DNA methylation and brain structure and function across the life course: A systematic review

MRI has enhanced our capacity to understand variations in brain structure and function conferred by the genome. We identified 60 studies that report associations between DNA methylation (DNAm) and human brain structure/function. Forty-three studies measured candidate loci DNAm; seventeen measured epigenome-wide DNAm. MRI features included region-of-interest and whole-brain structural, diffusion and functional imaging features. The studies report DNAm-MRI associations for: neurodevelopment and neurodevelopmental disorders; major depression and suicidality; alcohol use disorder; schizophrenia and psychosis; ageing, stroke, ataxia and neurodegeneration; post-traumatic stress disorder; and socio-emotional processing. Consistency between MRI features and differential DNAm is modest. Sources of bias: variable inclusion of comparator groups; different surrogate tissues used; variation in DNAm measurement methods; lack of control for genotype and cell-type composition; and variations in image processing. Knowledge of MRI features associated with differential DNAm may improve understanding of the role of DNAm in brain health and disease, but caution is required because conventions for linking DNAm and MRI data are not established, and clinical and methodological heterogeneity in existing literature is substantial.


Introduction
A growing number of studies have investigated associations between epigenetic signatures and neuroimaging markers of human disease, behaviour and cognition. This area of research is motivated by the understanding that epigenetic processes contribute to brain development, ageing and disease, and they may mediate interaction between genomic predisposition and environmental pressures that modify brain structure and function.
The term epigenetics refers to a set of molecular mechanisms that modulate the function of the genome in different cell types without altering the genome itself. DNA methylation (DNAm) is one such mechanism whereby a methyl group is covalently added to cytosine residues in a Cytosine-phosphate-Guanine context (CpG). This is thought to alter the accessibility of a locus to transcriptional machinery or modifying proteins, thereby influencing gene expression. DNA methylation may also occur after an initial change in gene expression and function as a form of longer-term control (Jones, 2012).
The pathogenesis of several genetic diseases with neurological phenotypes involve DNAm dysregulation. Rett syndrome is a severe neurological disorder characterised by regression of acquired skills, stereotypic movements, microcephaly, seizures, and intellectual disability. It results from a loss of function mutation in the X-linked methyl-CpG binding protein 2 gene (MECP2), which is a chromatin associated protein required for normal neuronal function throughout life. The loss of function mutation in MECP2 serves to reduce the binding affinity of the MECP2 protein to methylated DNA with effects on regulation of gene expression at transcriptional and post-transcriptional https://doi.org/10.1016/j.neubiorev.2020.03.007 Received 12 August 2019; Received in revised form 3 March 2020; Accepted 5 March 2020 levels (Lyst and Bird, 2015). Angelman syndrome is characterised by cognitive impairment, movement or balance disorder, typical abnormal behaviours, and impaired speech and language. It arises from a mutation or deletion of the maternal ubiquitin protein ligase 3A gene (UBE3A) and the neuronal tissue specific paternal imprinting of the gene that silences the paternal allele (Rangasamy et al., 2013). Fragile X syndrome arises due to CGG triplet expansion within the FMR1 gene promoter which becomes hypermethylated and results in a reduction of gene expression (Wijetunge et al., 2013). In addition to these genetic neurological disorders associated with DNAm dysregulation, variation in DNAm has been implicated in the pathogenesis of a number of complex neurodevelopmental, neurological and psychiatric diseases including autism, schizophrenia, and problems associated with trauma and stress (De Rubeis et al., 2014;Grayson and Guidotti, 2013;Klengel et al., 2014;Oberlander et al., 2008;Palumbo et al., 2018).
DNAm may be modifiable by environmental factors including physiological and emotional stress, child abuse, nutritional deprivation, and other lifestyle factors that operate from fetal life to old age (Cao-Lei et al., 2015;Del Blanco and Barco, 2018;Gouin et al., 2017;Heijmans et al., 2008;Joehanes et al., 2016;McGowan et al., 2009;Provenzi et al., 2018;Vidal et al., 2014). As such, it has been proposed that environmentally-induced changes in DNAm may enable short-term survival adaptation, but may also induce adaptations that contribute to impaired development of neural networks and increased risk of pathology (Gluckman et al., 2010;Hoffmann et al., 2017).
Structural, diffusion and functional brain MRI enable the parsing of complex behavioural traits and diseases onto quantitative indicators of brain structure and function. Such imaging endophenotypes have been used to demonstrate the impact of common genetic variants on brain health and disease (Elliott et al., 2018), but the role of DNAm in contributing to brain structure and / or function, as opposed to being a consequence of changes in structure or function, is less certain. To obtain a comprehensive overview of the extent to which differential DNAm associates human behaviour and disease we performed a systematic review of studies that have analysed DNAm with quantitative brain MRI data. By doing so we aimed to provide a comprehensive summary of what is known about DNAm-MRI associative relationships across the life course, and to identify image features associated with differential DNAm. Finally, we captured sources of heterogeneity in the extant literature to better inform the development of methods and conventions for analysing DNAm with MRI data.

Materials and methods
We performed a systematic literature search based on the PRISMA framework, according to a pre-registered protocol on PROSPERO (CRD42018090928).

Search strategy
Scopus, Web of Science, MEDLINE and EMBASE (via OVID) were searched in March 2018 to identify studies that integrated DNA methylation and human in vivo MRI (for search strategy see supplementary material). The search was composed of three OR clusters that contained terms relating to the key domains, which were combined with AND: MRI neuroimaging AND Brain AND DNA methylation. Bibliographies were searched for additional studies. There were no language restrictions on the search.

