Genome-Wide Patterns of Genetic Distances Reveal Candidate Loci Contributing to Human Population-Specific Traits

Authors

  • João Pedro de Magalhães,

    Corresponding author
    1. Integrative Genomics of Ageing Group, Institute of Integrative Biology, University of Liverpool, Liverpool, UK
    Search for more papers by this author
  • Alex Matsuda

    1. Escola Superior de Biotecnologia, Porto, Portugal
    Search for more papers by this author
    • Present address: Departments of Anesthesia and Medicine, and Division of Cardiovascular Medicine, Brigham and Women's Hospital, Harvard Medical School, Boston, MA, USA.


João Pedro de Magalhães, Biosciences Building, Crown Street, Liverpool L69 7ZB, UK. Tel: +44 151 7954517; Fax: +44 151 7954408; E-mail: jp@senescence.info

Summary

Modern humans originated in Africa before migrating across the world with founder effects and adaptations to new environments contributing to their present phenotypic diversity. Determining the genetic basis of differences between populations may provide clues about our evolutionary history and may have clinical implications. Herein, we develop a method to detect genes and biological processes in which populations most differ by calculating the genetic distance between modern populations and a hypothetical ancestral population. We apply our method to large-scale single nucleotide polymorphism (SNP) data from human populations of African, European and Asian origin. As expected, ancestral alleles were more conserved in the African populations and we found evidence of high divergence in genes previously suggested as targets of selection related to skin pigmentation, immune response, senses and dietary adaptations. Our genome-wide scan also reveals novel candidates for contributing to population-specific traits. These include genes related to neuronal development and behavior that may have been influenced by cultural processes. Moreover, in the African populations, we found a high divergence in genes related to UV protection and to the male reproductive system. Taken together, these results confirm and expand previous findings, providing new clues about the evolution and genetics of human phenotypic diversity.

Introduction

The evolution of human populations is a fascinating but controversial topic. Arguably the only point of consensus is the Out-of-Africa hypothesis, the idea that modern humans originated in Africa and then migrated into the rest of the world (Stringer & McKie, 1996; Tishkoff & Kidd, 2004; Handley et al., 2007). Several studies using mitochondrial DNA or polymorphic Alu insertions support the Out-of-Africa hypothesis and suggest bottlenecks in the exodus of modern humans from Africa (Cann et al., 1987; Batzer et al., 1994; Watkins et al., 2001; Watkins et al., 2003). Population bottlenecks during recent human evolution caused founder effects which contributed to genetic and phenotypic differences between populations (Tishkoff & Kidd, 2004; Ramachandran et al., 2005). Moreover, as human populations adapted to new environments around the world, including new climates, pathogens, and food sources, they became subject to unique evolutionary pressures (Tang et al., 2007; Barreiro et al., 2008). Cultural and even social differences may have also driven the evolution of population-specific traits (Laland et al., 2010). The evolutionary history of each population thus shapes its genetics and a combination of factors, including chance and natural selection, contributed to the present phenotypic diversity of human populations.

There is great interest in determining the genetic basis for differences between populations. They may provide clues about the evolutionary history of our species. Moreover, in biomedical research and from a clinical perspective, knowledge of genetic differences between populations are relevant as they could aid in the diagnosis and treatment of several conditions, particularly now in the era of personalized medicine (Burchard et al., 2003; Tishkoff & Kidd, 2004).

Recent large-scale projects have provided data on millions of genetic variants–in particular single nucleotide polymorphisms (SNPs)–from human populations originating in different geographical locations. A number of methods have been developed to detect positive selection in the human genome and applied to large-scale SNP data (Akey et al., 2002; Weir et al., 2005; Voight et al., 2006; Myles et al., 2007; Sabeti et al., 2007; Tang et al., 2007; Williamson et al., 2007; Barreiro et al., 2008; Myles et al., 2008; Johansson & Gyllensten, 2008; Coop et al., 2009; Pickrell et al., 2009; Chen et al., 2010). These and other previous studies have identified a number of candidate loci under selection in human populations, some of which have been shown to be phenotypically relevant. Briefly, one of the best studied loci is the lactase gene (LCT). The ability to digest lactose in adulthood is higher in Northern European populations, which has been suggested as an adaptation accompanying dairy farming. This hypothesis is supported by evidence of recent positive selection in LCT in Europeans (Bersaglieri et al., 2004). Skin pigmentation is another well-studied adaptation with several genes involved in skin pigmentation–such as SLC45A2 and SLC24A5–exhibiting evidence of recent positive selection in Europeans (Graf et al., 2005; Lamason et al., 2005; Myles et al., 2007). Finally, the Duffy blood group locus (DARC), a membrane protein in which a mutation confers resistant to vivax malaria, has long been considered a likely target of natural selection in African populations with high incidence of vivax malaria (Hamblin & Di Rienzo, 2000). Therefore, to name just three well-studied examples, lactose tolerance, skin pigmentation, and resistance to vivax malaria are three key phenotypic differences between populations likely driven by adaptations of human populations to new environments.

Studies focusing on population differences, such as those using the fixation index (Fst) as a measure of population differentiation and methods employing haplotype structure like extended haplotype homozygosity, by and large focus on detecting signatures of natural selection. While natural selection played a major role in determining differences between modern human populations, however, chance and demographic effects also played a crucial role. Indeed, random genetic drift and founder effects contributed to population differences that may have implications in medical genetics (Tishkoff & Kidd, 2004; Hartl & Clark, 2007). In this work, we wanted to take advantage of large-scale SNP datasets to determine candidate loci of population-specific traits without assumptions of whether genes were targets of selection or merely diverged by chance between populations. Therefore, we develop a simple method to infer population-specific genetic divergence by calculating the genetic distance between modern populations and a hypothetical ancestral population.

We apply our model-free method to large-scale SNP data from human populations of African, European, and Asian geographic origins. Initially we employ data obtained by Perlegen Sciences (Hinds et al., 2005) and later validate these results using data from the International HapMap Project (Frazer et al., 2007). Our genome-wide scan allowed us to identify outliers, genes and biological functions in which human populations most diverge from the ancestral state, which include several previously known candidates of recent positive selection as well as loci not previously identified as targets of selection but that may be related to population-specific phenotypes and disease susceptibility.

