Agreement in Breast Cancer Classification
between Microarray and Quantitative Reverse
Transcription PCR from Fresh-Frozen and
Formalin-Fixed, Paraffin-Embedded Tissues
Michael Mullins,1Laurent Perreard,2John F. Quackenbush,1Nicholas Gauthier,1
Steven Bayer,1Matthew Ellis,3Joel Parker,4Charles M. Perou,5Aniko Szabo,6and
Philip S. Bernard1,2*
Background: Microarray studies have identified differ-
ent molecular subtypes of breast cancer with prognostic
significance. To transition these classifications into the
clinical laboratory, we have developed a real-time quan-
titative reverse transcription (qRT)-PCR assay to diag-
nose the biological subtypes of breast cancer from
fresh-frozen (FF) and formalin-fixed, paraffin-embed-
ded (FFPE) tissues.
Methods: We used microarray data from 124 breast
samples as a training set for classifying tumors into 4
HER2?/ER?, basal-like, and normal-like. We used the
training set data in 2 different centroid-based algo-
rithms to predict sample class on 35 breast tumors (test
set) procured as FF and FFPE tissues (70 samples). We
classified samples on the basis of large and minimized
gene sets. We used the minimized gene set in a real-time
qRT-PCR assay to predict sample subtype from the FF
and FFPE tissues. We evaluated primer set performance
between procurement methods by use of several mea-
sures of agreement.
Results: The centroid-based algorithms were in com-
plete agreement in classification from FFPE tissues by
use of qRT-PCR and the minimized “intrinsic” gene set
(40 classifiers). There was 94% (33 of 35) concordance
between the diagnostic algorithms when comparing
subtype classification from FF tissue by use of microar-
ray (large and minimized gene set) and qRT-PCR data.
We found that the ratio of the diagonal SD to the
dynamic range was the best method for assessing agree-
ment on a gene-by-gene basis.
Conclusions: Centroid-based algorithms are robust
classifiers for breast cancer subtype assignment across
platforms and procurement conditions.
© 2007 American Association for Clinical Chemistry
Expression-based classifications are important for deter-
mining risk of relapse and making treatment decisions in
breast cancer (1–4). Classifications are often developed
using microarray data and then further validated on the
same or different platforms by use of minimized gene
sets. For instance, van’t Veer et al. (4) and van de Vijver et
al. (5) used microarray data in training and test sets to
validate a 70-gene signature that predicts relapse in
early-stage estrogen receptor (ER)7-positive and ER-neg-
ative tumors. In addition, Paik et al. (2) developed a
16-gene classifier that predicts relapse in ER-positive
1Department of Pathology, University of Utah School of Medicine, Salt
Lake City, UT.
2The ARUP Institute for Clinical and Experimental Pathology, Salt Lake
3Siteman Cancer Center, Washington University, St. Louis, MO.
4Constella Group, Durham, NC.
5Departments of Genetics and Pathology and Laboratory Sciences,
Lineberger Comprehensive Cancer Center, University of North Carolina at
Chapel Hill, Chapel Hill, NC.
6Department of Oncological Sciences, Huntsman Cancer Institute, Salt
Lake City, UT.
* Address correspondence to this author at: Huntsman Cancer Institute,
2000 Circle of Hope, Suite 3345, Salt Lake City, UT 84112-5550. Fax 801-585-
9872; e-mail email@example.com.
Received Received December 5, 2006; accepted May 3, 2007.
Previously published online at DOI: 10.1373/clinchem.2006.083725
7Nonstandard abbreviations: ER, estrogen receptor; qRT, quantitative
reverse transcription; FFPE, formalin-fixed, paraffin-embedded; PAM, predic-
tion analysis of microarray; SSP, single sample predictor; FF, fresh-frozen; Cy,
cyanine; m, diagonal bias; d, diagonal spread; dsd, diagonal SD; rd, diagonal
correlation; ccc, concordance correlation coefficient; DWD, distance weighted
discrimination; IHC, immunohistochemistry.
