Comparison of normalization methods with microRNA microarray
You-Jia Huaa,b,c,1, Kang Tua,c,1, Zhong-Yi Tanga,c,1, Yi-Xue Lia,⁎, Hua-Sheng Xiaoa,b,⁎
aBioinformatics Center, The Center of Functional Genomics, Key Lab of System Biology, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences,
Shanghai 200031, People's Republic of China
bNational Engineering Center for Biochip at Shanghai, Shanghai 201203, People's Republic of China
cGraduate School of the Chinese Academy of Sciences, Shanghai 200031, People's Republic of China
A B S T R A C TA R T I C L EI N F O
Received 6 December 2007
Accepted 1 April 2008
Available online 2 June 2008
MicroRNAs (miRNAs) are a group of RNAs that play important roles in regulating gene expression and protein
translation. In a previous study, we established an oligonucleotide microarray platform to detect miRNA
expression. Because it contained only hundreds of probes, data normalization was difficult. In this study, the
microarray data for eight miRNAs extracted from inflamed rat dorsal root ganglion (DRG) tissue were
normalized using 15 methods and compared with the results of real-time polymerase chain reaction. It was
found that the miRNA microarray data normalized by the print-tip loess method were the most consistent
with results from real-time polymerase chain reaction. Moreover, the same pattern was also observed in 14
different types of rat tissue. This study compares a variety of normalization methods and will be helpful in
the preprocessing of miRNA microarray data.
Crown Copyright © 2008 Published by Elsevier Inc. All rights reserved.
MicroRNAs (miRNAs), a large family of small, ~22-nt, noncoding
RNAs, have been identified by cloning or prediction in genomes of
dozens of species. Relevant information has been published in a
database . MiRNAs regulate a large number of genes in animals and
plants. In vertebrates, miRNAs mostly repress the translation of target
genes by binding to 3′ untranslated regions, and sometimes cleave the
mRNAs of those genes [2,3]. However, in plants, almost all of the
miRNAs cleave their target mRNAs, while a few repress transcription
[4,5]. MiRNAs are very important regulators of such biological
processes as development [6,7], cellular differentiation [8,9], and
tumor generation [10,11]. Many techniques have been used to study
miRNA expression, such as microarray, RT-PCR , Northern blotting
, and in situ hybridization. MiRNA microarray has been found to be
a global analysis tool fordetecting miRNAexpression. There have been
many microarray experiments on the relationship between miRNAs
and metabolism, cancer, development, cell fate acquisition, and tissue
differentiation;however, in mostof thesestudies,analysiswasaccom-
panied by little or no normalization. For example, Liu and Calin et al.
their miRNA microarrays on its median; Baskerville and Bartel ,
established, robust, microarray-based technique  was used to measure
the expression of 172 miRNAs in DRG after CFA-induced inflammation and
number of miRNAs and compared their microarray expression, as normal-
miRNA microarray data normalized with the print-tip loess method are
highly consistent with real-time PCR results.
Rat miRNA microarray development and the data on rat DRG from
CFA-induced inflammation model and different normal rat tissues
A rat miRNA microarray was developed that contained 172 rat
miRNA precursor sequences and 14 control miRNAs. All probes were
40 nt long, and located close to the 3′ end of each miRNA precursor.
Most of the probes contained mature miRNA sequences. For all
microarray slides, RNA samples were labeled with Cy5; Cy3-tagged
spike-in oligonucleotides were used for internal normalization. The
rat miRNA microarray was used to study miRNA expression of rat DRG
from complete Freund's adjuvant (CFA)-induced inflammation model
animals and normal rat tissues. Two sets of miRNA microarray data
were obtained. One comprised 14 rat tissues, and the other included
the time course of CFA-induced rat DRG inflammation. Experiments
were repeated two and four times, respectively. Real-time PCR was
used to validate the miRNA microarray data. A total of eight miRNAs
(rno-mir-103-2, rno-mir-128b, rno-mir-135b, rno-mir-140, rno-mir-
Genomics 92 (2008) 122–128
Abbreviations: miRNA, microRNA; RT-PCR, reverse transcription polymerase chain
reaction; DRG, dorsal root ganglion; CFA, complete Freund's adjuvant.
⁎ Corresponding authors. H.-S. Xiao is to be contacted at National Engineering Center
for Biochip at Shanghai, Shanghai 201203, People's Republic of China.
E-mail addresses: email@example.com (Y.-X. Li), firstname.lastname@example.org
1Y.-J.H., K.T., and Z.-Y.T. contributed equally to this work.
