New Research In
Physical Sciences
Social Sciences
Featured Portals
Articles by Topic
Biological Sciences
Featured Portals
Articles by Topic
- Agricultural Sciences
- Anthropology
- Applied Biological Sciences
- Biochemistry
- Biophysics and Computational Biology
- Cell Biology
- Developmental Biology
- Ecology
- Environmental Sciences
- Evolution
- Genetics
- Immunology and Inflammation
- Medical Sciences
- Microbiology
- Neuroscience
- Pharmacology
- Physiology
- Plant Biology
- Population Biology
- Psychological and Cognitive Sciences
- Sustainability Science
- Systems Biology
Evidence of directional and stabilizing selection in contemporary humans
Edited by Aravinda Chakravarti, Johns Hopkins University School of Medicine, Baltimore, MD, and approved November 3, 2017 (received for review May 1, 2017)
This article has a correction. Please see:

Significance
Combining high-throughput molecular genetic data with extensive phenotyping enables the direct study of natural selection in humans. We see firsthand how and at what rates contemporary human populations are evolving. Here we demonstrate that the genetic variants associated with several traits, including age at first birth in females and body-mass index in males, are also associated with reproductive success. In addition, for several traits, we demonstrate that individuals at either extreme of the phenotypic range have reduced fitness—the hallmark of stabilizing selection. Overall, the data are indicative of a moving optimum model for contemporary evolution of human quantitative traits.
Abstract
Modern molecular genetic datasets, primarily collected to study the biology of human health and disease, can be used to directly measure the action of natural selection and reveal important features of contemporary human evolution. Here we leverage the UK Biobank data to test for the presence of linear and nonlinear natural selection in a contemporary population of the United Kingdom. We obtain phenotypic and genetic evidence consistent with the action of linear/directional selection. Phenotypic evidence suggests that stabilizing selection, which acts to reduce variance in the population without necessarily modifying the population mean, is widespread and relatively weak in comparison with estimates from other species.
Natural selection can strongly affect patterns of phenotypic variation. This fact has led to considerable interest in understanding how natural selection and other evolutionary forces combined to shape the allelic spectrum underlying variation within and between populations. Most of this work has focused on searching the genome for signatures of past selective events (1). Yet selection fundamentally acts on phenotypes, not genotypes. Therefore, the relationships between phenotypes and fitness must be studied in contemporary populations to observe natural selection directly. In doing so, we can gain insights about the direction and magnitude of phenotypic evolution. Theoretically, such observations allow one to predict future evolutionary change, and they can serve as points of comparison with inferences of selection obtained from other sources of data. Here we report observational evidence that is consistent with the action of natural selection in a contemporary human population.
Directional selection results in a covariance between the trait and fitness and can lead to changes in the mean value of a trait in a population (2⇓–4). Further, if phenotypic variation for the trait is caused by genetic factors, then directional selection can result in changes in the genetic composition of a population. Phenotypes may also be subject to stabilizing selection or disruptive selection, which are both nonlinear forms of selection. The key distinction between stabilizing and disruptive selection is whether the relationship between fitness and a phenotype is concave down or up, respectively. Stabilizing selection, which is commonly invoked in theoretical studies of quantitative traits (5⇓⇓⇓–9), will tend to reduce phenotypic variation while disruptive selection will tend to increase it. In a seminal paper on the direct study of natural selection, Lande and Arnold (10) put forth a statistical framework by which the magnitude of both directional and nonlinear selection could be estimated from observational data via regression of fitness onto phenotypes and their squared values.
Application of the Lande and Arnold (10) framework to human populations has yielded evidence consistent with the action of directional selection on physiology, life-history, and body-size traits in both pre- and postindustrial societies (11). While important differences between the studied populations exist (12, 13), a few interesting trends have emerged. Multiple studies have suggested that directional selection has acted to lower the age at first birth in females (14⇓⇓⇓⇓–19), increase the age at menopause (14, 17), increase weight in females (17, 18, 20), and decrease height in females (13, 17, 18, 20, 21) in contemporary postindustrial populations.
Direct evidence for the action of stabilizing selection in humans is scarcer. Birth weight is one reported example of a human trait under stabilizing selection (22), although the intensity of selection has decreased in postindustrial societies (23). A twin study of female reproductive life-history traits showed evidence for a phenotypic optimum for age at menarche (14). Additionally, phenotypic evidence has been presented that is indicative of the simultaneous action of directional and stabilizing selection on height in the Dutch (24). However, a recent study in the contemporary United States found no evidence for any nonlinear selection (25)—although sample size may have limited the power to detect such effects. While selection acts on phenotypes, evolution requires genetic variation. The genetic covariance between a phenotype and fitness determines the expected evolutionary change (2⇓–4) of that phenotype in a population. Genetic covariances between traits can be estimated from pedigree information or directly from molecular genetic data (26).