Clinical Chemistry 53:7
tumors by use of quantitative reverse transcription (qRT)-
PCR on formalin-fixed, paraffin-embedded (FFPE) tis-
sues. Furthermore, Perou et al. (3) and Sorlie et al. (6)
showed that hierarchical clustering of microarray data
separates breast tumors into different “biological” sub-
types (luminal, HER2?/ER?, basal-like, and normal-like)
and that these subtypes are prognostic. The biological
classification has been validated on multiple patient co-
horts by use of cross-platform microarray analyses and
Although there are few genes in common between
those used to determine the biological subtypes and those
used in other classifications for breast cancer prognosis,
the different tests identify similar properties that predict
tumor behavior (1). The classification for biological sub-
types is based on hierarchical clustering, a major differ-
ence between it and other classifications for breast cancer.
The unsupervised nature of hierarchical clustering is
effective for discovery (10), but it is not suitable for
predicting a new sample’s class since dendrogram asso-
ciations can change when new data are introduced. How-
ever, it is possible to classify samples within the frame-
methods (7, 11–13). For instance, Tibshirani et al. (13)
showed that the nearest shrunken centroid method, used
in prediction analysis of microarray (PAM), can classify
samples as accurately as statistical approaches such as
artificial neural networks. In addition, Hu et al. (7) used
another simple centroid method called single sample
predictor (SSP) to classify subtypes of breast cancer.
We have shown that a minimized intrinsic gene set can
be used in a qRT-PCR assay to recapitulate the microarray
classification of breast cancer subtypes (8). In this study,
we refined our minimal gene set by using data from Hu et
al. (7) and compared 2 centroid-based methods for our
breast cancer classification across platforms (microarray
and qRT-PCR) and procurement methods [fresh-frozen
(FF) and FFPE]. In addition, we performed a gene-by-gene
analysis of the PCR data to compare agreement between
FF and FFPE tissues. Our methods have general applica-
tion to developing other multigene qRT-PCR tests for
Materials and Methods
tissue procurement and processing
All tissues and data used in this study were collected and
handled in compliance with federal and institutional
guidelines. Breast samples received in pathology were
flash-frozen in liquid nitrogen and stored at ?80 °C. We
procured samples at the University of North Carolina at
Chapel Hill, Thomas Jefferson University, University of
Chicago, and University of Utah. The 159 breast samples
included a 124-sample microarray training set and a
35-sample test set profiled by microarray and real-time
qRT-PCR (FF and FFPE). Total RNA from FF samples was
isolated using the RNeasy Midi Kit (Qiagen) and treated
on-column with DNase I to eliminate contaminating
DNA. The RNA was stored at ?80 °C until used for
We compared each FF sample in the test set to the
clinical FFPE tissue block. We used a hematoxylin and
eosin–stained slide to confirm the presence of ?50%
tumor and prepared 20-?m cuts with a microtome. Tissue
blocks were 1 to 5 years old (i.e., early-age FFPE). The
FFPE cut was deparaffinized in Hemo-De (Scientific
Safety Solvents) and washed with 100% ethanol. We
isolated total RNA by use of the High Pure RNA Paraffin
Kit (Roche Molecular Biochemicals). We followed manu-
facturer’s instructions for RNA extraction except that the
reagents were increased 2-fold for the 1st proteinase K
digestion. Samples were treated with TURBO DNA-free
(Ambion, #1906) and stored at ?80 °C until cDNA
first-strand cDNA synthesis
We performed cDNA synthesis for each sample in 40-?L
total volume reactions containing 600 ng total RNA.
Total RNA was first mixed with 2 ?L gene-specific
mixture containing 55 primers (each antisense primer
at 1 ?mol/L) and 2 ?L of 10 mmol/L dNTP Mix
(10 mmol/L each dATP, dGTP, dCTP, dTTP at pH 7).
Reagents were heated at 65 °C for 5 min in a PTC-100
Thermal Cycler (MJ Research) and briefly centrifuged. We
added the following reagents to each tube: 8 ?L of 5?
