Epigenetics applied to child and adolescent mental health: Progress, challenges and opportunities

Abstract Background Epigenetic processes are fast emerging as a promising molecular system in the search for both biomarkers and mechanisms underlying human health and disease risk, including psychopathology. Methods In this review, we discuss the application of epigenetics (specifically DNA methylation) to research in child and adolescent mental health, with a focus on the use of developmentally sensitive datasets, such as prospective, population‐based cohorts. We look back at lessons learned to date, highlight current developments in the field and areas of priority for future research. We also reflect on why epigenetic research on child and adolescent mental health currently lags behind other areas of epigenetic research and what we can do to overcome existing barriers. Results To move the field forward, we advocate for the need of large‐scale, harmonized, collaborative efforts that explicitly account for the time‐varying nature of epigenetic and mental health data across development. Conclusion We conclude with a perspective on what the future may hold in terms of translational applications as more robust signals emerge from epigenetic research on child and adolescent mental health.


INTRODUCTION
Half of mental illnesses are established before the age of 18 years, often manifesting first in childhood as emotional, behavioural and neurodevelopmental problems (Solmi et al., 2022). This points to early life as a critical window of opportunity for timely detection, prevention and intervention. Although numerous pre-and postnatal risk factors have already been identified (e.g., parental psychopathology, socio-economic hardship, childhood adversities; Barker et al., 2018), associations with child and adolescent mental health outcomes are far from straightforward, with equifinality (i.e., multiple risks associating with the same outcome) and multifinality (the same tively easy and cost-effective to quantify on a large scale. DNAm involves the addition of methyl molecules to DNA base pairs, typically in the context of cytosine-guanine (CpG) dinucleotides. Studies have shown that DNAm: (1) is partly under genetic control (Min et al., 2021); (2) is sensitive to environmental influences beginning in utero (e.g., dietary, chemical and psychosocial exposures (Cowley et al., 2018;González-Peña et al., 2021;Rijlaarsdam et al., 2017)); (3) is temporally dynamic, playing an essential role in (neuro)development ; and that (4) aberrations in DNAm associate with a wide range of health outcomes, including psychiatric disorders (Liu et al., 2018). As a result, DNAm has gained much interest in the search for both biomarkers and mechanisms underlying GED interplay on psychopathology. In this review, we discuss some of the complexities and unique opportunities of studying DNAm in the context of child and adolescent mental health.
We provide readers with a view on major lessons learned, current developments, and emerging topics in this rapidly growing field.
Considering how understudied this area is relative to, for example, epigenetics applied to adult health, we borrow some of the concepts and findings from other research areas, reflecting on potential implications for child and adolescent mental health. We conclude with recommendations for moving the field forward and a perspective on what the future may hold in terms of translational applications for research and clinical practice in child and adolescent mental health.

LOOKING BACK: LESSONS LEARNED
Human epigenetic research has seen tremendous growth over recent years. Here, we describe three key lessons that we have learned from this research, and what implications they have for the application of DNAm to child and adolescent mental health (Figure 1.1).

Ways in which DNAm does and does not vary
One of the main lessons we have learned from existing research is that DNAm is a highly dynamic process, and that different factors contribute to this variability. Topping the list of factors is arguably tissue and cell-type: for most DNAm sites, cross-tissue variability exceeds that of inter-individual variability within the same tissue (Hannon et al., 2015). This is particularly consequential for the study of brain-based phenotypes (e.g., mental health traits), given that DNAm measured from easily accessible tissues (e.g., blood) may show limited correspondence to those in the target tissue of interest (i.e., brain; Bakulski et al., 2016). Even within the brain, DNAm patterns can vary between different regions and cell-types (Edgar et al., 2017;Rizzardi et al., 2019). Another important factor is time: in a recent study pooling longitudinal, epigenome-wide data from over 2000 individuals, we found that more than half of DNAm sites change significantly during the first two decades of life, and can do so in a non-linear way (i.e., at different rates across development ;Mulder et al., 2020). By comparison, studies indicate only modest variability across individuals for many measured DNAm sites, leading to a continuing debate about the usefulness of including seemingly nonvariable sites in analyses. For example, DNAm patterns in many regions remain static (e.g., those related to cell-type differentiation and identity), and it is estimated that only about 20% of all methylation sites are variable (Ziller et al., 2013).
How is this shaping where we are going?