Materials and Methods

Data Sources and Selection

SNP data from the Perlegen and HapMap datasets were used in this study. The Perlegen dataset consists of genome-wide SNPs from three different populations of three different geographic regions: 23 African Americans (Afr-Am), 24 European Americans (Eur-Am), and 24 Han Chinese (Han-Ch) (Hinds et al., 2005). Only class A Perlegen data were used. This consists of SNPs randomly identified by Perlegen Sciences and thus minimizes any ascertainment bias. HapMap phase II data were also obtained from populations of African, European, and Asian origins: 90 individuals of European ancestry living in Utah, USA (CEU), 90 individuals from the Yoruba in Ibadan, Nigeria (YRI), 45 individuals from Beijing, China (CHB), and 45 individuals from Tokyo, Japan (JPT). Data from Chinese and Japanese populations were combined to derive a single Asian population (ASN) with the frequencies of each SNP calculated from the combined population.

SNPs, allele frequencies, and reference human and ancestral alleles were downloaded from SPSmart (Amigo et al., 2008) which in turn aggregates data from other sources such as dbSNP (http://www.ncbi.nlm.nih.gov/projects/SNP/). A hypothetical homozygous ancestral population was constructed from the ancestral state of SNPs, which come from dbSNP (Amigo et al., 2008) and are primarily inferred from the chimpanzee genome. SNPs for which no ancestral allele was available were not considered, though this occurred for only ∼5% of SNPs. Only autosomal chromosomes were analyzed.

Overall, 1,175,330 SNPs from Perlegen class A were used in this work. Of these, 873,180 were found in HapMap phase II and were employed for the complementary analyses using the HapMap dataset.

Chromosomal locations and gene symbols were obtained from Ensembl (http://www.ensembl.org/).

Estimating Genetic Distances

Genetic distances were calculated using the average genetic distance (DA), as described (Nei, 1987; Hughes et al., 2008). There is a vast literature on methods to calculate genetic distances between populations, though previous results have suggested strong correlations between different measures of genetic distances in large-scale SNP analyses for human populations (Akey et al., 2002). DA was chosen because it is a simple and intuitive measure that can be easily applied with our hypothetical ancestral population, as described below.

For a given biallelic SNP locus the genetic distance (d ) is given by

image

where x1 and y1 are the frequencies of the first allele and x2 and y2 are the frequencies of the second allele in a modern population (x1 and x2) and in a hypothetical ancestral population (y1 and y2). Because allele frequencies in the hypothetical ancestral population are always assumed to be 1 for the ancestral allele and 0 for the derived allele, d is effectively given by

image

where x is the frequency of the ancestral allele in the modern population and thus d will have a value between 0 and 1.

Average d values of SNPs in a gene give the average genetic distance (DA). Because the distribution of SNPs can vary between regions of the genome and nearby SNPs will have stronger linkage disequilibrium (LD), considering each SNP as an independent observation would be incorrect. Therefore, DA was first determined for blocks of 5 kb and then the DA of the gene was calculated from the average DA of all its blocks. The 5 kb window size was chosen because it was deemed to have an appropriate balance between number of SNPs and strong LD. Although the criteria is admittedly arbritary, SNPs within 5 kb of each other are in strong LD (Shifman et al., 2003; Hinds et al., 2005) and blocks of 5 kb have on average ∼3 (HapMap), ∼3.5 (Perlegen) SNPs per block. That said, the results using a window size of 2.5 kb or 10 kb were very consistent with those obtained with 5 kb (not shown).

Because the goal of this project was to detect functions and processes associated with population differences, a gene-centric approach was employed in which DA was calculated for each gene using SNPs within the gene's coding region plus SNPs 50 kb from the coding region on both sides. SNPs within 50 kb of a gene will tend to be under moderate LD (Shifman et al., 2003; Hinds et al., 2005).

As in other similar studies (Voight et al., 2006; Pickrell et al., 2009), only common SNPs were considered as these are more likely to differentiate populations. This meant only SNPs with a minor allele frequency >5% in at least one of the three human populations were used (Frazer et al., 2007). SNPs in which the minor alleles were different between two of the populations were also kept in the analysis, however.

Statistical Tests

Genes that have significantly higher DA values than the average would be candidates for specifying population differences (Fig. 1). To identify such genes in each of the modern human populations, the cumulative Poisson distribution was employed. First, the average DA for all blocks in all genes between the ancestral population and Afr-Am, Eur-Am, and Han-Ch was calculated to be, respectively, 0.190, 0.230, and 0.236. For the HapMap dataset, the average DA between the ancestral population and YRI, CEU, and ASN was, respectively, 0.184, 0.226, and 0.230. Based on genome-wide averages, the cumulative Poisson distribution was then used to calculate for each gene the probability that it has a higher DA in a given modern–ancestral population pair than would be expected by chance, after correcting for the number of blocks (nblocks) in the gene:

image

where inline image which represents the DA of the gene in the population pair over the average DA for all genes (inline image) multiplied by nblocks; λ is similarly estimated from the average DA of the two other modern populations for the gene over the average for all genes multiplied by nblocks while inline image (i.e., when DA= 1).

Figure 1.

Schema of our method to detect population-specific genetic divergence. The genetic distance (DA) between a hypothetical ancestral population and modern human populations–in the case of this work of African, European, and Asian origins–is calculated from the genome-wide SNP data (A). DA is then calculated for individual genes. Genes for which DA is larger in African (B), European (C), or Asian (D) populations than expected by chance are detected as population-specific genetic divergence and are thus candidates for specifying phenotypic differences in that population.

In effect, the k metric is equivalent to the number of derived alleles that are present in a given modern human population in a given gene corrected for the average across the genome. The P-value derived from the Poisson function represents the probability that a given gene in a population has a higher representation of derived alleles than of ancestral alleles when compared to other populations. It does not allow, however, a direct rejection of a null hypothesis. Instead, it is a metric designed to rank genes and detect the most extreme outliers that are potentially involved in population-specific phenotypes. When calculating P-values for Eurasians, DA is simply the average between Eur-Am and Han-Ch, while λ is estimated based only on the Afr-Am population. Overall, we calculated DA values for 17,344 genes, which had at least three blocks, using the Perlegen data and 17,038 genes using HapMap.

