Genotyping of extant and ancient populations has been used

Transcription

1 Maritime route of colonization of Europe Peristera Paschou a,1, Petros Drineas b,1, Evangelia Yannaki c, Anna Razou d, Katerina Kanaki d, Fotis Tsetsos a, Shanmukha Sampath Padmanabhuni a, Manolis Michalodimitrakis d, Maria C. Renda e, Sonja Pavlovic f, Achilles Anagnostopoulos c, John A. Stamatoyannopoulos g, Kenneth K. Kidd h, and George Stamatoyannopoulos g,2 a Department of Molecular Biology and Genetics, Democritus University of Thrace, Alexandroupolis, Greece; b Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY 12180; c Department of Hematology, George Papanicolaou Hospital, Thessaloniki, Greece; d Department of Forensic Medicine, University of Crete, Heraklion, Crete, Greece; e Unità di Ricerca P. Cutino, Ospedali Riuniti Villa Sofia-Cervello, Palermo, Italy; f Institute of Molecular Genetics and Genetic Engineering, University of Belgrade, Belgrade, Serbia; g Departments of Medicine and Genome Sciences, University of Washington, Seattle, WA 98195; and h Department of Genetics, Yale University School of Medicine, New Haven, CT Edited* by Yuet Wai Kan, University of California, San Francisco School of Medicine, San Francisco, CA, and approved May 9, 2014 (received for review November 7, 2013) The Neolithic populations, which colonized Europe approximately 9,000 y ago, presumably migrated from Near East to Anatolia and from there to Central Europe through Thrace and the Balkans. An alternative route would have been island hopping across the Southern European coast. To test this hypothesis, we analyzed genome-wide DNA polymorphisms on populations bordering the Mediterranean coast and from Anatolia and mainland Europe. We observe a striking structure correlating genes with geography around the Mediterranean Sea with characteristic east to west clines of gene flow. Using population network analysis, we also find that the gene flow from Anatolia to Europe was through Dodecanese, Crete, and the Southern European coast, compatible with the hypothesis that a maritime coastal route was mainly used for the migration of Neolithic farmers to Europe. Genotyping of extant and ancient populations has been used to address the question of the origins of the people of Europe. The genome of the present-day Europeans reflects merging of the Paleolithic settlers who colonized Europe 35,000 40,000 y before the present era (BPE) and the Neolithic people who started colonizing Europe approximately 9,000 y BPE. The Neolithic contribution to the gene pool of modern Europeans has been estimated with studies of extant European populations by using mitochondrial DNA, Y-chromosomal DNA, or nuclear DNA polymorphisms. Mitochondrial DNA studies estimate the Neolithic contribution to the maternal lineages of the modern Europeans to range between 10 and 20% (1). A contribution of approximately 22% was suggested by a study of Y-chromosome polymorphisms, which also found that the Neolithic contribution was more pronounced along the Mediterranean coast (2). Neolithic contributions of 50 70% were estimated with other methodologies (3 5), including highly polymorphic DNA markers (6). Clinal patterns of genetic diversity of autosomal (7 9) or Y-chromosomal (10) polymorphisms across Europe suggest that the Neolithic migrants originated from the Near East (7 9). It has been proposed that these Near Eastern migrants brought to Europe their new agricultural technologies (7 9, 11) and, perhaps, the Indo-European language (12). How did these Neolithic people reach Europe from the Near East? The geographic focus of the transition from foraging to the Neolithic way of life was the Levantine corridor, which extended from the Fertile Crescent to the southeastern sections of the central Anatolian basin (13). The Neolithic farmers could have taken three migration routes to Europe. One was by land to North-Eastern Anatolia and from there, through Bosporus and the Dardanelles, to Thrace and the Balkans (14, 15). A second route was a maritime route from the Aegean Anatolian coast to the Mediterranean islands and the coast of Southern Europe (12, 14 18). The third was from the Levantine coast to the Aegean islands and Greece (19). Navigation across the Mediterranean was active during the Early Neolithic and Upper Paleolithic (16 18) as illustrated by the finding of obsidian from the island of Milos in Paleolithic sites of the Greek mainland (19, 20) and the early colonization of Sardinia, Corsica, and Cyprus (18, 21 23). If a maritime route was used by the Neolithic farmers who settled Europe, their first stepping stones into Europe were the islands of Dodecanese and Crete. The Dodecanese is very close to the Aegean coast of Anatolia, whereas the west-most Dodecanesean islands are very close to Crete. Crete hosts one of the oldest Neolithic settlements of Europe in the site of Knossos, established 8,500 9,000 y BPE (24, 25), and the inhabitants of the island established the first advanced European civilization starting approximately 5,000 BPE. To obtain insights on the question of migrations to Europe, we analyzed genome-wide autosomal single nucleotide polymorphisms (SNPs) from a dataset of 32 populations. This dataset includes population samples from the islands of Crete and Dodecanese, one from Cappadocia in Central Anatolia, three subpopulations from different regions of mainland Greece, 14 other populations from Southern and Northern Europe, five populations from the Near East, and seven from North Africa. In addition to established methods for genetics analysis, we use a population genetics network approach that can define pathways of gene flow between populations. Our data are compatible with the hypothesis that a maritime route connecting Anatolia and Southern Europe through Dodecanese and Crete was the main route used by the Neolithic migrants to reach Europe. Results Genes Mirror Geography Across the Mediterranean Basin. We first used principal components analysis (PCA) to visualize the genotypic distances between studied populations (Fig. 1; also see SI Appendix, Significance The question of colonization of Europe by Neolithic people of the Near East and their contribution to the farming economy of Europe has been addressed with extensive archaeological studies and many genetic investigations of extant European and Near Eastern populations. Here, we use DNA polymorphisms of extant populations to investigate the patterns of gene flow from the Near East to Europe. Our data support the hypothesis that Near Eastern migrants reached Europe from Anatolia. A maritime route and island hopping was mainly used by these Near Eastern migrants to reach Southern Europe. Author contributions: G.S. designed research; P.P., P.D., E.Y., A.R., K.K., M.M., M.C.R., S.P., A.A., and K.K.K. performed research; E.Y., A.R., K.K., M.M., M.C.R., S.P., and A.A. performed population studies; K.K.K. contributed data; P.P., P.D., F.T., and S.S.P. analyzed data; and P.P., P.D., J.A.S., and G.S. wrote the paper. The authors declare no conflict of interest. *This Direct Submission article had a prearranged editor. Freely available online through the PNAS open access option. 1 P.P. and P.D. contributed equally to this work. 2 To whom correspondence should be addressed. gstam@u.washington.edu. This article contains supporting information online at /pnas /-/DCSupplemental. GENETICS PNAS June 24, 2014 vol. 111 no

2 Figs. S1 and S2 and Tables S1 and S2). Populations on the southern and northern coasts of the Mediterranean, appear to be separated by the geographic barrier of the Mediterranean Sea. The role of the Mediterranean Sea as a barrier for gene flow among populations was also supported by our analysis using the BARRIER software (26), which implements Monmonier s maximum difference algorithm (SI Appendix, SI Methods and Results and Fig. S3). In accordance with this finding, notice, in Fig. 1B, that the PCA distribution of the populations closely resembles the geographic map of the countries circling the Mediterranean Sea. On this PCA map of populations, the east coast of the Mediterranean Sea is appropriately occupied by the Palestinians and the Lebanese Druze. Yemenites and Bedouins branch out from the Mediterranean populations and are closer to the populations of the Near East. Fig. 2 further illustrates the considerable resemblance of the PCA projection of the genotypes on the 2D space to the geographic map of the European Mediterranean coast. The east to west cline from Near East and Anatolia across the southern Mediterranean coast fits with the hypothesis of early population movements from the Near East to Europe (7 9). The populations of the European Mediterranean Coast connect with the Near East although Anatolia (Cappadocia). In fact, the closest populations to Anatolia are those of Crete and Dodecanese rather than the populations of the Balkans or Northern Greece. When considering Europe, Anatolia, and the Near East using PCA (Fig. 1C), a clear gradient is again observed, with populations from Northern and Central Europe connecting to Anatolia and the Near East via Southern Europe and through the bridge of the islands of Dodecanese and Crete. Three population tests (f3 statistics) as described in ref. 27 did not provide any evidence of bidirectional admixture or population splits along this line of stepping stones connecting Anatolia to Southeastern Europe (SI Appendix, SI Methods and Results). The correlation of geographic coordinates of the Mediterranean populations to the top two principal components (as shown in SI Appendix, Fig.S2D) is very high; the Pearson correlation coefficient is 0.66 between latitude (north to south axis) and the first principal component, and 0.81 between longitude (east to west axis) and the second A B Yemenites Bedouin Druze EigenSNP 2 Cappadocia Crete Southern Europe Sardina Palestinians Egypt Libya Northern Africa * * Basque * C Northern Europe Southern Europe Crete Cappadocia Central Europe Libya Egypt Palestinians Druze Northern Africa Algeria Basque Bedouin Chuvash Crete Denmark Dodecanese Druze East Rumelia Egypt Finland France Hungary Ireland Italy Cappadocia Libya Macedonia North Morocco South Morocco Mozabite SE Laconia Palestine Peloponnese Russia Sahara OCC Sardinia Serbia Sicily Tunisia Tuscany Yemenite EigenSNP 1 Fig. 1. Genes mirror geography around the Mediterranean coast. (A) Geographic distribution of the populations included in this study. (B) Projection on top two principal components of samples from 25 populations genotyped on 75,000 genome-wide autosomal SNPs. A clear cline is observed with Anatolia (Cappadocia) connected to Southern Europe through the bridge of the islands of the Dodecanese and Crete. Bedouins and Yemenites drift toward Central- South Asia. No apparent gene flow between Northern Africa and the Southern coast of Europe is observed. (C) Projection on top two principal components of samples from 30 populations genotyped on the same set of SNPs presented in B. Northern European populations have now been added; Bedouins and Yemenites were removed. The cline now continues through Central and Northern Europe Paschou et al.