Screening and study selection
Search results were imported to EndNote X8 prior to removal of duplicates and retrieval of full texts prior to import to Covidence software (https://www.covidence.org/home). Screening was carried out by ENWW and DQS independently and studies were included if they met the following criteria: report of primary results; recruited human participants of any age with in vivo neuroimaging and DNAm; investigated the relationship between brain imaging using MRI and epigenome-wide or loci specific DNAm. Exclusion criteria: no original data reported (reviews, abstracts, letters, and grey literature as defined by Grey Literature International Committee Guidance); no test of association between brain imaging findings and DNA methylation; analysis was carried out in animals or in post-mortem human tissue; DNAm-MRI analyses of central nervous system malignancy.

Data extraction
ENWW extracted data for all included studies and DQS duplicated extraction for a subset of 25 %. Extraction included: participant characteristics, study design, method of DNAm ascertainment, MRI modality and image feature(s) of interest, and statistical method, consideration of cell sub-type composition includes cell sorting techniques or statistical methods to estimate and control for cell sub type composition and consideration of genotype or ancestry. The total number of participants per phenotypic category were estimated. Due to participant overlap in some studies, the largest study population reported was used to estimate the total number of participants.

Risk of bias assessment
In the absence of a validated quality assessment tool for studies linking DNAm with MRI data, we extracted data on study characteristics that might affect risk of bias: design, presence / absence of a comparator group; DNA source; candidate or epigenome-wide approach for evaluating differential DNAm; ascertainment of cell type composition; consideration of genotype; image processing methodology; and region-of-interest versus whole brain MRI analysis, and considered their prevalence in the epigenetic-neuroimaging literature.

Data synthesis
We provide a narrative synthesis structured around the type of association between DNAm and image feature, target population characteristics, tabulated, and categorised by disease. Fig. 1 is the PRISMA flow chart of study identification and inclusion. Our search strategy yielded 3442 in total: 3438 resulting from the search and an additional 4 papers were identified post search and were included. After removal of duplicates and exclusions at screening, 336 articles were eligible for full text evaluation. Of the full text articles that were screened for eligibility for inclusion, 276 were removed due to lack of eligibility. In total, 60 studies were included in this review (estimated 6775 participants): 34 (57 %) included a control or comparator group. Table 1 summarises diseases and conditions in phenotypic categories, number of participants per category, MRI modality, and method used to analyse DNAm (epigenome wide or candidate loci).

DNA methylation
All studies used surrogate tissues to probe methylation status: blood (n = 46); saliva (n = 12); one study reported using both blood and saliva, and one reported using blood/saliva/buccal cells (Ismaylova et al., 2017;Nikolova et al., 2014). The majority of studies did not report adjustments for cell type composition of samples used to estimate DNAm (n = 48).
Forty-three studies used a candidate gene analysis approach and 17 performed an epigenome-wide association study (EWAS). The majority of included studies that used a candidate gene approach measured methylation using bisulfite pyrosequencing (n = 36). Three studies measured methylation at a candidate gene using the EpiTYPER® system (Haas et al., 2016;Shelton et al., 2017Shelton et al., , 2016. One study carried out a restriction enzyme digest and sequencing to measure DNAm at a candidate site (McMillan et al., 2015). Three studies measured DNAm using Illumina 450k arrays, and used the output to inform candidate gene DNAm-MRI analyses Walton et al., 2014;Wiemerslage et al., 2017). Overall nineteen candidate loci were studied. SLC6A4 (Na+/Cl-Dependent Serotonin Transporter) was the gene of interest in 11 studies; BDNF (member of nerve growth factor family), FKBP5 (member of the immunophilin family) and OXTR (oxytocin receptor) were selected as candidate genes in four studies; NR3C1 (glucocorticoid receptor) in three studies; FMR1 (FMRP Translational Regulator 1), COMT (enzyme involved in catecholamine degradation) and DAT/SLC6A3 (Na + /K + dependent dopamine transporter) in two studies. The following genes were selected candidates in one study: OXT (encodes oxytocin and neurophysin 1 precursor protein), C9orf72 (thought to be a transcriptional regulator in the brain), CACNA1C (voltage gated calcium channel), DRD2 (dopamine receptor), KLF13 (transcription regulator), NCAPH2/LMF2 (encoding both a non-SMC subunit of the condensin II complex and a lipase maturation factor protein from the same locus), PPM1G (member of the PP2C family of Ser/Thr protein phosphatases), HTR3A (member of the ligand-gated ion channel receptor superfamily), SKA2 (Spindle And Kinetochore  E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 Associated Complex Subunit 2), SOD (superoxide dismutase) and TESC (tescalcin -regulates cell pH by controlling H+/Na + exchange across the plasma membrane).
Of studies which report DNAm at candidate genes, two reported that the sample used to measure DNAm was mononuclear blood cells (lymphocytes and monocytes) (Chouliaras et al., 2018;Ursini et al., 2011). None of the studies that measured DNAm by bisulfite pyrosequencing reported correction for cell type composition. 12 of the 20 studies which used Illumina arrays performed a correction for cell type composition. Only one study reported direct counting of cell types included in their samples (Freytag et al., 2017).

