Main

Globally, human populations show structured genetic diversity as a result of geographical dispersion, selection and drift. Understanding this variation can provide insights into evolutionary processes that shape both human adaptation and variation in disease susceptibility1. Although the Hapmap Project2 and the 1000 Genomes Project3 have greatly enhanced our understanding of genetic variation globally, the characterization of African populations remains limited. Other efforts examining African genetic diversity have been limited by variant density and sample sizes in individual populations4, or have focused on isolated groups, such as hunter gatherers (HG)5,6, limiting relevance to more widespread populations across Africa.

The African Genome Variation Project (AGVP) is an international collaboration that expands on these efforts by systematically assessing genetic diversity among 1,481 individuals from 18 ethno-linguistic groups from sub-Saharan Africa (SSA) (Fig. 1 and Supplementary Methods Tables 1 and 2) with the HumanOmni2.5M genotyping array and whole-genome sequences (WGS) from 320 individuals (Supplementary Methods Table 2). Importantly, the AGVP has evolved to help develop local resources for public health and genomic research, including strengthening research capacity, training, and collaboration across the region. We envisage that data from this project will provide a global resource for researchers, as well as facilitate genetic studies in Africa7.

Figure 1: Populations studied in the AGVP.
figure 1

a, 18 African populations studied in the AGVP including 2 populations from the 1000 Genomes Project. (The term ‘Ethiopia’ encompasses the Oromo, Amhara and Somali ethno-linguistic groups.) b, c, ADMIXTURE analysis of these 18 populations alone (n = 1,481) (b) and in a global context (n = 3,904) (c). Each colour represents a different ancestral cluster, with clusters 2–6 represented along the y-axis in b and clusters 2–18 represented in c. K = 6 and K = 18 were the most likely clusters on ADMIXTURE analysis. ADMIXTURE analysis suggests substructure between North, East, West and South Africa. Studying these populations in the context of Eurasian and African HG populations suggest extensive Eurasian and HG admixture across Africa.

PowerPoint slide

Population structure in SSA

On examining 2.2 million variants, we found modest differentiation among SSA populations (mean pairwise FST 0.019) (Supplementary Methods and Supplementary Table 1). Differentiation among the Niger-Congo language groups—the predominant linguistic grouping across Africa was noted to be modest (mean pairwise FST 0.009) (Supplementary Table 1), providing evidence for the ‘Bantu expansion’—a recent population expansion and movement throughout SSA originating in West Africa around 3,000 to 5,000 years ago8.

We identified 29.8 million single-nucleotide polymorphisms (SNPs) from Ethiopian, Zulu and Bagandan WGS (Extended Data Fig. 1 and Supplementary Methods). A substantial proportion of unshared (11%–23%) and novel (16%–24%) variants were observed, with the highest proportion among Ethiopian populations (Extended Data Fig. 1). The high proportion of unshared variation among populations recapitulates the need for large-scale sequencing across Africa, including among genetically divergent populations.

We used principal component analysis to explore relationships among AGVP populations (Extended Data Figs 2, 3, 4, 5, Supplementary Figs 1 and 2). PC1 appeared to represent a cline extending from West and East African populations towards Ethiopian populations, possibly suggesting Eurasian gene flow, while PC2 separated West African and South/East African populations (Extended Data Fig. 2). Inclusion of the 1000 Genomes Project, North African and Khoe-San (Khoisan) populations in principal component analysis (Extended Data Figs 3, 4, 5, and Supplementary Figs 1 and 2) suggested possible HG ancestry among southern Niger-Congo groups—highlighted by clustering towards the Khoe-San, in addition to confirming a cline towards Eurasian populations. ‘Unsupervised’ (that is, without including known information on individual ancestry) ADMIXTURE9 (https://www.genetics.ucla.edu/software/admixture/) analysis including the 1000 Genomes Project and Human Origins data sets (Fig. 1), also supported evidence for substantial Eurasian and HG ancestry in SSA (Fig. 1 and Extended Data Fig. 6).

To assess the effect of gene flow on population differentiation in SSA, we masked Eurasian ancestry across the genome (Supplementary Methods and Supplementary Note 6). This markedly reduced population differentiation, as measured by a decline in mean pairwise FST from 0.021 to 0.015 (Supplementary Note 6), suggests that Eurasian ancestry has a substantial impact on differentiation among SSA populations. We speculate that residual differentiation between Ethiopian and other SSA populations after masking Eurasian ancestry (pairwise FST = 0.027) may be a remnant of East African diversity pre-dating the Bantu expansion10.

