Height is a complex trait influenced by genes and a variety of environmental factors, including diet and the prenatal environment³. Heritability estimates suggest that ≥ 80% of variation in height may be genetically determined^3,4. Rare mutations with large effects on height in Mendelian syndromes have been identified in FBN1, FGFR3, GH1, EVC1, and GPC3 (MIM 154700, 134934, 262400, 604831, and 312870). Despite the high heritability, numerous candidate gene and linkage studies to identify loci influencing height in individuals of “normal stature” have been inconclusive⁵. Overall, variation in human height is likely to be polygenic and heterogeneous. Recently, the first GWAS of height¹ identified common variants in HMGA2 associated with normal variation in height, both in adults (p=4 × 10⁻¹⁶) and children (p=1 × 10⁻⁶); the variants account for a small fraction (~0.3%) of the overall variation in height.

To identify additional genetic variants associated with height, we analyzed genome-wide SNP data on 2,371 Finns from the FUSION study⁶ and 4,298 Sardinians from the SardiNIA study⁷ (Table 1). The two samples were originally genotyped with distinct sets of markers. We used genotype imputation methods^6,8 to facilitate comparison of the two studies and evaluate association between height and ~2.28 million common genetic variants. After verifying the overall accuracy of imputed genotypes in a subset of markers, we conducted within-study analyses using a rapid variance components-based association test⁹ and then carried out a meta-analysis of the two studies (Supplementary Figure 1).

Our results provided confirmatory evidence of association with rs1042725 and rs7968682, the recently reported common variants in HMGA2 (p=.031 and .0093, respectively, both in the same direction as the original report: Supplementary Table 1)¹. The five loci showing the most significant evidence of association in our study are shown in Supplementary Table 2. To our knowledge, common variants in these loci have not been associated with height previously.

The genes near our strongest signal (p<2 × 10⁻⁷), located on chromosome 20, have a plausible biological role in human height. Rare variants in growth differentiation factor 5 (GDF5) have been associated with disorders of skeletal development (see below), and BFZB (also known as C20orf44 and UQCC) encodes a ZIC-binding protein repressed by basic fibroblast growth factor¹⁰.

We pursued the chromosome 20 signal because it was the single best result in our initial scan, the surrounding region accounts for 40 of the 50 lowest p-values in our meta-analysis, and it overlaps an osteoarthritis susceptibility locus². We focused our follow-up efforts for this locus on SNP rs6060369 (p = 9.7 × 10⁻⁷ in the initial meta-analysis, Table 2) because it exhibited the strongest evidence for association among all Affymetrix Mapping 500K SNPs, which had been genotyped in the majority of our GWAS samples; we favored genotyped over imputed SNPs for initial follow-up. The SNP most significantly associated with height in the meta-analysis of imputed HapMap SNPs was rs725908, which is in strong LD with rs6060369 (r² = 0.96, Supplementary Table 2). SNP rs6060369 was initially imputed in the FUSION GWAS, but direct genotyping sustained strong evidence of association (meta-analysis p=1.5 × 10⁻⁶). In the GWAS samples, each copy of the C allele at rs6060369 was associated with an increase in height of 0.3 to 0.7 cm (accounting for 0.3–0.6% of the variance in height, after adjustment for age and sex).

Motivated by previous reports of sex differences at osteoarthritis-associated loci¹¹, we carried out an analysis stratified by sex. The stratified analyses show strong evidence of association in both males and females with no evidence of heterogeneity (Supplementary Table 3). We did not detect significant association of the SNP with other anthropometric measures (p>.05 for weight, body mass index, waist circumference, hip circumference, and waist-to-hip ratio).

We next tested the association of rs6060369 with height in nine follow-up samples, comprising 23,684 individuals of European ancestry and 3,860 African American individuals (Table 2). Six of the samples provided significant evidence of association (p<.05) and the other three showed a trend (p<.20) in the same direction (Table 2). The p-value for association in all 27,544 follow-up samples was 1.05 × 10⁻¹¹, and in all 34,213 GWAS and follow-up samples combined was 2.22 × 10⁻¹⁶ (Figure 1). In the follow-up samples, each copy of the C allele at rs6060369 was associated with an increase in height of 0.2 – 0.6 cm (Table 2), and overall we estimate the SNP explains 0.3–0.5% of the variance in height, both in males and females (Supplementary Table 3). Further independent evidence for association between rs6060369 and adult height comes from the 1958 British Birth Cohort for which rs6060369 is associated with height measured at 44–45 years of age (p = .0046, explaining 0.5% of the variance) (http://www.b58cgene.sgul.ac.uk/, accessed October 2007).

Open in a separate window

Figure 1

Evidence of (A) association with height and (B) linkage disequilibrium around GDF5 and BFZB/C20orf44/UQCC

All genotyped or imputed SNPs in the SardiNIA and FUSION GWA studies are plotted with association p-values (additive test) compared to genomic position in NCBI Build 35 (gray circles, red circles for labeled SNPs). Yellow squares indicate p-values for SNPs analyzed in a portion of follow up samples (FUSION Stage 1 and 2, SardiNIA, DGI, and ARIC studies). The green triangle indicates rs6060369 analyzed in all GWAS and follow up samples. Patterns of linkage disequilibrium (r²) for two of the HapMap populations (CEU, Utah residents with European ancestry, and YRI, Yoruba)³⁰ are plotted and colored with a 15 color red-green-blue scale, to represent values ranging from high (red), to intermediate (green), to low (blue).

