The genomic footprint of social stratification in admixing American populations
Abstract
Cultural and socioeconomic differences stratify human societies and shape their genetic structure beyond the sole effect of geography. Despite mating being limited by sociocultural stratification, most demographic models in population genetics often assume random mating. Taking advantage of the correlation between sociocultural stratification and the proportion of genetic ancestry in admixed populations, we sought to infer the former process in the Americas. To this aim, we define a mating model where the individual proportions of the genome inherited from Native American, European, and sub-Saharan African ancestral populations constrain the mating probabilities through ancestry-related assortative mating and sex bias parameters. We simulate a wide range of admixture scenarios under this model. Then, we train a deep neural network and retrieve good performance in predicting mating parameters from genomic data. Our results show how population stratification, shaped by socially constructed racial and gender hierarchies, has constrained the admixture processes in the Americas since the European colonization and the subsequent Atlantic slave trade.
Editor's evaluation
In this important study, the authors develop a neural network to investigate assortative mating and sex-bias in admixed populations from the Americas. Applying their method to modern-day human genomes, they estimate sex-biased admixture and ancestry-based assortative mating. The evidence supporting their claims is solid, and their results will be of interest to population geneticists, anthropologists, and those interested in the history of the Americas.
https://doi.org/10.7554/eLife.84429.sa0Introduction
Ancestry-related assortative mating and sex bias in admixture models
Assortative mating is the phenomenon whereby mates resemble each other more than would occur under random mating. Through the lens of population genetics, assortative mating entails the correlation of genetic variants between mates (Versluys et al., 2021). Assortative mating is often a consequence of a subdivided population structure, observed in most human populations (Cavalli-Sforza and Feldman, 1981; Nagoshi et al., 1990; Sebro and Risch, 2012; Sebro et al., 2017). In many species, assortative mating is also the product of active mate choice, an adaptive behavior by which individuals choose a genetically similar mate by their phenotype (Merrill et al., 2019; Versluys et al., 2021). Whether or not active mate choice significantly takes place in human populations is unclear, as it is challenging to discern the effect it might have beyond that of population structure (Eshel and Cavalli-Sforza, 1982; Xie et al., 2015; Sebro and Risch, 2012; Abdellaoui et al., 2014; Sebro et al., 2017).
Geography structures populations of most species, including humans. Individuals separated by shorter distances interact more than individuals separated by longer distances or geographical features like watercourses or mountain ranges, which imply limited mating as the geographical barriers increase (Wright, 1943; Malecot, 1948; Kimura and Weiss, 1964; Cavalli-Sforza et al., 1996; Novembre et al., 2008). The reduced gene-flow between geographical groups leads to differentiated processes of genetic drift and results in distinguishable genetic profiles known as genetic ancestries (Mathieson and Scally, 2020; Lewis et al., 2022; Coop, 2022).
In human populations, culture also shapes the population structure. Socioeconomic and cultural barriers, which might have a certain permeability, limit the interaction between human groups. These social groups with a common language, religion, socioeconomic status, etc. define overlapping subpopulations where mating takes place mostly within them (Manni, 2010; Campbell, 2015; Matsumae et al., 2021).
In the case of migration, two or more populations cohabit in the same location and eventually admix to become subpopulations of a newly admixed population. The admixture process does not take place randomly but it is constrained by barriers set by socioeconomic and cultural differences between subpopulations with their own distinguishable profile of genetic ancestries (Risch et al., 2009; Nagoshi et al., 1990; Sebro et al., 2017). These barriers are socially constructed and, particularly in colonial contexts, their permeability is often politically restricted (McLean, 2021).
In admixed individuals, the genetic ancestry related to each source population can be tracked along the genome and expressed in individual-based proportions.
Therefore, in recently admixed populations, the population structure driven by culture and socioeconomic differences is associated with differences in the proportions of genetic ancestry. As a consequence, the proportion of genetic ancestry between mates correlate. This phenomenon is defined as ancestry-related assortative mating (Burrell and Disotell, 2009; Bryc et al., 2010; Norris et al., 2019). In addition, individuals might show a preference towards mating partners of the opposite sex with lower or higher ancestry proportions, which is defined as ancestry-related sex bias (Goldberg and Rosenberg, 2015).
In admixed populations, the length of the continuous ancestry tracts is widely used to infer the time since admixture under the assumption of random mating. During gametogenesis in admixed individuals, recombination breaks down continuous ancestry tracts inherited from each of the source populations of the admixture event into smaller alternate fragments at each generation. Consequently, the length of the continuous ancestry tracts reflects how many generations ago the source populations migrated across the geographical barriers that prevented them to mate (Gravel, 2012; Hellenthal et al., 2014; Chintalapati et al., 2022). Herein, we argue that the tract length information can measure the non-randomness of mating associated with genetic ancestry and, therefore, it can also monitor the permeability of socioeconomic and cultural barriers between subpopulations with different genetic ancestries (Zaitlen et al., 2017).
Some of the methods to date admixture can discern multiple pulses of migration. Only a few of them have addressed complex admixture histories such as the fluctuation of unbalanced migrations of males and females from two source populations (Laurent et al., 2022). However, almost all these approaches assume random mating in the admixed population, overlooking the effect of population stratification in the population structure. To our knowledge, few studies have modeled ancestry-related assortative mating during admixture, although limited to two source populations (Goldberg et al., 2020; Kim et al., 2021).
Beyond analytical modeling, population genetics studies have also measured ancestry-related assortative mating through the correlation of genetic ancestry proportions between mates (Bryc et al., 2010; Korunes et al., 2022; Arauna et al., 2022). Non-random mating can also be monitored through deviations of the observed heterozygosity from Hardy-Weinberg equilibrium expected values (Crow and Felsenstein, 1968). Thus, when information on genetic ancestry of mating couples is not available, it is still possible to infer ancestry-related assortative mating through the comparison of the genetic ancestry of the two homologous chromosomes of the individuals (Norris et al., 2019). However, these approaches can only infer the mating patterns from the last generations.
