Epigenome-wide DNA methylation in externalizing behaviours: A review and combined analysis

DNA methylation (DNAm) is one of the most frequently studied epigenetic mechanisms facilitating the interplay of genomic and environmental factors, which can contribute to externalizing behaviours and related psychiatric disorders. Previous epigenome-wide association studies (EWAS) for externalizing behaviours have been limited in sample size, and, therefore, candidate genes and biomarkers with robust evidence are still lacking. We 1) performed a systematic literature review of EWAS of attention-deficit/hyperactivity disorder (ADHD)- and aggression-related behaviours conducted in peripheral tissue and cord blood and 2) combined the most strongly associated DNAm sites observed in individual studies (p < 10 -3 ) to identify candidate genes and biological systems for ADHD and aggressive behaviours. We observed enrichment for neuronal processes and neuronal cell marker genes for ADHD. Astrocyte and granulocytes cell markers among genes annotated to DNAm sites were relevant for both ADHD and aggression-related behaviours. Only 1 % of the most significant epigenetic findings for ADHD/ADHD symptoms were likely to be directly explained by genetic factors involved in ADHD. Finally, we discuss how the field would greatly benefit from larger sample sizes and harmonization of assessment instruments.


Externalizing behaviour
Externalizing behaviours are defined as any behaviour that manifests in one's outward behaviour and reflects a person's negative interaction with the external world. Externalizing behaviour is linked to impairment of impulse control such as risk-taking, impulsive, and aggressive behaviour. These behaviours become maladaptive when they cause impairment in daily life functioning, as seen in externalizing disorders. Highly prevalent and comorbid (Dick et al., 2005) externalizing disorders include attention-deficit/hyperactivity disorder (ADHD; population prevalence of 5 % (Polanczyk and Rohde, 2007)), conduct disorder (CD; population prevalence up to 3 % (Fairchild et al., 2019)), and oppositional defiant disorder (ODD; population prevalence up to 10 % (Barr and Dick, 2019;Nock et al., 2007). 30-50 % of children with ADHD also are diagnosed with CD and/or ODD (Gnanavel et al., 2019), and one third of children with CD also are diagnosed with ADHD (L. Masi, 2015).
High comorbidity between these disorders could be explained by shared genetic influences. Twin studies have shown that externalizing behaviour is highly heritable (up to 80 %) (Faraone and Larsson, 2019;Hicks et al., 2004), and several genetic risk variants associated with ADHD (Demontis et al., 2019) and CD symptoms (Salvatore and Dick, 2018) have been identified. The externalizing phenotypes, in particular ADHD, antisocial behaviour, and aggression can be captured in a single, highly heritable externalizing factor (Kendler et al., 2003). Genetic correlation between ADHD and aggression-related behaviours is high (74 %), and first genes associated with the comorbid phenotypes have also recently been identified (Demontis et al., 2021). The onset of symptoms for all these disorders is during childhood (Fairchild et al., 2019;Faraone et al., 2015;Nock et al., 2007). In addition to genetic influences, studies showed the influence of environmental adversity on the development of externalizing behaviours (Hicks et al., 2013;Waller et al., 2017;Weeland et al., 2015). Environmental factors implicated include low birth weight , toxin exposureincluding tobacco (Schwenke et al., 2018;Sciberras et al., 2017;Williams and Ross, 2007), prenatal stress (Palladino et al., 2019), and early attachment deprivation (Roskam et al., 2014). The interplay between genetic and environmental factors contributing to externalizing behaviours can be mediated by epigenetic modifications. Epigenetic mechanisms that could reflect this interplay in externalizing behaviours may include DNA methylation, miRNA expression, and histone modifications (Babenko et al., 2015).

