Bene Israel cluster with Indian populations but as a distinct group

We genotyped Bene Israel individuals and combined the data with 14 other Jewish populations from worldwide Diaspora previously genotyped using the same array [12,14]. We applied various quality control (QC) steps on these samples, resulting with 18 individuals of the Bene Israel community together with 347 samples from the other 14 Jewish populations. We also applied the same QC steps to a different dataset with samples from 18 different Indian populations (96 individuals) [17], as well as HapMap3 populations, that were genotyped previously on the same array [17] and merged the two datasets. We also merged the data with additional populations from the HGDP panel [19]: three non-Jewish Middle Eastern populations (Druze, Bedouin and Palestinians), and nine Pakistani populations (S1 Table). The Middle-Eastern populations were used to distinguish between Middle-Eastern and Jewish specific ancestry while the Pakistani populations were used to represent populations that are geographically located between India and the Middle-East. In addition, some Pakistani populations are also part of the ANI-ASI admixture, with a relatively large ANI component as compared to the Indian populations [17,18] and therefore also represent this ancient admixture. As the merging with HGDP resulted in considerable reduction in number of SNPs available for analysis, we only considered this merged dataset for some analyses.

PCA (Principal Component Analysis) of the merged dataset, including four HapMap populations (YRI, CEU, CHB and JPT; total 873 individuals) showed that Jewish populations cluster together with Europeans and Middle-Eastern populations, while Indian and Pakistani populations form their own cluster, between East-Asians and Jews/Europeans (Fig 1A). The Bene Israel population clustered with the Indian and Pakistani populations, similar to the results of a previous study [13]. Bene Israel was the closest to the Jewish/Middle-Eastern/European cluster as compared to all other Indian populations, while several Pakistani populations were even closer to that cluster than Bene Israel (Fig 1A). When focusing only on Jewish, Middle-Eastern, Pakistani and Indian samples (Fig 1B), the first PC separated between Indian/Pakistani and Jewish/Middle-Eastern populations while the second PC spanned the Jewish populations. Members of non-Jewish Middle-Eastern populations were located within the Jewish cluster, near Middle-Eastern Jewish populations. Indian populations were ordered based on their ANI-ASI admixture [17,18] such that populations with higher ANI proportion were closer to Jews in general, as expected from the Middle-Eastern origins of the latter [12–15]. Bene Israel members were located closely to members of other Indian populations but were also the closest to samples from Jewish populations among all Indian samples, while some Pakistani populations were more similar to Jewish and Middle-Eastern populations (Fig 1B). PCA with only Bene Israel, Indian and Pakistani population showed the ANI-ASI incline spanning both Indian and Pakistani populations, while Bene Israel were near populations with high ANI component, but slightly off the incline, perhaps suggesting a different origin. (Fig 1C). PCA of only Bene Israel, Pakistani, Middle-Eastern and Jewish populations showed the separation of Bene Israel members from other Jewish populations and the fact that some Pakistani populations were closer than Bene Israel to Jewish populations (Fig 1D). To avoid bias in PCA due to differences in number of samples between populations [20], we repeated the analysis while limiting the number of samples from each population to 4, and obtained similar results (S1 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Principal Component Analysis of Jewish, Indian and worldwide populations.

Each panel presents the top two principal components for a set of populations that include Bene Israel together with: (A) Jewish, Indian, Pakistani, Middle Eastern and four worldwide HapMap populations (CEU, CHB, JPT and YRI); (B) Jewish, Middle-Eastern, Pakistani and Indian populations; (C) Indian and Pakistani populations; (D) Jewish, Middle-Eastern and Pakistani populations. Abbreviations of Jewish populations: Bene Israel (Bene), Algerian Jews (ALGJ), Ashkenazi Jews (ASHJ), Djerban Jews (DJEJ), Georgian Jews (GEOJ), Greek Jews (GRKJ), Iranian Jews (IRNJ), Iraqi Jews (IRQJ), Italian Jews (ITAJ), Libyan Jews (LIBJ), Moroccan Jews (MORJ), Syrian Jews (SYRJ), Tunisian Jews (TUNJ), Yemenite Jews (YMNJ).

https://doi.org/10.1371/journal.pone.0152056.g001

We also examined the relation between Bene Israel and other Indian and Jewish populations using the F_ST statistic, which measures genetic drift between populations based on differences in allele frequencies [17,21] (S2 Fig and S2 Table). This analysis revealed the isolation and genetic drift of Bene Israel from both Indian and Jewish populations: While the average F_ST between pairs of Jewish populations was 0.011, the average F_ST between Bene Israel and other Jewish populations was significantly higher (0.04, Wilcoxon rank sum P-value = 1.97e-9). Similarly, while the average F_ST between different Indian populations was 0.011, the mean F_ST between Bene Israel and other Indian populations was significantly higher (0.033, P-value = 8.63e-12, Wilcoxon rank sum test).

