in HeLa
cells. Total RNA was harvested at the indicated time points after
treatment with Actinomycin D, an inhibitor of transcription.
An analysis of circHIPK3 and HIPK3 mRNA revealed that the
circRNA isoform was highly stable, with transcript half-life
exceeding 24 h, whereas the associated linear transcript exhibited
half-life of o4 h (Fig. 3c). Resistance to digestion with RNase R
exonuclease further confirmed that this RNA specie is circular in
form (Fig. 3d). qRT–PCR analysis of nuclear and cytoplasmic
circHIPK3 RNA and fluorescence in situ hybridization (FISH)
against circHIPK3 demonstrated that the circular form of HIPK3
preferentially localized in the cytoplasm (Fig. 3e,f). Taken
together, our results show that circHIPK3 is an abundant and
stable circRNA expressed in different human cells.
The formation of circHIPK3 from HIPK3 exon2. We next
investigated the mechanism by which circHIPK3 is formed. An
analysis of the flanking introns of HIPK3 Exon2 showed highly
complementary Alu repeats with 28 short interspersed elements
in the intron upstream of HIPK3 Exon2 and 51 short interspersed
elements downstream from Exon3 (Fig. 4a). Among these repeats,
two primate-specific Alu elements in an inverted orientation
Copies per 20pg
0
200
400
600
800
Brain
Heart
Lung
Colon
Liver
Stomach
Relative expression
Relative expression
0.0
0.3
0.6
0.9
1.2
0.0
0.3
0.6
0.9
1.2
Mock
RNase R
circHIPK3 HIPK3 mRNA circHIPK3
circHIPK3
circHIPK3
HIPK3 mRNA
Cytoplasm
Nuclear
0 h 4 h 8 h 12 h
24 h
0.0
0.5
1.0
1.5
HIPK3 mRN
A
HIPK3 mRNA
Fold change
circHIPK3
DAPI
a
b
d
e
f
c
HIPK3
circHIPK3
circHIPK3.2 (30)
circHIPK3 (1880)
circHIPK3.5 (2)
circHIPK3.3 (7)
circHIPK3.4 (6)
8 kb
GCAG GTAT
ACAG GTA G
Figure 3 | Characterization of circHIPK3 RNA in human cells. (a) The genomic loci of five circRNAs in HIPK3 gene. The supported unique reads were
presented. The expression of circHIPK3 was validated by RT–PCR followed by sanger sequencing. Arrows represent divergent primers binding to the
genome region of circHIPK3. (b) Absolute quantification for circHIPK3 and HIPK3 mRNA in six human normal tissues. (c) qRT–PCR for the abundance of
circHIPK3 and HIPK3 mRNA in HeLa cells treated with Actinomycin D at the indicated time points. (d) qRT–PCR for the abundance of circHIPK3 and HIPK3
mRNA in HeLa cells treated with RNase R. The amount of circHIPK3 and HIPK3 mRNA were normalized to the value measured in the mock treatment.
(e) qRT–PCR data indicating the abundance of circHIPK3 and HIPK3 mRNA in either the cytoplasm or nucleus of HeLa cells. The amounts of circHIPK3
and HIPK3 mRNA were normalized to the value measured in the cytoplasm. Data in (c–e) are the means
±
s.e.m. of three experiments. (f) RNA
fluorescence in situ hybridization for circHIPK3. Nuclei were stained with 4,6-diamidino-2-phenylindole (DAPI). Scale bar, 5 mm.
NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11215 ARTICLE
NATURE COMMUNICATIONS | 7:11215 | DOI: 10.1038/ncomms11215 | www.nature.com/naturecommunications 5