Despite these efforts, we still lack a rigorous and robust method to shed light onto the patterns of ancestry-related non-random mating across all generations in which the admixture process extends. More specifically, we are in need of a comprehensive model of ancestry-related assortative mating and sex bias, two parameters which have been rarely jointly modeled in population genetics.
Social stratification and population structure in the Americas
Among human populations, the admixing populations from the Americas are of special interest in admixture studies. We consider them as admixing populations, because their genetics is shaped by an ongoing admixture process of three differentiated continental ancestries that started five centuries ago, constrained by a strong social structure.
At the end of the 15th century, European powers initiated the colonization process in the lands inhabited by Native Americans. In this frame, European colonizers enslaved more than 10 million people brought from sub-Saharan Africa (Eltis, 2018). As a result of this historical event, the populations of the Americas are the outcome of the admixture of Native American, European, and sub-Saharan genetic ancestries (Salzano and Bortolini, 2005; Bedoya et al., 2006; Wang et al., 2008; Moreno-Estrada et al., 2013; Gravel et al., 2013; Ruiz-Linares et al., 2014; Adhikari et al., 2017; Adhikari et al., 2016; Ongaro et al., 2019).
After the abolition of slavery, most of these populations remained stratified based on the socioeconomic status and according to hierarchical notions of racial difference. Some of them have even experienced explicit segregation policies long after the abolition that aimed to prevent mating between subpopulations from different origins and maintain socioeconomic stratification (Douglass, 1882; Du Bois, 1935; Davis, 1981).
In Latin America, in addition to segregation, European colonial powers and creole elites implemented eugenicist policies under the frame of mestizaje/mestiçagem. Mestizaje/mestiçagem refers to the process of admixture of Native American, European, and sub-Saharan ancestries in the context of the European colonization. It is, therefore, associated to the mixture across hierarchical differences understood as ‘racial,’ differences of class, and differences of gender. Since mid-nineteenth century, Latin American nation-building elites have aimed to associate mestizaje/mestiçagem to an equalizing process, by claiming that it overcomes and blurs the socioeconomic differences related to ‘race.’ However, critics have argued that the mestizaje/mestiçagem’s notion of hybridity inherently entails the idea of its constitutive origins and the hierarchies that order those origins. In this sense, mestizaje/mestiçagem attaches greater value to the interactions that move towards whiteness and masculinity and lower value to those that move towards blackness or indigeneity, and femininity (Wade, 2017; Wade et al., 2020; Abel, 2022).
By analyzing the impact of the European colonization in the population structure through mating, we aim to evaluate the stratification related to the genetic ancestry not only quantifying the population subdivision but also measuring the genetic ancestry asymmetry between males and females in mating. Following this approach, we conceptualize a novel mechanistic mating model that explicitly integrates ancestry-related assortative mating and sex bias jointly, through an intersectional approach derived from the interrelated hierarchies observed in the admixture process (Crenshaw, 1989; Crenshaw, 1991; Wade, 2017). We consider a three-way admixture scenario mirroring the demography of the admixing American populations. We build and train a deep neural network to infer non-random mating parameters using extensive synthetic data. We deploy this network to genomic data from admixing American populations sequenced as part of the 1000 Genomes Project 1000 (Auton et al., 2015) and quantify the extent of ancestry-related assortative mating and sex bias. Finally, we discuss racial and gender hierarchies as inferred from their footprint on genetic structure.
Results and discussion
We report our results in three sections: (i) the novel mating model and framework for simulations, (ii) the performance of the neural network, and (iii) the inference of the ancestry-related mating probabilities for admixing American populations.
An ancestry-related mating model
We present an admixture model defined by the mating probabilities of all possible male and female couples, set by their ancestry proportion difference. For each ancestry, the ancestry-related sex bias (SB) and the ancestry-related assortative mating (AM) parameters determine the mating probability of each couple as a function of the difference in the ancestry proportion between male and female. We assume that the differences in the ancestry proportions within the mating couples follow a Normal distribution that translates into the mating probabilities (Figure 1).
The expected value of this Normal distribution defines SB, while AM is modeled as being inversely proportional to the variance (see Methods and materials for mathematical details). For a given ancestry, positive SB values indicate that couples where males have a higher proportion than females of this ancestry have more chances of mating, with the opposite pattern for negative SB values. Conversely, AM modulates the decay of mating probability when the difference of ancestry proportion within the couple moves away from the expected value set by the SB parameter. Therefore, when SB is close to zero and AM is high, a couple with similar ancestry proportions have much higher probability of mating than a couple with substantial differences in the ancestry proportions. If AM is low, a couple with similar ancestry proportions only have a slightly higher probability to mate than a couple with substantial differences in the ancestry proportions. The AM parameter measures the non-randomness of mating associated to a genetic ancestry. This includes both positive assortative mating -genetic similarity between mates- (when SB is zero) and negative assortative mating -genetic dissimilarity between mates- (when is not zero). This approach allows accounting for the male-female way of negative assortative mating through SB parameter. Figure 1B shows how a sample of individuals in generations 4 and 7 mate based on the mating probabilities set by the example values of AM and SB defined in Figure 1A and Figure 1—figure supplement 1.
We focus on the case of three-way ancestry, a model that describes the admixture of the populations of the Americas and their triple genetic ancestry: Native American, European, and sub-Saharan African. We define two alternative models, referred to as the One Pulse model and the Two Pulses model. The simpler One Pulse model assumes one migration event occurring 19 generations ago and includes five independent parameters: AM1, AM2, AM3, SB1, and SB2 for sub-Saharan African (1), Native American (2), and European (3) ancestries (Figure 1A). In the Two Pulses model, an additional parameter (the Gene Flow Rate 9 generations ago -GFR-) determines the fraction of the gene pool arriving in a second migration from each source population 9 generations ago (at generation 10) (Figure 1—figure supplement 1). In both models we assume a continuous admixture process that starts 19 generations ago, knowing that the populations analysed trace the first contact of Native American and European populations in the first half of 16th century (Sánchez-Albornoz, 1977; Thornton, 1987) and assuming a generation time of 26.9 years (Wang et al., 2023). In contrast with the approaches that aim to find an admixture date assuming random mating, we assume that the admixture process starts with the contact, and it is continuous and modulated with the mating parameters that we aim to infer.
