Introduction

Tropical forests are characterized by a high diversity of plants, with tall trees, dense canopies and low light penetration [1]. Their climate is generally warm with minimum temperatures well above the freezing point and mean annual rainfall above 1,000 mm [2]. Despite being one of the most productive environments of the world, tropical forests provide only few resources for humans [3]. Indeed, in these environments plants invest most of their energy in structure maintenance and not into the reproductive organs that are the most edible parts for humans and their prey species [2]. In addition, the instability of food resources in response to the high seasonality of rainfall raises the costs of foraging, further reducing its capacity to support human life [2,3].

Besides food limitation, other characteristics of tropical forests may also contribute to the hostility of these environments. For instance, tropical areas harbor on average 70% higher human pathogen diversity as compared to more temperate areas [4]. As a consequence, infant and child mortality rates among tropical forest dwellers should be high [5]. Moreover, the small differences between air and skin relative humidities and high temperature, coupled with little air movement, make sweat production and evaporation difficult in tropical forests, potentially compromising thermoregulation [6].

Due to the hostility of this environment, it is unclear whether humans have ever subsisted in tropical forests without depending on external resources, such as agriculture or possible exchanges with neighboring populations. Evidence of societies living in such harsh conditions is scarce for contemporary modern humans [2], as well as for early Homo [7]. Nonetheless, it is possible that humans have developed recent biological adaptations to tropical forests. A few examples of such adaptations have indeed been documented, the most well-known being the pygmy phenotype, defined by Perry and Dominy [6] as small human body size (mean adult male height < 155 cm). These authors argue that short-statured individuals may have advantages to cope with food limitation, thermoregulation, and mobility hardship in a dense forest and, with few exceptions, are thus found in hunter-gatherer populations living in tropical rainforests of Africa, Asia, Oceania, and the Americas [6]. However, it has also been suggested that this phenotype could be a by-product of selection for early onset of reproduction [8], which could enable populations to overcome problems related to their life history and increased mortality [9].

To investigate whether tropical forest dwellers have developed specific biological adaptations to this harsh environment, we searched for genome-wide signals of positive selection in populations from the Americas and Africa, specifically aiming at identifying convergent evolution signals, that is a significant signal of positive selection occurring at the same genomic region or biological pathway in populations belonging to two distinct evolutionary lineages. To that effect, we investigated populations living in tropical forests and others, genetically related, living outside these environments using publicly available genome-wide SNP data and a robust and sensitive F_ST-based method for inference of positive selection that explicitly includes a convergent selection model.

Subjects and Methods

Results

According to the F_ST-based Bayesian approach implemented via BayeScan [20, 21] 1,482, 1,222, 1,579, and 1,365 outlier SNPs were identified in populations sets PS1 to PS4 respectively, ranging from 0.21 to 0.27% of the SNPs analyzed. From these, 75 SNPs are significant regardless of which PS is analyzed. There were no cases in which balancing selection could be inferred, i.e. α never reached a significant negative value, which may be due to the characteristics of the genetic system employed (e.g. low polymorphism [20] or a possible ascertainment bias), to the lack of power to detect markers under balancing selection [23], or to the low number of populations analyzed [20], but not necessarily to the absence of this phenomenon in the evolutionary history of these populations.

Using a threshold of 50 kb, 568, 474, 620, and 517 genes were associated with those significant SNPs from PS1 to PS4 respectively, of which 57 genes were found to be co-occurring in all four PSs (namely ABLIM3, ACSS2, AKAP6, ANKRD26, ARHGEF10, ATIC, BCAT1, C20orf111, C2orf73, CBLN1, CNTN4, CNTNAP5, COL22A1, CPA5, CRTC3, CWH43, DCUN1D4, DHCR7, EPHB4, FAM188B, FKBP6, GALNT16, GLIS3, GLRB, GPC6, GRIK2, HLA-DPA1, HMG20B, HSF2, IQGAP1, KIAA1598, KLHL29, LRRC66, MASTL, MPST, NAALADL2, NRG1, NRP2, NSUN5, PARK2, PKIB, PPP2R2C, RABGGTB, RAD51B, RBFOX1, RBM9, RBMS3, RFX3, ROBO2, SCP2, SGCB, SHISA6, SPAG16, SPATA13, SPATA18, ST6GAL1, and TRG@). We confirmed with a random permutation test (N = 50,000) that the number of outlier SNPs (75) and genes (57) shared by all population sets is higher than expected by chance (p-value < 2e-5).

