Proportioning whole-genome single-nucleotide-polymorphism diversity for the identification of geographic population structure and genetic ancestry

Abstract

The identification of geographic population structure and genetic ancestry on the basis of a minimal set of genetic markers is desirable for a wide range of applications in medical and forensic sciences. However, the absence of sharp discontinuities in the neutral genetic diversity among human populations implies that, in practice, a large number of neutral markers will be required to identify the genetic ancestry of one individual. We showed that it is possible to reduce the amount of markers required for detecting continental population structure to only 10 single-nucleotide polymorphisms (SNPs), by applying a newly developed ascertainment algorithm to Affymetrix GeneChip Mapping 10K SNP array data that we obtained from samples of globally dispersed human individuals (the Y Chromosome Consortium panel). Furthermore, this set of SNPs was able to recover the genetic ancestry of individuals from all four continents represented in the original data set when applied to an independent, much larger, worldwide population data set (Centre d'Etude du Polymorphisme Humain-Human Genome Diversity Project Cell Line Panel). Finally, we provide evidence that the unusual patterns of genetic variation we observed at the respective genomic regions surrounding the five most informative SNPs is in agreement with local positive selection being the explanation for the striking SNP allele-frequency differences we found between continental groups of human populations.

Figure 1

Percentage of information explained when the number of markers that are ascertained from 8,491 SNPs by use of the genetic algorithm based on the informativeness of assignment index (I_n) is increased from 1 to 10, given four continental groups and the YCC panel (see main text for details). The 95% CI of each SNP combination was computed by resampling the same number of chromosomes from the populations and computing I_n 1,000 times.

Figure 2

STRUCTURE analysis of the YCC samples, with K=2, 3, or 4 groups, performed using genotypes of the 10 most informative SNPs ascertained using the genetic algorithm with the total YCC data. STRUCTURE analyses were computed using a model without admixture (A) and a model with admixture (B). Each analysis was repeated five times, after a Markov chain–Monte Carlo (MCMC) burning period of 50,000 and considering the next 200,000 MCMC iterations. In all five runs, good mixing was observed, and similar results were found in accordance with the model used. The natural logarithm of the estimated probability of the data (lnp) is as follows. In panel A, for K=2, lnp=-762.2; for K=3, lnp=-629.2; and, for K=4, lnp=-557.4. In panel B, for K=2, lnp=-764.9; for K=3, lnp=-631.2; and, for K=4, lnp=-559.5.

Figure 3

MDS plot based on the I_n matrix computed between pairs of populations by use of the genotypes of the 10 most informative SNPs in the 51 population samples from CEPH-HGDP. Four clusters of population can be identified: (i) sub-Saharan African populations, (ii) American populations, (iii) Eastern Asian and Oceanian populations, and (iv) European, Middle Eastern, North African, and Central/South Asian populations.

Figure 4

STRUCTURE analysis of the CEPH-HGDP samples, with K=2, 3, 4, or 5 groups, performed using genotypes of the 10 most informative SNPs ascertained using the genetic algorithm with the total YCC data. Two different STRUCTURE analyses were computed: a population model without admixture (A) and a population model with admixture (B). Each analysis was repeated five times after an MCMC burning period of 100,000 and considering the next 10,000 MCMC iterations. In all five runs, good mixing was observed, and similar results were found in accordance with the model used. The lnp, assuming K groups, is as follows. In panel A, for K=2, lnp=-11,801.2; for K=3, lnp=-10,977.3; for K=4, lnp=-10,279.2; and, for K=5, lnp=-10,324.9. In panel B, for K=2, lnp=-11,886.2; for K=3, lnp=-11,070.6; for K=4, lnp=-10,345.5; and, for K=5, lnp=-10,456.9. Cen. Af. Rep. = Central African Republic; S. Afr. = South Africa.

