Human Facial Shape and Size Heritability and
Joanne B. Cole,* Mange Manyama,
Jacinda R. Larson,
Denise K. Liberton,
Tracey M. Ferrara,*
Sheri L. Riccardi,* Mao Li,** Washington Mio,** Ophir D. Klein,
Stephanie A. Santorico,*
and Richard A. Spritz*
*Human Medical Genetics and Genomics Program, and †††Department of Pediatrics, University of Colorado School of Medicine,
Aurora, Colorado 80045, †Department of Anatomy, Catholic University of Health and Allied Sciences, Mwanza, TZ-18, Tanzania,
‡Department of Anatomy and Cell Biology, and §McCaig Institute for Bone and Joint Health, Alberta Children’s Hospital Research
Institute, Cumming School of Medicine, University of Calgary, T2N 1N4, Canada, **Department of Mathematics, Florida State
University, Tallahassee, Florida 32304, ††Department of Orofacial Sciences, ‡‡Department of Pediatrics, and §§Program in
Craniofacial Biology, University of California, San Francisco, California 94143, and ***Department of Mathematical and Statistical
Science, University of Colorado Denver, Colorado 80202
ABSTRACT The human face is an array of variable physical features that together make each of us unique and distinguishable. Striking
familial facial similarities underscore a genetic component, but little is known of the genes that underlie facial shape differences.
Numerous studies have estimated facial shape heritability using various methods. Here, we used advanced three-dimensional imaging
technology and quantitative human genetics analysis to estimate narrow-sense heritability, heritability explained by common genetic
variation, and pairwise genetic correlations of 38 measures of facial shape and size in normal African Bantu children from Tanzania.
Speciﬁcally, we ﬁt a linear mixed model of genetic relatedness between close and distant relatives to jointly estimate variance
components that correspond to heritability explained by genome-wide common genetic variation and variance explained by uncaptured
genetic variation, the sum representing total narrow-sense heritability. Our signiﬁcant estimates for narrow-sense heritability of speciﬁc
facial traits range from 28 to 67%, with horizontal measures being slightly more heritable than vertical or depth measures. Furthermore,
for over half of facial traits, .90% of narrow-sense heritability can be explained by common genetic variation. We also ﬁnd high absolute
genetic correlation between most traits, indicating large overlap in underlying genetic loci. Not surprisingly, traits measured in the same
physical orientation (i.e., both horizontal or both vertical) have high positive genetic correlations, whereas traits in opposite orientations
have high negative correlations. The complex genetic architecture of facial shape informs our understanding of the intricate relationships
among different facial features as well as overall facial development.
KEYWORDS heritability; facial shape; facial size; morphometrics; complex traits
HUMAN appearance is comprised of a remarkably variable
set of physical traits. Of all externally visible character-
istics, facial appearance is both the most morphologically
variable and the most distinctive and recognizable. Facial
appearance involves a major genetic component, with each
of the many structural features that deﬁne facial shape and
appearance themselves likely determined by a multiplicity of
genes, with environmental variables such as nutrition and
environmental toxins, exerting increasing inﬂuence over time
(Fitzgerald et al. 2010). Nevertheless, the striking similarity
of facial appearance within families, often across many gen-
erations, suggests that certain key genes exert particularly
large effects on facial shape and appearance.
Facial shape is measured in various ways, including speciﬁc
linear measurements between deﬁned morphological points
as well as complex quantitative measurements of the entire
face. Previous estimates of the heritability of facial shape
distances and angles were principally derived by direct mea-
surements between common facial morphometric landmarks
on human faces, cephalograms, and skulls. These estimates
Copyright © 2017 by the Genetics Society of America
Manuscript received June 27, 2016; accepted for publication December 8, 2016;
published Early Online December 14, 2016.
Supplemental material is available online at www.genetics.org/lookup/suppl/doi:10.
Corresponding author: University of Colorado School of Medicine, Anschutz Medical
Campus, Rm. 3100, Mail-stop 8300, 12800 East 19th Ave., Aurora, CO 80045.
Genetics, Vol. 205, 967–978 February 2017 967
vary widely; in general, facial height dimensions tend to be
more heritable than width (Manfredi et al. 1997; Carson
2006; Amini and Borzabadi-Farahani 2009; AlKhudhairi
and AlKoﬁde 2010), in contrast with the rest of the skull,
for which heritability of width tends to be greater than for
height (Martínez-Abadías et al. 2009a,b, 2012).
The genetic architecture of facial shape variation has been
studied more extensively in mice than in humans. In the mouse,
measures of craniofacial morphology are highly heritable (Leamy
1977; Klingenberg and Leamy 2001; Klingenberg et al. 2001;
Percival et al. 2016). Further, the mouse skull is highly inte-
grated in terms of phenotypic and genetic correlations (Leamy
1977; Cheverud 1982). Genetic and environmental correlations
also tend to be similar, likely due to their basis in similar de-
velopmental connections among traits (Cheverud 1982).
Morphological assessment of facial variation has typically
required manual landmarking, an approach that is slow, labor
intensive, and error prone; complicating its application to large-
scale studies as well as comparisonsacrossmultiplestudies.Asan
advance toward standardized, replicable phenotyping of human
facial traits, we combined advanced three-dimensional (3D)
imaging technology with a novel automated landmarking
method (Li et al. 2016) to derive precise, detailed, and informa-
tive facial phenotypes from 29 standard facial morphometric
landmarks (Supplemental Material, Table S1) (Bookstein 1997).
Our study, based on Bantu children from Tanzania, avoids
facial shape variation that occurs later in life due to injury,
weight gain, and disease. Moreover,these African children are
very lean, with minimal variation related to facial adiposity,
and furthermore have signiﬁcant occult relatedness, provid-
ing the opportunity to formally analyze the heritability of
facial shape phenotypes in this population.
To assess heritability, we analyzed genotypes of .15 million
common SNPs with minor allele frequencies .1% using
Genome-wide Complex Trait Analysis (GCTA) (Yang et al.