To identify candidate pathways, processes, and functions in each population, Gene Ontology (GO) annotation, which describes how gene products behave in a cellular context (Ashburner et al., 2000), was employed. The cumulative Poisson distribution described earlier was employed to derive P-values for GO categories. For a given GO category, DA and nblocks were calculated from the sum of DA and nblocks from all genes associated with that category. In other words, each GO category can be viewed as the concatenated sequence of all genes associated with it, as in other studies (de Magalhaes & Church, 2007). A complementary functional enrichment analysis of genes at the P < 0.05 cutoff was performed using the Database for Annotation, Visualization and Integrated Discovery (DAVID, Huang da et al., 2009).

A false discovery rate (FDR) analysis was performed to estimate the number of both genes and GO categories expected to be significant by chance. Simulations–using the Fisher-Yates shuffle algorithm to randomly permute SNP allelic frequencies of each block–were performed to determine, by chance, how many genes and GO categories below a given P-value threshold one is expected to find. Based on these simulations, an FDR-adjusted P-value (q-value), defined as the number of expected false positives over the number of significant results (Storey & Tibshirani, 2003), was estimated for each gene and each GO category.

The algorithm was implemented in the Perl language. Statistical tests were performed in Microsoft Excel 2007 and in SPSS 17 (SPSS Inc., Chicago, IL, USA).

Results

Algorithm to Detect Population-Specific Genetic Differentiation

The aim of this work is to identify genes and processes that are candidates for explaining phenotypic variation between populations. To detect such genes, the underlying principle behind our method is that loci that contribute to population-specific traits will have genetic patterns that are more distinct from states found in ancestral populations. Therefore, we first calculated for all genes the average genetic distance (DA) between each modern human population studied in this work and a hypothetical ancestral population (Fig. 1). We first employed SNP data from Perlegen Sciences genotyped from Afr-Am, Eur-Am, and Han-Ch individuals. Since our goal was to identify functions and processes that differ between populations, only SNPs within 50 kb of coding regions were included in our calculations of DA. Ancestral alleles were used to derive a hypothetical homozygous ancestral population (see Materials and Methods).

Our conceptually simple method is based on detecting genes that have a larger DA in a given population than expected by chance (Fig. 1). Therefore, to calibrate our measurements, we determined the genome-wide genetic distances between the modern Perlegen populations and our hypothetical ancestral population. As expected, our results show that DA between the ancestral population and Afr-Am is shorter than for Eur-Am and Han-Ch, respectively 0.190, 0.230, and 0.236. The differences in DA between all the populations were highly significant (P < 0.001 using a Wilcoxon signed-rank test). These results are in line with many previous findings showing that the genetic distance between Eurasian populations and Africans is larger than the differences between Eurasian populations (Hughes et al., 2008), in accordance with the African origin of modern humans. DA was consistent between chromosomes (the relative standard deviation for the three populations was ∼2%), and thus genome-wide DA values were used as reference for detecting outliers.

Because the distribution of SNPs is not consistent across the genome, DA was calculated from the average of the genetic distance of individual SNPs in blocks of 5 kb. As further detailed in the Methods, the DA of a given gene was then calculated from the average DA of its 5 kb blocks. To detect candidate genes in each population, we employed the cumulative Poisson distribution to calculate P-values and identify genes in which DA is larger than expected by chance–after adjusting for the number of blocks in each gene. Expected DA values are estimated from the other population pairs to account for variation across the genome (see Materials and Methods). The P-value derived from the Poisson distribution, however, does not allow a direct rejection of a null hypothesis but rather provides a metric to rank candidate genes.

Contrary to many previous similar works (in particular methods based on haplotype structure), our method is not a test for positive selection. Phenotypically relevant differences between populations could be due to chance events like neutral drift and population bottlenecks that are not under selection. Our method thus aims to detect candidate loci for population-specific traits without any intrinsic assumption of their origin. As further discussed ahead, however, it is plausible that the most striking differences between populations found using our method were driven by selection and can thus be seen as candidate regions for recent positive selection.

Genes that Differ Most between Populations

The method described allowed us to detect genes in our three studied populations and in Eurasians (Eur-Am + Han-Ch) for which DA is disproportionally higher in a given population when compared to the others, after correcting for the averages across the whole genome. At P < 0.05, we detected 74 genes in Afr-Am, 98 in Eur-Am, 147 in Han-Ch, and 218 genes in Eurasians (see Tables S1–S4). Simulations were performed to estimate that, by chance, 34 genes were expected in Afr-Am, Eur-Am, and Han-Ch, and 52 in Eurasians. Although it is encouraging to detect more genes than expected by chance, it is important to note that population demographic effects are not considered in our statistical test. Besides, since our method is not a test for selection it does not attempt to exclude random events. In fact, genes relevant to population-specific differences could occur purely by chance. This is why, as indicated, our P-value is a metric designed to detect and rank outliers, which can be considered candidates for population-specific differences.

As expected, a large number of the top genes detected using our method had been reported as candidates of selection by previous studies. LCT, for example, had a larger genetic distance in Europeans than in other populations (Fig. 2). Further examples of candidates of selection in Europeans that were rediscovered using our method include genes associated with skin pigmentation like SLC45A2 and SLC24A5 (Graf et al., 2005; Lamason et al., 2005; Barreiro et al., 2008), RAB3GAP1 and BCAS3 (Sabeti et al., 2007), MYEF2 and ARHGAP26 (Johansson & Gyllensten, 2008), and ZMYM4 (Pickrell et al., 2009) (Tables 1 and S2). DARC was the only candidate gene under selection in Africans rediscovered using our method (Table 1). Many previous candidates of selection in Asian populations were rediscovered, however, including XKR6 (Deng et al., 2008), ABCC11, SYTL3, and RTTN (Barreiro et al., 2008), EDAR, HERC1, and SLC30A9 (Sabeti et al., 2007) (Tables 1 and S3). Interestingly, some genes previously identified as candidates in Europeans or Asians, such as PPARD (Voight et al., 2006), DOCK4 (Johansson & Gyllensten, 2008), TOP2B, and DNAH6 (Tang et al., 2007), were rediscovered in our study as significant in Eurasians (see Table S4). Among the top genes in Eurasians was also C21orf34 (P= 0.023; Table S4) which has been previously associated with selection in populations outside of Africa (Pickrell et al., 2009). As far as we could tell, there were no contradictory findings of genes being detected by our method as candidates in one population having been previously reported to be under selection in another population included in our study. Overall, these results demonstrate the power of our method to detect biologically relevant results.