0888-7543/$ – see front matter. Crown Copyright © 2008 Published by Elsevier Inc. All rights reserved.
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/ygeno
143, rno-mir-148b, rno-mir-200b, and rno-mir-203) were selected to
test the accuracy of microarrays.
After background subtraction, the signal of each miRNA was
averaged. Coefficients of correlation between microarray replicates
were greater than 0.9. The average signal ranged from 1016 to 2945,
and average background ranged from 205 to 308. A probe set with a
signal-to-background ratio greater than 3 was considered “present.”
The present call rate among all the microarrays ranged from 36 to 74%.
Comparison of results obtained using 15 methods for normalization of
miRNA microarray data with real-time PCR data
We compared the raw microarray data for the CFA model with
real-time PCR data. The results revealed that the correlation between
the non-normalized microarray data and the real-time PCR data was
quite low (Fig. 1), ranging from –0.66 to 0.54 (Table 1). The raw
intensitiesof the positive andnegativecontrols couldnotbe separated
completely by hierarchical clustering (Figs. 2A and C). As shown in
Figs. 2B and D, after normalization, positive and negative controls
were almost completely separated from each other. This result
indicates the importance of appropriate normalization for miRNA
Next, we compared the performance of 15 normalization methods,
using the real-time PCR data as the “gold standard.” Both Pearson and
Spearman coefficients of correlation between the normalized micro-
array data and the real-time PCR results were calculated for each
normalization method (Fig. 3). Fig. 3A illustrates that for miRNA-203,
Pearson's coefficient of correlation between real-time PCR and
microarray data normalized by print-tip loess was the highest. This
result was confirmed by the results for all the other miRNAs tested, for
which the average correlation coefficient was 0.4 (Fig. 3B). Table 1 lists
all Pearson's correlation coefficients. Among the 15 normalization
Fig. 1. In the CFA-induced inflammation model, the log 2 ratio of the relative expression level of rno-mir-128b in (A) real-time PCR data, (B) print-tip loess-normalized microarray
data, and (C) non-normalized microarray data. ⁎Pb0.05; ⁎⁎Pb0.01; ⁎⁎⁎Pb0.001.
Pearson's correlation coefficients between real-time PCR data and data obtained with 15 normalization methods for eight miRNAs
Methodmir-140mir-128b mir-103-2 mir-135b mir-148bmir-143mir-200b mir-203
Y.-J. Hua et al. / Genomics 92 (2008) 122–128
methods, 8 were designed for two-channel microarrays and 7 for one-
channel microarrays. Fig. 4 illustrates that, on the whole, the two-
channel normalization methods were clearly better than the one-
channel methods. This means that that Cy3 channel, which consists of
spike-in heterogeneous oligonucleotides, is very important for system
correlation, and should be used in normalization procedures. As a
positive correlation between the Cy3 and Cy5 signals on each spot is
generally expected, it may be necessary to use the Cy5/Cy3 ratio
instead of raw intensities (Fig. 3). Among the eight two-channel
normalization methods, print-tip loess had the highest correlation
(Fig. 3 and Table 1). For example, in the CFA model, rno-miR-128b was
markedly upregulated, especially on Days 0.5 and 14 after CFA
injection, as shown in the print-tip loess-normalized microarray
data, as well as in the real-time PCR data (Fig.1). However, in the non-
normalized microarray data, rno-miR-128b appeared to be slightly
downregulated, especially on Day 4 (Fig.1). Details of the technique of
print-tip loess normalization are given in Fig. 5. There were a total of
six subarrays or blocks (2 rows×3 columns) in each microarray. The
three columns were technical triplicates. Each M value is normalized
by subtracting the corresponding value on the tip-group loess curve
from the raw data. The normalized values are the log ratios after
subtraction of the residuals of the print-tip loess regression ,
suggesting there was an M value excursion with respect to the Avalue
for most spots in each microarray before normalization (Fig. 4A), and
there was also a two-channel signal system error on each spot with
respect to its corresponding block (Fig. 4A). This system error for each
block was well eliminated from raw data by print-tip loess (Fig. 4B),
and the hypothesis of loess normalization was valid for each print-tip
analyzed the expression of one miRNA (rno-mir-203), which was
measured in 14 rat normal tissues using both microarray and real-time
PCR (Fig. 5). Apparently, print-tip loess normalization increased data
comparability between the two platforms, as can be seen in Fig. 5.