DNA methylation
One of the most frequently studied epigenetic mechanisms, facilitating the interplay of genomic and environmental factors in the regulation of gene expression and behaviour, is DNA methylation (DNAm) (Kubota et al., 2012). This epigenetic modification involves a methyl group being attached to the carbon-5 position of a cytosine (5mC) in CpG content (Bird, 1986), thereby altering the chemical properties of the DNA molecule and making it less accessible for transcription factors. Although there are exceptions, in general, hypermethylation of CpG-sites in the promoter region of a gene leads to gene repression, affecting the transcriptional profile of a cell and ultimately behaviour of the individual (Moore et al., 2013). DNAm patterns are cell type-specific (Smith and Meissner, 2013), and thus, study results can greatly vary depending on the origin of the tissue analysed (Lin et al., 2018). This cell type-specificity also implies that findings obtained in different tissues should be interpreted primarily in the context of this tissue. Associations between DNAm profiles and externalizing behavior can reflect causal effects on behavior, but could also represent consequences of the behavior analyzed. Alternatively, DNAm profiles could be biomarkers for causal factors, which independently of DNAm affect externalizing problems. Interpretation of the DNAm profiles has previously been critically appraised by Walton and colleagues (Walton et al., 2019). DNAm is, among others, influenced by ageing (Horvath and Raj, 2018), which means that results from childhood studies are not directly comparable to results from studies involving adults. Different epigenetic factors might play a role in behaviour at different developmental stages; for genetic factors, such differential involvement has also been reported e.g., for ADHD, where specific genetic influences act at different periods throughout the lifespan . Confounding factors, which should be adjusted for in DNAm studies, therefore include age; additionally, sex, ethnicity, smoking, and relevant environmental factors, as well as batch effects should be taken into account. Single nucleotide polymorphisms (SNPs) can also directly influence the extent of methylation at a specific CpG-site; such 'methylation quantitative trait loci (mQTLs)' provide an example how genetic variation can influence gene expression levels (Gaunt et al., 2016).
There are multiple ways to study DNAm. The field started out with candidate gene approaches. Those candidate gene studies of DNAm in externalizing behaviour focussed on genes involved in neurotransmitter systems targeted by stimulatory drugs, as well as genes involved in neurodevelopmental processes. Classical target genes include receptors and transporters involved in the dopaminergic and serotonergic neurotransmitter systems, more specifically, DAT1/SLC6A3, DRD4, DRD5, HTR1B, and 5-HTT/SLC6A2. However, the results from this literature are inconsistent. For example, four out of five studies reported hypermethylation of the DRD4 receptor, but all indicating different significant DNAm sites (i.e., CpG-sites). This might be related to large variation between studies in sample size (median N = 182, range 100 -426), study population (Hamza et al., 2019), and coverage of sites. A previous review of epigenetic candidate genes for ADHD concluded that a deeper understanding of the epigenetic mechanisms in ADHD is needed (Hamza et al., 2019). As with genetic association studies, epigenetic candidate genes have not (often) been identified as genes with significant effects in subsequent hypothesis-generating approaches. Hypothesis-generating ways of studying epigenetic mechanisms of externalizing behaviour, such as the epigenome-wide DNAm association studies (EWAS), are therefore important to bring progress to this field. The currently most frequently used methods for EWAS involve epigenome-wide arrays and bisulfite sequencing, where unmethylated cytosines are converted into uracil (Frommer et al., 1992). The arrays used for epigenome-wide methylation studies (EWAS) cover a selection of CpG-sites distributed across promoter regions, 5' untranslated regions (UTRs), first exons, gene body, and 3' UTRs of 99 % of the RefSeq annotated genes (Bibikova et al., 2011;Mansell et al., 2019). The two most frequently used arrays (450K and EPIC) differ from each other in terms of number (450,000 CpG-sites versus 850,000 CpG-sites) and distribution of coverage (mostly promoter regions on the 450K array versus broader coverage of regulatory elements on the EPIC array); 90 % of the content of the 450K array is also covered on the EPIC array (Bibikova et al., 2011;Mansell et al., 2019). Recently, an increasing body of literature on EWAS in externalizing behaviours has come available, but attempts to integrate the published data are welcome to create a more comprehensive overview of what has been discovered in the field.

Review outline
This review provides a systematic overview of existing EWAS in externalizing behaviours. Previous reviews, published in 2018 and 2019, only focussed on the two EWAS available for ADHD at that time (Dall'Aglio et al., 2018;Hamza et al., 2019). Since then, eleven EWAS have been published, and we thus aimed at carrying out a systematic and comprehensive review of EWAS and confounding factors in externalizing behaviours, also including EWAS in animals. Importantly, we additionally performed a combined analysis of the meta-statistics of EWAS studies for ADHD and aggression-related behaviours to identify overlapping CpG-sites and genes, gene ontology (GO) term and cell type enrichments, and the contribution of genetic factors to DNAm in externalizing behaviour. Finally, we performed a qualitative assessment of the contribution of our approach to the field and extracted recommendations from this, including recommendations on sample size, confounders, coverage of arrays, cell type-and tissue-specificity, and the current state of EWAS in animals.

Literature search
We performed an extensive literature search using Pubmed® with predefined search terms for human studies (Fig. 1) following the PRISMA guidelines (Moher et al., 2009) on 27/03/2020. Thereafter, all published studies were added manually until January 28, 2022. Search strategies were divided into ADHD and related traits, aggressive, antisocial and risk-taking behaviours, and aggression-related disorders. The exact search term and the exclusion criteria can be found in the Supplementary Text.ll All studies were given a score, based on three questions from the GRADE criteria (Guyatt et al., 2008): 1) Did the current study correct for the most relevant confounders? 2) Is the sample size sufficient to have a precise effect size? (N = 1000, based on power calculations (Mansell et al., 2019), see also our Discussion section about this point). 3) Is there a replication attempt? The maximum score a single paper could reach was 3.