MRI modalities
The majority of studies used 3 T MRI scanners (n = 42); five studies used 1.5 T scanners; five included participants scanned at 1.5 T and 3 T scanners; and eight studies did not report field strength. Many included studies carried out functional MRI (n = 29), of which three studies also carried out structural analysis. Structural MRI, to investigate volumes, cortical morphology, and conventional clinical MRI measures, was used in 26 studies. Diffusion MRI to measure properties of white matter microstructure was used in nine studies, and MR angiography was used in one. Table 2 shows 10 MRI features that are associated with differential DNAm in two or more studies. The image features are: hippocampal volume (n = 11); hippocampal functional connectivity/activity (n = 4); amygdala functional connectivity/activity (n = 9); PFC functional connectivity/activity (n = 8); regional FA measures (n = 6); cortical thickness (n = 6); regional diffusion metrics (AxD/MD/RD) (n = 5); global FA (n = 2); and cortical volume (n = 2). For most of these, differential DNAm occurred at different loci. However, amygdala functional connectivity/activity was associated with differential DNAm in SLC6A4 in six studies, and OXTR in two studies. PFC connectivity/ activity was associated with DNAm in BDNF in two studies. Hippocampal volume was associated with FKBP5 DNAm in two studies. Activation within the prefrontal cortex in working memory tasks were associated with COMT DNAm in two studies. The studies all employed a candidate gene approach to DNAm analysis and directions of association were mixed across the studies.

MRI features associated with DNAm across studies
Six studies tested an association between SLC6A4 methylation and amygdala reactivity. Four out of five studies that reported associations between SLC6A4 methylation and the emotional face processing task assessed by fMRI in response to threat stimuli found positive associations with amygdala reactivity and connectivity (Ismaylova et al., 2018;Nikolova et al., 2014;Schneider et al., 2018;Swartz et al., 2017). One study investigated resting state functional connectivity and reported a positive association between SLC6A4 methylation and amygdala coupling with the salience network (Muehlhan et al., 2015). However, in a further study that investigated SLC6A4 methylation and brain activity in the context of both resting state connectivity and the emotional face processing task, there were no significant associations reported with regard to the amygdala either with regards to whole brain analysis or amygdala region-of-interest analysis (Ismaylova et al., 2017). The sixth study that reported an association between SLC6A4 and amygdala connectivity was also task-based fMRI, and used a visual emotional attention shifting paradigm . These papers assessed DNAm within the promoter region of the SLC6A4 gene, with the exception of Schneider et al., 2018, where a retrotransposon element, AluJb, was studied.
Two studies reported significant associations between OXTR methylation and amygdala connectivity (Puglia et al., 2015;Ziegler et al., 2015). Each study used a different fMRI task paradigm: the first used a task employing words involving social phobia relevant verbal stimuli in a cohort of patients with social anxiety disorder (Ziegler et al., 2015); and the second used an emotional face-matching block-design task in a group of healthy study participants (Puglia et al., 2015). Both studies measured DNAm in blood but each targeted a different region of the OXTR gene in bisulfite sequencing. Ziegler et al., 2015 targeted exon 3 of the OXTR gene due to previous work demonstrating an association between methylation in this region and social cognition (Unternaehrer et al., 2012;Ziegler et al., 2015). Puglia et al., 2015 targeted a single CpG site at position -934 relative to the transcription start site, where methylation had previously been associated with autism spectrum disorder (Gregory et al., 2009;Puglia et al., 2015). The directions of associations reported in these two studies were mixed, and may be explained by the variations in study design and participant characteristics. Two studies reported significant associations between BDNF methylation and PFC functional connectivity Ursini et al., 2016). Moser et al. performed DNAm-MRI analysis in a combined group of analysis of study participants with PTSD, sub-threshold PTSD and healthy control women, and assessed neural response to a Modified Crowell Procedure . The second study reported BDNF DNAm associations with PFC activity during a working memory task in a healthy group of study participants (Ursini et al., 2016). Both of these studies measured DNAm at different regions of the BDNF gene. Ursini et al., 2016 measured BDNF DNAm in blood at a region containing the rs6265 SNP, which lies in a coding region of the gene, which may interact with environmental factors to modulate risk for schizophrenia, while Moser et al., 2015 measured BDNF promoter DNAm in saliva, where methylation has been associated with trauma associated psychiatric disease in children Ursini et al., 2016). The directions of associations reported in these studies were mixed.
Two studies reported associations between FKBP5 methylation and hippocampal volume (Klengel et al., 2013;Resmini et al., 2016). In study participants with childhood trauma related PTSD there was a negative correlation between FKBP5 methylation in blood and right hippocampal head volume (Klengel et al., 2013). In patients with Cushing's syndrome there was a positive association between FKBP5 methylation in blood and bilateral hippocampal volume (Resmini et al., 2016). Both of these studies measured DNAm at glucocorticoid responsive regions of the FKBP5 gene, but report significant associations in different regions. Klengel et al., 2013 reported significant associations with intron 7, while Resmini et al., 2016 reports a significant association with a CpG site within intron 2 (Klengel et al., 2013;Resmini et al., 2016).
Two studies reported associations between COMT DNAm and PFC activation during working memory tasks (Ursini et al., 2011;Walton et al., 2014). The two studies assessed brain activity during different working memory tasks: the n-back task (Ursini et al., 2011); and the Sternberg Item Recognition task (Walton et al., 2014). The COMT gene encodes two isoforms of the catechol-O-methyltransferase enzyme: one soluble and one membrane bound, each with their own promoters, with the membrane bound isoform being the form most commonly found in the brain. Positive associations between MB-COMT promoter DNAm in blood and activity in the left dlPFC and vlPFC (Broca's area 45) during the Sternberg Item Recognition task were found in a group of participants including healthy controls and schizophrenia patients (Walton et al., 2014). In the second study of healthy participants significant negative associations were reported between methylation at COMT rs4680 SNP in blood and activity in the bilateral PFC (BA 45/14 and BA47) during the n-back task in participants homozygous for the Val/ Val allele at the rs4680 COMT SNP (Ursini et al., 2011). The difference in direction of associations reported by these two studies may be explained by the variations in study design and participant characteristics. Table 3 summarises eight studies that focused on neurodevelopment: typical development; the impact of early life stress on young adulthood; BDNF polymorphism in early life; ADHD; preterm brain injury; hippocampal sclerosis in mesial temporal lobe epilepsy; birthweight discordance between twins and its impact on cortical anatomy.  Of the 67 genes with intrapair probe DNAm which associated with volume discordance in the cerebellum, 49 were negatively associated. Of the 48 genes associated with the striatum, 38 were negatively associated, while of the 78 associated with the thalamus 44 were negatively associated.

