Homo Sapiens: Anatomically Modern Humans (AMH)

Fig. 2: Genetic diversity across Africa.




a, Selected populations from the 1000 Genomes Project and the African Diversity Reference Panel18,20 illustrate diversity from western, eastern and southern Africa. We chose representative ethnic groups from each region (bold labels) to build parameterized models, including the newly genetically sequenced Nama populations from South Africa, Mende from Sierra Leone, Gumuz, Oromo and Amhara from Ethiopia, British individuals and a Neanderthal from Vindija Cave, Croatia. b,c, Principal component analysis highlights the range of genetic divergence anchored by western African, Nama, Gumuz and British individuals between principal components (PC) 1 and 2 (b), and 1 and 3 (c). Percentages show variance explained by each principal component. Colours represent the groups shown in bold in a. d, ADMIXTURE analysis using K = 4 principal components reveals signatures of recent gene flow in Africa that reflect colonial-period migration into the Nama, back-to-Africa gene flow among some Ethiopians, and Khoe-San admixture in the Zulu population.


We therefore aim to discriminate between a broader set of demographic models by studying the genomes of contemporary populations. We take as our starting point four models (single-population expansion, single-population expansion with regional persistence, archaic hominin admixture and multi-regional evolution; Fig. 1) using 290 genomes of individuals from southern, eastern and western Africa, as well as Eurasia. By including geographically and genetically diverse populations across Africa, we infer demographic models that explain more features of genetic diversity in more populations than previously reported. These analyses confirm the inadequacy of tree-like models and provide an opportunity to directly evaluate a wide range of alternative models.

We inferred detailed demographic histories using 4x–8x whole-genome sequencing data for four diverse African populations, comprising the Nama (Khoe-San from South Africa, newly presented here; see Supplementary Information section 1.2 for ethical and practical aspects of participant recruitment), the Mende (from Sierra Leone; from phase 3 of the 1000 Genomes Project18), the Gumuz (recent descendants of a hunter-gatherer group from Ethiopia20,21) and eastern African agriculturalists (Amhara and Oromo from Ethiopia20). The Amhara and Oromo populations, despite speaking distinct Afro-Asiatic languages, are highly genetically similar21,22 so we combined the two groups for a larger sample size (Fig. 2). We also included British individuals from the 1000 Genomes Project in our demographic models as a representative source of back-to-Africa gene flow and recent colonial admixture in South Africa. Finally, we used a high-coverage ancient Neanderthal genome from Vindija Cave in Croatia23 to account for gene flow from Neanderthals into people from outside Africa, and gauge the relative time depth of divergence, assuming that Neanderthals diverged 550 ka from a common stem. We computed one- and two-locus statistics for which the expectation within and across populations can be computed efficiently and that are well suited for both low- and high-coverage genomes2,24. Using a maximum-likelihood inference framework, we then fitted to these statistics a family of parameterized demographic models that involve population splits, size changes, continuous and variable migration rates and punctuated admixture events, to learn about the nature of the population structure over the past million years.

A Late Pleistocene common ancestry
We started with a model of geographical expansion from a single ancestral, unstructured source followed by migration between populations, without allowing for a contribution from an African archaic hominin lineage (Fig. 1a) or population structure before the expansion (Fig. 1d). As expected2, this first model was a poor fit to the data qualitatively (Supplementary Fig. 10) and quantitatively (log-likelihood (LL) ≈ −189,300; Supplementary Table 3). We next explored a suite of parameterized models in which population structure predates the differentiation of contemporary groups (Supplementary Information section 3). Depending on the parameters, these encompassed models allowing for ancestral reticulation, such as fragmentation-and-coalescence or meta-population models (Fig. 1b), archaic hominin admixture (Fig. 1c) and African multi-regionalism (Fig. 1d). The recent expansion and the African multi-regional models (Fig. 1a,d) have the same topology, so interpretation of the model depends on the specified or inferred divergence times.

Regardless of the model choice for early epochs, maximum-likelihood inference of human demographic history for the past 150 kyr was remarkably robust. In a reticulated model, we use ‘divergence’ between populations to mean the time of their most recent shared ancestry. The earliest divergence among contemporary human populations differentiates the southern African Nama population from the other African groups at 110−135 ka, with low to moderate levels of subsequent gene flow (Table 1). In none of the high-likelihood models that we explored was the divergence between Nama and other populations earlier than around 140 ka. We conclude that geographical patterns of contemporary H. sapiens population structure probably arose during MIS 5. Although we do find evidence for earlier population structure in Africa, contemporary populations cannot be easily mapped onto the more ancient ‘stem’ groups because only a small proportion of drift between contemporary populations can be attributed to drift between stems (Fig. 4, Supplementary Information section 5.2 and Supplementary Figs. 16–19).