ADMIXTURE analysis suggests Bene Israel members have Middle-Eastern ancestry

ADMIXTURE [22] assigns for each individual its proportion in any of a set of hypothetical ancestral populations and hence can reveal relations between different populations. We used this tool on our dataset for varying values of K (the number of hypothetical ancestral populations) on the same set of 873 individuals from the first PCA (Fig 2 and S3 Fig). For K = 3, we observed three clusters: East Asian, Sub-Saharan African and Middle Eastern/European. Indian and Pakistani populations were mainly composed from Middle-Eastern/European and East-Asian components. The proportion of the Middle-Eastern/European component in the Indian and Pakistani populations was highly correlated to their ANI component in the ANI-ASI admixture [17] (R = 0.98, P-value<e-16, Spearman correlation). Among the Indian populations, Bene Israel had the highest proportion of Middle-Eastern/European component, but it was comparable to that of some other Pakistani populations. At K = 4 an Indian cluster emerged which reflects the ASI component in these populations and most Indian and Pakistani populations were composed from this component and the Middle-Eastern/European component, while some of them also had an East-Asian component. Again, Bene Israel and three Pakistani populations (Balochi, Brahui and Makrani) had the highest proportions of Middle-Eastern/European component among Indian and Pakistani populations. At K = 5, the European/Middle-Eastern cluster was divided into two clusters: European (reflected by the European population CEU) and Middle-Eastern (reflected by Jewish and Middle-Eastern populations). Importantly, Bene Israel population exhibited a different trend as compared to other Indian populations: While the ANI component of Indian populations was now mainly reflected in the European component, Bene Israel showed a significantly higher proportion of a Middle-Eastern component (mean 29.5% as compared to less than 14% in all members of other Indian populations, Wilcoxon rank sum P-value = 1.75e-12). Nevertheless, some Pakistani populations showed similar or even higher proportions of the Middle-Eastern component (e.g., mean of 33% for Makrani). At K = 6, which provided the best fit based on cross-validation, a new cluster emerged which was mainly found in North-African Jews (Djerban, Libyan and Tunisian Jews). At K = 7 a new cluster represented Iranian Jews, but was also present in larger fractions in other Middle-Eastern and Asian populations, including Bene Israel. Interestingly, at K = 8 Bene Israel formed their own hypothetical ancestral component, marking again the uniqueness of this population and its deviation from other populations. This ancestral component was present, in minute proportions, in many Indian and Pakistani populations but also in some Middle-Eastern populations (Jewish and non-Jewish).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. ADMIXTURE analysis for Jewish, Indian, Pakistani, Middle Eastern (Druze, Bedouin and Palestinians) and representative HapMap (CEU, YRI, JPT and CHB) populations.

K, the number of clusters, varies from K = 3 to K = 8. We colored some of the populations names based on the following groups: Bene Israel (purple), Jews (red), Indian (blue) and Middle-Eastern (green). See also S3 Fig.

https://doi.org/10.1371/journal.pone.0152056.g002

Identity-by-descent analysis suggest Bene Israel members are related to Jewish populations

The results based on PCA and ADMIXTURE, show that Bene Israel is more closely-related to Middle-Eastern and Jewish populations as compared to all other Indian populations examined here. However, the same claim cannot be made when compared to some Pakistani populations. Therefore, the question remains: Does Bene Israel have Jewish or Middle-Eastern ancestry that is not shared by other Asian populations? Next, we analyzed the relations between populations based on Identity-by-descent (IBD) sharing of their individuals. IBD segments shared by two individuals represent a segment inherited from a common ancestor. Higher IBD sharing, and specifically of long segments, suggests a more recent common ancestor without intervening recombination [23]. Following a similar previous analysis of the Jewish populations examined here (except Bene Israel) [12,14], we used GERMLINE [24] to detect IBD segments between individuals and defined the IBD sharing between individuals to be the total length (in cM units) of IBD segments shared between the two individuals. IBD sharing between populations was defined as the average IBD sharing of unrelated individuals from these populations. As expected from previous results, each Jewish population exhibited significant higher IBD sharing with other Jewish populations than with Indian populations and each Indian population exhibited higher IBD sharing with other Indian populations than with Jewish populations (Wilcoxon P-value<0.05 for all populations; Fig 3A). Having these two IBD clusters of Indian and Jewish populations, we observed that compared to all Jewish populations, Bene Israel had the highest IBD sharing with all Indian populations, and compared to all Indian populations, Bene Israel had the highest IBD sharing with all Jewish populations (S4A Fig). Furthermore, the only population with no significant IBD sharing between the two clusters of Jewish and Indian populations was Bene Israel: (mean IBD sharing = 18.24 cM vs. 18.19 cM with Indian and Jewish populations, respectively. P-value = 0.61; Wilcoxon rank sum test; Fig 3B). Middle-Eastern Jews (specifically Georgian, Iraqi, Syrian and Iranian Jews) showed significantly higher average IBD sharing as compared to all other Jewish populations examined (Wilcoxon test P-value <e-10 for all pairs). Interestingly, the closest Indian populations to Bene Israel (Velama, Lodi and Bhil) were not those with the highest ANI component.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. IBD sharing between and within Jewish, Indian, Pakistani and Middle-Eastern populations.

