Interactions between glucocorticoid treatment and cis-regulatory polymorphisms contribute to cellular response phenotypes.
ABSTRACT Glucocorticoids (GCs) mediate physiological responses to environmental stress and are commonly used as pharmaceuticals. GCs act primarily through the GC receptor (GR, a transcription factor). Despite their clear biomedical importance, little is known about the genetic architecture of variation in GC response. Here we provide an initial assessment of variability in the cellular response to GC treatment by profiling gene expression and protein secretion in 114 EBV-transformed B lymphocytes of African and European ancestry. We found that genetic variation affects the response of nearby genes and exhibits distinctive patterns of genotype-treatment interactions, with genotypic effects evident in either only GC-treated or only control-treated conditions. Using a novel statistical framework, we identified interactions that influence the expression of 26 genes known to play central roles in GC-related pathways (e.g. NQO1, AIRE, and SGK1) and that influence the secretion of IL6.
- [show abstract] [hide abstract]
ABSTRACT: Cushing's syndrome results from sustained pathologic hypercortisolism caused by excessive corticotropin (ACTH) secretion by tumors in the pituitary gland (Cushing's disease, 70%) or elsewhere (15%), or by ACTH-independent cortisol secretion from adrenal tumors (15%). The clinical features are variable, and no single pattern is seen in all patients. Those features most specific for Cushing's syndrome include abnormal fat distribution, particularly in the supraclavicular and temporal fossae, proximal muscle weakness, wide purple striae, and decreased linear growth with continued weight gain in a child. Patients with characteristics of glucocorticoid excess should be screened with measurements of saliva or urine cortisol or dexamethasone suppression testing. The diagnosis of Cushing's syndrome should be followed by the measurement of plasma ACTH concentration to determine whether the hypercortisolism is ACTH-independent. In ACTH-dependent patients, bilateral inferior petrosal sinus sampling with measurement of ACTH before and after administration of ACTH-releasing hormone most accurately distinguishes pituitary from ectopic ACTH secretion. Surgical resection of tumor is the optimal treatment for all forms of Cushing's syndrome; bilateral adrenalectomy, medical treatment, or radiotherapy are sought in inoperable or recurrent cases. The medical treatment of choice is ketoconazole. The prognosis is better for Cushing's disease and benign adrenal causes of Cushing's syndrome than adrenocortical cancer and malignant ACTH-producing tumors.The American journal of medicine 01/2006; 118(12):1340-6. · 4.47 Impact Factor
- Annals of the Rheumatic Diseases 07/1949; 8(2):97-104. · 9.11 Impact Factor
- [show abstract] [hide abstract]
ABSTRACT: Glucocorticoids (GC) induce cell cycle arrest and apoptosis in lymphoid cells, and therefore constitute a central component in the treatment of lymphoid malignancies, particularly childhood acute lymphoblastic leukemia (ALL). In spite of its clinical significance and considerable efforts in many laboratories, however, the molecular basis of GC-induced apoptosis and the clinically important resistance phenomenon remains poorly defined. The anti-leukemic GC effects are critically dependent upon sufficient expression of the GC receptor (GR) throughout the response. In ALL cell lines, this is associated with, and may depend upon, GR autoinduction. In corresponding in vitro models, GC resistance frequently results from mutations in the GR gene and/or deficient regulation of its expression. The downstream components of the pathway, i.e., the GC-regulated genes responsible for cell death induction, have been studied by microarray-based comparative expression profiling, resulting in identification of a considerable number of GC-regulated candidate genes. Their possible function in the death response is, however, still undefined. One model predicts direct regulation of the apoptotic machinery, e.g., components of the "Bcl-2 rheostat", while a complementary hypothesis suggests deleterious GC effects on essential cellular functions, such as metabolism, production of and/or response to oxygen radicals, general transcription/translation, pH and volume control, etc. These regulatory effects may entail cell death, particularly if maintained for sufficient time through GR autoinduction. The latter form of cell death may occur even in the absence of functional apoptotic machinery (e.g., when caspases are blocked), but in this case appears to entail a more necrotic morphology. Taken together, GC may induce different types of cell death through distinct molecular pathways, depending on the cellular context. GC resistance might frequently result from defective GR expression, perhaps the most efficient means to target multiple antileukemic pathways.The Journal of Steroid Biochemistry and Molecular Biology 03/2005; 93(2-5):153-60. · 3.98 Impact Factor
Interactions between Glucocorticoid Treatment and
Cis-Regulatory Polymorphisms Contribute to Cellular
Joseph C. Maranville1., Francesca Luca1., Allison L. Richards1, Xiaoquan Wen2, David B. Witonsky1,
Shaneen Baxter1, Matthew Stephens1,2*, Anna Di Rienzo1*
1Department of Human Genetics, The University of Chicago, Chicago, Illinois, United States of America, 2Department of Statistics, The University of Chicago, Chicago,
Illinois, United States of America
Glucocorticoids (GCs) mediate physiological responses to environmental stress and are commonly used as pharmaceuticals.
GCs act primarily through the GC receptor (GR, a transcription factor). Despite their clear biomedical importance, little is
known about the genetic architecture of variation in GC response. Here we provide an initial assessment of variability in the
cellular response to GC treatment by profiling gene expression and protein secretion in 114 EBV-transformed B lymphocytes
of African and European ancestry. We found that genetic variation affects the response of nearby genes and exhibits
distinctive patterns of genotype-treatment interactions, with genotypic effects evident in either only GC-treated or only
control-treated conditions. Using a novel statistical framework, we identified interactions that influence the expression of 26
genes known to play central roles in GC-related pathways (e.g. NQO1, AIRE, and SGK1) and that influence the secretion of
Citation: Maranville JC, Luca F, Richards AL, Wen X, Witonsky DB, et al. (2011) Interactions between Glucocorticoid Treatment and Cis-Regulatory Polymorphisms
Contribute to Cellular Response Phenotypes. PLoS Genet 7(7): e1002162. doi:10.1371/journal.pgen.1002162
Editor: Greg Gibson, Georgia Institute of Technology, United States of America
Received April 8, 2011; Accepted May 15, 2011; Published July 7, 2011
Copyright: ? 2011 Maranville 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: FL was supported by a Blanceflor-Foundation Grant for post-graduate studies and by an AHA post-doctoral fellowship (0825792G). JCM was supported
by an NIH Genetics and Regulation Training Grant (T32 GM007197) to the University of Chicago and by an AHA pre-doctoral fellowship (11PRE4960001). This work
was supported by NIH grants DK56670 and DK-56670-S1 to ADR and NIH grant HG02585 to MS. 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 declared that no competing interests exist.
* E-mail: email@example.com (MS); firstname.lastname@example.org (ADR)
. These authors contributed equally to this work.
Glucocorticoids (GCs) are steroid hormones that mediate
homeostatic responses to environmental stressors through the
regulation of critical physiological processes (e.g. immune
response, energy metabolism and blood pressure (reviewed in
)). Owing to early observations of the anti-inflammatory
properties  of cortisol (i.e. the endogenous GC in humans),
synthetic GCs are widely used as pharmaceuticals for inflamma-
tory and autoimmune diseases (e.g. asthma  and rheumatoid
arthritis ). GCs are also used in the treatment of several types of
cancer , most notably lymphoid malignancies , due to their
pro-apoptotic activities and for symptomatic relief. While there is
evidence for a substantial genetic contribution [7–12], and for
inter-ethnic differences in drug response [13,14], little is known
about the genetic architecture of variation in GC response within
and between human populations.
Genetic effects on GC action couldprovide a mechanismfora vast
array of gene-environment interactions, which could have major
implications for human phenotypic variation. In fact, evidence of
such interactions has been observed for numerous traits relevant to
GCs including obesity , cardiovascular disease  and asthma
. With few exceptions (e.g. a regulatory polymorphism in the
promoter of IL6 ), little is currently known about the mechanisms
that underlie gene-environment interactions. If not properly
accounted for, these interactions can complicate efforts to identify
genetic and environmental factors associated with disease risk.
Furthermore, identifying genetic variation that interacts with
pharmaceutical treatments like GCs, which are a specific subset of
environmental factors, is of particular interest from a clinical
perspective and constitutes the primary goal of pharmacogenetics.
As GCs act largely by inducing changes in the expression
of target genes , regulatory polymorphisms are likely to
contribute to variation in response. The initial steps of the GC
response pathway are mediated by the GC receptor (GR) and
interacting transcription factors. GC binding allows the GR to
translocate from the cytoplasm to the nucleus, where it regulates
gene expression through at least two distinct mechanisms. The GR
can either drive the assembly of novel transcriptional regulatory
complexes at target genes, or inhibit regulatory complexes, such as
NFkB , that are already active at target genes. Some direct
GR target genes are, in turn, transcription factors that regulate
downstream target genes.
