A strategy for full interrogation of prognostic gene expression patterns: exploring the biology of diffuse large B cell lymphoma.

Abstract

Gene expression profiling yields quantitative data on gene expression used to create prognostic models that accurately predict patient outcome in diffuse large B cell lymphoma (DLBCL). Often, data are analyzed with genes classified by whether they fall above or below the median expression level. We sought to determine whether examining multiple cut-points might be a more powerful technique to investigate the association of gene expression with outcome.
We explored gene expression profiling data using variable cut-point analysis for 36 genes with reported prognostic value in DLBCL. We plotted two-group survival logrank test statistics against corresponding cut-points of the gene expression levels and smooth estimates of the hazard ratio of death versus gene expression levels. To facilitate comparisons we also standardized the expression of each of the genes by the fraction of patients that would be identified by any cut-point. A multiple comparison adjusted permutation p-value identified 3 different patterns of significance: 1) genes with significant cut-point points below the median, whose loss is associated with poor outcome (e.g. HLA-DR); 2) genes with significant cut-points above the median, whose over-expression is associated with poor outcome (e.g. CCND2); and 3) genes with significant cut-points on either side of the median, (e.g. extracellular molecules such as FN1).
Variable cut-point analysis with permutation p-value calculation can be used to identify significant genes that would not otherwise be identified with median cut-points and may suggest biological patterns of gene effects.

Full text

Page 1

A Strategy for Full Interrogation of Prognostic Gene
Expression Patterns: Exploring the Biology of Diffuse
Large B Cell Lymphoma
Lisa M. Rimsza1*, Joseph M. Unger2, Margaret E. Tome1, Michael L. LeBlanc2
1Department of Pathology, University of Arizona, Tucson, Arizona, United States of America, 2Division of Public Health Sciences, Fred Hutchinson Cancer Research
Center, Seattle, Washington, United States of America
Abstract
Background: Gene expression profiling yields quantitative data on gene expression used to create prognostic models that
accurately predict patient outcome in diffuse large B cell lymphoma (DLBCL). Often, data are analyzed with genes classified
by whether they fall above or below the median expression level. We sought to determine whether examining multiple cut-
points might be a more powerful technique to investigate the association of gene expression with outcome.
Methodology/Principal Findings: We explored gene expression profiling data using variable cut-point analysis for 36 genes
with reported prognostic value in DLBCL. We plotted two-group survival logrank test statistics against corresponding cut-
points of the gene expression levels and smooth estimates of the hazard ratio of death versus gene expression levels. To
facilitate comparisons we also standardized the expression of each of the genes by the fraction of patients that would be
identified by any cut-point. A multiple comparison adjusted permutation p-value identified 3 different patterns of
significance: 1) genes with significant cut-point points below the median, whose loss is associated with poor outcome (e.g.
HLA-DR); 2) genes with significant cut-points above the median, whose over-expression is associated with poor outcome
(e.g. CCND2); and 3) genes with significant cut-points on either side of the median, (e.g. extracellular molecules such as