Figure 2.

Genetic distance (DA) in the lactase locus between a hypothetical ancestral population and African American (Afr-Am, in light blue), European American (Eur-Am, in red), and Han Chinese populations (Han-Ch, in green). A rolling average is used across blocks in the chromosomal region. The location of the lactase gene (LCT) is represented in dark blue.

Table 1.  Selection of the most highly divergent genes in each population.
Gene1ChrStartEndDA Anc-Afr2DA Anc-Eur3DA Anc-Han4P-value
  1. Anc-Afr, Ancestral-African; Anc-Eur, Ancestral-European; Anc-Han, Ancestral-Chinese.

  2. 1Genes highlighted in bold have been previously shown to be under selection using other methods (see text for references).

  3. 2DA between ancestral and African American population.

  4. 3DA between ancestral and European American population.

  5. 4DA between ancestral and Han Chinese population.

Genes divergent in Eur-Am
 KCNH721628861631634532740.1950.3620.2570.000
 MYEF21546168921463078500.2070.4900.2080.000
 ZMYM4135457155357101310.1870.3400.1650.002
 SLC24A51546150461462718810.2230.4100.2010.003
 MYST41076205346765126450.1440.3040.2130.004
 ATP6V1H854740669549684030.2080.4350.2830.007
 PAWR1278459878786589210.1720.3550.2270.007
 BMP2K479866556801023630.2410.3790.2470.008
 LCT21362118851363612200.1180.2670.1510.008
Genes divergent in Han-Ch
 EXOC6B272206621729566850.1810.2720.4120.000
 POLR3B121052256191054781040.1880.2120.4300.000
 C6orf17361266529461267614470.1430.1810.6420.000
 HERC11561637871619632000.1930.2200.3990.000
 THADA243261479437266890.2120.2790.3830.001
 RTTN1865772025660739420.1840.3190.4100.002
 XKR6810741075111462580.1660.2000.3130.003
Genes divergent in Afr-Am
 TRIM10630177705302866900.4080.2440.2660.003
 SYT1412081281612084542590.1540.1080.1090.008
 CADM311573580401574895560.1990.1360.1190.010
 SLC35D2998072834982357970.3960.2820.2210.015
 DARC11573897211574929140.2090.1260.1410.017
Genes divergent in Eurasians (but not highly divergent in Eur-Am or Han-Ch)
 CSPP1868089157683210480.2740.6640.7790.001
 RB11347725884480040270.1320.3190.2640.001
 KCNT211944115361948941220.1680.3010.3240.001
 EPHA6397966311990002350.1780.3190.2990.001
 ZEB11031598148319067400.1710.3550.3420.001

Although it is encouraging to verify that many of the genes detected by our method had been suggested as candidates of selection previously, these results are merely confirmatory. One major interest of this analysis is the identification of new genes and processes that may help explain population differences. Interestingly, several genes were identified by our method as candidates for explaining population differences for the first time (to our knowledge). These new candidate genes for specifying population-specific traits are described below.

Top genes in Europeans included MYST4, a histone acetyltransferase involved in gametogenesis (McGraw et al., 2007), a potassium voltage-gated channel (KCNH7) as well as poorly studied genes like ATP6V1H and BMP2K (Table 1). In Asians, the top genes were EXOC6B, an exocyst complex component, POLR3B, C6orf173, and THADA, a gene associated with thyroid adenoma (Rippe et al., 2003). In Afr-Am, we detected fewer genes and even top genes had higher P-values (Tables 1 and S1). A few genes in the TRIM cluster on chromosome 6 were detected, close to the major histocompatibility complex. Other genes detected in Afr-Am included CADM3, SLC35D2, and SYT14. Finally, a number of candidate genes not significant among Eur-Am or Han-Chn were detected among Eurasians (Table 1). Among the top genes were retinoblastoma 1 (RB1), CSPP1, and KCNT2 which encodes a potassium channel.

A selection of the most significant genes for each population is shown in Table 1. Tables with significant genes, at the P < 0.05 threshold, for each population are available in Tables S1–S4 with the full dataset available online (http://genomics.senescence.info/evolution/human_populations.html).

If DA values were to follow a neutral model then a similar distribution would be expected for genic (defined as SNPs within 50 kb of a coding region) and nongenic regions (defined as all SNPs not in genic regions), as argued by others (Voight et al., 2006; Barreiro et al., 2008). To test this assumption, we determined extreme DA values between modern populations by calculating the difference between DA values in two populations (e.g., DA in Afr-Am –DA in Eur-Am) for all blocks in genic and nongenic regions. Most blocks had a DA difference between population pairs in the range of 0.01–0.15 with <1% of extreme blocks with a difference above 0.5. Interestingly, however, the proportion of extreme blocks was considerably higher in genic than in nongenic regions (Fig. 3). These results show that far more extreme DA values are observed in genic regions, suggesting that the patterns observed are not random and that at least some of the results observed in terms of differentiated genes between populations are due to selection.

Figure 3.

Comparison between the distribution of differences in genetic distance (DA) between pairs of the studied modern populations in genic versus nongenic regions. The percentage of blocks with a DA difference between two populations in genic regions is divided by the percentage in nongenic regions and plotted according to the DA difference. Pairwise comparisons are shown between African American and European American (Eur-Afr, in red), Han Chinese and European American (Chn-Eur, in green), and Han Chinese populations and African American (Chn-Afr, in purple).

Candidate Processes and Functions Related to Population-Specific Phenotypes

In addition to identifying new candidate genes, the goal of this work is to identify processes, functions, and mechanisms that may be related to population-specific phenotypes. Therefore, we performed an analysis using GO annotation (Ashburner et al., 2000). As detailed in the Methods, DA values were calculated from the concatenated sequences of genes associated with each GO category. Because this analysis looks for deviations from general patterns by combining values from different genes often located in different regions, significant GO categories are more likely explained by selection acting on specific processes than patterns in individual genes. Simulations were performed to estimate FDRs and P-values adjusted to what would be expected by chance (q-values, see Materials and Methods).