Population admixture in SSA

Formal tests for admixture (the three population test or f3 statistic)11, confirmed widespread Eurasian and HG admixture in SSA (Supplementary Tables 2 and 3). Quantification of admixture (Supplementary Table 4, Supplementary Methods and Supplementary Notes 3 and 4) indicated substantial Eurasian ancestry in many African populations (ranging from 0% to 50%), with the greatest proportion in East Africa (Fig. 2 and Supplementary Table 4). Similarly, HG admixture ranged from 0% to 23%, being greatest among Zulu and Sotho (Fig. 2 and Supplementary Table 5).

Figure 2: Dating and proportion of Eurasian and HG admixture among African populations.
figure 2

The proportion and distribution of Eurasian and HG admixture among different populations across Africa, with approximate dating of admixture using MALDER (code was provided by J. Pickrell; see Supplementary Information).

PowerPoint slide

We found evidence for historically complex and regionally distinct admixture with multiple HG and Eurasian populations across SSA (Fig. 2 and Supplementary Note 5). Specifically, ancient Eurasian admixture was observed in central West African populations (Yoruba; 7,500–10,500 years ago), old admixture among Ethiopian populations (2,400–3,200 years ago) consistent with previous reports10,12, and more recent complex admixture in some East African populations (150–1,500 years ago) (Fig. 2, Extended Data Fig. 7 and Supplementary Note 5). Our finding of ancient Eurasian admixture corroborates findings of non-zero Neanderthal ancestry in Yoruba, which is likely to have been introduced through Eurasian admixture and back migration, possibly facilitated by greening of the Sahara desert during this period13,14.

We also find evidence for complex and regionally distinct HG admixture across SSA (Fig. 2, Extended Data Figs 7 and Supplementary Note 5), with ancient gene flow (9,000 years ago) among Igbo and more recent admixture in East and South Africa (multiple events ranging from 100 years ago to 3,000 years ago), broadly consistent with historical movements reflecting the Bantu expansion. An exploration of the likeliest sources of admixture in our data suggested that HG admixture in Igbo was most closely represented by modern day Khoe-San populations rather than by rainforest HG populations (Supplementary Note 5). Given limited archaeological and linguistic evidence for the presence of Khoe-San populations in West Africa, this extant HG admixture might represent ancient populations, consistent with the presence of mass HG graves from the early Holocene period comprising skeletons with distinct morphological features15, and with evidence of HG rock art dating to this period in the western Sahara16,17. In East Africa, our analyses suggested that Mbuti rainforest HG populations most closely represented ancient HG mixing populations (Supplementary Note 5), with admixture dating to 3,000 years ago, suggesting that HG ancestry here is likely to be older than previously reported18. The primary source of HG admixture in Zulu and Sotho populations was from Khoe-San populations (Fig. 2 and Supplementary Note 5), consistent with linguistic assimilation of click consonants among these populations.

Positive selection in SSA

We examined highly differentiated SNPs between European and African populations, as well as among African populations to gain insights into loci that may have undergone selection in response to local adaptive forces (Supplementary Methods). To account for confounding due to Eurasian admixture, we also conducted analyses after masking Eurasian ancestry (Supplementary Methods and Supplementary Note 6).

On examining locus-specific Europe–Africa differentiation, enrichment of loci known to be under positive selection was observed among the most differentiated sites (P = 1.4 × 10−31). Furthermore, there was statistically significant enrichment for gene variants among these, indicating that this differentiation is unlikely to have arisen purely from random drift (P = 0.0002). Additionally, we found no evidence for background selection as the primary driver of differentiation among these loci (Supplementary Note 7).

In addition to genes known to be under positive selection (for example, SLC24A5, SLC45A2 and OCA219,20, LARGE21 and CYP3A4/5) (Supplementary Fig. 3), we found evidence of differentiation in novel gene regions, including one implicated in malaria (for chemokine receptor 1, CR1) (Extended Data Fig. 8). CR1 carries the Knops blood group antigens and has previously been implicated in malaria susceptibility22 and severity23, with evidence suggesting positive selection in malaria-endemic regions24 (Extended Data Fig. 8). We also identified highly differentiated variants within genes involved in osmoregulation (ATP1A1 and AQP2) (Extended Data Fig. 8). Deregulation of AQP2 expression and loss-of-function mutations in ATP1A1 have been associated with essential and secondary hypertension, respectively25,26. Climatic adaptive changes in these gene regions could potentially provide a biological basis for the high burden of hypertension and differences in salt sensitivity observed in SSA27.

