The first wave of large-scale, high-density genome-wide
association (GWA) studies has improved our understanding
of the genetic basis of many complex traits1. For several
diseases, including type 1 (Refs 2,3) and type 2 diabetes4–9,
inflammatory bowel disease10–14, prostate cancer15–20 and
breast cancer21–23, there has been rapid expansion in the
numbers of loci implicated in predisposition. For others,
such as asthma24, coronary heart disease25–27 and atrial
fibrillation28, fewer novel loci have been found, although
opportunities for mechanistic insights are equally prom-
ising. Several common variants influencing important
continuous traits, such as lipids7,29–31, height32–35 and
fat mass36–38, have also been found. An updated list of
published GWA studies can be found at the National
Cancer Institute (NCI)-National Human Genome
Research Institute (NHGRI)’s catalog of published
genome-wide association studies.
These findings are providing valuable clues to the
allelic architecture of complex traits in general. At the
same time, many methodological and technical issues
that are relevant to the successful prosecution of large-
scale association studies have been addressed. However,
despite understandable celebration of these achieve-
ments, sober reflection reveals many challenges ahead.
Compelling signals have been found, often highlighting
previously unsuspected biology, but, for most of the
traits studied, known variants explain only a fraction of
observed familial aggregation39, limiting the potential
for early application to determine individual disease
risk. Because current technology surveys only a lim-
ited subset of potentially relevant sequence variation,
this should come as no surprise. Much work remains
to obtain a complete inventory of the variants at each
locus that contribute to disease risk and to define the
molecular mechanisms through which these variants
operate. The ultimate objectives — full descriptions of
the susceptibility architecture of major biomedical traits
and translation of the findings into clinical practice —
With completion of the initial wave of GWA scans, it
is timely to consider the status of the field. This Review
considers each major step in the implementation of a
GWA scan, highlighting areas where there is an emerg-
ing consensus over the ingredients for success, and those
aspects for which considerable challenges remain.
Subject ascertainment and design
Although there is a growing focus on the application
of GWA methodologies to population-based cohorts,
most published GWA studies have featured case–
control designs, which raise issues related to the optimal
selection of both case and control samples.
*Wellcome Trust Centre for
Human Genetics, University
of Oxford, Oxford, UK.
Correspondence to M.I.M
Published online 9 April 2008
studies in which a dense array
of genetic markers, which
captures a substantial
proportion of common
variation in genome sequence,
is typed in a set of DNA
samples that are informative
for a trait of interest. The aim is
to map susceptibility effects
through the detection of
associations between genotype
frequency and trait status.
Genome-wide association studies
for complex traits: consensus,
uncertainty and challenges
Mark I. McCarthy*‡, Gonçalo R. Abecasis§, Lon R. Cardon*||, David B. Goldstein¶,
Julian Little#, John P. A. Ioannidis* *‡‡ and Joel N. Hirschhorn§§||||¶¶
Abstract | The past year has witnessed substantial advances in understanding the
genetic basis of many common phenotypes of biomedical importance. These advances
have been the result of systematic, well-powered, genome-wide surveys exploring the
relationships between common sequence variation and disease predisposition. This
approach has revealed over 50 disease-susceptibility loci and has provided insights into
the allelic architecture of multifactorial traits. At the same time, much has been learned
about the successful prosecution of association studies on such a scale. This Review
highlights the knowledge gained, defines areas of emerging consensus, and describes
the challenges that remain as researchers seek to obtain more complete descriptions
of the susceptibility architecture of biomedical traits of interest and to translate the
information gathered into improvements in clinical management.
356 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
An association study design in
which the primary comparison
is between a group of
individuals (cases), ascertained
for the phenotype of interest
and that are presumed to have
a high prevalence of
susceptibility alleles for that
trait, and a second group
(controls), not ascertained for
the phenotype and considered
likely to have a lower
prevalence of such alleles.
Bias arising from the fact that
the samples ascertained for the
study (particularly controls)
might not be representative of
the wider population that they
are purported to represent.
Bias resulting from the failure to
correctly assign individuals to
the relevant group in a case–
control study; for example, the
presence of some individuals
who meet the criteria for being
cases in a population-based
The presence in study samples
of individuals with different
ancestral and demographic
histories: if cases and controls
differ with respect to these
features, markers that are
informative for them might be
confounded with disease
status and lead to spurious
Case selection. The principal issues with regard to case
ascertainment revolve around the extent to which selec-
tion should be driven by manoeuvres that are designed
to improve study power through enrichment for spe-
cific disease-predisposing alleles. These include efforts
to minimize phenotypic heterogeneity or to focus on
extreme and/or familial cases (defined, for example,
by early age of onset or ascertainment from multiplex
pedigrees). Because the genetic architecture of most
complex traits remains poorly understood, the value of
such efforts is hard to predict. In most circumstances,
and particularly when the total GWA sample size has
financial or operational constraints, efforts to enrich
case selection are likely to improve power. However,
there are situations in which selection of familial
cases or extreme individuals might have the opposite
Control selection. optimal selection of control samples
remains more controversial, although the accumulating
empirical data indicate that many commonly expressed
concerns have been overstated. The Wellcome Trust
Case Control Consortium (WTCCC) study was able
to demonstrate the effectiveness of a ‘common control’
design in which 3,000 uK controls were compared
with 2,000 cases from each of 7 different diseases1. The
WTCCC also assuaged concerns about the potential
for selection bias when using non-population-based
controls1. Comparison of the genome-wide genotypic
distributions from the two constituents of the WTCCC
common-control resource (one derived from a popula-
tion-based birth cohort, the other from opportunistic
sampling of blood donors) revealed no excess of sig-
nificant associations, indicating that ascertainment,
selection and survival biases were, in this situation at
least, having minimal impact on genotype distributions.
Although each prospective control sample must be
critically evaluated, these findings suggest that a broad
range of ascertainment schemes are compatible with
one consequence of the common-control design is
the potential loss of power that is associated with the
inability to exclude latent diagnoses of the phenotype
of interest through intensive screening of controls.
Fortunately, the consequences of misclassification bias are
modest unless the trait is common, and any loss of power
is recoverable by increasing the sample size (BOX 1).
For common traits, such as obesity and hypertension, in
which the effect of misclassification on power is great-
est1, one remedy involves adopting a more stringent case
definition, for example, based on early age of onset or
ascertainment of a more extreme phenotype, while
still excluding monogenic cases. Although the most
powerful strategy for a given fixed sample size involves
a ‘hypernormal’ control group, it might be difficult to
identify such individuals without introducing inadvert-
ent selection effects. For instance, selecting extremely
low-weight individuals as controls for a case–control
study of obesity could result in overrepresentation of
alleles primarily associated with chronic medical dis-
ease or nicotine addiction rather than weight regulation
Other case–control design issues. Four other issues loom
large in the design of case–control studies. The first is
sample size, and with this issue the consensus view
is clear: the more samples the better1,34,35,38. The initial
wave of GWA studies has shown that, with rare excep-
tions, the effect sizes resulting from common SNP
associations are modest, and that sample sizes in the
thousands are essential1.
The second issue relates to the propensity for latent
population substructure (population stratification and
cryptic relatedness) to inflate the type 1 error rate
and generate spurious claims of association around
variants that are informative for that substructure42,43.
The evidence emerging from GWA studies is reassur-
ing: as long as cases and controls are well matched for
broad ethnic background, and measures are taken to
identify and exclude individuals whose GWA data
reveal substantial differences in genetic background,
the impact of residual substructure on type 1 error
seems modest1. Several statistical tools exist to detect
and adjust for residual stratification42,44, and invento-
ries of markers that are informative for the detection of
ethnic substructure are a useful by-product of current
scans1,45–47. These approaches can be used to adjust for
substructure even in populations with quite diverse
antecedents (such as european-descent populations
in North America)46,47 and with negligible impact on
power48. Analysis in African-descent populations is
complicated by their greater haplotypic diversity and
fine-scale geographical structure49, and by the exten-
sive admixture demonstrated by African-descent
populations that are resident in europe and North
America. Furthermore, it is important to note that the
tools mentioned above (particularly genomic-control
approaches44) correct for ‘average’ genome-wide meas-
ures of ethnic admixture, and will not always eliminate
spurious associations immediately adjacent to markers
that are strongly informative about ancestry.
‡Oxford Centre for Diabetes, Endocrinology and Metabolism, University of Oxford,
§Center for Statistical Genetics, Department of Biostatistics, University of Michigan,
Ann Arbor, Michigan 48109, USA.
||Fred Hutchinson Cancer Research Center, Seattle, Washington 98109, USA.
¶Center for Population Genomics and Pharmacogenetics, Duke Institute for Genomic
Sciences and Policy, Duke University Medical Center, Duke University, Durham,
North Carolina 27708, USA.
#Department of Epidemiology and Community Medicine, University of Ottawa,
Ottawa, Ontario, Canada.
**Clinical and Molecular Epidemiology Unit, Department of Hygiene and
Epidemiology, University of Ioannina School of Medicine, and Biomedical Research
Institute, Foundation for Research and Technology-Hellas, Ioannina 45110 Greece.
‡‡Department of Medicine, Tufts University School of Medicine, Boston,
Massachusetts 02111, USA.
§§Division of Genetics and Endocrinology and Program in Genomics, Children’s
Hospital, Boston, Massachusetts 02115, USA.
||||Department of Genetics, Harvard Medical School, Boston, Massachusetts 02115, USA.
¶¶Broad Institute at MIT and Harvard, Cambridge, Massachusetts 02142, USA.