FN1).
Conclusions/Significance: Variable cut-point analysis with permutation p-value calculation can be used to identify
significant genes that would not otherwise be identified with median cut-points and may suggest biological patterns of
gene effects.
Citation: Rimsza LM, Unger JM, Tome ME, LeBlanc ML (2011) A Strategy for Full Interrogation of Prognostic Gene Expression Patterns: Exploring the Biology of
Diffuse Large B Cell Lymphoma. PLoS ONE 6(8): e22267. doi:10.1371/journal.pone.0022267
Editor: Luwen Zhang, University of Nebraska – Lincoln, United States of America
Received February 2, 2011; Accepted June 21, 2011; Published August 4, 2011
Copyright: ? 2011 Rimsza et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: Funding for this work was provided by NIH grant R01 CA90998 (PI LeBlanc) and American Cancer Society grant RSG0605501LIB (PI Rimsza). The funders
had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have read the journal’s policy and have the following conflicts: the reagents used in this project were donated free of charge
from High Throughput Genomics. This does not alter the authors’ adherence to all the PLoS ONE policies on sharing data and materials.
* E-mail: lrimsza@email.arizona.edu
Introduction
Diffuse large B cell lymphoma (DLBCL) is an aggressive disease
witha variable outcome. In order to quantify patientrisk, numerous
biomarkers have been identified that can be detected with a variety
of methods. We recently described the use of a quantitative nuclease
protection assay (qNPA) to measure gene expression levels from
formalin fixed, paraffin embedded (FFPE) tissue blocks of DLBCL
[1]. In a subsequent study of CHOP and rituximab-CHOP (R-
CHOP) treated DLBCL cases, qNPA results for many genes were
significantly associated with overall survival [2]. Initial data analysis
was performed by categorizing patients into expression levels above
and below the median level of expression. The best selected 2-
variable model predicting overall survival in DLBCL was the
combinationofthemajorhistocompatibility(MHC)classIIantigen,
HLA-DRB, and the cell cycle associated gene, MYC. In agreement
with the literature, these results implicated lack of immunosurveil-
lance and increased cell proliferation as important features that
characterize the most aggressive B cell lymphomas [3–8].
We then further explored the relationship between expression
levels and survival for these two genes. We plotted the score test
statistic (logrank test statistic) from Cox regression for the
association of gene expression quantile and survival, where gene
expression was converted to a binary variable with cut-points
defined along a continuous spectrum of low to high expression [9].
For HLA-DRB, the highest logrank statistic chi-square value
indicating the most significant cut-point of gene expression was at
the 20thpercentile. Many other cut-points were also significant [2].
This observation was in keeping with previous data demonstrating
that there is a smooth non-linear association of MHC Class II
expression levels as related to patient risk, with small incremental
decreases in expression corresponding to increases in the hazard
ratio of death with a sharp increase in hazard at lower levels of
expression [10]. For MYC the most significant cut-point was at the
80thpercentile of expression. While the 80thpercentile was the
optimal cut-point, there was a wide range of cut-point values that
were also nominally significant [2]. This has biological implica-
tions for MYC, suggesting that incremental increases in MYC
PLoS ONE | www.plosone.org1August 2011 | Volume 6 | Issue 8 | e22267