Analysis of combined studies
To identify overlap in EWAS signals, an analysis combining the top hits of individual EWAS was performed. Out of the 29 studies included, 26 studies performed an EWAS in human participants. Ten studies reporting on the largest unique cohorts were used for the combined analysis, six for ADHD-related traits and four for aggression-related phenotypes. A combined analysis on externalizing behaviours (i.e., combining ADHD-and aggression-related behaviour) was not possible given the large overlap between study cohorts. Studies included in the combined analysis can be found in Table 1. CpG-sites with a p < 1 * 10 -3 for association with externalizing behaviour in the original research articles were extracted from the relevant studies; if needed the authors were contacted. We chose this rather lenient p-value threshold, since effect sizes of EWAS findings for behaviour are expected to be comparable to those seen in genome-wide association studies (GWAS), as is e. g., observed by the variance explained of CpG sites and SNPs in an EWAS and GWAS for cognitive functions (Davies et al., 2016;McCartney et al., 2022). At the chosen p-value threshold, we were able to detect CpG-sites with an R 2 = 0.08 with 80 % power (compared to an R2 =0.19 for the epigenome-wide significance threshold (p = 6 *10 -8 )). In studies of limited sample size, like the median sample size of ~200 in the studies included here, effects explaining such amounts of variance would not reach epigenome-wide significance. Moreover, there is often convergence on biological pathways from sites with individually weaker effects in multifactorial traits (Shohat et al., 2021). It is this multifactorial aspect of disorders and the convergence of biological pathways what we wanted to detect in our analysis by including CpG-sites with p-values below the epigenome-wide significance threshold. For comparison, however, we also report on the findings from p-value thresholds of p = 1 * 10 -4 and p = 1 * 10 -5 . For all overlapping probes and genes, we tested whether the overlap observed was greater than expected by chance. For this, we calculated the null probability to detect overlap and compared the actual overlap with a 1-sample proportions test with continuity corrections. We performed Gene Ontology (GO) term, cell type marker, and mQTL enrichments on the CpG-site lists. Details on these enrichments can be found in the Supplementary Text.

Results
We identified 14 and 12 studies on ADHD-and aggression-related behaviour in the existing scientific literature, respectively. These studies are summarized in Table 1, and descriptions of the individual studies can be found in the Supplementary Text.
Of the ADHD studies, seven studies investigated clinically diagnosed cases and controls, and seven studies investigated symptom scores or trajectories. Three studies were performed in adults (>21 years), and the other studies focussed on children and adolescents, of which six also investigated DNAm profiles in cord blood, leading to an investigated age range of 3 -44 years and a sample size range of 24 -4,689. Only three studies used the EPIC array, all the others used the 450K array with less coverage. For EWAS in externalizing behaviour, there is a risk of bias by confounding factors, including age, ethnicity, smoking, and cell type composition, of which the latter one can also be captured by latent variables. Other possibly relevant covariates are medication status of patients and socio-economic status. In Table 1, the confounders corrected for are presented for each of the studies. According to the recommended epigenome-wide significance threshold for studies using these arrays (Saffari et al., 2018), only four out of the fourteen studies reported significant single CpG-sites, but none of the findings overlapped between the studies (de Vocht et al., 2018;Neumann et al., 2020;van Dongen et al., 2019;Walton et al., 2017).
Of the aggression-related studies, six studies investigated aggressionrelated symptom scores in the general population, and six studies used a case-control design to investigate pathological forms of aggression. Three studies made use of the EPIC array, two studies made use of the 450K array, two studies made use of methylated DNA immunoprecipitation (meDIP) (Mohn et al., 2009), and one study made use of Hpall-MethSeq (Zawada et al., 2016). Age of participants ranged from 4 to 70 years, with a sample size range of 20 -14,190. Confounders corrected for can be found in Table 1. The number of significant findings reported in aggression-related EWAS, according to the recommended thresholds, was higher than for ADHD. Seven studies reported epigenome-wide significant single CpG-sites    Preferred Reporting Items for Systematic reviews and Meta-Analysis (PRISMA) flowchart of the literature search on epigenome-wide association studies of externalizing behaviour. After removing duplicates, the titles and abstracts of 244 articles were screened. The following studies were excluded: studies focussing on aggressive forms of nonexternalizing diseases, studies that did not use epigenome-wide approaches, and studies validating new methods but not reporting individual findings for externalizing behaviours. This resulted in the inclusion of 26 studies in the qualitative synthesis. Ten studies were used in the quantitative combined analysis.    (Chiocchetti et al., 2022) Abbreviations: 450K=Infinium HumanMethylation450 bead array; ADHD=attention-deficit/hyperactivity disorder; ADHD-RS=attention-deficit/hyperactivity disorder-rating scale; ASEBA=Achenbach system of empirically based assessment; ASR=adult self-report; BASC=behavioural assessment system for children; BMI=body mass index; CAARS=Conner's adults ADHD rating scale; CAADID=Conner's adult ADHD diagnostic interview for DSM-IV; CBCL=child behaviour checklist; CP=conduct problems; CU traits=callous-unemotional traits; DAWBA=development and well-being assessment; DIVA=diagnostic interview for ADHD in adults; DSM-IV=diagnostic and statistical manual of mental disorders -IV; DSM-V=diagnostic and statistical manual of mental disorders -V; EAS=emotionality, activity, and sociability; EPIC= Illumina Infinium EPIC Human Methylation microarray; FBB-ADHD=Fremdbeurteilungsbogen (parent rating scale for ADHD); FDR=false discovery rate; G1/G2 =generation 1/generation 2; GHQ-28 =general health questionnaire-28; GWAS=genome-wide Provencal et al., 2014), although it has to be noted that some of these studies show inflated test statistics. No overlap between study-specific significant findings was observed. This shows clearly that the replicability of DNAm findings is low for both ADHD and aggressive behaviour, so far. Furthermore, out of all EWAS on the externalizing behaviours, only one study identified a potential role for DNA methylation in one of the previously defined candidate genes, DRD4 (Cecil et al., 2018b). Only ten out of the 23 studies attempted a replication in the primary paper (Table 1), showing that performing replication studies is not yet common practice in DNAm studies; such studies would be important to increase robustness of findings. The number of EWAS in animal models identified in the literature was very limited as only three studies were found. In zebrafish, two research groups performed an EWAS for chemical-induced hyperactivity, using either whole genome bisulfite sequencing (WGBS) in whole fish (Olsvik et al., 2019) or meDIP in sperm cells (Carvan et al., 2017). Among the significant findings, both studies identified pcdh1g3 and nrxn3b to be associated with hyperactivity. Pcdh1g3 belongs to the family of protocadherins, and this gene family has homologues that were also identified in human EWAS for ADHD (Mooney et al., 2020;Peter et al., 2016;van Dongen et al., 2019). WGBS in the whole brain of bees identified TRP channels and genes involved in stress hormone signalling (calcitonin receptor and diuretic hormone (ADH)) to be differentially methylated two hours after the resident-intruder test in the most aggressive bees, compared to baseline (Herb et al., 2018). Given the methodological differences between the different animal studies and between those and the human studies, we could not proceed towards further integration of the animal work with the human EWAS in the form of these combined analyses.