Here, we provide an initial view of the genetic architecture of
variation in the GC-mediated regulation of transcription and
protein secretion. To accomplish this, we measured the expression
of 13,232 genes and the secretion levels of 10 proteins in paired
aliquots, one treated with the synthetic GC dexamethasone (dex)
PLoS Genetics | www.plosgenetics.org 1July 2011 | Volume 7 | Issue 7 | e1002162
and one treated with the vehicle for dex (EtOH) as a control, in a
panel of 114 densely genotyped HapMap B-lymphocytes trans-
formed with Epstein-Barr Virus (EBV), commonly known as
lymphoblastoid cell lines (LCLs). This panel included 57 Yoruba
(YRI) from Nigeria and 57 Toscani (TSI) from Italy. EBV
transformation proceeds, in part, by mimicking CD40 activation
and ultimately leads to cellular proliferation through a variety of
mechanisms, including the activation of the NFkB signaling
pathway . Given their activated state, LCLs are a suitable
system for studying the immunorepressive effects of GCs.
Additionally, some regulatory variants that affect GC response
in LCLs may be shared with other cell types, as observed for
baseline expression [22–24].
GCs have widespread effects on transcriptional
We found that 4,568 genes were differentially expressed, at a
FDR,0.01 (p,0.003), following treatment with GCs (8 h, 1 uM
dexamethasone), corresponding to ,32% of the expressed genes.
This number is similar to that observed in a recent study of
equivalent sample size in osteoblasts treated with GCs , but
larger than previous studies that used much smaller samples (often
a single cell line; e.g. ). This suggests that large sample sizes are
necessary to identify many GC target genes. Accordingly, we
found that sub-sampling data from our full panel of LCLs
dramatically reduced the number of differentially expressed genes,
especially at genes with inter-individual variation in transcriptional
response (Figure S1). It should be noted that tests of differential
expression rely on magnitude of transcriptional response and its
consistency across individuals. Because our main goal is to identify
the genetic basis of variation in response, we did not limit our
mapping analyses (see below) to the differentially expressed genes.
equal numbers of up and down-regulated genes. Up-regulated genes
were enriched for GC-related biological processes including cellular
response to stimulus (p=4.161026, FDR=7.561025) and cell cycle
(p=1.461025, FDR=2.661024), consistent with GC regulation of
lymphocyte proliferation. Down-regulated genes were enriched for
immune response genes (p=1.1610210, FDR=4.661029) and for
genes involved in the positive regulation of I-kappaB kinase/NF-
kappaB cascade (p=3.361025, FDR=3.361024), consistent with the
immunorepressive role of GCs.
To explore the extent of tissue-specificity in the transcriptional
response to GCs, we compared our data to the results in osteoblasts
. We found a significant overlap between the genes differentially
expressed inLCLsand in osteoblasts (p=4.8610213), butonly 28%
of genes differentially expressed in our study are differentially
expressed (p,0.05) in osteoblasts. This likely reflects some amount
of tissue specificity, although other factors are likely to contribute
(e.g. incomplete power , differences in duration of treatment).
We measured and corrected for multiple factors related to EBV-
transformation that have been previously shown to be associated
with gene expression patterns at baseline [27,28] (e.g. EBV copy
number). Unlike baseline expression, these factors showed hardly
any evidence for an effect on transcriptional response (see Table
S1); nonetheless, we corrected for them in all subsequent analyses.
No evidence for trans-acting genetic effects on
transcriptional response to GCs
Many of the proteins involved in the GC-mediated regulation of
transcription are well characterized (i.e. GR and interacting
transcription factors). Genetic variants that impact the function of
these regulatory proteins are likely to influence transcriptional
response at several, and potentially many, downstream genes.
Consequently, the genes that encode these proteins are candidate
expression quantitative trait loci (eQTLs) acting in trans to
modulate the transcriptional response to GCs. However, ge-
nome-wide tests for trans eQTLs suffer from a tremendous multiple
testing burden. Therefore, to reduce the number of tests being
performed, we first examined only these candidate genes for
response eQTLs. We used simple linear regression to test for an
association between log fold change in expression at each
expressed gene in the genome and genotype at all HapMap SNPs
within 100 kb of the gene that encodes the GR (NR3C1), and
found no significant evidence of association at a FDR,0.2
(Figure 1a, Figure S4a-S4b). Similarly, as the GR interacts with
other transcription factors in the regulation of target gene
transcription, we also tested all HapMap SNPs within 100 kb of
34 genes that encode transcription factors known to interact with
the GR (listed in Materials and Methods ). Here again, we
found no evidence for an effect of genetic variation at these loci on
the transcriptional response to GCs at a FDR,0.2 (Figure 1b,
Evidence for local effects on transcriptional response to
We then performed an unbiased, genome-wide scan for genetic
variation associated with GC response. Specifically, we tested for an
association between every HapMap SNP and log fold change at
every gene. While this analysis did not reveal any significant
associations at a FDR,0.2, we found that the top association was
between log fold change at C1orf106 and genotype at an intronic
SNP (rs4915463, p=8.4610211, FDR,0.67). Given the proximity
of the associated SNP to the C1orf106 locus and work by others
highlighting the impact of cis-acting regulatory polymorphisms on
baseline expression [30,31], we then focused our analyses on
HapMapSNPs near eachof the 12,619expressed,autosomal genes.
We found the strongest signal when we tested for an association
between log fold change at each gene and all SNPs within 100 kb
Glucocorticoids (GCs) are steroid hormones produced by
the human body in response to environmental stressors.
Despite their key role as physiological regulators and
widely administered pharmaceuticals, little is known about
the genetic basis of inter-individual and inter-ethnic
variation in GC response. As GC action is mediated by
the regulation of gene expression, we profiled transcript
abundance and protein secretion in EBV-transformed B
lymphocytes from a panel of 114 individuals, including
those of both African and European ancestry. Combining
these molecular traits with genome-wide genetic data, we
found that genotype-treatment interactions at polymor-
phisms near genes affected GC regulation of expression for
26 genes and of secretion for IL6. A novel statistical
approach revealed that these interactions could be
distinguished into distinct types, with some showing
genotypic effects only in GC-treated samples and others
showing genotypic effects only in control-treated samples,
with differing phenotypic and molecular interpretations.
The insights into the genetic basis of variation in GC
response and the statistical tools for identifying gene-
treatment interactions that we provide will aid future
efforts to identify genetic predictors of response to this
and other treatments.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org2July 2011 | Volume 7 | Issue 7 | e1002162
(compared to either a genome-wide scan or testing SNPs within
500 kb of each gene, Figure 1c). This analysis revealed local
response eQTLs for 8 genes at a FDR,0.1 (Figure 2). These
included genes previously shown to play important roles in GC-
related biological processes, including regulation of immune
response (MT1X  and MFGE8 ) and cell cycle progression
(e.g. BIRC3 ). These also included NQO1, a gene previously
shown to affect variation in response to GC pharmaceutical
Distinguishing types of genotype-treatment interactions
that influence response
Visual examination of the genes in Figure 2 indicates that
different genes show qualitatively different patterns. For some
genes, a genotypic effect is evident either in only the GC-treated
condition (C1orf106, NQO1, C9orf5, MFGE8, and BIRC3) or only
the control-treated condition (MT1X). For others, an effect is
evident in both, but differs between the two conditions (DNAJC5G
and MS4A7). These different patterns may have different
mechanistic and phenotypic interpretations, but are not distin-
guished by the test of log fold change, and so researchers have
previously been forced to identify such patterns post hoc (e.g. ).
To address this, we developed a novel statistical framework that
explicitly compares and identifies these different patterns of
In brief, our method explicitly compares five different models
relating each SNP to phenotypic measurements in the two
treatment conditions (GC and control):
1. Null model: no association between genotype and phenotype in
2. No-interaction model: genotype is associated with phenotype in
both conditions, with the same effect in each condition.
3. GC-only model: genotype is associated with phenotype in GC-
treated samples, but not in control-treated samples.
4. Control-only model: genotype is associated with phenotype in
control-treated samples, but not in GC-treated samples.