2010, 2011). We estimated narrow-sense heritability (h
heritability explained by common genetic variation (h
pairwise genetic correlations of 38 facial phenotypes, incorpo-
rating close family structures into a joint linear mixed model
with two variance components, one representing the genetic
relatedness between close relatives and the other representing
the genetic relatedness between all individuals in the study
(Zaitlen et al. 2013). The phenotypic variance was then parti-
tioned into variance explained by common genetic variation,
variance explained by close genetic relationships but not com-
mon genetic variation, and the remaining residual variance
explained by the environment.
We found that facial shape and size phenotypes are highly
heritable, and additionally are highly genetically correlated, and
that a large fraction of the genetic component of facial differences
can be explained by common variation genome-wide.Our ﬁndings
help elucidate the complex genetic relationships and pathways
underlying facial shape, augment basic biological understanding
of facial development, enable better modeling of facial shape
based on genetic correlations, and may assist in delineation and
diagnosis of facial dysmorphism syndromes.
Materials and Methods
Samples and data were collected from 3631 Bantu African
children aged 3–21 over a 3-year period in the Mwanza region
of Tanzania. Subjects with a known birth defect or a relative
with known orofacial cleft or facial birth defect were excluded.
Additional data collected were age, sex, height, weight, head
circumference, school, and detailed parental and grandparen-
tal ethnicity, and tribe information. Subjects with non-Bantu
tribal ancestry in one or more grandparents were excluded
from the study. Written informed consent was obtained from
all study subjects or their parents as appropriate.
3D facial imaging and automated landmarking
3D images were obtained using the Creaform MegaCapturor 3D
photogrammetry imaging system. Each subject was imaged
twice at six standard positions. Meshes were reconstructed at
the highest possible resolution and assembled using Inspeck
software to form a complete 3D mesh of the face (Figure 1).
A set of 29 morphometric facial landmarks (Figure 1 and
Table S1) was applied to each individual facial mesh using a
novel automated landmarking method (Li et al. 2016). Brieﬂy,
this method morphs a 29-landmark template (created from a
training set of manually landmarked faces) to each individual
face through the guidance of anchor points deﬁned by the local
curvature features of the face. Presenting the same image to
the automated landmarking system twice generates identical
landmark positions. Errors, when they occur, tend to be fairly
large, readily detectible, and easily removed.
Derivation of phenotypic variables
Landmarks were subjected to Procrustes superimposition for
geometric morphometric analysis (Bookstein 1997; Dryden
and Mardia 1998; Mitteroecker and Gunz 2009). Superﬁcial
artifacts (smiling, squinting, mouth open, lateral nasal defor-
mity, etc.) were identiﬁed by manual quality inspection of
each facial mesh, and were corrected using a multiple linear
model in which all factors and their interactions were con-
sidered. For each correction, we determined the signiﬁcance
of the artifact using a Procrustes distance permutation test on
age- and size-adjusted data. We performed the corrections
jointly using a linear model in Rto avoid overcorrection
caused by overlap among the artifacts. Additionally, we de-
termined whether the corrections affected biological signals
such as the estimates of ontogenetic or static allometry and
the heritabilities of the traits. For all artifacts, we performed
canonical variate analysis both before and after to visualize
the effect of the correction (Figure S1 and Figure S2).
An additional artifact in 3D photogrammetry is “skew,”
deﬁned as coordinated asymmetric displacement of land-
marks due either to variation in assembly of the facial views
to produce the assembled mesh or from parallax. Most land-
marks are affected by skew when it occurs, but individuals
are not equally affected. Therefore, we regressed the land-
mark data on the principal component (PC) scores for PCs
968 J. B. Cole et al.
corresponding to skew variation. Skew corrections have min-
imal effects on most subjects but graded effects on those to-
ward the ends of the skew PC. Figure S2 shows morphs that
correspond to the extreme skew values in the sample. Skew
corrections were applied to the linear distances as this im-
proved heritability, but not to the multivariate measures as this
did not affect heritability. We validated artifact and skew cor-
rections by determining their inﬂuences on the estimates of
allometric shape variation and on heritabilities.
We obtained 25 linear distance phenotypes representing
heights, widths, and depths of different facial structures from
the fully corrected landmark coordinates after restoring size.
Size was calculated as centroid size, as is standard in geo-
metric morphometrics (Mitteroecker and Gunz 2009). Table
S2 lists means and standard deviations for all linear distances
and centroid size.
Multivariate measures were calculated from artifact but
not skew-corrected data. We removed variation related to age
and size in symmetrized landmark data using multiple mul-
tivariate regression and centering the residuals on the sample
mean (Klingenberg and Zimmermann 1992; Klingenberg
1998). Symmetrizing the data removes facial asymmetry var-
iation. While there is biological variation in asymmetry, this
did not signiﬁcantly affect the ﬁrst 10 PCs, aside from the
skew artifacts discussed above. The age-shape relationship
is nonlinear, particularly when an analysis includes both
young children and adults. Within our sample, however,
polynomial ﬁts for age or centroid size did not signiﬁcantly
improve the ﬁt(Figure S3). As our sample mostly excludes
the major shape changes occurring early in facial growth and
the slowing of changes that occur in late ontogeny, linear
regression sufﬁciently captures the shape variation associ-
ated with age and centroid size. Shape variation related to
size, or static allometry, was estimated using the regression
scores corresponding to size independent of age. We
obtained the scores for the ﬁrst10PCsbasedontheage
and size-standardized landmark coordinate data (Figure
S4). All references to PCs in the results refer to the PC
scores derived from the phenotypic data.
We used Klingenberg’s permutation test for Escofﬁer’sRV
coefﬁcient to identify the set of spatially contiguous land-
marks that maximized the ratio of covariation among them-
selves to covariation with landmarks outside of that set
(Klingenberg 2009). This method does not consider overlap-
ping determinants of covariation structure (Hallgrímsson
et al. 2009), but it reveals sets of strongly covarying, spatially
adjacent landmarks. The resulting set, deﬁned by the nasal
region and upper lip (Figure S5), was subjected to separate
Procrustes alignment, and a principal component analysis
(PCA). The ﬁrst PC (40% of variance) served as a measure
of variation within this “module.”
