Evolution and Mechanisms of Plasticity in Wild Baboons (Papio cynocephalus)

by

Amanda Jeanne Lea

University Program in Ecology Duke University

Date:______Approved:

______Jenny Tung, Co-supervisor

______Susan Alberts, Co-supervisor

______Anne Pusey

______Barbara Engelhardt Martin

______Susan Murphy

Dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the University Program in Ecology in the Graduate School of Duke University

2017

ABSTRACT

Evolution and Mechanisms of Plasticity in Wild Baboons (Papio cynocephalus)

by

Amanda Jeanne Lea

University Program in Ecology Duke University

Date:______Approved:

______Jenny Tung, Co-supervisor

______Susan Alberts, Co-supervisor

______Anne Pusey

______Barbara Engelhardt Martin

______Susan Murphy

An abstract of a dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the University Program in Ecology in the Graduate School of Duke University

2017

Copyright by Amanda Jeanne Lea 2017

Abstract

In many species, early life experiences have striking effects on health, reproduction, and survival in adulthood. Thus, early life conditions shape a range of evolutionarily relevant traits, and in doing so alter the genotype-phenotype relationship and the phenotypic distribution on which selection acts. Because of the key role early life effects play in generating variation in fitness-related traits, understanding their evolution and mechanistic basis is crucial. To gain traction on these topics, my dissertation draws on ecological, demographic, and genomic data from a long-term study population of wild baboons in Amboseli, Kenya to address two themes: (i) the adaptive significance of early life effects and (ii) the molecular mechanisms that connect early life experiences with later life traits. In service to the second theme, I also (iii) develop a laboratory method for understanding the role of one particular mechanism—

DNA methylation—in translating environmental inputs into phenotypic variation. In chapter one, I empirically test two competing explanations for how early life effects evolve, providing novel insight into the evolution of developmental plasticity in a long- lived species. In chapter two, I address the degree to which ecological effects on fitness- related traits in a wild baboon population are potentially mediated by changes in DNA methylation. Finally, in chapter three, I develop a high-throughput assay to improve our knowledge of the phenotypic relevance of changes in the epigenome. Together, this

iv

work provides insight into the genes and mechanisms involved in sensing and responding to environmental variation, and more generally addresses several key gaps in our understanding of how environmental inputs are translated into evolutionarily relevant trait variation.

v

Contents

Abstract ...... iv

List of Tables ...... xi

List of Figures ...... xii

Acknowledgements ...... xiii

1. Introduction ...... 1

1.1 Background ...... 1

1.2 Chapter 1: The adaptive significance of developmental plasticity ...... 3

1.2.1 Key questions and current gaps in knowledge ...... 3

1.2.2 Empirical tests of developmental constraints versus predictive response models ...... 6

1.3 Chapter 2: The molecular mechanisms that mediate developmental plasticity ...... 8

1.3.1 Key questions and current gaps in knowledge ...... 8

1.3.2 Environmental effects on the epigenome in natural populations ...... 11

1.4 Functional effects of variation in DNA methylation ...... 13

1.4.1 Key questions and current gaps in knowledge ...... 13

1.4.2 Methods for causally testing the effects of DNA methylation on gene expression ...... 15

2. Developmental constraints in a wild primate ...... 18

2.1 Introduction ...... 18

2.2 Materials and Methods ...... 24

2.2.1 Study subjects, fertility data, and social status data ...... 24

vi

2.2.2 Rainfall data and the 2009 drought ...... 25

2.2.3 Defining low-quality environments in early life and adulthood ...... 25

2.2.4 Defining high-quality environments in early life and adulthood...... 27

2.2.5 Within-female analyses: interaction between early life and adult conditions on female fertility ...... 27

2.2.6 Between-condition analyses: testing for early life effects on adult fertility in a given adult environment ...... 30

2.2.7 Sensitivity to low and high rainfall in early life and adulthood ...... 31

2.2.8 Effects of maternal dominance rank on female fertility ...... 32

2.3 Results ...... 34

2.3.1 Fertility declines during the 2009 drought were greater for females born in low-quality environments than for females born in high-quality environments ...... 34

2.3.2 Sensitivity of female fertility to both extreme low and extreme high rainfall .. 38

2.3.3 Protective effects of high social status in early life ...... 41

2.4 Discussion ...... 42

2.4.1 Support for the developmental constraints model in wild baboons ...... 42

2.4.2 Factors that influence the relationship between early life ecology and adult fertility ...... 45

2.4.3 Conclusions ...... 47

3. Resource base influences genome-wide DNA methylation levels in wild baboons ..... 49

3.1 Introduction ...... 49

3.2 Materials and methods ...... 54

3.2.1 Study subjects and sample collection ...... 54

vii

3.2.2 Generation and processing of genome-wide DNA methylation data ...... 56

3.2.3 Testing for differences in DNA methylation levels at individual CpG sites .... 58

3.2.4 Enrichment of differentially methylated sites by genomic annotation ...... 59

3.2.5 Testing for differences in DNA methylation levels at metabolic pathways ..... 62

3.2.6 Identification of differentially methylated regions (DMRs) ...... 63

3.2.7 Testing the effects of PFKP promoter methylation on gene expression levels . 63

3.2.8 Investigating the stability or plasticity of DNA methylation levels for individuals that switched between resource bases ...... 65

3.3 Results ...... 67

3.3.1 Genome-wide DNA methylation levels contain a signature of resource base . 67

3.3.2 Sites associated with resource base are enriched in functionally important regions of the genome ...... 69

3.3.3 Resource base-associated CpG sites are enriched in specific biological pathways ...... 70

3.3.4 DMRs occur more often than expected by chance, and near a key metabolic gene ...... 71

3.3.5 Individuals that switched between resource bases more closely resembled wild-feeding individuals, regardless of the direction of the switch ...... 73

3.4 Discussion ...... 75

3.4.1 The functional relevance of differential methylation at resource base- associated sites ...... 75

3.4.2 Stability and plasticity in the epigenetic signature of resource base ...... 78

4. Genome-wide quantification and prediction of DNA methylation-dependent regulatory activity ...... 82

4.1 Introduction ...... 82

viii

4.2 Results ...... 85

4.2.1 mSTARR-seq identifies regions with known regulatory activity ...... 85

4.2.2 Methylation dependent activity is predictable based on sequence characteristics alone ...... 89

4.2.3 mSTARR-seq identifies transcription factor families affected by DNA methylation status ...... 92

4.2.4 mSTARR-seq explains heterogeneity in DNA methylation level-gene expression level correlations in vivo ...... 96

4.3 Discussion ...... 98

4.4 Materials and methods ...... 100

4.4.1 pmSTARRseq1 vector design ...... 100

4.4.2 Generation of plasmid libraries for mSTARR-seq ...... 101

4.4.3 Cell culture, plasmid transfection, and cell harvesting ...... 103

4.4.4 Isolation and preparation of mRNA derived from pmSTARRseq1 ...... 104

4.4.5 Preparation of plasmid DNA for DNA-seq ...... 106

4.4.6 Low-level data processing ...... 106

4.4.7 Identification of enhancers and methylation dependent (MD) enhancers ...... 107

4.4.8 Annotation of analyzed mSTARR-seq fragments ...... 109

4.4.9 In silico MspI digest ...... 111

4.4.10 Random forests classification ...... 112

4.3.11 Transcription factor binding motif enrichment analyses ...... 113

4.4.12 Correlations between DNA methylation and gene expression levels in primary cells ...... 114

ix

References ...... 116

Biography ...... 136

x

List of Tables

Table 1. Results from a generalized linear mixed effects model predicting resumption of cycling...... 36

Table 2. Results from a generalized linear mixed effects model predicting conception ... 37

xi

List of Figures

Figure 1: Consecutive low rainfall years magnified the severity of the 2009 drought...... 23

Figure 2: Females born in low-quality environments were less likely to resume cycling and conceive during the 2009 drought than in high-quality years...... 35

Figure 3: Interactions between adult and early life environments under alternative definitions of low-quality environment ...... 39

Figure 4: Fertility during the 2009 drought: females born in low-quality ecological conditions were protected by high-quality social environments...... 42

Figure 5: Study design ...... 55

Figure 6: Resource base influenced genome-wide DNA methylation levels ...... 68

Figure 7: Sites affected by resource base were enriched in functionally important regions of the genome...... 70

Figure 8: Wild-feeding baboons exhibited consistently higher levels of DNA methylation at the phosphofructokinase (PFKP) promoter, where methylation suppresses gene expression in reporter gene assays ...... 72

Figure 9: Individuals that switched resource base more closely resembled lifelong wild- feeding individuals, regardless of the direction of the switch...... 74

Figure 10: mSTARR-seq identifies regions with known regulatory activity...... 86

Figure 11: mSTARR-seq identifies methylation dependent (MD) enhancers, and MD activity is predicted by CpG density ...... 90

Figure 12: Random forests accurately predict MD versus non MD activity based on sequence features ...... 92

Figure 13: HOMER identifies transcription factors enriched in MD enhancers...... 94

Figure 14: mSTARR-seq data explains heterogeneity in previously published in vivo data sets ...... 95

xii

Acknowledgements

Many people provided support and insight while I completed this work, and to them I am very thankful. First, thanks to the staff of the Amboseli Baboon Research

Project (ABRP), including Bernard, Gideon, Longida, Nkii, Moonyoi, Vivian, and Tim. I am especially thankful to the ABRP senior field team – Raphael Mututua, Serah Sayialel, and Kinyua Warutere – who were involved in collecting essentially every biological sample and piece of data I was lucky enough to work with from the Amboseli population. They are skilled observers and wonderful people that I learned so much from. I am also forever grateful to the directors of the ABRP, especially Jeanne Altmann, for what they have created and maintained, and for allowing me to play a small part in their vision. Acknowledgements are also due to the Institute of Primate Research in

Kenya, the Kenya Wildlife Service, the National Museums of Kenya, Kenya’s National

Council for Science and Technology, members of the Amboseli–Longido pastoralist communities, Tortilis Camp, and Ker & Downey Safaris for their assistance in Kenya.

All of the work presented here is the result of close collaborations with so many talented scientists. In particular, I would like to acknowledge the contributions of Jeanne

Altmann, Beth Archie, Chris Vockley, Tina Del Carpio, Luis Barreiro, and Tim Reddy to work presented in this thesis. In addition, Niki Learn, Marcus Theus, Xiang Zhou, Paul

Durst, Tauras Vilgalys, Yingying Zhang, Mercy Akinyi, Ruth Nyakundi, Peter Mareri,

Thomas Kariuki, and Fred Nyundo have been essential collaborators on other work I

xiii

completed as a graduate student. Tawni Voyles, Amanda Shaver, and Mike Yuan also provided key laboratory support, and Sayan Mukerjee deserves special acknowledgement for being such a willing and effective statistics teacher. My colleagues in the Alberts and Tung labs at Duke University, especially Noah Snyder-Mackler,

Courtney Fitzpatrick, and Mercy Akinyi, have been vital in shaping and expanding my views on animal behavior, genomics, the scientific method, and many non-scientific aspects of the world around us. In addition, they have been incredibly generous with their time and friendship.

Outside of the world of science, I owe thanks to my parents, Pat and Dave Lea, and my sister, Michelle Lea, for their support and interest in my work (despite their confusion about what it is I do for a living). I also owe a great deal to Paul Durst, who has kept me sane over the course of graduate school and who always helps me keep some perspective and a smile on my face.

Thanks to my committee members, Susan Alberts, Jenny Tung, Susan Murphy,

Anne Pusey, and Barbara Engelhardt. This group of phenomenal female scientists has been both personally and professionally inspiring, and I am forever grateful for their constant support, encouragement, and wisdom. In particular, thanks to Susan and Jenny for investing so much time and effort in making me the scientist I am today, I cannot begin to describe how much I have learned and grown, and how grateful I am for their mentorship. Susan has had an enormous impact on the way I approach and

xiv

conceptualize scientific questions, and on the way I communicate on the page, in the classroom, and anywhere there is a scientific discussion to be had. Thanks finally to

Jenny for her deeply important role in all the work presented here, for training me to think critically and clearly, and for encouraging my interests in so many diverse areas of science; her mentorship has helped me realize that the new and unknown can be fun, and it has been a joy and an honor to learn, explore, and discover with her guidance.

xv

1. Introduction

1.1 Background

Early life environments can profoundly shape traits related to both human health and Darwinian fitness (Pryce et al. 2005; Monaghan 2008; Gluckman et al.

2008; Pesonen & Räikkönen 2012; McDade 2012; Sheriff & Love 2012; French &

Carp 2016). For example, humans exposed to famine in utero exhibit higher rates of obesity, heart disease, and schizophrenia in adulthood than siblings conceived under better conditions (Jones 1994; Roseboom et al. 2006). In addition, children exposed to multiple sources of early adversity experience more physiological wear-and-tear by midlife, and live shorter lives on average than unexposed individuals (Felitti et al. 1998; Barboza Solís et al. 2015). Similar effects of early life conditions are observed in non-human mammals. In red deer and Asian elephants, individuals born during ecologically challenging periods experience faster reproductive senescence than individuals born during better times

(Nussey et al. 2007; Mumby et al. 2015). Further, early life social adversity predicts increased mortality risk in two natural populations of female baboons

(Silk et al. 2009; Tung et al. 2016). Together, these studies provide just a few striking examples of the impact early life environments can have on later life traits in mammalian species. 1

The capacity of genetically similar individuals to produce variable phenotypes is known as ‘phenotypic plasticity’; phenotypic plasticity that depends on the early life environment is termed ‘developmental plasticity’.

Because the impact of early conditions can be so dramatic, with potent effects on adult health, reproduction, and survival, developmental plasticity is of central interest to multiple disciplines (Gluckman et al. 2005a, 2008; Waterland & Michels

2007; Kuzawa & Quinn 2009; Monaghan et al. 2012; Hostinar & Gunnar 2013).

Researchers in medicine, public health, psychology, and sociology seek to understand the link between early conditions and adult health because of its relevance to the treatment of environmentally-induced disease and the prevention of health disparities. Concurrently, evolutionary biologists aim to decipher the impact of early environments on fitness-related traits, because this knowledge will advance our understanding of the evolution of complex traits and the selection pressures that shape them.

Because of the key role early life effects play in generating variation in health, reproduction, and survival, understanding their evolution and mechanistic basis is crucial. To gain traction on these topics, my dissertation draws on ecological, demographic, and genomic data from a long-term study population of wild baboons in Amboseli, Kenya (Alberts & Altmann 2012) to 2

address two major themes: (i) the adaptive significance of early life effects (Lea et al. 2014, 2015a), and (ii) the molecular mechanisms that connect early life experiences with later life traits (Lea et al. 2016a). In service to these questions, I have also (iii) developed new computational and laboratory tools for understanding the role of one particular molecular mechanism—DNA methylation—in causally translating environmental inputs into changes in gene regulation (Lea et al. 2015b, 2016b). Below, I provide background and context for each of these three research foci and highlight the intellectual or methodological gaps my research has attempted to address.

1.2 Chapter 1: The adaptive significance of developmental plasticity

1.2.1 Key questions and current gaps in knowledge

Dobzhansky’s message—that “nothing in biology makes sense except in the light of evolution”—has been taken to heart by researchers interested in developmental plasticity (Dobzhansky 1973). The result has been a plethora of explanations for how natural selection has produced a trait, namely early environmental sensitivity, that can sometimes lead to detrimental health or fitness outcomes. These explanations fall into two broad categories, each of which assumes that developmental plasticity has been finely tuned through

3

adaptive evolution. However, the two categories of models differ in whether they view plastic responses as anticipatory versus determined largely by immediate constraints.

Developmental constraints models propose that, in resource-limited environments, developing organisms make tradeoffs to protect critical functions

(e.g., investing in brain development at the expense of growth). These tradeoffs may improve the organism’s chance of survival in early life, but will reduce somatic quality in the long-term and compromise adult health (Grafen 1988;

Lindström 1999; Monaghan 2008; Ellison 2014). In other words, natural selection may produce finely-tuned tradeoffs that attempt to optimize overall fitness

(relative to individuals that exhibit no plasticity), but these tradeoffs may have negative long-term consequences for individual health (relative to individuals who did not experience resource limitation). In contrast, predictive models describe a situation in which environmental cues in early life predict the adult environment; hence organisms evolve the ability to adjust their phenotype during development to maximize fitness in the predicted adult environment

(Bateson et al. 2004; Gluckman & Hanson 2004b; Gluckman et al. 2005a; Del

Giudice et al. 2011; Wells 2012). If the early cue induces the predictive adaptive

4

response, but fails to correctly predict the adult environment, the resulting mismatch will result in poor health.

Despite the conceptual impact of these models, relatively few studies have attempted to empirically test the predictions of developmental constraints versus predictive models, especially in humans and other long-lived species (but see

(Hayward & Lummaa 2013; Hayward et al. 2013; Douhard et al. 2014)). From a human health perspective, understanding whether predictive plasticity is expected to be common or rare in long-lived species (such as humans) is essential for knowing what the optimal adult environment ‘looks like’ for each individual, and what mitigation strategies are best employed in the face of early adversity.

Under a predictive model, health is maximized when early and adult environments are concordant, suggesting, for instance, that manipulating adult diet or lifestyle could mitigate the effects of undernourishment in early life. In contrast, if adult phenotypes arise from developmental constraints, matching the adult environment to the early one will be unproductive or even detrimental; hence intervention efforts should focus on improving early conditions. From an evolutionary biology perspective, understanding the distribution of predictive plasticity versus constraint-induced plasticity in nature will help us understand the tradeoffs organisms make under resource-limited conditions, as well as the 5

selective pressures that determine variation in tradeoffs across species.

1.2.2 Empirical tests of developmental constraints versus predictive response models

A critical test of predictive versus constraints models requires comparing the fitness of individuals born in high-quality environments with those born in low-quality environments, when both sets of individuals experience both high- and low- quality conditions as adults (Monaghan 2008; Hayward & Lummaa 2013; Uller et al. 2013; Douhard et al.

2014; Lea et al. 2015a). Under a predictive model, fitness will be maximized when individuals encounter matched early life and adult environments, whereas under a constraints model, individuals born in high-quality environments will consistently outperform individuals born in low-quality conditions. Experiments that satisfy this

‘fully factorial design’ have now been conducted in dozens of short-lived insect, amphibian, reptile, and plant species, and a recent meta-analysis of this literature concluded there was stronger support for constraints than predictive plasticity (Uller et al. 2013). However, in a few striking cases, experimental manipulations of early conditions has produced evidence that is strongly suggestive of predictive plasticity

(e.g., in recent studies of red squirrels (Dantzer et al. 2013) and zebra finches (Mariette &

Buchanan 2016), though these experiments do not fully adhere to the factorial design approach).

In contrast to experimental approaches employed in animals and plants, inferences about the evolution of developmental plasticity in humans have been largely 6

based on cross-population or between-cohort comparisons rather than more powerful within-individual tests (Stanner & Yudkin 2001; Li et al. 2010; Schulz 2010). This limitation exists because identifying human populations that are appropriate for performing within-individual tests is challenging (as is obtaining individual-based, longitudinal data from any long-lived species). Consequently, in spite of enthusiasm for the idea that mismatches between early and later life environments produce pathology in humans, evidence in support of this conclusion is very limited. Indeed, the only critical tests of predictive models in humans that we know of, in pre-industrial Finnish populations, find no support for predictive adaptive responses. Instead, these studies find that pre-industrial Finns born in poor environments exhibit fertility detriments, rather than enhancements, when they re-encounter challenging environmental conditions in adulthood (Hayward & Lummaa 2013; Hayward et al. 2013).

Long-term studies of long-lived mammals offer great potential for performing fully factorial tests, especially in cases where longitudinal data exist or are being collected (Douhard et al. 2014; Lea et al. 2015a). In chapter 1, we perform such a test by leveraging ecological and demographic data from a well-studied population of yellow baboons (Alberts & Altmann 2012). In particular, we examine fertility outcomes

(resumption of cycling and conception) when females naturally encounter environments that both match and mismatch their early life conditions. We find that females born in low-quality ecological conditions suffer long-term fertility costs relative to individuals

7

born in high-quality conditions (Lea et al. 2015a). Our results support developmental constraints models, and agree with the only other ‘fully factorial’ tests that have been conducted in long-lived mammals (Hayward & Lummaa 2013; Douhard et al. 2014).

Further, these results support predictions from a growing body of theoretical work: namely, that when the environment varies stochastically on a timescale shorter than the organism’s generation time, early life conditions will be a poor predictor of the adult environment and predictive plasticity should rarely be favored by natural selection

(Rickard & Lummaa 2007; Wells 2012; Nettle et al. 2013; Botero et al. 2015). This situation is commonly encountered by long-lived animals (Hayward & Lummaa 2013; Hayward et al. 2013; Douhard et al. 2014; Lea et al. 2015a), and we should therefore expect predictive plasticity to be rare in these species. However, more empirical tests from wild systems are desperately needed to interpret the evolutionary history of developmental plasticity in a comparative framework, and to understand the specific selective forces shaping plasticity evolution.

1.3 Chapter 2: The molecular mechanisms that mediate developmental plasticity

1.3.1 Key questions and current gaps in knowledge

Despite a rich history of studies documenting the relationship between early environmental variation and phenotypic variation across species, we know surprisingly little about the molecular mechanisms that mediate these effects.

8

Consequently, it remains unclear how early life experiences are stored into physiological memory to produce phenotypic diversity from otherwise static gene sequences.

Thus far, work on the molecular basis of developmental plasticity has focused heavily on ‘epigenetic’ mechanisms, a class of environmentally or developmentally-sensitive mechanisms that can alter gene expression levels without changing the underlying DNA sequence (e.g., DNA methylation

(Weaver et al. 2004; Murgatroyd et al. 2009; Dominguez-Salas et al. 2014; Tobi et al.

2014), histone modifications (Weaver et al. 2004; Xie et al. 2013; Simola et al. 2016), chromatin accessibility, and non-coding RNAs (Gapp et al. 2014)). Of these mechanisms, DNA methylation has by far received the most attention, for several key reasons. First, genome-wide patterns of DNA methylation are largely laid down in utero and during the first years of life, and establishment of these patterns can be environmentally sensitive (Weaver et al. 2004; Murgatroyd et al.

2009; Elliott et al. 2010; Faulk & Dolinoy 2011). Once established, DNA methylation often remains stable across cell divisions and throughout life

(though later life experiences can also remodel DNA methylation, e.g., (Barrès et al. 2012; Tung et al. 2012; Pacis et al. 2015)). Second, environmentally-induced changes in DNA methylation can alter gene expression levels, providing an 9

avenue through which they could impact organism-level traits (Feil & Fraga

2011; Jones 2012). For example, groundbreaking work in laboratory mice found that maternal diet during pregnancy influences offspring methylation near the agouti gene, which in turn affects agouti gene expression, fur color, body mass, and susceptibility to diabetes (Klebig et al. 1995; Wolff et al. 1998; Waterland &

Jirtle 2003). In a second study, rat pups that received poor maternal care exhibited increased methylation near the hippocampal glucocorticoid receptor gene (NR3C1), which in turn is thought to decrease NR3C1 expression and increase stress reactivity (Weaver et al. 2004).

Studies like these have generated substantial interest in the connection between early life experiences, variation in DNA methylation levels, and fitness or health outcomes (Jirtle & Skinner 2007; Zhang & Meaney 2010; Kappeler &

Meaney 2010; Feil & Fraga 2011; Tung & Gilad 2013; Verhulst et al. 2016).

However, most studies on this topic thus far have focused on either extreme early life challenges (primarily in laboratory rodents, captive rhesus macaques, or humans) or on a handful of candidate genes (Weaver et al. 2004; Heijmans et al. 2008; McGowan et al. 2009; Murgatroyd et al. 2009; Naumova et al. 2012;

Provençal et al. 2012; Tobi et al. 2014). Consequently, our understanding of the evolutionary relevance of environmental epigenetic effects remains limited. In 10

addition, because we do not understand the degree to which responses to stressors in lab animals recapitulate natural stressors in unmanipulated populations, we do not understand whether these systems are appropriate and generalizable models. Finally, genome-wide perspectives are important for providing an unbiased view of the genes involved in sensing and responding to environmental variation. Thus, expanding studies of DNA methylation to include diverse organisms, a range of naturally occurring environmental inputs, and genome-wide data, is key to understanding the role epigenetic mechanisms play in developmental plasticity.

1.3.2 Environmental effects on the epigenome in natural populations

Genomic protocols for measuring genome-wide epigenetic marks are rapidly becoming more cost-efficient and streamlined (Adey & Shendure 2012;

Buenrostro et al. 2013; Picelli et al. 2014), making them increasingly applicable to natural populations. In particular, high-throughput bisulfite sequencing protocols, which rely on the differential sensitivity of methylated versus unmethylated cytosines to the chemical sodium bisulfite, have become a popular approach for measuring genome-wide DNA methylation levels in wild systems

(Lea et al. 2016b; Verhoeven et al. 2016). Current protocols require low amounts of

DNA, avoid the use of species-specific arrays, and can be applied to organisms 11

without a reference genome (Klughammer et al. 2015). In addition, statistical tools for analyzing bisulfite sequencing data now make it possible to analyze datasets that contain complex covariates or genetic structure (Lea et al. 2015b,

2016b). The opening of this new frontier is exciting because work in many well- studied natural systems has already identified components of the early environment that impact fitness-related traits (Altmann 1991; Kruuk et al. 1999;

Onyango et al. 2008; Russell & Lummaa 2009; Hamel et al. 2009; English et al.

2013). Thus, the stage is set for work that probes the mechanistic correlates of these effects.

In chapter 2, we take advantage of one such study system to explore the consequences of a well-documented ecological contrast on DNA methylation levels. In particular, we use reduced representation bisulfite sequencing (RRBS) to profile genome-wide DNA methylation levels in whole blood collected from two sets of wild baboons in the Amboseli ecosystem of Kenya: (i) ‘wild-feeding’ baboons that foraged naturally in a savanna environment and (ii) ‘Lodge’ baboons that had ready access to spatially concentrated human food scraps, resulting in high feeding efficiency and low daily travel distances. Previous work has shown that these differences in lifestyle and diet lead to dramatic phenotypic differences: Lodge animals grow faster and exhibit higher insulin, cholesterol, and body fat levels compared to wild-feeding animals

12

(Muruthi et al. 1991; Altmann et al. 1993; Kemnitz et al. 2002). In total, we identify over a thousand CpG sites that are differentially methylated between the two feeding groups, providing some of the first evidence that ecological variation shapes the epigenome in a wild mammal (Lea et al. 2016a). Further, we show that some of the most extreme epigenetic changes occur at genes linked to obesity and body fat accumulation in humans and laboratory rodents, supporting the role of DNA methylation in translating signals from the environment into variation in evolutionarily relevant traits (Lea et al.