The use of molecular genetic data has multiple advantages over traditional sources of data for the study of contemporary selection (19, 25, 27). The most obvious advantage is the availability of data; genetic data from large samples of unrelated individuals are increasingly accessible to many researchers. Another advantage comes from the ability to control for possible cultural transmission of traits, which is generally confounded with genetics in observational studies because parents pass both on to their offspring (11). This issue can be partially mitigated by accounting for population structure (28) and geography in samples of unrelated individuals.
In the first attempt to use SNP-array data to study contemporary natural selection on complex traits, Tropf et al. (19) found a negative genetic correlation between relative lifetime reproductive success (rLRS)—the individual lifetime reproductive success divided by the mean—and age at first birth, using a bivariate linear mixed-modeling approach (26, 29). However, Beauchamp (25) noted that the bivariate analyses are underpowered with modest sample sizes and chose to analyze genetic predictors derived from the results of independent large genome-wide association studies (GWAS). Significant negative correlations between polygenic prediction scores for female educational attainment and rLRS have been found in the populations of the contemporary United States (25) and Iceland (27). But reliance on external GWAS summary statistics limits analyses to traits that have already been thoroughly characterized at the genetic level.
Here, we analyze the phenotypic and genetic correlates of rLRS in the UK Biobank (UKB). The UKB is a large population-based prospective study of the genetic and environmental determinants of aging-related disease (30). The dataset consists of over 500,000 individuals from the United Kingdom who have been genotyped at common SNPs and clinically phenotyped for many different traits. These data provide paired genotype and phenotype samples large enough to accurately measure additive genetic correlation between many heritable complex traits (31).
First, we apply the Lande and Arnold (10) framework through regression analyses of the relationship between a suite of phenotypes and a proxy for fitness, rLRS, in 217,728 females and 158,638 males. Then, the genetic data available from 157,807 female and 115,902 male unrelated samples are used to estimate genetic correlations between the phenotypes and rLRS through linkage disequilibrium (LD)-score regression analysis (32, 33). This analysis was supported by the observation that rLRS had a low, but measurable heritability. Our analyses replicate the main results of other recent studies (19, 25, 27) and uncover a host of other significant genetic correlations with rLRS. We also report estimates of quadratic relationships with rLRS, which may be interpreted as evidence consistent with stabilizing or disruptive selection, informing efforts to model the processes that maintain heritable variation in human complex traits (34⇓⇓⇓⇓⇓⇓⇓⇓–43). Our observations are consistent with the action of weak directional and stabilizing selection and limited disruptive selection in the UK Biobank population.
Phenotypic Observations
We estimate linear (
The
(A and B) Scatterplot showing the magnitude of (A) linear selection gradients
Predicted relative fitness as a function of height. Linear and quadratic selection gradients were converted into parameters of a Gaussian fitness function. Using the parameterized Gaussian fitness function, relative fitness values across the observed phenotypic range are predicted and shown by solid red (female) and dashed black (male) lines. The population means are indicated by vertical solid red (female) and dashed black (male) lines. Histograms of female (red) and male (gray) phenotypes are overlaid with an axis on the right-hand side. The horizontal dashed line indicates a relative predicted fitness of 1.