Overlapping results across different EWAS of ADHD
To study results of differentially methylated CpG-sites overlapping across individual EWAS in more depth, we extracted from the six studies described in Table 1 a total of 2811, 292, and 36 CpG-sites with an association p < 1 * 10 -3 , p < 1 * 10 -4 , and p < 1 * 10 -5 , respectively (Table S1). Two of the CpG-sites (cg02136725, cg05475109) from the p < 1 * 10 -3 list was found in two out of the six studies showing the same direction of effect; this overlap was larger than expected by chance (p < 0.005) ( Table S2). The overlapping CpG-sites could be annotated to the genes ARID4A (AT-Rich Interaction Domain 4A) and FLJ31306 (Proteasome 20 S Subunit Alpha 3 Antisense RNA 1), and an intergenic region on chromosome 8, respectively. Probe annotation to genes prior to overlap analysis showed 95 genes to overlap between at least two studies, which is more than expected by chance, even when taking gene size bias into account (p < 0.0032), except for the largest gene PTPRN2 (p = 0.0628) (Table S3). Among those 95, GNG7 (G Protein Subunit Gamma 7, encoding a transmembrane signalling protein), TNXB (Tenascin XB), ADAMTS2 (ADAM Metallopeptidase With Thrombospondin Type 1 Motif 2), ATP11A (ATPase Phospholipid Transporting 11A), BRD2 (Bromodomain Containing 2), CACNA1H (Calcium Voltage-Gated Channel Subunit Alpha1H), and SDK2 (Sidekick Cell Adhesion Molecule 2) were seen in three out of six studies. At the more stringent pvalue threshold of p < 1 * 10 -4 , we identified overlap between two studies with the same direction of effect for ACOXL (Acyl-CoA Oxidase-Like Protein).
investigated whether the genes from the combined CpG-list were enriched for brain and blood cell markers. We identified an enrichment for neurons (BH-adjusted p = 0.023) and astrocytes (BH-adjusted p = 0.021), and granulocytes (BH-adjusted p = 0.0095) ( Fig. 2A), independent of DNA methylation sites specific to granulocyte cell type composition (Salas et al., 2018). To exclude potential bias due to differential probe distribution, we compared the distribution of probes across marker genes for different blood cell types and brain cell types (Fig. S1). Granulocyte marker genes did not have a different probe distribution compared to the other blood cell marker genes, suggesting that the result is not driven by such bias (Fig. S1A). Differential probe distribution among brain cell marker genes did not seem to drive the findings for astrocyte and neuron enrichment either, as these marker genes did not show a different distribution of probes compared to marker genes for other brain cell types (Fig. S1B). No enrichment was found for the CpG-lists derived using more stringent p-value thresholds.
To investigate whether the differential methylation observed in the combined EWAS results reflected effects directly related to genetic risk variants for ADHD (Demontis et al., 2019), we determined whether SNPs associated with ADHD in a GWAS acted as (blood or brain) mQTLs for the top CpG-sites (Table S5). Using a threshold of p < 10 -3 for association with ADHD, we included 39,581 single nucleotide polymorphisms (SNPs) from the most recent ADHD GWAS (Demontis et al., 2019) in our analysis. Of the 2811 CpG-sites in the EWAS combined analysis for this same p-value threshold, 496 (18 %) formed 41,910 mQTL pairs for blood. Among those mQTL pairs, 255 SNPs associated with ADHD at p < 1 * 10 -3 in the ADHD GWAS influenced 9 unique CpG-sites. One of these CpG-sites (cg21582582) was also found to be an mQTL in the developing brain for SNPs in the gene DCUN1D1 (Defective In Cullin Neddylation 1 Domain Containing 1). This gene has not been linked to neurodevelopmental, behaviour, or psychiatric disorders to date. Using different p-value thresholds for both the CpG-list and the GWAS list yielded similar results, where never more than 1 % of CpG sites was known to be influenced by a SNP (Table S5). This might imply that less than 1 % of the EWAS findings are directly influenced by SNP risk factors for ADHD, although conclusive evidence should be gathered through genome-wide approaches, taking cell type-specificity into account.