2016a).

1.4 Functional effects of variation in DNA methylation

1.4.1 Key questions and current gaps in knowledge

Ultimately, both evolutionary and medical researchers aim to identify causal connections between early life conditions, changes in DNA methylation levels, and variation in fitness-related traits. As described above, most work so far has focused on addressing the first part of this causal cascade – the effect of early life experiences on variation in DNA methylation levels. In contrast, substantially less attention has been paid to whether these environmentally- induced epigenetic changes have any downstream effects on gene regulation.

Establishing this second link is important, because not all changes in DNA methylation actually affect gene regulation within the cell. For example, previous

13

studies have experimentally removed DNA methylation marks at specific CpGs within candidate gene promoters, and have asked how these targeted manipulations affect gene expression. Surprisingly, this work has revealed that, even for genes known to have DNA methylation-sensitive gene expression, not all CpG sites have equivalent (or even detectable) effects (Wang et al. 2012;

Maeder et al. 2013). Further, while DNA methylation in enhancer or promoter regions is canonically thought to repress the expression of nearby genes by inhibiting transcription factor binding (Jones 2012), in vitro experiments have shown that DNA methylation can both inhibit or facilitate transcription factor binding, depending on the transcription factor of interest (Nan et al. 1998; Perini et al. 2005; Rishi et al. 2010; Zhu et al. 2016). Together, these results suggest extensive heterogeneity in the relationship between DNA methylation and gene expression variation that we have only begun to understand.

Unraveling this relationship and understanding when variation in DNA methylation will translate into downstream effects on gene regulation is crucial.

Specifically, doing so will allow us to interpret the likely phenotypic impact of environmental effects on the epigenome, as well as other sources of variation such as age, disease status, or genotype that are of both biomedical and evolutionary interest. Further, identifying the loci where DNA methylation 14

causally affects gene expression will provide fundamental insight into the biology of gene regulation.

1.4.2 Methods for causally testing the effects of DNA methylation on gene expression

Defining the locus-specific, causal effects of DNA methylation on gene regulation in high-throughput has remained challenging due to a lack of appropriate methodological tools. Thus far, attempts to understand the causal relationship between changes in DNA methylation and changes in gene expression have relied on in vitro experiments outside the cellular context (Mann et al. 2013; O’Malley et al. 2016), on large-scale manipulations of the epigenome that make it impossible to pinpoint locus-by-locus effects (Weaver et al. 2004), or on low-throughput epigenome editing techniques (Thakore et al. 2016). For example, methods that rely on Zinc finger proteins (Rivenbark et al. 2012), transcription activator-like (TAL) effectors (Maeder et al. 2013), or deactivated

Cas9 (Liu et al. 2016; Vojta et al. 2016) can now deliver enzymes involved in methylation or demethylation to specific locations in the genome and thus directly manipulate endogenous DNA methylation levels. Following such experimental manipulations, gene expression levels or other molecular phenotypes can be measured to understand the causal effects of DNA

15

methylation at the locus of interest. Unfortunately, current implementations of these assays are costly and labor intensive, and work in this area has consequently focused on one or a handful of candidate genes. To establish predictive models for the DNA methylation-gene expression relationship and/or validate large sets of differentially methylated regions, an alternative, high- throughput approach is needed.

In chapter 3, we develop and apply ‘mSTARR-seq,’ an approach that combines genome-scale methods for quantifying gene regulatory activity in vivo

(i.e., STARR-seq (Arnold et al. 2013)) with enzymatic manipulation of DNA methylation marks (Klug & Rehli 2006a). The resulting assay can simultaneously assess enhancer activity at hundreds of thousands of loci, as well as the impact of

DNA methylation on this activity, in a single experiment. To demonstrate the utility of our method, we clone over 750,000 human DNA fragments into specially designed CpG-free reporter vectors and ask (i) whether each fragment has regulatory activity (i.e., can act as an enhancer in vivo) and (ii) if this activity is causally altered by the addition of methyl groups to CpG sites. We then use these data to identify genomic features (e.g., endogenous methylation levels, chromatin states, or transcription factor binding sites) that predict methylation- dependent regulatory activity. Together, our results provide a novel, broadly 16

applicable tool for testing the functional consequences of differential DNA methylation on a genome-wide scale. Further, they provide an important window into the conditions under which differential methylation likely ‘matters’ to the organism—setting the stage for more functionally informed analyses of how DNA methylation mediates environmental effects on phenotypic variation.

17

2. Developmental constraints in a wild primate

2.1 Introduction

Early life environments can dramatically shape a range of adult phenotypes, from molecular control of gene expression and stress reactivity (Weaver et al. 2004), to immune function (Chen et al. 2011) and disease risk (Roseboom et al. 2006). For example, humans subjected to famine in utero exhibit higher rates of obesity, heart disease, and schizophrenia in adulthood than siblings conceived under normal conditions (Jones

1994; Roseboom et al. 2006). In addition, resource competition in early life predicts faster reproductive senescence in wild red deer (Nussey et al. 2007) and explains ~35-55% of the variation in composite measures of fitness in adult sheep, deer, and goats (Hamel et al. 2009). However, while a relationship between early environments and fitness-related traits is well documented across species, the selective import of early environment- induced plasticity remains unclear.

Two influential hypotheses have been proposed to explain the evolution of plasticity in response to the early environment. The predictive adaptive response (PAR) model hypothesizes that animals have evolved to adjust their phenotype during development in anticipation of predicted adult conditions (Gluckman et al. 2005a;

Gluckman et al. 2005b). Individuals that encounter adult environments similar to their early conditions are predicted to gain a selective advantage, whereas animals that experience a mismatch between early life and adult environments should suffer a fitness

18

cost (Gluckman et al. 2005b; a; Monaghan 2008). The PAR model is well-known as an explanation for the long-term effects of in utero nutrition on later metabolic disease (e.g.,

Bateson et al. 2004). However, PARs can also include a much broader set of associations between early life conditions and later life traits, including cases that extend beyond gestation to later developmental periods (e.g., stress reactivity in rats in response to post-natal maternal behavior, Meaney 2001; see Gluckman et al. 2005b; Spencer et al.

2006 for additional examples).

Perhaps the most widely known examples of putative PARs come from studies of human cohorts that experienced scarcity in early life, but resource-rich environments in adulthood. These cohorts often exhibited elevated rates of cardiovascular disease and metabolic syndrome in adulthood, a pattern consistent with PAR-related mismatches

(Barker et al. 1993; Hales & Barker 2001; Gluckman et al. 2005a; Wells 2007). However, in these cases we usually do not know whether cohorts that experienced scarcity in early life would have been advantaged if resources were limited in adulthood, a test required to distinguish PAR from the alternative ‘developmental constraints’ model (also known as the ‘silver spoon’ hypothesis; Grafen 1988; Monaghan 2008). This alternative model predicts a simple relationship between early environmental quality and adult fitness: individuals born in high-quality environments experience a fitness advantage regardless of the adult environment.

19

Because developmental constraints and PAR predict identical outcomes when conditions in adulthood are favorable–individuals from high-quality early environments perform better–the same datasets have been invoked as support for both hypotheses

(Barker et al. 1993; Gluckman et al. 2005a; Wells 2007). In addition, associations between early life conditions and adult physiology or health are often interpreted as evidence for

PARs (Gluckman and Hanson 2004a; Bol et al. 2010; Kemp et al. 2012, see discussion in

Hayward et al. 2013). However, the PAR and developmental constraints models can only be distinguished by comparing fitness-related traits in individuals from high- and low-quality early environments, when each of these sets of individuals experience both high and low-quality adult conditions.

Indeed, many authors have argued that this is the necessary empirical test to distinguish between the PAR and developmental constraints models (Rickard &

Lummaa 2007; Monaghan 2008; Wells 2012; Uller et al. 2013), and a number of such tests have now been conducted in laboratory settings. A recent meta-analysis of studies in plants, arthropods, fish, and reptiles concluded that evidence for the PAR hypothesis is generally weak (Uller et al. 2013). However, because the PAR model is frequently invoked in the context of human health and evolution (Gluckman & Hanson 2004a; b;

Gluckman et al. 2005a), research on mammalian species—particularly natural populations—is needed to further assess its generality. To date, only one study has investigated the fitness effects of matched and mismatched early life and adult

20

environments in wild mammals. In this study, Douhard and colleagues analyzed data from two populations of wild roe deer, one living in a resource-rich environment and one living in a resource-poor environment (Douhard et al. 2014). In the resource-rich environment, the authors found no evidence in support of PAR. However, in the resource-poor environment, deer born in low-quality years sometimes outperformed those born in high-quality years—a pattern consistent with PAR, but also with viability selection. Specifically, viability selection may eliminate all but the most robust offspring during low-quality years, leading to high adult survival rates for individuals born in poor conditions (Douhard et al. 2014).

Here, we take advantage of data from a long-term study of wild baboons (Papio cynocephalus) in Kenya’s Amboseli ecosystem, to further evaluate the support for the

PAR versus developmental constraints hypotheses. Animals in this population experience considerable environmental variability, particularly in annual rainfall levels, and this variation has known consequences for fertility-related traits in females (Alberts

& Altmann 2003; Beehner et al. 2006). Importantly, individuals in this population are observed from birth throughout their lives, and measures of fertility are therefore available for the same females across ecologically variable years. Using these longitudinal data, we were thus able to examine fertility outcomes, for the same female, when her adult conditions both matched and mismatched her early life environment.

21

We focused our primary analysis on an extreme climatic event that occurred in

2009, when the Amboseli ecosystem experienced the worst drought recorded in over four decades (Figure 1). We treated this period as a ‘natural experiment’ that allowed us to ask whether early life conditions explained variation in females’ reproductive responses to a severe ecological challenge. Specifically, we asked how a given female’s fertility changed during the 2009 drought relative to her fertility during a high-quality year. To do so, we analyzed each female’s ability to (i) resume cycling after postpartum amenorrhea or (ii) conceive an offspring during the 2009 drought and during at least one

“high-quality year” (i.e., a year with average rainfall). We then explicitly tested whether a female’s fertility was highest when her adult conditions matched her early life conditions (note that we define early life as the 365 days following birth, and hence adopt the definition of the PAR model described by Gluckman et al. 2005b). Under the

PAR hypothesis, females that experienced high-quality early life conditions should perform better during high-quality adult years, and females born during dry, low- quality years should perform better during low-quality drought years. In contrast, under the developmental constraints hypothesis, females born during low-quality years should be disadvantaged relative to females born in high-quality years, regardless of the adult environment.

22

700

600

500

400

300 Cumulative Rainfall (mm) over 365 days (mm) over Rainfall Cumulative 200

1980 1985 1990 1995 2000 2005 2010

Hydrological year

Figure 1: Consecutive low rainfall years magnified the severity of the 2009 drought. Cumulative rainfall by hydrological year (1-Nov to 31-Oct) is plotted for all years recorded by the ABRP (1977 – 2011; note that our study focused on females born between 1985 and 2004, and examined fertility measures for these females in the years 1992-2011, when they were multiparous adults). The middle 40% and 90% of the cumulative rainfall distribution are highlighted in blue and purple, respectively. 2009 was the driest hydrological year on record and was preceded by another extremely dry year (sequence circled in gray).

Finally, we also tested whether high early life social status could mitigate the effects of early life ecological adversity. This analysis was motivated by observations

23

that high maternal social status predicts faster offspring growth and earlier offspring maturation in the Amboseli baboons (Altmann & Alberts 2005; Charpentier et al. 2008).

We therefore predicted that high maternal dominance rank might offset the costs associated with early ecological adversity, illustrating the highly multidimensional environment experienced by individuals in natural populations.

2.2 Materials and Methods

2.2.1 Study subjects, fertility data, and social status data

Study subjects were 50 adult female baboons from a long-term study population in Kenya’s Amboseli ecosystem (Alberts & Altmann 2012). Subjects were born between

1985 and 2004, and were individually recognized and monitored from birth. Subjects were observed on a near-daily basis, and reproductive status (pregnant, lactating, cycling, and, if cycling, stage/day of the estrous cycle) was scored retrospectively for each female in adulthood based on observations of the sexual skin and the paracallosal skin (Beehner et al. 2006; Gesquiere et al. 2007). Resumption of cycling was determined through observations of swelling of the sexual skin following a period of post-partum amenorrhea. Conception was documented through failure of both menstruation and sexual swelling after the luteal phase, as well as a subsequent diagnostic change of the paracallosal skin from black to pink. The first day of swelling deturgescence (size decrease) in the previous estrous cycle was taken as the conception date (Beehner et al.

2006). Female dominance hierarchies were constructed monthly for every social group in

24

the study population based on the outcomes of dyadic aggressive encounters; ordinal dominance ranks were assigned to every adult female based on these hierarchies

(Hausfater 1974).

2.2.2 Rainfall data and the 2009 drought

Precipitation was measured in millimeters using a rain gauge that was read daily, across all hydrological years beginning in 1977. Hydrological years in Amboseli begin in November, at the end of an annual 5-month dry season (June-October), and continue until the end of the following October. During the 2009 hydrological year

(November 2008-October 2009), we observed the lowest cumulative rainfall in the history of the long-term study. In addition, the 2008 hydrological year had the fourth- lowest cumulative rainfall in our records, heightening the effects of low rainfall during

2009 (Figure 1). The 2009 drought became the focus of our analysis of low-quality environments in adulthood (see below).

2.2.3 Defining low-quality environments in early life and adulthood

To identify low-quality environments in early life, we calculated cumulative rainfall over the first year of life for every female that reached reproductive maturity in our population (N = 289 females), including the 50 subjects of this study. Females in our dataset experienced a wide range of cumulative rainfall values over the first year of life of life (mean = 334.29; range = 151.4-767.0 mm; for comparison an arid desert is defined by <250 mm annual rainfall; Noy-Meir 1973). We defined years with cumulative rainfall

25

values within the bottom 30% of this distribution as low-quality early life environments.

We focused primarily on the first year of life as our early environment of interest, because nutrition during this time period has well-described effects on adult fitness in

Amboseli (Altmann 1991). However, because environmental conditions in utero can also be important, we performed parallel analyses that considered rainfall during the gestational and perinatal period (i.e., where the early life environment was defined by rainfall over the 365 days following conception, which includes gestation and the 1st six months of life).

We defined the 10-month dry period from 1-January to 31-October 2009 as ‘the

2009 drought’, and focused on this period as the primary low-quality environment of interest for adult females. We focused on this 10-month period because of its extreme properties: during these 10 months, only 72.7 mm of rain fell, compared to a mean of

224.28 ± 104.03 s.d. mm during these months in 34 other years. We excluded two months of the 2009 hydrological year (November and December 2008) from our analyses because 68.2 mm of rain fell during this period (i.e., almost as much rain as the entire 10 month drought period). This level of rainfall is low, but within one standard deviation of the mean rainfall for November – December periods in Amboseli (120.30 ± 66.48 mm), in contrast to the much more extreme drought conditions in the subsequent 10 months of

2009.

26

2.2.4 Defining high-quality environments in early life and adulthood.

We initially defined high-quality environments as years with rainfall in the middle 40% of the hydrological year distribution (across all years for which rainfall data was available, 1977-2011; Figure 1). Importantly, the distribution of rainfall values for high-quality early life and high-quality adult years defined in the manner described here were statistically indistinguishable (Kolmogorov-Smirnov test, D = 0.073, P = 0.996).

Defining high-quality environments as those with rainfall in the middle 40% of the distribution excluded very high rainfall years from the definition of high-quality early life and adult environments. While low precipitation has well-documented negative effects on baboon behavior (Alberts et al. 2005), life history (Charpentier et al.

2008), and fertility (Beehner et al. 2006), the effects of very high rainfall on baboons in

Amboseli are unknown and may include negative outcomes (e.g., flooding and thermoregulatory stress). We tested this a priori hypothesis by exploring how our results varied depending on the threshold we used to define high-quality conditions (see below,

Sensitivity to low and high rainfall in early life and adulthood).

2.2.5 Within-female analyses: interaction between early life and adult conditions on female fertility

We examined fertility for females born in either a low-quality environment or high-quality environment, restricting our dataset to females that experienced both the

2009 drought and at least one high-quality year as a multiparous adult. Nulliparous females exhibit marked differences in fertility characteristics, so we excluded them

27

(Altmann 1980; Gesquiere et al. 2007). For the remaining set of females (N = 50 females,

N = 172 female-years), we collated records of resumption of cycling and of conception based on long-term records. We considered other fertility measures, in particular the production of weanlings, Females that successfully weaned an infant represented a subset of the females that resumed cycling—specifically, the subset that resumed cycling after weaning a live infant as opposed to the subset that resumed cycling after losing an infant. Thus, analyzing weaning success would involve partitioning the analysis of cycle resumption into its two components: successfully weaning an infant versus losing an infant. While infants in our dataset experienced an overall infant mortality rate of ~22%, only a small number of infant deaths died before weaning in any given year. This translated into too little variation in infant survival/weaning success to perform a third, parallel analysis.

We modeled resumption of cycling and conception separately, using generalized linear mixed models with a binomial error structure and a logit link function. Models were fit using the function lmer in the R package lme4, and the significance of each fixed effect was estimated from the Wald Z-statistic (R Development Core Team 2012; Bates et al. 2014). In both models, our outcome variable was a binary variable indicating whether a given female resumed cycling or conceived during a 10-month period. Both models also included (i) a binary fixed effect variable classifying the 10-month period as drought

(1-January 2009 to 31-October 2009) or as high-quality (1-Jan to 31-Oct of a high-quality

28

year); (ii) a binary fixed effect variable classifying the female’s early life environment as low-quality or high-quality; and (iii) a random effect of female identity. We interpreted a significant main effect of current rainfall as evidence that prevailing conditions influenced reproductive outcomes (either resumption of cycling or conception). We interpreted a significant interaction between early and current rainfall conditions as evidence that females born in low-quality versus high-quality environments were differentially affected by ecological conditions in adulthood.

In both the resumption of cycling models and conception models, we controlled for three variables known to affect fertility in Amboseli females (Alberts & Altmann

2003; Beehner et al. 2006; Altmann et al. 2010). (i) The age of the female at the start of the time period, modeled as both a linear and a quadratic effect (after Beehner et al. 2006).

All females in our dataset were born in the study population, and ages were thus known to within a few days’ error and modeled as a continuous variable. (ii) The total number of individuals in the female’s social group at the start of the time period, which influences levels of resource competition and thus potentially fertility. (iii) An ordinal number describing the social status (i.e., dominance rank) of the female at the start of the time period (Hausfater 1974). All three variables were incorporated in our models as fixed effects.

In addition, because a female’s ability to conceive during a 10-month time period will depend on her immediate reproductive history, we also controlled for several

29

measures of “reproductive readiness.” In both models, we controlled for: (i) a continuous variable indicating the number of days since the female’s last live birth; and

(ii) a binary variable denoting whether the female had a dependent infant (< 1 year) that died during the 10-month period (because females that lose a dependent infant tend to quickly cease lactating and resume cycling). In our model for resumption of cycling, we also included a binary variable denoting whether the female had cycled at the start of the 10-month period. This variable controlled for the fact that females who are already cycling are less likely to cease cycling and then resume cycling again during the 10- month period. Similarly, in our conception model, we included a binary variable denoting whether the female was already pregnant at the beginning of the 10-month period.

2.2.6 Between-condition analyses: testing for early life effects on adult fertility in a given adult environment

Our dataset was restricted to females born in low-quality or high-quality environments, who experienced both drought and high-quality conditions as adults. In the analyses described above, we took advantage of this design to test whether within- individual differences in sensitivity to the adult environment were dependent on early life ecological circumstances. These within-individual analyses therefore indicate whether females born in low-quality environments showed a larger negative response to adverse environments in adulthood, compared to females born in high-quality environments.

However, these analyses do not directly compare females born in high-quality 30

environments to females born in low-quality environments, controlling for adult environment. Thus, we also conducted between-condition analyses that directly compared all females born in low-quality environments to all females born in high-quality environments, in both 2009 (N = 51 females) and in randomly selected high-quality years

(N = 105 females). For these between-condition analyses, we did not constrain the data set to females who had experienced both the 2009 drought and a high-quality year as a multiparous adult.

2.2.7 Sensitivity to low and high rainfall in early life and adulthood

To assess the sensitivity of fertility traits to low rainfall conditions, we explored how the magnitude of the interaction effect between early and current rainfall varied depending on the thresholds we used to define low-quality environments. We set thresholds in increments of 1% between the lowest 10% and the lowest 30% of the distribution of annual rainfall, for both early and adult environments. For every threshold value between the lowest 10% and the lowest 30% of the rainfall distribution, we estimated the interaction effect from the generalized linear mixed effects model for conception probabilities described above. This analysis allowed us to assess whether the interaction between current and early rainfall environments remained significant (i) when our definition of a low-quality adult environment included years that were not as severe as the 2009 drought, and (ii) when our definition of low-quality early

31

environments included years of varying severity. For this analysis, we retained the same definition of high-quality years (i.e. the middle 40% of the distribution).

We also tested the effects of very high rainfall on fertility outcomes, which we suspected might be costly to Amboseli females as well. Here, we varied the thresholds we used to define high-quality early and adult environments and estimated the interaction effect from a generalized linear mixed effects model predicting conception.

Specifically, our main analyses defined high-quality environments as years falling between the 30th and 70th percentiles of the early life or adult rainfall distribution; in our sensitivity analyses, this upper cutoff ranged from the 70th to 100th percentile (in increments of 1%), thus including progressively more extreme high rainfall years.

2.2.8 Effects of maternal dominance rank on female fertility

To assess the effects of early life social status on fertility during the 2009 drought, we measured maternal dominance rank, defined as the dominance rank of the mother at the offspring’s birth. Dominance rank is maternally inherited in baboons (Hausfater

1974; Hausfater et al. 1982), and thus current dominance rank and maternal dominance were collinear in our dataset (r2 = 0.67). We therefore tested whether maternal dominance rank accounted for additional variation in fertility after accounting for the effect of an individual’s current dominance rank. Specifically, we extracted the residuals from a generalized linear model predicting resumption of cycling or conception during the 2009 drought (as a function of age, group size, the focal female’s own dominance

32

rank, and all reproductive readiness measures, N = 50). We then used a linear model to ask, for females born in low-quality environments (N = 14), whether maternal dominance rank was correlated with residual variation in cycle resumption or conception probabilities during the 2009 drought. We note that the analysis of residuals can produce biased estimates of effect sizes (Darlington & Smulders 2001; Freckleton

2002). In this case, our analyses may underestimate the effect of interest because true maternal dominance rank effects could be masked by first taking into account the focal individual’s current dominance rank; thus, our analyses are likely to be biased towards type II error (false negative) rather than type I error. We chose to conduct such an analysis because we were interested in testing for maternal social status effects specifically for females born in low-quality environments during the 2009 drought only

(N = 14). As a post hoc contrast, we also repeated this procedure for females born in high- quality environments (N = 36).

All analyses described in the Materials and Methods were conducted in R version 2.15.0 (R Core Development Team 2012).

33

2.3 Results

2.3.1 Fertility declines during the 2009 drought were greater for females born in low-quality environments than for females born in high-quality environments

Overall, females were less likely to resume cycling after postpartum amenorrhea and less likely to conceive during the 2009 drought than during high-quality years

(estimates for high-quality years versus the 2009 drought: ! = 2.58 ± 0.83 s.e., p = 0.002, for resumption of cycling; ! = 3.49 ± 1.34 s.e., p = 0.008 for conception, N = 172 female- years). This estimate translates to a striking 24% decrease in the probability of conception, for the average female, during the 2009 drought relative to high-quality years.

However, our within-female analysis indicated that the cost to fertility associated with the 2009 drought varied across females. For females born in high-quality environments, conception rates were less affected by the 2009 drought than for females born in low-quality environments (interaction between early and adult environments:

! = -3.01 ± 1.49 s.e., p = 0.043). A similar trend was observed for resumption of cycling

(! = -1.70 ± 0.93 s.e., p = 0.069; Figure 2, Tables 1 and 2). Differences in sensitivity to drought conditions were not explained by genetic relatedness: females born in high- quality environments were not more related than females born in low-quality environments.

34

A B 22

2 1

1 1 0

0 0 N = 36 Bandwidth 0.37 1

N = 14 Bandwidth = 0.5 −

1 -1 Residualvariation

− density.default(x = dnd$resids_cyc) 2

density.default(x = dd$resids_cyc, bw = 0.5) density.default(x = dd$resids_cyc, −

2 Morelikely resumeto cycling − -2

2009 drought High-quality years 2009 drought High-quality years

0.5 0.0 0.5 −

0.5 0.0 0.5

-0.5 0 0.5 − -0.5 0 0.5

Density Females born in low-qualityDensity years Females born in high-quality years C D 2 2 2 1 1 1

0 0 0 N = 14 Bandwidth 0.25

N = 36 Bandwidth = 0.3258

1 1 -1 Residualvariation

− −

density.default(x = dnd$resids_conc) Morelikely conceiveto

2 2 density.default(x = dd$resids_conc, bw = 0.25) density.default(x = dd$resids_conc, -2

−

2009 drought High-quality years − 2009 drought High-quality years

0.5 0.0 0.5 −

0.5 0.0 0.5

-0.5 0 0.5 − -0.5 0 0.5 Density

Females born in low-qualityDensity years Females born in high-quality years

Figure 2: Females born in low-quality environments were less likely to (A) resume cycling and (C) conceive during the 2009 drought than in high-quality years. Females born in high-quality environments were (B) equally likely to resume cycling and (D) only slightly less likely to conceive during the 2009 drought. Plots are density distributions of residuals from linear mixed effects models predicting resumption of cycling and conception, respectively, and controlling for individual characteristics, reproductive history, and demographic factors that influence fertility (N = 50). Residual values are shown on the y-axis and density on the x-axis. The mean value for each distribution is indicated by a black horizontal line.

35

Table 1. Results from a generalized linear mixed effects model predicting resumption of cycling.