5. General interaction model: genotype is associated with
phenotype in both conditions, but with different effects in
Figure 1. Quantile-quantile plots summarizing the results from tests for genetic variation associated with log fold change in
expression. a) Test of all HapMap SNPs within 100 kb of the gene that encodes GR. b) Test of all HapMap SNPs within transcription factors that
interact with the GR. c) Results from genome-wide scan (all HapMap SNPs) are compared to scan limited to SNPs within 500 kb and 100 kb of each
gene. Observed p-values are shown as black dots. P-values from permutations are shown as grey dots. Results from 100 permutations are shown for
a) and b), and results from 3 permutations are shown for c).
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org3 July 2011 | Volume 7 | Issue 7 | e1002162
Figure 2. Patterns of interaction between genotype and GC treatment that underlie associations with log fold change for each of
the 8 genes where log fold change is significantly associated with genotype at a SNP within 100 kb. Plots on the left show the effect of
genotype on log2fold change, with genotype coded as copies of the minor allele. In plots on the right, each small dot corresponds to an individual
and is color coded based on genotype (red=homozygous for major allele, purple=heterozygous, and blue=homozygous for minor allele). Large
dots represent genotypic means. Associations are classified based on the configuration of genotypic effects in the two conditions, including a)
genotypic effects only in GC-treated samples, b) genotypic effects only in control-treated samples and c) genotypic effects in both conditions that
differ. On the right is a cartoon showing patterns of interaction in two-dimensional space with expression after GC-treatment on the y-axis and
expression after control-treatment on the x-axis. Each of the three dots on each line corresponds to the mean value for a genotype class
(red=homozygous for major allele, purple=heterozygous, and blue=homozygous for minor allele). Genotype is coded as copies of the minor allele.
Relative expression corresponds to log2-transformed microarray intensities.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org4 July 2011 | Volume 7 | Issue 7 | e1002162
For each SNP, we computed a likelihood ratio, or Bayes Factor
(BF) that measures the relative support in the data for each model
1–4. These BFs take account of the paired nature of the data, and
the correlations between measurements in the same LCL in
different conditions. We used a hierarchical model  to
combine information both across SNPs in each gene region, and
across genes, ultimately computing a posterior probability for each
gene that it follows each of the models 1–4, i.e. that it is affected by
a polymorphism that follows that model. We used these posterior
probabilities both to identify high-confidence eQTLs of each type,
and to estimate false discovery rates among eQTLs exceeding any
given posterior probability threshold. This method is broadly
applicable to the study of any gene-environment interactions with
paired phenotype measurements.
Using this novel framework we identified 26 genes with high-
confidence interactions (posterior probability of interaction.0.7,
FDR=0.10) between GC treatment and eQTLs. These interac-
tion eQTLs included 7 of the 8 response eQTLs identified by
mapping log fold change. The remainder generally showed strong,
but not genome-wide significant (FDR,0.10), association with log
fold change (see Table S2 and Figure S2). The larger number of
interactions identified compared with mapping log fold change (26
versus 8), therefore, reflects an increase in power that comes from
explicitly considering different plausible interaction scenarios. Of
these, the majority (18 of 26) showed strongest support for GC-
only interactions, with the remainder (8 of 26) showing strongest
support for control-only eQTLs. Only one interaction between
treatment and genotype identified through mapping the log fold
change was not identified by the Bayesian hierarchical model
(DNAJC5G). This interaction was the least significant of the 8
identified by mapping log fold change, and although it also shows
some signal in the Bayesian analysis (BF for general interaction
versus null=4.96102, BF for general interaction versus no-
interaction=6.46103), the signal was insufficient to outweigh the
low prior probability of a general interaction estimated by the
hierarchical model (prior=0.001, see Table S4).
The Bayesian hierarchical model revealed eQTLs at genes with
clear biological relevance to GC-related biological processes that
were not identified through mapping the log fold change. These
include additional genes involved in the regulation of immune
response (e.g. CST7  or NLRP2 ) and cell cycle progression
(e.g. PDGFRL ), well-established GC targets, such as serum
and glucocorticoid regulated kinase 1 (SGK1 ), and previously
unknown GC target genes. For example, we found a control-only
eQTL for multiple coagulation factor deficiency 2 (MDCF2),
which is involved in the production of pro-coagulation factors
. The effect of GCs on coagulation is controversial , but
has been suggested to play a role in their therapeutic effects on
diseases such as asthma .
Given that cortisol regulates a variety of physiological processes
relevant to numerous diseases, we compared our eQTL results to
those from genome-wide association studies collected as a part of
the GWAS catalog . We found that a GC-only eQTL for
AIRE (rs762421) was associated with risk of the Crohn’s disease
. AIRE encodes a potent repressor of autoimmunity and can
cause severe autoimmune disease when mutated . In addition
to its role in removing autoreactive T cells in the thymus, AIRE
also plays a role in B-cell mediated immune response . We
found that the putative risk allele (rs762421-G) is associated with
the down-regulation of AIRE expression by GCs. This allele may
confer increased susceptibility to this autoimmune disease by
allowing GCs to decrease AIRE expression.
In addition to these interacting polymorphisms, our analysis
identified a much larger number of genes (6,813 genes) affected by
no-interaction eQTLs (posterior probability.0.7; FDR=0.16). In
other words, transcript levels at these genes depend on the eQTL
genotype, but the magnitude of transcriptional response does not
(i.e. model 2, see Figure S10). Our observation that the vast
majority of cis-acting regulatory polymorphisms with identical
genotypic effects across treatment conditions is consistent with
findings in osteoblasts treated with GCs  and in yeast ,
suggesting that this may reflect a general biological trend, rather
than a feature specific to our treatment and experimental system.
We compared the distribution of minor allele frequencies between
the eQTLs following these three models and did not observe any
significant differences (Figure S9).
The results reported above come from using a hierarchical
model, which combines information across SNPs within each
gene. One limitation of this hierarchical model is that it allows at
most one eQTL per gene. This may cause it to miss interacting
SNPs in genes that contain both interacting and non-interacting
eQTLs, and for this reason the probabilities on interacting models
may be underestimated. (More generally this feature could cause
apparent discrepancies between the results from the hierarchical
model and the log fold change analysis, although this does not
seem to be the case in the results above.) To assess whether this
limitation might have led us to miss some strong interaction signals
we also performed a SNP-level analysis using the BF (for
interaction models 2–4 vs non-interaction models 0–1) computed
for each SNP. This analysis identified 247 SNPs, in 120 distinct
genes, with BF exceeding 103, although none exceeding 105, that
are candidates for being interacting eQTLs (Table S5).
Validation of cis-regulatory polymorphisms with
To determine whether GC-only and control-only eQTLs
represented regulatory polymorphisms with treatment-specific
genotypic effects, we assayed treatment-dependent allelic imbalance
using quantitative real time PCR in heterozygotes. This assay also
asks whether local eQTLs act in cis, as alleles at cis-regulatory
polymorphisms, by definition, affect target gene transcription only
on the chromosome on which theyreside.Among the 26 interaction
eQTLs, we chose five at random among those for which a common
coding SNP could be reliably genotyped. We assayed three genes
with evidence of GC-only eQTLs (C9orf5, LSG1, and MFGE8). We
found significant allelic imbalance, with allelic effects in the same
direction as predicted by the eQTL mapping results, in GC-treated
samples, butnot incontrol-treated samples, forallofthem(Table1).
Table 1. eQTL validation by allele-specific qRT–PCR.
Gene eQTL Model
C9orf5 GC-only0.12 0.041
C12orf45 control-only 0.090.1740.220.017
SRD5A2 control-only0.01 0.4450.950.118
Effects represent the natural log of the allelic ratio of qRT-PCR measurements of
mRNA abundance, with the allele associated with increased expression in eQTL
experiments as the numerator. P-values are from t-tests comparing allelic ratios
between heterozygotes and homozygotes (as a control for technical sources of
imbalance) at the eQTL. Significant p-values (p-value,0.05) are indicated in bold.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org5 July 2011 | Volume 7 | Issue 7 | e1002162
We also performed allelic imbalance assays on 2 of the 8 control-
only eQTLs (SRD5A2, C12orf45). We found significant allelic
imbalance at C12orf45 only in the control-treated samples. While
not significant at p,0.05, we observed a pattern consistent with a
control-only eQTL at SRD5A2. Our failure to fully validate all 5
assayed eQTLs by allelic imbalance could reflect some level of false
positive identifications of eQTL interactions, but may also reflect
incomplete power of the allelic imbalance assay.
eQTL replication in an independent study
We compared our results with those from an independent GC
response eQTL mapping study in LCLs derived from asthma
patients (W. Qui and K. Tantisira, personal communication). We
found that 4 of the 9 interaction eQTLs that we identified, and
that were tested in both studies, showed significant associations
with log fold change in this independent dataset (p,0.05,
C1orf106, LSG1, CST7, and MS4A7), and an additional 2 showed
suggestive associations (p,0.1, SYT17 and BIRC3). This overlap is
highly significant (p=8.561024). Importantly, the overlap for
single-treatment eQTLs is much greater than that for response
eQTLs: all of the top 10 eQTLs identified by Qiu et al (2011) in
each treatment condition were replicated in our data (p,0.05),
while only 1 of the top 10 eQTLs for log fold change was
replicated. This contrast highlights the known statistical challenge
of mapping gene-environment interactions.
