1 the Rise of Neuroimaging Epigenetics: a Systematic Review of Studies Examining Associations Between DNA Methylation and Brain

Home , Imaging genetics, Psychiatric genetics

The rise of neuroimaging epigenetics:

A systematic review of studies examining associations between

DNA methylation and brain imaging

Esther Walton1,2, Vince Calhoun3, Bas Heijmans4, Paul M. Thompson5 & Charlotte A.M. Cecil4,6,7

1. Department of Psychology, University of Bath, Bath, United Kingdom 2. MRC Integrative Epidemiology Unit, Bristol Medical School, University of Bristol, United Kingdom 3. Tri-institutional Center for Translational Research in Neuroimaging and Data Science, Georgia State University, Georgia Institute of Technology, Emory University, Atlanta, United States 4. Molecular Epidemiology, Dept. of Biomedical Data Sciences, Leiden University Medical Centre, Leiden, the Netherlands 5. Imaging Genetics Center, Mark and Mary Stevens Neuroimaging & Informatics Institute, Keck School of Medicine of the University of Southern California, Marina del Rey, CA, USA 6. Department of Child and Adolescent Psychiatry/Psychology, Erasmus University Medical Center Rotterdam, the Netherlands 7. Department of Epidemiology, Erasmus University Medical Center Rotterdam, the Netherlands

Corresponding author: Dr Esther Walton, Department of Psychology, University of Bath, Claverton Down, Bath, BA2 7AY, [email protected], phone: +44 1225 38 6563

Acknowledgements: This work was funded from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 707404 to CC and grant agreement No 848158 (EarlyCause) to EW and CC. PMT is funded in part by NIH grants U54 EB020403, R01 MH116147, R56 AG058854, P41 EB015922, R01 MH111671 and a Zenith Grant from the Alzheimer’s Association.

Disclosures: PMT received grant support from Biogen, Inc. (Boston, USA) for research unrelated to the topic of this manuscript.

Abstract

Epigenetic mechanisms, such as DNA methylation (DNAm), have gained increasing attention in the field of neuroimaging as a potential biomarker of – or mechanism mediating – genetic and environmental influences on the brain. Yet, the extent to which DNAm associates with individual differences in the brain – the most relevant organ for the study of psychiatric disorders – is currently unclear.

We systematically reviewed research combining structural or functional neuroimaging measures with DNAm to provide an overview of the current state-of-the-art in this new field of research and discuss current challenges.

We identified 78 articles, published between 2011 – 2019. Most studies investigated DNAm-brain associations in the context of psychiatric and behavioural outcomes (76%), often based on adult (77%) or clinical samples (46%). Only a few studies focussed on risk exposures (21%), developmental periods other than adulthood (23%) or analyzed repeated measures of DNAm or neuroimaging (5%). Studies were highly heterogeneous in design (longitudinal versus cross-sectional), sample characteristics and methods used (candidate-driven versus genome-wide) with relatively few shared practices and common standards. Sample sizes were generally low to moderate (median n=99).

On the basis of the strength and weaknesses of existing studies, we recommend how best to address current challenges, including the need for collaborative science to increase comparability across studies. We also advocate for the use of large, prospective, paediatric cohorts with repeated measures of methylation and imaging to draw conclusions about the directionality of associations between these measures.

1 Introduction

Large-scale consortium efforts as well as advances in technology and methods have made it increasingly feasible to map the relationship between brain measures and a wide range of mental health outcomes in unprecedented detail. These neural correlates are shaped by both genetic and environmental influences: twin-based heritability estimates range between 60-80% for brain volume with slightly lower values for white matter integrity (14-64%), resting state functional connectivity (15-42%) and task-based functional MRI (40-65%).1,2 Although both genetic and environmental factors (e.g. stress, substance use) clearly influence the brain, we do not yet understand how these factors jointly shape brain structure and function, and downstream psychiatric risk.

Epigenetic mechanisms that regulate gene expression, such as DNA methylation (DNAm, Box 1), have gained increasing attention in the neuroimaging field as a potential biomarker of – or mechanism mediating – genetic and environmental influences on the brain. Yet, the extent to which peripheral DNAm, the focus of this review, associates with differences in the brain – the most relevant organ for the study of psychiatric disorders – is currently unclear. The nascent field of neuroimaging epigenetics (Box 1) aims to fill this gap by uniting scientists from different disciplines, such as psychiatry, neuroimaging and biology. Considering the emerging character of the field, there is a large degree of heterogeneity in methods used (e.g., candidate versus genome-wide), study designs (prospective versus cross-sectional), sample characteristics, developmental periods investigated and health outcomes examined, as well as conclusions drawn from these studies (e.g. in terms of directionality and causality).

Here, we systematically review emerging research combining structural or functional neuroimaging measures with DNAm. Our aim is three-fold: (i) to provide an overview of the current state-of-the-art in this new research field and highlight the most promising findings; (ii) to discuss current challenges; and (iii) to recommend how best to address these in future.

2 Methods

In this review, we focus on DNAm in relation to the following neuroimaging modalities: Structural magnetic resonance imaging (MRI), structural connectivity, functional MRI during a task or during rest (resting-state functional connectivity) and position emission tomography (PET; Box 1).

