Identifying Selected Regions from Heterozygosity and Divergence Using a Light-Coverage Genomic Dataset from Two Human Populations Authors: Taras K. Oleksyk, Kai Zhao, Francisco M. De La Vega, Dennis A. Gilbert, Stephen J. O'Brien, Michael W. Smith Published: 2008, PLOS ONE DOI: 10.1371/journal.pone.0001712 ABSTRACT When a selective sweep occurs in the chromosomal region around a target gene in two populations that have recently separated, it produces three dramatic genomic consequences: 1) decreased multi-locus heterozygosity in the region; 2) elevated or diminished genetic divergence (F ST ) of multiple polymorphic variants adjacent to the selected locus between the divergent populations, due to the alternative fixation of alleles; and 3) a consequent regional increase in the variance of F ST (S 2 F ST ) for the same clustered variants, due to the increased alternative fixation of alleles in the loci surrounding the selection target. In the first part of our study, to search for potential targets of directional selection, we developed and validated a resampling-based computational approach; we then scanned an array of 31 different-sized moving windows of SNP variants (5–65 SNPs) across the human genome in a set of European and African American population samples with 183,997 SNP loci after correcting for the recombination rate variation. The analysis revealed 180 regions of recent selection with very strong evidence in either population or both. In the second part of our study, we compared the newly discovered putative regions to those sites previously postulated in the literature, using methods based on inspecting patterns of linkage disequilibrium, population divergence and other methodologies. The newly found regions were cross-validated with those found in nine other studies that have searched for selection signals. Our study was replicated especially well in those regions confirmed by three or more studies. These validated regions were independently verified, using a combination of different methods and different databases in other studies, and should include fewer false positives. The main strength of our analysis method compared to others is that it does not require dense genotyping and therefore can be used with data from population-based genome SNP scans from smaller studies of humans or other species. INTRODUCTION Patterns of allelic variation in the genome are shaped by successes and failures of genes influenced by evolutionary forces acting throughout population history. When a genetic variant becomes adaptive, populations experience changes in allele frequencies that reflect the strength and recurrence of the selective pressure(s). By identifying these residual footprints of genomic evolutionary processes in the form of “signatures of selection,” we hope to gain valuable insight into the evolutionary past of a species. A principal selection signature involves the local reduction in variation within the selected gene, as well as in adjacent SNP variants, around the selected chromosomal region known as “selective seep” [1] . Further, when two isolated populations are examined, one of which underwent strong selection in the past but the other did not, the frequencies of the selected SNP and adjacent alleles will often be more different between the populations than expected under the assumption of neutral genetic drift [2] – [4] . In addition, selection affects chromosomal segments, not simply individual SNPs, thus creating complex patterns of allele frequencies in the regions immediately surrounding the targeted site. In this study, we explore patterns of reduced heterozygosity and elevated between-population allele differentiation to identify strong selection signatures in the human genome. The method can also be applied to other diploid species when light-coverage SNP allele frequency genome scans of similar magnitude become available. To address these aspects as well as to explore a new approach, we designed and tested a strategy for revealing footprints of recent selection first by simulation and then based upon a sample of 183,993 SNP markers genotyped in 45 European Americans and 45 African Americans ( www.allsnp.com ). This dataset was chosen since it is independent and has never been used to estimate selection signatures. It also represents a modest database size that is smaller than the sizes of current whole-genome genotyping human population studies based on existing genotyping technologies [5] . We used minimum information that would likely be available from such databases and searched for selective signatures by analysing three parameters observed for each SNP: 1) heterozygosity in European Americans (H EA ); 2) heterozygosity in African Americans (H AA ); and 3) F ST between the corresponding individual SNPs in the two populations (F ST ). Centered on each available SNP, 31 arrays (or windows) including 5–65 SNPs were sampled along all human chromosomes except Y to evaluate each window for: 1) average SNP