We also tested 15 of our interaction eQTLs (i.e. all eQTLs
tested in both studies) for an association with response to GC
therapy in 172 asthma patients (W. Qui and K. Tantisira, personal
communication). We found that a GC-only eQTL for TNIP1 was
p=2.561023, Bonferroni-corrected p=0.037). TNIP1 has an
established role in the immune response, as it encodes a protein
that inhibits NFkB  and contains polymorphisms that have
been associated with risk of systemic lupus erythematosus .
Between-population variation in GC response
We observed substantial allele frequency differences between
populations at many of the putative interaction expression
quantitative trait nucleotides (eQTNs), defined as the most strongly
associated SNP for each gene. Furthermore, differences in allele
frequency at these eQTNs were predictive of differences in
average transcriptional response between populations (r2=0.33,
p=5.361023, Figure 3a). This demonstrates that these eQTNs
contribute to differences in response between populations, and so
and in drug response. It also provides independent supporting
evidence that these eQTNs interact with GC treatment.
In some cases, allele frequency differences may explain why
genes respond to GC treatment only in individuals of one
population. For example, we observed that the GC-only eQTL
allele associated with up-regulation of the detoxification enzyme
NAD(P)H:quinone oxidoreductase 1 (NQO1) was extremely rare
outside equatorial African populations (Figure 3b), likely causing
the observed lack of NQO1 response in TSI LCLs, and the strong
up-regulation in many YRI LCLs (Figure 3c). This result may be
of particular relevance to ethnic disparities in leukemia patient
response to GCs, as alleles that reduce NQO1 enzymatic activity
have been associated with decreased response to a chemotherapy
regime that included GCs in patients with acute lymphoblastic
[51,52] and acute myeloid leukemia .
In an effort to identify additional genes with differences in
average transcriptional response between populations, we applied
the same statistical framework described above to test for
interactions between population (rather than genotype) and GC
treatment. Using this approach, we identified 258 genes with
FDR=0.128) between populations; of these, 130 were up-
regulated by GC treatment while 128 were down-regulated. We
found a consistent pattern across genes, with a tendency for
stronger up-regulation in YRI LCLs at 78% of up-regulated genes
with population differences in response (Figure S3). Interacting
eQTLs are enriched among genes with population differences in
response compared to all expressed genes (odds ratio=6.0,
p=5.461023) while no-interaction eQTLs are not enriched (odds
IL6 secretion is affected by a GC-only cis-regulatory
The attenuation of the immune response by GCs is partially
mediated by decreased secretion of pro-inflammatory molecules.
We measured the secreted levels of 9 pro-inflammatory proteins
(IL1a, IL6, IL8, IP10, MDC, Rantes, TNFa, TNFb) and 1 anti-
inflammatory protein (IL10). Five pro-inflammatory proteins
showed significant differential secretion in response to GCs in
LCLs (TNFa, TNFb, Rantes, IP10 and IL1a –Table S3); all five
showed lower secretion levels in the presence of GC, consistent
with the immune-repressive role of GCs. To identify genetic
variation that influences GC-mediated regulation of protein
secretion, we tested HapMap SNPs for association with log fold
change in secretion at each protein. Similar to our eQTL results,
we found significant associations (at a FDR,0.2) only when we
limited our search to SNPs near the genes that encode each
protein (i.e. we found no significant associations in genome-wide
or a candidate gene analysis). Testing SNPs within 100 kb of each
cytokine, we found a significant association between secretion
response at IL6 and genotype at a SNP ,56 kb downstream
(rs10225286, p=1.961024, FDR=0.1, Figure 4). Because this
SNP did not show strong evidence of an effect on IL6
transcriptional response, we propose that it affects secretion
through a mechanism independent of mRNA levels or that it
affects transcriptional response at a different treatment time point.
Here, we report a genome-wide scan for genetic variation that
influences the GC-mediated regulation of transcription and
protein secretion. The cellular response to GCs depends on a
well-characterized set of regulatory proteins (i.e. the GR and
interacting proteins). This provided us with a set of strong
candidate loci to perform trans-eQTL mapping tests. Despite this,
we found no evidence for trans-acting factors. In contrast, the
strongest signal from an unbiased genome-wide scan was a SNP
associated with transcriptional response at a nearby gene, and
even more eQTLs were revealed when we limited our analysis to
SNPs within 100 kb of each gene. Numerous studies have tested
genetic variation within or near the GR and interacting
transcription factors for association with patient response to GC
treatment. These studies have found mostly rare functional
polymorphisms that are unlikely to explain most heritable
variation in GC response (reviewed in ). Furthermore, rare
polymorphisms in the GR have dramatic phenotypic effects (e.g.
extreme hypoglycemia and hypertension ), as expected for a
master regulator that influences all downstream processes. Instead
of genetic variants in master regulators, our results suggest that cis-
regulatory polymorphisms that interact with GC treatment at
target genes could play an important role in GC response, as first
suggested based on observations at the SGK1 gene . These
findings suggest that future attempts to identify genetic variation
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org6July 2011 | Volume 7 | Issue 7 | e1002162
associated with clinical response to GCs may benefit from focusing
on likely cis-regulatory polymorphisms that impact response at
individual GC target genes, instead of testing master regulators of
the GC response pathway.
We found that associations between genotype and transcrip-
tional response could be discriminated into distinct categories
based on the configuration of genotypic effects across treatment
conditions. These categories likely correspond to specific genetic
Figure 3. Population differences in transcriptional responses and allele frequency differences at an interaction eQTLs. a) Observed
population differences in log-fold change (y-axis) are plotted against predictions based on genotypic effects and differences in allele frequency
(x-axis). Genetic predicted values are significantly correlated with observed differences in response (r2=0.33, p=5.361023). b) The global distribution
of the C allele at rs689459, which is associated with up-regulation of NQO1 expression by GCs. c) The effect of the population-specific GC-only eQTL at
NQO1 by population. In plots on the far right, each small dot corresponds to an individual and is color-coded based on genotype (red=homozygous
for G allele, purple=heterozygous, and blue=homozygous for C allele). Large dots represent genotypic means. Genotype is coded as copies of the C
allele at rs689459. Relative expression corresponds to log2-transformed microarray intensities.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org7 July 2011 | Volume 7 | Issue 7 | e1002162
mechanisms. GC-only eQTLs may reflect polymorphisms that
influence the binding of transcription factors that are only active in
the presence of GC treatment (e.g. the GR and interacting
transcription factors). In support of this hypothesis, we found that
GC-only eQTLs tended to affect up-regulated genes (13 of 18).
Although the causative polymorphism may not be among the
genotyped SNPs, we found examples of GC-only eQTLs where
most of the signal centered on a SNP that disrupts a predicted GR
binding site, such as the eQTN for C9orf5 (rs10816772, p for motif
Control-only eQTLs are compatible with a variety of
mechanisms. For example, they may reflect polymorphisms that
disrupt the binding of regulatory complexes, like NFkB, that are
directly inhibited by the GR (e.g. through protein-protein
interaction). Consistent with this, we found examples of control-
only eQTLs where most of the signal centered on a SNP that
disrupts a predicted binding site for a transcription factor directly
inhibited by GR, such as the eQTN for FBXL6 (rs10448143,
matrix and core similarity for NFkB.0.9). Direct inhibition of
transcription factors by GR generally leads to down-regulation of
target genes. However, we found equal numbers of control-only
eQTLs affecting up-regulated and down-regulated genes (4 of
each), so additional mechanisms must explain some fraction of
control-only eQTLs. These may include genetic effects on
regulatory elements that are indirectly inhibited by GC treatment
(e.g. through GR competition for access to DNA by another
transcription factor) or polymorphisms that affect transcriptional
response at secondary targets.