2.1 Search strategy

PubMed/MEDLINE and Embase databases were systematically searched to identify studies that investigated the relationship between DNAm and brain imaging measures. Abstract, titles and keywords were searched for the terms “methylation”, “epigen*”, “MRI”, “magnetic resonance imaging”, “PET”, and “position emission tomography”. We excluded animal studies and studies of cancer, using the exclusion terms “mouse”, “rodent”, “cancer”, “glioblastoma”. Studies had to be peer-reviewed journal articles in English and published or in press by March 11, 2019; no quality criteria with respect to the study were applied (e.g. study design, statistical power, data generation, data analysis) to obtain a complete overview of the current state of the field. For full search terms, see Supplementary Materials (SM), section 1.1. Two authors (CC and EW) independently screened the studies. Any ambiguities were discussed and resolved between the authors.

Box 1

Glossary: Common terms in the field of Neuroimaging Epigenetics

Neuroimaging Epigenetics: a field of research that aims to uncover the relationship between epigenetic patterns and the brain, and to establish the extent to which epigenetic marks in diverse tissues may enhance our ability to explain or trace genetic and environmental influences on brain and behaviour.

Epigenetics

DNA methylation (DNAm): one of the most commonly studied epigenetic mechanisms, affecting transcriptional activity via the addition of a methyl group to the cytosine base in the DNA (often in the context of a cytosine-guanine (CpG) sequence). It is modulated by both genetic and environmental factors, can be reversible and change dynamically with age (Horvath, 2013; Ramchandani, Bhattacharya, Cervoni, & Szyf, 1999). Furthermore, disruptions in methylation have been linked to a range of poor health outcomes (including all-cause mortality (Marioni et al., 2015)) and psychiatric disorders 55. Importantly, DNAm is highly tissue-specific, which poses a key challenge in the field of neuroimaging epigenetics 26,27,29.

Epigenetic age: a measure of biological ageing, derived from DNAm. Several of such ageing biomarkers, also referred to as 'epigenetic clocks', have been developed. For example, the 'Horvath clock' is based on 353 CpGs and can be applied to most tissues types 3, while the 'Hannum clock' is specific to blood tissue 4. Epigenetic age acceleration (EAA) is most commonly defined as the difference between an individual’s chronological and biological age (i.e. a difference > 0 suggesting an acceleration of biological ageing compared to one's chronological age).

Brain Imaging

Magnetic Resonance Imaging (MRI): a method that uses a strong magnetic field and radio waves to produce an image of the brain.

Structural MRI: a measure of brain morphology such as cortical thickness, surface area or volume of either the whole brain or individual regions of interest (ROIs).

Structural connectivity: based on diffusion MRI or derived measures such as fractional anisotropy, structural connectivity measures water diffusion in brain tissue, giving insights into the structural properties of white matter tracts and aspects of tissue microstructure.

Functional MRI: a measure of blood oxygen-level dependent signal in the brain to provide estimates of brain activity during a task or during rest (resting-state functional connectivity).

Position emission tomography (PET): a measure of regional metabolic activity, perfusion, or molecular composition in the brain, using radioactive tracers.

Box 1. Glossary of common terms used in neuroimaging epigenetics.

2.2 Study selection

Searches returned n=459 studies in Embase and n=491 studies in PubMed/MEDLINE. An additional 15 studies were identified through other sources (e.g., reference lists, Google Scholar alert), resulting in a total of n=604 studies to be screened after duplicates were removed. Study abstracts were then screened, after which n=492 studies were excluded. This left a total of n=112 full texts to be assessed. Of these, n=34 texts were excluded (for a PRISMA diagram, see SM Figure 1), resulting in a total of n=78 studies to be discussed in the current review.

2.3 Data extraction

We extracted the following data from studies that met our eligibility criteria: sample size, sample type (e.g., clinical, population-based, convenience sample), age group (neonates, children, adolescents, adults, elderly), study design (e.g., cross-sectional, longitudinal with respect to the methylation and/or neuroimaging measures), epigenetic approach (genome-wide [e.g., Illumina array focusing on up to 850,000 CpG sites] versus candidate-driven [e.g., candidate gene/s, epigenetic age acceleration (EAA) based on the Horvath3 or Hannum method4, see also Box 1]), neuroimaging approach (region-of-interest (ROI)-based versus whole brain or global measures; diffusion MRI, PET, resting state functional connectivity, functional MRI, structural MRI [e.g. measures of volume, thickness or surface]), tissue type, number of time points DNAm or neuroimaging data was collected, whether replication attempts in an independent cohort were made, whether mediation analyses were included, and whether additional biological markers, behavioural outcomes or exposures were measured.

3 Results

We identified 78 articles, published between 2011 – 2019. The number of articles per year increased steadily to 20 published articles in 2018 (Figure 1A). Across all years, the majority of studies focused on candidate epigenetic markers (including candidate genes and EAA), but the proportion of genome-wide studies increased from 0% to 30% in 2018. Most studies (76%) investigated DNAm-brain associations in the context of psychiatric and behavioural outcomes, such as depression (n=15), schizophrenia (n=13) or neurodegenerative or ageing-related traits (n=6; Table 1, right panel), but several publications were based on the same datasets. In contrast, only 21% of studies focused on environmental influences such as childhood adversity (n=7) or other forms of stress (n=6). Almost half of all studies (n=35) included genetic data in their analysis (of which four were based on a twin or family design), suggesting a large interest in the unique vs joint effects of genetic and epigenetic variation on brain-based phenotypes. For a full list and details of publications, see SM Table 1.

