The myostatin gene, also called “growth and differentiation factor 8” (GDF8 [MIM 601788]) encodes a negative regulator of skeletal-muscle growth.³ First described in the mouse, myostatin is expressed in different muscles throughout the body, both during early development and in adults. Mouse null mutants are significantly larger than wild-type animals, with 200%–300% more skeletal-muscle mass because of an increase in the number of myocytes (hyperplasy) and an increase in the size of muscle fibers (hypertrophy).³ A similar phenotype is seen in some breeds of double-muscled cattle that also have myostatin mutations.^4,5 A loss-of-function mutation in the myostatin gene (a missplicing change in IVS1:G378A) has been associated with muscle hypertrophy in a human subject,⁶ and myostatin expression levels have been shown to be inversely correlated with muscle mass in healthy and HIV-infected subjects.⁷ These data suggest that myostatin acts in a similar fashion among all mammals. Here, we tested the hypothesis that patterns of human nucleotide variation at GDF8 have been shaped by positive natural selection.

We resequenced the complete coding sequence of GDF8, including partial flanking intron sequences and the 5′ upstream cis-promoter region in human panels of 76 African Americans and 70 Europeans. DNA samples were collected from residents of Pittsburgh, after receipt of written informed consent. PCR amplification primers were designed to target the entire coding sequence for each GDF8 exon (I–III) and the putative cis-promoter region for GDF8, located ~500 bp upstream of the ATG start codon.⁸ DNA sequencing was performed on an ABI 3700 automated DNA sequencer with use of amplification oligonucleotides as primers. (Primers and PCR conditions are available on request). Sequences were inspected by eye and were aligned for each individual, with use of the program SEQUENCHER (Gene Codes). The sequence for each individual was submitted to ^GenBank under accession numbers DQ927046–DQ927191.

In the resequenced panels, we detected eight SNPs in the coding sequence, one indel polymorphism in intron 1, and two SNPs in the 5′ upstream cis-promoter region (fig. 1A). Surprisingly, nucleotide diversity in the coding sequence (π=0.038%) is four times higher than in the noncoding sequence (π=0.009%) (table A1). Five of the coding SNPs cause nonconservative amino acid replacement changes, whereas the three remaining coding SNPs cause silent changes (table 1).

Under the neutral model of molecular evolution, the ratio of replacement:silent changes is expected to be the same within and between species (McDonald-Kreitman test).¹¹ To test this null hypothesis for GDF8, we compared the polymorphism within humans with divergence between humans and other mammals (table B1). Between species, we observed many more silent than replacement mutations, as is typical for most genes. However, within humans, we observed more replacement than silent mutations, and this difference is statistically significant (table 2). This pattern is unusual for human genes and, together with observations described below, suggests that the high proportion of amino acid variation in humans is due to positive natural selection.

To further explore whether the patterns seen at GDF8 in humans are unexpected, we compared GDF8 sequences, obtained from public databases, among mammals and other vertebrates. We estimated the average ratio of replacement:silent substitutions per site (d_N:d_S) across 15 mammalian lineages, using maximum likelihood¹² (table B1). In the absence of functional constraint, the nonsynonymous substitution rate (d_N) is expected to be equal to the synonymous substitution rate (d_S), whereas d_N:d_S<1 is indicative of purifying selection. The average d_N:d_S ratio across the mammalian phylogeny is 0.10, which suggests that GDF8 has been under strong constraint throughout much of mammalian evolution. Furthermore, most pairwise interspecific comparisons with humans show even higher levels of constraint. For example, between mouse and human d_N:d_S=0.05 for GDF8, well below the median genomewide value of 0.12.¹³ Thus, the pattern of five replacement and three silent changes within humans stands out as being exceptional.

Next, we looked at the amino acid sites of each of the five human replacement polymorphisms across all vertebrate species with available GDF8 sequences (a total of 20 species, including fish, birds, and mammals [table B1]). The five sites that are polymorphic in humans are remarkably conserved over evolutionary timescales. The ancestral amino acid states associated with the alleles Ala55Thr (G163A), Arg65His (G194A), and Asp103Asn (G307A) are conserved among all vertebrates; Lys153Arg (A2246G) is conserved among all taxa except fish (Danio rerio), and Met129Arg (T2174G) is conserved among all mammals except bovines. This high level of conservation for individual residues suggests that the mutations in humans have functional consequences.