3 Fig. 2. Genetic structure of populations along the Southern European Coast in relation to Anatolia. (A) PCA plot of 10 populations from the Southern European coast and Cappadocia. The first principal component reveals the East to West cline in genetic variation along the Southern Coast of Europe and Mediterranean islands. Basques and Sardinians appear isolated relatively to the remaining studied populations. (B) Structure of the Southern European populations and Cappadocia excluding the more remote Basques and Sardinians. principal component (P < 10 5 as defined by Mantel test, see SI Appendix, Table S3 for geographic coordinates of the populations; this analysis does not even rotate or rescale the data, as was done in ref. 28). Thus, the PCA analysis supports a migration pathway from Anatolia to Europe through island hopping in the Aegean Sea. To further analyze the relationships between the populations of the Mediterranean basin through an independent methodology, we used ADMIXTURE, an unsupervised ancestry-inference algorithm (Fig. 3; also see SI Appendix,Fig.S4andS5). ADMIXTURE analysis on an LD-pruned dataset of 72,951 SNPs leads to essentially the same conclusions as PCA. The Near East and Anatolia are connected to Southern Europe through Crete and Dodecanese supporting the hypothesis of a pathway to Europe through the coastal Mediterranean route. ADMIXTURE analysis also illustrated the distinction between the populations of the North African coast and those of the European Mediterranean coast, and their connection through the populations of the Near East (Fig. 3). Thus, both the results of the PCA and the ADMIXTURE analysis are compatible with a maritime route of migrations from Anatolia/Near East to Southern Europe, in which Crete and the Dodecanese were used as the stepping stones by the migrating populations. Network Analysis to Identify Pathways of Gene Flow Between Populations. To further test hypotheses on the routes of colonization of Southern Europe, we developed an approach for network-based population analysis. In doing so, we essentially attempted to reconstruct the pathways of gene flow between Near East, Anatolia, Mediterranean populations, and Northern/Central European populations as captured by PCA or ADMIXTURE. The goal was the creation of networks connecting the populations under study. To form a network of related populations, we leveraged the fact that both PCA and ADMIXTURE essentially reduce the dimensionality of the original dataset by expressing each sample as either (i) a linear combination of the top few eigenvectors (in the case of PCA) or (ii) as percentages of ancestry from a small number of typically unknown ancestral populations (in the case of ADMIXTURE). To be more precise, in our study, each individual sample is originally described with respect to 75,194 SNPs; mathematically, this observation is equivalent to saying that the sample lies in a 75,194-dimensional subspace. After applying PCA or ADMIXTURE to the dataset, all samples are described with respect to K coefficients. K ranges between one and eight in all our results. Thus, the output of PCA or ADMIXTURE lies in a K-dimensional subspace, with K << 75,194. In the case of PCA, these coefficients correspond to the projections of each sample in the top K principal components (SI Appendix, Fig. S6), whereas in the case of ADMIXTURE, the coefficients correspond to the percentages for ancestry of each sample with respect to the K ancestral, but unknown, populations. It is worth indicating that PCA and ADMIXTURE are very different statistical techniques: PCA assumes the existence of a small number of pairwise orthogonal components that explain the variance in the data, whereas ADMIXTURE is a model-based ancestry inference technique. (See SI Appendix for more details and a discussion of appropriate measures of distance for the low-dimensional PCA and ADMIXTURE subspaces.) To form the networks, we first identified the top 10 nearest neighbors of each individual in the PCA or ADMIXTURE K-dimensional subspace, with the additional constraint that these nearest neighbors must not lie in the population of origin of the target individual. More specifically, we start by computing the l 2 (Euclidean) distance (in the case of PCA) or the l 1 (total variation) distance (in the case of ADMIXTURE) between the target individual and every other individual in our dataset. (See SI Appendix for a precise definition of the distance metric.) This procedure is repeated for all individuals in the dataset. Intuitively, our measure of distance dist(x,y) between population X and population Y is the number of nearest neighbors that individuals in population X have in population Y. (A minor fine tuning can be used to mitigate the effect of different population sizes in our sample; see SI Appendix for details). The network formation algorithm can be described as follows: For each pair of populations X and Y, we compute both distances, dist(x,y) and dist(y,x). We create an edge between X and Y (thus claiming that the two populations are neighbors of each other) if min {dist(x,y),dist(y,x)} > 0, and we assign as a weight of the respective edge the value min{dist(x,y),dist(y,x)}. It is worth noting that our choice is quite conservative, because we assume that populations X and Y are related if, and only if, both X and Y have nearest neighbors in each other. Population Network Analysis Supports a Maritime Path for the Colonization of Europe. Results of the network formation algorithm based on PCA are visualized with the Cytoscape software (Fig. 4A). In this figure, an edge between two populations shows that the two populations share genetically related individuals, with thicker edges indicating a larger number of genetically close individuals. The resulting networks (Fig. 4A; also see SI Appendix, Figs. S6 and S7) clearly indicate a path from the Near East populations (Palestinian, Druze) to Anatolia (Cappadocia), and from there to the islands of Dodecanese, and Crete. The connections between Crete and the rest of Greece (South East- Laconia Peloponnese Macedonia) as well as the populations of Sicily and Italy are evident. Analyses were performed by using the top three to seven PCs (SI Appendix, Figs. S6 and S7) and results were robust regardless of the number of principal components used. The geographic proximity and partial overlap in the PCA of Crete and Sicily is also compatible with gene flow from Crete to Italy and to Southern Europe through population movements along the Southern Mediterranean coast. Results of our network analysis remain robust even when the independent methodology of ADMIXTURE is used to infer population distances. The resulting networks based on ADMIXTURE GENETICS Paschou et al. PNAS June 24, 2014 vol. 111 no

4 Centrality Statistics Support the Role of Islands in Connecting Anatolia to Europe. As part of our network analysis using Cytoscape, we also computed centrality statistics of the nodes in our network. The color of the network nodes in Fig. 4 denotes the socalled closeness centrality with warmer colors representing more important nodes. Crete and Dodecanese are among the most central nodes in the network. More specifically, the closeness centrality is a classical network metric indicating the total distance from a node in a network to all other nodes. It is often regarded as a measure of how long it will take to spread information (in this case, genetic material) to other nodes in the network in a sequential manner. Crete is always the highest or the second highest ranking node in all networks that we formed with respect to this statistic, and Cappadokia is always among the top five nodes. Similarly, if we consider the so-called betweenness centrality, which quantifies the number of times a node in the network acts as a bridge along the shortest path between two other nodes, Crete and Cappadocia share the number one and two spots in almost all networks that we formed. East Rumelia, the Peloponnese, and the Dodecanese also rank among the top 12 nodes in most of the networks that we formed, indicating the pivotal role of the populations that were collected and included in our analysis in the migration from Near East and Anatolia to Europe. Fig. 3. Population structure around the Mediterranean basin. A modelbased, unsupervised ancestry analysis approach (ADMIXTURE) was used to analyze populations on 72,951 (LD-pruned) genome-wide autosomal SNPs (K = 2 8). Two separate East to West clines are observed along the Northern African and Southern European Mediterranean coasts. outputs are shown in Fig. 4B (also see SI Appendix, Fig. S8). Both the PCA and the ADMIXTURE-based networks tell a consistent story and the pathway of gene flow from Anatolia to Southern Europe through the islands of Crete and Dodecanese is apparent. Results are robust, both with respect to the fact that two different approaches have been used (PCA and ADMIXTURE) and with respect to the dimensionality reduction parameter K, which varies between 3 and 8 for ADMIXTURE, and 3 and 7 for PCA. Network analysis using the Fst distance metric (SI Appendix,Fig.S9)does not provide very high resolution, but is similar in patterns to the networks obtained based on ADMIXTURE and PCA distances. Simulations of a stepping stone model of migrations around the Mediterranean using IBDSim (29) (SI Appendix, SI Methods and Results) and investigation of simulated genotypes through PCA and our population network analysis (SI Appendix, Figs. S10 and S11 and Table S4), provide additional support to our hypothesis and leads to patterns that are strikingly similar to the actual observed patterns and relationships among populations, with Crete acting as a hub that connects Anatolia to Southeastern Europe. Phylogenetic Analysis. We constructed a rooted neighbor-joining tree for our sample of 32 populations from North Africa, the Near East, Anatolia, and Southern and Northern Europe. The phylogenetic tree (SI Appendix, Fig. S12) was calculated from a matrix of Fst distances among populations, using the algorithms implemented in PHYLIP and SmartPCA. Results are compatible with the clusters identified through PCA and ADMIXTURE analysis, and the identified lineages are largely concordant with the findings from our population network analysis. The populations from North Africa form a separate cluster branching out from South-West Asia. The closest European branches to the population of Cappadocia from Anatolia are those representing the islands of the Dodecanese and Crete. In turn, Crete connects to the most southern tip of mainland Greece (Southeast Laconia) as well as Sicily and the remaining populations from Northern and Central Europe that form a separate cluster. Thus, phylogenetic analysis also points to the central role of Crete and the Dodecanese in connecting Anatolia to Southern Europe. Similar results are obtained through analysis of our dataset with TreeMix (30) and Neighbor-Net (31). Both algorithms aim to construct a phylogenetic graph, improving the fit of a simple tree by allowing more than one path between populations. TreeMix and Neighbor-Net graphs underline gene flow from Anatolia to Southern Europe through the islands of the Dodecanese and Crete, without evidence for additional migrations from Anatolia through the Balkans (SI Appendix, Figs. S13 and S14 and Table S5). Discussion In historical times, there have been three major invasions of South Eastern Europe from the direction of the Near East but no evidence of major migratory events and gene flow. The Persians dominated South Western Asia in the fifth century BC: They established satrapies in Asia Minor and invaded Europe, but they were stopped by the Greeks (32). The Arabs attempted multiple invasions during the seventh and eighth centuries AD, but they were stopped by the Byzantines (33). An Arab tribe originating from Andalusia established in Crete a pirate state in the ninth century, but they were exterminated by the Byzantines 140 y later, and they left no traces of settlement in the island other than the name of their seat of power in the town of Chandax (33). The Turks invaded Asia Minor starting the 11th century and occupied the Balkans in the subsequent three centuries, but any Turks and converts to Islam left from Greek territories with the population exchanges that took place in the 20th century (34); the origin of the Turkish tribes was the central Asia. Seljuk Turks settled in Anatolia in the 12th century AD; Paschou et al.

