Multiethnic genetic association studies improve power for locus discovery.

Abstract

To date, genome-wide association studies have focused almost exclusively on populations of European ancestry. These studies continue with the advent of next-generation sequencing, designed to systematically catalog and test low-frequency variation for a role in disease. A complementary approach would be to focus further efforts on cohorts of multiple ethnicities. This leverages the idea that population genetic drift may have elevated some variants to higher allele frequency in different populations, boosting statistical power to detect an association. Based on empirical allele frequency distributions from eleven populations represented in HapMap Phase 3 and the 1000 Genomes Project, we simulate a range of genetic models to quantify the power of association studies in multiple ethnicities relative to studies that exclusively focus on samples of European ancestry. In each of these simulations, a first phase of GWAS in exclusively European samples is followed by a second GWAS phase in any of the other populations (including a multiethnic design). We find that nontrivial power gains can be achieved by conducting future whole-genome studies in worldwide populations, where, in particular, African populations contribute the largest relative power gains for low-frequency alleles (<5%) of moderate effect that suffer from low power in samples of European descent. Our results emphasize the importance of broadening genetic studies to worldwide populations to ensure efficient discovery of genetic loci contributing to phenotypic trait variability, especially for those traits for which large numbers of samples of European ancestry have already been collected and tested.

Full text

Page 1

Multiethnic Genetic Association Studies Improve Power
for Locus Discovery
Sara L. Pulit1,2, Benjamin F. Voight2, Paul I. W. de Bakker1,2,3,4*
1Division of Genetics, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, Massachusetts, United States of America, 2Program in
Medical and Population Genetics, Broad Institute of Harvard and Massachusetts Institute of Technology, Cambridge, Massachusetts, United States of America,
3Department of Medical Genetics, University Medical Center Utrecht, Utrecht, The Netherlands, 4Julius Center for Health Sciences and Primary Care, University Medical
Center Utrecht, Utrecht, The Netherlands
Abstract
To date, genome-wide association studies have focused almost exclusively on populations of European ancestry. These
studies continue with the advent of next-generation sequencing, designed to systematically catalog and test low-frequency
variation for a role in disease. A complementary approach would be to focus further efforts on cohorts of multiple
ethnicities. This leverages the idea that population genetic drift may have elevated some variants to higher allele frequency
in different populations, boosting statistical power to detect an association. Based on empirical allele frequency
distributions from eleven populations represented in HapMap Phase 3 and the 1000 Genomes Project, we simulate a range
of genetic models to quantify the power of association studies in multiple ethnicities relative to studies that exclusively
focus on samples of European ancestry. In each of these simulations, a first phase of GWAS in exclusively European samples
is followed by a second GWAS phase in any of the other populations (including a multiethnic design). We find that nontrivial
power gains can be achieved by conducting future whole-genome studies in worldwide populations, where, in particular,
African populations contribute the largest relative power gains for low-frequency alleles (,5%) of moderate effect that
suffer from low power in samples of European descent. Our results emphasize the importance of broadening genetic
studies to worldwide populations to ensure efficient discovery of genetic loci contributing to phenotypic trait variability,
especially for those traits for which large numbers of samples of European ancestry have already been collected and tested.
Citation: Pulit SL, Voight BF, de Bakker PIW (2010) Multiethnic Genetic Association Studies Improve Power for Locus Discovery. PLoS ONE 5(9): e12600.
doi:10.1371/journal.pone.0012600
Editor: Michael Nicholas Weedon, Peninsula Medical School, United Kingdom
Received June 21, 2010; Accepted August 10, 2010; Published September 8, 2010
Copyright: ? 2010 Pulit et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted
use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: The authors gratefully acknowledge support from the National Institutes of Health/National Institute of Mental Health (R01MH084676). The funders
had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: pdebakker@rics.bwh.harvard.edu
Introduction
Over the past four years, genome-wide association studies
(GWAS) have started to reveal the genetic underpinnings of
complex traits and common diseases, yet only a modest fraction of
the heritability can be attributed to the collection of associated
variants discovered to date [1]. Even though impressive sample
sizes have been assembled through collaborative efforts, the
statistical power to discover susceptibility loci is limited by a
number of factors [2].
One limitation of the first wave of GWAS is the almost exclusive
interrogation of common variation with limited coverage of alleles
in the lower end of the frequency spectrum. Many low-frequency
alleles have not been ascertained, and even for those that were
catalogued in the dbSNP database, SNPs on the genome-wide
microarrays tag low-frequency variants only poorly through
pairwise linkage disequilibrium [3]. The second limitation is that
most GWAS to date have primarily studied samples of European
descent [4], with low power to detect association for alleles of low
frequency compared to more common alleles.
A compelling illustration of this second limitation was provided
by the recent discovery of the association at the KCNQ1 locus with
type 2 diabetes risk (with an estimated odds ratio =1.2) in two
contemporary GWAS in East-Asian population samples [5,6]. The
associated SNP (rs2283228) has a minor allele frequency of ,40%
in East-Asian samples but a minor allele frequency of ,5% in
European samples [7]. At this allele frequency, the association had
little power to be discovered (at p,561028) in the series of GWAS
in European-derived samples performed ahead of the two East-
Asian studies [8,9,10,11,12], even though the SNP was well tagged
in European samples. After its initial discovery in the East-Asian
samples, the association was successfully replicated as a pre-
specified hypothesis at a more liberal significance threshold. This
finding raises the question to what extent power is affected by the
focus on samples of European ancestry, and how power to detect
association of genetic variation segregating at low frequency in
European populations could be heightened by broadening GWAS
efforts to a more diverse set of populations.
To this end, we evaluate here the relative benefit of performing
future genome studies in worldwide populations (‘‘second phase’’),
preceded by an initial GWAS in a European population (‘‘first
phase’’). Using the empirically observed allele frequencies in
population samples represented by HapMap Phase 3 and the 1000
Genomes Project, we quantify the impact of frequency differences
between populations on the power to find novel association of
modest effect (GRR #1.5), assuming that genome-wide associa-
tion results are combined in the two GWAS phases. We also
address the implications of such allele frequency differences for
PLoS ONE | www.plosone.org1September 2010 | Volume 5 | Issue 9 | e12600