heterozygosity in European Americans (H` EA ); 2) average SNP heterozygosity in African Americans (H` AA ); and 3) variance of F ST among the adjacent SNPs (S 2 F ST ). We used S 2 F ST instead of the F ST mean estimator from each group of adjacent SNPs, since a measure of F ST mean across an array of loci would be more sensitive to those alleles that reached fixation in the form of the opposite allele, and less sensitive for those fixed in the same direction while variance captures this alternation. Several previous analyses of selective signatures have appeared which have been based either on decreased heterozygosity (H), population differentiation (F ST ), extended linkage disequilibrium, and even the premise that certain modern hereditary disease alleles were adaptive sometime in the past [3] , [4] , [6] – [15] . Discovered selection candidate regions included genes involved in development, immune defenses, reproduction, nutrition, behavior and other functions [16] . Although several of these regions have been discovered with multiple approaches (e.g., CCR5 , FY, LCT, G6PD, FOXP2 and others; Table S1 , Notes S1 ), other provocative regions have not, raising issues around the context of different algorithms and approaches, the strength and mode of selection, the timing of imputed selective events, the influence of study design, and the validity of unreplicated regions. To test the validity of our approach, the results of our scan was applied to previously nominated regions to explore how well this method validated previous discoveries. Finally, we incorporated nine other genome-wide or chromosome-wide attempts to find signatures of selection that included whole-genome searches for signatures of selection either by searching for the high values of local genomic divergence alone [4] or in combination with the allelic frequency spectrum [17] , [18] , looking for gene neighborhoods exhibiting extended linkage disequilibrium alone [9] , [19] , [20] or in combination with local genomic divergence [21] , or by examining an aberrant frequency spectrum [22] , [23] . We assessed the ten studies, including our own, and evaluated our findings in a multiple study comparison. RESULTS AND DISCUSSION The resampling-based approach compared values of population-based statistics such as heterozygosity and population divergence in a number of adjacent SNPs, to the distribution of values obtained by the resampling of the same number of SNPs from random locations along each chromosome. The strategy for selection detection involved the following six steps (see also Figure 1 and Materials and Methods ): 1) regional levels of heterozygosity (H`) and population divergence (S 2 F ST ) were estimated by choosing adjacent SNPs at the beginning of a single chromosome and computing H` EA , H` AA and S 2 F ST for each SNP in the two populations. 2) The SNP window was moved one SNP to the right and the same parameters were computed, and then again for the next group of five SNPs, until the results for 10,864 groups (for chromosome 1, see Table S2 ) were determined. 3) The same computation but using windows of 7, 9, 11 … 65 SNPs (i.e., a total of 31 SNP window sizes from 5 to 65 SNPs in length), was assessed across each chromosome. A baseline distribution to which these estimates are compared, was developed by choosing five (then 7, 9, 11 ... 65) random SNPs sampled 100,000 times across each chromosome. We analyzed odd numbers to center the sampling window on a SNP. The upper window size was dictated by limits in computational capabilities. 4) A distribution of empirical values was obtained and fractionally ranked relative to the randomly sampled expectations ( Figure 1 ). This process was then repeated 100 times to reach a total 10 million values. By combining these values, a chromosome-wide distribution of the mean fractional ranks (014 people in both populations, and heterozygosity [H]>0 in at least one population). The distribution of heterozygosity was skewed towards the larger values in both populations as expected because SNPs were intentionally selected for genotyping with a minor allele frequency (MAF) >0.05 or other strong evidence of their polymorphic status. Most loci were highly polymorphic; 68.53% of loci in African Americans and 63.79% of loci in European Americans had H>45%, while only in 6.83% of loci in African Americans and 8.55% of all loci the levels of heterozygosity were below 5%. At the same time, overall average heterozygosity is higher in European Americans (0.34) than in African Americans (0.33). Chromosome X shows the highest levels of F ST (0.201) and is almost twice as differentiated as the autosomes (0.125). There was a significant difference in the distribution of H AA , H EA , or F ST among the chromosomes overall, as well as among the autosomes in an analysis of variance (p<0.0001, Proc GLM, SAS 9.1). While some of this significance can be attributed to the known ascertainment bias, the difference remains significant when accounting