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 357
© 2008 Nature Publishing Group
evidence — typically gained
from analysis of GWA data —
that, despite allowance for
known family relationships,
individuals in the study sample
have residual, non-trivial
degrees of relatedness, which
can violate the independence
assumptions of standard
A suite of analytical approaches
in which association testing is
performed within families: such
approaches offer protection
from population substructure
effects but at the price of
The third issue concerns the relative merits of family-
based and case–control association methods. Although
family-based association methods provide a robust strategy
for dealing with stratification, this typically comes at
the cost of reduced power50. Given the ease with which
GWA data enable the detection of, and correction for,
population substructure42,44, this particular justification
has become less persuasive. Nevertheless, there are many
valuable clinical resources (for example, isolates) for
which pedigree information can be usefully exploited.
one option for the efficient use of family data in such a
setting is to restrict high-density scanning to a subset of
pedigree members and then use information on patterns
of chromosomal segregation derived from low-density
Box 1 | The impact of selection by phenotype among controls on power and sample size
In a case–control study, the manner in which the controls are ascertained (with respect to the phenotype of interest)
has implications for the power of the study and for sample size. The panels on the left show estimates of power for a
sample size of 2,000 cases and 2,000 controls and α (p value) = 10–6. Those on the right show the sample sizes (that is,
the number of case–control pairs) that are required for 80% power at the same threshold. In the upper panels, the
disease of interest has a population prevalence of 5% (so that cases are ascertained purely from the top 5% of the
population distribution); in the lower panels, the population prevalence is 20%. In each panel, power or sample size
estimates are shown for a range of control selection thresholds, that is, the trait-distribution threshold that is used to
define the controls. Under scenario A, controls are ascertained from the full distribution (that is, population-based
controls): a proportion (5% or 20%) will meet the criteria for being cases. Under scenario B, controls are ascertained
only if they cannot be cases: they come from the residual part (bottom 80% or 95%) of the distribution. Under scenario
C, hypernormal controls have been selected exclusively from the lowest 5% of the distribution. Each panel considers
four potential susceptibility loci. Tracks in blue denote loci that account for 0.25% of overall trait variance, tracks in red
denote loci that account for 1%. Light red and light blue symbols denote that the variant responsible is common
(overall allele frequency 30%), red and dark blue symbols denote that the variant is rare (1%).
As expected, in all settings, scenario C is the most powerful strategy for given overall case–control sample size, and
scenario A is the least powerful strategy. When the disease prevalence is modest (5%; upper panels), the distinctions
between scenarios A and B are not large, and it will often be easier to increase sample size than to undertake detailed
phenotypic examination of the controls to exclude latent cases. When the disease prevalence is higher (20%; lower
panels), misclassification is more prevalent under scenario A, the adverse consequences of using population-based
controls are more marked, and the advantages of using hypernormal controls (scenario C), if available, are most obvious.
Nature Reviews | Genetics
0 0.1 0.2 0.30.40.5 0.60.70.80.91.00 0.10.2 0.3 0.40.50.60.7 0.80.9 1.0
0 0.10.2 0.30.40.5 0.6 0.7 0.80.91.0
0 0.10.2 0.3 0.40.5 0.6 0.70.80.91.0
Trait prevalence 5%
Trait prevalence 20%
Control selection threshold
Control selection threshold
Control selection threshold
Control selection threshold
α = 10–6; 2,000 cases; 2,000 controls 80% power; α = 10–6
QTL variance 0.25%
QTL allele frequency 30%
QTL variance 0.25%
QTL allele frequency 1%
QTL variance 1%
QTL allele frequency 30%
QTL variance 1%
QTL allele frequency 1%
358 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
The phenomenon whereby a
single allele can affect several
distinct aspects of the
phenotype of an organism,
often traits not previously
thought to be mechanistically
(LD). The nonrandom allocation
of alleles at nearby variants to
individual chromosomes as a
result of recent mutation,
genetic drift or selection,
manifest as correlations
between genotypes at closely
Copy number variant
(CNV). A class of DNA
sequence variant (including
deletions and duplications) in
which the result is a departure
from the expected diploid
representation of DNA
DnA pooling approaches
Association studies that are
conducted using estimates of
allele frequencies derived from
pools of DNA compiled from
multiple subjects rather than
individual DNA samples.
genotyping in the remaining members to propagate
genotypes through the family51.
The fourth question relates to the potential to use
historical control genotypes to substitute for, or supple-
ment, newly typed controls in future GWA studies. The
risks associated with this (particularly inflation of the
type 1 error) will clearly depend on the extent to which
there are disparities between the new cases and historical
controls with respect to population origins, DNA format
(whole-genome amplified DNA versus native DNA as
well as storage conditions)52,53 and genotyping imple-
mentation (platform, genotyping centre, generation of
chip or allele-calling software). The limits of acceptable
divergence are not yet known, but it seems safest that
studies intending to use historical control data also
type a sample of ethnically matched controls (including
a subset of the historical control samples, if available)
using the same assay as for the newly defined cases. This
should allow the detection of systematic effects that are
attributable to non-disease-related differences between
the historical and new data. An alternative approach
would involve the re-genotyping of any interesting GWA
signals in all samples using a dedicated assay.
From case–control to cohort studies. Increasingly, the
GWA approach is being extended from analysis of
case–control samples to population-based cohorts54–56.
Although typically underpowered for dichotomous phe-
notypes (given limited cases for any given disease), such
cohorts often offer a rich tapestry of longitudinal meas-
ures for a wide range of quantitative traits, and lifestyle
and exposure data can enable an evaluation of the joint
effects of genes and environment. These studies promise
new insights into the genetic basis of continuous traits
and enhanced opportunities for revealing pleiotropy57,
although low power remains an issue — especially
for the detection of non-additive gene–environment
interactions58,59. Whereas GWA data meta-analysis is
the obvious solution to overcome restrictions of sam-
ple size, such procedures are often complicated by the
lack of standardization that characterizes the meas-
urement of many key continuous biological traits and
Marker selection and assay design. The debates about
marker selection that dominated early discussions of
GWA approaches have now boiled down to choices
between a limited range of commodity genome-wide
chips61,62. This is not the place for a detailed discussion
of the relative merits of specific array-designs other than
to point out that, genotype for genotype, designs that
take linkage disequilibrium (lD) structure into account
when defining content will achieve greater coverage,
in the index population at least. However, they will also
be more vulnerable to loss of coverage when assays fail,
or when typing samples whose genetic ancestry differs
from that of the reference panel or panels used to guide
marker selection61,62. Although greater feature density,
which allows more variants to be typed in a single array,
increases coverage, this does not necessarily equate to
greater power. When funding is limited, but the avail-
ability of samples is not, overall power might be maxi-
mized by typing more individuals with a less dense and
less costly array. In populations of non-African ancestry,
some of the drive towards ever-increasing array density
has been blunted by the capability to impute genotypes
at untyped loci (discussed later)63,64.
A new option in array selection comes with the inclu-
sion on contemporary commodity arrays of additional
probes that are designed to type copy number variants
(CNvs)65. However, because the global inventory of
CNvs remains incomplete66, and with limited empirical
data currently available, the extent to which the cur-
rent round of products (such as the Affymetrix 6.0 and
Illumina Human1M arrays) captures the structural vari-
ome remains unclear. However, their use should provide
early insights into the contribution such variants make to
common phenotypes of biomedical importance67.
For any given study, the final choice of array platform
is often a pragmatic one, based not only on the number
and ethnic origin of the samples, and the overall research
objectives, but also on factors such as cost, array delivery
schedules and available genotyping capacity.
DNA pooling. The costs of well-powered GWA stud-
ies have reignited interest in the value of DNA pooling
approaches as a means to conduct more economical
genome-wide surveys for association68. However, even
though several studies have been completed69, the falling
costs of commodity genotyping and the intrinsic limita-
tions of the pooling approach (reduced power, loss of
individual genotype data and difficulties ensuring equi-
molar representation of samples) mean that the future of
this approach, for GWA studies at least, is uncertain.
Obtaining robust genotype data. experience from the
first wave of GWA studies has demonstrated that scru-
pulous attention to detail is required throughout because
each stage is fraught with the potential for error and
bias1,52,53,70,71. Many of these errors and biases have the
potential to generate extreme values for the association
test statistic; if uncorrected, these can dominate the tails
of the distribution, such that interesting true associa-
tions become lost in a sea of spurious signals. efforts to
prevent, detect and eradicate sources of bias and error
therefore remain a high priority in GWA studies1, despite
continuing improvements in genotyping performance.
These efforts start with careful attention to the qual-
ity and accurate quantification of the starting DNA.
Differences in extraction methods between cases
and controls can be an important source of bias52,53.
Implementation of genotyping-performance metrics allows
poorly performing arrays to be targeted for re-analysis
and deficient samples to be selected for replacement1.
Given the scale of data generation, conversion of raw
experimental data into genotypes has necessitated the
development of automated methods. Indeed, the very
idea of assigning a discrete genotype call has increasingly
been replaced by measures of the posterior probability of
each possible genotype, given the observed data1. Several
algorithms have been developed for defining the three
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 359
© 2008 Nature Publishing Group
If patterns of missing data are
nonrandom with respect to
both genotype and trait status,
then analysis of the available
genotypes can result in
misleading associations where
none truly exists.
Signal intensity (cluster)
Plots of raw intensity data for
individual variants that are
generated by the genotyping
platform and represent the
extent to which the various
genotypes can be
discriminated: these provide a
useful visual diagnostic for the
genotyping data quality.
(HWe). A theoretical
description of the relationship
between genotype and allele
frequencies that is based on
expectation in a stable
population undergoing random
mating in the absence of
selection, new mutations and
gene flow: in the context of
genetic studies, departures
from equilibrium can be used
to highlight genotyping errors.
(Q-Q plot). In the context of
GWA studies, a Q-Q plot is a
diagnostic plot that compares
the distribution of observed
test statistics with the
distribution expected under
contingency-table test for
association that is well suited
to the detection of trends
across ordinal categories
(in this case, genotypes).
A school of statistics that uses
p values and combines them
with hypothesis testing to
genotype clusters at a given SNP, and for assigning geno-
type calls to each1,53,72–74, with each generation of soft-
ware heralding steady improvements in both accuracy
and call rate. This last point is crucial because there is
no room for complacency regarding SNPs that display
low call rates1,53. Case–control studies, in particular, are
vulnerable to informative missingness, whereby spurious
associations arise through differences in the patterns of
missing data with respect to genotype.
Considerations such as these have highlighted a
tension between stringency and call rate, which has pre-
vented the promulgation of quality-control thresholds
with a universal value for defining the set of ‘clean’ geno-
types for analysis. If researchers aim to maximize accu-
racy by setting the threshold for calling genotypes too
high, the consequence for many SNPs will be a low call
rate (such that some true signals are discarded) and/or
a rise in type 1 error (due to informative missingness).
However, if a lower threshold is favoured, as some groups
have done1, call rates will be preserved (and informative
missingness minimized) at the expense of accuracy. Those
who adopt this more liberal strategy accept that a non-
trivial number of poorly performing SNPs will survive the
quality-control process and might be disproportionately
represented among the most extreme association signals.
If resources are not to be wasted in fruitless validation
and replication studies, it becomes essential to subject
interesting signals to individual reviews of quality-control
parameters (especially visualization of the signal intensity
(cluster) plots) (BOX 2) before proceeding. one corollary is
that GWA data-sharing efforts should extend to the pro-
vision of raw signal data as well as ‘finished’ genotypes.
once armed with a set of called genotypes, the
final phase of quality control beckons. experience has
shown that most SNPs showing extreme departures
from Hardy–Weinberg equilibrium (HWe) in controls can
be safely discarded1, although lesser (but nevertheless
quite marked) departures are to be expected under the
null hypothesis, given the number of tests performed.