Many associations, few replicated
One of the strongest and most robust associations to emerge from population-based epigenetic studies is that of DNAm with smoking exposure. The largest epigenome-wide association study (EWAS) Peripheral DNAm: Growing support as a biological marker; mixed evidence as a mediator Early interest in DNAm in the context of mental health was largely focussed on its potential role as a mediator of environmental (and more recently also genetic) influences on psychiatric outcomes.
However, increased awareness of the 'tissue issue' has cast more doubt on the biological plausibility of peripheral DNAm as a mechanism underlying psychiatric risk. While experimental models have evidenced multiple ways in which DNAm in blood and brain tissue can be linked , it remains difficult to evaluate these models in humans. Soberingly, a recent large-scale study applying Mendelian randomization found limited evidence for a causal role of blood-based DNAm in neuropsychiatric disorders (Min et al., 2021).
Although cross-tissue variability makes mechanistic discoveries challenging, it does not undercut the potential of DNAm as a biological marker for disease prediction, stratification and diagnosis. Indeed, peripheral DNAm patterns are already being used to estimate a range of exposures, traits and health outcomes (e.g., age, smoking, BMI; McCartney et al., 2018) based on algorithms trained from large datasets, and to detect certain diseases sooner and more accurately than conventional diagnostic methods, leading to improved clinical care (Chen, Zang, et al., 2020;Priesterbach-Ackley et al., 2020). These applications have been slower to permeate the field of mental health, likely due to the more limited availability of sufficiently powered datasets and challenges with psychiatric phenotypes, such as heterogeneity in clinical presentation and assessment approaches. Nevertheless, the prospect of methylation-based profiling of neurodevelop-tnqh_9;mental and psychiatric conditions appears increasingly possible.
How is this shaping where we are going?

CURRENT DEVELOPMENTS IN THE FIELD: HIGHLIGHTS
In this section, we highlight research areas that are gaining increasing traction and lending new insights into the relationship between DNAm and (mental) health. Where possible, we refer specifically to findings on child and adolescent psychopathology, but also note research in adjacent fields that could be applied within a developmental context in the future (Figure 1.2).