The ﬁnal phenotype values then adjusted for biological
covariates as follows: multivariate measures including all PCs
and allometry were adjusted for sex; linear distances were
adjusted for age, sex, and centroid size (after age-sex adjust-
ment); and centroid size was adjusted for age and sex. In this
way, all multivariate measures and linear distances were
corrected for age, sex, and size prior to downstream analysis.
Phenotypic correlations used throughout are based on phe-
notype residual correlations on a subset of unrelated individuals.
All morphometric analyses were done in MorphoJ (Klingenberg
2011) or in Rusing the Geomorph (Adams and Otárola-Castillo
2013; Adams et al. 2014) and Morpho (Schlager 2016) packages.
Genome-wide genotyping and quality control
Genome-wide genotyping, quality control, and imputation of
3480 study subjects used in these analyses were described
previously (Cole et al. 2016). The ﬁnal postquality-control
Figure 1 3D facial scan with annotated
landmarks. Landmarks annotated are
deﬁned in Table S1.
Heritability of Human Facial Shape 969
dataset included 3480 individuals with complete phenotyp-
ing information and imputed genotypes at .15 million
markers with INFO scores .0.30. A sensitivity analysis com-
paring heritability estimates from the full imputed data set
and from a data set of only genotyped markers demonstrated
no bias in using imputed genotypes (Figure S6).
were estimated using GCTA software (Yang
et al. 2010, 2011). We ﬁt a joint linear mixed model with two
variance components for each phenotype as described (Zaitlen
et al. 2013). Brieﬂy, one variance component represented close
relatives only, in which any pairwise genetic correlation in the
full genetic relatedness matrix as calculated by GCTA ,0.05
was set to 0. Speciﬁcally, our variance component highlighting
close relationships contained 4425 pairwise relationships $0.05
from a total of 2937 individuals in each triangle half of the
matrix. While this variance component represents a small pro-
portion of all pairwise relationships in our sample (0.07%), it
includes 84% of our total sample. Therefore the “missing heri-
tability”explained by this additional variance component of
close relatives is based on a large number of independent nu-
clear families, and thus is not biased by a small number of large
families (Figure S7). Heritability estimates obtained using dif-
ferent relatedness thresholds demonstrated that 0.05 was both
unbiased and conservative when compared to 0.00 and 0.025.
The other variance component represented the full genetic
relatedness matrix for all individuals. The joint model uses
4,307,014 more pairwise comparisons from 3480 individuals
to estimate h
than the traditional GCTA unrelateds only model
(n= 1869), using their recommended relatedness threshold
cutoff value of 0.025 (Table S3)(Yanget al. 2010). GCTA uses
restricted maximum likelihood to estimate the variance of each
component, the sum of which represents total genetic variance,
used for calculating h
. By default, GCTA estimates that escape
from the parameter space were set to 1.0 310
variance. To adjust our signiﬁcance threshold for multiple testing
of correlated phenotypes, we performed PCA of the 38 phenotype
residuals in unrelated individualstodeterminethenumber
of effectively independent phenotypes. The ﬁrst 11 eigen-
vectors had eigenvalues .1, making them each representative
of at least one phenotype. We divided the traditional P,0.05
threshold by 11, making our signiﬁcance threshold P,0.0045.
To calculate heritabilities and genetic variances for each
landmark, we obtained the genetic and phenotypic variance-
covariance matrices for the symmetrized landmarks. This
leaves out one dimension for midline landmarks and treats
the landmark coordinates for both sides as a single variable.
Heritabilities are calculated as the ratio of genetic to pheno-
typic variance for each landmark coordinate. We then used
a heatmap method to visualize these variance components
across the face. Here, each component is represented as a
vector that has a length proportional to the variance compo-
nent, an origin at each landmark, and a direction parallel to a
vector that connects each landmark to the landmark centroid.
These vectors are then used to produce a face morph using the
thin-plate-spline method that is then superimposed on the
unmorphed mean face to generate a heatmap.
Genetic correlations between all phenotypes were also esti-
mated using GCTA software (Yang et al. 2011). We elimi-
nated PC8 and LS_STO, for which the joint heritability LRT
models were not signiﬁcant (Table 2), from the genetic cor-
relation matrix. Due to nonconvergence of such a large pa-
rameter space, 148/630 bivariate analyses failed the joint
analysis and instead, for the sake of having a complete ge-
netic correlation matrix, were ﬁt by the standard GCTA ap-
proach, using a single genetic relatedness matrix of only
unrelated individuals (n=1869). Of those 148 bivariate
models, 36 failed the single component model of unrelateds
due to constraining genetic or environmental components.
To construct a complete genetic correlation matrix, we esti-
mated these 36 values based on both univariate and bivariate
models of the traits affected. If the bivariate model’s genetic
component was constrained or the univariate model’s herita-
bility estimate was less than the SE of that estimate, we set
the genetic covariance to 1.0 310
ance. If the bivariate model’s environmental component was
constrained or the univariate model’s heritability was essentially
equal to one, we set the genetic covariance as equal to the total
phenotypic covariance between the two traits. Therefore, the
genetic correlation matrix represents only the shared genetic
correlation which can be explained by .15 million common
variants. Table S4 includes all h
genetic correlation estimates,
SE, and models in which those values were derived.
Phenotype data were deposited in FaceBase (https://www.
facebase.org/; accession number: FB00000667.01). Genotype
data were deposited in the Database of Genotypes and Pheno-
types (dbGaP) (http://www.ncbi.nlm.nih.gov/gap; accession
number: phs000622.v1.p1). This study was carried out with
overall approval and oversight of the Colorado Multiple Insti-
tutional Review Board (protocol #09-0731), was additionally
approved by the institutional review boards of the University
of Calgary, Florida State University, the University of California
San Francisco, and the Catholic University of Health and Allied
Sciences (Mwanza, Tanzania), and was carried out with the ap-
proval of the National Institute for Medical Research (Tanzania).