for the density of the SNPs assayed on different chromosomes (p<0.0001). GENOME SCAN FOR SELECTION, ANALYTIC METHODS A database of allelic frequencies, number of samples genotyped, and the physical positions for each of 183,997 SNPs from European (n = 45) and African Americans (n = 45) developed by Applied Biosystems was evaluated. All individuals were evaluated on a single 96-well plate. SNPs chosen for validation in this dataset were likely to be the more common ones, since criteria of known frequencies or multiple lines of evidence were used to choose them for the assay development [34] , [39] . Only those SNP allele frequency estimates with sample sizes over 15 or more individuals in each population were analyzed. Recombination-based distances (cM) were interpolated from a linear regression centered on the closest known marker and three neighboring markers flanking it on either side (cM distances were from an NCBI update [Build 35] of the combined linkage-physical map of the human genome [28] ). In a few local regions (no more than one or two with errant data points per chromosome), the local regression was unable to reliably predict the position of the central marker. The errant data was dropped from analysis so that the predictive regressions were based on the remaining markers. The expected heterozygosity for each locus in both European and African Americans was calculated as follows [40] : We computed the unbiased estimate of F ST as described by Weir and Cockerham [41] and in a manner similar to Akey et al. [42] . For i populations (where i =  1, 2, …, s), frequency of the allele A in the subpopulation i is denoted as p Ai and sample size in each population as n i . Given this, F ST can be calculated in the following equation: where MSG is the observed mean square error for loci within populations, MSP is the observed mean square errors between the populations, and n c is the average sample size across samples that also incorporate the variance in sample sizes over the populations [41] : To calculate the distributions of genetic frequencies within and between the two populations, we sampled loci sequentially along each chromosome with the variable size frames in increasing size order, using only the odd numbers ( Figure 1A top). The smallest frame included n = 5 data points, the next was n = 7, and the largest frame size included 65 points, resulting in 31 odd-numbered sizes of sliding frames. For each sliding frame of size N, moving median of expected heterozygosity across the sampling frame for each of the two human populations in this study (H` AA for African Americans and H` EA for European Americans), as well as a moving variance S 2 F ST for the same region, was calculated using PROC EXPAND (SAS 9.1): Distributions of H` AA , H` EA , and S 2 F ST , were estimated by resampling each chromosome for each of the 31 odd-number sizes from 5 to 65 ( Figure 1A , bottom). The unrestricted random sampling (URS) option in PROC SURVEYSELECT (SAS 9.1) was used in the resampling process. COALESCENT SIMULATIONS To address the conjecture of using S 2 F ST as a tool to detect selection and the power of our computational approach, we used simulations with a single selected site positioned among 261 markers along a simulated 2.61 Mb-long chromosome. The model utilized marker density close to that in our experimental dataset: one marker for each 10 kb. We also assumed an average mutation rate of μ  = 10 −8 per site. The models varied in selection, strength, and intensity of recombination. These simulations were performed using SelSim [24] , assuming equal sampling (48 chromosomes) and a complete separation between the two populations. Values of λ were calculated for the control or ancestral, and the selected or derived population. We calculated maximum λ values for 100 replicate “chromosomes” by the same resampling-based algorithm in four different parameter combinations, alternating high and low selection coefficients versus high and low recombination rates. Selection parameter (s) ranged from neutral to 0.3% to 3% and recombination intensity per site (r) from 0.3 to 0.6, assuming an 10,000 effective population size for each population (maximum allowed by the program). The simulation scheme emulates one previously utilized to model population substructure of dog breeds where Pollinger et al. [43] showed significant heterozygosity and F ST effects after a selective sweep. The SelSim model assumes the initial mutation to be rare [24] , and may overestimate the selection signal if selection started on a mutation that reached significant frequency [43] .Our study used SelSim for “proof of principle” purposes to demonstrate the patterns of our measurements: a decrease in heterozygosity, and particularly the increase in the S 2 F ST measure we described. The exact strength and the extent of the selection signatures under different conditions are beyond the scope of the SelSim program and this report. In the simulation presently employed, 48 