At P < 0.05, 21 categories were significant in Afr-Am, 14 in Eur-Am, and 32 in Han-Ch. A selection of significant categories that distanced themselves in each population are indicated in Tables 2–5 with all significant results available in Tables S5–S8. Similarly to the results observed for individual genes, several categories are related to those previously reported before as candidates for selection in human populations (Akey et al., 2002; Voight et al., 2006; Williamson et al., 2007; Barreiro et al., 2008; Pickrell et al., 2009). These include categories related to senses, like taste (GO:0008527), which is among the top categories in Afr-Am, or audition (GO:0042491) and olfaction (GO:0007608 and GO:0004984) in Han-Ch, as well as categories related to immune response (e.g., GO:0042613 in Afr-Am and GO:0009405 in Han-Ch). Categories possibly related to dietary adaptations, like unsaturated fatty acid biosynthetic process (GO:0006636 in Afr-Am), were also among top categories.

Table 2.  Selected top GO categories in Afr-Am (q < 0.1 and n≥ 3).
GO IDNameGO type1n GenesDA Anc-Afr2DA Anc-Eur3DA Anc-Han4q-Value
  1. Anc-Afr, Ancestral-African; Anc-Eur, Ancestral-European; Anc-Han, Ancestral-Chinese.

  2. 1C: Cellular component; F: Molecular function; P: Biological process.

  3. 2DA between ancestral and African American population.

  4. 3DA between ancestral and European American population.

  5. 4DA between ancestral and Han Chinese population.

GO:0005882Intermediate filamentC980.2090.2410.239<0.01 
GO:0006636Unsaturated fatty acid biosynthetic processP40.2610.2610.217<0.01 
GO:0005922Connexon complexC170.1850.1950.1970.025
GO:0008527Taste receptor activityF100.2490.2690.2440.047
GO:0016458Gene silencingP30.2280.2040.2150.068
GO:0042613MHC class II protein complexC130.1960.2010.2090.060
GO:0019902Phosphatase bindingF50.1970.2160.2200.062
GO:0006816Calcium ion transportP870.1930.2270.2290.061
GO:0019717SynaptosomeC340.1930.2240.2240.058
GO:0030539Male genitalia developmentP80.2290.2400.2250.083
GO:0007635Chemosensory behaviorP60.2220.2320.2150.088
GO:0050909Sensory perception of tasteP210.2220.2460.2390.082
Table 3.  Selected top GO categories in Eur-Am (q < 0.1 and n≥ 3).
GO IDNameGO type1n GenesDA Anc-Afr2DA Anc-Eur3DA Anc-Han4q-Value
  1. Anc-Afr, Ancestral-African; Anc-Eur, Ancestral-European; Anc-Han, Ancestral-Chinese.

  2. 1C: Cellular component; F: Molecular function; P: Biological process.

  3. 2DA between ancestral and African American population.

  4. 3DA between ancestral and European American population.

  5. 4DA between ancestral and Han Chinese population.

GO:0016568Chromatin modificationP980.1780.2360.231<0.01
GO:0007276Gamete generationP60.1860.3150.261<0.01
GO:0016573Histone acetylationP70.1810.2660.223<0.01
GO:0016407Acetyltransferase activityF90.1800.2590.216<0.01
GO:0008135Translation factor activity, nucleic acid bindingF40.2190.3650.297<0.01
GO:0004707MAP kinase activityF120.1780.2430.215<0.01
GO:0043353Enucleate erythrocyte differentiationP40.1650.2790.2490.011
GO:0045651Positive regulation of macrophage differentiationP40.1500.2750.2460.012
GO:0007518Myoblast cell fate determinationP30.2320.3440.2760.090
Table 4.  Selected top GO categories in Han-Ch (q < 0.1 and n≥ 3).
GO IDNameGO type1n GenesDA Anc-Afr2DA Anc-Eur3DA Anc-Han4q-value
  1. Anc-Afr, Ancestral-African; Anc-Eur, Ancestral-European; Anc-Han, Ancestral-Chinese.

  2. 1C: Cellular component; F: Molecular function; P: Biological process.

  3. 2DA between ancestral and African American population.

  4. 3DA between ancestral and European American population.

  5. 4DA between ancestral and Han Chinese population.

GO:0000145ExocystC50.1900.2470.314<0.01
GO:0006904Vesicle docking during exocytosisP170.1880.2310.269<0.01
GO:0048011Nerve growth factor receptor signaling pathwayP50.2020.2560.328<0.01
GO:0042491Auditory receptor cell differentiationP60.1970.2460.303<0.01
GO:0003682Chromatin bindingF690.1890.2350.254<0.01
GO:0007608Sensory perception of smellP3030.1940.2230.2420.023
GO:0004842Ubiquitin-protein ligase activityF1120.1860.2340.2470.023
GO:0008017Microtubule bindingF360.2060.2500.2760.023
GO:0004984Olfactory receptor activityF2960.1950.2220.2420.024
GO:0006512Ubiquitin cycleP3370.1820.2250.2350.025
GO:0007131Meiotic recombinationP180.1820.2230.2510.025
GO:0009405PathogenesisP60.2200.2970.3480.027
GO:0048699Generation of neuronsP40.2270.2540.3160.048
GO:0030246Carbohydrate bindingF100.1700.2120.2490.051
GO:0007610BehaviorP150.1890.2510.2730.063
GO:0000724Double-strand break repair via homologous recombinationP110.1780.2330.2620.065
GO:0001942hair follicle developmentP30.1910.2260.2920.080
GO:0005947Mitochondrial alpha-ketoglutarate dehydrogenase complexC40.1740.2400.2800.083
Table 5.  Selected top GO categories in Eurasians (q < 0.05 and n≥ 3).
GO IDNameGO type1n GenesDA Anc-Afr2DA Anc-Eur3DA Anc-Han4q-Value
  1. Anc-Afr, Ancestral-African; Anc-Eur, Ancestral-European; Anc-Han, Ancestral-Chinese.

