Populations speaking languages belonging to the Niger-Congo phylum have been the object of several studies, some of which aimed to assess the patterns associated with the diffusion of Bantu languages [13, 21, 28, 29, 44, 45]. This is the phylum containing the highest number of languages worldwide and genealogical classification of its families is still under debate . However, there is a consensus on the fact that western Atlantic and Mande are more ancient than central Benue-Congo and Bantu branches, while the emergence of Kordofanian remains unclear [26, 46, 47]. When autosomal variation is analysed, only a slight substructure among the populations belonging to the entire phylum is observed . By increasing both the number of populations and the geographical coverage, we were able to obtain new insights into the relations among Niger-Congo speakers.
The populations included in our dataset speak languages belonging to several sub-branches of the NC family (see Additional file 3: Figure S1b) and are scattered through a vast area of sub-Saharan Africa, which mainly includes two habitats: the savannah and the rainforest. Roughly speaking, the first prevails in the region from Senegal to Northern Cameroon while the second characterizes most of the areas corresponding to Southern Cameroon, Gabon and Congo. Climatic studies have shown that after the phenomenon known as the Younger Dryas (11.5 ± 0.25 ka B.P; [48, 49]), the climatic conditions in the sub-Saharan region became less arid and the distribution and density of the rainforest have remained stable for the last 9.5 ky . The peopling of the sub-Saharan region is likely to have increased since then and the populations here considered have probably been in contact within the same time frame.
Given the shared traditional patrilocal habit of the populations under study, we were able to focus on mtDNA variation as the source of genetic information for microevolutionary inference. By combining a multivariate approach with the test of specific migration patterns, we were able to detect a complex structure among the populations under study, which seems to be better explained by the effect of local environmental factors rather than the internal linguistic complexity of the NC phylum.
After testing three migratory models (Figure 3), we observed that the stepping-stone model better describes the distribution of mtDNA variation throughout the whole region. This may indicate a general tendency of women to spread out from their villages with the intensity of the migration decreasing with distance, so that only neighbouring groups share common genetic variation. The isolation by distance (IBD) pattern observed in our sample is in agreement with previous studies which showed that geographic distances better explain genetic differences among human populations than ethnic affiliations [19, 51].
Apart from this general indication, the analysis of mtDNA variation allowed us to identify two main groups quite clearly, with the rainforest populations being more structured and diverse than the savannah groups. In fact, the former populations are characterized by higher values of molecular measures of within-population diversity (see for example the MNPD in Table 1), larger genetic distances and phylogenetic trees with longer branches, and a lower proportion of different haplotypes (corresponding to Central in Table 1, and to clusters 2,7 and 4 in Figure 2). The analysis of genetic structure detected the main signal of differentiation in this group, separating clusters 4, 2 and 7 from the others. The two groups also show a significant difference in the distribution of their Fs values, with rainforest populations showing a less negative average (one tailed t-test for mean comparison, p-value = 2.3e-10) as well as including 5 out of the 7 populations with non-significant Fs values (Table 1), suggesting a less important role of demographic expansions in their evolutionary history. The Fu's test, and other statistics relying on haplotype frequencies, were found to be more sensitive for detecting expansions on nonrecombining genomic regions than Tajima's D and other tests . This signature of genetic drift could have been enhanced by the reduced effective population size of the mtDNA compared to autosomal loci, but it is unlikely to have generated the non random genetic structure observed here.
The signature of IBD detected within the savannah region is higher than the one in the rainforest, and indicates, together with the observations of a lower degree of isolation among the former, that the migratory patterns are more straightforward to interpret in the savannah than in the forest. Therefore, we may conclude that although geographic factors have a role in both areas, for the savannah this can be simply described as a linear correlation between physical and genetic distances, while for the rainforest the role played by environmental factors is probably more complex. This conclusion highlights the usefulness of explicit geographic models in trying to understand human genetic diversity, which has been previously suggested by Ray and Excoffier (2009) .
As an important evolutionary consideration, we should take into account the possibility that differences in Fu's statistical values between savannah and rainforest could be also explained by the role of selection. However, although the worldwide distribution of mtDNA lineages has been proposed to be driven by selective processes related to temperature changes, the geographic region here analysed appears to be quite homogeneous for this putative temperature effect . In future studies, researchers should consider that other climatic parameters which are different in the savannah and rainforest environments have yet to be explored.