chromosomes carried the selected mutation (derived) and 48 chromosomes did not (ancestral), approximating the numbers of individuals we examined in each population. If one of the populations was recently separated from the other and selection acted on a mutation that arose in an ancestral population, the partial selective sweep is an approximation of the true process with the assumption that most of the neutral variation appeared before the introduction of the selected mutation. Therefore, the set of haplotypes that carry the selected mutation contains a subsample of neutral variation in the ancestral population, and can be contrasted with the set carrying the ancestral allele. SCANNING THE GENOME FOR SELECTED REGIONS Each chromosome was sampled 100,000 times and fractional ranks, ranging from 0 to 1, determined for the sliding windows (described above; Figure 1B, C ). The procedure was repeated 100 times, and fractional ranks for each SNP sliding window were averaged. Extreme values of fractional rank were predicted using curvilinear regression (with r 2 consistently >0.99) to consistently extend estimates of λ below about 1×10 −7 attainable by fitting the observed data in the most extreme 2.5% SNP frames into their resampled distributions. These results were summarized by determining the lowest mean value of the fractional ranks of the sliding windows centered on that SNP (λ) for each H` AA , H` EA , and S 2 F ST . The values of λ(H` AA ), λ(H` EA ), and λ(S 2 F ST ) are evaluated relative to each other and are plotted against the cM position for each chromosome (illustrated in Figure 1D top). Diagrams of expected putative regions of recent selection in African Americans, European Americans, and in both populations, as well as of the ancestral selection, are illustrated for λ in Figure 1D (bottom). SUPPORTING INFORMATION Lowest mean fractional rank values (λ) of H AA , H EA , and S 2 F ST for regions implicated previously as selection targets. Only those genes indicating recent positive selection in humans or humans and primates are shown. Significant values are underlined. (0.10 MB DOC) Table S1 Notes (0.03 MB DOC) Percentages of the SNPs and portion of chromosomes sampled (0.05 MB DOC) Locations of putative selected sites, range of their extent, and λ for H` AA and H` AA and S 2 F ST , as well as the genes located within these regions. (0.36 MB DOC) Locations of the cross-validated regions. (0.95 MB DOC) Locations of all the discovered regions and sites used in this scan and other studies (3.00 MB DOC) Individual graphs of 180 putative selection regions. Peak numbers correspond to Table S3 where the lowest mean rank value (λ), locations, and genes included in the putative selection regions are also shown. The vertical scale corresponds to the negative logarithm of the λ(H AA ) (green line), λ(H EA ) (blue line), and λ(S 2 F ST ) (red line). A putative selected site is identified where two of the peaks overlap (as in Figure 2A , bottom). The horizontal scale indicates location in cM. The locations of SNPs are represented by the black hash marks on the top of the graph. The extent of genes is represented by the horizontal blue line. Chromosome numbers (Chrom), location in cM, the most significant window size (out of 30 possible, windowsize), type of selection (SelcType), name of the closest gene (Gene), and the distance to it from the central location in cM (Dist) are all listed in the heading above each graph. The circle in the middle indicates the central location of the putative region. The range of the selected region is indicated by the two black vertical lines. The names of the genes included in the selected regions are listed from left to right in Table S3 . (1.29 MB PDF) Figure S1 Notes (0.02 MB DOC) (0.03 MB DOC) REFERENCES [1] Maynard Smith J Haigh J 1974 The hitchhiking effect of a favorable gene. Genet Res 23 23 35 [2] Sabeti PC Schaffner SF Fry B Lohmueller J Varilly P 2006 Positive natural selection in the human lineage. Science 312 1614 1620 [3] Lewontin RC Krakauer J 1973 Distribution of gene frequency as a test of the theory of the selective neutrality of polymorphisms. Genetics 74 175 195 [4] Akey JM Zhang G Zhang K Jin L Shriver MD 2002 Interrogating a high-density SNP map for signatures of natural selection. Genome Res 12 1805 1814 [5] Engle LJ Simpson CL Landers JE 2006 Using high-throughput SNP technologies to study cancer. Oncogene 25 1594 1601 [6] Karl SA Avise JC 1992 Balancing selection at allozyme loci in oysters: implications from nuclear RFLPs. Science 256 100 102 [7] Berry A Kreitman M 1993 Molecular analysis of an allozyme cline: alcohol dehydrogenase in Drosophila melanogaster on the east coast of North America. Genetics 134 869 893 [8] Taylor MF Shen Y Kreitman ME 1995 A population genetic test of selection at the molecular level. Science 270 1497 1499 [9] Huttley GA Smith MW Carrington M O'Brien SJ 1999 A scan for linkage disequilibrium across the human genome. Genetics 152 1711 1722 [10] Hamblin