Our goal is to predict AM and SB (and GFR for the Two Pulses model) for the admixing American populations sampled in the 1000 genomes project (African Caribbeans in Barbados, ACB; African Ancestry in South-West USA, ASW; Colombians in Medellín, CLM; Mexicans in Los Angeles, MXL; Peruvians in Lima, PEL; and Puerto Ricans in Puerto Rico, PUR). To do so, we aim to compare the continuous ancestry tract lengths profile obtained from a local ancestry analysis performed on this data to the tract lengths profile issued from simulated data for each population with known combinations of the mating parameters.
For 10,000 random combinations of AM and SB parameters (and GFR for the Two Pulses model) for each population we simulate, forward-in-time, a range of admixture scenarios. The contribution of each genetic ancestry to the gene pool of the simulated admixed population is equivalent to the observed ancestry proportions after the local ancestry analysis on the real data. We simulate 22 autosomal chromosomes and the X chromosome for each individual at each generation, keeping track of the local genetic ancestry at each chromosomal region (Figure 1, Figure 2A). This approach serves a dual purpose: (i) to simulate the mating as a function of the genome-wide ancestry proportions of all males and females, based on the mating probabilities set by the AM and SB parameters of the mating model (Figure 1); (ii) to generate the continuous ancestry tract lengths profile as an output after the last simulated generation, which counts the number of fragments within each of the 22 windows defined by a length interval in cM, in a logarithmic scale (Figure 2B; see Materials and methods).
A deep neural network to estimate mating parameters efficiently
To infer all parameters in our model, we train a deep neural network for each population. By exploring the entire parameter space of AM and SB parameters (and for the Two Pulses model), we feed simulated continuous ancestry tract length profiles to a deep neural network consisting of fully-connected layers (Figure 2C).
The network sufficiently learn the weights for all parameters without overfitting over 40 epochs, as shown by the decay of the loss function (mean squared error) (Figure 3—figure supplement 1). We observe low mean squared error on the testing set for all parameters (Figure 3). Similarly, we appreciate a high correlation between true and predicted values, as shown by R2 values and the confusion matrices, at testing (Figure 3—figure supplement 2).
The trained network exhibits better predictions for AM parameters than for an parameters across all ancestries, populations, and migration models. Interestingly, the higher complexity of the Two Pulses migration model does not produce a higher mean squared error or lower R2 values for any of the tested parameters. In fact, the mean of the mean squared error for the Two Pulses model is only marginally higher than the mean of the mean squared error for the simpler One Pulse model (Figure 3, Figure 3—figure supplement 1, Figure 3—figure supplement 2).
The performance of the Local Ancestry analysis might constrain the shape of the continuous ancestry tract length profile and the accuracy of the predictions of the mating parameters. RFMix has been the state-of-the-art method for Local Ancestry method during the last decade. Recently, Gnomix gained popularity after claiming much lower computational requirements and similar or even better accuracy (Hilmarsson et al., 2021). We performed a Local Ancestry analysis with Gnomix and we generated the continuous ancestry tract lengths profile after it. The Gnomix tract length profile showed higher values in the shortest tracts window than the RFMix one (Figure 2—figure supplements 1–6). Gravel, 2012 alerted that short ancestry tracts are likely to have reduced accuracy rates in Local Ancestry inference, whereas longer tracts can be detected with increased confidence.
To evaluate possible biases introduced by the wrong assignation of isolated windows in the Local Ancestry inference, we tested the performance and the estimations of a Neural Network using a modified tract lengths profile for both RFMix and Gnomix. We modified the tract lengths profile by removing, or not, the window corresponding to the shortest tracts, or by dividing, or not, each value of the histogram by the total number of tracts in the Autosomes and in the X Chromosome. The tract lengths profile with All windows and Divided by total sum of tracts presents a low value of generalized variance (GV) between RFMix and Gnomix estimations and a low MSE value in the testing of the Neural Network (Tables 7 and 8, Figure 4—figure supplements 3–7). This shows that a matrix normalization of the tract length profile is enough to reduce the effects of the eventual biases introduced by the Local Ancestry method and points that the Neural Network is able to capture relevant information from the profile shape.
The Native American and sub-Saharan genetic ancestries respectively shape the mating probabilities in Latin American and African American populations
We sought to test the occurrence and extent of assortative mating and sex bias in the admixing American populations from 1000 genomes. To predict AM and SB parameters (and GFR for the Two Pulses model) we deployed the trained neural network on the continuous ancestry tract length profiles of these populations were obtained after a local ancestry analysis.
In the One Pulse model, the Latin American populations (CLM, MXL, PEL, PUR) present a consistent pattern where the Native American ancestry shapes the mating probabilities, as the AM parameter associated to this ancestry is the highest in all populations. Thus, the differences in the Native American ancestry between males and females modulate the mating in Latin American populations, although the Native American ancestry is not the one observed in highest proportion in all of them. PEL, CLM, and MXL populations show stronger AM values than PUR. The high AM values are coupled with negative SB for CLM and PUR populations, indicating that females of high Native American ancestry are more likely to mate with males of lower Native American ancestry. CLM and PUR exhibit significant negative sex-biased admixture while MXL and PEL do not. Conversely, ASW population presents the highest AM in the sub-Saharan African ancestry. Paired with a positive SB value, these estimates indicate that males of high sub-Saharan African ancestry are more likely to mate with females of lower sub-Saharan African ancestry. Finally, ACB populations show similar AM values for the three ancestries, with no specific ancestry modulating the mating probability (Table 3, Figure 4A).
In the Two Pulses model, we allow for an additional migration pulse 9 generations ago (at generation 10) and we let the neural network predict the gene flow rate through the GFR parameter. Under this new scenario, as expected, AM values are much lower than their corresponding values under a One Pulse model. In fact, part of the population structure that is modeled as social stratification in the One Pulse model is now modeled by gene flow from an additional migration event. Both models reflect similar admixture dynamics, where Native American and sub-Saharan African genetic components take longer to homogenize across the individuals of Latin American and African American populations, respectively (Figure 4B).