(A) Average IBD sharing between different populations. For each population we measured its average IBD with Bene Israel (purple) and all other Jewish (red) and Indian (blue) populations. (B) IBD sharing of Bene Israel with other Jewish and Indian populations. (C) Average IBD sharing between different populations. For each population from panel A, with the addition of Pakistani and Middle-Eastern populations, we measured its average IBD with Bene Israel (purple) and all other Jewish (red) and Indian (blue) populations. (D) IBD sharing of Bene Israel with other Jewish, Indian, Pakistani and non-Jewish Middle-Eastern populations. Analyses in panels C-D were performed on the dataset merged with HGDP dataset that contained smaller number of SNPs, and therefore the differences in IBD sharing values. (E) IBD sharing within populations.

https://doi.org/10.1371/journal.pone.0152056.g003

Next, we used the merged dataset with Middle-Eastern and Pakistani populations for a lower-resolution IBD analysis. Still, Bene Israel showed higher IBD sharing with Middle-Eastern Jewish populations in this dataset as well (Fig 3C). Importantly, while the PCA, and ADMIXTURE showed that some Pakistani populations were more similar to Middle-Eastern and Jewish populations than Bene Israel, no Pakistani population showed as high IBD sharing with Jewish populations as compared to Bene Israel (Fig 3C). In addition, we used this merged dataset to examine whether the relatively high IBD sharing of Bene Israel with Jewish populations and mainly Middle-Eastern Jews is Jewish specific or Middle-Eastern in general. The IBD sharing of Bene Israel and non-Jewish Middle-Eastern populations was lower than their sharing with all other Jewish populations (Fig 3C; S4B Fig), implying that the link between Bene Israel and Jewish populations is at least in part Jewish-specific and not only Middle-Eastern. Although there were differences between the IBD sharing in the two datasets due to the different set of SNPs, there was overall significant correlation between the ranking of IBD sharing of Jewish (R = 0.85, P-value = 1.36e-4; Spearman correlation) and Indian (R = 0.52, P-value = 0.03, Spearman correlation) populations in the two datasets. Higher sharing with Middle-Eastern Jewish populations was also observed when we restricted the analysis to longer segments of IBD that reflect a more recent ancestor (S5 Fig). While we did not see a clear trend for Indian populations showing higher IBD with Bene Israel, the higher-resolution analysis showed Velama to have the highest IBD sharing also in respect to longer IBD segments (S5 Fig).

Bene Israel as an admixed population of Jewish and Indian ancestral populations

Motivated by the above results, we next examined whether the Bene Israel community was an admixture of Indian and Jewish ancestral populations, using two different approaches as implemented in the ALDER [25] and GLOBETROTTER [26] tools. Given a putative admixed population and two populations that are taken as surrogates for the true ancestral populations, ALDER computes an admixture linkage disequilibrium (LD) statistic in the admixed population and uses it to examine whether the population is indeed an admixture of the ancestral populations [25]. In most cases of one Jewish and one Indian population taken as surrogate ancestral populations, there was a consistent and significant evidence for Bene Israel being admixture between these two populations (147 out of 252 possible pairs. S3 Table and S6 Fig; the only population that did not show a significant evidence for being an ancestral population for Bene Israel was the Indian Kashmiri Pandit). Repeating the same analysis but with pairs of Indian populations or pairs of Jewish populations, as well as replacing Bene Israel with any other Indian or Jewish population did not find any pair with significant and consistent evidence for admixture, suggesting that the observed admixture for Bene Israel was not reflecting ANI-ASI admixture but a unique admixture between Jewish and Indian populations. ALDER admixture estimated time varied across the 147 significant pairs of populations, from ~19 (Iraqi Jews and Satnami) to ~33 (Georgian Jews and Mala) generations ago (650–1050 years ago, assuming 29 years per generation [18,27]) with an average of ~25 generations (~820 years) ago (S2 Table). These estimations place the admixture between a Jewish and Indian population well after the estimated time for the ANI-ASI admixtures of Indian populations (64–144 generations ago [18]) and after the establishment of many Jewish Diasporas [15] (Fig 4A). Turning to admixture proportions estimations based on ALDER, those estimated for Indian populations, varying between 44% (Vaish) and 20.2% (Kharia) were generally higher than that of Jewish populations, varying between 23% (Georgian Jews) and 15.5% (Libyan Jews; Fig 4B). When repeating ALDER analysis using the merged dataset with non-Jewish Middle-Eastern populations, the results were less significant, as expected by the smaller number of markers, but still many pairs of one Jewish/Middle-Eastern population and one Indian population were significant. Importantly, the results were more significant for Jewish populations as compared to non-Jewish Middle-Eastern populations: While 3 of the 17 (17.6%) Jewish/Middle-Eastern populations examined were non-Jewish, only 8 of the 113 (7.1%) significant pairs contained non-Jewish population, and all other pairs contained a Jewish population (S4 Table and S7 Fig). Furthermore, when we replaced Bene Israel with Pakistani populations, there was no evidence for any Pakistani population being an admixed population with both Jewish and Indian ancestry. This result further emphasizes that the admixture detected by ALDER is not the ANI-ASI admixture.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Bene Israel as an admixed population with Indian and Jewish ancestry.