5 A Mozabite North_Morocco Algeria Egypt South_Morocco Libya Sahara_OCC B Palestinian Druze Tunisia Mozabite N_Morocco Algeria Palestinian S_Morocco Libya Sahara_OCC Egypt Druze Italy Chuvash Dodecanese Sicily East Serbia Tuscany Rumelia Hungary Crete Russia Finland Cappadocia SE_Laconia Macedonia France Denmark Peloponnese Ireland Cappadocia Basque Sardinia Dodecanese Italy Russia Finland Sicily Tuscany Crete East Hungary Chuvash Rumelia SE_Laconia Serbia Denmark Peloponnese Macedonia France Ireland Basque Sardinia Fig. 4. A coastal route of colonization of Europe. (A) A network analysis and visualization of the connections among 30 populations as revealed by the top five principal components. The network was formed by identifying nearest neighbors of each individual outside its populations of origin. Thicker edges represent stronger genetic relationships between pairs of populations, whereas warmer colors indicate high centrality of the respective nodes. The route connecting North Africa, Middle East, and Anatolia via the islands of the Dodecanese, and Crete to the rest of Europe, is apparent. (B) A network analysis and visualization of connections among 30 populations as revealed by ADMIXTURE with K = 5. Results are very similar to those in Fig. 4A, despite the fact that Admixture is a very different technique to extract ancestry information. Our networks are robust to the use of additional principal components or larger values of the ADMIXTURE parameter K for their formation (SI Appendix, Figs.S7andS8). however, the Anatolian Cappadocians we included in this study belong to the population that have kept the religion and the language of the pre-seljuk Cappadocians and, therefore, most likely carry the genetic makeup of the ancient Anatolians. The only important gene flows from Near East to Europe must have occurred in prehistoric times and, as genetic evidence suggests, the most prominent migrations should have occurred during the Neolithic. The idea that the Neolithic was introduced to Europe through coastal routes of colonization has been proposed by several archaeologists (12, 16, 17, 19, 22, 35). The earliest Neolithic sites with developed agricultural economies in Europe dated BPE are found in Greece (19, 36, 37). The general features of material culture of the Greek Neolithic (14, 19, 36) and the genetic features of the preserved crops and associated weeds of the earliest Greek Neolithic sites point to Near Eastern origins (38). How these Near Eastern migrants reached Greece is a matter of speculation. One route of migration was by land from Central to Northeast Anatolia and from there to Southern Balkans through Bosporus, the Dardanelles, and Thrace (14, 15, 39). This migration route is less likely because archaeological evidence (19, 36, 40, 41) including 14 C dating (19, 40, 41) suggests that the Neolithic sites in Thrace and Macedonia are younger than those of mainland Greece, an unexpected finding if the Neolithic migrants who colonized Greece arrived there from the north. Other models suggest that waves of the Near-Eastern migrants reached Greece by sailing either from the Aegean Anatolian coast (12, 14, 16, 17, 22, 35) or from the Levantine coast (19, 36). Our data support the Anatolian rather than the Levantine route because they consistently show the Aegean islands to be connected to the Near East through Anatolia. Archaeological evidence from Greek and Near Eastern and Anatolian Neolithic sites suggests that multiple waves of Neolithic migrants reached Greece and Southern Europe. Most likely multiple routes were used in these migrations but, as our data show, the maritime route and island hopping was prominent. Our findings also suggest that to the west of Greece, the Neolithic reached Sicily and Italy by sea, as it has been suggested by archaeologists (12, 42). Studies of extant European and Near Eastern populations using multiple autosomal genetic polymorphisms have established the presence of clinal distributions of allelic frequencies (4 10, 43, 44). These clines in gene frequencies have been attributed to the geographically gradual merging of the gene pools of the Neolithic Near East migrants with the gene pools of the existing Paleolithic population of Europe. The correlation of clinal gene frequencies with the archaeological record of the spread of agriculture in Europe lead to the suggestion that it was the migration of Neolithic populations from the Near East that led to the spread of agriculture in Europe (7). The underlying hypothesis is that the development of agriculture triggered marked population growth and produced demographic pressures that resulted in dispersion of the Neolithic populations to new regions (7 9, 11). The rate of dispersion from the Near East to Western Europe has been estimated to approximately km/y (44). A faster rate of dispersion is expected if maritime routes were used for the colonization of Southern Europe. Indeed, archaeological evidence suggests that farming spread faster in Southern Europe (12, 42, 45) and radiocarbon measurements in Neolithic sites are compatible with very rapid colonization of the west Mediterranean by Neolithic migrants (46, 47). Although the Southeastern Mediterranean islands seem to have acted as a bridge from Anatolia to Southern Europe, the relatively small degree of gene flow between the African and the European coasts shows that the Mediterranean Sea also had a barrier function as also suggested with studies of mtdna polymorphisms (48). Thus, the Mediterranean seems to have facilitated the migrations of Neolithic farmers along its Southern European coast but it mostly acted as an isolating factor between its European and African coasts. Materials and Methods Samples. We collected a total of 202 samples from nine populations that were genotyped on two different platforms (SI Appendix, Table S1). In our sample collection process from the Greek subpopulations, we extracted DNA from blood samples of individuals that were at least 70 y old and self-reported that all four grandparents originated from the target population. We expect that because of our sample selection process, our data reflect the genetic structure of the Greek subpopulations four generations before present. We combined our data with four additional datasets to study population structure around the Mediterranean basin as well as Northern Europe. Thus, we produced a dataset of 964 samples from 32 populations, genotyped on 75,194 SNPs. More specifically, we used additional data from (i) the Human Genome Diversity Panel (49), (ii) the HapMap Phase III Project (50), (iii) publicly available data on Northern African populations that were first GENETICS Paschou et al. PNAS June 24, 2014 vol. 111 no

