Abstract

   Attention-deficit and hyperactivity disorder (ADHD) is a common
   childhood disorder with a substantial genetic component. However, the
   extent to which epigenetic mechanisms play a role in the etiology of
   the disorder is unknown. We performed epigenome-wide association
   studies (EWAS) within the Pregnancy And Childhood Epigenetics (PACE)
   Consortium to identify DNA methylation sites associated with ADHD
   symptoms at two methylation assessment periods: birth and school age.
   We examined associations of both DNA methylation in cord blood with
   repeatedly assessed ADHD symptoms (age 4–15 years) in 2477 children
   from 5 cohorts and of DNA methylation at school age with concurrent
   ADHD symptoms (age 7–11 years) in 2374 children from 9 cohorts, with 3
   cohorts participating at both timepoints. CpGs identified with nominal
   significance (p < 0.05) in either of the EWAS were correlated between
   timepoints (ρ = 0.30), suggesting overlap in associations; however, top
   signals were very different. At birth, we identified nine CpGs that
   predicted later ADHD symptoms (p < 1 × 10^–7), including ERC2 and
   CREB5. Peripheral blood DNA methylation at one of these CpGs
   (cg01271805 in the promoter region of ERC2, which regulates
   neurotransmitter release) was previously associated with brain
   methylation. Another (cg25520701) lies within the gene body of CREB5,
   which previously was associated with neurite outgrowth and an ADHD
   diagnosis. In contrast, at school age, no CpGs were associated with
   ADHD with p < 1 × 10^−7. In conclusion, we found evidence in this study
   that DNA methylation at birth is associated with ADHD. Future studies
   are needed to confirm the utility of methylation variation as biomarker
   and its involvement in causal pathways.

   Subject terms: Psychiatric disorders, Genetics

Introduction

   Attention-deficit and hyperactivity disorder (ADHD) is a common
   neurodevelopmental disorder characterized by impulsivity, excessive
   activity, and attention problems. Symptoms often become apparent during
   school age with a world-wide prevalence of 5–7.5%^[90]1. Genetic
   heritability is estimated between 64 and 88%^[91]2,[92]3. Additionally,
   several environmental factors are suspected to impact ADHD, e.g.,
   prenatal maternal smoking or lead exposure^[93]4–[94]7. However, the
   genetics and environmental pathways contributing to ADHD risk remain
   unclear. Possibly, DNA methylation, an epigenetic mechanism regulating
   gene expression, may mediate genetic or environmental effects.

   Several studies have investigated DNA methylation in relation to ADHD
   diagnoses or symptoms using candidate approaches or epigenome-wide
   association studies (EWAS) in peripheral blood and saliva
   tissue^[95]8,[96]9. A leading hypothesis concerning the etiology of
   ADHD suggests that deficiencies in the dopamine system of the brain
   impact ADHD development^[97]4,[98]10. Consequently, candidate studies
   have focused on genes related to dopamine function. For instance, DNA
   methylation alterations in DRD4^[99]11–[100]13, DRD5^[101]12, and
   DAT1^[102]12,[103]14 genes have been associated with ADHD, though not
   consistently^[104]15. Beyond the candidate gene approach, three studies
   tested DNA methylation across the whole genome. One study performed an
   EWAS with saliva samples in school-aged children using a case–control
   design^[105]16. The study identified differentially methylated probes
   in VIPR2, a gene expressed in the caudate and previously associated
   with psychopathology. Another EWAS investigated cord and peripheral
   blood DNA methylation at birth and at 7 years of age^[106]17. At birth,
   13 probes located in SKI, ZNF544, ST3GAL3, and PEX2 were associated
   with ADHD trajectories from age 7 to 15 years, but the methylation
   status of these probes at age 7 was not associated with ADHD
   cross-sectionally. An EWAS in adults with ADHD failed to find any
   differentially methylated sites in peripheral blood^[107]18.

   Large multi-center epigenome-wide studies, which allow for increased
   power and generalizability, are lacking for childhood. Here we
   performed the first epigenome-wide prospective meta-analysis to
   identify DNA methylation sites associated with childhood ADHD symptoms
   in cohorts from the Pregnancy And Childhood Epigenetics (PACE)
   Consortium^[108]19. As DNA methylation changes over time^[109]20, so
   could potential associations with ADHD symptoms. On the one hand, one
   might expect that DNA methylation levels measured around the same time
   as ADHD symptoms would show the largest associations, as these might
   represent the immediate effects on symptoms or consequences of ADHD. On
   the other hand, causes of ADHD may be found early in childhood or even
   prenatally. Thus methylation levels at birth may be more relevant than
   later methylation profiles, as suggested by an earlier EWAS^[110]21.
   Since it is unclear when DNA methylation is most relevant for ADHD
   symptoms, we tested DNA methylation both at birth using cord blood and
   at school age (age 7–9 years) using DNA derived from peripheral whole
   blood. In the analyses of cord blood methylation, the aim was to
   explain ADHD symptoms between ages 4 and 15 years. Many participating
   cohorts assessed ADHD repeatedly, and we employed a repeated-measures
   design to increase precision. Furthermore, we utilized data in
   childhood to examine cross-sectional DNA methylation patterns
   associated with ADHD symptoms at school age.