The sliding-window procedure yielded 440, 399, 505, and 471 SNPs with a significant signal of positive selection for each PS (S1 File). From these, 19 are found to be functional non-synonymous mutations (missense) in one or two PSs (S2 Table). There was no instance of a significant SNP defined as nonsense, frameshift, or splice-site mutation.

Outlier SNPs are significantly enriched (Fisher exact test, p-value < 0.022) for genic SNPs in comparison to the whole HGDP SNP set. On average 72.8% of the outlier SNPs after the sliding-window procedure are located in or less than 50 kb away from protein-coding regions considering all four PSs, while for all SNPs in the HGDP database this proportion is only 63.6%. This enrichment is not due to potential LD between outlier SNPs, as it still holds when randomly sampling one outlier SNP and one non-outlier SNP per 100 kb window (N = 1000, mean p-value = 0.02, 90% CI [2.3e-4, 9.1e-2]). Moreover, all outliers were located at maximum 45.8 kb apart from a gene or from a putative regulatory element indicated by a hypersensibility to DNaseI.

Fig. 1 shows the Manhattan plots of the physical position of the SNPs in the genome and their estimated q-values for the different PSs, as well as the best model for selection (color-coded) in the sliding-window approach. It was possible to identify seven clusters of outlier SNPs co-occurring in all four PSs (indicated by gray-shaded boxes at Fig. 1 and at S1–S6 Figs.). Their coordinates, size, and genic content are described in Table 1.

Download:

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

Fig 1. Manhattan plots of the physical position of SNPs (x-axis) and corresponding q-values in log10 scale (y-axis) for inferring selection.

Based on a sliding-window approach, SNPs are color-coded according to the best supported model of selection, namely neutrality (black or grey), in Africa (blue), in the Americas (green), and in both continents (convergent evolution, red) with a False Discovery Rate of 0.1 (dashed line). Different sets of populations were used in the analysis yielding four different population sets (PS1–4). Congruent clusters of outlier SNPs considering all four PSs are highlighted with a grey box.

https://doi.org/10.1371/journal.pone.0121557.g001

Download:

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

Table 1. Coordinates, size, and genic content of the seven clusters of outlier SNPs occurring in all four population sets (PSs) considered in the analysis.

https://doi.org/10.1371/journal.pone.0121557.t001

For the first two clusters, the most likely model of selection for the majority of outlier SNPs is positive selection in the Americas; however, for a few SNPs convergent evolution cannot be ruled out (Fig. 1). While the first cluster includes the gene SCP2 (S1 Fig.), the second cluster lies in an inter-genic region (S2 Fig.). The next two clusters of significant SNPs occur on the same chromosome and selection in Africa is the best-supported model in most cases, although convergent selection is likely for a few SNPs (S3 Fig.). Cluster 3 is associated with gene CWH43, while cluster 4 includes SNPs that fall into genes such as DCUN1D4, LRRC66, SGCB, and SPATA18 and their vicinity. Convergent evolution is the best-supported model of selection for the majority of SNPs of the remaining three clusters (S4–S6 Figs.). Cluster 5 is associated with the following three genes: CCL28, C5orf34, and NNT; which may or may not present a significant signal of selection depending on the analyzed PS (S4 Fig.). Cluster 6 includes genes HSF2 and PKIB for all PSs, and the analysis of PS1 and PS3 revealed an additional significant SNP located at SERINC1. The last cluster (no. 7) includes two SNPs, one located in a FKBP6 intron and the other located 0.12 kb apart from NSUN5 and 3.55 kb from TRIM50. The abovementioned 14 genes that include or are close (<50 kb) to SNPs that present signals of positive selection in the sliding-windows approach for all four PSs are described in Table 2.