Study population and phenotypes
The study population consisted of 3480 normal African Bantu
children and adolescents ages 3–21 from the Mwanza region of
Tanzania. Over 70% of subjects were aged 7–12, and 44.4%
were male and 55.6% female (Figure 2). As described previ-
ously (Cole et al. 2016), PCA of population substructure dem-
onstrated minimal genetic clustering and an analysis of ﬁxation
index demonstrated no apparent subgroups by school or tribe.
970 J. B. Cole et al.
For each subject, we captured 3D facial scans, applied
29 standard facial morphometric landmarks (Figure 1 and
Table S1) (Bookstein 1997), derived 38 facial shape pheno-
types based on the landmarks, carried out genome-wide SNP
genotyping, and used these data to estimate h
each phenotype. These facial phenotypes represent several
different classes, including 3D summary variables in the form
of PCs derived from PCA of the whole face and PCA of the
most highly correlated landmarks positioned around the
midface, interlandmark linear distances, and global mea-
sures of overall facial size and the relationship between
sizeandshape(Table1).Inaddition, for each subject we
obtained height and weight, from which we calculated
body mass index (BMI). Analysis of these data showed that
the mean BMI of the study population was 1 SD below the
S8). Furthermore, the correlations between all age- and
sex-adjusted facial traits with age- and sex-adjusted BMI
in unrelated individuals are small (r
=20.18 to 0.14),
suggesting that BMI is not a confounding factor in our
Heritability of facial phenotypes
signiﬁcantly heritable (P,0.0045), with h
most heritable facial traits include PC7, representing nasal root
shape and mouth width (h
= 66.9%, SE = 7.2%); total facial
width (T_R_T_L) (h
= 66.2%, SE = 7.5%); the allometric vari-
= 64.3%, SE = 7.2%); centroid size (h
SE = 7.6%); and nasion to midendocanthion distance
= 63.9%, SE = 7.5%). Furthermore, by com-
bining genetic variance across all 10 orthogonal PCs, which
explain .87% of total shape variation captured by sparse
landmarking (Figure S4), we obtained a single global esti-
mate of total facial shape of h
= 50.1%. Previous studies
have suggested that vertical measures have greater heritabil-
ity than horizontal measures (Manfredi et al. 1997; Carson
2006; Amini and Borzabadi-Farahani 2009; AlKhudhairi and
AlKoﬁde 2010). However, we observed a trend toward horizon-
talfacialmeasureshavinggreaterheritability than vertical mea-
sures. Figure 4 depicts h
estimates for the 25 linear
distances, clustered by physical orientation and phenotypic cor-
relation. The three facial depth measurements, lower facial
depth (GN_T), midfacial depth (SN_T), and upper facial depth
(N_T), share the tragion landmark and have very similar h
= 48.6%, SE = 7.6%; h
= 48.7%, SE = 7.6%;
= 51.2%, SE = 7.6%; respectively), but exhibit very
different phenotypic correlations (GN_T:SN_T = 0.31,
GN_T:N_T = 20.04, and SN_T:N_T = 0.67). Similarly,
the three horizontal eye measurements, inner canthal dis-
tance (EN_R_EN_L), outer canthal distance (EX_R_EX_L),
and average palpebral ﬁssure length (EN_EX), likewise share
some landmarks in common and have fairly similar h
= 41.1%, SE = 7.6%; h
= 52.2%, SE = 7.6%;
= 56.6%, SE = 7.8%; respectively), but exhibit very
different phenotypic correlations (EN_R_EN_L:EX_R_EX_L =
0.49, EN_R_EN_L:EN_EX = 20.03, and EX_R_EX_L:EN_EX =
0.82). In contrast, three midfacial horizontal measures, mouth
Figure 2 Study age distribution by sex.
Heritability of Human Facial Shape 971
width (CH_R_CH_L), philtrum width (CPH_R_CPH_L), and
subnasal width (SBAL_R_SBAL_L), share no overlapping
landmarks, have similar h
= 33.7%, SE = 7.7%; and h
= 37.2%, SE =
8.0%; respectively), and also exhibit fairly high phenotypic
correlations (CH_R_CH_L:CPH_R_CPH_L = 0.55, CH_R_CH_L:
SBAL_R_SBAL_L = 0.62, and CPH_R_CPH_L:SBAL_R_SBAL_L =
0.48). It appears that facial traits of similar orientation
that either share overlapping morphological points or
have high phenotypic correlations are inﬂuenced by ad-
ditive genetic effects and environmental effects to similar
Figure 5 shows the anatomic distribution of phenotypic
and genetic variance as well as heritability by landmark. The
pattern is consistent with Figure 3 and Figure 4 in that the
landmarks deﬁning facial width and the orbital region tend
to have higher genetic variances and heritabilities.