(A) Bene Israel as an admixed population of Indian and Jewish origin. Admixture time (~19–33 generations ago) is after the ANI-ASI admixture of Indian populations (64–144 generations ago). The ANI side is associated with Eurasian populations while ASI is associated with indigenous Andaman Island people (e.g., Onge). Dotted lines correspond to admixture between populations. This is a schematic overview of Bene Israel genetic history and branch lengths are not proportional to actual time. (B) ALDER admixture proportions estimations for Indian and Jewish populations being ancestral populations of the Bene Israel community. Values (with standard errors) are based on ALDER analysis with one-reference population. See also S2 and S3 Tables. (C) GLOBETROTTER estimations for Indian (55%) and Jewish (45%) clusters proportions in Bene Israel admixture. The contribution of each population in each cluster is also presented.

https://doi.org/10.1371/journal.pone.0152056.g004

In addition, we performed f₄-based analysis [17] to test whether Bene Israel are closer to Jews than to non-Jewish Middle-Eastern populations (Materials and Methods). We found that Middle-Eastern Jewish populations were closer to Bene Israel as compared to other Middle-Eastern populations examined (Druze, Bedouin and Palestinians). Non-Middle-Eastern Jewish populations were still closer to Bene Israel as compared to Bedouin and Palestinians, but not as compared to Druze (S8 Fig). These results further support the hypothesis that the non-Indian ancestry of Bene Israel is Jewish specific, likely from a Middle-Eastern Jewish population.

In addition, we also applied GLOBETROTTER on our dataset. GLOBETROTTER assigns haplotype segments of the admixed populations to different populations and uses the co-distribution of such segments from different populations to detect and infer admixture. In comparison to ALDER, which focuses on a pair of putative ancestral populations, it considers all populations and assigns ancestry component to all of them simultaneously. Importantly, GLOBETROTTER found evidence for admixture in the similar time range suggested by ALDER –27.7±0.43 generations ago. Admixture proportion estimation was 55% Indian and 45% Jewish. The main contribution from the Jewish cluster was from Greek, Yemenite and Middle-Eastern Jewish populations while the Indian cluster was mainly composed of populations with a high ASI component (Fig 4C). This may suggest that the ancestral Indian population had high ASI component. However, some of the ANI component of the ancestral Indian population may have been captured by the Jewish cluster, and resulting in the Indian side containing a higher ASI component. If the latter is true, GLOBETROTTER proportions estimation for the Jewish side (45%) is an overestimation of the true proportion as it also contains some of the ANI component from the Indian side.

We note here that GLOBETROTTER's original study analyzed, among 95 worldwide populations, "Indian Jews" [26]. However, this group included members from both the Bene Israel and Cochin Jewish communities (four samples from each population [13]), and none of the Jewish populations examined here was used for the analysis. Nevertheless and reassuringly, they reported an admixture event occurring approximately 20 generations ago, with one side being Indian while the other side related to the Middle-East (e.g., South-Italians and Jordanians), which may partially reflect the admixture we report here for Bene Israel.

High endogamy and founder event in Bene Israel

We now turn to examine the post-admixture population structure of Bene Israel. Both Jewish populations [14,23], as well as Indian populations after the ANI-ASI admixture [17,18], show high endogamy. We found that while Jewish populations showed higher IBD sharing as compared to Indian populations (P-value = 4.84e-4, Wilcoxon rank sum test) the Bene Israel population exhibited a level that was almost as twice as high as any other of these populations (Fig 3D). Similarly, Bene Israel exhibits higher total length of homozygous segments (Fig 5A) and lower heterozygosity (S9 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Founder events in Bene Israel population.

(A) Total lengths of runs of homozygosity (ROH) in Jewish, Indian and HapMap populations. The larger variance in ROH values in some Indian populations is due to smaller sample size. (B-C) Autocorrelation in Bene Israel pairs, as a function of the genetic distance, after subtracting the autocorrelation between Bene Israel and other (B) Jewish and (C) Indian populations. Blue and red lines correspond to the fitted curve based on a single and two founder events, respectively.

https://doi.org/10.1371/journal.pone.0152056.g005

These results can suggest not only endogamy but also a genetic bottleneck or a founder event where the contemporary Bene Israel population descended from a small number of ancestors. To directly examine this hypothesis, we used an allele-sharing statistic that measures the autocorrelation of allele sharing between individuals within a population and subtracts the cross-population autocorrelation to remove the ancestral autocorrelation effect. The decay of this statistic with genetic distance can verify if and when a founder event has happened [17,28]. We applied this method to our dataset, using either all Jewish or all Indian populations for cross-population autocorrelation calculation and fitted it to one or two founder events (Fig 5B and 5C). When fitted to a single founder event, analysis suggested it occurred 16 (using Jewish populations) and 14 generations ago (using Indian populations). Fitting to two founder events, which was slightly better, suggested a first founder event 30 generations ago followed by a second event 12 generations ago (using Jewish populations), and a founder event 26 generations ago followed by a second event 9 generations ago (using Indian populations). The first of these two events fits within the timescale of the admixture estimated above and may reflect the founding of this population in the admixing of Jews and Indians. The estimated time of 14–16 generations ago of a single founder event may be the average of these two founder events. The founder event and the genetic drift associated with it are reflected in several other results: by the relatively high F_ST values between Bene Israel and other Indian and Jewish populations (S2 Fig) and by ADMIXTURE analysis which revealed that at K = 8 Bene Israel form their own distinct cluster, though this population sample encompasses only 18 individuals out of 873 (Fig 2).