Figure 1. Research studies in the field of neuroimaging epigenetics. A) Number of published studies between 2011 and 2019. B) Sample sizes across studies.

3.1 Study design and sample characteristics

Sample characteristics

Although both DNAm and neuroimaging are high-dimensional data types, sample sizes across studies were on average modest, with a median sample size of 99 (range 14 – 713; Figure 1B). Sample size tended to increase continuously over the years (rho=0.18) both for candidate and genome-wide studies. The largest of the identified studies reported an association between EAA (based on the Horvath and the Hannum method) and white matter hyperintensities in a sample of 713 Adults, explaining up to 8% of the variance in this neural phenotype.5

Studies primarily focused on clinical samples (46%), followed by population-based cohorts (26%), high-risk groups (15%), convenience or community samples such as college students (8%), and twin and family samples (5%).

The majority of research investigated DNAm-brain associations in the context of psychiatric diseases that are most commonly diagnosed in adulthood. As such, only a minority of studies (23%) focused on developmental periods other than adulthood, e.g., birth (n=4), childhood (n=3), adolescence (n=7) and old age (n=4). Interestingly, two studies based on the same sample carried out post-mortem ex-vivo neuroimaging and brain-tissue DNAm in participants followed longitudinally during the last 20 years of their lifetime.6,7

Summary Table (N = 78)

General characteristics N % DNAm and brain characteristics N % Additional variables examined N % Design DNAm approach Genetic influences Cross-sectional 66 85% Candidate 59 76% No 43 55% Prospective/longitudinal 12 15% Genome-wide 22 28% Yes 35 45% Sample type DNAm tissue Environmental influences Clinical 36 46% Peripheral blood 60 77% No 62 79% Population/Cohort study 20 26% Saliva 16 21% Yes 16 21% High-risk 12 15% Cord blood/tissue 3 4% Of which: Other (e.g. convenience, community, etc) 6 8% Buccal 3 4% Childhood adversity 7 44% Genetically-informed (e.g. twin, pedigree) 4 5% Brain (postmortem) 3 4% Other stress-related exposure 6 38% Developmental period Most investigated genes (≥4 studies shown; n = 51) Biological markers Adulthood 60 77% SERT 13 17% No 65 83% Adolescence 7 9% OXT/R 8 10% Yes 13 17% Neonatal 4 5% FKBP5 7 9% Of which: Old age 4 5% NR3C1 4 5% mRNA/gene expression levels 10 77% Childhood 3 4% BDNF 4 5% Serum/plasma protein levels 5 38% Postmortem 1 1% Repeated measures 0% Neuroimaging measure Psychiatric and behavioral outcomes No 74 95% Brain morphology 35 45% No 19 24% Yes 4 5% fMRI 29 37% Yes 59 76% Of which: Diffusion MRI 12 15% Of which: Repeated DNAm 3 75% Functional connectivity 8 10% Depression 15 26% Repeated neuroimaging 2 50% PET 4 5% Schizophrenia 13 22% Neurodegenerative & Ageing-related 6 10% Replication Neuroimaging approach Post-traumatic stress disorder 6 10% No 72 92% Region of Interest (ROI) 43 55% Cognitive and social processes 5 9% Yes 6 8% Whole brain 36 46% Addictions 5 9% Global 9 12% Other 12 21%

Table 1. Summary of results. Note: some categories may add up to > 100% as certain studies met multiple criteria at once (e.g., studies of multiple tissues or multiple biological markers).

Design

Most studies – although sometimes longitudinal in their overall design - featured a cross- sectional set-up, with DNAm and neuroimaging measures obtained at the same time point (85%). Of those that measured these variables at different time points, most obtained DNAm first and then data on brain structure or function. Follow-up period ranged from weeks (most often in studies with neonatal samples8,9) to decades.10 Eight studies tested mediation in their analysis, mostly based on cross-sectional data. For instance, Sadeh et al. 11 reported statistical evidence for mediation of the association between PTSD and cortical thickness via SKA2 DNAm in blood, within a sample of 145 trauma-exposed veterans. Only one study investigated DNAm as a statistical mediator in the association between genetics and brain structure.12

Repeated measures (DNAm or MRI)

Four studies measured either DNAm or MRI (or both) repeatedly, and none more than twice.10,13–15 For instance, with respect to repeated measures of MRI, McMillan et al. 14 assessed grey matter density in 20 individuals, who screened positively for a C9orf72 repeat expansion, and followed up 11 of these participants one year later. The authors reported that C9orf72 promoter hypermethylation was linked to reduced longitudinal decline in grey matter regions, which they interpreted as a potential protective factor against grey matter loss and decline in verbal recall.

Swartz et al. 15 measured both DNAm in the SERT gene and fMRI twice over a 2-year period in 87 adolescents and reported that an increase in SERT DNAm over time was associated with an increase in amygdala reactivity to fearful faces during the same time period.