Download:

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

Table 2. Genes with signals of positive selection suggesting human adaptations to tropical forests in Africa and the Americas.

https://doi.org/10.1371/journal.pone.0121557.t002

Five of the seven clusters showing signs of positive selection with the HGDP data sets also show signals of positive selection in the supplementary analysis. The exceptions are cluster 3 on chromosome 4, which was not covered by any SNP in the smaller data set, and cluster 7 on chromosome 7, for which the signal of positive selection disappears after replacing Pima by Zapotec. In the supplementary analysis, no additional clusters with signal of positive selection were found in all SPSs. The type of natural selection—convergent evolution or selection in a single continent—inferred from the supplementary data set is generally concordant with the analysis done on HGDP data, except for cluster 6 on chromosome 6, where selection in Africa is more important in the supplementary analysis. After performing the sliding-window approach, cluster 1 appears for SPS1 and SPS3, clusters 4 and 5 for all SPSs, and cluster 6 for SPS2 and SPS4.

Fifteen genomic regions had been found to be associated with the pygmy phenotype by means of covariation between allele frequencies and body height in Africa in another study [35]. We found signals of positive selection in the African continent in four of these regions. By comparing Mbuti to West Africa, Mendizabal et al. [35] identified two genomic regions as candidate for selection. One of them, located in the long arm of chromosome 10, is among those with the highest q-values in our study (Fig. 1). It includes 14 outlier SNPs in the following 11 genes: P4HA1, NUDT13, ECD, FAM149B1, DNAJC9, MRPS16, ANXA7, ZMYND17, USP54, PPP3CB, and TTC18. Two outlier SNPs in this region—rs2271904 and rs4294502—are both non-synonymous mutations in ECD and TTC18 respectively. The second region is defined by two SNPs (rs7174731 and rs7181518) in TRIP4. In a third region, we found two genes with positive selection signals considering West Africans in comparison to Biaka instead of Mbuti: USP46 and MLL3. An additional region suggested by Mendizabal et al. [35] to be associated with the pygmy phenotype presents signals of convergent adaptation using all four different combinations of populations in our study. They found significant SNPs in NNT, while we observed a more diffuse significant signal also including CCL28 in some cases (S4 Fig.).

As selection is acting on phenotypes instead of genotypes, we applied a gene set analysis [30] to detect pathways involved in adaption to living in tropical forests. We found 43 pathways scoring a q-value below 20% for at least one of the population sets and selection models (S7 Fig.), most of them scoring significant for selection in Africa. Among those we find a large cluster of 22 gene sets involved in immune response, nine of which show signals of convergent selection, and a cluster of 11 gene sets involved in RNA Polymerase III transcription initiation. In order to eliminate redundancy, we assigned overlapping genes to the highest scoring pathway. After this ‘pruning’ procedure, 14 pathways remain significant, including “Apoptosis” and “Cholesterol Biosynthesis” among other immune and nervous-system related pathways (Fig. 2). Four of these pathways present a consistent signal of positive selection across different PSs or selection models, namely “PD1-signaling”, “IL 12-mediated signaling events”, “RNA Polymerase III Transcription Initiation From Type 3 Promoter”, and “Chemokine receptors bind chemokines”. Note that taking a q-value threshold of 0.1 would still result in 8 significant gene sets, of which 3 show multiple signals of adaptation (data not shown).

Download:

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

Fig 2. Gene sets significantly enriched for selection signals.

Nodes are gene sets that score a q-value ≤ 20% after pruning in the gene set enrichment test (See S7 Fig. for the equivalent plot regarding the results before pruning). Gene sets are connected when at least one of them shares >33% of its genes with the other set before pruning. The size of a node scales with the size of the gene set. Each quadrant of a node represents results on one of the four selection scores, while the four rings in a node correspond to the four population sets (different combination of populations; see Methods). The color of each quarter of a ring corresponds to the significance of the test result (dark colored: q≤10%, light colored: 10%<q≤20%, white: q>20%).