Under both models of migration, the effect that SB has on the mating probabilities depends on the AM values, as lower AM values imply a lower effect of SB on the mating. In case of low AM, individuals are less constrained in their mating by their ancestry and, therefore, the effect of SB is less prominent.
To evaluate the similarity of the footprints left by either assortative mating or gene flow due to migration, we tested how a neural network trained to predict GFR could predict GFR from data with no gene-flow due to migration but only assortative mating. And, in parallel, we tested how a neural network trained to predict AM could predict AM from data with no assortative mating but only gene flow due to migration. We obtain strong correlations between simulated AM and predicted GFR, and between simulated GFR and predicted AM, which points that both demographic events could lead to similar admixture dynamics and would leave resembling genomic footprints in the population (Figure 4—figure supplements 1 and 2).
We next sought to test whether observed genomic data is more compatible with a One Pulse or Two Pulses migration model. To this aim, we calculated the composite likelihood ratio to compare the fit of simulated continuous ancestry tract length profile under the predicted values of AM and SB (and GFR for the Two Pulses model) to the empirical data. In ASW, CLM, MXL, and PEL populations the Two Pulses model has greater support than the One Pulse model. Results also show that, in all these populations, the Two Pulses model is more supported by the continuous ancestry tract length profile from the autosomal chromosomes than the profile from the X chromosome. (Figure 4C).
Discussion
We tackle the analysis of the ancestry-related non-random mating driven by social structure through a mating model, which allows us to globally evaluate the forces that modulate population structure but also study the effect of these population dynamics at the individual level. Our results show evidence of ancestry-related sex bias and assortative mating in American admixed populations. In Latin Americans, the proportion of Native American ancestry of men and women shape the mating probabilities and, therefore, the genetic structure of the population. By contrast, in African Americans, the sub-Saharan African ancestry modulates mating. Below, we evaluate the performance of our pipeline in discerning between migration and assortative mating and we explore how the next steps could incorporate more complex admixture scenarios. Finally, we discuss the significance of these results and the importance of our approach in studying social stratification.
Discerning geographical barriers and social barriers
The disentanglement of social barriers from geographical barriers is a major challenge to face when addressing the complexity of admixture. Non-random mating patterns associated with social structure and gene-flow after multiple migration pulses might leave similar footprints on the genome of a population. In our analyses, the admixture dynamics are modeled either by ancestry-related assortative mating (both in the One Pulse model and the Two Pulses model) or migration pulses (in the Two Pulses model). In both cases, the ancestry modulating the admixture dynamics takes longer to homogenize across the individuals of the admixing population. We have shown that an earlier admixture process under higher assortative mating might be interpreted as a later migration pulse under random mating or lower assortative mating. This has important implications for admixture dating methods that assume random mating, which can underestimate the generations from the admixture event if the admixture process started earlier and took place under assortative mating.
Despite the similarity of the genomic footprint left by both demographic events the distinction between both scenarios has important implications for understanding social processes. The continuous ancestry tract length profile issued from our mating model is different from that issued from an admixture model where the admixed population only receives constant gene-flow from migration of non-admixed source populations to model the population structure. In the latter approach, as defined in Goldberg and Rosenberg, 2015; Laurent et al., 2022, after migration the non-admixed source populations introduce full-length non-recombined chromosomes with a single ancestry in the admixed populations and shape an identifiable pattern. In our approach, the continuous ancestry tract length profile is capturing this pattern to discern between ancestry-related assortative mating and gene-flow due to migration. Although this pattern characterized by low or non-recombined continuous ancestry tracts in a few individuals could remain hidden in a population mean continuous ancestry tract length profile, it should be detectable with an individual-based one.
Alternatively, a population structure correlated with genetic ancestry could also be modeled with an island model starting from non-admixed panmictic subpopulations with migration rates between them split by sex. Our approach resembles this model but considers a more realistic population continuously structured by genetic ancestry, rather than discrete (although permeable) subpopulation panmictic demes. In this sense, we address population structure through a mating model, which helps to focus the discussion on the effect at individual level of the dynamics that stratify the population.
A versatile mating model to accommodate a wider range of admixture scenarios
The different patterns inferred in autosomal and X chromosomes suggest that some of the complexity of admixture is still not explained by our models. Inferences based on autosomal chromosomes have a greater support for a demographic scenario involving multiple migration events than inferences on the X chromosome. Taken together, these trends suggest a need to expand the model by adding either the possibility of sex-biased migrations or changes in SB and AM through time. In addition, African-American populations might have a complex genetic history involving on one hand male-biased sub-Saharan migration and on the other hand an admixture female-biased in the sub-Saharan ancestry. However, our current model can only accommodate this demographic scenario with a single SB parameter, and the results regarding these populations should be interpreted with caution.
The flexibility of our methodology provides the basis to disentangle more complex scenarios of admixture in a further approach. The plasticity of forward-in-time simulators allows for additional complexity in the admixture model, by modifying the AM and SB parameters through time and modeling multiple migration pulses and population growth.
Machine learning has the potential to infer a large number of parameters issued from complex models of admixture. However, this step likely implies a shift from a population-based continuous ancestry tract length profile to an individual-based one, which needs to be linked to a redesign of the architecture of the neural network. Other architectures such as convolutional neural networks and generative models have been recently deployed to infer introgression, structure, admixture proportions, and post-admixture selection from population genomic data (Gower et al., 2021; Wang et al., 2021; Meisner and Albrechtsen, 2022; Hamid et al., 2022). We caution that local ancestry analysis is a sensitive up-stream step to our method. Particularly, a wrong ancestry prediction in isolated windows that breaks up longer continuous ancestry tracts can substantially affect the tract lengths profile. This might happen despite a high proportion of windows with the genetic ancestry accurately predicted that translates to an overall good performance of the local ancestry method.