Estimate P Explanation of effects (P < 0.10) (Intercept) 1.223 0.506

Age -0.348 0.167 Age2 0.015 0.137 Current rainfall During the 2009 drought, females were less likely to resume 2.578 0.002 environment cycling. Early rainfall 0.785 0.318 environment Females born in low-quality environments were more likely to Current * early resume cycling during high-quality years than during the 2009 -1.695 0.069 rainfall environment drought. Females born in high-quality environments were more robust to changes in prevailing ecological conditions. st

3 Females cycling on January 1 were less likely to resume

6 Cycling -1.619 0.001 cycling over the following 10 months. Females that had recently given birth were less likely to Time since live birth 0.003 0.008 resume cycling. Infant death 0.722 0.227 Dominance rank 0.019 0.529 Group size -0.006 0.527

Table 2. Results from a generalized linear mixed effects model predicting conception.

Estimate P Explanation of effects (P < 0.10) (Intercept) -5.872 0.024

Age 1.255 0.001

Age2 -0.054 0.001 Very young and very old females were less likely to conceive. Current rainfall 3.493 0.009 During the drought, females were less likely to conceive. environment Early rainfall 1.562 0.249 environment Females born in low-quality environments were more likely to Current * early conceive during high-quality years than during the 2009 rainfall 3.015 0.043 drought. Females born in high-quality environments were more environment robust to changes in prevailing ecological conditions. 3

7 Females pregnant on January 1st were less likely to conceive Pregnant 6.729 <0.001 over the following 10 months. Time since live Females that had recently given birth were less likely to 0.005 <0.001 birth conceive. Following a recent infant death, females were more likely to Infant death 6.087 <0.001 conceive. Dominance rank 0.040 0.341 Group size 0.043 0.001 In larger groups, females were less likely to conceive.

When we shifted our definition of the early environment to include gestation and the first 6 months of life rather than the first year of life, we found less evidence that early rainfall influenced fertility in an adult environment-dependent manner.

Specifically, we did not detect evidence of an interaction effect on cycle resumption or conception probabilities (! = 0.167 ± 0.782 s.e., p = 0.830 for resumption of cycling; ! =

0.195 ± 0.993 s.e., p = 0.844 for conception).

Finally, our between-condition analysis, examining fertility in the 2009 drought alone, suggests that females born in high-quality environments were more likely to conceive than females born in low-quality environments (t = -1.646, p = 0.107). The direction of this trend was the same, although weak, for resumption of cycling (t = -

1.139, p = 0.263). However, the two groups were completely indistinguishable for both fertility measures when they were compared in high-quality adult environments

(resumption of cycling: p > 0.10 for 97% of tests; conception: p > 0.10 for 100% of tests).

Together, our results suggest that the long-term costs of early ecological adversity become most apparent during periods of ecological challenge later in life.

2.3.2 Sensitivity of female fertility to both extreme low and extreme high rainfall

We found that the interaction between early and adult environmental conditions remained significant, and was similar in magnitude, across a range of thresholds for defining low-quality early environments (lowest 10th to lowest 30th percentile of low rainfall years, Figure 3). In contrast, only the driest 10% of adult years yielded 38

significant, qualitatively similar results to our main analysis. These years equate to the driest 3 years experienced by females in our dataset, with ≤ 188.5 mm of cumulative rainfall (hydrological years 2000, 2008 and 2009). Together, our results suggest that relatively modest droughts in early life can impose long-term costs on female fertility, but that such costs are strongly expressed only during severe droughts during adulthood.

● ● ● ● ● 5

● ● ● ● ● ● ● ● ● ● ● ● ● ● 4 ● ● ● ● ● ● ● ● ● ● ● ● ● ●

3 ●

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 ● ● ● ● ● ● ● 0−30 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0−25 ● ● ●

abs(output_df3$V1) ● ● ● ● ● ● ● ● ● ● ● ● 0−20 ● ● ● ● output_df3$V4 ● ●

1 0−15 early environment Magnitude of the interaction effect interaction the of Magnitude ● ● ● 0−10 0−5

0 0 0 0−5 0−10 0−15 0−20 0−25 0−30 Percentiles of the first year of life rainfall Percentiles of the hydrological year rainfall distribution defining the output_df3$V3 distribution defining the adult environment

Figure 3: Interactions between adult and early life environments under alternative definitions of low-quality environment. The magnitude (i.e., the absolute value) of the interaction effect between adult and early environments on female conception is 39

plotted on the Z axis, as a function of different definitions of a “low-quality” environment. Percentiles used as cutoffs for low-quality environments are plotted on the X axis and Y axis, respectively. Colors and symbols are used for contrast and to delineate data points associated with a given low-quality adult year. For cases in which model estimates did not change as a result of changes in the definition of the adult environment (because no new data were included in the analysis), points are not plotted. The light blue points represent the interaction between early and adult conditions when 2009 was the sole low-quality year in adulthood. The dark blue points represent the same interaction when low-quality years in adulthood included both 2009 and the next driest year, and so on. For the two most extreme definitions of low-quality adult years (light blue and dark blue points: rainfall in the lowest 10% of hydrological years), the interaction effect was significant across all definitions of a low-quality early year (P<0.05), indicating that female fertility is more sensitive to low rainfall in early life than to low rainfall in adult life. The black arrow points to the estimate reported in the main text (i.e., where 2009 is the only low-quality adult year and low-quality early life is defined as the lowest 30% of the distribution for cumulative rainfall in the first year of life).

Similarly, changing our definition of high-quality environments to include very high rainfall years weakened the evidence for an early life-adult life interaction, suggesting that very high rainfall years are qualitatively different from normal rainfall years, and may be stressful for female baboons. Specifically, when extremely high rainfall years (rainfall >88th percentile of years) were included in the definition of high- quality adult environments, fertility during the 2009 drought and “high-quality” years became more similar, attenuating the evidence for an interaction effect between low- quality and high-quality environments. Likewise, for any given definition of adult environmental conditions, the magnitude of the interaction effect declined as more high- rainfall years were included in the definition of high-quality early life environments.

Together, these results suggest that female fertility is highest during moderate rainfall

40

periods in Amboseli, in support of our choice to define high-quality years based on the middle of the overall distribution.

2.3.3 Protective effects of high social status in early life

Finally, we hypothesized that another important component of the developmental environment, maternal social status, might buffer the effects of early life drought on adult fertility. Indeed, for females born in low-quality environments (N =

14), fertility was significantly less depressed by the 2009 drought for females born to high-ranking mothers compared to females born to low-ranking mothers (resumption of cycling: β = -0.112± 0.050 s.e., R2 = 0.296, p = 0.044; conception: β = -0.065 ± 0.027 s.e. p =

0.033, Figure 4). In contrast, maternal dominance rank did not significantly affect the probability of conception (β = 0.031 ± 0.023 s.e., p = 0.178) or cycle resumption (β = 0.008 ±

0.025 s.e., p = 0.751) for females born in high-quality environments (N = 36). Thus, high maternal dominance rank provided important protection for females exposed to early life ecological adversity, but did not affect adult fertility for females born in high-quality ecological conditions.

41

A B

1.0 1.0 probability

0.5 0.5 probability

0.0 0.0

-0.5 -0.5 Residual variation in conception in variation Residual Residual variation in cycle resumption cycle in variation Residual

-1.0 -1.0

Residual variation in conception in variation Residual Residual variation in resumption of cycling of resumption in variation Residual -1.5 -1.5

1 4 7 10 1 4 7 10 Maternal dominance rank Maternal dominance rank Maternal dominance rank Maternal dominance rank

Figure 4: Fertility during the 2009 drought: females born in low-quality ecological conditions were protected by high-quality social environments. Females born to high ranking mothers in low-quality environments were more likely to resume cycling (A) and conceive (B) during the 2009 drought than females born to low ranking mothers in low-quality environments. The best-fit lines from a linear regression of maternal dominance rank on the residuals of a model relating resumption of cycling or conception probability to ecological and reproductive readiness variables (N = 14) are plotted with a black dashed line. The mean performance for females born during high-quality years during the drought are indicated with arrows.

2.4 Discussion

2.4.1 Support for the developmental constraints model in wild baboons

Overall, our analyses do not support the PAR hypothesis: individuals born in low-quality environments did not outperform other individuals when they encountered similar low-quality conditions in adulthood. Instead, our within-female analysis

42

revealed that females born in low-quality environments experienced an estimated mean decrease in conception probabilities of 60.0% during the 2009 drought. In contrast, females born in high-quality environments experienced an estimated mean decrease in conception probabilities of only 10.2%—an almost six-fold difference in effect sizes.

Notably, our between-condition analysis (which focuses on differences between females born in different environments, instead of the degree to which individual females are affected by the adult environment) indicated that the two sets of females were statistically indistinguishable in high-quality adult environments, at least with respect to the fertility traits we investigated here. Taken together, these results indicate that females born in high-quality environments were more resilient to drought conditions than females born in low-quality environments, but they were not advantaged over females born in poor conditions during normal years. Thus, females born in high-quality environments may experience a long-term advantage over females born in low-quality environments, but only if the population is exposed to adverse conditions.

Consequently, early life effects are most accurately assessed by measuring fitness effects across a wide range of adult conditions (Nussey et al. 2007; Hamel et al. 2009).

We concentrated on fertility-related outcomes in our study, rather than mortality, because annual death rates are low in baboons, as in most primates (Morris et al. 2011).

Specifically, the small number of adult female deaths that occurred in the study population in 2009 (N = 13 out of 100 adult females) were too few to assess the effects of

43

early life conditions on survival during the 2009 drought. Thus, while our findings exclude a PAR explanation for fertility-related fitness components in the Amboseli baboons, the PAR model could, in principle, hold for survival-related fitness components—a possibility that demands further testing.

Our results are consistent with studies of both wild and laboratory animals that also report consistent advantages to individuals born in high-quality environments

(Reid et al. 2006; Nussey et al. 2007; Uller et al. 2013). In addition, they are consistent with tests of the PAR versus developmental constraints models in preindustrial humans and wild roe deer (although these studies did not use within-subjects designs, as presented here; Hayward and Lummaa 2013; Douhard et al. 2014). For example,

Douhard and collegues found that, in a resource-rich habitat, wild female roe deer born in high-quality years experienced a lifelong fitness advantage over individuals born in low-quality years. Similarly, Hayward and Lummaa (2013) demonstrated that, in a preindustrial Finnish population, poor early life environments were associated with increased later life mortality. In both cases, these patterns held regardless of the adult environment. Notably, some support for the PAR hypothesis was identified by Douhard and collegues in a second population of wild roe deer, which inhabited a more marginal environment. Specifically, females born in low-quality years were more likely to survive in harsh adult conditions than females born in high-quality years (Douhard et al. 2014).

However, the authors attributed this result to increased viability selection on deer born

44

in poor environments rather than PAR: under poor early-life conditions only the most robust roe deer survived to adulthood. Because we focused on within-individual comparisons of performance in both low-quality and high-quality years in adulthood, viability selection could not have influenced our results. This difference in study design may account for the absence of support for PAR in our study.

2.4.2 Factors that influence the relationship between early life ecology and adult fertility

While our findings are largely consistent with the developmental constraints hypothesis, we found that these constraints were contingent on at least two factors. First, we found that the negative consequences of early adversity were highly contingent on the adult environment. Specifically, we found that the long-term consequences of early adversity were only detectable in the worst conditions during adulthood, consistent with evidence from laboratory mice (Giovanoli et al. 2013) and humans (Hayward et al.

2013). For example, in a pre-industrial human population, individuals born during periods of low-crop yield exhibited lower survival and fertility during periods of famine compared to individuals born during periods of high-crop yield (Hayward et al. 2013).

Thus, the effects of early life adversity may, in certain cases, remain undetected unless individuals face challenging circumstances later in life—a possibility with important ramifications for understanding resilience.

Second, we found that the severity of early life drought effects was contingent on an individual’s early life social environment, highlighting the multifaceted nature of the 45

early environment. We estimate that, for females born in low-quality environments, each improvement of one maternal rank position would increase resumption of cycling probability and conception probability during severe drought years by 1.8% and 1.3%, respectively. Given that 14 rank positions separate the highest- and lowest-ranking females in an average social group, and given that severe drought conditions occur approximately once each decade (Figure 1), differences in female fertility during droughts could create a significant lifetime advantage for the offspring of high-ranking mothers. This fertility advantage could arise through several mechanisms. For example, offspring of high-ranking mothers may have increased access to resources, which could protect them from nutritional deficits during dry years (Whitten 1983; Hayward et al.

2013). Alternatively, high-ranking females could experience reduced social stress

(Sapolsky 2005), which could have cascading effects on their offspring’s physiology

(Onyango et al. 2008). Interestingly, similar protective effects of high-quality early life social environments have also been described in humans. Individuals that grew up in low socioeconomic status (SES) households, but experienced high levels of maternal warmth and affection, did not develop the chronic inflammation typically associated with a low SES childhood (Chen et al. 2011). Social buffering against other sources of adversity may thus have a long evolutionary history in the primate lineage.

46

2.4.3 Conclusions

Our results join other recent studies (Hayward et al. 2013; Uller et al. 2013;

Douhard et al. 2014) to motivate a reconsideration of widely accepted PAR explanations for early life effects on later life phenotypes, including high disease rates in human cohorts born during famines. The more common situation appears to be that individuals born in low-quality environments experience a lifelong disadvantage relative to individuals born in high-quality early environments. This disadvantage appears to be most acute when the adult environment deteriorates, emphasizing the need to study early life effects under variable environments and across the life course.

We do not believe, however, that the PAR model should be discarded altogether.

It has provided a valuable framework for investigating the evolution of early life effects, and may still prove to be important under a constrained set of conditions (or for some fitness components but not others). Recent studies have attempted to outline the settings in which PARs are predicted to evolve. Specifically, several authors have argued that

PARs should be favored when the environmental cue that triggers alternative phenotypic development is a reliable indicator of the adult environment (Rickard &

Lummaa 2007; Wells 2012; Nettle et al. 2013). This condition clearly does not hold for rainfall in Amboseli (Figure 1), and may rarely be the case for long-lived species. Thus, it is not surprising that the best examples of PARs have been found in short lived species such as Daphnia (Boersma et al. 1998), locusts (Applebaum & Heifetz 1999), and voles

47

(Lee 1988; Horton & Stetson 1992). In contrast, tests of the PAR model in long lived species (e.g., wild baboons, deer, preindustrial humans; Hayward et al. 2013; Hayward and Lummaa 2013; Douhard et al. 2014) have either rejected PAR explanations or identified weak evidence at best. If PARs do prove to be more frequent in short lived species than long lived species, it will be somewhat ironic that the model was originally proposed to account for early life effects in humans—a long-lived species that evolved in a highly variable savanna environment. Studies that investigate the relationship between the evidence for PARs, species life history, and environmental variability will be essential for addressing this question. Until it is resolved, however, we argue that a developmental constraints explanation, rather than PAR, should be treated as the null model in explaining the evolution of early life effects, particularly for long-lived organisms.

48

3. Resource base influences genome-wide DNA methylation levels in wild baboons

3.1 Introduction

Despite a rich history of studies documenting the relationship between ecological and phenotypic variation in natural populations, we know surprisingly little about the molecular mechanisms that mediate these effects. Insight into these mechanisms is important for understanding how natural phenotypic variation emerges and how organisms cope with environmental change. Genomic approaches can contribute to these questions by identifying the genes and pathways involved in sensing and responding to selectively relevant ecological variation. For example, studies in eusocial insects have highlighted the impact of nutrient exposure on genome-wide gene regulation, as well as its contribution to the emergence of distinct castes within a hive or colony (Kucharski et al. 2008; Foret et al. 2012). At the same time, other studies have identified rapid gene expression responses to song in zebra finch (Drnevich et al. 2012;

Whitney et al. 2014); a strong genome-wide signature of social status in hierarchical primates (Tung et al. 2012); and widespread transcriptional changes associated with mate choice in fish (Cummings et al. 2008). Together, such work points to a fundamental role for gene regulation in mediating physiological responses to environmental inputs.

By altering the expression of genes in the genome, gene regulatory mechanisms permit a range of phenotypic values to arise from an otherwise static genome sequence.

49

Mounting evidence suggests that epigenetic marks, particularly DNA methylation (the best studied to date), play a significant role in mediating the gene regulatory response to environmental conditions. DNA methylation refers to the covalent addition of a methyl group to a cytosine base, and, in mammals, occurs most often at CG dinucleotides (known as ‘CpG sites’). CpG sites are strongly enriched in regulatory sequences (e.g., gene promoters, gene bodies, and CpG-dense regions known as ‘CpG islands’) where changes in methylation can impact the expression of nearby genes. For example, DNA methylation in promoter or enhancer regions can repress gene expression by interfering with transcription factor binding, or by recruiting proteins that induce changes in chromatin accessibility (Klose & Bird 2006; Weber et al. 2007).

Meanwhile, gene body methylation is often associated with increased gene expression, and is thought to aid in transcriptional elongation (Jones 2012).

Genome-wide patterns of DNA methylation are first established during development, and once established, are faithfully transmitted across cell divisions throughout the life of the organism. However, environmental conditions can affect this process, either during development itself (when epigenetic patterns are known to be particularly sensitive: Tobi et al. 2009; Feil & Fraga 2011; Faulk & Dolinoy 2011) or later in life, when changes in DNA methylation help coordinate the cellular response to new environmental stimuli (Guo et al. 2011; Barrès et al. 2012; Pacis et al. 2015). Thus, changes in DNA methylation are thought to provide an avenue through which environmental

50

inputs can stably alter gene expression levels and, as a consequence, mediate environmental effects on organism-level traits (Meaney & Szyf 2005; Jirtle & Skinner

2007; Feil & Fraga 2011).

The relationship between DNA methylation and diet (both caloric intake and dietary content) is particularly well-studied in this regard (Van den Veyver 2002;

Heijmans et al. 2008; Carone et al. 2010; Moleres et al. 2013; Tobi et al. 2014). For example, in laboratory mice, maternal diet during pregnancy predicts offspring fur color and susceptibility to diabetes—a relationship mediated by its stable effects on offspring methylation near the agouti gene, which in turn changes agouti gene expression (Klebig et al. 1995; Wolff et al. 1998; Waterland & Jirtle 2003). Long-lasting effects of diet and resource availability (which we refer to in combination as ‘resource base’) also affect

DNA methylation patterns in humans. In Gambian populations that experience dramatic seasonal fluctuations in food availability, season of conception predicts offspring DNA methylation levels at several metastable epialleles (loci that show consistent, stable epigenetic patterns across tissues: Waterland et al. 2010; Dominguez-

Salas et al. 2014). Similarly, individuals conceived during the Dutch Hunger Winter, a severe war-time famine in the Netherlands, have been shown to exhibit stable differences in DNA methylation levels at both individual growth-related genes (e.g.,

IGF2, INSIGF, and IL10) and on a genome-wide scale (Heijmans et al. 2008; Tobi et al.

2009, 2014). These studies also suggest that diet effects on DNA methylation can be

51

acutely sensitive to timing: long-term changes in DNA methylation levels were only detectable in individuals exposed to the Dutch Hunger winter during the periconception period, but not later in pregnancy (Heijmans et al. 2008; Tobi et al. 2014, but see Tobi et al.

2009).

Variation in resource base, both during development and later in life, is also important in wild mammal populations, where it exerts potent effects on both fertility and mortality components of fitness (Altmann 1991; Gaillard et al. 2000; Beehner et al.

2006; Nussey et al. 2007; Revitali et al. 2009; Hamel et al. 2009). However, in contrast to human populations or lab model organisms, the role of DNA methylation in mediating these effects has not been investigated, leaving questions about the scope and timing of epigenetic sensitivity to the environment unanswered. To address this gap, we profiled genome-wide DNA methylation levels in a long-term study population of wild baboons in the Amboseli region of Kenya (Alberts & Altmann 2012). Specifically, we compared

DNA methylation patterns in ‘wild-feeding baboons’ to those in ‘Lodge group baboons’.

Wild-feeding baboons walked 4-6 km per day, foraging in a dry savanna environment on widely distributed foods. In contrast, while Lodge group baboons resided in the same savanna ecosystem, they had access to spatially concentrated human food scraps.

Lodge group baboons were therefore able to feed more efficiently and travel shorter distances each day to meet their caloric requirements (Muruthi et al. 1991; Bronikowski

& Altmann 1996). In addition, they experienced reduced seasonal and annual variance

52

in resource availability compared to wild-feeding animals. Previous work in Amboseli has documented striking behavioral and physiological differences between these groups. Lodge animals, who expended considerably less energy to achieve the same caloric intake as wild-feeding animals, exhibited higher serum insulin, cholesterol, and body fat levels compared to their wild-feeding counterparts (Muruthi et al. 1991;

Altmann et al. 1993; Kemnitz et al. 2002). Further, Lodge juveniles grew faster and matured earlier than wild-feeding animals, suggesting that the combination of more stable resource availability, higher feeding efficiency, and shorter travel distances translated into measurable fitness advantages (Altmann & Alberts 2005).

Here, we investigated whether environmentally induced changes in DNA methylation levels might contribute to the known phenotypic differences between

Lodge and wild-feeding animals. To do so, we investigated three sets of questions: (i) do the differences in resource base-associated with the Lodge versus wild-feeding conditions significantly predict DNA methylation levels?; (ii) are sites that are differentially methylated by resource base likely to be functionally important?; and (iii) is the signature of resource base stable or plastic over time, when environmental conditions change? For the third question, we drew on samples from eight male baboons that switched from either the Lodge to wild-feeding condition or from the wild-feeding to Lodge condition as a consequence of natal dispersal. We then asked whether pre-

53

dispersal (i.e., early life) or post-dispersal (i.e., adult) resource base left a stronger signature on genome-wide DNA methylation patterns.

3.2 Materials and methods

3.2.1 Study subjects and sample collection

All study subjects were members of a long-term study population of yellow baboons (Papio cynocephalus) that has been monitored by the Amboseli Baboon Research

Project (ABRP) for over four decades (Alberts & Altmann 2012). Our study focused on

69 animals from the ABRP study population, including: (i) 39 baboons that resided in a wild-feeding group from birth until the time of sampling (29 males and 11 females); (ii)

22 baboons that either resided in Lodge group from birth until the time of sampling (11 males and 7 females) or gained access to the Lodge resource base early in their lives (4 females born prior to monitoring of Lodge group in 1982); (iii) 3 males that were born in

Lodge group and dispersed into a wild-feeding group following reproductive maturation, and were sampled in a wild-feeding group as adults; and (iv) 5 males that were presumed to have been born in a wild-feeding group and that dispersed into

Lodge group following reproductive maturation, and were sampled in Lodge (Figure 5).

For the 3 males that were born in Lodge group and sampled in a wild-feeding group, their early histories and dispersal events were directly observed. For the 5 males that were presumed born in wild-feeding groups but were sampled in Lodge group, we inferred their early histories based on two pieces of evidence. First, adult male baboons

54

very rarely remain in their natal group to reproduce (Pusey & Packer 1987), and genetic analysis suggested that these adult males were unrelated to members of Lodge group other than their offspring (Altmann et al. 1996). Second, there were very few social groups associated with human food sources in the ecosystem at the time these males matured, suggesting wild-feeding origins for all or most of these animals.

‘Wild-feeding’ environment

Switched from Lodge to wild-feeding (n = 3)

Lifelong wild- feeding individuals (n = 39) ‘Lodge’ environment

Switched from wild- Lifelong Lodge feeding to Lodge (n = 5) group individuals (n = 22)

Figure 5: Study design. The correspondence between resource base and all study subjects is depicted here. Our main differential methylation analyses focused on individuals that spent their entire life (from birth until the time of sampling) in a wild-feeding social group (n = 39: white circle) or in the Lodge group (n = 22: gray circle). Our analyses of the stability versus plasticity of DNA methylation levels focused on individuals that switched resource base after natal dispersal (n = 8: switching baboons in the center).

To investigate epigenetic differences between Lodge and wild-feeding animals, we combined genome-wide DNA methylation data from a previous study on statistical

55

methods for differential DNA methylation analysis (n = 50 individuals; Lea et al. 2015b) with additional DNA methylation data generated to study resource base effects in this study (n = 19 individuals). All data were derived from whole blood samples collected by the ABRP between 1989 and 2011, following well-established procedures (Altmann et al.

1996; Tung et al. 2009, 2011, 2015). Briefly, animals were immobilized by an anesthetic- bearing dart delivered through a hand-held blow gun, and, following immobilization, were quickly transferred to a processing site for blood sample collection. Following sample collection, study subjects were allowed to regain consciousness in a covered holding cage until they were fully recovered from the effects of the anesthetic. Upon recovery, study subjects were released near their social group and closely monitored.

Blood samples were stored at the field site or at an ABRP-affiliated lab at the University of Nairobi until they were transported to the United States.

3.2.2 Generation and processing of genome-wide DNA methylation data

To measure genome-wide DNA methylation levels, we used a cost-effective, high-throughput sequencing approach known as reduced representation bisulfite sequencing (RRBS) (Meissner et al. 2008; Gu et al. 2011; Boyle et al. 2012). RRBS relies on two key steps: (i) digestion of genomic DNA with the enzyme Msp1, which produces

DNA fragments that begin and/or end with an informative CpG site; and (ii) treatment of Msp1-digested DNA with the chemical sodium bisulfite, which leaves methylated cytosines intact but converts unmethylated cytosines to uracil (and ultimately thymine 56

after PCR). Following high-throughput sequencing and mapping of all reads to a reference genome, CpG site-specific DNA methylation levels can be estimated as the ratio of reads read as cytosine (reflecting an originally methylated version of the base) to the total number of mapped reads (reflecting both methylated and unmethylated versions of the base, i.e., reads read as either cytosine or thymine).

To construct RRBS libraries, we followed the protocol of Boyle and colleagues

(Boyle et al. 2012). For each individual, we created a barcoded library from 180 ng of blood extracted baboon DNA, combined with 1 ng of unmethylated lambda phage DNA to assess the efficiency of the bisulfite conversion. Each sample was sequenced to a mean depth (± SD) of 27.25 ± 13.62 million reads on the Illumina HiSeq 2000 platform.

Following sequencing, we removed adapter contamination, low-quality bases, and bases artificially introduced during library construction using the program Trim Galore!

(Krueger 2015). We then used the program BSMAP (Xi & Li 2009) to map the trimmed reads to the olive baboon genome (Panu 2.0), and to extract the methylated read count and total read count for each individual and CpG site. Before performing differential methylation analyses, we filtered out constitutively hypermethylated and hypomethylated sites from our data set, as well as invariable sites and sites with low levels of mean coverage. Importantly, filtering for hypomethylation should reduce potential biases introduced from mapping yellow baboon RRBS reads to the olive baboon genome.

