Sizing up human height variation

Peter M Visscher
Genome-wide association studies have identified many variants affecting susceptibility to disease. Now, three
studies use this approach to study adult height variation in a combined sample size of 63,000 individuals and
report a total of 54 validated variants influencing this trait.
Human height is a classical quantitative trait.
Genetic studies, pioneered by Francis Galton
and Ronald Fisher more than 90 years ago1,2,
show a clear pattern of family resemblance,
consistent with a polygenic additive model
of inheritance. But how many is poly, and
what are the effect sizes of individual variants? In this issue, three consortia of research
groups report a total of 54 loci affecting
height variation in the population, identified using genome-wide association studies
of hundreds of thousands of genetic markers genotyped on a total of 63,000 people
measured for height35.

Genetic linkage studies look for a correlation between phenotypic and genotypic
similarity within families. Many such studies
have been reported for height, implicating
many (some would say all) genomic regions,
but the resolution of these studies is low, and
they have not resulted in the identification
of genetic variants that explain the linkage
signals. Past studies of candidate genes have
also not succeeded in explaining familial
resemblance for height. Indeed, similar
observations can be made for the studies of
many other complex traits, including those
of common diseases.

Quantitative genetics of height

Height is a highly heritable trait that is easy
to measure. It is a model trait for quantitative genetics and developmental biology, it
is associated with disease, including cancer6,
and it is a predictor of social outcomes in
life. In populations of European descent,
the average height is 178 cm for males and
165 cm for females, with a s.d. of 7 cm.
Its heritability is 0.8, which means that
within a population, about 80% of the variation in height among individuals is due to
genetic factors7. If we knew all the variants
that are responsible for genetic variation
and summed their effects within each person, then the difference between the top 5%
and bottom 5% of the population would be
about 26 cm, roughly the size of an adult
head. Mutations in certain genes can cause
extreme short or tall stature, but such mutations are rare and do not contribute to normal variation in the population.

Genome-wide association era

Enter the era of genome-wide association
studies (GWAS). These studies take a systematic unbiased approach by interrogating
the entire genome for associations between
common gene variants (single nucleotide
polymorphisms or SNPs) and a phenotype.
GWAS have been facilitated by the HapMap
project8, which quantified the total number
and genome locations of SNP markers that
need to be genotyped in order to detect an
association between common genetic variants and a trait in a hypothesis-free genome
scan. Further, development of commercial
SNP chips that allow rapid genotyping of
hundreds of thousands of common SNPs
have made this approach possible. The basic
design of a GWAS is to associate the trait of
interestfor example, disease susceptibility
or a quantitative traitwith SNPs on the
chip, usually by carrying out a statistical test
for each SNP in turn and then following up
on the best hits by genotyping them in one
or more independent samples for statistical validation. GWAS have been successful
in finding previously unknown loci associated with a wide range of diseases9 and other

Peter M. Visscher is at the Queensland Institute

of Medical Research, Brisbane, Queensland
4029, Australia.
e-mail: [email protected]

NATURE GENETICS | VOLUME 40 | NUMBER 5 | MAY 2008

100
90
80

Power of detection (%)

2008 Nature Publishing Group http://www.nature.com/naturegenetics

NEWS AND VIEWS

n = 5,000
n = 10,000
n = 20,000
n = 40,000

70
60
50
40
30
20
10
0

0.1

0.2

0.3

0.4

0.5

Effect size (% variance explained)

Figure 1 Statistical power of detection in GWAS

for variants that explain 0.10.5% of the variation
at a type I error rate of 5 107 (calculated using
the Genetic Power Calculator15). Shown is the
power to detect a variant with a given effect size,
assuming this type I error rate, which is typical for
a GWAS with a sample size of n = 5,00040,000.

phenotypes. Two previous GWAS, from two

of the three groups that report findings in
the current issue, each reported a single
newly identified validated variant affecting
height10,11, with each locus explaining a very
small proportion of the phenotypic variance
(0.3% to 0.5%).
The new studies in this issue35 followed
a multistage design in which the first stage
was used to select the most promising SNPs
and the later stages were used for validation.
The validation stage is important because,
when over 500,000 variants are tested, as
was done in the current studies, many will
be statistically significant by chance (about
25,000 if a standard type I error rate of 0.05
is used). The studies employed very large
samples sizes, from 14,00034,000 in the
test stage to 6,00020,000 in the validation stages. The total number of SNP chips
used across these studies was over 63,000,

489

 2008 Nature Publishing Group http://www.nature.com/naturegenetics

NEWS AND VIEWS

representing a total investment of roughly
$30 million. However, much, if not most,
of the genotyping was done by other studies investigating disease, and the researchers cleverly piggy-backed on these studies,
taking advantage of the fact that many such
studies have included measures on height.
Across the three studies, 95 SNPs were
taken forward for validation and 54 survived
stringent significance testing. These SNPs
were robustly associated with height variation in the general population. Reassuringly,
SNPs in the previously associated genes
HMGA2 and GDF5-UQCC were again
identified. SNPs in three genes (ZBTB38,
HHIP, HMGA2) were found to be associated with height in all three studies, and
a total of seven genes were implicated in
two of the three studies. The three studies
found an impressive total of 40 previously
unknown variants. The average effect size
per increasing allele was 0.4 cm, or 0.8
cm between the two homozygous classes.
The reason why the overlap in loci is modest is presumably because of the stringent
significance testing used by the researchers
and because the effect sizes are small. The
probability (power) of detecting a variant
of small effect when using a genome-wide
significance threshold is far from perfect,
even when using fairly large sample sizes
(Fig. 1). Lowering the significance threshold would increase the power but would also
yield more false positives. It is a challenge of
GWAS designs to get this balance right.
In the three height studies35, the sample
sizes of the test stages were approximately
14,000, 16,000 and 34,000. As shown in
Figure 1, a GWAS with a sample size of
10,000, with an assumed type I error rate
of 5 107, has 29% power to detect a variant that explains 0.2% of the variance.
Therefore, the chance that two independent
studies of this size both detect this variant is
only 0.292 = 0.08. For sample sizes of 20,000,
the probability of detecting the variant in
two studies is 0.81. Therefore, one explanation for why there was only modest overlap
in the detected loci between the studies is
statistical power at the chosen stringent type
I error rates.
What have we learned about the nature of
quantitative trait variation for height from
these studies? At a first glance it looks quite
simple: variation is explained by many variants of small effects, with no evidence for
interactions between alleles, either within
loci (dominance) or between loci (epistasis), and there are no strong differences in
effects between males and females. These
observations are consistent with patterns of