The different categories of interactions identified by our method
may also have distinct phenotypic interpretations. Polymorphisms
with GC-only effects on expression are likely to directly affect the
action of the GC-activated regulatory machinery. In contrast,
polymorphisms with control-only effects have no impact on the
cellular processes in the presence of GCs, but may still affect
phenotype by influencing variation in a ‘pre-treatment’ state. For
example, genetic effects on pro-inflammatory cytokine levels prior to
GC exposure could affect the amount of time cells take to reach the
optimal, lower levelsrequired to effectively suppress inflammation. In
summary, control-only QTLs may contribute more to variation in
underlying disease mechanisms, while GC-only QTLs may contrib-
ute to variation in GC pharmacodynamics. However, we also note
that, given their lower rates of validation and replication, there may
be a higher false positive rate for control-only eQTLs.
Inter-ethnic differences in GC response have been observed
clinically [13,14], and the prevalence of many GC-regulated
physiological traits differs across human populations . By
combining association mapping with comparisons between
populations, our study allowed a direct assessment of the
genetic basis of population differences in the cellular response to
GCs. We found that ancestry had substantial and systematic
effects on the transcriptional response to GCs, with a tendency
for stronger up-regulation after GC treatment in YRI LCLs.
Possible causes of such patterns include: non-genetic ‘confound-
ers’ (e.g. differences in immortalization procedure ), trans-
acting alleles that increase response and are at higher frequency
in YRI, or multiple, independent cis-acting alleles that increase
response in YRI at up-regulated genes. Our data favor the last
explanation. It seems unlikely that non-genetic ‘confounders’
explain all or most of the population differences, as we found
that the measured ‘confounders’ showed limited evidence of
effects on transcriptional response or differences between
populations (Figure S5). Although we cannot exclude the
possibility that population differences reflect a trans-acting
eQTL with differences in allele frequency, we found little
support for this explanation. Instead, we found evidence
suggesting that population differences may reflect differences
in allele frequency at cis-regulatory polymorphisms, as genes
with population differences in response were more likely to have
local interaction eQTLs. The possibility that a stronger response
in YRI reflects differences in allele frequency at cis-regulatory
polymorphisms is particularly interesting from an evolutionary
perspective, as differences in allele frequency acting in a
consistent direction (i.e. increasing GC responsiveness) across
multiple independent QTLs are usually interpreted as evidence
of polygenic adaptation [59–61].
In addition to these biological insights, we contribute novel
statistical methodology for mapping response phenotypes and
identifying gene-environment interactions. These methods are
applicable for any setting contrasting genotypic effects between
two conditions (with paired measurements), including pharmaco-
genetic studies of clinical response to drug therapy (e.g. ) and,
especially, functional genomic studies of genetic effects on
treatment response similar to the one presented here. These
methods provide a more powerful alternative to mapping a
measure of response (e.g. log fold change), which fails to
distinguish among different types of interactions, or comparing
Figure 4. Secretion QTL for IL6. Genotype is coded as copies of the minor allele. Log fold change in secretion corresponds to the difference
between dex and control of covariate-corrected, quantile-normalized, log2-transformed estimates of relative quantity from ELISA assays. In plot on
the right, each small dot corresponds to an individual and is color-coded based on genotype (red=homozygous for major allele,
purple=heterozygous, and blue=homozygous for minor allele). Large dots represent genotypic means.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org8July 2011 | Volume 7 | Issue 7 | e1002162
results from mapping separately in each condition, which ignores
the paired nature of the data.
In summary, this study provides an initial characterization of the
genetic basis of variation within and between human populations for
a key physiological regulator and commonly administered pharma-
ceutical. The biological insights and statistical tools presented here
extend our current understanding of the genetic basis of variation in
response to GCs, and will aid future efforts to characterize the
genetics of response to this and other treatments.
Materials and Methods
Cell culture and dexamethasone treatment
All cellular experiments described were conducted in lympho-
blastoid cell lines (LCLs), B lymphocytes transformed with Epstein-
Barr virus, that were collected as a part of the International
HapMap project. LCLs were thawed and passed once in RPMI
media supplemented with 15% fetal bovine serum, then washed
twice with phosphate-buffered saline and moved to RPMI media
supplemented with 15% charcoal-stripped fetal bovine serum.
After one passage in media with charcoal-stripped fetal bovine
serum (corresponding to a minimum culturing time of 5 days),
LCLs were seeded in the evening at a density of 56105cells/ml.
After an overnight incubation, LCLs were treated with 1026M
dexamethasone, and an equal amount of vehicle solution (solution
composed of 1% ethanol and 99% cell culture media) as a negative
control for treatment. For each LCL, one set of dex and control
aliquots was treated for 8 hours (to quantify mRNA abundance)
and the other for 24 hours (to assay inflammatory protein
secretion). The study design is depicted in Figure S6. LCLs were
thawed, cultured and treated in batches completely balanced by
treatment, population, technician and time of day. For quality
control purposes, biological replicates were performed for one
batch of four cell lines and both expression and treatment response
were highly replicable (Figure S7). Collection of all samples took 4
RNA extraction and array hybridization
For each expression study described in the preliminary data,
total RNA was extracted from each cell culture sample using the
QIAgen RNeasy Plus mini kit, and was found to be of high quality.
RNA was extracted from all 240 samples over the course of 5 days.
Total RNA was then reverse transcribed into cDNA, labeled,
hybridized to Illumina HumanHT-12 v3 Expression BeadChips
and scanned at the Southern California Genotyping Consortium
(SCGC: http://scgc.genetics.ucla.edu/) at the University of
California at Los Angeles. Each RNA sample was hybridized to
two separate arrays (i.e. in two technical replicates). To avoid
batch effects on RNA measurements, all 480 microarrays were
hybridized within 7 days. Summary data (e.g. mean intensity of
each probe across within-array replicates) were obtained using the
BeadStudio software (Illumina) at the SCGC. The microarray data
has been deposited in the Gene Expression Omnibus (GEO),
www.ncbi.nlm.nih.gov/geo, under accession number GSE29342.
Low-level analysis of microarray data
Low-level microarray analysis was performed using the
Bioconductor software package LUMI  in R (http://www.
r-project.org). We used applied variance stabilizing transformation
 to all arrays, removed probes with intensities indistinguishable
from background fluorescence levels in all samples (leaving 23,700
expressed probes), and performed quantile normalization across all
arrays. Probes were annotated by mapping to the RNA sequences
from RefSeq using BLAT. To avoid ambiguity in the source of
a signal due to cross-hybridization of similar RNA species, probes
that mapped to multiple genes were excluded from further
analyses. Probes that contained one or more HapMap SNPs were
also removed from further analyses to avoid spurious associations
between expression measurements and SNPs in linkage disequi-
Measurement and correction for confounders
To avoid spurious results and to reduce noise due to potential
confounders,we measured severalcovariatesrelevant to LCLbiology
including: EBV genome copy number, growth rate and mitochon-
drial genome copy number. EBV and mitochondrial genome copy
number were assessed using Taqman Gene Expression Assays (Assay
# Hs02596867_s1 for mitochondria and Pa03453399_s1 for EBV).
RNaseP was used as an endogenous control for both assays. We then
used linear regression to remove the effects of these potential
confounders at each gene and confounder-corrected data were used
in all subsequent analyses.
Identification of differentially expressed genes
In order to identify genes that, on average across individuals,
changed expression levels upon treatment with GCs, we
performed multiple linear regression at each gene with treatment
as the covariate of interest while taking other measured covariates
into account. To reduce the effects of outliers, microarray intensity
values were quantile normalized to a N(0,1) distribution across all
samples (treated and untreated). We used the distribution of p-
values observed when sample labels are permuted (ten permuta-
tions were used), an empirical estimate of the p-value distribution
under the null, to estimate the false discovery rate (FDR). We used
the online tool DAVID [65,66] to identify biological categories
enriched among differentially expressed genes, using all genes
expressed in LCLs (based on microarray data) as a background.
We used all HapMap SNPs for all mapping experiments
described. As TSI LCLs were only typed for phase III SNPs, we
used the CEU population sample to impute genotypes at all
HapMap phase I and II SNPs. Similarly, we imputed SNPs for
phase III YRI LCLs based on the YRI LCLs included in phase I
and II. Imputation was performed using BIMBAM , which
infers missing genotypes based on correlations between missing
and typed genotypes observed in samples where all genotypes are
typed. QTL mapping results were not qualitatively different if
using imputed or genotyped SNPs.
