We sequenced one 100-kb ENCODE region-ENr123 (hg18: Chr12 38,826,477-38,926,476) in four different Andhra Pradesh ethnic groups representing three castes, Brahmin, Yadava, and Mala/Madiga, and one tribal group Irula (Figure 1a). We chose ENr123 because it has a low gene density and should represent a selectively neutral region (gene density of 3.1% and non-exonic conservation rate of 1.7%). Among the 92 individuals that passed quality-control steps, a total of 453 SNPs were identified, corresponding to a SNP density of one SNP per 221 bp. To determine the accuracy of the newly identified SNPs, we carried out additional experiments using the Roche 454 sequencing platform to validate the Indian-specific SNPs in individuals with heterozygous genotypes (see Materials and methods for details). The validation results showed that the genotypes of new SNPs have a high confirmation rate (approximately 80% for heterozygous SNPs). For alleles that have been seen only once in the dataset, the confirmation rate is greater than 85% (Supplemental Table S1 in Additional file 1).

To generate a comparable dataset, we applied the same SNP calling criteria on 722 HapMap individuals who were sequenced using the same protocol in the ENCODE3 project 29. We then merged these two datasets (four Indian populations and eight HapMap populations (CEU, CHB, CHD, GIH, JPT, LWK, TSI, and YRI)) to obtain a final data set that consists of 1,484 SNPs in 722 individuals from 12 populations (see Materials and methods for SNP merging and filtering details).

Among the 1,484 total SNPs, 234 (15.8%) are specific to Indian populations (four Andhra Pradesh populations and the HapMap northern Indian GIH; Figure 1b). For Indian individuals, the average number of specific SNPs per individual is 1.5. This number is lower than in HapMap African individuals (2.4 SNPs), but higher than both HapMap European (1.3 SNPs) and HapMap East Asian individuals (1.1 SNPs). This result suggests that higher autosomal genetic diversity is harbored in Indian samples compared to other HapMap Eurasian samples. Among the 453 SNPs in the four newly sequenced south Indian populations, 137 (30%) are not present in any HapMap populations (Figure 1c), including one novel non-synonymous singleton variant (Supplemental text in Additional file 1).

Because many genetic diversity measurements are influenced by sample size, we normalized the sample size of each group by randomly selecting a subset of HapMap individuals to match the sample size of the Indians. For convenience, we denote four groups of populations (African, East Asian, European, and Indian) as 'continental groups'. For continental groups, 152 unrelated individuals were randomly selected from HapMap African, European, and East Asian samples, respectively (matching the 152 Indian individuals in the dataset). At the population level, 24 individuals were randomly selected from each HapMap population, and all individuals from south Indian populations were included in the analyses. After sample size normalization, we measured genetic diversity using various summary statistics, including the number of segregating sites (S), Watterson's θ estimator, nucleotide diversity (π), and observed SNP heterozygosity (H) for each population and continental group (Table 1). We also evaluated the haplotype diversity in each group by averaging the haplotype heterozygosity in ten 10-kb non-overlapping windows and tested the neutrality of the region using the Tajima's D test. The Tajima's D test result was consistent with neutrality, providing no evidence for either positive or balancing selection in this region (Table 1), as expected given the low gene density in this region.

At the population level, π and H indicate that some Indian populations have diversity levels comparable to or even higher than those of HapMap African populations. Specifically, Mala/Madiga, Yadava, and Irula have the highest π among all populations (84.46 π 10^-5, 88.94 π 10^-5, and 82.77 π 10^-5, respectively). In contrast, Brahmins and HapMap GIH have lower diversity levels, comparable to HapMap European and East Asian populations (Table 1). Due to small sample sizes, the confidence intervals of π for all populations overlap. However, at the continental level, Indians have significantly higher nucleotide diversity than Europeans and East Asians, although θ and haplotype diversity are similar among the three groups (Table 1). Removal of unconfirmed genotypes in Indian individuals does not change the results (Supplemental text and Supplemental Table S3 in Additional file 1).

Several studies have shown that heterozygosity decreases with increasing distance from eastern Africa, presumably due to multiple bottlenecks that human populations experienced during the migration 2230. Among non-Indian populations, we observed a significant negative correlation between H and the distance to eastern Africa (Figure 2; r = -0.77, P = 0.04). However, when the Indian populations were included, the correlation became non-significant (r = -0.33, P = 0.29). This lack of correlation is due to large variation in H among the Indian populations (60.02 π 10^-5 in Brahmins to 95.12 π 10^-5 in the Irula). This result demonstrates great variation in diversity among groups within India.

