Content uploaded by Gary Latham
Author content
All content in this area was uploaded by Gary Latham on Feb 22, 2016
Content may be subject to copyright.
Normalization of microRNA expression levels in
quantitative RT-PCR assays: Identification of suitable
reference RNA targets in normal and cancerous
human solid tissues
HEIDI J. PELTIER and GARY J. LATHAM
Asuragen, Inc., Austin, Texas 78744, USA
ABSTRACT
Proper normalization is a critical but often an underappreciated aspect of quantitative gene expression analysis. This study
describes the identification and characterization of appropriate reference RNA targets for the normalization of microRNA
(miRNA) quantitative RT-PCR data. miRNA microarray data from dozens of normal and disease human tissues revealed
ubiquitous and stably expressed normalization candidates for evaluation by qRT-PCR. miR-191 and miR-103, among others,
were found to be highly consistent in their expression across 13 normal tissues and five pair of distinct tumor/normal adjacent
tissues. These miRNAs were statistically superior to the most commonly used reference RNAs used in miRNA qRT-PCR
experiments, such as 5S rRNA, U6 snRNA, or total RNA. The most stable normalizers were also highly conserved across flash-
frozen and formalin-fixed paraffin-embedded lung cancer tumor/NAT sample sets, resulting in the confirmation of one well-
documented oncomir (let-7a), as well as the identification of novel oncomirs. These findings constitute the first report
describing the rigorous normalization of miRNA qRT-PCR data and have important implications for proper experimental design
and accurate data interpretation.
Keywords: miRNA; normalization; RT-PCR; biomarker; lung cancer; oncomir
INTRODUCTION
MicroRNA represents an important new class of regulatory
biomolecules, with roles in processes as diverse as early
development, cell proliferation and differentiation, apo-
ptosis, fat metabolism, and oncogenesis (for reviews, see
Ambros 2004; Bartel 2004; Esquela-Kerscher and Slack
2006; Calin and Croce 2006a). Indeed, tumor-associated,
differentially expressed miRNAs, termed ‘‘oncomirs,’’ have
been found in many different cancers and the number of
known oncomirs is growing rapidly (Calin and Croce 2006a;
Esquela-Kerscher and Slack 2006). Moreover, miRNAs show
promise in both diagnostic and therapeutic applications
(Wu et al. 2006; Szafranska et al. 2007). Although the small
size of miRNA (17–25 nucleotides [nt]) creates challenges
for their detection, recent innovative adaptations of existing
technologies for gene expression profiling, such as micro-
arrays and quantitative RT-PCR (qRT-PCR), have emerged
that make large-scale characterizations of miRNA expres-
sion patterns possible (Krichevsky et al. 2003; Nelson et al.
2004; Thomson et al. 2004). The accuracy of these methods
for expression analysis, however, is critically dependent on
proper normalization of the data inasmuch as inappropriate
normalization of qRT-PCR data can lead to incorrect con-
clusions (Tricarico et al. 2002; Bas et al. 2004).
The goal of many miRNA qRT-PCR expression experi-
ments is to identify differences between two groups of
samples, often a ‘‘normal’’ control and a ‘‘disease’’ speci-
men. Thus, the purpose of normalization is to remove as
much variation as possible between groups except for that
difference that is a consequence of the disease state itself.
Yet there are many other sources of variation in such
experiments aside from disease-specific differential expres-
sion of a particular RNA. These sources may be technical,
such as differences in sample procurement, stabilization,
RNA extraction, and target quantification, or biological,
reflecting sample-to-sample inconsistencies in cellular sub-
populations or even differences in bulk transcriptional
activity. Ideally, a normalizer is a single nucleic acid that
rna9399 Peltier and Latham ARTICLE RA
Reprint requests to: Gary J. Latham, Asuragen, Inc., 2150 Woodward
Street, Austin, TX 78744, USA; e-mail: glatham@asuragen.com; fax: (512)
681-5201.
Article published online ahead of print. Article and publication date are
at http://www.rnajournal.org/cgi/doi/10.1261/rna.939908.
844 RNA (2008), 14:844–852. Published by Cold Spring Harbor Laboratory Press. Copyright Ó2008 RNA Society.
JOBNAME: RNA 14#5 2008 PAGE: 1 OUTPUT: Saturday April 5 11:36:05 2008
csh/RNA/152280/rna9399
exhibits invariant expression across all samples, is expressed
along with the target in the cells of interest, and demon-
strates equivalent storage stability, extraction, and quanti-
fication efficiency as the target of interest. In reality, such a
normalizer does not exist (Vandesompele et al. 2002). In
the case of mRNA, multiple published reports have argued
for combinations of transcripts or ribosomal RNA (rRNA)
normalizers as part of an empirical strategy to minimize
unwanted variation (Vandesompele et al. 2002; Andersen
et al. 2004; Pfaffl et al. 2004; Szabo et al. 2004).
The identification of RNA normalizers is not possible
without first normalizing the data that is establishing a
reference baseline of expression. This paradox has been
addressed using pairwise measures (Vandesompele et al.
2002) and model-based estimates of expression variation
(Andersen et al. 2004), among other approaches (Pfaffl et al.
2004; Szabo et al. 2004). Here we describe the use of two
of these statistical methods to identify appropriate refer-
ence RNA species for miRNA qRT-PCR studies. As a class,
miRNAs pose a significant challenge for normalization.
First, miRNAs represent perhaps only 0.01% of the mass
amount of total RNA in a sample, and this fraction can vary
significantly across different samples (Liang et al. 2007).