Bene Israel admixture has been sex-biased

Lastly, we examined whether the ancestry of Bene Israel has been sex-biased using the Q ratio [29]. In a population with equal size of males and females, there are three copies of the X chromosome for every four copies of each autosome and therefore the expected genetic drift on the autosomes is 3/4 of the genetic drift on chromosome X, though this ratio is affected by many additional factors [29–31]. We found a significantly (P-value = 5.11e-6, Wilcoxon rank sum test) lower ratio between Bene Israel and Jewish populations (Fig 6 and S5 Table; mean = 0.58) than between Bene Israel and Indian populations (Fig 6 and S5 Table; mean = 0.73; See also S4 Fig). This entails that the Jewish contribution to Bene Israel has been smaller than otherwise expected for the X chromosome, which points to more male than female Jewish ancestors contributing to the formation of Bene Israel, consistent with findings from previous studies based on Y chromosomal analysis suggesting a paternal link between Bene Israel and Middle-Eastern populations [13]. As most of the Bene Israel samples in our dataset were women, we did not have enough power to analyze the Y chromosome, but mtDNA analysis revealed common Indian haplogroups (M, [32]), consistent with previous studies [8,11,13,15] and the sex-bias we discovered above, while only a few samples had the H haplogroup which is common in Europe and in the Middle-East [33] but also present in lower frequencies in some Indian populations [34] (S6 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Autosome to chromosome X genetic drift ratio between Bene Israel and other populations.

The line represents the 3/4 ratio expected under the null hypothesis of similar demography of males and females in respect to the other populations examined (Jews and Indians, in our case). See also S5 Table.

https://doi.org/10.1371/journal.pone.0152056.g006

Recruitment of Bene Israel individuals

Recruitment of Bene Israel members occurred at Sheba Medical Center in Tel Hashomer, Israel (20 subjects) and at the Indian Bene Israel Community in Ramla, Israel (8 subjects) following approval of the study protocol and consent form by the Sheba Medical Center Helsinki Ethics Committee and the Director General of the Israeli Ministry of Health. All subjects provided written informed consent. The 20 samples collected at Sheba Medical Center were taken from individuals (who came for prenatal or oncogenetic counseling) identifying themselves as Indian Jews, rather than necessarily Bene Israel and therefore were either Cochin Jews or Bene Israel, the two large Indian Jewish communities in Israel. Although all individuals were recruited in Israel, where the vast majority of Indian Jews now live, they did not mix with non-Indian Jewish populations: all individuals recruited reported that their four grandparents belonged to the same Jewish community, similar to other Jewish populations analyzed in the current and in previous works [12,14].

In order to distinguish between Bene Israel and Cochin Jews in the samples collected in Sheba Medical Center, we used the SMARTPCA tool from the EIGENSOFT software [41] and ADMIXTURE [22]. Briefly, we added to these 20 samples additional samples with known population of origin: 8 Bene Israel samples (from Ramla, as described above), 20 Cochin Jews samples (taken from the National Laboratory for the Genetics of Israeli Populations) and 36 Ashkenazi Jews samples [12]. Projection of these samples on the first two principal components (PCs) clearly showed three main clusters: Ashkenazi Jews, Cochin Jews and Bene Israel (S10 Fig). Similarly, ADMIXTURE analysis (with K = 3) suggested three main corresponding clusters (S11 Fig). We labeled an Indian Jewish sample as either Bene Israel or Cochin Jewish if the ADMIXTURE estimated the fraction of the corresponding inferred cluster in this sample was at least 95% and if this was also visibly reflected in the PC analysis.

Following this procedure, eleven Indian Jews from Sheba Medical Center were labeled as Bene Israel. One of the Bene Israel samples was later removed in the quality control (QC) steps performed on the dataset, as described below, resulting in a total 18 Bene Israel samples used in the current analysis.

In addition to that, we repeated PCA (S12 Fig) and ADMIXTURE (S13 Fig) analyses presented in the main text while using only the samples collected in Ramla and not those collected in Sheba Medical Center, obtaining similar results as those described in the main text for all Bene Israel individuals.

Further support to our Bene Israel and Cochin Jews labeling was obtained from another independent source. Previously, Behar et al. [13] genotyped various worldwide Jewish populations, including Cochin Jews and Bene Israel (four members from each of these two populations). We merged our dataset with that dataset which was genotyped on a different platform (Illumina Human610-Quad v1.0 BeadChip), resulting in 50,483 shared SNPs. Projection of the first two PCs showed that our labeling was in accordance with the labeling of their samples (S14 Fig).

Cochin Jews have a different history [2], as also exhibited here in the above analyses and in previous genetic studies as well [11], and therefore they were not included in the current study. However, in parallel to this study, we pursued similar analyses to those described here to Cochin Jews samples. The results of these analyses will be published separately (Waldman, Biddanda, Dubrovsky, Campbell, Oddoux, Friedman, Atzmon, Halperin, Ostrer, Keinan; manuscript in preparation).