Only one study investigated the effect of repeated measures of DNAm on MRI at the genome-wide level.10 Using data from 109 young adults, the authors reported that DNAm change in the SP6 gene over a 7-year period from birth to childhood was prospectively linked to an increased amygdala:hippocampal volume ratio in early adulthood.

Replication

Six studies included a replication step in an independent cohort and none performed a meta-analysis of multiple cohorts. For instance, Freytag et al. 16 performed independent component analysis on 450k DNAm data in blood in a sample of 514 healthy young adults. One of these DNAm components, enriched for processes related to immune function and inflammation, was associated with cortical thickness after correcting for age. This finding was replicated in an independent sample of 423 patients with depression and 205 controls, suggesting that blood may be particularly useful in detecting inflammation-related brain changes.

Nikolova et al. 17 reported an effect of SERT promoter DNAm measured in saliva on amygdala reactivity in 80 young adults and replicated this finding not only in an independent cohort of 323 individuals, but also based on a different tissue (blood) and developmental period (11-15 year olds). Of note, the effect based on blood tissue in the replication study explained more variance (10.4%) than the effect in the discovery cohort (6.7%), which provides little evidence of a false positive finding due to the ‘winner’s curse’.

3.2 DNA methylation

Approach

The majority of studies (76%) focused on candidate epigenetic markers, including candidate genes, selected DNAm sites (ranging from n=1 to several hundred sites; sometimes followed up based on an initial epigenome-wide association study; EWAS) or measures of EAA (Box 1). EAA (in all studies residualized or controlled for chronological age) was investigated predominantly in studies interested in stress or age-related cognitive decline.5,16,18–21 Most studies reported a negative association between Horvath’s EAA and brain morphology, such as global white matter tract integrity 20 and left hippocampal volume.19 Negative associations were also reported between Hannum’s EAA and integrity of the corpus callosum measures via diffusion MRI.21 Positive associations were observed between mean diffusivity and global fractional anisotropy and Hannum’s EAA, but not Horvath’s EAA,18 and between white matter hyperintensities and both measures of EAA, although effects were slightly larger for Hannum’s EAA.5 No association was found between either EAA measure and cortical thickness.16 The heterogeneity in these findings reflect the current uncertainty in the field of epigenetic ageing overall.22

Studies that followed a candidate gene approach focused primarily on SERT (n=13), followed by OXT/R (n=8), FKBP5 (n=7), NR3C1 (n=4), and BDNF (n=4). Findings related to the three most commonly investigated genes (SERT, FKBP5, OXTR) were generally inconsistent. For details, see SM section 2.1. Of those studies following a genome-wide approach, most conducted a probe-level EWAS. However, in some of these studies, the number of DNAm sites were reduced by up to 90% before analysis, for example, based on the extent of DNAm variability.8,23 Other methodological approaches besides EWASes included sparse multiple canonical correlation analysis, weighted gene correlation network analysis (WGCNA), (parallel) independent component analysis (ICA), and gene- set enrichment analysis (SM Table 1).

Tissue

Blood was by far the most studied tissue (n=60), followed by saliva (n=16), buccal (n=3) and post-mortem brain tissue (n=3, including replication datasets). Very few studies measured DNAm in more than one tissue type. For instance, Ismaylova et al. 24 related SERT DNAm in blood, saliva and buccal samples to grey matter volume, functional connectivity and fMRI data. The authors reported a positive association between DNAm in all tissue types and frontal grey matter volume, no associations in any tissue with neural response to emotional stimuli, and links to functional connectivity for blood and buccal samples, but not for saliva. This is only partially in line with the study by Nikolova et al. 17 discussed above, which found an effect of SERT DNAm measured in blood, but not in saliva, on neural activity. Using parallel ICA on grey matter volumes and DNAm measured in blood and saliva, Lin et al. 12 were able to replicate DNAm association across both peripheral tissues with frontal, temporal and occipital regions, but not with cerebellar areas.

3.3 MRI

Measures

We observed a balance between studies focusing on brain morphology (45%) and functional activity during a task (37%), followed by diffusion MRI (15%) and functional connectivity (10%). We identified only 4 studies that used PET imaging. Of those studies with a focus on brain morphology, most examined volumetric measures of brain structure, followed by cortical thickness. Most functional MRI studies used tasks related to emotion processing (fear, stress, emotional faces) with only a few paradigms related to other processes, such as working memory, mentalizing or reward.

Approach

Akin to the predominant adoption of a candidate gene approach observed for DNAm, most MRI studies (55%) followed a region-of-interest (ROI) approach, primarily focusing on the hippocampus or the amygdala, or considered a single global measure of brain structure (e.g., total cortical thickness). This decision might have been driven by sample size considerations, as average sample size of studies following a ROI approach were slightly lower to those employing whole-brain analyses (median nROI =84 vs nwhole_brain=167). It also seemed that a ROI approach was favored in exchange to keep the dimensionality high on the DNAm side, as only one of all whole-brain studies also implemented an EWAS approach, compared to eight ROI-based studies. Of note, this one study 8 described a new method to account and test for unknown covariates and model assumptions in a high dimensional DNAm-MRI setting.

4 Discussion and Recommendations

We systematically reviewed research linking DNAm to neuroimaging measures. We identified a total of 78 studies – a number that has been steadily increasing since the first published study in 2011 – emphasizing the growing interest in this new field of research. In the following, we focus on three points: comparability across studies, cross-tissue considerations and directionality.