Appropriate thresholds for any given study will depend
on the sample size and overall data quality, and might best
be defined by using the observed distribution of HWe
statistics (the WTCCC used this approach to set a thresh-
old at an exact p < 5.7 x 10–7 in controls)1. In any event,
HWe is an imprecise tool for quality control purposes.
Tests of departure from HWe are underpowered for the
detection of genotyping error75,76, whereas overenthusi-
astic use of HWe as a quality criterion can prove to be
counterproductive given that modest disequilibrium (in
cases particularly) can be a signature of true association.
Recent studies have emphasized the importance of
detecting individuals whose GWA data reveal an ancestry
that is discrepant with self-described labels, allowing them
to be removed from consideration1 or analysed separately.
Similarly, sample integrity needs to be confirmed using
recorded gender or previously obtained genotypes.
Inadvertent sample duplication and swaps, cross-contami-
nation and cryptic relatedness43 are frequently revealed by
analysis of GWA data, and, if unresolved, would violate
assumptions of statistical independence and introduce
misclassification effects. It is straightforward to identify
first- and second-degree relatives at least and exclude
one member of each pair from analysis. In the presence
of more remote relationships, options include explicit
modelling of those relationships, or adjustment through
This sequence of manoeuvres has proved challenging
enough for experienced groups given the size of the data
sets concerned. However, all groups working with such
data, whether generated themselves or downloaded from
the web, need to appreciate the intricacies of data quality
control if they are to avoid potential misinterpretation.
The comments above refer to the implementation of SNP-
based scans, but the task of converting intensity traces
and SNP-based data into multi-allelic CNv genotypes is
far more demanding and is only now being tackled65.
Analysis and interpretation
Diagnostic plots. A limited number of formats have
emerged as standard tools for representing the data
emerging from a GWA scan, with the quantile-quantile
plot (Q-Q plot) among the most widely used52,77 (BOX 2).
These plots help to indicate whether the study has gener-
ated more significant results than expected by chance and
to put such findings in context. undetected population
stratification or cryptic relatedness result in deviation
from the null across the entire distribution, whereas
large-effect susceptibility loci generate deviations at the
highly significant end of the range.
Single-point analyses. In most situations, the most power-
ful tool for the analysis of GWA data has been a single-
point, one degree of freedom test of association, such as the
Cochran–Armitage test. Such tests allow comparison of
the genotype distributions of cases and controls at each
SNP in turn, and can be conducted with or without
adjustment for relevant covariates, such as the principal
components of population substructure42. Although the
Cochran–Armitage method directly tests only one of sev-
eral possible genetic models, it has the merit of being robust
to modest deviations from additivity on the logistic scale
(at least to those most likely to be biologically relevant).
Furthermore, in situations in which the true model at the
causal variant is non-additive, even modest departures
from perfect lD will result in greatly reduced power to
detect that non-additivity at nearby variants: in the GWA
context therefore, in situations when few causal variants
will be directly typed, the additive model is likely to per-
form well. Whereas the use of alternative models (general,
dominant or recessive) could result in enhanced detection
of some signals78, the use of multiple correlated tests also
complicates computation of type 1 error rates and can
reduce the efficiency of subsequent follow-up efforts.
Historically, interpretation of genetic association find-
ings has adopted the standard frequentist approach to the
evaluation of significance. From such a view-point, GWA
results are compared against a single criterion of genome-
wide significance. Although several benchmarks have
been proposed, in european-descent GWA studies opin-
ion is coalescing around the need to adjust for 1–2 million
independent tests, which results in a target α (p value) of
~5 x 10–8 (Refs 49,79,80). However, such an approach fails
360 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
Box 2 | Visualization of genome-wide association data
Quantile-quantile (Q-Q) plots provide a visual summary of the distribution
of the observed test statistics generated by a genome-wide association
(GWA) study52,77. Typically, a single test statistic (for case–control studies,
a chi-squared (χ2) comparison of absolute genotype counts) is calculated
for each variant passing quality control. In panels a–d, the blue line
denotes expectation under the null and red circles indicate idealized test
results from hypothetical GWA data, generated under four scenarios: in
panel a the observed data conforms closely to expectation indicating
little evidence for association; in panel b inflation of the observed findings
across the distribution is seen, indicative of population stratification or
cryptic relatedness; in panel c there is similar evidence of population
substructure, but some suggestion of an excess of strong associations; in
panel d there is little evidence of substructure, but compelling evidence
for an excess of disease associations.
Signal intensity (cluster) plots provide diagnostics at the level of
individual SNPs. Typically, raw data from the genotyping platform is
plotted along two axes (one for each allele) to define clusters of data
corresponding to the three genotype groups. Panels e–h display idealized
plots based on ~200 genotypes. In panel e the three clusters are well
defined and individual genotypes are accurately called (as shown by the
three colours). In panel f the clusters are well defined, but an error in
allele calling has led to two clusters being assigned the same genotype.
In panel g significant overlap between clusters is likely to result in failure
to call certain genotypes (shown in open symbols in panel h). In this
example, all failed genotypes are either heterozygotes or homozygotes
for the green allele: this generates biased estimates of genotype
frequencies, which can result in spurious association signals owing to
Finally, genome-wide Manhattan plots display GWA findings with
respect to their genomic positions, highlighting signals of particular
interest. In panel i, an example from the type 2 diabetes component of the
Welcome Trust Case Control Consortium study1,5, the strongest
associations are seen on chromosomes 10 (transcription factor 7-like 2;
TCF7L2), 16 (fat mass and obesity associated; FTO) and 6 (CDK5 regulatory
subunit associated protein 1-like 1; CDKAL1). Additional strong signals
on chromosomes 1, 2 and 12 did not replicate.
0 10 20 30 40 50 60 70 80
Signal intensity X (arbitary units)
15 20 25 30
15 2025 30
15 2025 30
0 10 20 30 40 50 60 70 80
Signal intensity X (arbitary units)
010 20 30 40 50 60 70 80
Signal intensity X (arbitary units)
0 1020 30 40 50 60 70 80
Signal intensity X (arbitary units)
Nature Reviews | Genetics
Signal intensity Y (arbitary units) Signal intensity Y (arbitary units) Signal intensity Y (arbitary units)Signal intensity Y (arbitary units)
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 361
© 2008 Nature Publishing Group
The Bayesian alternative to
approaches to hypothesis
testing, essentially equivalent
to likelihood ratio tests: prior
and posterior information are
combined in a ratio that
measures the strength of the
evidence in favour of one
model rather than the other.
The probability that a reported
association between a genetic
variant and a trait of interest is
Association methods that rely
on the relationship between
the distribution of estimated
haplotype frequencies and trait
status, rather than each
individual variant in turn.
A set of approaches for filling in
missing genotype data using a
sparse set of genotypes (for
example, from a GWA scan)
and a scaffold of linkage
disequilibrium relationships (as
provided by the HapMap).
to account for factors such as the power of a study and the
number of likely true positives, which are important com-
ponents of any comprehensive evaluation of GWA find-
ings81. Consider a small case–control GWA performed
for a condition for which there is only weak evidence for
genetic involvement: any associations found are unlikely
to be genuine, whatever the p value obtained. Such
limitations have excited interest in Bayesian approaches,
which incorporate information on the likely number
of true associations, and the power of a given study to
detect associations of a given magnitude, into estimates of
credibility81,82. These methods, which generate measures
such as Bayes’ factors, or the false-positive report probability
rather than p values, avoid a single threshold for genome-
wide significance but depend on the ability to assign
plausible probabilities to each of the alternate hypoth-
eses (see Ref.1 for further discussion). Nevertheless,
both approaches conclude that only very low p values
equate to strong evidence for association (that is, are
associated with a low false-positive rate). Crucially, most
GWA signals that have attained such significance levels
have subsequently been confirmed by replication1.
Multi-marker analyses. As expected, given the incom-
plete coverage of common variation that is provided by
contemporary GWA platforms61,62, a modest boost to
power can be provided by computational approaches that
improve the detection of associations that are attribut-
able to variants that have not themselves been directly
typed83. Haplotype-based methods and imputation methods
are related but complementary approaches to achieve
this. In situations when haplotype-based analyses83,84
reveal evidence for association that exceeds that of any
directly typed SNP in the vicinity (after allowance for the
increased dimensionality), one can invoke either an effect
that is directly attributable to the haplotype (that is, inde-
pendent causal cis effects at multiple SNPs) or the expla-
nation that the haplotype tags more efficiently than any
individual genotyped SNP, an as yet untyped aetiological
variant. Imputation methods63,64 rely on information from
sets of resequenced and/or densely genotyped individuals
to infer missing genotypes at untyped variants. Because
data from the International HapMap Consortium49 are
typically used as the reference, imputation methods have
proved most powerful in recovering associations and
causal effects attributable to HapMap SNPs that are not
included on commercial arrays. Importantly, the use of
such methods is not restricted to samples drawn from
HapMap reference populations85.
Challenges. Immediate challenges in this area are numer-
ous. The imminent arrival of large-scale genome-wide
CNv data65 has focused attention on the development of
methods that are suited to the specific features of such
data (multiallelic, semi-quantitative and probably more
error prone) and methods that facilitate the integration of
CNv and SNP information. There is work to be done to
understand how best to incorporate the intrinsic uncer-
tainty that is associated with genotype calls derived from
both direct and, in particular, imputed data63,64, and in the
development of analysis tools that take account of, and/or
estimate, information on population history, which will
prove especially valuable for studies in genetic isolates86.
evaluation of the contribution of rare variants to com-
mon disease susceptibility raises issues related to detection
(rare variants are poorly captured by the standard GWA
arrays87) and functional assessment (the sheer number
of such variants and the limited power to test them for
association88). Finally, there is a need for improved meth-
ods to estimate the joint effects of multiple genes and/or
environmental exposures on disease predisposition. Such
analyses raise both computational and statistical issues,
related to the scale and complexity of the data and the
large number of hypotheses that could be addressed89.
Validation and replication
The importance of replication. The use of small sam-
ples, which are underpowered to detect loci of realistic
effect size, and over-liberal declarations of association
are the main reasons why so few of the complex-trait
associations that were claimed in the pre-GWA era
proved genuine81,90. This history, together with the high
dimensionality of GWA studies, their vulnerability to a
range of errors and biases, and the modest effect sizes
to be anticipated for most complex-trait susceptibility
alleles, help to explain the pre-eminent role of replication
in the evaluation of GWA findings91,92.