Figure 5

STRUCTURE analysis of each of the four groups detected in the HGDP-CEPH populations by previous STRUCTURE analysis (see main text) that considers models without admixture (A) and with admixture (B) and assumes K=2. A certain degree of population (sub)structure can be observed only in the case of American populations, but it disappears when three groups are considered (data not shown). Each analysis was repeated five times, after an MCMC burning period of 200,000 and considering the next 200,000 MCMC iterations. In all five runs, good mixing was observed, and similar results were found in accordance with the model used. The lnp, assuming K=2, is as follows. In panel A, for sub-Saharan Africa, lnp=-958.3; for America, lnp=-1,048.1; for East Asia and Oceania, lnp=-3,262.0; and, for Europe, the Middle East, Central/South Asia, and North-Africa, lnp=-5,321.5. In panel B, for sub-Saharan Africa, lnp=-946.7; for America, lnp=-1,057.4; for East Asia and Oceania, lnp=-3,263.5; and, for Europe, the Middle East, Central/South Asia, and North-Africa, lnp=-5,433.1.

Figure 6

BAPS 3.2 clustering results for K=2, 3, 4, and 5 groups in the HGDP-CEPH panel by use of the 10 most informative SNPs ascertained using the genetic algorithm with the YCC data. Each column represents an individual. The log (marginal likelihood) for K=2 groups is −11,687.5; for K=3, −10,832.6; for K=4, −10,164.8, and, for K=5, −10,024.32.

Figure 7

Sliding-window and haplotype analyses performed on the genomic region that includes SNP rs952718 and the ABCA12 gene. A, Sliding-window plot of the mean value observed for each window (the gene is represented by a black bar). B, Associated P value for comparison with an empirical distribution based on >10,000 genes (see main text). The P=.05 cutoff is represented by a black line. C, Bifurcation plots of the main core haplotypes in the three populations considered. D, Extended homozygosity versus genomic distance to the core haplotype. The region of the core haplotype was selected on the basis of the largest region that was statistically significant in the sliding-window analysis (from rs6758257 to rs6753310; see main text for details).

Figure 8

Sliding-window and haplotype analyses performed on the genomic region that includes SNP rs722869 and the VRK1 gene. A, Sliding-window plot of the mean value observed for each window (the gene is represented by a black bar). B, Associated P value for comparison with an empirical distribution based on >10,000 genes (see main text). The P=.05 cutoff is represented by a black line. C, Bifurcation plots of the main core haplotypes in the three populations considered. D, Extended homozygosity versus genomic distance to the core haplotype. The region of the core haplotype was selected on the basis of the largest region that was statistically significant in the sliding-window analysis (from rs1957137 to rs17191471; see main text for details).

Figure 9

Sliding-window and haplotype analyses performed on the genomic region that includes SNP rs1858465. A, Sliding-window plot of the mean value observed for each window. B, Associated P value for comparison with an empirical distribution based on >10,000 genes (see main text). The P=.05 cutoff is represented by a black line. C, Bifurcation plots of the main core haplotypes in the three populations considered. D, Extended homozygosity versus genomic distance to the core haplotype. The region of the core haplotype was selected on the basis of the largest region that was statistically significant in the sliding-window analysis (from rs2137476 to rs1398515; see main text for details).

Figure 10

Sliding-window and haplotype analyses performed on the genomic region that includes SNP rs1344870. A, Sliding-window plot of the mean value observed for each window. B, Associated P value for comparison with an empirical distribution based on >10,000 genes (see main text). The P=.05 cutoff is represented by a black line. C, Bifurcation plots of the main core haplotypes in the three populations considered. D, Extended homozygosity versus genomic distance to the core haplotype. The region of the core haplotype was selected on the basis of the largest region that was statistically significant in the sliding-window analysis (from rs2335092 to rs1898300; see main text for details).

Figure 11

Sliding-window and haplotype analyses performed on the genomic region that includes SNP rs1876482 (1 of the 10 most informative SNPs identified), which is located in the LOC442008 gene, by use of Perlegene data. A, Sliding-window plot of the mean value observed for each window (the gene is represented by a black bar). B, Associated P value for comparison with an empirical distribution based on >10,000 genes (see main text). The P=.05 cutoff is represented by a black line. C, Bifurcation plots of the main core haplotypes in the three populations considered. D, Extended homozygosity versus genomic distance to the core haplotype. The region of the core haplotype was selected on the basis of the largest region that was statistically significant in the sliding-window analysis (from rs12619554 to rs4832712; see main text for details). Note the high frequency of the third haplotype in the case of Asian populations and the slow decay of the EHH of that haplotype compared with the other haplotypes both within and between populations.