Thus, the normalizer of choice should mirror any whole-
sale changes in the global miRNA population. Further, a
normalizer should have similar purification properties as
miRNA. However, the extraction efficiency of miRNA from
samples is very different than for much longer RNAs.
Therefore, we chose to evaluate a panel of small RNAs—
miRNA, ribosomal RNA, and small nuclear RNA
(snRNA)—in pursuit of the most stable
species in both normal and diseased
human tissues. In all cases, methods for
RNA extraction were used that are
known to efficiently recover both small
and large RNA populations. Analyses of
microarray data of RNA isolated from
dozens of tissues revealed a set of refer-
ence miRNA candidates that were further
assessed across 13 distinct human solid
tissues,and,separately,fivepairoftumor
and normal adjacent tissue (NAT). This
analysis was then extended to include 12
pair of frozen lung cancer (LuCa) and
NATaswellas16pairofformalin-fixed
paraffin-embedded (FFPE) LuCa/NAT.
RESULTS
A global strategy for the selection
of candidate RNA normalizers
Our first goal was to develop a process
for identifying the most stable targets as
normalizers in miRNA qRT-PCR stud-
ies (Fig. 1). Beginning with miRNA microarray expression
data, four criteria were used to select candidate normalizer
miRNAs in the tissues of interest: (1) the miRNA must be
highly expressed in most, if not all, of the samples; (2) the
miRNA must be consistently expressed, as measured by
the modified z-score (Supplemental Table 1); (3) only one
representative from a given miRNA family should be
considered; and (4) the miRNA must be a target of a
commercially available qRT-PCR assay at the time of the
work. Using these filters, data from the microarray content
representing 287 human miRNAs resulted in the selection
of 10–15 miRNAs for evaluation by qRT-PCR, depending
on the sample set (see Supplemental Table 1 for more in-
formation). In addition, 5S rRNA (121 nt) and U6 snRNA
(RNU6B, 45 nt) were included in some cases owing to their
purported expression stability and use in several published
miRNA qRT-PCR studies (Takamizawa et al. 2004; Choong
et al. 2007; Corney et al. 2007; Pineles et al. 2007).
Determination of the most stable miRNA normalizers
in normal human tissues
Ideally, normalizers should be empirically validated for
each sample type planned for qRT-PCR. The conventional
wisdom is to procure 10–20 representatives of the sam-
ple(s) of interest and evaluate the stability of a set of pos-
sible normalizers within the relevant sample group(s). We
refer to this approach as a ‘‘vertical scan’’ (Fig. 1), since the
same panel of candidate normalizers is analyzed ‘‘down’’
multiple representatives of the same sample type. This
FIGURE 1. Strategy for the identification of stable targets for normalizing microRNA qRT-
PCR data.
Normalization of miRNA qRT-PCR data
www.rnajournal.org 845
JOBNAME: RNA 14#5 2008 PAGE: 2 OUTPUT: Saturday April 5 11:36:06 2008
csh/RNA/152280/rna9399
Fig. 1 live 4/C
strategy is contrasted with a ‘‘horizontal scan,’’ which is
the evaluation of global stability of putative normalizers
‘‘across’’ many different sample types. We initiated our
studies with a ‘‘horizontal scan’’ of normal human solid
tissues to determine if miRNA, rRNA (5S), or snRNA (U6)
demonstrated uniform expression stability across very
different biological backgrounds. A plot of the raw C
t
values by normal tissue type for each of the 12 RNA targets
selected is shown in Supplemental Figure 1, and the raw
data file for this and all other experiments described in this
study are provided in the supplemental data.
Two different algorithms were used to assess the variance
in expression levels: geNorm (Vandesompele et al. 2002)
and NormFinder (Andersen et al. 2004). As shown in
Figure 2A, both statistical methods ranked the RNA targets
similarly from most to least stable, with an excellent
correlation in raw stability values. miR-191 was the most
consistently expressed miRNA, followed by miR-93, miR-
106a, miR-17–5p, and miR-25. In contrast, U6 and 5S, two
commonly used normalizers for miRNA qRT-PCR experi-
ments (Takamizawa et al. 2004; Pineles et al. 2007; Corney
et al. 2007; Choong et al. 2007), were the least stable (Fig. 2;
Supplemental Table 2). Indeed, the difference in stability
between miR-191 and 5S was a standard deviation of nearly
61C
t
(cycle threshold), or a difference of 62-fold. Nor-
malization to total RNA mass was also evaluated, but this
reference approach ranked behind miR-191 and miR-93 in
stability (Fig. 2B).
Determination of the most stable miRNA normalizers
in a panel of cancer and NAT samples
qRT-PCR experimental designs typically include compar-
isons of normal and diseased tissue. As a result, we
performed a ‘‘horizontal scan’’ of five
sets of tumor and NAT sample RNA
(lymphoid, colon, prostate, lung, and
esophagus) (see also Supplemental
Table 3) to evaluate the expression
stabilities of a set of 12 miRNAs selected
using the microarray gating criteria
described above. The subsequent qRT-
PCR data were analyzed both as two
discrete groups (tumor and NAT), and
one combined group (tumor + NAT).
Interestingly, miR-191 was ranked as
the most stable RNA by geNorm, and
the second most stable by NormFinder
for the NAT group, but not for the
tumor group (Supplemental Table 4).
miR-17–5p and miR-25 also ranked
among the four most stable targets by
both algorithms for the NAT group.