Materials and methods

   This study comprises a birth methylation EWAS and a school-age
   methylation EWAS described successively below.

Birth methylation EWAS

Participants

   Five cohorts (Avon Longitudinal Study of Parents and Children
   (ALSPAC)^[111]22–[112]24, Generation R (GENR)^[113]25, INfancia y Medio
   Ambiente (INMA)^[114]26, Newborn Epigenetic Study
   (NEST)^[115]27,[116]28, and Prediction and prevention of preeclampsia
   and intrauterine growth restriction (PREDO)^[117]29) in the PACE
   consortium had information on DNA methylation in cord blood and ADHD
   symptoms. These cohorts have a combined sample size of 2477 (Table
   [118]1). Participants were mostly of European ancestry, except for
   NEST, an American cohort that also included participants of African
   ancestry. In NEST, separate EWAS were conducted for participants
   identifying as black or white to account for ancestry heterogeneity
   statistically in a random-effects meta-analysis. We also performed a
   sensitivity analyses with European ancestry children only. Parents gave
   informed consent for their children’s participation and local ethics
   committees approved the study protocols. See Supplementary Information
   [119]1 for full cohort descriptions.

Table 1.

   Cohort characteristics.
   Cohort Ancestry/ethnicity n Methylation age ADHD age Instrument (age)
   Standardized regression coefficients BACON estimates
   33% 50% 66% λ Inflation Bias
   Birth EWAS
   ALSPAC European 714 0 8, 11, 14, 15 DAWBA −0.21 0.25 0.89 1.60 1.10
   0.37
   GENR European 1191 0 6, 8,10 CBCL (6,10), Conners (8) −0.48 0.01 0.53
   1.51 1.20 0.05
   INMA European 325 0 7, 9 Conners (7), CBCL (9) −1.37 −0.40 0.43 0.80
   0.87 −0.19
   NEST Black 55 0 5 BASC −3.50 −0.03 3.63 1.16 1.10 0.00
   NEST White 56 0 5 BASC −2.54 −0.09 2.36 0.80 0.92 −0.01
   PREDO European 136 0 5 Conners −1.55 −0.25 1.20 1.45 0.95 0.21
   META – 2477 – – – −0.37 0.02 0.42 1.86 1.10 0.01
   School-age EWAS
   ALSPAC European 651 7 8 DAWBA −0.61 −0.10 0.54 1.09 1.00 −0.08
   GENR European 395 10 10 CBCL −0.93 −0.00 0.98 1.00 0.97 −0.01
   GLAKU European 215 12 12 CBCL −0.79 0.31 1.50 0.92 0.96 0.13
   HELIX European 1034 8 8 CBCL −0.26 0.47 1.40 1.11 0.98 0.28
   HELIX Pakistani 79 7 7 CBCL −1.66 1.86 5.48 0.98 0.96 0.26
   Meta – 2374 – – – −0.24 0.14 0.62 0.96 0.92 0.14
   [120]Open in a new tab

   n Number of participants, 33%, 50%, 66% quartiles of regression
   coefficient distribution, λ inflation of p values, Inflation inflation
   of p values due to suspected bias, Bias trend toward negative/positive
   distribution of regression coefficients due to suspected bias.

DNA methylation and quality control (QC)

   DNA methylation in cord blood was measured using the Illumina Infinium
   HumanMethylation450K BeadChip (Table [121]S1). Methylation levels
   outside of the lower quartile minus 3 × interquartile or upper quartile
   plus 3 × interquartile range were removed. Each cohort ran the EWAS
   separately according to a pre-specified harmonized analysis plan. The
   distribution of the regression estimates and p values were examined for
   each cohort and pooled results. Deviations from a normal distribution
   of regression estimates or a higher number of low p values than
   expected by chance may be signs of residual confounding or the result
   of a true poly-epigenetic signal. To help in interpretation of the
   results, we used the BACON method^[122]30. BACON analyzes the
   distribution of regression coefficients and estimates an empirical null
   distribution. Results can then be compared against the empirical null,
   which already includes biases, rather than the theoretical null. We
   excluded CpG probes, which were available in <4 cohorts; <1000
   participants; and allosomal probes, due to the complex interpretation
   of dosage compensation.