Allowing for continuous migration between the stem populations substantially improves the fits relative to zero migration between stems (LL ≈ −101,600 compared with −107,700 in the merger model, Supplementary Tables 6 and 7; and LL ≈ −115,300 versus −126,500 in the continuous migration model, Supplementary Tables 4 and 5). With continuous migration between stems, population structure extends back to more than 1 million years ago (Table 1). Migration between the stems in these models is moderate, with a fraction of migrant lineages (m) in each generation estimated as m = 6.3 × 10−5–1.3 × 10−4. For comparison, this is similar to the inferred migration rates between connected contemporary populations over the past 50 ka (Table 1). This ongoing (or at least, periodic) gene flow qualitatively distinguishes these models from previously proposed archaic hominin admixture models (Fig. 1c), as the early branches remain closely related and each branch contributes large amounts to all contemporary populations (Fig. 4). Because of this relatedness, only 1–4% of genetic differentiation among contemporary populations can be traced back to this early population structure (Supplementary Information section 5.2).

Fig. 4: Structure among stems is weak and present-day structure is generally recent.



a–d, From the best-fit models of our two parameterizations (a,b, continuous migration; c,d, merger with stem migration), we predicted differentiation and shared drift between populations at past time points. a,c, We computed expected pairwise differences Hi,j between individuals sampled from populations i and j existing at time t. b,d, To understand how drift between stems explains contemporary structure, we computed the proportion α2 of drift between pairs of sampled contemporary populations (here the Nama and Mende) that aligns with drift between past populations (see Supplementary Information section 5.2 for details and additional comparisons in Supplementary Figs. 16–19). Both models infer deep population structure with modest contributions to contemporary genetic differentiation. Most present-day differentiation dates back to the past 100 kyr.

Under the continuous-migration model, one of the two stems (stem 1) diverges into lineages leading to contemporary populations in western, southern and eastern Africa, and the other (stem 2) contributes variable ancestry to those populations. This migration from stem 2 is highest with the Mende (m = 1.6 × 10−4) compared with the Nama and populations from eastern Africa (m = 5.9 × 10−5 and 3.1 × 10−5, respectively), with migration allowed to occur until 5 ka. A sampled lineage from the Nama, Mende and Gumuz have probabilities of being in stem 2 at the time of stem 1 expansion (135 ka) of approximately 0.145, 0.20 and 0.130, respectively, although these probabilities change over time, precluding the notion of a fixed admixture proportion.

By contrast, in the multiple-merger model, stem populations merge with varying proportions to form the different contemporary groups. We observe a sharp bottleneck in stem 1 down to Ne = 100 after the split of the Neanderthal branch. This represents the lower bound allowed in our optimization (an Ne of 100), although the size of this bottleneck is poorly constrained (95% confidence interval 100–851). After a long period of exchange with stem 2, stem 1 then fractures into stem 1E and stem 1S at 478 ka. The timing of this divergence was also poorly constrained (95% confidence interval 276– 478 ka). These populations evolve independently until 119 ka (101−125 ka) when stem 1S and stem 2 combine to form the ancestors of the Nama, with proportions of 30% and 70%, respectively. Similarly, stem 1E and stem 2 combine in equal proportions (50% each) to form the ancestors of the western and eastern Africans (and thus also all individuals who later disperse during the out-of-Africa event). Finally, the Mende receive a large additional pulse of gene flow from stem 2, replacing 19% (18–21%) of their population 25 ka (22−26 ka). The later stem 2 contribution to the western African Mende resulted in better model fits (∆LL ≈ 60,000). This may indicate that an ancestral stem 2 population occupied western or central Africa, broadly speaking. The differing proportions in the Nama and eastern Africans may also indicate a geographical separation of stem 1S in southern Africa and stem 1E in eastern Africa.

To assess the robustness of the inferred models to analysis and reference population choices, Supplementary Information sections 6 and 7 include reanalyses with changes in the European and West African populations, as well as the recombination maps, filtering strategies and parameter optimization strategies. Although we find some differences in the inferred parameters (see Supplementary Information sections 7.1.1 and 7.2), the best-fit models across all reanalyses are quantitatively consistent.