These results were consistent with the
finding that miR-191, miR-17–5p, and
miR-25 were among the most stable
miRNAs in the normal tissue ‘‘horizon-
tal scan’’ (Supplemental Table 2), even
though the NAT group was a unique
sample set. In contrast, miR-103 was the
most stable RNA in the tumor group.
When the two sample groups were
combined, both geNorm and Norm-
Finder identified miR-191 and miR-25
as the most stable pair of normalizers.
One important difference between the
geNorm and NormFinder methods is
that NormFinder allows for the desig-
nation of sample groups, and thus
determination of both intragroup and
intergroup (i.e., tumor versus NAT)
variation. In this specific case, however,
FIGURE 2. geNorm and NormFinder analyses of qRT-PCR data from a ‘‘horizontal scan’’ of
13 normal human tissues. (A) Correlation of the geNorm M value and the NormFinder
stability value for the 12 RNA targets evaluated. (B) The average standard deviation across all
tissue samples when normalized to (1) geometric mean (GeoMean) of miR-191 and miR-93,
(2) miR-191, (3) total RNA mass, (4) miR-16, (5) let-7a, (6) U6 snRNA, and (7) 5S rRNA.
Peltier and Latham
846 RNA, Vol. 14, No. 5
JOBNAME: RNA 14#5 2008 PAGE: 3 OUTPUT: Saturday April 5 11:36:40 2008
csh/RNA/152280/rna9399
Fig. 2 live 4/C
the intragroup variation—the only type of variation that
geNorm measures—dominated the total observed varia-
tion, and thus both algorithms converged to the same pair
of normalizers (here, miR-191 and miR-25).
Determination of the most stable miRNA normalizers
in a panel of LuCa and NAT samples
The finding that specific miRNAs could be relatively stably
expressed across different normal and cancerous tissue
prompted us to evaluate miRNA normalizers in a ‘‘vertical
scan’’ of LuCa paired with NAT samples. Two independent
sample sets were evaluated: a panel of 12 flash-frozen LuCa
tumor/NAT pairs (Supplemental Table 5), and a discrete
collection of 16 FFPE LuCa tumor/NAT pairs (Supplemen-
tal Table 6). The rationale for this experimental design was
twofold: (1) Assess the robustness of normalizer selection
to different sample preparation methods using the same
tissue type. (2) Determine the reproducibility in quantify-
ing differential miRNA expression across the two sample
sets following normalization using the independently
determined most stable RNAs. Again, the miRNA targets
were selected from array data using the gating criteria
described above, with the exception that 5S was also
included as a comparative reference. For both frozen and
FFPE sample sets, NormFinder alone was used to identify
the most stable RNA targets, given its capability to assess
overall stability using distinct intragroup and intergroup
measures of variability.
In the frozen LuCa sample pairs, miR-191 was deter-
mined to be the most stable single miRNA (Supplemen-
tal Table 7). The most stable combination of RNAs was
miR-103 and let-7a. Closer inspection of the data reveals
why miR-191 was not included in the combination of the
two best normalizers: miR-191, though uniformly expressed
in the respective tumor and NAT groups, was slightly under-
expressed in the tumor set, whereas the combination of
miR-103 and let-7a offered similarly low intragroup vari-
ability, but opposing intergroup differences that summed to
essentially zero. Thus, consistent with the results presented
above, miR-191 and miR-103 were among the most in-
variant reference RNAs (out of 16 total targets) across the 12
pairs of tumor and NAT tissues. 5S, by comparison, was one
of the most unstable targets (ranked 13th out of 16 targets).
A subset of the most stable miRNAs from the frozen
LuCa experiment was then evaluated in 16 pairs of FFPE
LuCa tumor and NAT samples. The most stable RNA was
miR-103, followed by miR-191 (Supplemental Table 8).
The most stable pair of targets was miR-17–5p and miR-24.
Strikingly, six of the seven miRNAs common to the frozen
and FFPE LuCa datasets followed the same rank order of
stability (Supplemental Table 9). The lone exception was
miR-103, which ranked first in the FFPE LuCa set (out of
eight total targets), but fourth in the frozen LuCa set (out
of 16 total targets).
A case study for normalized differential expression
measurements by miRNA qRT-PCR: Quantification
of let-7a in nonsmall cell LuCa
An additional goal of this study was to determine how the
independent selection of most stable normalizers would
impact the quantification of miRNAs in normal versus
tumor group comparisons. We selected let-7a as a case
study for the following reasons: (1) several independent
studies have reported that let-7a is down-regulated in LuCa
(Takamizawa et al. 2004; Johnson et al. 2005; Yanaihara
et al. 2006; Inamura et al. 2007); (2) a published study has
measured let-7a expression in LuCa using the same qRT-
PCR assay format described here (Inamura et al. 2007),
permitting a quantitative comparison; and (3) a plausible
biological model to support the consequences of let-7a
down-regulation in LuCa tumors has been successfully
tested (Johnson et al. 2005).
Consequently, we evaluated let-7a expression in both the
frozen and FFPE LuCa sample sets when normalized to
RNA targets representing both relatively stable and unstable
expression. As shown in Figure 3A, let-7a expression was
1.2-fold lower in tumors than it was in NAT when
normalized to miR-191, although this modest change was
not statistically significant (P= 0.182). When let-7a was
normalized to the most stable pair that did not include let-
7a (that is, miR-191 and miR-24), the fold change was
comparable (1.4-fold), but also statistically significant
(P= 0.007). In fact, this degree of reduced expression of
let-7a in tumors was similar to that reported by Inamura
et al. (2007) (approximately 1.8-fold) using qRT-PCR.