Terminology in this area can be confusing. Here,
we use ‘technical validation’ to refer specifically to the
reanalysis of original GWA samples using a second
genotyping platform. Technical validation allows early
detection of technical errors in typing or imputation that
might have generated a spurious association signal, and
is an important prelude to large-scale efforts to evalu-
ate selected signals in additional independent samples
(referred to here as ‘replication’).
Best practice. The aim of validation and replication is to
determine which of the findings arising from the primary
GWA reflect true reproducible associations. Accordingly,
the focus is not merely to provide additional evidence to
support or refute the original association, but also the
systematic appraisal of potential sources of error and bias
that could have been responsible. Hence, it is important
to use independent replication samples, and to use dis-
tinct genotyping assays to expose technical artefacts.
Credibility is increased when multiple investigative groups
find the same association in independent samples.
Claims of replication should be reserved for findings
involving the same allele or haplotype (or an established
proxy thereof), the same phenotype and the same genetic
model as the original signal. otherwise the risk of spurious
claims of association is increased by the testing of multiple
hypotheses. This has an important bearing on decisions
about the extent to which early replication efforts at a
given signal should focus exclusively on the index variant,
as opposed to including additional adjacent SNPs93. The
inclusion of adjacent SNPs runs the risk of generating a
profusion of apparent associations around spurious GWA
signals — complicating interpretation of the evidence
for replication — and can involve a substantial waste of
genotyping effort around the many false-positive signals.
362 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
In most instances, therefore, it seems wise first to obtain
definitive evidence of association at the index variant. So
that failure to replicate can be meaningfully interpreted,
the samples used in early rounds of replication should be
broadly similar (with respect to ethnicity and ascertain-
ment) to those of the original study. of course, as loci are
validated as truly causal, extension into multiple ethnici-
ties is highly desirable to test the generalizability and con-
sistency of the proposed association, although such efforts
will probably entail additional efforts at variant discovery
given ethnic differences in lD.
Multi-stage designs. Several recent reports have empha-
sized the potential advantages of multi-stage designs, in
which signals from an initial, first-stage GWA are used
to define a subset of SNPs that are retyped in additional
second-stage samples94–96. Such designs have been seen as
an effective way of retaining power while reducing geno-
typing costs. However, the substantial price differential
between commodity and custom genotyping means that
those cost benefits can be less dramatic than comparisons
of genotype numbers alone would suggest. In circum-
stances in which the second-stage and GWA samples have
similar provenance, so that the prospects for appreciable
genetic heterogeneity between the stages are low, the
best-powered analytical strategy involves a joint analysis
(in which the distribution of test statistics across the data
from both stages combined is considered), rather than
the conventional replication design (which considers the
second-stage results in isolation)94. Such considerations
blur the boundaries of where exactly replication starts, but
whichever analytical approach is taken, confirmation in
many independent samples is important and it is the over-
all strength of the evidence of association that matters.
Power, sample size and heterogeneity. An appreciation of
power and sample size is central to the design and interpre-
tation of appropriate replication studies. Studies that lack
the power to offer convincing support or refutation of the
original finding can generate misleading inferences when
considered in isolation, although combinations of such
studies might be of value provided that all suitable studies
have been included. Calculations of replication sample
size need to consider the so-called ‘winner’s curse’ effect,
whereby the original study will typically overestimate
the true effect size97. Replication efforts that fail to make
such allowance will probably be underpowered98,99.
If well-performed replication studies confirm the
original findings, then the evidence in favour of associa-
tion is enhanced (unless of course both the original and
replication studies have succumbed to the same errors).
Interpretation of a failure to replicate is more difficult. If
it is clear that the replication studies were well powered
and well performed, and that there is genuine divergence
between the effect-size estimates (for example, no overlap
of 95% confidence intervals), then there are two possible
explanations. either the original finding was wrong, or
the difference in findings is attributable to some source of
heterogeneity100,101. The list of potential causes of hetero-
geneity is long: it includes variable patterns of lD between
the genotyped SNP and untyped causal alleles (although
this is unlikely if the samples are of similar ancestry); dif-
ferences in the distribution, frequency or effect size of the
causal alleles at a given locus (due to, for example, drift or
selection, or differences in case ascertainment); and the
impact of non-additive interactions with other genetic
variants or environmental exposures.
We should be wary of appealing to heterogeneity as
a rationale for failure to replicate, as over-eagerness to
deploy such an explanation would mean that no report
of association could ever be refuted. Nevertheless, there
are established instances in which the effects of proven
associations can vary substantially across studies, so
clearly circumstances exist where it can be justified.
The role of variants in the fat mass and obesity associ-
ated (FTO) gene on the risk of diabetes and obesity (BOX 3)
provides one such example, illustrating how a clear case
for heterogeneity can be made, especially when the initial
association finding has been robustly replicated and the
source of the heterogeneity is apparent1,5,36,37. Some sources
of heterogeneity can be directly evaluated, including dif-
ferences in lD structure between populations (particu-
larly once the true causal allele has been identified), clear
variation in case ascertainment that can be correlated
with effect size or a highly significant interaction with a
well-measured covariate. other sources of heterogeneity,
such as an unmeasured or poorly defined environmental
exposure, will be difficult or impossible to demonstrate.
The number of associations for which heterogeneity has
an important role cannot be reliably estimated from the
evidence to date: because our inventory of ‘proven’ com-
plex-trait associations is heavily biased towards variants
that have shown consistent replication across studies (this
being such an important factor in determining proof),
there is likely to be a substantial under-representation
of causal loci that feature appreciable between-study
Box 3 | Informative heterogeneity
Replication of results is an essential step in establishing that the associations
revealed by initial genome-wide association (GWA) studies are genuine. Failure to
replicate (in an otherwise well-performed and well-powered study) is usually
interpreted as indicating that the initial finding was spurious. However, failure to
replicate can also result from substantive differences between the discovery and
replication studies, for example, with respect to sample ascertainment.
A recent example is provided by the highly significant association between
variants in the fat mass and obesity associated (FTO) gene and type 2 diabetes that
was first detected in a UK GWA scan (odds ratio for diabetes ~1.27, p = 2 x 10–8) and
subsequently strongly replicated in other UK samples (odds ratio for diabetes ~1.22,
p = 5 x 10–7)1,5,36. However, this association could not be replicated in several well-
powered diabetes GWA scans4,6–8. The explanation for these divergent findings
derives from the fact that FTO was shown to influence diabetes risk through a
primary effect on weight regulation, such that the diabetes risk allele is associated
with higher fat mass, weight and body mass index36,37. In the UK studies, marked
differences in weight between diabetic cases and non-diabetic controls meant that
differences in FTO genotype frequencies were observed in diabetes case–control
analyses. Several other diabetes GWA scans had explicitly targeted case selection
on relatively lean individuals4,7 (to remove the confounding effects of obesity),
thereby abolishing the differences in weight between the diabetic cases and
controls, and with it the between-group differences in FTO genotype distributions.
Rather than dismissing the FTO association with type 2 diabetes as a failure of
replication, identification of the source of the heterogeneity (what might be termed
informative heterogeneity) provided an explanation for the discrepant findings and
highlighted the likely mechanism of its action.
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 363
© 2008 Nature Publishing Group
heterogeneity102. Moreover, estimation of between-study
heterogeneity is highly imprecise when the number of
studies is limited (less than 10) and/or their individual
power is low. Systematic efforts to define the impact of
gene–gene and gene–environment interactions, well-
powered GWA studies in multiple ethnicities and com-
prehensive exploration of variation around strong signals
of association should all help to address this point.
Finding additional susceptibility genes
Meta-analysis. Individual GWAs are underpowered to
detect all but the biggest effects, and the susceptibility
variants identified so far are probably only a subset of
the loci that would be detectable using this approach
if power were increased39. Joint (meta) analysis of data
from comparable GWA scans9,34,35,38,103 provides a low-
cost approach to enhance power for both main and
joint (gene–gene and gene–environment) effects, obtain
in silico replication, inform SNP selection for subsequent
replication efforts and explore potential sources of heter-
ogeneity. In type 2 diabetes, joint analysis of three GWA
scans (4,700 cases and 5,700 controls when combined)5–7,9
was central to the robust identification of several novel
susceptibility variants, and allowed confirmation of
the role of previously known susceptibility variants
of modest effect size, such as those in KCNJ11 (potassium
inwardly-rectifying channel, subfamily J, member 11)
and PPARG (peroxisome proliferator-activated
As increasingly large data sets are deployed to find
signals for which smaller samples were underpowered,
the average effect size of the novel variants that are dis-
covered will decrease. For some purposes, such as pre-
dictive diagnostics, this has implications for their likely
translational impact. even so, it is worth remembering
that owing to imperfect lD, a modest finding from the
initial GWA might ultimately lead to fine mapping of
variants with considerably larger effect sizes. When it
comes to insights into disease pathogenesis, however,
locus effect size is almost immaterial: even loci with
modest effects, once they are established as genuine, can
reveal novel causal mechanisms. Thus, discovery efforts
are likely to continue to bear fruit as long as the clinical,
logistical and financial resources are available to support
them, and, as genotyping costs fall, ever larger studies
will become possible.
The technical aspects of successful data combination
have been widely discussed in the meta-analysis litera-
ture100,104. Clearly, researchers seeking to obtain an unbi-
ased estimate of the significance of a given association
are duty-bound to be as inclusive as possible90. Although
it might be tempting to exclude particular studies on the
basis of some perceived failure of quality or potential het-
erogeneity, any such decisions must be carefully justified.
In the presence of heterogeneity, random-effects models
provide more appropriate estimates of overall effect size,
because heterogeneity violates the basic assumption of
fixed-effects analysis100. Although summary-level data
will suffice for many meta-analysis purposes, access to
primary individual-level data allows for more sophisti-
cated reanalysis, including the capacity to undertake hap-
lotype and conditional analyses, to perform imputation,
to examine the joint effects of genes and environment,
and to explore phenotypic heterogeneity.
Reintroducing biology. So far, most efforts at replication
have concentrated, not unreasonably, on the signals for
which the statistical evidence is strongest. However, the
efficient identification of additional susceptibility loci
with more modest effect sizes might benefit from the
integration of statistical evidence with some assessment
of functional candidacy. Several prioritization strategies
for defining variant subsets with increased ‘prior’ odds for
association can be envisaged, embracing aspects of bio-
logical candidacy (for example, variants mapping to genes
that interact with, or lie within the same pathways as, those
Box 4 | Strategies for resequencing within genome-wide association signals
Because genome-wide association (GWA) studies directly genotype only a small
proportion of the variants that segregate within the population examined, it is unlikely
that the causal variant(s) will be among those for which genotype data are available.