Social stratification by racial and gender hierarchies
The ultimate aim of our approach is to infer social stratification in the Americas from the analysis of the population genetic structure. To this end, we model ancestry-related assortative mating mediating the extent of the effect of ancestry-related sex bias. We infer how assortative mating and sex bias shape the mating dynamics of the population and we evaluate how the dimensions of inequality associated to these parameters impact at individual level, thanks to the mating model framework.
We have defined this model from an intersectional perspective, which understands that racial, gender and class hierarchies are mutually constituted (Davis, 1981; Gonzalez, 1984; Hooks, 1984; Crenshaw, 1989; Crenshaw, 1991; Collins, 1990; Carneiro, 1995; McCall, 2005; Hancock, 2007; Viveros Vigoya, 2016). Besides, from Decolonial Feminism, authors have stressed the significance of the frame of the European Colonization in redefining the concept of gender in America and have pointed to sexuality and mating as one of the contexts where racial, gender, and class inequalities manifest with more visibility (Stolcke, 1992; Lugones, 2007; Lugones, 2008; Viveros Vigoya, 2009; Viveros Vigoya, 2016). Specifically, in our approach, we consider that racial stratification intensifies gender inequalities during mating. At the individual level, we conceive that the effects of these social hierarchies on mating depend on the relative position of the subject in racial and gender axes of inequality respect to the position of the other individuals of the population, and therefore on the context (Yuval-Davis, 2006; Anthias, 2013; Jorba and Rodó-Zárate, 2019).
In this framework, we tackle the hypotheses about the role of mestizaje/mestiçagem in the inequality of a society developed by Wade, 2017; Wade et al., 2020 through a genetic study. Wade states that mestizaje/mestiçagem is a highly ambivalent discourse and set of practices, as it promotes and facilitates interactions across hierarchical differences of ‘race,’ class, and gender, but simultaneously reinforces those hierarchies. Studying the interaction of racial and gender hierarchies together, our approach allows us to investigate the dynamics of mestizaje/mestiçagem and challenge the oversimplified understanding that greater mixture translates into a decrease of discrimination and inequality. As such, high levels of mixture cannot be understood as a sign of equality or low discrimination if it takes place with strong gender-biased patterns. Instead, this scenario provides evidence for a deep interaction between racial and gender hierarchies that shapes the social structure of the population.
In the specific case of the Latin American populations considered here, the prejudices and biases related to gender associated with Native American ancestry constrain the dynamics within the society and determine the mating, shaping the population genetic structure. In CLM and PUR populations, females with a higher proportion of Native American ancestry tend to mate with males of lower Native American Ancestry. Interestingly, the populations that in the Two Pulses model present a higher migration pulse from Native American populations 9 generations ago (MXL and PEL) do not present a significant negative sex bias. This additional gene-flow probably took place under weaker sex-biased patterns and shifted the observed parameter towards non-negative values. This is consistent with historical records reporting an increased pressure on Native populations to culturally assimilate towards a white/mestizo norm starting in the 19th century, which spurred internal migrations to urban areas and the loss of indigenous languages in places like Mexico and Peru (Viqueira, 2010; Telles, 2014).
Our results support the idea that Mestizaje/mestiçagem operates through racial and gender hierarchies and it is accompanied by a gradual dilution of the sociocultural elements associated with non-European genetic ancestries into the admixing population.
Conclusion
To have a broader perspective of how racial and gender hierarchies operate across American societies, further approaches should analyze a wider dataset with more diverse, representative and carefully sampled populations. Specifically, the inclusion of socioeconomic variables in the sampling would allow us to evaluate how class hierarchies interrelate with racial and gender hierarchies. In addition, the transition to a more complex admixture model that monitors changes in the AM and parameters and includes sex unbalanced migrations should provide the possibility to evaluate how racial and gender hierarchies have changed through history in different regions of the Americas.
In conclusion, an interdisciplinary approach that incorporates up-to-date insights from social sciences is essential to conceptualize population genetic models that aim to evaluate genetic structure driven by social stratification. Further studies might expand the analysis of population stratification by exploiting the full potential of machine learning in population genetics. An intersectional perspective that jointly addresses the effects of racial, gender, and class hierarchies on population structure will be key to understanding the genetics of the admixing populations of the Americas.
Materials and methods
Use of genetic ancestry categories
Request a detailed protocolHuman genetic diversity is a continuum and does not show discrete groups of individuals. Therefore, race is not a proxy for human genetic variation. Similarly, genetic ancestry labels are not distinct categories in which the genetic diversity of a population can be intrinsically grouped. Instead, the inference of genetic ancestry should be used in a hypotheses-driven approach framed in a specific time horizon (Mathieson and Scally, 2020; Coop, 2022; Lewis et al., 2022). In any case, researchers should disclose the process by which they selected and assigned group labels and the rationale for any grouping of samples (National Academies of Sciences, Engineering, and Medicine, 2023).
Here, we work with genetic ancestry categories that refer to the part of the genome inherited from each of the three populations involved in the admixture process in the Americas in the particular framework of the European colonization (Native American, European, and sub-Saharan African). We use these genetic ancestry categories because we hypothesize that they are correlated with the socially constructed racial and gender hierarchies that shape the population structure in the Americas.
A mating model defined by ancestry-related assortative mating and sex bias
Request a detailed protocolWe derived a mechanistic model where AM and SB parameters constrain the mating probabilities as a function of the difference in the ancestry proportion between male and female, for each ancestry. We assume that the probability of mating follows a normal distribution where SB for each ancestry is defined as the expected value of the difference in the ancestry proportion within mating couples, while is modeled as being inversely proportional to its variance (Figure 1).
Model definition
Request a detailed protocolConsider an admixed finite population from isolated source populations comprised of females and males. Assume that for each individual we have a vector of inferred ancestry proportions for each source population , so that . We consider a random variable for mating as a realization of the event between a female and male .