However, if either 5S (P< 0.005) or total RNA (P< 0.0004)
were used as normalizers, the fold change for let-7a down-
regulation was roughly twice (2.2 to 2.8-fold) that
measured after normalization with the most stable refer-
ence miRNAs. Most alarmingly, normalization to miR-30d
reported twofold higher, rather than lower, let-7a expres-
sion in tumors with statistical confidence (P= 0.008). This
misleading result is a consequence of the systematically
lower expression of miR-30d in LuCa.
Next, a subset of the most stable miRNAs was evaluated
in the FFPE LuCa sample set (Fig. 3B). In this case,
normalization to miR-103 resulted in 1.5-fold lower
expression for let-7a in tumor (P= 0.01). Similarly,
normalization to the most stable pair, miR-17–5p and
miR-24, revealed 1.5-fold lower expression in tumors (P=
0.02). In contrast, if let-7a were normalized to less stable
targets, such as miR-25, miR-16, or even total RNA,
expression was interpreted as either lower (miR-25) or
higher (miR-16, or total RNA) in tumors, with none of the
outcomes reaching statistical significance. Another impor-
tant finding was that normalization to total RNA failed to
correct for large variations in expression as observed with
some samples. For example, three samples (see starred
samples in Fig. 3B) produced unexpectedly high C
t
values
Normalization of miRNA qRT-PCR data
www.rnajournal.org 847
JOBNAME: RNA 14#5 2008 PAGE: 4 OUTPUT: Saturday April 5 11:37:30 2008
csh/RNA/152280/rna9399
in either the tumor or the NAT RNA. Normalization to
total RNA cannot correct for inadvertent errors in total
RNA quantification, or the effects of inhibitors that reduce
the efficiency of the RT and/or PCR steps that can produce
such anomalies. In contrast, endogenous reference miRNAs
are coamplified in the same sample, and thus can help
correct data outliers that are sometimes observed when
working with real-world clinical specimens.
As a result, normalization to the most stable pair of
miRNAs for the frozen and FFPE sample panels, respec-
tively, both supported the statistically significant conclusion
of down-regulation of let-7a in lung tumors, consistent
with literature reports (Takamizawa et al. 2004; Johnson
et al. 2005; Yanaihara et al. 2006; Inamura et al. 2007).
Moreover, the quantification of differential expression was
highly reproducible between the independent frozen (1.4-
fold) and FFPE (1.5-fold) sample sets (Fig. 3).
Normalization with the most stable miRNAs identifies
novel oncomirs in LuCa
Having determined the most invariant miRNA normalizers
in LuCa tissue, we next measured the expression differences
associated with each of the RNAs that were evaluated in
the frozen sample collection. The 16 RNA targets were indi-
vidually normalized to: (1) the most stable pair or targets,
(2) the most stable single target, or (3) 5S rRNA. The fold
change and P-value was determined in each case (Table 1).
Normalization to stable RNA targets identified z30%
more targets that were differentially expressed at statistical
FIGURE 3. Differential expression of hsa-let-7a in human lung cancer: impact of normalizer stability on qRT-PCR expression analyses from both
flash-frozen and FFPE tumor and normal adjacent tumor tissues. (A) Hsa-let-7a differential expression in 12 flash-frozen lung cancer NAT pair as
normalized to (1) geometric mean (GeoMean) of miR-191 and miR-24, (2) miR-191, (3) miR-103, (4) 5S rRNA, (5) total RNA, and (6) miR-30d.
The dotted line indicates the respective average differential expression value. Negative values indicate reduced let-7a expression in lung cancer
samples and positive values indicate increased let-7a expression in lung cancer samples. (B) Hsa-let-7a differential expression in 16 FFPE lung
cancer NAT pairs as normalized to (1) geometric mean (GeoMean) of miR-17–5p and miR-24, (2) miR-103, (3) miR-191, (4) miR-16, (5) miR-
25, and (6) total RNA. The dotted line indicates the respective average differential expression value. (
*
) Indicates an outlier. Negative values
indicate reduced let-7a expression in lung cancer samples and positive values indicate increased let-7a expression in lung cancer samples.
Peltier and Latham
848 RNA, Vol. 14, No. 5
JOBNAME: RNA 14#5 2008 PAGE: 5 OUTPUT: Saturday April 5 11:37:31 2008
csh/RNA/152280/rna9399
Fig. 3 live 4/C
confidence (P< 0.05), and, overall, a much more balanced
group of up- and down-regulated targets, as would be
expected from this random sampling of miRNA targets.
Moreover, two miRNAs, miR-30d and miR-221, were
determined to be >2-fold differentially expressed in the
LuCa samples, a result supported by a separate analysis of
the FFPE LuCa sample panel (data not shown). We note
miR-30d has been previously reported to be reduced in
LuCa (Volinia et al. 2006); however, miR-221, which was
2.2-fold up-regulated in our study, has not been described
as differentially expressed in LuCa. Importantly, miR-221
would not have been identified as differentially expressed if
5S had been used as a normalizer (P= 0.39).
DISCUSSION
The goal of normalization is to minimize data variation that
can mask or exaggerate biologically meaningful changes,
thereby increasing the accuracy of expression measurements.