Another caveat of the present study may be the a priori definition of population units, based on the sampling location and the languages spoken by the individuals. We are aware that such a definition may lead to an approximation in the estimate of the spatial distribution of allele frequencies, since each population is considered as a sampling point. In the present case, we believe that, despite the vast geographical area covered by our dataset, the homogeneous nature of sampling helps overcoming this limitation and is allowing a reliable representation of the distribution of maternal lineages.
The complexity of the migratory patterns observed here is further emphasized by a discontinuity detected between clusters 7 and 2, which overlaps with a broad area of the rainforest region (encompassing Cameroon, Gabon, Congo and Central African Republic) where the sampling coverage is fairly homogeneous. Cultural factors do not seem to offer an explanation for this separation. In fact, the populations composing the two clusters speak languages that are closely related, within the Narrow Bantu family and show no major differences in their subsistence economy. On the other hand, environmental factors could have played a role if one considers that the rainforest habitat may decrease the intensity of gene flow among populations after their initial settlement in deforested areas, making migration more difficult. Another discontinuity in the pattern (between clusters 1 and 6) overlapped with a gap in the sampling coverage of the dataset under study, corresponding to the area of Guinea, the Ivory Cost and Liberia, where tropical rainforest vegetation generally prevails. In the absence of these samples, any further inference on the validity of the observed discontinuity would be very speculative. However, their analysis could contribute to a more exhaustive testing of the influence of different environments on the intensity of migrations among human populations.
Considering all the previous observations, we suggest that farming rainforest populations have probably undergone a local, more recent, and less intense demographic expansion than other food producer populations of the Niger-Congo phylum, which has been previously observed in Gabon through the analysis of Y chromosome lineages . Evidence of ancient peopling should also be taken into account when interpreting genetic data. In fact, central Africa is characterized by a well-defined succession of Middle Stone Age industries while western Africa seems to have been populated at very low densities until 10–12 kya [47, 55]. Rainforest farmers have also been shown to share both recent and ancient genetic backgrounds with hunter-gatherer populations [56–60].
It is interesting to note the unexpected association observed in cluster 6 where populations of nomadic shepherds from Northern Cameroon (Tali, Tupuri and FulbeC; see also MDS plot in Figure 1) were grouped together with Western groups. Complex relationships among Cameroon ethnic groups have already been reported in previous studies [21, 24, 29, 61]. Although the intermediate model we tested was not the best supported by the analysis, it actually detected high migration rates from cluster 6 to clusters 3 and 4 (data not shown). Mixed hierarchical models of migration combined with a better knowledge of the nomadic routes followed by these populations would be worth investigating in order to clarify our findings.
Focusing on the genetic variation of Niger-Congo-speaking populations, our results highlighted a stronger structure among the populations settled in the Central area, which correspond to the Bantu-speaking groups. In fact, populations settled in Nigeria and Ghana (clusters 3 and 1) and Guinea Bissau and Senegal (clusters 6 and 5), which present a high linguistic diversity, seem to be characterized by a rather continuous gene flow and show smaller inter-population differences. This contradicts the expectations described above, based on linguistic data, of a recent demic expansion from the area of Nigeria-Cameroon towards Central Southern and Eastern Africa, and an earlier diffusion from Western to Central Africa .
As a general conclusion, language does not seem to be the main predictor for the distribution of genetic variation among Niger-Congo-speaking populations. Despite the general belief that language is transmitted by migrating women, genetic analyses have repeatedly shown its preferential correlation with paternal rather than maternal genetic variation [12, 13, 21].
Unfortunately, we were unable to find a reliable approach for the definition of linguistic distances. Comprehensive classifications based on a quantitative measure of lexical similarities are only available for the Narrow Bantu languages, and not systematically for other Niger-Congo branches (Koen Bostoen, personal communication). Since in this dataset only 28/85 populations belong to the Narrow Bantu family, we decided to avoid this approach in order not to introduce interpretation biases due to inaccurate or questionable linguistic classification.
Even though the genetic clusters here reported cannot be considered as random mating units, the picture presented in our study suggests that, in particular thanks to female-biased movements, gene flow occurs among human populations speaking very different languages.
The analysis of paternal patterns of migration would be useful to shed light on the substructure and the random mating areas among patrilocal populations, while autosomal and X-chromosomal data could be productively investigated to explore whether sex-biased movements are detectable in the distribution of genome variation.