We calculate the probability of mating between a female and male as
where and are the probabilities of either female or a male to start a mating event that will have one child as the outcome.
is the probability of a female to mate with a male given the ancestry proportions of both female and male, once the female has already been chosen to initiate the mating.
is, therefore, the probability of a male to mate with a female given the ancestry proportions of both female and male, once the male has already been chosen to initiate the mating.
Mating probability of a couple
Request a detailed protocolIn the most basic model, all individuals can be assumed to have the same probability to start the mating event independent of their ancestry:
Once either a female or a male initiates the mating, the mating probability of each possible couple is defined as a function of the ancestry proportions of this individual and the ancestry proportions of each individual of the other sex ( or for respectively a male or a female that initiates the mating).
This probability is described by a multivariate normal distribution defined by a mean vector μ, related to SB, and covariance matrix , related to AM. This multivariate normal distribution consists of S variables defined as , related to each ancestry . They account for the difference of the ancestry proportion in each mating couple. The final mating probability for an individual and a possible mate is relative to the sum of all the probabilities for all the possible mates for this individual:
and:
where μ is the vector of the expected means of the ancestry proportion differences (i.e. for ancestry ) which defines SB for each ancestry. The diagonal of is the vector , where the variance of is inversely proportional to AM for each ancestry (Figure 1).
The sum of the mean vector (i.e. the sum SB parameters for all the ancestries) is zero (). In addition, is not full rank (). Consequently, the multivariate density function can be represented with dimensions, which has independent parameters.
The case of three ancestries, S=3
Request a detailed protocolWhen S=3, the multivariate normal distribution is equivalent to a two-dimensional multivariate normal distribution, which has five independent parameters (2 in μ and 3 in ):
where can be defined by the variances of the three-dimensional multivariate normal distribution, including :
The mating model for three ancestries is set by sex bias for ancestries 1 and 2 (SB1, SB2) and the assortative mating for ancestries 1, 2, and 3 (AM1, AM2, AM3). Therefore, SB for ancestries is defined as follows:
and AM for ancestries :
This arbitrary parameterisation has been chosen in order for the the assortative mating parameter AMs to cover the full spectrum of meaningful values taking values from 0 to 1 on a logarithmic scale, where 0 is random mating and 1 is very strong assortative mating.
Simulations
Request a detailed protocolWe performed 10,000 simulations per population (CLM, MXL, PEL, PUR, ACB, ASW from 1000 genomes) following the mating model described above using SLiM (Haller et al., 2019). In each simulation, AM and SB parameters (and GFR for the Two Pulses model) were independently sampled from a uniform distribution. We simulated the 22 autosomal chromosomes and the X chromosome for each individual. We tracked their real local ancestry by recording the source population from which each genomic fragment is inherited. At each generation, we simulated mating based on the mating model as a function of the genetic ancestry proportions of the individuals. After the mating of two individuals, we simulated recombination in the gametogenesis of the offspring and the progressive break down of the continuous ancestry tracts, using the local recombination probabilities from the genetic map provided in Delaneau et al., 2019. We ran a total of 19 generations, mirroring the time range from the beginning of the colonization to present day. For computational purposes, we scaled down by a factor of 1000 the lengths and recombination rates of the genome.
For the One Pulse demographic model, we simulated a constant population size of 1000 and we set initial gene-flow proportions equivalent to the observed genetic ancestry proportions for each population Table 1. For the Two Pulses model, we split the same size of migrant population from each source population in two migration waves at generation 0 and 10 as a function of the GFR parameter.
The continuous ancestry tract lengths profile
Request a detailed protocolThe continuous ancestry tract length profile is a statistic that is commonly used to date admixture events, assuming random mating. However, here we exploited the information summarised with this statistic to assess the gene-flow related to both migration and assortative mating. In addition, we included the population continuous ancestry tract length profiles of both autosomes and X chromosome to provide to the neural network information that can be used to predict sex bias. While both sexes contribute equally to the autosomal genepool, females and males contribute 2/3 and 1/3, respectively, to the X chromosome genepool. This asymmetric inheritance between autosomes and X chromosome combined with local ancestry information is highly informative of the complexity of sex-biased admixture histories (Goldberg and Rosenberg, 2015).
We calculated the continuous ancestry tract length profile on simulated data, for each individual, by counting the number of tracts for each length bin, greater than or equal to the lower threshold and lower than the upper threshold, defined by a vector of break points in a logarithmic scale, in centiMorgan (cM): . These breakpoints define a total of 22 length windows, which is a compromise of the RFMix resolution in the local ancestry analysis (0.1 cM) and a limited number of windows. Then, we obtain the continuous ancestry tract length profile for each individual. Finally, we perform the mean across the individuals of the same population as the permutation-invariant function to use it as input of he neural network. For real empirical data, we run a local ancestry analysis to split the genome of each individual into the fragments inherited from Native American, European, and sub-Saharan African ancestries to obtain the length of the continuous ancestry tracts (Figure 2A). To do this, we performed an RFMix analysis with RFMix v1.5.4 (Maples et al., 2013) with the following options: -w 0.1 G 19 -e 3. We used as target populations the six admixed populations of the Americas present in the 1000 genomes data (African Caribbeans in Barbados, ACB; African Ancestry in SW USA, ASW; Colombians in Medellín, CLM; Mexicans in Los Angeles, MXL; Peruvians in Lima, PEL; and Puerto Ricans in Puerto Rico, PUR) using the 30 x coverage data 1000 (Auton et al., 2015; Byrska-Bishop et al., 2022).