The choice of a normalization strategy, however, is anything
but trivial. In microarray studies, normalization methods
usually invoke sophisticated, population-based approaches
that exploit abundance measures of hundreds, if not
thousands, of genes. This luxury is absent in qRT-PCR,
where assays are typically designed to target a few to dozens
of targets, and limitations in optical bandwidth constrain
the number of multiplexed targets that can be quantified in
a single tube to five or less. Other considerations, such as
cost and the availability of sample material, may also restrict
the number of normalizers that can be evaluated.
Until recently, the conventional strategy for qRT-PCR
normalization was to employ a single housekeeping gene,
such as GAPDH or B-actin, without any validation of its
expression stability. This approach is ill-advised as such
genes can vary by 10-fold or more across different samples
(Warrington et al. 2000). Vandesompele et al. (2002) first
proposed an empirical approach to selecting suitable
normalizers for mRNA qRT-PCR, advocating that one to
three reference genes should be validated for each sample
type and corresponding sample preparation method. Yet,
to our knowledge, no study has been published that has
systematically evaluated normalization targets in miRNA
qRT-PCR assays, despite the surge of interest in miRNA
identification and quantification.
To date, only a handful of miRNA qRT-PCR studies
have described some form of normalization, using RNA
targets such as 5S (Takamizawa et al. 2004; Pineles et al.
2007), U6 (Choong et al. 2007; Corney et al. 2007; Shell
et al. 2007), 18S (Iorio et al. 2007), or miR-16 or let-7a
(Mattie et al. 2006). Rigorous experimental justification for
the selection of these targets in these publications has been
lacking. Importantly, our data suggests that total RNA is
inferior to miRNAs such as miR-191, miR-103, and miR-
17–5p in more than one biological context. In fact, 5S or
U6 were the two least stable RNA species in a panel of 12
RNA targets evaluated across 13 discrete normal human
tissues (Fig. 2). The fact that all of the miRNAs that were
tested were superior to these two commonly used normal-
izers may in part reflect the effectiveness of the miRNA
microarray selection criteria (Fig. 1).
Interestingly, the most stable miRNAs identified in the
normal tissue survey were also among the most stable RNA
targets in experiments with other sample sets. For example,
miR-191, miR-25, and miR-17–5p were three of the four
most stable RNAs in the NAT group from a panel of five
distinct cancer tissues, consistent with their stability in the
normal tissue survey. In tumor samples, miR-103 was the
most invariant RNA, whereas the combination of miR-191
and miR-25 was the most stable pair.
TABLE 1. Quantitative consequences of normalizing miRNA qRT-
PCR data to stable miRNA versus less stable rRNA targets: results
from a gene expression case study of flash-frozen lung cancer and
normal adjacent tumor samples
Negative values indicate reduced expression in lung cancer
samples (highlighted in green), whereas positive values (high-
lighted in red) indicate increased expression in lung cancer
samples. Pvalues <0.05 are shown in bold.
Normalization of miRNA qRT-PCR data
www.rnajournal.org 849
JOBNAME: RNA 14#5 2008 PAGE: 6 OUTPUT: Saturday April 5 11:38:32 2008
csh/RNA/152280/rna9399
Tab. 1 liv e 4 / C
To extend these findings, we next evaluated RNA
stability in 12 pairs of flash-frozen LuCa/NAT tissues, and,
separately, in 16 pairs of FFPE LuCa/NAT tissues. The rank
order of RNA stabilities for those seven RNAs in common
to both sample sets were extremely consistent, with six out
of seven miRNAs ordering the same from most to least
stable. This conservation in rank order between the two
unique sample sets was noteworthy, since each of the
reference RNAs demonstrated uniform expression in the
independent frozen LuCa set, and were clustered within a
relatively small range of expression stabilities. Consistent
with previous data, miR-191 and miR-103 were the most
stable single reference targets in both sample sets. Thus, the
miRNA stability ranking is robust to the sample treatment
and processing for human lung tissue.
The practical consequences of miRNA normalization
were then evaluated using let-7a as a case study, given the
evidence from multiple groups that let-7a is down-regulated
in lung tumors (Takamizawa et al. 2004; Johnson et al. 2005;
Yanaihara et al. 2006; Inamura et al. 2007). The results
revealed that the reduced expression of let-7a in LuCa was
determined with quantitative consistency (<10% difference
in fold change) and statistical confidence (P< 0.05) across
both the frozen and FFPE sample sets when the most stable
pair of reference miRNAs were used to normalize the data in
each case. Additionally, normalization to less stable refer-
ence species caused both quantitative (i.e., magnitude of
fold change) and qualitative (i.e., direction of fold change)
errors in let-7a differential expression. Worse yet, inappro-
priate normalization can support statistical confidence in
the wrong conclusion (see miR-30d panel in Fig. 3A, which
suggests that let-7a expression is up-regulated in LuCa).
There are several important implications from these
findings. First, large differences in expression between
normal and disease groups may be tolerant to poor
normalization, but small differences certainly will not. It
has been suggested that the capability of miRNAs to
regulate multiple targets within the same pathway can
amplify their biological effects (Calin and Croce 2006b).