Imputation methods63,64 allow association analysis to be extended to a larger set of
variants (HapMap SNPs for which reliable imputation is possible), but this still comprises a
minority of all common SNPs, and representation of rare variants is poorer still87. Targeted
resequencing allows recovery of a more complete inventory of sequence variation
within regions of interest, and enables systematic fine-mapping efforts to identify those
putatively causal variants with the strongest effects on disease susceptibility.
For any given region, the extent and scope of resequencing efforts depends on the
strategic goal. Consider the example of the fat mass and obesity associated (FTO) region
that is implicated in obesity and type 2 diabetes (see the figure)1,5,36. Association analysis
using directly typed (black) and imputed (grey) SNPs localized the association signal to a
47 kb region (A–B) defined by flanking recombination hot spots (red trace), within intron
1 of the FTO gene. If the goal is to identify common causal variants that could underlie
this index signal, it should suffice to resequence this ~50 kb interval in ~200 individuals
(a total of ~10 Mb). If the aim is, instead, to identify all common variants contributing to
regional susceptibility (including those independent of the index signal, so-far missed
owing to incomplete coverage) then all sequence that is relevant to gene expression or
function (arbitrarily, C–D) needs to be considered (~600 kb, hence a total of ~120 Mb).
Should the aim extend to identification of rare variants with independent effects on
disease risk, recovery of the full allelic spectrum of sequence variants will require deep
resequencing in ~500–1000 individuals: in the first instance, such deep resequencing
efforts might be preferentially targeted to exonic and conserved non-coding sequence.
AKTIP, AKT interacting protein; CHD9, chromodomain helicase DNA binding protein 9;
IRX3, iroquois homeobox 3; IRX5, iroquois homeobox 5; RBL2, retinoblastoma-like 2;
RPGRIP1L, retinitis pigmentosa GTPase regulator interacting protein 1-like.
Nature Reviews | Genetics
cM per Mb
cM from hit SNP
51.5 52.052.5 53.0 53.5
Chromosomal position (Mb)
364 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
that were previously implicated in susceptibility to the dis-
ease of interest) and genomic annotation (for example, all
variants capturing variation in microRNA target sites).
Other populations. Among the GWA scans that have
already been performed there has been a strong bias
towards samples of North european origin1–38. There
are excellent reasons for extending analyses to samples
from populations with differing mutational and demo-
graphical histories, including other major ethnic groups49
and population isolates86. each such sample offers new
opportunities in terms of detectable susceptibility vari-
ants, generating ethnic-specific patterns with respect to
the identity of the loci themselves, the frequency and
lD relationships of disease-susceptibility alleles, and the
presence of additional genetic or environmental factors
with which they might interact. Studies in other popula-
tions are therefore capable of revealing novel suscepti-
bility loci and aetiological pathways, as well as assisting
with the fine-mapping of causal variants within those loci
already confirmed105. Given the large sample sizes that are
needed to achieve these ends, the ascertainment of GWA
and replication sample sets from diverse ethnic groups
is a priority.
Other sequence variants. Finally, it is worth reiterating
that, so far, GWA scans have focused almost exclusively
on the detection of effects that are attributable to com-
mon SNPs. The first wave of GWA arrays offered limited
power to capture structural variants106 and rare variants
of any type87. The arrival of tools that are better-suited
to the direct evaluation of these variant types is likely to
provide the most immediate source of novel susceptibil-
ity loci, although each brings its own set of technical and
Following-up confirmed signals
The confirmed signals emerging from GWA scans and
subsequent replication efforts are just that — associa-
tion signals. The causal variants will only occasionally
be among those that are directly typed in GWA scans,
and the interval within which the aetiological variant(s)
are expected to lie (typically defined in terms of flank-
ing recombination hot spots) can be sizeable, often con-
taining several genes1,24. even worse, because it seems
likely that many complex-trait susceptibility effects are
mediated through remote regulatory elements107, the
coding exons of the susceptibility gene could lie well
beyond the interval of maximal association.
Resequencing and fine mapping. The task of moving
from confirmed association signal to complete enu-
meration of the pattern of causal variants at a given
locus poses significant challenges. on occasion, useful
shortcuts to exhaustive fine mapping might be avail-
able. The set of associated variants might include a
number with particularly strong biological credentials
— a non-synonymous coding SNP in a compelling bio-
logical candidate, for example. Alternatively, clues can
be gathered from expression studies108–110: the cluster of
associated variants might display strong cis associations
with expression of one of the nearby genes, transcript
levels of which are themselves associated with the phe-
notype of interest13,24. Although caution is warranted
in placing weight on in silico or in vitro functional data
(the concordance between epidemiological associa-
tions and measurable functional effects has historically
been poor111), such findings can provide a rapid route
towards direct functional confirmation of the implicated
In the absence of such insights, further progress will
generally require exhaustive examination of the region,
first, to generate a comprehensive inventory of regional
variation, and second, to use fine-mapping approaches to
define the signals with the strongest statistical claims112.
Despite recent advances in sequencing and genotyping,
both are daunting tasks.
Resequencing plans need to be tailored to the
desired objective (BOX 4). Implementation raises a host
of unanswered issues related to optimal study design
and data interpretation. There are numerous choices to
be made: first, about the samples to be resequenced, that
is, the balance of HapMap individuals versus disease
cases and whether to favour cases carrying the known
susceptibility variant or haplotype; second, the depth of
resequencing to be undertaken, which defines the allele
spectrum recovered; and finally, the merits of extend-
ing the search for rare variants beyond well-annotated
sequence, given the difficulties associated with obtain-
ing robust evidence for a statistical or functional effect
for rare variants in unannotated sequence (BOX 4).
Related challenges lie in the development of effi-
cient strategies for fine mapping that take into account
the desire both to discriminate between highly corre-
lated variants (to determine which variants are causal)
and to search for additional, independent signals.
Given the relatively high price of custom genotyping,
Box 5 | Challenges in following-up confirmed associations
Dramatic advances in identifying gene variants that influence human complex traits
have yet to be accompanied by consistent progress in understanding the mechanisms
by which these variants influence disease. For most associations, systematic efforts to
identify the underlying causal variant or variants have yet to be reported. There is a
clear need to establish tools — both bioinformatic and experimental — to support
efforts to map confirmed associations and thereby maximize the biological
information gathered from GWA studies. On the bioinformatics side, for example,
tools are required to display association data in the context of the increasingly rich
functional annotation of the genome (for example, the WGAViewer126).
On the experimental side, the availability of genome-wide profiles of gene
expression and alternate splicing across a range of human tissues and/or cell lines,
alongside dense genotype data from the same samples, would be a valuable
resource. Even for major cell types these do not currently exist108–110. The paucity of
such information for multiple brain tissues and defined populations of primary
lymphocytes, for example, impedes progress in defining causal mechanisms in
neuropsychiatric, infectious and autoimmune diseases. In some situations, we might
expect causal variants to have quite marked effects on expression phenotypes; in
other situations, their molecular consequences will be subtle, or restricted in terms
of space (for example, tissue specific) or time (for example, to particular periods of
development). Although it is straightforward to resequence the exons of a gene
of interest to screen for non-synonymous coding variants, establishing whether a
variant some hundreds of kilobases from the gene could be exerting modest effects
on the expression of specific transcripts in particular tissues is far more challenging,
and far less amenable to large-scale genome-wide analysis.
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 365
© 2008 Nature Publishing Group
An analytical approach that
allows one to test for a causal
relationship between two
phenotypes that show
observational associations, but
are subject to confounding:
makes use of the random
segregation of susceptibility
alleles at meiosis to explore
causality in a model that is
freed from most sources of
exhaustive fine-mapping efforts across multiple sig-
nals can cost far more than the original GWA scan.
Multi-stage approaches to fine mapping (homing
in on causal variants through successive rounds of
genotyping, whereby ever fewer SNPs are typed in
increasingly large samples) might seem efficient but
they can involve costly, time-consuming investment
in successive assay designs. Given inter-ethnic differ-
ences in patterns of lD and mutational histories (see
above), samples from other ethnicities (especially
those of African origin) should be extremely valuable
tools for fine mapping, and considerable effort is being
expended to identify samples on the scale required for
Function and translation. once all the statistical evi-
dence has been extracted and putative causal variants
are identified, substantial challenges remain (BOX 5).
Prominent among these is the need to obtain func-
tional confirmation that the variants implicated are
truly causal, and to reconstruct the molecular and
physiological mechanisms whereby they have an impact
on the phenotype of interest. Because many of the
complex-trait susceptibility variants so far identified map
to sequence of unknown function that is some distance
from the nearest coding sequence5–9,15,17,25–27, the design
of contextually appropriate functional assays is far from
straightforward. Improved tools for the functional anno-
tation of the genome (for example, from the eNCoDe
project)107 will be essential, but generating such annota-
tions for all tissues and/or cells of biomedical relevance
remains a monumental task.
It is also important to move beyond the selected
samples used in the discovery phase and evaluate the
impact of the variants on unbiased population sam-
ples. From an epidemiological perspective, consensus
guidelines have been proposed to grade the strength of
the epidemiological evidence for a given association113.
These take into account the extent of the evidence, the
consistency of the replication and the protection from
bias in the accumulated data. Population-based studies
will enable research to establish whether, using studies of
the joint effects of genes and environment and Mendelian
randomization approaches114, genetic discoveries can be
used to pinpoint modifiable environmental exposures.
Such analyses will require very large study samples for
which both genetic and environmental exposures are
Finally, the ultimate objective of genetic research
lies in the translation of the findings into advances in
clinical care (BOX 6). The mechanistic insights generated
by gene discovery might identify new therapeutic tar-
gets and lead to novel pharmaceutical and preventative
approaches. In addition, there is growing expectation
that individual patterns of genetic predisposition will
be of value in health-care delivery (personalized medi-
cine). For many, but not all115, diseases the modest effect
sizes of the variants emerging from GWA studies limits
the degree to which this is possible (BOX 7). Higher pen-
etrance, lower frequency variants that are not detected
by current GWA approaches but are amenable to new
high-throughput sequencing efforts might prove more
valuable in this respect116.