In contrast to a recent study (25), 12 traits in females and 14 traits in males have a significant nonlinear selection gradient estimate (
Fig. 1B shows that, unlike many of the
Fig. 1B shows that age at menopause, fluid intelligence score, and age at first birth (AFB) all have a positive
A multiple-regression analysis provided a more conservative perspective on the phenotypic correlates of rLRS. Due to multicollinearity (SI Appendix, Figs. S4 and S5) and nonoverlapping missing data, we had to choose only a subset of traits for the multiple regression. The full multiple-regression results are included in Dataset S1 and are summarized in SI Appendix, Table S1. In males, the estimates of
To further explore the relationship between AFB, EA, and rLRS we fitted a reduced multiple-regression model with EA, AFB, and their interaction. In the reduced model, all three terms (two linear and one interaction) were highly significant (SI Appendix, Table S1). As in the initial multiple regression, the direction of association for EA is positive in the reduced model. In addition, the interaction term between EA and AFB is strongly positive (
The estimates of
Genetic Correlations with rLRS
The phenotypic results are consistent with the action of natural selection, but for adaptation to occur there must be effects on the genetic level. To this end, we analyzed genetic data from 157,807 female and 115,902 male unrelated samples. Estimates of the genetic correlations between several traits and rLRS,
Theory predicts that traits highly correlated with fitness will have low heritability (47). As expected, rLRS has a low but significant SNP heritability in the UKB dataset, which means that we have power to detect strong genetic correlations. Specifically, the LD-score regression estimates of
Bar plots showing genetic correlations between a selection of traits and rLRS for females (red) and males (blue). Traits were selected on the basis of being marginally significant (P
The estimated genetic correlation with rLRS was significant for several anthropometric traits. For example, the estimates of
BMI provides another important example of evidence for directional selection on an anthropometric trait. We estimate that the
The genetic correlation estimate for AFB in females was the strongest observed in our study. We estimate that the
Another interesting aspect of the observed negative directional selection on AFB is that it would suggest selection for increased female reproductive lifespan. However, the evidence is less clear when we compare the results on AFB to other female reproductive life-history traits such as the age at menarche (AAM) and age at menopause (AMP). In fact, we estimate that the genetic correlation with rLRS is positive for AAM (
Estimation of the genetic evidence of nonlinear selection was not performed because of lack of statistical power. Theory predicts that the additive genetic variance for a squared phenotype is likely to be very small and, when present, is confounded with genetic control of phenotypic variability. In addition, the empirical heritability estimates for squared phenotypes are small (SI Appendix, Fig. S19). Despite the lack of power, a polygenic predictor for height, constructed from a meta-analysis of the Genetic Investigation of Anthropometric Traits (GIANT)–UKB joint dataset, did show a marginally significant negative quadratic regression coefficient in females (see SI Appendix for details).
Discussion
Estimates of linear and quadratic selection gradients were obtained via simple linear regression of a broad set of phenotypes onto a proxy for fitness. The results suggest that many traits measured in the UKB are under the influence of directional and stabilizing selection. However, many of the selection gradient estimates were not significant in a multiple-regression setting, implicating apparent selection (46). However, the population genetic architecture of a trait may still be modified by apparent selection.
For example, the direction of association between female EA and rLRS is positive in the multiple regression, which opposes results from our single-trait regressions, genetic correlation analyses, and multiple other published results (16, 25, 27, 48⇓–50). Our findings lead to the prediction that variants with a positive effect on female EA would decrease in frequency over time even if variance in EA itself does not directly cause variance in reproductive success. Consistent with this prediction, recent work demonstrated that the mean polygenic score for EA has declined over time in the Icelandic population (27), but also suggest that this trend may be explained by factors like female AFB.
Consistent with previous studies, our results support a hypothesis of strong negative selection on female AFB (14⇓⇓⇓⇓–19, 51⇓⇓–54). We also observed a small but positive relationship between AMP in females and rLRS on the phenotypic level, which agrees with previous results (11, 14, 17, 54). However, we find support for a negative genetic relationship between rLRS and AMP. Further, both genetic and phenotypic data suggest a positive correlation between AAM and rLRS. Thus, it is unclear whether the total reproductive lifespan is positively or negatively correlated with rLRS in our data. As larger samples from diverse populations become available, we may gain a more clear view of the selective forces acting on reproductive traits in contemporary humans.
There is clear evidence for correlation between rLRS and several anthropometric traits. Our findings are consistent with previous reports of selection for increased BMI (17, 25). Additionally, the data suggest that the relationship between rLRS and height is more negative in females than in males, which agrees with other results in the literature (18, 24, 25).
Our estimates are conditional on survival to postreproductive ages, so the intensity of selection could be different for traits that strictly influence survival. Birth weight is a classic example of a trait under strong stabilizing selection, where high and low birth weights are correlated with reduced survival in both males and females (22). Yet we find no evidence for stabilizing selection on birth weight in males and only a mrginally significant estimate of
There are a few other important caveats and limitations to our present analyses. All of our results are conditional on the suite of phenotypes that we have measured; there is a real possibility that there are unmeasured phenotypes that drive or confound some of our results. This issue is related to the phenomenon of apparent selection and should always be kept in mind when studying phenotypic selection (46). In addition, the genetic correlations are estimated using common SNP markers (minor allele frequency > 0.01), which may be a source of bias because the genetic variants with deleterious effects on fitness are likely to be rare and thus absent from our analyses. However, this should simply reduce the power of our analyses. Further, there is evidence that the population of the UKB may not be perfectly representative of the whole population of the United Kingdom (55). The potential ascertainment bias (heathy participant bias) in the UKB is important to consider and may have a quantitative effect on our estimates, but the bias is not likely to be large enough (55) to disrupt the conclusions of our work in a qualitative way.