Genetic basis of observed heritability
For 22 of the 38 facial phenotypes analyzed, .90% of the
narrow-sense heritability (h
) can be explained by the effects
of common genetic variation (h
). However, for a number
of other traits, common variation (h
) accounts for ,50% of
; indicating signiﬁcant additional genetic contributions be-
yond common variants that can be imputed for Africans from
the Illumina HumanOmni 2.5-8 array, which captures 54% of
common African variation (minor allele frequency .1%;
exome-8.pdf). However, we caution that these conclusions
Table 1 The 38 facial phenotypes derived from landmarks on 3D facial scans
Phenotype abbreviation Physical description
PC1 upper facial height, midfacial width
PC2 overall facial height, lower facial height
PC3 upper and middle facial width
PC4 width of the nose, mandible height
PC5 nose shape, height of the mouth
PC6 nasal width, maxillary prognathism
PC7 nasal root shape, mouth width
PC8 cheek protrusion
PC9 midface protrusion, upper facial height
P10 chin height, nasion protrusion
PC1 from a PCA of the midfacial landmark
midfacial landmark network around the nose and mouth
Centroid size facial size
Allometry variation in shape due to size
AL_R_AL_L nasal width
AC_PRN nasal ala length (average)
CH_R_CH_L mouth width
CPH_R_CPH_L philtrum width
EN_EX palpebral ﬁssure length (average)
EN_R_EN_L inner canthal width
EX_R_EX_L outer canthal width
GN_T lower facial depth (average)
LI_SL cutaneous lower lip height
LS_STO upper vermilion height
N_GN morphological facial height
N_MEN nasion to midendocanthion
N_PRN nasal bridge length
N_SN nasal height
N_STO upper facial height
N_T upper facial depth (average)
SBAL_R_SBAL_L subnasal width
SN_GN lower facial height
SN_LS philtrum length
SN_PRN nasal protrusion
SN_STO upper lip height
SN_T midfacial depth (average)
STO_LI lower vermilion height
STO_SL lower lip height
T_R_T_L facial width
Linear distances are the distance between two landmarks (e.g., AL_R and AL_L).
972 J. B. Cole et al.
are based on estimates of h
that have high SE (Table
2). These traits include centroid size, nasion to midendo-
canthion (N_MEN), palpebral ﬁssure length (EN_EX), PC5
representing nose shape and height of the mouth, PC8
representing cheek protrusion, and morphological facial
To elucidate underlying genetic relationships between
different facial traits, we estimated pairwise genetic correla-
tions between all signiﬁcantly heritable traits (Table S4) and
constructed a genetic correlation matrix of all signiﬁcantly
heritable linear distances (Figure 6). Due to lack of power
to detect h
in joint bivariate models between all traits
(Visscher et al. 2014), these genetic correlations are based
on the genetic covariance calculated from .15 million com-
mon variants and not total genetic covariance, resulting in
higher SE. Although not all genetic correlation estimates are
signiﬁcantly different from 0 (P,0.05), Figure 6 depicts
striking patterns of shared heritability among distinct facial
traits. A number of horizontal measurements mostly deﬁned
by nonoverlapping landmarks have high positive genetic cor-
relations with each other. These phenotypes include palpebral
ﬁssure length (EN_EX), outer canthal width (EX_R_EX_L),
facial width (T_R_T_L), mouth width (CH_R_CH_L), subnasal
width (SBAL_R_SBAL_L), philtrum width (CPH_R_CPH_L),
and MidfaceModPC1). We observed a similar pattern of
high positive genetic correlations, though to a lesser extent,
among midline vertical measurements. These include upper
lip height (SN_STO), morphological facial height (N_GN),
upperfacialheight(N_STO),PC2 representing both overall
and lower facial height, lower lip height (STO_SL), and
philtrum length (SN_LS). Importantly, the horizontal and
vertical measurements exhibit large negative genetic corre-
lationswitheachother,indicating that phenotypic variation
along both horizontal and vertical measurements are largely
caused by the same genetic variation acting to increase one
direction while decreasing the other. In a simple sense, this
means that the same alleles that cause an individual to have
a broad face also cause that individual to have a short face,
and vice versa.
We report here the ﬁrst estimates of both heritability and
genetic correlation of facial shape phenotypes derived from
3D facial scans and true genome-wide genetic correlations
between 1000s of individuals. Facial scans provide far more
accurate measurements than previous approaches based on
direct manual measurements between prominent facial
features (Ozsoy et al. 2009). Furthermore, direct calcula-
tion of genome sharing from genome-wide data are more
accurate than kinship coefﬁcients used in traditional her-
itability analyses, which represent the average genetic
sharing for any given relationship and not the actual ge-
netic correlation for any speciﬁc pair of relatives (Hayes
et al. 2009).
Our analysis, carried out in Bantu children from Tanza-
nia, provides the opportunity to assess heritability of facial
shape and size in a young, lean population. The choice of
population is both a strength and a limitation of this anal-
ysis. In any population, heritability is determined by a
combination of genetic variance and environmental inﬂu-
ences. Variation in facial adiposity, for instance, is small in
this population while it may be large in others (Figure S8)
(Cole et al. 2016). The focus on children creates the need
to adjust for age and size but also avoids facial shape
changes that occur later in life due to injury, weight gain,
Not surprisingly, we found that many quantitative facial-
shape phenotypes, derived with high accuracy from 3D facial
Figure 3 Heritability of 38 facial traits.