Our association signal lies in a 136 kb stretch of linkage disequilibrium (LD) that contains two genes, GDF5 and BFZB (Figure 1, Supplementary Table 4). BFZB is expressed in differentiating chondrocytes¹², starting at early stages of mesenchymal cell proliferation¹³. Studies in mouse embryonic stem cells have shown that BFZB is down-regulated on addition of FGF2 (bFGF)¹⁰, which acts in concert with bone morphogenic proteins and several Hox gene products to initiate and promote morphogenesis and growth of the skeleton. Thus, BFZB appears to be involved in a network of FGF2-regulated growth control. GDF5 is a member of the TGF-beta superfamily involved in bone growth and differentiation, both in adult and embryonic tissues^14,15. GDF5 expression is typically restricted to the primordial cartilage of long bones, with little expression in the vertebrae and ribs¹⁴. Mutations in GDF5 are associated with several disorders of skeletal development (MIM 201250, 200700, 112600, 113100, 228900, 185800 and 186500). Recombinant human GDF5 has been shown to restore vertebral disc height in a rabbit disc degeneration model, perhaps through enhanced extracellular matrix production¹⁵. Other nearby genes do not appear to be involved in chondrocyte differentiation, bone growth, or development, but the locus that includes GDF5 and BFZB is also interesting because it was highlighted as showing very strong evidence for selection in a genome-wide search for regions that underwent recent selection¹⁶. The target of selection is presently unknown.

A SNP located in the 5’ untranslated region of GDF5, rs143383, is strongly associated with osteoarthritis^2,17 and is estimated to be in very strong LD with rs6060369 in the HapMap, FUSION, and SardiNIA samples (r²=.83–.90). The SNP appears to influence GDF5 expression^2,17 and, we reasoned, might be a causal variant. Thus, we analyzed this SNP in our screening samples and a subset of our follow-up samples. The rs143383 A allele previously associated with increased risk of osteoarthritis was associated with decreased height in our studies (p=5.01 × 10⁻¹² versus p=4.08 × 10⁻⁹ for rs6060369 in the same subset of samples; Figure 1, Table 3). Analysis stratified by gender shows strong association in both males and females (Supplementary Table 5).

Interestingly, the ARIC African American samples that showed only a trend toward association with rs6060369 (p=.17) showed significant evidence of association with rs143383 (p=.011), illustrating the utility of studying different ancestral groups in the fine-mapping of complex disease genes^18,19. In the ARIC African American samples, even when rs6060369 was included in a regression model, rs143383 remained marginally associated with height (p=.034, estimated additive effect = 0.579 cm), while the association of rs6060369 disappeared (p=.92, estimated additive effect = −0.019 cm) when conditioning on rs143383. These results suggest that GDF5 is more likely to influence height, although other nonsynonymous SNPs present in GDF5 and BFZB may affect function instead or as well.

Miyamoto and colleagues demonstrated that the A allele of rs143383 is associated with decreased GDF5 transcriptional activity in chondrogenic cells². Decreased expression of GDF5 would logically lead to decreased limb bone growth, consistent with decreased height, as we observed. Decreased transcription of GDF5 may influence the amount of cartilage of the vertebrae, limb proportions, or joint angles, resulting both in a modest decrease in stature and susceptibility to osteoarthritis.

To evaluate the impact of osteoarthritis as a confounding factor, we repeated the association analysis restricted to younger individuals (age < 40). In SardiNIA, we analyzed 1,964 individuals and confirmed the association (p=.0018 for rs6060369, and p=.015 for rs143383) with effect size similar to estimates as for the combined sample (0.70 cm per copy of the C allele for rs6060369). In the Old Order Amish, the younger subgroup of 891 individuals showed a trend toward association of allele C with increased height (0.60 cm per copy of the C allele at rs6060369), but no significant association (p = .86), likely due to low statistical power.

To compare the evidence of association with length of long bones compared to vertebrae and skull, we tested rs6060369 and rs143383 for evidence of association with sitting height, which was measured in the ARIC and BWHHS studies. In ARIC European Americans and the BWHHS British sample, similar evidence of association was observed for both standing and sitting height. In ARIC African Americans, only rs143383 was significantly associated (p<.05) with height, and it was associated only with standing height, not with sitting height (Table 3), perhaps suggesting a stronger effect on long bones than on vertebrae.

Multiple regression analysis of our data suggests that a single common variant in the region may underlie the evidence of association. Specifically, multiple regression analysis in GWAS samples showed that after including rs6060369, rs143383, or rs725908 as a covariate, other association signals in the region become non-significant. One of these common variants or another nearby unmeasured variant in LD may influence height through expression of GDF5 ^2,17. However, either or both GDF5 and BFZB could be affected. Thoroughly evaluating the contribution of this locus to human height will require re-sequencing to catalog all genetic variants in the region and genotyping to evaluate their effects.

Combined, the variants identified here and previously reported in HMGA2 account for <1% of the variance in height, so that most of the 80% of variation in height estimated to be genetic remains unexplained. Our GWAS provides suggestive evidence for several other loci influencing height. For example, after excluding SNPs within 250 kb from GDF5, we observed a slight excess of SNPs with p-value <10⁻⁵ (38 observed versus 23 expected, Supplementary Figure 2). Still, it appears likely that many of the common variants influencing height will have only small effects. Follow-up of additional SNPs in larger meta-analyses will be necessary to define these variants²⁰, which may also be relevant not only to normal variation in height but also to musculoskeletal or other diseases.