Genetic mapping of log-fold change in expression
We tested for association between all HapMap SNPs and
transcriptional response at each gene, using log fold change in
expression (GC-treated over control-treated expression) as a
measure of response. For our candidate gene-based scan for
trans-acting eQTLs that influenced response, we tested all
HapMap SNPs within 500 kb and 100 kb (in two separate sets
of analyses) of genes encoding the GR and transcription factors
that interact with the GR. Interacting transcription factors include
the genes that encode the components of the NFkB complex, AP1,
Oct1, Oct2, CREB, ETS1, STAT3, STAT5, STAT6, C/EBP,
TFIID, T-bet, PU.1/Spi-1, Smad3, Smad4, Smad6, COUP-TFII,
IRF3, STIP1, Hic5/Ara55, and nTrip6 . P-values calculated
with permutated genotype labels were used as an empirical null
distribution. In order to maintain the correlation structure across
genes, the same permutation seed was used for all genes in both
candidate gene tests and the genome-wide scan. Ten permutations
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org9July 2011 | Volume 7 | Issue 7 | e1002162
were performed for the test of variation within 500 kb, 100
permutations were used for the test of variation within 100 kb and
3 permutations were used for the genome-wide scan. For mapping
log fold change at SNPs within 500 kb or 100 kb of each gene,
permutation seeds were set separately at each gene. Association
tests were performed using a combination of Python, the R
statistical package and the genetic association mapping program
Bayesian regression for identifying genetic associations
and interaction with treatment
We developed a novel Bayesian statistical framework for genetic
association analysis in settings where measurements are available
on the same individuals in two different conditions (in our case,
GC-treated and control-treated). Our methods extend and impro-
ve the methods from Barber et al. (2009) to explicitly consider
‘‘qualitative interaction’’ models where genetic variants are associ-
ated with measurements in only one of the two conditions. Our
method takes into account both sample pairing and the intra-
individual correlation of measurements under the two conditions.
We describe our method in greater detail in Text S1. These
methods are implemented in software called BRIdGE (Bayesian
Regression for Identifying Gene-Environment interactions), which
is available on the Stephens and the Di Rienzo laboratories’ web
Allele-specific quantitative PCR on cDNA to assay allelic
We used TaqMan quantitative genotyping assays to test for allelic
imbalance at coding SNPs in LD with eQTLs that interacted with
GC treatment. Imbalanced expression of the two coding alleles is an
independent line of evidence for a cis-acting regulatory polymor-
phism and for the configuration of the effect in the two treatment
conditions (i.e. the interaction model). Total RNA from an aliquot of
the same culture samples used to hybridize microarrays (this was a
separate RNA extraction as that used to hybridize microarrays) was
synthesized into cDNA using the High-Capacity cDNA Reverse
Transcription Kit (Applied Biosystems, Foster City, CA) according to
the manufacturer’s protocol. Taqman SNP Genotyping Assays were
used to quantify relative mRNA abundance of each allele on an ABI
PRISM 7900HT Sequence Detection System. To account for
differences between the two fluorochromes, a standard curve was
built for each of the two alleles using serial dilutions of a genomic
DNA from an individual that was heterozygous at the coding SNP.
For each assay, we calculated the natural log-ratio between the two
different alleles. The numerator of this ratio was always the allele
associated with increased expression in the corresponding treatment
condition. Within each treatment, we quantile normalized allelic log-
ratios and used a one-tailed t-test to identify significant differences in
average allelic log-ratios between heterozygotes and homozygotes (as
an empirical null distribution of allelic log-ratios) at the eQTL.
Overlap with other genetic association studies
We compared our eQTL results to multiple genetic association
studies including Qiu et al. (2011) and those in the GWAS catalog.
For each interaction eQTL, we compared evidence at the most
associated SNP in our data when it was tested in both studies.
When the most associated SNP was not tested in the comparison
dataset, we identified the best proxy SNP for each eQTL among
those tested in both studies. To ensure that the best proxy SNP
captured the pattern at the original eQTL, we required the proxy
SNP to show strong evidence of association for the same eQTL
model as the original eQTL (BF for association.500 and posterior
probability for model.0.5).
Comparing transcriptional response between
We contrasted the transcriptional response to GCs between
YRI and TSI LCLs. Differences in transcriptional response
between populations will result in differences in average expression
levels that differ depending on treatment, as opposed to GC-
independent population differences that will be identical in both
treatments. As this is analogous to gene-environment interactions,
we used the same statistical framework to identify genes with
differences in transcriptional response between populations (see
Bayesian regression for identifying genetic associations and
interaction with treatment in Text S1). Covariate-corrected
expression levels were quantile normalized across individuals
(both YRI and TSI) for each gene to reduce the effect of outliers.
As population differences at the phenotypic level may reflect
population differences in response following a consistent pattern
across many genes, we identified the direction of population
differences at each gene in terms of log-fold change.
Quantification of inflammatory markers in the cell culture
medium and identification of secretion QTLs that interact
with GC treatment
A multianalyte ELISA assay (Millipore) was performed on the
culture medium of the cell aliquots treated for 24 hours. The assay
was performed at the Flow Cytometry Facility at the University of
Chicago, according to the manufacturer instructions. Two
technical replicates were run for each sample. Samples were
assayed in batches balanced by treatment and population. For
each analyte, the average quantity across technical replicates was
calculated and used for all subsequent analyses. The correlation
structure between paired aliquots for each sample (GC and
control) was visually inspected (Figure S8). A small subset of
samples with low quantity detected showed no correlation between
GC and control aliquots because of noise in the measurement at
low concentrations. Consequently, these samples were excluded
from downstream analyses. Secretion levels were highly correlated
across proteins, likely representing a latent factor that generally
affects secretion levels. To remove the effect of this latent factor,
we used linear regression to correct secretion levels at each protein
by secretion levels at all other measured proteins.
We identified 4,568 differentially expressed genes including up-
and down- regulated genes. b) Sub-sampling shows that the
number of differentially expressed genes identified is a function of
sample size. Larger sample sizes tend to identify genes with c) more
variable and d) smaller responses. Aberrantly high, relative to the
overall trend, median coefficients of variation and low log-fold
changes at the smallest sample size (n=4) likely reflect increased
sampling noise due to a very small sample size.
Transcriptional response to GC treatment in LCLs. a)
tions with treatment is compared to the traditional frequentist test
for association with response (i.e. association between genotype
and log-fold change). The quantiles of the observed p-value
distribution (minimum p-value per gene) are plotted against
expected quantiles (based on 10 permutations). The p-value
threshold corresponding to a FDR,0.1 is marked by a horizontal
Results from Bayesian method for mapping interac-
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org10 July 2011 | Volume 7 | Issue 7 | e1002162
grey line. Genes with significant interaction eQTLs identified by
the Bayesian multivariate method at a posterior.0.7 are shown in
red for the GC-only model and in blue for the control-only model.
Interactions identified by our method include both top hits from
the frequentist analysis and a number of additional genes,
indicating that our approach provides increased in power to
detect polymorphisms that interact with treatment and affect
at up-regulated genes. Strength of response, measured by log-fold
changes (GC/control), is depicted by intensity and color, with red
corresponding to lower log-fold change and blue corresponding to
higher log-fold change, for a) up-regulated and b) down-regulated
genes with significant population differences in transcriptional
response. Rows represent genes and columns represent individu-
als. The vertical black lines represent the separation between the
A stronger response to GCs is observed in YRI LCLs
and log-fold change in expression (GC/control) are plotted against
p-values from permutations, representing expectations under the
null, for a) SNPs within 500 kb (only SNP with minimum p-value
is plotted for each gene) of the gene that encodes the GR and b)
SNPs within 500 kb of transcription factors known to interact with
the GR. c) SNPs within 500 kb of each of the interacting
transcription factors and d) within 100 kb of each of the
interacting transcription factors.
Observed p-values for association between genotype
distribution of factors known to affect LCL biology. We do not
observe significant differences in a) EBV genome copy number
(p=0.824) or b) growth rate (p=0.477). c) We did observe a
significantly higher level of mitochondrial genome copy number
among YRI LCLs (p=0.0463).
Differences between TSI and YRI LCLs in the
For each of 116 LCLs, one aliquot was treated with the synthetic
GC dexamethasone and another aliquot was treated with the
vehicle for dexamethasone (EtOH) as a treatment control. Two
sets of paired aliquots were treated for each LCL, one for 8 hours
and the other for 24 hours. RNA was extracted from aliquots
treated for 8 hours and hybridized to two replicate arrays (for a
total of 4 arrays hybridized per LCL). Supernatant from the
aliquots treated for 24 hours were used to assay protein secretion.
The study design used in this experiment is shown.