Expression of rno-miR-203 was low in olfactory bulb and heart, among
14 tissues, as indicated by both the print-tip loess-normalized
microarray data and the real-time PCR data. However, in the non-
normalized microarray data, the miRNA appeared to be highly
expressed in these two tissues. This shows that print-tip loess normal-
ization can efficiently correct systemic bias in miRNA microarrays.
Microarray is a powerful tool for high-throughput detection of
gene and miRNA expression. However, miRNA microarray has some
unique characteristics such as much fewer spots, so the normalization
methods commonly used for other types of microarrays (e.g., whole-
genome gene expression microarray) may not be appropriate. Several
articles discussing this problem have been published. The aim of this
study was to evaluate a variety of available normalization methods
and choose the one that performs best on miRNA microarray.
In the study described in this article, we designed the miRNA
microarray probes and labeling method according to Liu . The
probes of the miRNA microarray were based on the sequences of
Fig. 2. Clustering of microarraycontrol signals from: (A) raw data in miRNA tissue expressionprofiles; (B) print-tip loess-normalized data in miRNA tissue expressionprofiles; (C) raw
data for time course of CFA-induced inflammation of DRG; and (D) print-tip loess-normalized data for time course of CFA-induced inflammation of DRG. Red color denotes high
expression, and green color denotes low expression. Probes beginning with “tRNA” are positive controls, and probes beginning with “ath” are negative controls. B. brain stem;
C, cortex; D, DRG; H, heart; Hc, hippocampus; Ht, hypothalamus; K, kidney; Li, liver; Lu, lung; M, muscle; Ob, olfactory bulb; Sc, spinal cord; Sp, spleen; T, testicle.
Y.-J. Hua et al. / Genomics 92 (2008) 122–128
miRNA precursors, which included mature sites. This means that the
microarray could detect precursor and mature miRNAs. Our probes
had undergone BLAST alignment to the rat Refseq database, avoiding
or reducing nonspecific hybridization to other RNA molecules. Our
previous study indicated that mRNA has little cross-hybridization
effect on the miRNA microarray .
We observed low consistency between non-normalized micro-
array data and real-time PCR data in this study, suggesting that direct
use of microarray data without normalization is unreliable.
We compared 15 normalization methods using microarray data
and real-time PCR data. The results for both data sets showed that
two-channel data normalization is better than one-channel or no
normalization, and also demonstrated that Cy3 channel (signals of
spike-in oligonucleotides for internal control) is very important for
normalization. This is because unwanted spot effects, such as probe
concentration, shape, and size, can be eliminated by using the two-
channel intensities together.
There are many normalization methods for two-channel micro-
array data, such as loess, median, and positive control. Positive control
normalization uses the signals of positive controls (also called
“housekeeping genes”) as a standard for normalization. It is based
on the hypothesis that the expression levelof each housekeeping gene
should be invariable in different tissues or under different environ-
mental conditions. But this hypothesis is not always valid, because the
expression of some housekeeping genes may vary in different tissues.
The median method adjusts the median value of the Cy5/Cy3 log 2
ratio of all the microarrays to 0. It can eliminate systematic bias in
signals between microarrays, but cannot eliminate the bias on each
microarray . However, the loess method, which is a nonparametric
regression method, can efficiently eliminate the systematic bias in
Fig. 3. Spearman's rank correlation coefficients and Pearson's correlationcoefficients, which were calculated for the 15 normalization methods (including no normalization) and real-
time PCR. (A) Spearman's rank correlation coefficients of rno-mir-203 expression level were sorted by their values. The x axis denotes the type of method, and the y axis shows the
value of each Spearman's rank relative coefficient. (B) Clustering of the Pearson's correlation coefficients of expression level to eight miRNAs in the microarray. (C) Results of sorting
the average relative coefficients of all the miRNAs in (B) by their expression level, reflecting the average coincidence between microarray data after normalization and real-time PCR
data for eight miRNAs. The x axis denotes the normalization method, and the y axis shows the average value of the Pearson's correlation coefficients for eight miRNAs.
Y.-J. Hua et al. / Genomics 92 (2008) 122–128
signals on each microarray, but is not fit for between-array normal-
ization . Print-tip loess is a well-tested, general-purpose normal-
ization method that has provided good results on a wide range of
microarrays . Another improved method, scalePrintTipMAD,
theoretically based on scale normalization, has a high requirement
for “scale consistency.” Despite the characteristics (such as much
fewer spots), miRNA microarray is processed in the same way as other
oligonucleotide microarrays: fabrication, reverse transcription of
Fig. 4. (A) Before normalization and (B) after print-tip loess normalization. Each spot denotes the M value (A) and A value (B) of each signal, and each curve denotes the loess
regression curve of each block (or subarray) in the array. Six blocks (2×3) were marked as their row number followed by their column number. Then the M value of each spot was
checked against the regression curve.