6 released by Henn and coworkers (51), and (iv) data from the Kidd Laboratory at Yale University (allele frequencies for these data are available via the ALFRED database) (SI Appendix, Table S2). PCA and ADMIXTURE. We used our own MatLab implementation of PCA (52, 53) (see SI Appendix for details). Before running ADMIXTURE, we pruned the SNPs to remove SNPs in high LD by using a windowed approach and a value of r 2 equal to 0.8. Correlation Between PCA and Geographic Coordinates. We estimated the correlation between geographic coordinates (SI Appendix, Table S3) and the top two eigenvectors emerging from PCA. For each population in our sample, we approximated its location of origin either using information provided to us by the individuals that collected the respective sample, or by using a capital city that is relatively close to the population under study. The correlation between geographic coordinates and the eigenvectors was computed by converting both the geographic coordinates vector and the eigenvectors to z scores, and then computing the Pearson correlation coefficient. A Mantel test was run to estimate statistical significance. Network Analysis. To better understand the connection between the populations included in our study, we performed a network analysis on the results of PCA and ADMIXTURE. To form the networks, we identified the top few nearest neighbors of each sample by representing each sample with respect to the top K coefficients returned by PCA or ADMIXTURE, and then computing the distance of each sample to all other samples, under the additional constraint that these neighbors should not belong to the same population of origin as the sample itself. Once a network whose nodes correspond to populations and whose edges correspond to connections between populations, as described above, is formed, we visualize it by using the Cytoscape software package (see SI Appendix for details). ACKNOWLEDGMENTS. We thank F. Sakellaridi, E. Papadaki, I.Adamopoulos, K. Farmaki, M. Tsironis, A. Mariolis, and P. Kaloyannidis for their assistance during the field study; A. Papadopoulou, N. Psatha, and N. Zogas for technical assistance; and N. Patterson and I. Lazaridis for helpful discussions. This work partially was supported by National Institutes of Health grants (to G.S. and K.K.) and National Science Foundation grants (to P.D.) and cofunded by the European Union (European Social Fund) and National Resources under the Operational Programme Education and Lifelong Learning Action GENOMAP.GR, ARISTEIA II Programme, NSRF (to P.P.). 1. Richards M, et al. (2000) Tracing European founder lineages in the Near Eastern mtdna pool. Am J Hum Genet 67(5): Semino O, et al. (2000) The genetic legacy of Paleolithic Homo sapiens sapiens in extant Europeans: A Y chromosome perspective. Science 290(5494): Barbujani G, Bertorelle G (2001) Genetics and the population history of Europe. Proc Natl Acad Sci USA 98(1): Cavalli-Sforza LL, Minch E (1997) Paleolithic and Neolithic lineages in the European mitochondrial gene pool. Am J Hum Genet 61(1): Chikhi L, Nichols RA, Barbujani G, Beaumont MA (2002) Y genetic data support the Neolithic demic diffusion model. Proc Natl Acad Sci USA 99(17): Belle EMS, Landry P-A, Barbujani G (2006) Origins and evolution of the Europeans genome: Evidence from multiple microsatellite loci. Proc Biol Sci 273(1594): Menozzi P, Piazza A, Cavalli-Sforza L (1978) Synthetic maps of human gene frequencies in Europeans. Science 201(4358): Ammerman AJ, Cavalli-Sforza LL (1984) The Neolithic Transition and the Genetics of Populations in Europe (Princeton Univ Press, Princeton, NJ), pp Sokal RR, Oden NL, Wilson C (1991) Genetic evidence for the spread of agriculture in Europe by demic diffusion. Nature 351(6322): Rosser ZH, et al. (2000) Y-chromosomal diversity in Europe is clinal and influenced primarily by geography, rather than by language. Am J Hum Genet 67(6): Childe VG (1950) The Dawn of European Civilization (Routledge and Kegan Paul, London), pp Renfrew C (1987) Archaeology and Language (Jonathan Cape, London), pp Harris DR (1996) The Origins and Spread of Agriculture and Pastoralism in Eurasia (UCL Press, London), pp Hammond N (1976) Migrations and Invasions in Greece and Adjacent Areas (Noyes Press, Park Ridge, NJ), pp Özdogan M (2011) Archaeological evidence on the westward expansion of farming communities from eastern Anatolia to the Aegean and the Balkans. Curr Anthropol 52(S4):S415 S Broodbank C (2006) The origins and early development of Mediterranean maritime activity. J Mediterr Archaeol 19(2): Davison K, Dolukhanov P, Sarson GR, Shukurov A (2006) The role of waterways in the spread of the Neolithic. J Archaeol Sci 33(5): Cherry JF (1981) Pattern and process in the earliest colonization of the Mediterranean Islands. Proc Prehist Soc 47: Perlès C (2001) The Early Neolithic in Greece (Cambridge Univ Press, Cambridge, UK), pp Runnels CN (1995) Review of Aegean prehistory IV: The stone age of Greece from the Palaeolithic to the advent of the Neolithic. Am J Archaeol 99(4): Vigne J-D, et al. (2012) First wave of cultivators spread to Cyprus at least 10,600 y ago. Proc Natl Acad Sci USA 109(22): Cherry JF (1990) The first colonization of the Mediterranean islands: A review of recent research. J Mediterr Archaeol 3(2): Knapp AB (2013) The Archaeology of Cyprus (Cambridge World Archaeology, New York), pp Efstratiou N, Karestsou A, Banou ES, Margomenou D (2004) Knossos: Palace, City, State, eds Cadogan G, Hatzaki E, Vasilakis A (British School at Athens, London), pp Evans JD (1994) Knossos: Labyrinth of History, eds Evely D, Hughes-Brock H, Momigliano N (British School at Athens, London), pp Manni F, Guérard E, Heyer E (2004) Geographic patterns of (genetic, morphologic, linguistic) variation: How barriers can be detected by using Monmonier s algorithm. Hum Biol 76(2): Patterson N, et al. (2012) Ancient admixture in human history. Genetics 192(3): Novembre J, et al. (2008) Genes mirror geography within Europe. Nature 456(7218): Leblois R, Estoup A, Rousset F (2009) IBDSim: A computer program to simulate genotypic data under isolation by distance. Mol Ecol Resour 9(1): Pickrell JK, Pritchard JK (2012) Inference of population splits and mixtures from genome-wide allele frequency data. PLoS Genet 8(11):e Bryant D, Moulton V (2004) Neighbor-net: An agglomerative method for the construction of phylogenetic networks. Mol Biol Evol 21(2): Herodotus (1925) (Harvard Univ Press, Cambridge, MA). 33. Vasiliev AA (1952) History of the Byzantine Empire (Univ Wisconsin Press, Madison, Madison, WI), Vol 1, pp , (1925) Greece and Turkey - Convention concerning the exchange of Greek and Turkish populations and protocol. League of Nations Treaty Series 32: Broodbank C, Strasser TF (1991) Migrant farmers and the Neolithic colonization of Crete. Antiquity 65(247): Demoule J-P, Perlès C (1993) The Greek Neolithic: A new review. J World Prehist 7(4): Van Andel TH, Runnels CN (1995) The earliest farmers in Europe. Antiquity 69(264): Colledge S, Conolly J, Shennan S (2004) Archaeobotanical evidence for the spread of farming in the eastern Mediterranean. Curr Anthropol 45(Supp):S35 S Özdogan M (1997) The beginning of Neolithic economies in southeastern Europe: An Anatolian perspective. J Eur Archaeol 5(2): Cauwe N, Dolukhanov P, Kozlowski J, Van Berg PL (2007) Le Néolithique en Europe (Armand Collin, Paris), pp Andreou S, Fotiadis M, Kotsakis K (1996) Review of Aegean Prehistory V: The Neolithic and Bronze Age of Northern Greece. Am J Archaeol 100(3): Phillips P (1975) Early Farmers of West Mediterranean Europe (Hutchinson Univ Library, London), pp Sokal R, Menozzi P (1982) Spatial autocorrelations of HLA frequencies in Europe support demic diffusion of early farmers. Am Nat 119(1): Pinhasi R, Fort J, Ammerman AJ (2005) Tracing the origin and spread of agriculture in Europe. PLoS Biol 3(12):e Whittle A (1994) The first farmers. The Oxford Illustrated Prehistory of Europe, ed Cunliffe B (Oxford Univ Press, Oxford), pp Zilhão J (2001) Radiocarbon evidence for maritime pioneer colonization at the origins of farming in west Mediterranean Europe. Proc Natl Acad Sci USA 98(24): Zilhao Z (2000) From the Mesolithic to the Neolithic in the Iberian Peninsula. Europe s First Farmers, ed Price TD (Cambridge Univ Press, Cambridge, UK), pp Simoni L, Gueresi P, Pettener D, Barbujani G (1999) Patterns of gene flow inferred from genetic distances in the Mediterranean region. Hum Biol 71(3): Li JZ, et al. (2008) Worldwide human relationships inferred from genome-wide patterns of variation. Science 319(5866): The International HapMap 3 Consortium (2010) Integrating common and rare genetic variation in diverse human populations. Nature 467(7311): Henn BM, et al. (2012) Genomic Ancestry of North Africans Supports Back-to-Africa Migrations. PLoS Genetics 8(1):e Paschou P, et al. (2007) PCA-correlated SNPs for structure identification in worldwide human populations. PLoS Genet 3(9): Paschou P, et al. (2007) Intra- and interpopulation genotype reconstruction from tagging SNPs. Genome Res 17(1): Paschou et al.