First-Strand Buffer [250 mmol/L Tris-HCl (pH 8.3 at room
temperature), 375 mmol/L KCl, 15 mmol/L MgCl2], 2 ?L
of 0.1 mol/L dithiothreitol, 2 ?L RNase Out (Invitrogen),
and 2 ?L SuperScript III polymerase (200 units/?L). The
reaction was thoroughly mixed by pipetting and incu-
bated at 55 °C for 45 min followed by 15 min at 70 °C for
enzyme inactivation. After cDNA synthesis, samples were
purified with the QIAquick PCR Purification Kit (Qiagen).
We adjusted the samples to a final concentration of
1.25 mg/L cDNA with Tris-EDTA (10 mmol/L Tris-HCl,
pH 8.0, 0.1 mmol/L EDTA).
primer design and optimization
We designed primers with Roche LightCycler Probe De-
sign Software 2.0. We obtained reference gene sequences
through National Center for Biotechnology Information
LocusLink and found optimal primer sites with the aid of
Evidence Viewer (http://www.ncbi.nlm.nih.gov). We se-
lected primer sets to avoid known insertions/deletions
and mismatches while including all isoforms possible.
Amplicons were limited to 60 to 100 bp in length because
of the degraded condition of the FFPE mRNA. When
possible, RNA-specific amplicons were localized between
exons spanning large introns (?1 kb). Finally, we used
National Center for Biotechnology Information BLAST to
verify gene target specificity of each primer set. Primer
sequences are presented in Table 1 in the Data Supple-
ment that accompanies the online version of this article at
ers were synthesized by Operon, resuspended in Tris-
Mullins et al.: Profiling Breast Cancer from Fresh and FFPE Tissues
EDTA to a final concentration of 60 ?mol/L, and stored at
?80 °C. We assessed each new FFPE primer set for
performance through qRT-PCR runs with 3 serial 10-fold
dilutions of reference cDNA in duplicate and 2 no-
template control reactions. Primers were verified for use
when they fulfilled the following criteria: (a) target cross-
ing point ?30 in 10 ng reference cDNA; (b) PCR efficiency
?1.75; (c) no primer-dimers in presence of template as
determined through postamplification melting curve
analysis; and (d) no primer-dimers in negative template
control before cycle 40.
We carried out PCR amplification on the Roche LightCy-
cler 2.0. Each reaction contained 2 ?L cDNA (2.5 ng) and
18 ?L PCR master mix with the following final concen-
trations of reagents: 1 unit Platinum Taq, 50 mmol/L
Tris-HCl (pH 9.1), 1.6 mmol/L (NH4)2SO4, 400 mg/L
BSA, 4 mmol/L MgCl2, 0.2 mmol/L dATP, 0.2 mmol/L
dCTP, 0.2 mmol/L dGTP, 0.6 mmol/L dUTP, 1:40 000
dilution of SYBR Green I dye (Molecular Probes), and 0.4
?mol/L of both forward and reverse primers for the
selected target. The PCR was done with an initial dena-
turation step at 94 °C for 90 s and then 50 cycles of
denaturation (94 °C, 3 s), annealing (58 °C, 6 s), and
extension (72 °C, 6 s). Fluorescence acquisition (530 nm)
was taken once each cycle at the end of the extension
phase. After PCR, we initiated a postamplification melt-
ing curve program by heating to 94 °C for 15 s, cooling to
58 °C for 15 s, and slowly increasing the temperature
(0.1 °C/s) to 95 °Cwhile
Each PCR run contained a no-template control, a
calibrator reference in triplicate, and each sample in
duplicate. The calibrator reference sample comprised 3
breast cancer cell lines (MCF7, SKBR3, and ME16C) and
Stratagene Universal Human Reference RNA (Stratagene)
represented in equal parts. The target crossing point,
defined as the cycle at which the fluorescence of a sample
rises above the background, was automatically calculated
for each reaction by Roche LightCycler Software 4.0. For
relative quantification, we imported an external efficiency
curve (Eff ? 1.89) and set the calibrator at 10 ng for each
gene. To correct for differences in sample quality and
cDNA input, we adjusted copy numbers to the arithmetic
mean of 5 housekeeper genes [ACTB8(?-actin), PSMC4
(proteasome 26S subunit, ATPase, 4), PUM1 (pumilio
homolog 1, Drosophila), MRPL19 (mitochondrial ribo-
somal protein L19), and SF3A1 (splicing factor 3a, subunit
1, 120 kDa)]. Values from replicate samples were aver-
aged, and data were log2transformed. Raw copy numbers
(i.e., not housekeeper-adjusted) for all genes analyzed are
provided in Table 2 in the online Data Supplement.