To create three reference populations we first combined 1000 genomes with HGDP genomes (Bergström et al., 2020). We ran an unsupervised ADMIXTURE analysis (Alexander et al., 2009), from which we used the individuals with a proportion higher than 0.99 of one of the ancestries as reference populations for the RFMix analyses. For Native American ancestry (NAT): 6 Colombian, 12 Karitiana, 13 Maya, 13 Pima, 8 Surui, 2 MXL, and 19 PEL (these PEL and MXL individuals are also included in the target population). For European ancestry (EUR): 23 Basque, 12 BergamoItalian, 28 French, 15 Orcadian, 28 Sardinian, 8 Tuscan, 98 CEU, 91 GBR, 98 IBS. For Sub-Saharan African ancestry (AFR): 8 BantuKenya, 8 BantuSouthAfrica, 22 Biaka, 21 Mandenka, 13 Mbuti, 6 San, 22 Yoruba, 3 ACB, 1 ASW, 99 ESN, 102 GWD, 45 LWK, 85 MSL, 107 YRI. To avoid biases introduced by haploid X chromosomes of males, we only used the X chromosome RFMix output from females to generate the continuous ancestry tract length profile. We considered a tract the concatenation of contiguous 0.1 cM fragments with maximum posterior probability of being inherited from one of the three ancestries. We used the same break points to count the fragments in each length bin used in the simulations: (in cM) to obtain the continuous ancestry tract length profile for each individual (Figure 2B). We bootstrapped the fragment length profile of each individual by resampling 1000 times with replacement each ancestry histogram and assuming its mean. Then, we computed the mean across individuals of the same population of the continuous ancestry tract length profile to have a single matrix for each population equivalent to the output of simulations used to train the neural network. We obtained a three-dimensional matrix (22 length windows, three ancestries, two autosomal/X chromosome) that we use as the input to the trained neural network to predict the AM and SB parameters (and GFR for the Two Pulses model) (Figure 2C).
We also performed a Local Ancestry analysis with Gnomix, training a new model using the same reference populations as in the RFMix analysis (Hilmarsson et al., 2021). Like in the RFMix continuous ancestry tract length profile, we only used the X chromosome gnomix output from females. Once generated the tract lengths profile we estimated the mating parameters (Table 2, Tables 4 and 6, Figure 2—figure supplements 1–6). To evaluate possible biases in the mating paramaters estimation caused by errors in the Local Ancestry inference, we tested the performance of a transformed tract lengths profile for both RFMix and Gnomix. We modified the tract lengths profile by removing or not the window corresponding to the shortest tracts, or by dividing or not each value of the histogram by the total number of tracts in the Autosomes or in the X Chromosome. For each of the four combinations of the modifications of the tract lengths profile, we trained 1000 times the Neural Network to predict the mating parameters from either RFMix or Gnomix tract lengths profile (Tables 3–8, Figure 2, Figure 4—figure supplements 3–7). We then evaluated the performance of both estimations and the correlation between the results obtained from both tract lengths profiles. To do it, we analyzed the Generalized Variance of the bi-variate distribution, where the coordinates of each of the 1000 points are the predictions of a mating parameter (e.g. AM1) for one of the 1000 trained neural networks from both RFMix and Gnomix tract lengths profiles. The Generalized Variance is the determinant of the covariance matrix. Thus, it has lower values when the covariance between RFMix and Gnomix is higher and the variance within each of RFMix and Gnomix-based predictions is lower. Although all the mating parameters where estimated for the four modifications of the tract length profile, the profile with All windows and Divided by total sum of tracts was prioritized for downstream analysis, as it shows both low GV and low MSE values.
Neural network
Neural network architecture
Request a detailed protocolWe built a deep neural network comprised of four common fully-connected layers with 512, 256, 128, and 64 units, respectively, and ReLU activation functions. To avoid overfitting, we included a dropout layer with a rate of 0.2 after the last common layer. The network separates into five branches, each one for an independent parameter. Each branch forms a fully-connected layer with 32 units and ReLU activation functions followed by dropout with a rate of 0.2, and a final fully-connected output layer with a sigmoid activation function. There were five parameter branches for the One Pulse model (AM1, AM2, AM3, SB1, and SB2) and three extra parameter branches for the Two Pulses model (GFR1, GFR2, GFR3) (Figure 2C). In total the One Pulse model has 251,141 trainable weights and the Two Pulses model 263,819.
We used Adam as the optimizer and Mean Squared Error as the loss function. We rescale AM and SB from 0 to 1 to equally weight both parameters during learning. We trained the neural network for 40 epochs, a batch size of 64 with a validation split of 0.2 from the training and validation dataset. The training and validation dataset was a random 0.8 sample of the dataset comprising 10,000 matrices of the continuous ancestry tract lengths profile and we kept the remaining 0.2 for testing. We used Keras in Python to design and train the neural network (Chollet, 2015).
All the code is available at GitHub (copy archived at Mas-Sandoval, 2024).
Data availability
The current manuscript uses already published data, so no data have been generated for this manuscript. The code used for the computational analyses is made available at the address stated in the methods.
References
-
Admixture in Latin AmericaCurrent Opinion in Genetics & Development 41:106–114.https://doi.org/10.1016/j.gde.2016.09.003
-
The genetic diversity of the americasAnnual Review of Genomics and Human Genetics 18:277–296.https://doi.org/10.1146/annurev-genom-083115-022331
-
Fast model-based estimation of ancestry in unrelated individualsGenome Research 19:1655–1664.https://doi.org/10.1101/gr.094052.109
-
The genomic landscape of contemporary western Remote OceaniansCurrent Biology 32:4565–4575.https://doi.org/10.1016/j.cub.2022.08.055
-
Do languages and genes correlate?Language Dynamics and Change 5:202–226.https://doi.org/10.1163/22105832-00502007
-
BookCultural Transmission and Evolution: A Quantitative ApproachPrinceton University Press.https://doi.org/10.1515/9780691209357
-
BookThe History and Geography of Human GenesPrinceton University Press.https://doi.org/10.1515/9780691187266
-
BookBlack Feminist Thought: Knowledge, Consciousness, and the Politics of Empowerment.Boston: Unwin Hyman.
-
BookDemarginalizing the Intersection of Race and Sex: A Black Feminist Critique of Antidiscrimination Doctrine, Feminist Theory, and Antiracist PoliticsUniversity of Chicago Legal Forum.