Jewish dataset and genotyping

We included in this study a Jewish dataset containing samples from additional 14 Jewish populations, which were collected as described previously [12,14] (S1 Table). All samples, including the Bene Israel samples in this study, were genotyped on the Affymetrix 6.0 array (Affy v6) at the genomic facility at Albert Einstein College of Medicine. Compared to previous studies [12,14], we used an updated version (1.5.5) of the Birdseed tool in the Birdsuite software [42] to recall the genotypes again. Samples with ambiguous gender (based on genotyping) were removed from the analysis. For the current study, where we focused on Bene Israel, we did not include Ethiopian Jews as previous studies have shown that they do not group tightly with other Jewish populations [13,14].

Indian dataset

We also incorporated a dataset of Indian populations based on the study of Reich et al. [17] where the samples were genotyped in the same array (Affymetrix 6.0). We removed Indian populations living in islands and not the mainland (Great Andamanese and Onge), and those defined by Reich et al. [17] as genetic outliers (Aonaga, Nysha and Siddi, Chenchu). In addition, we removed Srivastava from our analysis, as after QC steps only one sample from this population was left, resulting in the inclusion of 18 Indian populations in our analysis (S1 Table). The dataset generated by Reich et al. [17] contained, in addition to the Indian populations, samples from HapMap3 [43] populations and these samples were also used (after QC, described below) for phasing and for some of the analyses (S1 Table).

Human Genome Diversity Project dataset

For some of the analyses we also included data for non-Jewish Middle-Eastern populations (Bedouin, Druze and Palestinian), and nine different Pakistani populations (Kalash, Balochi, Brahui, Makrani, Sindhi, Pathan, Burusho, Hazara and Uygur). Middle-Eastern populations were used to distinguish between Middle-Eastern and Jewish specific genetic attributes while Pakistani populations were taken as populations with high ANI component that reside geographically between India and the Middle-East. We incorporated these samples by merging our dataset with data from the Human Genome Diversity Project (HGDP) [44] genotyped on the Affymetrix GeneChip Human Mapping 500K [19]. This dataset included five unrelated samples from each of these populations, except Makrani and Sindhi with four samples each that passed QC steps (S1 Table).

Dataset merging and Quality Control

After removing SNPs with low call rate (<95%) from the two datasets (Jewish and Indian), we merged them together (including the HapMap samples in the Indian dataset [17]) and removed individuals as follows:

1. Relatives. Following Campbell et al. [14], which analyzed most of the Jewish populations present in the current study, two individuals were considered related if their total autosomal identity by descent (IBD) sharing was larger than 800 cM and if they shared at least 10 segments with length of at least 10 cM (see below how IBD sharing was calculated). To remove as few related individuals as possible while maintaining only unrelated individuals in our dataset, we constructed a graph whose vertices were individuals, and two individuals were connected if they were defined (according to the above criteria) as related. We then tried to find a maximal independent set (i.e., a maximal set of unrelated individuals) in this graph using a greedy algorithm [45].
2. Genetic outliers. We used the SMARTPCA program [41] to detect genetic outliers, with default parameters for genetic outlier removal. We removed individuals further than six standard deviations from the mean in any of the top ten eigenvectors over five iterations. This analysis was done for each population alone, based on autosomal SNPs.

The merged dataset following these QC steps included 513,581 and 25,379 autosomal and X chromosome (in the non-pseudoautosomal regions) single nucleotide polymorphisms (SNPs), respectively, for 461 individuals from 33 Jewish and Indian groups. Additional 844 samples from 11 HapMap3 populations were also available in this dataset, resulting in total 1305 samples. Further merging with the HGDP dataset (for some analyses) consists of 1363 samples with 304,973 shared autosomal SNPs.

The number of samples from each population is shown in S1 Table.

A set of filtered SNPs based on linkage disequilibrium (LD) was used in the following analyses: PCA, F_ST, ADMIXTURE, runs-of-homozygosity and heterozygosity. For each pair of SNPs showing LD of r²>0.5 we considered only one representative (using SMARTPCA's [41] r2thresh and killr2 flags). This filtering was done for each analysis alone, depending on LD in the specific set of populations used in the analysis. Other analyses were performed on the full datasets described above.

Identity-by-descent analysis

We phased the data with the BEAGLE software (version 3.3.2) [46] and extracted shared identity-by-descent (IBD) segments with GERMLINE (version 1.51) [24], using the same parameters described in Campbell et al. [14] which analyzed most of the Jewish populations examined here. To reduce the rate of false positive IBD segments, only segments with length of at least 3 cMs were considered for analysis. Similar to previous studies [12,23], we ignored regions with low informative content. Specifically, using non-overlapping windows (of 1 MB or 1 cM) we ignored all regions with SNP density of less than 100 SNPs per cM or per MB. Genetic positions were obtained from the HapMap genetic map (downloaded from: ftp://ftp.ncbi.nlm.nih.gov/hapmap/recombination/2011-01_phaseII_B37/).