Reporting and deposition
Because wide availability of data is central to the success
of many of these endeavours, there has been substan-
tial investment in structures to support data-sharing
between investigators, such as the National Institutes
of Health (NIH)-supported dbGAP programme117, and
the european Bioinformatics Institute-based european
Genotyping Archive. Moreover, the NIH recently
announced a policy for sharing of data obtained in NIH
supported or conducted GWA studies. Although aggre-
gate or summary data (genotype frequencies and asso-
ciation p values by group as provided by the WTCCC
Box 6 | Clinical translation of findings from GWA studies
Recent successes in the identification of susceptibility variants that underlie many
important biomedical phenotypes has increased confidence that this information
can be translated into clinically beneficial improvements in management. There
are two principal routes through which such translation might be effected (see
In the first, identification of novel causal pathways provides new opportunities for
clinical advances of generic benefit to all those suffering from (or at risk of) the
disease concerned. This might involve identification of therapeutic targets within
causal pathways, leading to novel therapeutic agents for treatment and/or
prevention. Identification of causal pathways should bolster efforts to identify
biomarkers, allowing improved disease prediction and monitoring of disease
progression and treatment response. Sometimes, genetic discoveries can highlight
important environmental contributors to disease, enabling public-health-based
disease prevention measures. Note that even modest genotype–phenotype
associations (provided they are confirmed as genuine through extensive replication
and functional studies) can offer significant new translational opportunities through
the identification of novel modifiable pathways.
The second translational route lies through using knowledge of individual patterns
of disease predisposition (for example, through genetic profiling) to develop more
personalized approaches to disease management. The major limitation here, for
most complex traits, is that the variants so far identified explain only a small
proportion of individual variation in disease risk39. Consequently, for most
individuals (all but the small proportion who have inherited extremely high or low
numbers of susceptibility alleles for a given disease), genetic profiling using
currently available markers provides limited information on disease risk beyond that
available from conventional risk factors. If genetic profiling is to become widely
applied in clinical practice, we need to: improve the accuracy of risk prediction
through identification of additional susceptibility variants; demonstrate, by
prospective studies, that profiling results in beneficial modifications of medical care
and/or personal responsibility; and establish an appropriate regulatory environment
for the use of such tests.
Nature Reviews | Genetics
Identification of susceptibility variants
Novel biological insights Improved measures of individual
Clinical advances Personalized medicine
366 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
and the National Cancer Institute Cancer Genetic
Markers of Susceptibility (CGeMS) studies, for exam-
ple1,118) provide an excellent and convenient substrate
for many data-sharing purposes, individual-level data
(including raw signal-intensity information) provide a
more valuable resource that supports a wider range of
analyses (see above).
lack of transparency and incomplete reporting of
both data and study methods have raised concerns in
many health research fields119–121, and poor reporting
has been associated with biased estimates of effect122.
The importance of transparency was emphasized by
the NCI-NHGRI Working Group on Replication in
Association Studies91. To help remedy this problem, some
groups have advocated the development of evidence-
based reporting guidelines similar to the consolidated
standards of reporting trials (CoNSoRT) guidelines
for the reporting of randomized clinical trials123. In
particular, the epidemiology community has recently
developed a reporting guidance (strengthening the
reporting of observational studies in epidemiology;
STRoBe) for cross-sectional, case–control and cohort
studies124: extension of this guidance to genetic associa-
tion studies is in an advanced stage of development (see
the Human Genome epidemiology Network (HuGeNet)
In addition to formal deposition mechanisms,
there has been a proliferation of investigator networks,
designed to exploit the potential for novel signal dis-
covery through integration of data from large numbers
of GWA scans that are informative for traits of inter-
est1,9,34,35,38,103,125. For phenotypes such as height and body
mass index, which are widely and reliably recorded in
many cohorts undergoing GWA analysis, aggregate
data from several tens of thousands of scans has facili-
tated detection of variants with small effects34,35,38. Such
data-sharing efforts will be vital for the success of the
coming wave of cohort-based GWAs, each of which
will allow the analysis of large numbers of phenotypes.
In this situation, focused replication of the best signals
for each trait becomes increasingly unattractive, and
enlarging sample size through accumulation of addi-
tional GWA data seems to be the most feasible option,
at least for initial discovery and validation efforts.
Management of these large, dynamic and overlapping
consortia can be challenging, particularly with respect
to ensuring that all those involved (particularly jun-
ior investigators) obtain appropriate credit for their
Summary and conclusions
The past year has seen a remarkable shift in our capac-
ity to dissect the genetic basis of common diseases and
continuous traits of biomedical significance. The GWA
approach has proven itself extremely well-suited to the
identification of common SNP-based variants with
modest to large effects on phenotype. Careful imple-
mentation and appropriate interpretation has resulted in
discoveries that have proven more robust than many had
anticipated. Growing numbers of novel susceptibility loci
have been identified, shedding light on the fundamental
mechanisms that influence disease predisposition, and
much is being learned about the complex relationships
between changes in genome sequence and phenotypic
However, we are far from the end of this particular
voyage, and recent discoveries are nothing more than
initial forays into the terra incognita of our genomes. We
remain unable to explain more than a small proportion
of observed familial clustering for most multifactorial
traits, a fact that emphasizes the need to extend analy-
sis to a more complete range of potential susceptibility
variants, and to support more explicit modelling of the
joint effects of genes and environment. Many of
the greatest challenges to be faced in the years ahead
lie not so much in the identification of the association
signals themselves, but in defining the molecular mecha-
nisms through which they influence disease risk and/or
Box 7 | Low-frequency variants and disease susceptibility
Genome wide association (GWA) studies are proving adept at identifying common
variants contributing to the inherited component of common diseases. Almost all
such variants seem to have modest effect sizes and, even when combined, their
impact on overall population variance and predictive power is limited127. There is a
marked disparity between the extent of overall familial aggregation observed for
many common diseases and that attributable to variants identified to date. In type 2
diabetes, known variants collectively account for a sibling relative risk of ~1.07 in
Europeans, way below the overall figure (~3) from epidemiological studies4–9.
Although the identification of additional common risk variants (at the already
identified loci, and the others that are yet to be found) will explain some of this deficit,
one emerging hypothesis anticipates that a significant proportion of this ‘missing
heritability’ will be attributable to low-frequency variants with intermediate
penetrance effects, which have been largely refractory to conventional
Consider a hypothetical variant with a minor allele frequency of 1% and an allelic odds
ratio of 3. Given a disease prevalence of 5%, the penetrance of the risk homozygote
(~45%) is too low to support Mendelian segregation and detection by traditional linkage
approaches. At the same time, the low risk-allele frequency means low detectability by
GWA1,87. Yet this variant has a stronger effect on familial risk than most known common
susceptibility variants: a locus-specific sibling relative risk of 1.038 comfortably exceeds
that of the strongest diabetes-susceptibility effects — that of transcription factor 7-like 2
(TCF7L2), which is strongly associated to diabetes, is approximately 1.025. As few as
thirty such variants across the genome would jointly generate a sibling relative risk >3,
and offer impressive predictive power (a discriminative accuracy of 77%)127. Novel
resequencing technologies, allied to large-scale association testing, provide the
potential to identify and characterize variants with these properties and evaluate their
contribution to disease risk. In the first instance, such efforts are likely to be targeted to
genes already implicated in disease susceptibility.
Nature Reviews | Genetics
Hard to identify
Highly unusual for
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 367
© 2008 Nature Publishing Group
Wellcome Trust Case Control Consortium. Genome-
wide association study of 14,000 cases of seven
common diseases and 3,000 shared controls. Nature
447, 661–678 (2007).
In this study, high density, genome-wide association
data on 17,000 individuals identified many novel
complex-trait susceptibility loci and explored key
methodological and technical issues relevant to the
Todd, J. A. et al. Robust associations of four new
chromosome regions from genome-wide analyses of
type 1 diabetes. Nature Genet. 39, 857–864 (2007).
Hakonarson, H. et al. A genome-wide association
study identifies KIAA0350 as a type 1 diabetes gene.
Nature 448, 591–594 (2007).
Sladek, R. et al. A genome-wide association study
identifies novel risk loci for type 2 diabetes. Nature
445, 881–885 (2007).
Zeggini, E. et al. Replication of genome-wide
association signals in UK samples reveals risk loci for
type 2 diabetes. Science 316, 1336–1341 (2007).
Scott, L. J. et al. A genome-wide association study of
type 2 diabetes in Finns detects multiple susceptibility
variants. Science 316, 1341–1345 (2007).
Diabetes Genetics Initiative. Genome-wide association
analysis identifies loci for type 2 diabetes and
triglyceride levels. Science 316, 1331–1336 (2007).
Steinthorsdottir, V. et al. A variant in CDKAL1
influences insulin response and risk of type 2 diabetes.
Nature Genet. 39, 770–775 (2007).
Zeggini, E., Scott, L. J., Saxena, R., Voight, B. &
DIAGRAM Consortium. Meta-analysis of genome-wide
association data and large-scale replication identifies
several additional susceptibility loci for type 2 diabetes.
Nature Genet. 30 Mar 2008 (doi:10.1038/nrg.120).
10. Parkes, M. et al. Sequence variants in the autophagy
gene IRGM and multiple other replicating loci
contribute to Crohn’s disease susceptibility. Nature
Genet. 39, 830–832 (2007).
11. Duerr, R. H. et al. A genome-wide association study
identifies IL23R as an inflammatory bowel disease
gene. Science 314, 1461–1463 (2006).
12. Rioux, J. D. et al. Genome-wide association study
identifies new susceptibility loci for Crohn disease and
implicates autophagy in disease pathogenesis. Nature
Genet. 39, 596–604 (2007).
13. Libioulle, C. et al. novel Crohn disease locus identified
by genome-wide association maps to a gene desert on
5p13.1 and modulates expression of PTGER4. PLoS
Genet. 3, e58 (2007).
14. Hampe, J. et al. A genome-wide association scan of
nonsynonymous SnPs identifies a susceptibility
variant for Crohn disease in ATG16L1. Nature Genet.
39, 207–211 (2007).
15. Gudmundsson, J. et al. Genome-wide association
study identifies a second prostate cancer susceptibility
variant at 8q24. Nature Genet. 39, 631–637 (2007).
16. Gudmundsson, J. et al. Two variants on chromosome
17 confer prostate cancer risk, and the one in TCF2
protects against type 2 diabetes. Nature Genet. 39,
This paper is one of the clearest demonstrations so
far of the potential for pleiotropy: the same
variants in TCF2 influence risk to both type 2
diabetes and prostate cancer.
17. Yeager, M. et al. Genome-wide association study of
prostate cancer identifies a second risk locus at 8q24.
