We are all familiar with the strong genetic control of faces seen in the almost indistinguishable facial appearance of identical twins, similarities within families and populations, as well as shared facial characteristics in medical syndromes. Nevertheless, our understanding of the genetic architecture of normal-range human facial variation has remained largely uninvestigated, until recently.
Cole et al., 2016 , and Shaffer et al., 2016 , present us with additional insights into the complex puzzle that is facial genetics. These two studies along with one other, Ahidkari et al., 2016 , represent the second round of genome-wide association studies (GWAS) on facial morphology; the first round appeared in 2012 [4,5]. Although genes affecting the development of the face have been studied for decades using pedigree-based analyses, molecular biology, and animal models, the discovery of genetic variants affecting normal-range facial variation using GWAS is still in its infancy. The progression of facial GWAS has been impeded by several factors, including a lack of appropriate 3-D facial scanning systems, the significant efforts required to collect participants, subsequent rigorous statistical analyses, and the lack of methods for functional follow-up analyses . The two facial GWAS from 2012 provided a handful of associated genes, one of which, PAX3, was significant in both studies (Fig 1). Evidence confirming a role for PAX3 is presented in the Shaffer et al. report, which also replicated three other genes from the 2012 GWAS: CACNA2D3, C5orf50, and PRDM16 . The failure to replicate the other associations could result from many factors in study design, such as consistent facial phenotyping as noted by  and population background as noted by . Future studies bear the responsibility to investigate earlier findings for replication, and with these two new facial GWAS, this list has tripled in length.
These two GWAS provide some new and exciting loci, datasets, and approaches. The increased list of craniofacial genes brings us closer to a better understanding of facial development and disease. They also increase our optimism in realizing unresolved medical and forensic applications that have compelling social implications. Although many genetic conditions that involve clinically significant (atypical) patterns of facial development have been mapped, and these studies have arguably established most of what we know about facial genetics, many individual cases remain undiagnosed. A better understanding of the genetics of normal-range craniofacial development can and should help in the delineation of which genes underlie these conditions. Generally, this process has been inverted, with information on the genetic determinants of disorders serving to help identify candidate genes for investigations into normal-range facial variation . Understanding the genetic architecture of normal-range facial variation would also assist efforts to predict faces from DNA , a forensic application that has generated much public interest.
One major strength of the two new studies is the availability of the facial shape and genotype data through FaceBase (www.facebase.org) and dbGaP (www.ncbi.nlm.nih.gov/gap), respectively. Both studies also provide examples for how to deal with imperfections and confounders typically encountered in 3-D facial datasets. For example, they deal with different scanning equipment, facial imaging and phenotyping quality control, and the issue of manual versus automated landmarking. Shaffer et al.  used inter-landmark distances that are known to be clinically relevant indicators. In contrast, Cole et al.  used both inter-landmark distances and geometric morphometrics, an approach to shape analysis that has been used extensively in evolutionary and developmental biology. Alternatively, Adhikari et al.  used an observer-based facial phenotyping approach where faces are subjectively assessed by human observers either into classes or on an ordinal scale based on the presence or absence of certain characteristics. Each phenotyping approach has its own advantages, but it is challenging to relate different kinds of facial measurements. More complete descriptions of how factors like particular SNPs affect facial variation, as displayed in Fig 1, will provide a stronger basis for making comparisons between the effects uncovered in one association study to those found in the next GWAS as well as comparisons between normal-range effects and the effects modeled in craniofacial disease.
The two new studies are ambitious in recruiting participants whose ages span from young  to a wide range . Facial growth and development and, later in life, facial aging are both complex and non-linear phenomena. Therefore, associations found in children may not be found in adults and vice versa. However, quoting Cole et al.  in their response to reviewers, “…that does not negate the relevance of studying facial shape in children, in which loci that contribute to developmental aspects of facial shape may be more engaged and relevant in children (with still-developing faces) than in adults.” Furthermore, in their response to reviewers, Shaffer et al.  said, “Failure to completely account for the effects of age variation on traits is highly unlikely to result in false positive genome-wide signals. That is, age cannot act as an epidemiological confounder because age does not alter the constitutional genome…” Also of note is the use of non-European populations, such as a cohort of East African children , and an admixed Latin-American sample .