Page 2

expression portend a worse prognosis with the sharpest increase in
risk at higher levels of expression.
In the current study, we went on to perform this variable cut-
point analysis on 36 genes to determine whether we could identify
genes that might have significant cut-points other than the median
and how that might be a factor in the reported discrepancies in
prognostic value of genes by different investigators and techniques.
Materials and Methods
Ethics Statement
The project was approved by the University of Arizona
Institutional Review Board (IRB) according to the principles
expressed in the Declaration of Helsinki. The University of
Arizona IRB specifically waived the need for informed consent for
this project.
Patient groups and mRNA data
We used the mRNA levels determined using qNPA (ArrayPla-
teRAssay, High Throughput Genomics, Tucson, AZ) as described
previously [1,2]. Briefly, unstained FFPE sections of 209 DLBCL,
previously treated with CHOP-like regimens (N=93) or R-CHOP
(N=116) were subjected to the qNPA procedure. This process
begins with cell lysis followed by exposure to specifically designed
probe sets that bind to the target mRNA of interest. S1 nuclease is
used to degrade all single stranded RNA and the surviving probes
are identified by binding to linker probes and detection probes on
the ArrayPlateRfollowed by chemiluminescence and imaging. The
study set of cases included FFPE blocks from cases of de novo,
previously untreated DLBCL which had also been a part of 2
larger case series using gene expression profiling of snap frozen
biopsies from patients treated with CHOP or R-CHOP and then
later in a study of ArrayPlateRgene expression technique on the
corresponding FFPE blocks [2,11,12]. The customized ArrayPla-
teRassay had been designed to assess the expression levels of 36
prognostic genes identified in DLBCL by different research groups
and published in the literature. A list of the genes, their function (if
known), and the reference from which they were chosen are listed
in Table 1. All research was conducted under an IRB (human
subjects committee) approved protocol from the University of
Arizona. We obtained expression measurements with $95%
success on all but 3 genes, and with #80% success on only one
gene (HTR2B).
Variable cut-point and smooth hazard regression analysis
Variable cut-point (or split-point) analysis was performed on all
36 genes in order to discriminate between groups of patients with
the most significant differences in overall survival. This statistical
technique calculates the score test statistic from a Cox model
(analogous to the logrank statistic) at a continuous spectrum of cut-
points on the gene expression variable [13]. (Typically the
maximum statistic is often used to define best split of patients.)
In the plot (Figure 1), the vertical axis corresponds to the score
statistic on the standard normal scale. To adjust for the evaluation
of the large number of cut-point models, permutation sampling is
used to control the family-wise type 1 error for each gene. The
permutation p-values presented in the cut-point plots are based on
1000 samples, and the horizontal line on each plot corresponds to
the 90thpercentile of the sampled permutation distribution of the
maximum test statistic. Therefore, a cut-point statistical test
reaching above the horizontal line has a permutation adjusted p-
value of ,0.10 [9]. Note that the 90thpercentile horizontal lines
for the genes are at approximately 2.5 for most gene expression
variables; if there were no adjustment for multiple comparisons, a
value of 1.64 would correspond to a p-value of 0.1. Without this
adjustment there would be the tendency to falsely believe
moderately large test statistics correspond to real association,
when observed associations could simply be due to the large
number of cut-point models that have been investigated. In
addition, to control statistical variability, a minimum possible
subgroup size of 10% of total patients was set for each analysis.
Since our previous test of panel-wide interaction between the
CHOP and R-CHOP groups had shown no significance, we
combined the 2 data sets for purposes of the current analysis [2].
However, the cut-point technique adjusted for treatment group
(CHOP versus R-CHOP) as a main effect in the relative risk
regression model, since R-CHOP is well known to be associated
with improved survival. The cut-point technique also allows for
more general adjustment of an existing prognostic model to assess
the statistical significance of the addition of a new gene expression
variable and cut-point. Analyses presented are based on overall
survival, where overall survival is defined as the time from study
registration until death. Patients without an observed death time
are censored at the last known time under follow-up.
While the cut-point evaluation allows the assessment of
statistical significance of multiple partitions of a gene expression
variable, it does not directly lead to an estimate of the underlying
regression function representing how gene expression is associated
with survival. Therefore, we also used hazard regression modeling
(based on a B-spline basis) to calculate smooth estimates of the
hazard function for each gene [14]. An alternative estimation
strategy for smooth hazard regression functions is by local
likelihood [15]. In addition, we transformed each gene expression
variable to be approximately uniformly distributed to make the
analysis consistent with the cut-point analysis, which only depends
on the rank of the gene expression variables. As done for the cut-
point analysis, we adjusted for the two treatment groups (CHOP
versus R-CHOP) via main effect in the relative risk regression
model.
While our combination of cut-point analysis and smooth hazard
regression modeling is useful for interpreting individual effects of a
small set of continuous biological measurements, such as gene
expression with censored survival patient outcome, there are other
related statistical methods available for multivariable modeling
and subgroup analysis. For instance, with respect to smooth
regression modeling, there has been considerable study of
generalized additive models, which consist of additive combina-
tions of smooth univariate regression functions. For deriving
subgroups in the context of many variables, the cut-point methods
we proposed can be utilized recursively to cut-up or partition the
data on multiple covariates to construct regression trees [16].
There is an extenstive discussion of other statistical or machine
learning algorithms in Hastie et al. [17]. Due to the complexity of
some of the multivariable models, their use is often better applied
to patient prognostic predictions or subgroup stratification rather
than probing the interpretation and clinical impact of individual
gene expression measurements. In addition, alternatives to the
smooth hazard regression models based on locally estimated
quantiles of the survival distribution can be helpful for exploring
gene effects [18]; however, we chose the hazard based methods for
our exploration of DLBCL gene expression data given the
relatively modest sample size. In addition, hazard regression
methods tend to achieve better variance control in such cases.
Results
We first generated a series of graphs for each of the 13 genes
with significant logrank statistic (Z-value) (Fig. 1). The different
DLBCL Gene Expression Variable Cut-Point Analysis
PLoS ONE | www.plosone.org2 August 2011 | Volume 6 | Issue 8 | e22267