As a result, even relatively small changes in miRNA
expression may be biologically significant. For this reason,
rigorous normalization of miRNA data may be even more
critical than that of other RNA functional classes. In fact,
stringent normalization may be required in some cases to
enable a detailed understanding of miRNA biology, and it
may be critical for the potential development of reliable
diagnostic assays. Second, a single miRNA normalizer may
be sufficient in some experimental situations, but more
than one may be required to produce accurate data that can
be interpreted with confidence. Last, the consistency of the
results between frozen and FFPE LuCa sample sets suggests
that normalization of miRNA qRT-PCR assays using
reference miRNA can be extremely robust. This is a
particularly remarkable result given that the two tissue
sources were procured independently of one another, and,
further, that FFPE samples present significant sources of
variation in RNA expression profiling compared to flash-
frozen samples (Xi et al. 2007). For example, FFPE samples
are subjected to harsh chemical fixation and high temper-
ature during the embedding process—procedures that both
damage RNA species and increase the variability in RNA
representation following extraction and amplification. The
observed conservation in miRNA stability across frozen
and FFPE sample groups is a testament to their utility in
normalizing qRT-PCR data using very different tissue
preparation methods. It has been speculated that miRNA,
owing to their small size, can survive this harsh tissue
processing with less variability than mRNA (Xi et al. 2007).
Finally, we recognize that the effort, cost, and sample
requirements necessary for the experimental selection of
miRNA normalizers is not always possible. In such cases,
the data presented here suggests that miR-191, miR-103,
and/or miR-17–5p, are likely more reasonable choices than
the default use of U6, 5S, or even a more or less randomly
selected miRNA such as miR-16. However, it is clear that
empirical validation is the optimal strategy for ensuring
accurate miRNA quantification by qRT-PCR.
MATERIALS AND METHODS
Total RNA samples
Samples sets included: (1) 13 individual normal flash-frozen
human tissue RNAs (Fig. 1), (2) five flash-frozen human tumor/
NAT RNA pairs (Fig. 1; Supplemental Table 3), (3) 12 flash-frozen
human lung tumor/NAT RNA pairs (Supplemental Table 5), and
(4) 16 FFPE human lung tumor/NAT RNA pairs. Samples derived
from human patients were acquired from commercial suppliers
by Asuragen’s Tissue Procurement Group in compliance with
the regulations as outlined in 45 CFR 46 and other regulatory
guidance (Supplemental Table 6).
Accurate and sensitive miRNA quantification is only possible
when methods are used that efficiently isolate small RNA species.
For this reason, FirstChoice Total RNA samples (Ambion),
certified to contain small RNAs (miRNA, siRNA, and snRNA)
as well as large RNAs (rRNA, mRNA, and tRNA), were used to
generate microRNA expression profiling and real-time RT-PCR
data unless otherwise noted. Total RNA from the 12 pair of flash-
frozen lung tissues was extracted using mirVana miRNA Isolation
Kit (Ambion) following the manufacturer’s protocol to ensure the
recovery of small RNA. Total RNA from 16 pair of FFPE lung
tissues was isolated using RecoverAll Total Nucleic Acid Isolation
Kit (Ambion) as described by Doleshal et al. (in press).
RNA concentrations were verified by measuring absorbance
(A260) on the NanoDrop Spectrophotometer ND-1000 (Nano-
Drop) and total RNA profiles were assessed on the Agilent 2100
bioanalyzer (Agilent Technologies) with equal mass loadings of
100 ng per sample onto the RNA 6000 Nano LabChip kit. The
Eukaryote Total RNA Nano assay on the 2100 bioanalyzer expert
software reported 28S/18S ratios in the typical range of 1.1–1.8 for
RNA isolated from flash-frozen tissues. RNA profiles for FFPE
RNA samples were not assessed on the bioanalyzer.
Peltier and Latham
850 RNA, Vol. 14, No. 5
JOBNAME: RNA 14#5 2008 PAGE: 7 OUTPUT: Saturday April 5 11:39:04 2008
csh/RNA/152280/rna9399
Selection of candidate targets for normalization
miRNA array expression profiling data sets were prepared
and generated in-house using mirVana miRNA Bioarrays V1
(Ambion) as previously described (Shingara et al. 2005). Within
each array data set, miRNA targets were further standardized by a
modified z-score ranking as described in Supplemental Table 1.
Following the modified z-score filtering, each normalization
candidate was transformed to a quantity as outlined by the
authors of geNorm and NormFinder (Vandesompele et al. 2002;
Andersen et al. 2004), creating a list of targets ranked accordingly
relative to a stability value (Supplemental Table 1). In addition, 5S
and U6 TaqMan assays were included based on their historical
use for Northern blot normalization. Both assays for 5S and U6
had comparable reproducibility and performance to the miRNA
TaqMan assays.
Real-time RT-PCR
qRT-PCR was performed in duplicate, including minus reverse
transcription (RT) controls to assess genomic DNA and non-
template controls that ensured a lack of signal in the assay
background. The RT reaction consisted of 1.0 mL103RT Buffer
(Ambion), 1.0 mL dNTPs 2.5 mM each (Ambion), 0.5 mL53RT
Primer (Applied Biosystems), 0.1 mL RNase Inhibitor Protein
40 U/mL (Ambion), 0.1 mL wt-MMLV-RT 100 U/ul (Ambion),
and 500 pg total RNA in a final volume of 10 mL. Reactions were
incubated on a 384-well GeneAmp PCR System 9700 at 16°C for
30 min, 42°C for 30 min, then 85°C for 5 min. Following the RT
step, 2 mL of the RT product was transferred into a 15 mL PCR
consisting of 1.5 mL103PCR Buffer (Invitrogen), 1.5 mL50mM
Magnesium Chloride (Invitrogen), 1.5 mL dNTP’s 2.5 mM each
(Ambion), 0.3 mL203TaqMan Assay (Applied Biosystems),
0.3 mL503ROX Standard (Ambion), and 0.1 mL Platinum Taq 5
U/mL (Invitrogen). PCR cycling began with template denaturation
and hot start Taq activation at 95°C for 1 min, then 40 cycles of
95°C for 5 sec, and 60°C for 30 sec performed on a 7900HT Fast
Real-Time PCR System with data collected during each cycle at
the 60°C extension step with 7900HT SDS v2.3 (Applied Bio-
systems). Threshold and baselines were manually determined with
thresholds typically set between 0.05 to 0.1 paired with a baseline
starting at 1 – 3 C
t
s and finishing at 15 – 17 C
t
s.
qRT-PCR data analysis
The qRT-PCR results were imported into Microsoft Excel and the
average value of duplicate C
t
values converted to quantities for
geNorm and NormFinder analysis (Vandesompele et al. 2002;
Andersen et al. 2004). The assessment of let-7a expression
comparing different targets as normalizers in LuCa was deter-
mined by the ddCt comparative threshold (DDC
t
) method.