We analyzed all samples by use of DNA microarray
(Agilent Human A1, Agilent Human A2, and Agilent
custom oligonucleotide microarrays). We labeled and
hybridized RNA for microarray analysis with the Agilent
low RNA input linear amplification reagent set (http://
as described in Hu et al. (14). Only RNA from FF tissue
was used for microarray experiments. Each sample was
assayed vs a common reference that was Stratagene’s
Human Universal Reference total RNA enriched with
equal amounts of RNA from the MCF7 and ME16C cell
lines. Microarray hybridizations were carried out on
Agilent Human oligonucleotide microarrays by using
2 ?g cyanine 3 (Cy3)-labeled reference sample and 2 ?g
Cy5-labeled experimental sample.
We scanned all microarrays by use of an Axon Scanner
4000B (Axon Instruments). We analyzed the image files
with GenePix Pro 4.1 (Axon Instruments) and uploaded
them into the UNC Microarray Database at the University
of North Carolina at Chapel Hill (https://genome.
unc.edu/), where a Lowess normalization procedure was
performed to adjust the Cy3 and Cy5 channels (15).
Microarray data for this study have been submitted to
Gene Expression Omnibus (http://www.ncbi.nlm.nih.
gov/geo/) under accession no. GSE6130.
clinical immunohistochemistry and pcr
At the time of diagnosis, samples were scored for protein
expression of ER, progesterone receptor, and HER2/neu
by use of standard operating procedures established at
each institution. Nuclei staining ?10% positive were
considered positive for ER and progesterone receptor.
Staining and scoring criteria for HER2 were according to
HercepTest™ (Dako). For quantitative PCR to determine
DNA copy number of the ERBB2 (v-erb-b2 erythroblastic
leukemia viral oncogen homolog 2) gene, we used a
clinical assay from ARUP Laboratories (catalog no.
selecting genes for real-time qRT-PCR
The real-time qRT-PCR assay consisted of 5 housekeeper
genes (16), 5 proliferation genes for risk stratification of
the luminal (ER-positive) tumors (8), and 40 intrinsic
genes important for distinguishing biological subtypes of
breast cancer (7). We statistically selected the minimal 40
intrinsic classifiers from a larger 1393 intrinsic gene set
previously reported in Hu et al. (7) by use of minimiza-
tion methods described by Dudoit and Fridlyand (17).
Briefly, we used a semisupervised classification method
in which samples are hierarchically clustered and as-
signed subtypes on the basis of the sample-associated
dendrogram (7, 11–13). We designated samples luminal,
8Human genes: ACTB, ?-actin; PSMC4, proteasome 26S subunit, ATPase,
4; PUM1, pumilio homolog 1, Drosophila; MRPL19, mitochondrial ribosomal
protein L19; SF3A1, splicing factor 3a, subunit 1, 120 kDa; ERBB2, v-erb-b2
erythroblastic leukemia viral oncogen homolog 2; ESR1, estrogen receptor 1;
IGBP1, immunoglobulin binding protein 1.
Clinical Chemistry 53, No. 7, 2007
HER2?/ER?, basal-like, or normal-like. We identified the
best class distinguishers according to the ratio of between-
group to within-group sums of squares. We performed a
10-fold cross-validation by using a nearest centroid clas-
sifier and testing overlapping gene sets of varying sizes.
We selected the smallest gene set that provided the
highest class prediction accuracy compared with the
classifications made by the complete microarray-based
intrinsic gene set.
assessing qRT-PCR agreement between ff and
We analyzed 35 matched FF and FFPE samples (70
samples total) by qRT-PCR using the same primer sets.