57

3.2.3 Testing for differences in DNA methylation levels at individual CpG sites

We first tested for a relationship between resource base and DNA methylation levels using data generated from individuals who had spent most or all of their lives

(prior to sampling) in a wild-feeding group (n = 39) or the Lodge group (n = 22). To do so, we used the binomial mixed effects approach implemented in the program MACAU

(Lea et al. 2015b). This approach allowed us to control for kinship in our data set and to work directly with the raw count data – two features that maximize power in bisulfite sequencing data sets. Specifically, for each CpG site, we used MACAU to model DNA methylation levels as a function of the fixed effects of resource base (Lodge or wild- feeding), sex, age of the animal, number of years since blood sample collection, and bisulfite conversion rate. We also included a random effect that accounts for genetic relatedness among individuals (Lea et al. 2015b). For each CpG site tested, we extracted the p-value associated with the resource base term and corrected for multiple hypothesis testing using the false discovery rate (FDR) approach implemented in the R package qvalue (Storey & Tibshirani 2003; Dabney & Storey 2015). We considered a CpG site to be differentially methylated by resource base (referred to below as ‘resource base- associated’) if it passed a 10% FDR threshold.

Because DNA methylation patterns are highly cell type-specific (Reinius et al.

2012; Roadmap Epigenomics Consortium et al. 2015), we also investigated whether differences in whole blood cell type composition between Lodge and wild-feeding

58

animals could confound our analysis. To do so, we drew on two datasets: (i) cell type proportion data generated from manual counts of Giemsa-stained blood smears (for 15

Lodge and 20 wild-feeding animals); and (ii) genome-wide, cell type-specific DNA methylation data from a previous study of human whole blood (Reinius et al. 2012; Jaffe

2015). We used these data to first test whether resource base predicted cell type composition, and then to investigate whether resource base-associated sites were more likely to exhibit cell type-specific DNA methylation patterns, which would suggest a potential confound.

3.2.4 Enrichment of differentially methylated sites by genomic annotation

We hypothesized that, if shifts in DNA methylation are part of a coordinated regulatory response to resource base, these epigenetic changes should be biased towards regions of the genome that control gene expression, and should be targeted towards genes involved in similar biological processes. To test these hypotheses, we evaluated whether our data were consistent with four predictions. Specifically, we expected resource base-associated sites to be: 1) over-represented in putatively functional gene regulatory elements (i.e., gene bodies, promoters, CpG islands, CpG island shores or enhancers) and under-represented in regions of the genome with no known regulatory function; 2) over-represented in chromatin states associated with active gene transcription and under-represented in chromatin states associated with gene repression; 3) more likely to fall in or near genes expressed in whole blood, compared to 59

genes not expressed in blood; and 4) enriched near genes involved in coherent biological pathways and processes.

To test prediction (1), we used publicly available annotation tracks for the olive baboon genome to assign each resource base-associated CpG site to one of the following categories: gene body, promoter, CpG island, CpG island shore, H3K4me1-marked enhancer or unannotated. For each category, we used Fisher’s exact test to test for significant over- or under-enrichment of resource base-associated sites relative to chance expectations. Importantly, we defined chance expectations based on the CpG sites that we actually profiled in our data set, which are themselves enriched for putatively functional regions of the genome. Thus, significant over- or under-enrichment of resource base- associated sites would indicate that resource base-associated sites are even more likely to fall in a given genomic compartment than other sites captured by the

RRBS protocol.

To test prediction (2), we drew on chromatin state annotation data generated by the NIH Roadmap Epigenomics Project for human peripheral blood mononuclear cells.

Chromatin states are defined by combinations of histone marks (acetylation or methylation) and provide information about the transcriptional activity and regulatory element function of the associated DNA. For example, actively expressed gene bodies are associated with a chromatin state defined by H3K36me3 marks, whereas repressed genes are associated with a chromatin state defined by H3K27me3 marks. Importantly,

60

histone marks tend to be highly conserved between humans and closely related primates, such as chimpanzees and rhesus macaques (Zhou et al. 2014); the baboon lineage diverged from the human lineage at the same time as rhesus macaques, supporting the overall accuracy of using Roadmap Epigenomics chromatin states here.

We therefore assigned each resource base-associated CpG site to one of 15 chromatin states. We then investigated the degree to which resource base-associated sites were over- or under-enriched in each chromatin state. As above, we again used Fisher’s exact tests against the background of the CpG sites included in our RRBS data set (not compared to the whole genome).

To test predictions (3) and (4), which rely on gene-level information, we first assigned each CpG site to a particular gene if it occurred in the gene body or within 10 kb of the gene transcription start site (TSS) or end site. To test prediction (3), we then categorized all genes as either not expressed in whole blood or blood-expressed (based on whether they were included in a whole-blood RNA-seq data set also from the

Amboseli baboons (Tung et al. 2015)). We used Fisher’s exact test to ask whether CpG sites assigned to blood-expressed genes were more likely to be differentially methylated by resource base, compared to CpG sites assigned to unexpressed genes. Finally, to test prediction (4), we performed categorical enrichment analysis using publicly available gene annotations (Kyoto Encyclopedia of Genes and Genomes (KEGG), (Ogata et al.

1999)) and the GeneTrail analysis software (Backes et al. 2007). Here, we focused only on

61

genes associated with differentially methylated sites (as defined above) and tested for over-representation of genes that fall within specific pathways, compared to chance expecta- tions. To do so, we used hypergeometric tests followed by FDR correction

(Benjamini&Hochberg 1995).

3.2.5 Testing for differences in DNA methylation levels at metabolic pathways

Given the known differences in diet, activity patterns, and physiology between

Lodge and wild-feeding animals, we were particularly interested in whether metabolism-related genes showed an epigenetic signature of resource base. To specifically address this question, we focused on CpG sites near (in the gene body or within 10 kb of the gene TSS or TES) genes involved in 36 KEGG pathways related to the metabolism of food or to energy balance (Ogata et al. 1999). These pathways were chosen a priori because of their relevance to the phenotypic differences between wild-feeding and Lodge animals. We used the R package ‘GlobalTest’ (Goeman et al. 2004) to ask whether linearly transformed methylation levels from CpG sites near genes involved in metabolism-related pathways displayed a signature of resource base. This approach asks whether samples with similar DNA methylation patterns (at predefined sets of sites) also have similar resource base labels (i.e., Lodge or wild-feeding), using a framework similar to penalized logistic regression. Thus, the level of analysis is shifted from individual CpG sites to sets of CpG sites associated with putatively similar functions, allowing us to specifically test pathway-based predictions (Goeman et al. 2004). We 62

corrected all GlobalTest p-values for multiple hypothesis testing using the R package qvalue (Dabney & Storey 2015).

3.2.6 Identification of differentially methylated regions (DMRs)

Spatially contiguous stretches of differentially methylated sites (often termed

‘differentially methylated regions’, or DMRs) are more likely to have functional effects on gene expression than differentially methylated sites that occur in isolation (Lister et al. 2009; Hansen et al. 2011; Jaffe et al. 2012). To identify DMRs in our data set, we focused on resource base-associated sites (detected at a 10% FDR) that had at least one other measured CpG site within a 2 kb window centered on the focal site (following the precedent for window size used in Lister et al. 2009 and Hansen et al. 2012). For sites that met this criterion, we counted the absolute number of nearby sites that also exhibited evidence for differential methylation, at a less conservative 20% FDR threshold. We defined DMRs as a cluster of at least 3 resource base-associated sites. We chose this cutoff because clusters of this size were extremely unlikely to occur by chance in permuted data. Specifically, despite relaxing the FDR threshold for identifying CpG sites close to the original resource base-associated sites, our criteria for identifying DMRs results in a relatively stringent FDR threshold of 6.5% FDR. Finally, we collapsed any

DMRs with overlapping boundaries into a single, longer DMR.

3.2.7 Testing the effects of PFKP promoter methylation on gene expression levels

Our analyses revealed one particularly large DMR at the promoter region of the 63

phosphofructokinase gene (PFKP). This DMR stretched across 192 CpG sites, including

30 sites associated with resource base at a 10% FDR. Because PFKP is involved in the rate-limiting step of glycolysis and has been previously implicated in obesity-related traits (Ehrich et al. 2005; Scuteri et al. 2007), we were interested in understanding whether

PFKP promoter methylation alone was sufficient to drive differences in gene expression.

This relationship is implicitly assumed by arguments linking environmental variation to phenotypic variation via epigenetic mechanisms, but is rarely tested in practice.

To test this hypothesis, we used an experimental reporter gene assay in which we cloned 817 bp of the PFKP promoter (containing 72 CpG sites) into a CpG-free vector backbone that contains the luciferase reporter gene (pCpGL, Klug & Rehli 2006). After growing up the PFKP-pCpGL construct in competent E. coli GT115 cells (InvivoGen), we subjected the purified plasmid to one of three treatments: (i) methylation of all 72 CpGs in the PFKP promoter region via treatment with M.SssI (a methyltransferase that targets all CG sequence motifs, resulting in a completely methylated PFKP promoter); (ii) methylation of 13 CpGs in the PFKP promoter via treatment with HhaI (a methyltransferase that targets only CGCG sequence motifs, resulting in a partially methylated PFKP promoter); and (iii) a mock treatment (water substituted for the methyltransferase enzyme, resulting in a completely unmethylated PFKP promoter).

We transfected four replicates of each treatment condition into the human K562 myeloid cell line and incubated the transfected cells for 24 hours (n=12 total transfection

64

experiments). To control for transfection efficiency, a vector containing Renilla luciferase was transfected in parallel. Post-incubation, cells were assayed for luciferase activity using a dual-luciferase reporter assay kit (Promega), and luciferase expression was normalized using measures of co-transfected Renilla activity. Finally, we tested for an effect of DNA methylation at the PFKP promoter on luciferase gene expression using pairwise Wilcoxon rank sum tests.

3.2.8 Investigating the stability or plasticity of DNA methylation levels for individuals that switched between resource bases

Finally, we tested two alternative hypotheses about DNA methylation patterns in the 8 males that switched resource base as a consequence of natal dispersal. First, we hypothesized that if resource base exerts stable, long-term effects on DNA methylation patterns, switching individuals should resemble their natal group members (Lodge or wild-feeding) rather than the group to which they belonged at the time of sampling.

Alternatively, if resource base-associated DNA methylation patterns are largely plastic in response to prevailing conditions, we hypothesized that switching individuals should exhibit DNA methylation patterns that reflect their resource base at the time of sampling rather than their natal group.

To differentiate between these two possibilities, we built a support vector machine (SVM) classifier, a machine-learning approach used for class prediction from high dimensional data (Cortes & Vapnik 1995). This classifier used DNA methylation data to distinguish between individuals that spent all (or the vast majority) of their lives 65

in either a wild-feeding group or in Lodge group (n = 61 individuals). As predictive features for this model, we included the 334,840 CpG sites that were not associated with age, sex, bisulfite conversion rate, or sample age at a nominal p-value of 0.05. We chose this global approach (rather than using significantly differentially methylated sites only) because it allowed us to include sites that may be truly affected by resource base, but did not pass the genome-wide significance threshold in the site-by-site analysis.

Additionally, using all sites ensured that the model classification accuracy was not biased by using features that had already been associated with the response variable in a previous analysis of the same data set (doing so can result in erroneously high classification accuracy even from completely random data: Hastie et al. 2009). Because

SVMs cannot work on binomially distributed count data, we linearly transformed our data before building the SVM.

Finally, we used the resulting SVM to ask whether individuals that switched resource base more closely resembled their pre-switch or post-switch conspecifics. To do so, we used the fitted model to predict the resource base of the 8 individuals that dispersed between groups (using DNA methylation data from these 8 individuals, for the same 334,840 CpG sites).

66

3.3 Results

3.3.1 Genome-wide DNA methylation levels contain a signature of resource base

We found that DNA methylation patterns in our full data set (i.e., after quality control, but before filtering for constitutively hypermethylated, hypomethylated, or invariant sites) recapitulated typical patterns observed in mammalian genomes.

Specifically, most of the genome was hypermethylated, with the exception of H3K4me1- marked enhancers, promoters, and CpG islands. Further, DNA methylation levels near the transcription start sites (TSS) of expressed genes were inversely related to their expression levels.

After filtering, we investigated DNA methylation levels at over half a million

CpG sites in the baboon genome (n = 535,996 sites). As expected when using RRBS (Gu et al. 2011; Boyle et al. 2012), many of these sites occurred in CpG-rich regions, particularly gene bodies (224,553 sites), promoters (25,730 sites), CpG islands (57,461 sites), and CpG island shores (117,226 sites). Further, this dataset encompasses many putatively functional regions of the genome, as at least one CpG was measured in 66% of genes,

28% of promoters, 40% of CpG islands, 44% of CpG island shores, and 11% of enhancers.

Within our filtered set, we identified 1,014 sites (at a 10% FDR) that were differentially methylated between lifelong wild-feeding and Lodge group animals (Figure 6). We did not detect significant effects of sex on DNA methylation levels, or significant sex by resource base interaction effects, consistent with previous studies that have identified

67 1.0

weak or no sex effects in human blood (Eckhardt et al. 2006; Lam et al. 2012). Also in line

with previous studies (Tobi et al. 2014), we did not observe strong directional bias for the 0.8

1,014 differentially methylated sites. Importantly, our analyses of cell type composition

and cell type-specific DNA methylation data indicate that our results are unlikely to be

explained by cell type heterogeneity effects. 0.6 y 0.4 0.2

Real data (10% FDR) -log10 pvalue, resource base effect Real data (NS)

0.0 Permuted data

0.0 0.2 0.4 0.6 0.8 1.0 -log 10 pvalue, uniform distribution x

Figure 6: Resource base influenced genome-wide DNA methylation levels. QQ-plot comparing the cumulative distribution of p-values from a uniform distribution against the cumulative distribution of (i) p-values generated from our main model, which tests for effects of resource base on site-specific DNA methylation levels (plotted in light and dark blue); and (ii) p-values generated from the same model when resource base values (Lodge or wild-feeding) were permuted. The observed

68

deviation from the x=y line suggests a pervasive effect of resource base on DNA methylation levels in the real data. Note that deviation from the x=y line at higher p- values (beginning around p=0.1) suggests that more sites are affected by resource base than we could detect in our sample size of 61 individuals (thus, our non-significant set of sites is a mix of true positives and true negatives). In contrast, when resource base values were permuted across individuals, the resulting p-values were roughly uniformly distributed, as expected, and fell along the x=y line.

3.3.2 Sites associated with resource base are enriched in functionally important regions of the genome

CpG sites associated with differences in resource base were highly nonrandomly distributed in the genome. Specifically, they were enriched in putative enhancers (p =

9.70 x 10-3) and gene promoters (p = 3.66 x 10-3), and underrepresented in functionally unannotated regions of the genome (p = 8.96 x 10-10; Figure 7). Further, they were more likely to occur near genes expressed in whole blood than near unexpressed genes (odds ratio = 1.51, p = 5.49 x 10-7). More fine-grained analyses of chromatin states also indicated a role for differentially methylated sites in the active regulation of genes: differentially methylated sites were more likely to occur in chromatin states associated with active gene transcription in blood cells, including chromatin states designated as ‘active TSS’ (p

= 2.78 x 10-4), ‘flanking active TSS’ (p = 1.44 x 10-2), ‘strong transcription’ (p = 5.24 x 10-3), and ‘enhancer’ (p = 1.28 x 10-3). In contrast, differentially methylated sites were strongly under-enriched in a chromatin state indicative of gene repression (‘repressed polycomb’, p = 5.49 x 10-3).

69

A Active B C Near blood- E_TssA E_ expressed genes TSS Strong D_Tx H3K4me1 transcription D_Promoter enhancers EnhancersC_Enh

C_Enhancers_H3K4me1Promoters Flanking B_TssAFlnk active TSS Uncategorized A_Unannotated Repressed regions A_ReprPC Polycomb

−20 0 20 0 50 100 Percent Percent enrichment Percent Percent enrichmentenrichment H3K4me3H3K4me1 H3K9me3 (over chance expectations) (over chance expectations) H3K36me3 H3K27me3

Figure 7: Sites affected by resource base were enriched in functionally important regions of the genome. (A) Sites associated with resource base were more likely to occur within 10 kb of genes expressed in whole blood, in regions homologous to H3K4me1-marked enhancers in humans, and in gene promoters. They were significantly under-represented in regions of the genome with no known functional role. (B) Sites associated with resource base were also non-randomly distributed across chromatin states. “Active TSS” and “Flanking Active TSS” reflect the transcription start sites of actively expressed genes; “Strong transcription” reflects gene bodies of highly expressed genes; “Enhancers” reflects regulatory elements that interact with promoters of expressed genes; and “Repressed Polycomb” reflects the promoter and gene bodies of silenced/unexpressed genes. In both (A) and (B), only regions/states with significant over/underenrichment are shown (C) Histone marks associated with each of the chromatin states presented in panel B (based on Roadmap Epigenomics data (Roadmap Epigenomics Consortium et al. 2015)). Each chromatin state is defined by the presence (blue square) or absence (white square) of the histone modifications shown below.

3.3.3 Resource base-associated CpG sites are enriched in specific biological pathways

We observed two pieces of evidence that resource base-associated CpG sites were concentrated in specific biological pathways. First, while no KEGG pathways were enriched at a 10% FDR threshold, a more relaxed 20% FDR threshold revealed that resource base-associated sites were enriched near genes involved in 5 KEGG pathways:

70

the T cell receptor and B cell receptor signaling pathways, axon guidance, phosphatidylinositol signaling, and insulin signaling. Second, our GlobalTest analyses revealed patterns of differential methylation associated with carbohydrate metabolism

(galactose: p = 1.29 x 10-5; fructose and mannose: p = 1.79 x 10-4; and the glycolysis pathway: p = 2.94 x 10-4); amino acid metabolism (glycine, serine, and threonine: p = 4.21 x 10-3; tryptophan: p = 0.018); insulin signaling (p = 0.012); and the breakdown of other dietary components (propanoate, p = 0.018; all GlobalTest results reported a 10% FDR threshold). For pathways that included the PFKP gene (specifically, the three carbohydrate metabolism pathways listed above), the observed effect of resource base on DNA methylation levels appears to have been driven almost entirely by differential methylation at PFKP.

3.3.4 DMRs occur more often than expected by chance, and near a key metabolic gene

We identified 87 2-kb windows that met our criteria for differentially methylated regions, compared to only 6 such windows observed on average in permuted data

(equivalent to a 6.5% FDR). These 87 windows collapsed into 29 distinct, longer DMRs, the largest of which fell within the promoter region of an insulin sensitive gene that encodes the rate-limiting enzyme in glycolysis (Lo et al. 2013; Webb et al. 2015). For 90% of the 192 sites we tested in this region, PFKP was more highly methylated in wild- feeding individuals than in Lodge group baboons (Figure 8; no site was significantly

71

more highly methylated in Lodge group animals), suggesting that PFKP expression may be down-regulated in less resource-rich environments.

A 1.5 Higher in wild-feeding p < 0.05 1.5 individuals FDR < 0.10 ● 1.0

● ● 1.0 ● ● ● ● ● ●● ● ● ●●●

0.5 ● ● ● ● ● ● ● ● ● ● ●●●●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●● ●●● ● ● ● ●● ● ● ● ●● ● 0.5 0.0 Beta (effect size) Beta (effect

Beta (effect size) Beta(effect * Higher in Lodge group 0.5 0.0 − individuals Beta Beta (effect size) 1.0 −

-0.5 −1500 −1000 −500 0 500 1000 1500 DistanceDistance from from the the transcription transcription start start site site (bp)

-1.0 B -1500 -1000 -500 0 ** 500 1000 1500 Distance from the transcription start site ● * 0.4 ●

●

●

● 0.2 ●

● ● ● Gene expression level level Gene expression

● (Ratio firefly/renilla luciferase) Geneexpression level

(Ratiofirefly/renilla luciferase) ● 0.0 ● No CpGs Some CpGs All CpGs methylated methylated methylated

Figure 8: Wild-feeding baboons exhibited consistently higher levels of DNA methylation at the phosphofructokinase (PFKP) promoter, where methylation suppresses gene expression in reporter gene assays. (A) The magnitude and direction of the effect of resource base on DNA methylation levels are plotted for all sites tested within 1.5 kb of the PFKP transcription start site. Sites with evidence for an effect of resource base at a nominal p-value of 0.05 are shown as green lines with open 72

green dots, and at a 10% FDR as green lines with filled red dots. All other sites are shown as gray lines. The first exon is denoted by a blue box, and the translation start site is denoted with a black asterisk. (B) Firefly luciferase expression levels (normalized by renilla luciferase expression levels) are plotted for 4 replicates per condition. Results from Wilcoxon signed rank tests are shown as follows: **p=0.014 for comparison between no CpGs methylated and all CpGs methylated; *p=0.057 comparison between no CpGs methylated and some CpGs methylated.

In support of this possibility, our reporter assay experiments confirmed that complete methylation of all CpGs in the PFKP promoter region (n = 72 CpGs in the region we tested) suppressed luciferase expression levels relative to fully unmethylated

PFKP promoter constructs (Wilcoxon rank sum test, W = 16, p = 0.014). Furthermore, even methylation of a minority of CpGs in the PFKP promoter (n = 13 CpGs) produced a graded reduction in gene expression levels, intermediate between the fully methylated and fully unmethylated versions (Wilcoxon rank sum test, W = 14, p = 0.057 for comparison between fully methylated and partially methylated constructs; Figure 8B).

3.3.5 Individuals that switched between resource bases more closely resembled wild-feeding individuals, regardless of the direction of the switch

Finally, we built a support vector machine (SVM) classifier (trained on 334,840

CpG sites) that discriminated between Lodge and wild-feeding individuals with 82% accuracy. We used this model to classify switching individuals based on their DNA methylation levels to test whether switching individuals would be (i) grouped with their natal conspecifics, suggesting that early life (i.e., pre-dispersal) resource base drives variation in DNA methylation levels; or (ii) grouped with their conspecifics at the time

73

of sampling, suggesting that prevailing conditions are more important than past history

(Figure 9).

A1.0 C1.0 H1: Epigenetic stability Predictions match natal group Results from real data 0.0 0.0 0.2

1.0 1.0 Predicted

SVM prediction SVM prediction Lodge group ● -1.0 -1.0 0.2 ●

0.0 0.0 ● ● Predicted ● 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 ● ● ● ● ● ●

0.0 ● Lodge group ● ● ● ● SVM predictionSVM

SVM prediction SVM prediction ●

-1.0 ● -1.0 ● ● ● ● ● Predicted ● ●

1 2 3 0.2 4 5 6 7 8 1 2 ● 3 4 5● 6 7 8 wild-feeding ●

● 0.2 ●

0.0 ● B Predicted − ● ● ● ● H2: Epigenetic plasticity ● ● Lodge group ● Predictions match current group SVM prediction ● ● predictionSVM ● ● ● ● predictionSVM ● Predicted1.0 ● 1.0 ●

0.4 ●

− ● ● wild feeding ● ● ●

0.0 ● ● ● 0.0 0.0 ● 0.2 ● ● − 1.0 1.0 ● ● SVM predictionSVM SVM prediction SVM prediction 0.6 − -1.0 ● -1.0 ● SVM prediction Predicted ● ● 0.0 0.0 SVM predictionSVM wild feeding 1 2 3 4 5 6 7 8 1 2 Individual3 4 switched5 6 from7 Lodge8 to wild-feeding1 2 3Individual4 5 switched6 7 from8 wild-feeding to Lodge ● 0.2

SVM prediction SVM prediction − -1.0 -1.0 0.4 − ● Figure1 92SVM prediction : Individuals3 4 5 6 that7 8switched resource1 2 base3 4 more5 6closely7 ●8 resembled lifelong wild-feedingpredictionSVM individuals, regardless of the direction of the switch. (A-B) Predicted ● and (C) observed results when an SVM classifier that distinguished between wild- 0.4

feeding and− Lodge individuals was applied to DNA methylation data from 0.6 − individuals that switched resource base at natal dispersal (for all plots, grey backgro1 und = 2predicted3 Lodge4 group,5 white background6 7 = predicted8 wild-feeding). We hypothesized that DNA methylation patterns in switching individuals would

consistently (A)0.6 resemble their early life group mates or (B) resemble their current − group mates. However, we observed (C) that regardless of the resource base history of switching individuals, they1 consistently2 3 resembled4 lifelong5 wild6 -feeding7 individuals8 (represented below as an SVM prediction value below 0). Results in C are shown as boxplots (distributions of predicted values for each individual) because we randomly subsampled our data (50 separate subsamples) to create balanced training sets before predicting the resource-base of switching individuals.

We did not find evidence in support of either of these hypotheses. Instead, we

found that the DNA methylation patterns of most switching males (7 of 8) were

74

consistently predicted to be more like wild-feeding animals. This pattern held regardless of whether we considered males that immigrated from a wild-feeding group to Lodge group, or males that immigrated from Lodge group to wild feeding groups. Only one male—an individual who moved from a wild feeding group into Lodge group—violated this pattern, primarily because our model could not consistently classify him with either wild-feeding or Lodge feeding individuals (Figure 9).

3.4 Discussion

Ecological variation, experienced throughout life, can have lasting and dramatic effects on trait variation. However, the molecular mechanisms that mediate these effects remain largely unexplored, especially in natural populations. Here, we present the first evidence that resource base — an environmental variable with known effects on activity patterns, growth rates, insulin levels, and body fat percentages in our study population

— influences DNA methylation levels in a wild mammal (Muruthi et al. 1991; Altmann et al. 1993; Kemnitz et al. 2002; Altmann & Alberts 2005). Specifically, we identified over a thousand differentially methylated CpG sites, as well as 29 differentially methylated regions, that differed between wild-feeding and Lodge group baboons. Our results support the importance of DNA methylation in translating signals from the environment into changes in gene regulation within cells.

3.4.1 The functional relevance of differential methylation at resource base-associated sites

Several pieces of evidence suggest that the changes in DNA methylation we 75

observed are targeted, coordinated, and likely to exert downstream effects on gene regulation. Specifically, differentially methylated sites were more likely to occur in promoter and enhancer regions; near genes expressed in blood, the tissue we sampled; and at stretches of DNA marked by transcriptionally active chromatin states. Further, differential methylation consistently occurred near genes involved in metabolism and insulin signaling, one of the known differences between Lodge and wild-feeding baboons in Amboseli (Kemnitz et al. 2002).

Because current methods for functional validation (e.g., reporter assay experiments) are not feasible on a genome-wide scale, we focused on validating the functional role of changes in DNA methylation at the largest identified DMR (in the promoter region of PFKP). However, several of the additional DMRs we identified fall near genes with relevance to metabolism and energy balance, and may also contribute to organism-level differences. For example, we identified a DMR in KCNIP4, where genetic variants have been previously associated with obesity-related traits (Comuzzie et al.