Neurodevelopment and neurodevelopmental disorders
(continued on next page) E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 Four of the eight studies included a comparison group, and four were case series. Two studies report data from the Growing up towards Healthy Outcomes in Singapore (GUSTO) cohort study Guillaume et al., 2018). Two studies used a candidate gene approach and both reported positive associations, but between different MRI features and different genes: the first between DNAm at the SLC6A4 promoter and cortical thickness in an uncontrolled case series of 102 children with ADHD (Park et al., 2015); and the second between DNAm at FKBP5 introns and dlPFC activation during the go no-go task in a case series of young adults with a history of early life stress exposure (Harms et al., 2017). The remaining six studies used the Illumina 450k array to measure DNAm, and all reported significant associations between DNAm at multiple loci and features including fractional anisotropy (FA), subcortical ROI volumes, cortical thickness, volume and surface area (Casey et al., 2017;Chen et al., 2015Chen et al., , 2018Guillaume et al., 2018;Long et al., 2017;Sparrow et al., 2016). Two studies reported associations between DNAm at SMOC2 and MRI brain imaging features Guillaume et al., 2018). Guillaume et al., 2018 performed an EWAS in neonates and report an association between DNAm at this gene and FA in the left ventrolateral thalamus (Guillaume et al., 2018). Chen et al., 2018 reported differential methylation and discordance in twin pairs discordant for ADHD diagnosis, and identified six probes associated with the SMOC2 gene which were differentially methylated and were associated with differences in striatum (among 48 probes that were differentially methylated with regard to striatum volume) . SMOC2 encodes SPARC-related modular calcium binding protein 2.
Two studies reported an association between neonatal water diffusion properties and DNAm (Guillaume et al., 2018;Sparrow et al., 2016). In a case control study of preterm infants and term born controls principal component analysis of saliva DNAm identified a component that was associated with shape of the right corticospinal tract (Sparrow et al., 2016); and in neonates from the GUSTO cohort study there was an association between umbilical cord blood DNAm at the SMOC2 gene and FA in the left ventrolateral thalamus (Guillaume et al., 2018). Table 4 describes 10 studies of MDD and 1 of suicidality. Seven studies had overlapping participants: five have some overlapping participants from an outpatient psychiatric clinic of Korea University Anam Hospital and community controls (Choi et al., 2015;Han et al., 2017a, b;Na et al., 2016;Won et al., 2016); and two other studies had overlapping participants Frodl et al., 2015). Nine studies included a comparator group and 2 were case series of individuals with familial risk for MDD.

Major depressive disorder (MDD) and suicidality
Ten of eleven studies used a candidate gene approach based on blood samples and one carried out EWAS on salivary DNA. Five candidate genes were studied in the context of MDD: SLC6A4 (n = 5), BDNF (n = 2), TESC (n = 1), FKBP5 (n = 1) and CACNA1C (n = 1). The single EWAS study reported an association between DNAm age and hippocampal volume, noting that DNAm mediated the effect of cortisol on hippocampal volume (Davis et al., 2017).
All four fMRI studies were task based. Two used the emotional face processing paradigm to investigate amygdala reactivity in relation to the SLC6A4AluJb element or its promoter, and both report significant correlations between activation and SLC6A4 DNAm among cases (Schneider et al., 2018;Swartz et al., 2017). Another study reported a correlation between DNAm at the SLC6A4 and limbic system activation during a visual emotional attention shifting task in an analysis which pooled MDD cases and controls . Together these studies suggest that SLC6A4 methylation positively correlates with limbic system reactivity in response to threat or negative emotional stimuli. However, these studies measured and reported SLC6A4 methylation at different sites, and analysed them differently. For example, Frodl et al. used a summary mean measure calculated from 11 CpG sites  (continued on next page) E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156    (continued on next page) E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 in the promoter, Swartz et al. calculated a residualised change score from 20 CpG sites, and Schneider et al. calculated mean methylation at the AluJb element of the promoter from 6 CpG sites Schneider et al., 2018;Swartz et al., 2017). The fourth fMRI study investigated brain responses to images that were associated with suicidal means in relation to DNAm in the CACNA1C transcription factor binding site, and reported differential brain activation in association with DNAm at two CpGs (Kim et al., 2017). Table 5 summarises 4 studies of alcohol use disorder, all of which investigated DNAm in candidate genes in relation to task-based fMRI. Two were case-control study designs and two were uncontrolled case series. Three candidate genes were assessed: DRD2, SLC6A3/DAT and PPM1G. Three different tasks were used: testing impulse control, reactivity to alcohol cues and processing of reward/punishment cues. Three of these studies measured DNAm in blood, and one in saliva. Three of the four studies reported positive correlations between functional activation in limbic system structures and DNAm among participants with alcohol use disorder (Bidwell et al., 2018;Ruggeri et al., 2015;Wiers et al., 2015). The remaining study found no significant correlations in the alcohol dependent cases, but did find negative correlations in functional activation in the nucleus accumbens and methylation at the SLC6A3 promoter in the context of a monetary incentive delay task (Muench et al., 2018). Table 6 summarises seven studies of schizophrenia and psychosis, all of which included a comparator group. Six of the included studies use data from participants in the Mind Clinical Imaging Consortium (MCIC) all of whom have MRI, DNAm and SNP data available (Deng et al., 2016;Hass et al., 2015;Hu et al., 2018;Liu et al., 2015;Walton et al., 2014;Wang et al., 2017).