We highlight three recommendations to move the field of neuroimaging epigenetics forward, building upon previous recommendations in the field of epigenetics.25–29

4.1 Comparability across studies

Overall, the studies discussed here provide evidence for associations between peripheral DNAm and brain structure or function. However, we note that studies were very heterogeneous with respect to the design (e.g. longitudinal versus cross-sectional), sample characteristics (e.g. clinical, population-based), tissue (e.g. blood, saliva) and methodological and statistical approaches utilized (e.g. candidate gene versus genome-wide; diverging sets of confounders) with relatively few shared practices and common standards, which might be seen in other, more mature fields (e.g., psychiatric genetics). Sample sizes were on the whole moderate (median n=99) and only 8% of studies included a replication step in an independent cohort.

Recommendation 1: Push towards collaborative science

We recommend, where possible, to harmonize DNAm and MRI pre-processing pipelines,30 incorporate meta-analytical approaches31 or to replicate the original study findings.32 This will help to increase sample size, maximize statistical power to detect small effects and weed out false positives. For instance, researchers interested in combining DNAm and neuroimaging may draw on the information provided in this review to identify potential replication samples (see Supplementary Table 1). A recent successful example of collaborative science comes from the ENIGMA-Epigenetics consortium which combined 3,337 samples from 11 cohorts.31 The authors examined epigenome- wide associations with three subcortical volumes and identified two CpGs linked to the hippocampus, each explaining 0.9% of the phenotypic variance. These results point to small effect sizes in the relationship between individual peripheral DNAm markers and subcortical volumes, which might indicate limited value of single CpGs as biomarkers for brain structure. However, small effects might still be of mechanistic relevance, providing evidence for causality (see recommendation 3). At the same time, we should further develop the use of polyepigenetic scores, multivariate data-dimension reduction techniques such as parallel ICA and investigate in greater depth the relationship between global measures such as EAA and brain ageing.

4.2 Cross-tissue correlation

Cross-tissue heterogeneity is a major concern in psychiatric epigenetic research.28 Studies of cross-tissue correlation based on biopsy or post-mortem samples in general do not find large overall comparability between tissues, and only a small proportion of DNAm sites show cross-tissue correspondence.33,34 However, in light of the limitations associated with biopsy samples (e.g., tissues displaying disease pathologies) or post-mortem samples (e.g., changing DNAm patterns due to post- mortem intervals)35, ex vivo neuroimaging can help to bridge the gap between ante-mortem MRI and post-mortem brain tissue studies, while in vivo PET and MRI imaging epigenetics can help to connect peripheral DNAm to in vivo brain-based processes throughout a person’s life. Equally important are concerns of cross-tissue heterogeneity across different peripheral tissues.29 For example, Ismaylova et al. 24 reported that buccal, rather than saliva or blood-based SERT DNAm, associated with fronto-limbic brain processes. Di Sante et al. 13 reported saliva-buccal correlations for FKBP5, but not for NR3C1, which remained stable over a two-year period. However, DNAm (in either gene, tissue or time point) was not related to hippocampal volume. These findings point

towards gene-specific associations and highlight the need to investigate which peripheral tissue might be most valuable to measure DNAm in. The choice of tissue might also address difficulties related to reduced compliance and heightened attrition when blood samples are required (compared to buccal swabs), especially in clinical populations or longitudinal samples.

Recommendation 2: Incorporate knowledge on cross-tissue DNAm correspondence Researchers should include information on cross-tissue correspondence when reporting results, either directly by measuring DNAm in multiple tissues, or indirectly by using available online resources (e.g. blood-brain comparison tools: http://epigenetics.iop.kcl.ac.uk/bloodbrain/; https://redgar598.shinyapps.io/BECon/). However, we acknowledge as argued elsewhere,28,36 that cross-tissue correlations are neither necessary nor sufficient for meaningful results. For example, it is conceivable that high levels of DNAm perturbations in blood are needed to have effects on brain structure or function, while only low levels of perturbations are required in brain tissue, or that the perturbation is tissue-specific because each tissue will experience an environmental factor to a different extent in quantity and quality. Moreover, if the correlation is primarily genetically-driven, it will be less meaningful for the study and interpretation of environmental influences. This would result in a situation in which meaningful biological conclusions can be drawn based on results in either tissue, despite a lack of cross-tissue correlation. Conversely, confounding could lead to DNAm cross-tissue correlations, but the effect of an exposure on DNA methylation in each tissue could be different. To investigate the degree to which these scenarios are likely, researchers could weigh their analysis by the amount of evidence for cross-tissue correspondence, perhaps after accounting for (cis-)genetic effects.12 This is similar to approaches suggested for imaging genetics, whereby enrichment methods can be applied leveraging SNP effects on imaging traits by genomic location, linkage disequilibrium or association with complex human traits.37 If cross-tissue correspondence is identified as a meaningful feature, we should obtain DNAm profiles in tissues that show best correspondence with the brain. Saliva or buccal cells may outperform blood as a proxy of brain- based DNAm as they originate from the same ectoderm germ layer as the brain,24,38 but we need more research to investigate this in depth.