P-values were determined by a two-tailed paired Student’s ttest
from the DC
t
values of tumor and NAT or C
t
values in the case of
normalization to total RNA (Fig. 3A,B). Normality of the data was
assessed by Lilliefor’s hypothesis test.
SUPPLEMENTAL DATA
Supplemental material can be found at http://www.rnajournal.
org.
ACKNOWLEDGMENTS
This work is supported in part by grant R44CA118785 from the
National Cancer Institute (NCI). We thank Ania Szafranska for
generous access to FFPE LuCa tumor/NAT samples, Tim Davison
for helpful statistical analysis discussions, and Bernie Andruss for
his critical review of manuscript.
Received November 26, 2007; accepted January 11, 2008.
REFERENCES
Ambros, V. 2004. The functions of animal microRNAs. Nature 431:
350–355.
Andersen, C.L., Jensen, J.L., and Orntoft, T.F. 2004. Normalization of
real-time quantitative reverse transcription-PCR data: A model-
based variance estimation approach to identify genes suited for
normalization, applied to bladder and colon cancer data sets.
Cancer Res. 64: 5245–5250.
Bartel, D.P. 2004. MicroRNAs: Genomics, biogenesis, mechanism,
and function. Cell 116: 281–297.
Bas, A., Forsberg, G., Hammarstro
¨m, S., and Hammarstro
¨m, M.L.
2004. Utility of the housekeeping genes 18S rRNA, beta-actin and
glyceraldehyde-3-phosphate-dehydrogenase for normalization in
real-time quantitative reverse transcriptase-polymerase chain reac-
tion analysis of gene expression in human T lymphocytes. Scand. J.
Immunol. 6: 566–573.
Calin, G.A. and Croce, C.M. 2006a. MicroRNA signatures in human
cancers. Nat. Rev. Cancer 6: 857–866.
Calin, G.A. and Croce, C.M. 2006b. MicroRNA–cancer connection:
The beginning of a new tale. Cancer Res. 15: 7390–7394.
Choong, M.L., Yang, H.H., and McNiece, I. 2007. MicroRNA
expression profiling during human cord blood-derived CD34 cell
erythropoiesis. Exp. Hematol. 4: 551–564.
Corney, D.C., Flesken-Nikitin, A., Godwin, A.K., Wang, W., and
Nikitin, A.Y. 2007. MicroRNA-34b and microRNA-34c are targets
of p53 and cooperate in control of cell proliferation and adhesion-
independent growth. Cancer Res. 18: 8433–8438.
Doleshal, M., Magotra, A.A., Choudhury, B., Cannon, B.D.,
Labourier, E., and Szafranska, A.E. 2008. Evaluation and valida-
tion of total RNA extraction methods for microRNA expression
analyses in formalin-fixed, paraffin-embedded tissues. J. Mol. Diag.
(in press).
Esquela-Kerscher, A. and Slack, F.J. 2006. Oncomirs—MicroRNAs
with a role in cancer. Nat. Rev. Cancer 6: 259–269.
Inamura, K., Togashi, Y., Nomura, K., Ninomiya, H., Hiramatsu, M.,
Satoh, Y., Okumura, S., Nakagawa, K., and Ishikawa, Y. 2007. Let-
7 microRNA expression is reduced in bronchioloalveolar carci-
noma, a non-invasive carcinoma, and is not correlated with
prognosis. Lung Cancer 58: 392–396.
Iorio, M.V., Visone, R., Di Leva, G., Donati, V., Petrocca, F.,
Casalini, P., Taccioli, C., Volinia, S., Liu, C.G., Alder, H., et al.
2007. MicroRNA signatures in human ovarian cancer. Cancer Res.
18: 8699–8707.
Johnson, S.M., Grosshans, H., Shingara, J., Byrom, M., Jarvis, R.,
Cheng, A., Labourier, E., Reinert, K.L., Brown, D., and Slack, F.J.
2005. RAS is regulated by the let-7 microRNA family. Cell 120:
635–647.
Krichevsky, A.M., King, K.S., Donahue, C.P., Khrapko, K., and
Kosik, K.S. 2003. A microRNA array reveals extensive regulation
of microRNAs during brain development. RNA 9: 1274–1281.
Liang, Y., Ridzon, D., Wong, L., and Chen, C. 2007. Characterization
of microRNA expression profiles in normal human tissues. BMC
Genomics 8: 166; doi: 10.1186/1471-2164-8-166.
Mattie, M.D., Benz, C.C., Bowers, J., Sensinger, K., Wong, L.,
Scott, G.K., Fedele, V., Ginzinger, D., Getts, R., and Haqq, C.