Nature Genet. 39, 645–649 (2007).
18. Thomas, G. et al. Multiple loci identified in a genome-
wide association study of prostate cancer. Nature
Genet. 40, 310–315 (2008).
19. Gudmundsson, J. et al. Common sequence variants
on 2p15 and Xp11.22 confer susceptibility to
prostate cancer. Nature Genet. 40, 281–283 (2008).
20. Eeles, R. A. et al. Multiple newly identified loci
associated with prostate cancer susceptibility. Nature
Genet. 40, 316–321 (2008).
21. Easton, D. F. et al. Genome-wide association study
identifies novel breast cancer susceptibility loci.
Nature 447, 1087–1093 (2007).
22. Hunter, D. J. et al. A genome-wide association study
identifies alleles in FGFR2 associated with risk of
sporadic postmenopausal breast cancer. Nature
Genet. 39, 870–874 (2007).
23. Stacey, S. n. et al. Common variants on chromosomes
2q35 and 16q12 confer susceptibility to estrogen
receptor-positive breast cancer. Nature Genet. 39,
24. Moffatt, M. F. et al. Genetic variants regulating
ORMDL3 expression contribute to the risk of
childhood asthma. Nature 448, 470–473 (2007).
25. Helgadottir, A. et al. A common variant on
chromosome 9p21 affects the risk of myocardial
infarction. Science 316, 1491–1493 (2007).
26. McPherson, R. et al. A common allele on chromosome
9 associated with coronary heart disease. Science
316, 1488–1491 (2007).
27. Samani, n. J. et al. Genomewide association analysis
of coronary artery disease. N. Engl. J. Med. 357,
28. Gudbjartsson, D. F. et al. Variants conferring risk of
atrial fibrillation on chromosome 4q25. Nature 448,
29. Willer, C. J. et al. newly identified loci that influence
lipid concentrations and risk of coronary artery
disease. Nature Genet. 40, 161–169 (2008).
30. Kathiresan, S. et al. Six new loci associated with blood
low-density lipoprotein cholesterol, high-density
lipoprotein cholesterol or triglycerides in humans.
Nature Genet. 40, 189–197 (2008).
31. Kooner, J. S. et al. Genome-wide scan identifies
variation in MLXIPL associated with plasma
triglycerides. Nature Genet. 40, 149–151 (2008).
32. Weedon, M. n. et al. A common variant of HMGA2 is
associated with adult and childhood height in the
general population. Nature Genet. 39, 1245–1250
This paper demonstrates the power of the GWA
approach to identify genes influencing continuous
biomedical phenotypes, in this case, height.
33. Sanna, S. et al. Common variants in the GDF5-UQCC
region are associated with variation in human height.
Nature Genet. 40, 198–203 (2008).
34. Weedon, M. n. et al. Genome-wide association
analysis identifies 20 loci that influence adult height.
Nature Genet. (in the press).
35. Lettre, G. et al. Genome-wide association studies
identify 10 novel loci for height and highlight new
biological pathways in human growth. Nature Genet.
(in the press).
36. Frayling, T. M. et al. A common variant in the FTO
gene is associated with body mass index and
predisposes to childhood and adult obesity. Science
316, 889–894 (2007).
37. Scuteri, A et al. Genome-wide association scans shows
genetic variants in the FTO gene are associated with
obesity-related traits. PLoS Genet. 3, e115 (2007).
38. Loos, R. J. F. et al. Association studies involving over
90,000 people demonstrate that common variants
near to MC4R influence fat mass, weight and risk of
obesity. Nature Genet. (in the press).
39. Altshuler, D. & Daly, M. Guilt beyond a reasonable
doubt. Nature Genet. 39, 813–815 (2007).
40. Li, M., Boehnke, M. & Abecasis, G. R. Efficient study
designs for test of genetic association using sibship
data and unrelated cases and controls. Am. J. Hum.
Genet. 78, 778–792 (2006).
41. Howson, J. M., Barratt, B.J., Todd, J. A. &
Cordell, H. J. Comparison of population- and family-
based methods for genetic association analysis in
the presence of interacting loci. Genet. Epidemiol. 29,
42. Price, A. L. et al. Principal components analysis
corrects for stratification in genome-wide association
studies. Nature Genet. 38, 904–909 (2006).
43. Voight, B. F. & Pritchard, J. K. Confounding from
cryptic relatedness in case–control association
studies. PLoS Genet.1, e32 (2005).
44. Zheng, G., Freidlin, B. & Gastwirth, J. L. Robust
genomic control for association studies. Am. J. Hum.
Genet. 78, 350–356 (2006).
45. Paschou, P. et al. PCA-correlated SnPs for structure
identification in worldwide human populations.
PLoS Genet. 3, e160 (2007).
46. Tian, C. et al. Analysis and application of European
genetic substructure using 300K SnP information.
PLoS Genet. 4, e4 (2008).
47. Price, A. L. et al. Discerning the ancestry of European
Americans in genetic association studies. PLoS Genet.
4, e236 (2008).
48. Fellay, J. et al. A whole-genome association study of
major determinants for host control of HIV-1. Science
317, 944–947 (2007).
49. International HapMap Consortium. A haplotype map
of the human genome. Nature 437, 1299–1320
50. Laird, n. M. & Lange, C. Family-based designs in the
age of large-scale gene-association studies. Nature
Rev. Genet. 7, 385–394 (2006).
51. Chen, W. M. & Abecasis, G. R. Family-based
association tests for genomewide association scans.
Am. J. Hum. Genet. 81, 913–926 (2007).
52. Clayton, D. G. et al. Population structure, differential
bias and genomic control in a large-scale, case–control
association study. Nature Genet. 37, 1243–1246
This paper presents a detailed description of the
potential for bias and error to complicate the
analysis of large-scale genetic association data.
53. Plagnol, V., Cooper, J. D., Todd, J. A. & Clayton D. G.
A method to address differential bias in genotyping in
large-scale association studies. PLoS Genet. 3, e74
54. Cupples, L. A. et al. The Framingham Heart Study
100k SnP genome-wide association study resource:
overview of 17 phenotype working group reports.
BMC Med. Genet. 8, S1 (2007).
55. Ridker, P. M. et al. Rationale, design, and
methodology of the Women’s Genome Health Study:
A genome-wide association study of more than
25,000 initially healthy American women. Clin. Chem.
54, 249–255 (2008).
56. Li, S. et al. The GLUT9 gene is associated with serum
uric acid levels in Sardinia and Chianti cohorts. PLoS
Genet. 3, e194 (2007).
57. Cordell, H. J. & Clayton, D. G. Genetic association
studies. Lancet 366, 1121–1131 (2005).
58. Wong, M. Y., Day, n. E., Luan, J. A., Chan, K.P &
Wareham, n. J. The detection of gene–environment
interaction for continuous traits: should we deal
with measurement error by bigger studies or
better measurement? Int. J. Epidemiol. 32, 51–57
59. Wong, M. Y., Day, n. E., Luan, J. A. & Wareham, n. J.
Estimation of magnitude in gene–environment
interactions in the presence of measurement error.
Stat. Med. 23, 987–998 (2004).
60. Burke, W., Khoury, M. J., Stewart, A., Zimmern, R. L.
& Bellagio Group. The path from genome-based
research to population health: development of
an international public health genomics network.
Genet. Med. 8, 451–458 (2006).
61. Barrett, J. C. & Cardon, L. R. Evaluating coverage of
genome-wide association studies. Nature Genet. 38,
62. Pe’er, I. et al. Evaluating and improving power in
whole-genome association studies using fixed marker
sets. Nature Genet. 38, 663–667 (2006).
63. Marchini, J., Howie, B., Myers, S., McVean, G. &
Donnelly, P. A new multipoint method for genome-
wide association studies by imputation of genotypes.
Nature Genet. 39, 906–913 (2007).
64. Servin, B. & Stephens, M. Imputation-based analysis
of association studies: candidate regions and
quantitative traits. PLoS Genet. 3, e114 (2007).
65. McCarroll, S. A. & Altshuler, D. M. Copy-number
variation and association studies of human disease.
Nature Genet. 39, S37–S42 (2007).
This paper gives an excellent summary of the
challenges to be addressed if large-scale genetic
association studies are to be extended to CNVs.
66. Scherer, S. W. et al. Challenges and standards in
integrating surveys of structural variation. Nature
Genet. 39, S7–S15 (2007).
67. Weiss, L. A. et al. Association between microdeletion
and microduplication at 16p11.2 and autism. N. Engl.
J. Med. 358, 667–675 (2008).
68. Sham, P., Bader, J. S., Craig, I., O’Donovan, M. &
Owen, M. DnA pooling: a tool for large-scale
association studies. Nature Rev. Genet. 3, 862–871
69. Cargill, M. et al. A large-scale genetic association
study confirms IL12B and leads to the identification of
IL23R as psoriasis-risk genes. Am. J. Hum. Genet. 80,
70. Wang, W. Y., Barratt, B. J., Clayton, D. G. &
Todd, J. A. Genome-wide association studies:
theoretical and practical concerns. Nature Rev.
Genet. 6, 109–118 (2005).
71. Hirschhorn, J. n. & Daly, M. J. Genome-wide
association studies for common diseases and complex
traits. Nature Rev. Genet. 6, 95–108 (2005).
72. nicolae, D.L,. Wu, X., Miyake, K. & Cox, n. J.
GEL: a novel genotype calling algorithm using
empirical likelihood. Bioinformatics 22,
73. Rabbee, n. & Speed, T. P. A genotype calling
algorithm for affymetrix SnP arrays. Bioinformatics
22, 7–12 (2006).
74. Xiao, Y., Segal, M. R., Yang, Y. H. & Yeh, R. F.
A multi-array multi-SnP genotyping algorithm for
Affymetrix SnP microarrays. Bioinformatics 23,
368 | MAy 2008 | voluMe 9
© 2008 Nature Publishing Group
75. Wittke-Thompson, J. K., Pluzhnikov, A. & Cox, n. J. Download full-text
Rational inferences about departures from Hardy–
Weinberg equilibrium. Am. J. Hum. Genet. 76,
76. Cox, D. G. & Kraft, P. Quantification of the power of
Hardy–Weinberg equilibrium testing to detect
genotyping error. Hum. Hered. 61, 10–14 (2006).
77. Smyth, D. J. et al. A genome-wide association study of
nonsynonymous SnPs identifies a type 1 diabetes
locus in the interferon-induced helicase (IFIH1) region.
Nature Genet. 38, 617–619 (2006).