7 Supplementary Material: Data New samples: We collected a total of 202 samples from nine populations that were genotyped on two different platforms (see Supplementary Table 1 below for details). In our sample collection process, we extracted DNA from blood samples of individuals that were at least 70 years old and self-reported that all four grandparents originated from the target population. We also took care to avoid relatives in our sampling. We expect that due to our sample selection process our data reflect the genetic structure of the target populations four generations before present. Supplementary Table 1: New samples collected and genotyped in our work. Sample size indicates the number of individuals in our study. Population Region Sample Size (N) Genotyping Platform Cappadocia Anatolia 10 Illumina OMNI 2.5 (2,379,855 SNPs) Crete South Europe 90 Illumina OMNI1-QUAD (1,134,514 SNPs) Dodecanese South Europe 10 Illumina OMNI 2.5 (2,379,855 SNPs) East Rumelia South Europe 12 Illumina OMNI 2.5 (2,379,855 SNPs) Macedonia South Europe 16 Illumina OMNI 2.5 (2,379,855 SNPs) Peloponnese South Europe 19 Illumina OMNI 2.5 (2,379,855 SNPs) Serbia South Europe 20 Illumina OMNI 2.5 (2,379,855 SNPs) Sicily South Europe 20 Illumina OMNI 2.5 (2,379,855 SNPs) South East Laconia South Europe 5 Illumina OMNI 2.5 (2,379,855 SNPs) 1

8 Additional Datasets: We combined our data with four additional datasets in order to study population structure around the Mediterranean basin as well as Northern Europe. More specifically, we used data from (i) the Human Genome Diversity Panel (HGDP) [45], (ii) the HapMap Phase III Project [46], (iii) publicly available data on Northern African populations that were first released by Henn et al [47], and (iv) data from the Kidd Lab at Yale University (allele frequencies for data from the Kidd Lab are available via the ALFRED database at Supplementary Table 2 presents a detailed list of the 762 additional samples from 23 populations that were used in our main analyses. Supplementary Table 2: An additional 762 samples from four other datasets were included in our main analyses. The genotyping platforms for each of the four additional datasets are described in the respective references. Sample size indicates the number of individuals in our study. Population Region Sample Size Reference (country of origin) (N) Algeria North Africa 19 Henn et al 2012 Egypt North Africa 19 Henn et al.2012 Libya North Africa 17 Henn et al.2012 North Morocco North Africa 18 Henn et al.2012 Sahara Occidental North Africa 18 Henn et al.2012 South Morocco North Africa 16 Henn et al.2012 Tunisia North Africa 18 Henn et al.2012 Bedouin South West Asia 47 HGDP Druze South West Asia 46 HGDP Mozabite South West Asia 28 HGDP Palestinian South West Asia 51 HGDP Yemen South West Asia 37 Kidd Lab Basque South Europe 44 Henn et al.2012, HGDP Italy South Europe 13 HGDP Sardinia South Europe 28 HGDP Tuscany South Europe 96 HAPMAP, HGDP Hungary Central Europe 41 Kidd Lab Chuvash East Europe 41 Kidd Lab Russia East Europe 46 HGDP, Kidd Lab France West Europe 29 HGDP Danes North Europe 43 Kidd Lab Finns North Europe 25 Kidd Lab Ireland North Europe 22 Kidd Lab 2

9 Supplementary Figure 1: Geographic map showing the locations of the populations in Supplementary Tables 1 and 2 (see also Figure 1A in the main text). 3

10 Combined Datasets: The main dataset that was used in all our analyses emerged after combining the populations in Supplementary Tables 1 and 2. After extracting common SNPs, merging and aligning genotypes, and applying basic quality control filters (see Supplementary Methods for details), this dataset consisted of 964 samples from 32 populations, genotyped on 75,194 SNPs. Supplementary Table 3: Approximate location of origin and geographic coordinates (longitude/latitude) for populations in our sample that surround the Mediterranean basin. Population Capital Latitude (North-South) Longitude (East-West) Sahara Occidental Laayoune South Morocco Agadir North Morocco Rabat Tunisia Tunis Mozabite Ghardaia Algeria Algiers Libya Benghazi Egypt Cairo Druze Damascus Palestine Gaza Cappadocia Kayseri East Rumelia Plovdiv Serbia Belgrade Macedonia Thessaloniki Peloponnese Tripoli SE Laconia Neapoli Laconias Dodecanese Rhodes Crete Heraclion Sicily Catania Italy Rome Tuscany Florence Sardinia Olbia Basque Bilbao

11 Supplementary Material: Methods and Results Merging genotypes from different sources: In order to merge datasets from the five different sources described in Supplementary Tables 1, 2, and 3, we had to pay particular attention to strand information and properly align common SNPs. Since such information was not always available, we chose to omit SNPs with reference alleles CG and AT in order to avoid ambiguity. Quality control: Despite the fact that the missing genotype rates in any of the five datasets that we analyzed in our work were quite low (invariably below 1%), we chose to perform an additional quality control check and remove any SNP that had a missing rate exceeding 20% in any of the populations under study. Our objective was to remove SNPs that might cause spurious artifacts in our analyses simply because they had many missing genotypes in one of the studied populations. Computing Fst: We computed Fst between all available populations using the SmartPCA tool from the Eigenstrat software package, as well as our own MatLab script that implements the Fst formula of the International HapMap Project. PCA: We used our own MatLab implementation of PCA, which mean-centers the data while ignoring missing entries. We first transformed the genotypic data to numeric values, without any loss of information, in order to apply linear algebraic analyses. Consider a dataset for a group of populations consisting of m subjects and assume that for each subject n biallelic SNPs have been assayed. Thus, we are given a table T, consisting of m rows and n columns. Each entry in the table is a genotype (pair of bases), ordered alphabetically. We transform this initial data table to an integer matrix A, which consists of m rows (one for each subject) and n columns (one for each SNP). Each entry of A will be set to +1, 0, -1; an entry could be empty in the case of missing genotypes. Let B 1 and B 2 be the bases that appear in the j-th SNP (in alphabetical order). If the genotypic information for the j-th SNP of the i-th individual is B 1 B 1 the (i,j)-th entry of A is set to +1; else if it is B 1 B 2 the (i,j)-th entry of A X is set to 0; else if it is B 2 B 2 the (i,j)-th entry of A is set to -1; see Paschou et al. 2007a [48], Paschou et al. 2007b [49] for more details. After forming the matrix A, we transform it to a mean-centered matrix M, by mean-centering each column (SNP) of A. It is worth noting that we ignore missing entries when forming M. Finally, in order to compute the principal components, it suffices to compute the Singular Value Decomposition of the matrix product MM T, which is an m-by-m matrix. Supplementary Figure 2 (five panels) shows PCA plots (in two or three dimensions) for subsets of the populations of Supplementary Tables 1 and 2. Panels (a) and (B) are PCA plots of all populations in Supplementary Tables 1 and 2, projected on the top two and three principal components. Panel (c) is the three-dimensional analog of main text Figure 1b. Panels (d) and (e) are PCA plots of populations around the Mediterranean basin, with Bedouin and Yemenites excluded. There is an obvious resemblance between the geography of the region and the twodimensional PCA plots. Different sample sizes and their effect on Principal Component Analysis: To test the effect of different sample sizes on PCA, we computed centroids of our input populations in Supplementary Tables 1 and 2 first using all available samples, then using only 20 samples per population (chosen uniformly at random), and finally using only 30 samples per population (chosen uniformly at random). We then correlated the centroids of the populations using all available samples with the centroids that were computed using 20 and 30 samples only. The 5

12 Pearson correlation coefficient for the first principal component between the centroids that were computed using all available samples and the centroids that were computed using 30 (respectively 20) samples per population exceeded (respectively 0.955) with a standard deviation (over 100 repetitions) of less than (resp ). The correlation for the second principal component was only slightly worse in the case of 20 samples per population: the Pearson correlation coefficient for the second principal component between the centroids that were computed using all available samples and the centroids that were computed using 30 (respectively 20) samples per population exceeded 0.93 (respectively 0.85) with a standard deviation (over 100 repetitions) of less than (respectively 0.02). Given these statistics, we opted to keep all available samples in our populations in order to capture as much of the population variance as possible. Correlation between PCA and geographic coordinates: In order to estimate the correlation between geographic coordinates and the top two eigenvectors emerging from PCA, we used the data in Supplementary Table 3. For each population in our sample, we approximated its location of origin either using information provided to us by the individuals that collected the respective sample, or by using a capital city that is relatively close to the population under study. The correlation between geographic coordinates and the eigenvectors was computed by converting both the geographic coordinate vector and the eigenvectors to z-scores, and then computing the Pearson correlation coefficient. We also ran a Mantel test in order to estimate the statistical significance of the result by permuting the geographic distance matrix between all pairs of populations (we note that we used the Haversine formula to accurately compute distance between populations using longitude and latitude information). After 10,000 permutations we did not get a single permutation whose Pearson correlation coefficient exceeded the observed ones, thus concluding that our observation is significant with a p-value at least Identification of genetic barriers: We used the BARRIER v2.2 software [26] to explore the genetic barriers in our dataset. BARRIER implements Monmonier s maximum difference algorithm in order to identify such barriers. First, using the geographic coordinates of the 23 populations surrounding the Mediterranean sea (see Supplementary Figure 2d for the PCA plot of these populations), we used BARRIER to plot the studied populations on the two dimensional plane (see Supplementary Figure 3a). Supplementary Figure 3a also shows the Delaunay triangulation and the Voronoi tessellation for the 23 populations. Recall that the Voronoi tessellation (blue lines) can be thought of as defining polygons that include all points on the plane that are closer to the centroid of the polygon under investigation than to the centroid of any other polygon. Since the centroids of the polygons in BARRIER correspond to the studied populations, we can consider the respective polygons as the areas in the plane that correspond to the respective population. If two polygons (and their respective populations) share an edge, then one might conclude that, based on the geographic distances, the respective populations are connected and thus some gene flow between the two populations could be expected. Indeed, the green edges in Supplementary Figure 3a (corresponding to the Delaunay triangulation) indicate connections between various pairs on populations based on geography; clearly, many connections exist that cross the Mediterranean sea: note that populations 1, 10, 12, 13, 14, 18, 22 all correspond to North African populations, yet, based solely on geography, they have multiple connections with South European populations. This picture changes drastically once genetic information from Principal Components Analysis is incorporated. In order to get statistical significance, we used a standard bootstrap procedure to 6