The distributions of
Stabilizing selection appears to be the more common form of nonlinear selection. The most common model of stabilizing selection used in evolutionary quantitative genetics is the Gaussian stabilizing selection model (56). One of the most important parameters of the Gaussian stabilizing selection model is the inverse selection intensity normalized by the phenotypic variance,
We have shown the power of combining high-throughput molecular genetic data with extensive phenotyping to study the ongoing dynamics of human evolution (11). Our work supports further study of a dynamic moving-optimum model for the evolution of complex traits in humans. Presently, we do not know whether the genetic architectures of complex traits are commensurate with equilibrium models parameterized by their contemporary selection gradients. If they are not, further research is needed to better understand how contemporary evolutionary forces differ from the ones that shaped the genetic architecture of the trait.
Materials and Methods
Phenotypic and genetic data were obtained from the UKB and may be accessed by all bona fide researchers from the UKB Access Management System. Only data from samples of self-reported white-British ancestry over the ages of 45 y for females and 50 y for men were used in all analyses, unless otherwise noted. Phenotypic analyses were performed using linear regression in R (57). Genetic correlations were calculated using LD-score regression software according to the protocol developed in ref. 33. Statistical significance was determined using Bonferroni-corrected P values at a family-wise error rate of 0.05. The Northwest Multicentre Research Ethics Committee approved the study and all participants in the UKB study provided written informed consent. For detailed descriptions of the data preparation and analyses, see SI Appendix, Supporting Materials and Methods.
Acknowledgments
This research was conducted using the UKB Resource under Project 12505. J.S., M.R.R., and P.M.V. are supported by the Australian Research Council (Discovery Project 160103860) and the Australian National Health and Medical Research Council (Grants 1078037 and 1113400). This work was supported by NIH Grant R01-GM115564 (to K.R.T.). This material is based upon work supported by the National Science Foundation (NSF) Graduate Research Fellowship Program under Grant DGE-1321846. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Footnotes
- ↵1To whom correspondence should be addressed. Email: peter.visscher{at}uq.edu.au.
Author contributions: J.S.S., M.R.R., K.R.T., and P.M.V. designed research; J.S.S. and J.S. performed research; J.S.S. and J.S. analyzed data; and J.S.S. and P.M.V. wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1707227114/-/DCSupplemental.
Published under the PNAS license.
References
- ↵
- ↵
- ↵
- ↵
- ↵.
- Haldane JBS
- ↵
- ↵.
- Kimura M
- ↵
- ↵
- ↵
- ↵
- ↵
- ↵
- ↵
- ↵
- ↵
- ↵.
- Byars SG,
- Ewbank D,
- Govindaraju DR,
- Stearns SC
- ↵.
- Stearns SC,
- Govindaraju DR,
- Ewbank D,
- Byars SG
- ↵
- ↵.
- Bailey SM,
- Garn SM
- ↵.
- Nettle D
- ↵
- ↵
- ↵.
- Stulp G,
- Barrett L,
- Tropf FC,
- Mills M
- ↵.
- Beauchamp JP
- ↵
- ↵.
- Kong A, et al.
- ↵
- ↵.
- Thompson R
- ↵
- ↵
- ↵
- ↵
- ↵
- ↵.
- Eyre-Walker A
- ↵
- ↵
- ↵
- ↵.
- Zuk O, et al.
- ↵.
- North TL,
- Beaumont MA
- ↵
- ↵.
- Uricchio LH,
- Zaitlen NA,
- Ye CJ,
- Witte JS,
- Hernandez RD
- ↵.
- Sanjak JS,
- Long AD,
- Thornton KR
- ↵
- ↵.
- Mostafavi H, et al.
- ↵.
- Johnson T,
- Barton N
- ↵
- ↵
- ↵
- ↵
- ↵.
- Kaar P,
- Jokela J,
- Helle T,
- Kojola I
- ↵.
- Helle S,
- Lummaa V,
- Jokela J
- ↵.
- Milot E, et al.
- ↵.
- Bolund E,
- Bouwhuis S,
- Pettay JE,
- Lummaa V
- ↵.
- Fry A, et al.
- ↵.
- Burger RR
- ↵.
- R Core Team

