Schizophrenia and psychosis
The majority of these studies measured blood DNAm on the Illumina HumanMethylation27 array (n = 4), and one study carried out a candidate gene analysis on blood MB-COMT DNAm. Five of these used the Illumina HumanMethylation27 array, and two carried out bisulfite pyrosequencing of two genes: OXTR and MB-COMT. fMRI was the most commonly used modality (n = 5), followed by sMRI (n = 3). Of these, one study used both modalities. fMRI tasks included: a sensorimotor task and the Sternberg Item Recognition Paradigm for working memory (Hu et al., 2018;Walton et al., 2014). Two studies did not report the fMRI paradigm (Deng et al., 2016;Wang et al., 2017). Table 7 summarises eight studies that investigated aspects of ageing, including healthy ageing, Alzheimer's disease, minor cognitive impairment, stroke, WMH burden and C9orf72 expansion and Fragile X Associated Tremor Ataxia Syndrome (FXTAS).

Ageing, stroke, ataxia and neurodegeneration
Imaging modalities used were as follows: dMRI (n = 2), sMRI (n = 5) d/sMRI (n = 1). All studies used blood samples to measure DNAm. Five studies carried out candidate gene analysis and three measured DNAm in an epigenome-wide manner. The candidate genes studied were: C9orf72 (n = 1), NCAPH2/LMF2 (n = 1), FMR1 (n = 2), SOD (n = 1). Two studies carried out a longitudinal analysis as well as a crosssectional analysis (McMillan et al., 2015;Swartz et al., 2017). One of these showed, in a group of 11 C9orf72 expansion patients, that there was more rapid grey matter atrophy in the right hippocampus, right thalamus and left middle frontal cortex where there was decreased methylation in the expansion region (McMillan et al., 2015). The other found in 87 study participants that increases in SLC6A4 promoter methylation corresponded to increases in amygdala reactivity in response to threat related stimuli in task-based fMRI (Swartz et al., 2017). Two studies that investigated associations between FMR1 methylation and brain imaging in women with FXTAS had overlapping participants (Shelton et al., 2017(Shelton et al., , 2016. The three EWAS studies investigated   E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 DNAm age in relation to its associations with brain imaging features (Chouliaras et al., 2018;Hodgson et al., 2017;Raina et al., 2017). Two of these three studies investigated dMRI: one found a negative association of FA with epigenetic age acceleration, while the other found a negative association between global mean diffusivity (MD) and Hannum age and a positive association between FA and Hannum age (Chouliaras et al., 2018;Hodgson et al., 2017). In an analysis with sMRI a positive association was found between WMH burden category and DNAm age acceleration (Raina et al., 2017). Table 8 summarises seven studies that were carried out in participants with PTSD (Table 6). Several studies had overlapping participants. Three studies report associations in an overlapping group of participants: PTSD cases and non-PTSD controls, where the case population was mothers who have experienced interpersonal violence (IPV) Schechter et al., 2017Schechter et al., , 2015. Two further studies had overlapping participants -a cohort of trauma exposed veterans Wolf et al., 2016). An additional group of veterans was also studied in relation to PTSD (McNerney et al., 2018). A further study recruited women who had experienced early life trauma to investigate associations between DNAm and sMRI features (Klengel et al., 2013). Five candidate genes were studied: NR3C1 (n = 2), SKA2 (n = 1), BDNF (n = 1), FKBP5 (n = 1), 5HT3A (n = 1). All three modalities were employed to investigate neuroimaging features: dMRI (n = 1), sMRI (n = 3), fMRI (n = 3). Overlap in study populations lay between two studies which measured DNAm using the Illumina 450k array Wolf et al., 2016). Wolf et al. (2016) carried out analysis of DNAm age while Sadeh et al. (2016) carried out a candidate gene analysis of SKA2 Wolf et al., 2016). McNerney et al. (2018) measured DNAm at the NR3C1 gene using bisulfite pyrosequencing (McNerney et al., 2018). All three of the articles investigating a population of mothers who have experienced IPV studied candidate genes: NR3C1, BDNF and 5HT3A Schechter et al., 2017Schechter et al., , 2015. All three used fMRI, with two using the same task (Modified Crowell procedure), and one using a unique task video stimulus depicting neutral, menacing, and male-female prosocial interactions . The two studies that reported on IPV PTSD mothers in relation to the modified Crowell procedure showed positive associations between DNAm at NR3C1 and BDNF and cluster activation in vmPFC OFC posterior cingulate cortex Schechter et al., 2015).
The three studies that investigated associations with OXTR methylation measured DNAm at two different positions: one in exon 3 and the other two at a CpG site -934 base pairs upstream of the TSS which has consistent methylation across blood and brain tissues (Jack et al., 2012;  E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 (Pruessner et al., 2000) • Negative association between FKBP5 (mean at intron 7 bin 2) DNAm and right hippocampal head volume in a subset of the cohort (n = 34 (continued on next page) E.N.W. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 Puglia et al., 2015;Rubin et al., 2016;Ziegler et al., 2015). While these studies all used fMRI, they also used different tasks: social phobia related verbal stimuli, the emotional face matching task, and the social perception task (where objects move in ways to suggest animacy or random movement). One study investigated OXTR DNAm in exon 3, which had previously been implicated in social cognition, in a patient population with social anxiety disorder (Ziegler et al., 2015). The other two investigated healthy participants; one of these was in the context of social perception and individual variability in social perception (displays of animacy) (Jack et al., 2012;Puglia et al., 2015). Five papers analysed associations between SLC6A4 methylation and neuroimaging features. They included fMRI (n = 4) and sMRI (n = 2). fMRI studies used resting state and/or emotional face processing tasks. Ismaylova et al. (2017) estimated DNAm from three different surrogate tissues: blood, saliva, and buccal cells and reported that right lateral parietal area (RLP) resting state connectivity with the lateral occipital cortex and frontal poles has a positive association with SLC6A4 DNAm (Ismaylova et al., 2017). Muehlhan et al. (2015) found positive associations between functional coupling between amygdala and nodes of the salience network with DNAm of the SLC6A4 promoter in blood (Muehlhan et al., 2015). In the emotional face processing task, two studies report positive association between SLC6A4 methylation and amygdala reactivity/connectivity in response to threat/fear related faces. Nikolova et al. reported a positive relationship between saliva DNAm at SLC6A4 and left amygdala reactivity to threat stimuli, while Ismaylova et al. reported that twins with higher saliva DNAm at SLC6A4 display greater connectivity between ACC and left amygdala (Ismaylova et al., 2018;Nikolova et al., 2014). The two studies that used sMRI both found positive associations between DNAm at SLC6A4 and grey matter volumes (Dannlowski et al., 2014;Ismaylova et al., 2017). However, they did not replicate each other by identifying the same brain regions as being associated even when using the same tissue to estimate DNAm.