13 compute 100 distance matrices by resampling individuals from our populations. To be precise, we used a ten-fold cross-validation procedure in order to construct these 100 distance matrices, where each matrix contains all pairwise distances between the centroids of the 23 studied populations, as computed via Principal Components Analysis on the sampled individuals. The BARRIER software accepts the bootstrap distance matrices as input in order to run Monmonier s algorithm in order to compute genetic barriers between the studied populations. Supplementary Figure 3b clearly shows (observe the thick red lines) strong genetic barriers separating the North African populations from the South European populations, as expected by the presence of the Mediterranean sea. Estimating population admixture: We used the ADMIXTURE v1.22 software for all our admixture analyses. The parameter K (number of ancestral populations) ranged between two and eight in all our analyses. Prior to running ADMIXTURE, we pruned the SNPs in order to remove SNPs in high LD. Towards that end, we used the PLINK software and pruned SNPs using a windowed approach and a value of r 2 equal to 0.8. We used DISTRUCT v.1.1 and CLUMPP v to visualize the output of ADMIXTURE. Supplementary Figure 4 shows an ADMIXTURE plot of all populations in Supplementary Tables 1 and 2, for all values of K between two and eight. Supplementary Figure 5 shows an ADMIXTURE plot of all populations in Supplementary Tables 1 and 2, excluding the Bedouins and the Yemenites, for all values of K between two and eight. The data in Supplementary Figure 5, namely the percentages of origin of each individual with respect to the K (unknown) ancestral populations, were used in our ADMIXTURE based network formation algorithms (see Supplementary Figure 8). In order to account for potential variance in multiple runs of ADMIXTURE, we performed 20 runs of ADMIXTURE for each value of K between two and eight. We focus our discussion on the dataset used in main text Figure 3; the conclusions for the datasets used in Supplementary Figures 4 and 5 are identical. Each of the 20 runs (for a fixed value of K) was performed with a different random seed using the s time parameter of ADMIXTURE in order to initialize the random seed generator using the current time of the machine; this is often considered to be a good approximation to true randomness. We observed that for all values of K the results were very highly correlated. More specifically, recall that ADMIXTURE returns, for each sample in the dataset, K coefficients indicating percentage of ancestry in each of the K (assumed) ancestral populations. First of all, for a particular value of K, the average Pearson correlation coefficient among the reported ADMIXTURE coordinates and the coordinates returned in the 20 replicates ranged from 0.94 for K=2 to 0.86 for K=8. This already indicates that the correlation between the various replicates of ADMIXTURE is very high (a similar observation was made by [Henn et al. (2011) Proc Natl Acad Sci USA 108(13): ], who reported that multiple runs of ADMIXTURE only resulted in minor changes and simply chose to report one run; in a subsequent paper, [47] only performed a single run of ADMIXTURE). The situation is actually even more favorable, once we look closer at the data: more specifically, for each of the 20 runs (say run j, for some j between one and 20), we computed the Pearson correlation coefficient between the reported coefficients for each of the K ancestral populations (as shown in main text Figure 3) and their best fit in the j-th run (we implemented a matching algorithm to perform the alignment between the K ancestral populations in the reported run and the j-th replicate). We noticed that, for all values of K between two and six, over 70% of the replicates were essentially identical to the reported one (average Pearson correlation coefficient exceeded 0.98), while for K equal to seven or eight approximately 50% of the replicates were essentially identical to the reported one (average Pearson correlation coefficient approximately 0.95). Checking the 7

14 reported values of log-likelihood and delta (as reported by ADMIXTURE), we observed that for the cluster of identical repetitions convergence was better than for the other repetitions, thus indicating a better approximation to the true optimal value. Therefore, we chose to report one of the essentially identical runs. A similar analysis was performed in [Behar et al. (2010) Nature (8) 466 (7303):238-42], who used a similar method to account for the slight variation in the ADMIXTURE output (and then chose to report one of the identical ADMIXTURE outputs). f3 statistics: In order to test for bi-directional admixture and population splits along the populations that stand at the gateway to Europe, we ran three population tests as described by Patterson et al [27] and implemented in TreeMix [30]. Seven populations bridging Anatolia to Southern Europe were included in the analysis (Cappadocia, Dodecanese, Crete, SE Laconia, Peloponnese, Macedonia, East Rumelia), and f3 statistics were calculated for all possible triplets of populations. After removal of SNPs in LD (at r 2 >0.5) from our initial dataset of over 650,000 SNPs, we were left with approximately 278,000 SNPs. When studying a triplet of populations (e.g., C; A,B) f3 statistics and the corresponding Z-score of significance can only be negative if population C has ancestry from populations related to both A and B. It is only in this case that paths exist between C and A as well as C and B that also take opposite drift directions. The observation of a significantly negative value of f3(c; A, B) is thus evidence of complex phylogeny in C. All triplets of populations are shown in our online supplementary material ( We only get marginally negative results for the Peloponnese population in Greece, indicating that this population might have resulted from some admixture from both North and South. All other tested triplets tested positive, indicating that there are simple paths connecting populations to each other along the course of migration from Anatolia into Europe. This is in concordance with our network analysis. Network analysis: To better understand the connection between the populations included in our study we performed a network analysis on the results of PCA and ADMIXTURE. We start by noting that both PCA and ADMIXTURE are essentially reducing the dimensionality of the original dataset by expressing each sample as either (a) a linear combination of the top few eigenvectors (in the case of PCA) or (b) as percentages of ancestry from a small number of typically unknown ancestral populations (in the case of ADMIXTURE). To be more precise, each individual sample is originally described with respect to 75,194 SNPs; mathematically, this is equivalent to saying that the sample lies in a 75,194-dimensional subspace. After applying PCA or ADMIXTURE to the dataset, all samples are described with respect to K coefficients; K ranges between one and eight in all our analyses. Mathematically, this is equivalent to saying that the output of PCA or ADMIXTURE lies in a K-dimensional subspace, with K <<< 75,194. It is worth highlighting that PCA and ADMIXTURE are unrelated dimensionality reduction techniques that make different assumptions regarding the underlying structure of the dataset. PCA is a very general algorithmic technique that recovers the latent structure of highdimensional data that lie in an approximately linear manifold, by leveraging the existence of a small number of pairwise orthogonal principal components that capture the linear structure of the dataset. ADMIXTURE, on the other hand, is a much more specialized model-based statistical analysis technique whose underlying model, to the best of our understanding, is only applicable to population genetics datasets. Its end goal is a (maximum likelihood based) estimation of individual ancestries from large-scale SNP datasets. An important difference regarding the output of the two techniques is that the output of PCA lies in a low-dimensional normed subspace as described above, where the metric of distance between two samples is the Euclidean distance, since PCA minimizes an objective function that is based on the Euclidean or l 2 norm. (Recall that the Euclidean or l 2 distance between a pair of points in a K-dimensional space is the square root of the sum of the squares of the coordinate-wise distances between the 8

15 two points.) However, while the output of ADMIXTURE also lies in a low-dimensional subspace, the relevant norm is not necessarily the Euclidean norm. As a matter of fact, the output of ADMIXTURE is essentially a probability distribution for each sample, indicating the percentage of origin of the sample with respect to the ancestral populations. Indeed, the coefficients returned by ADMIXTURE for each sample are positive and sum up to one, a fact that does not hold for PCA. Thus, a more appropriate choice of norm for ADMIXTURE s output is the l 1 -norm, which better captures the distance between two probability distributions. (Recall that the l 1 distance between a pair of points in a K-dimensional space is sum of the absolute values of the coordinate-wise distances between the two points.) In order to form a network of related populations using the output of PCA or ADMIXTURE, we first identify the top ten nearest neighbors of each individual in the PCA or ADMIXTURE dimensionally reduced subspace, with the additional constraint that these nearest neighbors must not lie in the population of origin of the target individual. To be more precise, consider an individual in population X; this individual is represented with respect to K coefficients, which are the output of PCA or ADMIXTURE. We start by computing the l 2 (Euclidean) distance in the case of PCA (or l 1 distance in the case of ADMIXTURE) between the target individual and every other individual in our dataset. Then, we identify the top ten nearest neighbors to the individual under study, with the constraint that these nearest neighbors do not belong to population X. (We note that the distances are only used in order to select the top ten nearest neighbors from the target individual; the numeric values of the distances of those neighbors and the target individual are ignored in the remainder of our network analysis.) This procedure is repeated for all individuals in the dataset. Assuming that we have n samples from p populations in our dataset, this procedure eventually returns an n-by-p table with rows corresponding to samples and columns corresponding to the populations under study. The (i,j)-th entry in the table denotes how many neighbors individual i has in population j. Clearly, by the construction of our nearest-neighbor identification algorithm, the (i,j)-th entry of the matrix is equal to zero if the j-th column of the table corresponds to the population of origin of the i-th sample. Given the above n-by-p table A, we are now ready to create a network of populations with edges connecting related populations. In order to describe our procedure to infer whether populations X and Y are connected, as well as the weight of the respective connection, we first need to introduce some notation. Let t be the cardinality of population X and let s be the cardinality of population Y. Let i 1,i 2,,i t be the row indices corresponding to samples in X. Let j Y be the column of A that corresponds to population Y and compute t 1 dist( Y, X ) Ai a jy s Similarly, if we use i 1,i 2,,i s to denote the row indices of the table A that correspond to samples in Y (recall that the cardinality of Y is s) and j X to denote the column of A that corresponds to population X, we can compute s 1 dist( Y, X ) Ai a jx t It should be obvious that, up to scaling, our distance metric simply counts the number of nearest neighbors that individuals of population X have in population Y. This is a simple and intuitive a 1 a 1 9