2012). In addition, we identified a DMR in the 5’ UTR of TPM1, where genetic variation has been associated with platelet count and volume (Soranzo et al. 2009; Gieger et al.

2011), both of which are biomarkers of obesity and metabolic syndrome (Coban et al.

2005; Jesri et al. 2005; Tavil et al. 2007). Together, these results points toward a model in which easy access to resources alters metabolic processes in Lodge group animals, at least in part through targeted changes in DNA methylation. Further work is needed to

76

assess the causal effects of changes in DNA methylation on gene expression at these loci, and more generally to improve methods for high-throughput, genome-wide functional validation.

By far, the most striking DMR we identified fell near a gene coding for an isoform of phosphofructokinase, which catalyzes the irreversible, committed step of glycolysis. Our reporter assay experiments indicate that lower levels of PFKP promoter methylation are sufficient to drive higher levels of PFKP expression (Figure 8). In combination with work in laboratory mice, these results suggest a possible avenue through which regulatory changes at PFKP may contribute to organism-level changes – namely, body fat mass accumulation – in Lodge animals. Specifically, several studies have demonstrated that mice who become obese on high carbohydrate diets exhibit increased levels of the phosphofructokinase enzyme relative to mice that did not become obese, but ingested similar numbers of calories (Yamini et al. 1992; Dourmashkin et al.

2005). Further, mice with experimentally reduced expression levels of the PFKM gene

(the muscle isoform of phosphofructokinase) exhibit greatly reduced levels of both lipogenesis (the process by which energy is stored as fat) and overall body fat (Getty-

Kaushik et al. 2010). Together, these studies argue that phosphofructokinase activity is stimulated by a high carbohydrate diet and consequently favors increased fat accumulation. Previous work in Amboseli has only tested for differences in protein content and overall energy intake between Lodge and wild-feeding animals (revealing

77

that protein intake is higher in wild-feeding animals, while energy intake does not differ: Muruthi et al. 1991). However, Lodge animals were likely to have ingested higher levels of simple carbohydrates as well, consistent with the model proposed above.

3.4.2 Stability and plasticity in the epigenetic signature of resource base

DNA methylation marks are largely established during development and subsequently carried across mitotic cell divisions. Consequently, most studies of diet or resource access effects on DNA methylation have focused on exposures during development (Wolff et al. 1998; Sinclair & Allegrucci 2007; Khulan et al. 2012). These studies have observed strong effects of maternal resource constraint during pregnancy on offspring methylation levels (Heijmans et al. 2008; Waterland et al. 2010; Tobi et al.

2014). Furthermore, they have emphasized the precise timing of maternal resource effects, which are sometimes limited to specific trimesters (Tobi et al. 2009, 2014).

The results of such studies have been widely interpreted as support for a ‘critical period’ model of environmental epigenetic effects, where environmental insults during development are primarily responsible for downstream effects on gene regulation

(Meaney & Szyf 2005; Faulk & Dolinoy 2011; Heim & Binder 2012). However, recent work has shown that DNA methylation levels are also affected by environmental conditions later in life, including adult socioeconomic status in humans (McGuinness et al. 2012; but see Lam et al. 2012), experimentally manipulated social status in captive rhesus macaques (Tung et al. 2012), and immune response to infection in humans (Marr 78

et al. 2014; McErlean et al. 2014; Pacis et al. 2015). These observations of epigenetic plasticity also extend to recent studies of energy balance and diet in humans: individuals that subsisted on a high fat diet for one week exhibited epigenetic changes at thousands of CpG sites compared to randomized controls (Jacobsen et al. 2012). Similarly, short- term exercise interventions induce widespread changes in DNA methylation levels

(Barrès et al. 2012; Rönn et al. 2013).

The current literature thus indicates that DNA methylation plays two complementary roles. In some cases, it encodes a signature of early life experience, producing stable effects on gene regulation that persist over time. In other cases, it continues to be dynamically regulated, allowing organisms to adjust their phenotypes to prevailing environmental conditions. Whereas the effects of resource availability have been largely studied in the context of the first role (Klebig et al. 1995; Wolff et al. 1998;

Heijmans et al. 2008; Carone et al. 2010; Tobi et al. 2014), we were able to take advantage of naturally occurring male dispersal to investigate both roles (Figure 5).

Surprisingly, our analyses revealed a lack of support for either long-term stability or global plasticity. Instead, we found that the DNA methylation patterns of switching individuals, whether originating from or immigrating into wild-feeding groups, almost universally resembled those of lifelong wild-feeding individuals (Figure

9). Wild-feeding individuals rely on widely distributed, seasonally available foods, and consequently experience greater seasonal and year-to-year variance in resource access

79

compared to Lodge animals. Our results thus indicate that more challenging environments, in terms of energy balance, may leave a stronger epigenetic signature than more favorable environments. This appears to hold whether exposure occurs earlier in life—consistent with long-term early life effects and developmental constraint models—or later in life (e.g., in post-dispersal males), consistent with a more sustained capacity for plasticity. Our results thus dovetail with recent work on both the evolutionary significance of early life effects (Lindström 1999; Gluckman et al. 2005a;

Botero et al. 2015), including in the Amboseli baboons (Lea et al. 2015a), as well as the possibility that epigenetic marks mechanistically mediate these effects (Weaver et al.

2004; Lam et al. 2012; Tobi et al. 2014). However, theoretical work is needed to connect the evolution of plasticity to expectations about the epigenetic patterns associated with different levels of adaptive plasticity (Furrow & Feldman 2014).

Finally, our findings indicate an important caveat for studies in ecological epigenetics: that inferences about stability versus plasticity may be contingent on the direction in which an environment changes. If we had only analyzed male baboons that transitioned from resource abundance to resource limitation (i.e., from the Lodge to wild-feeding resource base), our results would have supported complete plasticity. On the other hand, if we had focused only on males that transitioned from resource limitation to resource abundance, our data would have supported long-term stability.

Notably, the latter type of transition is the one that has been highlighted in studies of

80

human health, which have tended to emphasize the special importance of early life exposure (Gluckman & Hanson 2004; Schulz 2010). In contrast, our findings suggest that the specific experience of resource limitation may leave an epigenetic signature that transcends any critical window of exposure, at least in our system. Future research will be needed to assess the generality of these results. Nevertheless, we believe our results as a whole emphasize the importance—already acknowledged in evolutionary and behavioral ecology more generally—of taking a life course approach to ecological epigenetics. Organisms are a product of both their current environment and their past history, and we should expect the epigenetic patterns within their cells to reflect this combination. The more interesting outstanding questions are about the negotiation of this balance, including how it evolves.

81

4. Genome-wide quantification and prediction of DNA methylation-dependent regulatory activity

4.1 Introduction

DNA methylation—the covalent addition of methyl groups to cytosine bases—is a gene regulatory mechanism that plays a fundamental role in development (Smith &

Meissner 2013), disease susceptibility (El-Maarri 2005; Heyn & Esteller 2012), and the response to environmental conditions (Jirtle & Skinner 2007; Feil & Fraga 2011).

Genome-wide studies have revealed that variation in DNA methylation levels is associated with aging (Winnefeld & Lyko 2012; Day et al. 2013; Marioni et al. 2015), both short-term (Barrès et al. 2012; Pacis et al. 2015) and long-term (Weaver et al. 2004;

Dominguez-Salas et al. 2014; Tobi et al. 2014) reactions to environmental perturbations, and disease states, including cancer (Hansen et al. 2011; Hinoue et al. 2012; Aran et al.

2013), type-2 diabetes (Dayeh et al. 2014), and Alzheimer’s Disease (Bakulskia et al. 2012;

De Jager et al. 2014) . However, the vast majority of studies, especially at the population level, remain correlational. This is a major limitation, as experimental work has shown that not all changes in DNA methylation affect downstream gene regulation (Maeder et al. 2013; Andersson et al. 2014). Where this link does not exist, methylation variation is unlikely to impact organism-level trait variation. Identifying regions of the genome where DNA methylation might ‘matter’ for phenotypic variation requires the ability to systematically predict the causal relationship between locus-specific differential DNA methylation and gene expression levels on a genome-wide scale. 82

Currently available approaches cannot fill this gap, as they are either not locus- specific (e.g., 5-azacytidine or 5-aza-2'-deoxycytidine treatment to reduce global DNA methylation levels (Christman 2002)), too low throughput (e.g., editing of the endogenous epigenome (Rivenbark et al. 2012; Maeder et al. 2013; Liu et al. 2016) or manipulation of methyl marks in low-throughput episomal reporter assays (Klug &

Rehli 2006b)), or are performed outside the cellular context (e.g., in vitro assays of methylation-dependent transcription factor (TF) binding (Hu et al. 2013; Mann et al. 2013;

O’Malley et al. 2016)). These methods do, however, support a general interdependence between CpG methylation and gene regulation. DNA methylation levels have been shown to influence transcription initiation at select CpG-dense gene promoters (Klug &

Rehli 2006a; Weber et al. 2007; Jones 2012; Maeder et al. 2013; Rönn et al. 2013; Tobi et al.

2014); drive cell type-specific gene expression, especially in connection to enhancers that are methylated in a cell type-specific manner (Wiench et al. 2011; Andersson et al. 2014); and affect transcription factor binding (Hu et al. 2013; O’Malley et al. 2016). For example, a recent comparison of in vitro TF binding to methylated versus unmethylated DNA fragments estimated that 72% of Arabidopsis transcription factors are inhibited by DNA methylation (while an additional 4.3% of TFs preferentially bind methylated DNA)

(O’Malley et al. 2016). In vertebrates, recent studies have also found that that DNA methylation generally inhibits TF binding (Hu et al. 2013; Zhu et al. 2016), but in some cases can facilitate it (Nan et al. 1998; Perini et al. 2005; Rishi et al. 2010; Zhu et al. 2016).

83

For example, in vitro binding experiments have shown that many pioneer TFs such as forkhead box protein A (FOXA), GATA3, and GATA4 preferentially bind methylated

DNA (Hu et al. 2013). Pioneer TFs are responsible for binding and opening regions of heterochromatic DNA (which tend to be highly methylated (Roadmap Epigenomics

Consortium et al. 2015)), and it has thus been hypothesized that a preference for methylated DNA is a general property of pioneer TFs that facilitates their function (Zhu et al. 2016). However, due to a lack of genome-scale methods for causally querying the consequences of differential methylation in vivo, tests of this hypothesis remain limited.

Together, the above findings indicate extensive heterogeneity in the relationship between variation in DNA methylation and variation in gene regulation that we have only begun to understand (Zhu et al. 2016). Here, we report a new method, ‘mSTARR- seq,’ that can be used to investigate this heterogeneity by assaying the causal relationship between DNA methylation and regulatory activity in high-throughput, within a cellular context. mSTARR-seq combines genome-scale strategies for quantifying enhancer activity via self transcribing episomal reporter assays (i.e., STARR-seq (Arnold et al. 2013, 2014; Vanhille et al. 2015; Vockley et al. 2015, 2016)) with enzymatic manipulation of DNA methylation at millions of unique CpG sites. We couple mSTARR- seq to a new reporter plasmid (pmSTARRseq1) that is devoid of CpG sites and targets of bacterial methylation in the backbone, and demonstrate its potential by testing DNA methylation-dependent regulatory activity across 724,391 unique human DNA

84

fragments (covering 1.83 million CpG sites), in a single experiment. Our results enabled us to (i) build a predictive model for when CpG methylation can and cannot alter gene expression levels, (ii) identify known and novel transcription factors involved in DNA methylation-mediated gene repression or activation, and (iii) explain heterogeneity in the correlations between DNA methylation and gene expression levels in vivo. Together, they demonstrate how mSTARR-seq can be used to understand when, where, and under what conditions differential methylation is most likely to ‘matter’ for gene regulation, as well as the molecular mechanisms involved in methylation-dependent regulatory activity.

4.2 Results

4.2.1 mSTARR-seq identifies regions with known regulatory activity

To create the input library for our mSTARR-seq experiments, we combined MspI- digested genomic DNA, which strongly enriches for CpG-containing fragments, with randomly sheared DNA from the HapMap GM12878 cell line (in a mixture of 33% MspI- digested genomic DNA and 67% sheared DNA). We cloned the resulting fragment pool into pmSTARRseq1, a CpG-free reporter vector we engineered specifically for mSTARR- seq (Figure 10). Each fragment was inserted between a constitutive promoter and a poly- adenylation signal, such that regions with regulatory potential drive their own transcription in vivo (Arnold et al. 2013). We exposed the resulting plasmid library to either (i) the bacterial methyltransferase M.SssI in the presence of a methyl donor, which

85

methylates all CpG sites, or (ii) a sham treatment in which M.SssI was replaced with water, to maintain fully unmethylated constructs. We then transfected the unmethylated and methylated versions of the plasmid library into the ENCODE Tier 1 K562 myeloid cell line (n=6 unmethylated and 6 methylated replicates). Following a 48-hour incubation period, we extracted total RNA from each replicate and isolated and sequenced the plasmid-derived mRNA. To control for variation in the plasmid input libraries, we also sequenced the fragment inserts from each plasmid DNA pool (n=6 from each condition).

promoter ORF A poly-A signal intron B putative enhancer Histone_marks H3K4me1 unmethylated CpG H3K4me1_H3K27ac neither methylated CpG 2 pmSTARR seq1

0

-M.SssI +M.SssI

Log2(odds), Fisher's exact test Fisher's Log2(odds), -2

Insulator WeakTxn Repressed Heterochrom/lo TxnElongation Txn_Transition pmSTARR pmSTARR Repetitive/CNV1 Active_PromoterRepetitive/CNV2 Weak_Promoter Poised_PromoterWeak_Enhancer1Weak_Enhancer2 seq1 seq1 Strong_Enhancer1Strong_Enhancer2 ChromHMM_category 0.25 C Histone_marks H3K4me1 0.20 H3K4me1_H3K27ac neither 2

0.15

0.10 0

0.05 Density (arbitrary units) (arbitrary Density

Log2(odds), Fisher's exact test Fisher's Log2(odds), -2 0.00 mRNA-seq of plasmid-derived RNA DNA-seq of plasmid DNA 0.0 2.5 5.0 7.5 Insulator WeakTxn BS-seq of plasmid DNA Enhancer effectRepressed size TxnElongation Txn_Transition Heterochrom/lo Weak_Promoter Repetitive/CNV1 Poised_Promoter Active_Promoter Repetitive/CNV2 Weak_Enhancer1Weak_Enhancer2Strong_Enhancer1Strong_Enhancer2 ChromHMM_category Figure 10: mSTARR-seq identifies regions with known regulatory activity. (A) Protocol overview. Fragmented DNA is cloned into pmSTARRseq1 in high- 86

throughput, experimentally methylated or sham treated, and transfected into a cell line of interest. After a 48hr incubation period, plasmid DNA and mRNA derived from the plasmid are extracted and prepared for sequencing. (B) Regions with significant regulatory activity in the mSTARR-seq assay are enriched for chromatin state annotations defined by active marks (H3K4me1 and H3K27ac, colored orange). The y-axis depicts the log2 odds from a Fisher’s exact test for enrichment (or depletion) of mSTARR-seq identified enhancers in each of 12 ENCODE ChromHMM chromatin states. Positive y-axis values indicate enrichment and negative values indicate depletion (p<0.05 for all tests). (C) Effect sizes (corresponding to the estimated strength of each enhancer element; x-axis) are consistently larger for mSTARR-seq enhancers that occur in chromatin state annotations defined by active marks (H3K4me1 and H3K27ac; colored orange).

In total, we assayed ~3/4 of a million unique fragments in each replicate library

(mean ± SD = 759,725 ± 252,187 fragments per replicate library). This level of diversity is comparable to that observed in previous STARR-seq experiments, and far exceeds the diversity of all published massively parallel reporter assay (MPRA) experiments that rely on oligonucleotide synthesis to generate input fragments (Melnikov et al. 2012;

Arnold et al. 2013, 2014; Vanhille et al. 2015; Tewhey et al. 2016; Vockley et al. 2016). For subsequent analysis, we binned the genome into 200 bp non-overlapping intervals and counted read pileups from all plasmid mRNA and DNA derived libraries in each interval. We then filtered these regions to focus on the 277,896 intervals (covered by

724,391 unique fragments) that overlapped at least 1 mRNA read and 1 DNA read in ≥3 unmethylated condition replicates and ≥3 methylated condition replicates. This filtered data set represented 57% of fragments expected from an in silico MspI digest of the human genome, and 14% of the euchromatic genome in all (based on ChromHMM

ENCODE annotations of the K562 cell line).

87

We first identified regions where the abundance of plasmid-derived mRNA was significantly greater than the amount of input plasmid DNA, a pattern that indicates the capacity for self-transcribing enhancer activity (Arnold et al. 2013) (hereafter referred to as ‘enhancer activity’ or ‘regulatory activity’). We found that 9% of analyzed regions

(n=24,945 at a 10% false discovery rate) had significant regulatory activity either in the unmethylated condition, the methylated condition, or both. Intriguingly, we identified

41 regions that have significant enhancer activity when methylated (FDR<10%) but no statistical evidence for activity when unmethylated (FDR>20%). These regions would be completely missed by traditional STARR-seq or MPRA approaches.

As expected, the set of regions capable of enhancer activity was highly enriched for ENCODE chromatin states associated with H3K4me1 and H3K27ac, which mark active enhancers (enrichment for the two ‘strong enhancer’ chromatin states: Fisher’s exact test, odds=5.53 and 12.10, both p<10-15; enrichment for the ‘active promoter’ state: odds=3.32, p<10-15). Also as expected, mSTARR-seq enhancers were highly depleted in regions that lacked both marks (‘heterochromatin:’ odds=0.33, p<10-15; ‘insulator:’ odds=0.28, p<10-15; ‘repressed:’ odds=0.32, p<10-15; Figure 10). Enhancers that overlapped

H3K4me1 and H3K27ac-marked chromatin states consistently displayed the largest effect sizes (linear model, p<10-15; Figure 10), and those annotated as strong enhancers exhibited the strongest effects overall (p<10-15). In general, power to detect enhancer activity increased with larger query fragment sizes, such that the largest 10% of

88

fragments (>443 bp) were most strongly enriched for ENCODE-annotated enhancers

(odds=9.42, p<10-15). The shortest 10% of fragments (<203 bp, near the typical size of most synthesized MPRA oligos) showed substantially weaker enrichment (odds=4.71, p<10-15), suggesting that shorter fragments result in steric hindrance or eliminate binding sites key to functional enhancer activity.

4.2.2 Methylation dependent activity is predictable based on sequence characteristics alone

For regions identified as enhancers by mSTARR-seq (n=24,945 200 bp regions), we next asked whether the strength of enhancer activity could be altered by experimental addition of DNA methylation marks. To do so, we asked whether the magnitude of the difference between mRNA abundance and input plasmid DNA was significantly different in unmethylated versus experimentally methylated samples. We identified 2,143 enhancers with methylation-dependent (MD) activity (8.59% of those tested; 10% FDR), the majority of which (88%) are more active when unmethylated. In support of mSTARR-seq’s low false positive rate, we identified only 4 CpG-free regions with statistical support for MD activity, out of 941 CpG free enhancers tested. Further, although our statistical test was agnostic to whether methylation affected DNA input counts or counts from mSTARR-seq plasmid-derived RNA, the interactions we observed are clearly driven by differences in the RNA counts (DNA counts are highly consistent

89

between methylated and unmethylated conditions: Figure 11).

Figure 11: mSTARR-seq identifies methylation dependent (MD) enhancers, and MD activity is predicted by CpG density. (A) The mean difference in normalized counts (from limma) between unmethylated condition libraries and methylated condition libraries is plotted on the y-axis for comparisons of mRNA libraries (blue) and input plasmid DNA libraries (yellow). Values for every region identified as an MD enhancer (n=2143) are plotted (rank ordered by the magnitude of the y-axis value, for mRNA libraries). In all cases, the magnitude of the mRNA difference between conditions far exceeds the magnitude of the input plasmid DNA difference between conditions (which is consistently minimal). (B) The relationship between CpG density (number of CpG sites/fragment window length; x-axis) and the degree to which an mSTARR-seq enhancer is suppressed by methylation (measured as the difference in mRNA abundance between conditions; y-axis) is consistently positive (Spearman correlation, rho=0.246, p<10-15). All regions with significant enhancer activity in at least one condition are shown. Colors indicate whether each enhancer was significantly methylation dependent (FDR<10%). (C) Empirical cumulative distribution functions of fragment CpG density for regions identified as MD enhancers (red line) versus enhancers whose activity was unaffected by methylation (non-MD enhancers; blue line).

We next tested whether MD enhancer activity is tied to CpG density. Indeed, MD enhancers have significantly higher CpG densities than non-MD enhancers (Wilcoxon- signed rank test, W=3.51x107, p<10-15; Figure 11). Across all 24,945 regions with enhancer activity, CpG density alone explains 6.05% of the variation in the magnitude of methylation dependence (Spearman correlation, rho=0.246, p<10-15; Figure 11).

90

To investigate other features associated with MD regulatory activity, we used a data set of 89 genomic features and a random forests classifier to compare MD enhancers with greater activity when unmethylated (n=1863) against enhancers with no evidence for methylation dependence (n=5703; FDR>50% in the test for MD activity).

Our feature set included information about the endogenous chromatin state, endogenous chromatin accessibility, endogenous DNA methylation level, evolutionary conservation, and CpG density of each enhancer region, as well as information about putative TF binding sites for a list of 36 TFs. Overall, the random forests model was able to predict whether an enhancer region would be suppressed by methylation with 79% accuracy (Figure 12). In addition to the absolute number and density of CpG sites, 26 features were identified as key predictors in our random forest (based on two measures of variable importance, the mean decrease in accuracy and the Gini coefficient, at a 10%

FDR; Figure 12). Enhancers suppressed by DNA methylation were more likely to occur in regions of open chromatin, in ENCODE-annotated ‘active promoters’ and ‘strong enhancers’, and near genes highly expressed in K562s (relative to non MD enhancers); further, MD enhancers were more likely to contain binding sites for transcription factors such as SP4, NGFI-C, and ATF, all of which have CpG sites in their canonical binding motifs (Figure 12).

91 CpG number BACH2 Yes CpG density JUNDM2 6 Expression level SP1 No Methylation level SP4 5 4 Weak Promoter NGFI-C Insulator CTCF 2 DNase signal ATF Active Promoter CREB 0

Heterochromatin0 TCF3 Weak Txn STF1 -2 Repressed CST6 Weak Enhancer2 EHF -4 RNA Strong Enhancer2 LMO2 -5 (unmethylated-methylated) (unmethylated-methylated) -6 DNA Strong Enhancer1 HLF

counts normalized in difference Mean 0 500 1000 1500 2000 counts normalized in difference Mean 0.00 0 0.052 4 0.106 8 100.1512 14 0 2 4 6 8 10 12 14 Methylation dependent enhancer MeanCpG decrease density in accuracy Mean decrease in accuracy A B C 0.8 CpG number BACH2 BACH2 1.0 BACH2 Yes 0.8 CpG density JUNDM2JUNDM2 JUNDM2 No Expression level SP1 SP1 SP1 0.6 Methylation0.8 level SP4 SP4 SP4 0.6 Weak Promoter NGFI-C NGFI-C RFModel NGFI-C Model Insulator CTCF CTCF CTCF correct?predictionprediction0.6 DNase signal ATF ATF ATF 0.4 0.4 correctYescorrect Active Promoter CREBCREB CREB Heterochromatin TCF3 TCF3 TCF3 Binding motif incorrect 0.4 Noincorrect Weak Txn STF1 STF1 STF1 contains CpG

0.2 0.2 <= X axis value Repressed CST6 CST6 CST6 sites? 0.2 Proportionregionsof Weak Enhancer2 EHF EHF EHF Proportion of regions Proportion Non MD enhancerYes Yes

Proportion of regions with of regions Proportion Strong Enhancer2 LMO2 LMO2 LMO2 Proportion of regions Proportion MD enhancer No No 0.0 0.0 Strong Enhancer1 HLF HLF HLF 0.0 MD enhancernot MD enhancer MD enhancernotMD MDNon enhancer MD true_category 0.00 0.050 20.104 60.158 100.2012 14 0 20 24 046 26 848 10610 12812 1014 12 14 true_category Enhancer type CpGMeanMean density decrease decrease in accuracy MeanMeanMean decrease decrease decreaseMean in decrease accuracy in in accuracy accuracy in accuracy

Figure 12: Random forests accurately predict MD versus non MD activity based on sequence features. (A) The proportion of non MD and MD enhancers that were accurately classified by random forests. (B) Genomic features and (C) predicted transcription factor binding sites identified as important features by random forest analyses (at a 10% FDR). In both panels, the y-axis depicts the mean decrease in predictive accuracy between a fitted random forest including the focal variable and the same random forest when the predictor variable labels have been permuted. Note that variable importance scores are agnostic to whether the value of the predictor variable is higher in one class versus the other. In (C), bars are colored based on whether the position weight matrix (PWM) for the focal transcription factor includes a CpG site (PWMs were downloaded from http://centipede.uchicago.edu/SimpleMulti/).

4.2.3 mSTARR-seq identifies transcription factor families affected by DNA methylation status

The results of our random forests analysis, as well as previous work in in vitro and targeted in vivo systems, point toward inhibition of TF binding as a mechanism through which DNA methylation suppresses gene regulatory activity (Perini et al. 2005;

Wang et al. 2012; Hu et al. 2013; Domcke et al. 2015; Maurano et al. 2015; Stephens & Poon

2016). However, the degree to which specific TFs are consistently involved in methylation dependent regulatory activity remains largely unknown. To explore this question, we performed motif analysis to ask whether enhancers that were significantly affected by DNA methylation were enriched for known vertebrate TF binding motifs 92

compared to the background set of all regions with enhancer activity. Among the 1886

MD enhancers in which DNA methylation suppresses activity, we found 24 significantly enriched binding motifs (1% FDR), 15 of which belong to the ETS family (a 9.94x enrichment over chance; hypergeometric test p=9.35x10-16; Figure 13). Previous studies have reported that the binding of select ETS family TFs is strongly inhibited when CpGs within the binding site are fully methylated (namely for ETS1, ETS2, GABPα, ETV1)

(Yokomori et al. 1995; Umezawa et al. 1997; Lucas et al. 2009; Polansky et al. 2010; Cooper et al. 2015; Stephens & Poon 2016). It has been hypothesized (Stephens & Poon 2016) that this relationship is generally true for ‘Class I’ ETS family TFs (which bind the canonical motif ACCGGAAGT), but not ‘Class III’ ETS family TFs (whose binding motifs do not always include a CpG site (Wei et al. 2010)). In support of this idea, 12 of the 15 ETS TFs we identified belong to Class I, and none belong to Class III (the remaining 3 belong to

Class II, for which methylation-dependent binding is unexplored). Further, using data from a mouse study (Domcke et al. 2015) that mapped TF binding in the presence or absence of DNA methylation (by generating triple knockout stem cells with inactive

Dnmt1, Dnmt3a, and Dnmt3b methyltransferases), we find further support that binding of ETS family TFs is consistently suppressed by DNA methylation (Figure 14).