Agreement in the quantitative data was determined using
diagonal bias (m), diagonal spread (d), diagonal SD (dsd),
diagonal correlation (rd), and concordance correlation
In diagonal bias, a best-fitting line parallel to the
diagonal (slope ? 1) is made from a plot of the qRT-PCR
data (FF vs FFPE). Numerically, if (xi, yi), i ? 1, . . . ,
n denote the measurement pairs, then the best-fitting
line parallel to the diagonal is given by the following
y ? x ? y ? ? x ?
where x ? and y ? denote the sample means of the x and y
Then we calculate diagonal bias as:
m ?y ? ? x ?
The dsd was calculated as follows:
i ? 1
n ? 1
where diis the distance to the best fit line calculated as
di???yi? xi? ? ?y ? ? x ???
Let d represent the mean deviation from the best fit line
i ? 1
Diagonal correlation was used to determine the spread of
points around the diagonal line:
Var?X? ? Var?Y?
This method does not provide information about the
extent of deviation but allows measurements with differ-
ent units to be compared. Furthermore, if we let ? denote
the correlation coefficient and ?Xand ?Ythe respective
That is, the diagonal correlation penalizes the correlation
coefficient if there is a scale shift ?X? ?Y. We measured
the combined effect of the bias and scale shift by use of the
ccc proposed by Lin (18):
Var?X? ? Var?Y? ? ?Y?? X??2
assessing agreement between microarray and
qRT-PCR for classification
A breast cancer subtype predictor was developed in
ssp.pl) (13, 19). PAM and SSP are both nearest centroid
classifiers that use prototype samples in the training set
to develop centroids. Test samples are then assigned
the class of the nearest centroid as measured by Euclid-
ean distance. The major difference between the meth-
ods lies in how the centroids are constructed. SSP uses
a simple unstandardized centroid created from a subset
of genes identified during cross-validation, whereas
PAM creates standardized and shrunken or “de-
noised” centroids. The amount of shrinkage is deter-
mined in cross-validation. We used a training set with
prototype samples for luminal (64 samples), HER2?/
ER?(23 samples), basal-like (28 samples), and normal-
like (9 samples) subtypes. We classified an independent
test set (35 matched FF and FFPE samples) by use of the
large (1393 genes) and minimized (40 genes) versions of
the microarray intrinsic gene set (see Selecting Genes for
The qRT-PCR data from the test set were merged with
the microarray data of the training set before classification
by use of distance weighted discrimination (DWD), a
method that adjusts for systematic biases between differ-
ent platforms (20). The gold standard for classification of
the training and test samples was based on FF tissue RNA
and the classifications obtained when performing hierar-
chical clustering analysis using the 1393 gene intrinsic
gene set from microarray data.
assessment of qRT-PCR primer set performance
by comparing agreement between ff and ffpe
We evaluated the dataset of 35 matched FF and FFPE
tissues (70 samples) for 50 genes with the same PCR
conditions. Agreement between FF and FFPE tissues was
assessed for diagonal bias (m), diagonal correlation (rd)
diagonal SD (dsd), and ccc. Fig. 1 shows an agreement
Mullins et al.: Profiling Breast Cancer from Fresh and FFPE Tissues
plot for the relative quantification of the ER gene [ESR1
(estrogen receptor 1)] between FF and FFPE tissues. This
is a typical plot that was used to assess each classifier
gene. In the case of ESR1, there is a particularly large
dynamic range, and tumors are clearly divided into 2
populations. This separation highly associates with im-
munohistochemistry (IHC) status for ER, even without
normalization (see Fig. 1 in the online Data Supplement).
We have previously shown that ESR1 alone measured
from FF tissue has very high sensitivity and specificity by
use of ER status by IHC as the gold standard (8).
For each gene, the agreement between FF and FFPE
was analyzed using the raw data, housekeeper-normal-
ized data, and DWD-adjusted normalized data. Scatter
plots are provided in Fig. 2 in the online Data Supple-
ment, and values are presented in Table 3 in the online
Data Supplement. The line graphs in Fig. 2 show the
effects at each step of data processing. The raw (pre-
normalized) data show a negative bias for all genes, likely
due to lower RNA quality in the FFPE tissue (Fig. 2A).