To learn more about patterns of evolution at GDF8 in humans, we inferred haplotype phase for the diploid sequences.⁹ The two intermediate frequency polymorphisms—Ala55Thr and Lys153Arg (nucleotide mutations 163 and 2246, respectively)—reside on separate haplotypes (fig. 1A). The mutation at site 163 is associated with two haplotypes: A and K (referred to as “haplogroup 55”) (fig. 1A). Haplotype K bears a derived mutation in the GDF8 promoter region (site −762) that is found exclusively in association with the nonsynonymous change at site 163. The mutation at site 2246 is associated with six different haplotypes (haplotypes C, D, F, G, H, and J; referred to as “haplogroup 153”). Interestingly, all the low-frequency replacement mutations appear to be in association with a haplotype bearing the 2246 mutation (fig. 1B). We estimated the age of the mutations 2246 and 163, using long-range linked-polymorphism data (see YRI HapMap data below). The age estimates suggest that both alleles are relatively young and arose within the past 10,000 years (appendix C).

Alleles that have experienced recent positive selection may bear a signature of unusually long-range linkage disequilibrium with surrounding SNPs.^15,16 To test for this pattern at GDF8, we examined the SNP data from the International HapMap Project^,17 in the genomic region of GDF8. Both sites G163A and A2246G have been genotyped in the International HapMap Project (SNPs rs1805085 and rs1805086, respectively), and the minor-allele frequency for both SNPs in the Yoruban (YRI) panel from Ibadan, Nigeria, is 22%. We retrieved phased haplotypes spanning 300 kb roughly centered on GDF8 from the International HapMap Project (human genome build 16) for 60 individuals from the YRI panel. We similarly retrieved phased SNP data from the YRI panel at 10 additional anonymous genomic regions, each spanning ~300 kb from across chromosome 2. Nonoverlapping core haplotypes (restricted to a maximum size of 8 contiguous SNPs) were defined in each genomic region, and extended haplotype homozygosity (EHH) was calculated at increasing genetic distances (measured in centimorgans).¹⁵ We used the cores of the anonymous regions to generate an empirical distribution for the relationship between EHH and haplotype frequency. Statistical significance for departure of an EHH value within a frequency bin was determined for given GDF8 haplotypes relative to the empirical distribution data. All long-range haplotype analyses were conducted using ^SWEEP software, according to standard documentation.¹⁵ Using these data, we determined that the long-range haplotypes associated with haplogroups 153 and 55 exhibit a significant level of EHH relative to other core haplotypes up to ~0.2 cM away from GDF8 (fig. 2A) and relative to an empirical distribution of core haplotypes from other genomic loci (fig. 2B) (P=.01 and P=.03 for haplogroups 153 and 55, respectively). It is noteworthy that a recent genomic scan for positive selection in the human genome that was based on a modification of the EHH statistic (i.e., the integrated haplotype score [IHS]) failed to identify long-range haplotypes in the genomic region of GDF8 as extreme (top 1%) outliers.¹⁴ However, this method (the IHS statistic) lacks power to detect selection on derived haplotypes at population frequencies < 0.5,¹⁴ as is the case for GDF8 and some other loci known to be under selection (e.g., G6PD [MIM 305900]).

Interestingly, two replacement polymorphisms—G163A and A2246G (representing haplogroups 55 and 153, respectively)—are at relatively high frequency among African Americans (12% and 20%, respectively) and the YRI panel (22% each) but are at much lower frequencies among the Europeans sampled here (1% and 4%, respectively) (table 1) and in the HapMap panels from Europe (0% and 2%, respectively) and Asia (0% each). Furthermore, it is notable that >50% (39 of 76) of the African Americans and 75% (45 of 60) of the YRI individuals bear at least one of these replacement alleles (figs. (figs.1A1A and andA1).A1). To have a more complete view of the frequencies of G163A and A2246G in human populations, we genotyped 1,040 individuals in a worldwide panel that included samples from Africa, Europe, Asia, and the Americas (Human Genome Diversity Project [HGDP]–CEPH Diversity Panel).¹⁸ SNP genotyping of this global panel was performed using TaqMan Assays by Design (Applied Biosystems). A separate assay probe was designed for each of the polymorphisms (G163A and A2246G), and reactions were performed, following manufacturer’s protocols, on an ABI 7500 Real-Time PCR machine. All homozygotes of the rare alleles and any ambiguous calls were confirmed by PCR and direct resequencing. Results of the global survey show that one or both of mutations G163A or A2246G are found in populations of sub-Saharan Africa at a frequency >14%, whereas, in non-African populations, their frequencies are typically absent and only rarely exceed 5% (fig. 3 and table 3).