MT Di Rienzo A 2000 Detection of the signature of natural selection in humans: evidence from the Duffy blood group locus. Am J Hum Genet 66 1669 1679 [11] Hamblin MT Thompson EE Di Rienzo A 2002 Complex signatures of natural selection at the Duffy blood group locus. Am J Hum Genet 70 369 383 Epub 2001 Dec 2020 [12] Olson S 2002 Population genetics. Seeking the signs of selection. Science 298 1324 1325 [13] Sabeti PC Reich DE Higgins JM Levine HZ Richter DJ 2002 Detecting recent positive selection in the human genome from haplotype structure. Nature 419 832 837 [14] Tishkoff SA Verrelli BC 2003 Patterns of human genetic diversity: implications for human evolutionary history and disease. Annu Rev Genomics Hum Genet 4 293 340 [15] Vallender EJ Lahn BT 2004 Positive selection on the human genome. Hum Mol Genet 13 Spec No 2 R245 254 [16] Bamshad M Wooding SP 2003 Signatures of natural selection in the human genome. Nat Rev Genet 4 99 111 [17] Frazer KA Ballinger DG Cox DR Hinds DA Stuve LL 2007 A second generation human haplotype map of over 3.1 million SNPs. Nature 449 851 861 [18] Altshuler D Brooks LD Chakravarti A Collins FS Daly MJ 2005 A haplotype map of the human genome. Nature 437 1299 1320 [19] Voight BF Kudaravalli S Wen X Pritchard JK 2006 A map of recent positive selection in the human genome. PLoS Biol 4 e72 [20] Wang ET Kodama G Baldi P Moyzis RK 2006 Global landscape of recent inferred Darwinian selection for Homo sapiens. Proc Natl Acad Sci U S A 103 135 140 [21] Sabeti PC Varilly P Fry B Lohmueller J Hostetter E 2007 Genome-wide detection and characterization of positive selection in human populations. Nature 449 913 918 [22] Carlson CS Thomas DJ Eberle MA Swanson JE Livingston RJ 2005 Genomic regions exhibiting positive selection identified from dense genotype data. Genome Res 15 1553 1565 [23] Nielsen R Williamson S Kim Y Hubisz MJ Clark AG 2005 Genomic scans for selective sweeps using SNP data. Genome Res 15 1566 1575 [24] Spencer CC Coop G 2004 SelSim: a program to simulate population genetic data with natural selection and recombination. Bioinformatics 20 3673 3675 [25] Bersaglieri T Sabeti PC Patterson N Vanderploeg T Schaffner SF 2004 Genetic signatures of strong recent positive selection at the lactase gene. Am J Hum Genet 74 1111 1120 [26] Tishkoff SA Varkonyi R Cahinhinan N Abbes S Argyropoulos G 2001 Haplotype diversity and linkage disequilibrium at human G6PD: recent origin of alleles that confer malarial resistance. Science 293 455 462 [27] Spencer CCA Deloukas P Sarah H Mullikin J Myers S 2006 The Influence of Recombination on Human Genetic Diversity. PLoS Genet 2 e148 [28] Kong X Murphy K Raj T He C White PS 2004 A combined linkage-physical map of the human genome. Am J Hum Genet 75 1143 1148 [29] Teshima KM Coop G Przeworski M 2006 How reliable are empirical genomic scans for selective sweeps? Genome Research 16 702 712 [30] Kelley JL Madeoy J Calhoun JC Swanson W Akey JM 2006 Genomic signatures of positive selection in humans and the limits of outlier approaches. Genome Research 16 980 989 [31] Cavalli-Sforza LL 1966 Population structure and human evolution. Proc R Soc Lond B Biol Sci 164 362 379 [32] Parra EJ Marcini A Akey J Martinson J Batzer MA 1998 Estimating African American admixture proportions by use of population-specific alleles. Am J Hum Genet 63 1839 1851 [33] Smith MW Patterson N Lautenberger JA Truelove AL McDonald GJ 2004 A high-density admixture map for disease gene discovery in african americans. Am J Hum Genet 74 1001 1013 [34] De La Vega FM Dailey D Ziegle J Williams J Madden D 2002 New generation pharmacogenomic tools: a SNP linkage disequilibrium Map, validated SNP assay resource, and high-throughput instrumentation system for large-scale genetic studies. Biotechniques Suppl 48 50, 52, 54 [35] Cavalli-Sforza LL 2000 Genes, Peoples, and Languages. Seielstad M, translator. London, England University of California Press 228 [36] Begun DJ Holloway AK Stevens K Hillier LW Poh Y-P 2007 Population Genomics: Whole-Genome Analysis of Polymorphism and Divergence in Drosophila simulans. PLoS Biology 5 e310 [37] Dove A 2005 The SNPs are down: genotyping for the rest of us. Nat Meth 2 989 994 [38] Wright S 1951 The genetical structure of populations. Ann Eugen 15 323 354 [39] De la Vega FM Lazaruk KD Rhodes MD Wenz MH 2005 Assessment of two flexible and compatible SNP genotyping platforms: TaqMan SNP Genotyping Assays and the SNPlex Genotyping System. Mutat Res 573 111 135 [40] Nei M Roychoudhury AK 1974 Sampling variances of heterozygosity and genetic distance. Genetics 76 379 390 [41] Weir BS Cockerham CC 1984 Estimating F-statistics for the analysis of population sturucture. Evolution 38 1358 1370 [42] Tishkoff SA Williams SM 2002 Genetic analysis of african populations: human evolution and complex disease. Nature Reviews Genetics 3 611 621 [43] Pollinger JP Bustamante CD Fledel-Alon A Schmutz S Gray MM 2005 Selective sweep mapping of genes with large phenotypic effects. Genome Res 15 1809 1819