93

Figure 13: HOMER identifies transcription factors enriched in MD enhancers. TF enrichment for MD enhancers that are (A) more active when unmethylated and (B) more active when methylated. In both panels, the y-axis depicts the log2 odds from a Fisher’s exact test, comparing the frequency of each transcription factor binding site in the data subset of interest against the background set of all fragments with enhancer activity. Bars are colored based on their transcription factor family or subfamily, as defined by HOMER.

94

A B ● ● ● ●●● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ●● ● ●● ●● ● ● 1.5

15 ● ● ●● ● ● ●● ● ● ● ●● ● ● ●●

● ● ● ● ●●●● ● ●● ●●●●● ●● ●● ● ● ● ● ● ● ● ● 1.0 ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ●●● ●● ●● ● ● ● ● ●● ● ● ● ● ● ●●●● ●●● ●●● ●● ● ●●●● 0.5 ●●● ● ● ● ● ● ● ● ●● ● ● ● ●●● ●●● ● ●●●●●●●●● ● ● ● ● ● ● ● ● ● ● 10 ● ● ● ● ●●● ● ●● ● ●● ● ●● ● ● ● ●● ● ●● ●● ● ●● ●● ● ● ●● ● ● ● ●●●● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ●●●● ●●●● ● ●●●● ● ● ●● ● ●● ● ● ● ● ●●●● ● ●●●● ● ● ● ●● ● ●● ●●● ● ●● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ●●●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ● ● ● ●● ●● ● ●● ● ●●● ●●● ● ● ●●●●●●● ●● ●● ●●● ● ● ● ●● ● ●● ●●●●●●●●●● ●● ●● ●●● ● ●● ● ● ● ● ●● ●●● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ●●● ●●● ●● ●●● ●●●● ●● ●●●●● ● ●● ● ● ● ●●●●● ● ● ● ● ● ● ● ●●●● ●●●●● ● ●● ● ●●●● ● ●●●●●● ● ●●●●● ●● ● ● ● ● ●● ●●●●● ●● ●●●●● ●●● ● ●● ● ●●●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●●●● ●● ●●●● ●● ● ●●●●●●●● ●●●●● ●●●●● ●● ● ●●● ● ● ● ● ●● ● ● ● 0.0 ● ● ●●●●●● ●●● ● ● ●●●●●●●●●● ●●●● ● ●●●● ● ● ● ● ● ●● ● ● ● ●●●●●●●● ●●●●●● ●● ●● ●●● ●●● ● ●●● ●● ●●● ● ● ●● ● ●●● ●● ●●● ● ● ●●●●●●●●● ●●●● ●● ● ●●● ●● ●●● ●●●● ● ● ●●●●● ●● ●●●●●● ● ●●●●●●●●●●●●●●●● ●● ●● ●● ●● ● ●● ● ●● ● ●● ●●●●●●●●●●●●●●●● ● ● ● ● ● ● ● ●● ● ●● ●● ●●●●●●●●● ● ● ●●● ● ●● ●● ● ●● ● ● ●●●●●●●● ●●●●●●●●●● ●●● ●● ●●● ●●● ● ●● ●● ● ● ● ● ●● ●●●●●●●● ●●● ● ●●●●●● ● ● ● ●●● ● ●● ● ● ● ● ●● ● ●● ●●●●●●●● ●●● ● ●● ●●● ● ●● ● ● ●● ●●●●●●●●●● ● ● ●● ● ●●●● ● ●● ●● ●●●●● ●●●● ●●●●● ●● ● ● ● ● ● ● ● ●●●●●●●●●●●●●● ●● ● ●● ● ● ●●●●●●●●●●● ●●● ● ● ● ● ●● ●●● ●●●●●● ●●●●● ● ● ● ●● ● ● ●●●●●● ●●●● ●● ● ● ● ●●●● ●● ●●●●●●● ●●● ● ● ●●● ●●●●●●●● ●● ●●●●●●● ● ● ● ● ● ● ●●●●●●● ●●● ● ● ● ●●●●●●● ●●●●●●● 0.5 ●●● ●●●●● ● ● ● ●●●●● ● ●● ● ● ● ●●●●●●●● ●●●● ●●● ●●●●●●●● ●● ●●●●●●●● − ● ● ● ●● ●●● ●● ●● 5 ●● ● ● ● CpGs in MD enhancers value, ● ●●●●● ● ●● ●● ●●●●● ● ●●●●● ● ● ●● ● ●●● ● ●● ●●●●● ●● − ●●●●●●●●●●● ●●●● ● ●● ● ●●●●●●●●●●●●●●● ● ●● ● ● ●●● ● ● ●●●●●●●●●● ●●● ● ● ●●●●● ●● ● ● ●●● ●● ● ●●● ●●● ●●●●●●● ●● ●●●●● ●●● ● ●●● 1.0 ● ●●●●●● ● ●●●●●● ●●●●● − ●●●●● ●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ● ●●●●●●●●●●●●●●●●● GATA family log10 p ●●●●● ● ●●●●● AllAll CpGsCpGs ●●●●●● − ● ●●●●● specific / WT DHSs with motif (log2 enrichment) ● ●●● ●●●●●●● ● ●●●●● ● ETS family ●●●●●● − ● ●●●●● Promoter ●●●●●● Negative ●●●●● 1.5 ●●● ●●●● ● ●●●●●● ● ●●●● ● − ● FDR<0.1 ●●●● Negative in promoters

0 Negative in promoters TKO 0 10 20 30 40 50 0 5 10 15 TKO−specific DHSs with motif (%) −log10 p−value, CpGs in non MD enhancers

Figure 14: mSTARR-seq data explains heterogeneity in previously published in vivo data sets. (A) Data from a comparison of TF binding (identified using DNase-seq data) in wild type (WT) murine stem cells versus triple knockout (TKO) stem cells with inactive Dnmt1, Dnmt3a, and Dnmt3b methyltransferases. The x-axis depicts, for each of 1251 analyzed TF binding motifs, the percent of TKO-specific DNaseI hypersensitivity sites (that do not occur in WT) that contain a binding site for the focal TF. TFs with more positive x-axis values are frequently found in DNaseI hypersensitivity sites that occur in unmethylated genomes but not methylated genomes. The y-axis depicts the relative frequency of each TF motif in the set of DNaseI hypersensitivity sites that are TKO-specific versus the set found in WT (significance is based on a hypergeometric test comparing these ratios). TFs with more positive y-axis values have more prevalent binding in regions that are specifically open in unmethylated genomes (relative to the background frequency observed in WT methylated genomes). Notably, all tested ETS family motifs (red) appear to bind more in unmethylated genomes than methylated genomes, while GATA family TFs (blue) consistently show the opposite pattern. (B) QQ-plot comparing the cumulative distribution of p-values (supporting a correlation between DNA methylation levels and gene expression levels in human monocytes) for methylation dependent (MD) enhancers versus non MD enhancers. Colors indicate subsets of the data that encompass (i) all CpG sites that overlapped mSTARR-seq data (grey); (ii) only CpG sites in promoters; or (iii) CpG sites in promoters that exhibited a negative methylation-expression correlation. We observe the strongest enrichment of low p- values in MD enhancers relative to non MD enhancers when considering CpGs that

95

occurred in putative promoter regions and displayed the expected negative methylation-expression correlation.

Though 88% of mSTARR-seq enhancers are repressed when methylated, 12%

(n=257 regions) are significantly more active when methylated. Among this set, we identified 9 significantly enriched binding motifs compared to the background set of all enhancer regions (1% FDR). These motifs were enriched for TFs from the bHLH family

(5.44x enrichment, p=2.07x10-4) as well as the GATA subfamily of zinc finger TFs (16.11x enrichment, p=2.05x10-6; Figure 13). In agreement with these results, recent work has demonstrated that GATA3, GATA4, and several bHLH family TFs can bind methylated

DNA in vitro (Hu et al. 2013). Further, when binding of GATA family TFs is mapped in the presence or absence of DNA methylation (in the triple knockout framework described above (Domcke et al. 2015)), GATA binding is clearly more common in wild type/methylated genomes compared to unmethylated genomes (Figure 14). Together, our results thus support the hypothesis that the ability to bind methylated CpG sites is a general property of pioneer TFs (Zhu et al. 2016), including GATA family TFs. In particular, the strong affinity of GATA family TFs for methylated DNA may provide an anchor point for these proteins to open closed chromatin (Zhu et al. 2016).

4.2.4 mSTARR-seq explains heterogeneity in DNA methylation level- gene expression level correlations in vivo

DNA methylation is canonically thought to decrease the expression of nearby genes (especially when it occurs in enhancers and promoters). This relationship is

96

supported by the correlation between DNA methylation and gene expression measured within the same sample (rho=-0.295) (Lam et al. 2012). However, across samples, DNA methylation exhibits a highly heterogeneous relationship with gene expression: individuals who exhibit higher levels of DNA methylation sometimes have lower levels of gene expression, but at many sites, the opposite is true or the two measures are orthogonal (Meissner et al. 2008; Lister et al. 2009; Bell et al. 2011; Hansen et al. 2012; Lam et al. 2012). To investigate whether mSTARR-seq data can explain this variation, we drew on a publicly available data set containing paired genome-wide DNA methylation level and gene expression level data for 1202 human monocyte samples (Reynolds et al.

2014). When considering the full set of CpG sites that overlapped with our mSTARR-seq data, we observed that significant (FDR<10%) DNA methylation-gene expression correlations in monocytes were only weakly enriched in mSTARR-seq MD enhancers versus non-MD enhancers (Fisher’s exact test, odds=1.16, p=0.016; Figure 14). However, for CpG sites in promoter regions (i.e., those that fell within the 2 kb region upstream of the TSS), MD enhancer activity strongly predicts more significant DNA methylation- expression correlations in vivo in primary cells (odds=1.52, p<10-15). This relationship was even stronger when considering CpGs that occurred in putative promoter regions and displayed the expected negative correlation between DNA methylation and gene expression levels (odds=2.03, p=3.38x10-4).

97

4.3 Discussion

Here, we present a novel high-throughput assay for performing genome-scale tests of methylation-dependent enhancer activity. Our results indicate that regulatory elements with methylation-dependent enhancer activity are to some degree predictable from sequence alone, although the degree to which DNA methylation levels modulate gene expression within MD enhancers varies considerably. Our results also explain site- by-site variation in DNA methylation-gene expression correlations calculated from primary cells obtained in a completely different setting (a community-based population study of atherosclerosis), demonstrating their direct relevance to an in vivo biological setting. mSTARR-seq thus provides an efficient, cost-effective strategy for validating the functional importance of candidate differentially methylated sites and regions. The information obtained from classical STARR-seq (Arnold et al. 2013), which is informative about a sequence’s capacity to act as an enhancer, is also generated “for free” as a by- product of the protocol (e.g., data from the unmethylated condition is analogous to

STARR-seq data). Notably, our data strongly suggest that enhancer activity is most reliably detected and most concordant with in vivo chromatin state annotations when larger fragments are tested (relationship between fragment size and enrichment for

‘strong enhancer’ chromatin states: rho=0.722, p=4.80x10-4). This pattern is consistent with the typical size of enhancer elements and may explain, at least in part, recent observations that lentiviral-mediated incorporation of short query fragments (171 bp)

98

recapitulate ENCODE annotations better than episomal MPRA assays (Inoue et al. 2016).

In addition, our results shed light on the biology of transcription factor-DNA binding. Specifically, motif analysis identifies a large set of DNA methylation-sensitive transcription factors, many of which exhibit the canonical pattern in which methylated

DNA is associated with reduced regulatory activity. Some of our results confirm previously hypothesized patterns based on low-throughput electrophoretic mobility shift assays or in vitro TF-DNA binding outside the cellular context. For example, our results identify ETS family TFs as strongly DNA methylation-sensitive (Stephens &

Poon 2016). In addition, they provide intriguing support for the recent argument that pioneer TFs (such as GATA family TFs) have affinity for methylated DNA, which may help initiate local chromatin remodeling by allowing other co-acting regulatory factors to bind (Zhu et al. 2016). Importantly, the few tests of this hypothesis to date have only investigated whether pioneer TFs can bind naked methylated DNA, in the absence of a full cellular complement of competing TFs and possible alternative binding sites. Using mSTARR-seq, we contribute key evidence that TF binding to methylated DNA likely functionally affects downstream gene regulation.

mSTARR-seq can be readily adapted to other experimental designs by varying the input DNA fragment library (e.g., via targeted capture or chromatin immunoprecipitation (Vanhille et al. 2015; Vockley et al. 2016)), the transfected cell types, or the cellular environment (e.g., through in vitro stimulation of transfected cells). We

99

therefore expect that it will facilitate screening and dissection of the causal relationship between DNA methylation and gene expression levels in many contexts. In doing so, mSTARR-seq can help address key questions about both cell type, genotype, or environment-dependent MD activity and the molecular mechanisms that shape this relationship.

4.4 Materials and methods

4.4.1 pmSTARRseq1 vector design

We constructed a self-transcribing CpG-free vector that followed the general design of the STARR-seq vector (Arnold et al. 2013) but was built from Invivogen’s pCpGfree-promoter-Lucia plasmid, which contains an R6K origin of replication and a

Zeocin resistance gene. Specifically, we replaced the region between the NsiI and NheI digest sites with a synthesized sequence containing the following elements: a CpG -free

EF1 promoter (Invivogen), a CpG free version of a synthetic intron (pIRESpuro3,

Clontech), an ORF (the CpG-free CFP::Sh gene from Invivogen), and a cluster of MluI digest sites for screening. The MluI cluster is flanked by restriction enzyme digest sites for SpeI and NcoI to enable high-throughput replacement of this sequence with putative enhancer elements. To avoid any introduction of bacterial methylation, we designed the plasmid to be devoid of Dam methyltransferase targets (GATC), and we only replicated plasmids in a Dcm methyltransferase deficient bacterial strain (GT115, Invivogen).

100

4.4.2 Generation of plasmid libraries for mSTARR-seq

We isolated genomic DNA from the GM12878 cell line (QIAGEN, Blood & Cell

Culture DNA Mini Kit) and fragmented the genomic DNA via sonication (Bioruptor

Standard, Diagenode) or MspI digest (New England BioLabs). We size-selected both the sheared and MspI-digested DNA to a size range of ~300-700bp using gel electrophoresis followed by extraction with the QIAquick Gel Extraction Kit (QIAGEN) and Agencourt

AMPure XP bead cleanups (Beckman Coulter). We used the NEBNext DNA Library

Prep Master Mix Set for Illumina kit to create libraries from 1-2ug of either sheared or

MspI-digested DNA (completing 3-4 replicate library preps for input type). We followed the manufacturer’s instructions except for the final PCR amplification step, during which we used the following primers (forward: actaaagtctagagcttgtaACACTCTTTCCCTACACG, reverse: gaagtggctggctgaattccGTGACTGGAGTTCAGACG). These primers add a 15 base pair extension to the Illumina adapters to facilitate directional high-throughput cloning.

We cloned each library (3-4 replicates per input type) into the mSTARR-seq vector backbone (replacing the cluster of MluI digest sites) using Gibson assembly.

Specifically, we linearized the mSTARR-seq vector by digesting with SpeI and NcoI (New

England BioLabs), running the reaction product on a 1% agarose gel, and isolating the backbone fragment (3.875kb) with the QIAquick gel extraction (QIAGEN). We then assembled the mSTARR vector (1 ug) to an aliquot of MspI or sheared Illumina library

101

(284 ng) using 50 ul of NEBuilder HiFi DNA Assembly Master Mix in 100 ul reaction volumes. We completed 4 and 3 large scale assemblies using MspI-digested and sheared

DNA libraries, respectively. We incubated each of the 7 reactions for 60 min at 50C, and purified the resulting reaction using a 1X Agencourt AMPure XP bead (Beckman

Coulter) cleanup and an elution volume of 7.5 ul.

Each 7.5 ul aliquot of purified, assembled product (n=7 large scale assemblies) was transformed into 300 ul of electrocompetent GT115 E. coli cells (GT115 cells were initially purchased from Invivogen, and a custom electrocompetent version of the strain was prepared by Intact Genomics; note that standard E. coli cells cannot be used in this protocol due to the origin of replication specific to the mSTARR-seq plasmid). We used the BioRad Gene Pulser Xcell system for electroporation and the E.coli 2 mm gap 3 kv program. Following electroporation, each of the 7 pools of 300 ul of cells was allowed to recover in 10 mL of SOC for 1 hr at 37C. Each pool was then transferred to a flask containing 300 ml LB with Zeocin (100 ug/mL) and grown overnight for 12 hours.

Plasmids were then extracted with the QIAGEN Plasmid Plus Maxi Kit, according to the manufacturer’s instructions.

We pooled plasmid DNA derived from MspI-digested library transformations

(n=4) and sheared library transformations (n=3), and mixed these two pools in a 1:2 ratio prior to experimental methylation and transfection. We split the resulting aggregate plasmid pool into 12 replicates and performed 6 ‘methylated condition’

102

methyltransferase reactions (37 C overnight incubation of 60 ug of plasmid DNA with 50 ul of NEB buffer 2, 5 ul of 32 uM SAM, 9 ul of 20 U/ul M.SssI (New England Biolabs) in a

500 ul volume) and 6 ‘unmethylated condition’ reactions in which the M.SssI enzyme was replaced with water. Following incubation, each replicate was purified using a 0.8X

Agencourt AMPure XP bead cleanup (Beckman Coulter).

4.4.3 Cell culture, plasmid transfection, and cell harvesting

K562 cells were cultured in RPMI 1640 media (Gibco) supplemented with 10%

FBS (Sigma Aldrich) and 1% Penicillin/Streptomycin solution (Sigma Aldrich) at 37 C in a 5% CO2 incubator. We transfected each replicate plasmid pool (n=6 methylated and 6 unmethylated replicates) into 20 million cells. Each of the 12 replicate transfections was performed with Lipofectamine 3000 (ThermoFisher Scientific) according to the manufactures’ instructions with reagent quantities scaled accordingly (40 ug of DNA,

110 ul of Lipofectamine 3000, and 100 ul of P3000 per 20 million cells).

Total RNA and plasmid DNA was isolated from the K562 cells 48 hours post transfection. Prior to cell lysis, each replicate cell population was pelleted, washed with

1X PBS, and then incubated at 37 C for 10 min in 3 mL RPMI 1640 containing 1 mL of

Turbo DNase (Ambion) per 36 mL. Cells were then pelleted and washed with 1X PBS, and an aliquot of ~2 million unlysed cells per replicate was set aside for later plasmid extraction (QIAprep Spin Miniprep Kit). The remaining ~18 million cells were pelleted and lysed in 2 mL of Buffer RLT (QIAGEN).

103

4.4.4 Isolation and preparation of mRNA derived from pmSTARRseq1

Total RNA was extracted using the QIAGEN RNeasy Midi kit, and the polyA+

RNA fraction was isolated from 75 ug of total RNA using Dynabeads Oligo dT25

(Invitrogen) and eluted in 25 ul of 10 mM Tris-HCl. Each isolated mRNA sample was then DNase treated by adding 1 ul of Turbo DNase, 3 ul of Turbo DNase buffer, and 1 ul of water, followed by a 30 min incubation at 37C and purification using the RNeasy

MiniElute Cleanup Kit (QIAGEN). Purified mRNA was eluted in 25ul of RNase free water, after which reverse transcription was performed using Super Script III Reverse

Transcriptase (Invitrogen) following the manufacturer’s recommended protocol for higher yield. First strand cDNA synthesis was performed with a primer specific to the mSTARR-seq reporter plasmid (CAAACTCATCAATGTATCTTATCATG), including the optional RNase H step to remove RNA complementary to the cDNA. Following the

RNase H incubation, cDNA from each replicate was purified and concentrated using a

2X Agencourt AMPure XP bead cleanup (Beckman Coulter). Each sample was eluted in

20 ul of nuclease free water.

We amplified the cDNA obtained from reverse transcription for Illumina sequencing using a two step, nested PCR strategy. In the first PCR, we amplified each cDNA pool (n=12 replicates, 6 methylated and 6 unmethylated) using 2 primers specific to the mSTARR-seq reporter plasmid (forward:

GGGCCAGCTGTTGGGGTG*T*C*C*A*C (3’ end is protected by phophorothioate

104

bonds), reverse: CAAACTCATCAATGTATCTTATCATG). The forward primer spans the splice junction of the synthetic intron and specifically amplifies the reporter cDNA without amplifying any residual, unspliced plasmid DNA left in the reaction mixture.

Each PCR reaction contained 2.5 ul of the forward and reverse primers at 10 uM, 25 ul of

NEBNext High-Fidelity 2X PCR Master Mix, and 20 ul of cDNA. Cycling conditions were as follows: 98°C for 2 min; followed by 15 cycles of 98°C for 20 s, 65°C for 30 s,

72°C for 120 s; followed by a final 5 min extension at 72C for 5 min. PCR products were purified using a 2X Agencourt AMPure XP bead cleanup (Beckman Coulter) and eluted in 20 µl of EB buffer (QIAGEN).

The entire purified PCR product (20 ul) from each reaction (n=12 replicates, 6 methylated and 6 unmethylated) served as the template for the second PCR, which contained 25 ul of NEBNext High-Fidelity 2X PCR Master Mix, 2.5 ul of universal primer, and 2.5 ul of an indexed primer (NEBNext Multiplex Oligos for Illumina).

Cycling conditions were as follows: 98°C for 2 min; followed by 10 cycles of 98°C for 20 s, 65°C for 30 s, 72°C for 120 s; followed by a final 5 min extension at 72C for 5min. PCR products were purified using a 1.2X Agencourt AMPure XP bead cleanup (Beckman

Coulter) and final libraries were quantified on an Agilent DNA High Sensitivity Chip

(Agilent Bioanalyzer 2100). Each library was sequenced on the Illumina 4000 platform using 100 bp PE reads.

105

4.4.5 Preparation of plasmid DNA for DNA-seq

To measure the relative abundance of each DNA fragment in the input plasmid libraries, we created DNA-seq libraries from plasmid DNA (extracted from each replicate pool of K562s after transfection and 48 hr incubation). 10 ng of plasmid DNA was used as the input for 50 ul PCRs each containing 25 ul of NEBNext High-Fidelity 2X

PCR Master Mix, 2.5 ul of universal primer, and 2.5 ul of an indexed primer (NEBNext

Multiplex Oligos for Illumina). Cycling conditions were as follows: 98°C for 2 min; followed by 10 cycles of 98°C for 20 s, 65°C for 30 s, 72°C for 120 s; followed by a final 5 min extension at 72C for 5 min. PCR products were purified using a 1X Agencourt

AMPure XP bead cleanup (Beckman Coulter) and final libraries were quantified on an

Agilent DNA High Sensitivity Chip (Agilent Bioanalyzer 2100). Each library was sequenced on the Illumina 4000 platform using 100 bp PE reads.

4.4.6 Low-level data processing

Following sequencing, we removed adapter contamination and low-quality bases from each DNA-seq and mRNA-seq library using the program Trim Galore!

(Krueger 2015). We mapped the trimmed reads to the human reference genome (hg38) using BWA (Li & Durbin 2009), and retained only uniquely mapped reads. In the main text, we report that we assayed ~3/4 of a million unique fragments, which was estimated by counting the number of RNA fragments with unique start and end positions within each replicate (and taking the average across all 12 replicates). For all downstream

106

analyses, we focused on counts of the number of mapped reads in each library (n=12

DNA-seq and 12 mRNA-seq libraries) that overlapped 200 bp non-overlapping genomic intervals. The non-overlapping intervals were created with the BEDtools (Quinlan &

Hall 2010) function ‘makewindows’ and pileups for each library were determined with the ‘coverage’ function. Before analysis, we removed 200 bp regions that were not covered (i.e., had zero mRNA counts) in three or more replicates per condition (note, we did not threshold on DNA counts, because regions with non-zero mRNA counts always had non zero DNA counts). This filtering resulting in a matrix of counts for 277,896 200 bp regions, measured across 12 mRNA libraries and 12 DNA libraries.

4.4.7 Identification of enhancers and methylation dependent (MD) enhancers

Using the set of 277,896 200 bp regions that met our criteria for analysis, we first identified regions that exhibited self-transcribing regulatory activity (more mRNA relative to DNA input) in either the umethylated condition, the methylated condition, or both conditions. To do so, we first normalized the count matrix using the function

‘voomWithQualityWeights’ from the R package ‘limma’ (Law et al. 2014). For each 200 bp region, we then ran a nested model on the normalized values that estimated differences between mRNA and input DNA abundance within each condition:

!! = ! + !!!! ∗ ! !! = 0 + !!!! ∗ ! !! = 1 + !! (1) where !! is the normalized count value for sample !, !! is the sample type (input DNA or mRNA), and ! is an indicator variable for condition (where !! = 0 or !! = 1 indicates a 107

sample from the unmethylated or methylated condition, respectively). !! and !! are thus estimates of the strength of the differences between mRNA and input DNA abundance in the unmethylated and methylated conditions, respectively, where beta values > 0 indicate that mRNA abundance is greater than input DNA abundance. ! is the model intercept and !! denotes model error. We extracted the p-values associated with !! and

!! from each model, and corrected for multiple hypothesis testing using the R function

‘qvalue’ (Dabney & Storey 2015). We considered a region to have enhancer activity if !!,

!!, or both were greater than 0 and had an FDR-corrected p-value less than 0.1.

Using these criteria, we identified 24,945 regions with enhancer activity. We next tested for methylation dependent (MD) activity within this set by adding an interaction between condition (unmethylated or methylated) and sample type (input DNA or mRNA):

!! = ! + !!!! + !!!! + (!! x !!)!!!! + !! (2) where !!, !! !!, and !! are defined as in equation 1. !! is the effect of sample type (input

DNA or mRNA), !! is the effect of condition (unmethylated or methylated), and !!!! is the estimate of the interaction between the two variables (denoted as (!! x !!)). We extracted the p-value associated with !!!! from each model, and corrected for multiple hypothesis testing using the R function ‘qvalue’(Dabney & Storey 2015). We considered a region to have methylation dependent enhancer activity if the p-value estimate for !!!! placed it under an FDR of 10%.