https://doi.org/10.1371/journal.pone.0121557.g002

Discussion

Tropical forests are believed to be very harsh habitats for human beings [3]. In addition to being almost deprived from energy-rich food and edible plants [2], these environments are very propitious for the development of diseases [4] and might also compromise thermoregulation [6]. In this work we used a Bayesian method [20, 21] to identify SNPs in a genome-wide dataset showing overly large or low extent of differentiation between tropical and non-tropical forest populations. We sought to infer signals of convergent evolution by comparing native populations from tropical forests (Biaka, Mbuti, Karitiana, and Surui) with genetically related populations living elsewhere (Mandenka, Yoruba, and Pima) combined in four different population sets (PSs). The analysis suggested some SNPs, genes, and a few biological pathways in which convergence could be inferred. These outlier SNPs are found to be enriched for genic loci even after correcting for a potential LD bias. Those few cases that fall within an intergenic region, such as Cluster 2 (S1 Fig.), could always be associated (distance < 50 kb) to a putative regulatory element described by the ENCODE project. These results suggest that regulatory regions could also be involved in some recent human adaptations [36]. Genomic regions where the same signal of selection could be identified by employing different combinations of populations (clusters 1–7 on Fig. 1) suggest that the inferred signal is due to environmental selective pressures and adaptation rather than to the demographic history of the populations and those regions are the main focus of our discussion. For five out of the six clusters with sufficient SNP coverage in the smaller supplementary dataset, we observed significant SNPs with the same type of selection—convergence or selection in one continent—as observed in the analysis of the HGDP data. The exception is cluster 7 where the signal of convergent adaptation does not hold when we replace Pima by Zapotec. In this case, selection in Africa is the best-supported model. Based on the agreement between results generated with the two datasets, results for clusters 1, 2, and 4–6 should thus be considered as very robust, but results obtained for cluster 7 should be considered with some caution. We can see two reasons for an absence of congruent results for cluster 7. First, it could be that our supplementary analysis did not have enough power to detect signal given the lower number of SNPs available in Zapotec. Else, our initial results for cluster 7 could be due to the own evolutionary history of the Pima and have nothing to do with adaptation to tropical forests. This could be due to an environmental pressure shared between West African populations and Pima, but not present in Zapotec. It appears difficult to distinguish between these hypotheses without having more samples genotyped for a larger number of SNPs.

In a previous study involving Native Americans, Hünemeier et al. [37] suggested that ABCA1, a gene encoding a cholesterol efflux regulatory protein, was evolving under positive selection due to limited food resources that Native Americans encountered during their history. We also found significant signals of positive selection in genes that are related to lipid circulation and metabolism, such as SCP2 and CWH43 (Table 1) and in a genetic pathway related to cholesterol biosynthesis (Fig. 2 and S7 Fig.). Besides playing a role in nutrition, these genes could also be involved in immunological response, since cholesterol plays an important role in various infectious processes such as virus invasion and replication [38] as well as in resistance against malaria [39].

In this regard, Sabeti et al. [40] already noticed a preponderance of genes related to the immune system in available genome-wide scans for positive selection. This prevalence was further confirmed by Williamson et al. [41], López Herráez et al. [42], and Daub et al. [30]. Besides the two genes mentioned above—SCP2 and CWH43, which have a possible role in immunity—the protein encoded by CCL28 modulates immunity to HIV infection and skin-related inflammatory diseases. Additionally, the selective pressure of this class of genes may also be important at the multi-genic level, since a network of 22 pathways involved in immune response presents signals of positive selection (S7 Fig.). After pruning, two of these pathways remain significant for more than one PS, namely “PD-1 signaling” and “IL 12-mediated signaling events”. This indicates that they both have independently a strong selective signal, which would deserve further investigations for their role in adaptations to tropical environments.

Another category of genes that frequently presents signals of positive selection is fertility, more specifically, spermatozoid development [40]. In this regard, FKBP6, a male fertility factor, is found here to be potentially evolving under positive selection in our analyses (S6 Fig.), but this observation has to be taken with caution as this gene is not identified as an outlier in our supplementary analysis when Ameridian forest populations are compared to Zapotec instead of Pima.

Heat-shock transcription factors, such as that encoded by HSF2 (Cluster 6; S5 Fig.), are activated by stress and respond to elevated temperatures. One of the consequences of inefficient thermoregulation is the increase of body temperature. The observed positive selection signals at SNPs found in this gene could be due to an adaptation to the tropics, initiating gene(s) transcription in response to high body temperatures. Another study with African-, European-American, and Chinese populations also found a significant signal of positive selection in heat shock genes [41], suggesting that this category might have some importance in human adaptation to different environments worldwide.