2006. Optimized high-throughput microRNA expression profiling
Normalization of miRNA qRT-PCR data
www.rnajournal.org 851
JOBNAME: RNA 14#5 2008 PAGE: 8 OUTPUT: Saturday April 5 11:39:05 2008
csh/RNA/152280/rna9399
provides novel biomarker assessment of clinical prostate and
breast cancer biopsies. Mol. Cancer 5: 24; doi: 10.1186/1476-
4598-5-24.
Nelson, P.T., Baldwin, D.A., Scearce, L.M., Oberholtzer, J.C.,
Tobias, J.W., and Mourelatos, Z. 2004. Microarray-based, high-
throughput gene expression profiling of microRNAs. Nat. Methods
1: 155–161.
Pfaffl, M.W., Tichopad, A., Prgomet, C., and Neuvians, T.P. 2004.
Determination of stable housekeeping genes, differentially regu-
lated target genes and sample integrity: BestKeeper—Excel-based
tool using pair-wise correlations. Biotechnol. Lett. 6: 509–515.
Pineles, B.L., Romero, R., Montenegro, D., Tarca, A.L., Han, Y.M.,
Kim, Y.M., Draghici, S., Espinoza, J., Kusanovic, J.P., Mittal, P.,
et al. 2007. Distinct subsets of microRNAs are expressed differen-
tially in the human placentas of patients with preeclampsia.
Am. J. Obstet. Gynecol. 3: 261.e1–e6.
Shell, S., Park, S.M., Radjabi, A.R., Schickel, R., Kistner, E.O.,
Jewell, D.A., Feig, C., Lengyel, E., and Peter, M.E. 2007. Let-7
expression defines two differentiation stages of cancer. Proc. Natl.
Acad. Sci. 104: 11400–11405.
Shingara, J., Keiger, K., Shelton, J., Laosinchai-Wolf, W., Powers, P.,
Conrad, R., Brown, D., and Labourier, E. 2005. An optimized
isolation and labeling platform for accurate microRNA expression
profiling. RNA 11: 1461–1470.
Szabo, A., Perou, C.M., Karaca, M., Perreard, L., Quackenbush, J.F.,
and Bernard, P.S. 2004. Statistical modeling for selecting house-
keeper genes. Genome Biol. 5: R59; doi: 10.1186/gb-2004-5-8-r59.
Szafranska, A.E., Davison, T.S., John, J., Cannon, T., Sipos, B.,
Maghnouj, A., Labourier, E., and Hahn, S.A. 2007. MicroRNA
expression alterations are linked to tumorigenesis and non-neoplastic
processes in pancreatic ductal adenocarcinoma. Oncogene 26: 1–11.
Takamizawa, J., Konishi, H., Yanagisawa, K., Tomida, S., Osada, H.,
Endoh, H., Harano, T., Yatabe, Y., Nagino, M., Nimura, Y., et al.
2004. Reduced expression of the let-7 microRNAs in human lung
cancers in association with shortened postoperative survival.
Cancer Res. 11: 3753–3756.
Thomson, J.M., Parker, J., Perou, C.M., and Hammond, S.M. 2004.
Custom microarray platform for analysis of microRNA gene
expression. Nat. Methods 1: 47–53.
Tricarico, C., Pinzani, P., Bianchi, S., Paglierani, M., Distante, V.,
Pazzagli, M., Bustin, S.A., and Orlando, C. 2002. Quantitative
real-time reverse transcription polymerase chain reaction:
Normalization to rRNA or single housekeeping genes is inap-
propriate for human tissue biopsies. Anal. Biochem. 309: 293–
300.
Vandesompele, J., De Preter, K., Pattyn, F., Poppe, B., Van Roy, N.,
De Paepe, A., and Speleman, F. 2002. Accurate normalization of
real-time quantitative RT-PCR data by geometric averaging of mul-
tiple internal control genes. Genome Biol. 3: RESEARCH0034.1;
doi: 10.1186/gb-2007-3-8-research0034.
Volinia, S., Calin, G.A., Liu, C.G., Ambs, S., Cimmino, A.,
Petrocca, F., Visone, R., Iorio, M., Roldo, C., Ferracin, M., et al.
2006. A microRNA expression signature of human solid tumors
defines cancer gene targets. Proc. Natl. Acad. Sci. 103: 2257–2261.
Warrington, J.A., Nair, A., Mahadevappa, M., and Tsyganskaya, M.
2000. Comparison of human adult and fetal expression and
identification of 535 housekeeping/maintenance genes. Physiol.
Genomics 3: 143–147.
Wu, W., Sun, M., Zou, G.M., and Chen, J. 2006. MicroRNA and
cancer. Current status and prospective. Int. J. Cancer 120: 953–960.
Xi, Y., Nakajima, G., Gavin, E., Morris, C.G., Kudo, K., Hayashi, K.,
and Ju, J. 2007. Systematic analysis of microRNA expression of
RNA extracted from fresh frozen and formalin-fixed paraffin-
embedded samples. RNA 13: 1668–1674.
Yanaihara, N., Caplen, N., Bowman, E., Seike, M., Kumamoto, K.,
Yi, M., Stephens, R.M., Okamoto, A., Yokota, J., Tanaka, T., et al.
2006. Unique microRNA molecular profiles in lung cancer
diagnosis and prognosis. Cancer Cell 9: 189–198.
Peltier and Latham
852 RNA, Vol. 14, No. 5
JOBNAME: RNA 14#5 2008 PAGE: 9 OUTPUT: Saturday April 5 11:39:06 2008
csh/RNA/152280/rna9399