The bar plot represents h
(blue), and total h
blue) with error bars for all 38 facial
phenotypes analyzed. Bars that ap-
parently have no missing h
dicate that h
narrow-sense heritability of that pheno-
type can be explained fully by common
Heritability of Human Facial Shape 973
scans, are highly heritable. Furthermore, most of these quan-
titative facial phenotypes can be explained by common genetic
variants across the genome. In particular, based on h
horizontal measurements including facial width (T_R_T_L; h
66.2%, SE = 7.5%), nasal width (AL_R_AL_L; h
SE = 7.6%), outer canthalwidth(EX_R_EX_L;h
SE = 7.6%), and palpebral ﬁssure length (EN_EX; h
SE = 7.8%) appear to be among the most heritable facial
features; contrary to ﬁndings of previous heritability
studies of the face. There are several obvious potential
explanations for this difference. First, heritability of some
facial attributes may be population speciﬁc, driven by dif-
ferent underlying genetic variants in different populations;
thus reﬂecting differing underlying biological bases of
facial shape and size. Second, the present study popula-
tion, Tanzanian Bantu children, is a much leaner popula-
tion than has been studied previously. BMI in our cohort
of Tanzanian Bantu children is signiﬁcantly lower than
world standards (Figure S8)(deOniset al. 2004) and is
uncorrelated with our measures of facial shape after
adjusting for age, sex, and centroid size (see Materials
and Methods). Linear measures, particularly horizontal dis-
tances, collected in populations with higher BMI, may be
more affected by excess subcutaneous fat, reﬂecting a
greater environmental component, and thus proportion-
ately smaller genetic component. Third, genetic inﬂu-
ences on horizontal facial distances may be proportionately
greater at younger ages, as in our cohort of children and ado-
lescents ages 3–21, whereas these distances may become pro-
portionately more affected by environmental components
with increasing age. This model of age-related shape differ-
ences ﬁts well with what is already known about how the face
matures and morphs into the adult form. As the face reaches
adult shape at 16 years of age (as determined in males of
Table 2 Heritability of 38 facial traits
) LRT P-value
PC7 0.669 0.138 0.669 0.072 1.00 310
T_R_T_L 0.521 0.138 0.662 0.075 1.00 310
Allometry 0.643 0.132 0.643 0.072 1.00 310
Centroid Size 0.277 0.134 0.641 0.076 3.66 310
N_MEN 0.260 0.134 0.639 0.075 3.89 310
AL_R_AL_L 0.623 0.131 0.623 0.076 1.00 310
PC4 0.604 0.131 0.604 0.075 5.55 310
PC2 0.579 0.139 0.579 0.074 1.00 310
EX_R_EX_L 0.421 0.141 0.566 0.076 8.11 310
N_PRN 0.456 0.142 0.544 0.075 3.94 310
EN_EX 0.208 0.140 0.522 0.078 4.35 310
N_T 0.419 0.136 0.512 0.076 4.52 310
PC5 0.211 0.138 0.491 0.077 6.33 310
N_STO 0.443 0.140 0.490 0.076 2.22 310
GN_T 0.487 0.140 0.487 0.076 2.97 310
SN_LS 0.486 0.130 0.486 0.077 5.26 310
SN_T 0.469 0.139 0.486 0.076 2.97 310
PC3 0.308 0.139 0.478 0.078 2.03 310
PC1 0.477 0.140 0.477 0.076 5.76 310
PC8 0.074 0.137 0.471 0.079 2.50 310
N_SN 0.244 0.137 0.456 0.075 1.95 310
PC9 0.431 0.125 0.452 0.076 1.16 310
MidfaceModPC1 0.433 0.138 0.433 0.078 2.23 310
N_GN 0.159 0.137 0.426 0.078 1.11 310
EN_R_EN_L 0.392 0.142 0.411 0.076 2.09 310
AC_PRN 0.311 0.140 0.410 0.079 1.47 310
SN_GN 0.239 0.139 0.386 0.079 8.70 310
CH_R_CH_L 0.378 0.137 0.378 0.077 7.99 310
SBAL_R_SBAL_L 0.373 0.134 0.373 0.080 2.76 310
LI_SL 0.177 0.134 0.342 0.077 5.61 310
SN_PRN 0.242 0.139 0.340 0.074 1.37 310
CPH_R_CPH_L 0.337 0.126 0.337 0.077 3.45 310
STO_LI 0.324 0.139 0.324 0.075 2.48 310
SN_STO 0.314 0.131 0.314 0.079 1.57 310
PC10 0.291 0.140 0.291 0.080 0.000570 1.000 0.860
STO_SL 0.283 0.134 0.283 0.078 0.000102 1.000 0.863
PC6 0.169 0.131 0.169 0.077 0.0152 1.000 1.10
LS_STO 0.076 0.116 0.076 0.077 0.259 1.000 1.61
, and the proportion of narrow-sense heritability explained by common genetic variants (h
), all with SE.
LRT P-value for the joint model vs. the null model (H
LRT P-value was reported as 0, indicating it was less than the GCTA limit 1 310
974 J. B. Cole et al.
European descent), the midface undergoes a strong vertical
expansion and becomes relatively taller than the rest of the
face (Bastir et al. 2006).
The 10 PCs displayed similar heritabilities as the linear
distances. PCs represent axes of covariation within the data
and most combine variation from multiple if not most land-
marks. A limitation of PCA is that the assumption that each PC
is orthogonal to the previous may not map well onto the
underlying biological determinants of covariation structure.
In the absence of knowledge about those determinants, how-
ever, PCA is a widely accepted and rational approach to
multivariate data. In this context, each PC represents a distinct
facial shape transformation that emerges from the covariance
structure of the data and can be treated as a univariate trait.
Interestingly, global facial size appears to be among the
most heritable of facial traits. Allometry, a measure of the
variation in shape due to size, has h
of 64.3% (SE =
7.2%). Centroid size, a measure of overall face size, has
of 64.1% (SE = 7.6%). These ﬁndings indicate that
there may be a strong genetic basis underlying global size
of the face and how size drives shape variation; whereas
facial shape per se, irrespective of size, may be somewhat
more inﬂuenced by environmental factors. Although our
ﬁndings also indicate that the majority of facial shape
variation can be explained by the effects of common ge-
netic variation, there were several facial phenotypes, in-
cluding centroid size, for which h
did not explain the
majority of h
. Potential explanations for such missing
heritability include variants not in linkage disequilibrium
with variants on our array, rare causal genetic variants,
uncharacterized structural variation, and epistatic ef-
fects. Our genome-wide association studies (GWAS) of
Figure 4 Heritability of linear distances by measurement
orientation. The bar plot represents h
(yellow), missing h
(blue), and total h
(yellow + blue) with error bars for
25 linear distances. Traits are ﬁrst clustered by orientation,
then by facial structure with between-trait phenotypic cor-
relations seen in the colored matrix in the bottom half of
Heritability of Human Facial Shape 975
these same facial traits in Africans identiﬁed two loci
that were signiﬁcantly associated with either centroid
size or allometry (Cole et al. 2016); traits with high h
but variable h
. While these heritability estimates sup-
port an overall genetic contribution, the speciﬁcesti-
mate of h
does not provide information on the magnitude
of effect of contributing loci, and thus is not necessarily
an indicator of GWAS success. Irrespective of the spe-
estimates for centroid size and allometry, our
GWAS of 6300 individuals had the power to detect ge-
netic determinants with relatively large effect sizes for
We observed high positive genetic correlations among
variables that represent similar orientations on the face,
and rather high negative genetic correlations among var-
iables that represent different orientations. The highest
genetic correlations, of horizontal measures across the
facial midline, likely correspond to related genetic effects
on biological relationships underlying facial structure
during development, in which the two sides of the face
meet and fuse at the midline (Sperber et al. 2001). The
negative genetic correlations we observe between horizon-
tal and vertical facial measures are consistent with overall
phenotypic correlations between these measures. Further-
more, high positive and negative genetic correlations be-
tween a wide array of facial traits support the presence of
aﬁnite set of underlying genes involved in overall facial
Finally, our analysis of genetic variance and covariance
structure shows that genetic variation in the face is both
highly integrated and modular. Integration refers to the
developmentally based tendencies for traits to covary
(Hallgrímsson et al. 2009; Klingenberg 2013), while
modularity refers to suites of traits connected by devel-
opment (Wagner et al. 2007). We ﬁnd a large number of
both positive and negative correlations among traits,
attesting to highly structured patterns of variation. For
craniofacial morphology more generally, somatic growth,
Figure 5 Distribution of variance components across the face. (A–C) The anatomical distribution of phenotypic and genetic variances as well as
heritabilities is shown. These are represented as heatmaps based on a thin-plate-spine morph as described in Materials and Methods. (D) The heritability
estimates or the Procrustes-superimposed symmetrized landmarks are shown. (E) The vectors, magniﬁed 10-fold, used to generate the heritability
heatmap are depicted.