Epigenetic timing effects on neurodevelopmental outcomes
While still rare, the increased availability of birth cohorts with repeated epigenetic data in the same individuals has recently made it possible to explore key developmental aspects of the relationship be- that DNAm profiles at birth (cord blood) associate more strongly with certain neurodevelopmental problems, particularly ADHD symptoms, than DNAm measured cross-sectionally during childhood (whole blood)-a 'timing effect' initially observed in single cohorts  and recently confirmed via multi-cohort meta-analysis . Top DNAm sites at birth implicate, among others, genes involved in neural functions (e.g., myelination, neurotransmitter release). The most notable example is ST3GAL3: common variation in this gene has also been identified as a top GWAS hit for ADHD Klein et al., 2019), rare mutations of ST3GAL3 associate with cognitive and motor developmental delays (Khamirani et al., 2021), and ST3GAL3 knockout in mice results in profound cognitive deficits and hyperactivity due to myelination disruption (Rivero et al., 2021). Similar epigenetic timing effects (i.e., where prospective associations at birth show overall a stronger signal in EWAS results than cross-sectional associations in childhood) have also been observed for other neurodevelopmental phenotypes (e.g., social communication deficits (Rijlaarsdam et al., 2021)), but not for broader child mental (e.g., general psychopathology (Rijlaarsdam et al., 2022), sleep problems (Sammallahti et al., 2022)) or physical (e.g., BMI; Vehmeijer et al., 2020) health outcomes, despite studies using largely overlapping data, which points to a degree of phenotypic specificity. A detailed overview of epigenetic timing effects in the context of neurodevelopmental conditions such as ADHD, outstanding questions and research priorities in this area can be found elsewhere (Cecil & Nigg, 2022).
Why is the discovery of epigenetic timing effects meaningful? In addition to highlighting the dynamic nature of associations between DNAm and neurodevelopmental outcomes, this finding has two major implications: (1) it supports the potential of DNAm as an early pre-symptomatic marker of neurodevelopmental risk, and (2) it suggests that, to benefit from this potential marker, the timing of DNAm assessment could be crucial-the DNAm risk signal captured at birth may no longer be detectable when DNAm is measured later in life (Walton, 2019). Timing effects may also explain some of the seemingly inconsistent findings in the literature, as studies have sampled DNAm at widely different ages. Advanced approaches capable of handling large-scale epigenetic data at repeated time points (e.g., structured life-course modelling, structural equation modelling, and time-course analyses (Brown et al., 2020;Dunn et al., 2019;Hill et al., 2019;Simons et al., 2017)) are needed to further characterize and disentangle these timing effects, although separating true temporal signals from technical sources of variation in longitudinal data (e.g., batch effects) will be challenging. In future, studies will also need to establish whether a signal similar to the one observed in cord blood could be obtained from other neonatal peripheral tissues, such as neonatal blood spots, which are routinely collected from heel pricks during the first week of life in many countries and are already widely used for screening and diagnostic purposes. Ultimately, epigenetic signals at birth could inform strategies for improved early risk detection (e.g., by integrating DNAm markers in multi-modal assessment tools including other known risk factors), and shed light on biological correlates underlying neurodevelopmental risk.
Differentiating between biological and chronological age: Epigenetic clocks number of 'epigenetic clocks' have been developed that can predict chronological age (Horvath & Raj, 2018), the pace of ageing (Belsky et al., 2020) or declining health and mortality (Lu et al., 2019). In adults, epigenetic age acceleration (i.e., residual or differences scores, where DNAm-estimated age outpaces chronological age) has been associated with a myriad of factors, including sociodemographic characteristics (e.g., male sex, low socio-economic status), unhealthy behaviours (e.g., smoking, alcohol use), poor health outcomes (e.g., obesity, cancer, heart disease), all-cause mortality and, less consistently, brain outcomes (e.g., brain health, schizophrenia and depression, lower cognitive ability, total brain volume, cortical thinning, and greater vascular lesions in old age (Oblak et al., 2021)). For a comprehensive review of the applications of DNAm clocks, related challenges and recommendations, see Bell et al. (2019).
In contrast to adult studies, the application and significance of epigenetic clocks during development is far less clear. One challenge is methodological: most clocks are trained primarily on adult samples using wide age ranges and show less accuracy in paediatric samples (Sanders et al., 2022). While certain clocks have been specifically developed in paediatric samples, these focus either on cord blood to estimate gestational age at birth, or peripheral blood/ buccal cells in childhood and adolescence, complicating efforts to characterize and integrate epigenetic age measures across these stages of development (Wang & Zhou, 2021

Integration on multiple levels
As the field matures, it is becoming increasingly clear that we need to move towards integrative research to better capture the complexity of DNAm and its relationship to mental health.