Discussion
This systematic review of 60 studies involving approximately 6775 participants suggests that differential DNAm may be associated with MRI features of brain structure and / or function for conditions / diseases within the following categories: neurodevelopment and neurodevelopmental disorders; MDD and suicidality; alcohol use disorder; schizophrenia and psychosis; ageing, stroke, ataxia and neurodegeneration; post-traumatic stress disorder; healthy adults and socio-emotional processing. We found that 10 image features were associated with differential DNAm in two or more of the 60 studies. However, quantitative synthesis of DNAm-MRI associations was not possible due to the lack of consistency in DNAm findings, and / or heterogeneity in image features across studies. While the results presented here could provide new insights into the role of DNAm in health and disease across the life course, caution is required because the clinical and methodological heterogeneity of included studies was large. We identified the following sources of methodological heterogeneity and potential sources of bias: variable inclusion and characterisation of comparator groups in study designs; use of different tissues as a surrogate for brain; variation in methods for estimating DNAm and ascertainment of cell type composition; lack of control for genotype; and variations in image processing methodology and selected MRI features.

Study populations and comparator group validity
Twenty-six studies (43 %) did not include a comparator group, and some others recruited cases and controls but combined the groups in DNAm-MRI analysis or restricted DNAm-MRI analysis to one group only. These issues limit inference about disease association. Several aspects of study design that are fundamental to good epidemiology should be adopted in future DNAm-MRI studies to reduce potential  • VBM -whole brain.
• VBM8/SPM8 • Positive association between AluJb DNAm in the SLC6A4 gene and bilateral hippocampal grey matter volume in ROI analysis in both samples. In sample 1 peak coordinates cluster extended into the bilateral amygdala. This was not the case in sample 2, but was the case in the combined sample.
• Whole brain analysis: positive association clusters where a positive association is also found: bilateral insula, right putamen, amygdala and caudate, and left superior occipital gyrus, bilateral hippocampus.
• No areas showed a negative association between DNAm rates and grey matter volume.
• Negative association between ICA2 and F6 (identified through exploratory factor analysis (EFA) of variation in cortical thicknesses); F6 accounts for 4% of variance in cortical thickness measures -spatial pattern comprising mainly temporal areas (loadings > 0.3) with highest loadings in L and R temporal poles and entorhinal cortices. Wheater, et al. Neuroscience and Biobehavioral Reviews 113 (2020) 133-156 sources of bias and enable maximum inference. These include detailed descriptions of participant selection, the population from which they were selected, and the method used to select. Not only is this important for understanding case definition, but also because research participation is influenced by social class, education, and ethnicity, and some of these factors may influence DNAm (Stafford et al., 2013). The same descriptive standards should be applied to comparator groups to ensure clinical validity, and attention should be given to matching comparator groups for ethnicity and age because both affect DNAm (Fraser et al., 2012;Heyn et al., 2013;Horvath, 2013;Talens et al., 2010).

Surrogate tissues for brain DNAm and cell composition
Surrogate tissues include blood and saliva buccal cell samples. This is important when interpreting data because of tissue and cell heterogeneity in DNAm patterns. Buccal cells, a major cell type found in saliva, have been proposed to have more validity than blood as a surrogate tissue for brain since they have a common embryological cell lineage, both being derived from the ectoderm germ layer (Smith et al., 2015). One study that ascertained DNAm from three sample types (blood, saliva, buccal -buccal cells and saliva samples are collected differently) found that DNAm obtained from non-blood surrogate tissues most strongly associated with brain processes in living humans in the context of fMRI study (Ismaylova et al., 2017). This is consistent with a previous study in postmortem brain tissue in animal models, which found that saliva and brain are more similar in their DNA methylation patterns than are blood and brain (Lowe et al., 2013). Surrogate tissues have heterogeneous cell compositions which can impact the DNAm signature of the sample: blood comprises lymphocytes, neutrophils, basophils, monocytes and eosinophils; and saliva contains predominantly buccal cells but can include leukocytes. Crucially, the compositions of both surrogates are influenced by disease status. Hence, adjustment for cell composition should be considered in DNAm analysis.