(red), representing duplicate RNA hybridizations, tend to be larger
than correlations between randomly drawn pairs of arrays (grey),
suggesting a limited contribution from RNA hybridization to
variation in these measurements. b) Pair-wise correlations of
expression levels between biological replicates are always larger
(red) than randomly drawn pairs of LCLs (grey), similarly
suggesting that variation in cell culturing and treatment protocols
described here contribute little to variation in expression
measurements. c-d) Log-fold changes (GC/control) across the
4,568 differentially expressed genes compared between biological
replicates to assess the reproducibility of response. c) Correlations
between biological replicates (red) are higher than expected when
comparing randomly drawn pairs of LCLs (grey), suggesting that
variation from cell culturing and treatment does not explain the
majority of variation in response between LCLs. d) Correlations
a) Pair-wise correlations between technical replicates
between replicate pairs are shown for transcriptional response
across differentially expressed genes.
uals for each protein. Secretion levels represent log-transformed
ELISA measurements of protein quantity. Horizontal and vertical
lines in each plot indicate the threshold used to identify meaningful
Correlation between secretion levels across individ-
candidate eQTN categorized by model. No significant difference
is observed, based on Mann-Whitney U test, between control-only
and GC-only eQTNs (p=0.26), GC-only and no-interaction
eQTNs (p=0.94), or control-only and no-interaction eQTNs
Distribution of minor allele frequency for each
p-values from mapping the log fold change and log sum (sum of log
affected by no-interaction eQTLs. No deviation from null expecta-
tions are observed for association with log fold change, while a very
strong deviation from expectations under the null is observed for
association with the log sum.This isconsistentwith the stable effect of
eQTL genotype on expression at these genes. SNPs within 100 kb
were tested againstlog fold change at eachgene. Observed minimum
p-values per gene are shown as black dots. Minimum p-values from
permutations are shown as grey dots.
Quantile-quantile plots showing the distribution of
number, mitochondrial copy number, and growth rate.
Association between gene expression and EBV copy
eQTLs that interact with GC treatment.
Effects of GC treatment on cytokine secretion.
genes with an association following each model for all the
predictor variables we tested against gene expression.
Maximum likelihood estimates of the proportion of
SNP-based Bayesian analysis.
SNPs showing strongest signal for interaction from the
Supplementary Materials and Methods.
We thank Suzanne Conzen for helpful advice on the molecular biology of
glucocorticoid response, Yoav Gilad for sharing samples, Jonathan
Pritchard for helpful discussions about study design, Vivian Cheung for
helpful comments on the manuscript, Joe DeYoung and others at the
Southern California Genotyping Consortium (SCGC: http://scgc.genetics.
ucla.edu/) at the University of California at Los Angeles for microarray
hybridization and scanning services, and two anonymous reviewers for
Conceived and designed the experiments: JCM FL ADR. Performed the
experiments: JCM FL ALR SB. Analyzed the data: JCM FL. Contributed
reagents/materials/analysis tools: MS XW DBW. Wrote the paper: JCM
FL ADR MS.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org 11July 2011 | Volume 7 | Issue 7 | e1002162
1. Nieman LK, Ilias I (2005) Evaluation and treatment of Cushing’s syndrome.
Am J Med 118: 1340–1346.
2. Hench PS, Kendall EC, Slocumb CH, Polley HF (1949) Adrenocortical
Hormone in Arthritis : Preliminary Report. Ann Rheum Dis 8: 97–104.
3. Barnes NC (2007) The properties of inhaled corticosteroids: similarities and
differences. Prim Care Respir J 16: 149–154.
4. Hillier SG (2007) Diamonds are forever: the cortisone legacy. J Endocrinol 195:
5. Keith BD (2008) Systematic review of the clinical effect of glucocorticoids on
nonhematologic malignancy. BMC Cancer 8: 84.
6. Ploner C, Schmidt S, Presul E, Renner K, Schrocksnadel K, et al. (2005)
Glucocorticoid-induced apoptosis and glucocorticoid resistance in acute
lymphoblastic leukemia. J Steroid Biochem Mol Biol 93: 153–160.
7. Bateman ED, Boushey HA, Bousquet J, Busse WW, Clark TJ, et al. (2004) Can
guideline-defined asthma control be achieved? The Gaining Optimal Asthma
ControL study. Am J Respir Crit Care Med 170: 836–844.
8. Drazen JM, Silverman EK, Lee TH (2000) Heterogeneity of therapeutic
responses in asthma. Br Med Bull 56: 1054–1070.
9. Schwartz JT, Reuling FH, Feinleib M, Garrison RJ, Collie DJ (1972) Twin
heritability study of the effect of corticosteroids on intraocular pressure. J Med
Genet 9: 137–143.
10. Bray PJ, Cotton RG (2003) Variations of the human glucocorticoid receptor
gene (NR3C1): pathological and in vitro mutations and polymorphisms. Hum
Mutat 21: 557–568.
11. Kumsta R, Moser D, Streit F, Koper JW, Meyer J, et al. (2008) Characterization
of a glucocorticoid receptor gene (GR, NR3C1) promoter polymorphism reveals
functionality and extends a haplotype with putative clinical relevance. Am J Med
Genet B Neuropsychiatr Genet.
12. Hawkins GA, Lazarus R, Smith RS, Tantisira KG, Meyers DA, et al. (2009)
The glucocorticoid receptor heterocomplex gene STIP1 is associated with
improved lung function in asthmatic subjects treated with inhaled corticoste-
roids. J Allergy Clin Immunol.
13. Mattano LA, Jr., Sather HN, Trigg ME, Nachman JB (2000) Osteonecrosis as a
complication of treating acute lymphoblastic leukemia in children: a report from
the Children’s Cancer Group. J Clin Oncol 18: 3262–3272.
14. Chan M, Leung D, Szefler S, Spahn J (1998) Difficult-to-control asthma: clinical
characteristics of steroid-insensitive asthma. J Allergy Clin Immunol 101:
15. Hetherington MM, Cecil JE (2010) Gene-environment interactions in obesity.
Forum Nutr 63: 195–203.
16. Talmud PJ (2007) Gene-environment interaction and its impact on coronary
heart disease risk. Nutr Metab Cardiovasc Dis 17: 148–152.
17. Le Souef PN (2009) Gene-environmental interaction in the development of
atopic asthma: new developments. Curr Opin Allergy Clin Immunol 9:
18. Cole SW, Arevalo JM, Takahashi R, Sloan EK, Lutgendorf SK, et al. (2010)
Computational identification of gene-social environment interaction at the
human IL6 locus. Proc Natl Acad Sci U S A 107: 5681–5686.
19. Stahn C, Buttgereit F (2008) Genomic and nongenomic effects of glucocorti-
coids. Nat Clin Pract Rheumatol 4: 525–533.
20. Ray A, Prefontaine KE (1994) Physical association and functional antagonism
between the p65 subunit of transcription factor NF-kappa B and the
glucocorticoid receptor. Proc Natl Acad Sci U S A 91: 752–756.
21. Graham JP, Arcipowski KM, Bishop GA (2010) Differential B-lymphocyte
regulation by CD40 and its viral mimic, latent membrane protein 1. Immunol
Rev 237: 226–248.
22. Bullaughey K, Chavarria CI, Coop G, Gilad Y (2009) Expression quantitative
trait loci detected in cell lines are often present in primary tissues. Hum Mol
Genet 18: 4296–4303.
23. Ding J, Gudjonsson JE, Liang L, Stuart PE, Li Y, et al. (2010) Gene expression
in skin and lymphoblastoid cells: Refined statistical method reveals extensive
overlap in cis-eQTL signals. Am J Hum Genet 87: 779–789.
24. Nica AC, Parts L, Glass D, Nisbet J, Barrett A, et al. (2011) The architecture of
gene regulatory variation across multiple human tissues: the MuTHER study.
PLoS Genet 7: e1002003. doi:10.1371/journal.pgen.1002003.
25. Grundberg E, Adoue V, Kwan T, Ge B, Duan QL, et al. (2011) Global analysis
of the impact of environmental perturbation on cis-regulation of gene
expression. PLoS Genet 7: e1001279. doi:10.1371/journal.pgen.1001279.
26. Wu W, Zou M, Brickley DR, Pew T, Conzen SD (2006) Glucocorticoid receptor
activation signals through forkhead transcription factor 3a in breast cancer cells.
Mol Endocrinol 20: 2304–2314.
27. Stark AL, Zhang W, Mi S, Duan S, O’Donnell PH, et al. (2010) Heritable and
non-genetic factors as variables of pharmacologic phenotypes in lymphoblastoid
cell lines. Pharmacogenomics J.
28. Choy E, Yelensky R, Bonakdar S, Plenge RM, Saxena R, et al. (2008) Genetic
analysis of human traits in vitro: drug response and gene expression in
lymphoblastoid cell lines. PLoS Genet 4: e1000287. doi:10.1371/journal.