Gene-environment (G-E) integration
Current research mainly examines genetic or environmental effects on DNAm separately. This is problematic as environmental exposures may be genetically confounded, and in turn, genetic effects on DNAm may be environmentally modulated (i.e., potentially actionable).
Consideration of both G and E is thus essential to precisely identify influences on DNAm. This is supported by evidence that variation in DNAm is best explained by joint (i.e., additive and interactive) G-E effects, rather than G or E alone

Multi-omics and multi-phenotype integration
The extent to which statistically significant DNAm sites are also functionally relevant is often unclear. To address these challenges, studies have begun to integrate multiple layers of biological data.
Most commonly, this involves the use of transcriptomic data to test whether DNAm sites of interest associate with gene expression levels, either measured directly in peripheral tissues (with potentially limited relevance to the brain), or examined indirectly in the brain through the use of openly accessible resources. While many of these resources exist to help researchers functionally annotate and characterize findings (e.g., GTEx for gene expression (The GTEX Consortium, 2020); GoDMC for genetic effects on DNAm (Min et al., 2021); blood-brain comparison tools for cross-tissue concordance (Edgar et al., 2017)), these often lack sample context and celltype specific resolution. Another promising type of data integration, which is helping to clarify links between peripheral DNAm patterns and the brain in vivo, is the combination of DNAm and neuroimaging -although this only provides an indirect measure of peripheral-brain associations rather than a tool for directly inferring functional effects of DNAm on the brain. Besides lending biological insights, multi-omics integration may also help to achieve more powerful predictive models, as effect sizes from psychiatric EPIGENETICS APPLIED TO CHILD AND ADOLESCENT MENTAL HEALTH -5 of 14 EWASs are generally small, suggesting that DNAm patterns alone are likely to explain only limited variance in these phenotypes. Going forward, multivariate approaches will also be needed to account for the known co-occurrence of both (1) risk factors for psychopathology (e.g., parental psychopathology and childhood maltreatment); and (2) different domains of psychopathology themselves (e.g., internalizing and externalizing problems). In this respect, methods that are already established in the field of genetics, such as multi-trait GWAS (Wu et al., 2020) or genomic SEM (Grotzinger et al., 2019), could be extended for use with DNAm data.

DNAm in prediction
One of the key interests in DNAm lies in its potential as a predictor of health and disease risk. Indeed, the concept of 'methylation-based health profiling' has gained increased traction in recent years and is already demonstrating some success. For example, tools such as MethylDetectR (Hillary & Marioni, 2021) enable users to estimate a range of human traits (e.g., age, BMI), lifestyle characteristics (e.g., smoking and alcohol use) and biochemical variables (e.g., neurological and inflammatory proteins) based on peripheral DNAm alone. These DNAm-based estimates have a number of potential advantages: they allow users to obtain information on data that is not directly available in their dataset (e.g., proteomics; Gadd et al., 2022), they may offer more reliable information on certain variables (e.g., smoking) than more bias-prone traditional assessments (e.g., self-report; Bojesen et al., 2017), and importantly may perform better as predictors of disease risk. For example, DNAm-based estimates of BMI in adults have been found to predict risk for diabetes more strongly than BMI itself (Wahl et al., 2017). Although less mature, the application of predictive models to psychiatric epigenetics is beginning to bear fruit.

For example, adult studies have reported reproducible blood-based
DNAm 'signatures' of suicide risk (Clive et al., 2016), schizophrenia (Chen, Zang, et al., 2020), and future depression risk, showing greater explanatory power than models using genetic or clinical data alone (Clark et al., 2020). Furthermore, a large EWAS study in adults found that DNAm patterns in blood explain a substantial proportion of variance in general cognitive function (g), and that a methylationbased predictor derived from these results performs similarly to measured cognitive ability in predicting outcomes in independent samples, as well as generalizing across different age ranges and peripheral tissues Raffington et al., 2021).
Another study employing a sequencing-based approach recently identified a DNAm signature of trauma exposure in early adolescence, which predicted future psychiatric and health outcomes more strongly than self-reported trauma (van den Oord et al., 2022).
Importantly, the majority of these predictive DNAm sites were no longer associated with outcomes when measured again in adulthood, which provides further support for the temporally-dynamic nature of DNAm-mental health associations. It is noteworthy that area under the curve (AUC) estimates in these studies are not far off from wellestablished, clinically implemented predictive models, such as the Framingham Risk Score used to predict coronary heart disease (Tzoulaki et al., 2009). Whether similar prediction can be achieved for child and adolescent mental health outcomes is currently unclear. In the future, it would be interesting to see how well cord blood DNAm performs in predictive models of neurodevelopmental problems such as ADHD (e.g., compared to baseline models using more established risk factors), given the observed epigenetic timing effects described above.