78. Lettre, G., Lange, C. & Hirschhorn, J. n. Genetic model
testing and statistical power in population-based
association studies of quantitative traits. Genet.
Epidemiol. 31, 358–362 (2007).
79. Risch, n. & Merikangas, K. The future of genetic
studies of complex human diseases. Science 273,
80. Hoggart, C. J. et al. Genome-wide significance for
dense SnP and resequencing data. Genet. Epidemiol.
32, 179–185 (2008).
81. Wacholder, S., Chanock, S., Garcia-Closas, M.,
El Ghormli, L. & Rothman, n. Assessing the
probability that a positive report is false: an approach
for molecular epidemiology studies. J. Natl Cancer
Inst. 96, 434–442 (2004).
This is an influential paper setting out the rationale
for a Bayesian interpretation of genetic association
findings, focusing on methods for establishing the
confidence with which any given positive
association can be regarded.
82. Wakefield, J. A Bayesian measure of the probability
of false discovery in genetic epidemiology studies.
Am. J. Hum. Genet. 81, 208–227 (2007).
83. De Bakker, P. I. et al. Efficiency and power in genetic
association studies. Nature Genet. 37, 1217–1223
84. Morris, A. P. A flexible Bayesian framework for
modeling haplotype association with disease, allowing
for dominance effects of the underlying causative
variants. Am. J. Hum. Genet. 79, 679–694 (2006).
85. De Bakker, P. I. et al. Transferability of tag SnPs in
genetic association studies in multiple populations.
Nature Genet. 38, 1298–1303 (2006).
86. Service, S. et al. Magnitude and distribution of linkage
disequilibrium in population isolates and implications
for genome-wide association studies. Nature Genet.
38, 556–560 (2006).
87. Zeggini, E. et al. An evaluation of HapMap sample
size and tagging SnP performance in large-scale
empirical and simulated data sets. Nature Genet.
37, 1320–1322 (2005).
88. Easton, D. F. et al. A systematic genetic assessment
of 1,433 sequence variants of unknown clinical
significance in the BRCA1 and BRCA2 breast
cancer-predisposition genes. Am. J. Hum. Genet. 81,
89. Marchini, J., Donnelly, P. & Cardon, L. R.
Genome-wide strategies for detecting multiple loci
that influence complex diseases. Nature Genet. 37,
90. Hirschhorn, J.n., Lohmueller, K., Byrne, E. &
Hirschhorn, K. A comprehensive review of genetic
association studies. Genet. Med. 4, 45–61 (2002).
91. nCI-nHGRI Working Group on Replication in
Association Studies. Replicating genotype–phenotype
associations: what constitutes replication of a
genotype–phenotype association, and how best can it
be achieved? Nature 447, 655–660 (2007).
This feature article is a thoughtful summary of the
main issues relating to replication of genetic
92. Lohmueller, K. E., Pearce, C. L., Pike, M., Lander, E. S.
& Hirschhorn, J. n. Meta-analysis of genetic
association studies supports a contribution of
common variants to susceptibility to common disease.
Nature Genet. 33, 177–182 (2003).
93. Clarke, G. M., Carter, K. W., Palmer, L. J., Morris, A. P.
& Cardon, L. R. Fine mapping versus replication in
whole-genome association studies. Am. J. Hum. Genet.
81, 995–1007 (2007).
94. Skol, A. D., Scott, L. J., Abecasis, G. R. & Boehnke M.
Optimal designs for two-stage genome-wide association
studies. Genet. Epidemiol. 31, 766–788 (2007).
95. Wang, H., Thomas, D. C., Pe’er, I. & Stram, D. O.
Optimal two-stage genotyping designs for genome-
wide association scans. Genet. Epidemiol. 30,
96. Müller, H. H., Pahl, R. & Schäfer, H. Including
sampling and phenotyping costs into the optimization
of two stage designs for genome wide association
studies. Genet. Epidemiol. 31, 844–852 (2007).
97. Zollner, S. & Pritchard, J. K. Overcoming the winner’s
curse: estimating penetrance parameters from
case–control data. Am. J. Hum. Genet. 80, 605–615
98. Yu, K et al. Flexible design for following up positive
findings. Am. J. Hum. Genet. 81, 540–551 (2007).
99. Gorrochurn, P., Hodge, S. E., Heiman, G. A.,
Durner, M. & Greenberg, D. A. non-replication of
association studies: ‘pseudo-failures’ to replicate?
Genet. Med. 9, 325–331 (2007).
100. Ioannidis J. P., Patsopoulos, n. A. & Evangelou, E.
Heterogeneity in meta-analyses of genome-wide
association investigations. PLoS ONE 2, e841 (2007).
101. Ioannidis J. P. non-replication and inconsistency in the
genome-wide association setting. Hum. Hered. 64,
102. Moonesinghe, R., Khoury, M. J., Liu, T. &
Ioannidis, J. P. Required sample size and
nonreplicability thresholds for heterogeneous
genetic associations. Proc. Natl Acad. Sci. USA 105,
103. The GAIn Collaborative Research Group. new models
of collaboration in genome-wide association studies:
the Genetic Association Information network. Nature
Genet. 39, 1045–1051 (2007).
104. Egger, M., Schneider, M. & Davey Smith, G. Spurious
precision? Meta-analysis of observational studies.
BMJ 316, 140–144 (1998).
105. Helgason, A. et al. Refining the impact of TCF7L2 gene
variants on type 2 diabetes and adaptive evolution.
Nature Genet. 39, 218–225 (2007).
106. Locke, D. P., et al. Linkage disequilibrium and
heritability of copy-number polymorphisms within
duplicated regions of the human genome. Am. J. Hum.
Genet. 79, 275–290 (2006).
107. EnCODE Project Consortium. Identification and
analysis of functional elements in 1% of the human
genome by the EnCODE pilot project. Nature 447,
This is a detailed examination of the functional
annotation of a subset of the human genome,
which reveals the complexity of genomic
108. Stranger, B. et al. Population genomics of human
gene expression. Nature Genet. 39, 1217–1224
109. Dixon, A. L. et al. A genome-wide association study
of global gene expression. Nature Genet. 39,
110. Goring, H. H. et al. Discovery of expression QTLs
using large-scale transcriptional profiling in human
lymphocytes. Nature Genet. 39, 1208–1216 (2007).
111. Ioannidis, J. P. & Kavvoura, F. K. Concordance of
functional in vitro data and epidemiological
associations in complex disease genetics. Genet. Med.
8, 583–593 (2006).
112. Lowe, C. E. et al. Large-scale genetic fine mapping and
genotype–phenotype associations implicate
polymorphism in the IL2RA region in type 1 diabetes.
Nature Genet. 39, 1074–1082 (2007).
113. Ioannidis, J. P. et al. Assessment of cumulative
evidence on genetic associations: interim guidelines.
Int. J. Epidemiol. 37, 120–132 (2008).
114. Davey Smith, G. & Ebrahim, S. ‘Mendelian
randomization’: can genetic epidemiology contribute
to understanding environmental determinants of
disease? Int. J. Epidemiol. 32, 1–22 (2003).
115. Zheng, S. L. et al. Cumulative association of five
genetic variants with prostate cancer. N. Engl. J. Med.
358, 910–919 (2008).
116. Stratton, M. R. & Rahman, n. The emerging
landscape of breast cancer susceptibility. Nature
Genet. 40, 17–22 (2008).
117. Mailman, M. D. et al. The nCBI dbGaP database
of genotypes and phenotypes. Nature Genet. 39,
118. Zheng, S. L. et al. Association between two unlinked
loci at 8q24 and prostate cancer risk among European
Americans. J. Natl Cancer Inst. 99, 1499–1501 (2007).
119. Von Elm, E. & Egger, M. The scandal of poor
epidemiological research. BMJ 329, 868–869
120. Brazma, A. et al. Minimum information about a
microarray experiment (MIAME) — toward standards
for microarray data. Nature Genet. 29, 356–371
121. Altman, D. & Moher, D. Developing guidelines for
reporting healthcare research: scientific rationale and
procedures. Med. Clin. (Barc). 125, 8–13 (2005).
122. Gludd, L. L. Bias in clinical intervention research.
Am. J. Epidemiol. 163, 493–501 (2006).
123. Altman, D. G. et al. The revised COnSORT
statement for reporting randomized trials:
explanation and elaboration. Ann. Intern. Med.
134, 663–694 (2001).
124. Von Elm, E. et al. The strengthening the reporting of
observational studies in epidemiology (STROBE)
statement: guidelines for reporting observational
studies. Lancet 370, 1453–1457 (2007).
125. Seminara, D. et al. The emergence of networks in
human genome epidemiology: challenges and
opportunities. Epidemiology 18, 1–8 (2007).
126. Ge, D. et al. WGAViewer: a software for genomic
annotation of whole genome association studies.
Genome Res. 3 Mar 2008 (doi:10.1101/
127. Janssens, A. C. J.W, Gwinn, M., Subramonia-Iyer, S. &
Khoury, M. J. Does genetic testing really improve the
prediction of future type 2 diabetes? PLOS Med. 3,
Preparation of this article was supported by funding from the
European Commission to the MolPAGE Consortium (LSHG-
CT-2004-512066: MMcC) and by research grants from the
national Institutes for Health (nHGRI and nHLBI; GRA). We
thank our colleagues — particularly P. Donnelly, J. Marchini,
J. Barrett, E. Zeggini, C. Lindgren, M. Boehnke, F. Collins,
C. Spencer and D. Altshuler for discussions and the reviewers
for their comments.
Competing interests statement
The authors declare competing financial interests: see web
version for details.
Entrez Gene: http://www.ncbi.nlm.nih.gov/entrez/query.
CDKAL1 | FTO | KCNJ11 | PPARG | TCF7L2
The McCarthy Group homepage:
Catalog of published genome-wide association studies:
Consolidated standards of reporting trials (CONSORT):
ENCODE project: www.genome.gov/10005107
European Genotyping Archive (EGA):
Genetic Association Information Network (GAIN):
Human Genome Epidemiology Network (HuGeNet):
International HapMap Consortium: www.hapmap.org
National Cancer Institute’s cancer genetic markers of
susceptibility (CGEMS) study: http://cgems.cancer.gov
Policy for sharing of data obtained in NIH supported or
conducted GWA studies: http://grants.nih.gov/grants/
Strengthening the reporting of observational studies in
epidemiology (STROBE): http://www.strobe-statement.org
Wellcome Trust Case Control Consortium:
All links Are Active in the online pdf
NATuRe RevIeWS | genetics
voluMe 9 | MAy 2008 | 369
© 2008 Nature Publishing Group