Page 2

replicating bona fide associations (most discovered in European
samples) in different populations.
Overall, we show that there are substantial power gains to be
had by focusing on large multiethnic studies. Additionally, allele
frequency fluctuations between global populations and their
impact on power must be considered in replication studies.
Methods
We computed power to detect an association at genome-wide
significance (p,561028) using a theoretical model of log-additive
(multiplicative) effects, allowing us to perform all calculations
efficiently. For each SNP, we separately computed the non-
centrality x2parameter (NCP) in the first and second phase as a
function of risk allele frequency, case-control sample size, and
assumed effect size (genotype relative risk, GRR) (Text S1). In
phase 1, we used the empirically observed allele frequencies in
CEU (Utah residents with Northern and Western European
ancestry collected by the Centre d’Etude du Polymorphisme
Humain) to mimic the fact that most GWAS to date have tested
European samples. In phase 2, we used the allele frequencies in
any of the population panels represented in HapMap 3 or 1000
Genomes (see below). For multiethnic scenarios in phase 2, we
computed the NCPs per population. By summing the NCPs in all
populations in phases 1 and 2, we derived the asymptotic power
for a given SNP to reach a p-value of 561028(Text S1). For each
SNP, we averaged the power for both alleles simulated as the risk-
increasing allele (GRR .1). We repeated this procedure across all
1,440,616 SNPs present in HapMap 3 or 3,327,757 SNPs present
in 1000 Genomes, and report the average genome-wide power
across all SNPs, and the average power for SNP subsets stratified
by the minor allele frequency in CEU (where we refer to 1–5% as
low-frequency alleles). We varied the total sample size of each
scenario between 10,000, 20,000 and 80,000 samples (Table 1).
We varied the GRR between 1.1 and 1.5, following a
multiplicative risk model and assumed a fixed effect between
different populations in phases 1 and 2.
We note that our approach is not to be confused with a two-
stage replication design (as described, for example, in [13]) where
only a limited set of SNPs are taken from the phase 1 (based on
some p-value threshold) for additional testing in phase 2. Instead,
we leverage the idea that genome-wide association data sets are
combined in collaborative spirit, as is now routinely done for locus
discovery.
We obtain empirical allele frequencies for SNPs from two
sources: HapMap 3 and the 1000 Genomes Project. The HapMap
3 resource comprises genome-wide SNP data from 11 population
samples [14]: CEU (Utah residents with Northern and Western
European ancestry collected by the Centre d’Etude du Poly-
morphisme Humain), TSI (Toscans in Italy), CHB (Han Chinese
in Beijing, China), JPT (Japanese from Tokyo, Japan), CHD
(Chinese in Metropolitan Denver, Colorado), YRI (Yoruba in
Ibidan, Nigeria), MKK (Massai in Kinyawa, Kenya), LWK
(Luhya in Webuye, Kenya), ASW (African ancestry in southwest
USA), GIH (Gujarati Indians in Houston, Texas), and MXL
(Mexican ancestry in Los Angeles, California). In total, 1184