16 idea; a minor fine tuning is necessary in order to mitigate the effect of different population sizes in our sample. A number of comments will help the reader to better understand the above two metrics of distance between populations X and Y. First of all, our distance metric is not symmetric: mathematically, dist(x,y) is not necessarily equal to dist(y,x). Second, in order to mitigate the effect of different population sizes, we normalize the aforementioned count by dividing it by the square root of the population size. The goal of this normalization is to assign larger weights when nearest neighbors in smaller populations are identified. For example, identifying a nearest neighbor in a population Y that has only 10 samples should be weighted more than identifying a nearest neighbor in a population Y that has 100 samples. The choice of normalizing by the square root (more generally, a power less than one) of the sample size is a common tradeoff in order to mitigate the effects of different population sizes, while not overcompensating for such effects. Third, using the dimensionally reduced space (for various values of K) derived by PCA or ADMIXTURE in order to compute the nearest neighbors of each individual (instead of using all available genotypes and the allelic distance) has two advantages: first and foremost, it denoises the data by keeping only the most prominent and meaningful axes of variance and thus avoiding the statistical artifacts of the curse-of-dimensionality; second, it speeds up computations since the data now lie in a low-dimensional space. We are now ready to describe our network formation algorithm. For each pair of populations X and Y, we compute both distances: dist(x,y) and dist(y,x). We create an edge between X and Y (thus claiming that the two populations are neighbors of each other) if min{dist(x,y),dist(y,x)} > 0, and we assign as a weight of the respective edge the value min{dist(x,y),dist(y,x)}. It is worth noting that our choice is quite conservative, since we will assume that populations X and Y are related if and only if both X and Y have nearest neighbors in each other. This notion of edge weights handles situations where a population Y is far away from all other populations and thus dist(x,y)=0 for all populations X, but, by construction, there will exist some populations X so that dist(y,x)>0 (since each sample in Y will have some nearest neighbors in populations outside Y, even if Y is isolated). Finally, once a network whose nodes correspond to populations and whose edges correspond to connections between populations, as described above, is formed, we visualize it using the Cytoscape software package. We briefly discuss the number of nearest neighbors to be retained using our approach. It should be clear that our network formation algorithm has one free parameter, namely the number of nearest neighbors to retain in the first step of the process. We chose to set that value to ten, which is equal to 1/3 of the average number of samples per population in our dataset. (Supplementary Tables 1 and 2 show the number of samples per population; the average number of samples/population is equal to 30.1) We also experimented with the number of nearest neighbors set to eight, nine, eleven, and twelve, without observing noticeable differences in the resulting networks. Our network analysis using PCA is based on Supplementary Figures 2d and 2e, as well as their higher dimensional analogs, which of course cannot be plotted. Note that we included all 10

17 populations in Supplementary Tables 1 and 2, except for the Bedouins and the Yemenites, which form a separate cline towards Central-South Asia. Similarly, our network analyses using ADMIXTURE were based on Supplementary Figure 5, which again excludes the Bedouins and the Yemenites. First, in the case of PCA, we had to determine how many principal components are significant. Towards that end, we examined the distribution of singular values in Supplementary Figure 6 (both panels). Panel (a), which corresponds to the plots in Supplementary Figures 2d and 2e, shows that the singular values drop fast: the top three are clearly significant; then the fourth and the fifth one, as well as the sixth and the seventh one, form small clusters. A simple metric of significance (see the legend of Supplementary Figure 6) indicates that the top seven singular values are indeed significant. Adding the Bedouins and the Yemenites to our dataset results in a slight increase in the number of significant singular values, which is now equal to ten. Supplementary Figures 7 and 8 show the various networks that were formed using all values of K and either PCA or ADMIXTURE as the basis for computing nearest neighbors. In the case of PCA, we examined values between three and seven (the number of significant principal components). In the case of ADMIXTURE, we examined values of K between three and eight. Interestingly, all networks are highly similar, showing a distinct path from North Africa to Near East and Southern Europe via Cappadocia, Dodecanese, and Crete. The resulting networks, as visualized by Cytoscape, are also available online as.cys files and include numerous network statistics, as computed by Cytoscape (see Supplementary Online Material). Network analysis via Fst: We also performed a network analysis using the Fst metric as follows: we formed an edge between two populations if the respective Fst was below 0.08 (we chose this threshold value since it resulted in keeping the top 25% most significant Fst pairs while maintaining a connected network larger threshold values resulted in similar yet denser networks, while smaller values disconnected the network). The resulting network is shown in Supplementary Figure 9 and is similar in spirit to the networks formed via our metric that is based on PCA and ADMIXTURE. We believe that PCA and ADMIXTURE based networks capture in a much finer sense the connections between the analyzed populations; Fst is a much coarser statistic. Simulating a stepping-stone model: To further test the stepping-stone hypothesis regarding the migration of populations around the Mediterranean basin, we used IBDSim to generate simulated data. IBDSim [29] is a package for the simulation of genotypic data under isolation by distance. It is based on a backward generation by generation coalescent algorithm allowing the consideration of various isolation-by-distance models. We used IBDSim in order to generate 10,000 independent SNPs for ten populations that were placed on a one-dimensional lattice on coordinates that are representations of a subset of ten populations from our data. We set the effective population size to 30, the mutation rate to 0.5e-6, and the migration rate to 0.5. (See Supplementary Table 4, where ten consecutive -- from a geographic perspective -- populations around the Mediterranean basin were chosen; the table also indicates their geographic coordinates, the distance from one population to the next one, the normalized distance from one population to the next one, and the respective lattice points where the populations were placed. 11

18 More specifically, we treated the minimum distance between any two consecutive populations in our list as a single unit; this was the distance between Crete and the Dodecanese. Then, we divided each distance by the minimum distance, thus getting the sixth column of the table. Finally, we placed populations in the x-axis of a one-dimensional lattice by rounding the distances in the sixth column to the nearest integer in order to approximately respect the distances between two consecutive populations.) We used default values for the migration rate and the mutation rate and generated data using the stepping stone model, as implemented by IBDSim. Recall that in the stepping stone model, an individual can only move to an adjacent population. The resulting PCA plots and the corresponding network are shown in Supplementary Figures 10 and 11 and they are clearly reminiscent of patterns presented in Supplementary Figures 2 and 7 (PCA and network analysis of real data). Thus our simulations provide further support for our conclusions regarding the stepping stone hypothesis of population migrations around the Mediterranean with Crete acting as a hub that connects Anatolia to Southeastern Europe. Inferring phylogenetic trees: We used PHYLIP to infer phylogenetic trees using the output of SmartPCA (Fst distances between all pairs of available populations). Phylogenetic trees were constructed using the neighbor-joining method (NEIGHBOR tool in PHYLIP). DRAWGRAM and DRAWTREE were used in order to plot the resulting phylogenetic trees. In order to further verify our findings, we constructed a phylogenetic tree based on the Fst distance matrix. The resulting tree (see Supplementary Figure 12 that includes the Sub-Saharan African San population as an outgroup this is a South African population that has been made available by the HGDP consortium) clearly support the findings of PCA, ADMIXTURE, as well as our network analyses. TreeMix: To further verify the migration pathways and signals of admixture among the studied populations from Northern Africa, Middle East, Anatolia, and Europe, we additionally used TreeMix [30] to find a population graph that best describes the relationship between populations in the dataset by testing for gene flow between them. Genomewide allele frequency data is used to first find the maximum-likelihood tree of populations and then infer possible additional migration events by identifying populations that poorly fit this tree. The maximum likelihood tree of all populations included in this analysis is shown in Supplementary Figure 13a, with the respective residuals shown in Supplementary Figure 13b. (It is worth noting that this tree explains 96.3% of the variation of the data, as computed by TreeMix.) In concordance with all other analysis that we present, gene flow from Anatolia to Southern Europe appears to have occurred through the islands of the Dodecanese and Crete. Residuals were subsequently analyzed in order to identify pairs of populations that are more related to each other than is captured by this graph and corresponding migration events were inferred. We then sequentially added 10 migration events to reach a tree that captures approximately 98.5% of the data. Migration events are shown in Supplementary Table 5. TreeMix analysis supports some additional gene flow among populations from Northern Africa (e.g., from South Morocco to Egypt and Algeria), as well as a connection between South Morocco and Middle East. Furthermore, an indication of the complex genetic history of Sardinia is uncovered with connections to Northern Africa, Italy, and the Basques. 12