490

familial resemblance for height. However,

given the design and analysis used, there was
little statistical power to find evidence for
departures from this simple model. Not surprisingly, given the small effect sizes found,
there was no significant overlap between the
location of the associated variants and previously reported loci from linkage studies.
It remains a challenge to reconcile the findings of GWAS and linkage studies, because
the former suggest individual variants with
small effects, whereas the latter suggest
genomic regions with large effects within
pedigrees.
A return to candidate genes?
The GWAS approach, in general, is set to
identify associations with individual variants
that are unlikely to be causal. Nevertheless,
by looking at the genes in the vicinity of
the associated SNPs, one can look for biological pathways that are overrepresented.
Taking this approach, the height studies
report that their validated SNPs are in or
near genes involved in pathways related to
mesoderm development, skeletal development, mitosis, cancer, Hedgehog signaling
and chromatin remodeling. Variants in one
of the genes (ZBTB38) that was detected in
all three studies were shown to be associated with gene expression of that same gene
in blood and adipose tissue5, suggesting a
possible biological mechanism.
The authors also found a significant
number of the validated SNPs in genes that
could be considered candidate genes on
the basis of known mutant phenotypes in
humans or mice. But wasnt one of the rationales of the GWAS approach that previous
candidate gene studies hadnt worked? One
explanation for this (perceived) failure of
earlier candidate gene studies is that they
were severely underpowered, both in terms
of experimental sample size and SNP coverage. It is only by performing large, unbiased
GWAS that the relative importance of candidate genes can be quantified.
Looking to the future
The main conclusion emerging from the
current studies is that GWAS are able to
robustly identify common variants that are
associated with height but that the effect
sizes of individual variants are small, so that
very large sample sizes are needed to detect
associations reliably. Single laboratories are
unlikely to have sufficient sample sizes to do
powerful studies on their own, and the trend
in human complex trait mapping has been
to create consortia of research groups and
even consortia of consortia.

It remains unclear at this stage how much

genetic variation can be explained through
the GWAS approach. However, if the samples in these three studies were combined
together with other datasets that have been
collected on height and genome-wide SNP
data, then this question could be answered
empirically. Genome-wide studies on, say,
100,000 individuals, unthinkable only a few
years ago, will be soon be a reality. From
the limited amount of overlap in the genes
reported across the three studies, and the fact
that all three studies were very stringent in
their hypothesis testing, it seems that much
more variation is likely to be explained by
common variants of small effect.
There is still a long way to go from associated SNPs to causal variants. Resequencing
will uncover all variants in a given region,
not just the common ones, but it will
unearth so many that picking out the
responsible variant(s) will be a real challenge. Statistical association alone will not
provide the answer, and other sources of
information, such as predicted effects of
variants on proteins and evolutionary conservation12,13, may help.
Looking further to the future, if most
genetic variation for height can be explained
by association, even without knowing causal
variants, how can this knowledge be used?
Apart from gaining insights into the genetic
architecture of continuous traits and biological pathways involved in human height,
associations between variants and height can
in principle be used to predict height from
genetic data alone14, that is, without having
observed the phenotype. Such predictions
are useful in artificial selection programs
in agriculture and may also prove useful in
forensics and human medicine.
1. Fisher, R.A. Trans. Roy. Soc. Edin. 52, 399433
(1918).
2. Galton, F. Nature 33, 295298 (1886).
3. Weedon, M.N. et al. Nat. Genet. 40, 575583 (2008).
4. Lettre, G. et al. Nat. Genet. 40, 584591 (2008).
5. Gudbjartsson, D.F. et al. Nat. Genet. 40, 609615
(2008).
6. Gunnell, D. et al. Epidemiol. Rev. 23, 313342
(2001).
7. Visscher, P.M., Hill, W.G. & Wray, N.R. Nat. Rev. Genet.
9, 255266 (2008).
8. The International HapMap Consortium. Nature 437,
12991320 (2005).
9. The Wellcome Trust Case Control Consortium. Nature
447, 661678 (2007).
10. Sanna, S. et al. Nat. Genet. 40, 198203 (2008).
11. Weedon, M.N. et al. Nat. Genet. 39, 12451250
(2007).
12. Easton, D.F. et al. Am. J. Hum. Genet. 81, 873883
(2007).
13. Rebbeck, T.R., Spitz, M. & Wu, X. Nat. Rev. Genet. 5,
589597 (2004).
14. Wray, N.R., Goddard, M.E. & Visscher, P.M. Genome
Res. 17, 15201528 (2007).
15. Purcell, S., Cherny, S.S. & Sham, P.C. Bioinformatics
19, 149150 (2003).

VOLUME 40 | NUMBER 5 | MAY 2008 | NATURE GENETICS