ADHD symptoms

   ADHD symptoms were measured when children were aged 4–15 years
   (depending on the cohort) with parent-rated instruments, specifically
   the Behavior Assessment System for Children^[123]31, Child Behavior
   Checklist (CBCL)^[124]32,[125]33, Conners^[126]34 and the Development
   and Well-Being Assessment (DAWBA)^[127]35 (Table [128]S2). If a cohort
   had measured ADHD symptoms repeatedly (three cohorts), we used a mixed
   model (see “Statistical analysis”). The repeated-measures design
   increased the precision of the ADHD severity estimate and sample size,
   since missing data in an assessment can be handled with maximum
   likelihood. Given the variety of instruments used within and across
   cohorts, all ADHD scores were z-score standardized to enable
   meta-analysis.

Statistical analysis

   Cohorts with repeated ADHD assessment were analyzed using linear mixed
   models, with z-scores of ADHD symptoms as the outcome and methylation
   (in betas, ranging from 0 (unmethylated) to 1 (methylated)) as the main
   predictor. Each CpG probe was analyzed separately and pooled p values
   were adjusted for multiple correction using Bonferroni adjustment. We
   used a random intercept on the participant and batch level, to account
   for clustering due to repeated measures and batch effects. The
   following potential confounders were included as fixed effects:
   maternal age, educational level, smoking status (yes vs no during
   pregnancy), gestational age, sex, and estimated white blood cell
   proportions (Bakulski reference estimated with the Houseman
   method)^[129]36. Mixed models were fitted using restricted maximum
   likelihood. We used R^[130]37 with the lme4^[131]38 package to estimate
   the models. Cohorts with a single ADHD assessment wave used a model
   without random effects or batch level only.

   Meta-analysis was performed using the Han and Eskin random-effects
   model^[132]39. This model does not assume that true effects are
   homogeneous between cohorts; however, it does assume that null effects
   are homogeneous. This modified version of the random effect model has
   comparable power to a fixed-effects analysis, while better accounting
   for study heterogeneity, such as ancestry differences, in simulation
   studies^[133]39,[134]40. Genome-wide significance was defined at the
   Bonferroni-adjustment threshold of p < 1 × 10^–7, suggestive
   significance at p < 1 × 10^–5, and nominal significance at p < 0.05.

Follow-up analyses

   We performed several lookups of genome-wide significant probes. We used
   the BECon database^[135]41 to check the correlation between peripheral
   and brain methylation levels in postmortem tissue. To test genetic
   influence, we interrogated the genome-wide significant probes in
   MeQTL^[136]42 and twin heritability databases^[137]43. We also
   attempted to replicate genome-wide significant probes reported in a
   previous EWAS from the ALSPAC study^[138]17. For replication, we reran
   the meta-analysis without the ALSPAC cohort. To quantify the variance
   explained by genome-wide significant probes, we predicted ADHD scores
   at age 8 years in Generation R by all meta-analytically genome-wide
   significant probes. We applied 10-fold cross-validation with 100
   repetitions to improve generalizability and reduce bias from Generation
   R, which was part of the discovery.

   We examined whether any CpG sites associated with ADHD symptoms are
   also associated with prenatal maternal stress. As prenatal maternal
   stress is associated with child psychopathology with mixed evidence of
   affecting DNA methylation^[139]44,[140]45, DNA methylation may be a
   mediator of adverse prenatal stress effects. We operationalized
   prenatal maternal stress as in Rijlersdaam et al.^[141]45 by using a
   factor score reflecting life, contextual, personal stress, and
   interpersonal stress. One modification to the previous definition is
   that maternal education was not used in the computation of the factor
   score, as it had been included as covariate in the EWAS model. We first
   tested whether prenatal stress was associated with ADHD symptoms with
   an analysis model equivalent to the EWAS model, but instead of DNA
   methylation, the prenatal risk score was the main predictor. We then
   tested associations between prenatal stress as predictor and DNA
   methylation as outcome. p Values were obtained with the lmerTest
   package^[142]46. We estimated the prenatal stress associations only in
   the Generation R cohort.

Pathway analysis

   Pathway enrichment analysis were performed with the
   missMethylpackage^[143]47 on suggestive probes (p < 1 × 10^–5). We used
   as references gene ontology, Kyoto Encyclopedia of Genes and Genomes,