Page 3

Table 1. Prognostic genes tested1.
Name in original referenceAlternative namesqNPA nameReferenceFunction
BCL-6BCL6*Rosenwald
1/Lossos 6
Transcriptional repressor that controls
germinal center formation [26,27]
IMAGE 1334260 centerin/GCET1 (germinal center
B-cell expressed transcript 1)
SERPINA9* Rosenwald 2Serpin (serine protease inhibitor) [28]
IMAGE 814622GCET2 (germinal center B-cell
expressed transcript 2)/HGAL
(human germinal center-
associated lymphoma)
GCET2 Rosenwald 3Membrane-associated protein with a
putative role in signal transduction [29];
myosin-interacting protein that is a
putative inhibitor of cell migration [30]
HLA-DPaHLA-DPA1 Rosenwald 4Antigen presentation [31]
HLA-DQaHLA-DQA1Rosenwald 5 Antigen presentation [31]
HLA-DRa HLA-DRARosenwald 6Antigen presentation [31]
HLA-DRbHLA-DRB* Rosenwald 7Antigen presentation [31]
alpha-actininACTN1*Rosenwald 8 Non-muscle a-actinin isoform involved in
bundling actin filaments and attaching
them to focal adhesions; important for
cell motility [32]
collagen type III alpha1COL3A1* Rosenwald 9Type III fibrillar collagen; part of the
extracellular matrix in lymph nodes
[33,34]
connective tissue growth factorCTGF*Rosenwald 10Heparin and integrin binding protein
involved in extracellular matrix
remodeling [35]
fibronectin FN1* Rosenwald 11/Lossos 5 Extracellular integrin ligand involved in
cell adhesion [36]
KIAA0233 Piezo1FAM38A Rosenwald 12 Multipass transmembrane protein
involved in mechanotransduction and
regulation of integrin activation [37,38]
urokinase plasminogen
activator
Urokinase/uPA PLAU*Rosenwald 13 Serine protease that activates
plasminogen which results in extracellular
matrix degradation [39]
C-MYCMYC*Rosenwald 14 Transcription factor that controls
proliferation, growth, metabolism,
microRNAs and apoptosis [40]
E21G3 NucleosteminNS C20orf155Rosenwald 15 Nucleolar GTP-binding protein that
regulates cell cycle by regulating p53 and
maintains nucleolar structure [41,42]
NPM3 Nucleophosmin 3NPM3Rosenwald 16Nucleolar protein that inhibits ribosome
biogenesis and histone assembly and
enhances transcription [43,44]
BMP6Bone morphogenetic protein-6BMP6Rosenwald 17 Cytokine that regulates B-cell
lymphopoiesis [45]
LMO2LIM domain only-2 LMO2 Lossos1Transcription factor that regulates
erythropoiesis and angiogenesis [46,47]
BCL2BCL2Lossos 2Membrane bound protein that prevents
apoptosis [48]
SCYA3 MIP-1a(macrophage inflammatory
protein-1)
CCL3Lossos 3Chemokine that recruits cells to sites of
inflammation and inhibits hematopoietic
stem cell proliferation [49]
CCND2Cyclin D2CCND2* Lossos 4 Activator of cell cycle progression [50]
DRP2-dystrophin related
protein 2
DRP2 Shipp 1One of a class of structural proteins that
maintains membrane–associated
complexes at the points of intercellular
contact [51]
PRKACB-protein kinase
C beta 1
PKCbIIPRKCB1*Shipp 2 Serine/threonine-specific kinase that
plays a role in B-cell receptor signaling
and B-cell development [52]
H731-nuclear antigenProgrammed Cell Death 4 PDCD4*Shipp 3Protein translation initiation factor
inhibitor that is a putative context-
specific tumor suppressor [53,54]
DLBCL Gene Expression Variable Cut-Point Analysis
PLoS ONE | www.plosone.org3 August 2011 | Volume 6 | Issue 8 | e22267