108

4.4.8 Annotation of analyzed mSTARR-seq fragments

Each 200 bp region was originally assayed on one or more ~350 bp fragments

(and later binned into smaller non-overlapping intervals for analysis). Thus, focusing our annotations on the 200 bp analyzed window alone could miss adjacent genomic features that are ultimately responsible for observed enhancer or MD enhancer activity.

Therefore, for all annotations, we focused on the window between the start and end position of each assayed DNA fragment that contained a given 200 bp region. In cases where multiple overlapping DNA fragments contained a given 200 bp region, we used the genomic coordinates of the most downstream and most upstream start and end position, respectively. Consequently, when gathering all remaining genomic annotation, we used a pair of start and end coordinates for each 200 bp pair that spanned a mean of

442 bp (range 100-926bp), as expected from the fragment size distribution we cloned into the mSTARR-seq plasmid.

Using these start and end coordinates for each 200bp region, we calculated the proportion of each of 12 K562-annotated ENCODE ChromHMM classes covered by the fragments we analyzed

(http://hgdownload.cse.ucsc.edu/goldenPath/hg19/encodeDCC/wgEncodeBroadHmm/) using the BEDtools (Quinlan & Hall 2010) function ‘intersect’. We also estimated the proportion of fragments from an in silico MspI digest (see In silico MspI digest) that we captured. To calculate enrichment of chromatin states captured in our mSTARR-seq

109

enhancers, we assigned the fragment window associated with each 200 bp region to a particular chromatin state if >50% of the window overlapped that particular state and performed Fisher’s exact tests for each chromatin state.

To perform our random forests prediction, we gathered the following data for each fragment window associated with each analyzed 200 bp region:

(i) The expression level of the nearest gene for regions that were within 200 kb of a gene TSS or TES. Expression levels were summarized as FPKM estimates from RNA- seq performed on K562 cells (NCBI GEO accession GSE86747).

(ii) The average CpG methylation level of the fragment window in K562s.

Estimates were derived from ENCODE whole genome bisulfite sequencing data, after removing low coverage (<5x) CpG sites (NCBI GEO accession GSM958729).

(iii) The average DNase-seq signal for the fragment window in K562s, using measures of DNase-seq peak strength (downloaded from http://hgdownload.cse.ucsc.edu/goldenPath/hg38/database/wgEncodeRegDnaseUwK56

2Peak.txt.gz).

(iv) The average evolutionary conservation score of the region, using phastCons scores derived from a 44-way alignment across placental mammals

(http://hgdownload.cse.ucsc.edu/goldenPath/hg19/database/phastConsElements46wayP lacental.txt.gz).

110

(v) The absolute number of CpG sites and CpG density in the fragment window based on the human reference genome (hg38) sequence.

(vi) The presence or absence of a predicted binding site for each one of 260 transcription factors. For these annotations, we relied on publicly available coordinates for predicted transcription factor binding events in K562s. These coordinates were derived from the program CENTIPEDE, which uses position weight matrix information plus experimental data (DNase-seq data from K562s) to infer transcription factor binding sites with high specificity (http://centipede.uchicago.edu/SimpleMulti/).

4.4.9 In silico MspI digest

To obtain in silico MspI digest fragments, we simulated cut sites in the human genome (hg38) wherever the sequence CCGG was observed. We filtered the resulting fragments for sizes commonly retained during library preparation (100-500 bp), and then sampled 10 million fragments from this pool. Next, we mapped the first 100 bp and the last 100 bp of each fragment as a paired end library using the default paired-end mapping settings in BSMAP (Xi & Li 2009). We retained all uniquely mapping reads and extracted the most upstream and most downstream coordinate for each mapped read.

We then used the ‘intersect’ function from BEDtools (Quinlan & Hall 2010) to overlap this set of mapped fragments with the fragment windows associated with our set of

277,896 200 bp analyzed windows.

111

4.4.10 Random forests classification

To determine whether MD activity could be reliably predicted from a set of sequence and functional genomic features (see Annotation of analyzed mSTARR-seq fragments), we conducted random forest analyses (Liaw & Wiener 2002). We compared

MD enhancers with greater activity when unmethylated against enhancers with no evidence for methylation dependence (FDR>50% in the test for MD activity). Because random forests require a complete data set, we removed regions with missing values for any of the predictive features, resulting in a data set of n=1863 MD enhancers and n=5703 non-MD enhancers. Further, we pruned our random forests features to exclude any unnamed or non-human transcription factors, as well as transcription factors that were associated with fewer than 10 of the total 7566 analyzed enhancers.

Using the remaining 56 features, we used the R package ‘randomForest’ (Liaw &

Wiener 2002) to iteratively construct training sets comprised of approximately 2/3 of the original data, and test sets comprised of the remaining enhancers in the data set (also known as the ‘out of bag’ set). We grew 1000 classification trees and evaluated predictive accuracy using the out of bag test sets. To estimate the predictive value of individual features, we calculated the mean decrease in accuracy and the Gini coefficient when that feature was removed and estimated the significance of these contributions via comparison to the results of permutation analyses (1000 permutations of MD versus non-MD classification) implemented in the R package ‘rfPermute’ (Archer 2015). We

112

considered a variable to be significant if the mean decrease in accuracy and Gini coefficient estimates both passed a 10% empirical FDR (see Figure 12 for significant variables).

4.3.11 Transcription factor binding motif enrichment analyses

To ask whether binding sites for certain TFs were consistently associated with windows with MD enhancer activity, we used the ‘findMotifsGenome.pl’ script in the motif analysis program HOMER (Heinz et al. 2010) to test for enrichment of 309 known vertebrate binding motifs relative to the background set of fragment windows associated with all mSTARR-seq enhancers. We performed this analysis twice, focusing on test sets of fragment windows associated with MD enhancers that were more active when unmethylated and more active when methylated, respectively. We considered a

TF binding motif to be significantly enriched in each of the test sets if the motif was found on >5% of fragments in the background set and the motif passed a 1% FDR threshold (Benjamini & Hochberg 1995). Significantly enriched motifs are summarized in Figure 13. To test for enrichment of specific TF families among the set of significant motifs, we used hypergeometric tests to compare the proportion of significant motifs belonging to a particular TF family (e.g., ETS, GATA, etc.) to the proportion of all tested motifs that belong to that family.

113

4.4.12 Correlations between DNA methylation and gene expression levels in primary cells

To understand whether mSTARR-seq data can explain heterogeneity in in vivo

DNA methylation-gene expression correlations across individuals, in primary cells, we used a publicly available data set on 1202 monocyte samples (Reynolds et al. 2014). We downloaded paired genome-wide DNA methylation data (measured on the Illumina

HumanMethylation450 BeadChip) and gene expression data (measured on the Illumina

HumanHT-12 v4 Expression BeadChip) for each sample. DNA methylation data were provided as continuously varying, unbounded M-values (Du et al. 2010) and expression data were provided as raw signal intensity values, which we normalized using the voom function in the limma R package (Law et al. 2014)).

We filtered the DNA methylation data and expression data to remove probes with mean detection level p-values>0.05. We further filtered the DNA methylation data to remove probes with M-values corresponding to an average methylation level <0.1 or

>0.9, which represent constitutively hypo and hypermethylated regions, respectively.

This filtering left us with 15,247 probes measuring gene expression and 425,895 probes measuring DNA methylation. We next identified CpG sites that overlapped regions with significant enhancer activity in our mSTARR-seq assay. We associated each of these CpG sites to its closest gene and removed CpG sites from our analysis that were >100kb from any gene. In the remaining set of 2699 unique CpG sites (associated with 1002 unique genes), we used linear models to test for an association between methylation levels at

114

each filtered CpG site and gene expression levels. We extracted the p-value and estimated coefficient from each model, and considered a correlation to be significant if it passed a 10% FDR. Finally, we annotated each CpG based on whether it overlapped an

MD or non MD enhancer, and based on whether it occurred in the associated gene’s putative promoter region (i.e., within 2 kb region upstream of a gene’s TSS). We used

Fisher’s Exact Tests to investigate enrichment of CpG methylation-expression correlations in MD versus non-MD mSTARR-seq enhancers.

115

References

Adey A, Shendure J (2012) Ultra-low-input, tagmentation-based whole-genome bisulfite sequencing. Genome Research, 22, 1139–1143.

Alberts SC, Altmann J (2003) Intraspecific variability in fertility and offspring survival in a nonhuman primate: behavioral control of ecological and social sources. In: Offspring: The Biodemography of Fertility and Family Behavior (eds Wachter KW, Bulatao RA), pp. 140–169. National Academy Press, Washington, DC.

Alberts SC, Altmann J (2012) The Amboseli Baboon Research Project: 40 years of continuity and change. In: Long-Term Field Studies of Primates (eds Kappeler P, Watts DP), pp. 261–288. Springer, New York.

Alberts SC, Hollister-Smith J, Mututua RS et al. (2005) Seasonality and long term change in a savannah environment. In: Primate Seasonality: Implications for Human Evolution (eds Brockman DK, van Schaik CP), pp. 157–196. Cambridge University Press.

Altmann J (1980) Baboon mothers and infants. Harvard University Press, Cambridge.

Altmann S (1991) Diets of yearling female primates (Papio cynocephalus) predict lifetime fitness. Proceedings of the National Academy of Sciences of the United States of America, 88, 420–3.

Altmann J, Alberts SC (2005) Growth rates in a wild primate population: ecological influences and maternal effects. Behavioral Ecology and Sociobiology, 57, 490–501.

Altmann J, Alberts S, Haines S et al. (1996) Behavior predicts genetic structure in a wild primate group. Proceedings of the National Academy of Sciences of the United States of America, 93, 5797–5801.

Altmann J, Gesquiere L, Galbany J, Onyango PO, Alberts SC (2010) Life history context of reproductive aging in a wild primate model. Annals of the New York Academy of Sciences, 1204, 127–38.

Altmann J, Schoeller D, Altmann S a, Muruthi P, Sapolsky RM (1993) Body size and fatness of free-living baboons reflect food availability and activity levels. American Journal of Primatology, 30, 148–161.

Andersson R, Gebhard C, Miguel-Escalada I et al. (2014) An atlas of active enhancers across human cell types and tissues. Nature, 507, 455–61.

Applebaum S, Heifetz Y (1999) Density-dependent physiological phase in insects. 116

Annual review of entomology, 44, 317–41.

Aran D, Sabato S, Hellman A (2013) DNA methylation of distal regulatory sites characterizes dysregulation of cancer genes. Genome Biology, 14, R21.

Archer E (2015) rfPermute: estimate Permutation p-Values for Random Forest Importance Metrics. R package version 1.42.0.

Arnold CD, Gerlach D, Spies D et al. (2014) Quantitative genome-wide enhancer activity maps for five Drosophila species show functional enhancer conservation and turnover during cis-regulatory evolution. Nature Genetics, 46, 685–692.

Arnold CD, Gerlach D, Stelzer C et al. (2013) Genome-wide quantitative enhancer activity maps identified by STARR-seq. Science, 339, 1074–1077.

Backes C, Keller A, Kuentzer J et al. (2007) GeneTrail-advanced gene set enrichment analysis. Nucleic Acids Research, 35, 186–192.

Bakulskia K, Dolinoya D, Sartorb M et al. (2012) Genome-wide DNA methylation differences between late-onset Alzheimer’s disease and cognitively normal controls in human frontal cortex. J Alzheimers Dis, 29, 1–28.

Barboza Solís C, Kelly-Irving M, Fantin R et al. (2015) Adverse childhood experiences and physiological wear-and-tear in midlife: Findings from the 1958 British birth cohort. Proceedings of the National Academy of Sciences of the United States of America, 112, E738-46.

Barker DJP, Gluckman PD, Godfrey KM et al. (1993) Fetal nutrition and cardiovascular disease in adult life. Lancet, 341, 1421–2.

Barrès R, Yan J, Egan B et al. (2012) Acute exercise remodels promoter methylation in human skeletal muscle. Cell Metabolism, 15, 405–11.

Bates D, Maechler M, Bolker B, Walker S (2014) lme4: Linear mixed-effects models using Eigen and S4.

Bateson P, Barker D, Clutton-Brock T, Deb D (2004) Developmental plasticity and human health. Nature, 419–421.

Beehner JC, Onderdonk DA, Alberts SC, Altmann J (2006) The ecology of conception and pregnancy failure in wild baboons. Behavioral Ecology, 17, 741–750.

Bell JT, Pai A, Pickrell JK et al. (2011) DNA methylation patterns associate with genetic

117

and gene expression variation in HapMap cell lines. Genome Biology, 12, R10.

Benjamini Y, Hochberg Y (1995) Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society, 57, 289– 300.

Boersma M, Spaak P, De Meester L (1998) Predator-Mediated Plasticity in Morphology , Life History , and Behavior of Daphnia : The Uncoupling of Responses. The American Naturalist, 152, 237–48.

Bol V, Desjardins F, Reusens B, Balligand J-L, Remacle C (2010) Does early mismatched nutrition predispose to hypertension and atherosclerosis, in male mice? PLoS One, 5, e12656.

Botero C, Weissing FJ, Wright J, Rubenstein DR (2015) Evolutionary tipping points in the capacity to adapt to environmental change. Proceedings of the National Academy of Sciences, 112.

Boyle P, Clement K, Gu H, Smith Z (2012) Gel-free multiplexed reduced representation bisulfite sequencing for large-scale DNA methylation profiling. Genome Biology, 13, R92.

Bronikowski AM, Altmann J (1996) Foraging in a variable environment: weather patterns and the behavioral ecology of baboons. Behavioral Ecology and Sociobiology, 39, 11–25.

Buenrostro J, Giresi PG, Zaba L, Chang H, Greenleaf W (2013) Transposition of native chromatin for multimodal regulatory analysis and personal epigenomics Jason. Nature Methods, 10, 1213–1218.

Carone BR, Fauquier L, Habib N et al. (2010) Paternally induced transgenerational environmental reprogramming of metabolic gene expression in mammals. Cell, 143, 1084–96.

Charpentier MJE, Tung J, Altmann J, Alberts SC (2008) Age at maturity in wild baboons: genetic, environmental and demographic influences. Molecular Ecology, 17, 2026–40.

Chen E, Miller GE, Kobor MS, Cole SW (2011) Maternal warmth buffers the effects of low early-life socioeconomic status on pro-inflammatory signaling in adulthood. Molecular Psychiatry, 16, 729–37.

Christman JK (2002) 5-Azacytidine and 5-aza-2’-deoxycytidine as inhibitors of DNA methylation: mechanistic studies and their implications for cancer therapy.

118

Oncogene, 21, 5483–5495.

Coban E, Ozdogan M, Yazicioglu G, Akcit F (2005) The mean platelet volume in patients with obesity. International Journal of Clinical Practice, 59, 981–982.

Comuzzie AG, Cole S a., Laston SL et al. (2012) Novel Genetic Loci Identified for the Pathophysiology of Childhood Obesity in the Hispanic Population. PLoS ONE, 7.

Cooper CDO, Newman JA, Aitkenhead H, Allerston CK, Gileadi O (2015) Structures of the Ets protein DNA-binding domains of transcription factors Etv1, Etv4, Etv5, and Fev: Determinants of DNA binding and redox regulation by disulfide bond formation. Journal of Biological Chemistry, 290, 13692–13709.

Cortes C, Vapnik V (1995) Support-Vector Networks. Machine Learning, 273–297.

Cummings ME, Larkins-Ford J, Reilly CRL et al. (2008) Sexual and social stimuli elicit rapid and contrasting genomic responses. Proceedings of the Royal Society of London, Series B, 275, 393–402.

Dabney A, Storey J (2015) qvalue: Q-value estimation for false discovery rate control. R package version 1.43.0.

Dantzer B, Newman A, Boonstra R et al. (2013) Density triggers maternal hormones that increase adaptive offspring growth in a wild mammal. Science, 340, 1215–1217.

Darlington RB, Smulders T V. (2001) Problems with residual analysis. Animal Behaviour, 62, 599–602.

Day K, Waite LL, Thalacker-Mercer A et al. (2013) Differential DNA methylation with age displays both common and dynamic features across human tissues that are influenced by CpG landscape. Genome Biology, 14, R102.

Dayeh T, Volkov P, Salö S et al. (2014) Genome-wide Dna methylation analysis of human pancreatic islets from type 2 diabetic and non-diabetic donors identifies candidate genes that influence insulin secretion. PLoS Genetics, 10.

Dobzhansky T (1973) Nothing in Biology Makes Sense except in the Light of Evolution. The American Biology Teacher, 35, 125–129.

Domcke S, Bardet AF, Adrian Ginno P et al. (2015) Competition between DNA methylation and transcription factors determines binding of NRF1. Nature, 528, 575–579.

119

Dominguez-Salas P, Moore SE, Baker MS et al. (2014) Maternal nutrition at conception modulates DNA methylation of human metastable epialleles. Nature Communications, 5, 3746.

Douhard M, Plard F, Gaillard J-M et al. (2014) Fitness consequences of environmental conditions at different life stages in a long-lived vertebrate. Proceedings of the Royal Society of London, Series B, 281.

Dourmashkin JT, Chang G, Gayles EC et al. (2005) Different forms of obesity as a function of diet composition. International Journal of Obesity, 29, 1368–1378.

Drnevich J, Replogle KL, Lovell P et al. (2012) Impact of experience-dependent and - independent factors on gene expression in songbird brain. Proceedings of the National Academy of Sciences of the United States of America, 109, 17245–17252.

Du P, Zhang X, Huang C-C et al. (2010) Comparison of Beta-value and M-value methods for quantifying methylation levels by microarray analysis. BMC Bioinformatics, 11, 587.

Eckhardt F, Lewin J, Cortese R et al. (2006) DNA methylation profiling of human chromosomes 6, 20 and 22. Nature Genetics, 38, 1378–85.

Ehrich TH, Hrbek T, Kenney-hunt JP et al. (2005) Fine-Mapping Gene-by-Diet Interactions on Chromosome 13 in a LG/J x SM/J Murine Model of Obesity. Diabetes, 54, 1863–1872.

El-Maarri O (2005) DNA methylation and human disease. Nature Reviews Genetics, 544, 135–144.

Elliott E, Ezra-Nevo G, Regev L, Neufeld-Cohen A, Chen A (2010) Resilience to social stress coincides with functional DNA methylation of the CRF gene in adult mice. Nature Neuroscience, 13, 1351–3.

Ellison PT (2014) Evolutionary Tradeoffs. Evolution, medicine, and public health, 93.

English S, Huchard E, Nielsen JF, Clutton-Brock TH (2013) Early growth, dominance acquisition and lifetime reproductive success in male and female cooperative meerkats. Ecology and Evolution, 3, 4401–4407.

Faulk C, Dolinoy DC (2011) Timing is everything: The when and how of environmentally induced changes in the epigenome of animals. Epigenetics, 6, 791– 797.

120

Feil R, Fraga MF (2011) Epigenetics and the environment: emerging patterns and implications. Nature Reviews Genetics, 13, 97–109.

Felitti V, Anda R, Nordenberg D et al. (1998) Relationship of childhood abuse and household dysfunction to many of the leading causes of death in adults. The Adverse Childhood Experiences (ACE) Study. American Journal of Preventative Medicine, 14, 245–258.

Foret S, Kucharski R, Pellegrini M et al. (2012) DNA methylation dynamics, metabolic fluxes, gene splicing, and alternative phenotypes in honey bees. Proceedings of the National Academy of Sciences, 109, 4968–4973.

Freckleton R (2002) On the misuse of residuals in ecology: regression of residuals vs . multiple regression. Journal of Animal Ecology, 542–545.

French JA, Carp SB (2016) Early-life social adversity and developmental processes in nonhuman primates. Current Opinion in Behavioral Sciences, 7, 40–46.

Furrow RE, Feldman MW (2014) Genetic Variation and the Evolution of Epigenetic Regulation. Evolution, 68, 673–683.

Gaillard J, Festa-Bianchet M, Yoccoz N, Loison A, Toigo C (2000) Temporal variation in fitness components and population dynamics of large herbivores. Annual Review of Ecology and Systematics, 31, 367–393.

Gapp K, Jawaid A, Sarkies P et al. (2014) Implication of sperm RNAs in transgenerational inheritance of the effects of early trauma in mice. Nature neuroscience, 17, 667–9.

Gesquiere LR, Wango EO, Alberts SC, Altmann J (2007) Mechanisms of sexual selection: sexual swellings and estrogen concentrations as fertility indicators and cues for male consort decisions in wild baboons. Hormones and Behavior, 51, 114–25.

Getty-Kaushik L, Viereck JC, Goodman JM et al. (2010) Mice Deficient in Phosphofructokinase-M Have Greatly Decreased Fat Stores. Obesity, 18, 434–440.

Gieger C, Kühnel B, Radhakrishnan a et al. (2011) New gene functions in megakaryopoiesis and platelet formation. Nature, 480, 201–208.

Giovanoli S, Engler H, Engler A et al. (2013) Stress in puberty unmasks latent neuropathological consequences of prenatal immune activation in mice. Science, 339, 1095–1099.

Del Giudice M, Ellis BJ, Shirtcliff E a. (2011) The Adaptive Calibration Model of stress

121

responsivity. Neuroscience and Biobehavioral Reviews, 35, 1562–1592.

Gluckman PD, Hanson MT (2004a) The developmental origins of the metabolic syndrome. Trends in Endocrinology and Metabolism, 15, 183–7.

Gluckman PD, Hanson M a T (2004b) Living with the past: evolution, development, and patterns of disease. Science, 305, 1733–6.

Gluckman PPD, Hanson M, Cooper C, Thornburg KL (2008) Effect of in utero and early- life conditions on adult health and disease. The New England Journal of Medicine, 359, 61–73.

Gluckman PD, Hanson MT, Spencer HG (2005a) Predictive adaptive responses and human evolution. Trends in Ecology and Evolution, 20, 527–33.

Gluckman PD, Hanson M a, Spencer HG, Bateson P (2005b) Environmental influences during development and their later consequences for health and disease: implications for the interpretation of empirical studies. Proceedings of the Royal Society of London, Series B, 272, 671–7.

Goeman JJ, Van de Geer S, De Kort F, van Houwellingen HC (2004) A global test for groups fo genes: Testing association with a clinical outcome. Bioinformatics, 20, 93– 99.

Grafen A (1988) On the uses of data on lifetime reproductive success. In: Reproductive success (ed Clutton-Brock TH), pp. 454–471. University of Chicago Press, Chicago.

Gu H, Smith ZD, Bock C et al. (2011) Preparation of reduced representation bisulfite sequencing libraries for genome-scale DNA methylation profiling. Nature Protocols, 6, 468–81.

Guo JU, Ma DK, Mo H et al. (2011) Neuronal activity modifies the DNA methylation landscape in the adult brain. Nature Neuroscience, 14, 1345–1351.

Hales CN, Barker DJ (2001) The thrifty phenotype hypothesis. British Medical Bulletin, 60, 5–20.

Hamel S, Gaillard J-M, Festa-Bianchet M, Côté SD (2009) Individual quality, early-life conditions, and reproductive success in contrasted populations of large herbivores. Ecology, 90, 1981–95.

Hansen K, Langmead B, Irizarry R (2012) BSmooth : from whole genome bisulfite sequencing reads to differentially methylated regions. Genome Biology, 13, R83.

122

Hansen KD, Timp W, Bravo HC et al. (2011) Increased methylation variation in epigenetic domains across cancer types. Nature Genetics, 43, 768–75.

Hastie T, Tibshirani R, Friedman J (2009) The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, New York, NY.

Hausfater G (1974) Dominance and reproduction in baboons (Papio cynocephalus): A quantitative analysis. University of Chicago.

Hausfater G, Altmann J, Altmann S (1982) Long-Term Consistency of Dominance Relations Among Female Baboons (Papio cynocephalus). Science, 217, 752–5.

Hayward AD, Lummaa V (2013) Testing the evolutionary basis of the predictive adaptive response hypothesis in a preindustrial human population. Evolution, Medicine, and Public Health, 2013, 106–17.

Hayward AD, Rickard IJ, Lummaa V (2013) Influence of early-life nutrition on mortality and reproductive success during a subsequent famine in a preindustrial population. Proceedings of the National Academy of Sciences of the United States of America, 110, 13886–91.

Heijmans BT, Tobi EW, Stein AD et al. (2008) Persistent epigenetic differences associated with prenatal exposure to famine in humans. Proceedings of the National Academy of Sciences of the United States of America, 105, 17046–9.

Heim C, Binder EB (2012) Review of human studies on sensitive periods, gene- environment interactions, and epigenetics. Experimental Neurology, 233, 102–111.

Heinz S, Benner C, Spann N et al. (2010) Simple Combinations of Lineage-Determining Transcription Factors Prime cis-Regulatory Elements Required for Macrophage and B Cell Identities. Molecular Cell, 38, 576–589.

Heyn H, Esteller M (2012) DNA methylation profiling in the clinic: applications and challenges. Nature Reviews Genetics, 13, 679–92.

Hinoue T, Weisenberger DJ, Lange CPE et al. (2012) Genome-scale analysis of aberrant DNA methylation in colorectal cancer. Genome Research, 22, 271–282.

Horton TH, Stetson MH (1992) Maternal transfer of photoperiodic information in rodents. Animal Reproduction Science, 30, 29–44.

Hostinar CE, Gunnar MR (2013) The Developmental Effects of Early Life Stress: An Overview of Current Theoretical Frameworks. Current directions in psychological

123

science, 22, 400–406.

Hu S, Wan J, Su Y et al. (2013) DNA methylation presents distinct binding sites for human transcription factors. eLife, 2013, 1–16.

Inoue F, Kircher M, Martin B et al. (2016) A systematic comparison reveals substantial differences in chromosomal versus episomal encoding of enhancer activity. bioRxiv, 61606.

Jacobsen SC, Brøns C, Bork-Jensen J et al. (2012) Effects of short-term high-fat overfeeding on genome-wide DNA methylation in the skeletal muscle of healthy young men. Diabetologia, 55, 3341–3349.

Jaffe AE (2015) FlowSorted.Blood.450k: Illumina HumanMethylation data on sorted blood cell populations. R package version 1.5.1.

Jaffe AE, Murakami P, Lee H et al. (2012) Bump hunting to identify differentially methylated regions in epigenetic epidemiology studies. International Journal of Epidemiology, 41, 200–209.