976 J. B. Cole et al.
chondrocranial growth, and brain growth are known to
drive such patterns of integrated variation in both mouse
and human crania (Cooper et al. 2004; Hallgrímsson et al.
2006, 2009; Marcucio et al. 2011; Martínez-Abadías et al.
2012). Here, both facial size and facial shape allometry ex-
hibit a pattern of genetic correlations with facial measures
that capture aspects of facial height, midfacial width, and
lower facial prognathism. This pattern of genetic correla-
tions likely reﬂects the overall inﬂuence of somatic growth
on facial shape, forming a developmentally based module
within the face. Further work integrating developmental
studies with results such as these will shed light on the
mechanistic basis for the structure of variation in the face.
This work was funded by grants from the National Institutes of
Health under the National Institute of Dental and Craniofacial
Research (NIDCR) FaceBase Initiative (http://www.nidcr.nih.
gov/; NIDCR DE-020054 to R.A.S.), the Center for Inherited
Disease Research (http://www.cidr.jhmi.edu/; HG006829 to
R.A.S.), the National Institute of Justice (http://www.nij.
gov/Pages/welcome.aspx; 2013-DN-BX-K005 to R.A.S.),
and the National Science and Engineering Council Discov-
ery Grant (http://www.nserc-crsng.gc.ca/index_eng.asp;
DG#238992-12 to B.H.). The funders had no role in study
design, data collection and analysis, decision to publish,
or preparation of the manuscript.
Figure 6 Pairwise genetic correlation matrix of the linear distances. Genetic correlation was calculated from .15 million common genetic variants.
Traits that have high positive genetic correlations with each other are shown in blue, indicating that the same genetic loci alter the magnitude of those
traits in the same direction. Traits that have high negative genetic correlations with each other are shown in red, indicating that the same genetic loci are
contributing to each phenotype in opposite directions, increasing one while decreasing the other. Genetic correlation estimates that are signiﬁcantly
different (P,0.05) from 0 to +1 or 0 to 21 are marked with sand genetic correlation estimates that are signiﬁcantly different (P,0.05) than 0, +1,
and 21 are marked with s.
Heritability of Human Facial Shape 977
Adams, D. C., and E. Otárola-Castillo, 2013 Geomorph: an R
package for the collection and analysis of geometric morphom-
etric shape data. Methods Ecol. Evol. 4: 393–399.
Adams, D. C., M. L. Collyer, E. Otarola-Castillo, and E. Sherratt,
2014 Geomorph: Software for geometric morphometric anal-
yses. R package version 2.1. https://cran.r-project.org/web/
AlKhudhairi, T. D., and E. A. AlKoﬁde, 2010 Cephalometric cra-
niofacial features in Saudi parents and their offspring. Angle
Orthod. 80: 1010–1017.
Amini, F., and A. Borzabadi-Farahani, 2009 Heritability of dental
and skeletal cephalometric variables in monozygous and dizy-
gous Iranian twins. Orthod. Waves 68: 72–79.
Bastir, M., A. Rosas, and P. O’Higgins, 2006 Craniofacial levels
and the morphological maturation of the human skull.
J. Anat. 209: 637–654.
Bookstein, F. L., 1997 Morphometric Tools for Landmark Data:
Geometry and Biology. Cambridge University Press, Cambridge,
Carson, E. A., 2006 Maximum likelihood estimation of human cra-
niometric heritabilities. Am. J. Phys. Anthropol. 131: 169–180.
Cheverud, J. M., 1982 Phenotypic, genetic, and environmental
integration in the cranium. Evolution 36: 499–516.
Cole, J. B., M. F. Manyama, E. Kimwaga, J. Mathayo, J. R. Larson
et al., 2016 Genomewide association study of African children
identiﬁes association of SCHIP1 and PDE8A with facial size and
shape. PLoS Genet. 12: e1006174.
Cooper, D., J. Matyas, M. Katzenberg, and B. Hallgrimsson,
2004 Comparison of microcomputed tomographic and micro-
radiographic measurements of cortical bone porosity. Calcif. Tis-
sue Int. 74: 437–447.
de Onis, M., T. M. A. Wijnhoven, and A. W. Onyango, 2004 Worldwide
practices in child growth monitoring. J. Pediatr. 144: 461–465.
Dryden, I. L., and K. V. Mardia, 1998 Statistical Shape Analysis.