Fig. 5. Relative expression level of rno-mir-203 in rat tissue expression profiles, in (A) real-time PCR data, (B) print-tip loess-normalized microarray data, and (C) non-normalized
microarray data. B, brain stem; C, cortex; D, DRG;H, heart; Hc, hippocampus; Ht, hypothalamus; K, kidney; Li, liver; Lu, lung; M, muscle; Ob, olfactory bulb; Sc, spinal cord; Sp, spleen;
Y.-J. Hua et al. / Genomics 92 (2008) 122–128
samples, and hybridization. Because of its universality, print-tip loess
may perform better in miRNA microarray than other methods.
Print-tip loess performed better than all the other normalization
methods on our data sets. The fact that print-tip loess is better than
the median and loess methods (Fig. 3C) illustrates that miRNA
microarray has two characteristics: (1) there is a system excursion
of log ratio relative to the Avalue; (2) there is a system excursion with
respect to each block. The method of scalePrintTipMAD, which
additionally requires “scale consistency” in different print-tip groups,
does not have as good an effect as print-tip loess. In general, fewer
spots may lead to lower consistency. So this method is not fit for
miRNA microarray because of the limited number of probes.
Materals and methods
Tissue preparation and total RNA isolation
A total of 70 adult male Sprague–Dawley rats (body weight, 200–250 g) were used
to prepare the DRG tissues from the CFA-induced inflammation model animals. The
subcutaneous injection of 200 μL of CFA was made with a sterile tuberculin syringe into
the palmar surface of the terminal phalanx of the third digit of the left hindpaw of
Sprague–Dawley rats. The rats were allowed to survive 0.5, 2, 4, 7, and 14 days (10 rats
per group). Subcutaneous injections and postinjection animal care were carried out in
accordance with the policy of the Society for Neuroscience (USA) on the use of animals
in neuroscience research and the guidelines of the Committee for Research and Ethic
Issues of the International Association for the Study of Pain. The experiments were
approved by the Committee of Use of Laboratory Animals and Common Facility,
Institute of Neuroscience, Chinese Academy of Sciences. We kept the animals under
deep anesthesia for ~1 h after the CFA injection to minimize pain. All animals were kept
in a standard environment with close monitoring and postinjection care. Animals with
inflammation and 10 normal rats were anesthetized with sodium pentobarbital (60 mg/
kg), and the tissues were dissected.
A total of 10 Sprague–Dawley male rats (body weight, 200–250 g) were used to
prepare 14 types of normal tissues. Seven neural tissues (olfactory bulb, cortex, hip-
pocampus, brain stem, hypothalamus, spinal cord, and DRG) and seven nonneural tissues
(heart, lung, muscle, spleen, testicle, kidney, and liver) were collected from each rat.
Total RNAs of all the samples were extracted with Trizol (Invitrogen, Grand Island,
NY, USA) according to the manufacturer's protocol with the following modifications:
threefold ethanol was add to the supernatant for precipitation; and after RNA isolation,
the washing step with ethanol was not performed.
A rat miRNA microarray was used to profile miRNA expression in DRG and other
tissues. A total of 172 rat miRNA precursor sequences with annotated active sites were
selected for oligonucleotide design. These sequences corresponded to rat miRNAs
published in the miRNA Registry (http://www.sanger.ac.uk/Software/Rfam/mirna; v7.0,
accessed July 2005). These miRNA microarrays contain gene-specific oligonucleotide
probes generated from 172 rat miRNAs and 14 control miRNAs (8 rat tRNAs for positive
control and 6 Arabidopsis thaliana miRNAs for negative control). BLAST alignment was
performed for all of the sequences with the corresponding genome at http://www.ncbi.
nlm.nih.gov, and the hairpin structures were analyzed at http://www.bioinfo.rpi.edu/
applications/mfold/old/rna. All probes were 40 nt long, and were dissolved in 150 mM
phosphate acid buffer (pH 7.5–8.0). The final concentration of the probes was 25 pmol/
μL. Thereafter, a certain concentration of spike-in heterogeneous oligonucleotide
sequence was interfused in all solutions, including both probes and controls. Fullmoon
Biosystem oligonucleotide slides (Fullmoon Biosystem, Sunnyvale, CA, USA) were used,
and the miRNA microarray was fabricated with a GeneMachine OmniGrid 100
Microarrayer (Gene Machine, Rochester, MN, USA) in 1×2-pin and 12×8-spot
configurations of each subarray in triplicate. For each microarray, there were six
subarrays arranged in two rows and three columns (in triplicate for each probe). The
humidity was 75%, and the temperature was 20 °C. After printing, slides were hydrated
over night in saturated salt solution, and then UV crosslinked at 600 mJ/cm2(CL1000,
UVP LLC, Upland, CA, USA).