In contrast, overall differentiation among African populations was modest (maximum masked FST = 0.19) (Supplementary Fig. 4) and only 56/1,237 sites remained in the tail distribution after masking (Supplementary Methods, Supplementary Table 6). This suggests that a large proportion of differentiation observed among African populations could be due to Eurasian admixture, rather than adaptation to selective forces (Supplementary Note 6). Genes known to be under selection were notably enriched among the most differentiated loci after masking of Eurasian ancestry (P = 2.3 × 10−16). Among the 56 loci robust to Eurasian ancestry masking (Supplementary Table 6), we identified several loci known to be under selection (Extended Data Fig. 8), including a highly differentiated variant (rs1378940) in the CSK gene region implicated in hypertension in genome-wide association studies (GWAS)28. The major allele of rs1378940 among Africans was in complete linkage disequilibrium with the risk allele of the GWAS SNP rs1378942 (ref. 29), with the frequency of this allele highly correlated with latitude (r = −0.67), providing support for local adaptation in response to temperature as a possible mechanism for hypertension (Supplementary Fig. 5)30,31.

Comparing populations residing in endemic and non-endemic infectious disease regions (Supplementary Methods), we identified several loci associated with infectious disease susceptibility and severity. As well as the known sickle-cell locus related to malaria, this approach identified additional signals for genes potentially under selection, including the PKLR region32, RUNX333, the haptoglobin locus, CD16334, IL1035,36, CFH, and the CD28-ICOS-CLTA4 locus (Supplementary Table 7 and Extended Data Fig. 8)37. Similar comparisons for Lassa fever identified the known LARGE gene, as well as candidates associated with viral entry and immune response, including in the Histocompatibility Leukocyte Antigen region, DC-SIGN/DC-SIGNR38 (also known as CD209/CLEC4M), RNASEL, CXCR6, IFIH139 and OAS2/3 regions (Supplementary Table 7). For trypanosomiasis, we identified APOL140, as well as several loci implicated in immune response and binding to trypanosoma, including FAS, FASLG41,42, IL23R43, SIGLEC6 and SIGLEC12 (Supplementary Table 7)44. For trachoma, we identified signals in ABCA1 and CXCR6, which may be important for the growth of the parasite and host immune response, respectively (Supplementary Table 7)45,46.

Designing medical genetics studies in Africa

To inform the design of genomic studies in Africa, we addressed the following questions: (1) How well do current genotype arrays perform in African populations using existing reference panels for imputation? (2) Can these genotype arrays and reference panels identify and fine-map association signals in populations across Africa? (3) Can we improve imputation accuracy in African populations using a new African reference panel? and (4) What are the most cost-effective designs for large-scale GWAS in Africa?

The 1000 Genomes Project phase I integrated panel provided reasonably accurate imputation into the Illumina Omni 2.5M array in all populations (Supplementary Note 10). However, imputation accuracy was lower among Sotho, Zulu and Afro-Asiatic populations, possibly reflecting poor representation of some African haplotypes (including Khoe-San haplotypes) within the 1000 Genomes Project panel. These findings suggest that improvements in imputation accuracy across diverse population groups may require larger and more diverse reference panels.

We assessed the reproducibility and potential for fine-mapping association signals within Africa and globally at several disease susceptibility loci (Supplementary Methods, Supplementary Table 8 and Extended Data Fig. 9). Current genotype arrays and imputation panels allowed for identification of relevant association signals at most loci across populations in SSA, demonstrating that association signals are reproducible across populations in SSA (Extended Data Fig. 9 and Supplementary Figs 7–18). African populations are likely to provide better fine-mapping resolution around the causal locus (Supplementary Table 8). We highlight one example here: the sickle-cell anaemia locus (HBB)47, which is under positive selection owing to the protection the sickle cells confer against severe malaria. This locus showed marked heterogeneity in association signals across populations, reflecting different linkage disequilibrium patterns and allele frequencies among populations in SSA (Supplementary Figs 9 and 10). This pattern is probably the result of independent selection sweeps at this locus in different parts of Africa, leading to differences in hitchhiking rare haplotypes that attained high frequencies among different populations48. This suggests that these signatures are recent and occurred during or after the Bantu expansion, consistent with the hypothesis that the advent of agriculture and increased malaria transmission may have resulted in increased selection pressure49. However, in contrast to previous reports47, we show that association signals even at such highly differentiated loci can be captured with dense genotype data using existing reference panels for imputation, despite individual population groups not being fully represented in these. This suggests that, instead of large-scale population-specific sequencing across Africa, what is needed is a broad sequencing approach, targeted at capturing widespread haplotype diversity.