Overlapping results across different EWAS of aggression-related behaviour
The combined analysis of the differentially methylated top hits from four EWAS of aggression-related behaviour was performed on 2104, 183, and 21 CpG-sites with an association p < 1 * 10 -3 , p < 1 * 10 -4 , and p < 1 * 10 -5 , respectively (Table S6). Exact CpG-locations were not available for one study (Chiocchetti et al., 2022). Two CpG-sites overlapped between two out of the three studies (cg16226644 and cg14422759), and these were annotated to B3GALT4 (Beta-1,3-Galactosyltransferase 4, a membrane-bound glycoprotein) and RTN4R (Reticulon 4 Receptor, an oligodendrocyte myelin glycoprotein). However, the latter CpG-site showed different directions of effect in the two studies (Table S7). After annotating CpG-sites to genes, 62 genes showed overlap in at least two studies, and CAPN2 (Calpain 2, a calcium-activated neutral protease) and HDAC4 (Histone Deacetylase 4, involved in gene regulation via MEF2C and MEF2D transcription factors) were found in three out of four studies (Table S8). This overlap was larger than expected by chance (p < 0.0016). However, among CAPN2 and HDAC4, only the CpG-sites annotated to HDAC4 also showed a consistent direction of effect in all three studies. When using the more stringent p-value threshold of p < 1 * 10 -4 , CSNK1E (Casein kinase I isoform epsilon) was identified in two studies. GO-term enrichment analysis for these overlapping genes showed no enrichment after FDR-correction. The unadjusted findings were related to GO-terms indicating involvement in, among others, regulation of neuroinflammatory response, neuron migration glutamate receptor signalling, and histone H3-K9 methylation (Table S9). Enrichment for cell type markers showed enrichment for astrocytes (BH-adjusted p = 0.0066) and granulocytes (BH-adjusted p = 0.013; Fig. 2B). No enrichment was found for the CpG-lists derived using more stringent p-value thresholds.

Discussion
In this review we set out to summarize the current literature on EWAS in externalizing behaviours. Moreover, we analysed the DNAm sites most strongly associated with ADHD-and aggression-related behaviours, which enabled us to identify and prioritize valid candidate genes and biological mechanisms that were not reported in the individual EWAS, possibly due to lack of power. Interestingly, we were able to identify candidate epigenetic patterns divergent for ADHD-and Fig. 2. Analysis of the combined most strongly associated DNAm sites for A) ADHD and B) aggression. Brain and blood cell type marker enrichment. For each cell type, the -log10(adjusted) p-value is plotted. Numbers in brackets reflect the number of cell markers that were present in the gene list annotated to the CpG-sites out of all the cell markers.