samples have been genotyped on the Illumina Human-1M and
Affymetrix 6.0 arrays, and SNP genotypes merged and processed
for quality control. Our analysis is restricted to the allele
frequencies of 1,440,616 SNPs that are QC-passing and
polymorphic in at least one of the eleven populations (referred
to as the consensus data). To remove bias due to different sample
sizes of these populations, we randomly selected 50 unrelated
founders (100 unique chromosomes, dictated by MXL, the
smallest sample) from each population in the consensus genotype
data to compute unbiased allele frequencies.
We also used frequency estimates from low-coverage sequenc-
ing data in four population panels (CEU, CHB, JPT, and YRI)
generated in Pilot 1 of the 1000 Genomes Project [15]. The data
we use here was generated in 60 founder individuals of CEU, 30 of
CHB, 30 of JPT, and 59 of YRI, resulting in three analysis panels
of similar sample size (CEU, CHB+JPT and YRI).
As a complementary analysis, we performed a power analysis
for a specific set of 182 unique SNPs (189 reported associations; 7
SNPs reported for multiple diseases) that have recently been
described as significantly associated to 26 complex diseases (taken
from http://www.genome.gov/GWAStudies/, accessed on 28
September 2009). For most of these associations, the effect sizes
range from 1.1 to 1.3. Based on the observed allele frequencies in
each of the HapMap 3 populations, we computed the sample size
required to replicate the association at 80% power (at a nominal
p,0.05) for the reported risk allele, taking the reported odds ratio
as the GRR in a multiplicative model. This analysis reflects the
testing of a specific hypothesis (instead of discovery across the
whole genome) based on bona fide associations reported by GWAS
performed to date.
Results
We compared three main scenarios where the key difference
was the total sample size (Table 1). First we explored the simple
scenario where phases 1 and 2 are performed in 5,000 samples
from the CEU population panel (Table S1). Power was, as
expected, consistently lower across the entire range of effect sizes
and allele frequencies compared to a sample size of 20,000 (Table
S2) and 80,000 (Table S3).
At a sample size of 20,000, the aggregate power across all 1.4
million SNPs considered was 82% at an effect size (GRR) of 1.5,
but fell to 8% at a GRR of 1.1 (Table S2). Across the different
effect sizes, power was largely driven by the power contribution of
common alleles (MAF .10%). At a modest GRR of 1.2, the
power of common alleles was 87% and rose to nearly 100% at a
Table 1. Overview of the three main GWAS scenarios.
Total
Sample SizePhase 1Phase 2
Cases/controlsHapMap panelCases/controlsHapMap panel
10,0002,500/2,500CEU 2,500/2,500CEU, TSI, CHB, CHD, JPT, YRI, MKK, LWK, ASW, MXL, or GIH
20,0005,000/5,000CEU5,000/5,000CEU, TSI, CHB, CHD, JPT, YRI, MKK, LWK, ASW, MXL, or GIH
80,00010,000/10,000 CEU30,000/30,000CEU, TSI, CHB, CHD, JPT, YRI, MKK, LWK, ASW, MXL, or GIH, or CEU+CHB+YRI
or CEU+CHB+YRI+MXL+GIH+ASW or CEU+CHB+YRI+MXL+GIH+ASW
doi:10.1371/journal.pone.0012600.t001
Power of Multiethnic GWAS
PLoS ONE | www.plosone.org2 September 2010 | Volume 5 | Issue 9 | e12600