DNAm and causal inference
The past years have seen major developments in the application of epidemiological methods to epigenetic research ( DNAm patterns in blood are more likely a consequence of than a cause for BMI, consistent with findings from longitudinal observational data (Reed et al., 2020). So far, very few epigenetic studies have applied MR to child and adolescent brain-based phenotypes.

NEXT FRONTIERS
What is on the horizon for epigenetic research on (child and adolescent) mental health? To push the boundaries of what is currently possible, we first need to reach a fuller understanding of epigenetic data itself (Figure 1.3).

Mapping the covariance structure of DNAm
One property of DNAm that we still know little about is its 'internal structure' (i.e., patterns of covariance). The field of population genetics has made great strides in defining linkage disequilibrium (LD) in genetic data, enabling key developments such as the imputation of genome-wide data from a limited set of measured SNPs, improved polygenic score analyses as well as genetic heritability and correlation estimations based on summary statistics (Allegrini et al., 2022;Bulik-Sullivan et al., 2015).

Normative modelling and the rise of 'chronoepigenetics'
The time-varying nature of DNAm is another property that must be  (Oh & Petronis, 2021). Although disruptions in these cycles have been implicated in ageing and disease risk, little is known about their association with mental health outcomes, pointing to an interesting avenue for future research.

New ways of addressing the 'tissue issue'
Clearly, tissue and cell-type heterogeneity in DNAm remains a major challenge for epigenetic research, and we will need to keep improving the ways we take this heterogeneity into account from study design (e.g., biomarker vs. mechanistic research) to data analysis and interpretation. Current studies typically rely on algorithms to estimate and adjust for cell-type proportions. These algorithms, however, are imperfect and account for a limited set of cell-types. In future, such panels could be expanded, enabling us to better capture developmental changes in cell-type composition (e.g., including multipotent cells found in cord blood at birth, but scarcely present in peripheral blood later in life), as well as to extend recent methods for performing cell-type specific EWAS from bulk tissue (Rahmani et al., 2019). It will also be important to evaluate whether prior knowledge of cross-tissue concordance may be used to improve signal in psychiatric epigenetic studies, for example, by selecting or weighing DNAm sites based on blood-brain correlations, or prioritizing regions that show high inter-individual variability in combination with low cross-tissue variability (Gunasekara et al., 2019).

Moving beyond (cg) DNAm
Finally, it is important to note that (cg) DNAm is only one of multiple types of epigenetic factors, which likely play a role in neurodevelopment and show potential as markers or mediators of psychiatric risk. These include other types of DNAm marks found to be enriched in the brain (e.g., hydroxymethylcytosine; Spiers et al., 2017) as well as histone modifications implicated in several neurodevelopmental and psychiatric conditions, including autism spectrum disorder (Tseng et al., 2022). Further, experimental studies increasingly point to circulating microRNAs as a potential mechanism underlying intergenerational transmission of phenotypes, including stress-related physiological and behavioural alterations (Lempradl, 2020), while population-based studies in adults (Mens et al., 2021) are beginning to reveal their potential as biomarkers of disease. Currently, these types of data are still rare in paediatric studies. In future, large-scale profiling of multiple epigenetic marks during development will be needed to characterize their independent and joint contribution to child and adolescent mental health.