Valid candidate genes and biological systems for externalizing behaviours based on combined analysis
In the existing literature, 26 EWAS on externalizing behaviour (14 on ADHD (symptoms) and 12 on aggression-related behaviour) were found to be present. Based on our combined analysis of EWAS, we identified ten candidate genes for ADHD: GNG7, ADAMTS2, TNXB, ATP11A, BRD2, CACNA1H, SDK2, ACOXL, FLJ30306, and ARID4A. Multiple levels of evidence support these genes as candidate genes for ADHD (Supplementary Text). Interestingly, these candidate genes show different expression patterns throughout the bodyincluding peripheral tissues, where DNAm was assessed, and the brain (Table S10). GO-term enrichment analyses revealed that genes across EWAS of ADHD in material isolated from peripheral blood and saliva pointed to processes related to neuronal functions, and we hypothesize neurons, astrocytes, and granulocytes to be likely involved in the ADHD phenotype. Some studies have reported evidence for higher granulocyte levels in individuals with ADHD compared to healthy controls (Avcil, 2018;Wang et al., 2018). On top of that, comorbidity exists between ADHD and allergies (Instanes et al., 2018); granulocytes play an important role in the aetiology of allergies. Interestingly, they are also increasingly recognized to mediate the immune response in the brain (Pflieger et al., 2018). Altogether, these genes and biological pathways share overlapping functions of dopamine signalling, inflammation processes and nervous system development, which are putative underlying mechanisms of ADHD-related behaviours. Lastly, we found indications that less than 1 % of the top findings of the combined analysis was directly related to genetic effects of SNPs associated with ADHD based on the most recent ADHD GWAS (Demontis et al., 2019), suggesting that non-genomic effects of DNAm are also relevant for ADHD and complementary information can be obtained from genetic and DNAm studies of ADHD. It should be noted, however, that more precise information should be acquired through genome-wide, cell type-specific approaches. The percentage of CpGs related to genetic variance (18 % in our study) through mQTLs is comparable with one reported for another phenotype, i.e., 19 % for educational attainment (van Dongen et al., 2018).
Promising, valid candidate genes for aggression-related behaviour identified through combined analysis are B3GALT4 and RTN4R, both having overlapping CpG-sites in two out of three EWAS in our combined analysis. Beta-1,3-Galactosyltransferase 4 (B3GALT4) is a membranebound glycoprotein, that plays essential roles in cell signalling and development of the brain (Schnaar, 2016). Reticulon 4 Receptor (RTN4R) is a receptor for myelin-associated glycoprotein. Furthermore, CAPN2 and HDAC4 are candidate genes for aggression-related behaviour, given that these genes were identified in each of the three aggression EWAS source studies. Supporting evidence for these candidate genes can be found in the Supplementary Text. Lastly, we found an enrichment of granulocyte and astrocyte cell markers among the genes annotated to aggression-related differentially methylated CpG-sites. All putative candidate genes for aggression are related to brain development, but also share associations with ADHD-related behaviour in the literature (Supplementary Text). The overlap for granulocyte enrichment in both ADHD and aggression-related behaviour is interesting in the apparent absence of overlapping genes between the two in our analyses; it suggests that there might be convergence of biological pathways through different epigenetic mechanisms. Moreover, the enrichment results indicate a potential role for astrocytes in aggression-related behaviour, as also suggested by rodent studies in which impaired astrocytic function, through interference with a variety of mechanisms (e.g., decreased extracellular glutamate clearance by astrocytes), was linked to more aggressive behaviours (Kalinine et al., 2014;Karpati et al., 2019;Madadi et al., 2019). This seems promising, because results obtained from surrogate tissues are in line with the findings from more directly relevant tissue, suggesting that these biomarkers may be linked to potential biological mechanisms.
In summary, although ADHD and aggression both belong to the externalizing spectrum, different biological enrichments and cellular substrates implicated through EWAS were observed. This is an intriguing difference and important addition compared to GWAS studies (Demontis et al., 2021), which have shown substantial overlap between the externalizing behaviours.