  2. 1C: Cellular component; F: Molecular function; P: Biological process.

  3. 2DA between ancestral and African American population.

  4. 3DA between ancestral and European American population.

  5. 4DA between ancestral and Han Chinese population.

GO:0000145ExocystC50.19010.24650.3138<0.01
GO:0007243Protein kinase cascadeP640.17540.22740.2354<0.01
GO:0045651Positive regulation of macrophage differentiationP40.15010.27470.2459<0.01
GO:0015908Fatty acid transportP40.19310.3250.3358<0.01
GO:0016568Chromatin modificationP980.17820.23570.2308<0.01
GO:0043550Regulation of lipid kinase activityP30.15570.27410.2531<0.01
GO:0004842Ubiquitin-protein ligase activityF1120.18630.23430.2468<0.01
GO:0006091Generation of precursor metabolites and energyP690.19920.25740.2682<0.01
GO:0043353Enucleate erythrocyte differentiationP40.1650.27870.2487<0.01
GO:0006512Ubiquitin cycleP3370.1820.22520.2349<0.01
GO:0030693Caspase activityF150.20220.2850.2868<0.01
GO:0030168Platelet activationP160.18670.25310.2624<0.01
GO:0048185Activin bindingF30.15650.25320.2577<0.01
GO:0004128Cytochrome-b5 reductase activityF50.16650.25490.2647<0.01
GO:0001515Opioid peptide activityF50.17740.2790.2708<0.01
GO:0015871Choline transportP30.210.30270.3442<0.01

Among Afr-Am (Table 2), it is interesting to note male genitalia development (GO:0030539) as a candidate GO category. Another novel (to our knowledge) category was UV protection (GO:0009650), even though it was borderline significant with q= 0.11 (and P= 0.049; Table S5).

In Eur-Am, the top categories were related to chromatin and histone acetylation (GO:0016568 and GO:0016573), in line with previous results (Voight et al., 2006). Another top category was gamete generation (GO:0007276), though again selection in genes related to reproduction has been found in similar such studies (Akey et al., 2002; Voight et al., 2006). In this context, mating behavior (GO:0007617) also caught our attention (P= 0.040 and q= 0.11; Tables S6).

Categories related to exocytosis were the top categories in Han-Ch (GO:0000145 and GO:0006904), which is consistent with the genic analysis given that an exocyst component was the top gene in this population. Another top category was meiotic recombination (GO:0007131). Of interest was also the mitochondrial alpha-ketoglutarate dehydrogenase complex (GO:0005947) involved in oxidative stress and neurodegeneration (Gibson et al., 2005). As detailed in Table 4, other top categories included ubiquitin-protein ligase activity (GO:0004842) and behavior (GO:0007610).

In Eurasians, 56 categories passed P < 0.05. Top categories included fatty acid transport (GO:0015908) and caspase activity (GO:0030693), thus mostly recapitulating the categories for other populations and previous studies (Table 5).

Strikingly, categories related to nervous system development were top candidates in all three populations: synaptosome (GO:0019717) in Afr-Am, nerve growth factor receptor signaling pathway (GO:0048011), and generation of neurons (GO:0048699) in Han-Ch, and though slightly below our statistical threshold (q= 0.12) dendritic spine (GO:0043197) in Eur-Am.

We also performed a functional enrichment analysis among significant genes (P < 0.05) for each population using DAVID (Huang da et al., 2009). Although some functions and processes were found enriched, such as exocytosis in Han-Ch, the only significant category after FDR correction was zinc finger region among Afr-Am (Tables S17–S20). These results show that our GO enrichment methodology, that in essence gives a different weight to each gene contributing to the GO categories, outperforms more traditional approaches.

Comparison with Patterns from HapMap Data

Although encouraged by our findings, we were concerned about the population admixture in the African American group and whether this could have been a cause for the lower number of significant genes in this group. Besides, the particular history of African Americans (e.g., slavery) could have shaped the patterns observed in our results. Therefore, we repeated our analysis using HapMap phase II data, which like Perlegen includes populations from African (YRI), European (CEU), and Asian (ASN) origins but with the African population originating in Nigeria (Frazer et al., 2007). For this analysis we only considered HapMap SNPs in class A Perlegen data to minimize any ascertainment bias, thus the total number of SNPs was lower than when using the Perlegen data. Data for Chinese and Japanese populations was merged into a single Asian population (see Materials and Methods).

The results from HapMap were largely in accordance with the results obtained using the Perlegen dataset. Briefly, the DA between the ancestral population and YRI was again shorter than for CEU and ASN, respectively 0.184, 0.226, and 0.230. We found slightly fewer statistically significant genes, however, for European and Asian populations than using the Perlegen dataset, respectively, 84 and 132 at P < 0.05 (Tables S10 and S11), and more were expected by chance (65.6) than in the Perlegen analysis. Besides, the top genes for CEU and ASN were among the significant genes already identified in the Perlegen data for Eur-Am and Han-Ch. As for GO categories (Tables S14 and S15), the top categories largely recapitulated the results for Eur-Am and Han-Ch. Therefore, while it is encouraging that the results obtained from the Perlegen populations were corroborated in CEU and ASN, the analysis of HapMap did not substantially add to our previous findings. Our results for all genes, however, are available in the Supporting Information.

For the African population the results obtained using HapMap data were better than those using Perlegen. At P < 0.05, 170 genes were detected in YRI when 65.6 would be expected by chance. Even at P < 0.01, 29 genes were identified (2.7 would be expected by chance) which is a considerable improvement over the results obtained using Perlegen data. Of these 29 genes, 11 had been identified in the Perlegen analysis. Perhaps the admixture in the Perlegen Afr-Am population lowers significance levels and the number of significant genes. The top gene in YRI was KIAA1267, a poorly studied gene that is in the same region of the microtubule-associated protein tau (MAPT), also among our candidates. One gene re-identified was RSBN1 since it was previously shown to have a signal for selection in YRI (Voight et al., 2006). Another gene that caught our attention was STRBP, the spermatid perinuclear RNA binding protein. This gene was also detected in the Perlegen analysis but with only P= 0.042 when in YRI the signal is much stronger (P= 0.001). Other genes detected with strong evidence for higher divergence in YRI include CADM3, ANKRD17, COX18, GJB2, and POU1F1 (Tables 6 and S9).