De Jager PL, Srivastava G, Lunnon K et al. (2014) Alzheimer’s disease: early alterations in brain DNA methylation at ANK1, BIN1, RHBDF2 and other loci. Nature Neuroscience, 17, 1156–1163.

Jesri A, Okonofua EC, Egan BM (2005) Platelet and white blood cell counts are elevated in patients with the metabolic syndrome. Journal of Clinical Hypertension, 7, 705-711- 713.

Jirtle RL, Skinner MK (2007) Environmental epigenomics and disease susceptibility. Nature reviews. Genetics, 8, 253–62.

Jones P (1994) Schizophrenia after prenatal exposure to the Dutch hunger winter of 1944- 1945. Archives of General Psychiatry, 51, 333–4.

Jones P (2012) Functions of DNA methylation: islands, start sites, gene bodies and beyond. Nature Reviews Genetics, 13, 484–92.

Kappeler L, Meaney MJ (2010) Epigenetics and parental effects. BioEssays : news and reviews in molecular, cellular and developmental biology, 32, 818–27.

Kemnitz JW, Sapolsky RM, Altmann J et al. (2002) Effects of food availability on serum insulin and lipid concentrations in free-ranging baboons. American Journal of Primatology, 57, 13–9.

124

Kemp MW, Kallapur SG, Jobe AH, Newnham JP (2012) Obesity and the developmental origins of health and disease. Journal of Paediatrics and Child Health, 48, 86–90.

Khulan B, Cooper WN, Skinner BM et al. (2012) Periconceptional maternal micronutrient supplementation is associated with widespread gender related changes in the epigenome: A study of a unique resource in the Gambia. Human Molecular Genetics, 21, 2086–2101.

Klebig ML, Wilkinson JE, Geisler JG, Woychik RP (1995) Ectopic expression of the agouti gene in transgenic mice causes obesity, features of type II diabetes, and yellow fur. Proceedings of the National Academy of Sciences of the United States of America, 92, 4728– 32.

Klose RJ, Bird AP (2006) Genomic DNA methylation: the mark and its mediators. Trends in Biochemical Sciences, 31, 89–97.

Klug M, Rehli M (2006a) Functional analysis of promoter CpG methylation using a CpG- free luciferase reporter vector. Epigenetics, 1, 127–130.

Klug M, Rehli M (2006b) Functional Analysis of Promoter CpG Methylation Using a CpG-Free Luciferase Reporter Vector. Epigenetics, 1, 127–130.

Klughammer J, Datlinger P, Printz D et al. (2015) Differential DNA Methylation Analysis without a Reference Genome. Cell Reports, 13, 2621–2633.

Krueger F (2015) Trim Galore!

Kruuk LE, Clutton-Brock TH, Rose KE, Guinness FE (1999) Early determinants of lifetime reproductive success differ between the sexes in red deer. Proceedings. Biological sciences / The Royal Society, 266, 1655–61.

Kucharski R, Maleszka J, Foret S, Maleszka R (2008) Nutritional control of reproductive status in honeybees via DNA methylation. Science, 319, 1827–1830.

Kuzawa CW, Quinn E a. (2009) Developmental Origins of Adult Function and Health: Evolutionary Hypotheses. Annual Review of Anthropology, 38, 131–147.

Lam LL, Emberly E, Fraser HB et al. (2012) Factors underlying variable DNA methylation in a human community cohort. Proceedings of the National Academy of Sciences of the United States of America, 109, 17253–60.

Law C, Chen Y, Shi W, Smyth G (2014) Voom! Precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biology, 15, R29.

125

Lea AJ, Altmann J, Alberts SC, Tung J (2015a) Developmental constraints in a wild primate. The American Naturalist, 185, 809–821.

Lea AJ, Altmann J, Alberts SC, Tung J (2016a) Resource base influences genome-wide DNA methylation levels in wild baboons (Papio cynocephalus). Molecular Ecology, 25, 1681–1696.

Lea AJ, Learn NH, Theus MJ, Altmann J, Alberts SC (2014) Complex sources of variance in female dominance rank in a nepotistic society. Animal Behaviour, 94, 87–99.

Lea A, Tung J, Zhou X (2015b) A flexible, efficient binomial mixed model for identifying differential DNA methylation in bisulfite sequencing data. PLoS Genetics, 11, e1005650.

Lea A, Vilgalys T, Durst P, Tung J (2016b) Maximizing ecological and evolutionary insight from bisulfite sequencing data sets. bioRxiv, 1–27.

Lee M (1988) Vole infant development is influenced by maternal photoperiodic history perinatally. American Journal of Physiology, 255, 831–838.

Li H, Durbin R (2009) Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics, 25, 1754–60.

Li Y, He Y, Qi L et al. (2010) Exposure to the Chinese famine in early life and the risk of hyperglycemia and type 2 diabetes in adulthood. Diabetes, 59, 2400–2406.

Liaw A, Wiener M (2002) Classification and regression by randomForest. R News, 2, 18– 22.

Lindström J (1999) Early development and fitness in birds and mammals. Trends in Ecology & Evolution, 14, 343–348.

Lister R, Pelizzola M, Dowen R, Hawkins R (2009) Human DNA methylomes at base resolution show widespread epigenomic differences. Nature, 462.

Liu XS, Wu H, Ji X et al. (2016) Editing DNA Methylation in the Mammalian Genome. Cell, 167, 233–247.e17.

Lo KA, Labadorf A, Kennedy NJ et al. (2013) Analysis of In Vitro Insulin-Resistance Models and Their Physiological Relevance to InVivo Diet-Induced Adipose Insulin Resistance. Cell Reports, 5, 259–270.

Lucas ME, Crider KS, Powell DR, Kapoor-Vazirani P, Vertino PM (2009) Methylation-

126

sensitive regulation of TMS1/ASC by the Ets factor, GA-binding protein. Journal of Biological Chemistry, 284, 14698–14709.

Maeder ML, Angstman JF, Richardson ME et al. (2013) Targeted DNA demethylation and activation of endogenous genes using programmable TALE-TET1 fusion proteins. Nature Biotechnology, 31, 1137–42.

Mann IK, Chatterjee R, Zhao J et al. (2013) CG methylated microarrays identify a novel methylated sequence bound by the CEBPB | ATF4 heterodimer that is active in vivo. Genome Research, 988–997.

Mariette MM, Buchanan KL (2016) Prenatal acoustic communication programs offspring for high posthatching temperatures in a songbird. Science, 353, 812–814.

Marioni RE, Shah S, McRae AF et al. (2015) DNA methylation age of blood predicts all- cause mortality in later life. Genome Biology, 16, 1–12.

Marr AK, MacIsaac JL, Jiang R et al. (2014) Leishmania donovani Infection Causes Distinct Epigenetic DNA Methylation Changes in Host Macrophages. PLoS Pathogens, 10, e1004419.

Maurano MT, Wang H, John S et al. (2015) Role of DNA Methylation in Modulating Transcription Factor Occupancy. Cell Reports, 12, 1184–1195.

McDade TW (2012) Early environments and the ecology of inflammation. Proceedings of the National Academy of Sciences of the United States of America, 109 Suppl, 17281–8.

McErlean P, Jr SF, Costa FF et al. (2014) Human rhinovirus infection causes different DNA methylation changes in nasal epithelial cells from healthy and asthmatic subjects. BMC Medical Genomics, 7, 37.

McGowan PO, Sasaki A, D’Alessio AC et al. (2009) Epigenetic regulation of the glucocorticoid receptor in human brain associates with childhood abuse. Nature Neuroscience, 12, 342–8.

McGuinness D, McGlynn LM, Johnson PCD et al. (2012) Socio-economic status is associated with epigenetic differences in the pSoBid cohort. International Journal of Epidemiology, 41, 151–60.

Meaney M (2001) Maternal care, gene expression, and the transmission of individual differences in stress reactivity across generations. Annual Review of Neuroscience, 1161–1192.

127

Meaney M, Szyf M (2005) Environmental programming of stress responses through DNA methylation: life at the interface between a dynamic environment and a fixed genome. Dialogues in Clinical Neuroscience, 3, 103–123.

Meissner A, Mikkelsen TS, Gu H et al. (2008) Genome-scale DNA methylation maps of pluripotent and differentiated cells. Nature, 454, 766–70.

Melnikov A, Murugan A, Zhang X et al. (2012) Systematic dissection and optimization of inducible enhancers in human cells using a massively parallel reporter assay. Nature biotechnology, 30, 271–7.

Moleres A, Campión J, Milagro FI et al. (2013) Differential DNA methylation patterns between high and low responders to a weight loss intervention in overweight or obese adolescents: the EVASYON study. FASEB journal: official publication of the Federation of American Societies for Experimental Biology, 27, 2504–12.

Monaghan P (2008) Early growth conditions, phenotypic development and environmental change. Philosophical transactions of the Royal Society of London. Series B, Biological sciences, 363, 1635–45.

Monaghan P, Heidinger BJ, D’Alba L, Evans NP, Spencer K a (2012) For better or worse: reduced adult lifespan following early-life stress is transmitted to breeding partners. Proceedings. Biological sciences / The Royal Society, 279, 709–14.

Morris WF, Altmann J, Brockman DK et al. (2011) Low demographic variability in wild primate populations: fitness impacts of variation, covariation, and serial correlation in vital rates. The American Naturalist, 177, E14-28.

Mumby HS, Mar KU, Hayward AD, Htut W, Htut-aung Y (2015) Elephants born in the high stress season have faster reproductive ageing. Scientific Reports, 5, 1–11.

Murgatroyd C, Patchev A V, Wu Y et al. (2009) Dynamic DNA methylation programs persistent adverse effects of early-life stress. Nature Neuroscience, 12, 1559–66.

Muruthi P, Altmann J, Altmann S (1991) Resource base, parity, and reproductive condition affect females’ feeding time and nutrient intake within and between groups of a baboon population. Oecologia, 467–472.

Nan X, Ng HH, Johnson CA et al. (1998) Transcriptional repression by the methyl-CpG- binding protein MeCP2 involves a histone deacetylase complex. Nature, 393, 386– 389.

Naumova OY, Lee M, Koposov R et al. (2012) Differential patterns of whole-genome

128

DNA methylation in institutionalized children and children raised by their biological parents. Development and Psychopathology, 24, 143–55.

Nettle D, Frankenhuis W, Rickard I (2013) The evolution of predictive adaptive responses in human life history. Proceedings of the Royal Society of London, Series B, 280, 13–43.

Nussey DH, Kruuk LEB, Morris A, Clutton-Brock TH (2007) Environmental conditions in early life influence ageing rates in a wild population of red deer. Current Biology, 17, R1000-1.

O’Malley RC, Huang SSC, Song L et al. (2016) Cistrome and Epicistrome Features Shape the Regulatory DNA Landscape. Cell, 165, 1280–1292.

Ogata H, Goto S, Sato K et al. (1999) KEGG: Kyoto encyclopedia of genes and genomes. Nucleic Acids Research, 27, 29–34.

Onyango PO, Gesquiere LR, Wango EO, Alberts SC, Altmann J (2008) Persistence of maternal effects in baboons: Mother’s dominance rank at son’s conception predicts stress hormone levels in subadult males. Hormones and Behavior, 54, 319–24.

Pacis A, Tailleux L, Lambourne J et al. (2015) Bacterial Infection Remodels the DNA Methylation Landscape of Human Dendritic Cells. Genome Research.

Perini G, Diolaiti D, Porro A, Della Valle G (2005) In vivo transcriptional regulation of N-Myc target genes is controlled by E-box methylation. Proceedings of the National Academy of Sciences of the United States of America, 102, 12117–22.

Pesonen A-K, Räikkönen K (2012) The lifespan consequences of early life stress. Physiology & behavior, 106, 722–7.

Picelli S, Björklund AK, Reinius B et al. (2014) Tn5 transposase and tagmentation procedures for massively scaled sequencing projects. Genome Research, 24, 2033– 2040.

Polansky JK, Schreiber L, Thelemann C et al. (2010) Methylation matters: Binding of Ets- 1 to the demethylated Foxp3 gene contributes to the stabilization of Foxp3 expression in regulatory T cells. Journal of Molecular Medicine, 88, 1029–1040.

Provençal N, Suderman MJ, Guillemin C et al. (2012) The signature of maternal rearing in the methylome in rhesus macaque prefrontal cortex and T cells. The Journal of Neuroscience, 32, 15626–15642.

129

Pryce CR, Rüedi-Bettschen D, Dettling AC et al. (2005) Long-term effects of early-life environmental manipulations in rodents and primates: Potential animal models in depression research. Neuroscience and biobehavioral reviews, 29, 649–74.

Pusey A, Packer C (1987) Dispersal and philopatry. In: Primate Societies (eds Smuts B, Cheney D, Seyfarth RM, Wrangham RW, Struhsaker T), pp. 150–166. University of Chicago Press, Chicago, IL.

Quinlan AR, Hall IM (2010) BEDTools: A flexible suite of utilities for comparing genomic features. Bioinformatics, 26, 841–842.

Reid JM, Bignal EM, Bignal S, McCracken DI, Monaghan P (2006) Spatial variation in demography and population growth rate: the importance of natal location. The Journal of Animal Ecology, 75, 1201–11.

Reinius LE, Acevedo N, Joerink M et al. (2012) Differential DNA methylation in purified human blood cells: implications for cell lineage and studies on disease susceptibility. PloS One, 7, e41361.

Revitali MANP, Ima MAL, Eserve PELM, Elt DOAK (2009) Population dynamics of two sympatric rodents in a variable environmental rainfall resource availability and predation.pdf. Ecology, 90, 1996–2006.

Reynolds LM, Taylor JR, Ding J et al. (2014) Age-related variations in the methylome associated with gene expression in human monocytes and T cells. Nature Communications, 5, 5366.

Rickard IJ, Lummaa V (2007) The predictive adaptive response and metabolic syndrome: challenges for the hypothesis. Trends in Endocrinology and Metabolism, 18, 94–9.

Rishi V, Bhattacharya P, Chatterjee R et al. (2010) CpG methylation of half-CRE sequences creates C/EBPalpha binding sites that activate some tissue-specific genes. Proceedings of the National Academy of Sciences of the United States of America, 107, 20311–6.

Rivenbark AG, Stolzenburg S, Beltran AS et al. (2012) Epigenetic reprogramming of cancer cells via targeted DNA methylation. Epigenetics, 7, 350–60.

Roadmap Epigenomics Consortium, Kundaje A, Meuleman W et al. (2015) Integrative analysis of 111 reference human epigenomes. Nature, 518, 317–330.

Rönn T, Volkov P, Davegårdh C et al. (2013) A six months exercise intervention influences the genome-wide DNA methylation pattern in human adipose tissue.

130

PLoS Genetics, 9, e1003572.

Roseboom T, de Rooij S, Painter R (2006) The Dutch famine and its long-term consequences for adult health. Early Human Development, 82, 485–91.

Russell AF, Lummaa V (2009) Maternal effects in cooperative breeders: from hymenopterans to humans. Philosophical transactions of the Royal Society of London. Series B, Biological sciences, 364, 1143–1167.

Sapolsky RM (2005) The influence of social hierarchy on primate health. Science, 308, 648–52.

Schulz LC (2010) The Dutch Hunger Winter and the developmental origins of health and disease. Proceedings of the National Academy of Sciences of the United States of America, 107, 16757–8.

Scuteri A, Sanna S, Chen WM et al. (2007) Genome-wide association scan shows genetic variants in the FTO gene are associated with obesity-related traits. PLoS Genetics, 3, 1200–1210.

Sheriff MJ, Love OP (2012) Determining the adaptive potential of maternal stress. Ecology Letters, 1–10.

Silk JB, Beehner JC, Bergman TJ et al. (2009) The benefits of social capital: close social bonds among female baboons enhance offspring survival. Proceedings of the Royal Society of London, Series B, 276, 3099–104.

Simola DF, Graham RJ, Brady CM et al. (2016) Epigenetic (re)programming of caste- specific behavior in the ant Camponotus floridanus. Science, 351, aac6633.

Sinclair K, Allegrucci C (2007) DNA methylation, insulin resistance, and blood pressure in offspring determined by maternal periconceptional B vitamin and methionine status. Proceedings of the National Academy of Sciences of the United States of America, 104, 19351–19356.

Smith ZD, Meissner A (2013) DNA methylation: roles in mammalian development. Nature Reviews Genetics, 14, 204–220.

Soranzo N, Spector TD, Mangino M et al. (2009) A genome-wide meta-analysis identifies 22 loci associated with eight hematological parameters in the HaemGen consortium. Nature Genetics, 41, 1182–1190.

Stanner S a, Yudkin JS (2001) Fetal programming and the Leningrad Siege study. Twin

131

research, 4, 287–292.

Stephens DC, Poon GMK (2016) Differential sensitivity to methylated DNA by ETS- family transcription factors is intrinsically encoded in their DNA-binding domains. Nucleic Acids Research, 44, 8671–8681.

Storey JD, Tibshirani R (2003) Statistical significance for genomewide studies. Proceedings of the National Academy of Sciences of the United States of America, 100, 9440–5.

Tavil Y, Sen N, Yazici HU et al. (2007) Mean platelet volume in patients with metabolic syndrome and its relationship with coronary artery disease. Thrombosis Research, 120, 245–250.

Team RDC (2012) R: A language and environment for statistical computing.

Tewhey R, Kotliar D, Park DS et al. (2016) Direct Identification of Hundreds of Expression-Modulating Variants using a Multiplexed Reporter Assay. Cell, 165, 1519–1529.

Thakore PI, Black JB, Hilton IB, Gersbach CA (2016) Editing the epigenome: technologies for programmable transcription and epigenetic modulation. Nature Methods, 13, 127–137.

Tobi EW, Goeman JJ, Monajemi R et al. (2014) DNA methylation signatures link prenatal famine exposure to growth and metabolism. Nature Communications, 5, 1–13.

Tobi EW, Lumey LH, Talens RP et al. (2009) DNA methylation differences after exposure to prenatal famine are common and timing- and sex-specific. Human Molecular Genetics, 18, 4046–53.

Tung J, Akinyi MY, Mutura S et al. (2011) Allele-specific gene expression in a wild nonhuman primate population. Molecular Ecology, 20, 725–39.

Tung J, Archie EA, Altmann J, Alberts S (2016) Cumulative early life adversity predicts longevity in wild baboons. Nature Communications, 7, 11181.

Tung J, Barreiro LB, Johnson ZP et al. (2012) Social environment is associated with gene regulatory variation in the rhesus macaque immune system. Proceedings of the National Academy of Sciences of the United States of America, 109, 6490–5.

Tung J, Gilad Y (2013) Social environmental effects on gene regulation. Cellular and molecular life sciences, 70, 4323–39.

132

Tung J, Primus A, Bouley AJ et al. (2009) Evolution of a malaria resistance gene in wild primates. Nature, 460, 388–91.

Tung J, Zhou X, Alberts SC, Stephens M, Gilad Y (2015) The genetic architecture of gene expression levels in wild baboons. eLife, 4, 1–22.

Uller T, Nakagawa S, English S (2013) Weak evidence for anticipatory parental effects in plants and animals. Journal of Evolutionary Biology, 26, 2161–70.

Umezawa A, Yamamoto H, Rhodes K et al. (1997) Methylation of an ETS site in the intron enhancer of the keratin 18 gene participates in tissue-specific repression. Molecular and Cellular Biology, 17, 4885–94.

Vanhille L, Griffon A, Maqbool MA et al. (2015) High-throughput and quantitative assessment of enhancer activity in mammals by CapStarr-seq. Nature communications, 6, 6905.

Verhoeven KJF, VonHoldt BM, Sork VL (2016) Epigenetics in ecology and evolution: what we know and what we need to know. Molecular Ecology, 25, 1631–1638.

Verhulst EC, Mateman AC, Zwier M V. et al. (2016) Evidence from pyrosequencing indicates that natural variation in animal personality is associated with DRD4 DNA methylation. Molecular Ecology, 25, 1801–1811.

Van den Veyver IB (2002) Genetic effects of methylation diets. Annual Review of Nutrition, 22, 255–82.

Vockley CM, D’Ippolito AM, McDowell IC et al. (2016) Direct GR Binding Sites Potentiate Clusters of TF Binding across the Human Genome. Cell, 166, 1269– 1281.e19.

Vockley CM, Guo C, Majoros WH et al. (2015) Massively parallel quantification of the regulatory effects of noncoding genetic variation in a human cohort. Genome Research, 25, 1206–1214.

Vojta A, Dobrinić P, Tadić V et al. (2016) Repurposing the CRISPR-Cas9 system for targeted DNA methylation. Nucleic acids research, 1–14.

Wang H, Maurano MT, Qu H et al. (2012) Widespread plasticity in CTCF occupancy linked to DNA methylation. Genome Research, 22, 1680–1688.

Waterland RA, Jirtle RL (2003) Transposable elements: targets for early nutritional effects on epigenetic gene regulation. Molecular and cellular biology, 23, 5293–300.

133

Waterland R a, Kellermayer R, Laritsky E et al. (2010) Season of conception in rural gambia affects DNA methylation at putative human metastable epialleles. PLoS Genetics, 6, e1001252.

Waterland R a, Michels KB (2007) Epigenetic epidemiology of the developmental origins hypothesis. Annual review of nutrition, 27, 363–88.

Weaver ICG, Cervoni N, Champagne F a et al. (2004) Epigenetic programming by maternal behavior. Nature Neuroscience, 7, 847–54.

Webb BA, Forouhar F, Szu F et al. (2015) Structures of human phosphofructokinase-1 and atomic basis of cancer-associatedmutations. Nature.

Weber M, Hellmann I, Stadler MB et al. (2007) Distribution, silencing potential and evolutionary impact of promoter DNA methylation in the human genome. Nature Genetics, 39, 457–66.

Wei G-H, Badis G, Berger MF et al. (2010) Genome-wide analysis of ETS-family DNA- binding in vitro and in vivo. The EMBO journal, 29, 2147–60.

Wells JCK (2007) Environmental quality, developmental plasticity and the thrifty phenotype: a review of evolutionary models. Evolutionary Bioinformatics Online, 3, 109–20.

Wells JCK (2012) A critical appraisal of the predictive adaptive response hypothesis. International Journal of Epidemiology, 41, 229–35.

Whitney O, Pfenning AR, Howard JT et al. (2014) Core and region-enriched networks of behaviorally regulated genes and the singing genome. Science, 27.

Whitten PL (1983) Diet and dominance among female vervet monkeys. American Journal of Primatology, 5, 139–159.

Wiench M, John S, Baek S et al. (2011) DNA methylation status predicts cell type-specific enhancer activity. The EMBO journal, 30, 3028–39.

Winnefeld M, Lyko F (2012) The aging epigenome: DNA methylation from the cradle to the grave. Genome biology, 13, 165.

Wolff GL, Kodell RL, Moore SR, Cooney C (1998) Maternal epigenetics and methyl supplements affect agouti gene expression in Avy/a mice. American Societies for Experimental Biology, 12, 949–57.

134

Xi Y, Li W (2009) BSMAP: whole genome bisulfite sequence MAPping program. BMC Bioinformatics, 10, 232.

Xie L, Korkmaz KS, Braun K, Bock J (2013) Early life stress-induced histone acetylations correlate with activation of the synaptic plasticity genes Arc and Egr1 in the mouse hippocampus. Journal of Neurochemistry, 125, 457–464.

Yamini S, Carswell N, Michaelis OE, Szepesi B (1992) Adaptation in enzyme (metabolic) pathways to obesity, carbohydrate diet and to the occurrence of NIDDM in male and female SHR/N-cp rats. International Journal of Obesity and Related Metabolic Disorders, 16, 765—774.

Yokomori N, Kobayashi R, Moore R, Sueyoshi T, Negishi M (1995) A DNA methylation site in the male-specific P450 (Cyp 2d-9) promoter and binding of the heteromeric transcription factor GABP. Molecular and Cellular Biology, 15, 5355–5362.

Zhang T-Y, Meaney MJ (2010) Epigenetics and the environmental regulation of the genome and its function. Annual review of psychology, 61, 439–66, C1-3.

Zhou X, Cain C, Myrthil M et al. (2014) Epigenetic Modifications are Associated with Inter-species Gene Expression Variation in Primates. Genome Biology, 15, 1–19.

Zhu H, Wang G, Qian J (2016) Transcription factors as readers and effectors of DNA methylation. Nature Reviews Genetics, 17, 551–65.

135

Biography

Amanda Jeanne Lea was born on January 27, 1987 in San Mateo, California. She completed her undergraduate Bachelor of Science degree in Ecology and Evolutionary

Biology from the University of California: Los Angeles in 2009. During graduate school,

Amanda received the following fellowships: a National Science Foundation Graduate

Research Fellowship, a James B. Duke Graduate Fellowship, a National Evolutionary

Synthesis Center and Triangle Center for Evolutionary Medicine Fellowship, a Duke

University Summer Research Fellowship, and a Duke University Biology Department

Fellowship. She has been the recipient of grant funding from the Duke University

Graduate School, the Biology and Evolutionary Anthropology departments of Duke

University, Sigma Xi, the National Science Foundation, and the Leakey Foundation.

She has also been an author on the following scientific articles: “Pervasive effects of aging on gene expression in wild wolves” (Molecular Biology and Evolution 33: 1967-

1978); “Resource base influences genome-wide DNA methylation levels” (Molecular

Ecology 25: 1681-1696); “A flexible, efficient binomial mixed model for identifying differential DNA methylation in bisulfite sequencing data” (PLoS Genetics 11: e1005650); “Genome-wide evidence reveals that African golden jackals are a distinct species” (Current Biology 25: 2158-2165); “Developmental constraints in a wild primate.” (The American Naturalist 185: 809-21); “Disease and urbanization drive genetic change in urban bobcat populations” (Evolutionary Applications 8: 75-92);

136

“Complex sources of variance in female dominance rank in a nepotistic society” (Animal

Behaviour 94: 87-99); “Beyond masking: endangered Stephen’s kangaroo rats respond to traffic noise with footdrumming” (Biological Conservation 150: 53-58); “Ontogenetic and sex differences influence alarm call response: a meta-analysis” (Ethology 117: 839-851);

“Heightened risk reveals state-dependent anti-predator responses in marmots”

(Behavioral Ecology and Sociobiology 65: 1525-1533); “Heritable victimization and the benefits of agonistic relationships” (Proceedings of the National Academy of Sciences

107: 21587-21592); “Heritability of anti-predator traits: vigilance and locomotor performance in marmots” (Journal of Evolutionary Biology 23: 879-887); and

“Heterospecific eavesdropping in a nonsocial species” (Behavioral Ecology 19: 1041-

1046).

137