Main challenges and opportunities for the reviewed EWAS of externalizing behaviour
After reviewing and analysing the existing literature, we identified main challenges and opportunities for EWAS in externalizing behaviour. In this following section, we describe these issues and make recommendations for future studies. Most EWAS included in this review made use of commercially available arrays. Based on a permutation approach, a threshold of 9.42 * 10 -8 has been proposed to call significance in this type of EWAS (Mansell et al., 2019). However, due to the small sample sizes included in most studies, many of them reported 'suggestively significant' findings with a p-value threshold of 1 * 10 -5 . Indeed, the median sample size for EWAS on externalizing behaviour is 198 participants. The probability to detect sites with 2 % difference in DNAm is only 60 % in studies with N = 500 participants and further decreases to 12 % with a sample size of N = 100, indicating that we face a sample size-driven power issue (Mansell et al., 2019). This is a clear limitation of previous EWAS to be overcome in future studies involving larger sample sizes. Meta-analyses of these existing datasets could possibly provide us with more in-depth information on the epigenetic patterns in externalizing behaviour. Meta-EWAS would be particularly useful if cohorts were to use harmonized quality control and analysis pipelines to minimize heterogeneity. Moreover, the research field would highly benefit from replication studies in independent study cohorts (Kraft et al., 2009).
Almost all studies assessed externalizing behaviours using different instruments and designs, which may also have contributed to the limited overlap in findings. Harmonized use of phenotyping instruments and clearer distinctions and subtyping of the assessed behaviour may facilitate identification of associated differential methylation in future studies. For example, for aggression-related behaviours van Dongen and colleagues used self-reported aggression measures likely reflecting reactive aggression (van Dongen et al., 2019). Several others used proactive aggression measures (Cecil et al., 2018b;Guillemin et al., 2014;Provencal et al., 2014). Additionally, Cecil and co-workers observed that physical aggression in their cohort was associated with ethnicity (Cecil et al., 2018b), which might confound the association results of the measured DNAm profiles.
Confounding factors, such as age, sex, and smoking, should be carefully addressed in EWAS for externalizing behaviour. The studies discussed in this review do differ in terms of the set of confounders corrected for. Although DNAm is thought to be a relatively stable epigenetic modification (Dolinoy et al., 2007), methylation does change e.g., with age. Therefore, it is important to study DNAm in both children and adults, and to realize that combining childhood and adult studies can provide validation/generalization, but not strict replication. Moreover, studying cord and peripheral blood in the context of externalizing behaviours can be highly valuable, as findings in cord and peripheral blood require different interpretations. DNAm associations in cross-sectional studies of peripheral blood may reflect both causes and/or consequences of externalizing behaviours or its correlates. DNAm associations in cord blood, on the other hand, might reflect a sensitivity to develop externalizing behaviour over time, and could potentially be used as predictors (Cecil and Nigg, 2022). Also, an unequal sex distribution for ADHD exists, especially in children (Martin et al., 2018), and an interesting question is whether this bias might be (partially) driven by DNAm. Mooney and co-workers indeed found that DNAm was influenced by an interaction of sex with ADHD diagnosis (Mooney et al., 2020). Most published EWAS have included both sexes to be able to generalize the results over the complete population, but only few studies have taken possible interaction effects into account. Considering both sexes, it would be of great value to also investigate potential sex differences; for ADHD, a previous study concluded that molecular genetic analyses of autosomal common variants largely do not explain the sex bias in ADHD prevalence (Martin et al., 2018). One may want to consider epigenetics as a source of such sex differences. In line with this, sex chromosomes are currently often removed from EWAS because of methodological difficulties, but possibly include important additional information (Bonvicini et al., 2016;Li et al., 2020b). Lastly, smoking is an important confounding factor, but not all of the reviewed studies took this completely into account. Especially the studies of early childhood did not include smoking as a covariate, although it is known that maternal smoking can influence DNAm in the child (Joubert et al., 2012). Therefore, a measure of tobacco exposure should optimally always be incorporated in the statistical EWAS model for sensitivity analysis to find out whether the phenotype is mediated by prenatal tobacco exposure (Zeilinger et al., 2013).
Moreover, the different studies used different arrays, and these did not always cover interesting genes completely. Although the coverage of DNAm arrays has increased over the years (from 27,000 sites to 850,000 sites), by far not all 28 million CpG-sites in the genome are covered. In genetic studies, missing data can be imputed, based on the linkage disequilibrium (LD) between genetic variants. The idea to also impute DNAm values to reach broader coverage than the current arrays is very tempting, and a recent study showed so-called methylation haplotype blocks exists, 147,888 of them were observed (Guo et al., 2017). Some efforts have already been made to impute CpG sites found on the EPIC array from 450K data (Li et al., 2020a), and also to impute low-quality methylation estimates within whole-genome bisulfite sequencing (Zou et al., 2018), and these efforts could possibly reduce the costs of EWAS and provide us with broader coverage of DNAm across the genome in future studies.
It should be considered that all the human EWAS of externalizing behaviours reviewed in this study were performed in material derived from either blood or saliva, not allowing for direct inference of brain DNAm levels of these genes. Attempts have been made to compare DNAm levels across different tissue sources (Braun et al., 2019;Hannon et al., 2015), which shows that DNAm profiles are highly cell type-specific (Murata et al., 2019;Titus et al., 2017). This cell type specificity raises another problem. The investigated tissues often consist of a mixture of cells, and therefore meaningful epigenetic signals can get 'diluted' if not present in all cell types and even may get cancelled out across cell types. Investigation of cell type-specific DNAm can be a solution, either by isolating the cell types of interest or by using statistical methods like epigenomic deconvolution (Chan et al., 2020).
The reviewed EWAS do not tell us whether observed associated DNAm profiles are causal for externalizing behaviours or a consequence of a lifestyle associated with externalizing behaviour (Walton et al., 2019). In the current review, only two studies focussed solely on environmental factors: prenatal paracetamol exposure (Gervin et al., 2017), and early childhood malnutrition (Peter et al., 2016) (Supplementary Text). One other way to investigate whether environmental factors may have a causal relationship with externalizing behaviours via DNAm, is via two-step Mendelian randomization (MR). In the field of smoking and cancer, this method has already advanced knowledge greatly (Battram et al., 2019). First efforts of applying MR in the context of neurodevelopmental disorders and mental health are also being made, and it was suggested that, rather than a causal factor, DNAm in the BDNF gene is a consequence of childhood seizures .