4.3 Directionality

The studies reviewed here broadly support an association between peripheral DNAm and brain structure or function. Yet, based on these findings, it is unclear whether DNAm markers present a cause or consequence of brain alterations or are due to other confounding variables. Disentangling the direction of effect is extremely difficult, as most imaging epigenetics research is based on cross-sectional, case-control samples. In line with a model proposed by Aberg et al.,39 peripheral epigenetic patterns might mirror those in brain tissue, possibly due to a common cause. Here, an exposure (such as smoking) might impact both peripheral DNAm as well as brain traits, resulting in a non-causal association between the two (“Proxy” model in Figure 2). However, brain pathology may have downstream effects on peripheral processes affecting DNAm. For instance, changes in hypothalamic structure or function may impair metabolic or hormonal processes, which then leave a signature in peripheral DNAm (“Signature” model in Figure 2). In this case, the DNAm signature would not be causal for brain-based phenotypes or related psychiatric traits, but a downstream consequence of them. Last, the reverse could also be envisaged whereby peripheral processes, such as increased inflammation or vascular events, could alter brain traits via changes in DNAm (“Mechanism” model in Figure 2). In all these scenarios, DNAm markers could be used as a biomarker of disease (or risk thereof), but only in the “Mechanism” model could we use peripheral DNAm as an interventional target, so long as the brain is the causal tissue for the disease. These

scenarios do not need to be mutually exclusive (e.g. associations might be bi-directional, where DNAm impacts brain function, which in turn affects DNAm) and animal studies so far provide independent evidence for all three scenarios.40–42 Causal inference methods such as Mendelian Randomization (MR) provide an opportunity to assess causality in human data,43,44 but to our knowledge no study to date has used MR to investigate causal links between DNAm and MRI phenotypes.

Figure 2. Possible directional links between methylomic patterns in peripheral and brain-tissue.

Recommendation 3: the use of prospective, pediatric cohorts with repeated measures of methylation and imaging

To disentangle directionality of effects, we need studies that collect repeated measures of DNAm and / or MRI data over time. In addition, relating DNAm to MRI patterns early in life could shed light on the role of epigenetic and brain variation in the onset and persistence of mental health disorders. This ideally requires (i) prospective data in young individuals beginning before the onset of symptoms; (ii) the availability of repeated measures of both DNAm and brain imaging at different developmental periods; and (iii) longitudinal follow-ups into adulthood. There are few cohorts worldwide with a study design that can address these questions (see Table 2 for a non-exhaustive list of cohorts). Using this set-up within the ALSPAC cohort, we were able to show, for example, that DNAm at birth is more predictive of later ADHD symptoms (measured repeatedly between ages 7 to 15) than DNAm measured at age 7.45 In relation to the brain, we could also show that the majority of DNAm sites at birth and age 7, are not stably predictive of amygdala:hippocampal volume, measured at age 18.10

To further strengthen causal inference or in cases where no such data is available, two- sample MR 46 can be applied, ideally using developmentally and tissue-specific genetic instruments (e.g. single nucleotide polymorphisms that are associated with early brain development). For example, using data derived from prefrontal cortex tissue, Hatcher et al. 47 applied MR and found that DNAm may putatively mediate effects of genetic variants on traits, such as schizophrenia. Moreover, it is important to keep an open mind as to the specific epigenetic mechanism involved. DNA methylation changes may well not be mechanistically causal themselves but mark other

regulatory processes like changes in transcription factor binding or in histone modifications. In this case, a focus on instrumenting DNAm would falsely lead to negative findings.

Cohort Age at Years of Repeated Repeated Sample Reference baseline follow-up measures of Measures of size* DNAm MRI (number of (number of time points) time points) Generation R Birth 14y Y (n = 3) Y (n = 3) Up to Kooijman et n=9778 al. 48

ALSPAC Birth 24y Y (n=7) N Up to Boyd et al. 49 15,458 IMAGEN 14y 8y Y (n = 2) Y (n = 3) Up to Schumann 2463 et al. 50

Brazilian High 6-14y 6y Y (n = 3) Y (n = 3) Up to Salum et al., Risk Cohort 2,512 51

FinnBrain Birth 5y N Y (n = 2) Up to Karlsson et 3,808 al. 52

GUSTO Birth 3y Y (n = 10) Y (n = 3) Up to Soh et al., 53 1,176 DCHS Birth 6y N Y (n = 3) Up to Donald et al. 1,143 54

Table 2. Prospective, paediatric cohorts with repeated measures of methylation and/or imaging * Sample sizes refer to whole cohort, not limited to subsample with DNAm or MRI.

5. Conclusion

Neuroimaging epigenetics constitutes a promising, growing field of research building on a recent drive towards more dimensional, multi-system approaches to mental health. To contribute substantially to advancements in mental health research, we advocate for i) larger samples sizes to account for likely small effects and the high-dimensionality of epigenetic and brain data; ii) the use of longitudinal population-based studies with repeated assessments to minimize sample bias and address the issue of directionality and iii) increased scientific rigor, including replication, cross-tissue considerations, sensitivity and specificity analyses.

References

1. Colclough, G. L. et al. The heritability of multi-modal connectivity in human brain activity. Elife 6,

e20178 (2017).