Candidate gene versus epigenome-wide approaches
Forty-three included studies in this review used a candidate gene analysis approach, with candidacy most often rooted in knowledge or hypotheses about the biological underpinnings of the target disease. The majority of candidate gene studies focused on promoter regions, which risks neglect of other important regulatory regions, because it limits research to currently understood genomic elements that have a clear role in gene expression regulation.
Although there are inherent sources of bias in the candidate approach, it has been suggested that candidate gene studies can provide a useful starting point for hypothesis construction for example by investigating candidates that have emerged from the GWAS literature (Nikolova and Hariri, 2015). However, we did not find that candidate selection was based on the GWAS literature in the majority of studies. For example, SLC6A4 was the candidate gene in eleven studies. SLC6A4 encodes a serotonin transporter responsible for serotonin re-uptake from the synapse, and is one of the targets for the serotonin reuptake inhibitor class of antidepressants. Despite the importance of this gene product in the pharmacology of depression, recent GWAS studies report no association between candidate polymorphisms at SLC6A4 (or a number of other candidate genes) and major depression (Border et al., 2019). This suggests that candidate genes that encode important pharmacological targets are not necessarily aetiologically significant. This is not to say that the methylation status of these genes is irrelevant: for example, it could be useful for investigating medication response, or as a biomarker. Frodl et al. carried out an DNAm-MRI analysis with medication status as a covariate, and but found that this did not modify their findings regarding SLC6A4 methylation and brain function in MDD patients and controls, while Booij et al. have reported differential methylation at the SLC6A4 promoter in association SSRI exposure Frodl et al., 2015). In a human population of MDD no difference was found in methylation between medication free subjects and those on medication in a post mortem study (Sabunciyan et al., 2012). These uncertainties suggest that further research into association between medication exposure and DNAm is warranted, and that hypothesis formation around candidate genes should be explicit in terms of theorised mechanism of disease or biomarker development. Further, mechanistic studies are crucial to aid our understanding of the interaction between DNAm and disease or therapy.
The rationales given for the selection of candidate genes for DNAm-MRI analysis were varied. Several studies selected candidates based on prior unbiased analyses in relation to phenotype (Kim et al., 2017;Ruggeri et al., 2015;Sadeh et al., 2016). For example, both PP1MG and SKA2 were identified in EWAS analyses (Ruggeri et al., 2015;Sadeh et al., 2016), and CACNA1C was identified from a hypothesis free Methyl-Seq (Kim et al., 2017). Other candidate genes were chosen because they have a strong genetic link to the patient population being studied, such as C9orf72 expansion in ALS-clinic patients and FMR1 in FXTAS patients (McMillan et al., 2015;Shelton et al., 2017Shelton et al., , 2016. For the most part candidate genes were selected because their gene products are considered key components of a pathway of relevance to a condition or phenotype studied. NR3C1 and FKBP5, the genes that encode the glucocorticoid receptor and FK506 binding protein (which regulates glucocorticoid receptor sensitivity), are both involved in cortisol signalling and were selected in studies investigating stress response or stress related neuropsychiatric conditions such as PTSD or MDD, or in Cushing's Syndrome where cortisol is known to be dysregulated. BDNF, which encodes for brain derived neurotrophic factor, was studied in association to MDD, PTSD and working memory. COMT encodes catechol-O-methyltransferase, an enzyme that degrades catecholamines such as dopamine, epinephrine and norepinephrine, was the candidate gene in studies investigating DNAm-MRI in the context of working memory in schizophrenia and in healthy participants. OXTR encodes the receptor for oxytocin and was selected in studies that investigated an aspect of emotional or social processing. SLC6A4 encodes the serotonin reuptake transporter and is a target for the serotonin reuptake inhibitor class of anti-depressants. It was the selected candidate gene in studies of MDD and ADHD patients and emotional processing, and resting state amygdala activity.
Seventeen studies used an EWAS approach. There was no overlap between differentially methylated regions identified using this method versus the selected candidate gene approach for any of the phenotypic categories, which highlights the value of epigenome-wide approaches for unbiased investigation of the putative role of DNAm in brain structure and / or function. However, one study which employed an epigenome-wide approach to DNAm analysis identified COMT and SLC6A3 as being negatively associated with right amygdala volume in neonates homozygous for BDNF rs6265 Met/Met allele . Both of these genes were selected as candidates in other studies but associations were reported in different populations with regards to age and clinical background, and with different image features (Muench et al., 2018;Ursini et al., 2011;Walton et al., 2014;Wiers et al., 2015). Many of the studies that utilised epigenome wide methods for measuring DNAm did not go on to report on individual loci. In future studies whole genome-based approaches could play an important role in generating novel candidates and pathways for future research that could provide mechanistic insight into pathogenesis.