For each pair of unrelated individuals we calculated the total length of autosomal IBD sharing. Given two populations, the average IBD sharing of these two populations was defined as the average IBD sharing between all pairs of individuals from these populations. Similarly, the average IBD sharing within a population was defined as the average IBD sharing between all pairs from this population. In addition, we also calculated the average IBD sharing between a population and the group of all Jewish populations, by averaging the IBD sharing between this population and each of the other Jewish populations. We repeated a similar procedure to calculate average IBD sharing between a population and the group of all Indian populations. As this analysis was done to compare other populations to Bene Israel, the Bene Israel population was not considered to be either Jewish or Indian in this analysis.

We obtained empirical estimations for the distribution of average IBD sharing between populations by sampling 10,000 times ten Bene Israel individuals and measuring the average IBD between these individuals and members of other populations. These distributions were used to obtain standard error estimates for the average IBD sharing between Bene Israel and these populations. In addition, we used these distributions to examine whether Middle-Eastern Jewish populations show higher IBD sharing with Bene Israel as compared to other Jewish populations. This was done by comparing the average IBD sharing distributions of two Jewish populations (one of them being Georgian, Iraqi, Syrian or Iranian Jews) using Wilcoxon rank sum test.

Principal Component Analysis

Principal component analysis (PCA) was performed using the SMARTPCA program [41] on the Jewish, Indian, and several additional worldwide populations. As the number of samples from each population can affect the results of PCA [20] and as there were more samples in each of the Jewish populations as compared to each of the Indian populations, we repeated PCA using not more than four samples from each population selected randomly (using the popsizelimit flag in SMARTPCA).

F_ST

We calculated population differentiation based on differences in allele frequencies between each pair of populations using the F_ST statistic, following the definitions described previously in Reich et al. [17]. Specifically, let p_i be the frequency of a variant in a biallelic SNP in two populations (i = 1,2) and define q_i = 1-p_i. We defined F_ST as where and

To generalize this measure to more than a single SNP, we followed the "ratio of averages" approach [47], where N and D were averaged separately and only then their ratio was taken. Thus, let N_i, D_i be the above definitions for SNP i, then for a set of SNPs S, F_ST was defined as:

Similar to Reich et al. [17], the following estimators where used for N and D: and

Where a_i and b_i are the allele counts of the two alleles, and n_i = a_i+b_i. We calculated F_ST separately for the autosomes and for the X chromosome.

ADMIXTURE

ADMIXTURE [22], a STRUCTURE [48] like algorithm, assigns for each individual its proportion in any of K hypothetical ancestral populations, and therefore can reveal relations between different populations. ADMIXTURE (version 1.2) analysis was performed with default parameters and varying values of K (from K = 3 to K = 10), with 200 bootstrap replicates. We ran ADMIXTURE with the dataset merged with the HGDP populations and included, in addition to the 33 Jewish and Indian populations, the following populations: CEU, YRI, CHB, JPT Druze, Bedouin and Palestinians, and nine Pakistani populations (total 873 unrelated samples). The dataset was filtered based on LD, resulting in 174,421 autosomal SNPs. ADMIXTURE's cross validation procedure was used to determine the K that fits the data best.

Inferring admixture proportions and time

We applied two tools to examine the hypothesis that Bene Israel population was an admixed population and to infer admixture proportions and time: ALDER (version 1.03) [25] (version 1.03) and GLOBETROTTER (downloaded in March 2015) [26]. A detailed description is found in the original publications of these tools, while we provide a brief description in the following.

(1) ALDER: Given a putative admixed population and two surrogate populations taken as a proxy for the presumed ancestral populations, ALDER uses admixture LD statistic to look for evidence for admixture (assuming a single admixture event). For each pair of SNPs ALDER calculates a statistic being the covariance of these two SNPs in the admixed population, weighted by the allele frequency differences between the two reference populations. Exploring the behavior of this admixture LD statistic as a function of the genetic distance between the two SNPs can imply whether the population is admixed or not.

ALDER fits the statistics curve to an exponential function where n is the number of generations since admixture and d is the genetic distance (in Morgans). In addition to the test of admixture using two reference populations, ALDER examines evidence for admixture using only one surrogate population as a reference, with the admixed population serving as a proxy for the second population.

We considered a pair of populations as a candidate for being the ancestral populations for a certain population if all three ALDER results (two one-reference admixture LD and a two-reference admixture LD analyses) were significant and the estimated time of decay was consistent between the three.

In addition to time of admixture, ALDER also estimates admixture proportions from the amplitude of the exponential curve. This is done both in the one-reference version of ALDER (estimating the lower bound of admixture proportion of that population) as well as in the two-reference version. As the populations examined here are taken as a proxy for the true mixing populations, the admixture proportions suggested are lower bounds [25]. A caveat in the two-reference version admixture proportion estimation is that ALDER does not determine to which population to assign the admixture proportion estimation α (i.e., it does not distinguish between α and 1-α). Therefore, we used min(α,1-α) as a lower bound for the admixture proportion of the Jewish population in each significant pair. To determine α from the output of ALDER two-reference population test, f₂ values (representing genetic drift between the two populations [17]) are needed [25] and these were calculated using MixMapper [49].