To study the relationship among populations, we first performed principal components analysis (PCA) on the genetic distances between populations using the normalized dataset. When all populations are included in the analysis, the first principal component (PC1) accounts for 93% of the total variance and separates African and non-African populations (Supplemental Figure S1 in Additional file 1). In PCA of only Eurasian populations, PC1 separates Indian populations from European and East Asian populations, and PC2 separates European and Asian populations (Figure 3). Among Indian populations, the tribal Irula and HapMap GIH have the shortest distance to East Asian populations while Brahmin has the largest distance. The northern Indian GIH population diverges from south Indians and its closest relationship is with HapMap TSI populations. This observation is consistent with the general genetic cline in India observed in previous studies 1331. We also performed PCA and ADMIXTURE analysis at the individual level (Supplemental Figure S2 in Additional file 1). Because of the relatively small size of our dataset, individuals are not tightly clustered as seen in studies with genome-wide data 192223. The African individuals are separated from the Eurasian individuals, but Eurasian individuals from different populations are not separated into distinct clusters.

Next, we examined the divergence between Indian and non-Indian populations using pairwise F_ST estimates. In comparing major continental groups, India and Europe have the smallest F_ST value (Table 2). At the individual population level, however, Indian populations show varying affinities to other Eurasian populations: the Indian tribal population (Irula) shows closer affinity to HapMap East Asian populations while the HapMap GIH and the Brahmin show a closer relationship to HapMap European populations. The Mala/Madiga and Yadava show a similar distance to the HapMap European and East Asian populations (Table 3). Among Indian populations (Supplemental Table S2 in Additional file 1), the smallest F_ST value is between Yadava and Mala/Madiga (0.1%), and the largest F_ST value is between HapMap GIH and the tribal Irula (10.4%).

The complete sequence data allow us to obtain an accurate derived-allele frequency (DAF) spectrum. At both the continental and population levels, the DAF spectra in our dataset are characterized by a high proportion of low-frequency SNPs, as expected for sequencing data (Supplemental text and Supplemental Figure S3 in Additional file 1). Based on the DAF spectra, we are able to infer the parameters associated with Indian population history, such as the divergence time, effective size, and migration rate between populations using the program ∂a∂i (Diffusion Approximation for Demographic Inference) 32.

Because ∂a∂i can simultaneously infer population parameters in models involving three populations, we first estimated the parameters associated with the out-of-Africa event using the African continental group and two continental Eurasian groups. We started from a simplified three-population divergence model based on the out-of-Africa model described in ∂a∂i 32 and assessed the model-fitting improvement of adding different parameters to the model (Supplemental text in Additional file 1). Our results suggest that allowing exponential growth in the Eurasian continental groups substantially improves the model. On the other hand, allowing migrations among groups provides little improvement in the data-model fitting, suggesting that little gene flow occurred between the continental groups (Supplemental Figure S5 in Additional file 1). Therefore, we inferred the parameters from the three-population out-of-Africa model, allowing exponential growth in the Eurasian groups but no migration among groups (Figure 4a). Under this model, a one-time change in African population size occurs at time T_Af before any population divergence, and the population size changes from the ancestral population size N_A to N_Af in Africa. At time T_B the Eurasian ancestral population with a population size of N_B diverges from the African population, while the African population size N_Af remains constant until the present. The two Eurasian groups split from the ancestral population N_B at time T_1-2, with initial population sizes of N_{1_0} and N_{2_0}, respectively. Both populations experience exponential population size changes from the time of divergence to reach the current population sizes N₁ and N₂.