In moving forward, it is worthwhile to reflect briefly on a somewhat distracting obsession in GWAS, namely, achieving genome-wide statistical significance. In a European cohort, the significance threshold is 5*10^-8, while in an African cohort, due to shorter haplotype blocks, an even more stringent threshold (2.5*10^-8) is used . The current trend in GWAS is to increase the number of SNPs (up to 10M) throughout the genome, increasing the odds of including relevant SNPs, which is made possible thanks to increased ease of genotyping and genotype imputation. However, this is not the only way forward. Both of these new facial GWAS illustrate the use of contemporary genetic methodology such as meta-analyses and linear mixed models. In association studies, linear mixed models are emerging as a method of choice  where recent promising advancements have been made . However, in striking contrast, facial phenotypes are still reduced to a limited series of a priori simplifications that are analyzed individually. However, the solution is not simply to increase the number of facial measurements, increasing the odds of including relevant phenotypic traits. The computational burden and, more importantly, the considerations of statistical power make this approach inefficient.
An alternative approach involves identifying phenotypic measurements taken from the same facial subunit or module and using these measures jointly in multivariate association analyses. The human body is organized into multiple organ systems, each of which is oriented towards a specific task in accordance to its function, anatomy, and embryological origin. Facial shape, likewise, represents a complex system that is experimentally and developmentally separable into several modules—for instance, bone and muscle cells for structural support and movement; skin cells for transpiration, protection, and temperature regulation; and retinal cells for image processing. Combining linear measurements, e.g., using factor analysis, can facilitate the description of unmeasured facial traits. Simply measuring and combining two sides of a triangle also captures information on the unmeasured third side. There is no need to also measure and investigate the third side of the triangle separately. The primary challenge lies in defining relevant groupings of measurements. Combining either too many or unrelated measurements in a multivariate analysis will reduce statistical power. Thus, defining which measures are related is not trivial. Further research on facial anatomy and phenotypic correlation, heritability and genetic correlation, modularity and integration, as well as adaptation and evolution will provide additional clues for grouping facial traits. The work and data presented in these new facial GWAS already allow testing of such an approach.
1. Cole JB, Manyama M, Kimwaga E, Mathayo J, Larson JR, Libertib DK, et al. (2016) Genomewide association study of african children identifies association of SHIP1 and PDE8A with facial size and shape. PLoS Genet 12(8): e1006174.
2. Shaffer JR, Orlova E, Lee MK, Leslie EJ, Raffensperger ZD, Heike CL et al. (2016) Genome-wide association study reveals multiple loci influencing normal human facial morphology. PLoS Genet 12(8): e1006149.
3. Adhikari K, Fuentes-Guajardo M, Quinto-Sánchez M, Mendoza-Revilla J, Chacón-Duque JC, et al. (2016) A genome-wide association scan implicates DCHS2, RUNX2, GLI3, PAX1 and EDAR in human facial variation. Nature communications 7: 11616. doi: 10.1038/ncomms11616 27193062
4. Liu F, van der Lijn F, Schurmann C, Zhu G, Chakravarty MM, et al. (2012) A genome-wide association study identifies five loci influencing facial morphology in Europeans. PLoS Genet 8: e1002932. doi: 10.1371/journal.pgen.1002932 23028347
5. Paternoster L, Zhurov AI, Toma AM, Kemp JP, St Pourcain B, et al. (2012) Genome-wide Association Study of Three-Dimensional Facial Morphology Identifies a Variant in PAX3 Associated with Nasion Position. The American Journal of Human Genetics 90: 478–485. doi: 10.1016/j.ajhg.2011.12.021 22341974
6. Prescott SL, Srinivasan R, Marchetto MC, Grishina I, Narvaiza I, et al. (2015) Enhancer divergence and cis-regulatory evolution in the human and chimp neural crest. Cell 163: 68–83. doi: 10.1016/j.cell.2015.08.036 26365491
7. Claes P, Liberton DK, Daniels K, Rosana KM, Quillen EE, et al. (2014) Modeling 3D facial shape from DNA. PLoS Genet 10: e1004224. doi: 10.1371/journal.pgen.1004224 24651127
9. Pe'er I, Yelensky R, Altshuler D, Daly MJ (2008) Estimation of the multiple testing burden for genomewide association studies of nearly all common variants. Genetic epidemiology 32: 381–385. doi: 10.1002/gepi.20303 18348202
10. Yang J, Zaitlen NA, Goddard ME, Visscher PM, Price AL (2014) Advantages and pitfalls in the application of mixed-model association methods. Nature genetics 46: 100–106. doi: 10.1038/ng.2876 24473328
11. Joo JWJ, Hormozdiari F, Han B, Eskin E (2016) Multiple testing correction in linear mixed models. Genome biology 17: 1.