(2) GLOBETROTTER: Given a putative admixed population and a set of populations (some of them may be a proxy for the presumed ancestral populations), GLOBETROTTER examines whether the putative admixed population is an admixed population of some of the populations from that set [26]. As GLOBETROTTER is based on haplotypes, we phased the data using BEAGLE [46]. GLOBETROTTER algorithm requires several steps. First, the chromosomes of each individual in the admixed population are broken into "chunks" where each chunk is assigned, based on similarity, to a single individual from one of the other populations. This step, implemented in the CHROMOPAINTER [50] tool results in "coloring" of the chromosomes of admixed individuals with different populations. Second, for each pair of populations a curve, which quantifies each genetic distance how often a pair of haplotype chunks separated by this distance come from each pair of populations, is produced. Similar to ALDER, the decay rates of these curves are used to examine whether admixture event happened and to infer its time, while the amplitude of the curve is used to infer the contributing populations and their proportions. In case of evidence for admixture, GLOBETROTTER also examines whether the data fits better single exponential decay (i.e., single admixture event) or a mixture of exponential decays (i.e., several admixture events or continuous admixture over a longer period). In case of admixture, GLOBETROTTER suggests two main clusters of admixture, each may be composed of several populations, which together represent the genetic structure of the ancestral population.

We used CHROMOPAINTER (version 2) and ran GLOBETROTTER on Bene Israel and the Jewish and Indian populations, using 100 bootstrap replicates to obtain standard error estimates for the admixture time.

Time estimates

We converted the number of generations into years by assuming 29 years per generation for such recent history [18,27] and that individuals genotyped in the current study were born circa 1950 CE. Thus, if n is the number of generations since admixture, we convert it to the year 1950-29(n+1) (CE). Changes in generation lengths estimations will scale the time estimations proportionally.

Homozygosity and Heterozygosity estimations

We used PLINK (version 1.07) [51] to identify runs-of-homozygosity (ROH)–autozygous segments in the genome. We used the following flags in PLINK: "—homozyg—homozyg-window-kb 1000—homozyg-window-snp 100—homozyg-window-het 1—homozyg-window-missing 5—homozyg-snp 100—homozyg-kb 1000".

The heterozygosity score of an individual was defined as the fraction of the heterozygous SNPs among all autosomal SNPs (after LD-pruning).

Estimating founder event time

We used allele sharing autocorrelation for estimating time of founder event, along the lines suggested by recent studies [17,28]. Specifically, for each pair of individuals from the population, and for each autosomal SNP, we measure the number of alleles these individuals share: zero, one or two. When both of the individuals are heterozygous for the SNP, we consider them as sharing one allele (to account for haplotype phasing ambiguity). Thus, each SNP is represented by a vector where each entry in the vector corresponds to a pair of individuals and the value of that entry is the number of shared alleles between these two individuals. Next, a Pearson correlation coefficient is calculated between the vectors for each pairs of SNPs (referred as allele sharing autocorrelation). To remove the effect of ancestral allele sharing autocorrelation, we subtract the cross-population allele sharing using this population and a different population. To infer the founder event, we plot the autocorrelation vs. genetic distance and fit the curve to the exponential equation where t represents the number of generations since the founder event and D is the genetic distance (in Morgans) between the two SNPs [17,28].

We applied this method for Bene Israel and calculated allele sharing autocorrelation between each pair of SNPs less than 30 cM apart. We partitioned the values into 0.1 cM bins and considered the mean of each bin. To consider two founder events, we fitted the decay to an equation of the form where t₁ and t₂ were the times (in generations) since the two founder events. Fitting was done by non-linear least squares, using the nls function in R [52]. Evaluation between the single and two founder events models was done by comparing the sum of residuals of each of the models.

Sex-biased population differentiation

To examine sex-biased demography, we calculated a statistic presented by Keinan et al. [29]: It estimates differentiation in allele frequencies (measured by F_ST) between two populations for the autosomes (F_ST^AUTO) and for the X chromosome (F_ST^X) to estimate a ratio:

Q captures the relative genetic drift between the X chromosome and the autosomes. Under several assumptions [29], if the effective population size of males and females has been equal since the two populations split, Q is expected to be 3/4, the ratio of effective population size of the X chromosome to the autosomes in this case. A significant deviation from 3/4 may suggest sex-biased demography since population split.

mtDNA analysis

mtDNA genotypes were used to assign to each of the Bene Israel samples a mtDNA haplogroup based on HaploGrep classification [53].

f₄ analysis

We performed f4 test [17,36] of the type f₄(YRI, Bene Israel; J, ME) where J and ME are Jewish and Middle-Eastern populations, respectively. Assuming YRI is an outgroup for all other populations (Bene Israel, J and ME), If ME is an outgroup in respect to Bene Israel and J, f₄(YRI, Bene Israel; J, ME) is expected to be negative, while if J is an outgroup in respect to Bene Israel and ME, it is expected to be positive [17,36]. For the analysis we also assume that the Indian ancestry of Bene Israel does not affect the results (i.e., the split between the Indian ancestors of Bene Israel and Jewish populations occurred before the split between Jewish and Middle-Eastern populations).