Ten micrograms of total RNA was added to the reverse transcript reaction mix in a
final volume of 11.5 μL, containing 1 μg of [3′-(N)8-(A)3-Cy5-5′] oligonucleotide primer.
The mixture was incubated for 10 min at 70 °C and chilled on ice. With the mixture on
ice, 2 μL of 10× first-strand buffer,1 μL of 5 mM unlabeled dNTP mix,1.5 μL of 1 mM Cy5-
dCTP, 1 μL of RNase inhibitor, and 3 μL of SuperScript II RNaseHˉ reverse transcriptase
(200 units/μL, Invitrogen) were mixed; the final volume was 20 μL. The mixture was
incubated for 2 h at 42 °C and then for 10 min at 70 °C. After incubation for first-strand
cDNA synthesis, 2 μL of 2.5 N NaOH was added to the first-strand reaction mix and the
reaction was incubated at 37 °C for 15 min to denature the RNA/DNA hybrids and
degrade RNA templates. Then, 10 μL of 2 N Hepes was added to neutralize the reaction
mix. The cDNA targets were purified with the QIAquick Nucleotide Removal Kit
(Qiagene, Valencia, CA, USA). The slides were hybridized in 6× SSPE/5× Denhardt with
5 μg Cy3-tagged complementary sequence of spike in heterogeneous oligonucleotide,
which would be used as the standard for data normalization at 42 °C for 16 h, and then
washed in Lotion I (2× SSC/0.5% SDS) at 42 °C for 15 min, Lotion II (1×SSC/0.1% SDS) at
42 °C for 10 min, Lotion III (0.1× SSC) at room temperature for 5 min, and deionized
distilled waterat room temperature for 1–2 min. Processed slides werescanned with an
Agilent Scanner (Santa Clara,CA, USA)with the laser set to633and 545 nm, at power 80
and PMT 100 settings, and a scan resolution of 10 μm.
Real-time quantitative PCR
Real-time quantitative PCR was performed according to standard protocols on an
Applied Biosystem 7000 Sequence Detection System (Applied Biosystems, Foster City,
CA USA). Five micrograms of total RNA from each sample was reverse transcribed to
SYBR green PCR master mix (Applied Biosystems), 0.5 μL of Rox (Applied Biosystems), 5
pmol of each primer, and water to bring the final volume to 25 μL. The reactions were
amplified for 15 s at 95 °C and 1 min at 60 °C for 45 cycles. The thermal denaturation
protocol was run at the end of the PCR to determine the number of products present in
the reaction. U6 snRNA (U6) was used as an internal control. All reactions were run in
triplicate and included notemplateand no reverse transcription as negativecontrols for
each gene. The cycle number at which the reaction crossed an arbitrarily placed
threshold (CT) was determined for each gene, and the relative amount of each miRNA to
U6 RNA was described using 2–ΔCT, where ΔCT=(CT miRNA– CT U6RNA).
5′ and 3′ primers
Reverse: 5′- CTCCGCCGTCATCATTACC-3′
Our microarrays were hybridized with Cy5-labeled RNA samples and Cy3-tagged spike in
oligonucleotide sequence as internal controls, simultaneously. After microarray scanning
(Agilent scanner) and image reading (ImaGene), background was subtracted from signal for
each spot. As only Cy5 channel signal was related to the experimental aim, both the two-
only) were tested. Each normalization method was performed by calling corresponding
two-dimensional spatial location normalization (twoD) , within-print-tip-group intensity-
dependent location normalization (print-tip loess) , within-print-tip-group intensity-
dependent location normalization followed by within-print-tip-group scale normalization
using the median absolute deviation (scalePrintTipMAD) , positive control normalization
(log ratio.housekeeping), global transformation using variance stabilizing normalization (vsn),
and no normalization (none). One-channel data normalization methods included: quantile
normalization (cy5.quantiles) , cubic splines normalization (cy5.qspine) , local
polynomial regression fitting normalization (cy5.loess) , robust quantile normalization
(cy5.quantiles.robust) , positive control normalization (cy5.housekeeping), global transfor-
these methods were evaluated by calculating Pearson and Spearman  coefficients of
correlation between the normalized microarray data and the real-time PCR data, respectively.