TAKING STOCK: WHY IS THE APPLICATION OF EPIGENETICS TO CHILD AND ADOLESCENT MENTAL HEALTH LAGGING BEHIND?
As readers will have likely noticed, many of the new findings and developments highlighted in this review do not originate directly from the field of child and adolescent mental health, but rather from EPIGENETICS APPLIED TO CHILD AND ADOLESCENT MENTAL HEALTH -7 of 14 fields adjacent to it. This reflects a broader trend in the literature: despite tremendous growth in psychiatric (and more broadly healthrelated) epigenetics over the past 2 decades, the application of DNAm to child and adolescent mental health continues to account for only a fraction of this work (see Figure 2 (Yengo et al., 2022). Many child and adolescent psychiatric disorders are more heritable than height, but PRSs generally explain less than 5% of their variance Li & He, 2021). This can likely be explained by the much lower sample sizes (typically under 100,000 participants) and higher measurement errors of the discovery GWASs. The low variance explained casts doubt on the clinical utility of current psychiatric PRSs. However, some have argued that they may already be sufficient to identify extreme cases, aid in differential diagnosis and improve treatment response (Fullerton & Nurnberger, 2019).
Could adding information on DNAm bring these applications a step closer to clinical practice? In addition to common variants, DNAm may capture genetic effects that are not measured by SNP arrays, such as rare variants, as well as gene-environment correlations and interactions. Furthermore, DNAm may act as a 'biological record' of environmental exposures, leading to more reliable assessments (e.g., cg05575921 methylation vs. self-reported smoking) and greater predictive power compared to alternative measurement approaches (e.g., MRS of trauma exposure predicting psychiatric risk better than self-reported trauma). Unlike genetic data, the timevarying nature of DNAm also offers possibilities (and unique challenges) for tracking disease status and health over time, which could be particularly useful for early risk detection, patient stratification and response to treatment. In this respect, there is already much interest in utilizing epigenetic clocks in adulthood as markers of healthy ageing, which may be extended earlier in life to evaluate healthy development. Whereas these epigenetic clocks may be used as broad health markers; they may not fully capture disease-specific pathological mechanisms. Epigenetic predictors trained on specific outcomes could provide more nuanced information about particular health profiles to inform diagnoses and guide decision-making in more concrete clinical settings. Regarding diagnosis, it is notable that tools relying on 'epi-signatures' from peripheral blood have already been developed for a wide range of Mendelian neurodevelopmental diseases, demonstrating utility for brain-based disorders (Aref-Eshghi et al., 2020). Whether epigenetic-based tools could one day be used to improve diagnostic accuracy of child and adolescent psychiatric conditions-and whether the benefits of such tools would outweigh potential risks and ethical concerns-is an important topic for future research.
At the same time, MRS development will likely face the same challenge of insufficient discovery sample sizes as for PRSs in the past. Individual effect sizes of DNAm sites are not appreciably larger than SNP effects, but sample sizes of EWAS are many magnitudes lower compared to GWAS. Furthermore, unlike GWAS, we have yet to reach a consensus regarding the use of standardized pipelines for pre-processing of DNAm arrays, including which method to choose for data normalization and batch correction-an important step for maximising comparability between studies and reducing noise due to technical variation in EWAS meta-analyses.
As mentioned in previous sections, while it is possible to impute unmeasured SNPs from genotyping arrays, this is more challenging for DNAm arrays. As such, EWAS studies and potential downstream applications, including MRS development, are confined to measured probes, which represent only a small fraction of DNAm sites on the genome. Lastly, reverse causality may limit the application of cross-sectional data to develop methylation-based predictive tools, and confounding may provide misleading therapeutic targets.
In conclusion, the field is still in its infancy, and concrete translational applications remain a distant goal. Nevertheless, DNAm continues to hold unique potential as a biological system for biomarker discovery and mechanistic insights into the aetiology of child and adolescent psychiatric disorders. Looking to the future, increases in sample sizes-via collaborative science, harmonization efforts and better use of existing data-in combination with a focus on developmentally-sensitive, longitudinal study designs will be crucial to move the field forward and leverage this potential.
EPIGENETICS APPLIED TO CHILD AND ADOLESCENT MENTAL HEALTH -9 of 14