Much of the bias was corrected by normalization to the
housekeeper genes and DWD adjustment. As expected,
DWD had a significant effect on bias (m) but did not affect
other measurements of agreement (Fig. 2, B and C).
Genes with the highest diagonal correlation between
FF and FFPE usually had the largest dynamic range in
expression (e.g., ESR1, TFF3, COX6C, and FBP1). House-
keeper genes and other genes with low variability in
expression [IGBP1 (immunoglobulin binding protein 1)]
had the lowest diagonal correlation since they form more
of a cloud than a line around the diagonal. The house-
keeper genes all had high agreement in terms of having
low variability in expression across samples in the FF and
The ccc considers both bias and scale shift when
determining agreement. The median ccc between FF and
FFPE for the raw data of the 45 genes (housekeepers
excluded) was 0.28. Normalization to housekeepers raised
the ccc median to 0.48, and adjusting with DWD brought
the median to 0.61. A comparison of the ccc value to the
ratio of the dsd over the dynamic range identified many
of the same primer sets as good (or poor) performers from
the FFPE-derived samples.
breast cancer subtype classification of test
set by use of pam and ssp
Hierarchical clustering of the 124-sample training set by
use of the minimized intrinsic gene set identified from Hu
et al. (7) shows 4 distinct classes representing luminal,
HER2?/ER?, basal-like, and normal-like (see Fig. 5 in the
online Data Supplement). We developed centroid classi-
fiers from the microarray expression data by use of PAM
and SSP (7, 13, 19). We made class predictions on the test
set by use of microarray (large and minimized “intrinsic”
Log−expression in the FF sample
e l p
h t n i n
o i s s
e r p
Fig. 1. Agreement plot between FF and FFPE for the ER gene.
We analyzed gene expression in 35 breast tumors procured as FF and FFPE
tissues, using the same conditions on the matched samples for reverse
transcription and PCR. A best-fit line (dashed) is compared with the ideal line
(solid), and the distance between them is the diagonal bias (m). The distance of
each point to the best-fit line is represented as di.
Fig. 2. Line graphs showing the effects of data processing across different methods of assessing agreement between FF and FFPE tissues.
The raw data (Raw), housekeeper-normalized data (Norm), and DWD-adjusted normalized data are shown for diagonal bias (A), concordance correlation (B), and diagonal
SD (C). The raw (prenormalized) data show a negative bias for all genes, likely because of lower RNA quality in the FFPE tissue. Much of the bias is corrected by
normalization to the housekeeper genes and DWD adjustment. Although DWD had a significant effect on bias, it did not affect the other measurements of agreement.
Clinical Chemistry 53, No. 7, 2007
sets) and qRT-PCR data (see Table 4 in the online Data
Supplement). Each individual microarray (large and min-
imized) and PCR dataset was DWD merged with the
training set before subtype class prediction.
Agreement in classification between large and minimized mi-
croarray gene sets. Of 35 samples, 33 (94%) were classified
the same between PAM and SSP when using the large
intrinsic microarray dataset for classification. In both
discrepant cases, IHC data agreed with the PAM classifi-
cation. There was the same agreement (94%) when per-
forming the analysis with the minimized version of the
microarray data. Interestingly, there was 1 sample that
was called HER2?/ER?by both PAM and SSP when
using the large microarray dataset but called basal-like by
both methods when using the minimized microarray
dataset. Additional analysis of this sample by quantitative
PCR showed no DNA amplification of HER2/ERBB2
Agreement in classification between FF and FFPE. By qRT-
PCR, there was 97% (34 of 35) concordance between FF
and FFPE using PAM and 91% (32 of 35) concordance
using SSP. There was 94% (33 of 35) concordance between
the diagnostic algorithms from FF tissue and complete
agreement in classification from FFPE tissue. Because the
FFPE samples were obtained from the clinical block, it is
likely that there was a higher tumor percentage in those
samples than in the matched FF sample, which could
affect the agreement. Indeed, 2 of the 3 discrepancies in
classification made by SSP occurred when the FF tissue
sample was classified as normal-like (microarray and
PCR) and the FFPE sample was classified as luminal
(PCR). These samples were ER positive by IHC and likely
luminal. The only discrepancy in PAM was in a sample
classified as normal-like from FF tissue and luminal from
Overall concordance across methods. Overall, PAM diag-
nosed 33 of 35 samples (94%) the same across microarray
and qRT-PCR, whereas SSP diagnosed 30 of 35 samples
(86%) the same across platforms and procurement meth-
ods. Discrepancies were of several types, including lumi-
nal tumors classified as normal-like, HER2?/ER?tumors
classified as luminal, and basal-like tumors classified as
Translating large-scale microarray studies into clinical
tests requires several critical steps, including gene set
minimization, cross-platform validation, and develop-
ment of robust classification algorithms.