Page 3

GRR of 1.5. For lower frequency variants (MAF #5%), power to
detect an association remained low. For a GRR of 1.1 or 1.2,
power for SNPs with 1–5% frequency was essentially zero, but
rose to 65% when GRR was 1.5. Increasing the effect size reduced
the minimum allele frequency at which an association could be
detected. For a variant of 3% frequency, there was virtually no
power to detect at an effect size of 1.2 but almost 60% power to
detect it at a GRR of 1.4.
For a sample size of 80,000 samples, power to detect an
association of a lower frequency allele (MAF #5%) remained low
at smaller effect sizes. Power is ,3% at GRR=1.1 (but ,89% for
common variants with the same effect size), and did not rise above
the 80% power threshold until a GRR of 1.3 (Table S3).
Next we compared the two European samples (CEU and TSI)
in phase 2 and found that the aggregate estimates for both
populations were very similar for all scenarios considered. A power
gain in TSI could be observed for lower-frequency variants at
smaller total sample sizes (10,000 and 20,000), including variants
that appear monomorphic in CEU (which by definition cannot
contribute power). This effect, however, was balanced by a power
loss in TSI for the largest sample size (80,000), due to the total
number of monomorphic SNPs in TSI (241,688 SNPs compared
to 199,821 monomorphic SNPs in CEU; Figure S1) and the
amplification of their inability to contribute power by the large
number of samples.
Performing phase 2 of the GWAS in a non-European
population panel demonstrated a marked power gain for alleles
with low frequency in CEU (MAF #5%), in a trend that remained
true with increasing sample size or greater effect sizes (Figure 1
and Figure S2 for alleles of 5–10% frequency in CEU). Any of
the African population panels (YRI, MKK, or LWK) and the
admixed African-American (ASW) panel provided the greatest
power gain as compared to a GWAS performed in CEU samples.
For a sample size of 20,000, the power to detect low-frequency
variants (MAF #5%) with a GRR of 1.2 was close to zero in CEU,
but about 40% in YRI. At a GRR of 1.4, power was about 50% in
CEU but 80% in the African population samples. These results
demonstrate that there is a notable gain in power for this class of
variants by moving into an African population sample for phase 2.
Power for alleles of low frequency in CEU also increased by
performing phase 2 in samples from one of the East-Asian
populations (CHB, CHD, or JPT) though the gain was not as
significant as that provided by moving to an African population
sample or a sample from the admixed ASW panel in phase 2. The
other admixed population panels (GIH, MXL) also improved
power, but to a lesser degree than one of the East-Asian panels.
For a sample size of 80,000, performing phase 2 of the GWAS
in a non-European population sample continued to yield a net
gain in power, but only for those alleles that have low frequency in
CEU and for smaller effect sizes (Figure 2). For larger effect sizes
(GRR .1.3), polymorphic SNPs in CEU (especially those at lower
frequency where power can still be improved) continued to gain in
power, while the power for many of the other populations reached
a plateau due to a comparatively larger slice of apparently
monomorphic SNPs (Figure S1), an effect due to ascertainment
bias of the SNP genotyping platforms used in HapMap 3.
Figure 1. Power to detect association for lower-frequency alleles (# #5%) in CEU based on HapMap 3 data. Power is given for various
individual population panels (CEU, TSI, YRI, MKK, LWK, and ASW), a panel with major continental representation (CEU+CHB+YRI) and a cosmopolitan
panel with major continental representation and admixed populations (GIH, MXL, and ASW) interrogated in phase 2, aggregated over those alleles
that have lower frequency (1–5%) in CEU.
doi:10.1371/journal.pone.0012600.g001
Power of Multiethnic GWAS
PLoS ONE | www.plosone.org3 September 2010 | Volume 5 | Issue 9 | e12600