-
Accurate, scalable and integrative haplotype estimationNature Communications 10:5436.https://doi.org/10.1038/s41467-019-13225-y
-
Assortative mating by population of origin in a mechanistic model of admixtureTheoretical Population Biology 134:129–146.https://doi.org/10.1016/j.tpb.2020.02.004
-
SLiM 3: Forward genetic simulations beyond the wright–fisher modelMolecular Biology and Evolution 36:632–637.https://doi.org/10.1093/molbev/msy228
-
Skin deep: The decoupling of genetic admixture levels from phenotypes that differed between source populationsAmerican Journal of Physical Anthropology 175:406–421.https://doi.org/10.1002/ajpa.24261
-
BookMalecot. — Les Mathématiques de l’HéréditéPublications de la Société Linnéenne de Lyon.
-
RFMix: a discriminative modeling approach for rapid and robust local-ancestry inferenceAmerican Journal of Human Genetics 93:278–288.https://doi.org/10.1016/j.ajhg.2013.06.020
-
SoftwareAssortative-mating, version swh:1:rev:0e2b36b0bb87a3ad4f6127b6e686d4a0ba6e2ac1Software Heritage.
-
Genetic dissection of assortative mating behaviorPLOS Biology 17:e2005902.https://doi.org/10.1371/journal.pbio.2005902
-
BookUsing Population Descriptors in Genetics and Genomics Research: A New Framework for an Evolving FieldWashington, DC: The National Academies Press.https://doi.org/10.17226/26902
-
The genomic impact of european colonization of the americasCurrent Biology 29:3974–3986.https://doi.org/10.1016/j.cub.2019.09.076
-
BookThe evolution and Genetics of Latin American populationsCambridge Studies in Biological and Evolutionary Anthropology.
-
Structured mating: Patterns and implicationsPLOS Genetics 13:e1006655.https://doi.org/10.1371/journal.pgen.1006655
-
BookAmerican Indian Holocaust and Survival: A Population History since 1492University of Oklahoma Press.
-
Why do we pick similar mates, or do we?Biology Letters 17:20210463.https://doi.org/10.1098/rsbl.2021.0463
-
La sexualización de la raza y la racialización de la sexualidad en el contexto latinoamericano actualRevista Latinoamericana de Estudios de Familia 1:63–81.
-
BookConvivial Constellations in Latin AmericaNew York, NY: Routledge.https://doi.org/10.4324/9780367817176
-
Automatic inference of demographic parameters using generative adversarial networksMolecular Ecology Resources 21:2689–2705.https://doi.org/10.1111/1755-0998.13386
-
Human generation times across the past 250,000 yearsScience Advances 9:eabm7047.https://doi.org/10.1126/sciadv.abm7047
-
Intersectionality and Feminist PoliticsEuropean Journal of Women’s Studies 13:193–209.https://doi.org/10.1177/1350506806065752
Article and author information
Author details
Funding
Leverhulme Trust (RPG-2018-208)
- Matteo Fumagalli
- Alex Mas-Sandoval
Horizon 2020 (865356)
- Alex Mas-Sandoval
National Institutes of Health (R15HG011528)
- Sara Mathieson
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
We thank Sarah Abel and Andres Ruiz-Linares for carefully reading the manuscript and for their insightful discussion regarding the implications of the findings. We would also like to thank Flora Jay for her valuable feedback on the methods.
MF and AMS are funded by The Leverhulme Research Project Grant RPG-2018–208. AMS is also funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 865356). SM is funded in part by the NIH grant R15HG011528. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or other funding sources.
Copyright
© 2023, Mas-Sandoval et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 1,425
- views
-
- 208
- downloads
-
- 3
- citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading
-
- Evolutionary Biology
- Genetics and Genomics
It is well established that several Homo sapiens populations experienced admixture with extinct human species during their evolutionary history. Sometimes, such a gene flow could have played a role in modulating their capability to cope with a variety of selective pressures, thus resulting in archaic adaptive introgression events. A paradigmatic example of this evolutionary mechanism is offered by the EPAS1 gene, whose most frequent haplotype in Himalayan highlanders was proved to reduce their susceptibility to chronic mountain sickness and to be introduced in the gene pool of their ancestors by admixture with Denisovans. In this study, we aimed at further expanding the investigation of the impact of archaic introgression on more complex adaptive responses to hypobaric hypoxia evolved by populations of Tibetan/Sherpa ancestry, which have been plausibly mediated by soft selective sweeps and/or polygenic adaptations rather than by hard selective sweeps. For this purpose, we used a combination of composite-likelihood and gene network-based methods to detect adaptive loci in introgressed chromosomal segments from Tibetan WGS data and to shortlist those presenting Denisovan-like derived alleles that participate to the same functional pathways and are absent in populations of African ancestry, which are supposed to do not have experienced Denisovan admixture. According to this approach, we identified multiple genes putatively involved in archaic introgression events and that, especially as regards TBC1D1, RASGRF2, PRKAG2, and KRAS, have plausibly contributed to shape the adaptive modulation of angiogenesis and of certain cardiovascular traits in high-altitude Himalayan peoples. These findings provided unprecedented evidence about the complexity of the adaptive phenotype evolved by these human groups to cope with challenges imposed by hypobaric hypoxia, offering new insights into the tangled interplay of genetic determinants that mediates the physiological adjustments crucial for human adaptation to the high-altitude environment.
-
- Evolutionary Biology
Signs of ageing become apparent only late in life, after organismal development is finalized. Ageing, most notably, decreases an individual’s fitness. As such, it is most commonly perceived as a non-adaptive force of evolution and considered a by-product of natural selection. Building upon the evolutionarily conserved age-related Smurf phenotype, we propose a simple mathematical life-history trait model in which an organism is characterized by two core abilities: reproduction and homeostasis. Through the simulation of this model, we observe (1) the convergence of fertility’s end with the onset of senescence, (2) the relative success of ageing populations, as compared to non-ageing populations, and (3) the enhanced evolvability (i.e. the generation of genetic variability) of ageing populations. In addition, we formally demonstrate the mathematical convergence observed in (1). We thus theorize that mechanisms that link the timing of fertility and ageing have been selected and fixed over evolutionary history, which, in turn, explains why ageing populations are more evolvable and therefore more successful. Broadly speaking, our work suggests that ageing is an adaptive force of evolution.