It is generally accepted that the pygmy phenotype might have evolved as an adaptation to life in dense tropical forests, to thermoregulation, and to food scarcity [6] or as a by-product of selection for early onset of reproduction [8]. Our research design enables the comparison of two African pygmy populations with two other non-pygmy populations from the same continent, from which inferences can be drawn about the differences in selective pressures that African pygmy and non-pygmy populations are subjected to. Clusters 3 and 4 (S3 Fig.) are the two main regions where we found signals for positive selection in Africa. The first cluster presents a gene involved in lipid metabolism (CWH43, discussed above) and the second includes the genes DCUN1D4, LRRC66, and SGCB, which are found in a region known to be associated to severe limb-girdle muscular Duchenne-like dystrophy [43].

Other regions of interest for the study of the pygmy phenotype are those four in which we found positive selection signals and are associated with the pygmy phenotype according to another study [35]. From the genes found in this region, three are notable for being also associated with height, bone development or the pygmy phenotype in previous studies. Those genes are PPP3CB, which encodes a subunit of calcineurin, a protein that regulates bone formation by osteoblast differentiation [44]; TRIP4, a gene with positive selection in pygmy populations [45] that relates to a category (thyroid hormone receptor) that was suggested to be associated with this phenotype [42]; and MLL3, involved in histone modification, a category associated with height in a genome-wide association study [46].

In general, non-convergent positive selection signals were more pervasive than signals of convergent evolution at the SNP, gene and pathway levels. This suggests that the similar selective pressures imposed by the tropical forest environment in Africa and the Americas have targeted different genes and pathways. This could be a result of different genetic background of the populations from these two continents and also a result of the differences in the time-scales that adaptation took place, since the Americas were peopled more recently than Africa [16–18]. As it is particularly true for the pygmy phenotype [45], putative adaptations to the tropical forests could be the result of selection acting on different genetic targets even though their result might be similar at the phenotypic level. Nonetheless, due to the power limitation associated with detecting convergent evolution [21], this model cannot be ruled out for those regions where at least one SNP was assigned to it among other SNPs assigned to a model of positive selection in one continent, which is the case of Clusters 1 to 4. Moreover, the scarcity of convergence examples could also be a result of the SNP density of the array employed, since the power to detect selection is determined by physical linkage between the sampled SNPs and the causal variant. In addition, due to different time-scales that such events took place, in populations such as the Africans linkage is more likely to be disrupted than in a younger population such as the Amerindians, which makes it necessary to use a denser SNP-array in order to rule out false-negatives in African populations. In other words, we might be missing some actual convergence signals due to a lack of sufficient linkage between the target region and the sampled SNPs in Africa and convergence might thus be underestimated in this continent. Therefore, it is likely that additional genes evolving under convergent evolution in these populations could be found by using different methods and genetic systems.

Conclusions

The F_ST-based hierarchical Bayesian method used in our study enabled us to detect a number of regions with positive selection, suggesting that the following biological functions and pathways may play a role in human adaptations to tropical forest: lipid metabolism, immunology, body development, and heat stress response. The same signals found in different population sets suggest that they are due to environmental adaptation rather than to the demographic history of the sampled populations. Further refinement of these analyses with e.g. full genome or exome sequence information could reveal which particular mutations are responsible for these adaptations. Moreover, the few cases in which convergent evolution could be inferred contrast with the larger amount of genes with non-convergent positive selection signals, suggesting that Africans and Amerindians may have followed different routes to adapt to similar environmental selective pressures.

Supporting Information

Acknowledgments

The authors would like to thank Drs. Maria Luiza Petzl-Erler, Eduardo Tarazona Santos, Nelson J. R. Fagundes and the three anonymous reviewers for their valuable comments on previous versions of the manuscript, Prof. Andres Ruiz-Linares for kindly making available the Zapotec dataset, and Prof. Sandro L. Bonatto for his contribution to the project. Detailed results regarding the list with all outlier SNPs before and after the sliding-window approach, their nearest gene (if located in a genic region), q-value, and best-supported model of selection as well as the table with gene set enrichment results are available upon request.