Page 4

We also examined the utility of two different multiethnic samples
in phase 2. One multiethnic sample comprised the major
continental population panels (CEU+YRI+CHB) while the second
comprised the major continental panels in addition to the admixed
panels (CEU+YRI+CHB+GIH+MXL+ASW). Although the power
gain for alleles of lower frequency in CEU was noticeable in both
multiethnic panels, the best power was obtained for the ‘‘cosmo-
politan’’ panel with representation from the major continental
groups as well as the admixed populations (Figures 1 and 2).
During the course of this study, we had access to low-coverage
pilot data from the 1000 Genomes Project, which allowed us to
replicate these relationships across 3,327,757 detected variable
sites with estimated allele frequencies in the CEU, CHB, JPT and
YRI panels across the whole genome, essentially free of
ascertainment bias. Overall, the power estimates based on these
allele frequency distributions were unchanged (Tables S4, S5
and S6), compared to the HapMap 3-based results, also for a
lowered discovery threshold of p,161028to reflect increased
variation in African samples (data not shown).
Consistent with the HapMap 3-based results, power to detect
an association was driven primarily by common variation and
power to detect lower-frequency variants remained limited for
GWAS scenarios exclusively in European samples (Tables S4,
S5 and S6). For lower-frequency variants (MAF #5%), the 80%
power threshold was not reached until the sample size was
80,000 and the effect size was 1.3 (Figure 3), exactly as was
observed in the HapMap 3 dataset. Performing phase 2 in
African ancestry samples increased power to detect alleles of
lower frequency in CEU (Figure 4 and Figure S3 for alleles of
5–10% frequency in CEU). Power was also increased by
including samples of East-Asian ancestry in phase 2, but to a
lesser extent than samples of African ancestry. Following a
multiethnic approach by combining samples from the CEU,
CHB, JPT and YRI panels, we see again a power improvement
for lower frequency alleles (in CEU). At an effect size of 1.2 and
a total sample size of 80,000, utilizing the multiethnic approach
in phase 2 increased power to detect low-frequency variants in
CEU to 85%, in comparison to 60% power when only European
samples are tested in phase 2. These results indicate a potential
gain in power to detect alleles of lower frequency in European
samples by performing a second phase of GWAS in samples of
African descent (Figures 3 and 4).
To illustrate the impact of allele frequency differences between
populations, we focus on two specific scenarios: one GWAS in
which phase 1 and phase 2 are performed in 10,000 European
(CEU) samples (20,000 total samples), and a second GWAS in
which phase 1 tests 10,000 European (CEU) samples and phase 2
tests 10,000 African (YRI) samples (Figure 5). Of all .3 million
SNPs with allele frequency estimates in CEU and YRI from 1000
Genomes, we can identify four categories: (1) SNPs that reached
power $80% in both scenarios (65.6% of all .3 million SNPs), (2)
SNPs with power $80% in only the European GWAS (6.5%), (3)
SNPs with power $80% in only the GWAS with YRI samples in
phase 2 (9.3%), and (4) SNPs that do not reach 80% power in
either scenario (18.6%). Most of the power gain (3.3%) for the
CEU+YRI design (relative to CEU+CEU) is due to alleles that are
15–40% higher in frequency in YRI compared to CEU. For these
alleles that are 15–40% more common in YRI, the net gain in
Figure 2. Power as a function of allele frequency (# #5%) in CEU based on HapMap 3 data. For a sample size of 80,000 and a modest effect
size (GRR of 1.1 and 1.2), power is given for CEU, CHB, YRI, and two multiethnic panels (‘‘major continental’’, CEU+CHB+YRI, and ‘‘cosmopolitan’’,
CEU+CHB+YRI+ASW+GIH+MXL) in phase 2. Including non-European samples in phase 2 improves power to detect an association for alleles that have
lower frequency in CEU.
doi:10.1371/journal.pone.0012600.g002
Power of Multiethnic GWAS
PLoS ONE | www.plosone.org4 September 2010 | Volume 5 | Issue 9 | e12600

Page 5

Figure 3. Power to detect association for lower-frequency alleles (# #5%) in CEU using 1000 Genomes Project data. Power is given for
three individual panels (CEU, CHB+JPT, YRI), and a multiethnic panel (CEU+CHB+JPT+YRI) in phase 2, aggregated over those alleles that have lower
frequency (1–5%) in CEU.
doi:10.1371/journal.pone.0012600.g003
Figure 4. Power as a function of allele frequency (# #5%) in CEU using 1000 Genomes Project data. For a sample size of 80,000 and
modest effect size (GRR of 1.1 and 1.2), power is given for three individual panels (CEU, CHB+JPT, YRI) and a multiethnic panel in phase 2. Including
non-European samples in phase 2 improves power to detect an association for alleles that have lower frequency in CEU.
doi:10.1371/journal.pone.0012600.g004
Power of Multiethnic GWAS
PLoS ONE | www.plosone.org5 September 2010 | Volume 5 | Issue 9 | e12600