The inferred parameters between continental groups, along with confidence intervals (CIs) for each parameter, are shown in Table 4. When the mutation rate is set at 1.48 π 10^-8 per base pair per generation (see Materials and methods for mutation rate estimate), the ancestral population size is estimated to be between 13,000 and 14,000 for all models (Table 4). The African effective population size estimates (N_Af, 18,036 to 18,976; CI, 15,077 to 22,673) are comparable to the size of the Eurasian ancestral population (N_B, 12,624 to 21,371; CI, 7,360 to 32,843). At the time of the Eurasian population divergence, the population sizes of the two Eurasian continental groups in each model (N_{1_0} and N_{2_0}) are consistently smaller than the African and the Eurasian ancestral population sizes, with one exception for the estimated European population size (25,543; CI 6,101 to 29,016) in the Africa-East Asia-Europe model. These results suggest that the Eurasian population experienced population bottlenecks at the time of their divergence. Among Eurasians, East Asians have the smallest effective population size at the time of divergence (approximately 1,500; CI, 779 to 3,703; Table 4). The divergence time estimates between Africans and non-Africans range from 88.4 to 111.5 kya and the CIs of all three estimates overlapped, consistent with the existence of a single ancestral Eurasian population. The three non-African continental groups diverged from each other more recently than 40 kya: East Asians were separated from Indians (39.3 kya; CI, 29.7 to 59.1) and Europeans (39.2 kya; CI, 29.8 to 55.8) before the divergence of Indians and Europeans (26.6 kya; CI, 20.1 to 40.8). Overall, these results support a scenario in which the ancestors of the Indian, European, and East Asian individuals left Africa in one major migration event, and then diverged from one another more than 40,000 years later.

To further examine the population history among Eurasian populations, we constructed a four-population model containing all four continental groups (Figure 4b). Because parameters from only three populations can be estimated by ∂a∂i at the same time, we fixed the parameters of the out-of-Africa epoch (N_Af , N_B, T_Af, and T_B) in the model based on the parameters estimated from the three-population model with the highest likelihood (Africa-East Asia-European), as described in ∂a∂i 32. A model comparison again suggests that adding migrations to the model does not substantially improve the model-fitting (Supplemental text and Supplemental Figure S6 in Additional file 1). Therefore, migrations were excluded from the model to reduce the number of inferred parameters and to improve the speed of computation. Among the three population divergence scenarios, two models ('East Asia first' and 'India first') showed similar maximum likelihood values (-1,278.9 and -1,278.7, respectively), indicating comparable fitting to the data. In contrast, the 'Europe first' model has a substantially lower maximum likelihood value (-1,280.7), suggesting that this model is less plausible. The estimated parameters for the 'East Asia first' and the 'India first' models are shown in Table 5. Consistent with the three-population models, the 'East Asia first' mode estimates that East Asians diverged from the ancestral Eurasian population approximately 44 kya, and Europeans and Indians diverged approximately 24 kya. Interestingly, the 'India first' model suggests that the divergence time among the three continental groups are similar, with Indians diverging only 0.2 kya before Europeans and East Asians. Under this model, the initial population size of the Indian population (N_{1_0}, 11,410; CI, 4,568 to 28,665) is comparable to the Eurasian ancestral population size (N_B, 12,345), consistent with the high diversity we observed in these Indian samples.

When individual populations are analyzed, the patterns are largely consistent with the results from continental groups (Supplemental text and Supplemental Table S4 in Additional file 1). The CIs around the parameters are generally larger, indicating a loss of power due to the smaller sample sizes of the individual populations compared to the continental groups.

Ninety-four individuals from three caste groups and one tribal group from Andhra Pradesh, India were sampled (Figure 1a). All samples belong to the Dravidian language family and were collected as unrelated individuals as described previously 4546. All studies of South Indian populations were performed with approval of the Institutional Review Board of the University of Utah and Andhra University, India. To sequence the ENCODE region ENr123, we used the same sets of primers that were used for the ENCODE3 project for PCR amplification and the same Sanger sequencing. Next, we obtained the sequence of 722 HapMap individuals from the ENCODE3 project 29 and performed SNP calling using the same SNP discovery pipeline 47. This experimental design allowed us to directly compare genetic variation patterns observed in these Indian populations with those observed in the HapMap populations studied by ENCODE3 29. The sequence traces of the Indian samples generated from this study can be accessed at NCBI trace archive 48 by submitting the query: center_project = 'RHIDZ'.

After the SNP-calling process, two individuals with less than 80% call rates were removed from the dataset (one Brahmin and one Yadava). The SNP calls from the remaining 92 samples that passed quality control were then combined with the SNP calls from eight HapMap non-admixed populations studied by ENCODE3, including individuals from the Centre d'Etude du Polymorphisme Humain collection in Utah, USA, with ancestry from Northern and Western Europe (CEU), Han Chinese in Beijing, China (CHB), Japanese in Tokyo, Japan (JPT), Yoruba in Ibadan, Nigeria (YRI), Chinese in Metropolitan Denver, CO, USA (CHD), Gujarati Indians in Houston, TX, USA (GIH), Luhya in Webuye, Kenya (LWK), and Toscani in Italy (TSI), to create a final dataset containing 722 individuals from 12 populations.