Epigenome-wide studies in animals
In addition to statistical approaches, studies in animals could also help understanding whether DNAm is in the causal path of an observed behaviour. EWAS in animals enable identification of gene-environment interplay that leads to altered DNAm levels in multiple tissues and allow links to behavioural changes. A wide range of animal models can be used to study and manipulate externalizing behaviour, such as aggression (Herb et al., 2018;Walker et al., 2017) and hyperactivity (Olsvik et al., 2019). An ideal experimental set-up for EWAS would be to expose either a cohort from an outbred strain or animals from different inbred strains to relevant (e.g., early life) stressors. In this way, it is possible to perform controlled studies, in which researchers have access to relevant brain tissue in different phases of life, and the opportunity to interfere with specific genes and biological pathways.
To date, EWAS for externalizing behaviour in animals has been extremely limited (Carvan et al., 2017;Herb et al., 2018;Olsvik et al., 2019), and the full potential of animal model organisms has not been exploited yet. The two studies performed in zebrafish (Carvan et al., 2017;Olsvik et al., 2019) showed overlapping findings located in pcdh1g3 and nrxn3b. The first gene has homologues found in human EWAS of ADHD and is highly expressed in the developing nervous system (Mooney et al., 2020;Peter et al., 2016;van Dongen et al., 2019). The latter gene has human homologues that are involved in neuropsychiatric disorders, including ADHD (Kasem et al., 2018). Lastly, using a resident-intruder test in bees to introduce aggression, WGBS in whole brains of bees showed TRP channels and genes involved in stress hormone signalling, i.e., the calcitonin receptor and diuretic hormone (ADH) to be differentially methylated two hours after the test in the most aggressive bees, compared to baseline (Herb et al., 2018). None of the genes prioritized in the animal studies were top hits in one of the human EWAS on aggression-related behaviour; however, the ADH receptor has been identified in one of the human aggression EWAS , indicating that the ADH system might be involved in aggressive behaviour in both insects and humans.
One example of how animal studies can be used to investigate the causal role of specific epigenetic changes in the regulation of behaviour relates to the cognate receptor NR3C1 of glucocorticoids (CORT), the main hypothalamic-pituitary-adrenocortical stress hormone. It is particularly relevant in this context given that stress is an important risk factor for externalizing behaviours (Haller et al., 2014). Indeed, multiple human studies, including one EWAS previously described , have shown that NR3C1 methylation correlates with externalizing behaviour (Cicchetti and Handley, 2017;Dadds et al., 2015;Heinrich et al., 2015;Liu et al., 2021;Radtke et al., 2015). In chicken, CORT exposure during embryogenesis was shown to increase DNAm of NR3C1 in the brain in the same direction as observed in human blood and, importantly, chickens with elevated NR3C1 methylation exhibited increased levels of aggressive behaviour (Ahmed et al., 2014). In addition, rats with elevated CORT responsiveness upon a stressor were shown to display enhanced levels of aggression (Walker et al., 2017). Importantly, the causal implication of these findings has been supported by studies in multiple model organisms showing that NR3C1 activation leads to increased aggression, while blockade of the receptor diminishes aggression (Dunlap et al., 2011;Papilloud et al., 2019;Schjolden et al., 2009). In the future, the development of tools capable of targeting specific DNAm sites for specific genes will be instrumental to underscoring the precise role of gene methylation changes in specific brain regions and cell types for the modulation of externalizing behaviours.

Conclusions
Results from studies described in this review show that EWAS of externalizing behaviours have the potential to provide more insight into the underlying causes and the interplay between genome and environment in externalizing behaviours. Our combined analyses, providing superior study power, identified several candidate genes (e.g., GNG7, TNXB, HDAC4, B3GALT4, and RTN4R), and cellular substrates for the externalizing behaviours. Importantly, we detected different biological enrichments and cellular substrates for ADHD and aggression-related behaviours. We also found that genetic and epigenetic studies provide highly complementary information. However, the field of EWAS in externalizing behaviours is clearly still in its early stages, and especially the inconsistency of findings, due to relatively small sample sizes, phenotypic heterogeneity, and methodological differences in study design and analysis strategy limit replicability of findings. Given the findings revealed in these combined analyses, we argue that pursuing epigenetic approaches in externalizing behaviours are needed. To do this, the field will greatly benefit from larger sample sizes, harmonization of assessment instruments, and meta-analysis efforts. We emphasize that more advanced methods to extract information from existing data and the integration across complementary approaches including the use of animal models will enhance our understanding of biological mechanisms involved in the interplay between genome and environment in externalizing behaviours.

Competing Interests
BF received educational speaking fees from Medice. All other authors have no potential conflict of interest.