Several centroid-based algorithms have been devel-
oped for predicting sample subtypes from microarray
data (13, 17, 19, 21). Programs that are simple and intui-
tive in design, such as linear discriminant analysis, are
preferred owing to their transparency (19). PAM adds a
feature selection to linear discriminant analysis in which
t-statistics are computed for each gene to determine its
contribution to the assigned subtypes (13). The t-statistics
for each gene are then ranked, and the gene set can be
minimized by selecting the top genes that provide a
minimal false discovery value. The main difference be-
tween SSP and PAM is that PAM shrinks the centroid
toward the overall mean for classification. Here we di-
rectly compared PAM with SSP by use of the large
microarray dataset applied in Hu et al. (7) and also a
minimized version. On this dataset, PAM performed
slightly better than SSP for classification across gene sets
and conditions, although both methods performed well.
Determining agreement between methods is a complex
issue. Cronin et al. (22) used Pearson correlation to show
that the genes with the highest correlation in microarray
maintained their association with qRT-PCR. They used
short amplicons and control housekeeper genes in the
qRT-PCR assay to correct biases between FF and FFPE
tissues. Although correlation provides information about
the linearity and slope (positive or negative correlation) of
the data, it does not indicate the amount of bias, scale
shift, or data spread. These additional measurements are
helpful in determining whether the discrepancies in the
data can be compensated for experimentally (e.g., house-
keeper genes) or by use of software algorithms.
We found that the most useful analyses for assessing
PCR primer set performance across FF and FFPE tissues
were the ccc, the diagonal SD, and the dynamic range.
Genes with a large dynamic range often had high corre-
lation and were good classifiers across conditions, even
with relatively large diagonal SDs. Although genes with a
small dynamic range can be good classifiers, the measure-
ment may not be as reproducible if there is a large amount
of variation. Thus, we found that the best assessment of a
classifier was using a ratio of the diagonal SD to the
Translating an assay from microarray to qRT-PCR
provides a 2nd level of gene validation and allows the test
to be used on archived FFPE tissue blocks from clinical
trials or on samples submitted for routine diagnostics
(2, 22). This study demonstrates that a qRT-PCR assay for
the biological subtypes of breast cancer can be used with
a centroid-based classifier to predict tumor type from
FFPE tissues. The assay has application in the clinical
laboratory for prognosis in breast cancer.
Grant/funding support: This work was supported by Na-
tional Cancer Institute Grants R33-CA97769-01 (to P.S.B.) and
P50-CA58223-09A1 (to C.M.P.).
Financial disclosures: None declared.
Acknowledgments: We appreciate the help of the core facil-
ities for tissue procurement at the participating institutions.
We thank Carlynn Willmore-Payne and Joseph A. Holden for
their technical expertise.
Mullins et al.: Profiling Breast Cancer from Fresh and FFPE Tissues
References Download full-text
1. Fan C, Oh DS, Wessels L, Weigelt B, Nuyten DS, Nobel AB, et al.
Concordance among gene-expression-based predictors for breast
cancer. N Engl J Med 2006;355:560–9.