Interindividual variability, effect of genotype, and temporal stability
Sources of variability that are crucial for understanding of how changes in DNAm associate with endophenotypes include interindividual variability which may be stochastic, genetic or environmental in origin, and temporal stability (Lancaster et al., 2018). Twin studies are useful for separating genetic and environmental contributions to variability, and several are included in the DNAm-MRI literature to date (Casey et al., 2017;Chen et al., 2018;Ismaylova et al., 2018). Twenty-four studies controlled for genotypic effects, for example by covarying for single nucleotide polymorphisms (SNPs), or for the first PC of the variance in genome wide genetic data. This is because variation in DNAm may be explained partly by common genetic variation, such as a SNP that removes a methylation site. Therefore, genotyping in DNAm-MRI analyses is necessary for determining whether epigenetic associations with brain structure and function are genotypespecific, or whether associations are independent of genotype.
The stability of some epialleles varies over time with dynamic methylation and demethylation at CpG sites occurring in typical development and ageing, and in response to environmental exposures (Lancaster et al., 2018). Only fifteen studies reported the timing of DNAm sampling in respect to MRI acquisition; we suggest that this information is included in future studies because it is required to compare study populations and interpret phenotypic associations. Longitudinal investigation could be used to establish the temporal stability of methylation at a genomic region or site, the impact of medication, and for assessing potential causal relationships.
All of the included studies carried out cross-sectional analysis and two additionally carried out longitudinal analysis. Three other studies performed mediation analysis to assess causal relationships (Casey et al., 2017;Freytag et al., 2017;Harms et al., 2017). Studies designed to investigate association or causation should report the magnitude of differential DNAm between groups since the described changes in DNAm are often small and the biological relevance of these remains uncertain. Ultimately however, differential DNAm-MRI data will need to be interpreted in the wider context of mechanistic studies of gene expression in order to make inferences about causality. For example, experiments that demonstrate changes in gene expression associated with the differential methylation patterns reported from DNAm-MRI studies would be an important first step in dissecting causality, although it is important to note that changes in DNAm may actually occur as a consequence of changes in gene expression.

Neuroimaging considerations
Structural and functional quantitative neuroimaging features have been included in this review, assessed using sMRI, fMRI and dMRI. Others have emphasised the importance of selecting image features that represent a deficient or pathological phenotype when investigating associations between DNAm and MRI (Lancaster et al., 2018). Five DNAm-MRI associations recurred in two or more studies (SLC6A4 and amygdala reactivity; OXTR and amygdala reactivity; BDNF and PFC function; FKBP5 and hippocampal volume; COMT and PFC function), and a narrative synthesis of these demonstrates that there were mixed directions of associations, which may be explained by the variation in where in a genomic region DNAm was measured and in which tissue, in tasks employed during fMRI and in participant characteristics. A variety of fMRI paradigms have been used to probe brain function in this field, such as resting state and emotional processing, and dMRI studies have used different diffusion gradient encoding schemes and strengths (bvalues) to measure the mobility of water molecules in vivo. A further layer of complexity is added in terms of differences in image processing pipelines such as CIVET (http://www.bic.mni.mcgill.ca/ ServicesSoftware/CIVET), FreeSurfer (https://surfer.nmr.mgh.harvard. edu) and SPM (https://www.fil.ion.ucl.ac.uk/spm) for sMRI; and voxel based methods e.g. TBSS (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/TBSS) or tractography methods e.g. probabilistic neighbourhood tractography (http://www.tractor-mri.org.uk/tractography) for analysis of dMRI (Ad-Dab'bagh et al., 2006;Anblagan et al., 2015;Bastin et al., 2010;Clayden et al., 2007Clayden et al., , 2006Fischl, 2012;Smith et al., 2006). In the future, it will be necessary to form consensus and harmonisation about the optimal acquisition protocols and analysis pipelines to allow findings from different groups to be combined more readily. Initiatives such as the ENIGMA consortium which aims to understand brain structure, function and disease using brain imaging and genetic data from groups around the world are an important step in this direction.

Strengths and limitations
This systematic review has several strengths. It was based on a predefined protocol and followed standard guidelines with rigorous screening of > 3,400 articles and without language or publication year restrictions. Our analysis was based on 60 studies examining healthy individuals and a wide range of neurological and psychiatric diseases in more than 6,000 patients across the life course. This provided a detailed overview of a growing body of literature linking DNAm and MRI data types, and enabled us to identify the key sources of methodological variation that should be addressed as standards and conventions in DNAm-MRI analyses become established.
Our study also has limitations. First, there was substantial clinical and methodological heterogeneity which prohibited quantitative synthesis for any single phenotype category, or MRI feature. Second, in the absence of validated tools for assessing quality and risk of bias in DNAm-MRI analyses, we were unable to evaluate these issues quantitively.

Conclusion
There is some evidence of association between differential DNAm and human brain structure and / or function across the life course. However, varied methodologies used to acquire and analyse DNAm-MRI data prevent quantitative synthesis. The development of standards and conventions for studies linking DNAm with MRI data is required, with particular focus on: detailed case and comparator group definition; surrogate tissue type; adjustment for cell composition and common genetic variation; consistent approaches to DNAm estimation using whole genome approaches; and use of image acquisition protocols, analysis pipelines and image feature selection that best support pooled analyses that are likely to be required to achieve statistical power for EWAS-neuroimaging studies. In summary, this review has found that: differential DNAm is associated with image features of brain structure and function in health and disease across the life course using data from over 6,000 individuals; evidence that differential DNAm is associated with specific image features is modest; and it has identified specific sources of sample and methodological heterogeneity in existing DNAm-MRI analyses. We anticipate that identification of these will expedite rational development of analytic methods in this emerging field so that researchers might more rapidly design studies that support causal inference.