Table 6.  Selection of the most highly divergent genes in the African (Yoruba) population.
Gene1ChrStartEndDA Anc-YRI2DA Anc-CEU3DA Anc-ASN4P-value
  1. Anc-YRI, Ancestral-African; Anc-CEU, Ancestral-European; Anc-ASN, Ancestral-Asian.

  2. 1Genes highlighted in bold have been previously shown to be under selection using other methods (see text for references).

  3. 2DA between ancestral and African (Yoruba) population.

  4. 3DA between ancestral and European population.

  5. 4DA between ancestral and Asian population.

KIAA12671741413181416759430.4960.3450.5420.001
CADM311573580401574895560.2270.1290.1040.001
STRBP91248771391251207180.2920.1660.2190.001
SCMH1141215461415303750.3320.2280.2450.001
C5orf34543472567436009440.3300.1440.0290.001
ANKRD17474109369743933660.1620.0690.0590.002
BTRC101030538151033570600.3430.2100.3350.002
DARC11573897211574929140.2410.1190.1220.002
SYT1412081281612084542590.1930.1260.1080.002
CEP571195113290952555020.2150.1390.1380.003
COX18474089280742043360.1830.0470.0320.003
MAPT1741277624415115470.4710.3410.5220.004
C9orf5691250173031251190750.2970.1140.1520.004
GJB21319609605197151140.2890.1390.1520.005
RSBN111140559771142065930.2890.2100.1710.006
POU1F1387341473874584270.3040.2200.2350.010

A larger number of GO categories (64 when 3.3 would be expected by chance) were also detected in YRI (Table S13). The results were largely consistent, however, with the Perlegen analysis with by and large the same type of categories being detected. Categories that caught our attention in the Perlegen analysis were re-identified such as UV protection (GO:0009650) and male genitalia development (GO:0030539) as well as categories related to senses, immune responses, dietary adaptations, and nervous system development (Table S13). In the latter category type, axon (GO:0030424) and neuron differentiation (GO:0030182) were identified, which had not been found in the Perlegen analysis. Other categories in YRI not identified before that might be of interest were hair follicle morphogenesis (GO:0031069), artery morphogenesis (GO:0048844), and blood vessel remodeling (GO:0001974).

Full lists with significant GO categories at the P < 0.05 threshold for all HapMap populations are available in Tables S13–S16.

Discussion

There is a vast literature on methods to detect signatures of selection in human populations (Akey et al., 2002; Weir et al., 2005; Voight et al., 2006; Sabeti et al., 2007; Tang et al., 2007; Williamson et al., 2007; Myles et al., 2008; Coop et al., 2009Pickrell et al., 2009). Our method is not necessarily more powerful than other methods, such as those employing Fst or haplotype structure, to detect selection. The main novelty of our study is that we scan for genetic differences based not on patterns of selection but rather on the genetic distance between each population and an inferred ancestral population. Because of this broader focus on population differences our study provides a different approach. Although many of our results are confirmatory, our work also reveals new candidate regions. In fact, the large number of genes and GO categories that overlap with previous findings–including those few genes for which considerable evidence exists of a phenotypic role in population-specific traits like LCT, DARC, SLC45A2, and SLC24A5–demonstrate that our method can detect genes that are phenotypically relevant.

The shorter genetic distance between the African populations (Afr-Am and YRI) and the hypothetical ancestral population is consistent with previous analyses of Alu insertions suggesting that the ancestral state is more common in African populations (Watkins et al., 2001; Watkins et al., 2003). Our analysis showing the preservation of human ancestral alleles in an African population, however, is based on a much larger number of polymorphisms. It is consistent with the Out-of-Africa hypothesis since demographic effects like bottlenecks and adaptations to new environments likely drove genetic differentiation in Eurasians.

It is important to note that our method of calculating P-values is model-free and does not account for population demographic effects such as bottlenecks during the migration of populations out of Africa and into Eurasia. Considering the several known targets of selection detected using our method, it is plausible that many genes detected using our method could be driven by selection–and hence should be considered as candidates in future studies–but it is possible that genes are highly divergent in a given population due to demographic effects. Nonetheless, even for alleles that became more prevalent in a given population due to chance, it is possible that these are phenotypically relevant. Ample evidence suggests that founder effects contributed to population differences (Ramachandran et al., 2005). For example, population bottlenecks have been suggested to have an effect on human phenotypic variation based on analyses of cranial traits (Manica et al., 2007). Its focus on genetic differences that could have been driven by selection and by demographic effects is one of the distinct aspects of our method when compared to genome-wide scans for selection and is the reason why our algorithm is not intrinsically designed to detect selection.

One problem that may affect some of our candidate genes, as it has for similar studies, is genetic hitchhiking (Smith & Haigh, 1974). For example, among our top ranked genes in Eur-Am, RAB3GAP1 is located within less than 150 kb of LCT. LCT has a lower P-value and it is possible that–even though RAB3GAP1 has been identified as a candidate gene under selection by another study (Sabeti et al., 2007)–RAB3GAP1 has a strong signal due to its proximity to LCT. Another caveat of our study is that our ancestral population is artificial, since an ancestral homozygous population never actually existed, and this might contribute to increase the power of our method but also to generate false positives. Besides, we only employed data from three populations, which may not be fully representative of their geographical origins. In particular for genes and GO categories with the lowest signals, our results may not apply to all human populations in Africa, Europe, or Asia. Nonetheless, the consistent results obtained by applying our method to both Perlegen and HapMap datasets, which were derived from different sets of populations, increase our confidence that our algorithm can provide new insights on the migration of human populations from Africa, one of the most important evolutionary events of our species. With the recent explosion in large-scale SNP data for human populations, in particular driven by advances in DNA sequencing (de Magalhaes et al., 2010), our method could be useful to a broad range of studies and ours can be seen as only a proof-of-concept study. Given our method's balance of simplicity and precision it may be useful for many other studies, including those in nonhuman species.