2. Jansen, A. G., Mous, S. E., White, T., Posthuma, D. & Polderman, T. J. C. What Twin Studies Tell

Us About the Heritability of Brain Development, Morphology, and Function: A Review.

Neuropsychol Rev 25, 27–46 (2015).

3. Horvath, S. DNA methylation age of human tissues and cell types. Genome Biol. 14, R115 (2013).

4. Hannum, G. et al. Genome-wide methylation profiles reveal quantitative views of human aging

rates. Mol. Cell 49, 359–367 (2013).

5. Raina, A. et al. Cerebral white matter hyperintensities on MRI and acceleration of epigenetic

aging: the atherosclerosis risk in communities study. Clin Epigenetics 9, 21 (2017).

6. Gaiteri, C. et al. Gene expression and DNA methylation are extensively coordinated with MRI-

based brain microstructural characteristics. Brain Imaging and Behavior (2018)

doi:10.1007/s11682-018-9910-4.

7. Yu, L. et al. Association Between Brain Gene Expression, DNA Methylation, and Alteration of Ex

Vivo Magnetic Resonance Imaging Transverse Relaxation in Late-Life Cognitive Decline. JAMA

Neurol 74, 1473–1480 (2017).

8. Guillaume, B. et al. Improving mass-univariate analysis of neuroimaging data by modelling

important unknown covariates: Application to Epigenome-Wide Association Studies.

Neuroimage 173, 57–71 (2018).

9. Ou, X., Thakali, K. M., Shankar, K., Andres, A. & Badger, T. M. Maternal adiposity negatively

influences infant brain white matter development: Maternal Obesity and Infant Brain. Obesity

23, 1047–1054 (2015).

10. Walton, E. et al. Longitudinal epigenetic predictors of amygdala:hippocampus volume ratio. J

Child Psychol Psychiatry 58, 1341–1350 (2017).

11. Sadeh, N. et al. SKA2 methylation is associated with decreased prefrontal cortical thickness and

greater PTSD severity among trauma-exposed veterans. Mol. Psychiatry 21, 357–363 (2016).

12. Lin, D. et al. Cross-Tissue Exploration of Genetic and Epigenetic Effects on Brain Gray Matter in

Schizophrenia. Schizophr Bull 44, 443–452 (2018).

13. Di Sante, J. et al. Peripheral DNA methylation of HPA axis-related genes in humans: Cross-tissue

convergence, two-year stability and behavioural and neural correlates.

Psychoneuroendocrinology 97, 196–205 (2018).

14. McMillan, C. T. et al. C9orf72 promoter hypermethylation is neuroprotective: Neuroimaging and

neuropathologic evidence. Neurology 84, 1622–1630 (2015).

15. Swartz, J. R., Hariri, A. R. & Williamson, D. E. An epigenetic mechanism links socioeconomic

status to changes in depression-related brain function in high-risk adolescents. Mol. Psychiatry

22, 209–214 (2017).

16. Freytag, V. et al. A peripheral epigenetic signature of immune system genes is linked to

neocortical thickness and memory. Nat Commun 8, 15193 (2017).

17. Nikolova, Y. S. et al. Beyond genotype: serotonin transporter epigenetic modification predicts

human brain function. Nat. Neurosci. 17, 1153–1155 (2014).

18. Chouliaras, L. et al. Peripheral DNA methylation, cognitive decline and brain aging: pilot findings

from the Whitehall II imaging study. Epigenomics 10, 585–595 (2018).

19. Davis, E. G. et al. Accelerated DNA methylation age in adolescent girls: associations with

elevated diurnal cortisol and reduced hippocampal volume. Transl Psychiatry 7, e1223 (2017).

20. Hodgson, K. et al. Epigenetic Age Acceleration Assessed with Human White-Matter Images. The

Journal of Neuroscience 37, 4735–4743 (2017).

21. Wolf, E. J. et al. Accelerated DNA methylation age: Associations with PTSD and neural integrity.

Psychoneuroendocrinology 63, 155–162 (2016).

22. Bell, C. G. et al. DNA methylation aging clocks: challenges and recommendations. Genome

Biology 20, 249 (2019).

23. Casey, K. F. et al. Birth weight discordance, DNA methylation, and cortical morphology of

adolescent monozygotic twins. Human Brain Mapping 38, 2037–2050 (2017).

24. Ismaylova, E. et al. Serotonin transporter gene promoter methylation in peripheral cells in

healthy adults: Neural correlates and tissue specificity. Eur Neuropsychopharmacol 27, 1032–

1041 (2017).

25. Birney, E., Smith, G. D. & Greally, J. M. Epigenome-wide Association Studies and the

Interpretation of Disease -Omics. PLoS Genet 12, (2016).

26. Lappalainen, T. & Greally, J. M. Associating cellular epigenetic models with human phenotypes.

Nature Reviews Genetics 18, 441–451 (2017).

27. Michels, K. B. et al. Recommendations for the design and analysis of epigenome-wide

association studies. Nature Methods 10, 949–955 (2013).

28. Mill, J. & Heijmans, B. T. From promises to practical strategies in epigenetic epidemiology.

Nature Reviews Genetics 14, 585–594 (2013).

29. Teschendorff, A. E. & Relton, C. L. Statistical and integrative system-level analysis of DNA

methylation data. Nature Reviews Genetics 19, 129–147 (2018).