2. Paik S, Shak S, Tang G, Kim C, Baker J, Cronin M, et al. A
multigene assay to predict recurrence of tamoxifen-treated, node-
negative breast cancer. N Engl J Med 2004;351:2817–26.
3. Perou CM, Sorlie T, Eisen MB, van de Rijn M, Jeffrey SS, Rees CA,
et al. Molecular portraits of human breast tumours. Nature
4. van’t Veer LJ, Dai H, van de Vijver MJ, He YD, Hart AA, Mao M, et
al. Gene expression profiling predicts clinical outcome of breast
cancer. Nature 2002;415:530–6.
5. van de Vijver MJ, He YD, van’t Veer LJ, Dai H, Hart AA, Voskuil DW,
et al. A gene-expression signature as a predictor of survival in
breast cancer. N Engl J Med 2002;347:1999–2009.
6. Sorlie T, Perou CM, Tibshirani R, Aas T, Geisler S, Johnsen H, et
al. Gene expression patterns of breast carcinomas distinguish
tumor subclasses with clinical implications. Proc Natl Acad Sci
U S A 2001;98:10869–74.
7. Hu Z, Fan C, Oh DS, Marron JS, He X, Qaqish BF, et al. The
molecular portraits of breast tumors are conserved across mi-
croarray platforms. BMC Genomics 2006;7:96.
8. Perreard L, Fan C, Quackenbush JF, Mullins M, Gauthier NP,
Nelson E, et al. Classification and risk stratification of invasive
breast carcinomas using a real-time quantitative RT-PCR assay.
Breast Cancer Res 2006;8:R23.
9. Sorlie T, Tibshirani R, Parker J, Hastie T, Marron JS, Nobel A, et al.
Repeated observation of breast tumor subtypes in independent
gene expression data sets. Proc Natl Acad Sci U S A 2003;100:
10. Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis
and display of genome-wide expression patterns. Proc Natl Acad
Sci U S A 1998;95:14863–8.
11. Bair E, Tibshirani R. Semi-supervised methods to predict patient
survival from gene expression data. PLoS Biol 2004;2:E108.
12. Bullinger L, Dohner K, Bair E, Frohling S, Schlenk RF, Tibshirani R,
et al. Use of gene-expression profiling to identify prognostic
subclasses in adult acute myeloid leukemia. N Engl J Med
13. Tibshirani R, Hastie T, Narasimhan B, Chu G. Diagnosis of
multiple cancer types by shrunken centroids of gene expression.
Proc Natl Acad Sci U S A 2002;99:6567–72.
14. Hu Z, Troester M, Perou CM. High reproducibility using sodium
hydroxide-stripped long oligonucleotide DNA microarrays. Biotech-
15. Yang YH, Dudoit S, Luu P, Lin DM, Peng V, Ngai J, et al.
Normalization for cDNA microarray data: a robust composite
method addressing single and multiple slide systematic variation.
Nucleic Acids Res 2002;30:e15.
16. Szabo A, Perou CM, Karaca M, Perreard L, Quackenbush JF,
Bernard PS. Statistical modeling for selecting housekeeper
genes. Genome Biol 2004;5:R59.
17. Dudoit S, Fridlyand J. A prediction-based resampling method for
estimating the number of clusters in a dataset. Genome Biol
18. Lin LI. A concordance correlation coefficient to evaluate reproduc-
ibility. Biometrics 1989;45:255–68.
19. Dudoit S, Fridlyand J, Speed TP. Comparison of discrimination
methods for the classification of tumors using gene expression
data. J Am Stat Assoc 2002;97:77–87.
20. Benito M, Parker J, Du Q, Wu J, Xiang D, Perou CM, et al.
Adjustment of systematic microarray data biases. Bioinformatics
21. Dabney AR. Classification of microarrays to nearest centroids.
22. Cronin M, Pho M, Dutta D, Stephans JC, Shak S, Kiefer MC, et al.
Measurement of gene expression in archival paraffin-embedded
tissues: development and performance of a 92-gene reverse
transcriptase-polymerase chain reaction assay. Am J Pathol
Clinical Chemistry 53, No. 7, 2007