The signature of positive selection at the molecular level in humans is often weak, in part because of the relatively small long-term effective population size and low levels of standing variation. As a result, recent studies have focused on methods to detect selection on the basis of subtle aspects of the data, such as the decay of long-range linkage disequilibrium around a target of selection.^14,15,19 In contrast, the results presented here for GDF8 also reveal a strong signature of positive selection that is based on an excess of nonsynonymous polymorphism, as previously seen only for a few other genes in humans (e.g., G6PD and the major histocompatibility complex).^20,21 An earlier survey of human variation at GDF8 revealed three additional low-frequency replacement polymorphisms (but no additional silent polymorphisms) that were not detected in our current resequenced panels,¹⁰ which increases to eight the total number of replacement changes at GDF8 that have been reported to date (table 1) and contributes further to the observed excess of replacement polymorphisms among humans.

Although positive selection may increase the level of polymorphism at a locus, the excess of replacement changes seen at GDF8 could, in principle, alternatively be explained either by a recent relaxation of selective constraint or by the presence of slightly deleterious variants among humans.^22,23 However, both of these explanations seem unlikely for GDF8. Relaxed selection at GDF8 is improbable, in view of the overall strong conservation of this gene over deep evolutionary timescales and the major phenotypic effect associated with loss of function in humans.⁶ The two major replacement variants at GDF8 are unlikely to be slightly deleterious, because these are at relatively high frequency (up to 31%) in sub-Saharan Africans. Moreover, the atypical long-range haplotype conservation associated with haplogroups 153 and 55 suggests that these variants have rapidly increased in frequency. Although demographic processes associated with population bottlenecks and expansions may create long-range haplotype patterns that mimic a signature of selection,²⁴ this is not likely to be the case at myostatin, since the signature of selection is seen in the ancestral African population. Together, these data argue that some form of recent diversifying selection has played a significant role in shaping patterns of variation at GDF8.

The molecular positions of polymorphic residues 55 and 153 within the human GDF8 peptide allow us to speculate about the phenotypic consequences of these variants. Both residues are found within the propeptide region (residues 1–266) of GDF8. As is characteristic of other members of the transforming growth factor beta (TGF-β) superfamily, the GDF8 precursor peptide is cleaved into an (N-terminus) propeptide and a (C-terminus) mature peptide. The active form of myostatin is a homodimer of the mature peptide, which binds to extracellular activin type II receptors (ACTRIIB [MIM 602730]) to induce intracellular activation of SMAD proteins.²⁵ Importantly, the propeptide of GDF8 binds to the mature homodimer to form a latent myostatin complex and thus regulates GDF8 activity by preventing the homodimer from binding to its target receptors.²⁶ Concordantly, overexpression of GDF8 propeptide in transgenic mice causes muscle hypertrophy and hyperplasia similar to that in GDF8-null mutants.^26,27 Moreover, intraperitoneal administration of myostatin propeptide to Mdx mice (models for Duchenne muscular dystrophy) has been shown to rescue some of the muscular pathophysiological effects found in this mutant.²⁸ Interestingly, residue 55 is within a major inhibitory domain of the GDF8 propeptide (residues 42–115)²⁹ and therefore may influence the regulatory properties of the propeptide. In general, any mutations that increase the binding affinity between the propeptide and the mature peptide could generate a relative deficiency of myostatin activity. One of the many possible adaptive implications of such an effect could be protection from muscle wasting in times of famine, a potentially recurrent phenomenon for early agricultural societies.³⁰