29. Kassel O, Herrlich P (2007) Crosstalk between the glucocorticoid receptor and
other transcription factors: molecular aspects. Mol Cell Endocrinol 275: 13–29.
30. Pickrell JK, Marioni JC, Pai AA, Degner JF, Engelhardt BE, et al.
Understanding mechanisms underlying human gene expression variation with
RNA sequencing. Nature 464: 768–772.
31. Stranger BE, Nica AC, Forrest MS, Dimas A, Bird CP, et al. (2007) Population
genomics of human gene expression. Nat Genet 39: 1217–1224.
32. Raymond AD, Gekonge B, Giri MS, Hancock A, Papasavvas E, et al. (2010)
Increased metallothionein gene expression, zinc, and zinc-dependent resistance
to apoptosis in circulating monocytes during HIV viremia. J Leukoc Biol 88:
33. Ait-Oufella H, Kinugawa K, Zoll J, Simon T, Boddaert J, et al. (2007)
Lactadherin deficiency leads to apoptotic cell accumulation and accelerated
atherosclerosis in mice. Circulation 115: 2168–2177.
34. Lepelletier Y, Zollinger R, Ghirelli C, Raynaud F, Hadj-Slimane R, et al. (2010)
Toll-like receptor control of glucocorticoid-induced apoptosis in human
plasmacytoid pre-dendritic cells (pDC). Blood.
35. Bolufer P, Barragan E, Collado M, Cervera J, Lopez JA, et al. (2006) Influence
of genetic polymorphisms on the risk of developing leukemia and on disease
progression. Leuk Res 30: 1471–1491.
36. Smith EN, Kruglyak L (2008) Gene-environment interaction in yeast gene
expression. PLoS Biol 6: e83. doi:10.1371/journal.pbio.0060083.
37. Veyrieras JB, Kudaravalli S, Kim SY, Dermitzakis ET, Gilad Y, et al. (2008)
High-resolution mapping of expression-QTLs yields insight into human gene
regulation. PLoS Genet 4: e1000214. doi:10.1371/journal.pgen.1000214.
38. Hamilton G, Colbert JD, Schuettelkopf AW, Watts C (2008) Cystatin F is a
cathepsin C-directed protease inhibitor regulated by proteolysis. EMBO J 27:
39. Petrilli V, Papin S, Tschopp J (2005) The inflammasome. Curr Biol 15: R581.
40. Guo FJ, Zhang WJ, Li YL, Liu Y, Li YH, et al. (2010) Expression and functional
characterization of platelet-derived growth factor receptor-like gene.
World J Gastroenterol 16: 1465–1472.
41. Lang F, Artunc F, Vallon V (2009) The physiological impact of the serum and
glucocorticoid-inducible kinase SGK1. Curr Opin Nephrol Hypertens 18:
42. Spreafico M, Peyvandi F (2009) Combined Factor V and Factor VIII
Deficiency. Semin Thromb Hemost 35: 390–399.
43. van Zaane B, Nur E, Squizzato A, Gerdes VE, Buller HR, et al. (2010)
Systematic review on the effect of glucocorticoid use on procoagulant, anti-
coagulant and fibrinolytic factors. J Thromb Haemost 8: 2483–2493.
44. Brims FJ, Chauhan AJ, Higgins B, Shute JK (2009) Coagulation factors in the
airways in moderate and severe asthma and the effect of inhaled steroids.
Thorax 64: 1037–1043.
45. Hindorff LA JH, Hall PN, Mehta JP, Manolio TA (2011) A Catalog of Published
Genome-Wide Association Studies.: Available at: www.genome.gov/gwastudies.
46. Barrett JC, Hansoul S, Nicolae DL, Cho JH, Duerr RH, et al. (2008) Genome-
wide association defines more than 30 distinct susceptibility loci for Crohn’s
disease. Nat Genet 40: 955–962.
47. Matsumoto M (2009) The role of autoimmune regulator (Aire) in the
development of the immune system. Microbes Infect 11: 928–934.
48. Lindh E, Lind SM, Lindmark E, Hassler S, Perheentupa J, et al. (2008) AIRE
regulates T-cell-independent B-cell responses through BAFF. Proc Natl Acad
Sci U S A 105: 18466–18471.
49. Zhou L, Mideros SX, Bao L, Hanlon R, Arredondo FD, et al. (2009) Infection
and genotype remodel the entire soybean transcriptome. BMC Genomics 10: 49.
50. Cohen S, Ciechanover A, Kravtsova-Ivantsiv Y, Lapid D, Lahav-Baratz S
(2009) ABIN-1 negatively regulates NF-kappaB by inhibiting processing of the
p105 precursor. Biochem Biophys Res Commun 389: 205–210.
51. Stanulla M, Dynybil C, Bartels DB, Dordelmann M, Loning L, et al. (2007) The
NQO1 C609T polymorphism is associated with risk of secondary malignant
neoplasms after treatment for childhood acute lymphoblastic leukemia: a
matched-pair analysis from the ALL-BFM study group. Haematologica 92:
52. Derbinski J, Schulte A, Kyewski B, Klein L (2001) Promiscuous gene expression
in medullary thymic epithelial cells mirrors the peripheral self. Nat Immunol 2:
53. Naoe T, Tagawa Y, Kiyoi H, Kodera Y, Miyawaki S, et al. (2002) Prognostic
significance of the null genotype of glutathione S-transferase-T1 in patients with
acute myeloid leukemia: increased early death after chemotherapy. Leukemia
54. van Rossum EF, Lamberts SW (2006) Glucocorticoid resistance syndrome: A
diagnostic and therapeutic approach. Best Pract Res Clin Endocrinol Metab 20:
55. McMahon SK, Pretorius CJ, Ungerer JP, Salmon NJ, Conwell LS, et al. (2010)
Neonatal complete generalized glucocorticoid resistance and growth hormone
deficiency caused by a novel homozygous mutation in Helix 12 of the ligand
binding domain of the glucocorticoid receptor gene (NR3C1). J Clin Endocrinol
Metab 95: 297–302.
56. Luca F, Kashyap S, Southard C, Zou M, Witonsky D, et al. (2009) Adaptive
variation regulates the expression of the human SGK1 gene in response to stress.
PLoS Genet 5: e1000489. doi:10.1371/journal.pgen.1000489.
57. Lin SX, Carnethon M, Szklo M, Bertoni A (2011) Racial/ethnic differences in
the association of triglycerides with other metabolic syndrome components: the
Multi-Ethnic Study of Atherosclerosis. Metab Syndr Relat Disord 9: 35–40.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org12July 2011 | Volume 7 | Issue 7 | e1002162
58. Akey JM, Biswas S, Leek JT, Storey JD (2007) On the design and analysis of
gene expression studies in human populations. Nat Genet 39: 807–808; author
59. Bullard JH, Mostovoy Y, Dudoit S, Brem RB (2010) Polygenic and directional
60. Fraser HB, Moses AM, Schadt EE (2010) Evidence for widespread adaptive
evolution of gene expression in budding yeast. Proc Natl Acad Sci U S A 107:
61. Orr HA (1998) Testing natural selection versus genetic drift in phenotypic
evolution using quantitative trait locus data. Genetics 149: 2099–2104.
62. Barber MJ, Mangravite LM, Hyde CL, Chasman DI, Smith JD, et al. (2010)
Genome-wide association of lipid-lowering response to statins in combined study
populations. PLoS ONE 5: e9763. doi:10.1371/journal.pone.0009763.
63. Du P, Kibbe WA, Lin SM (2008) lumi: a pipeline for processing Illumina
microarray. Bioinformatics 24: 1547–1548.
64. Lin SM, Du P, Huber W, Kibbe WA (2008) Model-based variance-stabilizing
transformation for Illumina microarray data. Nucleic Acids Res 36: e11.
65. Huang da W, Sherman BT, Lempicki RA (2009) Systematic and integrative
analysis of large gene lists using DAVID bioinformatics resources. Nat Protoc 4:
66. Dennis G, Jr., Sherman BT, Hosack DA, Yang J, Gao W, et al. (2003) DAVID:
Database for Annotation, Visualization, and Integrated Discovery. Genome Biol
67. Guan Y, Stephens M (2008) Practical issues in imputation-based association
mapping. PLoS Genet 4: e1000279. doi:10.1371/journal.pgen.1000279.
Glucocorticoids Interact with Cis-Polymorphisms
PLoS Genetics | www.plosgenetics.org13 July 2011 | Volume 7 | Issue 7 | e1002162