19 Neighbor-Net analysis: As a final verification of the migration pathways and signals of admixture among the studied populations, we performed a phylogenetic network analysis using the Neighbor-Net method [31] implemented in SplitsTree v4.13. Neighbor-Net is a distance based method for the construction of phylogenetic networks that is based on the neighborjoining algorithm. Recall that a fundamental property of a tree is that between every pair of populations there is a unique path connecting them, while a network or graph allows for multiple paths. The Neighbor-Net algorithm relaxes the neighbor-joining algorithm to construct a phylogenetic graph that aims to improve the fit to the data by allowing (if necessary) additional paths between pairs of populations. Comparing to our own network formation algorithms, it is worth noting that our approach is not based on neighbor-joining and does not attempt to fit any particular structure to the data. It simply picks the most informative connections in order to produce a graph of gene flow among populations. In order to run Neighbor-Net, we first computed the Fst matrix between all populations using EIGENSOFT on our dataset and used this matrix as the input for the Neighbor-Net algorithm. The Neighbor-Net method was used to produce an unrooted network, which was subsequently visualized using the EqualAngle method (Supplementary Figure 14). In concordance with all of our previous analysis, it is again clear that the Dodecanese and Crete play a pivotal role as a hub in gene flow connecting Anatolia to Southern Europe, and the rest of Europe. 13

20 Supplementary Material: Figures Supplementary Figure 2: PCA plots of (subsets of) the populations in Supplementary Tables 1 and 2. (a) PCA plot of all populations in Supplementary Tables 1 and 2; projection on top two Principal Components. The cline from Northern Africa to Northern Europe via Near East, Cappadocia, Dodecanese, and Crete is apparent. (b) PCA plot of all populations in Supplementary Tables 1 and 2; projection on top three Principal Components. (c) PCA plot of populations around the Mediterranean basin; projection on top three principal components. This plot is the three-dimensional version of main text Figure 1b. Bedouins and Yemenites are separated from the Mediterranean basin populations. (d) PCA plot of populations around the Mediterranean basin; projection on top two principal components. (Bedouin and Yemenites haven been excluded from this plot.) (e) PCA plot of populations around the Mediterranean basin; projection on top three principal components. (Bedouin and Yemenites have been excluded from this plot.) The Tunisian population is separated from the other populations in the third principal component. 14

21 Figure S2a 15

22 Figure S2b 16

23 Figure S2c 17

24 Figure S2d 18

25 Figure S2e 19

26 Supplementary Figure 3: Results from running the BARRIER v2.2 software. (a) Delaunay triangulation (green lines) and Voronoi tessellation (blue lines) of populations around the Mediterranean basin. (b) Genetic barriers were computed using 100 bootstrap distance matrices (computed via Principal Components Analysis and ten-fold cross-validation) and are indicated by red lines; thickness of the lines increases with the statistical significance of the respective barriers. Notice that strong genetic barriers separate the North African populations from the South European populations, as expected by the presence of the Mediterranean. Figure S3a Palestine Egypt Druze Figure S3b Cappadocia Libya Dodecanese Crete SE Laconia Peloponnese Tunisia Algeria Sicily Italy Sardinia North Morocco South Morocco Sahara OCC Basque 20

27 Supplementary Figure 4: ADMIXTURE plot of all populations in Supplementary Tables 1 and 2, for all values of K between two and eight. A clear gradient can be observed, from North Africa to Near East, South Europe (with a sub-gradient forming from Cappadocia, Dodecanese, Crete, and Peloponnesos), and, eventually, North Europe. K=2 K=3 K=4 21

28 K=5 K=6 K=7 K=8 22

29 Supplementary Figure 5: ADMIXTURE plot of all populations in Supplementary Tables 1 and 2, excluding the Bedouins and the Yemenites, for all values of K between two and eight. We again observe a clear gradient from North Africa to Near East, South Europe (with a subgradient forming from Cappadocia, Dodecanese, Crete, and Peloponnesos), and, eventually, North Europe. The percentages of origin of each individual with respect to the K (unknown) ancestral populations were used in our network formation algorithms (see Supplementary Figure 8). K=2 K=3 K=4 23

30 K=5 K=6 K=7 K=8 24

31 Supplementary Figure 6: (a) Plot of top 50 singular values of the covariance matrix formed using all samples, excluding the Bedouins and the Yemenites, in Supplementary Tables 1 and 2. The top principal components of this covariance matrix will be used in our PCA-based network formation algorithms (Supplementary Figure 7). Applying a simple metric of singular value significance (i.e., extracting principal components corresponding to singular values that exceed the average singular value by at least three standard deviations) identifies the top seven singular values as significant. (b) Plot of top 50 singular values of the covariance matrix formed using all samples, including the Bedouins and the Yemenites, in Supplementary Tables 1 and 2. Applying a simple metric of singular value significance (i.e., extracting principal components corresponding to singular values that exceed the average singular value by at least three standard deviations) identifies the top ten singular values as significant. Figure S6a 25

32 Figure S6b 26

33 Supplementary Figure 7: Networks formed using PCA and the top K (for all values of K between three and seven) principal components in order to identify nearest neighbors for each individual (see Supplementary Methods for details and Supplementary Figure 6a for the number of significant Principal Components). Warmer colors indicate nodes of high centrality for the whole network (as computed by Cytoscape), while thicker edges indicate strong connections (high genetic similarity between the respective populations). Different values of K result in highly similar networks, highlighting the robustness of our results. The path from Northern Africa to Southern Europe via Near East, Cappadocia, the Dodecanese, and, of course, Crete is obvious. Note that the panel corresponding to K=5 is identical to Figure 4a in the main text; we include it here as well to facilitate comparisons. PCA-based network, K=3 27

34 PCA-based network, K=4 PCA-based network, K=5 28

35 PCA-based network, K=6 PCA-based network, K=7 29

36 Supplementary Figure 8: Networks formed using ADMIXTURE with the parameter K (the number of ancestral populations) set between four and eight in order to identify nearest neighbors for each individual (see Supplementary Methods for details). Warmer colors indicate nodes of high centrality for the whole network (as computed and visualized by Cytoscape), while thicker edges indicate strong connections (high genetic similarity between the respective populations). Different values of K result in highly similar networks, highlighting the robustness of our results. Even more interestingly, the networks of Supplementary Figures 7 and 8 are again very similar, even though PCA is a model-free dimensionality reduction technique, while ADXMIXTURE is a model-based approach (see Supplementary Methods for a more detailed discussion of the two approaches). Once more, the path from Northern Africa to Southern Europe via Near East, Cappadocia, the Dodecanese, and, of course, Crete is obvious. Note that the panel corresponding to K=5 is identical to Figure 4b in the main text; we include it here as well to facilitate comparisons. ADMIXTURE-based network, K=4 30

37 ADMIXTURE-based network, K=5 ADMIXTURE-based network, K=6 31

38 ADMIXTURE-based network, K=7 ADMIXTURE-based network, K=8 32

39 Supplementary Figure 9: Network formed using Fst values between pairs of populations (see Supplementary Methods for details). Warmer colors indicate nodes of high centrality for the whole network (as computed by Cytoscape), while thicker edges indicate strong connections (high genetic similarity between the respective populations). 33

40 Supplementary Table 4: Geographic coordinates of the ten populations around the Mediterranean basin that were used as a basis for our simulations in Supplementary Figures 10 and 11. The table also indicates the distance from one population to the next one, the normalized distance from one population to the next one, and the respective lattice points where the populations were placed. We denoted the ten populations as pop1 through pop10 in Supplementary Figures 10 and 11. Population Latitude Longitude Distance (in kms) from previous population Normalized distance (divided by minimum distance) Coordinates in x-axis (rounded) Algeria (pop1) Tunisia (pop2) Libya (pop3) Egypt (pop4) Druze (pop5) Palestine (pop6) Cappadocia (pop7) Dodecanese (pop8) Crete (pop9) Peloponnese (pop10) 34

41 Supplementary Figure 10. PCA plot of ten simulated populations; see Supplementary Table 4 for a list of the corresponding populations around the Mediterannean basin that were used as a basis for our simulation. Note that the figure is reminiscent of our PCA plots of populations around the Mediterranean basin. 35

42 Supplementary Figure 11. Network formed by our algorithm using the PCA plot of data simulated under a stepping stone model (Supplementary Figure S10). In order to simulate the stepping stone model of migrations, distances between real populations around the Mediterranean were used (see Supplementary Table 4). Note that the network of the simulated data is in complete concordance with the network formed using real data for the respective populations. 36

43 Supplementary Figure 12. Phylogenetic tree formed using Fst distances between all populations studied. The results are concordant with a closer genetic relationship of Anatolia to the islands of Crete and the Dodecanese rather than to the Balkans. 37

44 Supplementary Figure 13a. The maximum-likelihood tree generated by TreeMix, capturing 96.27% of the data. In concordance with all other analysis that we present, gene flow from Anatolia to Southern Europe appears to have occurred through the islands of the Dodecanese and Crete. Drift parameter 38

45 Supplementary Figure 13b. Residuals of the maximum-likelihood phylogenetic tree generated by TreeMix and shown in Supplementary Figure 13a. Population pairs that show high residuals (i.e., are not perfectly captured by the graph in Supplementary Figure S13a), were subsequently analyzed in order to identify pairs of populations that are more related to each other and corresponding migration events were inferred. Supplementary Table 5 shows the top ten inferred migration events based on this residual plot. As implemented in TreeMix, the residual covariance between each pair of populations i and j is divided by the average standard error across all pairs. This scaled residual in then plotted in each cell (i,j). Colors are described in the palette on the right. Residuals above zero represent populations that are more closely related to each other in the data than in the best-fit tree and thus are candidates for admixture events. 39