We thank Xu Zhang's lab (Laboratory of Sensory System, Institute of
Y.-J. Hua et al. / Genomics 92 (2008) 122–128
Sciences)forprovidingallof theRNAsamples.Thisworkwassupported Download full-text
by the “863” program (2006AA020704) and by the “National Basic
Research Program of China” (2006CB910700, 2004CB720103,
 S. Griffiths-Jones, The microRNA Registry, Nucleic Acids Res. 32 (2004) D109–D111.
 P.H. Olsen, V. Ambros, The lin-4 regulatory RNA controls developmental timing in
Caenorhabditis elegans by blocking LIN-14 protein synthesis after the initiation of
translation, Dev. Biol. 216 (1999) 671–680.
 K. Seggerson, L. Tang, E.G. Moss, Two genetic circuits repress the Caenorhabditis
elegans heterochronic gene lin-28 after translation initiation, Dev. Biol. 243 (2002)
 G. Tang, B.J. Reinhart, D.P. Bartel, P.D. Zamore, A biochemical framework for RNA
silencing in plants, Genes Dev. 17 (2003) 49–63.
 B.J. Reinhart, et al., The 21-nucleotide let-7 RNA regulates developmental timing in
Caenorhabditis elegans, Nature 403 (2000) 901–906.
 A.L. Abbott, et al., The let-7 MicroRNA family members mir-48, mir-84, and mir-
241 function together to regulate developmental timing in Caenorhabditis elegans,
Dev. Cell Biol. 9 (2005) 403–414.
 C.Z. Chen, L. Li, H.F. Lodish, D.P. Bartel, MicroRNAs modulate hematopoietic lineage
differentiation, Science 303 (2004) 83–86.
 M. Lagos-Quintana, R.Rauhut, W. Lendeckel,T.Tuschl,Identification of novel genes
coding for small expressed RNAs, Science 294 (2001) 853–858.
 L. He, et al., A microRNA polycistron as a potential human oncogene, Nature 435
 T.D. Schmittgen, J. Jiang, Q. Liu, L. Yang, A high-throughput method to monitor the
expression of microRNA precursors, Nucleic Acids Res. 32 (2004) e43.
 J.J. Zhao, et al.,Genome-wide microRNA profiling inhuman fetal nervoustissues by
oligonucleotide microarray, Childs Nerv. Syst. 22 (2006) 1419–1425.
 C.G. Liu, et al., An oligonucleotide microchip for genome-wide microRNA profiling
in human and mouse tissues, Proc. Natl. Acad. Sci. U. S. A. 101 (2004) 9740–9744.
 G.A. Calin, et al., MicroRNA profiling reveals distinct signatures in B cell chronic
lymphocytic leukemias, Proc. Natl. Acad. Sci. U. S. A. 101 (2004) 11755–11760.
 G.A. Calin, et al., Human microRNA genes are frequently located at fragile sites and
 S. Baskerville, D.P. Bartel, Microarray profiling of microRNAs reveals frequent
coexpression with neighboring miRNAs and host genes, RNA 11 (2005) 241–247.
 R.Q. Liang, et al., An oligonucleotide microarray for microRNA expression analysis
based on labeling RNA with quantum dot and nanogold probe, Nucleic Acids Res.
33 (2005) e17.
 R Development Core Team, R: A language and environment for statistical
computing, R Foundation for Statistical Computing, 2006.
 Y.H. Yang, S. Dudoit, P. Luu, T.P. Speed, Normalization for cDNA microarray data.
 B.M. Bolstad, R.A. Irizarry, M. Astrand, T.P. Speed, A comparison of normalization
methods for high density oligonucleotide array data based on variance and bias,
Bioinformatics 19 (2003) 185–193.
 C. Workman, et al., A new non-linear normalization method for reducing
variability in DNA microarrayexperiments, Genome Biol. 3 (2002) (research0048).
 R.C. Gentleman, et al., Bioconductor: open software development for computa-
tional biology and bioinformatics, Genome Biol. 5 (2004) R80.
 M. Hollander, D.A. Wolfe, Nonparametric statistical inference, 1973.
 G.K. Smyth, T. Speed, Normalization of cDNA microarray data, Methods 31 (2003)
Y.-J. Hua et al. / Genomics 92 (2008) 122–128