John Wiley & Sons, Chichester, United Kingdom.
2010 Update on facial aging. Aesthet. Surg. J. 30: 11S–24S.
Hallgrímsson, B., J. J. Brown, A. F. Ford‐Hutchinson, H. D. Sheets,
M. L. Zelditch et al., 2006 The brachymorph mouse and the
developmental‐genetic basis for canalization and morphological
integration. Evol. Dev. 8: 61–73.
Hallgrímsson, B., H. Jamniczky, N. M. Young, C. Rolian, T. E.
Parsons et al., 2009 Deciphering the palimpsest: studying the
relationship between morphological integration and phenotypic
covariation. Evol. Biol. 36: 355–376.
Hayes, B. J., P. M. Visscher, and M. E. Goddard, 2009 Increased
accuracy of artiﬁcial selection by using the realized relationship
matrix. Genet. Res. 91: 47–60.
Klingenberg, C. P., 1998 Heterochrony and allometry: the analy-
sis of evolutionary change in ontogeny. Biol. Rev. Camb. Philos.
Soc. 73: 79–123.
Klingenberg, C. P., 2011 MorphoJ: an integrated software package
for geometric morphometrics. Mol. Ecol. Resour. 11: 353–357.
Klingenberg, C. P., 2009 Morphometric integration and modular-
ity in conﬁgurations of landmarks: tools for evaluating a priori
hypotheses. Evol. Dev. 11: 405–421.
Klingenberg, C. P., 2013 Cranial integration and modularity: in-
sights into evolution and development from morphometric data.
Hystrix 24: 43–58.
Klingenberg, C. P., and M. Zimmermann, 1992 Static, ontoge-
netic, and evolutionary allometry: a multivariate comparison
in nine species of water striders. Am. Nat. 140: 601–620.
Klingenberg, C. P., and L. J. Leamy, 2001 Quantitative genetics of
geometric shape in the mouse mandible. Evolution 55: 2342–
Klingenberg, C., L. Leamy, E. Routman, and J. Cheverud,
2001 Genetic architecture of mandible shape in mice: effects
of quantitative trait loci analyzed by geometric morphometrics.
Genetics 157: 785–802.
Laurie, C. C., K. F. Doheny, D. B. Mirel, E. W. Pugh, L. J. Bierut
et al., 2010 Quality control and quality assurance in genotypic
data for genome-wide association studies. Genet. Epidemiol. 34:
Leamy, L., 1977 Genetic and environmental correlations of
morphometric traits in random bred house mice. Evolution 31:
Li, M., J. B. Cole, M. Manyama, J. R. Larson, D. K. Liberton et al.,
2016 Rapid automated landmarking for morphometric analy-
sis of three dimensional facial scans. J. Anat. (in press).
Manfredi, C., R. Martina, G. B. Grossi, and M. Giuliani,
1997 Heritability of 39 orthodontic cephalometric parameters
on MZ, DZ twins and MN-paired singletons. Am. J. Orthod.
Dentofacial Orthop. 111: 44–51.
Marcucio, R. S., N. M. Young, D. Hu, and B. Hallgrimsson,
2011 Mechanisms that underlie co‐variation of the brain and
face. Genesis 49: 177–189.
Martínez-Abadías, N., M. Esparza, T. Sjøvold, R. González-José, M.
Santos et al., 2009a Heritability of human cranial dimensions:
comparing the evolvability of different cranial regions. J. Anat.
Martínez-Abadías, N., C. Paschetta, S. de Azevedo, M. Esparza, and
R. González-José, 2009b Developmental and genetic con-
straints on neurocranial globularity: insights from analyses of
deformed skulls and quantitative genetics. Evol. Biol. 36: 37–56.
Martínez-Abadías, N., P. Mitteroecker, T. E. Parsons, M. Esparza, T.
Sjøvold et al., 2012 The developmental basis of quantitative
craniofacial variation in humans and mice. Evol. Biol. 39: 554–
Mitteroecker, P., and P. Gunz, 2009 Advances in geometric
morphometrics. Evol. Biol. 36: 235–247.
2009 Method selection in craniofacial measurements: advantages
and disadvantages of 3D digitization method. J. Craniomaxillofac.
Surg. 37: 285–290.
Percival, C. J., D. K. Liberton, F. Pardo-Manuel De Villena, R. Spritz,
R. Marcucio et al., 2016 Genetics of murine craniofacial mor-
phology: diallel analysis of the eight founders of the Collaborative
Cross. J. Anat. 228(1): 96–112.
Schlager, S., 2016 Morpho: Calculations and visualizations re-
lated to geometric morphometrics. R package version 184.108.40.206.
Sperber, G. H., G. D. Guttmann, and S. M. Sperber,
2001 Craniofacial Development (Book for Windows & Macin-
tosh). BC Decker Inc, Hamilton, ON.
Visscher, P. M., G. Hemani, A. A. Vinkhuyzen, G.-B. Chen, S. H. Lee
et al., 2014 Statistical power to detect genetic (co)variance of
complex traits using SNP data in unrelated samples. PLoS
Genet. 10: e1004269.
Wagner, G. P., M. Pavlicev, and J. M. Cheverud, 2007 The road to
modularity. Nat. Rev. Genet. 8: 921–931.
Yang, J., B. Benyamin, B. P. McEvoy, S. Gordon, A. K. Henders et al.,
2010 Common SNPs explain a large proportion of the herita-
bility for human height. Nat. Genet. 42: 565–569.
Yang, J., S. H. Lee, M. E. Goddard, and P. M. Visscher,
2011 GCTA: a tool for genome-wide complex trait analysis.
Am. J. Hum. Genet. 88: 76–82.
Zaitlen, N., P. Kraft, N. Patterson, B. Pasaniuc, G. Bhatia et al.,
2013 Using extended genealogy to estimate components of
heritability for 23 quantitative and dichotomous traits. PLoS
Genet. 9: e1003520.
Communicating editor: N. R. Wray
978 J. B. Cole et al.