30. Thompson, P. M. et al. ENIGMA and the individual: Predicting factors that affect the brain in 35

countries worldwide. NeuroImage 145, 389–408 (2017).

31. Jia, T. et al. Epigenome-wide meta-analysis of blood DNA methylation and its association with

subcortical volumes: findings from the ENIGMA Epigenetics Working Group. Molecular

Psychiatry (2019) doi:10.1038/s41380-019-0605-z.

32. van Rooij, J. et al. Evaluation of commonly used analysis strategies for epigenome- and

transcriptome-wide association studies through replication of large-scale population studies.

Genome Biology 20, 235 (2019).

33. Hannon, E., Lunnon, K., Schalkwyk, L. & Mill, J. Interindividual methylomic variation across

blood, cortex, and cerebellum: implications for epigenetic studies of neurological and

neuropsychiatric phenotypes. Epigenetics 10, 1024–1032 (2015).

34. Walton, E. et al. Correspondence of DNA Methylation Between Blood and Brain Tissue and Its

Application to Schizophrenia Research. Schizophr Bull 42, 406–414 (2016).

35. Bakulski, K. M., Halladay, A., Hu, V. W., Mill, J. & Fallin, M. D. Epigenetic Research in

Neuropsychiatric Disorders: the “Tissue Issue”. Curr Behav Neurosci Rep 3, 264–274 (2016).

36. Walton, E., Relton, C. L. & Caramaschi, D. Using Openly Accessible Resources to Strengthen

Causal Inference in Epigenetic Epidemiology of Neurodevelopment and Mental Health. Genes

(Basel) 10, (2019).

37. Fan, C. C. et al. Beyond heritability: improving discoverability in imaging genetics. Hum Mol

Genet 27, R22–R28 (2018).

38. Smith, A. K. et al. DNA Extracted From Saliva for Methylation Studies of Psychiatric Traits:

Evidence Tissue Specificity and Relatedness to Brain. Am J Med Genet B Neuropsychiatr Genet 0,

36–44 (2015).

39. Aberg, K. A. et al. Testing two models describing how methylome-wide studies in blood are

informative for psychiatric conditions. Epigenomics 5, 367–377 (2013).

40. Ewald, E. R. et al. Alterations in DNA methylation of Fkbp5 as a determinant of blood-brain

correlation of glucocorticoid exposure. Psychoneuroendocrinology 44, 112–122 (2014).

41. Meng, Q. et al. Traumatic Brain Injury Induces Genome-Wide Transcriptomic, Methylomic, and

Network Perturbations in Brain and Blood Predicting Neurological Disorders. EBioMedicine 16,

184 (2017).

42. Wang, J. et al. Epigenetic modulation of inflammation and synaptic plasticity promotes resilience

against stress in mice. Nature Communications 9, 477 (2018).

43. Davey Smith, G. & Hemani, G. Mendelian randomization: genetic anchors for causal inference in

epidemiological studies. Hum Mol Genet 23, R89–R98 (2014).

44. Dekkers, K. F. et al. Blood lipids influence DNA methylation in circulating cells. Genome Biology

17, 138 (2016).

45. Walton, E. et al. Epigenetic profiling of ADHD symptoms trajectories: a prospective, methylome-

wide study. Mol. Psychiatry 22, 250–256 (2017).

46. Hemani, G. et al. The MR-Base platform supports systematic causal inference across the human

phenome. Elife 7, (2018).

47. Hatcher, C., Relton, C. L., Gaunt, T. R. & Richardson, T. G. Leveraging brain cortex-derived

molecular data to elucidate epigenetic and transcriptomic drivers of complex traits and disease.

Translational Psychiatry 9, 1–10 (2019).

48. Kooijman, M. N. et al. The Generation R Study: design and cohort update 2017. Eur J Epidemiol

31, 1243–1264 (2016).

49. Boyd, A. et al. Data Resource Profile: The ALSPAC birth cohort as a platform to study the

relationship of environment and health and social factors. Int J Epidemiol (2019)

doi:10.1093/ije/dyz063.

50. Schumann, G. et al. The IMAGEN study: reinforcement-related behaviour in normal brain

function and psychopathology. Molecular Psychiatry 15, 1128 (2010).

51. Salum, G. A. et al. High risk cohort study for psychiatric disorders in childhood: rationale, design,

methods and preliminary results. Int J Methods Psychiatr Res 24, 58–73 (2015).

52. Karlsson, L. et al. Cohort Profile: The FinnBrain Birth Cohort Study (FinnBrain). Int J Epidemiol 47,

15–16j (2018).

53. Soh, S.-E. et al. Insights from the Growing Up in Singapore Towards Healthy Outcomes (GUSTO)

cohort study. Ann. Nutr. Metab. 64, 218–225 (2014).

54. Donald, K. A. et al. Drakenstein Child Health Study (DCHS): investigating determinants of early

child development and cognition. BMJ Paediatrics Open 2, e000282 (2018).

55. Barker, E. D., Walton, E. & Cecil, C. A. M. Annual Research Review: DNA methylation as a

mediator in the association between risk exposure and child and adolescent psychopathology. J

Child Psychol Psychiatry 59, 303–322 (2018).