Despite the overlap with known results, it is noteworthy that our algorithm allowed us to detect several new genes and processes as candidates for explaining population differences. These could be targets for further study to better understand human evolutionary history and adaptation and the many new candidate genes identified using our method could lead to new insights. In particular, there were several genes in each of the populations with strong signals (Tables 1 and 6). These include KIAA1267, CADM3, STRBP, SCMH1, MAPT, and SYT14 in Africans and MYST4, KCNH7, and BMP2K in Europeans. KIAA1267 and MAPT are in the same region and have both been associated with neurodegenerative diseases (Tobin et al., 2008), though whether this is related to our results is unclear. STRBP could be another case of a gene related to reproduction under selection since evidence from mice suggests this gene is involved in spermatid RNA metabolism (Schumacher et al., 1995). Although not previously suggested to be under selection in human populations, the gene encoding the poorly studied bone morphogenic protein BMP2K has been previously suggested to be under selection in hominids (Arbiza et al., 2006). Finally, it is curious to note that MYST4 and NCOA2, also in our top results (Table S2), have both been known to form fusion proteins in acute myeloid leukemia (Murati et al., 2004), though whether this is related to their strong signals in our study is unknown.

Taken together, our results demonstrate that our method is statistically sound. Even though we did not simulate demographic effects because our goal is not to detect selection but rather phenotypic changes, we used an empirical distribution to demonstrate that the patterns observed are not random. Under neutrality, the percentage of blocks at varying levels of differentiation between population pairs would be expected to be the same for genic and nongenic regions since these are under the same demographic effects. Since selection preferentially targets genic regions (Voight et al., 2006; Barreiro et al., 2008), our results showing a higher percentage of extreme DA values in genic regions suggest that selection played a role in our observed results (Fig. 3).

Even if the results from our study are statistically sound, it is intrinsically difficult to demonstrate that even the genes with the strongest signals are phenotypically relevant. Nonetheless, our work provides important leads for further research avenues. Multiple candidate genes emerged in Han-Ch (Tables 1 and S3). Examples include EXOC6B, POLR3B, C6orf173, and THADA. The strong signal in GO categories related to exocytosis together with the fact that a component of the exocyst (EXOC6B) was the gene with the lowest P-value makes a strong case for selection in this protein complex. We speculate that the role of exocytosis in the nervous system–and neuronal development had strong signals in Han-Ch–may be responsible for these observations.

While the phenotypic role of individual genes in population differences may be difficult to assert, by looking at molecular pathways and processes we were able to detect many that may be related to population-specific traits and were perhaps under selection in specific populations. In fact, divergent GO categories are more likely to be caused by selection since they encompass concatenated sequences from multiple genes, usually in different genomic regions. Our results from GO categories are thus statistically robust and in line with results using other methods. They confirm previous findings suggesting selection in processes related to senses like taste, smell, and audition as well as immune response. Many of the categories we found related to lipid and fatty acid metabolism are likely the result of adaptations to new food sources, as reported by others (Voight et al., 2006). Therefore, several of our results are in accordance with previous findings indicating patterns of selection consistent with adaptation to new environments such as climates, pathogens, and sources of food.

Although our finding of GO categories related to reproduction as potential targets of selection has been reported before (Voight et al., 2006), we were intrigued to find evidence of higher divergence in genes related to male genitalia in Afr-Am. Genes related to male genitalia were also among our enriched GO categories in the HapMap African population, showing that these results are not due to the evolutionary history of African Americans. Because African Americans have been shown to have higher levels of testosterone than Caucasians (Ross et al., 1986; Winters et al., 2001), we speculate that these results could be phenotypically relevant.

We found evidence of divergence in multiple categories related to neuronal development in all three populations, in both Perlegen and HapMap analyses, yet more strongly in Han-Ch. Williamson et al. (2007) previously identified genes involved in nervous system development as candidates of recent selective sweeps, yet our results expand on these previous findings and provide additional genes and categories with evidence of being involved in population differentiation. These results are also in accordance with human evolutionary history. As populations migrated to new environments, adaptations in terms of behavior driven by climate, landscape, fauna, or flora might have resulted in adaptations at the level of the nervous system. These results may also be explained by cultural processes that could have played a role in shaping the human genome (Laland et al., 2010). Indeed, behavior was a significant GO category in Han-Ch which as far as we know is novel. Given the nature of our analysis, however, demographic effects alone could also explain these results.

Several other GO categories identified by us as candidates may provide clues about human evolution, adaptation, and genetics. To our knowledge, UV protection has not been associated with selection in African populations and, even though it may only be relevant to specific geographic regions, we found such evidence for both Afr-Am and YRI populations. We speculate that slightly deleterious derived alleles affecting UV protection drifted to higher frequencies in African populations given the protection from the sun offered by dark skin pigmentation.

In conclusion, we present a conceptually simple algorithm to detect genes and processes that most diverged in populations and apply it to data from three human populations from three continents. This work confirms and extends previous findings on the African origins of the human species, providing new candidate genes for specifying population-specific traits. Some of these genes may have been targets of selection while others may have been differentiated by chance. Our work also reveals putative processes under selection in each of the populations. Of particular note, our results showing evidence of differentiation in neuronal genes support the hypothesis that adaptations to new environments and/or cultural processes could have driven adaptations at the level of the nervous system. Besides, our results suggesting divergence in genes related to the male reproductive system in Africans are to our knowledge novel. These findings may help understand the genetics and evolution of human populations as well as the phenotypic and biological differences between them. Our method and results are therefore complementary and synergistic to those previously published and are one more step in understanding genetic differences between populations that could be important for human evolution and adaptation as well as medically relevant. A combination of methods will be necessary to understand the genetic basis of population phenotypic differences and our method is one more approach available to researchers.

Acknowledgements

The work of AM was supported by a fellowship from the Fundação Luso-Americana. JPM thanks the BBSRC (BB/H008497/1), the Ellison Medical Foundation, and a Marie Curie International Reintegration Grant within EC-FP7 for supporting work in his lab. Further thanks to Chris Tyler-Smith for useful suggestions; Steve Paterson, Harry Noyes, and Austin Hughes for comments on previous drafts; Jorge Amigo Lechuga for valuable assistance with SPSmart; and Domingos Magalhães for advice on the statistical analyses.

Ancillary