After merging the HapMap and the south Indian data sets, 112 loci that are fixed in all 12 populations were removed from the dataset. Thirteen tri-allelic SNPs were also removed because most analyses in this study are designed for bi-allelic SNPs. For SNPs that are fixed in certain populations, genotypes were filled-in using the hg18 reference allele because the reference allele information was used in the SNP calling process (that is, only genotypes that are different from the reference alleles are called as SNPs).

The Hardy-Weinberg equilibrium test was performed on each of the 12 populations, and P-values from each test were obtained and transformed to Z-scores. Twelve Z-scores were combined to a single Z-score and transformed to a single P-value for each SNP. Bonferroni correction was used, and 48 SNPs that failed the test at the 0.01 level (P < 0.01/1,532) were removed. The ancestral/derived allele states of each SNP were determined using the human/chimpanzee alignment obtained from the UCSC database (hg18 vs.panTro2 49). Minor-alleles of 17 SNPs were assigned as the derived allele because the derived allele could not be determined by human-chimpanzee alignments. Genotypes of all samples in the final dataset are available as a supplemental file on our website 50 under Published Data.

For the 137 SNPs that are specific to our samples (that is, not present in any HapMap populations), we performed a validation experiment using an independent platform (Roche 454). When the minor allele is present in more than five individuals at a given locus, five individuals with the heterozygous genotype were randomly selected for validation. Among the 137 SNPs, we successfully designed and assayed 119 SNPs in 211 individual experiments. For the validation pipeline, we used PCR to amplify regions around the variants using the same primers as those used in the initial variant detection pipeline. In order to make genotype calls on all experiments simultaneously and also to reduce the cost of Roche 454 sequencing, we pooled PCR reactions in ten different pools and each pool was sequenced using a quarter of a Roche Titanium 454 sequencing run. The analysis was done using the Atlas-SNP2 pipeline available at the BCM-HGSC 51. Reads from the 454 runs were anchored using BLAT 52 to a unique spot in the genome, followed with the refined alignments using the cross_match program 53. We required at least 50 reads mapped to the variant site to make a validation call and the fraction of reads with the variant to be >15% of all reads mapping to that site.

Sequence-analysis statistics (S, θ, π, H and Tajima's D), and the confidence intervals for θ and π were calculated using the Population Genetics and Evolution Toolbox 54 in MATLAB (version r2009a). To assess haplotype diversity, the dataset was phased using fastPHASE (version 1.2) 55 with imputation, and the phased dataset was separated into ten 10-kb non-overlapping windows. Haplotype heterozygosity was then calculated for each window, and the mean heterozygosity for each population/continental group was calculated. For the SNP heterozygosity/geographic distance correlation analysis, the great-circle distance between each population and Addis Ababa, Ethiopia, a proposed point of modern human origin 56, was calculated. For populations that were collected from places other than their origins, an approximate origin location was used, such as Beijing, China for CHD, and Gujarat, India for GIH. F_ST estimates between populations were calculated by the method described by Weir and Cockerham 57. Nei's genetic distances between populations were estimated from allele-frequency data as implemented in the PHYLIP software package 58 and PCA was performed using MATLAB.

Demographic history parameters were inferred using the program ∂a∂i (version 1.5.2) 59. Using a diffusion approximation to the allele-frequency spectrum, ∂a∂i implements a series of methods to infer population history based on sequence data. We compared three different three-population out-of-Africa models and three four-population out-of Africa models to test the effect of adding different parameters to the model (Supplemental text and Supplemental Figures S5 and S6 in Additional file 1). For the two models used in the final analysis, the python programs that were used to estimate the parameters, including the function calls, grid sizes, initial parameters, and parameter boundaries, are shown in Supplemental Figures S7 and S8 in Additional file 1. To ensure that the algorithm identified the optimal parameters, ten independent runs were performed on each model, and the parameter set with the highest likelihood was selected as the final result. For each model, 500 bootstrap replicates were performed on the dataset to obtain the confidence intervals. The per-generation mutation rate was estimated based on the human-chimpanzee divergence in this region (1.2%) using the method described in 43, with a generation time of 25 years, a human-chimpanzee speciation time of 6 million years ago, and a human-chimpanzee ancestral effective population size of 84,000 (averaged from the estimates from 606162).