To assess the utility of a larger and more diverse African reference panel for imputation, we generated a panel integrating the 1000 Genomes Project phase I and AGVP WGS panels (Supplementary Methods and Supplementary Note 9). Using this integrated panel, we observed marked improvements in imputation accuracy across the whole range of the allele frequency spectrum in specific populations poorly represented by the 1000 Genomes Project panel (Fig. 3 and Supplementary Note 11). These findings suggest that even common haplotypes in some SSA populations may not be sufficiently captured by existing panels, limiting our power to examine associations of common variants with disease. Importantly, given the specificity of the improvement in imputation accuracy, we infer that targeted sequencing of divergent populations representing a broad spectrum of haplotypes across Africa, including HG and North/East African haplotypes, rather than widespread population sequencing is likely to provide a more efficient strategy to improve imputation accuracy and a practicable GWAS framework in Africa.

Figure 3: Improvement in imputation accuracy with the AGVP WGS panel.
figure 3

The substantial improvement in imputation accuracy in some populations (Sotho), compared to minimal improvement in others (Igbo) with the addition of the AGVP WGS reference panel to the 1000 Genomes Project phase I reference panel (‘merged’) suggests poor representation of some haplotypes (for example, Khoe-San haplotypes in Sotho) in the 1000 Genomes Project reference panel alone (‘1000’). r2 is the correlation coefficient, representing the correlation between imputed and genotyped data, on masking each genotyped variant during imputation. MAF, minor allele frequency.

PowerPoint slide

We compared the utility of existing chip designs (2.5M Illumina) and ultralow-coverage WGS designs (0.5×, 1×, 2× coverage) to determine the optimal design for African GWAS. Sensitivity for common variation was >90% at all sequencing depths (Supplementary Note 12). Examining the effective sample size for a fixed budget50, we found the effective sample size was greater for all ultralow-coverage WGS and chip array designs compared with 4× WGS. When computational costs were accounted for (Supplementary Note 12), the HumanOmni2.5M array provided the greatest effective sample size supporting the development and large-scale use of efficient genotype arrays in Africa, where these have been underutilized.

We therefore sought to evaluate a potential chip design to tag common variation across a wider range of African populations (Supplementary Note 13). Importantly, we show that an array with one million genetic variants could capture >80% of common variation (minor allele frequency >5%) across the genome (Extended Data Fig. 10). These analyses suggest that designing a pan-African genotype array to effectively capture common genetic variation across Africa is feasible, and could greatly facilitate large-scale genomic studies in Africa.

Discussion

The marked haplotype diversity within Africa has important implications for the design of large-scale medical genomics studies across the region, as well as studies of population history and evolution. In this context, the AGVP is a resource that will facilitate a broad range of genomic studies in Africa and globally.

Although Africa is the most genetically diverse region in the world, we provide evidence for relatively modest differentiation among populations representing the major sub-populations in SSA, consistent with recent population movement and expansion across the region beginning around 5,000 years ago—the Bantu expansion8. Although the history of the Bantu expansion is probably complex, assessments of population admixture can provide new insights. We note historically complex and regionally distinct admixture with multiple HG and Eurasian populations across SSA, including ancient HG and Eurasian ancestry in West and East Africa and more recent complex HG admixture in South Africa. As well as explaining genetic differentiation among modern populations in SSA, these admixture patterns provide genetic evidence for early back-to-Africa migrations, the possible existence of extant HG populations in western Africa—compatible with archaeological evidence15, and patterns of gene flow consistent with the Bantu expansion, including genetic assimilation of populations resident across the region.

This admixture also has important implications for the assessment of differentiation and positive selection in Africa. Accounting for these elements, we have identified loci under positive selection that are linked with hypertension, malaria, and other pathogens. This provides a proof-of-concept for the ability of geographically widespread genetic data within Africa to identify loci under selection related to diverse environments.

Our evidence for the broad transferability of genetic association signals and their statistical refinement has important implications for medical genetic research in Africa. Importantly, we highlight that such studies are feasible and can be enabled through the development of more efficient genotype arrays and diverse WGS reference panels for accurate imputation of common variation. In this context, we describe a framework for a new pan-African genotype array that could directly facilitate large-scale genomic studies in Africa.

A critical next step is the large-scale deep sequencing of multiple and diverse populations across Africa, which should be integrated with ancient DNA data. This would enable us to identify and understand signals of ancient admixture, patterns of historical population movements, and to provide a comprehensive resource for medical genomic studies in Africa.