Page 4

cut-point values assessed for each gene are represented by the dots
along the connected line of chi-square values. The solid horizontal
line represents the 90thpercentile of the permutation distribution
of the maximal score statistics under the assumption the gene is not
associated with patient outcome (i.e., under the null hypothesis).
Given the exploratory nature of this analysis, all values with a
significance cut-off above the 90thpercentile line (type 1 error of
0.10) were considered significant. An overall p-value adjusted for
the permutation analysis is presented on each of the panels. Note
that only score statistics for cut-points that generate subgroups of
patients with $10% of the sample size were considered, since
smaller groups would probably not be considered useful clinically.
We think it is useful to plot the cut-point analysis against the
quantile of the gene expression distribution so that one could just
read what fraction of the sample is above or below the cut-point.
Thirteen out of the 36 genes (36%) had at least 1 significant cut-
point at p,0.10, including SERPINA9, HLA-DRB, ACTN1,
COL3A, CTGF, FN1, PLAU, MYC, BCL6, CCND2, PRKCB1,
PDCD4, and TLE1. Of these, 10 (77%) would have been
significant at a pre-specified cut-point at the median (SERPINA9,
ACTN1, COL3A, CTGF, PLAU, MYC, BCL6, CCND2,
PDCD4, and TLE1) and 3 genes (or a relative 23% of the 13
genes) would not have been significant (HLA-DRB, FN1, and
PRKCB1). Therefore, the median cut-point analysis would have
missed detecting the significance of a notable selection of genes.
Inspection of the graphs revealed patterns that allowed us to
classify the results into 3 different groups. The first group was
defined as those genes with the significant cut-points only below
the median. The second group was defined as genes with
significant cut-points only above the median. The third group
was defined as genes with significant cut-points above, below, or
including the median.
The single gene in the first category was HLA-DRB, with the
highest chi-square values all below the median and the most
significant cut-point at the 20thpercentile. This pattern is
consistent with a gene whose loss is associated with poor outcome.
The two genes that fell into the second category, showing
significant cut-points above the median gene expression values,
were CCND2 and PRKCB1. CCND2 is G1/S-specific regulator
of cyclin-dependent kinases, and PRKCB1 functions as a serine-
and threonine-specific protein kinase. This second pattern is
consistent with genes whose over-expression is associated with
poor outcome.
Ten genes fell into the third category, with significant cut-points
above and below the median gene expression values. The genes in
this category included ACTN1, COL3A, FN1, CTGF, PLAU,
Name in original referenceAlternative namesqNPA name Reference Function
39 UTR of unknown proteinMicrotubule-Associated
Protein 1B
MAP1BShipp 4Protein that stabilizes microtubules,
attaches other proteins to microtubules
and has a putative role in microvessicle
trafficking [55,56]
Transducin-like enhancer
protein 1
Groucho TLE1*Shipp 5 Transcriptional co-repressor involved in
differentiation of hematopoietic cells
[57,58]
Uncharacterizedcitrin SLC25A13Shipp 6 Mitochondrial inner membrane aspartate-
glutamate carrier that moves aspartate to
the cytosol and NADH reducing
equivalents into the mitochondria [59,60]
PDE4B Phosphodiesterase 4B,
cAMP-specific
PDE4B Shipp 7Phosphodiesterase that degrades cAMP
to inactivate cAMP signaling [61]
UncharacterizedUDP-Gal:betaGlcNAc
b-1,4-galactosyltransferase
polypeptide 1
B4GALT1 Shipp 8Enzyme that transfers galactose to
glycoproteins in a steriospecific manner;
galactoproteins are involved in immune
cell trafficking [62]
PRKCG Protein kinase C, gammaPRKCG Shipp 9 Serine/threonine–specific kinase
activated by lipid signals and reactive
oxygen species [63,64]
Oviductal glycoprotein MUC9 OVGP1Shipp 10Glycoprotein secreted by oviduct
epithelial cells under estrogen control
[65]
(MINO/NOR1) Mitogen induced
nuclear orphan receptor
NR4A3Shipp 11Nuclear hormone receptor that regulates
metabolism and inhibits leukemogenesis
in a ligand-independent manner [66,67]
Zinc-finger protein C2H2-150 ZNF212Shipp 12 Putative transcription factor [68]
5-Hydroxytryptamine 2B receptorHTR2B Shipp 13 Serotonin receptor isotype involved in
tumorigenesis [69,70]
CatalaseCAT Tome 1 Peroxisomal enzyme that metabolizes
H2O2[71]
Manganese superoxide dismutase SOD2Tome 2 Mitochondrial enzyme that metabolizes
superoxide [72]
1Last names with numbers refer to genes that are members of prognostic gene signatures previously reported in A. Rosenwald et al, I. Lossos et al, M. Shipp et al, and M.
Tome et al. [12], [73], [74], [75].
doi:10.1371/journal.pone.0022267.t001
Table 1.Cont.
DLBCL Gene Expression Variable Cut-Point Analysis
PLoS ONE | www.plosone.org4August 2011 | Volume 6 | Issue 8 | e22267

Page 5

DLBCL Gene Expression Variable Cut-Point Analysis
PLoS ONE | www.plosone.org5August 2011 | Volume 6 | Issue 8 | e22267